The Design and Implementation of the FreeBSD Operating System, Second Edition
Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)


[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]

FreeBSD/Linux Kernel Cross Reference
sys/ufs/ffs/ffs_softdep.c

Version: -  FREEBSD  -  FREEBSD-13-STABLE  -  FREEBSD-13-0  -  FREEBSD-12-STABLE  -  FREEBSD-12-0  -  FREEBSD-11-STABLE  -  FREEBSD-11-0  -  FREEBSD-10-STABLE  -  FREEBSD-10-0  -  FREEBSD-9-STABLE  -  FREEBSD-9-0  -  FREEBSD-8-STABLE  -  FREEBSD-8-0  -  FREEBSD-7-STABLE  -  FREEBSD-7-0  -  FREEBSD-6-STABLE  -  FREEBSD-6-0  -  FREEBSD-5-STABLE  -  FREEBSD-5-0  -  FREEBSD-4-STABLE  -  FREEBSD-3-STABLE  -  FREEBSD22  -  l41  -  OPENBSD  -  linux-2.6  -  MK84  -  PLAN9  -  xnu-8792 
SearchContext: -  none  -  3  -  10 

    1 /*-
    2  * Copyright 1998, 2000 Marshall Kirk McKusick.
    3  * Copyright 2009, 2010 Jeffrey W. Roberson <jeff@FreeBSD.org>
    4  * All rights reserved.
    5  *
    6  * The soft updates code is derived from the appendix of a University
    7  * of Michigan technical report (Gregory R. Ganger and Yale N. Patt,
    8  * "Soft Updates: A Solution to the Metadata Update Problem in File
    9  * Systems", CSE-TR-254-95, August 1995).
   10  *
   11  * Further information about soft updates can be obtained from:
   12  *
   13  *      Marshall Kirk McKusick          http://www.mckusick.com/softdep/
   14  *      1614 Oxford Street              mckusick@mckusick.com
   15  *      Berkeley, CA 94709-1608         +1-510-843-9542
   16  *      USA
   17  *
   18  * Redistribution and use in source and binary forms, with or without
   19  * modification, are permitted provided that the following conditions
   20  * are met:
   21  *
   22  * 1. Redistributions of source code must retain the above copyright
   23  *    notice, this list of conditions and the following disclaimer.
   24  * 2. Redistributions in binary form must reproduce the above copyright
   25  *    notice, this list of conditions and the following disclaimer in the
   26  *    documentation and/or other materials provided with the distribution.
   27  *
   28  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS ``AS IS'' AND ANY EXPRESS OR
   29  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
   30  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
   31  * IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY DIRECT, INDIRECT,
   32  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
   33  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
   34  * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
   35  * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR
   36  * TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
   37  * USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
   38  *
   39  *      from: @(#)ffs_softdep.c 9.59 (McKusick) 6/21/00
   40  */
   41 
   42 #include <sys/cdefs.h>
   43 __FBSDID("$FreeBSD: releng/10.0/sys/ufs/ffs/ffs_softdep.c 253974 2013-08-05 22:02:45Z mckusick $");
   44 
   45 #include "opt_ffs.h"
   46 #include "opt_quota.h"
   47 #include "opt_ddb.h"
   48 
   49 /*
   50  * For now we want the safety net that the DEBUG flag provides.
   51  */
   52 #ifndef DEBUG
   53 #define DEBUG
   54 #endif
   55 
   56 #include <sys/param.h>
   57 #include <sys/kernel.h>
   58 #include <sys/systm.h>
   59 #include <sys/bio.h>
   60 #include <sys/buf.h>
   61 #include <sys/kdb.h>
   62 #include <sys/kthread.h>
   63 #include <sys/ktr.h>
   64 #include <sys/limits.h>
   65 #include <sys/lock.h>
   66 #include <sys/malloc.h>
   67 #include <sys/mount.h>
   68 #include <sys/mutex.h>
   69 #include <sys/namei.h>
   70 #include <sys/priv.h>
   71 #include <sys/proc.h>
   72 #include <sys/rwlock.h>
   73 #include <sys/stat.h>
   74 #include <sys/sysctl.h>
   75 #include <sys/syslog.h>
   76 #include <sys/vnode.h>
   77 #include <sys/conf.h>
   78 
   79 #include <ufs/ufs/dir.h>
   80 #include <ufs/ufs/extattr.h>
   81 #include <ufs/ufs/quota.h>
   82 #include <ufs/ufs/inode.h>
   83 #include <ufs/ufs/ufsmount.h>
   84 #include <ufs/ffs/fs.h>
   85 #include <ufs/ffs/softdep.h>
   86 #include <ufs/ffs/ffs_extern.h>
   87 #include <ufs/ufs/ufs_extern.h>
   88 
   89 #include <vm/vm.h>
   90 #include <vm/vm_extern.h>
   91 #include <vm/vm_object.h>
   92 
   93 #include <geom/geom.h>
   94 
   95 #include <ddb/ddb.h>
   96 
   97 #define KTR_SUJ 0       /* Define to KTR_SPARE. */
   98 
   99 #ifndef SOFTUPDATES
  100 
  101 int
  102 softdep_flushfiles(oldmnt, flags, td)
  103         struct mount *oldmnt;
  104         int flags;
  105         struct thread *td;
  106 {
  107 
  108         panic("softdep_flushfiles called");
  109 }
  110 
  111 int
  112 softdep_mount(devvp, mp, fs, cred)
  113         struct vnode *devvp;
  114         struct mount *mp;
  115         struct fs *fs;
  116         struct ucred *cred;
  117 {
  118 
  119         return (0);
  120 }
  121 
  122 void 
  123 softdep_initialize()
  124 {
  125 
  126         return;
  127 }
  128 
  129 void
  130 softdep_uninitialize()
  131 {
  132 
  133         return;
  134 }
  135 
  136 void
  137 softdep_unmount(mp)
  138         struct mount *mp;
  139 {
  140 
  141 }
  142 
  143 void
  144 softdep_setup_sbupdate(ump, fs, bp)
  145         struct ufsmount *ump;
  146         struct fs *fs;
  147         struct buf *bp;
  148 {
  149 }
  150 
  151 void
  152 softdep_setup_inomapdep(bp, ip, newinum, mode)
  153         struct buf *bp;
  154         struct inode *ip;
  155         ino_t newinum;
  156         int mode;
  157 {
  158 
  159         panic("softdep_setup_inomapdep called");
  160 }
  161 
  162 void
  163 softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags)
  164         struct buf *bp;
  165         struct mount *mp;
  166         ufs2_daddr_t newblkno;
  167         int frags;
  168         int oldfrags;
  169 {
  170 
  171         panic("softdep_setup_blkmapdep called");
  172 }
  173 
  174 void 
  175 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
  176         struct inode *ip;
  177         ufs_lbn_t lbn;
  178         ufs2_daddr_t newblkno;
  179         ufs2_daddr_t oldblkno;
  180         long newsize;
  181         long oldsize;
  182         struct buf *bp;
  183 {
  184         
  185         panic("softdep_setup_allocdirect called");
  186 }
  187 
  188 void 
  189 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
  190         struct inode *ip;
  191         ufs_lbn_t lbn;
  192         ufs2_daddr_t newblkno;
  193         ufs2_daddr_t oldblkno;
  194         long newsize;
  195         long oldsize;
  196         struct buf *bp;
  197 {
  198         
  199         panic("softdep_setup_allocext called");
  200 }
  201 
  202 void
  203 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
  204         struct inode *ip;
  205         ufs_lbn_t lbn;
  206         struct buf *bp;
  207         int ptrno;
  208         ufs2_daddr_t newblkno;
  209         ufs2_daddr_t oldblkno;
  210         struct buf *nbp;
  211 {
  212 
  213         panic("softdep_setup_allocindir_page called");
  214 }
  215 
  216 void
  217 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
  218         struct buf *nbp;
  219         struct inode *ip;
  220         struct buf *bp;
  221         int ptrno;
  222         ufs2_daddr_t newblkno;
  223 {
  224 
  225         panic("softdep_setup_allocindir_meta called");
  226 }
  227 
  228 void
  229 softdep_journal_freeblocks(ip, cred, length, flags)
  230         struct inode *ip;
  231         struct ucred *cred;
  232         off_t length;
  233         int flags;
  234 {
  235         
  236         panic("softdep_journal_freeblocks called");
  237 }
  238 
  239 void
  240 softdep_journal_fsync(ip)
  241         struct inode *ip;
  242 {
  243 
  244         panic("softdep_journal_fsync called");
  245 }
  246 
  247 void
  248 softdep_setup_freeblocks(ip, length, flags)
  249         struct inode *ip;
  250         off_t length;
  251         int flags;
  252 {
  253         
  254         panic("softdep_setup_freeblocks called");
  255 }
  256 
  257 void
  258 softdep_freefile(pvp, ino, mode)
  259                 struct vnode *pvp;
  260                 ino_t ino;
  261                 int mode;
  262 {
  263 
  264         panic("softdep_freefile called");
  265 }
  266 
  267 int 
  268 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
  269         struct buf *bp;
  270         struct inode *dp;
  271         off_t diroffset;
  272         ino_t newinum;
  273         struct buf *newdirbp;
  274         int isnewblk;
  275 {
  276 
  277         panic("softdep_setup_directory_add called");
  278 }
  279 
  280 void 
  281 softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize)
  282         struct buf *bp;
  283         struct inode *dp;
  284         caddr_t base;
  285         caddr_t oldloc;
  286         caddr_t newloc;
  287         int entrysize;
  288 {
  289 
  290         panic("softdep_change_directoryentry_offset called");
  291 }
  292 
  293 void 
  294 softdep_setup_remove(bp, dp, ip, isrmdir)
  295         struct buf *bp;
  296         struct inode *dp;
  297         struct inode *ip;
  298         int isrmdir;
  299 {
  300         
  301         panic("softdep_setup_remove called");
  302 }
  303 
  304 void 
  305 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
  306         struct buf *bp;
  307         struct inode *dp;
  308         struct inode *ip;
  309         ino_t newinum;
  310         int isrmdir;
  311 {
  312 
  313         panic("softdep_setup_directory_change called");
  314 }
  315 
  316 void
  317 softdep_setup_blkfree(mp, bp, blkno, frags, wkhd)
  318         struct mount *mp;
  319         struct buf *bp;
  320         ufs2_daddr_t blkno;
  321         int frags;
  322         struct workhead *wkhd;
  323 {
  324 
  325         panic("%s called", __FUNCTION__);
  326 }
  327 
  328 void
  329 softdep_setup_inofree(mp, bp, ino, wkhd)
  330         struct mount *mp;
  331         struct buf *bp;
  332         ino_t ino;
  333         struct workhead *wkhd;
  334 {
  335 
  336         panic("%s called", __FUNCTION__);
  337 }
  338 
  339 void
  340 softdep_setup_unlink(dp, ip)
  341         struct inode *dp;
  342         struct inode *ip;
  343 {
  344 
  345         panic("%s called", __FUNCTION__);
  346 }
  347 
  348 void
  349 softdep_setup_link(dp, ip)
  350         struct inode *dp;
  351         struct inode *ip;
  352 {
  353 
  354         panic("%s called", __FUNCTION__);
  355 }
  356 
  357 void
  358 softdep_revert_link(dp, ip)
  359         struct inode *dp;
  360         struct inode *ip;
  361 {
  362 
  363         panic("%s called", __FUNCTION__);
  364 }
  365 
  366 void
  367 softdep_setup_rmdir(dp, ip)
  368         struct inode *dp;
  369         struct inode *ip;
  370 {
  371 
  372         panic("%s called", __FUNCTION__);
  373 }
  374 
  375 void
  376 softdep_revert_rmdir(dp, ip)
  377         struct inode *dp;
  378         struct inode *ip;
  379 {
  380 
  381         panic("%s called", __FUNCTION__);
  382 }
  383 
  384 void
  385 softdep_setup_create(dp, ip)
  386         struct inode *dp;
  387         struct inode *ip;
  388 {
  389 
  390         panic("%s called", __FUNCTION__);
  391 }
  392 
  393 void
  394 softdep_revert_create(dp, ip)
  395         struct inode *dp;
  396         struct inode *ip;
  397 {
  398 
  399         panic("%s called", __FUNCTION__);
  400 }
  401 
  402 void
  403 softdep_setup_mkdir(dp, ip)
  404         struct inode *dp;
  405         struct inode *ip;
  406 {
  407 
  408         panic("%s called", __FUNCTION__);
  409 }
  410 
  411 void
  412 softdep_revert_mkdir(dp, ip)
  413         struct inode *dp;
  414         struct inode *ip;
  415 {
  416 
  417         panic("%s called", __FUNCTION__);
  418 }
  419 
  420 void
  421 softdep_setup_dotdot_link(dp, ip)
  422         struct inode *dp;
  423         struct inode *ip;
  424 {
  425 
  426         panic("%s called", __FUNCTION__);
  427 }
  428 
  429 int
  430 softdep_prealloc(vp, waitok)
  431         struct vnode *vp;
  432         int waitok;
  433 {
  434 
  435         panic("%s called", __FUNCTION__);
  436 
  437         return (0);
  438 }
  439 
  440 int
  441 softdep_journal_lookup(mp, vpp)
  442         struct mount *mp;
  443         struct vnode **vpp;
  444 {
  445 
  446         return (ENOENT);
  447 }
  448 
  449 void
  450 softdep_change_linkcnt(ip)
  451         struct inode *ip;
  452 {
  453 
  454         panic("softdep_change_linkcnt called");
  455 }
  456 
  457 void 
  458 softdep_load_inodeblock(ip)
  459         struct inode *ip;
  460 {
  461 
  462         panic("softdep_load_inodeblock called");
  463 }
  464 
  465 void 
  466 softdep_update_inodeblock(ip, bp, waitfor)
  467         struct inode *ip;
  468         struct buf *bp;
  469         int waitfor;
  470 {
  471 
  472         panic("softdep_update_inodeblock called");
  473 }
  474 
  475 int
  476 softdep_fsync(vp)
  477         struct vnode *vp;       /* the "in_core" copy of the inode */
  478 {
  479 
  480         return (0);
  481 }
  482 
  483 void
  484 softdep_fsync_mountdev(vp)
  485         struct vnode *vp;
  486 {
  487 
  488         return;
  489 }
  490 
  491 int
  492 softdep_flushworklist(oldmnt, countp, td)
  493         struct mount *oldmnt;
  494         int *countp;
  495         struct thread *td;
  496 {
  497 
  498         *countp = 0;
  499         return (0);
  500 }
  501 
  502 int
  503 softdep_sync_metadata(struct vnode *vp)
  504 {
  505 
  506         return (0);
  507 }
  508 
  509 int
  510 softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor)
  511 {
  512 
  513         return (0);
  514 }
  515 
  516 int
  517 softdep_slowdown(vp)
  518         struct vnode *vp;
  519 {
  520 
  521         panic("softdep_slowdown called");
  522 }
  523 
  524 void
  525 softdep_releasefile(ip)
  526         struct inode *ip;       /* inode with the zero effective link count */
  527 {
  528 
  529         panic("softdep_releasefile called");
  530 }
  531 
  532 int
  533 softdep_request_cleanup(fs, vp, cred, resource)
  534         struct fs *fs;
  535         struct vnode *vp;
  536         struct ucred *cred;
  537         int resource;
  538 {
  539 
  540         return (0);
  541 }
  542 
  543 int
  544 softdep_check_suspend(struct mount *mp,
  545                       struct vnode *devvp,
  546                       int softdep_deps,
  547                       int softdep_accdeps,
  548                       int secondary_writes,
  549                       int secondary_accwrites)
  550 {
  551         struct bufobj *bo;
  552         int error;
  553         
  554         (void) softdep_deps,
  555         (void) softdep_accdeps;
  556 
  557         bo = &devvp->v_bufobj;
  558         ASSERT_BO_WLOCKED(bo);
  559 
  560         MNT_ILOCK(mp);
  561         while (mp->mnt_secondary_writes != 0) {
  562                 BO_UNLOCK(bo);
  563                 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp),
  564                     (PUSER - 1) | PDROP, "secwr", 0);
  565                 BO_LOCK(bo);
  566                 MNT_ILOCK(mp);
  567         }
  568 
  569         /*
  570          * Reasons for needing more work before suspend:
  571          * - Dirty buffers on devvp.
  572          * - Secondary writes occurred after start of vnode sync loop
  573          */
  574         error = 0;
  575         if (bo->bo_numoutput > 0 ||
  576             bo->bo_dirty.bv_cnt > 0 ||
  577             secondary_writes != 0 ||
  578             mp->mnt_secondary_writes != 0 ||
  579             secondary_accwrites != mp->mnt_secondary_accwrites)
  580                 error = EAGAIN;
  581         BO_UNLOCK(bo);
  582         return (error);
  583 }
  584 
  585 void
  586 softdep_get_depcounts(struct mount *mp,
  587                       int *softdepactivep,
  588                       int *softdepactiveaccp)
  589 {
  590         (void) mp;
  591         *softdepactivep = 0;
  592         *softdepactiveaccp = 0;
  593 }
  594 
  595 void
  596 softdep_buf_append(bp, wkhd)
  597         struct buf *bp;
  598         struct workhead *wkhd;
  599 {
  600 
  601         panic("softdep_buf_appendwork called");
  602 }
  603 
  604 void
  605 softdep_inode_append(ip, cred, wkhd)
  606         struct inode *ip;
  607         struct ucred *cred;
  608         struct workhead *wkhd;
  609 {
  610 
  611         panic("softdep_inode_appendwork called");
  612 }
  613 
  614 void
  615 softdep_freework(wkhd)
  616         struct workhead *wkhd;
  617 {
  618 
  619         panic("softdep_freework called");
  620 }
  621 
  622 #else
  623 
  624 FEATURE(softupdates, "FFS soft-updates support");
  625 
  626 /*
  627  * These definitions need to be adapted to the system to which
  628  * this file is being ported.
  629  */
  630 
  631 #define M_SOFTDEP_FLAGS (M_WAITOK)
  632 
  633 #define D_PAGEDEP       0
  634 #define D_INODEDEP      1
  635 #define D_BMSAFEMAP     2
  636 #define D_NEWBLK        3
  637 #define D_ALLOCDIRECT   4
  638 #define D_INDIRDEP      5
  639 #define D_ALLOCINDIR    6
  640 #define D_FREEFRAG      7
  641 #define D_FREEBLKS      8
  642 #define D_FREEFILE      9
  643 #define D_DIRADD        10
  644 #define D_MKDIR         11
  645 #define D_DIRREM        12
  646 #define D_NEWDIRBLK     13
  647 #define D_FREEWORK      14
  648 #define D_FREEDEP       15
  649 #define D_JADDREF       16
  650 #define D_JREMREF       17
  651 #define D_JMVREF        18
  652 #define D_JNEWBLK       19
  653 #define D_JFREEBLK      20
  654 #define D_JFREEFRAG     21
  655 #define D_JSEG          22
  656 #define D_JSEGDEP       23
  657 #define D_SBDEP         24
  658 #define D_JTRUNC        25
  659 #define D_JFSYNC        26
  660 #define D_SENTINEL      27
  661 #define D_LAST          D_SENTINEL
  662 
  663 unsigned long dep_current[D_LAST + 1];
  664 unsigned long dep_highuse[D_LAST + 1];
  665 unsigned long dep_total[D_LAST + 1];
  666 unsigned long dep_write[D_LAST + 1];
  667 
  668 static SYSCTL_NODE(_debug, OID_AUTO, softdep, CTLFLAG_RW, 0,
  669     "soft updates stats");
  670 static SYSCTL_NODE(_debug_softdep, OID_AUTO, total, CTLFLAG_RW, 0,
  671     "total dependencies allocated");
  672 static SYSCTL_NODE(_debug_softdep, OID_AUTO, highuse, CTLFLAG_RW, 0,
  673     "high use dependencies allocated");
  674 static SYSCTL_NODE(_debug_softdep, OID_AUTO, current, CTLFLAG_RW, 0,
  675     "current dependencies allocated");
  676 static SYSCTL_NODE(_debug_softdep, OID_AUTO, write, CTLFLAG_RW, 0,
  677     "current dependencies written");
  678 
  679 #define SOFTDEP_TYPE(type, str, long)                                   \
  680     static MALLOC_DEFINE(M_ ## type, #str, long);                       \
  681     SYSCTL_ULONG(_debug_softdep_total, OID_AUTO, str, CTLFLAG_RD,       \
  682         &dep_total[D_ ## type], 0, "");                                 \
  683     SYSCTL_ULONG(_debug_softdep_current, OID_AUTO, str, CTLFLAG_RD,     \
  684         &dep_current[D_ ## type], 0, "");                               \
  685     SYSCTL_ULONG(_debug_softdep_highuse, OID_AUTO, str, CTLFLAG_RD,     \
  686         &dep_highuse[D_ ## type], 0, "");                               \
  687     SYSCTL_ULONG(_debug_softdep_write, OID_AUTO, str, CTLFLAG_RD,       \
  688         &dep_write[D_ ## type], 0, "");
  689 
  690 SOFTDEP_TYPE(PAGEDEP, pagedep, "File page dependencies"); 
  691 SOFTDEP_TYPE(INODEDEP, inodedep, "Inode dependencies");
  692 SOFTDEP_TYPE(BMSAFEMAP, bmsafemap,
  693     "Block or frag allocated from cyl group map");
  694 SOFTDEP_TYPE(NEWBLK, newblk, "New block or frag allocation dependency");
  695 SOFTDEP_TYPE(ALLOCDIRECT, allocdirect, "Block or frag dependency for an inode");
  696 SOFTDEP_TYPE(INDIRDEP, indirdep, "Indirect block dependencies");
  697 SOFTDEP_TYPE(ALLOCINDIR, allocindir, "Block dependency for an indirect block");
  698 SOFTDEP_TYPE(FREEFRAG, freefrag, "Previously used frag for an inode");
  699 SOFTDEP_TYPE(FREEBLKS, freeblks, "Blocks freed from an inode");
  700 SOFTDEP_TYPE(FREEFILE, freefile, "Inode deallocated");
  701 SOFTDEP_TYPE(DIRADD, diradd, "New directory entry");
  702 SOFTDEP_TYPE(MKDIR, mkdir, "New directory");
  703 SOFTDEP_TYPE(DIRREM, dirrem, "Directory entry deleted");
  704 SOFTDEP_TYPE(NEWDIRBLK, newdirblk, "Unclaimed new directory block");
  705 SOFTDEP_TYPE(FREEWORK, freework, "free an inode block");
  706 SOFTDEP_TYPE(FREEDEP, freedep, "track a block free");
  707 SOFTDEP_TYPE(JADDREF, jaddref, "Journal inode ref add");
  708 SOFTDEP_TYPE(JREMREF, jremref, "Journal inode ref remove");
  709 SOFTDEP_TYPE(JMVREF, jmvref, "Journal inode ref move");
  710 SOFTDEP_TYPE(JNEWBLK, jnewblk, "Journal new block");
  711 SOFTDEP_TYPE(JFREEBLK, jfreeblk, "Journal free block");
  712 SOFTDEP_TYPE(JFREEFRAG, jfreefrag, "Journal free frag");
  713 SOFTDEP_TYPE(JSEG, jseg, "Journal segment");
  714 SOFTDEP_TYPE(JSEGDEP, jsegdep, "Journal segment complete");
  715 SOFTDEP_TYPE(SBDEP, sbdep, "Superblock write dependency");
  716 SOFTDEP_TYPE(JTRUNC, jtrunc, "Journal inode truncation");
  717 SOFTDEP_TYPE(JFSYNC, jfsync, "Journal fsync complete");
  718 
  719 static MALLOC_DEFINE(M_SENTINEL, "sentinel", "Worklist sentinel");
  720 
  721 static MALLOC_DEFINE(M_SAVEDINO, "savedino", "Saved inodes");
  722 static MALLOC_DEFINE(M_JBLOCKS, "jblocks", "Journal block locations");
  723 
  724 /* 
  725  * translate from workitem type to memory type
  726  * MUST match the defines above, such that memtype[D_XXX] == M_XXX
  727  */
  728 static struct malloc_type *memtype[] = {
  729         M_PAGEDEP,
  730         M_INODEDEP,
  731         M_BMSAFEMAP,
  732         M_NEWBLK,
  733         M_ALLOCDIRECT,
  734         M_INDIRDEP,
  735         M_ALLOCINDIR,
  736         M_FREEFRAG,
  737         M_FREEBLKS,
  738         M_FREEFILE,
  739         M_DIRADD,
  740         M_MKDIR,
  741         M_DIRREM,
  742         M_NEWDIRBLK,
  743         M_FREEWORK,
  744         M_FREEDEP,
  745         M_JADDREF,
  746         M_JREMREF,
  747         M_JMVREF,
  748         M_JNEWBLK,
  749         M_JFREEBLK,
  750         M_JFREEFRAG,
  751         M_JSEG,
  752         M_JSEGDEP,
  753         M_SBDEP,
  754         M_JTRUNC,
  755         M_JFSYNC,
  756         M_SENTINEL
  757 };
  758 
  759 static LIST_HEAD(mkdirlist, mkdir) mkdirlisthd;
  760 
  761 #define DtoM(type) (memtype[type])
  762 
  763 /*
  764  * Names of malloc types.
  765  */
  766 #define TYPENAME(type)  \
  767         ((unsigned)(type) <= D_LAST ? memtype[type]->ks_shortdesc : "???")
  768 /*
  769  * End system adaptation definitions.
  770  */
  771 
  772 #define DOTDOT_OFFSET   offsetof(struct dirtemplate, dotdot_ino)
  773 #define DOT_OFFSET      offsetof(struct dirtemplate, dot_ino)
  774 
  775 /*
  776  * Forward declarations.
  777  */
  778 struct inodedep_hashhead;
  779 struct newblk_hashhead;
  780 struct pagedep_hashhead;
  781 struct bmsafemap_hashhead;
  782 
  783 /*
  784  * Private journaling structures.
  785  */
  786 struct jblocks {
  787         struct jseglst  jb_segs;        /* TAILQ of current segments. */
  788         struct jseg     *jb_writeseg;   /* Next write to complete. */
  789         struct jseg     *jb_oldestseg;  /* Oldest segment with valid entries. */
  790         struct jextent  *jb_extent;     /* Extent array. */
  791         uint64_t        jb_nextseq;     /* Next sequence number. */
  792         uint64_t        jb_oldestwrseq; /* Oldest written sequence number. */
  793         uint8_t         jb_needseg;     /* Need a forced segment. */
  794         uint8_t         jb_suspended;   /* Did journal suspend writes? */
  795         int             jb_avail;       /* Available extents. */
  796         int             jb_used;        /* Last used extent. */
  797         int             jb_head;        /* Allocator head. */
  798         int             jb_off;         /* Allocator extent offset. */
  799         int             jb_blocks;      /* Total disk blocks covered. */
  800         int             jb_free;        /* Total disk blocks free. */
  801         int             jb_min;         /* Minimum free space. */
  802         int             jb_low;         /* Low on space. */
  803         int             jb_age;         /* Insertion time of oldest rec. */
  804 };
  805 
  806 struct jextent {
  807         ufs2_daddr_t    je_daddr;       /* Disk block address. */
  808         int             je_blocks;      /* Disk block count. */
  809 };
  810 
  811 /*
  812  * Internal function prototypes.
  813  */
  814 static  void softdep_error(char *, int);
  815 static  void drain_output(struct vnode *);
  816 static  struct buf *getdirtybuf(struct buf *, struct rwlock *, int);
  817 static  void clear_remove(void);
  818 static  void clear_inodedeps(void);
  819 static  void unlinked_inodedep(struct mount *, struct inodedep *);
  820 static  void clear_unlinked_inodedep(struct inodedep *);
  821 static  struct inodedep *first_unlinked_inodedep(struct ufsmount *);
  822 static  int flush_pagedep_deps(struct vnode *, struct mount *,
  823             struct diraddhd *);
  824 static  int free_pagedep(struct pagedep *);
  825 static  int flush_newblk_dep(struct vnode *, struct mount *, ufs_lbn_t);
  826 static  int flush_inodedep_deps(struct vnode *, struct mount *, ino_t);
  827 static  int flush_deplist(struct allocdirectlst *, int, int *);
  828 static  int sync_cgs(struct mount *, int);
  829 static  int handle_written_filepage(struct pagedep *, struct buf *);
  830 static  int handle_written_sbdep(struct sbdep *, struct buf *);
  831 static  void initiate_write_sbdep(struct sbdep *);
  832 static  void diradd_inode_written(struct diradd *, struct inodedep *);
  833 static  int handle_written_indirdep(struct indirdep *, struct buf *,
  834             struct buf**);
  835 static  int handle_written_inodeblock(struct inodedep *, struct buf *);
  836 static  int jnewblk_rollforward(struct jnewblk *, struct fs *, struct cg *,
  837             uint8_t *);
  838 static  int handle_written_bmsafemap(struct bmsafemap *, struct buf *);
  839 static  void handle_written_jaddref(struct jaddref *);
  840 static  void handle_written_jremref(struct jremref *);
  841 static  void handle_written_jseg(struct jseg *, struct buf *);
  842 static  void handle_written_jnewblk(struct jnewblk *);
  843 static  void handle_written_jblkdep(struct jblkdep *);
  844 static  void handle_written_jfreefrag(struct jfreefrag *);
  845 static  void complete_jseg(struct jseg *);
  846 static  void complete_jsegs(struct jseg *);
  847 static  void jseg_write(struct ufsmount *ump, struct jseg *, uint8_t *);
  848 static  void jaddref_write(struct jaddref *, struct jseg *, uint8_t *);
  849 static  void jremref_write(struct jremref *, struct jseg *, uint8_t *);
  850 static  void jmvref_write(struct jmvref *, struct jseg *, uint8_t *);
  851 static  void jtrunc_write(struct jtrunc *, struct jseg *, uint8_t *);
  852 static  void jfsync_write(struct jfsync *, struct jseg *, uint8_t *data);
  853 static  void jnewblk_write(struct jnewblk *, struct jseg *, uint8_t *);
  854 static  void jfreeblk_write(struct jfreeblk *, struct jseg *, uint8_t *);
  855 static  void jfreefrag_write(struct jfreefrag *, struct jseg *, uint8_t *);
  856 static  inline void inoref_write(struct inoref *, struct jseg *,
  857             struct jrefrec *);
  858 static  void handle_allocdirect_partdone(struct allocdirect *,
  859             struct workhead *);
  860 static  struct jnewblk *cancel_newblk(struct newblk *, struct worklist *,
  861             struct workhead *);
  862 static  void indirdep_complete(struct indirdep *);
  863 static  int indirblk_lookup(struct mount *, ufs2_daddr_t);
  864 static  void indirblk_insert(struct freework *);
  865 static  void indirblk_remove(struct freework *);
  866 static  void handle_allocindir_partdone(struct allocindir *);
  867 static  void initiate_write_filepage(struct pagedep *, struct buf *);
  868 static  void initiate_write_indirdep(struct indirdep*, struct buf *);
  869 static  void handle_written_mkdir(struct mkdir *, int);
  870 static  int jnewblk_rollback(struct jnewblk *, struct fs *, struct cg *,
  871             uint8_t *);
  872 static  void initiate_write_bmsafemap(struct bmsafemap *, struct buf *);
  873 static  void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *);
  874 static  void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *);
  875 static  void handle_workitem_freefile(struct freefile *);
  876 static  int handle_workitem_remove(struct dirrem *, int);
  877 static  struct dirrem *newdirrem(struct buf *, struct inode *,
  878             struct inode *, int, struct dirrem **);
  879 static  struct indirdep *indirdep_lookup(struct mount *, struct inode *,
  880             struct buf *);
  881 static  void cancel_indirdep(struct indirdep *, struct buf *,
  882             struct freeblks *);
  883 static  void free_indirdep(struct indirdep *);
  884 static  void free_diradd(struct diradd *, struct workhead *);
  885 static  void merge_diradd(struct inodedep *, struct diradd *);
  886 static  void complete_diradd(struct diradd *);
  887 static  struct diradd *diradd_lookup(struct pagedep *, int);
  888 static  struct jremref *cancel_diradd_dotdot(struct inode *, struct dirrem *,
  889             struct jremref *);
  890 static  struct jremref *cancel_mkdir_dotdot(struct inode *, struct dirrem *,
  891             struct jremref *);
  892 static  void cancel_diradd(struct diradd *, struct dirrem *, struct jremref *,
  893             struct jremref *, struct jremref *);
  894 static  void dirrem_journal(struct dirrem *, struct jremref *, struct jremref *,
  895             struct jremref *);
  896 static  void cancel_allocindir(struct allocindir *, struct buf *bp,
  897             struct freeblks *, int);
  898 static  int setup_trunc_indir(struct freeblks *, struct inode *,
  899             ufs_lbn_t, ufs_lbn_t, ufs2_daddr_t);
  900 static  void complete_trunc_indir(struct freework *);
  901 static  void trunc_indirdep(struct indirdep *, struct freeblks *, struct buf *,
  902             int);
  903 static  void complete_mkdir(struct mkdir *);
  904 static  void free_newdirblk(struct newdirblk *);
  905 static  void free_jremref(struct jremref *);
  906 static  void free_jaddref(struct jaddref *);
  907 static  void free_jsegdep(struct jsegdep *);
  908 static  void free_jsegs(struct jblocks *);
  909 static  void rele_jseg(struct jseg *);
  910 static  void free_jseg(struct jseg *, struct jblocks *);
  911 static  void free_jnewblk(struct jnewblk *);
  912 static  void free_jblkdep(struct jblkdep *);
  913 static  void free_jfreefrag(struct jfreefrag *);
  914 static  void free_freedep(struct freedep *);
  915 static  void journal_jremref(struct dirrem *, struct jremref *,
  916             struct inodedep *);
  917 static  void cancel_jnewblk(struct jnewblk *, struct workhead *);
  918 static  int cancel_jaddref(struct jaddref *, struct inodedep *,
  919             struct workhead *);
  920 static  void cancel_jfreefrag(struct jfreefrag *);
  921 static  inline void setup_freedirect(struct freeblks *, struct inode *,
  922             int, int);
  923 static  inline void setup_freeext(struct freeblks *, struct inode *, int, int);
  924 static  inline void setup_freeindir(struct freeblks *, struct inode *, int,
  925             ufs_lbn_t, int);
  926 static  inline struct freeblks *newfreeblks(struct mount *, struct inode *);
  927 static  void freeblks_free(struct ufsmount *, struct freeblks *, int);
  928 static  void indir_trunc(struct freework *, ufs2_daddr_t, ufs_lbn_t);
  929 ufs2_daddr_t blkcount(struct fs *, ufs2_daddr_t, off_t);
  930 static  int trunc_check_buf(struct buf *, int *, ufs_lbn_t, int, int);
  931 static  void trunc_dependencies(struct inode *, struct freeblks *, ufs_lbn_t,
  932             int, int);
  933 static  void trunc_pages(struct inode *, off_t, ufs2_daddr_t, int);
  934 static  int cancel_pagedep(struct pagedep *, struct freeblks *, int);
  935 static  int deallocate_dependencies(struct buf *, struct freeblks *, int);
  936 static  void newblk_freefrag(struct newblk*);
  937 static  void free_newblk(struct newblk *);
  938 static  void cancel_allocdirect(struct allocdirectlst *,
  939             struct allocdirect *, struct freeblks *);
  940 static  int check_inode_unwritten(struct inodedep *);
  941 static  int free_inodedep(struct inodedep *);
  942 static  void freework_freeblock(struct freework *);
  943 static  void freework_enqueue(struct freework *);
  944 static  int handle_workitem_freeblocks(struct freeblks *, int);
  945 static  int handle_complete_freeblocks(struct freeblks *, int);
  946 static  void handle_workitem_indirblk(struct freework *);
  947 static  void handle_written_freework(struct freework *);
  948 static  void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *);
  949 static  struct worklist *jnewblk_merge(struct worklist *, struct worklist *,
  950             struct workhead *);
  951 static  struct freefrag *setup_allocindir_phase2(struct buf *, struct inode *,
  952             struct inodedep *, struct allocindir *, ufs_lbn_t);
  953 static  struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t,
  954             ufs2_daddr_t, ufs_lbn_t);
  955 static  void handle_workitem_freefrag(struct freefrag *);
  956 static  struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long,
  957             ufs_lbn_t);
  958 static  void allocdirect_merge(struct allocdirectlst *,
  959             struct allocdirect *, struct allocdirect *);
  960 static  struct freefrag *allocindir_merge(struct allocindir *,
  961             struct allocindir *);
  962 static  int bmsafemap_find(struct bmsafemap_hashhead *, struct mount *, int,
  963             struct bmsafemap **);
  964 static  struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *,
  965             int cg, struct bmsafemap *);
  966 static  int newblk_find(struct newblk_hashhead *, struct mount *, ufs2_daddr_t,
  967             int, struct newblk **);
  968 static  int newblk_lookup(struct mount *, ufs2_daddr_t, int, struct newblk **);
  969 static  int inodedep_find(struct inodedep_hashhead *, struct fs *, ino_t,
  970             struct inodedep **);
  971 static  int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **);
  972 static  int pagedep_lookup(struct mount *, struct buf *bp, ino_t, ufs_lbn_t,
  973             int, struct pagedep **);
  974 static  int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t,
  975             struct mount *mp, int, struct pagedep **);
  976 static  void pause_timer(void *);
  977 static  int request_cleanup(struct mount *, int);
  978 static  int process_worklist_item(struct mount *, int, int);
  979 static  void process_removes(struct vnode *);
  980 static  void process_truncates(struct vnode *);
  981 static  void jwork_move(struct workhead *, struct workhead *);
  982 static  void jwork_insert(struct workhead *, struct jsegdep *);
  983 static  void add_to_worklist(struct worklist *, int);
  984 static  void wake_worklist(struct worklist *);
  985 static  void wait_worklist(struct worklist *, char *);
  986 static  void remove_from_worklist(struct worklist *);
  987 static  void softdep_flush(void);
  988 static  void softdep_flushjournal(struct mount *);
  989 static  int softdep_speedup(void);
  990 static  void worklist_speedup(void);
  991 static  int journal_mount(struct mount *, struct fs *, struct ucred *);
  992 static  void journal_unmount(struct mount *);
  993 static  int journal_space(struct ufsmount *, int);
  994 static  void journal_suspend(struct ufsmount *);
  995 static  int journal_unsuspend(struct ufsmount *ump);
  996 static  void softdep_prelink(struct vnode *, struct vnode *);
  997 static  void add_to_journal(struct worklist *);
  998 static  void remove_from_journal(struct worklist *);
  999 static  void softdep_process_journal(struct mount *, struct worklist *, int);
 1000 static  struct jremref *newjremref(struct dirrem *, struct inode *,
 1001             struct inode *ip, off_t, nlink_t);
 1002 static  struct jaddref *newjaddref(struct inode *, ino_t, off_t, int16_t,
 1003             uint16_t);
 1004 static  inline void newinoref(struct inoref *, ino_t, ino_t, off_t, nlink_t,
 1005             uint16_t);
 1006 static  inline struct jsegdep *inoref_jseg(struct inoref *);
 1007 static  struct jmvref *newjmvref(struct inode *, ino_t, off_t, off_t);
 1008 static  struct jfreeblk *newjfreeblk(struct freeblks *, ufs_lbn_t,
 1009             ufs2_daddr_t, int);
 1010 static  struct jtrunc *newjtrunc(struct freeblks *, off_t, int);
 1011 static  void move_newblock_dep(struct jaddref *, struct inodedep *);
 1012 static  void cancel_jfreeblk(struct freeblks *, ufs2_daddr_t);
 1013 static  struct jfreefrag *newjfreefrag(struct freefrag *, struct inode *,
 1014             ufs2_daddr_t, long, ufs_lbn_t);
 1015 static  struct freework *newfreework(struct ufsmount *, struct freeblks *,
 1016             struct freework *, ufs_lbn_t, ufs2_daddr_t, int, int, int);
 1017 static  int jwait(struct worklist *, int);
 1018 static  struct inodedep *inodedep_lookup_ip(struct inode *);
 1019 static  int bmsafemap_backgroundwrite(struct bmsafemap *, struct buf *);
 1020 static  struct freefile *handle_bufwait(struct inodedep *, struct workhead *);
 1021 static  void handle_jwork(struct workhead *);
 1022 static  struct mkdir *setup_newdir(struct diradd *, ino_t, ino_t, struct buf *,
 1023             struct mkdir **);
 1024 static  struct jblocks *jblocks_create(void);
 1025 static  ufs2_daddr_t jblocks_alloc(struct jblocks *, int, int *);
 1026 static  void jblocks_free(struct jblocks *, struct mount *, int);
 1027 static  void jblocks_destroy(struct jblocks *);
 1028 static  void jblocks_add(struct jblocks *, ufs2_daddr_t, int);
 1029 
 1030 /*
 1031  * Exported softdep operations.
 1032  */
 1033 static  void softdep_disk_io_initiation(struct buf *);
 1034 static  void softdep_disk_write_complete(struct buf *);
 1035 static  void softdep_deallocate_dependencies(struct buf *);
 1036 static  int softdep_count_dependencies(struct buf *bp, int);
 1037 
 1038 static struct rwlock lk;
 1039 RW_SYSINIT(softdep_lock, &lk, "Softdep Lock");
 1040 
 1041 #define TRY_ACQUIRE_LOCK(lk)            rw_try_wlock(lk)
 1042 #define ACQUIRE_LOCK(lk)                rw_wlock(lk)
 1043 #define FREE_LOCK(lk)                   rw_wunlock(lk)
 1044 
 1045 #define BUF_AREC(bp)                    lockallowrecurse(&(bp)->b_lock)
 1046 #define BUF_NOREC(bp)                   lockdisablerecurse(&(bp)->b_lock)
 1047 
 1048 /*
 1049  * Worklist queue management.
 1050  * These routines require that the lock be held.
 1051  */
 1052 #ifndef /* NOT */ DEBUG
 1053 #define WORKLIST_INSERT(head, item) do {        \
 1054         (item)->wk_state |= ONWORKLIST;         \
 1055         LIST_INSERT_HEAD(head, item, wk_list);  \
 1056 } while (0)
 1057 #define WORKLIST_REMOVE(item) do {              \
 1058         (item)->wk_state &= ~ONWORKLIST;        \
 1059         LIST_REMOVE(item, wk_list);             \
 1060 } while (0)
 1061 #define WORKLIST_INSERT_UNLOCKED        WORKLIST_INSERT
 1062 #define WORKLIST_REMOVE_UNLOCKED        WORKLIST_REMOVE
 1063 
 1064 #else /* DEBUG */
 1065 static  void worklist_insert(struct workhead *, struct worklist *, int);
 1066 static  void worklist_remove(struct worklist *, int);
 1067 
 1068 #define WORKLIST_INSERT(head, item) worklist_insert(head, item, 1)
 1069 #define WORKLIST_INSERT_UNLOCKED(head, item) worklist_insert(head, item, 0)
 1070 #define WORKLIST_REMOVE(item) worklist_remove(item, 1)
 1071 #define WORKLIST_REMOVE_UNLOCKED(item) worklist_remove(item, 0)
 1072 
 1073 static void
 1074 worklist_insert(head, item, locked)
 1075         struct workhead *head;
 1076         struct worklist *item;
 1077         int locked;
 1078 {
 1079 
 1080         if (locked)
 1081                 rw_assert(&lk, RA_WLOCKED);
 1082         if (item->wk_state & ONWORKLIST)
 1083                 panic("worklist_insert: %p %s(0x%X) already on list",
 1084                     item, TYPENAME(item->wk_type), item->wk_state);
 1085         item->wk_state |= ONWORKLIST;
 1086         LIST_INSERT_HEAD(head, item, wk_list);
 1087 }
 1088 
 1089 static void
 1090 worklist_remove(item, locked)
 1091         struct worklist *item;
 1092         int locked;
 1093 {
 1094 
 1095         if (locked)
 1096                 rw_assert(&lk, RA_WLOCKED);
 1097         if ((item->wk_state & ONWORKLIST) == 0)
 1098                 panic("worklist_remove: %p %s(0x%X) not on list",
 1099                     item, TYPENAME(item->wk_type), item->wk_state);
 1100         item->wk_state &= ~ONWORKLIST;
 1101         LIST_REMOVE(item, wk_list);
 1102 }
 1103 #endif /* DEBUG */
 1104 
 1105 /*
 1106  * Merge two jsegdeps keeping only the oldest one as newer references
 1107  * can't be discarded until after older references.
 1108  */
 1109 static inline struct jsegdep *
 1110 jsegdep_merge(struct jsegdep *one, struct jsegdep *two)
 1111 {
 1112         struct jsegdep *swp;
 1113 
 1114         if (two == NULL)
 1115                 return (one);
 1116 
 1117         if (one->jd_seg->js_seq > two->jd_seg->js_seq) {
 1118                 swp = one;
 1119                 one = two;
 1120                 two = swp;
 1121         }
 1122         WORKLIST_REMOVE(&two->jd_list);
 1123         free_jsegdep(two);
 1124 
 1125         return (one);
 1126 }
 1127 
 1128 /*
 1129  * If two freedeps are compatible free one to reduce list size.
 1130  */
 1131 static inline struct freedep *
 1132 freedep_merge(struct freedep *one, struct freedep *two)
 1133 {
 1134         if (two == NULL)
 1135                 return (one);
 1136 
 1137         if (one->fd_freework == two->fd_freework) {
 1138                 WORKLIST_REMOVE(&two->fd_list);
 1139                 free_freedep(two);
 1140         }
 1141         return (one);
 1142 }
 1143 
 1144 /*
 1145  * Move journal work from one list to another.  Duplicate freedeps and
 1146  * jsegdeps are coalesced to keep the lists as small as possible.
 1147  */
 1148 static void
 1149 jwork_move(dst, src)
 1150         struct workhead *dst;
 1151         struct workhead *src;
 1152 {
 1153         struct freedep *freedep;
 1154         struct jsegdep *jsegdep;
 1155         struct worklist *wkn;
 1156         struct worklist *wk;
 1157 
 1158         KASSERT(dst != src,
 1159             ("jwork_move: dst == src"));
 1160         freedep = NULL;
 1161         jsegdep = NULL;
 1162         LIST_FOREACH_SAFE(wk, dst, wk_list, wkn) {
 1163                 if (wk->wk_type == D_JSEGDEP)
 1164                         jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep);
 1165                 if (wk->wk_type == D_FREEDEP)
 1166                         freedep = freedep_merge(WK_FREEDEP(wk), freedep);
 1167         }
 1168 
 1169         rw_assert(&lk, RA_WLOCKED);
 1170         while ((wk = LIST_FIRST(src)) != NULL) {
 1171                 WORKLIST_REMOVE(wk);
 1172                 WORKLIST_INSERT(dst, wk);
 1173                 if (wk->wk_type == D_JSEGDEP) {
 1174                         jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep);
 1175                         continue;
 1176                 }
 1177                 if (wk->wk_type == D_FREEDEP)
 1178                         freedep = freedep_merge(WK_FREEDEP(wk), freedep);
 1179         }
 1180 }
 1181 
 1182 static void
 1183 jwork_insert(dst, jsegdep)
 1184         struct workhead *dst;
 1185         struct jsegdep *jsegdep;
 1186 {
 1187         struct jsegdep *jsegdepn;
 1188         struct worklist *wk;
 1189 
 1190         LIST_FOREACH(wk, dst, wk_list)
 1191                 if (wk->wk_type == D_JSEGDEP)
 1192                         break;
 1193         if (wk == NULL) {
 1194                 WORKLIST_INSERT(dst, &jsegdep->jd_list);
 1195                 return;
 1196         }
 1197         jsegdepn = WK_JSEGDEP(wk);
 1198         if (jsegdep->jd_seg->js_seq < jsegdepn->jd_seg->js_seq) {
 1199                 WORKLIST_REMOVE(wk);
 1200                 free_jsegdep(jsegdepn);
 1201                 WORKLIST_INSERT(dst, &jsegdep->jd_list);
 1202         } else
 1203                 free_jsegdep(jsegdep);
 1204 }
 1205 
 1206 /*
 1207  * Routines for tracking and managing workitems.
 1208  */
 1209 static  void workitem_free(struct worklist *, int);
 1210 static  void workitem_alloc(struct worklist *, int, struct mount *);
 1211 static  void workitem_reassign(struct worklist *, int);
 1212 
 1213 #define WORKITEM_FREE(item, type) \
 1214         workitem_free((struct worklist *)(item), (type))
 1215 #define WORKITEM_REASSIGN(item, type) \
 1216         workitem_reassign((struct worklist *)(item), (type))
 1217 
 1218 static void
 1219 workitem_free(item, type)
 1220         struct worklist *item;
 1221         int type;
 1222 {
 1223         struct ufsmount *ump;
 1224         rw_assert(&lk, RA_WLOCKED);
 1225 
 1226 #ifdef DEBUG
 1227         if (item->wk_state & ONWORKLIST)
 1228                 panic("workitem_free: %s(0x%X) still on list",
 1229                     TYPENAME(item->wk_type), item->wk_state);
 1230         if (item->wk_type != type && type != D_NEWBLK)
 1231                 panic("workitem_free: type mismatch %s != %s",
 1232                     TYPENAME(item->wk_type), TYPENAME(type));
 1233 #endif
 1234         if (item->wk_state & IOWAITING)
 1235                 wakeup(item);
 1236         ump = VFSTOUFS(item->wk_mp);
 1237         KASSERT(ump->softdep_deps > 0,
 1238             ("workitem_free: %s: softdep_deps going negative",
 1239             ump->um_fs->fs_fsmnt));
 1240         if (--ump->softdep_deps == 0 && ump->softdep_req)
 1241                 wakeup(&ump->softdep_deps);
 1242         KASSERT(dep_current[item->wk_type] > 0,
 1243             ("workitem_free: %s: dep_current[%s] going negative",
 1244             ump->um_fs->fs_fsmnt, TYPENAME(item->wk_type)));
 1245         dep_current[item->wk_type]--;
 1246         free(item, DtoM(type));
 1247 }
 1248 
 1249 static void
 1250 workitem_alloc(item, type, mp)
 1251         struct worklist *item;
 1252         int type;
 1253         struct mount *mp;
 1254 {
 1255         struct ufsmount *ump;
 1256 
 1257         item->wk_type = type;
 1258         item->wk_mp = mp;
 1259         item->wk_state = 0;
 1260 
 1261         ump = VFSTOUFS(mp);
 1262         ACQUIRE_LOCK(&lk);
 1263         dep_current[type]++;
 1264         if (dep_current[type] > dep_highuse[type])
 1265                 dep_highuse[type] = dep_current[type];
 1266         dep_total[type]++;
 1267         ump->softdep_deps++;
 1268         ump->softdep_accdeps++;
 1269         FREE_LOCK(&lk);
 1270 }
 1271 
 1272 static void
 1273 workitem_reassign(item, newtype)
 1274         struct worklist *item;
 1275         int newtype;
 1276 {
 1277 
 1278         KASSERT(dep_current[item->wk_type] > 0,
 1279             ("workitem_reassign: %s: dep_current[%s] going negative",
 1280             VFSTOUFS(item->wk_mp)->um_fs->fs_fsmnt, TYPENAME(item->wk_type)));
 1281         dep_current[item->wk_type]--;
 1282         dep_current[newtype]++;
 1283         if (dep_current[newtype] > dep_highuse[newtype])
 1284                 dep_highuse[newtype] = dep_current[newtype];
 1285         dep_total[newtype]++;
 1286         item->wk_type = newtype;
 1287 }
 1288 
 1289 /*
 1290  * Workitem queue management
 1291  */
 1292 static int max_softdeps;        /* maximum number of structs before slowdown */
 1293 static int maxindirdeps = 50;   /* max number of indirdeps before slowdown */
 1294 static int tickdelay = 2;       /* number of ticks to pause during slowdown */
 1295 static int proc_waiting;        /* tracks whether we have a timeout posted */
 1296 static int *stat_countp;        /* statistic to count in proc_waiting timeout */
 1297 static struct callout softdep_callout;
 1298 static int req_pending;
 1299 static int req_clear_inodedeps; /* syncer process flush some inodedeps */
 1300 static int req_clear_remove;    /* syncer process flush some freeblks */
 1301 static int softdep_flushcache = 0; /* Should we do BIO_FLUSH? */
 1302 
 1303 /*
 1304  * runtime statistics
 1305  */
 1306 static int stat_worklist_push;  /* number of worklist cleanups */
 1307 static int stat_blk_limit_push; /* number of times block limit neared */
 1308 static int stat_ino_limit_push; /* number of times inode limit neared */
 1309 static int stat_blk_limit_hit;  /* number of times block slowdown imposed */
 1310 static int stat_ino_limit_hit;  /* number of times inode slowdown imposed */
 1311 static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */
 1312 static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */
 1313 static int stat_inode_bitmap;   /* bufs redirtied as inode bitmap not written */
 1314 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
 1315 static int stat_dir_entry;      /* bufs redirtied as dir entry cannot write */
 1316 static int stat_jaddref;        /* bufs redirtied as ino bitmap can not write */
 1317 static int stat_jnewblk;        /* bufs redirtied as blk bitmap can not write */
 1318 static int stat_journal_min;    /* Times hit journal min threshold */
 1319 static int stat_journal_low;    /* Times hit journal low threshold */
 1320 static int stat_journal_wait;   /* Times blocked in jwait(). */
 1321 static int stat_jwait_filepage; /* Times blocked in jwait() for filepage. */
 1322 static int stat_jwait_freeblks; /* Times blocked in jwait() for freeblks. */
 1323 static int stat_jwait_inode;    /* Times blocked in jwait() for inodes. */
 1324 static int stat_jwait_newblk;   /* Times blocked in jwait() for newblks. */
 1325 static int stat_cleanup_high_delay; /* Maximum cleanup delay (in ticks) */
 1326 static int stat_cleanup_blkrequests; /* Number of block cleanup requests */
 1327 static int stat_cleanup_inorequests; /* Number of inode cleanup requests */
 1328 static int stat_cleanup_retries; /* Number of cleanups that needed to flush */
 1329 static int stat_cleanup_failures; /* Number of cleanup requests that failed */
 1330 
 1331 SYSCTL_INT(_debug_softdep, OID_AUTO, max_softdeps, CTLFLAG_RW,
 1332     &max_softdeps, 0, "");
 1333 SYSCTL_INT(_debug_softdep, OID_AUTO, tickdelay, CTLFLAG_RW,
 1334     &tickdelay, 0, "");
 1335 SYSCTL_INT(_debug_softdep, OID_AUTO, maxindirdeps, CTLFLAG_RW,
 1336     &maxindirdeps, 0, "");
 1337 SYSCTL_INT(_debug_softdep, OID_AUTO, worklist_push, CTLFLAG_RW,
 1338     &stat_worklist_push, 0,"");
 1339 SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_push, CTLFLAG_RW,
 1340     &stat_blk_limit_push, 0,"");
 1341 SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_push, CTLFLAG_RW,
 1342     &stat_ino_limit_push, 0,"");
 1343 SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_hit, CTLFLAG_RW,
 1344     &stat_blk_limit_hit, 0, "");
 1345 SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_hit, CTLFLAG_RW,
 1346     &stat_ino_limit_hit, 0, "");
 1347 SYSCTL_INT(_debug_softdep, OID_AUTO, sync_limit_hit, CTLFLAG_RW,
 1348     &stat_sync_limit_hit, 0, "");
 1349 SYSCTL_INT(_debug_softdep, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW,
 1350     &stat_indir_blk_ptrs, 0, "");
 1351 SYSCTL_INT(_debug_softdep, OID_AUTO, inode_bitmap, CTLFLAG_RW,
 1352     &stat_inode_bitmap, 0, "");
 1353 SYSCTL_INT(_debug_softdep, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW,
 1354     &stat_direct_blk_ptrs, 0, "");
 1355 SYSCTL_INT(_debug_softdep, OID_AUTO, dir_entry, CTLFLAG_RW,
 1356     &stat_dir_entry, 0, "");
 1357 SYSCTL_INT(_debug_softdep, OID_AUTO, jaddref_rollback, CTLFLAG_RW,
 1358     &stat_jaddref, 0, "");
 1359 SYSCTL_INT(_debug_softdep, OID_AUTO, jnewblk_rollback, CTLFLAG_RW,
 1360     &stat_jnewblk, 0, "");
 1361 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_low, CTLFLAG_RW,
 1362     &stat_journal_low, 0, "");
 1363 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_min, CTLFLAG_RW,
 1364     &stat_journal_min, 0, "");
 1365 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_wait, CTLFLAG_RW,
 1366     &stat_journal_wait, 0, "");
 1367 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_filepage, CTLFLAG_RW,
 1368     &stat_jwait_filepage, 0, "");
 1369 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_freeblks, CTLFLAG_RW,
 1370     &stat_jwait_freeblks, 0, "");
 1371 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_inode, CTLFLAG_RW,
 1372     &stat_jwait_inode, 0, "");
 1373 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_newblk, CTLFLAG_RW,
 1374     &stat_jwait_newblk, 0, "");
 1375 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_blkrequests, CTLFLAG_RW,
 1376     &stat_cleanup_blkrequests, 0, "");
 1377 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_inorequests, CTLFLAG_RW,
 1378     &stat_cleanup_inorequests, 0, "");
 1379 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_high_delay, CTLFLAG_RW,
 1380     &stat_cleanup_high_delay, 0, "");
 1381 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_retries, CTLFLAG_RW,
 1382     &stat_cleanup_retries, 0, "");
 1383 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_failures, CTLFLAG_RW,
 1384     &stat_cleanup_failures, 0, "");
 1385 SYSCTL_INT(_debug_softdep, OID_AUTO, flushcache, CTLFLAG_RW,
 1386     &softdep_flushcache, 0, "");
 1387 
 1388 SYSCTL_DECL(_vfs_ffs);
 1389 
 1390 LIST_HEAD(bmsafemap_hashhead, bmsafemap) *bmsafemap_hashtbl;
 1391 static u_long   bmsafemap_hash; /* size of hash table - 1 */
 1392 
 1393 static int compute_summary_at_mount = 0;        /* Whether to recompute the summary at mount time */
 1394 SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW,
 1395            &compute_summary_at_mount, 0, "Recompute summary at mount");
 1396 
 1397 static struct proc *softdepproc;
 1398 static struct kproc_desc softdep_kp = {
 1399         "softdepflush",
 1400         softdep_flush,
 1401         &softdepproc
 1402 };
 1403 SYSINIT(sdproc, SI_SUB_KTHREAD_UPDATE, SI_ORDER_ANY, kproc_start,
 1404     &softdep_kp);
 1405 
 1406 static void
 1407 softdep_flush(void)
 1408 {
 1409         struct mount *nmp;
 1410         struct mount *mp;
 1411         struct ufsmount *ump;
 1412         struct thread *td;
 1413         int remaining;
 1414         int progress;
 1415 
 1416         td = curthread;
 1417         td->td_pflags |= TDP_NORUNNINGBUF;
 1418 
 1419         for (;;) {      
 1420                 kproc_suspend_check(softdepproc);
 1421                 ACQUIRE_LOCK(&lk);
 1422                 /*
 1423                  * If requested, try removing inode or removal dependencies.
 1424                  */
 1425                 if (req_clear_inodedeps) {
 1426                         clear_inodedeps();
 1427                         req_clear_inodedeps -= 1;
 1428                         wakeup_one(&proc_waiting);
 1429                 }
 1430                 if (req_clear_remove) {
 1431                         clear_remove();
 1432                         req_clear_remove -= 1;
 1433                         wakeup_one(&proc_waiting);
 1434                 }
 1435                 FREE_LOCK(&lk);
 1436                 remaining = progress = 0;
 1437                 mtx_lock(&mountlist_mtx);
 1438                 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp)  {
 1439                         nmp = TAILQ_NEXT(mp, mnt_list);
 1440                         if (MOUNTEDSOFTDEP(mp) == 0)
 1441                                 continue;
 1442                         if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
 1443                                 continue;
 1444                         progress += softdep_process_worklist(mp, 0);
 1445                         ump = VFSTOUFS(mp);
 1446                         remaining += ump->softdep_on_worklist;
 1447                         mtx_lock(&mountlist_mtx);
 1448                         nmp = TAILQ_NEXT(mp, mnt_list);
 1449                         vfs_unbusy(mp);
 1450                 }
 1451                 mtx_unlock(&mountlist_mtx);
 1452                 if (remaining && progress)
 1453                         continue;
 1454                 ACQUIRE_LOCK(&lk);
 1455                 if (!req_pending)
 1456                         msleep(&req_pending, &lk, PVM, "sdflush", hz);
 1457                 req_pending = 0;
 1458                 FREE_LOCK(&lk);
 1459         }
 1460 }
 1461 
 1462 static void
 1463 worklist_speedup(void)
 1464 {
 1465         rw_assert(&lk, RA_WLOCKED);
 1466         if (req_pending == 0) {
 1467                 req_pending = 1;
 1468                 wakeup(&req_pending);
 1469         }
 1470 }
 1471 
 1472 static int
 1473 softdep_speedup(void)
 1474 {
 1475 
 1476         worklist_speedup();
 1477         bd_speedup();
 1478         return speedup_syncer();
 1479 }
 1480 
 1481 /*
 1482  * Add an item to the end of the work queue.
 1483  * This routine requires that the lock be held.
 1484  * This is the only routine that adds items to the list.
 1485  * The following routine is the only one that removes items
 1486  * and does so in order from first to last.
 1487  */
 1488 
 1489 #define WK_HEAD         0x0001  /* Add to HEAD. */
 1490 #define WK_NODELAY      0x0002  /* Process immediately. */
 1491 
 1492 static void
 1493 add_to_worklist(wk, flags)
 1494         struct worklist *wk;
 1495         int flags;
 1496 {
 1497         struct ufsmount *ump;
 1498 
 1499         rw_assert(&lk, RA_WLOCKED);
 1500         ump = VFSTOUFS(wk->wk_mp);
 1501         if (wk->wk_state & ONWORKLIST)
 1502                 panic("add_to_worklist: %s(0x%X) already on list",
 1503                     TYPENAME(wk->wk_type), wk->wk_state);
 1504         wk->wk_state |= ONWORKLIST;
 1505         if (ump->softdep_on_worklist == 0) {
 1506                 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list);
 1507                 ump->softdep_worklist_tail = wk;
 1508         } else if (flags & WK_HEAD) {
 1509                 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list);
 1510         } else {
 1511                 LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list);
 1512                 ump->softdep_worklist_tail = wk;
 1513         }
 1514         ump->softdep_on_worklist += 1;
 1515         if (flags & WK_NODELAY)
 1516                 worklist_speedup();
 1517 }
 1518 
 1519 /*
 1520  * Remove the item to be processed. If we are removing the last
 1521  * item on the list, we need to recalculate the tail pointer.
 1522  */
 1523 static void
 1524 remove_from_worklist(wk)
 1525         struct worklist *wk;
 1526 {
 1527         struct ufsmount *ump;
 1528 
 1529         ump = VFSTOUFS(wk->wk_mp);
 1530         WORKLIST_REMOVE(wk);
 1531         if (ump->softdep_worklist_tail == wk)
 1532                 ump->softdep_worklist_tail =
 1533                     (struct worklist *)wk->wk_list.le_prev;
 1534         ump->softdep_on_worklist -= 1;
 1535 }
 1536 
 1537 static void
 1538 wake_worklist(wk)
 1539         struct worklist *wk;
 1540 {
 1541         if (wk->wk_state & IOWAITING) {
 1542                 wk->wk_state &= ~IOWAITING;
 1543                 wakeup(wk);
 1544         }
 1545 }
 1546 
 1547 static void
 1548 wait_worklist(wk, wmesg)
 1549         struct worklist *wk;
 1550         char *wmesg;
 1551 {
 1552 
 1553         wk->wk_state |= IOWAITING;
 1554         msleep(wk, &lk, PVM, wmesg, 0);
 1555 }
 1556 
 1557 /*
 1558  * Process that runs once per second to handle items in the background queue.
 1559  *
 1560  * Note that we ensure that everything is done in the order in which they
 1561  * appear in the queue. The code below depends on this property to ensure
 1562  * that blocks of a file are freed before the inode itself is freed. This
 1563  * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
 1564  * until all the old ones have been purged from the dependency lists.
 1565  */
 1566 int 
 1567 softdep_process_worklist(mp, full)
 1568         struct mount *mp;
 1569         int full;
 1570 {
 1571         int cnt, matchcnt;
 1572         struct ufsmount *ump;
 1573         long starttime;
 1574 
 1575         KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp"));
 1576         /*
 1577          * Record the process identifier of our caller so that we can give
 1578          * this process preferential treatment in request_cleanup below.
 1579          */
 1580         matchcnt = 0;
 1581         ump = VFSTOUFS(mp);
 1582         ACQUIRE_LOCK(&lk);
 1583         starttime = time_second;
 1584         softdep_process_journal(mp, NULL, full?MNT_WAIT:0);
 1585         while (ump->softdep_on_worklist > 0) {
 1586                 if ((cnt = process_worklist_item(mp, 10, LK_NOWAIT)) == 0)
 1587                         break;
 1588                 else
 1589                         matchcnt += cnt;
 1590                 /*
 1591                  * If requested, try removing inode or removal dependencies.
 1592                  */
 1593                 if (req_clear_inodedeps) {
 1594                         clear_inodedeps();
 1595                         req_clear_inodedeps -= 1;
 1596                         wakeup_one(&proc_waiting);
 1597                 }
 1598                 if (req_clear_remove) {
 1599                         clear_remove();
 1600                         req_clear_remove -= 1;
 1601                         wakeup_one(&proc_waiting);
 1602                 }
 1603                 /*
 1604                  * We do not generally want to stop for buffer space, but if
 1605                  * we are really being a buffer hog, we will stop and wait.
 1606                  */
 1607                 if (should_yield()) {
 1608                         FREE_LOCK(&lk);
 1609                         kern_yield(PRI_USER);
 1610                         bwillwrite();
 1611                         ACQUIRE_LOCK(&lk);
 1612                 }
 1613                 /*
 1614                  * Never allow processing to run for more than one
 1615                  * second. Otherwise the other mountpoints may get
 1616                  * excessively backlogged.
 1617                  */
 1618                 if (!full && starttime != time_second)
 1619                         break;
 1620         }
 1621         if (full == 0)
 1622                 journal_unsuspend(ump);
 1623         FREE_LOCK(&lk);
 1624         return (matchcnt);
 1625 }
 1626 
 1627 /*
 1628  * Process all removes associated with a vnode if we are running out of
 1629  * journal space.  Any other process which attempts to flush these will
 1630  * be unable as we have the vnodes locked.
 1631  */
 1632 static void
 1633 process_removes(vp)
 1634         struct vnode *vp;
 1635 {
 1636         struct inodedep *inodedep;
 1637         struct dirrem *dirrem;
 1638         struct mount *mp;
 1639         ino_t inum;
 1640 
 1641         rw_assert(&lk, RA_WLOCKED);
 1642 
 1643         mp = vp->v_mount;
 1644         inum = VTOI(vp)->i_number;
 1645         for (;;) {
 1646 top:
 1647                 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0)
 1648                         return;
 1649                 LIST_FOREACH(dirrem, &inodedep->id_dirremhd, dm_inonext) {
 1650                         /*
 1651                          * If another thread is trying to lock this vnode
 1652                          * it will fail but we must wait for it to do so
 1653                          * before we can proceed.
 1654                          */
 1655                         if (dirrem->dm_state & INPROGRESS) {
 1656                                 wait_worklist(&dirrem->dm_list, "pwrwait");
 1657                                 goto top;
 1658                         }
 1659                         if ((dirrem->dm_state & (COMPLETE | ONWORKLIST)) == 
 1660                             (COMPLETE | ONWORKLIST))
 1661                                 break;
 1662                 }
 1663                 if (dirrem == NULL)
 1664                         return;
 1665                 remove_from_worklist(&dirrem->dm_list);
 1666                 FREE_LOCK(&lk);
 1667                 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT))
 1668                         panic("process_removes: suspended filesystem");
 1669                 handle_workitem_remove(dirrem, 0);
 1670                 vn_finished_secondary_write(mp);
 1671                 ACQUIRE_LOCK(&lk);
 1672         }
 1673 }
 1674 
 1675 /*
 1676  * Process all truncations associated with a vnode if we are running out
 1677  * of journal space.  This is called when the vnode lock is already held
 1678  * and no other process can clear the truncation.  This function returns
 1679  * a value greater than zero if it did any work.
 1680  */
 1681 static void
 1682 process_truncates(vp)
 1683         struct vnode *vp;
 1684 {
 1685         struct inodedep *inodedep;
 1686         struct freeblks *freeblks;
 1687         struct mount *mp;
 1688         ino_t inum;
 1689         int cgwait;
 1690 
 1691         rw_assert(&lk, RA_WLOCKED);
 1692 
 1693         mp = vp->v_mount;
 1694         inum = VTOI(vp)->i_number;
 1695         for (;;) {
 1696                 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0)
 1697                         return;
 1698                 cgwait = 0;
 1699                 TAILQ_FOREACH(freeblks, &inodedep->id_freeblklst, fb_next) {
 1700                         /* Journal entries not yet written.  */
 1701                         if (!LIST_EMPTY(&freeblks->fb_jblkdephd)) {
 1702                                 jwait(&LIST_FIRST(
 1703                                     &freeblks->fb_jblkdephd)->jb_list,
 1704                                     MNT_WAIT);
 1705                                 break;
 1706                         }
 1707                         /* Another thread is executing this item. */
 1708                         if (freeblks->fb_state & INPROGRESS) {
 1709                                 wait_worklist(&freeblks->fb_list, "ptrwait");
 1710                                 break;
 1711                         }
 1712                         /* Freeblks is waiting on a inode write. */
 1713                         if ((freeblks->fb_state & COMPLETE) == 0) {
 1714                                 FREE_LOCK(&lk);
 1715                                 ffs_update(vp, 1);
 1716                                 ACQUIRE_LOCK(&lk);
 1717                                 break;
 1718                         }
 1719                         if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST)) ==
 1720                             (ALLCOMPLETE | ONWORKLIST)) {
 1721                                 remove_from_worklist(&freeblks->fb_list);
 1722                                 freeblks->fb_state |= INPROGRESS;
 1723                                 FREE_LOCK(&lk);
 1724                                 if (vn_start_secondary_write(NULL, &mp,
 1725                                     V_NOWAIT))
 1726                                         panic("process_truncates: "
 1727                                             "suspended filesystem");
 1728                                 handle_workitem_freeblocks(freeblks, 0);
 1729                                 vn_finished_secondary_write(mp);
 1730                                 ACQUIRE_LOCK(&lk);
 1731                                 break;
 1732                         }
 1733                         if (freeblks->fb_cgwait)
 1734                                 cgwait++;
 1735                 }
 1736                 if (cgwait) {
 1737                         FREE_LOCK(&lk);
 1738                         sync_cgs(mp, MNT_WAIT);
 1739                         ffs_sync_snap(mp, MNT_WAIT);
 1740                         ACQUIRE_LOCK(&lk);
 1741                         continue;
 1742                 }
 1743                 if (freeblks == NULL)
 1744                         break;
 1745         }
 1746         return;
 1747 }
 1748 
 1749 /*
 1750  * Process one item on the worklist.
