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
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("$FreeBSD$"); 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/racct.h> 73 #include <sys/rwlock.h> 74 #include <sys/stat.h> 75 #include <sys/sysctl.h> 76 #include <sys/syslog.h> 77 #include <sys/vnode.h> 78 #include <sys/conf.h> 79 80 #include <ufs/ufs/dir.h> 81 #include <ufs/ufs/extattr.h> 82 #include <ufs/ufs/quota.h> 83 #include <ufs/ufs/inode.h> 84 #include <ufs/ufs/ufsmount.h> 85 #include <ufs/ffs/fs.h> 86 #include <ufs/ffs/softdep.h> 87 #include <ufs/ffs/ffs_extern.h> 88 #include <ufs/ufs/ufs_extern.h> 89 90 #include <vm/vm.h> 91 #include <vm/vm_extern.h> 92 #include <vm/vm_object.h> 93 94 #include <geom/geom.h> 95 96 #include <ddb/ddb.h> 97 98 #define KTR_SUJ 0 /* Define to KTR_SPARE. */ 99 100 #ifndef SOFTUPDATES 101 102 int 103 softdep_flushfiles(oldmnt, flags, td) 104 struct mount *oldmnt; 105 int flags; 106 struct thread *td; 107 { 108 109 panic("softdep_flushfiles called"); 110 } 111 112 int 113 softdep_mount(devvp, mp, fs, cred) 114 struct vnode *devvp; 115 struct mount *mp; 116 struct fs *fs; 117 struct ucred *cred; 118 { 119 120 return (0); 121 } 122 123 void 124 softdep_initialize() 125 { 126 127 return; 128 } 129 130 void 131 softdep_uninitialize() 132 { 133 134 return; 135 } 136 137 void 138 softdep_unmount(mp) 139 struct mount *mp; 140 { 141 142 panic("softdep_unmount called"); 143 } 144 145 void 146 softdep_setup_sbupdate(ump, fs, bp) 147 struct ufsmount *ump; 148 struct fs *fs; 149 struct buf *bp; 150 { 151 152 panic("softdep_setup_sbupdate called"); 153 } 154 155 void 156 softdep_setup_inomapdep(bp, ip, newinum, mode) 157 struct buf *bp; 158 struct inode *ip; 159 ino_t newinum; 160 int mode; 161 { 162 163 panic("softdep_setup_inomapdep called"); 164 } 165 166 void 167 softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags) 168 struct buf *bp; 169 struct mount *mp; 170 ufs2_daddr_t newblkno; 171 int frags; 172 int oldfrags; 173 { 174 175 panic("softdep_setup_blkmapdep called"); 176 } 177 178 void 179 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 180 struct inode *ip; 181 ufs_lbn_t lbn; 182 ufs2_daddr_t newblkno; 183 ufs2_daddr_t oldblkno; 184 long newsize; 185 long oldsize; 186 struct buf *bp; 187 { 188 189 panic("softdep_setup_allocdirect called"); 190 } 191 192 void 193 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp) 194 struct inode *ip; 195 ufs_lbn_t lbn; 196 ufs2_daddr_t newblkno; 197 ufs2_daddr_t oldblkno; 198 long newsize; 199 long oldsize; 200 struct buf *bp; 201 { 202 203 panic("softdep_setup_allocext called"); 204 } 205 206 void 207 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 208 struct inode *ip; 209 ufs_lbn_t lbn; 210 struct buf *bp; 211 int ptrno; 212 ufs2_daddr_t newblkno; 213 ufs2_daddr_t oldblkno; 214 struct buf *nbp; 215 { 216 217 panic("softdep_setup_allocindir_page called"); 218 } 219 220 void 221 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 222 struct buf *nbp; 223 struct inode *ip; 224 struct buf *bp; 225 int ptrno; 226 ufs2_daddr_t newblkno; 227 { 228 229 panic("softdep_setup_allocindir_meta called"); 230 } 231 232 void 233 softdep_journal_freeblocks(ip, cred, length, flags) 234 struct inode *ip; 235 struct ucred *cred; 236 off_t length; 237 int flags; 238 { 239 240 panic("softdep_journal_freeblocks called"); 241 } 242 243 void 244 softdep_journal_fsync(ip) 245 struct inode *ip; 246 { 247 248 panic("softdep_journal_fsync called"); 249 } 250 251 void 252 softdep_setup_freeblocks(ip, length, flags) 253 struct inode *ip; 254 off_t length; 255 int flags; 256 { 257 258 panic("softdep_setup_freeblocks called"); 259 } 260 261 void 262 softdep_freefile(pvp, ino, mode) 263 struct vnode *pvp; 264 ino_t ino; 265 int mode; 266 { 267 268 panic("softdep_freefile called"); 269 } 270 271 int 272 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 273 struct buf *bp; 274 struct inode *dp; 275 off_t diroffset; 276 ino_t newinum; 277 struct buf *newdirbp; 278 int isnewblk; 279 { 280 281 panic("softdep_setup_directory_add called"); 282 } 283 284 void 285 softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize) 286 struct buf *bp; 287 struct inode *dp; 288 caddr_t base; 289 caddr_t oldloc; 290 caddr_t newloc; 291 int entrysize; 292 { 293 294 panic("softdep_change_directoryentry_offset called"); 295 } 296 297 void 298 softdep_setup_remove(bp, dp, ip, isrmdir) 299 struct buf *bp; 300 struct inode *dp; 301 struct inode *ip; 302 int isrmdir; 303 { 304 305 panic("softdep_setup_remove called"); 306 } 307 308 void 309 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 310 struct buf *bp; 311 struct inode *dp; 312 struct inode *ip; 313 ino_t newinum; 314 int isrmdir; 315 { 316 317 panic("softdep_setup_directory_change called"); 318 } 319 320 void 321 softdep_setup_blkfree(mp, bp, blkno, frags, wkhd) 322 struct mount *mp; 323 struct buf *bp; 324 ufs2_daddr_t blkno; 325 int frags; 326 struct workhead *wkhd; 327 { 328 329 panic("%s called", __FUNCTION__); 330 } 331 332 void 333 softdep_setup_inofree(mp, bp, ino, wkhd) 334 struct mount *mp; 335 struct buf *bp; 336 ino_t ino; 337 struct workhead *wkhd; 338 { 339 340 panic("%s called", __FUNCTION__); 341 } 342 343 void 344 softdep_setup_unlink(dp, ip) 345 struct inode *dp; 346 struct inode *ip; 347 { 348 349 panic("%s called", __FUNCTION__); 350 } 351 352 void 353 softdep_setup_link(dp, ip) 354 struct inode *dp; 355 struct inode *ip; 356 { 357 358 panic("%s called", __FUNCTION__); 359 } 360 361 void 362 softdep_revert_link(dp, ip) 363 struct inode *dp; 364 struct inode *ip; 365 { 366 367 panic("%s called", __FUNCTION__); 368 } 369 370 void 371 softdep_setup_rmdir(dp, ip) 372 struct inode *dp; 373 struct inode *ip; 374 { 375 376 panic("%s called", __FUNCTION__); 377 } 378 379 void 380 softdep_revert_rmdir(dp, ip) 381 struct inode *dp; 382 struct inode *ip; 383 { 384 385 panic("%s called", __FUNCTION__); 386 } 387 388 void 389 softdep_setup_create(dp, ip) 390 struct inode *dp; 391 struct inode *ip; 392 { 393 394 panic("%s called", __FUNCTION__); 395 } 396 397 void 398 softdep_revert_create(dp, ip) 399 struct inode *dp; 400 struct inode *ip; 401 { 402 403 panic("%s called", __FUNCTION__); 404 } 405 406 void 407 softdep_setup_mkdir(dp, ip) 408 struct inode *dp; 409 struct inode *ip; 410 { 411 412 panic("%s called", __FUNCTION__); 413 } 414 415 void 416 softdep_revert_mkdir(dp, ip) 417 struct inode *dp; 418 struct inode *ip; 419 { 420 421 panic("%s called", __FUNCTION__); 422 } 423 424 void 425 softdep_setup_dotdot_link(dp, ip) 426 struct inode *dp; 427 struct inode *ip; 428 { 429 430 panic("%s called", __FUNCTION__); 431 } 432 433 int 434 softdep_prealloc(vp, waitok) 435 struct vnode *vp; 436 int waitok; 437 { 438 439 panic("%s called", __FUNCTION__); 440 } 441 442 int 443 softdep_journal_lookup(mp, vpp) 444 struct mount *mp; 445 struct vnode **vpp; 446 { 447 448 return (ENOENT); 449 } 450 451 void 452 softdep_change_linkcnt(ip) 453 struct inode *ip; 454 { 455 456 panic("softdep_change_linkcnt called"); 457 } 458 459 void 460 softdep_load_inodeblock(ip) 461 struct inode *ip; 462 { 463 464 panic("softdep_load_inodeblock called"); 465 } 466 