 1751  */
 1752 static int
 1753 process_worklist_item(mp, target, flags)
 1754         struct mount *mp;
 1755         int target;
 1756         int flags;
 1757 {
 1758         struct worklist sentinel;
 1759         struct worklist *wk;
 1760         struct ufsmount *ump;
 1761         int matchcnt;
 1762         int error;
 1763 
 1764         rw_assert(&lk, RA_WLOCKED);
 1765         KASSERT(mp != NULL, ("process_worklist_item: NULL mp"));
 1766         /*
 1767          * If we are being called because of a process doing a
 1768          * copy-on-write, then it is not safe to write as we may
 1769          * recurse into the copy-on-write routine.
 1770          */
 1771         if (curthread->td_pflags & TDP_COWINPROGRESS)
 1772                 return (-1);
 1773         PHOLD(curproc); /* Don't let the stack go away. */
 1774         ump = VFSTOUFS(mp);
 1775         matchcnt = 0;
 1776         sentinel.wk_mp = NULL;
 1777         sentinel.wk_type = D_SENTINEL;
 1778         LIST_INSERT_HEAD(&ump->softdep_workitem_pending, &sentinel, wk_list);
 1779         for (wk = LIST_NEXT(&sentinel, wk_list); wk != NULL;
 1780             wk = LIST_NEXT(&sentinel, wk_list)) {
 1781                 if (wk->wk_type == D_SENTINEL) {
 1782                         LIST_REMOVE(&sentinel, wk_list);
 1783                         LIST_INSERT_AFTER(wk, &sentinel, wk_list);
 1784                         continue;
 1785                 }
 1786                 if (wk->wk_state & INPROGRESS)
 1787                         panic("process_worklist_item: %p already in progress.",
 1788                             wk);
 1789                 wk->wk_state |= INPROGRESS;
 1790                 remove_from_worklist(wk);
 1791                 FREE_LOCK(&lk);
 1792                 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT))
 1793                         panic("process_worklist_item: suspended filesystem");
 1794                 switch (wk->wk_type) {
 1795                 case D_DIRREM:
 1796                         /* removal of a directory entry */
 1797                         error = handle_workitem_remove(WK_DIRREM(wk), flags);
 1798                         break;
 1799 
 1800                 case D_FREEBLKS:
 1801                         /* releasing blocks and/or fragments from a file */
 1802                         error = handle_workitem_freeblocks(WK_FREEBLKS(wk),
 1803                             flags);
 1804                         break;
 1805 
 1806                 case D_FREEFRAG:
 1807                         /* releasing a fragment when replaced as a file grows */
 1808                         handle_workitem_freefrag(WK_FREEFRAG(wk));
 1809                         error = 0;
 1810                         break;
 1811 
 1812                 case D_FREEFILE:
 1813                         /* releasing an inode when its link count drops to 0 */
 1814                         handle_workitem_freefile(WK_FREEFILE(wk));
 1815                         error = 0;
 1816                         break;
 1817 
 1818                 default:
 1819                         panic("%s_process_worklist: Unknown type %s",
 1820                             "softdep", TYPENAME(wk->wk_type));
 1821                         /* NOTREACHED */
 1822                 }
 1823                 vn_finished_secondary_write(mp);
 1824                 ACQUIRE_LOCK(&lk);
 1825                 if (error == 0) {
 1826                         if (++matchcnt == target)
 1827                                 break;
 1828                         continue;
 1829                 }
 1830                 /*
 1831                  * We have to retry the worklist item later.  Wake up any
 1832                  * waiters who may be able to complete it immediately and
 1833                  * add the item back to the head so we don't try to execute
 1834                  * it again.
 1835                  */
 1836                 wk->wk_state &= ~INPROGRESS;
 1837                 wake_worklist(wk);
 1838                 add_to_worklist(wk, WK_HEAD);
 1839         }
 1840         LIST_REMOVE(&sentinel, wk_list);
 1841         /* Sentinal could've become the tail from remove_from_worklist. */
 1842         if (ump->softdep_worklist_tail == &sentinel)
 1843                 ump->softdep_worklist_tail =
 1844                     (struct worklist *)sentinel.wk_list.le_prev;
 1845         PRELE(curproc);
 1846         return (matchcnt);
 1847 }
 1848 
 1849 /*
 1850  * Move dependencies from one buffer to another.
 1851  */
 1852 int
 1853 softdep_move_dependencies(oldbp, newbp)
 1854         struct buf *oldbp;
 1855         struct buf *newbp;
 1856 {
 1857         struct worklist *wk, *wktail;
 1858         int dirty;
 1859 
 1860         dirty = 0;
 1861         wktail = NULL;
 1862         ACQUIRE_LOCK(&lk);
 1863         while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
 1864                 LIST_REMOVE(wk, wk_list);
 1865                 if (wk->wk_type == D_BMSAFEMAP &&
 1866                     bmsafemap_backgroundwrite(WK_BMSAFEMAP(wk), newbp))
 1867                         dirty = 1;
 1868                 if (wktail == 0)
 1869                         LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
 1870                 else
 1871                         LIST_INSERT_AFTER(wktail, wk, wk_list);
 1872                 wktail = wk;
 1873         }
 1874         FREE_LOCK(&lk);
 1875 
 1876         return (dirty);
 1877 }
 1878 
 1879 /*
 1880  * Purge the work list of all items associated with a particular mount point.
 1881  */
 1882 int
 1883 softdep_flushworklist(oldmnt, countp, td)
 1884         struct mount *oldmnt;
 1885         int *countp;
 1886         struct thread *td;
 1887 {
 1888         struct vnode *devvp;
 1889         int count, error = 0;
 1890         struct ufsmount *ump;
 1891 
 1892         /*
 1893          * Alternately flush the block device associated with the mount
 1894          * point and process any dependencies that the flushing
 1895          * creates. We continue until no more worklist dependencies
 1896          * are found.
 1897          */
 1898         *countp = 0;
 1899         ump = VFSTOUFS(oldmnt);
 1900         devvp = ump->um_devvp;
 1901         while ((count = softdep_process_worklist(oldmnt, 1)) > 0) {
 1902                 *countp += count;
 1903                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
 1904                 error = VOP_FSYNC(devvp, MNT_WAIT, td);
 1905                 VOP_UNLOCK(devvp, 0);
 1906                 if (error)
 1907                         break;
 1908         }
 1909         return (error);
 1910 }
 1911 
 1912 int
 1913 softdep_waitidle(struct mount *mp)
 1914 {
 1915         struct ufsmount *ump;
 1916         int error;
 1917         int i;
 1918 
 1919         ump = VFSTOUFS(mp);
 1920         ACQUIRE_LOCK(&lk);
 1921         for (i = 0; i < 10 && ump->softdep_deps; i++) {
 1922                 ump->softdep_req = 1;
 1923                 if (ump->softdep_on_worklist)
 1924                         panic("softdep_waitidle: work added after flush.");
 1925                 msleep(&ump->softdep_deps, &lk, PVM, "softdeps", 1);
 1926         }
 1927         ump->softdep_req = 0;
 1928         FREE_LOCK(&lk);
 1929         error = 0;
 1930         if (i == 10) {
 1931                 error = EBUSY;
 1932                 printf("softdep_waitidle: Failed to flush worklist for %p\n",
 1933                     mp);
 1934         }
 1935 
 1936         return (error);
 1937 }
 1938 
 1939 /*
 1940  * Flush all vnodes and worklist items associated with a specified mount point.
 1941  */
 1942 int
 1943 softdep_flushfiles(oldmnt, flags, td)
 1944         struct mount *oldmnt;
 1945         int flags;
 1946         struct thread *td;
 1947 {
 1948 #ifdef QUOTA
 1949         struct ufsmount *ump;
 1950         int i;
 1951 #endif
 1952         int error, early, depcount, loopcnt, retry_flush_count, retry;
 1953         int morework;
 1954 
 1955         loopcnt = 10;
 1956         retry_flush_count = 3;
 1957 retry_flush:
 1958         error = 0;
 1959 
 1960         /*
 1961          * Alternately flush the vnodes associated with the mount
 1962          * point and process any dependencies that the flushing
 1963          * creates. In theory, this loop can happen at most twice,
 1964          * but we give it a few extra just to be sure.
 1965          */
 1966         for (; loopcnt > 0; loopcnt--) {
 1967                 /*
 1968                  * Do another flush in case any vnodes were brought in
 1969                  * as part of the cleanup operations.
 1970                  */
 1971                 early = retry_flush_count == 1 || (oldmnt->mnt_kern_flag &
 1972                     MNTK_UNMOUNT) == 0 ? 0 : EARLYFLUSH;
 1973                 if ((error = ffs_flushfiles(oldmnt, flags | early, td)) != 0)
 1974                         break;
 1975                 if ((error = softdep_flushworklist(oldmnt, &depcount, td)) != 0 ||
 1976                     depcount == 0)
 1977                         break;
 1978         }
 1979         /*
 1980          * If we are unmounting then it is an error to fail. If we
 1981          * are simply trying to downgrade to read-only, then filesystem
 1982          * activity can keep us busy forever, so we just fail with EBUSY.
 1983          */
 1984         if (loopcnt == 0) {
 1985                 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
 1986                         panic("softdep_flushfiles: looping");
 1987                 error = EBUSY;
 1988         }
 1989         if (!error)
 1990                 error = softdep_waitidle(oldmnt);
 1991         if (!error) {
 1992                 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) {
 1993                         retry = 0;
 1994                         MNT_ILOCK(oldmnt);
 1995                         KASSERT((oldmnt->mnt_kern_flag & MNTK_NOINSMNTQ) != 0,
 1996                             ("softdep_flushfiles: !MNTK_NOINSMNTQ"));
 1997                         morework = oldmnt->mnt_nvnodelistsize > 0;
 1998 #ifdef QUOTA
 1999                         ump = VFSTOUFS(oldmnt);
 2000                         UFS_LOCK(ump);
 2001                         for (i = 0; i < MAXQUOTAS; i++) {
 2002                                 if (ump->um_quotas[i] != NULLVP)
 2003                                         morework = 1;
 2004                         }
 2005                         UFS_UNLOCK(ump);
 2006 #endif
 2007                         if (morework) {
 2008                                 if (--retry_flush_count > 0) {
 2009                                         retry = 1;
 2010                                         loopcnt = 3;
 2011                                 } else
 2012                                         error = EBUSY;
 2013                         }
 2014                         MNT_IUNLOCK(oldmnt);
 2015                         if (retry)
 2016                                 goto retry_flush;
 2017                 }
 2018         }
 2019         return (error);
 2020 }
 2021 
 2022 /*
 2023  * Structure hashing.
 2024  * 
 2025  * There are three types of structures that can be looked up:
 2026  *      1) pagedep structures identified by mount point, inode number,
 2027  *         and logical block.
 2028  *      2) inodedep structures identified by mount point and inode number.
 2029  *      3) newblk structures identified by mount point and
 2030  *         physical block number.
 2031  *
 2032  * The "pagedep" and "inodedep" dependency structures are hashed
 2033  * separately from the file blocks and inodes to which they correspond.
 2034  * This separation helps when the in-memory copy of an inode or
 2035  * file block must be replaced. It also obviates the need to access
 2036  * an inode or file page when simply updating (or de-allocating)
 2037  * dependency structures. Lookup of newblk structures is needed to
 2038  * find newly allocated blocks when trying to associate them with
 2039  * their allocdirect or allocindir structure.
 2040  *
 2041  * The lookup routines optionally create and hash a new instance when
 2042  * an existing entry is not found.
 2043  */
 2044 #define DEPALLOC        0x0001  /* allocate structure if lookup fails */
 2045 #define NODELAY         0x0002  /* cannot do background work */
 2046 
 2047 /*
 2048  * Structures and routines associated with pagedep caching.
 2049  */
 2050 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
 2051 u_long  pagedep_hash;           /* size of hash table - 1 */
 2052 #define PAGEDEP_HASH(mp, inum, lbn) \
 2053         (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
 2054             pagedep_hash])
 2055 
 2056 static int
 2057 pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp)
 2058         struct pagedep_hashhead *pagedephd;
 2059         ino_t ino;
 2060         ufs_lbn_t lbn;
 2061         struct mount *mp;
 2062         int flags;
 2063         struct pagedep **pagedeppp;
 2064 {
 2065         struct pagedep *pagedep;
 2066 
 2067         LIST_FOREACH(pagedep, pagedephd, pd_hash) {
 2068                 if (ino == pagedep->pd_ino && lbn == pagedep->pd_lbn &&
 2069                     mp == pagedep->pd_list.wk_mp) {
 2070                         *pagedeppp = pagedep;
 2071                         return (1);
 2072                 }
 2073         }
 2074         *pagedeppp = NULL;
 2075         return (0);
 2076 }
 2077 /*
 2078  * Look up a pagedep. Return 1 if found, 0 otherwise.
 2079  * If not found, allocate if DEPALLOC flag is passed.
 2080  * Found or allocated entry is returned in pagedeppp.
 2081  * This routine must be called with splbio interrupts blocked.
 2082  */
 2083 static int
 2084 pagedep_lookup(mp, bp, ino, lbn, flags, pagedeppp)
 2085         struct mount *mp;
 2086         struct buf *bp;
 2087         ino_t ino;
 2088         ufs_lbn_t lbn;
 2089         int flags;
 2090         struct pagedep **pagedeppp;
 2091 {
 2092         struct pagedep *pagedep;
 2093         struct pagedep_hashhead *pagedephd;
 2094         struct worklist *wk;
 2095         int ret;
 2096         int i;
 2097 
 2098         rw_assert(&lk, RA_WLOCKED);
 2099         if (bp) {
 2100                 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
 2101                         if (wk->wk_type == D_PAGEDEP) {
 2102                                 *pagedeppp = WK_PAGEDEP(wk);
 2103                                 return (1);
 2104                         }
 2105                 }
 2106         }
 2107         pagedephd = PAGEDEP_HASH(mp, ino, lbn);
 2108         ret = pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp);
 2109         if (ret) {
 2110                 if (((*pagedeppp)->pd_state & ONWORKLIST) == 0 && bp)
 2111                         WORKLIST_INSERT(&bp->b_dep, &(*pagedeppp)->pd_list);
 2112                 return (1);
 2113         }
 2114         if ((flags & DEPALLOC) == 0)
 2115                 return (0);
 2116         FREE_LOCK(&lk);
 2117         pagedep = malloc(sizeof(struct pagedep),
 2118             M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO);
 2119         workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp);
 2120         ACQUIRE_LOCK(&lk);
 2121         ret = pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp);
 2122         if (*pagedeppp) {
 2123                 /*
 2124                  * This should never happen since we only create pagedeps
 2125                  * with the vnode lock held.  Could be an assert.
 2126                  */
 2127                 WORKITEM_FREE(pagedep, D_PAGEDEP);
 2128                 return (ret);
 2129         }
 2130         pagedep->pd_ino = ino;
 2131         pagedep->pd_lbn = lbn;
 2132         LIST_INIT(&pagedep->pd_dirremhd);
 2133         LIST_INIT(&pagedep->pd_pendinghd);
 2134         for (i = 0; i < DAHASHSZ; i++)
 2135                 LIST_INIT(&pagedep->pd_diraddhd[i]);
 2136         LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
 2137         WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
 2138         *pagedeppp = pagedep;
 2139         return (0);
 2140 }
 2141 
 2142 /*
 2143  * Structures and routines associated with inodedep caching.
 2144  */
 2145 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
 2146 static u_long   inodedep_hash;  /* size of hash table - 1 */
 2147 #define INODEDEP_HASH(fs, inum) \
 2148       (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
 2149 
 2150 static int
 2151 inodedep_find(inodedephd, fs, inum, inodedeppp)
 2152         struct inodedep_hashhead *inodedephd;
 2153         struct fs *fs;
 2154         ino_t inum;
 2155         struct inodedep **inodedeppp;
 2156 {
 2157         struct inodedep *inodedep;
 2158 
 2159         LIST_FOREACH(inodedep, inodedephd, id_hash)
 2160                 if (inum == inodedep->id_ino && fs == inodedep->id_fs)
 2161                         break;
 2162         if (inodedep) {
 2163                 *inodedeppp = inodedep;
 2164                 return (1);
 2165         }
 2166         *inodedeppp = NULL;
 2167 
 2168         return (0);
 2169 }
 2170 /*
 2171  * Look up an inodedep. Return 1 if found, 0 if not found.
 2172  * If not found, allocate if DEPALLOC flag is passed.
 2173  * Found or allocated entry is returned in inodedeppp.
 2174  * This routine must be called with splbio interrupts blocked.
 2175  */
 2176 static int
 2177 inodedep_lookup(mp, inum, flags, inodedeppp)
 2178         struct mount *mp;
 2179         ino_t inum;
 2180         int flags;
 2181         struct inodedep **inodedeppp;
 2182 {
 2183         struct inodedep *inodedep;
 2184         struct inodedep_hashhead *inodedephd;
 2185         struct fs *fs;
 2186 
 2187         rw_assert(&lk, RA_WLOCKED);
 2188         fs = VFSTOUFS(mp)->um_fs;
 2189         inodedephd = INODEDEP_HASH(fs, inum);
 2190 
 2191         if (inodedep_find(inodedephd, fs, inum, inodedeppp))
 2192                 return (1);
 2193         if ((flags & DEPALLOC) == 0)
 2194                 return (0);
 2195         /*
 2196          * If we are over our limit, try to improve the situation.
 2197          */
 2198         if (dep_current[D_INODEDEP] > max_softdeps && (flags & NODELAY) == 0)
 2199                 request_cleanup(mp, FLUSH_INODES);
 2200         FREE_LOCK(&lk);
 2201         inodedep = malloc(sizeof(struct inodedep),
 2202                 M_INODEDEP, M_SOFTDEP_FLAGS);
 2203         workitem_alloc(&inodedep->id_list, D_INODEDEP, mp);
 2204         ACQUIRE_LOCK(&lk);
 2205         if (inodedep_find(inodedephd, fs, inum, inodedeppp)) {
 2206                 WORKITEM_FREE(inodedep, D_INODEDEP);
 2207                 return (1);
 2208         }
 2209         inodedep->id_fs = fs;
 2210         inodedep->id_ino = inum;
 2211         inodedep->id_state = ALLCOMPLETE;
 2212         inodedep->id_nlinkdelta = 0;
 2213         inodedep->id_savedino1 = NULL;
 2214         inodedep->id_savedsize = -1;
 2215         inodedep->id_savedextsize = -1;
 2216         inodedep->id_savednlink = -1;
 2217         inodedep->id_bmsafemap = NULL;
 2218         inodedep->id_mkdiradd = NULL;
 2219         LIST_INIT(&inodedep->id_dirremhd);
 2220         LIST_INIT(&inodedep->id_pendinghd);
 2221         LIST_INIT(&inodedep->id_inowait);
 2222         LIST_INIT(&inodedep->id_bufwait);
 2223         TAILQ_INIT(&inodedep->id_inoreflst);
 2224         TAILQ_INIT(&inodedep->id_inoupdt);
 2225         TAILQ_INIT(&inodedep->id_newinoupdt);
 2226         TAILQ_INIT(&inodedep->id_extupdt);
 2227         TAILQ_INIT(&inodedep->id_newextupdt);
 2228         TAILQ_INIT(&inodedep->id_freeblklst);
 2229         LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
 2230         *inodedeppp = inodedep;
 2231         return (0);
 2232 }
 2233 
 2234 /*
 2235  * Structures and routines associated with newblk caching.
 2236  */
 2237 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
 2238 u_long  newblk_hash;            /* size of hash table - 1 */
 2239 #define NEWBLK_HASH(fs, inum) \
 2240         (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
 2241 
 2242 static int
 2243 newblk_find(newblkhd, mp, newblkno, flags, newblkpp)
 2244         struct newblk_hashhead *newblkhd;
 2245         struct mount *mp;
 2246         ufs2_daddr_t newblkno;
 2247         int flags;
 2248         struct newblk **newblkpp;
 2249 {
 2250         struct newblk *newblk;
 2251 
 2252         LIST_FOREACH(newblk, newblkhd, nb_hash) {
 2253                 if (newblkno != newblk->nb_newblkno)
 2254                         continue;
 2255                 if (mp != newblk->nb_list.wk_mp)
 2256                         continue;
 2257                 /*
 2258                  * If we're creating a new dependency don't match those that
 2259                  * have already been converted to allocdirects.  This is for
 2260                  * a frag extend.
 2261                  */
 2262                 if ((flags & DEPALLOC) && newblk->nb_list.wk_type != D_NEWBLK)
 2263                         continue;
 2264                 break;
 2265         }
 2266         if (newblk) {
 2267                 *newblkpp = newblk;
 2268                 return (1);
 2269         }
 2270         *newblkpp = NULL;
 2271         return (0);
 2272 }
 2273 
 2274 /*
 2275  * Look up a newblk. Return 1 if found, 0 if not found.
 2276  * If not found, allocate if DEPALLOC flag is passed.
 2277  * Found or allocated entry is returned in newblkpp.
 2278  */
 2279 static int
 2280 newblk_lookup(mp, newblkno, flags, newblkpp)
 2281         struct mount *mp;
 2282         ufs2_daddr_t newblkno;
 2283         int flags;
 2284         struct newblk **newblkpp;
 2285 {
 2286         struct newblk *newblk;
 2287         struct newblk_hashhead *newblkhd;
 2288 
 2289         newblkhd = NEWBLK_HASH(VFSTOUFS(mp)->um_fs, newblkno);
 2290         if (newblk_find(newblkhd, mp, newblkno, flags, newblkpp))
 2291                 return (1);
 2292         if ((flags & DEPALLOC) == 0)
 2293                 return (0);
 2294         FREE_LOCK(&lk);
 2295         newblk = malloc(sizeof(union allblk), M_NEWBLK,
 2296             M_SOFTDEP_FLAGS | M_ZERO);
 2297         workitem_alloc(&newblk->nb_list, D_NEWBLK, mp);
 2298         ACQUIRE_LOCK(&lk);
 2299         if (newblk_find(newblkhd, mp, newblkno, flags, newblkpp)) {
 2300                 WORKITEM_FREE(newblk, D_NEWBLK);
 2301                 return (1);
 2302         }
 2303         newblk->nb_freefrag = NULL;
 2304         LIST_INIT(&newblk->nb_indirdeps);
 2305         LIST_INIT(&newblk->nb_newdirblk);
 2306         LIST_INIT(&newblk->nb_jwork);
 2307         newblk->nb_state = ATTACHED;
 2308         newblk->nb_newblkno = newblkno;
 2309         LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
 2310         *newblkpp = newblk;
 2311         return (0);
 2312 }
 2313 
 2314 /*
 2315  * Structures and routines associated with freed indirect block caching.
 2316  */
 2317 struct freeworklst *indir_hashtbl;
 2318 u_long  indir_hash;             /* size of hash table - 1 */
 2319 #define INDIR_HASH(mp, blkno) \
 2320         (&indir_hashtbl[((((register_t)(mp)) >> 13) + (blkno)) & indir_hash])
 2321 
 2322 /*
 2323  * Lookup an indirect block in the indir hash table.  The freework is
 2324  * removed and potentially freed.  The caller must do a blocking journal
 2325  * write before writing to the blkno.
 2326  */
 2327 static int
 2328 indirblk_lookup(mp, blkno)
 2329         struct mount *mp;
 2330         ufs2_daddr_t blkno;
 2331 {
 2332         struct freework *freework;
 2333         struct freeworklst *wkhd;
 2334 
 2335         wkhd = INDIR_HASH(mp, blkno);
 2336         TAILQ_FOREACH(freework, wkhd, fw_next) {
 2337                 if (freework->fw_blkno != blkno)
 2338                         continue;
 2339                 if (freework->fw_list.wk_mp != mp)
 2340                         continue;
 2341                 indirblk_remove(freework);
 2342                 return (1);
 2343         }
 2344         return (0);
 2345 }
 2346 
 2347 /*
 2348  * Insert an indirect block represented by freework into the indirblk
 2349  * hash table so that it may prevent the block from being re-used prior
 2350  * to the journal being written.
 2351  */
 2352 static void
 2353 indirblk_insert(freework)
 2354         struct freework *freework;
 2355 {
 2356         struct jblocks *jblocks;
 2357         struct jseg *jseg;
 2358 
 2359         jblocks = VFSTOUFS(freework->fw_list.wk_mp)->softdep_jblocks;
 2360         jseg = TAILQ_LAST(&jblocks->jb_segs, jseglst);
 2361         if (jseg == NULL)
 2362                 return;
 2363         
 2364         LIST_INSERT_HEAD(&jseg->js_indirs, freework, fw_segs);
 2365         TAILQ_INSERT_HEAD(INDIR_HASH(freework->fw_list.wk_mp,
 2366             freework->fw_blkno), freework, fw_next);
 2367         freework->fw_state &= ~DEPCOMPLETE;
 2368 }
 2369 
 2370 static void
 2371 indirblk_remove(freework)
 2372         struct freework *freework;
 2373 {
 2374 
 2375         LIST_REMOVE(freework, fw_segs);
 2376         TAILQ_REMOVE(INDIR_HASH(freework->fw_list.wk_mp,
 2377             freework->fw_blkno), freework, fw_next);
 2378         freework->fw_state |= DEPCOMPLETE;
 2379         if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE)
 2380                 WORKITEM_FREE(freework, D_FREEWORK);
 2381 }
 2382 
 2383 /*
 2384  * Executed during filesystem system initialization before
 2385  * mounting any filesystems.
 2386  */
 2387 void 
 2388 softdep_initialize()
 2389 {
 2390         int i;
 2391 
 2392         LIST_INIT(&mkdirlisthd);
 2393         max_softdeps = desiredvnodes * 4;
 2394         pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, &pagedep_hash);
 2395         inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
 2396         newblk_hashtbl = hashinit(max_softdeps / 2,  M_NEWBLK, &newblk_hash);
 2397         bmsafemap_hashtbl = hashinit(1024, M_BMSAFEMAP, &bmsafemap_hash);
 2398         i = 1 << (ffs(desiredvnodes / 10) - 1);
 2399         indir_hashtbl = malloc(i * sizeof(indir_hashtbl[0]), M_FREEWORK,
 2400             M_WAITOK);
 2401         indir_hash = i - 1;
 2402         for (i = 0; i <= indir_hash; i++)
 2403                 TAILQ_INIT(&indir_hashtbl[i]);
 2404 
 2405         /* initialise bioops hack */
 2406         bioops.io_start = softdep_disk_io_initiation;
 2407         bioops.io_complete = softdep_disk_write_complete;
 2408         bioops.io_deallocate = softdep_deallocate_dependencies;
 2409         bioops.io_countdeps = softdep_count_dependencies;
 2410 
 2411         /* Initialize the callout with an mtx. */
 2412         callout_init_mtx(&softdep_callout, &lk, 0);
 2413 }
 2414 
 2415 /*
 2416  * Executed after all filesystems have been unmounted during
 2417  * filesystem module unload.
 2418  */
 2419 void
 2420 softdep_uninitialize()
 2421 {
 2422 
 2423         callout_drain(&softdep_callout);
 2424         hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash);
 2425         hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash);
 2426         hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash);
 2427         hashdestroy(bmsafemap_hashtbl, M_BMSAFEMAP, bmsafemap_hash);
 2428         free(indir_hashtbl, M_FREEWORK);
 2429 }
 2430 
 2431 /*
 2432  * Called at mount time to notify the dependency code that a
 2433  * filesystem wishes to use it.
 2434  */
 2435 int
 2436 softdep_mount(devvp, mp, fs, cred)
 2437         struct vnode *devvp;
 2438         struct mount *mp;
 2439         struct fs *fs;
 2440         struct ucred *cred;
 2441 {
 2442         struct csum_total cstotal;
 2443         struct ufsmount *ump;
 2444         struct cg *cgp;
 2445         struct buf *bp;
 2446         int error, cyl;
 2447 
 2448         MNT_ILOCK(mp);
 2449         mp->mnt_flag = (mp->mnt_flag & ~MNT_ASYNC) | MNT_SOFTDEP;
 2450         if ((mp->mnt_kern_flag & MNTK_SOFTDEP) == 0) {
 2451                 mp->mnt_kern_flag = (mp->mnt_kern_flag & ~MNTK_ASYNC) | 
 2452                         MNTK_SOFTDEP | MNTK_NOASYNC;
 2453         }
 2454         MNT_IUNLOCK(mp);
 2455         ump = VFSTOUFS(mp);
 2456         LIST_INIT(&ump->softdep_workitem_pending);
 2457         LIST_INIT(&ump->softdep_journal_pending);
 2458         TAILQ_INIT(&ump->softdep_unlinked);
 2459         LIST_INIT(&ump->softdep_dirtycg);
 2460         ump->softdep_worklist_tail = NULL;
 2461         ump->softdep_on_worklist = 0;
 2462         ump->softdep_deps = 0;
 2463         if ((fs->fs_flags & FS_SUJ) &&
 2464             (error = journal_mount(mp, fs, cred)) != 0) {
 2465                 printf("Failed to start journal: %d\n", error);
 2466                 return (error);
 2467         }
 2468         /*
 2469          * When doing soft updates, the counters in the
 2470          * superblock may have gotten out of sync. Recomputation
 2471          * can take a long time and can be deferred for background
 2472          * fsck.  However, the old behavior of scanning the cylinder
 2473          * groups and recalculating them at mount time is available
 2474          * by setting vfs.ffs.compute_summary_at_mount to one.
 2475          */
 2476         if (compute_summary_at_mount == 0 || fs->fs_clean != 0)
 2477                 return (0);
 2478         bzero(&cstotal, sizeof cstotal);
 2479         for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
 2480                 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
 2481                     fs->fs_cgsize, cred, &bp)) != 0) {
 2482                         brelse(bp);
 2483                         return (error);
 2484                 }
 2485                 cgp = (struct cg *)bp->b_data;
 2486                 cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
 2487                 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
 2488                 cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
 2489                 cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
 2490                 fs->fs_cs(fs, cyl) = cgp->cg_cs;
 2491                 brelse(bp);
 2492         }
 2493 #ifdef DEBUG
 2494         if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
 2495                 printf("%s: superblock summary recomputed\n", fs->fs_fsmnt);
 2496 #endif
 2497         bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
 2498         return (0);
 2499 }
 2500 
 2501 void
 2502 softdep_unmount(mp)
 2503         struct mount *mp;
 2504 {
 2505 
 2506         MNT_ILOCK(mp);
 2507         mp->mnt_flag &= ~MNT_SOFTDEP;
 2508         if (MOUNTEDSUJ(mp) == 0) {
 2509                 MNT_IUNLOCK(mp);
 2510                 return;
 2511         }
 2512         mp->mnt_flag &= ~MNT_SUJ;
 2513         MNT_IUNLOCK(mp);
 2514         journal_unmount(mp);
 2515 }
 2516 
 2517 static struct jblocks *
 2518 jblocks_create(void)
 2519 {
 2520         struct jblocks *jblocks;
 2521 
 2522         jblocks = malloc(sizeof(*jblocks), M_JBLOCKS, M_WAITOK | M_ZERO);
 2523         TAILQ_INIT(&jblocks->jb_segs);
 2524         jblocks->jb_avail = 10;
 2525         jblocks->jb_extent = malloc(sizeof(struct jextent) * jblocks->jb_avail,
 2526             M_JBLOCKS, M_WAITOK | M_ZERO);
 2527 
 2528         return (jblocks);
 2529 }
 2530 
 2531 static ufs2_daddr_t
 2532 jblocks_alloc(jblocks, bytes, actual)
 2533         struct jblocks *jblocks;
 2534         int bytes;
 2535         int *actual;
 2536 {
 2537         ufs2_daddr_t daddr;
 2538         struct jextent *jext;
 2539         int freecnt;
 2540         int blocks;
 2541 
 2542         blocks = bytes / DEV_BSIZE;
 2543         jext = &jblocks->jb_extent[jblocks->jb_head];
 2544         freecnt = jext->je_blocks - jblocks->jb_off;
 2545         if (freecnt == 0) {
 2546                 jblocks->jb_off = 0;
 2547                 if (++jblocks->jb_head > jblocks->jb_used)
 2548                         jblocks->jb_head = 0;
 2549                 jext = &jblocks->jb_extent[jblocks->jb_head];
 2550                 freecnt = jext->je_blocks;
 2551         }
 2552         if (freecnt > blocks)
 2553                 freecnt = blocks;
 2554         *actual = freecnt * DEV_BSIZE;
 2555         daddr = jext->je_daddr + jblocks->jb_off;
 2556         jblocks->jb_off += freecnt;
 2557         jblocks->jb_free -= freecnt;
 2558 
 2559         return (daddr);
 2560 }
 2561 
 2562 static void
 2563 jblocks_free(jblocks, mp, bytes)
 2564         struct jblocks *jblocks;
 2565         struct mount *mp;
 2566         int bytes;
 2567 {
 2568 
 2569         jblocks->jb_free += bytes / DEV_BSIZE;
 2570         if (jblocks->jb_suspended)
 2571                 worklist_speedup();
 2572         wakeup(jblocks);
 2573 }
 2574 
 2575 static void
 2576 jblocks_destroy(jblocks)
 2577         struct jblocks *jblocks;
 2578 {
 2579 
 2580         if (jblocks->jb_extent)
 2581                 free(jblocks->jb_extent, M_JBLOCKS);
 2582         free(jblocks, M_JBLOCKS);
 2583 }
 2584 
 2585 static void
 2586 jblocks_add(jblocks, daddr, blocks)
 2587         struct jblocks *jblocks;
 2588         ufs2_daddr_t daddr;
 2589         int blocks;
 2590 {
 2591         struct jextent *jext;
 2592 
 2593         jblocks->jb_blocks += blocks;
 2594         jblocks->jb_free += blocks;
 2595         jext = &jblocks->jb_extent[jblocks->jb_used];
 2596         /* Adding the first block. */
 2597         if (jext->je_daddr == 0) {
 2598                 jext->je_daddr = daddr;
 2599                 jext->je_blocks = blocks;
 2600                 return;
 2601         }
 2602         /* Extending the last extent. */
 2603         if (jext->je_daddr + jext->je_blocks == daddr) {
 2604                 jext->je_blocks += blocks;
 2605                 return;
 2606         }
 2607         /* Adding a new extent. */
 2608         if (++jblocks->jb_used == jblocks->jb_avail) {
 2609                 jblocks->jb_avail *= 2;
 2610                 jext = malloc(sizeof(struct jextent) * jblocks->jb_avail,
 2611                     M_JBLOCKS, M_WAITOK | M_ZERO);
 2612                 memcpy(jext, jblocks->jb_extent,
 2613                     sizeof(struct jextent) * jblocks->jb_used);
 2614                 free(jblocks->jb_extent, M_JBLOCKS);
 2615                 jblocks->jb_extent = jext;
 2616         }
 2617         jext = &jblocks->jb_extent[jblocks->jb_used];
 2618         jext->je_daddr = daddr;
 2619         jext->je_blocks = blocks;
 2620         return;
 2621 }
 2622 
 2623 int
 2624 softdep_journal_lookup(mp, vpp)
 2625         struct mount *mp;
 2626         struct vnode **vpp;
 2627 {
 2628         struct componentname cnp;
 2629         struct vnode *dvp;
 2630         ino_t sujournal;
 2631         int error;
 2632 
 2633         error = VFS_VGET(mp, ROOTINO, LK_EXCLUSIVE, &dvp);
 2634         if (error)
 2635                 return (error);
 2636         bzero(&cnp, sizeof(cnp));
 2637         cnp.cn_nameiop = LOOKUP;
 2638         cnp.cn_flags = ISLASTCN;
 2639         cnp.cn_thread = curthread;
 2640         cnp.cn_cred = curthread->td_ucred;
 2641         cnp.cn_pnbuf = SUJ_FILE;
 2642         cnp.cn_nameptr = SUJ_FILE;
 2643         cnp.cn_namelen = strlen(SUJ_FILE);
 2644         error = ufs_lookup_ino(dvp, NULL, &cnp, &sujournal);
 2645         vput(dvp);
 2646         if (error != 0)
 2647                 return (error);
 2648         error = VFS_VGET(mp, sujournal, LK_EXCLUSIVE, vpp);
 2649         return (error);
 2650 }
 2651 
 2652 /*
 2653  * Open and verify the journal file.
 2654  */
 2655 static int
 2656 journal_mount(mp, fs, cred)
 2657         struct mount *mp;
 2658         struct fs *fs;
 2659         struct ucred *cred;
 2660 {
 2661         struct jblocks *jblocks;
 2662         struct vnode *vp;
 2663         struct inode *ip;
 2664         ufs2_daddr_t blkno;
 2665         int bcount;
 2666         int error;
 2667         int i;
 2668 
 2669         error = softdep_journal_lookup(mp, &vp);
 2670         if (error != 0) {
 2671                 printf("Failed to find journal.  Use tunefs to create one\n");
 2672                 return (error);
 2673         }
 2674         ip = VTOI(vp);
 2675         if (ip->i_size < SUJ_MIN) {
 2676                 error = ENOSPC;
 2677                 goto out;
 2678         }
 2679         bcount = lblkno(fs, ip->i_size);        /* Only use whole blocks. */
 2680         jblocks = jblocks_create();
 2681         for (i = 0; i < bcount; i++) {
 2682                 error = ufs_bmaparray(vp, i, &blkno, NULL, NULL, NULL);
 2683                 if (error)
 2684                         break;
 2685                 jblocks_add(jblocks, blkno, fsbtodb(fs, fs->fs_frag));
 2686         }
 2687         if (error) {
 2688                 jblocks_destroy(jblocks);
 2689                 goto out;
 2690         }
 2691         jblocks->jb_low = jblocks->jb_free / 3; /* Reserve 33%. */
 2692         jblocks->jb_min = jblocks->jb_free / 10; /* Suspend at 10%. */
 2693         VFSTOUFS(mp)->softdep_jblocks = jblocks;
 2694 out:
 2695         if (error == 0) {
 2696                 MNT_ILOCK(mp);
 2697                 mp->mnt_flag |= MNT_SUJ;
 2698                 mp->mnt_flag &= ~MNT_SOFTDEP;
 2699                 MNT_IUNLOCK(mp);
 2700                 /*
 2701                  * Only validate the journal contents if the
 2702                  * filesystem is clean, otherwise we write the logs
 2703                  * but they'll never be used.  If the filesystem was
 2704                  * still dirty when we mounted it the journal is
 2705                  * invalid and a new journal can only be valid if it
 2706                  * starts from a clean mount.
 2707                  */
 2708                 if (fs->fs_clean) {
 2709                         DIP_SET(ip, i_modrev, fs->fs_mtime);
 2710                         ip->i_flags |= IN_MODIFIED;
 2711                         ffs_update(vp, 1);
 2712                 }
 2713         }
 2714         vput(vp);
 2715         return (error);
 2716 }
 2717 
 2718 static void
 2719 journal_unmount(mp)
 2720         struct mount *mp;
 2721 {
 2722         struct ufsmount *ump;
 2723 
 2724         ump = VFSTOUFS(mp);
 2725         if (ump->softdep_jblocks)
 2726                 jblocks_destroy(ump->softdep_jblocks);
 2727         ump->softdep_jblocks = NULL;
 2728 }
 2729 
 2730 /*
 2731  * Called when a journal record is ready to be written.  Space is allocated
 2732  * and the journal entry is created when the journal is flushed to stable
 2733  * store.
 2734  */
 2735 static void
 2736 add_to_journal(wk)
 2737         struct worklist *wk;
 2738 {
 2739         struct ufsmount *ump;
 2740 
 2741         rw_assert(&lk, RA_WLOCKED);
 2742         ump = VFSTOUFS(wk->wk_mp);
 2743         if (wk->wk_state & ONWORKLIST)
 2744                 panic("add_to_journal: %s(0x%X) already on list",
 2745                     TYPENAME(wk->wk_type), wk->wk_state);
 2746         wk->wk_state |= ONWORKLIST | DEPCOMPLETE;
 2747         if (LIST_EMPTY(&ump->softdep_journal_pending)) {
 2748                 ump->softdep_jblocks->jb_age = ticks;
 2749                 LIST_INSERT_HEAD(&ump->softdep_journal_pending, wk, wk_list);
 2750         } else
 2751                 LIST_INSERT_AFTER(ump->softdep_journal_tail, wk, wk_list);
 2752         ump->softdep_journal_tail = wk;
 2753         ump->softdep_on_journal += 1;
 2754 }
 2755 
 2756 /*
 2757  * Remove an arbitrary item for the journal worklist maintain the tail
 2758  * pointer.  This happens when a new operation obviates the need to
 2759  * journal an old operation.
 2760  */
 2761 static void
 2762 remove_from_journal(wk)
 2763         struct worklist *wk;
 2764 {
 2765         struct ufsmount *ump;
 2766 
 2767         rw_assert(&lk, RA_WLOCKED);
 2768         ump = VFSTOUFS(wk->wk_mp);
 2769 #ifdef SUJ_DEBUG
 2770         {
 2771                 struct worklist *wkn;
 2772 
 2773                 LIST_FOREACH(wkn, &ump->softdep_journal_pending, wk_list)
 2774                         if (wkn == wk)
 2775                                 break;
 2776                 if (wkn == NULL)
 2777                         panic("remove_from_journal: %p is not in journal", wk);
 2778         }
 2779 #endif
 2780         /*
 2781          * We emulate a TAILQ to save space in most structures which do not
 2782          * require TAILQ semantics.  Here we must update the tail position
 2783          * when removing the tail which is not the final entry. This works
 2784          * only if the worklist linkage are at the beginning of the structure.
 2785          */
 2786         if (ump->softdep_journal_tail == wk)
 2787                 ump->softdep_journal_tail =
 2788                     (struct worklist *)wk->wk_list.le_prev;
 2789 
 2790         WORKLIST_REMOVE(wk);
 2791         ump->softdep_on_journal -= 1;
 2792 }
 2793 
 2794 /*
 2795  * Check for journal space as well as dependency limits so the prelink
 2796  * code can throttle both journaled and non-journaled filesystems.
 2797  * Threshold is 0 for low and 1 for min.
 2798  */
 2799 static int
 2800 journal_space(ump, thresh)
 2801         struct ufsmount *ump;
 2802         int thresh;
 2803 {
 2804         struct jblocks *jblocks;
 2805         int avail;
 2806 
 2807         jblocks = ump->softdep_jblocks;
 2808         if (jblocks == NULL)
 2809                 return (1);
 2810         /*
 2811          * We use a tighter restriction here to prevent request_cleanup()
 2812          * running in threads from running into locks we currently hold.
 2813          */
 2814         if (dep_current[D_INODEDEP] > (max_softdeps / 10) * 9)
 2815                 return (0);
 2816         if (thresh)
 2817                 thresh = jblocks->jb_min;
 2818         else
 2819                 thresh = jblocks->jb_low;
 2820         avail = (ump->softdep_on_journal * JREC_SIZE) / DEV_BSIZE;
 2821         avail = jblocks->jb_free - avail;
 2822 
 2823         return (avail > thresh);
 2824 }
 2825 
 2826 static void
 2827 journal_suspend(ump)
 2828         struct ufsmount *ump;
 2829 {
 2830         struct jblocks *jblocks;
 2831         struct mount *mp;
 2832 
 2833         mp = UFSTOVFS(ump);
 2834         jblocks = ump->softdep_jblocks;
 2835         MNT_ILOCK(mp);
 2836         if ((mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
 2837                 stat_journal_min++;
 2838                 mp->mnt_kern_flag |= MNTK_SUSPEND;
 2839                 mp->mnt_susp_owner = FIRST_THREAD_IN_PROC(softdepproc);
 2840         }
 2841         jblocks->jb_suspended = 1;
 2842         MNT_IUNLOCK(mp);
 2843 }
 2844 
 2845 static int
 2846 journal_unsuspend(struct ufsmount *ump)
 2847 {
 2848         struct jblocks *jblocks;
 2849         struct mount *mp;
 2850 
 2851         mp = UFSTOVFS(ump);
 2852         jblocks = ump->softdep_jblocks;
 2853 
 2854         if (jblocks != NULL && jblocks->jb_suspended &&
 2855             journal_space(ump, jblocks->jb_min)) {
 2856                 jblocks->jb_suspended = 0;
 2857                 FREE_LOCK(&lk);
 2858                 mp->mnt_susp_owner = curthread;
 2859                 vfs_write_resume(mp, 0);
 2860                 ACQUIRE_LOCK(&lk);
 2861                 return (1);
 2862         }
 2863         return (0);
 2864 }
 2865 
 2866 /*
 2867  * Called before any allocation function to be certain that there is
 2868  * sufficient space in the journal prior to creating any new records.
 2869  * Since in the case of block allocation we may have multiple locked
 2870  * buffers at the time of the actual allocation we can not block
 2871  * when the journal records are created.  Doing so would create a deadlock
 2872  * if any of these buffers needed to be flushed to reclaim space.  Instead
 2873  * we require a sufficiently large amount of available space such that
 2874  * each thread in the system could have passed this allocation check and
 2875  * still have sufficient free space.  With 20% of a minimum journal size
 2876  * of 1MB we have 6553 records available.
 2877  */
 2878 int
 2879 softdep_prealloc(vp, waitok)
 2880         struct vnode *vp;
 2881         int waitok;
 2882 {
 2883         struct ufsmount *ump;
 2884 
 2885         /*
 2886          * Nothing to do if we are not running journaled soft updates.
 2887          * If we currently hold the snapshot lock, we must avoid handling
 2888          * other resources that could cause deadlock.
 2889          */
 2890         if (DOINGSUJ(vp) == 0 || IS_SNAPSHOT(VTOI(vp)))
 2891                 return (0);
 2892         ump = VFSTOUFS(vp->v_mount);
 2893         ACQUIRE_LOCK(&lk);
 2894         if (journal_space(ump, 0)) {
 2895                 FREE_LOCK(&lk);
 2896                 return (0);
 2897         }
 2898         stat_journal_low++;
 2899         FREE_LOCK(&lk);
 2900         if (waitok == MNT_NOWAIT)
 2901                 return (ENOSPC);
 2902         /*
 2903          * Attempt to sync this vnode once to flush any journal
 2904          * work attached to it.
 2905          */
 2906         if ((curthread->td_pflags & TDP_COWINPROGRESS) == 0)
 2907                 ffs_syncvnode(vp, waitok, 0);
 2908         ACQUIRE_LOCK(&lk);
 2909         process_removes(vp);
 2910         process_truncates(vp);
 2911         if (journal_space(ump, 0) == 0) {
 2912                 softdep_speedup();
 2913                 if (journal_space(ump, 1) == 0)
 2914                         journal_suspend(ump);
 2915         }
 2916         FREE_LOCK(&lk);
 2917 
 2918         return (0);
 2919 }
 2920 
 2921 /*
 2922  * Before adjusting a link count on a vnode verify that we have sufficient
 2923  * journal space.  If not, process operations that depend on the currently
 2924  * locked pair of vnodes to try to flush space as the syncer, buf daemon,
 2925  * and softdep flush threads can not acquire these locks to reclaim space.
 2926  */
 2927 static void
 2928 softdep_prelink(dvp, vp)
 2929         struct vnode *dvp;
 2930         struct vnode *vp;
 2931 {
 2932         struct ufsmount *ump;
 2933 
 2934         ump = VFSTOUFS(dvp->v_mount);
 2935         rw_assert(&lk, RA_WLOCKED);
 2936         /*
 2937          * Nothing to do if we have sufficient journal space.
 2938          * If we currently hold the snapshot lock, we must avoid
 2939          * handling other resources that could cause deadlock.
 2940          */
 2941         if (journal_space(ump, 0) || (vp && IS_SNAPSHOT(VTOI(vp))))
 2942                 return;
 2943         stat_journal_low++;
 2944         FREE_LOCK(&lk);
 2945         if (vp)
 2946                 ffs_syncvnode(vp, MNT_NOWAIT, 0);
 2947         ffs_syncvnode(dvp, MNT_WAIT, 0);
 2948         ACQUIRE_LOCK(&lk);
 2949         /* Process vp before dvp as it may create .. removes. */
 2950         if (vp) {
 2951                 process_removes(vp);
 2952                 process_truncates(vp);
 2953         }
 2954         process_removes(dvp);
 2955         process_truncates(dvp);
 2956         softdep_speedup();
 2957         process_worklist_item(UFSTOVFS(ump), 2, LK_NOWAIT);
 2958         if (journal_space(ump, 0) == 0) {
 2959                 softdep_speedup();
 2960                 if (journal_space(ump, 1) == 0)
 2961                         journal_suspend(ump);
 2962         }
 2963 }
 2964 
 2965 static void
 2966 jseg_write(ump, jseg, data)
 2967         struct ufsmount *ump;
 2968         struct jseg *jseg;
 2969         uint8_t *data;
 2970 {
 2971         struct jsegrec *rec;
 2972 
 2973         rec = (struct jsegrec *)data;
 2974         rec->jsr_seq = jseg->js_seq;
 2975         rec->jsr_oldest = jseg->js_oldseq;
 2976         rec->jsr_cnt = jseg->js_cnt;
 2977         rec->jsr_blocks = jseg->js_size / ump->um_devvp->v_bufobj.bo_bsize;
 2978         rec->jsr_crc = 0;
 2979         rec->jsr_time = ump->um_fs->fs_mtime;
 2980 }
 2981 
 2982 static inline void
 2983 inoref_write(inoref, jseg, rec)
 2984         struct inoref *inoref;
 2985         struct jseg *jseg;
 2986         struct jrefrec *rec;
 2987 {
 2988 
 2989         inoref->if_jsegdep->jd_seg = jseg;
 2990         rec->jr_ino = inoref->if_ino;
 2991         rec->jr_parent = inoref->if_parent;
 2992         rec->jr_nlink = inoref->if_nlink;
 2993         rec->jr_mode = inoref->if_mode;
 2994         rec->jr_diroff = inoref->if_diroff;
 2995 }
 2996 
 2997 static void
 2998 jaddref_write(jaddref, jseg, data)
 2999         struct jaddref *jaddref;
 3000         struct jseg *jseg;
 3001         uint8_t *data;
 3002 {
 3003         struct jrefrec *rec;
 3004 
 3005         rec = (struct jrefrec *)data;
 3006         rec->jr_op = JOP_ADDREF;
 3007         inoref_write(&jaddref->ja_ref, jseg, rec);
 3008 }
 3009 
 3010 static void
 3011 jremref_write(jremref, jseg, data)
 3012         struct jremref *jremref;
 3013         struct jseg *jseg;
 3014         uint8_t *data;
 3015 {
 3016         struct jrefrec *rec;
 3017 
 3018         rec = (struct jrefrec *)data;
 3019         rec->jr_op = JOP_REMREF;
 3020         inoref_write(&jremref->jr_ref, jseg, rec);
 3021 }
 3022 
 3023 static void
 3024 jmvref_write(jmvref, jseg, data)
 3025         struct jmvref *jmvref;
 3026         struct jseg *jseg;
 3027         uint8_t *data;
 3028 {
 3029         struct jmvrec *rec;
 3030 
 3031         rec = (struct jmvrec *)data;
 3032         rec->jm_op = JOP_MVREF;
 3033         rec->jm_ino = jmvref->jm_ino;
 3034         rec->jm_parent = jmvref->jm_parent;
 3035         rec->jm_oldoff = jmvref->jm_oldoff;
 3036         rec->jm_newoff = jmvref->jm_newoff;
 3037 }
 3038 
 3039 static void
 3040 jnewblk_write(jnewblk, jseg, data)
 3041         struct jnewblk *jnewblk;
 3042         struct jseg *jseg;
 3043         uint8_t *data;
 3044 {
 3045         struct jblkrec *rec;
 3046 
 3047         jnewblk->jn_jsegdep->jd_seg = jseg;
 3048         rec = (struct jblkrec *)data;
 3049         rec->jb_op = JOP_NEWBLK;
 3050         rec->jb_ino = jnewblk->jn_ino;
 3051         rec->jb_blkno = jnewblk->jn_blkno;
 3052         rec->jb_lbn = jnewblk->jn_lbn;
 3053         rec->jb_frags = jnewblk->jn_frags;
 3054         rec->jb_oldfrags = jnewblk->jn_oldfrags;
 3055 }
 3056 
 3057 static void
 3058 jfreeblk_write(jfreeblk, jseg, data)
 3059         struct jfreeblk *jfreeblk;
 3060         struct jseg *jseg;
 3061         uint8_t *data;
 3062 {
 3063         struct jblkrec *rec;
 3064 
 3065         jfreeblk->jf_dep.jb_jsegdep->jd_seg = jseg;
 3066         rec = (struct jblkrec *)data;
 3067         rec->jb_op = JOP_FREEBLK;
 3068         rec->jb_ino = jfreeblk->jf_ino;
 3069         rec->jb_blkno = jfreeblk->jf_blkno;
 3070         rec->jb_lbn = jfreeblk->jf_lbn;
 3071         rec->jb_frags = jfreeblk->jf_frags;
 3072         rec->jb_oldfrags = 0;
 3073 }
 3074 
 3075 static void
 3076 jfreefrag_write(jfreefrag, jseg, data)
 3077         struct jfreefrag *jfreefrag;
 3078         struct jseg *jseg;
 3079         uint8_t *data;
 3080 {
 3081         struct jblkrec *rec;
 3082 
 3083         jfreefrag->fr_jsegdep->jd_seg = jseg;
 3084         rec = (struct jblkrec *)data;
 3085         rec->jb_op = JOP_FREEBLK;
 3086         rec->jb_ino = jfreefrag->fr_ino;
 3087         rec->jb_blkno = jfreefrag->fr_blkno;
 3088         rec->jb_lbn = jfreefrag->fr_lbn;
 3089         rec->jb_frags = jfreefrag->fr_frags;
 3090         rec->jb_oldfrags = 0;
 3091 }
 3092 
 3093 static void
 3094 jtrunc_write(jtrunc, jseg, data)
 3095         struct jtrunc *jtrunc;
 3096         struct jseg *jseg;
 3097         uint8_t *data;
 3098 {
 3099         struct jtrncrec *rec;
 3100 
 3101         jtrunc->jt_dep.jb_jsegdep->jd_seg = jseg;
 3102         rec = (struct jtrncrec *)data;
 3103         rec->jt_op = JOP_TRUNC;
 3104         rec->jt_ino = jtrunc->jt_ino;
 3105         rec->jt_size = jtrunc->jt_size;
 3106         rec->jt_extsize = jtrunc->jt_extsize;
 3107 }
 3108 
 3109 static void
 3110 jfsync_write(jfsync, jseg, data)
 3111         struct jfsync *jfsync;
 3112         struct jseg *jseg;
 3113         uint8_t *data;
 3114 {
 3115         struct jtrncrec *rec;
 3116 
 3117         rec = (struct jtrncrec *)data;
 3118         rec->jt_op = JOP_SYNC;
 3119         rec->jt_ino = jfsync->jfs_ino;
 3120         rec->jt_size = jfsync->jfs_size;
 3121         rec->jt_extsize = jfsync->jfs_extsize;
 3122 }
 3123 
 3124 static void
 3125 softdep_flushjournal(mp)
 3126         struct mount *mp;
 3127 {
 3128         struct jblocks *jblocks;
 3129         struct ufsmount *ump;
 3130 
 3131         if (MOUNTEDSUJ(mp) == 0)
 3132                 return;
 3133         ump = VFSTOUFS(mp);
 3134         jblocks = ump->softdep_jblocks;
 3135         ACQUIRE_LOCK(&lk);
 3136         while (ump->softdep_on_journal) {
 3137                 jblocks->jb_needseg = 1;
 3138                 softdep_process_journal(mp, NULL, MNT_WAIT);
 3139         }
 3140         FREE_LOCK(&lk);
 3141 }
 3142 
 3143 static void softdep_synchronize_completed(struct bio *);
 3144 static void softdep_synchronize(struct bio *, struct ufsmount *, void *);
 3145 
 3146 static void
 3147 softdep_synchronize_completed(bp)
 3148         struct bio *bp;
 3149 {
 3150         struct jseg *oldest;
 3151         struct jseg *jseg;
 3152 
 3153         /*
 3154          * caller1 marks the last segment written before we issued the
 3155          * synchronize cache.
 3156          */
 3157         jseg = bp->bio_caller1;
 3158         oldest = NULL;
 3159         ACQUIRE_LOCK(&lk);
 3160         /*
 3161          * Mark all the journal entries waiting on the synchronize cache
 3162          * as completed so they may continue on.
 3163          */
 3164         while (jseg != NULL && (jseg->js_state & COMPLETE) == 0) {
 3165                 jseg->js_state |= COMPLETE;
 3166                 oldest = jseg;
 3167                 jseg = TAILQ_PREV(jseg, jseglst, js_next);
 3168         }
 3169         /*
 3170          * Restart deferred journal entry processing from the oldest
 3171          * completed jseg.
 3172          */
 3173         if (oldest)
 3174                 complete_jsegs(oldest);
 3175 
 3176         FREE_LOCK(&lk);
 3177         g_destroy_bio(bp);
 3178 }
 3179 
 3180 /*
 3181  * Send BIO_FLUSH/SYNCHRONIZE CACHE to the device to enforce write ordering
 3182  * barriers.  The journal must be written prior to any blocks that depend
 3183  * on it and the journal can not be released until the blocks have be
 3184  * written.  This code handles both barriers simultaneously.
 3185  */
 3186 static void
 3187 softdep_synchronize(bp, ump, caller1)
 3188         struct bio *bp;
 3189         struct ufsmount *ump;
 3190         void *caller1;
 3191 {
 3192 
 3193         bp->bio_cmd = BIO_FLUSH;
 3194         bp->bio_flags |= BIO_ORDERED;
 3195         bp->bio_data = NULL;
 3196         bp->bio_offset = ump->um_cp->provider->mediasize;
 3197         bp->bio_length = 0;
 3198         bp->bio_done = softdep_synchronize_completed;
 3199         bp->bio_caller1 = caller1;
 3200         g_io_request(bp,
 3201             (struct g_consumer *)ump->um_devvp->v_bufobj.bo_private);
 3202 }
 3203 
 3204 /*
 3205  * Flush some journal records to disk.
 3206  */
 3207 static void
 3208 softdep_process_journal(mp, needwk, flags)
 3209         struct mount *mp;
 3210         struct worklist *needwk;
 3211         int flags;
 3212 {
 3213         struct jblocks *jblocks;
 3214         struct ufsmount *ump;
 3215         struct worklist *wk;
 3216         struct jseg *jseg;
 3217         struct buf *bp;
 3218         struct bio *bio;
 3219         uint8_t *data;
 3220         struct fs *fs;
 3221         int shouldflush;
 3222         int segwritten;
 3223         int jrecmin;    /* Minimum records per block. */
 3224         int jrecmax;    /* Maximum records per block. */
 3225         int size;
 3226         int cnt;
 3227         int off;
 3228         int devbsize;
 3229 
 3230         if (MOUNTEDSUJ(mp) == 0)
 3231                 return;
 3232         shouldflush = softdep_flushcache;
 3233         bio = NULL;
 3234         jseg = NULL;
 3235         ump = VFSTOUFS(mp);
 3236         fs = ump->um_fs;
 3237         jblocks = ump->softdep_jblocks;
 3238         devbsize = ump->um_devvp->v_bufobj.bo_bsize;
 3239         /*
 3240          * We write anywhere between a disk block and fs block.  The upper
 3241          * bound is picked to prevent buffer cache fragmentation and limit
 3242          * processing time per I/O.
 3243          */
 3244         jrecmin = (devbsize / JREC_SIZE) - 1; /* -1 for seg header */
 3245         jrecmax = (fs->fs_bsize / devbsize) * jrecmin;
 3246         segwritten = 0;
 3247         for (;;) {
 3248                 cnt = ump->softdep_on_journal;
 3249                 /*
 3250                  * Criteria for writing a segment:
 3251                  * 1) We have a full block.
 3252                  * 2) We're called from jwait() and haven't found the
 3253                  *    journal item yet.
 3254                  * 3) Always write if needseg is set.
 3255                  * 4) If we are called from process_worklist and have
 3256                  *    not yet written anything we write a partial block
 3257                  *    to enforce a 1 second maximum latency on journal
 3258                  *    entries.
 3259                  */
 3260                 if (cnt < (jrecmax - 1) && needwk == NULL &&
 3261                     jblocks->jb_needseg == 0 && (segwritten || cnt == 0))
 3262                         break;
 3263                 cnt++;
 3264                 /*
 3265                  * Verify some free journal space.  softdep_prealloc() should
 3266                  * guarantee that we don't run out so this is indicative of
 3267                  * a problem with the flow control.  Try to recover
 3268                  * gracefully in any event.
 3269                  */
 3270                 while (jblocks->jb_free == 0) {
 3271                         if (flags != MNT_WAIT)
 3272                                 break;
 3273                         printf("softdep: Out of journal space!\n");
 3274                         softdep_speedup();
 3275                         msleep(jblocks, &lk, PRIBIO, "jblocks", hz);
 3276                 }
 3277                 FREE_LOCK(&lk);
 3278                 jseg = malloc(sizeof(*jseg), M_JSEG, M_SOFTDEP_FLAGS);
 3279                 workitem_alloc(&jseg->js_list, D_JSEG, mp);
 3280                 LIST_INIT(&jseg->js_entries);
 3281                 LIST_INIT(&jseg->js_indirs);
 3282                 jseg->js_state = ATTACHED;
 3283                 if (shouldflush == 0)
 3284                         jseg->js_state |= COMPLETE;
 3285                 else if (bio == NULL)
 3286                         bio = g_alloc_bio();
 3287                 jseg->js_jblocks = jblocks;
 3288                 bp = geteblk(fs->fs_bsize, 0);
 3289                 ACQUIRE_LOCK(&lk);
 3290                 /*
 3291                  * If there was a race while we were allocating the block
 3292                  * and jseg the entry we care about was likely written.
 3293                  * We bail out in both the WAIT and NOWAIT case and assume
 3294                  * the caller will loop if the entry it cares about is
 3295                  * not written.
 3296                  */
 3297                 cnt = ump->softdep_on_journal;
 3298                 if (cnt + jblocks->jb_needseg == 0 || jblocks->jb_free == 0) {
 3299                         bp->b_flags |= B_INVAL | B_NOCACHE;
 3300                         WORKITEM_FREE(jseg, D_JSEG);
 3301                         FREE_LOCK(&lk);
 3302                         brelse(bp);
 3303                         ACQUIRE_LOCK(&lk);
 3304                         break;
 3305                 }
 3306                 /*
 3307                  * Calculate the disk block size required for the available
 3308                  * records rounded to the min size.
 3309                  */
 3310                 if (cnt == 0)
 3311                         size = devbsize;
 3312                 else if (cnt < jrecmax)
 3313                         size = howmany(cnt, jrecmin) * devbsize;
 3314                 else
 3315                         size = fs->fs_bsize;
 3316                 /*
 3317                  * Allocate a disk block for this journal data and account
 3318                  * for truncation of the requested size if enough contiguous
 3319                  * space was not available.
 3320                  */
 3321                 bp->b_blkno = jblocks_alloc(jblocks, size, &size);
 3322                 bp->b_lblkno = bp->b_blkno;
 3323                 bp->b_offset = bp->b_blkno * DEV_BSIZE;
 3324                 bp->b_bcount = size;
 3325                 bp->b_flags &= ~B_INVAL;
 3326                 bp->b_flags |= B_VALIDSUSPWRT | B_NOCOPY;
 3327                 /*
 3328                  * Initialize our jseg with cnt records.  Assign the next
 3329                  * sequence number to it and link it in-order.
 3330                  */
 3331                 cnt = MIN(cnt, (size / devbsize) * jrecmin);
 3332                 jseg->js_buf = bp;
 3333                 jseg->js_cnt = cnt;
 3334                 jseg->js_refs = cnt + 1;        /* Self ref. */
 3335                 jseg->js_size = size;
 3336                 jseg->js_seq = jblocks->jb_nextseq++;
 3337                 if (jblocks->jb_oldestseg == NULL)
 3338                         jblocks->jb_oldestseg = jseg;
 3339                 jseg->js_oldseq = jblocks->jb_oldestseg->js_seq;
 3340                 TAILQ_INSERT_TAIL(&jblocks->jb_segs, jseg, js_next);
 3341                 if (jblocks->jb_writeseg == NULL)
 3342                         jblocks->jb_writeseg = jseg;
 3343                 /*
 3344                  * Start filling in records from the pending list.
 3345                  */
 3346                 data = bp->b_data;
 3347                 off = 0;
 3348                 while ((wk = LIST_FIRST(&ump->softdep_journal_pending))
 3349                     != NULL) {
 3350                         if (cnt == 0)
 3351                                 break;
 3352                         /* Place a segment header on every device block. */
 3353                         if ((off % devbsize) == 0) {
 3354                                 jseg_write(ump, jseg, data);
 3355                                 off += JREC_SIZE;
 3356                                 data = bp->b_data + off;
 3357                         }
 3358                         if (wk == needwk)
 3359                                 needwk = NULL;
 3360                         remove_from_journal(wk);
 3361                         wk->wk_state |= INPROGRESS;
 3362                         WORKLIST_INSERT(&jseg->js_entries, wk);
 3363                         switch (wk->wk_type) {
 3364                         case D_JADDREF:
 3365                                 jaddref_write(WK_JADDREF(wk), jseg, data);
 3366                                 break;
 3367                         case D_JREMREF:
 3368                                 jremref_write(WK_JREMREF(wk), jseg, data);
 3369                                 break;
 3370                         case D_JMVREF:
 3371                                 jmvref_write(WK_JMVREF(wk), jseg, data);
 3372                                 break;
 3373                         case D_JNEWBLK:
 3374                                 jnewblk_write(WK_JNEWBLK(wk), jseg, data);
 3375                                 break;
 3376                         case D_JFREEBLK:
 3377                                 jfreeblk_write(WK_JFREEBLK(wk), jseg, data);
 3378                                 break;
 3379                         case D_JFREEFRAG:
 3380                                 jfreefrag_write(WK_JFREEFRAG(wk), jseg, data);
 3381                                 break;
 3382                         case D_JTRUNC:
 3383                                 jtrunc_write(WK_JTRUNC(wk), jseg, data);
 3384                                 break;
 3385                         case D_JFSYNC:
 3386                                 jfsync_write(WK_JFSYNC(wk), jseg, data);
 3387                                 break;
 3388                         default:
 3389                                 panic("process_journal: Unknown type %s",
 3390                                     TYPENAME(wk->wk_type));
 3391                                 /* NOTREACHED */
 3392                         }
 3393                         off += JREC_SIZE;
 3394                         data = bp->b_data + off;
 3395                         cnt--;
 3396                 }
 3397                 /*
 3398                  * Write this one buffer and continue.
 3399                  */
 3400                 segwritten = 1;
 3401                 jblocks->jb_needseg = 0;
 3402                 WORKLIST_INSERT(&bp->b_dep, &jseg->js_list);
 3403                 FREE_LOCK(&lk);
 3404                 pbgetvp(ump->um_devvp, bp);
 3405                 /*
 3406                  * We only do the blocking wait once we find the journal
 3407                  * entry we're looking for.
 3408                  */
 3409                 if (needwk == NULL && flags == MNT_WAIT)
 3410                         bwrite(bp);
 3411                 else
 3412                         bawrite(bp);
 3413                 ACQUIRE_LOCK(&lk);
 3414         }
 3415         /*
 3416          * If we wrote a segment issue a synchronize cache so the journal
 3417          * is reflected on disk before the data is written.  Since reclaiming
 3418          * journal space also requires writing a journal record this
 3419          * process also enforces a barrier before reclamation.
 3420          */
 3421         if (segwritten && shouldflush) {
 3422                 softdep_synchronize(bio, ump, 
 3423                     TAILQ_LAST(&jblocks->jb_segs, jseglst));
 3424         } else if (bio)
 3425                 g_destroy_bio(bio);
 3426         /*
 3427          * If we've suspended the filesystem because we ran out of journal
 3428          * space either try to sync it here to make some progress or
 3429          * unsuspend it if we already have.
 3430          */
 3431         if (flags == 0 && jblocks->jb_suspended) {
 3432                 if (journal_unsuspend(ump))
 3433                         return;
 3434                 FREE_LOCK(&lk);
 3435                 VFS_SYNC(mp, MNT_NOWAIT);
 3436                 ffs_sbupdate(ump, MNT_WAIT, 0);
 3437                 ACQUIRE_LOCK(&lk);
 3438         }
 3439 }
 3440 
 3441 /*
 3442  * Complete a jseg, allowing all dependencies awaiting journal writes
 3443  * to proceed.  Each journal dependency also attaches a jsegdep to dependent
 3444  * structures so that the journal segment can be freed to reclaim space.
 3445  */
 3446 static void
 3447 complete_jseg(jseg)
 3448         struct jseg *jseg;
 3449 {
 3450         struct worklist *wk;
 3451         struct jmvref *jmvref;
 3452         int waiting;
 3453 #ifdef INVARIANTS
 3454         int i = 0;
 3455 #endif
 3456 
 3457         while ((wk = LIST_FIRST(&jseg->js_entries)) != NULL) {
 3458                 WORKLIST_REMOVE(wk);
 3459                 waiting = wk->wk_state & IOWAITING;
 3460                 wk->wk_state &= ~(INPROGRESS | IOWAITING);
 3461                 wk->wk_state |= COMPLETE;
 3462                 KASSERT(i++ < jseg->js_cnt,
 3463                     ("handle_written_jseg: overflow %d >= %d",
 3464                     i - 1, jseg->js_cnt));
 3465                 switch (wk->wk_type) {
 3466                 case D_JADDREF:
 3467                         handle_written_jaddref(WK_JADDREF(wk));
 3468                         break;
 3469                 case D_JREMREF:
 3470                         handle_written_jremref(WK_JREMREF(wk));
 3471                         break;
 3472                 case D_JMVREF:
 3473                         rele_jseg(jseg);        /* No jsegdep. */
 3474                         jmvref = WK_JMVREF(wk);
 3475                         LIST_REMOVE(jmvref, jm_deps);
 3476                         if ((jmvref->jm_pagedep->pd_state & ONWORKLIST) == 0)
 3477                                 free_pagedep(jmvref->jm_pagedep);
 3478                         WORKITEM_FREE(jmvref, D_JMVREF);
 3479                         break;
 3480                 case D_JNEWBLK:
 3481                         handle_written_jnewblk(WK_JNEWBLK(wk));
 3482                         break;
 3483                 case D_JFREEBLK:
 3484                         handle_written_jblkdep(&WK_JFREEBLK(wk)->jf_dep);
 3485                         break;
 3486                 case D_JTRUNC:
 3487                         handle_written_jblkdep(&WK_JTRUNC(wk)->jt_dep);
 3488                         break;
 3489                 case D_JFSYNC:
 3490                         rele_jseg(jseg);        /* No jsegdep. */
 3491                         WORKITEM_FREE(wk, D_JFSYNC);
 3492                         break;
 3493                 case D_JFREEFRAG:
 3494                         handle_written_jfreefrag(WK_JFREEFRAG(wk));
 3495                         break;
 3496                 default:
 3497                         panic("handle_written_jseg: Unknown type %s",
 3498                             TYPENAME(wk->wk_type));
 3499                         /* NOTREACHED */
 3500                 }
 3501                 if (waiting)
 3502                         wakeup(wk);
 3503         }
 3504         /* Release the self reference so the structure may be freed. */
 3505         rele_jseg(jseg);
 3506 }
 3507 
 3508 /*
 3509  * Determine which jsegs are ready for completion processing.  Waits for
 3510  * synchronize cache to complete as well as forcing in-order completion
 3511  * of journal entries.
 3512  */
 3513 static void
 3514 complete_jsegs(jseg)
 3515         struct jseg *jseg;
 3516 {
 3517         struct jblocks *jblocks;
 3518         struct jseg *jsegn;
 3519 
 3520         jblocks = jseg->js_jblocks;
 3521         /*
 3522          * Don't allow out of order completions.  If this isn't the first
 3523          * block wait for it to write before we're done.
 3524          */
 3525         if (jseg != jblocks->jb_writeseg)
 3526                 return;
 3527         /* Iterate through available jsegs processing their entries. */
 3528         while (jseg && (jseg->js_state & ALLCOMPLETE) == ALLCOMPLETE) {
 3529                 jblocks->jb_oldestwrseq = jseg->js_oldseq;
 3530                 jsegn = TAILQ_NEXT(jseg, js_next);
 3531                 complete_jseg(jseg);
 3532                 jseg = jsegn;
 3533         }
 3534         jblocks->jb_writeseg = jseg;
 3535         /*
 3536          * Attempt to free jsegs now that oldestwrseq may have advanced. 
 3537          */
 3538         free_jsegs(jblocks);
 3539 }
 3540 
 3541 /*
 3542  * Mark a jseg as DEPCOMPLETE and throw away the buffer.  Attempt to handle
 3543  * the final completions.
 3544  */
 3545 static void
 3546 handle_written_jseg(jseg, bp)
 3547         struct jseg *jseg;
 3548         struct buf *bp;
 3549 {
 3550 
 3551         if (jseg->js_refs == 0)
 3552                 panic("handle_written_jseg: No self-reference on %p", jseg);
 3553         jseg->js_state |= DEPCOMPLETE;
 3554         /*
 3555          * We'll never need this buffer again, set flags so it will be
 3556          * discarded.
 3557          */
 3558         bp->b_flags |= B_INVAL | B_NOCACHE;
 3559         pbrelvp(bp);
 3560         complete_jsegs(jseg);
 3561 }
 3562 
 3563 static inline struct jsegdep *
 3564 inoref_jseg(inoref)
 3565         struct inoref *inoref;
 3566 {
 3567         struct jsegdep *jsegdep;
 3568 
 3569         jsegdep = inoref->if_jsegdep;
 3570         inoref->if_jsegdep = NULL;
 3571 
 3572         return (jsegdep);
 3573 }
 3574 
 3575 /*
 3576  * Called once a jremref has made it to stable store.  The jremref is marked
 3577  * complete and we attempt to free it.  Any pagedeps writes sleeping waiting
 3578  * for the jremref to complete will be awoken by free_jremref.
 3579  */
 3580 static void
 3581 handle_written_jremref(jremref)
 3582         struct jremref *jremref;
 3583 {
 3584         struct inodedep *inodedep;
 3585         struct jsegdep *jsegdep;
 3586         struct dirrem *dirrem;
 3587 
 3588         /* Grab the jsegdep. */
 3589         jsegdep = inoref_jseg(&jremref->jr_ref);
 3590         /*
 3591          * Remove us from the inoref list.
 3592          */
 3593         if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino,
 3594             0, &inodedep) == 0)
 3595                 panic("handle_written_jremref: Lost inodedep");
 3596         TAILQ_REMOVE(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps);
 3597         /*
 3598          * Complete the dirrem.
 3599          */
 3600         dirrem = jremref->jr_dirrem;
 3601         jremref->jr_dirrem = NULL;
 3602         LIST_REMOVE(jremref, jr_deps);
 3603         jsegdep->jd_state |= jremref->jr_state & MKDIR_PARENT;
 3604         jwork_insert(&dirrem->dm_jwork, jsegdep);
 3605         if (LIST_EMPTY(&dirrem->dm_jremrefhd) &&
 3606             (dirrem->dm_state & COMPLETE) != 0)
 3607                 add_to_worklist(&dirrem->dm_list, 0);
 3608         free_jremref(jremref);
 3609 }
 3610 
 3611 /*
 3612  * Called once a jaddref has made it to stable store.  The dependency is
 3613  * marked complete and any dependent structures are added to the inode
 3614  * bufwait list to be completed as soon as it is written.  If a bitmap write
 3615  * depends on this entry we move the inode into the inodedephd of the
 3616  * bmsafemap dependency and attempt to remove the jaddref from the bmsafemap.
 3617  */
 3618 static void
 3619 handle_written_jaddref(jaddref)
 3620         struct jaddref *jaddref;
 3621 {
 3622         struct jsegdep *jsegdep;
 3623         struct inodedep *inodedep;
 3624         struct diradd *diradd;
 3625         struct mkdir *mkdir;
 3626 
 3627         /* Grab the jsegdep. */
 3628         jsegdep = inoref_jseg(&jaddref->ja_ref);
 3629         mkdir = NULL;
 3630         diradd = NULL;
 3631         if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino,
 3632             0, &inodedep) == 0)
 3633                 panic("handle_written_jaddref: Lost inodedep.");
 3634         if (jaddref->ja_diradd == NULL)
 3635                 panic("handle_written_jaddref: No dependency");
 3636         if (jaddref->ja_diradd->da_list.wk_type == D_DIRADD) {
 3637                 diradd = jaddref->ja_diradd;
 3638                 WORKLIST_INSERT(&inodedep->id_bufwait, &diradd->da_list);
 3639         } else if (jaddref->ja_state & MKDIR_PARENT) {
 3640                 mkdir = jaddref->ja_mkdir;
 3641                 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir->md_list);
 3642         } else if (jaddref->ja_state & MKDIR_BODY)
 3643                 mkdir = jaddref->ja_mkdir;
 3644         else
 3645                 panic("handle_written_jaddref: Unknown dependency %p",
 3646                     jaddref->ja_diradd);
 3647         jaddref->ja_diradd = NULL;      /* also clears ja_mkdir */
 3648         /*
 3649          * Remove us from the inode list.
 3650          */
 3651         TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, if_deps);
 3652         /*
 3653          * The mkdir may be waiting on the jaddref to clear before freeing.
 3654          */
 3655         if (mkdir) {
 3656                 KASSERT(mkdir->md_list.wk_type == D_MKDIR,
 3657                     ("handle_written_jaddref: Incorrect type for mkdir %s",
 3658                     TYPENAME(mkdir->md_list.wk_type)));
 3659                 mkdir->md_jaddref = NULL;
 3660                 diradd = mkdir->md_diradd;
 3661                 mkdir->md_state |= DEPCOMPLETE;
 3662                 complete_mkdir(mkdir);
 3663         }
 3664         jwork_insert(&diradd->da_jwork, jsegdep);
 3665         if (jaddref->ja_state & NEWBLOCK) {
 3666                 inodedep->id_state |= ONDEPLIST;
 3667                 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_inodedephd,
 3668                     inodedep, id_deps);
 3669         }
 3670         free_jaddref(jaddref);
 3671 }
 3672 
 3673 /*
 3674  * Called once a jnewblk journal is written.  The allocdirect or allocindir
 3675  * is placed in the bmsafemap to await notification of a written bitmap.  If
 3676  * the operation was canceled we add the segdep to the appropriate
 3677  * dependency to free the journal space once the canceling operation
 3678  * completes.
 3679  */
 3680 static void
 3681 handle_written_jnewblk(jnewblk)
 3682         struct jnewblk *jnewblk;
 3683 {
 3684         struct bmsafemap *bmsafemap;
 3685         struct freefrag *freefrag;
 3686         struct freework *freework;
 3687         struct jsegdep *jsegdep;
 3688         struct newblk *newblk;
 3689 
 3690         /* Grab the jsegdep. */
 3691         jsegdep = jnewblk->jn_jsegdep;
 3692         jnewblk->jn_jsegdep = NULL;
 3693         if (jnewblk->jn_dep == NULL) 
 3694                 panic("handle_written_jnewblk: No dependency for the segdep.");
 3695         switch (jnewblk->jn_dep->wk_type) {
 3696         case D_NEWBLK:
 3697         case D_ALLOCDIRECT:
 3698         case D_ALLOCINDIR:
 3699                 /*
 3700                  * Add the written block to the bmsafemap so it can
 3701                  * be notified when the bitmap is on disk.
 3702                  */
 3703                 newblk = WK_NEWBLK(jnewblk->jn_dep);
 3704                 newblk->nb_jnewblk = NULL;
 3705                 if ((newblk->nb_state & GOINGAWAY) == 0) {
 3706                         bmsafemap = newblk->nb_bmsafemap;
 3707                         newblk->nb_state |= ONDEPLIST;
 3708                         LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk,
 3709                             nb_deps);
 3710                 }
 3711                 jwork_insert(&newblk->nb_jwork, jsegdep);
 3712                 break;
 3713         case D_FREEFRAG:
 3714                 /*
 3715                  * A newblock being removed by a freefrag when replaced by
 3716                  * frag extension.
 3717                  */
 3718                 freefrag = WK_FREEFRAG(jnewblk->jn_dep);
 3719                 freefrag->ff_jdep = NULL;
 3720                 jwork_insert(&freefrag->ff_jwork, jsegdep);
 3721                 break;
 3722         case D_FREEWORK:
 3723                 /*
 3724                  * A direct block was removed by truncate.
 3725                  */
 3726                 freework = WK_FREEWORK(jnewblk->jn_dep);
 3727                 freework->fw_jnewblk = NULL;
 3728                 jwork_insert(&freework->fw_freeblks->fb_jwork, jsegdep);
 3729                 break;
 3730         default:
 3731                 panic("handle_written_jnewblk: Unknown type %d.",
 3732                     jnewblk->jn_dep->wk_type);
 3733         }
 3734         jnewblk->jn_dep = NULL;
 3735         free_jnewblk(jnewblk);
 3736 }
 3737 
 3738 /*
 3739  * Cancel a jfreefrag that won't be needed, probably due to colliding with
 3740  * an in-flight allocation that has not yet been committed.  Divorce us
 3741  * from the freefrag and mark it DEPCOMPLETE so that it may be added
 3742  * to the worklist.
 3743  */
 3744 static void
 3745 cancel_jfreefrag(jfreefrag)
 3746         struct jfreefrag *jfreefrag;
 3747 {
 3748         struct freefrag *freefrag;
 3749 
 3750         if (jfreefrag->fr_jsegdep) {
 3751                 free_jsegdep(jfreefrag->fr_jsegdep);
 3752                 jfreefrag->fr_jsegdep = NULL;
 3753         }
 3754         freefrag = jfreefrag->fr_freefrag;
 3755         jfreefrag->fr_freefrag = NULL;
 3756         free_jfreefrag(jfreefrag);
 3757         freefrag->ff_state |= DEPCOMPLETE;
 3758         CTR1(KTR_SUJ, "cancel_jfreefrag: blkno %jd", freefrag->ff_blkno);
 3759 }
 3760 
 3761 /*
 3762  * Free a jfreefrag when the parent freefrag is rendered obsolete.
 3763  */
 3764 static void
 3765 free_jfreefrag(jfreefrag)
 3766         struct jfreefrag *jfreefrag;
 3767 {
 3768 
 3769         if (jfreefrag->fr_state & INPROGRESS)
 3770                 WORKLIST_REMOVE(&jfreefrag->fr_list);
 3771         else if (jfreefrag->fr_state & ONWORKLIST)
 3772                 remove_from_journal(&jfreefrag->fr_list);
 3773         if (jfreefrag->fr_freefrag != NULL)
 3774                 panic("free_jfreefrag:  Still attached to a freefrag.");
 3775         WORKITEM_FREE(jfreefrag, D_JFREEFRAG);
 3776 }
 3777 
 3778 /*
 3779  * Called when the journal write for a jfreefrag completes.  The parent
 3780  * freefrag is added to the worklist if this completes its dependencies.
 3781  */
 3782 static void
 3783 handle_written_jfreefrag(jfreefrag)
 3784         struct jfreefrag *jfreefrag;
 3785 {
 3786         struct jsegdep *jsegdep;
 3787         struct freefrag *freefrag;
 3788 
 3789         /* Grab the jsegdep. */
 3790         jsegdep = jfreefrag->fr_jsegdep;
 3791         jfreefrag->fr_jsegdep = NULL;
 3792         freefrag = jfreefrag->fr_freefrag;
 3793         if (freefrag == NULL)
 3794                 panic("handle_written_jfreefrag: No freefrag.");
 3795         freefrag->ff_state |= DEPCOMPLETE;
 3796         freefrag->ff_jdep = NULL;
 3797         jwork_insert(&freefrag->ff_jwork, jsegdep);
 3798         if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE)
 3799                 add_to_worklist(&freefrag->ff_list, 0);
 3800         jfreefrag->fr_freefrag = NULL;
 3801         free_jfreefrag(jfreefrag);
 3802 }
 3803 
 3804 /*
 3805  * Called when the journal write for a jfreeblk completes.  The jfreeblk
 3806  * is removed from the freeblks list of pending journal writes and the
 3807  * jsegdep is moved to the freeblks jwork to be completed when all blocks
 3808  * have been reclaimed.
 3809  */
 3810 static void
 3811 handle_written_jblkdep(jblkdep)
 3812         struct jblkdep *jblkdep;
 3813 {
 3814         struct freeblks *freeblks;
 3815         struct jsegdep *jsegdep;
 3816 
 3817         /* Grab the jsegdep. */
 3818         jsegdep = jblkdep->jb_jsegdep;
 3819         jblkdep->jb_jsegdep = NULL;
 3820         freeblks = jblkdep->jb_freeblks;
 3821         LIST_REMOVE(jblkdep, jb_deps);
 3822         jwork_insert(&freeblks->fb_jwork, jsegdep);
 3823         /*
 3824          * If the freeblks is all journaled, we can add it to the worklist.
 3825          */
 3826         if (LIST_EMPTY(&freeblks->fb_jblkdephd) &&
 3827             (freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
 3828                 add_to_worklist(&freeblks->fb_list, WK_NODELAY);
 3829 
 3830         free_jblkdep(jblkdep);
 3831 }
 3832 
 3833 static struct jsegdep *
 3834 newjsegdep(struct worklist *wk)
 3835 {
 3836         struct jsegdep *jsegdep;
 3837 
 3838         jsegdep = malloc(sizeof(*jsegdep), M_JSEGDEP, M_SOFTDEP_FLAGS);
 3839         workitem_alloc(&jsegdep->jd_list, D_JSEGDEP, wk->wk_mp);
 3840         jsegdep->jd_seg = NULL;
 3841 
 3842         return (jsegdep);
 3843 }
 3844 
 3845 static struct jmvref *
 3846 newjmvref(dp, ino, oldoff, newoff)
 3847         struct inode *dp;
 3848         ino_t ino;
 3849         off_t oldoff;
 3850         off_t newoff;
 3851 {
 3852         struct jmvref *jmvref;
 3853 
 3854         jmvref = malloc(sizeof(*jmvref), M_JMVREF, M_SOFTDEP_FLAGS);
 3855         workitem_alloc(&jmvref->jm_list, D_JMVREF, UFSTOVFS(dp->i_ump));
 3856         jmvref->jm_list.wk_state = ATTACHED | DEPCOMPLETE;
 3857         jmvref->jm_parent = dp->i_number;
 3858         jmvref->jm_ino = ino;
 3859         jmvref->jm_oldoff = oldoff;
 3860         jmvref->jm_newoff = newoff;
 3861 
 3862         return (jmvref);
 3863 }
 3864 
 3865 /*
 3866  * Allocate a new jremref that tracks the removal of ip from dp with the
 3867  * directory entry offset of diroff.  Mark the entry as ATTACHED and
 3868  * DEPCOMPLETE as we have all the information required for the journal write
 3869  * and the directory has already been removed from the buffer.  The caller
 3870  * is responsible for linking the jremref into the pagedep and adding it
 3871  * to the journal to write.  The MKDIR_PARENT flag is set if we're doing
 3872  * a DOTDOT addition so handle_workitem_remove() can properly assign
 3873  * the jsegdep when we're done.
 3874  */
 3875 static struct jremref *
 3876 newjremref(struct dirrem *dirrem, struct inode *dp, struct inode *ip,
 3877     off_t diroff, nlink_t nlink)
 3878 {
 3879         struct jremref *jremref;
 3880 
 3881         jremref = malloc(sizeof(*jremref), M_JREMREF, M_SOFTDEP_FLAGS);
 3882         workitem_alloc(&jremref->jr_list, D_JREMREF, UFSTOVFS(dp->i_ump));
 3883         jremref->jr_state = ATTACHED;
 3884         newinoref(&jremref->jr_ref, ip->i_number, dp->i_number, diroff,
 3885            nlink, ip->i_mode);
 3886         jremref->jr_dirrem = dirrem;
 3887 
 3888         return (jremref);
 3889 }
 3890 
 3891 static inline void
 3892 newinoref(struct inoref *inoref, ino_t ino, ino_t parent, off_t diroff,
 3893     nlink_t nlink, uint16_t mode)
 3894 {
 3895 
 3896         inoref->if_jsegdep = newjsegdep(&inoref->if_list);
 3897         inoref->if_diroff = diroff;
 3898         inoref->if_ino = ino;
 3899         inoref->if_parent = parent;
 3900         inoref->if_nlink = nlink;
 3901         inoref->if_mode = mode;
 3902 }
 3903 
 3904 /*
 3905  * Allocate a new jaddref to track the addition of ino to dp at diroff.  The
 3906  * directory offset may not be known until later.  The caller is responsible
 3907  * adding the entry to the journal when this information is available.  nlink
 3908  * should be the link count prior to the addition and mode is only required
 3909  * to have the correct FMT.
 3910  */
 3911 static struct jaddref *
 3912 newjaddref(struct inode *dp, ino_t ino, off_t diroff, int16_t nlink,
 3913     uint16_t mode)
 3914 {
 3915         struct jaddref *jaddref;
 3916 
 3917         jaddref = malloc(sizeof(*jaddref), M_JADDREF, M_SOFTDEP_FLAGS);
 3918         workitem_alloc(&jaddref->ja_list, D_JADDREF, UFSTOVFS(dp->i_ump));
 3919         jaddref->ja_state = ATTACHED;
 3920         jaddref->ja_mkdir = NULL;
 3921         newinoref(&jaddref->ja_ref, ino, dp->i_number, diroff, nlink, mode);
 3922 
 3923         return (jaddref);
 3924 }
 3925 
 3926 /*
 3927  * Create a new free dependency for a freework.  The caller is responsible
 3928  * for adjusting the reference count when it has the lock held.  The freedep
 3929  * will track an outstanding bitmap write that will ultimately clear the
 3930  * freework to continue.
 3931  */
 3932 static struct freedep *
 3933 newfreedep(struct freework *freework)
 3934 {
 3935         struct freedep *freedep;
 3936 
 3937         freedep = malloc(sizeof(*freedep), M_FREEDEP, M_SOFTDEP_FLAGS);
 3938         workitem_alloc(&freedep->fd_list, D_FREEDEP, freework->fw_list.wk_mp);
 3939         freedep->fd_freework = freework;
 3940 
 3941         return (freedep);
 3942 }
 3943 
 3944 /*
 3945  * Free a freedep structure once the buffer it is linked to is written.  If
 3946  * this is the last reference to the freework schedule it for completion.
 3947  */
 3948 static void
 3949 free_freedep(freedep)
 3950         struct freedep *freedep;
 3951 {
 3952         struct freework *freework;
 3953 
 3954         freework = freedep->fd_freework;
 3955         freework->fw_freeblks->fb_cgwait--;
 3956         if (--freework->fw_ref == 0)
 3957                 freework_enqueue(freework);
 3958         WORKITEM_FREE(freedep, D_FREEDEP);
 3959 }
 3960 
 3961 /*
 3962  * Allocate a new freework structure that may be a level in an indirect
 3963  * when parent is not NULL or a top level block when it is.  The top level
 3964  * freework structures are allocated without lk held and before the freeblks
 3965  * is visible outside of softdep_setup_freeblocks().
 3966  */
 3967 static struct freework *
 3968 newfreework(ump, freeblks, parent, lbn, nb, frags, off, journal)
 3969         struct ufsmount *ump;
 3970         struct freeblks *freeblks;
 3971         struct freework *parent;
 3972         ufs_lbn_t lbn;
 3973         ufs2_daddr_t nb;
 3974         int frags;
 3975         int off;
 3976         int journal;
 3977 {
 3978         struct freework *freework;
 3979 
 3980         freework = malloc(sizeof(*freework), M_FREEWORK, M_SOFTDEP_FLAGS);
 3981         workitem_alloc(&freework->fw_list, D_FREEWORK, freeblks->fb_list.wk_mp);
 3982         freework->fw_state = ATTACHED;
 3983         freework->fw_jnewblk = NULL;
 3984         freework->fw_freeblks = freeblks;
 3985         freework->fw_parent = parent;
 3986         freework->fw_lbn = lbn;
 3987         freework->fw_blkno = nb;
 3988         freework->fw_frags = frags;
 3989         freework->fw_indir = NULL;
 3990         freework->fw_ref = (MOUNTEDSUJ(UFSTOVFS(ump)) == 0 || lbn >= -NXADDR)
 3991                 ? 0 : NINDIR(ump->um_fs) + 1;
 3992         freework->fw_start = freework->fw_off = off;
 3993         if (journal)
 3994                 newjfreeblk(freeblks, lbn, nb, frags);
 3995         if (parent == NULL) {
 3996                 ACQUIRE_LOCK(&lk);
 3997                 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list);
 3998                 freeblks->fb_ref++;
 3999                 FREE_LOCK(&lk);
 4000         }
 4001 
 4002         return (freework);
 4003 }
 4004 
 4005 /*
 4006  * Eliminate a jfreeblk for a block that does not need journaling.
 4007  */
 4008 static void
 4009 cancel_jfreeblk(freeblks, blkno)
 4010         struct freeblks *freeblks;
 4011         ufs2_daddr_t blkno;
 4012 {
 4013         struct jfreeblk *jfreeblk;
 4014         struct jblkdep *jblkdep;
 4015 
 4016         LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) {
 4017                 if (jblkdep->jb_list.wk_type != D_JFREEBLK)
 4018                         continue;
 4019                 jfreeblk = WK_JFREEBLK(&jblkdep->jb_list);
 4020                 if (jfreeblk->jf_blkno == blkno)
 4021                         break;
 4022         }
 4023         if (jblkdep == NULL)
 4024                 return;
 4025         CTR1(KTR_SUJ, "cancel_jfreeblk: blkno %jd", blkno);
 4026         free_jsegdep(jblkdep->jb_jsegdep);
 4027         LIST_REMOVE(jblkdep, jb_deps);
 4028         WORKITEM_FREE(jfreeblk, D_JFREEBLK);
 4029 }
 4030 
 4031 /*
 4032  * Allocate a new jfreeblk to journal top level block pointer when truncating
 4033  * a file.  The caller must add this to the worklist when lk is held.
 4034  */
 4035 static struct jfreeblk *
 4036 newjfreeblk(freeblks, lbn, blkno, frags)
 4037         struct freeblks *freeblks;
 4038         ufs_lbn_t lbn;
 4039         ufs2_daddr_t blkno;
 4040         int frags;
 4041 {
 4042         struct jfreeblk *jfreeblk;
 4043 
 4044         jfreeblk = malloc(sizeof(*jfreeblk), M_JFREEBLK, M_SOFTDEP_FLAGS);
 4045         workitem_alloc(&jfreeblk->jf_dep.jb_list, D_JFREEBLK,
 4046             freeblks->fb_list.wk_mp);
 4047         jfreeblk->jf_dep.jb_jsegdep = newjsegdep(&jfreeblk->jf_dep.jb_list);
 4048         jfreeblk->jf_dep.jb_freeblks = freeblks;
 4049         jfreeblk->jf_ino = freeblks->fb_inum;
 4050         jfreeblk->jf_lbn = lbn;
 4051         jfreeblk->jf_blkno = blkno;
 4052         jfreeblk->jf_frags = frags;
 4053         LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jfreeblk->jf_dep, jb_deps);
 4054 
 4055         return (jfreeblk);
 4056 }
 4057 
 4058 /*
 4059  * Allocate a new jtrunc to track a partial truncation.
 4060  */
 4061 static struct jtrunc *
 4062 newjtrunc(freeblks, size, extsize)
 4063         struct freeblks *freeblks;
 4064         off_t size;
 4065         int extsize;
 4066 {
 4067         struct jtrunc *jtrunc;
 4068 
 4069         jtrunc = malloc(sizeof(*jtrunc), M_JTRUNC, M_SOFTDEP_FLAGS);
 4070         workitem_alloc(&jtrunc->jt_dep.jb_list, D_JTRUNC,
 4071             freeblks->fb_list.wk_mp);
 4072         jtrunc->jt_dep.jb_jsegdep = newjsegdep(&jtrunc->jt_dep.jb_list);
 4073         jtrunc->jt_dep.jb_freeblks = freeblks;
 4074         jtrunc->jt_ino = freeblks->fb_inum;
 4075         jtrunc->jt_size = size;
 4076         jtrunc->jt_extsize = extsize;
 4077         LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jtrunc->jt_dep, jb_deps);
 4078 
 4079         return (jtrunc);
 4080 }
 4081 
 4082 /*
 4083  * If we're canceling a new bitmap we have to search for another ref
 4084  * to move into the bmsafemap dep.  This might be better expressed
 4085  * with another structure.
 4086  */
 4087 static void
 4088 move_newblock_dep(jaddref, inodedep)
 4089         struct jaddref *jaddref;
 4090         struct inodedep *inodedep;
 4091 {
 4092         struct inoref *inoref;
 4093         struct jaddref *jaddrefn;
 4094 
 4095         jaddrefn = NULL;
 4096         for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref;
 4097             inoref = TAILQ_NEXT(inoref, if_deps)) {
 4098                 if ((jaddref->ja_state & NEWBLOCK) &&
 4099                     inoref->if_list.wk_type == D_JADDREF) {
 4100                         jaddrefn = (struct jaddref *)inoref;
 4101                         break;
 4102                 }
 4103         }
 4104         if (jaddrefn == NULL)
 4105                 return;
 4106         jaddrefn->ja_state &= ~(ATTACHED | UNDONE);
 4107         jaddrefn->ja_state |= jaddref->ja_state &
 4108             (ATTACHED | UNDONE | NEWBLOCK);
 4109         jaddref->ja_state &= ~(ATTACHED | UNDONE | NEWBLOCK);
 4110         jaddref->ja_state |= ATTACHED;
 4111         LIST_REMOVE(jaddref, ja_bmdeps);
 4112         LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_jaddrefhd, jaddrefn,
 4113             ja_bmdeps);
 4114 }
 4115 
 4116 /*
 4117  * Cancel a jaddref either before it has been written or while it is being
 4118  * written.  This happens when a link is removed before the add reaches
 4119  * the disk.  The jaddref dependency is kept linked into the bmsafemap
 4120  * and inode to prevent the link count or bitmap from reaching the disk
 4121  * until handle_workitem_remove() re-adjusts the counts and bitmaps as
 4122  * required.
 4123  *
 4124  * Returns 1 if the canceled addref requires journaling of the remove and
 4125  * 0 otherwise.
 4126  */
 4127 static int
 4128 cancel_jaddref(jaddref, inodedep, wkhd)
 4129         struct jaddref *jaddref;
 4130         struct inodedep *inodedep;
 4131         struct workhead *wkhd;
 4132 {
 4133         struct inoref *inoref;
 4134         struct jsegdep *jsegdep;
 4135         int needsj;
 4136 
 4137         KASSERT((jaddref->ja_state & COMPLETE) == 0,
 4138             ("cancel_jaddref: Canceling complete jaddref"));
 4139         if (jaddref->ja_state & (INPROGRESS | COMPLETE))
 4140                 needsj = 1;
 4141         else
 4142                 needsj = 0;
 4143         if (inodedep == NULL)
 4144                 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino,
 4145                     0, &inodedep) == 0)
 4146                         panic("cancel_jaddref: Lost inodedep");
 4147         /*
 4148          * We must adjust the nlink of any reference operation that follows
 4149          * us so that it is consistent with the in-memory reference.  This
 4150          * ensures that inode nlink rollbacks always have the correct link.
 4151          */
 4152         if (needsj == 0) {
 4153                 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref;
 4154                     inoref = TAILQ_NEXT(inoref, if_deps)) {
 4155                         if (inoref->if_state & GOINGAWAY)
 4156                                 break;
 4157                         inoref->if_nlink--;
 4158                 }
 4159         }
 4160         jsegdep = inoref_jseg(&jaddref->ja_ref);
 4161         if (jaddref->ja_state & NEWBLOCK)
 4162                 move_newblock_dep(jaddref, inodedep);
 4163         wake_worklist(&jaddref->ja_list);
 4164         jaddref->ja_mkdir = NULL;
 4165         if (jaddref->ja_state & INPROGRESS) {
 4166                 jaddref->ja_state &= ~INPROGRESS;
 4167                 WORKLIST_REMOVE(&jaddref->ja_list);
 4168                 jwork_insert(wkhd, jsegdep);
 4169         } else {
 4170                 free_jsegdep(jsegdep);
 4171                 if (jaddref->ja_state & DEPCOMPLETE)
 4172                         remove_from_journal(&jaddref->ja_list);
 4173         }
 4174         jaddref->ja_state |= (GOINGAWAY | DEPCOMPLETE);
 4175         /*
 4176          * Leave NEWBLOCK jaddrefs on the inodedep so handle_workitem_remove
 4177          * can arrange for them to be freed with the bitmap.  Otherwise we
 4178          * no longer need this addref attached to the inoreflst and it
 4179          * will incorrectly adjust nlink if we leave it.
 4180          */
 4181         if ((jaddref->ja_state & NEWBLOCK) == 0) {
 4182                 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref,
 4183                     if_deps);
 4184                 jaddref->ja_state |= COMPLETE;
 4185                 free_jaddref(jaddref);
 4186                 return (needsj);
 4187         }
 4188         /*
 4189          * Leave the head of the list for jsegdeps for fast merging.
 4190          */
 4191         if (LIST_FIRST(wkhd) != NULL) {
 4192                 jaddref->ja_state |= ONWORKLIST;
 4193                 LIST_INSERT_AFTER(LIST_FIRST(wkhd), &jaddref->ja_list, wk_list);
 4194         } else
 4195                 WORKLIST_INSERT(wkhd, &jaddref->ja_list);
 4196 
 4197         return (needsj);
 4198 }
 4199 
 4200 /* 
 4201  * Attempt to free a jaddref structure when some work completes.  This
 4202  * should only succeed once the entry is written and all dependencies have
 4203  * been notified.
 4204  */
 4205 static void
 4206 free_jaddref(jaddref)
 4207         struct jaddref *jaddref;
 4208 {
 4209 
 4210         if ((jaddref->ja_state & ALLCOMPLETE) != ALLCOMPLETE)
 4211                 return;
 4212         if (jaddref->ja_ref.if_jsegdep)
 4213                 panic("free_jaddref: segdep attached to jaddref %p(0x%X)\n",
 4214                     jaddref, jaddref->ja_state);
 4215         if (jaddref->ja_state & NEWBLOCK)
 4216                 LIST_REMOVE(jaddref, ja_bmdeps);
 4217         if (jaddref->ja_state & (INPROGRESS | ONWORKLIST))
 4218                 panic("free_jaddref: Bad state %p(0x%X)",
 4219                     jaddref, jaddref->ja_state);
 4220         if (jaddref->ja_mkdir != NULL)
 4221                 panic("free_jaddref: Work pending, 0x%X\n", jaddref->ja_state);
 4222         WORKITEM_FREE(jaddref, D_JADDREF);
 4223 }
 4224 
 4225 /*
 4226  * Free a jremref structure once it has been written or discarded.
 4227  */
 4228 static void
 4229 free_jremref(jremref)
 4230         struct jremref *jremref;
 4231 {
 4232 
 4233         if (jremref->jr_ref.if_jsegdep)
 4234                 free_jsegdep(jremref->jr_ref.if_jsegdep);
 4235         if (jremref->jr_state & INPROGRESS)
 4236                 panic("free_jremref: IO still pending");
 4237         WORKITEM_FREE(jremref, D_JREMREF);
 4238 }
 4239 
 4240 /*
 4241  * Free a jnewblk structure.
 4242  */
 4243 static void
 4244 free_jnewblk(jnewblk)
 4245         struct jnewblk *jnewblk;
 4246 {
 4247 
 4248         if ((jnewblk->jn_state & ALLCOMPLETE) != ALLCOMPLETE)
 4249                 return;
 4250         LIST_REMOVE(jnewblk, jn_deps);
 4251         if (jnewblk->jn_dep != NULL)
 4252                 panic("free_jnewblk: Dependency still attached.");
 4253         WORKITEM_FREE(jnewblk, D_JNEWBLK);
 4254 }
 4255 
 4256 /*
 4257  * Cancel a jnewblk which has been been made redundant by frag extension.
 4258  */
 4259 static void
 4260 cancel_jnewblk(jnewblk, wkhd)
 4261         struct jnewblk *jnewblk;
 4262         struct workhead *wkhd;
 4263 {
 4264         struct jsegdep *jsegdep;
 4265 
 4266         CTR1(KTR_SUJ, "cancel_jnewblk: blkno %jd", jnewblk->jn_blkno);
 4267         jsegdep = jnewblk->jn_jsegdep;
 4268         if (jnewblk->jn_jsegdep == NULL || jnewblk->jn_dep == NULL)
 4269                 panic("cancel_jnewblk: Invalid state");
 4270         jnewblk->jn_jsegdep  = NULL;
 4271         jnewblk->jn_dep = NULL;
 4272         jnewblk->jn_state |= GOINGAWAY;
 4273         if (jnewblk->jn_state & INPROGRESS) {
 4274                 jnewblk->jn_state &= ~INPROGRESS;
 4275                 WORKLIST_REMOVE(&jnewblk->jn_list);
 4276                 jwork_insert(wkhd, jsegdep);
 4277         } else {
 4278                 free_jsegdep(jsegdep);
 4279                 remove_from_journal(&jnewblk->jn_list);
 4280         }
 4281         wake_worklist(&jnewblk->jn_list);
 4282         WORKLIST_INSERT(wkhd, &jnewblk->jn_list);
 4283 }
 4284 
 4285 static void
 4286 free_jblkdep(jblkdep)
 4287         struct jblkdep *jblkdep;
 4288 {
 4289 
 4290         if (jblkdep->jb_list.wk_type == D_JFREEBLK)
 4291                 WORKITEM_FREE(jblkdep, D_JFREEBLK);
 4292         else if (jblkdep->jb_list.wk_type == D_JTRUNC)
 4293                 WORKITEM_FREE(jblkdep, D_JTRUNC);
 4294         else
 4295                 panic("free_jblkdep: Unexpected type %s",
 4296                     TYPENAME(jblkdep->jb_list.wk_type));
 4297 }
 4298 
 4299 /*
 4300  * Free a single jseg once it is no longer referenced in memory or on
 4301  * disk.  Reclaim journal blocks and dependencies waiting for the segment
 4302  * to disappear.
 4303  */
 4304 static void
 4305 free_jseg(jseg, jblocks)
 4306         struct jseg *jseg;
 4307         struct jblocks *jblocks;
 4308 {
 4309         struct freework *freework;
 4310 
 4311         /*
 4312          * Free freework structures that were lingering to indicate freed
 4313          * indirect blocks that forced journal write ordering on reallocate.
 4314          */
 4315         while ((freework = LIST_FIRST(&jseg->js_indirs)) != NULL)
 4316                 indirblk_remove(freework);
 4317         if (jblocks->jb_oldestseg == jseg)
 4318                 jblocks->jb_oldestseg = TAILQ_NEXT(jseg, js_next);
 4319         TAILQ_REMOVE(&jblocks->jb_segs, jseg, js_next);
 4320         jblocks_free(jblocks, jseg->js_list.wk_mp, jseg->js_size);
 4321         KASSERT(LIST_EMPTY(&jseg->js_entries),
 4322             ("free_jseg: Freed jseg has valid entries."));
 4323         WORKITEM_FREE(jseg, D_JSEG);
 4324 }
 4325 
 4326 /*
 4327  * Free all jsegs that meet the criteria for being reclaimed and update
 4328  * oldestseg.
 4329  */
 4330 static void
 4331 free_jsegs(jblocks)
 4332         struct jblocks *jblocks;
 4333 {
 4334         struct jseg *jseg;
 4335 
 4336         /*
 4337          * Free only those jsegs which have none allocated before them to
 4338          * preserve the journal space ordering.
 4339          */
 4340         while ((jseg = TAILQ_FIRST(&jblocks->jb_segs)) != NULL) {
 4341                 /*
 4342                  * Only reclaim space when nothing depends on this journal
 4343                  * set and another set has written that it is no longer
 4344                  * valid.
 4345                  */
 4346                 if (jseg->js_refs != 0) {
 4347                         jblocks->jb_oldestseg = jseg;
 4348                         return;
 4349                 }
 4350                 if ((jseg->js_state & ALLCOMPLETE) != ALLCOMPLETE)
 4351                         break;
 4352                 if (jseg->js_seq > jblocks->jb_oldestwrseq)
 4353                         break;
 4354                 /*
 4355                  * We can free jsegs that didn't write entries when
 4356                  * oldestwrseq == js_seq.
 4357                  */
 4358                 if (jseg->js_seq == jblocks->jb_oldestwrseq &&
 4359                     jseg->js_cnt != 0)
 4360                         break;
 4361                 free_jseg(jseg, jblocks);
 4362         }
 4363         /*
 4364          * If we exited the loop above we still must discover the
 4365          * oldest valid segment.
 4366          */
 4367         if (jseg)
 4368                 for (jseg = jblocks->jb_oldestseg; jseg != NULL;
 4369                      jseg = TAILQ_NEXT(jseg, js_next))
 4370                         if (jseg->js_refs != 0)
 4371                                 break;
 4372         jblocks->jb_oldestseg = jseg;
 4373         /*
 4374          * The journal has no valid records but some jsegs may still be
 4375          * waiting on oldestwrseq to advance.  We force a small record
 4376          * out to permit these lingering records to be reclaimed.
 4377          */
 4378         if (jblocks->jb_oldestseg == NULL && !TAILQ_EMPTY(&jblocks->jb_segs))
 4379                 jblocks->jb_needseg = 1;
 4380 }
 4381 
 4382 /*
 4383  * Release one reference to a jseg and free it if the count reaches 0.  This
 4384  * should eventually reclaim journal space as well.
 4385  */
 4386 static void
 4387 rele_jseg(jseg)
 4388         struct jseg *jseg;
 4389 {
 4390 
 4391         KASSERT(jseg->js_refs > 0,
 4392             ("free_jseg: Invalid refcnt %d", jseg->js_refs));
 4393         if (--jseg->js_refs != 0)
 4394                 return;
 4395         free_jsegs(jseg->js_jblocks);
 4396 }
 4397 
 4398 /*
 4399  * Release a jsegdep and decrement the jseg count.
 4400  */
 4401 static void
 4402 free_jsegdep(jsegdep)
 4403         struct jsegdep *jsegdep;
 4404 {
 4405 
 4406         if (jsegdep->jd_seg)
 4407                 rele_jseg(jsegdep->jd_seg);
 4408         WORKITEM_FREE(jsegdep, D_JSEGDEP);
 4409 }
 4410 
 4411 /*
 4412  * Wait for a journal item to make it to disk.  Initiate journal processing
 4413  * if required.
 4414  */
 4415 static int
 4416 jwait(wk, waitfor)
 4417         struct worklist *wk;
 4418         int waitfor;
 4419 {
 4420 
 4421         /*
 4422          * Blocking journal waits cause slow synchronous behavior.  Record
 4423          * stats on the frequency of these blocking operations.
 4424          */
 4425         if (waitfor == MNT_WAIT) {
 4426                 stat_journal_wait++;
 4427                 switch (wk->wk_type) {
 4428                 case D_JREMREF:
 4429                 case D_JMVREF:
 4430                         stat_jwait_filepage++;
 4431                         break;
 4432                 case D_JTRUNC:
 4433                 case D_JFREEBLK:
 4434                         stat_jwait_freeblks++;
 4435                         break;
 4436                 case D_JNEWBLK:
 4437                         stat_jwait_newblk++;
 4438                         break;
 4439                 case D_JADDREF:
 4440                         stat_jwait_inode++;
 4441                         break;
 4442                 default:
 4443                         break;
 4444                 }
 4445         }
 4446         /*
 4447          * If IO has not started we process the journal.  We can't mark the
 4448          * worklist item as IOWAITING because we drop the lock while
 4449          * processing the journal and the worklist entry may be freed after
 4450          * this point.  The caller may call back in and re-issue the request.
 4451          */
 4452         if ((wk->wk_state & INPROGRESS) == 0) {
 4453                 softdep_process_journal(wk->wk_mp, wk, waitfor);
 4454                 if (waitfor != MNT_WAIT)
 4455                         return (EBUSY);
 4456                 return (0);
 4457         }
 4458         if (waitfor != MNT_WAIT)
 4459                 return (EBUSY);
 4460         wait_worklist(wk, "jwait");
 4461         return (0);
 4462 }
 4463 
 4464 /*
 4465  * Lookup an inodedep based on an inode pointer and set the nlinkdelta as
 4466  * appropriate.  This is a convenience function to reduce duplicate code
 4467  * for the setup and revert functions below.
 4468  */
 4469 static struct inodedep *
 4470 inodedep_lookup_ip(ip)
 4471         struct inode *ip;
 4472 {
 4473         struct inodedep *inodedep;
 4474         int dflags;
 4475 
 4476         KASSERT(ip->i_nlink >= ip->i_effnlink,
 4477             ("inodedep_lookup_ip: bad delta"));
 4478         dflags = DEPALLOC;
 4479         if (IS_SNAPSHOT(ip))
 4480                 dflags |= NODELAY;
 4481         (void) inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, dflags,
 4482             &inodedep);
 4483         inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
 4484         KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked"));
 4485 
 4486         return (inodedep);
 4487 }
 4488 
 4489 /*
 4490  * Called prior to creating a new inode and linking it to a directory.  The
 4491  * jaddref structure must already be allocated by softdep_setup_inomapdep
 4492  * and it is discovered here so we can initialize the mode and update
 4493  * nlinkdelta.
 4494  */
 4495 void
 4496 softdep_setup_create(dp, ip)
 4497         struct inode *dp;
 4498         struct inode *ip;
 4499 {
 4500         struct inodedep *inodedep;
 4501         struct jaddref *jaddref;
 4502         struct vnode *dvp;
 4503 
 4504         KASSERT(ip->i_nlink == 1,
 4505             ("softdep_setup_create: Invalid link count."));
 4506         dvp = ITOV(dp);
 4507         ACQUIRE_LOCK(&lk);
 4508         inodedep = inodedep_lookup_ip(ip);
 4509         if (DOINGSUJ(dvp)) {
 4510                 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
 4511                     inoreflst);
 4512                 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number,
 4513                     ("softdep_setup_create: No addref structure present."));
 4514         }
 4515         softdep_prelink(dvp, NULL);
 4516         FREE_LOCK(&lk);
 4517 }
 4518 
 4519 /*
 4520  * Create a jaddref structure to track the addition of a DOTDOT link when
 4521  * we are reparenting an inode as part of a rename.  This jaddref will be
 4522  * found by softdep_setup_directory_change.  Adjusts nlinkdelta for
 4523  * non-journaling softdep.
 4524  */
 4525 void
 4526 softdep_setup_dotdot_link(dp, ip)
 4527         struct inode *dp;
 4528         struct inode *ip;
 4529 {
 4530         struct inodedep *inodedep;
 4531         struct jaddref *jaddref;
 4532         struct vnode *dvp;
 4533         struct vnode *vp;
 4534 
 4535         dvp = ITOV(dp);
 4536         vp = ITOV(ip);
 4537         jaddref = NULL;
 4538         /*
 4539          * We don't set MKDIR_PARENT as this is not tied to a mkdir and
 4540          * is used as a normal link would be.
 4541          */
 4542         if (DOINGSUJ(dvp))
 4543                 jaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET,
 4544                     dp->i_effnlink - 1, dp->i_mode);
 4545         ACQUIRE_LOCK(&lk);
 4546         inodedep = inodedep_lookup_ip(dp);
 4547         if (jaddref)
 4548                 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref,
 4549                     if_deps);
 4550         softdep_prelink(dvp, ITOV(ip));
 4551         FREE_LOCK(&lk);
 4552 }
 4553 
 4554 /*
 4555  * Create a jaddref structure to track a new link to an inode.  The directory
 4556  * offset is not known until softdep_setup_directory_add or
 4557  * softdep_setup_directory_change.  Adjusts nlinkdelta for non-journaling
 4558  * softdep.
 4559  */
 4560 void
 4561 softdep_setup_link(dp, ip)
 4562         struct inode *dp;
 4563         struct inode *ip;
 4564 {
 4565         struct inodedep *inodedep;
 4566         struct jaddref *jaddref;
 4567         struct vnode *dvp;
 4568 
 4569         dvp = ITOV(dp);
 4570         jaddref = NULL;
 4571         if (DOINGSUJ(dvp))
 4572                 jaddref = newjaddref(dp, ip->i_number, 0, ip->i_effnlink - 1,
 4573                     ip->i_mode);
 4574         ACQUIRE_LOCK(&lk);
 4575         inodedep = inodedep_lookup_ip(ip);
 4576         if (jaddref)
 4577                 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref,
 4578                     if_deps);
 4579         softdep_prelink(dvp, ITOV(ip));
 4580         FREE_LOCK(&lk);
 4581 }
 4582 
 4583 /*
 4584  * Called to create the jaddref structures to track . and .. references as
 4585  * well as lookup and further initialize the incomplete jaddref created
 4586  * by softdep_setup_inomapdep when the inode was allocated.  Adjusts
 4587  * nlinkdelta for non-journaling softdep.
 4588  */
 4589 void
 4590 softdep_setup_mkdir(dp, ip)
 4591         struct inode *dp;
 4592         struct inode *ip;
 4593 {
 4594         struct inodedep *inodedep;
 4595         struct jaddref *dotdotaddref;
 4596         struct jaddref *dotaddref;
 4597         struct jaddref *jaddref;
 4598         struct vnode *dvp;
 4599 
 4600         dvp = ITOV(dp);
 4601         dotaddref = dotdotaddref = NULL;
 4602         if (DOINGSUJ(dvp)) {
 4603                 dotaddref = newjaddref(ip, ip->i_number, DOT_OFFSET, 1,
 4604                     ip->i_mode);
 4605                 dotaddref->ja_state |= MKDIR_BODY;
 4606                 dotdotaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET,
 4607                     dp->i_effnlink - 1, dp->i_mode);
 4608                 dotdotaddref->ja_state |= MKDIR_PARENT;
 4609         }
 4610         ACQUIRE_LOCK(&lk);
 4611         inodedep = inodedep_lookup_ip(ip);
 4612         if (DOINGSUJ(dvp)) {
 4613                 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
 4614                     inoreflst);
 4615                 KASSERT(jaddref != NULL,
 4616                     ("softdep_setup_mkdir: No addref structure present."));
 4617                 KASSERT(jaddref->ja_parent == dp->i_number, 
 4618                     ("softdep_setup_mkdir: bad parent %ju",
 4619                     (uintmax_t)jaddref->ja_parent));
 4620                 TAILQ_INSERT_BEFORE(&jaddref->ja_ref, &dotaddref->ja_ref,
 4621                     if_deps);
 4622         }
 4623         inodedep = inodedep_lookup_ip(dp);
 4624         if (DOINGSUJ(dvp))
 4625                 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst,
 4626                     &dotdotaddref->ja_ref, if_deps);
 4627         softdep_prelink(ITOV(dp), NULL);
 4628         FREE_LOCK(&lk);
 4629 }
 4630 
 4631 /*
 4632  * Called to track nlinkdelta of the inode and parent directories prior to
 4633  * unlinking a directory.
 4634  */
 4635 void
 4636 softdep_setup_rmdir(dp, ip)
 4637         struct inode *dp;
 4638         struct inode *ip;
 4639 {
 4640         struct vnode *dvp;
 4641 
 4642         dvp = ITOV(dp);
 4643         ACQUIRE_LOCK(&lk);
 4644         (void) inodedep_lookup_ip(ip);
 4645         (void) inodedep_lookup_ip(dp);
 4646         softdep_prelink(dvp, ITOV(ip));
 4647         FREE_LOCK(&lk);
 4648 }
 4649 
 4650 /*
 4651  * Called to track nlinkdelta of the inode and parent directories prior to
 4652  * unlink.
 4653  */
 4654 void
 4655 softdep_setup_unlink(dp, ip)
 4656         struct inode *dp;
 4657         struct inode *ip;
 4658 {
 4659         struct vnode *dvp;
 4660 
 4661         dvp = ITOV(dp);
 4662         ACQUIRE_LOCK(&lk);
 4663         (void) inodedep_lookup_ip(ip);
 4664         (void) inodedep_lookup_ip(dp);
 4665         softdep_prelink(dvp, ITOV(ip));
 4666         FREE_LOCK(&lk);
 4667 }
 4668 
 4669 /*
 4670  * Called to release the journal structures created by a failed non-directory
 4671  * creation.  Adjusts nlinkdelta for non-journaling softdep.
 4672  */
 4673 void
 4674 softdep_revert_create(dp, ip)
 4675         struct inode *dp;
 4676         struct inode *ip;
 4677 {
 4678         struct inodedep *inodedep;
 4679         struct jaddref *jaddref;
 4680         struct vnode *dvp;
 4681 
 4682         dvp = ITOV(dp);
 4683         ACQUIRE_LOCK(&lk);
 4684         inodedep = inodedep_lookup_ip(ip);
 4685         if (DOINGSUJ(dvp)) {
 4686                 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
 4687                     inoreflst);
 4688                 KASSERT(jaddref->ja_parent == dp->i_number,
 4689                     ("softdep_revert_create: addref parent mismatch"));
 4690                 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
 4691         }
 4692         FREE_LOCK(&lk);
 4693 }
 4694 
 4695 /*
 4696  * Called to release the journal structures created by a failed dotdot link
 4697  * creation.  Adjusts nlinkdelta for non-journaling softdep.
 4698  */
 4699 void
 4700 softdep_revert_dotdot_link(dp, ip)
 4701         struct inode *dp;
 4702         struct inode *ip;
 4703 {
 4704         struct inodedep *inodedep;
 4705         struct jaddref *jaddref;
 4706         struct vnode *dvp;
 4707 
 4708         dvp = ITOV(dp);
 4709         ACQUIRE_LOCK(&lk);
 4710         inodedep = inodedep_lookup_ip(dp);
 4711         if (DOINGSUJ(dvp)) {
 4712                 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
 4713                     inoreflst);
 4714                 KASSERT(jaddref->ja_parent == ip->i_number,
 4715                     ("softdep_revert_dotdot_link: addref parent mismatch"));
 4716                 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
 4717         }
 4718         FREE_LOCK(&lk);
 4719 }
 4720 
 4721 /*
 4722  * Called to release the journal structures created by a failed link
 4723  * addition.  Adjusts nlinkdelta for non-journaling softdep.
 4724  */
 4725 void
 4726 softdep_revert_link(dp, ip)
 4727         struct inode *dp;
 4728         struct inode *ip;
 4729 {
 4730         struct inodedep *inodedep;
 4731         struct jaddref *jaddref;
 4732         struct vnode *dvp;
 4733 
 4734         dvp = ITOV(dp);
 4735         ACQUIRE_LOCK(&lk);
 4736         inodedep = inodedep_lookup_ip(ip);
 4737         if (DOINGSUJ(dvp)) {
 4738                 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
 4739                     inoreflst);
 4740                 KASSERT(jaddref->ja_parent == dp->i_number,
 4741                     ("softdep_revert_link: addref parent mismatch"));
 4742                 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
 4743         }
 4744         FREE_LOCK(&lk);
 4745 }
 4746 
 4747 /*
 4748  * Called to release the journal structures created by a failed mkdir
 4749  * attempt.  Adjusts nlinkdelta for non-journaling softdep.
 4750  */
 4751 void
 4752 softdep_revert_mkdir(dp, ip)
 4753         struct inode *dp;
 4754         struct inode *ip;
 4755 {
 4756         struct inodedep *inodedep;
 4757         struct jaddref *jaddref;
 4758         struct jaddref *dotaddref;
 4759         struct vnode *dvp;
 4760 
 4761         dvp = ITOV(dp);
 4762 
 4763         ACQUIRE_LOCK(&lk);
 4764         inodedep = inodedep_lookup_ip(dp);
 4765         if (DOINGSUJ(dvp)) {
 4766                 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
 4767                     inoreflst);
 4768                 KASSERT(jaddref->ja_parent == ip->i_number,
 4769                     ("softdep_revert_mkdir: dotdot addref parent mismatch"));
 4770                 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
 4771         }
 4772         inodedep = inodedep_lookup_ip(ip);
 4773         if (DOINGSUJ(dvp)) {
 4774                 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
 4775                     inoreflst);
 4776                 KASSERT(jaddref->ja_parent == dp->i_number,
 4777                     ("softdep_revert_mkdir: addref parent mismatch"));
 4778                 dotaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref,
 4779                     inoreflst, if_deps);
 4780                 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
 4781                 KASSERT(dotaddref->ja_parent == ip->i_number,
 4782                     ("softdep_revert_mkdir: dot addref parent mismatch"));
 4783                 cancel_jaddref(dotaddref, inodedep, &inodedep->id_inowait);
 4784         }
 4785         FREE_LOCK(&lk);
 4786 }
 4787 
 4788 /* 
 4789  * Called to correct nlinkdelta after a failed rmdir.
 4790  */
 4791 void
 4792 softdep_revert_rmdir(dp, ip)
 4793         struct inode *dp;
 4794         struct inode *ip;
 4795 {
 4796 
 4797         ACQUIRE_LOCK(&lk);
 4798         (void) inodedep_lookup_ip(ip);
 4799         (void) inodedep_lookup_ip(dp);
 4800         FREE_LOCK(&lk);
 4801 }
 4802 
 4803 /*
 4804  * Protecting the freemaps (or bitmaps).
 4805  * 
 4806  * To eliminate the need to execute fsck before mounting a filesystem
 4807  * after a power failure, one must (conservatively) guarantee that the
 4808  * on-disk copy of the bitmaps never indicate that a live inode or block is
 4809  * free.  So, when a block or inode is allocated, the bitmap should be
 4810  * updated (on disk) before any new pointers.  When a block or inode is
 4811  * freed, the bitmap should not be updated until all pointers have been
 4812  * reset.  The latter dependency is handled by the delayed de-allocation
 4813  * approach described below for block and inode de-allocation.  The former
 4814  * dependency is handled by calling the following procedure when a block or
 4815  * inode is allocated. When an inode is allocated an "inodedep" is created
 4816  * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
 4817  * Each "inodedep" is also inserted into the hash indexing structure so
 4818  * that any additional link additions can be made dependent on the inode
 4819  * allocation.
 4820  * 
 4821  * The ufs filesystem maintains a number of free block counts (e.g., per
 4822  * cylinder group, per cylinder and per <cylinder, rotational position> pair)
 4823  * in addition to the bitmaps.  These counts are used to improve efficiency
 4824  * during allocation and therefore must be consistent with the bitmaps.
 4825  * There is no convenient way to guarantee post-crash consistency of these
 4826  * counts with simple update ordering, for two main reasons: (1) The counts
 4827  * and bitmaps for a single cylinder group block are not in the same disk
 4828  * sector.  If a disk write is interrupted (e.g., by power failure), one may
 4829  * be written and the other not.  (2) Some of the counts are located in the
 4830  * superblock rather than the cylinder group block. So, we focus our soft
 4831  * updates implementation on protecting the bitmaps. When mounting a
 4832  * filesystem, we recompute the auxiliary counts from the bitmaps.
 4833  */
 4834 
 4835 /*
 4836  * Called just after updating the cylinder group block to allocate an inode.
 4837  */
 4838 void
 4839 softdep_setup_inomapdep(bp, ip, newinum, mode)
 4840         struct buf *bp;         /* buffer for cylgroup block with inode map */
 4841         struct inode *ip;       /* inode related to allocation */
 4842         ino_t newinum;          /* new inode number being allocated */
 4843         int mode;
 4844 {
 4845         struct inodedep *inodedep;
 4846         struct bmsafemap *bmsafemap;
 4847         struct jaddref *jaddref;
 4848         struct mount *mp;
 4849         struct fs *fs;
 4850 
 4851         mp = UFSTOVFS(ip->i_ump);
 4852         fs = ip->i_ump->um_fs;
 4853         jaddref = NULL;
 4854 
 4855         /*
 4856          * Allocate the journal reference add structure so that the bitmap
 4857          * can be dependent on it.
 4858          */
 4859         if (MOUNTEDSUJ(mp)) {
 4860                 jaddref = newjaddref(ip, newinum, 0, 0, mode);
 4861                 jaddref->ja_state |= NEWBLOCK;
 4862         }
 4863 
 4864         /*
 4865          * Create a dependency for the newly allocated inode.
 4866          * Panic if it already exists as something is seriously wrong.
 4867          * Otherwise add it to the dependency list for the buffer holding
 4868          * the cylinder group map from which it was allocated.
 4869          *
 4870          * We have to preallocate a bmsafemap entry in case it is needed
 4871          * in bmsafemap_lookup since once we allocate the inodedep, we
 4872          * have to finish initializing it before we can FREE_LOCK().
 4873          * By preallocating, we avoid FREE_LOCK() while doing a malloc
 4874          * in bmsafemap_lookup. We cannot call bmsafemap_lookup before
 4875          * creating the inodedep as it can be freed during the time
 4876          * that we FREE_LOCK() while allocating the inodedep. We must
 4877          * call workitem_alloc() before entering the locked section as
 4878          * it also acquires the lock and we must avoid trying doing so
 4879          * recursively.
 4880          */
 4881         bmsafemap = malloc(sizeof(struct bmsafemap),
 4882             M_BMSAFEMAP, M_SOFTDEP_FLAGS);
 4883         workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp);
 4884         ACQUIRE_LOCK(&lk);
 4885         if ((inodedep_lookup(mp, newinum, DEPALLOC | NODELAY, &inodedep)))
 4886                 panic("softdep_setup_inomapdep: dependency %p for new"
 4887                     "inode already exists", inodedep);
 4888         bmsafemap = bmsafemap_lookup(mp, bp, ino_to_cg(fs, newinum), bmsafemap);
 4889         if (jaddref) {
 4890                 LIST_INSERT_HEAD(&bmsafemap->sm_jaddrefhd, jaddref, ja_bmdeps);
 4891                 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref,
 4892                     if_deps);
 4893         } else {
 4894                 inodedep->id_state |= ONDEPLIST;
 4895                 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
 4896         }
 4897         inodedep->id_bmsafemap = bmsafemap;
 4898         inodedep->id_state &= ~DEPCOMPLETE;
 4899         FREE_LOCK(&lk);
 4900 }
 4901 
 4902 /*
 4903  * Called just after updating the cylinder group block to
 4904  * allocate block or fragment.
 4905  */
 4906 void
 4907 softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags)
 4908         struct buf *bp;         /* buffer for cylgroup block with block map */
 4909         struct mount *mp;       /* filesystem doing allocation */
 4910         ufs2_daddr_t newblkno;  /* number of newly allocated block */
 4911         int frags;              /* Number of fragments. */
 4912         int oldfrags;           /* Previous number of fragments for extend. */
 4913 {
 4914         struct newblk *newblk;
 4915         struct bmsafemap *bmsafemap;
 4916         struct jnewblk *jnewblk;
 4917         struct fs *fs;
 4918 
 4919         fs = VFSTOUFS(mp)->um_fs;
 4920         jnewblk = NULL;
 4921         /*
 4922          * Create a dependency for the newly allocated block.
 4923          * Add it to the dependency list for the buffer holding
 4924          * the cylinder group map from which it was allocated.
 4925          */
 4926         if (MOUNTEDSUJ(mp)) {
 4927                 jnewblk = malloc(sizeof(*jnewblk), M_JNEWBLK, M_SOFTDEP_FLAGS);
 4928                 workitem_alloc(&jnewblk->jn_list, D_JNEWBLK, mp);
 4929                 jnewblk->jn_jsegdep = newjsegdep(&jnewblk->jn_list);
 4930                 jnewblk->jn_state = ATTACHED;
 4931                 jnewblk->jn_blkno = newblkno;
 4932                 jnewblk->jn_frags = frags;
 4933                 jnewblk->jn_oldfrags = oldfrags;
 4934 #ifdef SUJ_DEBUG
 4935                 {
 4936                         struct cg *cgp;
 4937                         uint8_t *blksfree;
 4938                         long bno;
 4939                         int i;
 4940         
 4941                         cgp = (struct cg *)bp->b_data;
 4942                         blksfree = cg_blksfree(cgp);
 4943                         bno = dtogd(fs, jnewblk->jn_blkno);
 4944                         for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags;
 4945                             i++) {
 4946                                 if (isset(blksfree, bno + i))
 4947                                         panic("softdep_setup_blkmapdep: "
 4948                                             "free fragment %d from %d-%d "
 4949                                             "state 0x%X dep %p", i,
 4950                                             jnewblk->jn_oldfrags,
 4951                                             jnewblk->jn_frags,
 4952                                             jnewblk->jn_state,
 4953                                             jnewblk->jn_dep);
 4954                         }
 4955                 }
 4956 #endif
 4957         }
 4958 
 4959         CTR3(KTR_SUJ,
 4960             "softdep_setup_blkmapdep: blkno %jd frags %d oldfrags %d",
 4961             newblkno, frags, oldfrags);
 4962         ACQUIRE_LOCK(&lk);
 4963         if (newblk_lookup(mp, newblkno, DEPALLOC, &newblk) != 0)
 4964                 panic("softdep_setup_blkmapdep: found block");
 4965         newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp,
 4966             dtog(fs, newblkno), NULL);
 4967         if (jnewblk) {
 4968                 jnewblk->jn_dep = (struct worklist *)newblk;
 4969                 LIST_INSERT_HEAD(&bmsafemap->sm_jnewblkhd, jnewblk, jn_deps);
 4970         } else {
 4971                 newblk->nb_state |= ONDEPLIST;
 4972                 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
 4973         }
 4974         newblk->nb_bmsafemap = bmsafemap;
 4975         newblk->nb_jnewblk = jnewblk;
 4976         FREE_LOCK(&lk);
 4977 }
 4978 
 4979 #define BMSAFEMAP_HASH(fs, cg) \
 4980       (&bmsafemap_hashtbl[((((register_t)(fs)) >> 13) + (cg)) & bmsafemap_hash])
 4981 
 4982 static int
 4983 bmsafemap_find(bmsafemaphd, mp, cg, bmsafemapp)
 4984         struct bmsafemap_hashhead *bmsafemaphd;
 4985         struct mount *mp;
 4986         int cg;
 4987         struct bmsafemap **bmsafemapp;
 4988 {
 4989         struct bmsafemap *bmsafemap;
 4990 
 4991         LIST_FOREACH(bmsafemap, bmsafemaphd, sm_hash)
 4992                 if (bmsafemap->sm_list.wk_mp == mp && bmsafemap->sm_cg == cg)
 4993                         break;
 4994         if (bmsafemap) {
 4995                 *bmsafemapp = bmsafemap;
 4996                 return (1);
 4997         }
 4998         *bmsafemapp = NULL;
 4999 
 5000         return (0);
 5001 }
 5002 
 5003 /*
 5004  * Find the bmsafemap associated with a cylinder group buffer.
 5005  * If none exists, create one. The buffer must be locked when
 5006  * this routine is called and this routine must be called with
 5007  * the softdep lock held. To avoid giving up the lock while
 5008  * allocating a new bmsafemap, a preallocated bmsafemap may be
 5009  * provided. If it is provided but not needed, it is freed.
 5010  */
 5011 static struct bmsafemap *
 5012 bmsafemap_lookup(mp, bp, cg, newbmsafemap)
 5013         struct mount *mp;
 5014         struct buf *bp;
 5015         int cg;
 5016         struct bmsafemap *newbmsafemap;
 5017 {
 5018         struct bmsafemap_hashhead *bmsafemaphd;
 5019         struct bmsafemap *bmsafemap, *collision;
 5020         struct worklist *wk;
 5021         struct fs *fs;
 5022 
 5023         rw_assert(&lk, RA_WLOCKED);
 5024         KASSERT(bp != NULL, ("bmsafemap_lookup: missing buffer"));
 5025         LIST_FOREACH(wk, &bp->b_dep, wk_list) {
 5026                 if (wk->wk_type == D_BMSAFEMAP) {
 5027                         if (newbmsafemap)
 5028                                 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP);
 5029                         return (WK_BMSAFEMAP(wk));
 5030                 }
 5031         }
 5032         fs = VFSTOUFS(mp)->um_fs;
 5033         bmsafemaphd = BMSAFEMAP_HASH(fs, cg);
 5034         if (bmsafemap_find(bmsafemaphd, mp, cg, &bmsafemap) == 1) {
 5035                 if (newbmsafemap)
 5036                         WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP);
 5037                 return (bmsafemap);
 5038         }
 5039         if (newbmsafemap) {
 5040                 bmsafemap = newbmsafemap;
 5041         } else {
 5042                 FREE_LOCK(&lk);
 5043                 bmsafemap = malloc(sizeof(struct bmsafemap),
 5044                         M_BMSAFEMAP, M_SOFTDEP_FLAGS);
 5045                 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp);
 5046                 ACQUIRE_LOCK(&lk);
 5047         }
 5048         bmsafemap->sm_buf = bp;
 5049         LIST_INIT(&bmsafemap->sm_inodedephd);
 5050         LIST_INIT(&bmsafemap->sm_inodedepwr);
 5051         LIST_INIT(&bmsafemap->sm_newblkhd);
 5052         LIST_INIT(&bmsafemap->sm_newblkwr);
 5053         LIST_INIT(&bmsafemap->sm_jaddrefhd);
 5054         LIST_INIT(&bmsafemap->sm_jnewblkhd);
 5055         LIST_INIT(&bmsafemap->sm_freehd);
 5056         LIST_INIT(&bmsafemap->sm_freewr);
 5057         if (bmsafemap_find(bmsafemaphd, mp, cg, &collision) == 1) {
 5058                 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
 5059                 return (collision);
 5060         }
 5061         bmsafemap->sm_cg = cg;
 5062         LIST_INSERT_HEAD(bmsafemaphd, bmsafemap, sm_hash);
 5063         LIST_INSERT_HEAD(&VFSTOUFS(mp)->softdep_dirtycg, bmsafemap, sm_next);
 5064         WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
 5065         return (bmsafemap);
 5066 }
 5067 
 5068 /*
 5069  * Direct block allocation dependencies.
 5070  * 
 5071  * When a new block is allocated, the corresponding disk locations must be
 5072  * initialized (with zeros or new data) before the on-disk inode points to
 5073  * them.  Also, the freemap from which the block was allocated must be
 5074  * updated (on disk) before the inode's pointer. These two dependencies are
 5075  * independent of each other and are needed for all file blocks and indirect
 5076  * blocks that are pointed to directly by the inode.  Just before the
 5077  * "in-core" version of the inode is updated with a newly allocated block
 5078  * number, a procedure (below) is called to setup allocation dependency
 5079  * structures.  These structures are removed when the corresponding
 5080  * dependencies are satisfied or when the block allocation becomes obsolete
 5081  * (i.e., the file is deleted, the block is de-allocated, or the block is a
 5082  * fragment that gets upgraded).  All of these cases are handled in
 5083  * procedures described later.
 5084  * 
 5085  * When a file extension causes a fragment to be upgraded, either to a larger
 5086  * fragment or to a full block, the on-disk location may change (if the
 5087  * previous fragment could not simply be extended). In this case, the old
 5088  * fragment must be de-allocated, but not until after the inode's pointer has
 5089  * been updated. In most cases, this is handled by later procedures, which
 5090  * will construct a "freefrag" structure to be added to the workitem queue
 5091  * when the inode update is complete (or obsolete).  The main exception to
 5092  * this is when an allocation occurs while a pending allocation dependency
 5093  * (for the same block pointer) remains.  This case is handled in the main
 5094  * allocation dependency setup procedure by immediately freeing the
 5095  * unreferenced fragments.
 5096  */ 
 5097 void 
 5098 softdep_setup_allocdirect(ip, off, newblkno, oldblkno, newsize, oldsize, bp)
 5099         struct inode *ip;       /* inode to which block is being added */
 5100         ufs_lbn_t off;          /* block pointer within inode */
 5101         ufs2_daddr_t newblkno;  /* disk block number being added */
 5102         ufs2_daddr_t oldblkno;  /* previous block number, 0 unless frag */
 5103         long newsize;           /* size of new block */
 5104         long oldsize;           /* size of new block */
 5105         struct buf *bp;         /* bp for allocated block */
 5106 {
 5107         struct allocdirect *adp, *oldadp;
 5108         struct allocdirectlst *adphead;
 5109         struct freefrag *freefrag;
 5110         struct inodedep *inodedep;
 5111         struct pagedep *pagedep;
 5112         struct jnewblk *jnewblk;
 5113         struct newblk *newblk;
 5114         struct mount *mp;
 5115         ufs_lbn_t lbn;
 5116 
 5117         lbn = bp->b_lblkno;
 5118         mp = UFSTOVFS(ip->i_ump);
 5119         if (oldblkno && oldblkno != newblkno)
 5120                 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn);
 5121         else
 5122                 freefrag = NULL;
 5123 
 5124         CTR6(KTR_SUJ,
 5125             "softdep_setup_allocdirect: ino %d blkno %jd oldblkno %jd "
 5126             "off %jd newsize %ld oldsize %d",
 5127             ip->i_number, newblkno, oldblkno, off, newsize, oldsize);
 5128         ACQUIRE_LOCK(&lk);
 5129         if (off >= NDADDR) {
 5130                 if (lbn > 0)
 5131                         panic("softdep_setup_allocdirect: bad lbn %jd, off %jd",
 5132                             lbn, off);
 5133                 /* allocating an indirect block */
 5134                 if (oldblkno != 0)
 5135                         panic("softdep_setup_allocdirect: non-zero indir");
 5136         } else {
 5137                 if (off != lbn)
 5138                         panic("softdep_setup_allocdirect: lbn %jd != off %jd",
 5139                             lbn, off);
 5140                 /*
 5141                  * Allocating a direct block.