467 void 468 softdep_update_inodeblock(ip, bp, waitfor) 469 struct inode *ip; 470 struct buf *bp; 471 int waitfor; 472 { 473 474 panic("softdep_update_inodeblock called"); 475 } 476 477 int 478 softdep_fsync(vp) 479 struct vnode *vp; /* the "in_core" copy of the inode */ 480 { 481 482 return (0); 483 } 484 485 void 486 softdep_fsync_mountdev(vp) 487 struct vnode *vp; 488 { 489 490 return; 491 } 492 493 int 494 softdep_flushworklist(oldmnt, countp, td) 495 struct mount *oldmnt; 496 int *countp; 497 struct thread *td; 498 { 499 500 *countp = 0; 501 return (0); 502 } 503 504 int 505 softdep_sync_metadata(struct vnode *vp) 506 { 507 508 panic("softdep_sync_metadata called"); 509 } 510 511 int 512 softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor) 513 { 514 515 panic("softdep_sync_buf called"); 516 } 517 518 int 519 softdep_slowdown(vp) 520 struct vnode *vp; 521 { 522 523 panic("softdep_slowdown called"); 524 } 525 526 int 527 softdep_request_cleanup(fs, vp, cred, resource) 528 struct fs *fs; 529 struct vnode *vp; 530 struct ucred *cred; 531 int resource; 532 { 533 534 return (0); 535 } 536 537 int 538 softdep_check_suspend(struct mount *mp, 539 struct vnode *devvp, 540 int softdep_depcnt, 541 int softdep_accdepcnt, 542 int secondary_writes, 543 int secondary_accwrites) 544 { 545 struct bufobj *bo; 546 int error; 547 548 (void) softdep_depcnt, 549 (void) softdep_accdepcnt; 550 551 bo = &devvp->v_bufobj; 552 ASSERT_BO_WLOCKED(bo); 553 554 MNT_ILOCK(mp); 555 while (mp->mnt_secondary_writes != 0) { 556 BO_UNLOCK(bo); 557 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp), 558 (PUSER - 1) | PDROP, "secwr", 0); 559 BO_LOCK(bo); 560 MNT_ILOCK(mp); 561 } 562 563 /* 564 * Reasons for needing more work before suspend: 565 * - Dirty buffers on devvp. 566 * - Secondary writes occurred after start of vnode sync loop 567 */ 568 error = 0; 569 if (bo->bo_numoutput > 0 || 570 bo->bo_dirty.bv_cnt > 0 || 571 secondary_writes != 0 || 572 mp->mnt_secondary_writes != 0 || 573 secondary_accwrites != mp->mnt_secondary_accwrites) 574 error = EAGAIN; 575 BO_UNLOCK(bo); 576 return (error); 577 } 578 579 void 580 softdep_get_depcounts(struct mount *mp, 581 int *softdepactivep, 582 int *softdepactiveaccp) 583 { 584 (void) mp; 585 *softdepactivep = 0; 586 *softdepactiveaccp = 0; 587 } 588 589 void 590 softdep_buf_append(bp, wkhd) 591 struct buf *bp; 592 struct workhead *wkhd; 593 { 594 595 panic("softdep_buf_appendwork called"); 596 } 597 598 void 599 softdep_inode_append(ip, cred, wkhd) 600 struct inode *ip; 601 struct ucred *cred; 602 struct workhead *wkhd; 603 { 604 605 panic("softdep_inode_appendwork called"); 606 } 607 608 void 609 softdep_freework(wkhd) 610 struct workhead *wkhd; 611 { 612 613 panic("softdep_freework called"); 614 } 615 616 #else 617 618 FEATURE(softupdates, "FFS soft-updates support"); 619 620 static SYSCTL_NODE(_debug, OID_AUTO, softdep, CTLFLAG_RW, 0, 621 "soft updates stats"); 622 static SYSCTL_NODE(_debug_softdep, OID_AUTO, total, CTLFLAG_RW, 0, 623 "total dependencies allocated"); 624 static SYSCTL_NODE(_debug_softdep, OID_AUTO, highuse, CTLFLAG_RW, 0, 625 "high use dependencies allocated"); 626 static SYSCTL_NODE(_debug_softdep, OID_AUTO, current, CTLFLAG_RW, 0, 627 "current dependencies allocated"); 628 static SYSCTL_NODE(_debug_softdep, OID_AUTO, write, CTLFLAG_RW, 0, 629 "current dependencies written"); 630 631 unsigned long dep_current[D_LAST + 1]; 632 unsigned long dep_highuse[D_LAST + 1]; 633 unsigned long dep_total[D_LAST + 1]; 634 unsigned long dep_write[D_LAST + 1]; 635 636 #define SOFTDEP_TYPE(type, str, long) \ 637 static MALLOC_DEFINE(M_ ## type, #str, long); \ 638 SYSCTL_ULONG(_debug_softdep_total, OID_AUTO, str, CTLFLAG_RD, \ 639 &dep_total[D_ ## type], 0, ""); \ 640 SYSCTL_ULONG(_debug_softdep_current, OID_AUTO, str, CTLFLAG_RD, \ 641 &dep_current[D_ ## type], 0, ""); \ 642 SYSCTL_ULONG(_debug_softdep_highuse, OID_AUTO, str, CTLFLAG_RD, \ 643 &dep_highuse[D_ ## type], 0, ""); \ 644 SYSCTL_ULONG(_debug_softdep_write, OID_AUTO, str, CTLFLAG_RD, \ 645 &dep_write[D_ ## type], 0, ""); 646 647 SOFTDEP_TYPE(PAGEDEP, pagedep, "File page dependencies"); 648 SOFTDEP_TYPE(INODEDEP, inodedep, "Inode dependencies"); 649 SOFTDEP_TYPE(BMSAFEMAP, bmsafemap, 650 "Block or frag allocated from cyl group map"); 651 SOFTDEP_TYPE(NEWBLK, newblk, "New block or frag allocation dependency"); 652 SOFTDEP_TYPE(ALLOCDIRECT, allocdirect, "Block or frag dependency for an inode"); 653 SOFTDEP_TYPE(INDIRDEP, indirdep, "Indirect block dependencies"); 654 SOFTDEP_TYPE(ALLOCINDIR, allocindir, "Block dependency for an indirect block"); 655 SOFTDEP_TYPE(FREEFRAG, freefrag, "Previously used frag for an inode"); 656 SOFTDEP_TYPE(FREEBLKS, freeblks, "Blocks freed from an inode"); 657 SOFTDEP_TYPE(FREEFILE, freefile, "Inode deallocated"); 658 SOFTDEP_TYPE(DIRADD, diradd, "New directory entry"); 659 SOFTDEP_TYPE(MKDIR, mkdir, "New directory"); 660 SOFTDEP_TYPE(DIRREM, dirrem, "Directory entry deleted"); 661 SOFTDEP_TYPE(NEWDIRBLK, newdirblk, "Unclaimed new directory block"); 662 SOFTDEP_TYPE(FREEWORK, freework, "free an inode block"); 663 SOFTDEP_TYPE(FREEDEP, freedep, "track a block free"); 664 SOFTDEP_TYPE(JADDREF, jaddref, "Journal inode ref add"); 665 SOFTDEP_TYPE(JREMREF, jremref, "Journal inode ref remove"); 666 SOFTDEP_TYPE(JMVREF, jmvref, "Journal inode ref move"); 667 SOFTDEP_TYPE(JNEWBLK, jnewblk, "Journal new block"); 668 SOFTDEP_TYPE(JFREEBLK, jfreeblk, "Journal free block"); 669 SOFTDEP_TYPE(JFREEFRAG, jfreefrag, "Journal free frag"); 670 SOFTDEP_TYPE(JSEG, jseg, "Journal segment"); 671 SOFTDEP_TYPE(JSEGDEP, jsegdep, "Journal segment complete"); 672 SOFTDEP_TYPE(SBDEP, sbdep, "Superblock write dependency"); 673 SOFTDEP_TYPE(JTRUNC, jtrunc, "Journal inode truncation"); 674 SOFTDEP_TYPE(JFSYNC, jfsync, "Journal fsync complete"); 675 676 static MALLOC_DEFINE(M_SENTINEL, "sentinel", "Worklist sentinel"); 677 678 static MALLOC_DEFINE(M_SAVEDINO, "savedino", "Saved inodes"); 679 static MALLOC_DEFINE(M_JBLOCKS, "jblocks", "Journal block locations"); 680 static MALLOC_DEFINE(M_MOUNTDATA, "softdep", "Softdep per-mount data"); 681 682 #define M_SOFTDEP_FLAGS (M_WAITOK) 683 684 /* 685 * translate from workitem type to memory type 686 * MUST match the defines above, such that memtype[D_XXX] == M_XXX 687 */ 688 static struct malloc_type *memtype[] = { 689 M_PAGEDEP, 690 M_INODEDEP, 691 M_BMSAFEMAP, 692 M_NEWBLK, 693 M_ALLOCDIRECT, 694 M_INDIRDEP, 695 M_ALLOCINDIR, 696 M_FREEFRAG, 697 M_FREEBLKS, 698 M_FREEFILE, 699 M_DIRADD, 700 M_MKDIR, 701 M_DIRREM, 702 M_NEWDIRBLK, 703 M_FREEWORK, 704 M_FREEDEP, 705 M_JADDREF, 706 M_JREMREF, 707 M_JMVREF, 708 M_JNEWBLK, 709 M_JFREEBLK, 710 M_JFREEFRAG, 711 M_JSEG, 712 M_JSEGDEP, 713 M_SBDEP, 714 M_JTRUNC, 715 M_JFSYNC, 716 M_SENTINEL 717 }; 718 719 #define DtoM(type) (memtype[type]) 720 721 /* 722 * Names of malloc types. 723 */ 724 #define TYPENAME(type) \ 725 ((unsigned)(type) <= D_LAST ? memtype[type]->ks_shortdesc : "???") 726 /* 727 * End system adaptation definitions. 728 */ 729 730 #define DOTDOT_OFFSET offsetof(struct dirtemplate, dotdot_ino) 731 #define DOT_OFFSET offsetof(struct dirtemplate, dot_ino) 732 733 /* 734 * Internal function prototypes. 