 5142                  *
 5143                  * If we are allocating a directory block, then we must
 5144                  * allocate an associated pagedep to track additions and
 5145                  * deletions.
 5146                  */
 5147                 if ((ip->i_mode & IFMT) == IFDIR)
 5148                         pagedep_lookup(mp, bp, ip->i_number, off, DEPALLOC,
 5149                             &pagedep);
 5150         }
 5151         if (newblk_lookup(mp, newblkno, 0, &newblk) == 0)
 5152                 panic("softdep_setup_allocdirect: lost block");
 5153         KASSERT(newblk->nb_list.wk_type == D_NEWBLK,
 5154             ("softdep_setup_allocdirect: newblk already initialized"));
 5155         /*
 5156          * Convert the newblk to an allocdirect.
 5157          */
 5158         WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT);
 5159         adp = (struct allocdirect *)newblk;
 5160         newblk->nb_freefrag = freefrag;
 5161         adp->ad_offset = off;
 5162         adp->ad_oldblkno = oldblkno;
 5163         adp->ad_newsize = newsize;
 5164         adp->ad_oldsize = oldsize;
 5165 
 5166         /*
 5167          * Finish initializing the journal.
 5168          */
 5169         if ((jnewblk = newblk->nb_jnewblk) != NULL) {
 5170                 jnewblk->jn_ino = ip->i_number;
 5171                 jnewblk->jn_lbn = lbn;
 5172                 add_to_journal(&jnewblk->jn_list);
 5173         }
 5174         if (freefrag && freefrag->ff_jdep != NULL &&
 5175             freefrag->ff_jdep->wk_type == D_JFREEFRAG)
 5176                 add_to_journal(freefrag->ff_jdep);
 5177         inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep);
 5178         adp->ad_inodedep = inodedep;
 5179 
 5180         WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list);
 5181         /*
 5182          * The list of allocdirects must be kept in sorted and ascending
 5183          * order so that the rollback routines can quickly determine the
 5184          * first uncommitted block (the size of the file stored on disk
 5185          * ends at the end of the lowest committed fragment, or if there
 5186          * are no fragments, at the end of the highest committed block).
 5187          * Since files generally grow, the typical case is that the new
 5188          * block is to be added at the end of the list. We speed this
 5189          * special case by checking against the last allocdirect in the
 5190          * list before laboriously traversing the list looking for the
 5191          * insertion point.
 5192          */
 5193         adphead = &inodedep->id_newinoupdt;
 5194         oldadp = TAILQ_LAST(adphead, allocdirectlst);
 5195         if (oldadp == NULL || oldadp->ad_offset <= off) {
 5196                 /* insert at end of list */
 5197                 TAILQ_INSERT_TAIL(adphead, adp, ad_next);
 5198                 if (oldadp != NULL && oldadp->ad_offset == off)
 5199                         allocdirect_merge(adphead, adp, oldadp);
 5200                 FREE_LOCK(&lk);
 5201                 return;
 5202         }
 5203         TAILQ_FOREACH(oldadp, adphead, ad_next) {
 5204                 if (oldadp->ad_offset >= off)
 5205                         break;
 5206         }
 5207         if (oldadp == NULL)
 5208                 panic("softdep_setup_allocdirect: lost entry");
 5209         /* insert in middle of list */
 5210         TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
 5211         if (oldadp->ad_offset == off)
 5212                 allocdirect_merge(adphead, adp, oldadp);
 5213 
 5214         FREE_LOCK(&lk);
 5215 }
 5216 
 5217 /*
 5218  * Merge a newer and older journal record to be stored either in a
 5219  * newblock or freefrag.  This handles aggregating journal records for
 5220  * fragment allocation into a second record as well as replacing a
 5221  * journal free with an aborted journal allocation.  A segment for the
 5222  * oldest record will be placed on wkhd if it has been written.  If not
 5223  * the segment for the newer record will suffice.
 5224  */
 5225 static struct worklist *
 5226 jnewblk_merge(new, old, wkhd)
 5227         struct worklist *new;
 5228         struct worklist *old;
 5229         struct workhead *wkhd;
 5230 {
 5231         struct jnewblk *njnewblk;
 5232         struct jnewblk *jnewblk;
 5233 
 5234         /* Handle NULLs to simplify callers. */
 5235         if (new == NULL)
 5236                 return (old);
 5237         if (old == NULL)
 5238                 return (new);
 5239         /* Replace a jfreefrag with a jnewblk. */
 5240         if (new->wk_type == D_JFREEFRAG) {
 5241                 if (WK_JNEWBLK(old)->jn_blkno != WK_JFREEFRAG(new)->fr_blkno)
 5242                         panic("jnewblk_merge: blkno mismatch: %p, %p",
 5243                             old, new);
 5244                 cancel_jfreefrag(WK_JFREEFRAG(new));
 5245                 return (old);
 5246         }
 5247         if (old->wk_type != D_JNEWBLK || new->wk_type != D_JNEWBLK)
 5248                 panic("jnewblk_merge: Bad type: old %d new %d\n",
 5249                     old->wk_type, new->wk_type);
 5250         /*
 5251          * Handle merging of two jnewblk records that describe
 5252          * different sets of fragments in the same block.
 5253          */
 5254         jnewblk = WK_JNEWBLK(old);
 5255         njnewblk = WK_JNEWBLK(new);
 5256         if (jnewblk->jn_blkno != njnewblk->jn_blkno)
 5257                 panic("jnewblk_merge: Merging disparate blocks.");
 5258         /*
 5259          * The record may be rolled back in the cg.
 5260          */
 5261         if (jnewblk->jn_state & UNDONE) {
 5262                 jnewblk->jn_state &= ~UNDONE;
 5263                 njnewblk->jn_state |= UNDONE;
 5264                 njnewblk->jn_state &= ~ATTACHED;
 5265         }
 5266         /*
 5267          * We modify the newer addref and free the older so that if neither
 5268          * has been written the most up-to-date copy will be on disk.  If
 5269          * both have been written but rolled back we only temporarily need
 5270          * one of them to fix the bits when the cg write completes.
 5271          */
 5272         jnewblk->jn_state |= ATTACHED | COMPLETE;
 5273         njnewblk->jn_oldfrags = jnewblk->jn_oldfrags;
 5274         cancel_jnewblk(jnewblk, wkhd);
 5275         WORKLIST_REMOVE(&jnewblk->jn_list);
 5276         free_jnewblk(jnewblk);
 5277         return (new);
 5278 }
 5279 
 5280 /*
 5281  * Replace an old allocdirect dependency with a newer one.
 5282  * This routine must be called with splbio interrupts blocked.
 5283  */
 5284 static void
 5285 allocdirect_merge(adphead, newadp, oldadp)
 5286         struct allocdirectlst *adphead; /* head of list holding allocdirects */
 5287         struct allocdirect *newadp;     /* allocdirect being added */
 5288         struct allocdirect *oldadp;     /* existing allocdirect being checked */
 5289 {
 5290         struct worklist *wk;
 5291         struct freefrag *freefrag;
 5292 
 5293         freefrag = NULL;
 5294         rw_assert(&lk, RA_WLOCKED);
 5295         if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
 5296             newadp->ad_oldsize != oldadp->ad_newsize ||
 5297             newadp->ad_offset >= NDADDR)
 5298                 panic("%s %jd != new %jd || old size %ld != new %ld",
 5299                     "allocdirect_merge: old blkno",
 5300                     (intmax_t)newadp->ad_oldblkno,
 5301                     (intmax_t)oldadp->ad_newblkno,
 5302                     newadp->ad_oldsize, oldadp->ad_newsize);
 5303         newadp->ad_oldblkno = oldadp->ad_oldblkno;
 5304         newadp->ad_oldsize = oldadp->ad_oldsize;
 5305         /*
 5306          * If the old dependency had a fragment to free or had never
 5307          * previously had a block allocated, then the new dependency
 5308          * can immediately post its freefrag and adopt the old freefrag.
 5309          * This action is done by swapping the freefrag dependencies.
 5310          * The new dependency gains the old one's freefrag, and the
 5311          * old one gets the new one and then immediately puts it on
 5312          * the worklist when it is freed by free_newblk. It is
 5313          * not possible to do this swap when the old dependency had a
 5314          * non-zero size but no previous fragment to free. This condition
 5315          * arises when the new block is an extension of the old block.
 5316          * Here, the first part of the fragment allocated to the new
 5317          * dependency is part of the block currently claimed on disk by
 5318          * the old dependency, so cannot legitimately be freed until the
 5319          * conditions for the new dependency are fulfilled.
 5320          */
 5321         freefrag = newadp->ad_freefrag;
 5322         if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
 5323                 newadp->ad_freefrag = oldadp->ad_freefrag;
 5324                 oldadp->ad_freefrag = freefrag;
 5325         }
 5326         /*
 5327          * If we are tracking a new directory-block allocation,
 5328          * move it from the old allocdirect to the new allocdirect.
 5329          */
 5330         if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) {
 5331                 WORKLIST_REMOVE(wk);
 5332                 if (!LIST_EMPTY(&oldadp->ad_newdirblk))
 5333                         panic("allocdirect_merge: extra newdirblk");
 5334                 WORKLIST_INSERT(&newadp->ad_newdirblk, wk);
 5335         }
 5336         TAILQ_REMOVE(adphead, oldadp, ad_next);
 5337         /*
 5338          * We need to move any journal dependencies over to the freefrag
 5339          * that releases this block if it exists.  Otherwise we are
 5340          * extending an existing block and we'll wait until that is
 5341          * complete to release the journal space and extend the
 5342          * new journal to cover this old space as well.
 5343          */
 5344         if (freefrag == NULL) {
 5345                 if (oldadp->ad_newblkno != newadp->ad_newblkno)
 5346                         panic("allocdirect_merge: %jd != %jd",
 5347                             oldadp->ad_newblkno, newadp->ad_newblkno);
 5348                 newadp->ad_block.nb_jnewblk = (struct jnewblk *)
 5349                     jnewblk_merge(&newadp->ad_block.nb_jnewblk->jn_list, 
 5350                     &oldadp->ad_block.nb_jnewblk->jn_list,
 5351                     &newadp->ad_block.nb_jwork);
 5352                 oldadp->ad_block.nb_jnewblk = NULL;
 5353                 cancel_newblk(&oldadp->ad_block, NULL,
 5354                     &newadp->ad_block.nb_jwork);
 5355         } else {
 5356                 wk = (struct worklist *) cancel_newblk(&oldadp->ad_block,
 5357                     &freefrag->ff_list, &freefrag->ff_jwork);
 5358                 freefrag->ff_jdep = jnewblk_merge(freefrag->ff_jdep, wk,
 5359                     &freefrag->ff_jwork);
 5360         }
 5361         free_newblk(&oldadp->ad_block);
 5362 }
 5363 
 5364 /*
 5365  * Allocate a jfreefrag structure to journal a single block free.
 5366  */
 5367 static struct jfreefrag *
 5368 newjfreefrag(freefrag, ip, blkno, size, lbn)
 5369         struct freefrag *freefrag;
 5370         struct inode *ip;
 5371         ufs2_daddr_t blkno;
 5372         long size;
 5373         ufs_lbn_t lbn;
 5374 {
 5375         struct jfreefrag *jfreefrag;
 5376         struct fs *fs;
 5377 
 5378         fs = ip->i_fs;
 5379         jfreefrag = malloc(sizeof(struct jfreefrag), M_JFREEFRAG,
 5380             M_SOFTDEP_FLAGS);
 5381         workitem_alloc(&jfreefrag->fr_list, D_JFREEFRAG, UFSTOVFS(ip->i_ump));
 5382         jfreefrag->fr_jsegdep = newjsegdep(&jfreefrag->fr_list);
 5383         jfreefrag->fr_state = ATTACHED | DEPCOMPLETE;
 5384         jfreefrag->fr_ino = ip->i_number;
 5385         jfreefrag->fr_lbn = lbn;
 5386         jfreefrag->fr_blkno = blkno;
 5387         jfreefrag->fr_frags = numfrags(fs, size);
 5388         jfreefrag->fr_freefrag = freefrag;
 5389 
 5390         return (jfreefrag);
 5391 }
 5392 
 5393 /*
 5394  * Allocate a new freefrag structure.
 5395  */
 5396 static struct freefrag *
 5397 newfreefrag(ip, blkno, size, lbn)
 5398         struct inode *ip;
 5399         ufs2_daddr_t blkno;
 5400         long size;
 5401         ufs_lbn_t lbn;
 5402 {
 5403         struct freefrag *freefrag;
 5404         struct fs *fs;
 5405 
 5406         CTR4(KTR_SUJ, "newfreefrag: ino %d blkno %jd size %ld lbn %jd",
 5407             ip->i_number, blkno, size, lbn);
 5408         fs = ip->i_fs;
 5409         if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
 5410                 panic("newfreefrag: frag size");
 5411         freefrag = malloc(sizeof(struct freefrag),
 5412             M_FREEFRAG, M_SOFTDEP_FLAGS);
 5413         workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ip->i_ump));
 5414         freefrag->ff_state = ATTACHED;
 5415         LIST_INIT(&freefrag->ff_jwork);
 5416         freefrag->ff_inum = ip->i_number;
 5417         freefrag->ff_vtype = ITOV(ip)->v_type;
 5418         freefrag->ff_blkno = blkno;
 5419         freefrag->ff_fragsize = size;
 5420 
 5421         if (MOUNTEDSUJ(UFSTOVFS(ip->i_ump))) {
 5422                 freefrag->ff_jdep = (struct worklist *)
 5423                     newjfreefrag(freefrag, ip, blkno, size, lbn);
 5424         } else {
 5425                 freefrag->ff_state |= DEPCOMPLETE;
 5426                 freefrag->ff_jdep = NULL;
 5427         }
 5428 
 5429         return (freefrag);
 5430 }
 5431 
 5432 /*
 5433  * This workitem de-allocates fragments that were replaced during
 5434  * file block allocation.
 5435  */
 5436 static void 
 5437 handle_workitem_freefrag(freefrag)
 5438         struct freefrag *freefrag;
 5439 {
 5440         struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp);
 5441         struct workhead wkhd;
 5442 
 5443         CTR3(KTR_SUJ,
 5444             "handle_workitem_freefrag: ino %d blkno %jd size %ld",
 5445             freefrag->ff_inum, freefrag->ff_blkno, freefrag->ff_fragsize);
 5446         /*
 5447          * It would be illegal to add new completion items to the
 5448          * freefrag after it was schedule to be done so it must be
 5449          * safe to modify the list head here.
 5450          */
 5451         LIST_INIT(&wkhd);
 5452         ACQUIRE_LOCK(&lk);
 5453         LIST_SWAP(&freefrag->ff_jwork, &wkhd, worklist, wk_list);
 5454         /*
 5455          * If the journal has not been written we must cancel it here.
 5456          */
 5457         if (freefrag->ff_jdep) {
 5458                 if (freefrag->ff_jdep->wk_type != D_JNEWBLK)
 5459                         panic("handle_workitem_freefrag: Unexpected type %d\n",
 5460                             freefrag->ff_jdep->wk_type);
 5461                 cancel_jnewblk(WK_JNEWBLK(freefrag->ff_jdep), &wkhd);
 5462         }
 5463         FREE_LOCK(&lk);
 5464         ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno,
 5465            freefrag->ff_fragsize, freefrag->ff_inum, freefrag->ff_vtype, &wkhd);
 5466         ACQUIRE_LOCK(&lk);
 5467         WORKITEM_FREE(freefrag, D_FREEFRAG);
 5468         FREE_LOCK(&lk);
 5469 }
 5470 
 5471 /*
 5472  * Set up a dependency structure for an external attributes data block.
 5473  * This routine follows much of the structure of softdep_setup_allocdirect.
 5474  * See the description of softdep_setup_allocdirect above for details.
 5475  */
 5476 void 
 5477 softdep_setup_allocext(ip, off, newblkno, oldblkno, newsize, oldsize, bp)
 5478         struct inode *ip;
 5479         ufs_lbn_t off;
 5480         ufs2_daddr_t newblkno;
 5481         ufs2_daddr_t oldblkno;
 5482         long newsize;
 5483         long oldsize;
 5484         struct buf *bp;
 5485 {
 5486         struct allocdirect *adp, *oldadp;
 5487         struct allocdirectlst *adphead;
 5488         struct freefrag *freefrag;
 5489         struct inodedep *inodedep;
 5490         struct jnewblk *jnewblk;
 5491         struct newblk *newblk;
 5492         struct mount *mp;
 5493         ufs_lbn_t lbn;
 5494 
 5495         if (off >= NXADDR)
 5496                 panic("softdep_setup_allocext: lbn %lld > NXADDR",
 5497                     (long long)off);
 5498 
 5499         lbn = bp->b_lblkno;
 5500         mp = UFSTOVFS(ip->i_ump);
 5501         if (oldblkno && oldblkno != newblkno)
 5502                 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn);
 5503         else
 5504                 freefrag = NULL;
 5505 
 5506         ACQUIRE_LOCK(&lk);
 5507         if (newblk_lookup(mp, newblkno, 0, &newblk) == 0)
 5508                 panic("softdep_setup_allocext: lost block");
 5509         KASSERT(newblk->nb_list.wk_type == D_NEWBLK,
 5510             ("softdep_setup_allocext: newblk already initialized"));
 5511         /*
 5512          * Convert the newblk to an allocdirect.
 5513          */
 5514         WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT);
 5515         adp = (struct allocdirect *)newblk;
 5516         newblk->nb_freefrag = freefrag;
 5517         adp->ad_offset = off;
 5518         adp->ad_oldblkno = oldblkno;
 5519         adp->ad_newsize = newsize;
 5520         adp->ad_oldsize = oldsize;
 5521         adp->ad_state |=  EXTDATA;
 5522 
 5523         /*
 5524          * Finish initializing the journal.
 5525          */
 5526         if ((jnewblk = newblk->nb_jnewblk) != NULL) {
 5527                 jnewblk->jn_ino = ip->i_number;
 5528                 jnewblk->jn_lbn = lbn;
 5529                 add_to_journal(&jnewblk->jn_list);
 5530         }
 5531         if (freefrag && freefrag->ff_jdep != NULL &&
 5532             freefrag->ff_jdep->wk_type == D_JFREEFRAG)
 5533                 add_to_journal(freefrag->ff_jdep);
 5534         inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep);
 5535         adp->ad_inodedep = inodedep;
 5536 
 5537         WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list);
 5538         /*
 5539          * The list of allocdirects must be kept in sorted and ascending
 5540          * order so that the rollback routines can quickly determine the
 5541          * first uncommitted block (the size of the file stored on disk
 5542          * ends at the end of the lowest committed fragment, or if there
 5543          * are no fragments, at the end of the highest committed block).
 5544          * Since files generally grow, the typical case is that the new
 5545          * block is to be added at the end of the list. We speed this
 5546          * special case by checking against the last allocdirect in the
 5547          * list before laboriously traversing the list looking for the
 5548          * insertion point.
 5549          */
 5550         adphead = &inodedep->id_newextupdt;
 5551         oldadp = TAILQ_LAST(adphead, allocdirectlst);
 5552         if (oldadp == NULL || oldadp->ad_offset <= off) {
 5553                 /* insert at end of list */
 5554                 TAILQ_INSERT_TAIL(adphead, adp, ad_next);
 5555                 if (oldadp != NULL && oldadp->ad_offset == off)
 5556                         allocdirect_merge(adphead, adp, oldadp);
 5557                 FREE_LOCK(&lk);
 5558                 return;
 5559         }
 5560         TAILQ_FOREACH(oldadp, adphead, ad_next) {
 5561                 if (oldadp->ad_offset >= off)
 5562                         break;
 5563         }
 5564         if (oldadp == NULL)
 5565                 panic("softdep_setup_allocext: lost entry");
 5566         /* insert in middle of list */
 5567         TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
 5568         if (oldadp->ad_offset == off)
 5569                 allocdirect_merge(adphead, adp, oldadp);
 5570         FREE_LOCK(&lk);
 5571 }
 5572 
 5573 /*
 5574  * Indirect block allocation dependencies.
 5575  * 
 5576  * The same dependencies that exist for a direct block also exist when
 5577  * a new block is allocated and pointed to by an entry in a block of
 5578  * indirect pointers. The undo/redo states described above are also
 5579  * used here. Because an indirect block contains many pointers that
 5580  * may have dependencies, a second copy of the entire in-memory indirect
 5581  * block is kept. The buffer cache copy is always completely up-to-date.
 5582  * The second copy, which is used only as a source for disk writes,
 5583  * contains only the safe pointers (i.e., those that have no remaining
 5584  * update dependencies). The second copy is freed when all pointers
 5585  * are safe. The cache is not allowed to replace indirect blocks with
 5586  * pending update dependencies. If a buffer containing an indirect
 5587  * block with dependencies is written, these routines will mark it
 5588  * dirty again. It can only be successfully written once all the
 5589  * dependencies are removed. The ffs_fsync routine in conjunction with
 5590  * softdep_sync_metadata work together to get all the dependencies
 5591  * removed so that a file can be successfully written to disk. Three
 5592  * procedures are used when setting up indirect block pointer
 5593  * dependencies. The division is necessary because of the organization
 5594  * of the "balloc" routine and because of the distinction between file
 5595  * pages and file metadata blocks.
 5596  */
 5597 
 5598 /*
 5599  * Allocate a new allocindir structure.
 5600  */
 5601 static struct allocindir *
 5602 newallocindir(ip, ptrno, newblkno, oldblkno, lbn)
 5603         struct inode *ip;       /* inode for file being extended */
 5604         int ptrno;              /* offset of pointer in indirect block */
 5605         ufs2_daddr_t newblkno;  /* disk block number being added */
 5606         ufs2_daddr_t oldblkno;  /* previous block number, 0 if none */
 5607         ufs_lbn_t lbn;
 5608 {
 5609         struct newblk *newblk;
 5610         struct allocindir *aip;
 5611         struct freefrag *freefrag;
 5612         struct jnewblk *jnewblk;
 5613 
 5614         if (oldblkno)
 5615                 freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize, lbn);
 5616         else
 5617                 freefrag = NULL;
 5618         ACQUIRE_LOCK(&lk);
 5619         if (newblk_lookup(UFSTOVFS(ip->i_ump), newblkno, 0, &newblk) == 0)
 5620                 panic("new_allocindir: lost block");
 5621         KASSERT(newblk->nb_list.wk_type == D_NEWBLK,
 5622             ("newallocindir: newblk already initialized"));
 5623         WORKITEM_REASSIGN(newblk, D_ALLOCINDIR);
 5624         newblk->nb_freefrag = freefrag;
 5625         aip = (struct allocindir *)newblk;
 5626         aip->ai_offset = ptrno;
 5627         aip->ai_oldblkno = oldblkno;
 5628         aip->ai_lbn = lbn;
 5629         if ((jnewblk = newblk->nb_jnewblk) != NULL) {
 5630                 jnewblk->jn_ino = ip->i_number;
 5631                 jnewblk->jn_lbn = lbn;
 5632                 add_to_journal(&jnewblk->jn_list);
 5633         }
 5634         if (freefrag && freefrag->ff_jdep != NULL &&
 5635             freefrag->ff_jdep->wk_type == D_JFREEFRAG)
 5636                 add_to_journal(freefrag->ff_jdep);
 5637         return (aip);
 5638 }
 5639 
 5640 /*
 5641  * Called just before setting an indirect block pointer
 5642  * to a newly allocated file page.
 5643  */
 5644 void
 5645 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
 5646         struct inode *ip;       /* inode for file being extended */
 5647         ufs_lbn_t lbn;          /* allocated block number within file */
 5648         struct buf *bp;         /* buffer with indirect blk referencing page */
 5649         int ptrno;              /* offset of pointer in indirect block */
 5650         ufs2_daddr_t newblkno;  /* disk block number being added */
 5651         ufs2_daddr_t oldblkno;  /* previous block number, 0 if none */
 5652         struct buf *nbp;        /* buffer holding allocated page */
 5653 {
 5654         struct inodedep *inodedep;
 5655         struct freefrag *freefrag;
 5656         struct allocindir *aip;
 5657         struct pagedep *pagedep;
 5658         struct mount *mp;
 5659         int dflags;
 5660 
 5661         if (lbn != nbp->b_lblkno)
 5662                 panic("softdep_setup_allocindir_page: lbn %jd != lblkno %jd",
 5663                     lbn, bp->b_lblkno);
 5664         CTR4(KTR_SUJ,
 5665             "softdep_setup_allocindir_page: ino %d blkno %jd oldblkno %jd "
 5666             "lbn %jd", ip->i_number, newblkno, oldblkno, lbn);
 5667         ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page");
 5668         mp = UFSTOVFS(ip->i_ump);
 5669         aip = newallocindir(ip, ptrno, newblkno, oldblkno, lbn);
 5670         dflags = DEPALLOC;
 5671         if (IS_SNAPSHOT(ip))
 5672                 dflags |= NODELAY;
 5673         (void) inodedep_lookup(mp, ip->i_number, dflags, &inodedep);
 5674         /*
 5675          * If we are allocating a directory page, then we must
 5676          * allocate an associated pagedep to track additions and
 5677          * deletions.
 5678          */
 5679         if ((ip->i_mode & IFMT) == IFDIR)
 5680                 pagedep_lookup(mp, nbp, ip->i_number, lbn, DEPALLOC, &pagedep);
 5681         WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list);
 5682         freefrag = setup_allocindir_phase2(bp, ip, inodedep, aip, lbn);
 5683         FREE_LOCK(&lk);
 5684         if (freefrag)
 5685                 handle_workitem_freefrag(freefrag);
 5686 }
 5687 
 5688 /*
 5689  * Called just before setting an indirect block pointer to a
 5690  * newly allocated indirect block.
 5691  */
 5692 void
 5693 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
 5694         struct buf *nbp;        /* newly allocated indirect block */
 5695         struct inode *ip;       /* inode for file being extended */
 5696         struct buf *bp;         /* indirect block referencing allocated block */
 5697         int ptrno;              /* offset of pointer in indirect block */
 5698         ufs2_daddr_t newblkno;  /* disk block number being added */
 5699 {
 5700         struct inodedep *inodedep;
 5701         struct allocindir *aip;
 5702         ufs_lbn_t lbn;
 5703         int dflags;
 5704 
 5705         CTR3(KTR_SUJ,
 5706             "softdep_setup_allocindir_meta: ino %d blkno %jd ptrno %d",
 5707             ip->i_number, newblkno, ptrno);
 5708         lbn = nbp->b_lblkno;
 5709         ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta");
 5710         aip = newallocindir(ip, ptrno, newblkno, 0, lbn);
 5711         dflags = DEPALLOC;
 5712         if (IS_SNAPSHOT(ip))
 5713                 dflags |= NODELAY;
 5714         inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, dflags, &inodedep);
 5715         WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list);
 5716         if (setup_allocindir_phase2(bp, ip, inodedep, aip, lbn))
 5717                 panic("softdep_setup_allocindir_meta: Block already existed");
 5718         FREE_LOCK(&lk);
 5719 }
 5720 
 5721 static void
 5722 indirdep_complete(indirdep)
 5723         struct indirdep *indirdep;
 5724 {
 5725         struct allocindir *aip;
 5726 
 5727         LIST_REMOVE(indirdep, ir_next);
 5728         indirdep->ir_state |= DEPCOMPLETE;
 5729 
 5730         while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) {
 5731                 LIST_REMOVE(aip, ai_next);
 5732                 free_newblk(&aip->ai_block);
 5733         }
 5734         /*
 5735          * If this indirdep is not attached to a buf it was simply waiting
 5736          * on completion to clear completehd.  free_indirdep() asserts
 5737          * that nothing is dangling.
 5738          */
 5739         if ((indirdep->ir_state & ONWORKLIST) == 0)
 5740                 free_indirdep(indirdep);
 5741 }
 5742 
 5743 static struct indirdep *
 5744 indirdep_lookup(mp, ip, bp)
 5745         struct mount *mp;
 5746         struct inode *ip;
 5747         struct buf *bp;
 5748 {
 5749         struct indirdep *indirdep, *newindirdep;
 5750         struct newblk *newblk;
 5751         struct worklist *wk;
 5752         struct fs *fs;
 5753         ufs2_daddr_t blkno;
 5754 
 5755         rw_assert(&lk, RA_WLOCKED);
 5756         indirdep = NULL;
 5757         newindirdep = NULL;
 5758         fs = ip->i_fs;
 5759         for (;;) {
 5760                 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
 5761                         if (wk->wk_type != D_INDIRDEP)
 5762                                 continue;
 5763                         indirdep = WK_INDIRDEP(wk);
 5764                         break;
 5765                 }
 5766                 /* Found on the buffer worklist, no new structure to free. */
 5767                 if (indirdep != NULL && newindirdep == NULL)
 5768                         return (indirdep);
 5769                 if (indirdep != NULL && newindirdep != NULL)
 5770                         panic("indirdep_lookup: simultaneous create");
 5771                 /* None found on the buffer and a new structure is ready. */
 5772                 if (indirdep == NULL && newindirdep != NULL)
 5773                         break;
 5774                 /* None found and no new structure available. */
 5775                 FREE_LOCK(&lk);
 5776                 newindirdep = malloc(sizeof(struct indirdep),
 5777                     M_INDIRDEP, M_SOFTDEP_FLAGS);
 5778                 workitem_alloc(&newindirdep->ir_list, D_INDIRDEP, mp);
 5779                 newindirdep->ir_state = ATTACHED;
 5780                 if (ip->i_ump->um_fstype == UFS1)
 5781                         newindirdep->ir_state |= UFS1FMT;
 5782                 TAILQ_INIT(&newindirdep->ir_trunc);
 5783                 newindirdep->ir_saveddata = NULL;
 5784                 LIST_INIT(&newindirdep->ir_deplisthd);
 5785                 LIST_INIT(&newindirdep->ir_donehd);
 5786                 LIST_INIT(&newindirdep->ir_writehd);
 5787                 LIST_INIT(&newindirdep->ir_completehd);
 5788                 if (bp->b_blkno == bp->b_lblkno) {
 5789                         ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp,
 5790                             NULL, NULL);
 5791                         bp->b_blkno = blkno;
 5792                 }
 5793                 newindirdep->ir_freeblks = NULL;
 5794                 newindirdep->ir_savebp =
 5795                     getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0);
 5796                 newindirdep->ir_bp = bp;
 5797                 BUF_KERNPROC(newindirdep->ir_savebp);
 5798                 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
 5799                 ACQUIRE_LOCK(&lk);
 5800         }
 5801         indirdep = newindirdep;
 5802         WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
 5803         /*
 5804          * If the block is not yet allocated we don't set DEPCOMPLETE so
 5805          * that we don't free dependencies until the pointers are valid.
 5806          * This could search b_dep for D_ALLOCDIRECT/D_ALLOCINDIR rather
 5807          * than using the hash.
 5808          */
 5809         if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk))
 5810                 LIST_INSERT_HEAD(&newblk->nb_indirdeps, indirdep, ir_next);
 5811         else
 5812                 indirdep->ir_state |= DEPCOMPLETE;
 5813         return (indirdep);
 5814 }
 5815 
 5816 /*
 5817  * Called to finish the allocation of the "aip" allocated
 5818  * by one of the two routines above.
 5819  */
 5820 static struct freefrag *
 5821 setup_allocindir_phase2(bp, ip, inodedep, aip, lbn)
 5822         struct buf *bp;         /* in-memory copy of the indirect block */
 5823         struct inode *ip;       /* inode for file being extended */
 5824         struct inodedep *inodedep; /* Inodedep for ip */
 5825         struct allocindir *aip; /* allocindir allocated by the above routines */
 5826         ufs_lbn_t lbn;          /* Logical block number for this block. */
 5827 {
 5828         struct fs *fs;
 5829         struct indirdep *indirdep;
 5830         struct allocindir *oldaip;
 5831         struct freefrag *freefrag;
 5832         struct mount *mp;
 5833 
 5834         rw_assert(&lk, RA_WLOCKED);
 5835         mp = UFSTOVFS(ip->i_ump);
 5836         fs = ip->i_fs;
 5837         if (bp->b_lblkno >= 0)
 5838                 panic("setup_allocindir_phase2: not indir blk");
 5839         KASSERT(aip->ai_offset >= 0 && aip->ai_offset < NINDIR(fs),
 5840             ("setup_allocindir_phase2: Bad offset %d", aip->ai_offset));
 5841         indirdep = indirdep_lookup(mp, ip, bp);
 5842         KASSERT(indirdep->ir_savebp != NULL,
 5843             ("setup_allocindir_phase2 NULL ir_savebp"));
 5844         aip->ai_indirdep = indirdep;
 5845         /*
 5846          * Check for an unwritten dependency for this indirect offset.  If
 5847          * there is, merge the old dependency into the new one.  This happens
 5848          * as a result of reallocblk only.
 5849          */
 5850         freefrag = NULL;
 5851         if (aip->ai_oldblkno != 0) {
 5852                 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) {
 5853                         if (oldaip->ai_offset == aip->ai_offset) {
 5854                                 freefrag = allocindir_merge(aip, oldaip);
 5855                                 goto done;
 5856                         }
 5857                 }
 5858                 LIST_FOREACH(oldaip, &indirdep->ir_donehd, ai_next) {
 5859                         if (oldaip->ai_offset == aip->ai_offset) {
 5860                                 freefrag = allocindir_merge(aip, oldaip);
 5861                                 goto done;
 5862                         }
 5863                 }
 5864         }
 5865 done:
 5866         LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
 5867         return (freefrag);
 5868 }
 5869 
 5870 /*
 5871  * Merge two allocindirs which refer to the same block.  Move newblock
 5872  * dependencies and setup the freefrags appropriately.
 5873  */
 5874 static struct freefrag *
 5875 allocindir_merge(aip, oldaip)
 5876         struct allocindir *aip;
 5877         struct allocindir *oldaip;
 5878 {
 5879         struct freefrag *freefrag;
 5880         struct worklist *wk;
 5881 
 5882         if (oldaip->ai_newblkno != aip->ai_oldblkno)
 5883                 panic("allocindir_merge: blkno");
 5884         aip->ai_oldblkno = oldaip->ai_oldblkno;
 5885         freefrag = aip->ai_freefrag;
 5886         aip->ai_freefrag = oldaip->ai_freefrag;
 5887         oldaip->ai_freefrag = NULL;
 5888         KASSERT(freefrag != NULL, ("setup_allocindir_phase2: No freefrag"));
 5889         /*
 5890          * If we are tracking a new directory-block allocation,
 5891          * move it from the old allocindir to the new allocindir.
 5892          */
 5893         if ((wk = LIST_FIRST(&oldaip->ai_newdirblk)) != NULL) {
 5894                 WORKLIST_REMOVE(wk);
 5895                 if (!LIST_EMPTY(&oldaip->ai_newdirblk))
 5896                         panic("allocindir_merge: extra newdirblk");
 5897                 WORKLIST_INSERT(&aip->ai_newdirblk, wk);
 5898         }
 5899         /*
 5900          * We can skip journaling for this freefrag and just complete
 5901          * any pending journal work for the allocindir that is being
 5902          * removed after the freefrag completes.
 5903          */
 5904         if (freefrag->ff_jdep)
 5905                 cancel_jfreefrag(WK_JFREEFRAG(freefrag->ff_jdep));
 5906         LIST_REMOVE(oldaip, ai_next);
 5907         freefrag->ff_jdep = (struct worklist *)cancel_newblk(&oldaip->ai_block,
 5908             &freefrag->ff_list, &freefrag->ff_jwork);
 5909         free_newblk(&oldaip->ai_block);
 5910 
 5911         return (freefrag);
 5912 }
 5913 
 5914 static inline void
 5915 setup_freedirect(freeblks, ip, i, needj)
 5916         struct freeblks *freeblks;
 5917         struct inode *ip;
 5918         int i;
 5919         int needj;
 5920 {
 5921         ufs2_daddr_t blkno;
 5922         int frags;
 5923 
 5924         blkno = DIP(ip, i_db[i]);
 5925         if (blkno == 0)
 5926                 return;
 5927         DIP_SET(ip, i_db[i], 0);
 5928         frags = sblksize(ip->i_fs, ip->i_size, i);
 5929         frags = numfrags(ip->i_fs, frags);
 5930         newfreework(ip->i_ump, freeblks, NULL, i, blkno, frags, 0, needj);
 5931 }
 5932 
 5933 static inline void
 5934 setup_freeext(freeblks, ip, i, needj)
 5935         struct freeblks *freeblks;
 5936         struct inode *ip;
 5937         int i;
 5938         int needj;
 5939 {
 5940         ufs2_daddr_t blkno;
 5941         int frags;
 5942 
 5943         blkno = ip->i_din2->di_extb[i];
 5944         if (blkno == 0)
 5945                 return;
 5946         ip->i_din2->di_extb[i] = 0;
 5947         frags = sblksize(ip->i_fs, ip->i_din2->di_extsize, i);
 5948         frags = numfrags(ip->i_fs, frags);
 5949         newfreework(ip->i_ump, freeblks, NULL, -1 - i, blkno, frags, 0, needj);
 5950 }
 5951 
 5952 static inline void
 5953 setup_freeindir(freeblks, ip, i, lbn, needj)
 5954         struct freeblks *freeblks;
 5955         struct inode *ip;
 5956         int i;
 5957         ufs_lbn_t lbn;
 5958         int needj;
 5959 {
 5960         ufs2_daddr_t blkno;
 5961 
 5962         blkno = DIP(ip, i_ib[i]);
 5963         if (blkno == 0)
 5964                 return;
 5965         DIP_SET(ip, i_ib[i], 0);
 5966         newfreework(ip->i_ump, freeblks, NULL, lbn, blkno, ip->i_fs->fs_frag,
 5967             0, needj);
 5968 }
 5969 
 5970 static inline struct freeblks *
 5971 newfreeblks(mp, ip)
 5972         struct mount *mp;
 5973         struct inode *ip;
 5974 {
 5975         struct freeblks *freeblks;
 5976 
 5977         freeblks = malloc(sizeof(struct freeblks),
 5978                 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO);
 5979         workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp);
 5980         LIST_INIT(&freeblks->fb_jblkdephd);
 5981         LIST_INIT(&freeblks->fb_jwork);
 5982         freeblks->fb_ref = 0;
 5983         freeblks->fb_cgwait = 0;
 5984         freeblks->fb_state = ATTACHED;
 5985         freeblks->fb_uid = ip->i_uid;
 5986         freeblks->fb_inum = ip->i_number;
 5987         freeblks->fb_vtype = ITOV(ip)->v_type;
 5988         freeblks->fb_modrev = DIP(ip, i_modrev);
 5989         freeblks->fb_devvp = ip->i_devvp;
 5990         freeblks->fb_chkcnt = 0;
 5991         freeblks->fb_len = 0;
 5992 
 5993         return (freeblks);
 5994 }
 5995 
 5996 static void
 5997 trunc_indirdep(indirdep, freeblks, bp, off)
 5998         struct indirdep *indirdep;
 5999         struct freeblks *freeblks;
 6000         struct buf *bp;
 6001         int off;
 6002 {
 6003         struct allocindir *aip, *aipn;
 6004 
 6005         /*
 6006          * The first set of allocindirs won't be in savedbp.
 6007          */
 6008         LIST_FOREACH_SAFE(aip, &indirdep->ir_deplisthd, ai_next, aipn)
 6009                 if (aip->ai_offset > off)
 6010                         cancel_allocindir(aip, bp, freeblks, 1);
 6011         LIST_FOREACH_SAFE(aip, &indirdep->ir_donehd, ai_next, aipn)
 6012                 if (aip->ai_offset > off)
 6013                         cancel_allocindir(aip, bp, freeblks, 1);
 6014         /*
 6015          * These will exist in savedbp.
 6016          */
 6017         LIST_FOREACH_SAFE(aip, &indirdep->ir_writehd, ai_next, aipn)
 6018                 if (aip->ai_offset > off)
 6019                         cancel_allocindir(aip, NULL, freeblks, 0);
 6020         LIST_FOREACH_SAFE(aip, &indirdep->ir_completehd, ai_next, aipn)
 6021                 if (aip->ai_offset > off)
 6022                         cancel_allocindir(aip, NULL, freeblks, 0);
 6023 }
 6024 
 6025 /*
 6026  * Follow the chain of indirects down to lastlbn creating a freework
 6027  * structure for each.  This will be used to start indir_trunc() at
 6028  * the right offset and create the journal records for the parrtial
 6029  * truncation.  A second step will handle the truncated dependencies.
 6030  */
 6031 static int
 6032 setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno)
 6033         struct freeblks *freeblks;
 6034         struct inode *ip;
 6035         ufs_lbn_t lbn;
 6036         ufs_lbn_t lastlbn;
 6037         ufs2_daddr_t blkno;
 6038 {
 6039         struct indirdep *indirdep;
 6040         struct indirdep *indirn;
 6041         struct freework *freework;
 6042         struct newblk *newblk;
 6043         struct mount *mp;
 6044         struct buf *bp;
 6045         uint8_t *start;
 6046         uint8_t *end;
 6047         ufs_lbn_t lbnadd;
 6048         int level;
 6049         int error;
 6050         int off;
 6051 
 6052 
 6053         freework = NULL;
 6054         if (blkno == 0)
 6055                 return (0);
 6056         mp = freeblks->fb_list.wk_mp;
 6057         bp = getblk(ITOV(ip), lbn, mp->mnt_stat.f_iosize, 0, 0, 0);
 6058         if ((bp->b_flags & B_CACHE) == 0) {
 6059                 bp->b_blkno = blkptrtodb(VFSTOUFS(mp), blkno);
 6060                 bp->b_iocmd = BIO_READ;
 6061                 bp->b_flags &= ~B_INVAL;
 6062                 bp->b_ioflags &= ~BIO_ERROR;
 6063                 vfs_busy_pages(bp, 0);
 6064                 bp->b_iooffset = dbtob(bp->b_blkno);
 6065                 bstrategy(bp);
 6066                 curthread->td_ru.ru_inblock++;
 6067                 error = bufwait(bp);
 6068                 if (error) {
 6069                         brelse(bp);
 6070                         return (error);
 6071                 }
 6072         }
 6073         level = lbn_level(lbn);
 6074         lbnadd = lbn_offset(ip->i_fs, level);
 6075         /*
 6076          * Compute the offset of the last block we want to keep.  Store
 6077          * in the freework the first block we want to completely free.
 6078          */
 6079         off = (lastlbn - -(lbn + level)) / lbnadd;
 6080         if (off + 1 == NINDIR(ip->i_fs))
 6081                 goto nowork;
 6082         freework = newfreework(ip->i_ump, freeblks, NULL, lbn, blkno, 0, off+1,
 6083             0);
 6084         /*
 6085          * Link the freework into the indirdep.  This will prevent any new
 6086          * allocations from proceeding until we are finished with the
 6087          * truncate and the block is written.
 6088          */
 6089         ACQUIRE_LOCK(&lk);
 6090         indirdep = indirdep_lookup(mp, ip, bp);
 6091         if (indirdep->ir_freeblks)
 6092                 panic("setup_trunc_indir: indirdep already truncated.");
 6093         TAILQ_INSERT_TAIL(&indirdep->ir_trunc, freework, fw_next);
 6094         freework->fw_indir = indirdep;
 6095         /*
 6096          * Cancel any allocindirs that will not make it to disk.
 6097          * We have to do this for all copies of the indirdep that
 6098          * live on this newblk.
 6099          */
 6100         if ((indirdep->ir_state & DEPCOMPLETE) == 0) {
 6101                 newblk_lookup(mp, dbtofsb(ip->i_fs, bp->b_blkno), 0, &newblk);
 6102                 LIST_FOREACH(indirn, &newblk->nb_indirdeps, ir_next)
 6103                         trunc_indirdep(indirn, freeblks, bp, off);
 6104         } else
 6105                 trunc_indirdep(indirdep, freeblks, bp, off);
 6106         FREE_LOCK(&lk);
 6107         /*
 6108          * Creation is protected by the buf lock. The saveddata is only
 6109          * needed if a full truncation follows a partial truncation but it
 6110          * is difficult to allocate in that case so we fetch it anyway.
 6111          */
 6112         if (indirdep->ir_saveddata == NULL)
 6113                 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP,
 6114                     M_SOFTDEP_FLAGS);
 6115 nowork:
 6116         /* Fetch the blkno of the child and the zero start offset. */
 6117         if (ip->i_ump->um_fstype == UFS1) {
 6118                 blkno = ((ufs1_daddr_t *)bp->b_data)[off];
 6119                 start = (uint8_t *)&((ufs1_daddr_t *)bp->b_data)[off+1];
 6120         } else {
 6121                 blkno = ((ufs2_daddr_t *)bp->b_data)[off];
 6122                 start = (uint8_t *)&((ufs2_daddr_t *)bp->b_data)[off+1];
 6123         }
 6124         if (freework) {
 6125                 /* Zero the truncated pointers. */
 6126                 end = bp->b_data + bp->b_bcount;
 6127                 bzero(start, end - start);
 6128                 bdwrite(bp);
 6129         } else
 6130                 bqrelse(bp);
 6131         if (level == 0)
 6132                 return (0);
 6133         lbn++; /* adjust level */
 6134         lbn -= (off * lbnadd);
 6135         return setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno);
 6136 }
 6137 
 6138 /*
 6139  * Complete the partial truncation of an indirect block setup by
 6140  * setup_trunc_indir().  This zeros the truncated pointers in the saved
 6141  * copy and writes them to disk before the freeblks is allowed to complete.
 6142  */
 6143 static void
 6144 complete_trunc_indir(freework)
 6145         struct freework *freework;
 6146 {
 6147         struct freework *fwn;
 6148         struct indirdep *indirdep;
 6149         struct buf *bp;
 6150         uintptr_t start;
 6151         int count;
 6152 
 6153         indirdep = freework->fw_indir;
 6154         for (;;) {
 6155                 bp = indirdep->ir_bp;
 6156                 /* See if the block was discarded. */
 6157                 if (bp == NULL)
 6158                         break;
 6159                 /* Inline part of getdirtybuf().  We dont want bremfree. */
 6160                 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0)
 6161                         break;
 6162                 if (BUF_LOCK(bp,
 6163                     LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, &lk) == 0)
 6164                         BUF_UNLOCK(bp);
 6165                 ACQUIRE_LOCK(&lk);
 6166         }
 6167         rw_assert(&lk, RA_WLOCKED);
 6168         freework->fw_state |= DEPCOMPLETE;
 6169         TAILQ_REMOVE(&indirdep->ir_trunc, freework, fw_next);
 6170         /*
 6171          * Zero the pointers in the saved copy.
 6172          */
 6173         if (indirdep->ir_state & UFS1FMT)
 6174                 start = sizeof(ufs1_daddr_t);
 6175         else
 6176                 start = sizeof(ufs2_daddr_t);
 6177         start *= freework->fw_start;
 6178         count = indirdep->ir_savebp->b_bcount - start;
 6179         start += (uintptr_t)indirdep->ir_savebp->b_data;
 6180         bzero((char *)start, count);
 6181         /*
 6182          * We need to start the next truncation in the list if it has not
 6183          * been started yet.
 6184          */
 6185         fwn = TAILQ_FIRST(&indirdep->ir_trunc);
 6186         if (fwn != NULL) {
 6187                 if (fwn->fw_freeblks == indirdep->ir_freeblks)
 6188                         TAILQ_REMOVE(&indirdep->ir_trunc, fwn, fw_next);
 6189                 if ((fwn->fw_state & ONWORKLIST) == 0)
 6190                         freework_enqueue(fwn);
 6191         }
 6192         /*
 6193          * If bp is NULL the block was fully truncated, restore
 6194          * the saved block list otherwise free it if it is no
 6195          * longer needed.
 6196          */
 6197         if (TAILQ_EMPTY(&indirdep->ir_trunc)) {
 6198                 if (bp == NULL)
 6199                         bcopy(indirdep->ir_saveddata,
 6200                             indirdep->ir_savebp->b_data,
 6201                             indirdep->ir_savebp->b_bcount);
 6202                 free(indirdep->ir_saveddata, M_INDIRDEP);
 6203                 indirdep->ir_saveddata = NULL;
 6204         }
 6205         /*
 6206          * When bp is NULL there is a full truncation pending.  We
 6207          * must wait for this full truncation to be journaled before
 6208          * we can release this freework because the disk pointers will
 6209          * never be written as zero.
 6210          */
 6211         if (bp == NULL)  {
 6212                 if (LIST_EMPTY(&indirdep->ir_freeblks->fb_jblkdephd))
 6213                         handle_written_freework(freework);
 6214                 else
 6215                         WORKLIST_INSERT(&indirdep->ir_freeblks->fb_freeworkhd,
 6216                            &freework->fw_list);
 6217         } else {
 6218                 /* Complete when the real copy is written. */
 6219                 WORKLIST_INSERT(&bp->b_dep, &freework->fw_list);
 6220                 BUF_UNLOCK(bp);
 6221         }
 6222 }
 6223 
 6224 /*
 6225  * Calculate the number of blocks we are going to release where datablocks
 6226  * is the current total and length is the new file size.
 6227  */
 6228 ufs2_daddr_t
 6229 blkcount(fs, datablocks, length)
 6230         struct fs *fs;
 6231         ufs2_daddr_t datablocks;
 6232         off_t length;
 6233 {
 6234         off_t totblks, numblks;
 6235 
 6236         totblks = 0;
 6237         numblks = howmany(length, fs->fs_bsize);
 6238         if (numblks <= NDADDR) {
 6239                 totblks = howmany(length, fs->fs_fsize);
 6240                 goto out;
 6241         }
 6242         totblks = blkstofrags(fs, numblks);
 6243         numblks -= NDADDR;
 6244         /*
 6245          * Count all single, then double, then triple indirects required.
 6246          * Subtracting one indirects worth of blocks for each pass
 6247          * acknowledges one of each pointed to by the inode.
 6248          */
 6249         for (;;) {
 6250                 totblks += blkstofrags(fs, howmany(numblks, NINDIR(fs)));
 6251                 numblks -= NINDIR(fs);
 6252                 if (numblks <= 0)
 6253                         break;
 6254                 numblks = howmany(numblks, NINDIR(fs));
 6255         }
 6256 out:
 6257         totblks = fsbtodb(fs, totblks);
 6258         /*
 6259          * Handle sparse files.  We can't reclaim more blocks than the inode
 6260          * references.  We will correct it later in handle_complete_freeblks()
 6261          * when we know the real count.
 6262          */
 6263         if (totblks > datablocks)
 6264                 return (0);
 6265         return (datablocks - totblks);
 6266 }
 6267 
 6268 /*
 6269  * Handle freeblocks for journaled softupdate filesystems.
 6270  *
 6271  * Contrary to normal softupdates, we must preserve the block pointers in
 6272  * indirects until their subordinates are free.  This is to avoid journaling
 6273  * every block that is freed which may consume more space than the journal
 6274  * itself.  The recovery program will see the free block journals at the
 6275  * base of the truncated area and traverse them to reclaim space.  The
 6276  * pointers in the inode may be cleared immediately after the journal
 6277  * records are written because each direct and indirect pointer in the
 6278  * inode is recorded in a journal.  This permits full truncation to proceed
 6279  * asynchronously.  The write order is journal -> inode -> cgs -> indirects.
 6280  *
 6281  * The algorithm is as follows:
 6282  * 1) Traverse the in-memory state and create journal entries to release
 6283  *    the relevant blocks and full indirect trees.
 6284  * 2) Traverse the indirect block chain adding partial truncation freework
 6285  *    records to indirects in the path to lastlbn.  The freework will
 6286  *    prevent new allocation dependencies from being satisfied in this
 6287  *    indirect until the truncation completes.
 6288  * 3) Read and lock the inode block, performing an update with the new size
 6289  *    and pointers.  This prevents truncated data from becoming valid on
 6290  *    disk through step 4.
 6291  * 4) Reap unsatisfied dependencies that are beyond the truncated area,
 6292  *    eliminate journal work for those records that do not require it.
 6293  * 5) Schedule the journal records to be written followed by the inode block.
 6294  * 6) Allocate any necessary frags for the end of file.
 6295  * 7) Zero any partially truncated blocks.
 6296  *
 6297  * From this truncation proceeds asynchronously using the freework and
 6298  * indir_trunc machinery.  The file will not be extended again into a
 6299  * partially truncated indirect block until all work is completed but
 6300  * the normal dependency mechanism ensures that it is rolled back/forward
 6301  * as appropriate.  Further truncation may occur without delay and is
 6302  * serialized in indir_trunc().
 6303  */
 6304 void
 6305 softdep_journal_freeblocks(ip, cred, length, flags)
 6306         struct inode *ip;       /* The inode whose length is to be reduced */
 6307         struct ucred *cred;
 6308         off_t length;           /* The new length for the file */
 6309         int flags;              /* IO_EXT and/or IO_NORMAL */
 6310 {
 6311         struct freeblks *freeblks, *fbn;
 6312         struct worklist *wk, *wkn;
 6313         struct inodedep *inodedep;
 6314         struct jblkdep *jblkdep;
 6315         struct allocdirect *adp, *adpn;
 6316         struct fs *fs;
 6317         struct buf *bp;
 6318         struct vnode *vp;
 6319         struct mount *mp;
 6320         ufs2_daddr_t extblocks, datablocks;
 6321         ufs_lbn_t tmpval, lbn, lastlbn;
 6322         int frags, lastoff, iboff, allocblock, needj, dflags, error, i;
 6323 
 6324         fs = ip->i_fs;
 6325         mp = UFSTOVFS(ip->i_ump);
 6326         vp = ITOV(ip);
 6327         needj = 1;
 6328         iboff = -1;
 6329         allocblock = 0;
 6330         extblocks = 0;
 6331         datablocks = 0;
 6332         frags = 0;
 6333         freeblks = newfreeblks(mp, ip);
 6334         ACQUIRE_LOCK(&lk);
 6335         /*
 6336          * If we're truncating a removed file that will never be written
 6337          * we don't need to journal the block frees.  The canceled journals
 6338          * for the allocations will suffice.
 6339          */
 6340         dflags = DEPALLOC;
 6341         if (IS_SNAPSHOT(ip))
 6342                 dflags |= NODELAY;
 6343         inodedep_lookup(mp, ip->i_number, dflags, &inodedep);
 6344         if ((inodedep->id_state & (UNLINKED | DEPCOMPLETE)) == UNLINKED &&
 6345             length == 0)
 6346                 needj = 0;
 6347         CTR3(KTR_SUJ, "softdep_journal_freeblks: ip %d length %ld needj %d",
 6348             ip->i_number, length, needj);
 6349         FREE_LOCK(&lk);
 6350         /*
 6351          * Calculate the lbn that we are truncating to.  This results in -1
 6352          * if we're truncating the 0 bytes.  So it is the last lbn we want
 6353          * to keep, not the first lbn we want to truncate.
 6354          */
 6355         lastlbn = lblkno(fs, length + fs->fs_bsize - 1) - 1;
 6356         lastoff = blkoff(fs, length);
 6357         /*
 6358          * Compute frags we are keeping in lastlbn.  0 means all.
 6359          */
 6360         if (lastlbn >= 0 && lastlbn < NDADDR) {
 6361                 frags = fragroundup(fs, lastoff);
 6362                 /* adp offset of last valid allocdirect. */
 6363                 iboff = lastlbn;
 6364         } else if (lastlbn > 0)
 6365                 iboff = NDADDR;
 6366         if (fs->fs_magic == FS_UFS2_MAGIC)
 6367                 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
 6368         /*
 6369          * Handle normal data blocks and indirects.  This section saves
 6370          * values used after the inode update to complete frag and indirect
 6371          * truncation.
 6372          */
 6373         if ((flags & IO_NORMAL) != 0) {
 6374                 /*
 6375                  * Handle truncation of whole direct and indirect blocks.
 6376                  */
 6377                 for (i = iboff + 1; i < NDADDR; i++)
 6378                         setup_freedirect(freeblks, ip, i, needj);
 6379                 for (i = 0, tmpval = NINDIR(fs), lbn = NDADDR; i < NIADDR;
 6380                     i++, lbn += tmpval, tmpval *= NINDIR(fs)) {
 6381                         /* Release a whole indirect tree. */
 6382                         if (lbn > lastlbn) {
 6383                                 setup_freeindir(freeblks, ip, i, -lbn -i,
 6384                                     needj);
 6385                                 continue;
 6386                         }
 6387                         iboff = i + NDADDR;
 6388                         /*
 6389                          * Traverse partially truncated indirect tree.
 6390                          */
 6391                         if (lbn <= lastlbn && lbn + tmpval - 1 > lastlbn)
 6392                                 setup_trunc_indir(freeblks, ip, -lbn - i,
 6393                                     lastlbn, DIP(ip, i_ib[i]));
 6394                 }
 6395                 /*
 6396                  * Handle partial truncation to a frag boundary.
 6397                  */
 6398                 if (frags) {
 6399                         ufs2_daddr_t blkno;
 6400                         long oldfrags;
 6401 
 6402                         oldfrags = blksize(fs, ip, lastlbn);
 6403                         blkno = DIP(ip, i_db[lastlbn]);
 6404                         if (blkno && oldfrags != frags) {
 6405                                 oldfrags -= frags;
 6406                                 oldfrags = numfrags(ip->i_fs, oldfrags);
 6407                                 blkno += numfrags(ip->i_fs, frags);
 6408                                 newfreework(ip->i_ump, freeblks, NULL, lastlbn,
 6409                                     blkno, oldfrags, 0, needj);
 6410                         } else if (blkno == 0)
 6411                                 allocblock = 1;
 6412                 }
 6413                 /*
 6414                  * Add a journal record for partial truncate if we are
 6415                  * handling indirect blocks.  Non-indirects need no extra
 6416                  * journaling.
 6417                  */
 6418                 if (length != 0 && lastlbn >= NDADDR) {
 6419                         ip->i_flag |= IN_TRUNCATED;
 6420                         newjtrunc(freeblks, length, 0);
 6421                 }
 6422                 ip->i_size = length;
 6423                 DIP_SET(ip, i_size, ip->i_size);
 6424                 datablocks = DIP(ip, i_blocks) - extblocks;
 6425                 if (length != 0)
 6426                         datablocks = blkcount(ip->i_fs, datablocks, length);
 6427                 freeblks->fb_len = length;
 6428         }
 6429         if ((flags & IO_EXT) != 0) {
 6430                 for (i = 0; i < NXADDR; i++)
 6431                         setup_freeext(freeblks, ip, i, needj);
 6432                 ip->i_din2->di_extsize = 0;
 6433                 datablocks += extblocks;
 6434         }
 6435 #ifdef QUOTA
 6436         /* Reference the quotas in case the block count is wrong in the end. */
 6437         quotaref(vp, freeblks->fb_quota);
 6438         (void) chkdq(ip, -datablocks, NOCRED, 0);
 6439 #endif
 6440         freeblks->fb_chkcnt = -datablocks;
 6441         UFS_LOCK(ip->i_ump);
 6442         fs->fs_pendingblocks += datablocks;
 6443         UFS_UNLOCK(ip->i_ump);
 6444         DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks);
 6445         /*
 6446          * Handle truncation of incomplete alloc direct dependencies.  We
 6447          * hold the inode block locked to prevent incomplete dependencies
 6448          * from reaching the disk while we are eliminating those that
 6449          * have been truncated.  This is a partially inlined ffs_update().
 6450          */
 6451         ufs_itimes(vp);
 6452         ip->i_flag &= ~(IN_LAZYACCESS | IN_LAZYMOD | IN_MODIFIED);
 6453         error = bread(ip->i_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
 6454             (int)fs->fs_bsize, cred, &bp);
 6455         if (error) {
 6456                 brelse(bp);
 6457                 softdep_error("softdep_journal_freeblocks", error);
 6458                 return;
 6459         }
 6460         if (bp->b_bufsize == fs->fs_bsize)
 6461                 bp->b_flags |= B_CLUSTEROK;
 6462         softdep_update_inodeblock(ip, bp, 0);
 6463         if (ip->i_ump->um_fstype == UFS1)
 6464                 *((struct ufs1_dinode *)bp->b_data +
 6465                     ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1;
 6466         else
 6467                 *((struct ufs2_dinode *)bp->b_data +
 6468                     ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2;
 6469         ACQUIRE_LOCK(&lk);
 6470         (void) inodedep_lookup(mp, ip->i_number, dflags, &inodedep);
 6471         if ((inodedep->id_state & IOSTARTED) != 0)
 6472                 panic("softdep_setup_freeblocks: inode busy");
 6473         /*
 6474          * Add the freeblks structure to the list of operations that
 6475          * must await the zero'ed inode being written to disk. If we
 6476          * still have a bitmap dependency (needj), then the inode
 6477          * has never been written to disk, so we can process the
 6478          * freeblks below once we have deleted the dependencies.
 6479          */
 6480         if (needj)
 6481                 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list);
 6482         else
 6483                 freeblks->fb_state |= COMPLETE;
 6484         if ((flags & IO_NORMAL) != 0) {
 6485                 TAILQ_FOREACH_SAFE(adp, &inodedep->id_inoupdt, ad_next, adpn) {
 6486                         if (adp->ad_offset > iboff)
 6487                                 cancel_allocdirect(&inodedep->id_inoupdt, adp,
 6488                                     freeblks);
 6489                         /*
 6490                          * Truncate the allocdirect.  We could eliminate
 6491                          * or modify journal records as well.
 6492                          */
 6493                         else if (adp->ad_offset == iboff && frags)
 6494                                 adp->ad_newsize = frags;
 6495                 }
 6496         }
 6497         if ((flags & IO_EXT) != 0)
 6498                 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0)
 6499                         cancel_allocdirect(&inodedep->id_extupdt, adp,
 6500                             freeblks);
 6501         /*
 6502          * Scan the bufwait list for newblock dependencies that will never
 6503          * make it to disk.
 6504          */
 6505         LIST_FOREACH_SAFE(wk, &inodedep->id_bufwait, wk_list, wkn) {
 6506                 if (wk->wk_type != D_ALLOCDIRECT)
 6507                         continue;
 6508                 adp = WK_ALLOCDIRECT(wk);
 6509                 if (((flags & IO_NORMAL) != 0 && (adp->ad_offset > iboff)) ||
 6510                     ((flags & IO_EXT) != 0 && (adp->ad_state & EXTDATA))) {
 6511                         cancel_jfreeblk(freeblks, adp->ad_newblkno);
 6512                         cancel_newblk(WK_NEWBLK(wk), NULL, &freeblks->fb_jwork);
 6513                         WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk);
 6514                 }
 6515         }
 6516         /*
 6517          * Add journal work.
 6518          */
 6519         LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps)
 6520                 add_to_journal(&jblkdep->jb_list);
 6521         FREE_LOCK(&lk);
 6522         bdwrite(bp);
 6523         /*
 6524          * Truncate dependency structures beyond length.
 6525          */
 6526         trunc_dependencies(ip, freeblks, lastlbn, frags, flags);
 6527         /*
 6528          * This is only set when we need to allocate a fragment because
 6529          * none existed at the end of a frag-sized file.  It handles only
 6530          * allocating a new, zero filled block.
 6531          */
 6532         if (allocblock) {
 6533                 ip->i_size = length - lastoff;
 6534                 DIP_SET(ip, i_size, ip->i_size);
 6535                 error = UFS_BALLOC(vp, length - 1, 1, cred, BA_CLRBUF, &bp);
 6536                 if (error != 0) {
 6537                         softdep_error("softdep_journal_freeblks", error);
 6538                         return;
 6539                 }
 6540                 ip->i_size = length;
 6541                 DIP_SET(ip, i_size, length);
 6542                 ip->i_flag |= IN_CHANGE | IN_UPDATE;
 6543                 allocbuf(bp, frags);
 6544                 ffs_update(vp, 0);
 6545                 bawrite(bp);
 6546         } else if (lastoff != 0 && vp->v_type != VDIR) {
 6547                 int size;
 6548 
 6549                 /*
 6550                  * Zero the end of a truncated frag or block.
 6551                  */
 6552                 size = sblksize(fs, length, lastlbn);
 6553                 error = bread(vp, lastlbn, size, cred, &bp);
 6554                 if (error) {
 6555                         softdep_error("softdep_journal_freeblks", error);
 6556                         return;
 6557                 }
 6558                 bzero((char *)bp->b_data + lastoff, size - lastoff);
 6559                 bawrite(bp);
 6560 
 6561         }
 6562         ACQUIRE_LOCK(&lk);
 6563         inodedep_lookup(mp, ip->i_number, dflags, &inodedep);
 6564         TAILQ_INSERT_TAIL(&inodedep->id_freeblklst, freeblks, fb_next);
 6565         freeblks->fb_state |= DEPCOMPLETE | ONDEPLIST;
 6566         /*
 6567          * We zero earlier truncations so they don't erroneously
 6568          * update i_blocks.
 6569          */
 6570         if (freeblks->fb_len == 0 && (flags & IO_NORMAL) != 0)
 6571                 TAILQ_FOREACH(fbn, &inodedep->id_freeblklst, fb_next)
 6572                         fbn->fb_len = 0;
 6573         if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE &&
 6574             LIST_EMPTY(&freeblks->fb_jblkdephd))
 6575                 freeblks->fb_state |= INPROGRESS;
 6576         else
 6577                 freeblks = NULL;
 6578         FREE_LOCK(&lk);
 6579         if (freeblks)
 6580                 handle_workitem_freeblocks(freeblks, 0);
 6581         trunc_pages(ip, length, extblocks, flags);
 6582 
 6583 }
 6584 
 6585 /*
 6586  * Flush a JOP_SYNC to the journal.
 6587  */
 6588 void
 6589 softdep_journal_fsync(ip)
 6590         struct inode *ip;
 6591 {
 6592         struct jfsync *jfsync;
 6593 
 6594         if ((ip->i_flag & IN_TRUNCATED) == 0)
 6595                 return;
 6596         ip->i_flag &= ~IN_TRUNCATED;
 6597         jfsync = malloc(sizeof(*jfsync), M_JFSYNC, M_SOFTDEP_FLAGS | M_ZERO);
 6598         workitem_alloc(&jfsync->jfs_list, D_JFSYNC, UFSTOVFS(ip->i_ump));
 6599         jfsync->jfs_size = ip->i_size;
 6600         jfsync->jfs_ino = ip->i_number;
 6601         ACQUIRE_LOCK(&lk);
 6602         add_to_journal(&jfsync->jfs_list);
 6603         jwait(&jfsync->jfs_list, MNT_WAIT);
 6604         FREE_LOCK(&lk);
 6605 }
 6606 
 6607 /*
 6608  * Block de-allocation dependencies.
 6609  * 
 6610  * When blocks are de-allocated, the on-disk pointers must be nullified before
 6611  * the blocks are made available for use by other files.  (The true
 6612  * requirement is that old pointers must be nullified before new on-disk
 6613  * pointers are set.  We chose this slightly more stringent requirement to
 6614  * reduce complexity.) Our implementation handles this dependency by updating
 6615  * the inode (or indirect block) appropriately but delaying the actual block
 6616  * de-allocation (i.e., freemap and free space count manipulation) until
 6617  * after the updated versions reach stable storage.  After the disk is
 6618  * updated, the blocks can be safely de-allocated whenever it is convenient.
 6619  * This implementation handles only the common case of reducing a file's
 6620  * length to zero. Other cases are handled by the conventional synchronous
 6621  * write approach.
 6622  *
 6623  * The ffs implementation with which we worked double-checks
 6624  * the state of the block pointers and file size as it reduces
 6625  * a file's length.  Some of this code is replicated here in our
 6626  * soft updates implementation.  The freeblks->fb_chkcnt field is
 6627  * used to transfer a part of this information to the procedure
 6628  * that eventually de-allocates the blocks.
 6629  *
 6630  * This routine should be called from the routine that shortens
 6631  * a file's length, before the inode's size or block pointers
 6632  * are modified. It will save the block pointer information for
 6633  * later release and zero the inode so that the calling routine
 6634  * can release it.
 6635  */
 6636 void
 6637 softdep_setup_freeblocks(ip, length, flags)
 6638         struct inode *ip;       /* The inode whose length is to be reduced */
 6639         off_t length;           /* The new length for the file */
 6640         int flags;              /* IO_EXT and/or IO_NORMAL */
 6641 {
 6642         struct ufs1_dinode *dp1;
 6643         struct ufs2_dinode *dp2;
 6644         struct freeblks *freeblks;
 6645         struct inodedep *inodedep;
 6646         struct allocdirect *adp;
 6647         struct buf *bp;
 6648         struct fs *fs;
 6649         ufs2_daddr_t extblocks, datablocks;
 6650         struct mount *mp;
 6651         int i, delay, error, dflags;
 6652         ufs_lbn_t tmpval;
 6653         ufs_lbn_t lbn;
 6654 
 6655         CTR2(KTR_SUJ, "softdep_setup_freeblks: ip %d length %ld",
 6656             ip->i_number, length);
 6657         fs = ip->i_fs;
 6658         mp = UFSTOVFS(ip->i_ump);
 6659         if (length != 0)
 6660                 panic("softdep_setup_freeblocks: non-zero length");
 6661         freeblks = newfreeblks(mp, ip);
 6662         extblocks = 0;
 6663         datablocks = 0;
 6664         if (fs->fs_magic == FS_UFS2_MAGIC)
 6665                 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
 6666         if ((flags & IO_NORMAL) != 0) {
 6667                 for (i = 0; i < NDADDR; i++)
 6668                         setup_freedirect(freeblks, ip, i, 0);
 6669                 for (i = 0, tmpval = NINDIR(fs), lbn = NDADDR; i < NIADDR;
 6670                     i++, lbn += tmpval, tmpval *= NINDIR(fs))
 6671                         setup_freeindir(freeblks, ip, i, -lbn -i, 0);
 6672                 ip->i_size = 0;
 6673                 DIP_SET(ip, i_size, 0);
 6674                 datablocks = DIP(ip, i_blocks) - extblocks;
 6675         }
 6676         if ((flags & IO_EXT) != 0) {
 6677                 for (i = 0; i < NXADDR; i++)
 6678                         setup_freeext(freeblks, ip, i, 0);
 6679                 ip->i_din2->di_extsize = 0;
 6680                 datablocks += extblocks;
 6681         }
 6682 #ifdef QUOTA
 6683         /* Reference the quotas in case the block count is wrong in the end. */
 6684         quotaref(ITOV(ip), freeblks->fb_quota);
 6685         (void) chkdq(ip, -datablocks, NOCRED, 0);
 6686 #endif
 6687         freeblks->fb_chkcnt = -datablocks;
 6688         UFS_LOCK(ip->i_ump);
 6689         fs->fs_pendingblocks += datablocks;
 6690         UFS_UNLOCK(ip->i_ump);
 6691         DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks);
 6692         /*
 6693          * Push the zero'ed inode to to its disk buffer so that we are free
 6694          * to delete its dependencies below. Once the dependencies are gone
 6695          * the buffer can be safely released.
 6696          */
 6697         if ((error = bread(ip->i_devvp,
 6698             fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
 6699             (int)fs->fs_bsize, NOCRED, &bp)) != 0) {
 6700                 brelse(bp);
 6701                 softdep_error("softdep_setup_freeblocks", error);
 6702         }
 6703         if (ip->i_ump->um_fstype == UFS1) {
 6704                 dp1 = ((struct ufs1_dinode *)bp->b_data +
 6705                     ino_to_fsbo(fs, ip->i_number));
 6706                 ip->i_din1->di_freelink = dp1->di_freelink;
 6707                 *dp1 = *ip->i_din1;
 6708         } else {
 6709                 dp2 = ((struct ufs2_dinode *)bp->b_data +
 6710                     ino_to_fsbo(fs, ip->i_number));
 6711                 ip->i_din2->di_freelink = dp2->di_freelink;
 6712                 *dp2 = *ip->i_din2;
 6713         }
 6714         /*
 6715          * Find and eliminate any inode dependencies.
 6716          */
 6717         ACQUIRE_LOCK(&lk);
 6718         dflags = DEPALLOC;
 6719         if (IS_SNAPSHOT(ip))
 6720                 dflags |= NODELAY;
 6721         (void) inodedep_lookup(mp, ip->i_number, dflags, &inodedep);
 6722         if ((inodedep->id_state & IOSTARTED) != 0)
 6723                 panic("softdep_setup_freeblocks: inode busy");
 6724         /*
 6725          * Add the freeblks structure to the list of operations that
 6726          * must await the zero'ed inode being written to disk. If we
 6727          * still have a bitmap dependency (delay == 0), then the inode
 6728          * has never been written to disk, so we can process the
 6729          * freeblks below once we have deleted the dependencies.
 6730          */
 6731         delay = (inodedep->id_state & DEPCOMPLETE);
 6732         if (delay)
 6733                 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list);
 6734         else
 6735                 freeblks->fb_state |= COMPLETE;
 6736         /*
 6737          * Because the file length has been truncated to zero, any
 6738          * pending block allocation dependency structures associated
 6739          * with this inode are obsolete and can simply be de-allocated.
 6740          * We must first merge the two dependency lists to get rid of
 6741          * any duplicate freefrag structures, then purge the merged list.
 6742          * If we still have a bitmap dependency, then the inode has never
 6743          * been written to disk, so we can free any fragments without delay.
 6744          */
 6745         if (flags & IO_NORMAL) {
 6746                 merge_inode_lists(&inodedep->id_newinoupdt,
 6747                     &inodedep->id_inoupdt);
 6748                 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
 6749                         cancel_allocdirect(&inodedep->id_inoupdt, adp,
 6750                             freeblks);
 6751         }
 6752         if (flags & IO_EXT) {
 6753                 merge_inode_lists(&inodedep->id_newextupdt,
 6754                     &inodedep->id_extupdt);
 6755                 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0)
 6756                         cancel_allocdirect(&inodedep->id_extupdt, adp,
 6757                             freeblks);
 6758         }
 6759         FREE_LOCK(&lk);
 6760         bdwrite(bp);
 6761         trunc_dependencies(ip, freeblks, -1, 0, flags);
 6762         ACQUIRE_LOCK(&lk);
 6763         if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0)
 6764                 (void) free_inodedep(inodedep);
 6765         freeblks->fb_state |= DEPCOMPLETE;
 6766         /*
 6767          * If the inode with zeroed block pointers is now on disk
 6768          * we can start freeing blocks.
 6769          */  
 6770         if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
 6771                 freeblks->fb_state |= INPROGRESS;
 6772         else
 6773                 freeblks = NULL;
 6774         FREE_LOCK(&lk);
 6775         if (freeblks)
 6776                 handle_workitem_freeblocks(freeblks, 0);
 6777         trunc_pages(ip, length, extblocks, flags);
 6778 }
 6779 
 6780 /*
 6781  * Eliminate pages from the page cache that back parts of this inode and
 6782  * adjust the vnode pager's idea of our size.  This prevents stale data
 6783  * from hanging around in the page cache.
 6784  */
 6785 static void
 6786 trunc_pages(ip, length, extblocks, flags)
 6787         struct inode *ip;
 6788         off_t length;
 6789         ufs2_daddr_t extblocks;
 6790         int flags;
 6791 {
 6792         struct vnode *vp;
 6793         struct fs *fs;
 6794         ufs_lbn_t lbn;
 6795         off_t end, extend;
 6796 
 6797         vp = ITOV(ip);
 6798         fs = ip->i_fs;
 6799         extend = OFF_TO_IDX(lblktosize(fs, -extblocks));
 6800         if ((flags & IO_EXT) != 0)
 6801                 vn_pages_remove(vp, extend, 0);
 6802         if ((flags & IO_NORMAL) == 0)
 6803                 return;
 6804         BO_LOCK(&vp->v_bufobj);
 6805         drain_output(vp);
 6806         BO_UNLOCK(&vp->v_bufobj);
 6807         /*
 6808          * The vnode pager eliminates file pages we eliminate indirects
 6809          * below.
 6810          */
 6811         vnode_pager_setsize(vp, length);
 6812         /*
 6813          * Calculate the end based on the last indirect we want to keep.  If
 6814          * the block extends into indirects we can just use the negative of
 6815          * its lbn.  Doubles and triples exist at lower numbers so we must
 6816          * be careful not to remove those, if they exist.  double and triple
 6817          * indirect lbns do not overlap with others so it is not important
 6818          * to verify how many levels are required.
 6819          */
 6820         lbn = lblkno(fs, length);
 6821         if (lbn >= NDADDR) {
 6822                 /* Calculate the virtual lbn of the triple indirect. */
 6823                 lbn = -lbn - (NIADDR - 1);
 6824                 end = OFF_TO_IDX(lblktosize(fs, lbn));
 6825         } else
 6826                 end = extend;
 6827         vn_pages_remove(vp, OFF_TO_IDX(OFF_MAX), end);
 6828 }
 6829 
 6830 /*
 6831  * See if the buf bp is in the range eliminated by truncation.
 6832  */
 6833 static int
 6834 trunc_check_buf(bp, blkoffp, lastlbn, lastoff, flags)
 6835         struct buf *bp;
 6836         int *blkoffp;
 6837         ufs_lbn_t lastlbn;
 6838         int lastoff;
 6839         int flags;
 6840 {
 6841         ufs_lbn_t lbn;
 6842 
 6843         *blkoffp = 0;
 6844         /* Only match ext/normal blocks as appropriate. */
 6845         if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) ||
 6846             ((flags & IO_NORMAL) == 0 && (bp->b_xflags & BX_ALTDATA) == 0))
 6847                 return (0);
 6848         /* ALTDATA is always a full truncation. */
 6849         if ((bp->b_xflags & BX_ALTDATA) != 0)
 6850                 return (1);
 6851         /* -1 is full truncation. */
 6852         if (lastlbn == -1)
 6853                 return (1);
 6854         /*
 6855          * If this is a partial truncate we only want those
 6856          * blocks and indirect blocks that cover the range
 6857          * we're after.
 6858          */
 6859         lbn = bp->b_lblkno;
 6860         if (lbn < 0)
 6861                 lbn = -(lbn + lbn_level(lbn));
 6862         if (lbn < lastlbn)
 6863                 return (0);
 6864         /* Here we only truncate lblkno if it's partial. */
 6865         if (lbn == lastlbn) {
 6866                 if (lastoff == 0)
 6867                         return (0);
 6868                 *blkoffp = lastoff;
 6869         }
 6870         return (1);
 6871 }
 6872 
 6873 /*
 6874  * Eliminate any dependencies that exist in memory beyond lblkno:off
 6875  */
 6876 static void
 6877 trunc_dependencies(ip, freeblks, lastlbn, lastoff, flags)
 6878         struct inode *ip;
 6879         struct freeblks *freeblks;
 6880         ufs_lbn_t lastlbn;
 6881         int lastoff;
 6882         int flags;
 6883 {
 6884         struct bufobj *bo;
 6885         struct vnode *vp;
 6886         struct buf *bp;
 6887         struct fs *fs;
 6888         int blkoff;
 6889 
 6890         /*
 6891          * We must wait for any I/O in progress to finish so that
 6892          * all potential buffers on the dirty list will be visible.
 6893          * Once they are all there, walk the list and get rid of
 6894          * any dependencies.
 6895          */
 6896         fs = ip->i_fs;
 6897         vp = ITOV(ip);
 6898         bo = &vp->v_bufobj;
 6899         BO_LOCK(bo);
 6900         drain_output(vp);
 6901         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs)
 6902                 bp->b_vflags &= ~BV_SCANNED;
 6903 restart:
 6904         TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
 6905                 if (bp->b_vflags & BV_SCANNED)
 6906                         continue;
 6907                 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) {
 6908                         bp->b_vflags |= BV_SCANNED;
 6909                         continue;
 6910                 }
 6911                 if ((bp = getdirtybuf(bp, BO_LOCKPTR(bo), MNT_WAIT)) == NULL)
 6912                         goto restart;
 6913                 BO_UNLOCK(bo);
 6914                 if (deallocate_dependencies(bp, freeblks, blkoff))
 6915                         bqrelse(bp);
 6916                 else
 6917                         brelse(bp);
 6918                 BO_LOCK(bo);
 6919                 goto restart;
 6920         }
 6921         /*
 6922          * Now do the work of vtruncbuf while also matching indirect blocks.
 6923          */
 6924         TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs)
 6925                 bp->b_vflags &= ~BV_SCANNED;
 6926 cleanrestart:
 6927         TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) {
 6928                 if (bp->b_vflags & BV_SCANNED)
 6929                         continue;
 6930                 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) {
 6931                         bp->b_vflags |= BV_SCANNED;
 6932                         continue;
 6933                 }
 6934                 if (BUF_LOCK(bp,
 6935                     LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
 6936                     BO_LOCKPTR(bo)) == ENOLCK) {
 6937                         BO_LOCK(bo);
 6938                         goto cleanrestart;
 6939                 }
 6940                 bp->b_vflags |= BV_SCANNED;
 6941                 bremfree(bp);
 6942                 if (blkoff != 0) {
 6943                         allocbuf(bp, blkoff);
 6944                         bqrelse(bp);
 6945                 } else {
 6946                         bp->b_flags |= B_INVAL | B_NOCACHE | B_RELBUF;
 6947                         brelse(bp);
 6948                 }
 6949                 BO_LOCK(bo);
 6950                 goto cleanrestart;
 6951         }
 6952         drain_output(vp);
 6953         BO_UNLOCK(bo);
 6954 }
 6955 
 6956 static int
 6957 cancel_pagedep(pagedep, freeblks, blkoff)
 6958         struct pagedep *pagedep;
 6959         struct freeblks *freeblks;
 6960         int blkoff;
 6961 {
 6962         struct jremref *jremref;
 6963         struct jmvref *jmvref;
 6964         struct dirrem *dirrem, *tmp;
 6965         int i;
 6966 
 6967         /*
 6968          * Copy any directory remove dependencies to the list
 6969          * to be processed after the freeblks proceeds.  If
 6970          * directory entry never made it to disk they
 6971          * can be dumped directly onto the work list.
 6972          */
 6973         LIST_FOREACH_SAFE(dirrem, &pagedep->pd_dirremhd, dm_next, tmp) {
 6974                 /* Skip this directory removal if it is intended to remain. */
 6975                 if (dirrem->dm_offset < blkoff)
 6976                         continue;
 6977                 /*
 6978                  * If there are any dirrems we wait for the journal write
 6979                  * to complete and then restart the buf scan as the lock
 6980                  * has been dropped.
 6981                  */
 6982                 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) {
 6983                         jwait(&jremref->jr_list, MNT_WAIT);
 6984                         return (ERESTART);
 6985                 }
 6986                 LIST_REMOVE(dirrem, dm_next);
 6987                 dirrem->dm_dirinum = pagedep->pd_ino;
 6988                 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &dirrem->dm_list);
 6989         }
 6990         while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) {
 6991                 jwait(&jmvref->jm_list, MNT_WAIT);
 6992                 return (ERESTART);
 6993         }
 6994         /*
 6995          * When we're partially truncating a pagedep we just want to flush
 6996          * journal entries and return.  There can not be any adds in the
 6997          * truncated portion of the directory and newblk must remain if
 6998          * part of the block remains.
 6999          */
 7000         if (blkoff != 0) {
 7001                 struct diradd *dap;
 7002 
 7003                 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
 7004                         if (dap->da_offset > blkoff)
 7005                                 panic("cancel_pagedep: diradd %p off %d > %d",
 7006                                     dap, dap->da_offset, blkoff);
 7007                 for (i = 0; i < DAHASHSZ; i++)
 7008                         LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist)
 7009                                 if (dap->da_offset > blkoff)
 7010                                         panic("cancel_pagedep: diradd %p off %d > %d",
 7011                                             dap, dap->da_offset, blkoff);
 7012                 return (0);
 7013         }
 7014         /*
 7015          * There should be no directory add dependencies present
 7016          * as the directory could not be truncated until all
 7017          * children were removed.
 7018          */
 7019         KASSERT(LIST_FIRST(&pagedep->pd_pendinghd) == NULL,
 7020             ("deallocate_dependencies: pendinghd != NULL"));
 7021         for (i = 0; i < DAHASHSZ; i++)
 7022                 KASSERT(LIST_FIRST(&pagedep->pd_diraddhd[i]) == NULL,
 7023                     ("deallocate_dependencies: diraddhd != NULL"));
 7024         if ((pagedep->pd_state & NEWBLOCK) != 0)
 7025                 free_newdirblk(pagedep->pd_newdirblk);
 7026         if (free_pagedep(pagedep) == 0)
 7027                 panic("Failed to free pagedep %p", pagedep);
 7028         return (0);
 7029 }
 7030 
 7031 /*
 7032  * Reclaim any dependency structures from a buffer that is about to
 7033  * be reallocated to a new vnode. The buffer must be locked, thus,
 7034  * no I/O completion operations can occur while we are manipulating
 7035  * its associated dependencies. The mutex is held so that other I/O's
 7036  * associated with related dependencies do not occur.
 7037  */
 7038 static int
 7039 deallocate_dependencies(bp, freeblks, off)
 7040         struct buf *bp;
 7041         struct freeblks *freeblks;
 7042         int off;
 7043 {
 7044         struct indirdep *indirdep;
 7045         struct pagedep *pagedep;
 7046         struct allocdirect *adp;
 7047         struct worklist *wk, *wkn;
 7048 
 7049         ACQUIRE_LOCK(&lk);
 7050         LIST_FOREACH_SAFE(wk, &bp->b_dep, wk_list, wkn) {
 7051                 switch (wk->wk_type) {
 7052                 case D_INDIRDEP:
 7053                         indirdep = WK_INDIRDEP(wk);
 7054                         if (bp->b_lblkno >= 0 ||
 7055                             bp->b_blkno != indirdep->ir_savebp->b_lblkno)
 7056                                 panic("deallocate_dependencies: not indir");
 7057                         cancel_indirdep(indirdep, bp, freeblks);
 7058                         continue;
 7059 
 7060                 case D_PAGEDEP:
 7061                         pagedep = WK_PAGEDEP(wk);
 7062                         if (cancel_pagedep(pagedep, freeblks, off)) {
 7063                                 FREE_LOCK(&lk);
 7064                                 return (ERESTART);
 7065                         }
 7066                         continue;
 7067 
 7068                 case D_ALLOCINDIR:
 7069                         /*
 7070                          * Simply remove the allocindir, we'll find it via
 7071                          * the indirdep where we can clear pointers if
 7072                          * needed.
 7073                          */
 7074                         WORKLIST_REMOVE(wk);
 7075                         continue;
 7076 
 7077                 case D_FREEWORK:
 7078                         /*
 7079                          * A truncation is waiting for the zero'd pointers
 7080                          * to be written.  It can be freed when the freeblks
 7081                          * is journaled.
 7082                          */
 7083                         WORKLIST_REMOVE(wk);
 7084                         wk->wk_state |= ONDEPLIST;
 7085                         WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk);
 7086                         break;
 7087 
 7088                 case D_ALLOCDIRECT:
 7089                         adp = WK_ALLOCDIRECT(wk);
 7090                         if (off != 0)
 7091                                 continue;
 7092                         /* FALLTHROUGH */
 7093                 default:
 7094                         panic("deallocate_dependencies: Unexpected type %s",
 7095                             TYPENAME(wk->wk_type));
 7096                         /* NOTREACHED */
 7097                 }
 7098         }
 7099         FREE_LOCK(&lk);
 7100         /*
 7101          * Don't throw away this buf, we were partially truncating and
 7102          * some deps may always remain.
 7103          */
 7104         if (off) {
 7105                 allocbuf(bp, off);
 7106                 bp->b_vflags |= BV_SCANNED;
 7107                 return (EBUSY);
 7108         }
 7109         bp->b_flags |= B_INVAL | B_NOCACHE;
 7110 
 7111         return (0);
 7112 }
 7113 
 7114 /*
 7115  * An allocdirect is being canceled due to a truncate.  We must make sure
 7116  * the journal entry is released in concert with the blkfree that releases
 7117  * the storage.  Completed journal entries must not be released until the
 7118  * space is no longer pointed to by the inode or in the bitmap.
 7119  */
 7120 static void
 7121 cancel_allocdirect(adphead, adp, freeblks)
 7122         struct allocdirectlst *adphead;
 7123         struct allocdirect *adp;
 7124         struct freeblks *freeblks;
 7125 {
 7126         struct freework *freework;
 7127         struct newblk *newblk;
 7128         struct worklist *wk;
 7129 
 7130         TAILQ_REMOVE(adphead, adp, ad_next);
 7131         newblk = (struct newblk *)adp;
 7132         freework = NULL;
 7133         /*
 7134          * Find the correct freework structure.
 7135          */
 7136         LIST_FOREACH(wk, &freeblks->fb_freeworkhd, wk_list) {
 7137                 if (wk->wk_type != D_FREEWORK)
 7138                         continue;
 7139                 freework = WK_FREEWORK(wk);
 7140                 if (freework->fw_blkno == newblk->nb_newblkno)
 7141                         break;
 7142         }
 7143         if (freework == NULL)
 7144                 panic("cancel_allocdirect: Freework not found");
 7145         /*
 7146          * If a newblk exists at all we still have the journal entry that
 7147          * initiated the allocation so we do not need to journal the free.
 7148          */
 7149         cancel_jfreeblk(freeblks, freework->fw_blkno);
 7150         /*
 7151          * If the journal hasn't been written the jnewblk must be passed
 7152          * to the call to ffs_blkfree that reclaims the space.  We accomplish
 7153          * this by linking the journal dependency into the freework to be
 7154          * freed when freework_freeblock() is called.  If the journal has
 7155          * been written we can simply reclaim the journal space when the
 7156          * freeblks work is complete.
 7157          */
 7158         freework->fw_jnewblk = cancel_newblk(newblk, &freework->fw_list,
 7159             &freeblks->fb_jwork);
 7160         WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list);
 7161 }
 7162 
 7163 
 7164 /*
 7165  * Cancel a new block allocation.  May be an indirect or direct block.  We
 7166  * remove it from various lists and return any journal record that needs to
 7167  * be resolved by the caller.
 7168  *
 7169  * A special consideration is made for indirects which were never pointed
 7170  * at on disk and will never be found once this block is released.
 7171  */
 7172 static struct jnewblk *
 7173 cancel_newblk(newblk, wk, wkhd)
 7174         struct newblk *newblk;
 7175         struct worklist *wk;
 7176         struct workhead *wkhd;
 7177 {
 7178         struct jnewblk *jnewblk;
 7179 
 7180         CTR1(KTR_SUJ, "cancel_newblk: blkno %jd", newblk->nb_newblkno);
 7181             
 7182         newblk->nb_state |= GOINGAWAY;
 7183         /*
 7184          * Previously we traversed the completedhd on each indirdep
 7185          * attached to this newblk to cancel them and gather journal
 7186          * work.  Since we need only the oldest journal segment and
 7187          * the lowest point on the tree will always have the oldest
 7188          * journal segment we are free to release the segments
 7189          * of any subordinates and may leave the indirdep list to
 7190          * indirdep_complete() when this newblk is freed.
 7191          */
 7192         if (newblk->nb_state & ONDEPLIST) {
 7193                 newblk->nb_state &= ~ONDEPLIST;
 7194                 LIST_REMOVE(newblk, nb_deps);
 7195         }
 7196         if (newblk->nb_state & ONWORKLIST)
 7197                 WORKLIST_REMOVE(&newblk->nb_list);
 7198         /*
 7199          * If the journal entry hasn't been written we save a pointer to
 7200          * the dependency that frees it until it is written or the
 7201          * superseding operation completes.
 7202          */
 7203         jnewblk = newblk->nb_jnewblk;
 7204         if (jnewblk != NULL && wk != NULL) {
 7205                 newblk->nb_jnewblk = NULL;
 7206                 jnewblk->jn_dep = wk;
 7207         }
 7208         if (!LIST_EMPTY(&newblk->nb_jwork))
 7209                 jwork_move(wkhd, &newblk->nb_jwork);
 7210         /*
 7211          * When truncating we must free the newdirblk early to remove
 7212          * the pagedep from the hash before returning.
 7213          */
 7214         if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL)
 7215                 free_newdirblk(WK_NEWDIRBLK(wk));
 7216         if (!LIST_EMPTY(&newblk->nb_newdirblk))
 7217                 panic("cancel_newblk: extra newdirblk");
 7218 
 7219         return (jnewblk);
 7220 }
 7221 
 7222 /*
 7223  * Schedule the freefrag associated with a newblk to be released once
 7224  * the pointers are written and the previous block is no longer needed.
 7225  */
 7226 static void
 7227 newblk_freefrag(newblk)
 7228         struct newblk *newblk;
 7229 {
 7230         struct freefrag *freefrag;
 7231 
 7232         if (newblk->nb_freefrag == NULL)
 7233                 return;
 7234         freefrag = newblk->nb_freefrag;
 7235         newblk->nb_freefrag = NULL;
 7236         freefrag->ff_state |= COMPLETE;
 7237         if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE)
 7238                 add_to_worklist(&freefrag->ff_list, 0);
 7239 }
 7240 
 7241 /*
 7242  * Free a newblk. Generate a new freefrag work request if appropriate.
 7243  * This must be called after the inode pointer and any direct block pointers
 7244  * are valid or fully removed via truncate or frag extension.
 7245  */
 7246 static void
 7247 free_newblk(newblk)
 7248         struct newblk *newblk;
 7249 {
 7250         struct indirdep *indirdep;
 7251         struct worklist *wk;
 7252 
 7253         KASSERT(newblk->nb_jnewblk == NULL,
 7254             ("free_newblk: jnewblk %p still attached", newblk->nb_jnewblk));
 7255         KASSERT(newblk->nb_list.wk_type != D_NEWBLK,
 7256             ("free_newblk: unclaimed newblk"));
 7257         rw_assert(&lk, RA_WLOCKED);
 7258         newblk_freefrag(newblk);
 7259         if (newblk->nb_state & ONDEPLIST)
 7260                 LIST_REMOVE(newblk, nb_deps);
 7261         if (newblk->nb_state & ONWORKLIST)
 7262                 WORKLIST_REMOVE(&newblk->nb_list);
 7263         LIST_REMOVE(newblk, nb_hash);
 7264         if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL)
 7265                 free_newdirblk(WK_NEWDIRBLK(wk));
 7266         if (!LIST_EMPTY(&newblk->nb_newdirblk))
 7267                 panic("free_newblk: extra newdirblk");
 7268         while ((indirdep = LIST_FIRST(&newblk->nb_indirdeps)) != NULL)
 7269                 indirdep_complete(indirdep);
 7270         handle_jwork(&newblk->nb_jwork);
 7271         WORKITEM_FREE(newblk, D_NEWBLK);
 7272 }
 7273 
 7274 /*
 7275  * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep.
 7276  * This routine must be called with splbio interrupts blocked.
 7277  */
 7278 static void
 7279 free_newdirblk(newdirblk)
 7280         struct newdirblk *newdirblk;
 7281 {
 7282         struct pagedep *pagedep;
 7283         struct diradd *dap;
 7284         struct worklist *wk;
 7285 
 7286         rw_assert(&lk, RA_WLOCKED);
 7287         WORKLIST_REMOVE(&newdirblk->db_list);
 7288         /*
 7289          * If the pagedep is still linked onto the directory buffer
 7290          * dependency chain, then some of the entries on the
 7291          * pd_pendinghd list may not be committed to disk yet. In
 7292          * this case, we will simply clear the NEWBLOCK flag and
 7293          * let the pd_pendinghd list be processed when the pagedep
 7294          * is next written. If the pagedep is no longer on the buffer
 7295          * dependency chain, then all the entries on the pd_pending
 7296          * list are committed to disk and we can free them here.
 7297          */
 7298         pagedep = newdirblk->db_pagedep;
 7299         pagedep->pd_state &= ~NEWBLOCK;
 7300         if ((pagedep->pd_state & ONWORKLIST) == 0) {
 7301                 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
 7302                         free_diradd(dap, NULL);
 7303                 /*
 7304                  * If no dependencies remain, the pagedep will be freed.
 7305                  */
 7306                 free_pagedep(pagedep);
 7307         }
 7308         /* Should only ever be one item in the list. */
 7309         while ((wk = LIST_FIRST(&newdirblk->db_mkdir)) != NULL) {
 7310                 WORKLIST_REMOVE(wk);
 7311                 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
 7312         }
 7313         WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
 7314 }
 7315 
 7316 /*
 7317  * Prepare an inode to be freed. The actual free operation is not
 7318  * done until the zero'ed inode has been written to disk.
 7319  */
 7320 void
 7321 softdep_freefile(pvp, ino, mode)
 7322         struct vnode *pvp;
 7323         ino_t ino;
 7324         int mode;
 7325 {
 7326         struct inode *ip = VTOI(pvp);
 7327         struct inodedep *inodedep;
 7328         struct freefile *freefile;
 7329         struct freeblks *freeblks;
 7330 
 7331         /*
 7332          * This sets up the inode de-allocation dependency.
 7333          */
 7334         freefile = malloc(sizeof(struct freefile),
 7335                 M_FREEFILE, M_SOFTDEP_FLAGS);
 7336         workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount);
 7337         freefile->fx_mode = mode;
 7338         freefile->fx_oldinum = ino;
 7339         freefile->fx_devvp = ip->i_devvp;
 7340         LIST_INIT(&freefile->fx_jwork);
 7341         UFS_LOCK(ip->i_ump);
 7342         ip->i_fs->fs_pendinginodes += 1;
 7343         UFS_UNLOCK(ip->i_ump);
 7344 
 7345         /*
 7346          * If the inodedep does not exist, then the zero'ed inode has
 7347          * been written to disk. If the allocated inode has never been
 7348          * written to disk, then the on-disk inode is zero'ed. In either
 7349          * case we can free the file immediately.  If the journal was
 7350          * canceled before being written the inode will never make it to
 7351          * disk and we must send the canceled journal entrys to
 7352          * ffs_freefile() to be cleared in conjunction with the bitmap.
 7353          * Any blocks waiting on the inode to write can be safely freed
 7354          * here as it will never been written.
 7355          */
 7356         ACQUIRE_LOCK(&lk);
 7357         inodedep_lookup(pvp->v_mount, ino, 0, &inodedep);
 7358         if (inodedep) {
 7359                 /*
 7360                  * Clear out freeblks that no longer need to reference
 7361                  * this inode.
 7362                  */
 7363                 while ((freeblks =
 7364                     TAILQ_FIRST(&inodedep->id_freeblklst)) != NULL) {
 7365                         TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks,
 7366                             fb_next);
 7367                         freeblks->fb_state &= ~ONDEPLIST;
 7368                 }
 7369                 /*
 7370                  * Remove this inode from the unlinked list.
 7371                  */
 7372                 if (inodedep->id_state & UNLINKED) {
 7373                         /*
 7374                          * Save the journal work to be freed with the bitmap
 7375                          * before we clear UNLINKED.  Otherwise it can be lost
 7376                          * if the inode block is written.
 7377                          */
 7378                         handle_bufwait(inodedep, &freefile->fx_jwork);
 7379                         clear_unlinked_inodedep(inodedep);
 7380                         /* Re-acquire inodedep as we've dropped lk. */
 7381                         inodedep_lookup(pvp->v_mount, ino, 0, &inodedep);
 7382                 }
 7383         }
 7384         if (inodedep == NULL || check_inode_unwritten(inodedep)) {
 7385                 FREE_LOCK(&lk);
 7386                 handle_workitem_freefile(freefile);
 7387                 return;
 7388         }
 7389         if ((inodedep->id_state & DEPCOMPLETE) == 0)
 7390                 inodedep->id_state |= GOINGAWAY;
 7391         WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
 7392         FREE_LOCK(&lk);
 7393         if (ip->i_number == ino)
 7394                 ip->i_flag |= IN_MODIFIED;
 7395 }
 7396 
 7397 /*
 7398  * Check to see if an inode has never been written to disk. If
 7399  * so free the inodedep and return success, otherwise return failure.
 7400  * This routine must be called with splbio interrupts blocked.
 7401  *
 7402  * If we still have a bitmap dependency, then the inode has never
 7403  * been written to disk. Drop the dependency as it is no longer
 7404  * necessary since the inode is being deallocated. We set the
 7405  * ALLCOMPLETE flags since the bitmap now properly shows that the
 7406  * inode is not allocated. Even if the inode is actively being
 7407  * written, it has been rolled back to its zero'ed state, so we
 7408  * are ensured that a zero inode is what is on the disk. For short
 7409  * lived files, this change will usually result in removing all the
 7410  * dependencies from the inode so that it can be freed immediately.
 7411  */
 7412 static int
 7413 check_inode_unwritten(inodedep)
 7414         struct inodedep *inodedep;
 7415 {
 7416 
 7417         rw_assert(&lk, RA_WLOCKED);
 7418 
 7419         if ((inodedep->id_state & (DEPCOMPLETE | UNLINKED)) != 0 ||
 7420             !LIST_EMPTY(&inodedep->id_dirremhd) ||
 7421             !LIST_EMPTY(&inodedep->id_pendinghd) ||
 7422             !LIST_EMPTY(&inodedep->id_bufwait) ||
 7423             !LIST_EMPTY(&inodedep->id_inowait) ||
 7424             !TAILQ_EMPTY(&inodedep->id_inoreflst) ||
 7425             !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
 7426             !TAILQ_EMPTY(&inodedep->id_newinoupdt) ||
 7427             !TAILQ_EMPTY(&inodedep->id_extupdt) ||
 7428             !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
 7429             !TAILQ_EMPTY(&inodedep->id_freeblklst) ||
 7430             inodedep->id_mkdiradd != NULL || 
 7431             inodedep->id_nlinkdelta != 0)
 7432                 return (0);
 7433         /*
 7434          * Another process might be in initiate_write_inodeblock_ufs[12]
 7435          * trying to allocate memory without holding "Softdep Lock".
 7436          */
 7437         if ((inodedep->id_state & IOSTARTED) != 0 &&
 7438             inodedep->id_savedino1 == NULL)
 7439                 return (0);
 7440 
 7441         if (inodedep->id_state & ONDEPLIST)
 7442                 LIST_REMOVE(inodedep, id_deps);
 7443         inodedep->id_state &= ~ONDEPLIST;
 7444         inodedep->id_state |= ALLCOMPLETE;
 7445         inodedep->id_bmsafemap = NULL;
 7446         if (inodedep->id_state & ONWORKLIST)
 7447                 WORKLIST_REMOVE(&inodedep->id_list);
 7448         if (inodedep->id_savedino1 != NULL) {
 7449                 free(inodedep->id_savedino1, M_SAVEDINO);
 7450                 inodedep->id_savedino1 = NULL;
 7451         }
 7452         if (free_inodedep(inodedep) == 0)
 7453                 panic("check_inode_unwritten: busy inode");
 7454         return (1);
 7455 }
 7456 
 7457 /*
 7458  * Try to free an inodedep structure. Return 1 if it could be freed.
 7459  */
 7460 static int
 7461 free_inodedep(inodedep)
 7462         struct inodedep *inodedep;
 7463 {
 7464 
 7465         rw_assert(&lk, RA_WLOCKED);
 7466         if ((inodedep->id_state & (ONWORKLIST | UNLINKED)) != 0 ||
 7467             (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
 7468             !LIST_EMPTY(&inodedep->id_dirremhd) ||
 7469             !LIST_EMPTY(&inodedep->id_pendinghd) ||
 7470             !LIST_EMPTY(&inodedep->id_bufwait) ||
 7471             !LIST_EMPTY(&inodedep->id_inowait) ||
 7472             !TAILQ_EMPTY(&inodedep->id_inoreflst) ||
 7473             !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
 7474             !TAILQ_EMPTY(&inodedep->id_newinoupdt) ||
 7475             !TAILQ_EMPTY(&inodedep->id_extupdt) ||
 7476             !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
 7477             !TAILQ_EMPTY(&inodedep->id_freeblklst) ||
 7478             inodedep->id_mkdiradd != NULL ||
 7479             inodedep->id_nlinkdelta != 0 ||
 7480             inodedep->id_savedino1 != NULL)
 7481                 return (0);
 7482         if (inodedep->id_state & ONDEPLIST)
 7483                 LIST_REMOVE(inodedep, id_deps);
 7484         LIST_REMOVE(inodedep, id_hash);
 7485         WORKITEM_FREE(inodedep, D_INODEDEP);
 7486         return (1);
 7487 }
 7488 
 7489 /*
 7490  * Free the block referenced by a freework structure.  The parent freeblks
 7491  * structure is released and completed when the final cg bitmap reaches
 7492  * the disk.  This routine may be freeing a jnewblk which never made it to
 7493  * disk in which case we do not have to wait as the operation is undone
 7494  * in memory immediately.
 7495  */
 7496 static void
 7497 freework_freeblock(freework)
 7498         struct freework *freework;
 7499 {
 7500         struct freeblks *freeblks;
 7501         struct jnewblk *jnewblk;
 7502         struct ufsmount *ump;
 7503         struct workhead wkhd;
 7504         struct fs *fs;
 7505         int bsize;
 7506         int needj;
 7507 
 7508         rw_assert(&lk, RA_WLOCKED);
 7509         /*
 7510          * Handle partial truncate separately.
 7511          */
 7512         if (freework->fw_indir) {
 7513                 complete_trunc_indir(freework);
 7514                 return;
 7515         }
 7516         freeblks = freework->fw_freeblks;
 7517         ump = VFSTOUFS(freeblks->fb_list.wk_mp);
 7518         fs = ump->um_fs;
 7519         needj = MOUNTEDSUJ(freeblks->fb_list.wk_mp) != 0;
 7520         bsize = lfragtosize(fs, freework->fw_frags);
 7521         LIST_INIT(&wkhd);
 7522         /*
 7523          * DEPCOMPLETE is cleared in indirblk_insert() if the block lives
 7524          * on the indirblk hashtable and prevents premature freeing.
 7525          */
 7526         freework->fw_state |= DEPCOMPLETE;
 7527         /*
 7528          * SUJ needs to wait for the segment referencing freed indirect
 7529          * blocks to expire so that we know the checker will not confuse
 7530          * a re-allocated indirect block with its old contents.
 7531          */
 7532         if (needj && freework->fw_lbn <= -NDADDR)
 7533                 indirblk_insert(freework);
 7534         /*
 7535          * If we are canceling an existing jnewblk pass it to the free
 7536          * routine, otherwise pass the freeblk which will ultimately
 7537          * release the freeblks.  If we're not journaling, we can just
 7538          * free the freeblks immediately.
 7539          */
 7540         jnewblk = freework->fw_jnewblk;
 7541         if (jnewblk != NULL) {
 7542                 cancel_jnewblk(jnewblk, &wkhd);
 7543                 needj = 0;
 7544         } else if (needj) {
 7545                 freework->fw_state |= DELAYEDFREE;
 7546                 freeblks->fb_cgwait++;
 7547                 WORKLIST_INSERT(&wkhd, &freework->fw_list);
 7548         }
 7549         FREE_LOCK(&lk);
 7550         freeblks_free(ump, freeblks, btodb(bsize));
 7551         CTR4(KTR_SUJ,
 7552             "freework_freeblock: ino %d blkno %jd lbn %jd size %ld",
 7553             freeblks->fb_inum, freework->fw_blkno, freework->fw_lbn, bsize);
 7554         ffs_blkfree(ump, fs, freeblks->fb_devvp, freework->fw_blkno, bsize,
 7555             freeblks->fb_inum, freeblks->fb_vtype, &wkhd);
 7556         ACQUIRE_LOCK(&lk);
 7557         /*
 7558          * The jnewblk will be discarded and the bits in the map never
 7559          * made it to disk.  We can immediately free the freeblk.
 7560          */
 7561         if (needj == 0)
 7562                 handle_written_freework(freework);
 7563 }
 7564 
 7565 /*
 7566  * We enqueue freework items that need processing back on the freeblks and
 7567  * add the freeblks to the worklist.  This makes it easier to find all work
 7568  * required to flush a truncation in process_truncates().
 7569  */
 7570 static void
 7571 freework_enqueue(freework)
 7572         struct freework *freework;
 7573 {
 7574         struct freeblks *freeblks;
 7575 
 7576         freeblks = freework->fw_freeblks;
 7577         if ((freework->fw_state & INPROGRESS) == 0)
 7578                 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list);
 7579         if ((freeblks->fb_state &
 7580             (ONWORKLIST | INPROGRESS | ALLCOMPLETE)) == ALLCOMPLETE &&
 7581             LIST_EMPTY(&freeblks->fb_jblkdephd))
 7582                 add_to_worklist(&freeblks->fb_list, WK_NODELAY);
 7583 }
 7584 
 7585 /*
 7586  * Start, continue, or finish the process of freeing an indirect block tree.
 7587  * The free operation may be paused at any point with fw_off containing the
 7588  * offset to restart from.  This enables us to implement some flow control
 7589  * for large truncates which may fan out and generate a huge number of
 7590  * dependencies.
 7591  */
 7592 static void
 7593 handle_workitem_indirblk(freework)
 7594         struct freework *freework;
 7595 {
 7596         struct freeblks *freeblks;
 7597         struct ufsmount *ump;
 7598         struct fs *fs;
 7599 
 7600         freeblks = freework->fw_freeblks;
 7601         ump = VFSTOUFS(freeblks->fb_list.wk_mp);
 7602         fs = ump->um_fs;
 7603         if (freework->fw_state & DEPCOMPLETE) {
 7604                 handle_written_freework(freework);
 7605                 return;
 7606         }
 7607         if (freework->fw_off == NINDIR(fs)) {
 7608                 freework_freeblock(freework);
 7609                 return;
 7610         }
 7611         freework->fw_state |= INPROGRESS;
 7612         FREE_LOCK(&lk);
 7613         indir_trunc(freework, fsbtodb(fs, freework->fw_blkno),
 7614             freework->fw_lbn);
 7615         ACQUIRE_LOCK(&lk);
 7616 }
 7617 
 7618 /*
 7619  * Called when a freework structure attached to a cg buf is written.  The
 7620  * ref on either the parent or the freeblks structure is released and
 7621  * the freeblks is added back to the worklist if there is more work to do.
 7622  */
 7623 static void
 7624 handle_written_freework(freework)
 7625         struct freework *freework;
 7626 {
 7627         struct freeblks *freeblks;
 7628         struct freework *parent;
 7629 
 7630         freeblks = freework->fw_freeblks;
 7631         parent = freework->fw_parent;
 7632         if (freework->fw_state & DELAYEDFREE)
 7633                 freeblks->fb_cgwait--;
 7634         freework->fw_state |= COMPLETE;
 7635         if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE)
 7636                 WORKITEM_FREE(freework, D_FREEWORK);
 7637         if (parent) {
 7638                 if (--parent->fw_ref == 0)
 7639                         freework_enqueue(parent);
 7640                 return;
 7641         }
 7642         if (--freeblks->fb_ref != 0)
 7643                 return;
 7644         if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST | INPROGRESS)) ==
 7645             ALLCOMPLETE && LIST_EMPTY(&freeblks->fb_jblkdephd)) 
 7646                 add_to_worklist(&freeblks->fb_list, WK_NODELAY);
 7647 }
 7648 
 7649 /*
 7650  * This workitem routine performs the block de-allocation.
 7651  * The workitem is added to the pending list after the updated
 7652  * inode block has been written to disk.  As mentioned above,
 7653  * checks regarding the number of blocks de-allocated (compared
 7654  * to the number of blocks allocated for the file) are also
 7655  * performed in this function.
 7656  */
 7657 static int
 7658 handle_workitem_freeblocks(freeblks, flags)
 7659         struct freeblks *freeblks;
 7660         int flags;
 7661 {
 7662         struct freework *freework;
 7663         struct newblk *newblk;
 7664         struct allocindir *aip;
 7665         struct ufsmount *ump;
 7666         struct worklist *wk;
 7667 
 7668         KASSERT(LIST_EMPTY(&freeblks->fb_jblkdephd),
 7669             ("handle_workitem_freeblocks: Journal entries not written."));
 7670         ump = VFSTOUFS(freeblks->fb_list.wk_mp);
 7671         ACQUIRE_LOCK(&lk);
 7672         while ((wk = LIST_FIRST(&freeblks->fb_freeworkhd)) != NULL) {
 7673                 WORKLIST_REMOVE(wk);
 7674                 switch (wk->wk_type) {
 7675                 case D_DIRREM:
 7676                         wk->wk_state |= COMPLETE;
 7677                         add_to_worklist(wk, 0);
 7678                         continue;
 7679 
 7680                 case D_ALLOCDIRECT:
 7681                         free_newblk(WK_NEWBLK(wk));
 7682                         continue;
 7683 
 7684                 case D_ALLOCINDIR:
 7685                         aip = WK_ALLOCINDIR(wk);
 7686                         freework = NULL;
 7687                         if (aip->ai_state & DELAYEDFREE) {
 7688                                 FREE_LOCK(&lk);
 7689                                 freework = newfreework(ump, freeblks, NULL,
 7690                                     aip->ai_lbn, aip->ai_newblkno,
 7691                                     ump->um_fs->fs_frag, 0, 0);
 7692                                 ACQUIRE_LOCK(&lk);
 7693                         }
 7694                         newblk = WK_NEWBLK(wk);
 7695                         if (newblk->nb_jnewblk) {
 7696                                 freework->fw_jnewblk = newblk->nb_jnewblk;
 7697                                 newblk->nb_jnewblk->jn_dep = &freework->fw_list;
 7698                                 newblk->nb_jnewblk = NULL;
 7699                         }
 7700                         free_newblk(newblk);
 7701                         continue;
 7702 
 7703                 case D_FREEWORK:
 7704                         freework = WK_FREEWORK(wk);
 7705                         if (freework->fw_lbn <= -NDADDR)
 7706                                 handle_workitem_indirblk(freework);
 7707                         else
 7708                                 freework_freeblock(freework);
 7709                         continue;
 7710                 default:
 7711                         panic("handle_workitem_freeblocks: Unknown type %s",
 7712                             TYPENAME(wk->wk_type));
 7713                 }
 7714         }
 7715         if (freeblks->fb_ref != 0) {
 7716                 freeblks->fb_state &= ~INPROGRESS;
 7717                 wake_worklist(&freeblks->fb_list);
 7718                 freeblks = NULL;
 7719         }
 7720         FREE_LOCK(&lk);
 7721         if (freeblks)
 7722                 return handle_complete_freeblocks(freeblks, flags);
 7723         return (0);
 7724 }
 7725 
 7726 /*
 7727  * Handle completion of block free via truncate.  This allows fs_pending
 7728  * to track the actual free block count more closely than if we only updated
 7729  * it at the end.  We must be careful to handle cases where the block count
 7730  * on free was incorrect.
 7731  */
 7732 static void
 7733 freeblks_free(ump, freeblks, blocks)
 7734         struct ufsmount *ump;
 7735         struct freeblks *freeblks;
 7736         int blocks;
 7737 {
 7738         struct fs *fs;
 7739         ufs2_daddr_t remain;
 7740 
 7741         UFS_LOCK(ump);
 7742         remain = -freeblks->fb_chkcnt;
 7743         freeblks->fb_chkcnt += blocks;
 7744         if (remain > 0) {
 7745                 if (remain < blocks)
 7746                         blocks = remain;
 7747                 fs = ump->um_fs;
 7748                 fs->fs_pendingblocks -= blocks;
 7749         }
 7750         UFS_UNLOCK(ump);
 7751 }
 7752 
 7753 /*
 7754  * Once all of the freework workitems are complete we can retire the
 7755  * freeblocks dependency and any journal work awaiting completion.  This
 7756  * can not be called until all other dependencies are stable on disk.
 7757  */
 7758 static int
 7759 handle_complete_freeblocks(freeblks, flags)
 7760         struct freeblks *freeblks;
 7761         int flags;
 7762 {
 7763         struct inodedep *inodedep;
 7764         struct inode *ip;
 7765         struct vnode *vp;
 7766         struct fs *fs;
 7767         struct ufsmount *ump;
 7768         ufs2_daddr_t spare;
 7769 
 7770         ump = VFSTOUFS(freeblks->fb_list.wk_mp);
 7771         fs = ump->um_fs;
 7772         flags = LK_EXCLUSIVE | flags;
 7773         spare = freeblks->fb_chkcnt;
 7774 
 7775         /*
 7776          * If we did not release the expected number of blocks we may have
 7777          * to adjust the inode block count here.  Only do so if it wasn't
 7778          * a truncation to zero and the modrev still matches.
 7779          */
 7780         if (spare && freeblks->fb_len != 0) {
 7781                 if (ffs_vgetf(freeblks->fb_list.wk_mp, freeblks->fb_inum,
 7782                     flags, &vp, FFSV_FORCEINSMQ) != 0)
 7783                         return (EBUSY);
 7784                 ip = VTOI(vp);
 7785                 if (DIP(ip, i_modrev) == freeblks->fb_modrev) {
 7786                         DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - spare);
 7787                         ip->i_flag |= IN_CHANGE;
 7788                         /*
 7789                          * We must wait so this happens before the
 7790                          * journal is reclaimed.
 7791                          */
 7792                         ffs_update(vp, 1);
 7793                 }
 7794                 vput(vp);
 7795         }
 7796         if (spare < 0) {
 7797                 UFS_LOCK(ump);
 7798                 fs->fs_pendingblocks += spare;
 7799                 UFS_UNLOCK(ump);
 7800         }
 7801 #ifdef QUOTA
 7802         /* Handle spare. */
 7803         if (spare)
 7804                 quotaadj(freeblks->fb_quota, ump, -spare);
 7805         quotarele(freeblks->fb_quota);
 7806 #endif
 7807         ACQUIRE_LOCK(&lk);
 7808         if (freeblks->fb_state & ONDEPLIST) {
 7809                 inodedep_lookup(freeblks->fb_list.wk_mp, freeblks->fb_inum,
 7810                     0, &inodedep);
 7811                 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, fb_next);
 7812                 freeblks->fb_state &= ~ONDEPLIST;
 7813                 if (TAILQ_EMPTY(&inodedep->id_freeblklst))
 7814                         free_inodedep(inodedep);
 7815         }
 7816         /*
 7817          * All of the freeblock deps must be complete prior to this call
 7818          * so it's now safe to complete earlier outstanding journal entries.
 7819          */
 7820         handle_jwork(&freeblks->fb_jwork);
 7821         WORKITEM_FREE(freeblks, D_FREEBLKS);
 7822         FREE_LOCK(&lk);
 7823         return (0);
 7824 }
 7825 
 7826 /*
 7827  * Release blocks associated with the freeblks and stored in the indirect
 7828  * block dbn. If level is greater than SINGLE, the block is an indirect block
 7829  * and recursive calls to indirtrunc must be used to cleanse other indirect
 7830  * blocks.
 7831  *
 7832  * This handles partial and complete truncation of blocks.  Partial is noted
 7833  * with goingaway == 0.  In this case the freework is completed after the
 7834  * zero'd indirects are written to disk.  For full truncation the freework
 7835  * is completed after the block is freed.
 7836  */
 7837 static void
 7838 indir_trunc(freework, dbn, lbn)
 7839         struct freework *freework;
 7840         ufs2_daddr_t dbn;
 7841         ufs_lbn_t lbn;
 7842 {
 7843         struct freework *nfreework;
 7844         struct workhead wkhd;
 7845         struct freeblks *freeblks;
 7846         struct buf *bp;
 7847         struct fs *fs;
 7848         struct indirdep *indirdep;
 7849         struct ufsmount *ump;
 7850         ufs1_daddr_t *bap1 = 0;
 7851         ufs2_daddr_t nb, nnb, *bap2 = 0;
 7852         ufs_lbn_t lbnadd, nlbn;
 7853         int i, nblocks, ufs1fmt;
 7854         int freedblocks;
 7855         int goingaway;
 7856         int freedeps;
 7857         int needj;
 7858         int level;
 7859         int cnt;
 7860 
 7861         freeblks = freework->fw_freeblks;
 7862         ump = VFSTOUFS(freeblks->fb_list.wk_mp);
 7863         fs = ump->um_fs;
 7864         /*
 7865          * Get buffer of block pointers to be freed.  There are three cases:
 7866          * 
 7867          * 1) Partial truncate caches the indirdep pointer in the freework
 7868          *    which provides us a back copy to the save bp which holds the
 7869          *    pointers we want to clear.  When this completes the zero
 7870          *    pointers are written to the real copy.
 7871          * 2) The indirect is being completely truncated, cancel_indirdep()
 7872          *    eliminated the real copy and placed the indirdep on the saved
 7873          *    copy.  The indirdep and buf are discarded when this completes.
 7874          * 3) The indirect was not in memory, we read a copy off of the disk
 7875          *    using the devvp and drop and invalidate the buffer when we're
 7876          *    done.
 7877          */
 7878         goingaway = 1;
 7879         indirdep = NULL;
 7880         if (freework->fw_indir != NULL) {
 7881                 goingaway = 0;
 7882                 indirdep = freework->fw_indir;
 7883                 bp = indirdep->ir_savebp;
 7884                 if (bp == NULL || bp->b_blkno != dbn)
 7885                         panic("indir_trunc: Bad saved buf %p blkno %jd",
 7886                             bp, (intmax_t)dbn);
 7887         } else if ((bp = incore(&freeblks->fb_devvp->v_bufobj, dbn)) != NULL) {
 7888                 /*
 7889                  * The lock prevents the buf dep list from changing and
 7890                  * indirects on devvp should only ever have one dependency.
 7891                  */
 7892                 indirdep = WK_INDIRDEP(LIST_FIRST(&bp->b_dep));
 7893                 if (indirdep == NULL || (indirdep->ir_state & GOINGAWAY) == 0)
 7894                         panic("indir_trunc: Bad indirdep %p from buf %p",
 7895                             indirdep, bp);
 7896         } else if (bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize,
 7897             NOCRED, &bp) != 0) {
 7898                 brelse(bp);
 7899                 return;
 7900         }
 7901         ACQUIRE_LOCK(&lk);
 7902         /* Protects against a race with complete_trunc_indir(). */
 7903         freework->fw_state &= ~INPROGRESS;
 7904         /*
 7905          * If we have an indirdep we need to enforce the truncation order
 7906          * and discard it when it is complete.
 7907          */
 7908         if (indirdep) {
 7909                 if (freework != TAILQ_FIRST(&indirdep->ir_trunc) &&
 7910                     !TAILQ_EMPTY(&indirdep->ir_trunc)) {
 7911                         /*
 7912                          * Add the complete truncate to the list on the
 7913                          * indirdep to enforce in-order processing.
 7914                          */
 7915                         if (freework->fw_indir == NULL)
 7916                                 TAILQ_INSERT_TAIL(&indirdep->ir_trunc,
 7917                                     freework, fw_next);
 7918                         FREE_LOCK(&lk);
 7919                         return;
 7920                 }
 7921                 /*
 7922                  * If we're goingaway, free the indirdep.  Otherwise it will
 7923                  * linger until the write completes.
 7924                  */
 7925                 if (goingaway) {
 7926                         free_indirdep(indirdep);
 7927                         ump->um_numindirdeps -= 1;
 7928                 }
 7929         }
 7930         FREE_LOCK(&lk);
 7931         /* Initialize pointers depending on block size. */
 7932         if (ump->um_fstype == UFS1) {
 7933                 bap1 = (ufs1_daddr_t *)bp->b_data;
 7934                 nb = bap1[freework->fw_off];
 7935                 ufs1fmt = 1;
 7936         } else {
 7937                 bap2 = (ufs2_daddr_t *)bp->b_data;
 7938                 nb = bap2[freework->fw_off];
 7939                 ufs1fmt = 0;
 7940         }
 7941         level = lbn_level(lbn);
 7942         needj = MOUNTEDSUJ(UFSTOVFS(ump)) != 0;
 7943         lbnadd = lbn_offset(fs, level);
 7944         nblocks = btodb(fs->fs_bsize);
 7945         nfreework = freework;
 7946         freedeps = 0;
 7947         cnt = 0;
 7948         /*
 7949          * Reclaim blocks.  Traverses into nested indirect levels and
 7950          * arranges for the current level to be freed when subordinates
 7951          * are free when journaling.
 7952          */
 7953         for (i = freework->fw_off; i < NINDIR(fs); i++, nb = nnb) {
 7954                 if (i != NINDIR(fs) - 1) {
 7955                         if (ufs1fmt)
 7956                                 nnb = bap1[i+1];
 7957                         else
 7958                                 nnb = bap2[i+1];
 7959                 } else
 7960                         nnb = 0;
 7961                 if (nb == 0)
 7962                         continue;
 7963                 cnt++;
 7964                 if (level != 0) {
 7965                         nlbn = (lbn + 1) - (i * lbnadd);
 7966                         if (needj != 0) {
 7967                                 nfreework = newfreework(ump, freeblks, freework,
 7968                                     nlbn, nb, fs->fs_frag, 0, 0);
 7969                                 freedeps++;
 7970                         }
 7971                         indir_trunc(nfreework, fsbtodb(fs, nb), nlbn);
 7972                 } else {
 7973                         struct freedep *freedep;
 7974 
 7975                         /*
 7976                          * Attempt to aggregate freedep dependencies for
 7977                          * all blocks being released to the same CG.
 7978                          */
 7979                         LIST_INIT(&wkhd);
 7980                         if (needj != 0 &&
 7981                             (nnb == 0 || (dtog(fs, nb) != dtog(fs, nnb)))) {
 7982                                 freedep = newfreedep(freework);
 7983                                 WORKLIST_INSERT_UNLOCKED(&wkhd,
 7984                                     &freedep->fd_list);
 7985                                 freedeps++;
 7986                         }
 7987                         CTR3(KTR_SUJ,
 7988                             "indir_trunc: ino %d blkno %jd size %ld",
 7989                             freeblks->fb_inum, nb, fs->fs_bsize);
 7990                         ffs_blkfree(ump, fs, freeblks->fb_devvp, nb,
 7991                             fs->fs_bsize, freeblks->fb_inum,
 7992                             freeblks->fb_vtype, &wkhd);
 7993                 }
 7994         }
 7995         if (goingaway) {
 7996                 bp->b_flags |= B_INVAL | B_NOCACHE;
 7997                 brelse(bp);
 7998         }
 7999         freedblocks = 0;
 8000         if (level == 0)
 8001                 freedblocks = (nblocks * cnt);
 8002         if (needj == 0)
 8003                 freedblocks += nblocks;
 8004         freeblks_free(ump, freeblks, freedblocks);
 8005         /*
 8006          * If we are journaling set up the ref counts and offset so this
 8007          * indirect can be completed when its children are free.
 8008          */
 8009         if (needj) {
 8010                 ACQUIRE_LOCK(&lk);
 8011                 freework->fw_off = i;
 8012                 freework->fw_ref += freedeps;
 8013                 freework->fw_ref -= NINDIR(fs) + 1;
 8014                 if (level == 0)
 8015                         freeblks->fb_cgwait += freedeps;
 8016                 if (freework->fw_ref == 0)
 8017                         freework_freeblock(freework);
 8018                 FREE_LOCK(&lk);
 8019                 return;
 8020         }
 8021         /*
 8022          * If we're not journaling we can free the indirect now.
 8023          */
 8024         dbn = dbtofsb(fs, dbn);
 8025         CTR3(KTR_SUJ,
 8026             "indir_trunc 2: ino %d blkno %jd size %ld",
 8027             freeblks->fb_inum, dbn, fs->fs_bsize);
 8028         ffs_blkfree(ump, fs, freeblks->fb_devvp, dbn, fs->fs_bsize,
 8029             freeblks->fb_inum, freeblks->fb_vtype, NULL);
 8030         /* Non SUJ softdep does single-threaded truncations. */
 8031         if (freework->fw_blkno == dbn) {
 8032                 freework->fw_state |= ALLCOMPLETE;
 8033                 ACQUIRE_LOCK(&lk);
 8034                 handle_written_freework(freework);
 8035                 FREE_LOCK(&lk);
 8036         }
 8037         return;
 8038 }
 8039 
 8040 /*
 8041  * Cancel an allocindir when it is removed via truncation.  When bp is not
 8042  * NULL the indirect never appeared on disk and is scheduled to be freed
 8043  * independently of the indir so we can more easily track journal work.
 8044  */
 8045 static void
 8046 cancel_allocindir(aip, bp, freeblks, trunc)
 8047         struct allocindir *aip;
 8048         struct buf *bp;
 8049         struct freeblks *freeblks;
 8050         int trunc;
 8051 {
 8052         struct indirdep *indirdep;
 8053         struct freefrag *freefrag;
 8054         struct newblk *newblk;
 8055 
 8056         newblk = (struct newblk *)aip;
 8057         LIST_REMOVE(aip, ai_next);
 8058         /*
 8059          * We must eliminate the pointer in bp if it must be freed on its
 8060          * own due to partial truncate or pending journal work.
 8061          */
 8062         if (bp && (trunc || newblk->nb_jnewblk)) {
 8063                 /*
 8064                  * Clear the pointer and mark the aip to be freed
 8065                  * directly if it never existed on disk.
 8066                  */
 8067                 aip->ai_state |= DELAYEDFREE;
 8068                 indirdep = aip->ai_indirdep;
 8069                 if (indirdep->ir_state & UFS1FMT)
 8070                         ((ufs1_daddr_t *)bp->b_data)[aip->ai_offset] = 0;
 8071                 else
 8072                         ((ufs2_daddr_t *)bp->b_data)[aip->ai_offset] = 0;
 8073         }
 8074         /*
 8075          * When truncating the previous pointer will be freed via
 8076          * savedbp.  Eliminate the freefrag which would dup free.
 8077          */
 8078         if (trunc && (freefrag = newblk->nb_freefrag) != NULL) {
 8079                 newblk->nb_freefrag = NULL;
 8080                 if (freefrag->ff_jdep)
 8081                         cancel_jfreefrag(
 8082                             WK_JFREEFRAG(freefrag->ff_jdep));
 8083                 jwork_move(&freeblks->fb_jwork, &freefrag->ff_jwork);
 8084                 WORKITEM_FREE(freefrag, D_FREEFRAG);
 8085         }
 8086         /*
 8087          * If the journal hasn't been written the jnewblk must be passed
 8088          * to the call to ffs_blkfree that reclaims the space.  We accomplish
 8089          * this by leaving the journal dependency on the newblk to be freed
 8090          * when a freework is created in handle_workitem_freeblocks().
 8091          */
 8092         cancel_newblk(newblk, NULL, &freeblks->fb_jwork);
 8093         WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list);
 8094 }
 8095 
 8096 /*
 8097  * Create the mkdir dependencies for . and .. in a new directory.  Link them
 8098  * in to a newdirblk so any subsequent additions are tracked properly.  The
 8099  * caller is responsible for adding the mkdir1 dependency to the journal
 8100  * and updating id_mkdiradd.  This function returns with lk held.
 8101  */
 8102 static struct mkdir *
 8103 setup_newdir(dap, newinum, dinum, newdirbp, mkdirp)
 8104         struct diradd *dap;
 8105         ino_t newinum;
 8106         ino_t dinum;
 8107         struct buf *newdirbp;
 8108         struct mkdir **mkdirp;
 8109 {
 8110         struct newblk *newblk;
 8111         struct pagedep *pagedep;
 8112         struct inodedep *inodedep;
 8113         struct newdirblk *newdirblk = 0;
 8114         struct mkdir *mkdir1, *mkdir2;
 8115         struct worklist *wk;
 8116         struct jaddref *jaddref;
 8117         struct mount *mp;
 8118 
 8119         mp = dap->da_list.wk_mp;
 8120         newdirblk = malloc(sizeof(struct newdirblk), M_NEWDIRBLK,
 8121             M_SOFTDEP_FLAGS);
 8122         workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp);
 8123         LIST_INIT(&newdirblk->db_mkdir);
 8124         mkdir1 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS);
 8125         workitem_alloc(&mkdir1->md_list, D_MKDIR, mp);
 8126         mkdir1->md_state = ATTACHED | MKDIR_BODY;
 8127         mkdir1->md_diradd = dap;
 8128         mkdir1->md_jaddref = NULL;
 8129         mkdir2 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS);
 8130         workitem_alloc(&mkdir2->md_list, D_MKDIR, mp);
 8131         mkdir2->md_state = ATTACHED | MKDIR_PARENT;
 8132         mkdir2->md_diradd = dap;
 8133         mkdir2->md_jaddref = NULL;
 8134         if (MOUNTEDSUJ(mp) == 0) {
 8135                 mkdir1->md_state |= DEPCOMPLETE;
 8136                 mkdir2->md_state |= DEPCOMPLETE;
 8137         }
 8138         /*
 8139          * Dependency on "." and ".." being written to disk.
 8140          */
 8141         mkdir1->md_buf = newdirbp;
 8142         ACQUIRE_LOCK(&lk);
 8143         LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
 8144         /*
 8145          * We must link the pagedep, allocdirect, and newdirblk for
 8146          * the initial file page so the pointer to the new directory
 8147          * is not written until the directory contents are live and
 8148          * any subsequent additions are not marked live until the
 8149          * block is reachable via the inode.
 8150          */
 8151         if (pagedep_lookup(mp, newdirbp, newinum, 0, 0, &pagedep) == 0)
 8152                 panic("setup_newdir: lost pagedep");
 8153         LIST_FOREACH(wk, &newdirbp->b_dep, wk_list)
 8154                 if (wk->wk_type == D_ALLOCDIRECT)
 8155                         break;
 8156         if (wk == NULL)
 8157                 panic("setup_newdir: lost allocdirect");
 8158         if (pagedep->pd_state & NEWBLOCK)
 8159                 panic("setup_newdir: NEWBLOCK already set");
 8160         newblk = WK_NEWBLK(wk);
 8161         pagedep->pd_state |= NEWBLOCK;
 8162         pagedep->pd_newdirblk = newdirblk;
 8163         newdirblk->db_pagedep = pagedep;
 8164         WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list);
 8165         WORKLIST_INSERT(&newdirblk->db_mkdir, &mkdir1->md_list);
 8166         /*
 8167          * Look up the inodedep for the parent directory so that we
 8168          * can link mkdir2 into the pending dotdot jaddref or
 8169          * the inode write if there is none.  If the inode is
 8170          * ALLCOMPLETE and no jaddref is present all dependencies have
 8171          * been satisfied and mkdir2 can be freed.
 8172          */
 8173         inodedep_lookup(mp, dinum, 0, &inodedep);
 8174         if (MOUNTEDSUJ(mp)) {
 8175                 if (inodedep == NULL)
 8176                         panic("setup_newdir: Lost parent.");
 8177                 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
 8178                     inoreflst);
 8179                 KASSERT(jaddref != NULL && jaddref->ja_parent == newinum &&
 8180                     (jaddref->ja_state & MKDIR_PARENT),
 8181                     ("setup_newdir: bad dotdot jaddref %p", jaddref));
 8182                 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
 8183                 mkdir2->md_jaddref = jaddref;
 8184                 jaddref->ja_mkdir = mkdir2;
 8185         } else if (inodedep == NULL ||
 8186             (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
 8187                 dap->da_state &= ~MKDIR_PARENT;
 8188                 WORKITEM_FREE(mkdir2, D_MKDIR);
 8189                 mkdir2 = NULL;
 8190         } else {
 8191                 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
 8192                 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir2->md_list);
 8193         }
 8194         *mkdirp = mkdir2;
 8195 
 8196         return (mkdir1);
 8197 }
 8198 
 8199 /*
 8200  * Directory entry addition dependencies.
 8201  * 
 8202  * When adding a new directory entry, the inode (with its incremented link
 8203  * count) must be written to disk before the directory entry's pointer to it.
 8204  * Also, if the inode is newly allocated, the corresponding freemap must be
 8205  * updated (on disk) before the directory entry's pointer. These requirements
 8206  * are met via undo/redo on the directory entry's pointer, which consists
 8207  * simply of the inode number.
 8208  * 
 8209  * As directory entries are added and deleted, the free space within a
 8210  * directory block can become fragmented.  The ufs filesystem will compact
 8211  * a fragmented directory block to make space for a new entry. When this
 8212  * occurs, the offsets of previously added entries change. Any "diradd"
 8213  * dependency structures corresponding to these entries must be updated with
 8214  * the new offsets.
 8215  */
 8216 
 8217 /*
 8218  * This routine is called after the in-memory inode's link
 8219  * count has been incremented, but before the directory entry's
 8220  * pointer to the inode has been set.
 8221  */
 8222 int
 8223 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
 8224         struct buf *bp;         /* buffer containing directory block */
 8225         struct inode *dp;       /* inode for directory */
 8226         off_t diroffset;        /* offset of new entry in directory */
 8227         ino_t newinum;          /* inode referenced by new directory entry */
 8228         struct buf *newdirbp;   /* non-NULL => contents of new mkdir */
 8229         int isnewblk;           /* entry is in a newly allocated block */
 8230 {
 8231         int offset;             /* offset of new entry within directory block */
 8232         ufs_lbn_t lbn;          /* block in directory containing new entry */
 8233         struct fs *fs;
 8234         struct diradd *dap;
 8235         struct newblk *newblk;
 8236         struct pagedep *pagedep;
 8237         struct inodedep *inodedep;
 8238         struct newdirblk *newdirblk = 0;
 8239         struct mkdir *mkdir1, *mkdir2;
 8240         struct jaddref *jaddref;
 8241         struct mount *mp;
 8242         int isindir;
 8243 
 8244         /*
 8245          * Whiteouts have no dependencies.
 8246          */
 8247         if (newinum == WINO) {
 8248                 if (newdirbp != NULL)
 8249                         bdwrite(newdirbp);
 8250                 return (0);
 8251         }
 8252         jaddref = NULL;
 8253         mkdir1 = mkdir2 = NULL;
 8254         mp = UFSTOVFS(dp->i_ump);
 8255         fs = dp->i_fs;
 8256         lbn = lblkno(fs, diroffset);
 8257         offset = blkoff(fs, diroffset);
 8258         dap = malloc(sizeof(struct diradd), M_DIRADD,
 8259                 M_SOFTDEP_FLAGS|M_ZERO);
 8260         workitem_alloc(&dap->da_list, D_DIRADD, mp);
 8261         dap->da_offset = offset;
 8262         dap->da_newinum = newinum;
 8263         dap->da_state = ATTACHED;
 8264         LIST_INIT(&dap->da_jwork);
 8265         isindir = bp->b_lblkno >= NDADDR;
 8266         if (isnewblk &&
 8267             (isindir ? blkoff(fs, diroffset) : fragoff(fs, diroffset)) == 0) {
 8268                 newdirblk = malloc(sizeof(struct newdirblk),
 8269                     M_NEWDIRBLK, M_SOFTDEP_FLAGS);
 8270                 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp);
 8271                 LIST_INIT(&newdirblk->db_mkdir);
 8272         }
 8273         /*
 8274          * If we're creating a new directory setup the dependencies and set
 8275          * the dap state to wait for them.  Otherwise it's COMPLETE and
 8276          * we can move on.
 8277          */
 8278         if (newdirbp == NULL) {
 8279                 dap->da_state |= DEPCOMPLETE;
 8280                 ACQUIRE_LOCK(&lk);
 8281         } else {
 8282                 dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
 8283                 mkdir1 = setup_newdir(dap, newinum, dp->i_number, newdirbp,
 8284                     &mkdir2);
 8285         }
 8286         /*
 8287          * Link into parent directory pagedep to await its being written.
 8288          */
 8289         pagedep_lookup(mp, bp, dp->i_number, lbn, DEPALLOC, &pagedep);
 8290 #ifdef DEBUG
 8291         if (diradd_lookup(pagedep, offset) != NULL)
 8292                 panic("softdep_setup_directory_add: %p already at off %d\n",
 8293                     diradd_lookup(pagedep, offset), offset);
 8294 #endif
 8295         dap->da_pagedep = pagedep;
 8296         LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
 8297             da_pdlist);
 8298         inodedep_lookup(mp, newinum, DEPALLOC | NODELAY, &inodedep);
 8299         /*
 8300          * If we're journaling, link the diradd into the jaddref so it
 8301          * may be completed after the journal entry is written.  Otherwise,
 8302          * link the diradd into its inodedep.  If the inode is not yet
 8303          * written place it on the bufwait list, otherwise do the post-inode
 8304          * write processing to put it on the id_pendinghd list.
 8305          */
 8306         if (MOUNTEDSUJ(mp)) {
 8307                 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
 8308                     inoreflst);
 8309                 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number,
 8310                     ("softdep_setup_directory_add: bad jaddref %p", jaddref));
 8311                 jaddref->ja_diroff = diroffset;
 8312                 jaddref->ja_diradd = dap;
 8313                 add_to_journal(&jaddref->ja_list);
 8314         } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
 8315                 diradd_inode_written(dap, inodedep);
 8316         else
 8317                 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
 8318         /*
 8319          * Add the journal entries for . and .. links now that the primary
 8320          * link is written.
 8321          */
 8322         if (mkdir1 != NULL && MOUNTEDSUJ(mp)) {
 8323                 jaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref,
 8324                     inoreflst, if_deps);
 8325                 KASSERT(jaddref != NULL &&
 8326                     jaddref->ja_ino == jaddref->ja_parent &&
 8327                     (jaddref->ja_state & MKDIR_BODY),
 8328                     ("softdep_setup_directory_add: bad dot jaddref %p",
 8329                     jaddref));
 8330                 mkdir1->md_jaddref = jaddref;
 8331                 jaddref->ja_mkdir = mkdir1;
 8332                 /*
 8333                  * It is important that the dotdot journal entry
 8334                  * is added prior to the dot entry since dot writes
 8335                  * both the dot and dotdot links.  These both must
 8336                  * be added after the primary link for the journal
 8337                  * to remain consistent.