735 */ 736 static void check_clear_deps(struct mount *); 737 static void softdep_error(char *, int); 738 static int softdep_process_worklist(struct mount *, int); 739 static int softdep_waitidle(struct mount *, int); 740 static void drain_output(struct vnode *); 741 static struct buf *getdirtybuf(struct buf *, struct rwlock *, int); 742 static int check_inodedep_free(struct inodedep *); 743 static void clear_remove(struct mount *); 744 static void clear_inodedeps(struct mount *); 745 static void unlinked_inodedep(struct mount *, struct inodedep *); 746 static void clear_unlinked_inodedep(struct inodedep *); 747 static struct inodedep *first_unlinked_inodedep(struct ufsmount *); 748 static int flush_pagedep_deps(struct vnode *, struct mount *, 749 struct diraddhd *); 750 static int free_pagedep(struct pagedep *); 751 static int flush_newblk_dep(struct vnode *, struct mount *, ufs_lbn_t); 752 static int flush_inodedep_deps(struct vnode *, struct mount *, ino_t); 753 static int flush_deplist(struct allocdirectlst *, int, int *); 754 static int sync_cgs(struct mount *, int); 755 static int handle_written_filepage(struct pagedep *, struct buf *, int); 756 static int handle_written_sbdep(struct sbdep *, struct buf *); 757 static void initiate_write_sbdep(struct sbdep *); 758 static void diradd_inode_written(struct diradd *, struct inodedep *); 759 static int handle_written_indirdep(struct indirdep *, struct buf *, 760 struct buf**, int); 761 static int handle_written_inodeblock(struct inodedep *, struct buf *, int); 762 static int jnewblk_rollforward(struct jnewblk *, struct fs *, struct cg *, 763 uint8_t *); 764 static int handle_written_bmsafemap(struct bmsafemap *, struct buf *, int); 765 static void handle_written_jaddref(struct jaddref *); 766 static void handle_written_jremref(struct jremref *); 767 static void handle_written_jseg(struct jseg *, struct buf *); 768 static void handle_written_jnewblk(struct jnewblk *); 769 static void handle_written_jblkdep(struct jblkdep *); 770 static void handle_written_jfreefrag(struct jfreefrag *); 771 static void complete_jseg(struct jseg *); 772 static void complete_jsegs(struct jseg *); 773 static void jseg_write(struct ufsmount *ump, struct jseg *, uint8_t *); 774 static void jaddref_write(struct jaddref *, struct jseg *, uint8_t *); 775 static void jremref_write(struct jremref *, struct jseg *, uint8_t *); 776 static void jmvref_write(struct jmvref *, struct jseg *, uint8_t *); 777 static void jtrunc_write(struct jtrunc *, struct jseg *, uint8_t *); 778 static void jfsync_write(struct jfsync *, struct jseg *, uint8_t *data); 779 static void jnewblk_write(struct jnewblk *, struct jseg *, uint8_t *); 780 static void jfreeblk_write(struct jfreeblk *, struct jseg *, uint8_t *); 781 static void jfreefrag_write(struct jfreefrag *, struct jseg *, uint8_t *); 782 static inline void inoref_write(struct inoref *, struct jseg *, 783 struct jrefrec *); 784 static void handle_allocdirect_partdone(struct allocdirect *, 785 struct workhead *); 786 static struct jnewblk *cancel_newblk(struct newblk *, struct worklist *, 787 struct workhead *); 788 static void indirdep_complete(struct indirdep *); 789 static int indirblk_lookup(struct mount *, ufs2_daddr_t); 790 static void indirblk_insert(struct freework *); 791 static void indirblk_remove(struct freework *); 792 static void handle_allocindir_partdone(struct allocindir *); 793 static void initiate_write_filepage(struct pagedep *, struct buf *); 794 static void initiate_write_indirdep(struct indirdep*, struct buf *); 795 static void handle_written_mkdir(struct mkdir *, int); 796 static int jnewblk_rollback(struct jnewblk *, struct fs *, struct cg *, 797 uint8_t *); 798 static void initiate_write_bmsafemap(struct bmsafemap *, struct buf *); 799 static void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *); 800 static void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *); 801 static void handle_workitem_freefile(struct freefile *); 802 static int handle_workitem_remove(struct dirrem *, int); 803 static struct dirrem *newdirrem(struct buf *, struct inode *, 804 struct inode *, int, struct dirrem **); 805 static struct indirdep *indirdep_lookup(struct mount *, struct inode *, 806 struct buf *); 807 static void cancel_indirdep(struct indirdep *, struct buf *, 808 struct freeblks *); 809 static void free_indirdep(struct indirdep *); 810 static void free_diradd(struct diradd *, struct workhead *); 811 static void merge_diradd(struct inodedep *, struct diradd *); 812 static void complete_diradd(struct diradd *); 813 static struct diradd *diradd_lookup(struct pagedep *, int); 814 static struct jremref *cancel_diradd_dotdot(struct inode *, struct dirrem *, 815 struct jremref *); 816 static struct jremref *cancel_mkdir_dotdot(struct inode *, struct dirrem *, 817 struct jremref *); 818 static void cancel_diradd(struct diradd *, struct dirrem *, struct jremref *, 819 struct jremref *, struct jremref *); 820 static void dirrem_journal(struct dirrem *, struct jremref *, struct jremref *, 821 struct jremref *); 822 static void cancel_allocindir(struct allocindir *, struct buf *bp, 823 struct freeblks *, int); 824 static int setup_trunc_indir(struct freeblks *, struct inode *, 825 ufs_lbn_t, ufs_lbn_t, ufs2_daddr_t); 826 static void complete_trunc_indir(struct freework *); 827 static void trunc_indirdep(struct indirdep *, struct freeblks *, struct buf *, 828 int); 829 static void complete_mkdir(struct mkdir *); 830 static void free_newdirblk(struct newdirblk *); 831 static void free_jremref(struct jremref *); 832 static void free_jaddref(struct jaddref *); 833 static void free_jsegdep(struct jsegdep *); 834 static void free_jsegs(struct jblocks *); 835 static void rele_jseg(struct jseg *); 836 static void free_jseg(struct jseg *, struct jblocks *); 837 static void free_jnewblk(struct jnewblk *); 838 static void free_jblkdep(struct jblkdep *); 839 static void free_jfreefrag(struct jfreefrag *); 840 static void free_freedep(struct freedep *); 841 static void journal_jremref(struct dirrem *, struct jremref *, 842 struct inodedep *); 843 static void cancel_jnewblk(struct jnewblk *, struct workhead *); 844 static int cancel_jaddref(struct jaddref *, struct inodedep *, 845 struct workhead *); 846 static void cancel_jfreefrag(struct jfreefrag *); 847 static inline void setup_freedirect(struct freeblks *, struct inode *, 848 int, int); 849 static inline void setup_freeext(struct freeblks *, struct inode *, int, int); 850 static inline void setup_freeindir(struct freeblks *, struct inode *, int, 851 ufs_lbn_t, int); 852 static inline struct freeblks *newfreeblks(struct mount *, struct inode *); 853 static void freeblks_free(struct ufsmount *, struct freeblks *, int); 854 static void indir_trunc(struct freework *, ufs2_daddr_t, ufs_lbn_t); 855 static ufs2_daddr_t blkcount(struct fs *, ufs2_daddr_t, off_t); 856 static int trunc_check_buf(struct buf *, int *, ufs_lbn_t, int, int); 857 static void trunc_dependencies(struct inode *, struct freeblks *, ufs_lbn_t, 858 int, int); 859 static void trunc_pages(struct inode *, off_t, ufs2_daddr_t, int); 860 static int cancel_pagedep(struct pagedep *, struct freeblks *, int); 861 static int deallocate_dependencies(struct buf *, struct freeblks *, int); 862 static void newblk_freefrag(struct newblk*); 863 static void free_newblk(struct newblk *); 864 static void cancel_allocdirect(struct allocdirectlst *, 865 struct allocdirect *, struct freeblks *); 866 static int check_inode_unwritten(struct inodedep *); 867 static int free_inodedep(struct inodedep *); 868 static void freework_freeblock(struct freework *); 869 static void freework_enqueue(struct freework *); 870 static int handle_workitem_freeblocks(struct freeblks *, int); 871 static int handle_complete_freeblocks(struct freeblks *, int); 872 static void handle_workitem_indirblk(struct freework *); 873 static void handle_written_freework(struct freework *); 874 static void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *); 875 static struct worklist *jnewblk_merge(struct worklist *, struct worklist *, 876 struct workhead *); 877 static struct freefrag *setup_allocindir_phase2(struct buf *, struct inode *, 878 struct inodedep *, struct allocindir *, ufs_lbn_t); 879 static struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t, 880 ufs2_daddr_t, ufs_lbn_t); 881 static void handle_workitem_freefrag(struct freefrag *); 882 static struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long, 883 ufs_lbn_t); 884 static void allocdirect_merge(struct allocdirectlst *, 885 struct allocdirect *, struct allocdirect *); 886 static struct freefrag *allocindir_merge(struct allocindir *, 887 struct allocindir *); 888 static int bmsafemap_find(struct bmsafemap_hashhead *, int, 889 struct bmsafemap **); 890 static struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *, 891 int cg, struct bmsafemap *); 892 static int newblk_find(struct newblk_hashhead *, ufs2_daddr_t, int, 893 struct newblk **); 894 static int newblk_lookup(struct mount *, ufs2_daddr_t, int, struct newblk **); 895 static int inodedep_find(struct inodedep_hashhead *, ino_t, 896 struct inodedep **); 897 static int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **); 898 static int pagedep_lookup(struct mount *, struct buf *bp, ino_t, ufs_lbn_t, 899 int, struct pagedep **); 900 static int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t, 901 struct pagedep **); 902 static void pause_timer(void *); 903 static int request_cleanup(struct mount *, int); 904 static int softdep_request_cleanup_flush(struct mount *, struct ufsmount *); 905 static void schedule_cleanup(struct mount *); 906 static void softdep_ast_cleanup_proc(struct thread *); 907 static struct ufsmount *softdep_bp_to_mp(struct buf *bp); 908 static int process_worklist_item(struct mount *, int, int); 909 static void process_removes(struct vnode *); 910 static void process_truncates(struct vnode *); 911 static void jwork_move(struct workhead *, struct workhead *); 912 static void jwork_insert(struct workhead *, struct jsegdep *); 913 static void add_to_worklist(struct worklist *, int); 914 static void wake_worklist(struct worklist *); 915 static void wait_worklist(struct worklist *, char *); 916 static void remove_from_worklist(struct worklist *); 917 static void softdep_flush(void *); 918 static void softdep_flushjournal(struct mount *); 919 static int softdep_speedup(struct ufsmount *); 920 static void worklist_speedup(struct mount *); 921 static int journal_mount(struct mount *, struct fs *, struct ucred *); 922 static void journal_unmount(struct ufsmount *); 923 static int journal_space(struct ufsmount *, int); 924 static void journal_suspend(struct ufsmount *); 925 static int journal_unsuspend(struct ufsmount *ump); 926 static void softdep_prelink(struct vnode *, struct vnode *); 927 static void add_to_journal(struct worklist *); 928 static void remove_from_journal(struct worklist *); 929 static bool softdep_excess_items(struct ufsmount *, int); 930 static void softdep_process_journal(struct mount *, struct worklist *, int); 931 static struct jremref *newjremref(struct dirrem *, struct inode *, 932 struct inode *ip, off_t, nlink_t); 933 static struct jaddref *newjaddref(struct inode *, ino_t, off_t, int16_t, 934 uint16_t); 935 static inline void newinoref(struct inoref *, ino_t, ino_t, off_t, nlink_t, 936 uint16_t); 937 static inline struct jsegdep *inoref_jseg(struct inoref *); 938 static struct jmvref *newjmvref(struct inode *, ino_t, off_t, off_t); 939 static struct jfreeblk *newjfreeblk(struct freeblks *, ufs_lbn_t, 940 ufs2_daddr_t, int); 941 static void adjust_newfreework(struct freeblks *, int); 942 static struct jtrunc *newjtrunc(struct freeblks *, off_t, int); 943 static void move_newblock_dep(struct jaddref *, struct inodedep *); 944 static void cancel_jfreeblk(struct freeblks *, ufs2_daddr_t); 945 static struct jfreefrag *newjfreefrag(struct freefrag *, struct inode *, 946 ufs2_daddr_t, long, ufs_lbn_t); 947 static struct freework *newfreework(struct ufsmount *, struct freeblks *, 948 struct freework *, ufs_lbn_t, ufs2_daddr_t, int, int, int); 949 static int jwait(struct worklist *, int); 950 static struct inodedep *inodedep_lookup_ip(struct inode *); 951 static int bmsafemap_backgroundwrite(struct bmsafemap *, struct buf *); 952 static struct freefile *handle_bufwait(struct inodedep *, struct workhead *); 953 static void handle_jwork(struct workhead *); 954 static struct mkdir *setup_newdir(struct diradd *, ino_t, ino_t, struct buf *, 955 struct mkdir **); 956 static struct jblocks *jblocks_create(void); 957 static ufs2_daddr_t jblocks_alloc(struct jblocks *, int, int *); 958 static void jblocks_free(struct jblocks *, struct mount *, int); 959 static void jblocks_destroy(struct jblocks *); 960 static void jblocks_add(struct jblocks *, ufs2_daddr_t, int); 961 962 /* 963 * Exported softdep operations. 964 */ 965 static void softdep_disk_io_initiation(struct buf *); 966 static void softdep_disk_write_complete(struct buf *); 967 static void softdep_deallocate_dependencies(struct buf *); 968 static int softdep_count_dependencies(struct buf *bp, int); 969 970 /* 971 * Global lock over all of soft updates. 972 */ 973 static struct mtx lk; 974 MTX_SYSINIT(softdep_lock, &lk, "Global Softdep Lock", MTX_DEF); 975 976 #define ACQUIRE_GBLLOCK(lk) mtx_lock(lk) 977 #define FREE_GBLLOCK(lk) mtx_unlock(lk) 978 #define GBLLOCK_OWNED(lk) mtx_assert((lk), MA_OWNED) 979 980 /* 981 * Per-filesystem soft-updates locking. 982 */ 983 #define LOCK_PTR(ump) (&(ump)->um_softdep->sd_fslock) 984 #define TRY_ACQUIRE_LOCK(ump) rw_try_wlock(&(ump)->um_softdep->sd_fslock) 985 #define ACQUIRE_LOCK(ump) rw_wlock(&(ump)->um_softdep->sd_fslock) 986 #define FREE_LOCK(ump) rw_wunlock(&(ump)->um_softdep->sd_fslock) 987 #define LOCK_OWNED(ump) rw_assert(&(ump)->um_softdep->sd_fslock, \ 988 RA_WLOCKED) 989 990 #define BUF_AREC(bp) lockallowrecurse(&(bp)->b_lock) 991 #define BUF_NOREC(bp) lockdisablerecurse(&(bp)->b_lock) 992 993 /* 994 * Worklist queue management. 