 8338                  */
 8339                 add_to_journal(&mkdir2->md_jaddref->ja_list);
 8340                 add_to_journal(&jaddref->ja_list);
 8341         }
 8342         /*
 8343          * If we are adding a new directory remember this diradd so that if
 8344          * we rename it we can keep the dot and dotdot dependencies.  If
 8345          * we are adding a new name for an inode that has a mkdiradd we
 8346          * must be in rename and we have to move the dot and dotdot
 8347          * dependencies to this new name.  The old name is being orphaned
 8348          * soon.
 8349          */
 8350         if (mkdir1 != NULL) {
 8351                 if (inodedep->id_mkdiradd != NULL)
 8352                         panic("softdep_setup_directory_add: Existing mkdir");
 8353                 inodedep->id_mkdiradd = dap;
 8354         } else if (inodedep->id_mkdiradd)
 8355                 merge_diradd(inodedep, dap);
 8356         if (newdirblk) {
 8357                 /*
 8358                  * There is nothing to do if we are already tracking
 8359                  * this block.
 8360                  */
 8361                 if ((pagedep->pd_state & NEWBLOCK) != 0) {
 8362                         WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
 8363                         FREE_LOCK(&lk);
 8364                         return (0);
 8365                 }
 8366                 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk)
 8367                     == 0)
 8368                         panic("softdep_setup_directory_add: lost entry");
 8369                 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list);
 8370                 pagedep->pd_state |= NEWBLOCK;
 8371                 pagedep->pd_newdirblk = newdirblk;
 8372                 newdirblk->db_pagedep = pagedep;
 8373                 FREE_LOCK(&lk);
 8374                 /*
 8375                  * If we extended into an indirect signal direnter to sync.
 8376                  */
 8377                 if (isindir)
 8378                         return (1);
 8379                 return (0);
 8380         }
 8381         FREE_LOCK(&lk);
 8382         return (0);
 8383 }
 8384 
 8385 /*
 8386  * This procedure is called to change the offset of a directory
 8387  * entry when compacting a directory block which must be owned
 8388  * exclusively by the caller. Note that the actual entry movement
 8389  * must be done in this procedure to ensure that no I/O completions
 8390  * occur while the move is in progress.
 8391  */
 8392 void 
 8393 softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize)
 8394         struct buf *bp;         /* Buffer holding directory block. */
 8395         struct inode *dp;       /* inode for directory */
 8396         caddr_t base;           /* address of dp->i_offset */
 8397         caddr_t oldloc;         /* address of old directory location */
 8398         caddr_t newloc;         /* address of new directory location */
 8399         int entrysize;          /* size of directory entry */
 8400 {
 8401         int offset, oldoffset, newoffset;
 8402         struct pagedep *pagedep;
 8403         struct jmvref *jmvref;
 8404         struct diradd *dap;
 8405         struct direct *de;
 8406         struct mount *mp;
 8407         ufs_lbn_t lbn;
 8408         int flags;
 8409 
 8410         mp = UFSTOVFS(dp->i_ump);
 8411         de = (struct direct *)oldloc;
 8412         jmvref = NULL;
 8413         flags = 0;
 8414         /*
 8415          * Moves are always journaled as it would be too complex to
 8416          * determine if any affected adds or removes are present in the
 8417          * journal.
 8418          */
 8419         if (MOUNTEDSUJ(mp)) {
 8420                 flags = DEPALLOC;
 8421                 jmvref = newjmvref(dp, de->d_ino,
 8422                     dp->i_offset + (oldloc - base),
 8423                     dp->i_offset + (newloc - base));
 8424         }
 8425         lbn = lblkno(dp->i_fs, dp->i_offset);
 8426         offset = blkoff(dp->i_fs, dp->i_offset);
 8427         oldoffset = offset + (oldloc - base);
 8428         newoffset = offset + (newloc - base);
 8429         ACQUIRE_LOCK(&lk);
 8430         if (pagedep_lookup(mp, bp, dp->i_number, lbn, flags, &pagedep) == 0)
 8431                 goto done;
 8432         dap = diradd_lookup(pagedep, oldoffset);
 8433         if (dap) {
 8434                 dap->da_offset = newoffset;
 8435                 newoffset = DIRADDHASH(newoffset);
 8436                 oldoffset = DIRADDHASH(oldoffset);
 8437                 if ((dap->da_state & ALLCOMPLETE) != ALLCOMPLETE &&
 8438                     newoffset != oldoffset) {
 8439                         LIST_REMOVE(dap, da_pdlist);
 8440                         LIST_INSERT_HEAD(&pagedep->pd_diraddhd[newoffset],
 8441                             dap, da_pdlist);
 8442                 }
 8443         }
 8444 done:
 8445         if (jmvref) {
 8446                 jmvref->jm_pagedep = pagedep;
 8447                 LIST_INSERT_HEAD(&pagedep->pd_jmvrefhd, jmvref, jm_deps);
 8448                 add_to_journal(&jmvref->jm_list);
 8449         }
 8450         bcopy(oldloc, newloc, entrysize);
 8451         FREE_LOCK(&lk);
 8452 }
 8453 
 8454 /*
 8455  * Move the mkdir dependencies and journal work from one diradd to another
 8456  * when renaming a directory.  The new name must depend on the mkdir deps
 8457  * completing as the old name did.  Directories can only have one valid link
 8458  * at a time so one must be canonical.
 8459  */
 8460 static void
 8461 merge_diradd(inodedep, newdap)
 8462         struct inodedep *inodedep;
 8463         struct diradd *newdap;
 8464 {
 8465         struct diradd *olddap;
 8466         struct mkdir *mkdir, *nextmd;
 8467         short state;
 8468 
 8469         olddap = inodedep->id_mkdiradd;
 8470         inodedep->id_mkdiradd = newdap;
 8471         if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
 8472                 newdap->da_state &= ~DEPCOMPLETE;
 8473                 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
 8474                         nextmd = LIST_NEXT(mkdir, md_mkdirs);
 8475                         if (mkdir->md_diradd != olddap)
 8476                                 continue;
 8477                         mkdir->md_diradd = newdap;
 8478                         state = mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY);
 8479                         newdap->da_state |= state;
 8480                         olddap->da_state &= ~state;
 8481                         if ((olddap->da_state &
 8482                             (MKDIR_PARENT | MKDIR_BODY)) == 0)
 8483                                 break;
 8484                 }
 8485                 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0)
 8486                         panic("merge_diradd: unfound ref");
 8487         }
 8488         /*
 8489          * Any mkdir related journal items are not safe to be freed until
 8490          * the new name is stable.
 8491          */
 8492         jwork_move(&newdap->da_jwork, &olddap->da_jwork);
 8493         olddap->da_state |= DEPCOMPLETE;
 8494         complete_diradd(olddap);
 8495 }
 8496 
 8497 /*
 8498  * Move the diradd to the pending list when all diradd dependencies are
 8499  * complete.
 8500  */
 8501 static void
 8502 complete_diradd(dap)
 8503         struct diradd *dap;
 8504 {
 8505         struct pagedep *pagedep;
 8506 
 8507         if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
 8508                 if (dap->da_state & DIRCHG)
 8509                         pagedep = dap->da_previous->dm_pagedep;
 8510                 else
 8511                         pagedep = dap->da_pagedep;
 8512                 LIST_REMOVE(dap, da_pdlist);
 8513                 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
 8514         }
 8515 }
 8516 
 8517 /*
 8518  * Cancel a diradd when a dirrem overlaps with it.  We must cancel the journal
 8519  * add entries and conditonally journal the remove.
 8520  */
 8521 static void
 8522 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref)
 8523         struct diradd *dap;
 8524         struct dirrem *dirrem;
 8525         struct jremref *jremref;
 8526         struct jremref *dotremref;
 8527         struct jremref *dotdotremref;
 8528 {
 8529         struct inodedep *inodedep;
 8530         struct jaddref *jaddref;
 8531         struct inoref *inoref;
 8532         struct mkdir *mkdir;
 8533 
 8534         /*
 8535          * If no remove references were allocated we're on a non-journaled
 8536          * filesystem and can skip the cancel step.
 8537          */
 8538         if (jremref == NULL) {
 8539                 free_diradd(dap, NULL);
 8540                 return;
 8541         }
 8542         /*
 8543          * Cancel the primary name an free it if it does not require
 8544          * journaling.
 8545          */
 8546         if (inodedep_lookup(dap->da_list.wk_mp, dap->da_newinum,
 8547             0, &inodedep) != 0) {
 8548                 /* Abort the addref that reference this diradd.  */
 8549                 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
 8550                         if (inoref->if_list.wk_type != D_JADDREF)
 8551                                 continue;
 8552                         jaddref = (struct jaddref *)inoref;
 8553                         if (jaddref->ja_diradd != dap)
 8554                                 continue;
 8555                         if (cancel_jaddref(jaddref, inodedep,
 8556                             &dirrem->dm_jwork) == 0) {
 8557                                 free_jremref(jremref);
 8558                                 jremref = NULL;
 8559                         }
 8560                         break;
 8561                 }
 8562         }
 8563         /*
 8564          * Cancel subordinate names and free them if they do not require
 8565          * journaling.
 8566          */
 8567         if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
 8568                 LIST_FOREACH(mkdir, &mkdirlisthd, md_mkdirs) {
 8569                         if (mkdir->md_diradd != dap)
 8570                                 continue;
 8571                         if ((jaddref = mkdir->md_jaddref) == NULL)
 8572                                 continue;
 8573                         mkdir->md_jaddref = NULL;
 8574                         if (mkdir->md_state & MKDIR_PARENT) {
 8575                                 if (cancel_jaddref(jaddref, NULL,
 8576                                     &dirrem->dm_jwork) == 0) {
 8577                                         free_jremref(dotdotremref);
 8578                                         dotdotremref = NULL;
 8579                                 }
 8580                         } else {
 8581                                 if (cancel_jaddref(jaddref, inodedep,
 8582                                     &dirrem->dm_jwork) == 0) {
 8583                                         free_jremref(dotremref);
 8584                                         dotremref = NULL;
 8585                                 }
 8586                         }
 8587                 }
 8588         }
 8589 
 8590         if (jremref)
 8591                 journal_jremref(dirrem, jremref, inodedep);
 8592         if (dotremref)
 8593                 journal_jremref(dirrem, dotremref, inodedep);
 8594         if (dotdotremref)
 8595                 journal_jremref(dirrem, dotdotremref, NULL);
 8596         jwork_move(&dirrem->dm_jwork, &dap->da_jwork);
 8597         free_diradd(dap, &dirrem->dm_jwork);
 8598 }
 8599 
 8600 /*
 8601  * Free a diradd dependency structure. This routine must be called
 8602  * with splbio interrupts blocked.
 8603  */
 8604 static void
 8605 free_diradd(dap, wkhd)
 8606         struct diradd *dap;
 8607         struct workhead *wkhd;
 8608 {
 8609         struct dirrem *dirrem;
 8610         struct pagedep *pagedep;
 8611         struct inodedep *inodedep;
 8612         struct mkdir *mkdir, *nextmd;
 8613 
 8614         rw_assert(&lk, RA_WLOCKED);
 8615         LIST_REMOVE(dap, da_pdlist);
 8616         if (dap->da_state & ONWORKLIST)
 8617                 WORKLIST_REMOVE(&dap->da_list);
 8618         if ((dap->da_state & DIRCHG) == 0) {
 8619                 pagedep = dap->da_pagedep;
 8620         } else {
 8621                 dirrem = dap->da_previous;
 8622                 pagedep = dirrem->dm_pagedep;
 8623                 dirrem->dm_dirinum = pagedep->pd_ino;
 8624                 dirrem->dm_state |= COMPLETE;
 8625                 if (LIST_EMPTY(&dirrem->dm_jremrefhd))
 8626                         add_to_worklist(&dirrem->dm_list, 0);
 8627         }
 8628         if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum,
 8629             0, &inodedep) != 0)
 8630                 if (inodedep->id_mkdiradd == dap)
 8631                         inodedep->id_mkdiradd = NULL;
 8632         if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
 8633                 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
 8634                         nextmd = LIST_NEXT(mkdir, md_mkdirs);
 8635                         if (mkdir->md_diradd != dap)
 8636                                 continue;
 8637                         dap->da_state &=
 8638                             ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY));
 8639                         LIST_REMOVE(mkdir, md_mkdirs);
 8640                         if (mkdir->md_state & ONWORKLIST)
 8641                                 WORKLIST_REMOVE(&mkdir->md_list);
 8642                         if (mkdir->md_jaddref != NULL)
 8643                                 panic("free_diradd: Unexpected jaddref");
 8644                         WORKITEM_FREE(mkdir, D_MKDIR);
 8645                         if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
 8646                                 break;
 8647                 }
 8648                 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0)
 8649                         panic("free_diradd: unfound ref");
 8650         }
 8651         if (inodedep)
 8652                 free_inodedep(inodedep);
 8653         /*
 8654          * Free any journal segments waiting for the directory write.
 8655          */
 8656         handle_jwork(&dap->da_jwork);
 8657         WORKITEM_FREE(dap, D_DIRADD);
 8658 }
 8659 
 8660 /*
 8661  * Directory entry removal dependencies.
 8662  * 
 8663  * When removing a directory entry, the entry's inode pointer must be
 8664  * zero'ed on disk before the corresponding inode's link count is decremented
 8665  * (possibly freeing the inode for re-use). This dependency is handled by
 8666  * updating the directory entry but delaying the inode count reduction until
 8667  * after the directory block has been written to disk. After this point, the
 8668  * inode count can be decremented whenever it is convenient.
 8669  */
 8670 
 8671 /*
 8672  * This routine should be called immediately after removing
 8673  * a directory entry.  The inode's link count should not be
 8674  * decremented by the calling procedure -- the soft updates
 8675  * code will do this task when it is safe.
 8676  */
 8677 void 
 8678 softdep_setup_remove(bp, dp, ip, isrmdir)
 8679         struct buf *bp;         /* buffer containing directory block */
 8680         struct inode *dp;       /* inode for the directory being modified */
 8681         struct inode *ip;       /* inode for directory entry being removed */
 8682         int isrmdir;            /* indicates if doing RMDIR */
 8683 {
 8684         struct dirrem *dirrem, *prevdirrem;
 8685         struct inodedep *inodedep;
 8686         int direct;
 8687 
 8688         /*
 8689          * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.  We want
 8690          * newdirrem() to setup the full directory remove which requires
 8691          * isrmdir > 1.
 8692          */
 8693         dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
 8694         /*
 8695          * Add the dirrem to the inodedep's pending remove list for quick
 8696          * discovery later.
 8697          */
 8698         if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0,
 8699             &inodedep) == 0)
 8700                 panic("softdep_setup_remove: Lost inodedep.");
 8701         KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked"));
 8702         dirrem->dm_state |= ONDEPLIST;
 8703         LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext);
 8704 
 8705         /*
 8706          * If the COMPLETE flag is clear, then there were no active
 8707          * entries and we want to roll back to a zeroed entry until
 8708          * the new inode is committed to disk. If the COMPLETE flag is
 8709          * set then we have deleted an entry that never made it to
 8710          * disk. If the entry we deleted resulted from a name change,
 8711          * then the old name still resides on disk. We cannot delete
 8712          * its inode (returned to us in prevdirrem) until the zeroed
 8713          * directory entry gets to disk. The new inode has never been
 8714          * referenced on the disk, so can be deleted immediately.
 8715          */
 8716         if ((dirrem->dm_state & COMPLETE) == 0) {
 8717                 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
 8718                     dm_next);
 8719                 FREE_LOCK(&lk);
 8720         } else {
 8721                 if (prevdirrem != NULL)
 8722                         LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
 8723                             prevdirrem, dm_next);
 8724                 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
 8725                 direct = LIST_EMPTY(&dirrem->dm_jremrefhd);
 8726                 FREE_LOCK(&lk);
 8727                 if (direct)
 8728                         handle_workitem_remove(dirrem, 0);
 8729         }
 8730 }
 8731 
 8732 /*
 8733  * Check for an entry matching 'offset' on both the pd_dirraddhd list and the
 8734  * pd_pendinghd list of a pagedep.
 8735  */
 8736 static struct diradd *
 8737 diradd_lookup(pagedep, offset)
 8738         struct pagedep *pagedep;
 8739         int offset;
 8740 {
 8741         struct diradd *dap;
 8742 
 8743         LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
 8744                 if (dap->da_offset == offset)
 8745                         return (dap);
 8746         LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
 8747                 if (dap->da_offset == offset)
 8748                         return (dap);
 8749         return (NULL);
 8750 }
 8751 
 8752 /*
 8753  * Search for a .. diradd dependency in a directory that is being removed.
 8754  * If the directory was renamed to a new parent we have a diradd rather
 8755  * than a mkdir for the .. entry.  We need to cancel it now before
 8756  * it is found in truncate().
 8757  */
 8758 static struct jremref *
 8759 cancel_diradd_dotdot(ip, dirrem, jremref)
 8760         struct inode *ip;
 8761         struct dirrem *dirrem;
 8762         struct jremref *jremref;
 8763 {
 8764         struct pagedep *pagedep;
 8765         struct diradd *dap;
 8766         struct worklist *wk;
 8767 
 8768         if (pagedep_lookup(UFSTOVFS(ip->i_ump), NULL, ip->i_number, 0, 0,
 8769             &pagedep) == 0)
 8770                 return (jremref);
 8771         dap = diradd_lookup(pagedep, DOTDOT_OFFSET);
 8772         if (dap == NULL)
 8773                 return (jremref);
 8774         cancel_diradd(dap, dirrem, jremref, NULL, NULL);
 8775         /*
 8776          * Mark any journal work as belonging to the parent so it is freed
 8777          * with the .. reference.
 8778          */
 8779         LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list)
 8780                 wk->wk_state |= MKDIR_PARENT;
 8781         return (NULL);
 8782 }
 8783 
 8784 /*
 8785  * Cancel the MKDIR_PARENT mkdir component of a diradd when we're going to
 8786  * replace it with a dirrem/diradd pair as a result of re-parenting a
 8787  * directory.  This ensures that we don't simultaneously have a mkdir and
 8788  * a diradd for the same .. entry.
 8789  */
 8790 static struct jremref *
 8791 cancel_mkdir_dotdot(ip, dirrem, jremref)
 8792         struct inode *ip;
 8793         struct dirrem *dirrem;
 8794         struct jremref *jremref;
 8795 {
 8796         struct inodedep *inodedep;
 8797         struct jaddref *jaddref;
 8798         struct mkdir *mkdir;
 8799         struct diradd *dap;
 8800 
 8801         if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0,
 8802             &inodedep) == 0)
 8803                 return (jremref);
 8804         dap = inodedep->id_mkdiradd;
 8805         if (dap == NULL || (dap->da_state & MKDIR_PARENT) == 0)
 8806                 return (jremref);
 8807         for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir;
 8808             mkdir = LIST_NEXT(mkdir, md_mkdirs))
 8809                 if (mkdir->md_diradd == dap && mkdir->md_state & MKDIR_PARENT)
 8810                         break;
 8811         if (mkdir == NULL)
 8812                 panic("cancel_mkdir_dotdot: Unable to find mkdir\n");
 8813         if ((jaddref = mkdir->md_jaddref) != NULL) {
 8814                 mkdir->md_jaddref = NULL;
 8815                 jaddref->ja_state &= ~MKDIR_PARENT;
 8816                 if (inodedep_lookup(UFSTOVFS(ip->i_ump), jaddref->ja_ino, 0,
 8817                     &inodedep) == 0)
 8818                         panic("cancel_mkdir_dotdot: Lost parent inodedep");
 8819                 if (cancel_jaddref(jaddref, inodedep, &dirrem->dm_jwork)) {
 8820                         journal_jremref(dirrem, jremref, inodedep);
 8821                         jremref = NULL;
 8822                 }
 8823         }
 8824         if (mkdir->md_state & ONWORKLIST)
 8825                 WORKLIST_REMOVE(&mkdir->md_list);
 8826         mkdir->md_state |= ALLCOMPLETE;
 8827         complete_mkdir(mkdir);
 8828         return (jremref);
 8829 }
 8830 
 8831 static void
 8832 journal_jremref(dirrem, jremref, inodedep)
 8833         struct dirrem *dirrem;
 8834         struct jremref *jremref;
 8835         struct inodedep *inodedep;
 8836 {
 8837 
 8838         if (inodedep == NULL)
 8839                 if (inodedep_lookup(jremref->jr_list.wk_mp,
 8840                     jremref->jr_ref.if_ino, 0, &inodedep) == 0)
 8841                         panic("journal_jremref: Lost inodedep");
 8842         LIST_INSERT_HEAD(&dirrem->dm_jremrefhd, jremref, jr_deps);
 8843         TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps);
 8844         add_to_journal(&jremref->jr_list);
 8845 }
 8846 
 8847 static void
 8848 dirrem_journal(dirrem, jremref, dotremref, dotdotremref)
 8849         struct dirrem *dirrem;
 8850         struct jremref *jremref;
 8851         struct jremref *dotremref;
 8852         struct jremref *dotdotremref;
 8853 {
 8854         struct inodedep *inodedep;
 8855 
 8856 
 8857         if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 0,
 8858             &inodedep) == 0)
 8859                 panic("dirrem_journal: Lost inodedep");
 8860         journal_jremref(dirrem, jremref, inodedep);
 8861         if (dotremref)
 8862                 journal_jremref(dirrem, dotremref, inodedep);
 8863         if (dotdotremref)
 8864                 journal_jremref(dirrem, dotdotremref, NULL);
 8865 }
 8866 
 8867 /*
 8868  * Allocate a new dirrem if appropriate and return it along with
 8869  * its associated pagedep. Called without a lock, returns with lock.
 8870  */
 8871 static struct dirrem *
 8872 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
 8873         struct buf *bp;         /* buffer containing directory block */
 8874         struct inode *dp;       /* inode for the directory being modified */
 8875         struct inode *ip;       /* inode for directory entry being removed */
 8876         int isrmdir;            /* indicates if doing RMDIR */
 8877         struct dirrem **prevdirremp; /* previously referenced inode, if any */
 8878 {
 8879         int offset;
 8880         ufs_lbn_t lbn;
 8881         struct diradd *dap;
 8882         struct dirrem *dirrem;
 8883         struct pagedep *pagedep;
 8884         struct jremref *jremref;
 8885         struct jremref *dotremref;
 8886         struct jremref *dotdotremref;
 8887         struct vnode *dvp;
 8888 
 8889         /*
 8890          * Whiteouts have no deletion dependencies.
 8891          */
 8892         if (ip == NULL)
 8893                 panic("newdirrem: whiteout");
 8894         dvp = ITOV(dp);
 8895         /*
 8896          * If we are over our limit, try to improve the situation.
 8897          * Limiting the number of dirrem structures will also limit
 8898          * the number of freefile and freeblks structures.
 8899          */
 8900         ACQUIRE_LOCK(&lk);
 8901         if (!IS_SNAPSHOT(ip) && dep_current[D_DIRREM] > max_softdeps / 2)
 8902                 (void) request_cleanup(ITOV(dp)->v_mount, FLUSH_BLOCKS);
 8903         FREE_LOCK(&lk);
 8904         dirrem = malloc(sizeof(struct dirrem),
 8905                 M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO);
 8906         workitem_alloc(&dirrem->dm_list, D_DIRREM, dvp->v_mount);
 8907         LIST_INIT(&dirrem->dm_jremrefhd);
 8908         LIST_INIT(&dirrem->dm_jwork);
 8909         dirrem->dm_state = isrmdir ? RMDIR : 0;
 8910         dirrem->dm_oldinum = ip->i_number;
 8911         *prevdirremp = NULL;
 8912         /*
 8913          * Allocate remove reference structures to track journal write
 8914          * dependencies.  We will always have one for the link and
 8915          * when doing directories we will always have one more for dot.
 8916          * When renaming a directory we skip the dotdot link change so
 8917          * this is not needed.
 8918          */
 8919         jremref = dotremref = dotdotremref = NULL;
 8920         if (DOINGSUJ(dvp)) {
 8921                 if (isrmdir) {
 8922                         jremref = newjremref(dirrem, dp, ip, dp->i_offset,
 8923                             ip->i_effnlink + 2);
 8924                         dotremref = newjremref(dirrem, ip, ip, DOT_OFFSET,
 8925                             ip->i_effnlink + 1);
 8926                         dotdotremref = newjremref(dirrem, ip, dp, DOTDOT_OFFSET,
 8927                             dp->i_effnlink + 1);
 8928                         dotdotremref->jr_state |= MKDIR_PARENT;
 8929                 } else
 8930                         jremref = newjremref(dirrem, dp, ip, dp->i_offset,
 8931                             ip->i_effnlink + 1);
 8932         }
 8933         ACQUIRE_LOCK(&lk);
 8934         lbn = lblkno(dp->i_fs, dp->i_offset);
 8935         offset = blkoff(dp->i_fs, dp->i_offset);
 8936         pagedep_lookup(UFSTOVFS(dp->i_ump), bp, dp->i_number, lbn, DEPALLOC,
 8937             &pagedep);
 8938         dirrem->dm_pagedep = pagedep;
 8939         dirrem->dm_offset = offset;
 8940         /*
 8941          * If we're renaming a .. link to a new directory, cancel any
 8942          * existing MKDIR_PARENT mkdir.  If it has already been canceled
 8943          * the jremref is preserved for any potential diradd in this
 8944          * location.  This can not coincide with a rmdir.
 8945          */
 8946         if (dp->i_offset == DOTDOT_OFFSET) {
 8947                 if (isrmdir)
 8948                         panic("newdirrem: .. directory change during remove?");
 8949                 jremref = cancel_mkdir_dotdot(dp, dirrem, jremref);
 8950         }
 8951         /*
 8952          * If we're removing a directory search for the .. dependency now and
 8953          * cancel it.  Any pending journal work will be added to the dirrem
 8954          * to be completed when the workitem remove completes.
 8955          */
 8956         if (isrmdir)
 8957                 dotdotremref = cancel_diradd_dotdot(ip, dirrem, dotdotremref);
 8958         /*
 8959          * Check for a diradd dependency for the same directory entry.
 8960          * If present, then both dependencies become obsolete and can
 8961          * be de-allocated.
 8962          */
 8963         dap = diradd_lookup(pagedep, offset);
 8964         if (dap == NULL) {
 8965                 /*
 8966                  * Link the jremref structures into the dirrem so they are
 8967                  * written prior to the pagedep.
 8968                  */
 8969                 if (jremref)
 8970                         dirrem_journal(dirrem, jremref, dotremref,
 8971                             dotdotremref);
 8972                 return (dirrem);
 8973         }
 8974         /*
 8975          * Must be ATTACHED at this point.
 8976          */
 8977         if ((dap->da_state & ATTACHED) == 0)
 8978                 panic("newdirrem: not ATTACHED");
 8979         if (dap->da_newinum != ip->i_number)
 8980                 panic("newdirrem: inum %ju should be %ju",
 8981                     (uintmax_t)ip->i_number, (uintmax_t)dap->da_newinum);
 8982         /*
 8983          * If we are deleting a changed name that never made it to disk,
 8984          * then return the dirrem describing the previous inode (which
 8985          * represents the inode currently referenced from this entry on disk).
 8986          */
 8987         if ((dap->da_state & DIRCHG) != 0) {
 8988                 *prevdirremp = dap->da_previous;
 8989                 dap->da_state &= ~DIRCHG;
 8990                 dap->da_pagedep = pagedep;
 8991         }
 8992         /*
 8993          * We are deleting an entry that never made it to disk.
 8994          * Mark it COMPLETE so we can delete its inode immediately.
 8995          */
 8996         dirrem->dm_state |= COMPLETE;
 8997         cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref);
 8998 #ifdef SUJ_DEBUG
 8999         if (isrmdir == 0) {
 9000                 struct worklist *wk;
 9001 
 9002                 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list)
 9003                         if (wk->wk_state & (MKDIR_BODY | MKDIR_PARENT))
 9004                                 panic("bad wk %p (0x%X)\n", wk, wk->wk_state);
 9005         }
 9006 #endif
 9007 
 9008         return (dirrem);
 9009 }
 9010 
 9011 /*
 9012  * Directory entry change dependencies.
 9013  * 
 9014  * Changing an existing directory entry requires that an add operation
 9015  * be completed first followed by a deletion. The semantics for the addition
 9016  * are identical to the description of adding a new entry above except
 9017  * that the rollback is to the old inode number rather than zero. Once
 9018  * the addition dependency is completed, the removal is done as described
 9019  * in the removal routine above.
 9020  */
 9021 
 9022 /*
 9023  * This routine should be called immediately after changing
 9024  * a directory entry.  The inode's link count should not be
 9025  * decremented by the calling procedure -- the soft updates
 9026  * code will perform this task when it is safe.
 9027  */
 9028 void 
 9029 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
 9030         struct buf *bp;         /* buffer containing directory block */
 9031         struct inode *dp;       /* inode for the directory being modified */
 9032         struct inode *ip;       /* inode for directory entry being removed */
 9033         ino_t newinum;          /* new inode number for changed entry */
 9034         int isrmdir;            /* indicates if doing RMDIR */
 9035 {
 9036         int offset;
 9037         struct diradd *dap = NULL;
 9038         struct dirrem *dirrem, *prevdirrem;
 9039         struct pagedep *pagedep;
 9040         struct inodedep *inodedep;
 9041         struct jaddref *jaddref;
 9042         struct mount *mp;
 9043 
 9044         offset = blkoff(dp->i_fs, dp->i_offset);
 9045         mp = UFSTOVFS(dp->i_ump);
 9046 
 9047         /*
 9048          * Whiteouts do not need diradd dependencies.
 9049          */
 9050         if (newinum != WINO) {
 9051                 dap = malloc(sizeof(struct diradd),
 9052                     M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO);
 9053                 workitem_alloc(&dap->da_list, D_DIRADD, mp);
 9054                 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
 9055                 dap->da_offset = offset;
 9056                 dap->da_newinum = newinum;
 9057                 LIST_INIT(&dap->da_jwork);
 9058         }
 9059 
 9060         /*
 9061          * Allocate a new dirrem and ACQUIRE_LOCK.
 9062          */
 9063         dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
 9064         pagedep = dirrem->dm_pagedep;
 9065         /*
 9066          * The possible values for isrmdir:
 9067          *      0 - non-directory file rename
 9068          *      1 - directory rename within same directory
 9069          *   inum - directory rename to new directory of given inode number
 9070          * When renaming to a new directory, we are both deleting and
 9071          * creating a new directory entry, so the link count on the new
 9072          * directory should not change. Thus we do not need the followup
 9073          * dirrem which is usually done in handle_workitem_remove. We set
 9074          * the DIRCHG flag to tell handle_workitem_remove to skip the 
 9075          * followup dirrem.
 9076          */
 9077         if (isrmdir > 1)
 9078                 dirrem->dm_state |= DIRCHG;
 9079 
 9080         /*
 9081          * Whiteouts have no additional dependencies,
 9082          * so just put the dirrem on the correct list.
 9083          */
 9084         if (newinum == WINO) {
 9085                 if ((dirrem->dm_state & COMPLETE) == 0) {
 9086                         LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
 9087                             dm_next);
 9088                 } else {
 9089                         dirrem->dm_dirinum = pagedep->pd_ino;
 9090                         if (LIST_EMPTY(&dirrem->dm_jremrefhd))
 9091                                 add_to_worklist(&dirrem->dm_list, 0);
 9092                 }
 9093                 FREE_LOCK(&lk);
 9094                 return;
 9095         }
 9096         /*
 9097          * Add the dirrem to the inodedep's pending remove list for quick
 9098          * discovery later.  A valid nlinkdelta ensures that this lookup
 9099          * will not fail.
 9100          */
 9101         if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0)
 9102                 panic("softdep_setup_directory_change: Lost inodedep.");
 9103         dirrem->dm_state |= ONDEPLIST;
 9104         LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext);
 9105 
 9106         /*
 9107          * If the COMPLETE flag is clear, then there were no active
 9108          * entries and we want to roll back to the previous inode until
 9109          * the new inode is committed to disk. If the COMPLETE flag is
 9110          * set, then we have deleted an entry that never made it to disk.
 9111          * If the entry we deleted resulted from a name change, then the old
 9112          * inode reference still resides on disk. Any rollback that we do
 9113          * needs to be to that old inode (returned to us in prevdirrem). If
 9114          * the entry we deleted resulted from a create, then there is
 9115          * no entry on the disk, so we want to roll back to zero rather
 9116          * than the uncommitted inode. In either of the COMPLETE cases we
 9117          * want to immediately free the unwritten and unreferenced inode.
 9118          */
 9119         if ((dirrem->dm_state & COMPLETE) == 0) {
 9120                 dap->da_previous = dirrem;
 9121         } else {
 9122                 if (prevdirrem != NULL) {
 9123                         dap->da_previous = prevdirrem;
 9124                 } else {
 9125                         dap->da_state &= ~DIRCHG;
 9126                         dap->da_pagedep = pagedep;
 9127                 }
 9128                 dirrem->dm_dirinum = pagedep->pd_ino;
 9129                 if (LIST_EMPTY(&dirrem->dm_jremrefhd))
 9130                         add_to_worklist(&dirrem->dm_list, 0);
 9131         }
 9132         /*
 9133          * Lookup the jaddref for this journal entry.  We must finish
 9134          * initializing it and make the diradd write dependent on it.
 9135          * If we're not journaling, put it on the id_bufwait list if the
 9136          * inode is not yet written. If it is written, do the post-inode
 9137          * write processing to put it on the id_pendinghd list.
 9138          */
 9139         inodedep_lookup(mp, newinum, DEPALLOC | NODELAY, &inodedep);
 9140         if (MOUNTEDSUJ(mp)) {
 9141                 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
 9142                     inoreflst);
 9143                 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number,
 9144                     ("softdep_setup_directory_change: bad jaddref %p",
 9145                     jaddref));
 9146                 jaddref->ja_diroff = dp->i_offset;
 9147                 jaddref->ja_diradd = dap;
 9148                 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
 9149                     dap, da_pdlist);
 9150                 add_to_journal(&jaddref->ja_list);
 9151         } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
 9152                 dap->da_state |= COMPLETE;
 9153                 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
 9154                 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
 9155         } else {
 9156                 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
 9157                     dap, da_pdlist);
 9158                 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
 9159         }
 9160         /*
 9161          * If we're making a new name for a directory that has not been
 9162          * committed when need to move the dot and dotdot references to
 9163          * this new name.
 9164          */
 9165         if (inodedep->id_mkdiradd && dp->i_offset != DOTDOT_OFFSET)
 9166                 merge_diradd(inodedep, dap);
 9167         FREE_LOCK(&lk);
 9168 }
 9169 
 9170 /*
 9171  * Called whenever the link count on an inode is changed.
 9172  * It creates an inode dependency so that the new reference(s)
 9173  * to the inode cannot be committed to disk until the updated
 9174  * inode has been written.
 9175  */
 9176 void
 9177 softdep_change_linkcnt(ip)
 9178         struct inode *ip;       /* the inode with the increased link count */
 9179 {
 9180         struct inodedep *inodedep;
 9181         int dflags;
 9182 
 9183         ACQUIRE_LOCK(&lk);
 9184         dflags = DEPALLOC;
 9185         if (IS_SNAPSHOT(ip))
 9186                 dflags |= NODELAY;
 9187         inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, dflags, &inodedep);
 9188         if (ip->i_nlink < ip->i_effnlink)
 9189                 panic("softdep_change_linkcnt: bad delta");
 9190         inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
 9191         FREE_LOCK(&lk);
 9192 }
 9193 
 9194 /*
 9195  * Attach a sbdep dependency to the superblock buf so that we can keep
 9196  * track of the head of the linked list of referenced but unlinked inodes.
 9197  */
 9198 void
 9199 softdep_setup_sbupdate(ump, fs, bp)
 9200         struct ufsmount *ump;
 9201         struct fs *fs;
 9202         struct buf *bp;
 9203 {
 9204         struct sbdep *sbdep;
 9205         struct worklist *wk;
 9206 
 9207         if (MOUNTEDSUJ(UFSTOVFS(ump)) == 0)
 9208                 return;
 9209         LIST_FOREACH(wk, &bp->b_dep, wk_list)
 9210                 if (wk->wk_type == D_SBDEP)
 9211                         break;
 9212         if (wk != NULL)
 9213                 return;
 9214         sbdep = malloc(sizeof(struct sbdep), M_SBDEP, M_SOFTDEP_FLAGS);
 9215         workitem_alloc(&sbdep->sb_list, D_SBDEP, UFSTOVFS(ump));
 9216         sbdep->sb_fs = fs;
 9217         sbdep->sb_ump = ump;
 9218         ACQUIRE_LOCK(&lk);
 9219         WORKLIST_INSERT(&bp->b_dep, &sbdep->sb_list);
 9220         FREE_LOCK(&lk);
 9221 }
 9222 
 9223 /*
 9224  * Return the first unlinked inodedep which is ready to be the head of the
 9225  * list.  The inodedep and all those after it must have valid next pointers.
 9226  */
 9227 static struct inodedep *
 9228 first_unlinked_inodedep(ump)
 9229         struct ufsmount *ump;
 9230 {
 9231         struct inodedep *inodedep;
 9232         struct inodedep *idp;
 9233 
 9234         rw_assert(&lk, RA_WLOCKED);
 9235         for (inodedep = TAILQ_LAST(&ump->softdep_unlinked, inodedeplst);
 9236             inodedep; inodedep = idp) {
 9237                 if ((inodedep->id_state & UNLINKNEXT) == 0)
 9238                         return (NULL);
 9239                 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked);
 9240                 if (idp == NULL || (idp->id_state & UNLINKNEXT) == 0)
 9241                         break;
 9242                 if ((inodedep->id_state & UNLINKPREV) == 0)
 9243                         break;
 9244         }
 9245         return (inodedep);
 9246 }
 9247 
 9248 /*
 9249  * Set the sujfree unlinked head pointer prior to writing a superblock.
 9250  */
 9251 static void
 9252 initiate_write_sbdep(sbdep)
 9253         struct sbdep *sbdep;
 9254 {
 9255         struct inodedep *inodedep;
 9256         struct fs *bpfs;
 9257         struct fs *fs;
 9258 
 9259         bpfs = sbdep->sb_fs;
 9260         fs = sbdep->sb_ump->um_fs;
 9261         inodedep = first_unlinked_inodedep(sbdep->sb_ump);
 9262         if (inodedep) {
 9263                 fs->fs_sujfree = inodedep->id_ino;
 9264                 inodedep->id_state |= UNLINKPREV;
 9265         } else
 9266                 fs->fs_sujfree = 0;
 9267         bpfs->fs_sujfree = fs->fs_sujfree;
 9268 }
 9269 
 9270 /*
 9271  * After a superblock is written determine whether it must be written again
 9272  * due to a changing unlinked list head.
 9273  */
 9274 static int
 9275 handle_written_sbdep(sbdep, bp)
 9276         struct sbdep *sbdep;
 9277         struct buf *bp;
 9278 {
 9279         struct inodedep *inodedep;
 9280         struct mount *mp;
 9281         struct fs *fs;
 9282 
 9283         rw_assert(&lk, RA_WLOCKED);
 9284         fs = sbdep->sb_fs;
 9285         mp = UFSTOVFS(sbdep->sb_ump);
 9286         /*
 9287          * If the superblock doesn't match the in-memory list start over.
 9288          */
 9289         inodedep = first_unlinked_inodedep(sbdep->sb_ump);
 9290         if ((inodedep && fs->fs_sujfree != inodedep->id_ino) ||
 9291             (inodedep == NULL && fs->fs_sujfree != 0)) {
 9292                 bdirty(bp);
 9293                 return (1);
 9294         }
 9295         WORKITEM_FREE(sbdep, D_SBDEP);
 9296         if (fs->fs_sujfree == 0)
 9297                 return (0);
 9298         /*
 9299          * Now that we have a record of this inode in stable store allow it
 9300          * to be written to free up pending work.  Inodes may see a lot of
 9301          * write activity after they are unlinked which we must not hold up.
 9302          */
 9303         for (; inodedep != NULL; inodedep = TAILQ_NEXT(inodedep, id_unlinked)) {
 9304                 if ((inodedep->id_state & UNLINKLINKS) != UNLINKLINKS)
 9305                         panic("handle_written_sbdep: Bad inodedep %p (0x%X)",
 9306                             inodedep, inodedep->id_state);
 9307                 if (inodedep->id_state & UNLINKONLIST)
 9308                         break;
 9309                 inodedep->id_state |= DEPCOMPLETE | UNLINKONLIST;
 9310         }
 9311 
 9312         return (0);
 9313 }
 9314 
 9315 /*
 9316  * Mark an inodedep as unlinked and insert it into the in-memory unlinked list.
 9317  */
 9318 static void
 9319 unlinked_inodedep(mp, inodedep)
 9320         struct mount *mp;
 9321         struct inodedep *inodedep;
 9322 {
 9323         struct ufsmount *ump;
 9324 
 9325         rw_assert(&lk, RA_WLOCKED);
 9326         if (MOUNTEDSUJ(mp) == 0)
 9327                 return;
 9328         ump = VFSTOUFS(mp);
 9329         ump->um_fs->fs_fmod = 1;
 9330         if (inodedep->id_state & UNLINKED)
 9331                 panic("unlinked_inodedep: %p already unlinked\n", inodedep);
 9332         inodedep->id_state |= UNLINKED;
 9333         TAILQ_INSERT_HEAD(&ump->softdep_unlinked, inodedep, id_unlinked);
 9334 }
 9335 
 9336 /*
 9337  * Remove an inodedep from the unlinked inodedep list.  This may require
 9338  * disk writes if the inode has made it that far.
 9339  */
 9340 static void
 9341 clear_unlinked_inodedep(inodedep)
 9342         struct inodedep *inodedep;
 9343 {
 9344         struct ufsmount *ump;
 9345         struct inodedep *idp;
 9346         struct inodedep *idn;
 9347         struct fs *fs;
 9348         struct buf *bp;
 9349         ino_t ino;
 9350         ino_t nino;
 9351         ino_t pino;
 9352         int error;
 9353 
 9354         ump = VFSTOUFS(inodedep->id_list.wk_mp);
 9355         fs = ump->um_fs;
 9356         ino = inodedep->id_ino;
 9357         error = 0;
 9358         for (;;) {
 9359                 rw_assert(&lk, RA_WLOCKED);
 9360                 KASSERT((inodedep->id_state & UNLINKED) != 0,
 9361                     ("clear_unlinked_inodedep: inodedep %p not unlinked",
 9362                     inodedep));
 9363                 /*
 9364                  * If nothing has yet been written simply remove us from
 9365                  * the in memory list and return.  This is the most common
 9366                  * case where handle_workitem_remove() loses the final
 9367                  * reference.
 9368                  */
 9369                 if ((inodedep->id_state & UNLINKLINKS) == 0)
 9370                         break;
 9371                 /*
 9372                  * If we have a NEXT pointer and no PREV pointer we can simply
 9373                  * clear NEXT's PREV and remove ourselves from the list.  Be
 9374                  * careful not to clear PREV if the superblock points at
 9375                  * next as well.
 9376                  */
 9377                 idn = TAILQ_NEXT(inodedep, id_unlinked);
 9378                 if ((inodedep->id_state & UNLINKLINKS) == UNLINKNEXT) {
 9379                         if (idn && fs->fs_sujfree != idn->id_ino)
 9380                                 idn->id_state &= ~UNLINKPREV;
 9381                         break;
 9382                 }
 9383                 /*
 9384                  * Here we have an inodedep which is actually linked into
 9385                  * the list.  We must remove it by forcing a write to the
 9386                  * link before us, whether it be the superblock or an inode.
 9387                  * Unfortunately the list may change while we're waiting
 9388                  * on the buf lock for either resource so we must loop until
 9389                  * we lock the right one.  If both the superblock and an
 9390                  * inode point to this inode we must clear the inode first
 9391                  * followed by the superblock.
 9392                  */
 9393                 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked);
 9394                 pino = 0;
 9395                 if (idp && (idp->id_state & UNLINKNEXT))
 9396                         pino = idp->id_ino;
 9397                 FREE_LOCK(&lk);
 9398                 if (pino == 0) {
 9399                         bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc),
 9400                             (int)fs->fs_sbsize, 0, 0, 0);
 9401                 } else {
 9402                         error = bread(ump->um_devvp,
 9403                             fsbtodb(fs, ino_to_fsba(fs, pino)),
 9404                             (int)fs->fs_bsize, NOCRED, &bp);
 9405                         if (error)
 9406                                 brelse(bp);
 9407                 }
 9408                 ACQUIRE_LOCK(&lk);
 9409                 if (error)
 9410                         break;
 9411                 /* If the list has changed restart the loop. */
 9412                 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked);
 9413                 nino = 0;
 9414                 if (idp && (idp->id_state & UNLINKNEXT))
 9415                         nino = idp->id_ino;
 9416                 if (nino != pino ||
 9417                     (inodedep->id_state & UNLINKPREV) != UNLINKPREV) {
 9418                         FREE_LOCK(&lk);
 9419                         brelse(bp);
 9420                         ACQUIRE_LOCK(&lk);
 9421                         continue;
 9422                 }
 9423                 nino = 0;
 9424                 idn = TAILQ_NEXT(inodedep, id_unlinked);
 9425                 if (idn)
 9426                         nino = idn->id_ino;
 9427                 /*
 9428                  * Remove us from the in memory list.  After this we cannot
 9429                  * access the inodedep.
 9430                  */
 9431                 KASSERT((inodedep->id_state & UNLINKED) != 0,
 9432                     ("clear_unlinked_inodedep: inodedep %p not unlinked",
 9433                     inodedep));
 9434                 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST);
 9435                 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked);
 9436                 FREE_LOCK(&lk);
 9437                 /*
 9438                  * The predecessor's next pointer is manually updated here
 9439                  * so that the NEXT flag is never cleared for an element
 9440                  * that is in the list.
 9441                  */
 9442                 if (pino == 0) {
 9443                         bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize);
 9444                         ffs_oldfscompat_write((struct fs *)bp->b_data, ump);
 9445                         softdep_setup_sbupdate(ump, (struct fs *)bp->b_data,
 9446                             bp);
 9447                 } else if (fs->fs_magic == FS_UFS1_MAGIC)
 9448                         ((struct ufs1_dinode *)bp->b_data +
 9449                             ino_to_fsbo(fs, pino))->di_freelink = nino;
 9450                 else
 9451                         ((struct ufs2_dinode *)bp->b_data +
 9452                             ino_to_fsbo(fs, pino))->di_freelink = nino;
 9453                 /*
 9454                  * If the bwrite fails we have no recourse to recover.  The
 9455                  * filesystem is corrupted already.
 9456                  */
 9457                 bwrite(bp);
 9458                 ACQUIRE_LOCK(&lk);
 9459                 /*
 9460                  * If the superblock pointer still needs to be cleared force
 9461                  * a write here.
 9462                  */
 9463                 if (fs->fs_sujfree == ino) {
 9464                         FREE_LOCK(&lk);
 9465                         bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc),
 9466                             (int)fs->fs_sbsize, 0, 0, 0);
 9467                         bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize);
 9468                         ffs_oldfscompat_write((struct fs *)bp->b_data, ump);
 9469                         softdep_setup_sbupdate(ump, (struct fs *)bp->b_data,
 9470                             bp);
 9471                         bwrite(bp);
 9472                         ACQUIRE_LOCK(&lk);
 9473                 }
 9474 
 9475                 if (fs->fs_sujfree != ino)
 9476                         return;
 9477                 panic("clear_unlinked_inodedep: Failed to clear free head");
 9478         }
 9479         if (inodedep->id_ino == fs->fs_sujfree)
 9480                 panic("clear_unlinked_inodedep: Freeing head of free list");
 9481         inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST);
 9482         TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked);
 9483         return;
 9484 }
 9485 
 9486 /*
 9487  * This workitem decrements the inode's link count.
 9488  * If the link count reaches zero, the file is removed.
 9489  */
 9490 static int
 9491 handle_workitem_remove(dirrem, flags)
 9492         struct dirrem *dirrem;
 9493         int flags;
 9494 {
 9495         struct inodedep *inodedep;
 9496         struct workhead dotdotwk;
 9497         struct worklist *wk;
 9498         struct ufsmount *ump;
 9499         struct mount *mp;
 9500         struct vnode *vp;
 9501         struct inode *ip;
 9502         ino_t oldinum;
 9503 
 9504         if (dirrem->dm_state & ONWORKLIST)
 9505                 panic("handle_workitem_remove: dirrem %p still on worklist",
 9506                     dirrem);
 9507         oldinum = dirrem->dm_oldinum;
 9508         mp = dirrem->dm_list.wk_mp;
 9509         ump = VFSTOUFS(mp);
 9510         flags |= LK_EXCLUSIVE;
 9511         if (ffs_vgetf(mp, oldinum, flags, &vp, FFSV_FORCEINSMQ) != 0)
 9512                 return (EBUSY);
 9513         ip = VTOI(vp);
 9514         ACQUIRE_LOCK(&lk);
 9515         if ((inodedep_lookup(mp, oldinum, 0, &inodedep)) == 0)
 9516                 panic("handle_workitem_remove: lost inodedep");
 9517         if (dirrem->dm_state & ONDEPLIST)
 9518                 LIST_REMOVE(dirrem, dm_inonext);
 9519         KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd),
 9520             ("handle_workitem_remove:  Journal entries not written."));
 9521 
 9522         /*
 9523          * Move all dependencies waiting on the remove to complete
 9524          * from the dirrem to the inode inowait list to be completed
 9525          * after the inode has been updated and written to disk.  Any
 9526          * marked MKDIR_PARENT are saved to be completed when the .. ref
 9527          * is removed.
 9528          */
 9529         LIST_INIT(&dotdotwk);
 9530         while ((wk = LIST_FIRST(&dirrem->dm_jwork)) != NULL) {
 9531                 WORKLIST_REMOVE(wk);
 9532                 if (wk->wk_state & MKDIR_PARENT) {
 9533                         wk->wk_state &= ~MKDIR_PARENT;
 9534                         WORKLIST_INSERT(&dotdotwk, wk);
 9535                         continue;
 9536                 }
 9537                 WORKLIST_INSERT(&inodedep->id_inowait, wk);
 9538         }
 9539         LIST_SWAP(&dirrem->dm_jwork, &dotdotwk, worklist, wk_list);
 9540         /*
 9541          * Normal file deletion.
 9542          */
 9543         if ((dirrem->dm_state & RMDIR) == 0) {
 9544                 ip->i_nlink--;
 9545                 DIP_SET(ip, i_nlink, ip->i_nlink);
 9546                 ip->i_flag |= IN_CHANGE;
 9547                 if (ip->i_nlink < ip->i_effnlink)
 9548                         panic("handle_workitem_remove: bad file delta");
 9549                 if (ip->i_nlink == 0) 
 9550                         unlinked_inodedep(mp, inodedep);
 9551                 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
 9552                 KASSERT(LIST_EMPTY(&dirrem->dm_jwork),
 9553                     ("handle_workitem_remove: worklist not empty. %s",
 9554                     TYPENAME(LIST_FIRST(&dirrem->dm_jwork)->wk_type)));
 9555                 WORKITEM_FREE(dirrem, D_DIRREM);
 9556                 FREE_LOCK(&lk);
 9557                 goto out;
 9558         }
 9559         /*
 9560          * Directory deletion. Decrement reference count for both the
 9561          * just deleted parent directory entry and the reference for ".".
 9562          * Arrange to have the reference count on the parent decremented
 9563          * to account for the loss of "..".
 9564          */
 9565         ip->i_nlink -= 2;
 9566         DIP_SET(ip, i_nlink, ip->i_nlink);
 9567         ip->i_flag |= IN_CHANGE;
 9568         if (ip->i_nlink < ip->i_effnlink)
 9569                 panic("handle_workitem_remove: bad dir delta");
 9570         if (ip->i_nlink == 0)
 9571                 unlinked_inodedep(mp, inodedep);
 9572         inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
 9573         /*
 9574          * Rename a directory to a new parent. Since, we are both deleting
 9575          * and creating a new directory entry, the link count on the new
 9576          * directory should not change. Thus we skip the followup dirrem.
 9577          */
 9578         if (dirrem->dm_state & DIRCHG) {
 9579                 KASSERT(LIST_EMPTY(&dirrem->dm_jwork),
 9580                     ("handle_workitem_remove: DIRCHG and worklist not empty."));
 9581                 WORKITEM_FREE(dirrem, D_DIRREM);
 9582                 FREE_LOCK(&lk);
 9583                 goto out;
 9584         }
 9585         dirrem->dm_state = ONDEPLIST;
 9586         dirrem->dm_oldinum = dirrem->dm_dirinum;
 9587         /*
 9588          * Place the dirrem on the parent's diremhd list.
 9589          */
 9590         if (inodedep_lookup(mp, dirrem->dm_oldinum, 0, &inodedep) == 0)
 9591                 panic("handle_workitem_remove: lost dir inodedep");
 9592         LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext);
 9593         /*
 9594          * If the allocated inode has never been written to disk, then
 9595          * the on-disk inode is zero'ed and we can remove the file
 9596          * immediately.  When journaling if the inode has been marked
 9597          * unlinked and not DEPCOMPLETE we know it can never be written.
 9598          */
 9599         inodedep_lookup(mp, oldinum, 0, &inodedep);
 9600         if (inodedep == NULL ||
 9601             (inodedep->id_state & (DEPCOMPLETE | UNLINKED)) == UNLINKED ||
 9602             check_inode_unwritten(inodedep)) {
 9603                 FREE_LOCK(&lk);
 9604                 vput(vp);
 9605                 return handle_workitem_remove(dirrem, flags);
 9606         }
 9607         WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
 9608         FREE_LOCK(&lk);
 9609         ip->i_flag |= IN_CHANGE;
 9610 out:
 9611         ffs_update(vp, 0);
 9612         vput(vp);
 9613         return (0);
 9614 }
 9615 
 9616 /*
 9617  * Inode de-allocation dependencies.
 9618  * 
 9619  * When an inode's link count is reduced to zero, it can be de-allocated. We
 9620  * found it convenient to postpone de-allocation until after the inode is
 9621  * written to disk with its new link count (zero).  At this point, all of the
 9622  * on-disk inode's block pointers are nullified and, with careful dependency
 9623  * list ordering, all dependencies related to the inode will be satisfied and
 9624  * the corresponding dependency structures de-allocated.  So, if/when the
 9625  * inode is reused, there will be no mixing of old dependencies with new
 9626  * ones.  This artificial dependency is set up by the block de-allocation
 9627  * procedure above (softdep_setup_freeblocks) and completed by the
 9628  * following procedure.
 9629  */
 9630 static void 
 9631 handle_workitem_freefile(freefile)
 9632         struct freefile *freefile;
 9633 {
 9634         struct workhead wkhd;
 9635         struct fs *fs;
 9636         struct inodedep *idp;
 9637         struct ufsmount *ump;
 9638         int error;
 9639 
 9640         ump = VFSTOUFS(freefile->fx_list.wk_mp);
 9641         fs = ump->um_fs;
 9642 #ifdef DEBUG
 9643         ACQUIRE_LOCK(&lk);
 9644         error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp);
 9645         FREE_LOCK(&lk);
 9646         if (error)
 9647                 panic("handle_workitem_freefile: inodedep %p survived", idp);
 9648 #endif
 9649         UFS_LOCK(ump);
 9650         fs->fs_pendinginodes -= 1;
 9651         UFS_UNLOCK(ump);
 9652         LIST_INIT(&wkhd);
 9653         LIST_SWAP(&freefile->fx_jwork, &wkhd, worklist, wk_list);
 9654         if ((error = ffs_freefile(ump, fs, freefile->fx_devvp,
 9655             freefile->fx_oldinum, freefile->fx_mode, &wkhd)) != 0)
 9656                 softdep_error("handle_workitem_freefile", error);
 9657         ACQUIRE_LOCK(&lk);
 9658         WORKITEM_FREE(freefile, D_FREEFILE);
 9659         FREE_LOCK(&lk);
 9660 }
 9661 
 9662 
 9663 /*
 9664  * Helper function which unlinks marker element from work list and returns
 9665  * the next element on the list.
 9666  */
 9667 static __inline struct worklist *
 9668 markernext(struct worklist *marker)
 9669 {
 9670         struct worklist *next;
 9671         
 9672         next = LIST_NEXT(marker, wk_list);
 9673         LIST_REMOVE(marker, wk_list);
 9674         return next;
 9675 }
 9676 
 9677 /*
 9678  * Disk writes.
 9679  * 
 9680  * The dependency structures constructed above are most actively used when file
 9681  * system blocks are written to disk.  No constraints are placed on when a
 9682  * block can be written, but unsatisfied update dependencies are made safe by
 9683  * modifying (or replacing) the source memory for the duration of the disk
 9684  * write.  When the disk write completes, the memory block is again brought
 9685  * up-to-date.
 9686  *
 9687  * In-core inode structure reclamation.
 9688  * 
 9689  * Because there are a finite number of "in-core" inode structures, they are
 9690  * reused regularly.  By transferring all inode-related dependencies to the
 9691  * in-memory inode block and indexing them separately (via "inodedep"s), we
 9692  * can allow "in-core" inode structures to be reused at any time and avoid
 9693  * any increase in contention.
 9694  *
 9695  * Called just before entering the device driver to initiate a new disk I/O.
 9696  * The buffer must be locked, thus, no I/O completion operations can occur
 9697  * while we are manipulating its associated dependencies.
 9698  */
 9699 static void 
 9700 softdep_disk_io_initiation(bp)
 9701         struct buf *bp;         /* structure describing disk write to occur */
 9702 {
 9703         struct worklist *wk;
 9704         struct worklist marker;
 9705         struct inodedep *inodedep;
 9706         struct freeblks *freeblks;
 9707         struct jblkdep *jblkdep;
 9708         struct newblk *newblk;
 9709 
 9710         /*
 9711          * We only care about write operations. There should never
 9712          * be dependencies for reads.
 9713          */
 9714         if (bp->b_iocmd != BIO_WRITE)
 9715                 panic("softdep_disk_io_initiation: not write");
 9716 
 9717         if (bp->b_vflags & BV_BKGRDINPROG)
 9718                 panic("softdep_disk_io_initiation: Writing buffer with "
 9719                     "background write in progress: %p", bp);
 9720 
 9721         marker.wk_type = D_LAST + 1;    /* Not a normal workitem */
 9722         PHOLD(curproc);                 /* Don't swap out kernel stack */
 9723 
 9724         ACQUIRE_LOCK(&lk);
 9725         /*
 9726          * Do any necessary pre-I/O processing.
 9727          */
 9728         for (wk = LIST_FIRST(&bp->b_dep); wk != NULL;
 9729              wk = markernext(&marker)) {
 9730                 LIST_INSERT_AFTER(wk, &marker, wk_list);
 9731                 switch (wk->wk_type) {
 9732 
 9733                 case D_PAGEDEP:
 9734                         initiate_write_filepage(WK_PAGEDEP(wk), bp);
 9735                         continue;
 9736 
 9737                 case D_INODEDEP:
 9738                         inodedep = WK_INODEDEP(wk);
 9739                         if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC)
 9740                                 initiate_write_inodeblock_ufs1(inodedep, bp);
 9741                         else
 9742                                 initiate_write_inodeblock_ufs2(inodedep, bp);
 9743                         continue;
 9744 
 9745                 case D_INDIRDEP:
 9746                         initiate_write_indirdep(WK_INDIRDEP(wk), bp);
 9747                         continue;
 9748 
 9749                 case D_BMSAFEMAP:
 9750                         initiate_write_bmsafemap(WK_BMSAFEMAP(wk), bp);
 9751                         continue;
 9752 
 9753                 case D_JSEG:
 9754                         WK_JSEG(wk)->js_buf = NULL;
 9755                         continue;
 9756 
 9757                 case D_FREEBLKS:
 9758                         freeblks = WK_FREEBLKS(wk);
 9759                         jblkdep = LIST_FIRST(&freeblks->fb_jblkdephd);
 9760                         /*
 9761                          * We have to wait for the freeblks to be journaled
 9762                          * before we can write an inodeblock with updated
 9763                          * pointers.  Be careful to arrange the marker so
 9764                          * we revisit the freeblks if it's not removed by
 9765                          * the first jwait().
 9766                          */
 9767                         if (jblkdep != NULL) {
 9768                                 LIST_REMOVE(&marker, wk_list);
 9769                                 LIST_INSERT_BEFORE(wk, &marker, wk_list);
 9770                                 jwait(&jblkdep->jb_list, MNT_WAIT);
 9771                         }
 9772                         continue;
 9773                 case D_ALLOCDIRECT:
 9774                 case D_ALLOCINDIR:
 9775                         /*
 9776                          * We have to wait for the jnewblk to be journaled
 9777                          * before we can write to a block if the contents
 9778                          * may be confused with an earlier file's indirect
 9779                          * at recovery time.  Handle the marker as described
 9780                          * above.
 9781                          */
 9782                         newblk = WK_NEWBLK(wk);
 9783                         if (newblk->nb_jnewblk != NULL &&
 9784                             indirblk_lookup(newblk->nb_list.wk_mp,
 9785                             newblk->nb_newblkno)) {
 9786                                 LIST_REMOVE(&marker, wk_list);
 9787                                 LIST_INSERT_BEFORE(wk, &marker, wk_list);
 9788                                 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT);
 9789                         }
 9790                         continue;
 9791 
 9792                 case D_SBDEP:
 9793                         initiate_write_sbdep(WK_SBDEP(wk));
 9794                         continue;
 9795 
 9796                 case D_MKDIR:
 9797                 case D_FREEWORK:
 9798                 case D_FREEDEP:
 9799                 case D_JSEGDEP:
 9800                         continue;
 9801 
 9802                 default:
 9803                         panic("handle_disk_io_initiation: Unexpected type %s",
 9804                             TYPENAME(wk->wk_type));
 9805                         /* NOTREACHED */
 9806                 }
 9807         }
 9808         FREE_LOCK(&lk);
 9809         PRELE(curproc);                 /* Allow swapout of kernel stack */
 9810 }
 9811 
 9812 /*
 9813  * Called from within the procedure above to deal with unsatisfied
 9814  * allocation dependencies in a directory. The buffer must be locked,
 9815  * thus, no I/O completion operations can occur while we are
 9816  * manipulating its associated dependencies.
 9817  */
 9818 static void
 9819 initiate_write_filepage(pagedep, bp)
 9820         struct pagedep *pagedep;
 9821         struct buf *bp;
 9822 {
 9823         struct jremref *jremref;
 9824         struct jmvref *jmvref;
 9825         struct dirrem *dirrem;
 9826         struct diradd *dap;
 9827         struct direct *ep;
 9828         int i;
 9829 
 9830         if (pagedep->pd_state & IOSTARTED) {
 9831                 /*
 9832                  * This can only happen if there is a driver that does not
 9833                  * understand chaining. Here biodone will reissue the call
 9834                  * to strategy for the incomplete buffers.
 9835                  */
 9836                 printf("initiate_write_filepage: already started\n");
 9837                 return;
 9838         }
 9839         pagedep->pd_state |= IOSTARTED;
 9840         /*
 9841          * Wait for all journal remove dependencies to hit the disk.
 9842          * We can not allow any potentially conflicting directory adds
 9843          * to be visible before removes and rollback is too difficult.
 9844          * lk may be dropped and re-acquired, however we hold the buf
 9845          * locked so the dependency can not go away.
 9846          */
 9847         LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next)
 9848                 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL)
 9849                         jwait(&jremref->jr_list, MNT_WAIT);
 9850         while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL)
 9851                 jwait(&jmvref->jm_list, MNT_WAIT);
 9852         for (i = 0; i < DAHASHSZ; i++) {
 9853                 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
 9854                         ep = (struct direct *)
 9855                             ((char *)bp->b_data + dap->da_offset);
 9856                         if (ep->d_ino != dap->da_newinum)
 9857                                 panic("%s: dir inum %ju != new %ju",
 9858                                     "initiate_write_filepage",
 9859                                     (uintmax_t)ep->d_ino,
 9860                                     (uintmax_t)dap->da_newinum);
 9861                         if (dap->da_state & DIRCHG)
 9862                                 ep->d_ino = dap->da_previous->dm_oldinum;
 9863                         else
 9864                                 ep->d_ino = 0;
 9865                         dap->da_state &= ~ATTACHED;
 9866                         dap->da_state |= UNDONE;
 9867                 }
 9868         }
 9869 }
 9870 
 9871 /*
 9872  * Version of initiate_write_inodeblock that handles UFS1 dinodes.
 9873  * Note that any bug fixes made to this routine must be done in the
 9874  * version found below.
 9875  *
 9876  * Called from within the procedure above to deal with unsatisfied
 9877  * allocation dependencies in an inodeblock. The buffer must be
 9878  * locked, thus, no I/O completion operations can occur while we
 9879  * are manipulating its associated dependencies.
 9880  */
 9881 static void 
 9882 initiate_write_inodeblock_ufs1(inodedep, bp)
 9883         struct inodedep *inodedep;
 9884         struct buf *bp;                 /* The inode block */
 9885 {
 9886         struct allocdirect *adp, *lastadp;
 9887         struct ufs1_dinode *dp;
 9888         struct ufs1_dinode *sip;
 9889         struct inoref *inoref;
 9890         struct fs *fs;
 9891         ufs_lbn_t i;
 9892 #ifdef INVARIANTS
 9893         ufs_lbn_t prevlbn = 0;
 9894 #endif
 9895         int deplist;
 9896 
 9897         if (inodedep->id_state & IOSTARTED)
 9898                 panic("initiate_write_inodeblock_ufs1: already started");
 9899         inodedep->id_state |= IOSTARTED;
 9900         fs = inodedep->id_fs;
 9901         dp = (struct ufs1_dinode *)bp->b_data +
 9902             ino_to_fsbo(fs, inodedep->id_ino);
 9903 
 9904         /*
 9905          * If we're on the unlinked list but have not yet written our
 9906          * next pointer initialize it here.
 9907          */
 9908         if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) {
 9909                 struct inodedep *inon;
 9910 
 9911                 inon = TAILQ_NEXT(inodedep, id_unlinked);
 9912                 dp->di_freelink = inon ? inon->id_ino : 0;
 9913         }
 9914         /*
 9915          * If the bitmap is not yet written, then the allocated
 9916          * inode cannot be written to disk.
 9917          */
 9918         if ((inodedep->id_state & DEPCOMPLETE) == 0) {
 9919                 if (inodedep->id_savedino1 != NULL)
 9920                         panic("initiate_write_inodeblock_ufs1: I/O underway");
 9921                 FREE_LOCK(&lk);
 9922                 sip = malloc(sizeof(struct ufs1_dinode),
 9923                     M_SAVEDINO, M_SOFTDEP_FLAGS);
 9924                 ACQUIRE_LOCK(&lk);
 9925                 inodedep->id_savedino1 = sip;
 9926                 *inodedep->id_savedino1 = *dp;
 9927                 bzero((caddr_t)dp, sizeof(struct ufs1_dinode));
 9928                 dp->di_gen = inodedep->id_savedino1->di_gen;
 9929                 dp->di_freelink = inodedep->id_savedino1->di_freelink;
 9930                 return;
 9931         }
 9932         /*
 9933          * If no dependencies, then there is nothing to roll back.
 9934          */
 9935         inodedep->id_savedsize = dp->di_size;
 9936         inodedep->id_savedextsize = 0;
 9937         inodedep->id_savednlink = dp->di_nlink;
 9938         if (TAILQ_EMPTY(&inodedep->id_inoupdt) &&
 9939             TAILQ_EMPTY(&inodedep->id_inoreflst))
 9940                 return;
 9941         /*
 9942          * Revert the link count to that of the first unwritten journal entry.
 9943          */
 9944         inoref = TAILQ_FIRST(&inodedep->id_inoreflst);
 9945         if (inoref)
 9946                 dp->di_nlink = inoref->if_nlink;
 9947         /*
 9948          * Set the dependencies to busy.
 9949          */
 9950         for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
 9951              adp = TAILQ_NEXT(adp, ad_next)) {
 9952 #ifdef INVARIANTS
 9953                 if (deplist != 0 && prevlbn >= adp->ad_offset)
 9954                         panic("softdep_write_inodeblock: lbn order");
 9955                 prevlbn = adp->ad_offset;
 9956                 if (adp->ad_offset < NDADDR &&
 9957                     dp->di_db[adp->ad_offset] != adp->ad_newblkno)
 9958                         panic("%s: direct pointer #%jd mismatch %d != %jd",
 9959                             "softdep_write_inodeblock",
 9960                             (intmax_t)adp->ad_offset,
 9961                             dp->di_db[adp->ad_offset],
 9962                             (intmax_t)adp->ad_newblkno);
 9963                 if (adp->ad_offset >= NDADDR &&
 9964                     dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno)
 9965                         panic("%s: indirect pointer #%jd mismatch %d != %jd",
 9966                             "softdep_write_inodeblock",
 9967                             (intmax_t)adp->ad_offset - NDADDR,
 9968                             dp->di_ib[adp->ad_offset - NDADDR],
 9969                             (intmax_t)adp->ad_newblkno);
 9970                 deplist |= 1 << adp->ad_offset;
 9971                 if ((adp->ad_state & ATTACHED) == 0)
 9972                         panic("softdep_write_inodeblock: Unknown state 0x%x",
 9973                             adp->ad_state);
 9974 #endif /* INVARIANTS */
 9975                 adp->ad_state &= ~ATTACHED;
 9976                 adp->ad_state |= UNDONE;
 9977         }
 9978         /*
 9979          * The on-disk inode cannot claim to be any larger than the last
 9980          * fragment that has been written. Otherwise, the on-disk inode
 9981          * might have fragments that were not the last block in the file
 9982          * which would corrupt the filesystem.
 9983          */
 9984         for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
 9985              lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
 9986                 if (adp->ad_offset >= NDADDR)
 9987                         break;
 9988                 dp->di_db[adp->ad_offset] = adp->ad_oldblkno;
 9989                 /* keep going until hitting a rollback to a frag */
 9990                 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
 9991                         continue;
 9992                 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize;
 9993                 for (i = adp->ad_offset + 1; i < NDADDR; i++) {
 9994 #ifdef INVARIANTS
 9995                         if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
 9996                                 panic("softdep_write_inodeblock: lost dep1");
 9997 #endif /* INVARIANTS */
 9998                         dp->di_db[i] = 0;
 9999                 }
10000                 for (i = 0; i < NIADDR; i++) {
10001 #ifdef INVARIANTS
10002                         if (dp->di_ib[i] != 0 &&
10003                             (deplist & ((1 << NDADDR) << i)) == 0)
10004                                 panic("softdep_write_inodeblock: lost dep2");
10005 #endif /* INVARIANTS */
10006                         dp->di_ib[i] = 0;
10007                 }
10008                 return;
10009         }
10010         /*
10011          * If we have zero'ed out the last allocated block of the file,
10012          * roll back the size to the last currently allocated block.