995 * These routines require that the lock be held. 996 */ 997 #ifndef /* NOT */ DEBUG 998 #define WORKLIST_INSERT(head, item) do { \ 999 (item)->wk_state |= ONWORKLIST; \ 1000 LIST_INSERT_HEAD(head, item, wk_list); \ 1001 } while (0) 1002 #define WORKLIST_REMOVE(item) do { \ 1003 (item)->wk_state &= ~ONWORKLIST; \ 1004 LIST_REMOVE(item, wk_list); \ 1005 } while (0) 1006 #define WORKLIST_INSERT_UNLOCKED WORKLIST_INSERT 1007 #define WORKLIST_REMOVE_UNLOCKED WORKLIST_REMOVE 1008 1009 #else /* DEBUG */ 1010 static void worklist_insert(struct workhead *, struct worklist *, int); 1011 static void worklist_remove(struct worklist *, int); 1012 1013 #define WORKLIST_INSERT(head, item) worklist_insert(head, item, 1) 1014 #define WORKLIST_INSERT_UNLOCKED(head, item) worklist_insert(head, item, 0) 1015 #define WORKLIST_REMOVE(item) worklist_remove(item, 1) 1016 #define WORKLIST_REMOVE_UNLOCKED(item) worklist_remove(item, 0) 1017 1018 static void 1019 worklist_insert(head, item, locked) 1020 struct workhead *head; 1021 struct worklist *item; 1022 int locked; 1023 { 1024 1025 if (locked) 1026 LOCK_OWNED(VFSTOUFS(item->wk_mp)); 1027 if (item->wk_state & ONWORKLIST) 1028 panic("worklist_insert: %p %s(0x%X) already on list", 1029 item, TYPENAME(item->wk_type), item->wk_state); 1030 item->wk_state |= ONWORKLIST; 1031 LIST_INSERT_HEAD(head, item, wk_list); 1032 } 1033 1034 static void 1035 worklist_remove(item, locked) 1036 struct worklist *item; 1037 int locked; 1038 { 1039 1040 if (locked) 1041 LOCK_OWNED(VFSTOUFS(item->wk_mp)); 1042 if ((item->wk_state & ONWORKLIST) == 0) 1043 panic("worklist_remove: %p %s(0x%X) not on list", 1044 item, TYPENAME(item->wk_type), item->wk_state); 1045 item->wk_state &= ~ONWORKLIST; 1046 LIST_REMOVE(item, wk_list); 1047 } 1048 #endif /* DEBUG */ 1049 1050 /* 1051 * Merge two jsegdeps keeping only the oldest one as newer references 1052 * can't be discarded until after older references. 1053 */ 1054 static inline struct jsegdep * 1055 jsegdep_merge(struct jsegdep *one, struct jsegdep *two) 1056 { 1057 struct jsegdep *swp; 1058 1059 if (two == NULL) 1060 return (one); 1061 1062 if (one->jd_seg->js_seq > two->jd_seg->js_seq) { 1063 swp = one; 1064 one = two; 1065 two = swp; 1066 } 1067 WORKLIST_REMOVE(&two->jd_list); 1068 free_jsegdep(two); 1069 1070 return (one); 1071 } 1072 1073 /* 1074 * If two freedeps are compatible free one to reduce list size. 1075 */ 1076 static inline struct freedep * 1077 freedep_merge(struct freedep *one, struct freedep *two) 1078 { 1079 if (two == NULL) 1080 return (one); 1081 1082 if (one->fd_freework == two->fd_freework) { 1083 WORKLIST_REMOVE(&two->fd_list); 1084 free_freedep(two); 1085 } 1086 return (one); 1087 } 1088 1089 /* 1090 * Move journal work from one list to another. Duplicate freedeps and 1091 * jsegdeps are coalesced to keep the lists as small as possible. 1092 */ 1093 static void 1094 jwork_move(dst, src) 1095 struct workhead *dst; 1096 struct workhead *src; 1097 { 1098 struct freedep *freedep; 1099 struct jsegdep *jsegdep; 1100 struct worklist *wkn; 1101 struct worklist *wk; 1102 1103 KASSERT(dst != src, 1104 ("jwork_move: dst == src")); 1105 freedep = NULL; 1106 jsegdep = NULL; 1107 LIST_FOREACH_SAFE(wk, dst, wk_list, wkn) { 1108 if (wk->wk_type == D_JSEGDEP) 1109 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep); 1110 else if (wk->wk_type == D_FREEDEP) 1111 freedep = freedep_merge(WK_FREEDEP(wk), freedep); 1112 } 1113 1114 while ((wk = LIST_FIRST(src)) != NULL) { 1115 WORKLIST_REMOVE(wk); 1116 WORKLIST_INSERT(dst, wk); 1117 if (wk->wk_type == D_JSEGDEP) { 1118 jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep); 1119 continue; 1120 } 1121 if (wk->wk_type == D_FREEDEP) 1122 freedep = freedep_merge(WK_FREEDEP(wk), freedep); 1123 } 1124 } 1125 1126 static void 1127 jwork_insert(dst, jsegdep) 1128 struct workhead *dst; 1129 struct jsegdep *jsegdep; 1130 { 1131 struct jsegdep *jsegdepn; 1132 struct worklist *wk; 1133 1134 LIST_FOREACH(wk, dst, wk_list) 1135 if (wk->wk_type == D_JSEGDEP) 1136 break; 1137 if (wk == NULL) { 1138 WORKLIST_INSERT(dst, &jsegdep->jd_list); 1139 return; 1140 } 1141 jsegdepn = WK_JSEGDEP(wk); 1142 if (jsegdep->jd_seg->js_seq < jsegdepn->jd_seg->js_seq) { 1143 WORKLIST_REMOVE(wk); 1144 free_jsegdep(jsegdepn); 1145 WORKLIST_INSERT(dst, &jsegdep->jd_list); 1146 } else 1147 free_jsegdep(jsegdep); 1148 } 1149 1150 /* 1151 * Routines for tracking and managing workitems. 1152 */ 1153 static void workitem_free(struct worklist *, int); 1154 static void workitem_alloc(struct worklist *, int, struct mount *); 1155 static void workitem_reassign(struct worklist *, int); 1156 1157 #define WORKITEM_FREE(item, type) \ 1158 workitem_free((struct worklist *)(item), (type)) 1159 #define WORKITEM_REASSIGN(item, type) \ 1160 workitem_reassign((struct worklist *)(item), (type)) 1161 1162 static void 1163 workitem_free(item, type) 1164 struct worklist *item; 1165 int type; 1166 { 1167 struct ufsmount *ump; 1168 1169 #ifdef DEBUG 1170 if (item->wk_state & ONWORKLIST) 1171 panic("workitem_free: %s(0x%X) still on list", 1172 TYPENAME(item->wk_type), item->wk_state); 1173 if (item->wk_type != type && type != D_NEWBLK) 1174 panic("workitem_free: type mismatch %s != %s", 1175 TYPENAME(item->wk_type), TYPENAME(type)); 1176 #endif 1177 if (item->wk_state & IOWAITING) 1178 wakeup(item); 1179 ump = VFSTOUFS(item->wk_mp); 1180 LOCK_OWNED(ump); 1181 KASSERT(ump->softdep_deps > 0, 1182 ("workitem_free: %s: softdep_deps going negative", 1183 ump->um_fs->fs_fsmnt)); 1184 if (--ump->softdep_deps == 0 && ump->softdep_req) 1185 wakeup(&ump->softdep_deps); 1186 KASSERT(dep_current[item->wk_type] > 0, 1187 ("workitem_free: %s: dep_current[%s] going negative", 1188 ump->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1189 KASSERT(ump->softdep_curdeps[item->wk_type] > 0, 1190 ("workitem_free: %s: softdep_curdeps[%s] going negative", 1191 ump->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1192 atomic_subtract_long(&dep_current[item->wk_type], 1); 1193 ump->softdep_curdeps[item->wk_type] -= 1; 1194 free(item, DtoM(type)); 1195 } 1196 1197 static void 1198 workitem_alloc(item, type, mp) 1199 struct worklist *item; 1200 int type; 1201 struct mount *mp; 1202 { 1203 struct ufsmount *ump; 1204 1205 item->wk_type = type; 1206 item->wk_mp = mp; 1207 item->wk_state = 0; 1208 1209 ump = VFSTOUFS(mp); 1210 ACQUIRE_GBLLOCK(&lk); 1211 dep_current[type]++; 1212 if (dep_current[type] > dep_highuse[type]) 1213 dep_highuse[type] = dep_current[type]; 1214 dep_total[type]++; 1215 FREE_GBLLOCK(&lk); 1216 ACQUIRE_LOCK(ump); 1217 ump->softdep_curdeps[type] += 1; 1218 ump->softdep_deps++; 1219 ump->softdep_accdeps++; 1220 FREE_LOCK(ump); 1221 } 1222 1223 static void 1224 workitem_reassign(item, newtype) 1225 struct worklist *item; 1226 int newtype; 1227 { 1228 struct ufsmount *ump; 1229 1230 ump = VFSTOUFS(item->wk_mp); 1231 LOCK_OWNED(ump); 1232 KASSERT(ump->softdep_curdeps[item->wk_type] > 0, 1233 ("workitem_reassign: %s: softdep_curdeps[%s] going negative", 1234 VFSTOUFS(item->wk_mp)->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1235 ump->softdep_curdeps[item->wk_type] -= 1; 1236 ump->softdep_curdeps[newtype] += 1; 1237 KASSERT(dep_current[item->wk_type] > 0, 1238 ("workitem_reassign: %s: dep_current[%s] going negative", 1239 VFSTOUFS(item->wk_mp)->um_fs->fs_fsmnt, TYPENAME(item->wk_type))); 1240 ACQUIRE_GBLLOCK(&lk); 1241 dep_current[newtype]++; 1242 dep_current[item->wk_type]--; 1243 if (dep_current[newtype] > dep_highuse[newtype]) 1244 dep_highuse[newtype] = dep_current[newtype]; 1245 dep_total[newtype]++; 1246 FREE_GBLLOCK(&lk); 1247 item->wk_type = newtype; 1248 } 1249 1250 /* 1251 * Workitem queue management 1252 */ 1253 static int max_softdeps; /* maximum number of structs before slowdown */ 1254 static int tickdelay = 2; /* number of ticks to pause during slowdown */ 1255 static int proc_waiting; /* tracks whether we have a timeout posted */ 1256 static int *stat_countp; /* statistic to count in proc_waiting timeout */ 1257 static struct callout softdep_callout; 1258 static int req_clear_inodedeps; /* syncer process flush some inodedeps */ 1259 static int req_clear_remove; /* syncer process flush some freeblks */ 1260 static int softdep_flushcache = 0; /* Should we do BIO_FLUSH? */ 1261 1262 /* 1263 * runtime statistics 1264 */ 1265 static int stat_flush_threads; /* number of softdep flushing threads */ 1266 static int stat_worklist_push; /* number of worklist cleanups */ 1267 static int stat_blk_limit_push; /* number of times block limit neared */ 1268 static int stat_ino_limit_push; /* number of times inode limit neared */ 1269 static int stat_blk_limit_hit; /* number of times block slowdown imposed */ 1270 static int stat_ino_limit_hit; /* number of times inode slowdown imposed */ 1271 static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */ 1272 static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */ 1273 static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */ 1274 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */ 1275 static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */ 1276 static int stat_jaddref; /* bufs redirtied as ino bitmap can not write */ 1277 static int stat_jnewblk; /* bufs redirtied as blk bitmap can not write */ 1278 static int stat_journal_min; /* Times hit journal min threshold */ 1279 static int stat_journal_low; /* Times hit journal low threshold */ 1280 static int stat_journal_wait; /* Times blocked in jwait(). */ 1281 static int stat_jwait_filepage; /* Times blocked in jwait() for filepage. */ 1282 static int stat_jwait_freeblks; /* Times blocked in jwait() for freeblks. */ 1283 static int stat_jwait_inode; /* Times blocked in jwait() for inodes. */ 1284 static int stat_jwait_newblk; /* Times blocked in jwait() for newblks. */ 1285 static int stat_cleanup_high_delay; /* Maximum cleanup delay (in ticks) */ 1286 static int stat_cleanup_blkrequests; /* Number of block cleanup requests */ 1287 static int stat_cleanup_inorequests; /* Number of inode cleanup requests */ 1288 static int stat_cleanup_retries; /* Number of cleanups that needed to flush */ 1289 static int stat_cleanup_failures; /* Number of cleanup requests that failed */ 1290 static int stat_emptyjblocks; /* Number of potentially empty journal blocks */ 1291 1292 SYSCTL_INT(_debug_softdep, OID_AUTO, max_softdeps, CTLFLAG_RW, 1293 &max_softdeps, 0, ""); 1294 SYSCTL_INT(_debug_softdep, OID_AUTO, tickdelay, CTLFLAG_RW, 1295 &tickdelay, 0, ""); 1296 SYSCTL_INT(_debug_softdep, OID_AUTO, flush_threads, CTLFLAG_RD, 1297 &stat_flush_threads, 0, ""); 1298 SYSCTL_INT(_debug_softdep, OID_AUTO, worklist_push, CTLFLAG_RW, 1299 &stat_worklist_push, 0,""); 1300 SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_push, CTLFLAG_RW, 1301 &stat_blk_limit_push, 0,""); 1302 SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_push, CTLFLAG_RW, 1303 &stat_ino_limit_push, 0,""); 1304 SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_hit, CTLFLAG_RW, 1305 &stat_blk_limit_hit, 0, ""); 1306 SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_hit, CTLFLAG_RW, 1307 &stat_ino_limit_hit, 0, ""); 1308 SYSCTL_INT(_debug_softdep, OID_AUTO, sync_limit_hit, CTLFLAG_RW, 1309 &stat_sync_limit_hit, 0, ""); 1310 SYSCTL_INT(_debug_softdep, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, 1311 &stat_indir_blk_ptrs, 0, ""); 1312 SYSCTL_INT(_debug_softdep, OID_AUTO, inode_bitmap, CTLFLAG_RW, 1313 &stat_inode_bitmap, 0, ""); 1314 SYSCTL_INT(_debug_softdep, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, 1315 &stat_direct_blk_ptrs, 0, ""); 1316 SYSCTL_INT(_debug_softdep, OID_AUTO, dir_entry, CTLFLAG_RW, 1317 &stat_dir_entry, 0, ""); 1318 SYSCTL_INT(_debug_softdep, OID_AUTO, jaddref_rollback, CTLFLAG_RW, 1319 &stat_jaddref, 0, ""); 1320 SYSCTL_INT(_debug_softdep, OID_AUTO, jnewblk_rollback, CTLFLAG_RW, 1321 &stat_jnewblk, 0, ""); 1322 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_low, CTLFLAG_RW, 1323 &stat_journal_low, 0, ""); 1324 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_min, CTLFLAG_RW, 1325 &stat_journal_min, 0, ""); 1326 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_wait, CTLFLAG_RW, 1327 &stat_journal_wait, 0, ""); 1328 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_filepage, CTLFLAG_RW, 1329 &stat_jwait_filepage, 0, ""); 1330 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_freeblks, CTLFLAG_RW, 1331 &stat_jwait_freeblks, 0, ""); 1332 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_inode, CTLFLAG_RW, 1333 &stat_jwait_inode, 0, ""); 1334 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_newblk, CTLFLAG_RW, 1335 &stat_jwait_newblk, 0, ""); 1336 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_blkrequests, CTLFLAG_RW, 1337 &stat_cleanup_blkrequests, 0, ""); 1338 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_inorequests, CTLFLAG_RW, 1339 &stat_cleanup_inorequests, 0, ""); 1340 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_high_delay, CTLFLAG_RW, 1341 &stat_cleanup_high_delay, 0, ""); 1342 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_retries, CTLFLAG_RW, 1343 &stat_cleanup_retries, 0, ""); 1344 SYSCTL_INT(_debug_softdep, OID_AUTO, cleanup_failures, CTLFLAG_RW, 1345 &stat_cleanup_failures, 0, ""); 1346 SYSCTL_INT(_debug_softdep, OID_AUTO, flushcache, CTLFLAG_RW, 1347 &softdep_flushcache, 0, ""); 1348 SYSCTL_INT(_debug_softdep, OID_AUTO, emptyjblocks, CTLFLAG_RD, 1349 &stat_emptyjblocks, 0, ""); 1350 1351 SYSCTL_DECL(_vfs_ffs); 1352 1353 /* Whether to recompute the summary at mount time */ 1354 static int compute_summary_at_mount = 0; 1355 SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW, 1356 &compute_summary_at_mount, 0, "Recompute summary at mount"); 1357 static int print_threads = 0; 1358 SYSCTL_INT(_debug_softdep, OID_AUTO, print_threads, CTLFLAG_RW, 1359 &print_threads, 0, "Notify flusher thread start/stop"); 1360 1361 /* List of all filesystems mounted with soft updates */ 1362 static TAILQ_HEAD(, mount_softdeps) softdepmounts; 1363 1364 /* 1365 * This function cleans the worklist for a filesystem. 1366 * Each filesystem running with soft dependencies gets its own 1367 * thread to run in this function. The thread is started up in 1368 * softdep_mount and shutdown in softdep_unmount. They show up 1369 * as part of the kernel "bufdaemon" process whose process 1370 * entry is available in bufdaemonproc. 