10013          * We know that this last allocated block is a full-sized as
10014          * we already checked for fragments in the loop above.
10015          */
10016         if (lastadp != NULL &&
10017             dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) {
10018                 for (i = lastadp->ad_offset; i >= 0; i--)
10019                         if (dp->di_db[i] != 0)
10020                                 break;
10021                 dp->di_size = (i + 1) * fs->fs_bsize;
10022         }
10023         /*
10024          * The only dependencies are for indirect blocks.
10025          *
10026          * The file size for indirect block additions is not guaranteed.
10027          * Such a guarantee would be non-trivial to achieve. The conventional
10028          * synchronous write implementation also does not make this guarantee.
10029          * Fsck should catch and fix discrepancies. Arguably, the file size
10030          * can be over-estimated without destroying integrity when the file
10031          * moves into the indirect blocks (i.e., is large). If we want to
10032          * postpone fsck, we are stuck with this argument.
10033          */
10034         for (; adp; adp = TAILQ_NEXT(adp, ad_next))
10035                 dp->di_ib[adp->ad_offset - NDADDR] = 0;
10036 }
10037                 
10038 /*
10039  * Version of initiate_write_inodeblock that handles UFS2 dinodes.
10040  * Note that any bug fixes made to this routine must be done in the
10041  * version found above.
10042  *
10043  * Called from within the procedure above to deal with unsatisfied
10044  * allocation dependencies in an inodeblock. The buffer must be
10045  * locked, thus, no I/O completion operations can occur while we
10046  * are manipulating its associated dependencies.
10047  */
10048 static void 
10049 initiate_write_inodeblock_ufs2(inodedep, bp)
10050         struct inodedep *inodedep;
10051         struct buf *bp;                 /* The inode block */
10052 {
10053         struct allocdirect *adp, *lastadp;
10054         struct ufs2_dinode *dp;
10055         struct ufs2_dinode *sip;
10056         struct inoref *inoref;
10057         struct fs *fs;
10058         ufs_lbn_t i;
10059 #ifdef INVARIANTS
10060         ufs_lbn_t prevlbn = 0;
10061 #endif
10062         int deplist;
10063 
10064         if (inodedep->id_state & IOSTARTED)
10065                 panic("initiate_write_inodeblock_ufs2: already started");
10066         inodedep->id_state |= IOSTARTED;
10067         fs = inodedep->id_fs;
10068         dp = (struct ufs2_dinode *)bp->b_data +
10069             ino_to_fsbo(fs, inodedep->id_ino);
10070 
10071         /*
10072          * If we're on the unlinked list but have not yet written our
10073          * next pointer initialize it here.
10074          */
10075         if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) {
10076                 struct inodedep *inon;
10077 
10078                 inon = TAILQ_NEXT(inodedep, id_unlinked);
10079                 dp->di_freelink = inon ? inon->id_ino : 0;
10080         }
10081         /*
10082          * If the bitmap is not yet written, then the allocated
10083          * inode cannot be written to disk.
10084          */
10085         if ((inodedep->id_state & DEPCOMPLETE) == 0) {
10086                 if (inodedep->id_savedino2 != NULL)
10087                         panic("initiate_write_inodeblock_ufs2: I/O underway");
10088                 FREE_LOCK(&lk);
10089                 sip = malloc(sizeof(struct ufs2_dinode),
10090                     M_SAVEDINO, M_SOFTDEP_FLAGS);
10091                 ACQUIRE_LOCK(&lk);
10092                 inodedep->id_savedino2 = sip;
10093                 *inodedep->id_savedino2 = *dp;
10094                 bzero((caddr_t)dp, sizeof(struct ufs2_dinode));
10095                 dp->di_gen = inodedep->id_savedino2->di_gen;
10096                 dp->di_freelink = inodedep->id_savedino2->di_freelink;
10097                 return;
10098         }
10099         /*
10100          * If no dependencies, then there is nothing to roll back.
10101          */
10102         inodedep->id_savedsize = dp->di_size;
10103         inodedep->id_savedextsize = dp->di_extsize;
10104         inodedep->id_savednlink = dp->di_nlink;
10105         if (TAILQ_EMPTY(&inodedep->id_inoupdt) &&
10106             TAILQ_EMPTY(&inodedep->id_extupdt) &&
10107             TAILQ_EMPTY(&inodedep->id_inoreflst))
10108                 return;
10109         /*
10110          * Revert the link count to that of the first unwritten journal entry.
10111          */
10112         inoref = TAILQ_FIRST(&inodedep->id_inoreflst);
10113         if (inoref)
10114                 dp->di_nlink = inoref->if_nlink;
10115 
10116         /*
10117          * Set the ext data dependencies to busy.
10118          */
10119         for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
10120              adp = TAILQ_NEXT(adp, ad_next)) {
10121 #ifdef INVARIANTS
10122                 if (deplist != 0 && prevlbn >= adp->ad_offset)
10123                         panic("softdep_write_inodeblock: lbn order");
10124                 prevlbn = adp->ad_offset;
10125                 if (dp->di_extb[adp->ad_offset] != adp->ad_newblkno)
10126                         panic("%s: direct pointer #%jd mismatch %jd != %jd",
10127                             "softdep_write_inodeblock",
10128                             (intmax_t)adp->ad_offset,
10129                             (intmax_t)dp->di_extb[adp->ad_offset],
10130                             (intmax_t)adp->ad_newblkno);
10131                 deplist |= 1 << adp->ad_offset;
10132                 if ((adp->ad_state & ATTACHED) == 0)
10133                         panic("softdep_write_inodeblock: Unknown state 0x%x",
10134                             adp->ad_state);
10135 #endif /* INVARIANTS */
10136                 adp->ad_state &= ~ATTACHED;
10137                 adp->ad_state |= UNDONE;
10138         }
10139         /*
10140          * The on-disk inode cannot claim to be any larger than the last
10141          * fragment that has been written. Otherwise, the on-disk inode
10142          * might have fragments that were not the last block in the ext
10143          * data which would corrupt the filesystem.
10144          */
10145         for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
10146              lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
10147                 dp->di_extb[adp->ad_offset] = adp->ad_oldblkno;
10148                 /* keep going until hitting a rollback to a frag */
10149                 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
10150                         continue;
10151                 dp->di_extsize = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize;
10152                 for (i = adp->ad_offset + 1; i < NXADDR; i++) {
10153 #ifdef INVARIANTS
10154                         if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0)
10155                                 panic("softdep_write_inodeblock: lost dep1");
10156 #endif /* INVARIANTS */
10157                         dp->di_extb[i] = 0;
10158                 }
10159                 lastadp = NULL;
10160                 break;
10161         }
10162         /*
10163          * If we have zero'ed out the last allocated block of the ext
10164          * data, roll back the size to the last currently allocated block.
10165          * We know that this last allocated block is a full-sized as
10166          * we already checked for fragments in the loop above.
10167          */
10168         if (lastadp != NULL &&
10169             dp->di_extsize <= (lastadp->ad_offset + 1) * fs->fs_bsize) {
10170                 for (i = lastadp->ad_offset; i >= 0; i--)
10171                         if (dp->di_extb[i] != 0)
10172                                 break;
10173                 dp->di_extsize = (i + 1) * fs->fs_bsize;
10174         }
10175         /*
10176          * Set the file data dependencies to busy.
10177          */
10178         for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
10179              adp = TAILQ_NEXT(adp, ad_next)) {
10180 #ifdef INVARIANTS
10181                 if (deplist != 0 && prevlbn >= adp->ad_offset)
10182                         panic("softdep_write_inodeblock: lbn order");
10183                 if ((adp->ad_state & ATTACHED) == 0)
10184                         panic("inodedep %p and adp %p not attached", inodedep, adp);
10185                 prevlbn = adp->ad_offset;
10186                 if (adp->ad_offset < NDADDR &&
10187                     dp->di_db[adp->ad_offset] != adp->ad_newblkno)
10188                         panic("%s: direct pointer #%jd mismatch %jd != %jd",
10189                             "softdep_write_inodeblock",
10190                             (intmax_t)adp->ad_offset,
10191                             (intmax_t)dp->di_db[adp->ad_offset],
10192                             (intmax_t)adp->ad_newblkno);
10193                 if (adp->ad_offset >= NDADDR &&
10194                     dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno)
10195                         panic("%s indirect pointer #%jd mismatch %jd != %jd",
10196                             "softdep_write_inodeblock:",
10197                             (intmax_t)adp->ad_offset - NDADDR,
10198                             (intmax_t)dp->di_ib[adp->ad_offset - NDADDR],
10199                             (intmax_t)adp->ad_newblkno);
10200                 deplist |= 1 << adp->ad_offset;
10201                 if ((adp->ad_state & ATTACHED) == 0)
10202                         panic("softdep_write_inodeblock: Unknown state 0x%x",
10203                             adp->ad_state);
10204 #endif /* INVARIANTS */
10205                 adp->ad_state &= ~ATTACHED;
10206                 adp->ad_state |= UNDONE;
10207         }
10208         /*
10209          * The on-disk inode cannot claim to be any larger than the last
10210          * fragment that has been written. Otherwise, the on-disk inode
10211          * might have fragments that were not the last block in the file
10212          * which would corrupt the filesystem.
10213          */
10214         for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
10215              lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
10216                 if (adp->ad_offset >= NDADDR)
10217                         break;
10218                 dp->di_db[adp->ad_offset] = adp->ad_oldblkno;
10219                 /* keep going until hitting a rollback to a frag */
10220                 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
10221                         continue;
10222                 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize;
10223                 for (i = adp->ad_offset + 1; i < NDADDR; i++) {
10224 #ifdef INVARIANTS
10225                         if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
10226                                 panic("softdep_write_inodeblock: lost dep2");
10227 #endif /* INVARIANTS */
10228                         dp->di_db[i] = 0;
10229                 }
10230                 for (i = 0; i < NIADDR; i++) {
10231 #ifdef INVARIANTS
10232                         if (dp->di_ib[i] != 0 &&
10233                             (deplist & ((1 << NDADDR) << i)) == 0)
10234                                 panic("softdep_write_inodeblock: lost dep3");
10235 #endif /* INVARIANTS */
10236                         dp->di_ib[i] = 0;
10237                 }
10238                 return;
10239         }
10240         /*
10241          * If we have zero'ed out the last allocated block of the file,
10242          * roll back the size to the last currently allocated block.
10243          * We know that this last allocated block is a full-sized as
10244          * we already checked for fragments in the loop above.
10245          */
10246         if (lastadp != NULL &&
10247             dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) {
10248                 for (i = lastadp->ad_offset; i >= 0; i--)
10249                         if (dp->di_db[i] != 0)
10250                                 break;
10251                 dp->di_size = (i + 1) * fs->fs_bsize;
10252         }
10253         /*
10254          * The only dependencies are for indirect blocks.
10255          *
10256          * The file size for indirect block additions is not guaranteed.
10257          * Such a guarantee would be non-trivial to achieve. The conventional
10258          * synchronous write implementation also does not make this guarantee.
10259          * Fsck should catch and fix discrepancies. Arguably, the file size
10260          * can be over-estimated without destroying integrity when the file
10261          * moves into the indirect blocks (i.e., is large). If we want to
10262          * postpone fsck, we are stuck with this argument.
10263          */
10264         for (; adp; adp = TAILQ_NEXT(adp, ad_next))
10265                 dp->di_ib[adp->ad_offset - NDADDR] = 0;
10266 }
10267 
10268 /*
10269  * Cancel an indirdep as a result of truncation.  Release all of the
10270  * children allocindirs and place their journal work on the appropriate
10271  * list.
10272  */
10273 static void
10274 cancel_indirdep(indirdep, bp, freeblks)
10275         struct indirdep *indirdep;
10276         struct buf *bp;
10277         struct freeblks *freeblks;
10278 {
10279         struct allocindir *aip;
10280 
10281         /*
10282          * None of the indirect pointers will ever be visible,
10283          * so they can simply be tossed. GOINGAWAY ensures
10284          * that allocated pointers will be saved in the buffer
10285          * cache until they are freed. Note that they will
10286          * only be able to be found by their physical address
10287          * since the inode mapping the logical address will
10288          * be gone. The save buffer used for the safe copy
10289          * was allocated in setup_allocindir_phase2 using
10290          * the physical address so it could be used for this
10291          * purpose. Hence we swap the safe copy with the real
10292          * copy, allowing the safe copy to be freed and holding
10293          * on to the real copy for later use in indir_trunc.
10294          */
10295         if (indirdep->ir_state & GOINGAWAY)
10296                 panic("cancel_indirdep: already gone");
10297         if ((indirdep->ir_state & DEPCOMPLETE) == 0) {
10298                 indirdep->ir_state |= DEPCOMPLETE;
10299                 LIST_REMOVE(indirdep, ir_next);
10300         }
10301         indirdep->ir_state |= GOINGAWAY;
10302         VFSTOUFS(indirdep->ir_list.wk_mp)->um_numindirdeps += 1;
10303         /*
10304          * Pass in bp for blocks still have journal writes
10305          * pending so we can cancel them on their own.
10306          */
10307         while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
10308                 cancel_allocindir(aip, bp, freeblks, 0);
10309         while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0)
10310                 cancel_allocindir(aip, NULL, freeblks, 0);
10311         while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != 0)
10312                 cancel_allocindir(aip, NULL, freeblks, 0);
10313         while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != 0)
10314                 cancel_allocindir(aip, NULL, freeblks, 0);
10315         /*
10316          * If there are pending partial truncations we need to keep the
10317          * old block copy around until they complete.  This is because
10318          * the current b_data is not a perfect superset of the available
10319          * blocks.
10320          */
10321         if (TAILQ_EMPTY(&indirdep->ir_trunc))
10322                 bcopy(bp->b_data, indirdep->ir_savebp->b_data, bp->b_bcount);
10323         else
10324                 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
10325         WORKLIST_REMOVE(&indirdep->ir_list);
10326         WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, &indirdep->ir_list);
10327         indirdep->ir_bp = NULL;
10328         indirdep->ir_freeblks = freeblks;
10329 }
10330 
10331 /*
10332  * Free an indirdep once it no longer has new pointers to track.
10333  */
10334 static void
10335 free_indirdep(indirdep)
10336         struct indirdep *indirdep;
10337 {
10338 
10339         KASSERT(TAILQ_EMPTY(&indirdep->ir_trunc),
10340             ("free_indirdep: Indir trunc list not empty."));
10341         KASSERT(LIST_EMPTY(&indirdep->ir_completehd),
10342             ("free_indirdep: Complete head not empty."));
10343         KASSERT(LIST_EMPTY(&indirdep->ir_writehd),
10344             ("free_indirdep: write head not empty."));
10345         KASSERT(LIST_EMPTY(&indirdep->ir_donehd),
10346             ("free_indirdep: done head not empty."));
10347         KASSERT(LIST_EMPTY(&indirdep->ir_deplisthd),
10348             ("free_indirdep: deplist head not empty."));
10349         KASSERT((indirdep->ir_state & DEPCOMPLETE),
10350             ("free_indirdep: %p still on newblk list.", indirdep));
10351         KASSERT(indirdep->ir_saveddata == NULL,
10352             ("free_indirdep: %p still has saved data.", indirdep));
10353         if (indirdep->ir_state & ONWORKLIST)
10354                 WORKLIST_REMOVE(&indirdep->ir_list);
10355         WORKITEM_FREE(indirdep, D_INDIRDEP);
10356 }
10357 
10358 /*
10359  * Called before a write to an indirdep.  This routine is responsible for
10360  * rolling back pointers to a safe state which includes only those
10361  * allocindirs which have been completed.
10362  */
10363 static void
10364 initiate_write_indirdep(indirdep, bp)
10365         struct indirdep *indirdep;
10366         struct buf *bp;
10367 {
10368 
10369         indirdep->ir_state |= IOSTARTED;
10370         if (indirdep->ir_state & GOINGAWAY)
10371                 panic("disk_io_initiation: indirdep gone");
10372         /*
10373          * If there are no remaining dependencies, this will be writing
10374          * the real pointers.
10375          */
10376         if (LIST_EMPTY(&indirdep->ir_deplisthd) &&
10377             TAILQ_EMPTY(&indirdep->ir_trunc))
10378                 return;
10379         /*
10380          * Replace up-to-date version with safe version.
10381          */
10382         if (indirdep->ir_saveddata == NULL) {
10383                 FREE_LOCK(&lk);
10384                 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP,
10385                     M_SOFTDEP_FLAGS);
10386                 ACQUIRE_LOCK(&lk);
10387         }
10388         indirdep->ir_state &= ~ATTACHED;
10389         indirdep->ir_state |= UNDONE;
10390         bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
10391         bcopy(indirdep->ir_savebp->b_data, bp->b_data,
10392             bp->b_bcount);
10393 }
10394 
10395 /*
10396  * Called when an inode has been cleared in a cg bitmap.  This finally
10397  * eliminates any canceled jaddrefs
10398  */
10399 void
10400 softdep_setup_inofree(mp, bp, ino, wkhd)
10401         struct mount *mp;
10402         struct buf *bp;
10403         ino_t ino;
10404         struct workhead *wkhd;
10405 {
10406         struct worklist *wk, *wkn;
10407         struct inodedep *inodedep;
10408         uint8_t *inosused;
10409         struct cg *cgp;
10410         struct fs *fs;
10411 
10412         ACQUIRE_LOCK(&lk);
10413         fs = VFSTOUFS(mp)->um_fs;
10414         cgp = (struct cg *)bp->b_data;
10415         inosused = cg_inosused(cgp);
10416         if (isset(inosused, ino % fs->fs_ipg))
10417                 panic("softdep_setup_inofree: inode %ju not freed.",
10418                     (uintmax_t)ino);
10419         if (inodedep_lookup(mp, ino, 0, &inodedep))
10420                 panic("softdep_setup_inofree: ino %ju has existing inodedep %p",
10421                     (uintmax_t)ino, inodedep);
10422         if (wkhd) {
10423                 LIST_FOREACH_SAFE(wk, wkhd, wk_list, wkn) {
10424                         if (wk->wk_type != D_JADDREF)
10425                                 continue;
10426                         WORKLIST_REMOVE(wk);
10427                         /*
10428                          * We can free immediately even if the jaddref
10429                          * isn't attached in a background write as now
10430                          * the bitmaps are reconciled.
10431                          */
10432                         wk->wk_state |= COMPLETE | ATTACHED;
10433                         free_jaddref(WK_JADDREF(wk));
10434                 }
10435                 jwork_move(&bp->b_dep, wkhd);
10436         }
10437         FREE_LOCK(&lk);
10438 }
10439 
10440 
10441 /*
10442  * Called via ffs_blkfree() after a set of frags has been cleared from a cg
10443  * map.  Any dependencies waiting for the write to clear are added to the
10444  * buf's list and any jnewblks that are being canceled are discarded
10445  * immediately.
10446  */
10447 void
10448 softdep_setup_blkfree(mp, bp, blkno, frags, wkhd)
10449         struct mount *mp;
10450         struct buf *bp;
10451         ufs2_daddr_t blkno;
10452         int frags;
10453         struct workhead *wkhd;
10454 {
10455         struct bmsafemap *bmsafemap;
10456         struct jnewblk *jnewblk;
10457         struct worklist *wk;
10458         struct fs *fs;
10459 #ifdef SUJ_DEBUG
10460         uint8_t *blksfree;
10461         struct cg *cgp;
10462         ufs2_daddr_t jstart;
10463         ufs2_daddr_t jend;
10464         ufs2_daddr_t end;
10465         long bno;
10466         int i;
10467 #endif
10468 
10469         CTR3(KTR_SUJ,
10470             "softdep_setup_blkfree: blkno %jd frags %d wk head %p",
10471             blkno, frags, wkhd);
10472 
10473         ACQUIRE_LOCK(&lk);
10474         /* Lookup the bmsafemap so we track when it is dirty. */
10475         fs = VFSTOUFS(mp)->um_fs;
10476         bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL);
10477         /*
10478          * Detach any jnewblks which have been canceled.  They must linger
10479          * until the bitmap is cleared again by ffs_blkfree() to prevent
10480          * an unjournaled allocation from hitting the disk.
10481          */
10482         if (wkhd) {
10483                 while ((wk = LIST_FIRST(wkhd)) != NULL) {
10484                         CTR2(KTR_SUJ,
10485                             "softdep_setup_blkfree: blkno %jd wk type %d",
10486                             blkno, wk->wk_type);
10487                         WORKLIST_REMOVE(wk);
10488                         if (wk->wk_type != D_JNEWBLK) {
10489                                 WORKLIST_INSERT(&bmsafemap->sm_freehd, wk);
10490                                 continue;
10491                         }
10492                         jnewblk = WK_JNEWBLK(wk);
10493                         KASSERT(jnewblk->jn_state & GOINGAWAY,
10494                             ("softdep_setup_blkfree: jnewblk not canceled."));
10495 #ifdef SUJ_DEBUG
10496                         /*
10497                          * Assert that this block is free in the bitmap
10498                          * before we discard the jnewblk.
10499                          */
10500                         cgp = (struct cg *)bp->b_data;
10501                         blksfree = cg_blksfree(cgp);
10502                         bno = dtogd(fs, jnewblk->jn_blkno);
10503                         for (i = jnewblk->jn_oldfrags;
10504                             i < jnewblk->jn_frags; i++) {
10505                                 if (isset(blksfree, bno + i))
10506                                         continue;
10507                                 panic("softdep_setup_blkfree: not free");
10508                         }
10509 #endif
10510                         /*
10511                          * Even if it's not attached we can free immediately
10512                          * as the new bitmap is correct.
10513                          */
10514                         wk->wk_state |= COMPLETE | ATTACHED;
10515                         free_jnewblk(jnewblk);
10516                 }
10517         }
10518 
10519 #ifdef SUJ_DEBUG
10520         /*
10521          * Assert that we are not freeing a block which has an outstanding
10522          * allocation dependency.
10523          */
10524         fs = VFSTOUFS(mp)->um_fs;
10525         bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL);
10526         end = blkno + frags;
10527         LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) {
10528                 /*
10529                  * Don't match against blocks that will be freed when the
10530                  * background write is done.
10531                  */
10532                 if ((jnewblk->jn_state & (ATTACHED | COMPLETE | DEPCOMPLETE)) ==
10533                     (COMPLETE | DEPCOMPLETE))
10534                         continue;
10535                 jstart = jnewblk->jn_blkno + jnewblk->jn_oldfrags;
10536                 jend = jnewblk->jn_blkno + jnewblk->jn_frags;
10537                 if ((blkno >= jstart && blkno < jend) ||
10538                     (end > jstart && end <= jend)) {
10539                         printf("state 0x%X %jd - %d %d dep %p\n",
10540                             jnewblk->jn_state, jnewblk->jn_blkno,
10541                             jnewblk->jn_oldfrags, jnewblk->jn_frags,
10542                             jnewblk->jn_dep);
10543                         panic("softdep_setup_blkfree: "
10544                             "%jd-%jd(%d) overlaps with %jd-%jd",
10545                             blkno, end, frags, jstart, jend);
10546                 }
10547         }
10548 #endif
10549         FREE_LOCK(&lk);
10550 }
10551 
10552 /*
10553  * Revert a block allocation when the journal record that describes it
10554  * is not yet written.
10555  */
10556 int
10557 jnewblk_rollback(jnewblk, fs, cgp, blksfree)
10558         struct jnewblk *jnewblk;
10559         struct fs *fs;
10560         struct cg *cgp;
10561         uint8_t *blksfree;
10562 {
10563         ufs1_daddr_t fragno;
10564         long cgbno, bbase;
10565         int frags, blk;
10566         int i;
10567 
10568         frags = 0;
10569         cgbno = dtogd(fs, jnewblk->jn_blkno);
10570         /*
10571          * We have to test which frags need to be rolled back.  We may
10572          * be operating on a stale copy when doing background writes.
10573          */
10574         for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++)
10575                 if (isclr(blksfree, cgbno + i))
10576                         frags++;
10577         if (frags == 0)
10578                 return (0);
10579         /*
10580          * This is mostly ffs_blkfree() sans some validation and
10581          * superblock updates.
10582          */
10583         if (frags == fs->fs_frag) {
10584                 fragno = fragstoblks(fs, cgbno);
10585                 ffs_setblock(fs, blksfree, fragno);
10586                 ffs_clusteracct(fs, cgp, fragno, 1);
10587                 cgp->cg_cs.cs_nbfree++;
10588         } else {
10589                 cgbno += jnewblk->jn_oldfrags;
10590                 bbase = cgbno - fragnum(fs, cgbno);
10591                 /* Decrement the old frags.  */
10592                 blk = blkmap(fs, blksfree, bbase);
10593                 ffs_fragacct(fs, blk, cgp->cg_frsum, -1);
10594                 /* Deallocate the fragment */
10595                 for (i = 0; i < frags; i++)
10596                         setbit(blksfree, cgbno + i);
10597                 cgp->cg_cs.cs_nffree += frags;
10598                 /* Add back in counts associated with the new frags */
10599                 blk = blkmap(fs, blksfree, bbase);
10600                 ffs_fragacct(fs, blk, cgp->cg_frsum, 1);
10601                 /* If a complete block has been reassembled, account for it. */
10602                 fragno = fragstoblks(fs, bbase);
10603                 if (ffs_isblock(fs, blksfree, fragno)) {
10604                         cgp->cg_cs.cs_nffree -= fs->fs_frag;
10605                         ffs_clusteracct(fs, cgp, fragno, 1);
10606                         cgp->cg_cs.cs_nbfree++;
10607                 }
10608         }
10609         stat_jnewblk++;
10610         jnewblk->jn_state &= ~ATTACHED;
10611         jnewblk->jn_state |= UNDONE;
10612 
10613         return (frags);
10614 }
10615 
10616 static void 
10617 initiate_write_bmsafemap(bmsafemap, bp)
10618         struct bmsafemap *bmsafemap;
10619         struct buf *bp;                 /* The cg block. */
10620 {
10621         struct jaddref *jaddref;
10622         struct jnewblk *jnewblk;
10623         uint8_t *inosused;
10624         uint8_t *blksfree;
10625         struct cg *cgp;
10626         struct fs *fs;
10627         ino_t ino;
10628 
10629         if (bmsafemap->sm_state & IOSTARTED)
10630                 return;
10631         bmsafemap->sm_state |= IOSTARTED;
10632         /*
10633          * Clear any inode allocations which are pending journal writes.
10634          */
10635         if (LIST_FIRST(&bmsafemap->sm_jaddrefhd) != NULL) {
10636                 cgp = (struct cg *)bp->b_data;
10637                 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
10638                 inosused = cg_inosused(cgp);
10639                 LIST_FOREACH(jaddref, &bmsafemap->sm_jaddrefhd, ja_bmdeps) {
10640                         ino = jaddref->ja_ino % fs->fs_ipg;
10641                         if (isset(inosused, ino)) {
10642                                 if ((jaddref->ja_mode & IFMT) == IFDIR)
10643                                         cgp->cg_cs.cs_ndir--;
10644                                 cgp->cg_cs.cs_nifree++;
10645                                 clrbit(inosused, ino);
10646                                 jaddref->ja_state &= ~ATTACHED;
10647                                 jaddref->ja_state |= UNDONE;
10648                                 stat_jaddref++;
10649                         } else
10650                                 panic("initiate_write_bmsafemap: inode %ju "
10651                                     "marked free", (uintmax_t)jaddref->ja_ino);
10652                 }
10653         }
10654         /*
10655          * Clear any block allocations which are pending journal writes.
10656          */
10657         if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) {
10658                 cgp = (struct cg *)bp->b_data;
10659                 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
10660                 blksfree = cg_blksfree(cgp);
10661                 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) {
10662                         if (jnewblk_rollback(jnewblk, fs, cgp, blksfree))
10663                                 continue;
10664                         panic("initiate_write_bmsafemap: block %jd "
10665                             "marked free", jnewblk->jn_blkno);
10666                 }
10667         }
10668         /*
10669          * Move allocation lists to the written lists so they can be
10670          * cleared once the block write is complete.
10671          */
10672         LIST_SWAP(&bmsafemap->sm_inodedephd, &bmsafemap->sm_inodedepwr,
10673             inodedep, id_deps);
10674         LIST_SWAP(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr,
10675             newblk, nb_deps);
10676         LIST_SWAP(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, worklist,
10677             wk_list);
10678 }
10679 
10680 /*
10681  * This routine is called during the completion interrupt
10682  * service routine for a disk write (from the procedure called
10683  * by the device driver to inform the filesystem caches of
10684  * a request completion).  It should be called early in this
10685  * procedure, before the block is made available to other
10686  * processes or other routines are called.
10687  *
10688  */
10689 static void 
10690 softdep_disk_write_complete(bp)
10691         struct buf *bp;         /* describes the completed disk write */
10692 {
10693         struct worklist *wk;
10694         struct worklist *owk;
10695         struct workhead reattach;
10696         struct freeblks *freeblks;
10697         struct buf *sbp;
10698 
10699         /*
10700          * If an error occurred while doing the write, then the data
10701          * has not hit the disk and the dependencies cannot be unrolled.
10702          */
10703         if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0)
10704                 return;
10705         LIST_INIT(&reattach);
10706         /*
10707          * This lock must not be released anywhere in this code segment.
10708          */
10709         sbp = NULL;
10710         owk = NULL;
10711         ACQUIRE_LOCK(&lk);
10712         while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
10713                 WORKLIST_REMOVE(wk);
10714                 dep_write[wk->wk_type]++;
10715                 if (wk == owk)
10716                         panic("duplicate worklist: %p\n", wk);
10717                 owk = wk;
10718                 switch (wk->wk_type) {
10719 
10720                 case D_PAGEDEP:
10721                         if (handle_written_filepage(WK_PAGEDEP(wk), bp))
10722                                 WORKLIST_INSERT(&reattach, wk);
10723                         continue;
10724 
10725                 case D_INODEDEP:
10726                         if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
10727                                 WORKLIST_INSERT(&reattach, wk);
10728                         continue;
10729 
10730                 case D_BMSAFEMAP:
10731                         if (handle_written_bmsafemap(WK_BMSAFEMAP(wk), bp))
10732                                 WORKLIST_INSERT(&reattach, wk);
10733                         continue;
10734 
10735                 case D_MKDIR:
10736                         handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
10737                         continue;
10738 
10739                 case D_ALLOCDIRECT:
10740                         wk->wk_state |= COMPLETE;
10741                         handle_allocdirect_partdone(WK_ALLOCDIRECT(wk), NULL);
10742                         continue;
10743 
10744                 case D_ALLOCINDIR:
10745                         wk->wk_state |= COMPLETE;
10746                         handle_allocindir_partdone(WK_ALLOCINDIR(wk));
10747                         continue;
10748 
10749                 case D_INDIRDEP:
10750                         if (handle_written_indirdep(WK_INDIRDEP(wk), bp, &sbp))
10751                                 WORKLIST_INSERT(&reattach, wk);
10752                         continue;
10753 
10754                 case D_FREEBLKS:
10755                         wk->wk_state |= COMPLETE;
10756                         freeblks = WK_FREEBLKS(wk);
10757                         if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE &&
10758                             LIST_EMPTY(&freeblks->fb_jblkdephd))
10759                                 add_to_worklist(wk, WK_NODELAY);
10760                         continue;
10761 
10762                 case D_FREEWORK:
10763                         handle_written_freework(WK_FREEWORK(wk));
10764                         break;
10765 
10766                 case D_JSEGDEP:
10767                         free_jsegdep(WK_JSEGDEP(wk));
10768                         continue;
10769 
10770                 case D_JSEG:
10771                         handle_written_jseg(WK_JSEG(wk), bp);
10772                         continue;
10773 
10774                 case D_SBDEP:
10775                         if (handle_written_sbdep(WK_SBDEP(wk), bp))
10776                                 WORKLIST_INSERT(&reattach, wk);
10777                         continue;
10778 
10779                 case D_FREEDEP:
10780                         free_freedep(WK_FREEDEP(wk));
10781                         continue;
10782 
10783                 default:
10784                         panic("handle_disk_write_complete: Unknown type %s",
10785                             TYPENAME(wk->wk_type));
10786                         /* NOTREACHED */
10787                 }
10788         }
10789         /*
10790          * Reattach any requests that must be redone.
10791          */
10792         while ((wk = LIST_FIRST(&reattach)) != NULL) {
10793                 WORKLIST_REMOVE(wk);
10794                 WORKLIST_INSERT(&bp->b_dep, wk);
10795         }
10796         FREE_LOCK(&lk);
10797         if (sbp)
10798                 brelse(sbp);
10799 }
10800 
10801 /*
10802  * Called from within softdep_disk_write_complete above. Note that
10803  * this routine is always called from interrupt level with further
10804  * splbio interrupts blocked.
10805  */
10806 static void 
10807 handle_allocdirect_partdone(adp, wkhd)
10808         struct allocdirect *adp;        /* the completed allocdirect */
10809         struct workhead *wkhd;          /* Work to do when inode is writtne. */
10810 {
10811         struct allocdirectlst *listhead;
10812         struct allocdirect *listadp;
10813         struct inodedep *inodedep;
10814         long bsize;
10815 
10816         if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
10817                 return;
10818         /*
10819          * The on-disk inode cannot claim to be any larger than the last
10820          * fragment that has been written. Otherwise, the on-disk inode
10821          * might have fragments that were not the last block in the file
10822          * which would corrupt the filesystem. Thus, we cannot free any
10823          * allocdirects after one whose ad_oldblkno claims a fragment as
10824          * these blocks must be rolled back to zero before writing the inode.
10825          * We check the currently active set of allocdirects in id_inoupdt
10826          * or id_extupdt as appropriate.
10827          */
10828         inodedep = adp->ad_inodedep;
10829         bsize = inodedep->id_fs->fs_bsize;
10830         if (adp->ad_state & EXTDATA)
10831                 listhead = &inodedep->id_extupdt;
10832         else
10833                 listhead = &inodedep->id_inoupdt;
10834         TAILQ_FOREACH(listadp, listhead, ad_next) {
10835                 /* found our block */
10836                 if (listadp == adp)
10837                         break;
10838                 /* continue if ad_oldlbn is not a fragment */
10839                 if (listadp->ad_oldsize == 0 ||
10840                     listadp->ad_oldsize == bsize)
10841                         continue;
10842                 /* hit a fragment */
10843                 return;
10844         }
10845         /*
10846          * If we have reached the end of the current list without
10847          * finding the just finished dependency, then it must be
10848          * on the future dependency list. Future dependencies cannot
10849          * be freed until they are moved to the current list.
10850          */
10851         if (listadp == NULL) {
10852 #ifdef DEBUG
10853                 if (adp->ad_state & EXTDATA)
10854                         listhead = &inodedep->id_newextupdt;
10855                 else
10856                         listhead = &inodedep->id_newinoupdt;
10857                 TAILQ_FOREACH(listadp, listhead, ad_next)
10858                         /* found our block */
10859                         if (listadp == adp)
10860                                 break;
10861                 if (listadp == NULL)
10862                         panic("handle_allocdirect_partdone: lost dep");
10863 #endif /* DEBUG */
10864                 return;
10865         }
10866         /*
10867          * If we have found the just finished dependency, then queue
10868          * it along with anything that follows it that is complete.
10869          * Since the pointer has not yet been written in the inode
10870          * as the dependency prevents it, place the allocdirect on the
10871          * bufwait list where it will be freed once the pointer is
10872          * valid.
10873          */
10874         if (wkhd == NULL)
10875                 wkhd = &inodedep->id_bufwait;
10876         for (; adp; adp = listadp) {
10877                 listadp = TAILQ_NEXT(adp, ad_next);
10878                 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
10879                         return;
10880                 TAILQ_REMOVE(listhead, adp, ad_next);
10881                 WORKLIST_INSERT(wkhd, &adp->ad_block.nb_list);
10882         }
10883 }
10884 
10885 /*
10886  * Called from within softdep_disk_write_complete above.  This routine
10887  * completes successfully written allocindirs.
10888  */
10889 static void
10890 handle_allocindir_partdone(aip)
10891         struct allocindir *aip;         /* the completed allocindir */
10892 {
10893         struct indirdep *indirdep;
10894 
10895         if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
10896                 return;
10897         indirdep = aip->ai_indirdep;
10898         LIST_REMOVE(aip, ai_next);
10899         /*
10900          * Don't set a pointer while the buffer is undergoing IO or while
10901          * we have active truncations.
10902          */
10903         if (indirdep->ir_state & UNDONE || !TAILQ_EMPTY(&indirdep->ir_trunc)) {
10904                 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
10905                 return;
10906         }
10907         if (indirdep->ir_state & UFS1FMT)
10908                 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
10909                     aip->ai_newblkno;
10910         else
10911                 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
10912                     aip->ai_newblkno;
10913         /*
10914          * Await the pointer write before freeing the allocindir.
10915          */
10916         LIST_INSERT_HEAD(&indirdep->ir_writehd, aip, ai_next);
10917 }
10918 
10919 /*
10920  * Release segments held on a jwork list.
10921  */
10922 static void
10923 handle_jwork(wkhd)
10924         struct workhead *wkhd;
10925 {
10926         struct worklist *wk;
10927 
10928         while ((wk = LIST_FIRST(wkhd)) != NULL) {
10929                 WORKLIST_REMOVE(wk);
10930                 switch (wk->wk_type) {
10931                 case D_JSEGDEP:
10932                         free_jsegdep(WK_JSEGDEP(wk));
10933                         continue;
10934                 case D_FREEDEP:
10935                         free_freedep(WK_FREEDEP(wk));
10936                         continue;
10937                 case D_FREEFRAG:
10938                         rele_jseg(WK_JSEG(WK_FREEFRAG(wk)->ff_jdep));
10939                         WORKITEM_FREE(wk, D_FREEFRAG);
10940                         continue;
10941                 case D_FREEWORK:
10942                         handle_written_freework(WK_FREEWORK(wk));
10943                         continue;
10944                 default:
10945                         panic("handle_jwork: Unknown type %s\n",
10946                             TYPENAME(wk->wk_type));
10947                 }
10948         }
10949 }
10950 
10951 /*
10952  * Handle the bufwait list on an inode when it is safe to release items
10953  * held there.  This normally happens after an inode block is written but
10954  * may be delayed and handled later if there are pending journal items that
10955  * are not yet safe to be released.
10956  */
10957 static struct freefile *
10958 handle_bufwait(inodedep, refhd)
10959         struct inodedep *inodedep;
10960         struct workhead *refhd;
10961 {
10962         struct jaddref *jaddref;
10963         struct freefile *freefile;
10964         struct worklist *wk;
10965 
10966         freefile = NULL;
10967         while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
10968                 WORKLIST_REMOVE(wk);
10969                 switch (wk->wk_type) {
10970                 case D_FREEFILE:
10971                         /*
10972                          * We defer adding freefile to the worklist
10973                          * until all other additions have been made to
10974                          * ensure that it will be done after all the
10975                          * old blocks have been freed.
10976                          */
10977                         if (freefile != NULL)
10978                                 panic("handle_bufwait: freefile");
10979                         freefile = WK_FREEFILE(wk);
10980                         continue;
10981 
10982                 case D_MKDIR:
10983                         handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
10984                         continue;
10985 
10986                 case D_DIRADD:
10987                         diradd_inode_written(WK_DIRADD(wk), inodedep);
10988                         continue;
10989 
10990                 case D_FREEFRAG:
10991                         wk->wk_state |= COMPLETE;
10992                         if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE)
10993                                 add_to_worklist(wk, 0);
10994                         continue;
10995 
10996                 case D_DIRREM:
10997                         wk->wk_state |= COMPLETE;
10998                         add_to_worklist(wk, 0);
10999                         continue;
11000 
11001                 case D_ALLOCDIRECT:
11002                 case D_ALLOCINDIR:
11003                         free_newblk(WK_NEWBLK(wk));
11004                         continue;
11005 
11006                 case D_JNEWBLK:
11007                         wk->wk_state |= COMPLETE;
11008                         free_jnewblk(WK_JNEWBLK(wk));
11009                         continue;
11010 
11011                 /*
11012                  * Save freed journal segments and add references on
11013                  * the supplied list which will delay their release
11014                  * until the cg bitmap is cleared on disk.
11015                  */
11016                 case D_JSEGDEP:
11017                         if (refhd == NULL)
11018                                 free_jsegdep(WK_JSEGDEP(wk));
11019                         else
11020                                 WORKLIST_INSERT(refhd, wk);
11021                         continue;
11022 
11023                 case D_JADDREF:
11024                         jaddref = WK_JADDREF(wk);
11025                         TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref,
11026                             if_deps);
11027                         /*
11028                          * Transfer any jaddrefs to the list to be freed with
11029                          * the bitmap if we're handling a removed file.
11030                          */
11031                         if (refhd == NULL) {
11032                                 wk->wk_state |= COMPLETE;
11033                                 free_jaddref(jaddref);
11034                         } else
11035                                 WORKLIST_INSERT(refhd, wk);
11036                         continue;
11037 
11038                 default:
11039                         panic("handle_bufwait: Unknown type %p(%s)",
11040                             wk, TYPENAME(wk->wk_type));
11041                         /* NOTREACHED */
11042                 }
11043         }
11044         return (freefile);
11045 }
11046 /*
11047  * Called from within softdep_disk_write_complete above to restore
11048  * in-memory inode block contents to their most up-to-date state. Note
11049  * that this routine is always called from interrupt level with further
11050  * splbio interrupts blocked.
11051  */
11052 static int 
11053 handle_written_inodeblock(inodedep, bp)
11054         struct inodedep *inodedep;
11055         struct buf *bp;         /* buffer containing the inode block */
11056 {
11057         struct freefile *freefile;
11058         struct allocdirect *adp, *nextadp;
11059         struct ufs1_dinode *dp1 = NULL;
11060         struct ufs2_dinode *dp2 = NULL;
11061         struct workhead wkhd;
11062         int hadchanges, fstype;
11063         ino_t freelink;
11064 
11065         LIST_INIT(&wkhd);
11066         hadchanges = 0;
11067         freefile = NULL;
11068         if ((inodedep->id_state & IOSTARTED) == 0)
11069                 panic("handle_written_inodeblock: not started");
11070         inodedep->id_state &= ~IOSTARTED;
11071         if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) {
11072                 fstype = UFS1;
11073                 dp1 = (struct ufs1_dinode *)bp->b_data +
11074                     ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
11075                 freelink = dp1->di_freelink;
11076         } else {
11077                 fstype = UFS2;
11078                 dp2 = (struct ufs2_dinode *)bp->b_data +
11079                     ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
11080                 freelink = dp2->di_freelink;
11081         }
11082         /*
11083          * Leave this inodeblock dirty until it's in the list.
11084          */
11085         if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) == UNLINKED) {
11086                 struct inodedep *inon;
11087 
11088                 inon = TAILQ_NEXT(inodedep, id_unlinked);
11089                 if ((inon == NULL && freelink == 0) ||
11090                     (inon && inon->id_ino == freelink)) {
11091                         if (inon)
11092                                 inon->id_state |= UNLINKPREV;
11093                         inodedep->id_state |= UNLINKNEXT;
11094                 }
11095                 hadchanges = 1;
11096         }
11097         /*
11098          * If we had to rollback the inode allocation because of
11099          * bitmaps being incomplete, then simply restore it.
11100          * Keep the block dirty so that it will not be reclaimed until
11101          * all associated dependencies have been cleared and the
11102          * corresponding updates written to disk.
11103          */
11104         if (inodedep->id_savedino1 != NULL) {
11105                 hadchanges = 1;
11106                 if (fstype == UFS1)
11107                         *dp1 = *inodedep->id_savedino1;
11108                 else
11109                         *dp2 = *inodedep->id_savedino2;
11110                 free(inodedep->id_savedino1, M_SAVEDINO);
11111                 inodedep->id_savedino1 = NULL;
11112                 if ((bp->b_flags & B_DELWRI) == 0)
11113                         stat_inode_bitmap++;
11114                 bdirty(bp);
11115                 /*
11116                  * If the inode is clear here and GOINGAWAY it will never
11117                  * be written.  Process the bufwait and clear any pending
11118                  * work which may include the freefile.
11119                  */
11120                 if (inodedep->id_state & GOINGAWAY)
11121                         goto bufwait;
11122                 return (1);
11123         }
11124         inodedep->id_state |= COMPLETE;
11125         /*
11126          * Roll forward anything that had to be rolled back before 
11127          * the inode could be updated.
11128          */
11129         for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
11130                 nextadp = TAILQ_NEXT(adp, ad_next);
11131                 if (adp->ad_state & ATTACHED)
11132                         panic("handle_written_inodeblock: new entry");
11133                 if (fstype == UFS1) {
11134                         if (adp->ad_offset < NDADDR) {
11135                                 if (dp1->di_db[adp->ad_offset]!=adp->ad_oldblkno)
11136                                         panic("%s %s #%jd mismatch %d != %jd",
11137                                             "handle_written_inodeblock:",
11138                                             "direct pointer",
11139                                             (intmax_t)adp->ad_offset,
11140                                             dp1->di_db[adp->ad_offset],
11141                                             (intmax_t)adp->ad_oldblkno);
11142                                 dp1->di_db[adp->ad_offset] = adp->ad_newblkno;
11143                         } else {
11144                                 if (dp1->di_ib[adp->ad_offset - NDADDR] != 0)
11145                                         panic("%s: %s #%jd allocated as %d",
11146                                             "handle_written_inodeblock",
11147                                             "indirect pointer",
11148                                             (intmax_t)adp->ad_offset - NDADDR,
11149                                             dp1->di_ib[adp->ad_offset - NDADDR]);
11150                                 dp1->di_ib[adp->ad_offset - NDADDR] =
11151                                     adp->ad_newblkno;
11152                         }
11153                 } else {
11154                         if (adp->ad_offset < NDADDR) {
11155                                 if (dp2->di_db[adp->ad_offset]!=adp->ad_oldblkno)
11156                                         panic("%s: %s #%jd %s %jd != %jd",
11157                                             "handle_written_inodeblock",
11158                                             "direct pointer",
11159                                             (intmax_t)adp->ad_offset, "mismatch",
11160                                             (intmax_t)dp2->di_db[adp->ad_offset],
11161                                             (intmax_t)adp->ad_oldblkno);
11162                                 dp2->di_db[adp->ad_offset] = adp->ad_newblkno;
11163                         } else {
11164                                 if (dp2->di_ib[adp->ad_offset - NDADDR] != 0)
11165                                         panic("%s: %s #%jd allocated as %jd",
11166                                             "handle_written_inodeblock",
11167                                             "indirect pointer",
11168                                             (intmax_t)adp->ad_offset - NDADDR,
11169                                             (intmax_t)
11170                                             dp2->di_ib[adp->ad_offset - NDADDR]);
11171                                 dp2->di_ib[adp->ad_offset - NDADDR] =
11172                                     adp->ad_newblkno;
11173                         }
11174                 }
11175                 adp->ad_state &= ~UNDONE;
11176                 adp->ad_state |= ATTACHED;
11177                 hadchanges = 1;
11178         }
11179         for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) {
11180                 nextadp = TAILQ_NEXT(adp, ad_next);
11181                 if (adp->ad_state & ATTACHED)
11182                         panic("handle_written_inodeblock: new entry");
11183                 if (dp2->di_extb[adp->ad_offset] != adp->ad_oldblkno)
11184                         panic("%s: direct pointers #%jd %s %jd != %jd",
11185                             "handle_written_inodeblock",
11186                             (intmax_t)adp->ad_offset, "mismatch",
11187                             (intmax_t)dp2->di_extb[adp->ad_offset],
11188                             (intmax_t)adp->ad_oldblkno);
11189                 dp2->di_extb[adp->ad_offset] = adp->ad_newblkno;
11190                 adp->ad_state &= ~UNDONE;
11191                 adp->ad_state |= ATTACHED;
11192                 hadchanges = 1;
11193         }
11194         if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
11195                 stat_direct_blk_ptrs++;
11196         /*
11197          * Reset the file size to its most up-to-date value.
11198          */
11199         if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1)
11200                 panic("handle_written_inodeblock: bad size");
11201         if (inodedep->id_savednlink > LINK_MAX)
11202                 panic("handle_written_inodeblock: Invalid link count "
11203                     "%d for inodedep %p", inodedep->id_savednlink, inodedep);
11204         if (fstype == UFS1) {
11205                 if (dp1->di_nlink != inodedep->id_savednlink) { 
11206                         dp1->di_nlink = inodedep->id_savednlink;
11207                         hadchanges = 1;
11208                 }
11209                 if (dp1->di_size != inodedep->id_savedsize) {
11210                         dp1->di_size = inodedep->id_savedsize;
11211                         hadchanges = 1;
11212                 }
11213         } else {
11214                 if (dp2->di_nlink != inodedep->id_savednlink) { 
11215                         dp2->di_nlink = inodedep->id_savednlink;
11216                         hadchanges = 1;
11217                 }
11218                 if (dp2->di_size != inodedep->id_savedsize) {
11219                         dp2->di_size = inodedep->id_savedsize;
11220                         hadchanges = 1;
11221                 }
11222                 if (dp2->di_extsize != inodedep->id_savedextsize) {
11223                         dp2->di_extsize = inodedep->id_savedextsize;
11224                         hadchanges = 1;
11225                 }
11226         }
11227         inodedep->id_savedsize = -1;
11228         inodedep->id_savedextsize = -1;
11229         inodedep->id_savednlink = -1;
11230         /*
11231          * If there were any rollbacks in the inode block, then it must be
11232          * marked dirty so that its will eventually get written back in
11233          * its correct form.
11234          */
11235         if (hadchanges)
11236                 bdirty(bp);
11237 bufwait:
11238         /*
11239          * Process any allocdirects that completed during the update.
11240          */
11241         if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
11242                 handle_allocdirect_partdone(adp, &wkhd);
11243         if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
11244                 handle_allocdirect_partdone(adp, &wkhd);
11245         /*
11246          * Process deallocations that were held pending until the
11247          * inode had been written to disk. Freeing of the inode
11248          * is delayed until after all blocks have been freed to
11249          * avoid creation of new <vfsid, inum, lbn> triples
11250          * before the old ones have been deleted.  Completely
11251          * unlinked inodes are not processed until the unlinked
11252          * inode list is written or the last reference is removed.
11253          */
11254         if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) != UNLINKED) {
11255                 freefile = handle_bufwait(inodedep, NULL);
11256                 if (freefile && !LIST_EMPTY(&wkhd)) {
11257                         WORKLIST_INSERT(&wkhd, &freefile->fx_list);
11258                         freefile = NULL;
11259                 }
11260         }
11261         /*
11262          * Move rolled forward dependency completions to the bufwait list
11263          * now that those that were already written have been processed.
11264          */
11265         if (!LIST_EMPTY(&wkhd) && hadchanges == 0)
11266                 panic("handle_written_inodeblock: bufwait but no changes");
11267         jwork_move(&inodedep->id_bufwait, &wkhd);
11268 
11269         if (freefile != NULL) {
11270                 /*
11271                  * If the inode is goingaway it was never written.  Fake up
11272                  * the state here so free_inodedep() can succeed.
11273                  */
11274                 if (inodedep->id_state & GOINGAWAY)
11275                         inodedep->id_state |= COMPLETE | DEPCOMPLETE;
11276                 if (free_inodedep(inodedep) == 0)
11277                         panic("handle_written_inodeblock: live inodedep %p",
11278                             inodedep);
11279                 add_to_worklist(&freefile->fx_list, 0);
11280                 return (0);
11281         }
11282 
11283         /*
11284          * If no outstanding dependencies, free it.
11285          */
11286         if (free_inodedep(inodedep) ||
11287             (TAILQ_FIRST(&inodedep->id_inoreflst) == 0 &&
11288              TAILQ_FIRST(&inodedep->id_inoupdt) == 0 &&
11289              TAILQ_FIRST(&inodedep->id_extupdt) == 0 &&
11290              LIST_FIRST(&inodedep->id_bufwait) == 0))
11291                 return (0);
11292         return (hadchanges);
11293 }
11294 
11295 static int
11296 handle_written_indirdep(indirdep, bp, bpp)
11297         struct indirdep *indirdep;
11298         struct buf *bp;
11299         struct buf **bpp;
11300 {
11301         struct allocindir *aip;
11302         struct buf *sbp;
11303         int chgs;
11304 
11305         if (indirdep->ir_state & GOINGAWAY)
11306                 panic("handle_written_indirdep: indirdep gone");
11307         if ((indirdep->ir_state & IOSTARTED) == 0)
11308                 panic("handle_written_indirdep: IO not started");
11309         chgs = 0;
11310         /*
11311          * If there were rollbacks revert them here.
11312          */
11313         if (indirdep->ir_saveddata) {
11314                 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
11315                 if (TAILQ_EMPTY(&indirdep->ir_trunc)) {
11316                         free(indirdep->ir_saveddata, M_INDIRDEP);
11317                         indirdep->ir_saveddata = NULL;
11318                 }
11319                 chgs = 1;
11320         }
11321         indirdep->ir_state &= ~(UNDONE | IOSTARTED);
11322         indirdep->ir_state |= ATTACHED;
11323         /*
11324          * Move allocindirs with written pointers to the completehd if
11325          * the indirdep's pointer is not yet written.  Otherwise
11326          * free them here.
11327          */
11328         while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != 0) {
11329                 LIST_REMOVE(aip, ai_next);
11330                 if ((indirdep->ir_state & DEPCOMPLETE) == 0) {
11331                         LIST_INSERT_HEAD(&indirdep->ir_completehd, aip,
11332                             ai_next);
11333                         newblk_freefrag(&aip->ai_block);
11334                         continue;
11335                 }
11336                 free_newblk(&aip->ai_block);
11337         }
11338         /*
11339          * Move allocindirs that have finished dependency processing from
11340          * the done list to the write list after updating the pointers.
11341          */
11342         if (TAILQ_EMPTY(&indirdep->ir_trunc)) {
11343                 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
11344                         handle_allocindir_partdone(aip);
11345                         if (aip == LIST_FIRST(&indirdep->ir_donehd))
11346                                 panic("disk_write_complete: not gone");
11347                         chgs = 1;
11348                 }
11349         }
11350         /*
11351          * Preserve the indirdep if there were any changes or if it is not
11352          * yet valid on disk.
11353          */
11354         if (chgs) {
11355                 stat_indir_blk_ptrs++;
11356                 bdirty(bp);
11357                 return (1);
11358         }
11359         /*
11360          * If there were no changes we can discard the savedbp and detach
11361          * ourselves from the buf.  We are only carrying completed pointers
11362          * in this case.
11363          */
11364         sbp = indirdep->ir_savebp;
11365         sbp->b_flags |= B_INVAL | B_NOCACHE;
11366         indirdep->ir_savebp = NULL;
11367         indirdep->ir_bp = NULL;
11368         if (*bpp != NULL)
11369                 panic("handle_written_indirdep: bp already exists.");
11370         *bpp = sbp;
11371         /*
11372          * The indirdep may not be freed until its parent points at it.
11373          */
11374         if (indirdep->ir_state & DEPCOMPLETE)
11375                 free_indirdep(indirdep);
11376 
11377         return (0);
11378 }
11379 
11380 /*
11381  * Process a diradd entry after its dependent inode has been written.
11382  * This routine must be called with splbio interrupts blocked.
11383  */
11384 static void
11385 diradd_inode_written(dap, inodedep)
11386         struct diradd *dap;
11387         struct inodedep *inodedep;
11388 {
11389 
11390         dap->da_state |= COMPLETE;
11391         complete_diradd(dap);
11392         WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
11393 }
11394 
11395 /*
11396  * Returns true if the bmsafemap will have rollbacks when written.  Must
11397  * only be called with lk and the buf lock on the cg held.
11398  */
11399 static int
11400 bmsafemap_backgroundwrite(bmsafemap, bp)
11401         struct bmsafemap *bmsafemap;
11402         struct buf *bp;
11403 {
11404         int dirty;
11405 
11406         dirty = !LIST_EMPTY(&bmsafemap->sm_jaddrefhd) | 
11407             !LIST_EMPTY(&bmsafemap->sm_jnewblkhd);
11408         /*
11409          * If we're initiating a background write we need to process the
11410          * rollbacks as they exist now, not as they exist when IO starts.
11411          * No other consumers will look at the contents of the shadowed
11412          * buf so this is safe to do here.
11413          */
11414         if (bp->b_xflags & BX_BKGRDMARKER)
11415                 initiate_write_bmsafemap(bmsafemap, bp);
11416 
11417         return (dirty);
11418 }
11419 
11420 /*
11421  * Re-apply an allocation when a cg write is complete.
11422  */
11423 static int
11424 jnewblk_rollforward(jnewblk, fs, cgp, blksfree)
11425         struct jnewblk *jnewblk;
11426         struct fs *fs;
11427         struct cg *cgp;
11428         uint8_t *blksfree;
11429 {
11430         ufs1_daddr_t fragno;
11431         ufs2_daddr_t blkno;
11432         long cgbno, bbase;
11433         int frags, blk;
11434         int i;
11435 
11436         frags = 0;
11437         cgbno = dtogd(fs, jnewblk->jn_blkno);
11438         for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) {
11439                 if (isclr(blksfree, cgbno + i))
11440                         panic("jnewblk_rollforward: re-allocated fragment");
11441                 frags++;
11442         }
11443         if (frags == fs->fs_frag) {
11444                 blkno = fragstoblks(fs, cgbno);
11445                 ffs_clrblock(fs, blksfree, (long)blkno);
11446                 ffs_clusteracct(fs, cgp, blkno, -1);
11447                 cgp->cg_cs.cs_nbfree--;
11448         } else {
11449                 bbase = cgbno - fragnum(fs, cgbno);
11450                 cgbno += jnewblk->jn_oldfrags;
11451                 /* If a complete block had been reassembled, account for it. */
11452                 fragno = fragstoblks(fs, bbase);
11453                 if (ffs_isblock(fs, blksfree, fragno)) {
11454                         cgp->cg_cs.cs_nffree += fs->fs_frag;
11455                         ffs_clusteracct(fs, cgp, fragno, -1);
11456                         cgp->cg_cs.cs_nbfree--;
11457                 }
11458                 /* Decrement the old frags.  */
11459                 blk = blkmap(fs, blksfree, bbase);
11460                 ffs_fragacct(fs, blk, cgp->cg_frsum, -1);
11461                 /* Allocate the fragment */
11462                 for (i = 0; i < frags; i++)
11463                         clrbit(blksfree, cgbno + i);
11464                 cgp->cg_cs.cs_nffree -= frags;
11465                 /* Add back in counts associated with the new frags */
11466                 blk = blkmap(fs, blksfree, bbase);
11467                 ffs_fragacct(fs, blk, cgp->cg_frsum, 1);
11468         }
11469         return (frags);
11470 }
11471 
11472 /*
11473  * Complete a write to a bmsafemap structure.  Roll forward any bitmap
11474  * changes if it's not a background write.  Set all written dependencies 
11475  * to DEPCOMPLETE and free the structure if possible.
11476  */
11477 static int
11478 handle_written_bmsafemap(bmsafemap, bp)
11479         struct bmsafemap *bmsafemap;
11480         struct buf *bp;
11481 {
11482         struct newblk *newblk;
11483         struct inodedep *inodedep;
11484         struct jaddref *jaddref, *jatmp;
11485         struct jnewblk *jnewblk, *jntmp;
11486         struct ufsmount *ump;
11487         uint8_t *inosused;
11488         uint8_t *blksfree;
11489         struct cg *cgp;
11490         struct fs *fs;
11491         ino_t ino;
11492         int foreground;
11493         int chgs;
11494 
11495         if ((bmsafemap->sm_state & IOSTARTED) == 0)
11496                 panic("initiate_write_bmsafemap: Not started\n");
11497         ump = VFSTOUFS(bmsafemap->sm_list.wk_mp);
11498         chgs = 0;
11499         bmsafemap->sm_state &= ~IOSTARTED;
11500         foreground = (bp->b_xflags & BX_BKGRDMARKER) == 0;
11501         /*
11502          * Release journal work that was waiting on the write.
11503          */
11504         handle_jwork(&bmsafemap->sm_freewr);
11505 
11506         /*
11507          * Restore unwritten inode allocation pending jaddref writes.
11508          */
11509         if (!LIST_EMPTY(&bmsafemap->sm_jaddrefhd)) {
11510                 cgp = (struct cg *)bp->b_data;
11511                 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
11512                 inosused = cg_inosused(cgp);
11513                 LIST_FOREACH_SAFE(jaddref, &bmsafemap->sm_jaddrefhd,
11514                     ja_bmdeps, jatmp) {
11515                         if ((jaddref->ja_state & UNDONE) == 0)
11516                                 continue;
11517                         ino = jaddref->ja_ino % fs->fs_ipg;
11518                         if (isset(inosused, ino))
11519                                 panic("handle_written_bmsafemap: "
11520                                     "re-allocated inode");
11521                         /* Do the roll-forward only if it's a real copy. */
11522                         if (foreground) {
11523                                 if ((jaddref->ja_mode & IFMT) == IFDIR)
11524                                         cgp->cg_cs.cs_ndir++;
11525                                 cgp->cg_cs.cs_nifree--;
11526                                 setbit(inosused, ino);
11527                                 chgs = 1;
11528                         }
11529                         jaddref->ja_state &= ~UNDONE;
11530                         jaddref->ja_state |= ATTACHED;
11531                         free_jaddref(jaddref);
11532                 }
11533         }
11534         /*
11535          * Restore any block allocations which are pending journal writes.
11536          */
11537         if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) {
11538                 cgp = (struct cg *)bp->b_data;
11539                 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
11540                 blksfree = cg_blksfree(cgp);
11541                 LIST_FOREACH_SAFE(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps,
11542                     jntmp) {
11543                         if ((jnewblk->jn_state & UNDONE) == 0)
11544                                 continue;
11545                         /* Do the roll-forward only if it's a real copy. */
11546                         if (foreground &&
11547                             jnewblk_rollforward(jnewblk, fs, cgp, blksfree))
11548                                 chgs = 1;
11549                         jnewblk->jn_state &= ~(UNDONE | NEWBLOCK);
11550                         jnewblk->jn_state |= ATTACHED;
11551                         free_jnewblk(jnewblk);
11552                 }
11553         }
11554         while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkwr))) {
11555                 newblk->nb_state |= DEPCOMPLETE;
11556                 newblk->nb_state &= ~ONDEPLIST;
11557                 newblk->nb_bmsafemap = NULL;
11558                 LIST_REMOVE(newblk, nb_deps);
11559                 if (newblk->nb_list.wk_type == D_ALLOCDIRECT)
11560                         handle_allocdirect_partdone(
11561                             WK_ALLOCDIRECT(&newblk->nb_list), NULL);
11562                 else if (newblk->nb_list.wk_type == D_ALLOCINDIR)
11563                         handle_allocindir_partdone(
11564                             WK_ALLOCINDIR(&newblk->nb_list));
11565                 else if (newblk->nb_list.wk_type != D_NEWBLK)
11566                         panic("handle_written_bmsafemap: Unexpected type: %s",
11567                             TYPENAME(newblk->nb_list.wk_type));
11568         }
11569         while ((inodedep = LIST_FIRST(&bmsafemap->sm_inodedepwr)) != NULL) {
11570                 inodedep->id_state |= DEPCOMPLETE;
11571                 inodedep->id_state &= ~ONDEPLIST;
11572                 LIST_REMOVE(inodedep, id_deps);
11573                 inodedep->id_bmsafemap = NULL;
11574         }
11575         LIST_REMOVE(bmsafemap, sm_next);
11576         if (chgs == 0 && LIST_EMPTY(&bmsafemap->sm_jaddrefhd) &&
11577             LIST_EMPTY(&bmsafemap->sm_jnewblkhd) &&
11578             LIST_EMPTY(&bmsafemap->sm_newblkhd) &&
11579             LIST_EMPTY(&bmsafemap->sm_inodedephd) &&
11580             LIST_EMPTY(&bmsafemap->sm_freehd)) {
11581                 LIST_REMOVE(bmsafemap, sm_hash);
11582                 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
11583                 return (0);
11584         }
11585         LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next);
11586         if (foreground)
11587                 bdirty(bp);
11588         return (1);
11589 }
11590 
11591 /*
11592  * Try to free a mkdir dependency.
11593  */
11594 static void
11595 complete_mkdir(mkdir)
11596         struct mkdir *mkdir;
11597 {
11598         struct diradd *dap;
11599 
11600         if ((mkdir->md_state & ALLCOMPLETE) != ALLCOMPLETE)
11601                 return;
11602         LIST_REMOVE(mkdir, md_mkdirs);
11603         dap = mkdir->md_diradd;
11604         dap->da_state &= ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY));
11605         if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) {
11606                 dap->da_state |= DEPCOMPLETE;
11607                 complete_diradd(dap);
11608         }
11609         WORKITEM_FREE(mkdir, D_MKDIR);
11610 }
11611 
11612 /*
11613  * Handle the completion of a mkdir dependency.
11614  */
11615 static void
11616 handle_written_mkdir(mkdir, type)
11617         struct mkdir *mkdir;
11618         int type;
11619 {
11620 
11621         if ((mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)) != type)
11622                 panic("handle_written_mkdir: bad type");
11623         mkdir->md_state |= COMPLETE;
11624         complete_mkdir(mkdir);
11625 }
11626 
11627 static int
11628 free_pagedep(pagedep)
11629         struct pagedep *pagedep;
11630 {
11631         int i;
11632 
11633         if (pagedep->pd_state & NEWBLOCK)
11634                 return (0);
11635         if (!LIST_EMPTY(&pagedep->pd_dirremhd))
11636                 return (0);
11637         for (i = 0; i < DAHASHSZ; i++)
11638                 if (!LIST_EMPTY(&pagedep->pd_diraddhd[i]))
11639                         return (0);
11640         if (!LIST_EMPTY(&pagedep->pd_pendinghd))
11641                 return (0);
11642         if (!LIST_EMPTY(&pagedep->pd_jmvrefhd))
11643                 return (0);
11644         if (pagedep->pd_state & ONWORKLIST)
11645                 WORKLIST_REMOVE(&pagedep->pd_list);
11646         LIST_REMOVE(pagedep, pd_hash);
11647         WORKITEM_FREE(pagedep, D_PAGEDEP);
11648 
11649         return (1);
11650 }
11651 
11652 /*
11653  * Called from within softdep_disk_write_complete above.
11654  * A write operation was just completed. Removed inodes can
11655  * now be freed and associated block pointers may be committed.
11656  * Note that this routine is always called from interrupt level
11657  * with further splbio interrupts blocked.
11658  */
11659 static int 
11660 handle_written_filepage(pagedep, bp)
11661         struct pagedep *pagedep;
11662         struct buf *bp;         /* buffer containing the written page */
11663 {
11664         struct dirrem *dirrem;
11665         struct diradd *dap, *nextdap;
11666         struct direct *ep;
11667         int i, chgs;
11668 
11669         if ((pagedep->pd_state & IOSTARTED) == 0)
11670                 panic("handle_written_filepage: not started");
11671         pagedep->pd_state &= ~IOSTARTED;
11672         /*
11673          * Process any directory removals that have been committed.
11674          */
11675         while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
11676                 LIST_REMOVE(dirrem, dm_next);
11677                 dirrem->dm_state |= COMPLETE;
11678                 dirrem->dm_dirinum = pagedep->pd_ino;
11679                 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd),
11680                     ("handle_written_filepage: Journal entries not written."));
11681                 add_to_worklist(&dirrem->dm_list, 0);
11682         }
11683         /*
11684          * Free any directory additions that have been committed.
11685          * If it is a newly allocated block, we have to wait until
11686          * the on-disk directory inode claims the new block.
11687          */
11688         if ((pagedep->pd_state & NEWBLOCK) == 0)
11689                 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
11690                         free_diradd(dap, NULL);
11691         /*
11692          * Uncommitted directory entries must be restored.
11693          */
11694         for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
11695                 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
11696                      dap = nextdap) {
11697                         nextdap = LIST_NEXT(dap, da_pdlist);
11698                         if (dap->da_state & ATTACHED)
11699                                 panic("handle_written_filepage: attached");
11700                         ep = (struct direct *)
11701                             ((char *)bp->b_data + dap->da_offset);
11702                         ep->d_ino = dap->da_newinum;
11703                         dap->da_state &= ~UNDONE;
11704                         dap->da_state |= ATTACHED;
11705                         chgs = 1;
11706                         /*
11707                          * If the inode referenced by the directory has
11708                          * been written out, then the dependency can be
11709                          * moved to the pending list.
11710                          */
11711                         if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
11712                                 LIST_REMOVE(dap, da_pdlist);
11713                                 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
11714                                     da_pdlist);
11715                         }
11716                 }
11717         }
11718         /*
11719          * If there were any rollbacks in the directory, then it must be
11720          * marked dirty so that its will eventually get written back in
11721          * its correct form.
11722          */
11723         if (chgs) {
11724                 if ((bp->b_flags & B_DELWRI) == 0)
11725                         stat_dir_entry++;
11726                 bdirty(bp);
11727                 return (1);
11728         }
11729         /*
11730          * If we are not waiting for a new directory block to be
11731          * claimed by its inode, then the pagedep will be freed.
11732          * Otherwise it will remain to track any new entries on
11733          * the page in case they are fsync'ed.
11734          */
11735         free_pagedep(pagedep);
11736         return (0);
11737 }
11738 
11739 /*
11740  * Writing back in-core inode structures.
11741  * 
11742  * The filesystem only accesses an inode's contents when it occupies an
11743  * "in-core" inode structure.  These "in-core" structures are separate from
11744  * the page frames used to cache inode blocks.  Only the latter are
11745  * transferred to/from the disk.  So, when the updated contents of the
11746  * "in-core" inode structure are copied to the corresponding in-memory inode
11747  * block, the dependencies are also transferred.  The following procedure is
11748  * called when copying a dirty "in-core" inode to a cached inode block.
11749  */
11750 
11751 /*
11752  * Called when an inode is loaded from disk. If the effective link count
11753  * differed from the actual link count when it was last flushed, then we
11754  * need to ensure that the correct effective link count is put back.
11755  */
11756 void 
11757 softdep_load_inodeblock(ip)
11758         struct inode *ip;       /* the "in_core" copy of the inode */
11759 {
11760         struct inodedep *inodedep;
11761 
11762         /*
11763          * Check for alternate nlink count.
11764          */
11765         ip->i_effnlink = ip->i_nlink;
11766         ACQUIRE_LOCK(&lk);
11767         if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0,
11768             &inodedep) == 0) {
11769                 FREE_LOCK(&lk);
11770                 return;
11771         }
11772         ip->i_effnlink -= inodedep->id_nlinkdelta;
11773         FREE_LOCK(&lk);
11774 }
11775 
11776 /*
11777  * This routine is called just before the "in-core" inode
11778  * information is to be copied to the in-memory inode block.
11779  * Recall that an inode block contains several inodes. If
11780  * the force flag is set, then the dependencies will be
11781  * cleared so that the update can always be made. Note that
11782  * the buffer is locked when this routine is called, so we
11783  * will never be in the middle of writing the inode block 
11784  * to disk.
11785  */
11786 void 
11787 softdep_update_inodeblock(ip, bp, waitfor)
11788         struct inode *ip;       /* the "in_core" copy of the inode */
11789         struct buf *bp;         /* the buffer containing the inode block */
11790         int waitfor;            /* nonzero => update must be allowed */
11791 {
11792         struct inodedep *inodedep;
11793         struct inoref *inoref;
11794         struct worklist *wk;
11795         struct mount *mp;
11796         struct buf *ibp;
11797         struct fs *fs;
11798         int error;
11799 
11800         mp = UFSTOVFS(ip->i_ump);
11801         fs = ip->i_fs;
11802         /*
11803          * Preserve the freelink that is on disk.  clear_unlinked_inodedep()
11804          * does not have access to the in-core ip so must write directly into
11805          * the inode block buffer when setting freelink.
11806          */
11807         if (fs->fs_magic == FS_UFS1_MAGIC)
11808                 DIP_SET(ip, i_freelink, ((struct ufs1_dinode *)bp->b_data +
11809                     ino_to_fsbo(fs, ip->i_number))->di_freelink);
11810         else
11811                 DIP_SET(ip, i_freelink, ((struct ufs2_dinode *)bp->b_data +
11812                     ino_to_fsbo(fs, ip->i_number))->di_freelink);
11813         /*
11814          * If the effective link count is not equal to the actual link
11815          * count, then we must track the difference in an inodedep while
11816          * the inode is (potentially) tossed out of the cache. Otherwise,
11817          * if there is no existing inodedep, then there are no dependencies
11818          * to track.
11819          */
11820         ACQUIRE_LOCK(&lk);
11821 again:
11822         if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) {
11823                 FREE_LOCK(&lk);
11824                 if (ip->i_effnlink != ip->i_nlink)
11825                         panic("softdep_update_inodeblock: bad link count");
11826                 return;
11827         }
11828         if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink)
11829                 panic("softdep_update_inodeblock: bad delta");
11830         /*
11831          * If we're flushing all dependencies we must also move any waiting
11832          * for journal writes onto the bufwait list prior to I/O.
11833          */
11834         if (waitfor) {
11835                 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
11836                         if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
11837                             == DEPCOMPLETE) {
11838                                 jwait(&inoref->if_list, MNT_WAIT);
11839                                 goto again;
11840                         }
11841                 }
11842         }
11843         /*
11844          * Changes have been initiated. Anything depending on these
11845          * changes cannot occur until this inode has been written.
11846          */
11847         inodedep->id_state &= ~COMPLETE;
11848         if ((inodedep->id_state & ONWORKLIST) == 0)
11849                 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
11850         /*
11851          * Any new dependencies associated with the incore inode must 
11852          * now be moved to the list associated with the buffer holding
11853          * the in-memory copy of the inode. Once merged process any
11854          * allocdirects that are completed by the merger.
11855          */
11856         merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt);
11857         if (!TAILQ_EMPTY(&inodedep->id_inoupdt))
11858                 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt),
11859                     NULL);
11860         merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt);
11861         if (!TAILQ_EMPTY(&inodedep->id_extupdt))
11862                 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt),
11863                     NULL);
11864         /*
11865          * Now that the inode has been pushed into the buffer, the
11866          * operations dependent on the inode being written to disk
11867          * can be moved to the id_bufwait so that they will be
11868          * processed when the buffer I/O completes.
11869          */
11870         while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
11871                 WORKLIST_REMOVE(wk);
11872                 WORKLIST_INSERT(&inodedep->id_bufwait, wk);
11873         }
11874         /*
11875          * Newly allocated inodes cannot be written until the bitmap
11876          * that allocates them have been written (indicated by
11877          * DEPCOMPLETE being set in id_state). If we are doing a
11878          * forced sync (e.g., an fsync on a file), we force the bitmap
11879          * to be written so that the update can be done.
11880          */
11881         if (waitfor == 0) {
11882                 FREE_LOCK(&lk);
11883                 return;
11884         }
11885 retry:
11886         if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) != 0) {
11887                 FREE_LOCK(&lk);
11888                 return;
11889         }
11890         ibp = inodedep->id_bmsafemap->sm_buf;
11891         ibp = getdirtybuf(ibp, &lk, MNT_WAIT);
11892         if (ibp == NULL) {
11893                 /*
11894                  * If ibp came back as NULL, the dependency could have been
11895                  * freed while we slept.  Look it up again, and check to see
11896                  * that it has completed.
11897                  */
11898                 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0)
11899                         goto retry;
11900                 FREE_LOCK(&lk);
11901                 return;
11902         }
11903         FREE_LOCK(&lk);
11904         if ((error = bwrite(ibp)) != 0)
11905                 softdep_error("softdep_update_inodeblock: bwrite", error);
11906 }
11907 
11908 /*
11909  * Merge the a new inode dependency list (such as id_newinoupdt) into an
11910  * old inode dependency list (such as id_inoupdt). This routine must be
11911  * called with splbio interrupts blocked.
11912  */
11913 static void
11914 merge_inode_lists(newlisthead, oldlisthead)
11915         struct allocdirectlst *newlisthead;
11916         struct allocdirectlst *oldlisthead;
11917 {
11918         struct allocdirect *listadp, *newadp;
11919 
11920         newadp = TAILQ_FIRST(newlisthead);
11921         for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) {
11922                 if (listadp->ad_offset < newadp->ad_offset) {
11923                         listadp = TAILQ_NEXT(listadp, ad_next);
11924                         continue;
11925                 }
11926                 TAILQ_REMOVE(newlisthead, newadp, ad_next);
11927                 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
11928                 if (listadp->ad_offset == newadp->ad_offset) {
11929                         allocdirect_merge(oldlisthead, newadp,
11930                             listadp);
11931                         listadp = newadp;
11932                 }
11933                 newadp = TAILQ_FIRST(newlisthead);
11934         }
11935         while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) {
11936                 TAILQ_REMOVE(newlisthead, newadp, ad_next);
11937                 TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next);
11938         }
11939 }
11940 
11941 /*
11942  * If we are doing an fsync, then we must ensure that any directory
11943  * entries for the inode have been written after the inode gets to disk.
11944  */
11945 int
11946 softdep_fsync(vp)
11947         struct vnode *vp;       /* the "in_core" copy of the inode */
11948 {
11949         struct inodedep *inodedep;
11950         struct pagedep *pagedep;
11951         struct inoref *inoref;
11952         struct worklist *wk;
11953         struct diradd *dap;
11954         struct mount *mp;
11955         struct vnode *pvp;
11956         struct inode *ip;
11957         struct buf *bp;
11958         struct fs *fs;
11959         struct thread *td = curthread;
11960         int error, flushparent, pagedep_new_block;
11961         ino_t parentino;
11962         ufs_lbn_t lbn;
11963 
11964         ip = VTOI(vp);
11965         fs = ip->i_fs;
11966         mp = vp->v_mount;
11967         ACQUIRE_LOCK(&lk);
11968 restart:
11969         if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) {
11970                 FREE_LOCK(&lk);
11971                 return (0);
11972         }
11973         TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
11974                 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
11975                     == DEPCOMPLETE) {
11976                         jwait(&inoref->if_list, MNT_WAIT);
11977                         goto restart;
11978                 }
11979         }
11980         if (!LIST_EMPTY(&inodedep->id_inowait) ||
11981             !TAILQ_EMPTY(&inodedep->id_extupdt) ||
11982             !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
11983             !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
11984             !TAILQ_EMPTY(&inodedep->id_newinoupdt))
11985                 panic("softdep_fsync: pending ops %p", inodedep);
11986         for (error = 0, flushparent = 0; ; ) {
11987                 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
11988                         break;
11989                 if (wk->wk_type != D_DIRADD)
11990                         panic("softdep_fsync: Unexpected type %s",
11991                             TYPENAME(wk->wk_type));
11992                 dap = WK_DIRADD(wk);
11993                 /*
11994                  * Flush our parent if this directory entry has a MKDIR_PARENT
11995                  * dependency or is contained in a newly allocated block.
11996                  */
11997                 if (dap->da_state & DIRCHG)
11998                         pagedep = dap->da_previous->dm_pagedep;
11999                 else
12000                         pagedep = dap->da_pagedep;
12001                 parentino = pagedep->pd_ino;
12002                 lbn = pagedep->pd_lbn;
12003                 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE)
12004                         panic("softdep_fsync: dirty");
12005                 if ((dap->da_state & MKDIR_PARENT) ||
12006                     (pagedep->pd_state & NEWBLOCK))
12007                         flushparent = 1;
12008                 else
12009                         flushparent = 0;
12010                 /*
12011                  * If we are being fsync'ed as part of vgone'ing this vnode,
12012                  * then we will not be able to release and recover the
12013                  * vnode below, so we just have to give up on writing its
12014                  * directory entry out. It will eventually be written, just
12015                  * not now, but then the user was not asking to have it
12016                  * written, so we are not breaking any promises.
12017                  */
12018                 if (vp->v_iflag & VI_DOOMED)
12019                         break;
12020                 /*
12021                  * We prevent deadlock by always fetching inodes from the
12022                  * root, moving down the directory tree. Thus, when fetching
12023                  * our parent directory, we first try to get the lock. If
12024                  * that fails, we must unlock ourselves before requesting
12025                  * the lock on our parent. See the comment in ufs_lookup
12026                  * for details on possible races.
12027                  */
12028                 FREE_LOCK(&lk);
12029                 if (ffs_vgetf(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp,
12030                     FFSV_FORCEINSMQ)) {
12031                         error = vfs_busy(mp, MBF_NOWAIT);
12032                         if (error != 0) {
12033                                 vfs_ref(mp);
12034                                 VOP_UNLOCK(vp, 0);
12035                                 error = vfs_busy(mp, 0);
12036                                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
12037                                 vfs_rel(mp);
12038                                 if (error != 0)
12039                                         return (ENOENT);
12040                                 if (vp->v_iflag & VI_DOOMED) {
12041                                         vfs_unbusy(mp);
12042                                         return (ENOENT);
12043                                 }
12044                         }
12045                         VOP_UNLOCK(vp, 0);
12046                         error = ffs_vgetf(mp, parentino, LK_EXCLUSIVE,
12047                             &pvp, FFSV_FORCEINSMQ);
12048                         vfs_unbusy(mp);
12049                         vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
12050                         if (vp->v_iflag & VI_DOOMED) {
12051                                 if (error == 0)
12052                                         vput(pvp);
12053                                 error = ENOENT;
12054                         }
12055                         if (error != 0)
12056                                 return (error);
12057                 }
12058                 /*
12059                  * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps
12060                  * that are contained in direct blocks will be resolved by 
12061                  * doing a ffs_update. Pagedeps contained in indirect blocks
12062                  * may require a complete sync'ing of the directory. So, we
12063                  * try the cheap and fast ffs_update first, and if that fails,
12064                  * then we do the slower ffs_syncvnode of the directory.
12065                  */
12066                 if (flushparent) {
12067                         int locked;
12068 
12069                         if ((error = ffs_update(pvp, 1)) != 0) {
12070                                 vput(pvp);
12071                                 return (error);
12072                         }
12073                         ACQUIRE_LOCK(&lk);
12074                         locked = 1;
12075                         if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) {
12076                                 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) {
12077                                         if (wk->wk_type != D_DIRADD)
12078                                                 panic("softdep_fsync: Unexpected type %s",
12079                                                       TYPENAME(wk->wk_type));
12080                                         dap = WK_DIRADD(wk);
12081                                         if (dap->da_state & DIRCHG)
12082                                                 pagedep = dap->da_previous->dm_pagedep;
12083                                         else
12084                                                 pagedep = dap->da_pagedep;
12085                                         pagedep_new_block = pagedep->pd_state & NEWBLOCK;
12086                                         FREE_LOCK(&lk);
12087                                         locked = 0;
12088                                         if (pagedep_new_block && (error =
12089                                             ffs_syncvnode(pvp, MNT_WAIT, 0))) {
12090                                                 vput(pvp);
12091                                                 return (error);
12092                                         }
12093                                 }
12094                         }
12095                         if (locked)
12096                                 FREE_LOCK(&lk);
12097                 }
12098                 /*
12099                  * Flush directory page containing the inode's name.
12100                  */
12101                 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred,
12102                     &bp);
12103                 if (error == 0)
12104                         error = bwrite(bp);
12105                 else
12106                         brelse(bp);
12107                 vput(pvp);
12108                 if (error != 0)
12109                         return (error);
12110                 ACQUIRE_LOCK(&lk);
12111                 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0)
12112                         break;
12113         }
12114         FREE_LOCK(&lk);
12115         return (0);
12116 }
12117 
12118 /*
12119  * Flush all the dirty bitmaps associated with the block device
12120  * before flushing the rest of the dirty blocks so as to reduce
12121  * the number of dependencies that will have to be rolled back.
12122  *
12123  * XXX Unused?
12124  */
12125 void
12126 softdep_fsync_mountdev(vp)
12127         struct vnode *vp;
12128 {
12129         struct buf *bp, *nbp;
12130         struct worklist *wk;
12131         struct bufobj *bo;
12132 
12133         if (!vn_isdisk(vp, NULL))
12134                 panic("softdep_fsync_mountdev: vnode not a disk");
12135         bo = &vp->v_bufobj;
12136 restart:
12137         BO_LOCK(bo);
12138         ACQUIRE_LOCK(&lk);
12139         TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
12140                 /* 
12141                  * If it is already scheduled, skip to the next buffer.
12142                  */
12143                 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL))
12144                         continue;
12145 
12146                 if ((bp->b_flags & B_DELWRI) == 0)
12147                         panic("softdep_fsync_mountdev: not dirty");
12148                 /*
12149                  * We are only interested in bitmaps with outstanding
12150                  * dependencies.
12151                  */
12152                 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
12153                     wk->wk_type != D_BMSAFEMAP ||
12154                     (bp->b_vflags & BV_BKGRDINPROG)) {
12155                         BUF_UNLOCK(bp);
12156                         continue;
12157                 }
12158                 FREE_LOCK(&lk);
12159                 BO_UNLOCK(bo);
12160                 bremfree(bp);
12161                 (void) bawrite(bp);
12162                 goto restart;
12163         }
12164         FREE_LOCK(&lk);
12165         drain_output(vp);
12166         BO_UNLOCK(bo);
12167 }
12168 
12169 /*
12170  * Sync all cylinder groups that were dirty at the time this function is
12171  * called.  Newly dirtied cgs will be inserted before the sentinel.  This
12172  * is used to flush freedep activity that may be holding up writes to a
12173  * indirect block.
12174  */
12175 static int
12176 sync_cgs(mp, waitfor)
12177         struct mount *mp;
12178         int waitfor;
12179 {
12180         struct bmsafemap *bmsafemap;
12181         struct bmsafemap *sentinel;
12182         struct ufsmount *ump;
12183         struct buf *bp;
12184         int error;
12185 
12186         sentinel = malloc(sizeof(*sentinel), M_BMSAFEMAP, M_ZERO | M_WAITOK);
12187         sentinel->sm_cg = -1;
12188         ump = VFSTOUFS(mp);
12189         error = 0;
12190         ACQUIRE_LOCK(&lk);
12191         LIST_INSERT_HEAD(&ump->softdep_dirtycg, sentinel, sm_next);
12192         for (bmsafemap = LIST_NEXT(sentinel, sm_next); bmsafemap != NULL;
12193             bmsafemap = LIST_NEXT(sentinel, sm_next)) {
12194                 /* Skip sentinels and cgs with no work to release. */
12195                 if (bmsafemap->sm_cg == -1 ||
12196                     (LIST_EMPTY(&bmsafemap->sm_freehd) &&
12197                     LIST_EMPTY(&bmsafemap->sm_freewr))) {
12198                         LIST_REMOVE(sentinel, sm_next);
12199                         LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next);
12200                         continue;
12201                 }
12202                 /*
12203                  * If we don't get the lock and we're waiting try again, if
12204                  * not move on to the next buf and try to sync it.
12205                  */
12206                 bp = getdirtybuf(bmsafemap->sm_buf, &lk, waitfor);
12207                 if (bp == NULL && waitfor == MNT_WAIT)
12208                         continue;
12209                 LIST_REMOVE(sentinel, sm_next);
12210                 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next);
12211                 if (bp == NULL)
12212                         continue;
12213                 FREE_LOCK(&lk);
12214                 if (waitfor == MNT_NOWAIT)
12215                         bawrite(bp);
12216                 else
12217                         error = bwrite(bp);
12218                 ACQUIRE_LOCK(&lk);
12219                 if (error)
12220                         break;
12221         }
12222         LIST_REMOVE(sentinel, sm_next);
12223         FREE_LOCK(&lk);
12224         free(sentinel, M_BMSAFEMAP);
12225         return (error);
12226 }
12227 
12228 /*
12229  * This routine is called when we are trying to synchronously flush a
12230  * file. This routine must eliminate any filesystem metadata dependencies
12231  * so that the syncing routine can succeed.
12232  */
12233 int
12234 softdep_sync_metadata(struct vnode *vp)
12235 {
12236         int error;
12237 
12238         /*
12239          * Ensure that any direct block dependencies have been cleared,
12240          * truncations are started, and inode references are journaled.
12241          */
12242         ACQUIRE_LOCK(&lk);
12243         /*
12244          * Write all journal records to prevent rollbacks on devvp.
12245          */
12246         if (vp->v_type == VCHR)
12247                 softdep_flushjournal(vp->v_mount);
12248         error = flush_inodedep_deps(vp, vp->v_mount, VTOI(vp)->i_number);
12249         /*
12250          * Ensure that all truncates are written so we won't find deps on
12251          * indirect blocks.
12252          */
12253         process_truncates(vp);
12254         FREE_LOCK(&lk);
12255 
12256         return (error);
12257 }
12258 
12259 /*
12260  * This routine is called when we are attempting to sync a buf with
12261  * dependencies.  If waitfor is MNT_NOWAIT it attempts to schedule any
12262  * other IO it can but returns EBUSY if the buffer is not yet able to
12263  * be written.  Dependencies which will not cause rollbacks will always
12264  * return 0.
12265  */
12266 int
12267 softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor)
12268 {
12269         struct indirdep *indirdep;
12270         struct pagedep *pagedep;
12271         struct allocindir *aip;
12272         struct newblk *newblk;
12273         struct buf *nbp;
12274         struct worklist *wk;
12275         int i, error;
12276 
12277         /*
12278          * For VCHR we just don't want to force flush any dependencies that
12279          * will cause rollbacks.
12280          */
12281         if (vp->v_type == VCHR) {
12282                 if (waitfor == MNT_NOWAIT && softdep_count_dependencies(bp, 0))
12283                         return (EBUSY);
12284                 return (0);
12285         }
12286         ACQUIRE_LOCK(&lk);
12287         /*
12288          * As we hold the buffer locked, none of its dependencies
12289          * will disappear.
12290          */
12291         error = 0;
12292 top:
12293         LIST_FOREACH(wk, &bp->b_dep, wk_list) {
12294                 switch (wk->wk_type) {
12295 
12296                 case D_ALLOCDIRECT:
12297                 case D_ALLOCINDIR:
12298                         newblk = WK_NEWBLK(wk);
12299                         if (newblk->nb_jnewblk != NULL) {
12300                                 if (waitfor == MNT_NOWAIT) {
12301                                         error = EBUSY;
12302                                         goto out_unlock;
12303                                 }
12304                                 jwait(&newblk->nb_jnewblk->jn_list, waitfor);
12305                                 goto top;
12306                         }
12307                         if (newblk->nb_state & DEPCOMPLETE ||
12308                             waitfor == MNT_NOWAIT)
12309                                 continue;
12310                         nbp = newblk->nb_bmsafemap->sm_buf;
12311                         nbp = getdirtybuf(nbp, &lk, waitfor);
12312                         if (nbp == NULL)
12313                                 goto top;
12314                         FREE_LOCK(&lk);
12315                         if ((error = bwrite(nbp)) != 0)
12316                                 goto out;
12317                         ACQUIRE_LOCK(&lk);
12318                         continue;
12319 
12320                 case D_INDIRDEP:
12321                         indirdep = WK_INDIRDEP(wk);
12322                         if (waitfor == MNT_NOWAIT) {
12323                                 if (!TAILQ_EMPTY(&indirdep->ir_trunc) ||
12324                                     !LIST_EMPTY(&indirdep->ir_deplisthd)) {
12325                                         error = EBUSY;
12326                                         goto out_unlock;
12327                                 }
12328                         }
12329                         if (!TAILQ_EMPTY(&indirdep->ir_trunc))
12330                                 panic("softdep_sync_buf: truncation pending.");
12331                 restart:
12332                         LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
12333                                 newblk = (struct newblk *)aip;
12334                                 if (newblk->nb_jnewblk != NULL) {
12335                                         jwait(&newblk->nb_jnewblk->jn_list,
12336                                             waitfor);
12337                                         goto restart;
12338                                 }
12339                                 if (newblk->nb_state & DEPCOMPLETE)
12340                                         continue;
12341                                 nbp = newblk->nb_bmsafemap->sm_buf;
12342                                 nbp = getdirtybuf(nbp, &lk, waitfor);
12343                                 if (nbp == NULL)
12344                                         goto restart;
12345                                 FREE_LOCK(&lk);
12346                                 if ((error = bwrite(nbp)) != 0)
12347                                         goto out;
12348                                 ACQUIRE_LOCK(&lk);
12349                                 goto restart;
12350                         }
12351                         continue;
12352 
12353                 case D_PAGEDEP:
12354                         /*
12355                          * Only flush directory entries in synchronous passes.
12356                          */
12357                         if (waitfor != MNT_WAIT) {
12358                                 error = EBUSY;
12359                                 goto out_unlock;
12360                         }
12361                         /*
12362                          * While syncing snapshots, we must allow recursive
12363                          * lookups.
12364                          */
12365                         BUF_AREC(bp);
12366                         /*
12367                          * We are trying to sync a directory that may
12368                          * have dependencies on both its own metadata
12369                          * and/or dependencies on the inodes of any
12370                          * recently allocated files. We walk its diradd
12371                          * lists pushing out the associated inode.
12372                          */
12373                         pagedep = WK_PAGEDEP(wk);
12374                         for (i = 0; i < DAHASHSZ; i++) {
12375                                 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
12376                                         continue;
12377                                 if ((error = flush_pagedep_deps(vp, wk->wk_mp,
12378                                     &pagedep->pd_diraddhd[i]))) {
12379                                         BUF_NOREC(bp);
12380                                         goto out_unlock;
12381                                 }
12382                         }
12383                         BUF_NOREC(bp);
12384                         continue;
12385 
12386                 case D_FREEWORK:
12387                 case D_FREEDEP:
12388                 case D_JSEGDEP:
12389                 case D_JNEWBLK:
12390                         continue;
12391 
12392                 default:
12393                         panic("softdep_sync_buf: Unknown type %s",
12394                             TYPENAME(wk->wk_type));
12395                         /* NOTREACHED */
12396                 }
12397         }
12398 out_unlock:
12399         FREE_LOCK(&lk);
12400 out:
12401         return (error);
12402 }
12403 
12404 /*
12405  * Flush the dependencies associated with an inodedep.
12406  * Called with splbio blocked.
12407  */
12408 static int
12409 flush_inodedep_deps(vp, mp, ino)
12410         struct vnode *vp;
12411         struct mount *mp;
12412         ino_t ino;
12413 {
12414         struct inodedep *inodedep;
12415         struct inoref *inoref;
12416         int error, waitfor;
12417 
12418         /*
12419          * This work is done in two passes. The first pass grabs most
12420          * of the buffers and begins asynchronously writing them. The
12421          * only way to wait for these asynchronous writes is to sleep
12422          * on the filesystem vnode which may stay busy for a long time
12423          * if the filesystem is active. So, instead, we make a second
12424          * pass over the dependencies blocking on each write. In the
12425          * usual case we will be blocking against a write that we
12426          * initiated, so when it is done the dependency will have been
12427          * resolved. Thus the second pass is expected to end quickly.
12428          * We give a brief window at the top of the loop to allow
12429          * any pending I/O to complete.
12430          */
12431         for (error = 0, waitfor = MNT_NOWAIT; ; ) {
12432                 if (error)
12433                         return (error);
12434                 FREE_LOCK(&lk);
12435                 ACQUIRE_LOCK(&lk);
12436 restart:
12437                 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0)
12438                         return (0);
12439                 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
12440                         if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
12441                             == DEPCOMPLETE) {
12442                                 jwait(&inoref->if_list, MNT_WAIT);
12443                                 goto restart;
12444                         }
12445                 }
12446                 if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) ||
12447                     flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) ||
12448                     flush_deplist(&inodedep->id_extupdt, waitfor, &error) ||
12449                     flush_deplist(&inodedep->id_newextupdt, waitfor, &error))
12450                         continue;
12451                 /*
12452                  * If pass2, we are done, otherwise do pass 2.
12453                  */
12454                 if (waitfor == MNT_WAIT)
12455                         break;
12456                 waitfor = MNT_WAIT;
12457         }
12458         /*
12459          * Try freeing inodedep in case all dependencies have been removed.
12460          */
12461         if (inodedep_lookup(mp, ino, 0, &inodedep) != 0)
12462                 (void) free_inodedep(inodedep);
12463         return (0);
12464 }
12465 
12466 /*
12467  * Flush an inode dependency list.
12468  * Called with splbio blocked.
12469  */
12470 static int
12471 flush_deplist(listhead, waitfor, errorp)
12472         struct allocdirectlst *listhead;
12473         int waitfor;
12474         int *errorp;
12475 {
12476         struct allocdirect *adp;
12477         struct newblk *newblk;
12478         struct buf *bp;
12479 
12480         rw_assert(&lk, RA_WLOCKED);
12481         TAILQ_FOREACH(adp, listhead, ad_next) {
12482                 newblk = (struct newblk *)adp;
12483                 if (newblk->nb_jnewblk != NULL) {
12484                         jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT);
12485                         return (1);
12486                 }
12487                 if (newblk->nb_state & DEPCOMPLETE)
12488                         continue;
12489                 bp = newblk->nb_bmsafemap->sm_buf;
12490                 bp = getdirtybuf(bp, &lk, waitfor);
12491                 if (bp == NULL) {
12492                         if (waitfor == MNT_NOWAIT)
12493                                 continue;
12494                         return (1);
12495                 }
12496                 FREE_LOCK(&lk);
12497                 if (waitfor == MNT_NOWAIT)
12498                         bawrite(bp);
12499                 else 
12500                         *errorp = bwrite(bp);
12501                 ACQUIRE_LOCK(&lk);
12502                 return (1);
12503         }
12504         return (0);
12505 }
12506 
12507 /*
12508  * Flush dependencies associated with an allocdirect block.
12509  */
12510 static int
12511 flush_newblk_dep(vp, mp, lbn)
12512         struct vnode *vp;
12513         struct mount *mp;
12514         ufs_lbn_t lbn;
12515 {
12516         struct newblk *newblk;
12517         struct bufobj *bo;
12518         struct inode *ip;
12519         struct buf *bp;
12520         ufs2_daddr_t blkno;
12521         int error;
12522 
12523         error = 0;
12524         bo = &vp->v_bufobj;
12525         ip = VTOI(vp);
12526         blkno = DIP(ip, i_db[lbn]);
12527         if (blkno == 0)
12528                 panic("flush_newblk_dep: Missing block");
12529         ACQUIRE_LOCK(&lk);
12530         /*
12531          * Loop until all dependencies related to this block are satisfied.
12532          * We must be careful to restart after each sleep in case a write
12533          * completes some part of this process for us.
12534          */
12535         for (;;) {
12536                 if (newblk_lookup(mp, blkno, 0, &newblk) == 0) {
12537                         FREE_LOCK(&lk);
12538                         break;
12539                 }
12540                 if (newblk->nb_list.wk_type != D_ALLOCDIRECT)
12541                         panic("flush_newblk_deps: Bad newblk %p", newblk);
12542                 /*
12543                  * Flush the journal.
12544                  */
12545                 if (newblk->nb_jnewblk != NULL) {
12546                         jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT);
12547                         continue;
12548                 }
12549                 /*
12550                  * Write the bitmap dependency.
12551                  */
12552                 if ((newblk->nb_state & DEPCOMPLETE) == 0) {
12553                         bp = newblk->nb_bmsafemap->sm_buf;
12554                         bp = getdirtybuf(bp, &lk, MNT_WAIT);
12555                         if (bp == NULL)
12556                                 continue;
12557                         FREE_LOCK(&lk);
12558                         error = bwrite(bp);
12559                         if (error)
12560                                 break;
12561                         ACQUIRE_LOCK(&lk);
12562                         continue;
12563                 }
12564                 /*
12565                  * Write the buffer.
12566                  */
12567                 FREE_LOCK(&lk);
12568                 BO_LOCK(bo);
12569                 bp = gbincore(bo, lbn);
12570                 if (bp != NULL) {
12571                         error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
12572                             LK_INTERLOCK, BO_LOCKPTR(bo));
12573                         if (error == ENOLCK) {
12574                                 ACQUIRE_LOCK(&lk);
12575                                 continue; /* Slept, retry */
12576                         }
12577                         if (error != 0)
12578                                 break;  /* Failed */
12579                         if (bp->b_flags & B_DELWRI) {
12580                                 bremfree(bp);
12581                                 error = bwrite(bp);
12582                                 if (error)
12583                                         break;
12584                         } else
12585                                 BUF_UNLOCK(bp);
12586                 } else
12587                         BO_UNLOCK(bo);
12588                 /*
12589                  * We have to wait for the direct pointers to
12590                  * point at the newdirblk before the dependency
12591                  * will go away.
12592                  */
12593                 error = ffs_update(vp, 1);
12594                 if (error)
12595                         break;
12596                 ACQUIRE_LOCK(&lk);
12597         }
12598         return (error);
12599 }
12600 
12601 /*
12602  * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
12603  * Called with splbio blocked.
12604  */
12605 static int
12606 flush_pagedep_deps(pvp, mp, diraddhdp)
12607         struct vnode *pvp;
12608         struct mount *mp;
12609         struct diraddhd *diraddhdp;
12610 {
12611         struct inodedep *inodedep;
12612         struct inoref *inoref;
12613         struct ufsmount *ump;
12614         struct diradd *dap;
12615         struct vnode *vp;
12616         int error = 0;
12617         struct buf *bp;
12618         ino_t inum;
12619 
12620         ump = VFSTOUFS(mp);
12621 restart:
12622         while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
12623                 /*
12624                  * Flush ourselves if this directory entry
12625                  * has a MKDIR_PARENT dependency.
12626                  */
12627                 if (dap->da_state & MKDIR_PARENT) {
12628                         FREE_LOCK(&lk);
12629                         if ((error = ffs_update(pvp, 1)) != 0)
12630                                 break;
12631                         ACQUIRE_LOCK(&lk);
12632                         /*
12633                          * If that cleared dependencies, go on to next.
12634                          */
12635                         if (dap != LIST_FIRST(diraddhdp))
12636                                 continue;
12637                         if (dap->da_state & MKDIR_PARENT)
12638                                 panic("flush_pagedep_deps: MKDIR_PARENT");
12639                 }
12640                 /*
12641                  * A newly allocated directory must have its "." and
12642                  * ".." entries written out before its name can be
12643                  * committed in its parent. 
12644                  */
12645                 inum = dap->da_newinum;
12646                 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0)
12647                         panic("flush_pagedep_deps: lost inode1");
12648                 /*
12649                  * Wait for any pending journal adds to complete so we don't
12650                  * cause rollbacks while syncing.
12651                  */
12652                 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
12653                         if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
12654                             == DEPCOMPLETE) {
12655                                 jwait(&inoref->if_list, MNT_WAIT);
12656                                 goto restart;
12657                         }
12658                 }
12659                 if (dap->da_state & MKDIR_BODY) {
12660                         FREE_LOCK(&lk);
12661                         if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp,
12662                             FFSV_FORCEINSMQ)))
12663                                 break;
12664                         error = flush_newblk_dep(vp, mp, 0);
12665                         /*
12666                          * If we still have the dependency we might need to
12667                          * update the vnode to sync the new link count to
12668                          * disk.
12669                          */
12670                         if (error == 0 && dap == LIST_FIRST(diraddhdp))
12671                                 error = ffs_update(vp, 1);
12672                         vput(vp);
12673                         if (error != 0)
12674                                 break;
12675                         ACQUIRE_LOCK(&lk);
12676                         /*
12677                          * If that cleared dependencies, go on to next.
12678                          */
12679                         if (dap != LIST_FIRST(diraddhdp))
12680                                 continue;
12681                         if (dap->da_state & MKDIR_BODY) {
12682                                 inodedep_lookup(UFSTOVFS(ump), inum, 0,
12683                                     &inodedep);
12684                                 panic("flush_pagedep_deps: MKDIR_BODY "
12685                                     "inodedep %p dap %p vp %p",
12686                                     inodedep, dap, vp);
12687                         }
12688                 }
12689                 /*
12690                  * Flush the inode on which the directory entry depends.
12691                  * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
12692                  * the only remaining dependency is that the updated inode
12693                  * count must get pushed to disk. The inode has already
12694                  * been pushed into its inode buffer (via VOP_UPDATE) at
12695                  * the time of the reference count change. So we need only
12696                  * locate that buffer, ensure that there will be no rollback
12697                  * caused by a bitmap dependency, then write the inode buffer.
12698                  */
12699 retry:
12700                 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0)
12701                         panic("flush_pagedep_deps: lost inode");
12702                 /*
12703                  * If the inode still has bitmap dependencies,
12704                  * push them to disk.
12705                  */
12706                 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) == 0) {
12707                         bp = inodedep->id_bmsafemap->sm_buf;
12708                         bp = getdirtybuf(bp, &lk, MNT_WAIT);
12709                         if (bp == NULL)
12710                                 goto retry;
12711                         FREE_LOCK(&lk);
12712                         if ((error = bwrite(bp)) != 0)
12713                                 break;
12714                         ACQUIRE_LOCK(&lk);
12715                         if (dap != LIST_FIRST(diraddhdp))
12716                                 continue;
12717                 }
12718                 /*
12719                  * If the inode is still sitting in a buffer waiting
12720                  * to be written or waiting for the link count to be
12721                  * adjusted update it here to flush it to disk.
12722                  */
12723                 if (dap == LIST_FIRST(diraddhdp)) {
12724                         FREE_LOCK(&lk);
12725                         if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp,
12726                             FFSV_FORCEINSMQ)))
12727                                 break;
12728                         error = ffs_update(vp, 1);
12729                         vput(vp);
12730                         if (error)
12731                                 break;
12732                         ACQUIRE_LOCK(&lk);
12733                 }
12734                 /*
12735                  * If we have failed to get rid of all the dependencies
12736                  * then something is seriously wrong.
12737                  */
12738                 if (dap == LIST_FIRST(diraddhdp)) {
12739                         inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep);
12740                         panic("flush_pagedep_deps: failed to flush " 
12741                             "inodedep %p ino %ju dap %p",
12742                             inodedep, (uintmax_t)inum, dap);
12743                 }
12744         }
12745         if (error)
12746                 ACQUIRE_LOCK(&lk);
12747         return (error);
12748 }
12749 
12750 /*
12751  * A large burst of file addition or deletion activity can drive the
12752  * memory load excessively high. First attempt to slow things down
12753  * using the techniques below. If that fails, this routine requests
12754  * the offending operations to fall back to running synchronously
12755  * until the memory load returns to a reasonable level.
12756  */
12757 int
12758 softdep_slowdown(vp)
12759         struct vnode *vp;
12760 {
12761         struct ufsmount *ump;
12762         int jlow;
12763         int max_softdeps_hard;
12764 
12765         ACQUIRE_LOCK(&lk);
12766         jlow = 0;
12767         /*
12768          * Check for journal space if needed.
12769          */
12770         if (DOINGSUJ(vp)) {
12771                 ump = VFSTOUFS(vp->v_mount);
12772                 if (journal_space(ump, 0) == 0)
12773                         jlow = 1;
12774         }
12775         max_softdeps_hard = max_softdeps * 11 / 10;
12776         if (dep_current[D_DIRREM] < max_softdeps_hard / 2 &&
12777             dep_current[D_INODEDEP] < max_softdeps_hard &&
12778             VFSTOUFS(vp->v_mount)->um_numindirdeps < maxindirdeps &&
12779             dep_current[D_FREEBLKS] < max_softdeps_hard && jlow == 0) {
12780                 FREE_LOCK(&lk);
12781                 return (0);
12782         }
12783         if (VFSTOUFS(vp->v_mount)->um_numindirdeps >= maxindirdeps || jlow)
12784                 softdep_speedup();
12785         stat_sync_limit_hit += 1;
12786         FREE_LOCK(&lk);
12787         if (DOINGSUJ(vp))
12788                 return (0);
12789         return (1);
12790 }
12791 
12792 /*
12793  * Called by the allocation routines when they are about to fail
12794  * in the hope that we can free up the requested resource (inodes
12795  * or disk space).
12796  * 
12797  * First check to see if the work list has anything on it. If it has,
12798  * clean up entries until we successfully free the requested resource.
12799  * Because this process holds inodes locked, we cannot handle any remove
12800  * requests that might block on a locked inode as that could lead to
12801  * deadlock. If the worklist yields none of the requested resource,
12802  * start syncing out vnodes to free up the needed space.
12803  */
12804 int
12805 softdep_request_cleanup(fs, vp, cred, resource)
12806         struct fs *fs;
12807         struct vnode *vp;
12808         struct ucred *cred;
12809         int resource;
12810 {
12811         struct ufsmount *ump;
12812         struct mount *mp;
12813         struct vnode *lvp, *mvp;
12814         long starttime;
12815         ufs2_daddr_t needed;
12816         int error;
12817 
12818         /*
12819          * If we are being called because of a process doing a
12820          * copy-on-write, then it is not safe to process any
12821          * worklist items as we will recurse into the copyonwrite
12822          * routine.  This will result in an incoherent snapshot.
12823          * If the vnode that we hold is a snapshot, we must avoid
12824          * handling other resources that could cause deadlock.
12825          */
12826         if ((curthread->td_pflags & TDP_COWINPROGRESS) || IS_SNAPSHOT(VTOI(vp)))
12827                 return (0);
12828 
12829         if (resource == FLUSH_BLOCKS_WAIT)
12830                 stat_cleanup_blkrequests += 1;
12831         else
12832                 stat_cleanup_inorequests += 1;
12833 
12834         mp = vp->v_mount;
12835         ump = VFSTOUFS(mp);
12836         mtx_assert(UFS_MTX(ump), MA_OWNED);
12837         UFS_UNLOCK(ump);
12838         error = ffs_update(vp, 1);
12839         if (error != 0) {
12840                 UFS_LOCK(ump);
12841                 return (0);
12842         }
12843         /*
12844          * If we are in need of resources, consider pausing for
12845          * tickdelay to give ourselves some breathing room.
12846          */
12847         ACQUIRE_LOCK(&lk);
12848         process_removes(vp);
12849         process_truncates(vp);
12850         request_cleanup(UFSTOVFS(ump), resource);
12851         FREE_LOCK(&lk);
12852         /*
12853          * Now clean up at least as many resources as we will need.
12854          *
12855          * When requested to clean up inodes, the number that are needed
12856          * is set by the number of simultaneous writers (mnt_writeopcount)
12857          * plus a bit of slop (2) in case some more writers show up while
12858          * we are cleaning.
12859          *
12860          * When requested to free up space, the amount of space that
12861          * we need is enough blocks to allocate a full-sized segment
12862          * (fs_contigsumsize). The number of such segments that will
12863          * be needed is set by the number of simultaneous writers
12864          * (mnt_writeopcount) plus a bit of slop (2) in case some more
12865          * writers show up while we are cleaning.
12866          *
12867          * Additionally, if we are unpriviledged and allocating space,
12868          * we need to ensure that we clean up enough blocks to get the
12869          * needed number of blocks over the threshhold of the minimum
12870          * number of blocks required to be kept free by the filesystem
12871          * (fs_minfree).
12872          */
12873         if (resource == FLUSH_INODES_WAIT) {
12874                 needed = vp->v_mount->mnt_writeopcount + 2;
12875         } else if (resource == FLUSH_BLOCKS_WAIT) {
12876                 needed = (vp->v_mount->mnt_writeopcount + 2) *
12877                     fs->fs_contigsumsize;
12878                 if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE, 0))
12879                         needed += fragstoblks(fs,
12880                             roundup((fs->fs_dsize * fs->fs_minfree / 100) -
12881                             fs->fs_cstotal.cs_nffree, fs->fs_frag));
12882         } else {
12883                 UFS_LOCK(ump);
12884                 printf("softdep_request_cleanup: Unknown resource type %d\n",
12885                     resource);
12886                 return (0);
12887         }
12888         starttime = time_second;
12889 retry:
12890         if ((resource == FLUSH_BLOCKS_WAIT && ump->softdep_on_worklist > 0 &&
12891             fs->fs_cstotal.cs_nbfree <= needed) ||
12892             (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 &&
12893             fs->fs_cstotal.cs_nifree <= needed)) {
12894                 ACQUIRE_LOCK(&lk);
12895                 if (ump->softdep_on_worklist > 0 &&
12896                     process_worklist_item(UFSTOVFS(ump),
12897                     ump->softdep_on_worklist, LK_NOWAIT) != 0)
12898                         stat_worklist_push += 1;
12899                 FREE_LOCK(&lk);
12900         }
12901         /*
12902          * If we still need resources and there are no more worklist
12903          * entries to process to obtain them, we have to start flushing
12904          * the dirty vnodes to force the release of additional requests
12905          * to the worklist that we can then process to reap addition
12906          * resources. We walk the vnodes associated with the mount point
12907          * until we get the needed worklist requests that we can reap.
12908          */
12909         if ((resource == FLUSH_BLOCKS_WAIT && 
12910              fs->fs_cstotal.cs_nbfree <= needed) ||
12911             (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 &&
12912              fs->fs_cstotal.cs_nifree <= needed)) {
12913                 MNT_VNODE_FOREACH_ALL(lvp, mp, mvp) {
12914                         if (TAILQ_FIRST(&lvp->v_bufobj.bo_dirty.bv_hd) == 0) {
12915                                 VI_UNLOCK(lvp);
12916                                 continue;
12917                         }
12918                         if (vget(lvp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT,
12919                             curthread))
12920                                 continue;
12921                         if (lvp->v_vflag & VV_NOSYNC) { /* unlinked */
12922                                 vput(lvp);
12923                                 continue;
12924                         }
12925                         (void) ffs_syncvnode(lvp, MNT_NOWAIT, 0);
12926                         vput(lvp);
12927                 }
12928                 lvp = ump->um_devvp;
12929                 if (vn_lock(lvp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
12930                         VOP_FSYNC(lvp, MNT_NOWAIT, curthread);
12931                         VOP_UNLOCK(lvp, 0);
12932                 }
12933                 if (ump->softdep_on_worklist > 0) {
12934                         stat_cleanup_retries += 1;
12935                         goto retry;
12936                 }
12937                 stat_cleanup_failures += 1;
12938         }
12939         if (time_second - starttime > stat_cleanup_high_delay)
12940                 stat_cleanup_high_delay = time_second - starttime;
12941         UFS_LOCK(ump);
12942         return (1);
12943 }
12944 
12945 /*
12946  * If memory utilization has gotten too high, deliberately slow things
12947  * down and speed up the I/O processing.
12948  */
12949 extern struct thread *syncertd;
12950 static int
12951 request_cleanup(mp, resource)
12952         struct mount *mp;
12953         int resource;
12954 {
12955         struct thread *td = curthread;
12956         struct ufsmount *ump;
12957 
12958         rw_assert(&lk, RA_WLOCKED);
12959         /*
12960          * We never hold up the filesystem syncer or buf daemon.
12961          */
12962         if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF))
12963                 return (0);
12964         ump = VFSTOUFS(mp);
12965         /*
12966          * First check to see if the work list has gotten backlogged.
12967          * If it has, co-opt this process to help clean up two entries.
12968          * Because this process may hold inodes locked, we cannot
12969          * handle any remove requests that might block on a locked
12970          * inode as that could lead to deadlock.  We set TDP_SOFTDEP
12971          * to avoid recursively processing the worklist.
12972          */
12973         if (ump->softdep_on_worklist > max_softdeps / 10) {
12974                 td->td_pflags |= TDP_SOFTDEP;
12975                 process_worklist_item(mp, 2, LK_NOWAIT);
12976                 td->td_pflags &= ~TDP_SOFTDEP;
12977                 stat_worklist_push += 2;
12978                 return(1);
12979         }
12980         /*
12981          * Next, we attempt to speed up the syncer process. If that
12982          * is successful, then we allow the process to continue.
12983          */
12984         if (softdep_speedup() &&
12985             resource != FLUSH_BLOCKS_WAIT &&
12986             resource != FLUSH_INODES_WAIT)
12987                 return(0);
12988         /*
12989          * If we are resource constrained on inode dependencies, try
12990          * flushing some dirty inodes. Otherwise, we are constrained
12991          * by file deletions, so try accelerating flushes of directories
12992          * with removal dependencies. We would like to do the cleanup
12993          * here, but we probably hold an inode locked at this point and 
12994          * that might deadlock against one that we try to clean. So,
12995          * the best that we can do is request the syncer daemon to do
12996          * the cleanup for us.
12997          */
12998         switch (resource) {
12999 
13000         case FLUSH_INODES:
13001         case FLUSH_INODES_WAIT:
13002                 stat_ino_limit_push += 1;
13003                 req_clear_inodedeps += 1;
13004                 stat_countp = &stat_ino_limit_hit;
13005                 break;
13006 
13007         case FLUSH_BLOCKS:
13008         case FLUSH_BLOCKS_WAIT:
13009                 stat_blk_limit_push += 1;
13010                 req_clear_remove += 1;
13011                 stat_countp = &stat_blk_limit_hit;
13012                 break;
13013 
13014         default:
13015                 panic("request_cleanup: unknown type");
13016         }
13017         /*
13018          * Hopefully the syncer daemon will catch up and awaken us.
13019          * We wait at most tickdelay before proceeding in any case.
13020          */
13021         proc_waiting += 1;
13022         if (callout_pending(&softdep_callout) == FALSE)
13023                 callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2,
13024                     pause_timer, 0);
13025 
13026         msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0);
13027         proc_waiting -= 1;
13028         return (1);
13029 }
13030 
13031 /*
13032  * Awaken processes pausing in request_cleanup and clear proc_waiting
13033  * to indicate that there is no longer a timer running.
13034  */
13035 static void
13036 pause_timer(arg)
13037         void *arg;
13038 {
13039 
13040         /*
13041          * The callout_ API has acquired mtx and will hold it around this
13042          * function call.
13043          */
13044         *stat_countp += 1;
13045         wakeup_one(&proc_waiting);
13046         if (proc_waiting > 0)
13047                 callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2,
13048                     pause_timer, 0);
13049 }
13050 
13051 /*
13052  * Flush out a directory with at least one removal dependency in an effort to
13053  * reduce the number of dirrem, freefile, and freeblks dependency structures.
13054  */
13055 static void
13056 clear_remove(void)
13057 {
13058         struct pagedep_hashhead *pagedephd;
13059         struct pagedep *pagedep;
13060         static int next = 0;
13061         struct mount *mp;
13062         struct vnode *vp;
13063         struct bufobj *bo;
13064         int error, cnt;
13065         ino_t ino;
13066 
13067         rw_assert(&lk, RA_WLOCKED);
13068 
13069         for (cnt = 0; cnt <= pagedep_hash; cnt++) {
13070                 pagedephd = &pagedep_hashtbl[next++];
13071                 if (next > pagedep_hash)
13072                         next = 0;
13073                 LIST_FOREACH(pagedep, pagedephd, pd_hash) {
13074                         if (LIST_EMPTY(&pagedep->pd_dirremhd))
13075                                 continue;
13076                         mp = pagedep->pd_list.wk_mp;
13077                         ino = pagedep->pd_ino;
13078                         if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
13079                                 continue;
13080                         FREE_LOCK(&lk);
13081 
13082                         /*
13083                          * Let unmount clear deps
13084                          */
13085                         error = vfs_busy(mp, MBF_NOWAIT);
13086                         if (error != 0)
13087                                 goto finish_write;
13088                         error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp,
13089                              FFSV_FORCEINSMQ);
13090                         vfs_unbusy(mp);
13091                         if (error != 0) {
13092                                 softdep_error("clear_remove: vget", error);
13093                                 goto finish_write;
13094                         }
13095                         if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0)))
13096                                 softdep_error("clear_remove: fsync", error);
13097                         bo = &vp->v_bufobj;
13098                         BO_LOCK(bo);
13099                         drain_output(vp);
13100                         BO_UNLOCK(bo);
13101                         vput(vp);
13102                 finish_write:
13103                         vn_finished_write(mp);
13104                         ACQUIRE_LOCK(&lk);
13105                         return;
13106                 }
13107         }
13108 }
13109 
13110 /*
13111  * Clear out a block of dirty inodes in an effort to reduce
13112  * the number of inodedep dependency structures.
13113  */
13114 static void
13115 clear_inodedeps(void)
13116 {
13117         struct inodedep_hashhead *inodedephd;
13118         struct inodedep *inodedep;
13119         static int next = 0;
13120         struct mount *mp;
13121         struct vnode *vp;
13122         struct fs *fs;
13123         int error, cnt;
13124         ino_t firstino, lastino, ino;
13125 
13126         rw_assert(&lk, RA_WLOCKED);
13127         /*
13128          * Pick a random inode dependency to be cleared.
13129          * We will then gather up all the inodes in its block 
13130          * that have dependencies and flush them out.
13131          */
13132         for (cnt = 0; cnt <= inodedep_hash; cnt++) {
13133                 inodedephd = &inodedep_hashtbl[next++];
13134                 if (next > inodedep_hash)
13135                         next = 0;
13136                 if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
13137                         break;
13138         }
13139         if (inodedep == NULL)
13140                 return;
13141         fs = inodedep->id_fs;
13142         mp = inodedep->id_list.wk_mp;
13143         /*
13144          * Find the last inode in the block with dependencies.
13145          */
13146         firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
13147         for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
13148                 if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0)
13149                         break;
13150         /*
13151          * Asynchronously push all but the last inode with dependencies.
13152          * Synchronously push the last inode with dependencies to ensure
13153          * that the inode block gets written to free up the inodedeps.
13154          */
13155         for (ino = firstino; ino <= lastino; ino++) {
13156                 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0)
13157                         continue;
13158                 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
13159                         continue;
13160                 FREE_LOCK(&lk);
13161                 error = vfs_busy(mp, MBF_NOWAIT); /* Let unmount clear deps */
13162                 if (error != 0) {
13163                         vn_finished_write(mp);
13164                         ACQUIRE_LOCK(&lk);
13165                         return;
13166                 }
13167                 if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp,
13168                     FFSV_FORCEINSMQ)) != 0) {
13169                         softdep_error("clear_inodedeps: vget", error);
13170                         vfs_unbusy(mp);
13171                         vn_finished_write(mp);
13172                         ACQUIRE_LOCK(&lk);
13173                         return;
13174                 }
13175                 vfs_unbusy(mp);
13176                 if (ino == lastino) {
13177                         if ((error = ffs_syncvnode(vp, MNT_WAIT, 0)))
13178                                 softdep_error("clear_inodedeps: fsync1", error);
13179                 } else {
13180                         if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0)))
13181                                 softdep_error("clear_inodedeps: fsync2", error);
13182                         BO_LOCK(&vp->v_bufobj);
13183                         drain_output(vp);
13184                         BO_UNLOCK(&vp->v_bufobj);
13185                 }
13186                 vput(vp);
13187                 vn_finished_write(mp);
13188                 ACQUIRE_LOCK(&lk);
13189         }
13190 }
13191 
13192 void
13193 softdep_buf_append(bp, wkhd)
13194         struct buf *bp;
13195         struct workhead *wkhd;
13196 {
13197         struct worklist *wk;
13198 
13199         ACQUIRE_LOCK(&lk);
13200         while ((wk = LIST_FIRST(wkhd)) != NULL) {
13201                 WORKLIST_REMOVE(wk);
13202                 WORKLIST_INSERT(&bp->b_dep, wk);
13203         }
13204         FREE_LOCK(&lk);
13205 
13206 }
13207 
13208 void
13209 softdep_inode_append(ip, cred, wkhd)
13210         struct inode *ip;
13211         struct ucred *cred;
13212         struct workhead *wkhd;
13213 {
13214         struct buf *bp;
13215         struct fs *fs;
13216         int error;
13217 
13218         fs = ip->i_fs;
13219         error = bread(ip->i_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
13220             (int)fs->fs_bsize, cred, &bp);
13221         if (error) {
13222                 bqrelse(bp);
13223                 softdep_freework(wkhd);
13224                 return;
13225         }
13226         softdep_buf_append(bp, wkhd);
13227         bqrelse(bp);
13228 }
13229 
13230 void
13231 softdep_freework(wkhd)
13232         struct workhead *wkhd;
13233 {
13234 
13235         ACQUIRE_LOCK(&lk);
13236         handle_jwork(wkhd);
13237         FREE_LOCK(&lk);
13238 }
13239 
13240 /*
13241  * Function to determine if the buffer has outstanding dependencies
13242  * that will cause a roll-back if the buffer is written. If wantcount
13243  * is set, return number of dependencies, otherwise just yes or no.
13244  */
13245 static int
13246 softdep_count_dependencies(bp, wantcount)
13247         struct buf *bp;
13248         int wantcount;
13249 {
13250         struct worklist *wk;
13251         struct bmsafemap *bmsafemap;
13252         struct freework *freework;
13253         struct inodedep *inodedep;
13254         struct indirdep *indirdep;
13255         struct freeblks *freeblks;
13256         struct allocindir *aip;
13257         struct pagedep *pagedep;
13258         struct dirrem *dirrem;
13259         struct newblk *newblk;
13260         struct mkdir *mkdir;
13261         struct diradd *dap;
13262         int i, retval;
13263 
13264         retval = 0;
13265         ACQUIRE_LOCK(&lk);
13266         LIST_FOREACH(wk, &bp->b_dep, wk_list) {
13267                 switch (wk->wk_type) {
13268 
13269                 case D_INODEDEP:
13270                         inodedep = WK_INODEDEP(wk);
13271                         if ((inodedep->id_state & DEPCOMPLETE) == 0) {
13272                                 /* bitmap allocation dependency */
13273                                 retval += 1;
13274                                 if (!wantcount)
13275                                         goto out;
13276                         }
13277                         if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
13278                                 /* direct block pointer dependency */
13279                                 retval += 1;
13280                                 if (!wantcount)
13281                                         goto out;
13282                         }
13283                         if (TAILQ_FIRST(&inodedep->id_extupdt)) {
13284                                 /* direct block pointer dependency */
13285                                 retval += 1;
13286                                 if (!wantcount)
13287                                         goto out;
13288                         }
13289                         if (TAILQ_FIRST(&inodedep->id_inoreflst)) {
13290                                 /* Add reference dependency. */
13291                                 retval += 1;
13292                                 if (!wantcount)
13293                                         goto out;
13294                         }
13295                         continue;
13296 
13297                 case D_INDIRDEP:
13298                         indirdep = WK_INDIRDEP(wk);
13299 
13300                         TAILQ_FOREACH(freework, &indirdep->ir_trunc, fw_next) {
13301                                 /* indirect truncation dependency */
13302                                 retval += 1;
13303                                 if (!wantcount)
13304                                         goto out;
13305                         }
13306 
13307                         LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
13308                                 /* indirect block pointer dependency */
13309                                 retval += 1;
13310                                 if (!wantcount)
13311                                         goto out;
13312                         }
13313                         continue;
13314 
13315                 case D_PAGEDEP:
13316                         pagedep = WK_PAGEDEP(wk);
13317                         LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
13318                                 if (LIST_FIRST(&dirrem->dm_jremrefhd)) {
13319                                         /* Journal remove ref dependency. */
13320                                         retval += 1;
13321                                         if (!wantcount)
13322                                                 goto out;
13323                                 }
13324                         }
13325                         for (i = 0; i < DAHASHSZ; i++) {
13326 
13327                                 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
13328                                         /* directory entry dependency */
13329                                         retval += 1;
13330                                         if (!wantcount)
13331                                                 goto out;
13332                                 }
13333                         }
13334                         continue;
13335 
13336                 case D_BMSAFEMAP:
13337                         bmsafemap = WK_BMSAFEMAP(wk);
13338                         if (LIST_FIRST(&bmsafemap->sm_jaddrefhd)) {
13339                                 /* Add reference dependency. */
13340                                 retval += 1;
13341                                 if (!wantcount)
13342                                         goto out;
13343                         }
13344                         if (LIST_FIRST(&bmsafemap->sm_jnewblkhd)) {
13345                                 /* Allocate block dependency. */
13346                                 retval += 1;
13347                                 if (!wantcount)
13348                                         goto out;
13349                         }
13350                         continue;
13351 
13352                 case D_FREEBLKS:
13353                         freeblks = WK_FREEBLKS(wk);
13354                         if (LIST_FIRST(&freeblks->fb_jblkdephd)) {
13355                                 /* Freeblk journal dependency. */
13356                                 retval += 1;
13357                                 if (!wantcount)
13358                                         goto out;
13359                         }
13360                         continue;
13361 
13362                 case D_ALLOCDIRECT:
13363                 case D_ALLOCINDIR:
13364                         newblk = WK_NEWBLK(wk);
13365                         if (newblk->nb_jnewblk) {
13366                                 /* Journal allocate dependency. */
13367                                 retval += 1;
13368                                 if (!wantcount)
13369                                         goto out;
13370                         }
13371                         continue;
13372 
13373                 case D_MKDIR:
13374                         mkdir = WK_MKDIR(wk);
13375                         if (mkdir->md_jaddref) {
13376                                 /* Journal reference dependency. */
13377                                 retval += 1;
13378                                 if (!wantcount)
13379                                         goto out;
13380                         }
13381                         continue;
13382 
13383                 case D_FREEWORK:
13384                 case D_FREEDEP:
13385                 case D_JSEGDEP:
13386                 case D_JSEG:
13387                 case D_SBDEP:
13388                         /* never a dependency on these blocks */
13389                         continue;
13390 
13391                 default:
13392                         panic("softdep_count_dependencies: Unexpected type %s",
13393                             TYPENAME(wk->wk_type));
13394                         /* NOTREACHED */
13395                 }
13396         }
13397 out:
13398         FREE_LOCK(&lk);
13399         return retval;
13400 }
13401 
13402 /*
13403  * Acquire exclusive access to a buffer.
13404  * Must be called with a locked mtx parameter.
13405  * Return acquired buffer or NULL on failure.
13406  */
13407 static struct buf *
13408 getdirtybuf(bp, lock, waitfor)
13409         struct buf *bp;
13410         struct rwlock *lock;
13411         int waitfor;
13412 {
13413         int error;
13414 
13415         rw_assert(lock, RA_WLOCKED);
13416         if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) {
13417                 if (waitfor != MNT_WAIT)
13418                         return (NULL);
13419                 error = BUF_LOCK(bp,
13420                     LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, lock);
13421                 /*
13422                  * Even if we sucessfully acquire bp here, we have dropped
13423                  * lock, which may violates our guarantee.
13424                  */
13425                 if (error == 0)
13426                         BUF_UNLOCK(bp);
13427                 else if (error != ENOLCK)
13428                         panic("getdirtybuf: inconsistent lock: %d", error);
13429                 rw_wlock(lock);
13430                 return (NULL);
13431         }
13432         if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
13433                 if (lock == &lk && waitfor == MNT_WAIT) {
13434                         rw_wunlock(lock);
13435                         BO_LOCK(bp->b_bufobj);
13436                         BUF_UNLOCK(bp);
13437                         if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
13438                                 bp->b_vflags |= BV_BKGRDWAIT;
13439                                 msleep(&bp->b_xflags, BO_LOCKPTR(bp->b_bufobj),
13440                                        PRIBIO | PDROP, "getbuf", 0);
13441                         } else
13442                                 BO_UNLOCK(bp->b_bufobj);
13443                         rw_wlock(lock);
13444                         return (NULL);
13445                 }
13446                 BUF_UNLOCK(bp);
13447                 if (waitfor != MNT_WAIT)
13448                         return (NULL);
13449                 /*
13450                  * The lock argument must be bp->b_vp's mutex in
13451                  * this case.
13452                  */
13453 #ifdef  DEBUG_VFS_LOCKS
13454                 if (bp->b_vp->v_type != VCHR)
13455                         ASSERT_BO_WLOCKED(bp->b_bufobj);
13456 #endif
13457                 bp->b_vflags |= BV_BKGRDWAIT;
13458                 rw_sleep(&bp->b_xflags, lock, PRIBIO, "getbuf", 0);
13459                 return (NULL);
13460         }
13461         if ((bp->b_flags & B_DELWRI) == 0) {
13462                 BUF_UNLOCK(bp);
13463                 return (NULL);
13464         }
13465         bremfree(bp);
13466         return (bp);
13467 }
13468 
13469 
13470 /*
13471  * Check if it is safe to suspend the file system now.  On entry,
13472  * the vnode interlock for devvp should be held.  Return 0 with
13473  * the mount interlock held if the file system can be suspended now,
13474  * otherwise return EAGAIN with the mount interlock held.
13475  */
13476 int
13477 softdep_check_suspend(struct mount *mp,
13478                       struct vnode *devvp,
13479                       int softdep_deps,
13480                       int softdep_accdeps,
13481                       int secondary_writes,
13482                       int secondary_accwrites)
13483 {
13484         struct bufobj *bo;
13485         struct ufsmount *ump;
13486         int error;
13487 
13488         ump = VFSTOUFS(mp);
13489         bo = &devvp->v_bufobj;
13490         ASSERT_BO_WLOCKED(bo);
13491 
13492         for (;;) {
13493                 if (!TRY_ACQUIRE_LOCK(&lk)) {
13494                         BO_UNLOCK(bo);
13495                         ACQUIRE_LOCK(&lk);
13496                         FREE_LOCK(&lk);
13497                         BO_LOCK(bo);
13498                         continue;
13499                 }
13500                 MNT_ILOCK(mp);
13501                 if (mp->mnt_secondary_writes != 0) {
13502                         FREE_LOCK(&lk);
13503                         BO_UNLOCK(bo);
13504                         msleep(&mp->mnt_secondary_writes,
13505                                MNT_MTX(mp),
13506                                (PUSER - 1) | PDROP, "secwr", 0);
13507                         BO_LOCK(bo);
13508                         continue;
13509                 }
13510                 break;
13511         }
13512 
13513         /*
13514          * Reasons for needing more work before suspend:
13515          * - Dirty buffers on devvp.
13516          * - Softdep activity occurred after start of vnode sync loop
13517          * - Secondary writes occurred after start of vnode sync loop
13518          */
13519         error = 0;
13520         if (bo->bo_numoutput > 0 ||
13521             bo->bo_dirty.bv_cnt > 0 ||
13522             softdep_deps != 0 ||
13523             ump->softdep_deps != 0 ||
13524             softdep_accdeps != ump->softdep_accdeps ||
13525             secondary_writes != 0 ||
13526             mp->mnt_secondary_writes != 0 ||
13527             secondary_accwrites != mp->mnt_secondary_accwrites)
13528                 error = EAGAIN;
13529         FREE_LOCK(&lk);
13530         BO_UNLOCK(bo);
13531         return (error);
13532 }
13533 
13534 
13535 /*
13536  * Get the number of dependency structures for the file system, both
13537  * the current number and the total number allocated.  These will
13538  * later be used to detect that softdep processing has occurred.
13539  */
13540 void
13541 softdep_get_depcounts(struct mount *mp,
13542                       int *softdep_depsp,
13543                       int *softdep_accdepsp)
13544 {
13545         struct ufsmount *ump;
13546 
13547         ump = VFSTOUFS(mp);
13548         ACQUIRE_LOCK(&lk);
13549         *softdep_depsp = ump->softdep_deps;
13550         *softdep_accdepsp = ump->softdep_accdeps;
13551         FREE_LOCK(&lk);
13552 }
13553 
13554 /*
13555  * Wait for pending output on a vnode to complete.
13556  * Must be called with vnode lock and interlock locked.
13557  *
13558  * XXX: Should just be a call to bufobj_wwait().
13559  */
13560 static void
13561 drain_output(vp)
13562         struct vnode *vp;
13563 {
13564         struct bufobj *bo;
13565 
13566         bo = &vp->v_bufobj;
13567         ASSERT_VOP_LOCKED(vp, "drain_output");
13568         ASSERT_BO_WLOCKED(bo);
13569 
13570         while (bo->bo_numoutput) {
13571                 bo->bo_flag |= BO_WWAIT;
13572                 msleep((caddr_t)&bo->bo_numoutput,
13573                     BO_LOCKPTR(bo), PRIBIO + 1, "drainvp", 0);
13574         }
13575 }
13576 
13577 /*
13578  * Called whenever a buffer that is being invalidated or reallocated
13579  * contains dependencies. This should only happen if an I/O error has
13580  * occurred. The routine is called with the buffer locked.
13581  */ 
13582 static void
13583 softdep_deallocate_dependencies(bp)
13584         struct buf *bp;
13585 {
13586 
13587         if ((bp->b_ioflags & BIO_ERROR) == 0)
13588                 panic("softdep_deallocate_dependencies: dangling deps");
13589         if (bp->b_vp != NULL && bp->b_vp->v_mount != NULL)
13590                 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
13591         else
13592                 printf("softdep_deallocate_dependencies: "
13593                     "got error %d while accessing filesystem\n", bp->b_error);
13594         if (bp->b_error != ENXIO)
13595                 panic("softdep_deallocate_dependencies: unrecovered I/O error");
13596 }
13597 
13598 /*
13599  * Function to handle asynchronous write errors in the filesystem.
13600  */
13601 static void
13602 softdep_error(func, error)
13603         char *func;
13604         int error;
13605 {
13606 
13607         /* XXX should do something better! */
13608         printf("%s: got error %d while accessing filesystem\n", func, error);
13609 }
13610 
13611 #ifdef DDB
13612 
13613 static void
13614 inodedep_print(struct inodedep *inodedep, int verbose)
13615 {
13616         db_printf("%p fs %p st %x ino %jd inoblk %jd delta %d nlink %d"
13617             " saveino %p\n",
13618             inodedep, inodedep->id_fs, inodedep->id_state,
13619             (intmax_t)inodedep->id_ino,
13620             (intmax_t)fsbtodb(inodedep->id_fs,
13621             ino_to_fsba(inodedep->id_fs, inodedep->id_ino)),
13622             inodedep->id_nlinkdelta, inodedep->id_savednlink,
13623             inodedep->id_savedino1);
13624 
13625         if (verbose == 0)
13626                 return;
13627 
13628         db_printf("\tpendinghd %p, bufwait %p, inowait %p, inoreflst %p, "
13629             "mkdiradd %p\n",
13630             LIST_FIRST(&inodedep->id_pendinghd),
13631             LIST_FIRST(&inodedep->id_bufwait),
13632             LIST_FIRST(&inodedep->id_inowait),
13633             TAILQ_FIRST(&inodedep->id_inoreflst),
13634             inodedep->id_mkdiradd);
13635         db_printf("\tinoupdt %p, newinoupdt %p, extupdt %p, newextupdt %p\n",
13636             TAILQ_FIRST(&inodedep->id_inoupdt),
13637             TAILQ_FIRST(&inodedep->id_newinoupdt),
13638             TAILQ_FIRST(&inodedep->id_extupdt),
13639             TAILQ_FIRST(&inodedep->id_newextupdt));
13640 }
13641 
13642 DB_SHOW_COMMAND(inodedep, db_show_inodedep)
13643 {
13644 
13645         if (have_addr == 0) {
13646                 db_printf("Address required\n");
13647                 return;
13648         }
13649         inodedep_print((struct inodedep*)addr, 1);
13650 }
13651 
13652 DB_SHOW_COMMAND(inodedeps, db_show_inodedeps)
13653 {
13654         struct inodedep_hashhead *inodedephd;
13655         struct inodedep *inodedep;
13656         struct fs *fs;
13657         int cnt;
13658 
13659         fs = have_addr ? (struct fs *)addr : NULL;
13660         for (cnt = 0; cnt < inodedep_hash; cnt++) {
13661                 inodedephd = &inodedep_hashtbl[cnt];
13662                 LIST_FOREACH(inodedep, inodedephd, id_hash) {
13663                         if (fs != NULL && fs != inodedep->id_fs)
13664                                 continue;
13665                         inodedep_print(inodedep, 0);
13666                 }
13667         }
13668 }
13669 
13670 DB_SHOW_COMMAND(worklist, db_show_worklist)
13671 {
13672         struct worklist *wk;
13673 
13674         if (have_addr == 0) {
13675                 db_printf("Address required\n");
13676                 return;
13677         }
13678         wk = (struct worklist *)addr;
13679         printf("worklist: %p type %s state 0x%X\n",
13680             wk, TYPENAME(wk->wk_type), wk->wk_state);
13681 }
13682 
13683 DB_SHOW_COMMAND(workhead, db_show_workhead)
13684 {
13685         struct workhead *wkhd;
13686         struct worklist *wk;
13687         int i;
13688 
13689         if (have_addr == 0) {
13690                 db_printf("Address required\n");
13691                 return;
13692         }
13693         wkhd = (struct workhead *)addr;
13694         wk = LIST_FIRST(wkhd);
13695         for (i = 0; i < 100 && wk != NULL; i++, wk = LIST_NEXT(wk, wk_list))
13696                 db_printf("worklist: %p type %s state 0x%X",
13697                     wk, TYPENAME(wk->wk_type), wk->wk_state);
13698         if (i == 100)
13699                 db_printf("workhead overflow");
13700         printf("\n");
13701 }
13702 
13703 
13704 DB_SHOW_COMMAND(mkdirs, db_show_mkdirs)
13705 {
13706         struct jaddref *jaddref;
13707         struct diradd *diradd;
13708         struct mkdir *mkdir;
13709 
13710         LIST_FOREACH(mkdir, &mkdirlisthd, md_mkdirs) {
13711                 diradd = mkdir->md_diradd;
13712                 db_printf("mkdir: %p state 0x%X dap %p state 0x%X",
13713                     mkdir, mkdir->md_state, diradd, diradd->da_state);
13714                 if ((jaddref = mkdir->md_jaddref) != NULL)
13715                         db_printf(" jaddref %p jaddref state 0x%X",
13716                             jaddref, jaddref->ja_state);
13717                 db_printf("\n");
13718         }
13719 }
13720 
13721 #endif /* DDB */
13722 
13723 #endif /* SOFTUPDATES */

Cache object: fda6aab32a283239c7602fa43bbf33a8


[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]


This page is part of the FreeBSD/Linux Linux Kernel Cross-Reference, and was automatically generated using a modified version of the LXR engine.