1371 */ 1372 static int searchfailed; 1373 extern struct proc *bufdaemonproc; 1374 static void 1375 softdep_flush(addr) 1376 void *addr; 1377 { 1378 struct mount *mp; 1379 struct thread *td; 1380 struct ufsmount *ump; 1381 1382 td = curthread; 1383 td->td_pflags |= TDP_NORUNNINGBUF; 1384 mp = (struct mount *)addr; 1385 ump = VFSTOUFS(mp); 1386 atomic_add_int(&stat_flush_threads, 1); 1387 ACQUIRE_LOCK(ump); 1388 ump->softdep_flags &= ~FLUSH_STARTING; 1389 wakeup(&ump->softdep_flushtd); 1390 FREE_LOCK(ump); 1391 if (print_threads) { 1392 if (stat_flush_threads == 1) 1393 printf("Running %s at pid %d\n", bufdaemonproc->p_comm, 1394 bufdaemonproc->p_pid); 1395 printf("Start thread %s\n", td->td_name); 1396 } 1397 for (;;) { 1398 while (softdep_process_worklist(mp, 0) > 0 || 1399 (MOUNTEDSUJ(mp) && 1400 VFSTOUFS(mp)->softdep_jblocks->jb_suspended)) 1401 kthread_suspend_check(); 1402 ACQUIRE_LOCK(ump); 1403 if ((ump->softdep_flags & (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1404 msleep(&ump->softdep_flushtd, LOCK_PTR(ump), PVM, 1405 "sdflush", hz / 2); 1406 ump->softdep_flags &= ~FLUSH_CLEANUP; 1407 /* 1408 * Check to see if we are done and need to exit. 1409 */ 1410 if ((ump->softdep_flags & FLUSH_EXIT) == 0) { 1411 FREE_LOCK(ump); 1412 continue; 1413 } 1414 ump->softdep_flags &= ~FLUSH_EXIT; 1415 FREE_LOCK(ump); 1416 wakeup(&ump->softdep_flags); 1417 if (print_threads) 1418 printf("Stop thread %s: searchfailed %d, did cleanups %d\n", td->td_name, searchfailed, ump->um_softdep->sd_cleanups); 1419 atomic_subtract_int(&stat_flush_threads, 1); 1420 kthread_exit(); 1421 panic("kthread_exit failed\n"); 1422 } 1423 } 1424 1425 static void 1426 worklist_speedup(mp) 1427 struct mount *mp; 1428 { 1429 struct ufsmount *ump; 1430 1431 ump = VFSTOUFS(mp); 1432 LOCK_OWNED(ump); 1433 if ((ump->softdep_flags & (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1434 ump->softdep_flags |= FLUSH_CLEANUP; 1435 wakeup(&ump->softdep_flushtd); 1436 } 1437 1438 static int 1439 softdep_speedup(ump) 1440 struct ufsmount *ump; 1441 { 1442 struct ufsmount *altump; 1443 struct mount_softdeps *sdp; 1444 1445 LOCK_OWNED(ump); 1446 worklist_speedup(ump->um_mountp); 1447 bd_speedup(); 1448 /* 1449 * If we have global shortages, then we need other 1450 * filesystems to help with the cleanup. Here we wakeup a 1451 * flusher thread for a filesystem that is over its fair 1452 * share of resources. 1453 */ 1454 if (req_clear_inodedeps || req_clear_remove) { 1455 ACQUIRE_GBLLOCK(&lk); 1456 TAILQ_FOREACH(sdp, &softdepmounts, sd_next) { 1457 if ((altump = sdp->sd_ump) == ump) 1458 continue; 1459 if (((req_clear_inodedeps && 1460 altump->softdep_curdeps[D_INODEDEP] > 1461 max_softdeps / stat_flush_threads) || 1462 (req_clear_remove && 1463 altump->softdep_curdeps[D_DIRREM] > 1464 (max_softdeps / 2) / stat_flush_threads)) && 1465 TRY_ACQUIRE_LOCK(altump)) 1466 break; 1467 } 1468 if (sdp == NULL) { 1469 searchfailed++; 1470 FREE_GBLLOCK(&lk); 1471 } else { 1472 /* 1473 * Move to the end of the list so we pick a 1474 * different one on out next try. 1475 */ 1476 TAILQ_REMOVE(&softdepmounts, sdp, sd_next); 1477 TAILQ_INSERT_TAIL(&softdepmounts, sdp, sd_next); 1478 FREE_GBLLOCK(&lk); 1479 if ((altump->softdep_flags & 1480 (FLUSH_CLEANUP | FLUSH_EXIT)) == 0) 1481 altump->softdep_flags |= FLUSH_CLEANUP; 1482 altump->um_softdep->sd_cleanups++; 1483 wakeup(&altump->softdep_flushtd); 1484 FREE_LOCK(altump); 1485 } 1486 } 1487 return (speedup_syncer()); 1488 } 1489 1490 /* 1491 * Add an item to the end of the work queue. 1492 * This routine requires that the lock be held. 1493 * This is the only routine that adds items to the list. 1494 * The following routine is the only one that removes items 1495 * and does so in order from first to last. 1496 */ 1497 1498 #define WK_HEAD 0x0001 /* Add to HEAD. */ 1499 #define WK_NODELAY 0x0002 /* Process immediately. */ 1500 1501 static void 1502 add_to_worklist(wk, flags) 1503 struct worklist *wk; 1504 int flags; 1505 { 1506 struct ufsmount *ump; 1507 1508 ump = VFSTOUFS(wk->wk_mp); 1509 LOCK_OWNED(ump); 1510 if (wk->wk_state & ONWORKLIST) 1511 panic("add_to_worklist: %s(0x%X) already on list", 1512 TYPENAME(wk->wk_type), wk->wk_state); 1513 wk->wk_state |= ONWORKLIST; 1514 if (ump->softdep_on_worklist == 0) { 1515 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 1516 ump->softdep_worklist_tail = wk; 1517 } else if (flags & WK_HEAD) { 1518 LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list); 1519 } else { 1520 LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list); 1521 ump->softdep_worklist_tail = wk; 1522 } 1523 ump->softdep_on_worklist += 1; 1524 if (flags & WK_NODELAY) 1525 worklist_speedup(wk->wk_mp); 1526 } 1527 1528 /* 1529 * Remove the item to be processed. If we are removing the last 1530 * item on the list, we need to recalculate the tail pointer. 1531 */ 1532 static void 1533 remove_from_worklist(wk) 1534 struct worklist *wk; 1535 { 1536 struct ufsmount *ump; 1537 1538 ump = VFSTOUFS(wk->wk_mp); 1539 if (ump->softdep_worklist_tail == wk) 1540 ump->softdep_worklist_tail = 1541 (struct worklist *)wk->wk_list.le_prev; 1542 WORKLIST_REMOVE(wk); 1543 ump->softdep_on_worklist -= 1; 1544 } 1545 1546 static void 1547 wake_worklist(wk) 1548 struct worklist *wk; 1549 { 1550 if (wk->wk_state & IOWAITING) { 1551 wk->wk_state &= ~IOWAITING; 1552 wakeup(wk); 1553 } 1554 } 1555 1556 static void 1557 wait_worklist(wk, wmesg) 1558 struct worklist *wk; 1559 char *wmesg; 1560 { 1561 struct ufsmount *ump; 1562 1563 ump = VFSTOUFS(wk->wk_mp); 1564 wk->wk_state |= IOWAITING; 1565 msleep(wk, LOCK_PTR(ump), PVM, wmesg, 0); 1566 } 1567 1568 /* 1569 * Process that runs once per second to handle items in the background queue. 1570 * 1571 * Note that we ensure that everything is done in the order in which they 1572 * appear in the queue. The code below depends on this property to ensure 1573 * that blocks of a file are freed before the inode itself is freed. This 1574 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated 1575 * until all the old ones have been purged from the dependency lists. 1576 */ 1577 static int 1578 softdep_process_worklist(mp, full) 1579 struct mount *mp; 1580 int full; 1581 { 1582 int cnt, matchcnt; 1583 struct ufsmount *ump; 1584 long starttime; 1585 1586 KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp")); 1587 if (MOUNTEDSOFTDEP(mp) == 0) 1588 return (0); 1589 matchcnt = 0; 1590 ump = VFSTOUFS(mp); 1591 ACQUIRE_LOCK(ump); 1592 starttime = time_second; 1593 softdep_process_journal(mp, NULL, full ? MNT_WAIT : 0); 1594 check_clear_deps(mp); 1595 while (ump->softdep_on_worklist > 0) { 1596 if ((cnt = process_worklist_item(mp, 10, LK_NOWAIT)) == 0) 1597 break; 1598 else 1599 matchcnt += cnt; 1600 check_clear_deps(mp); 1601 /* 1602 * We do not generally want to stop for buffer space, but if 1603 * we are really being a buffer hog, we will stop and wait. 1604 */ 1605 if (should_yield()) { 1606 FREE_LOCK(ump); 1607 kern_yield(PRI_USER); 1608 bwillwrite(); 1609 ACQUIRE_LOCK(ump); 1610 } 1611 /* 1612 * Never allow processing to run for more than one 1613 * second. This gives the syncer thread the opportunity 1614 * to pause if appropriate. 1615 */ 1616 if (!full && starttime != time_second) 1617 break; 1618 } 1619 if (full == 0) 1620 journal_unsuspend(ump); 1621 FREE_LOCK(ump); 1622 return (matchcnt); 1623 } 1624 1625 /* 1626 * Process all removes associated with a vnode if we are running out of 1627 * journal space. Any other process which attempts to flush these will 1628 * be unable as we have the vnodes locked. 1629 */ 1630 static void 1631 process_removes(vp) 1632 struct vnode *vp; 1633 { 1634 struct inodedep *inodedep; 1635 struct dirrem *dirrem; 1636 struct ufsmount *ump; 1637 struct mount *mp; 1638 ino_t inum; 1639 1640 mp = vp->v_mount; 1641 ump = VFSTOUFS(mp); 1642 LOCK_OWNED(ump); 1643 inum = VTOI(vp)->i_number; 1644 for (;;) { 1645 top: 1646 if (inodedep_lookup(mp, inum, 0, &inodedep) == 0) 1647 return; 1648 LIST_FOREACH(dirrem, &inodedep->id_dirremhd, dm_inonext) { 1649 /* 1650 * If another thread is trying to lock this vnode 1651 * it will fail but we must wait for it to do so 1652 * before we can proceed. 1653 */ 1654 if (dirrem->dm_state & INPROGRESS) { 1655 wait_worklist(&dirrem->dm_list, "pwrwait"); 1656 goto top; 1657 } 1658 if ((dirrem->dm_state & (COMPLETE | ONWORKLIST)) == 1659 (COMPLETE | ONWORKLIST)) 1660 break; 1661 } 1662 if (dirrem == NULL) 1663 return; 1664 remove_from_worklist(&dirrem->dm_list); 1665 FREE_LOCK(ump); 1666 if (vn_start_secondary_write(NULL, &mp, V_NOWAIT)) 1667 panic("process_removes: suspended filesystem"); 1668 handle_workitem_remove(dirrem, 0); 1669 vn_finished_secondary_write(mp); 1670 ACQUIRE_LOCK(ump); 1671 } 1672 } 1673 1674 /* 1675 * Process all truncations associated with a vnode if we are running out 1676 * of journal space. This is called when the vnode lock is already held 1677 * and no other process can clear the truncation. This function returns 1678 * a value greater than zero if it did any work. 1679 */ 1680 static void 1681 process_truncates(vp) 1682 struct vnode *vp; 1683 { 1684 struct inodedep *inodedep; 1685 struct freeblks *freeblks; 1686 struct ufsmount *ump; 1687 struct mount *mp; 1688 ino_t inum; 1689 int cgwait; 1690 1691 mp = vp->v_mount; 1692 ump = VFSTOUFS(mp); 1693 LOCK_OWNED(ump); 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(ump); 1715 ffs_update(vp, 1); 1716 ACQUIRE_LOCK(ump); 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(ump); 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(ump); 1731 break; 1732 } 1733 if (freeblks->fb_cgwait) 1734 cgwait++; 1735 } 1736 if (cgwait) { 1737 FREE_LOCK(ump); 1738 sync_cgs(mp, MNT_WAIT); 1739 ffs_sync_snap(mp, MNT_WAIT); 1740 ACQUIRE_LOCK(ump); 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 KASSERT(mp != NULL, ("process_worklist_item: NULL mp")); 1765 /* 1766 * If we are being called because of a process doing a 1767 * copy-on-write, then it is not safe to write as we may 1768 * recurse into the copy-on-write routine. 1769 */ 1770 if (curthread->td_pflags & TDP_COWINPROGRESS) 1771 return (-1); 1772 PHOLD(curproc); /* Don't let the stack go away. */ 1773 ump = VFSTOUFS(mp); 1774 LOCK_OWNED(ump); 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(ump); 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(ump); 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 /* Sentinal could've become the tail from remove_from_worklist. */ 1841 if (ump->softdep_worklist_tail == &sentinel) 1842 ump->softdep_worklist_tail = 1843 (struct worklist *)sentinel.wk_list.le_prev; 1844 LIST_REMOVE(&sentinel, wk_list); 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 struct ufsmount *ump; 1859 int dirty; 1860 1861 if ((wk = LIST_FIRST(&oldbp->b_dep)) == NULL) 1862 return (0); 1863 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 1864 ("softdep_move_dependencies called on non-softdep filesystem")); 1865 dirty = 0; 1866 wktail = NULL; 1867 ump = VFSTOUFS(wk->wk_mp); 1868 ACQUIRE_LOCK(ump); 1869 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) { 1870 LIST_REMOVE(wk, wk_list); 1871 if (wk->wk_type == D_BMSAFEMAP && 1872 bmsafemap_backgroundwrite(WK_BMSAFEMAP(wk), newbp)) 1873 dirty = 1; 1874 if (wktail == NULL) 1875 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list); 1876 else 1877 LIST_INSERT_AFTER(wktail, wk, wk_list); 1878 wktail = wk; 1879 } 1880 FREE_LOCK(ump); 1881 1882 return (dirty); 1883 } 1884 1885 /* 1886 * Purge the work list of all items associated with a particular mount point. 1887 */ 1888 int 1889 softdep_flushworklist(oldmnt, countp, td) 1890 struct mount *oldmnt; 1891 int *countp; 1892 struct thread *td; 1893 { 1894 struct vnode *devvp; 1895 struct ufsmount *ump; 1896 int count, error; 1897 1898 /* 1899 * Alternately flush the block device associated with the mount 1900 * point and process any dependencies that the flushing 1901 * creates. We continue until no more worklist dependencies 1902 * are found. 1903 */ 1904 *countp = 0; 1905 error = 0; 1906 ump = VFSTOUFS(oldmnt); 1907 devvp = ump->um_devvp; 1908 while ((count = softdep_process_worklist(oldmnt, 1)) > 0) { 1909 *countp += count; 1910 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1911 error = VOP_FSYNC(devvp, MNT_WAIT, td); 1912 VOP_UNLOCK(devvp, 0); 1913 if (error != 0) 1914 break; 1915 } 1916 return (error); 1917 } 1918 1919 #define SU_WAITIDLE_RETRIES 20 1920 static int 1921 softdep_waitidle(struct mount *mp, int flags __unused) 1922 { 1923 struct ufsmount *ump; 1924 struct vnode *devvp; 1925 struct thread *td; 1926 int error, i; 1927 1928 ump = VFSTOUFS(mp); 1929 devvp = ump->um_devvp; 1930 td = curthread; 1931 error = 0; 1932 ACQUIRE_LOCK(ump); 1933 for (i = 0; i < SU_WAITIDLE_RETRIES && ump->softdep_deps != 0; i++) { 1934 ump->softdep_req = 1; 1935 KASSERT((flags & FORCECLOSE) == 0 || 1936 ump->softdep_on_worklist == 0, 1937 ("softdep_waitidle: work added after flush")); 1938 msleep(&ump->softdep_deps, LOCK_PTR(ump), PVM | PDROP, 1939 "softdeps", 10 * hz); 1940 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY); 1941 error = VOP_FSYNC(devvp, MNT_WAIT, td); 1942 VOP_UNLOCK(devvp, 0); 1943 ACQUIRE_LOCK(ump); 1944 if (error != 0) 1945 break; 1946 } 1947 ump->softdep_req = 0; 1948 if (i == SU_WAITIDLE_RETRIES && error == 0 && ump->softdep_deps != 0) { 1949 error = EBUSY; 1950 printf("softdep_waitidle: Failed to flush worklist for %p\n", 1951 mp); 1952 } 1953 FREE_LOCK(ump); 1954 return (error); 1955 } 1956 1957 /* 1958 * Flush all vnodes and worklist items associated with a specified mount point. 1959 */ 1960 int 1961 softdep_flushfiles(oldmnt, flags, td) 1962 struct mount *oldmnt; 1963 int flags; 1964 struct thread *td; 1965 { 1966 #ifdef QUOTA 1967 struct ufsmount *ump; 1968 int i; 1969 #endif 1970 int error, early, depcount, loopcnt, retry_flush_count, retry; 1971 int morework; 1972 1973 KASSERT(MOUNTEDSOFTDEP(oldmnt) != 0, 1974 ("softdep_flushfiles called on non-softdep filesystem")); 1975 loopcnt = 10; 1976 retry_flush_count = 3; 1977 retry_flush: 1978 error = 0; 1979 1980 /* 1981 * Alternately flush the vnodes associated with the mount 1982 * point and process any dependencies that the flushing 1983 * creates. In theory, this loop can happen at most twice, 1984 * but we give it a few extra just to be sure. 1985 */ 1986 for (; loopcnt > 0; loopcnt--) { 1987 /* 1988 * Do another flush in case any vnodes were brought in 1989 * as part of the cleanup operations. 1990 */ 1991 early = retry_flush_count == 1 || (oldmnt->mnt_kern_flag & 1992 MNTK_UNMOUNT) == 0 ? 0 : EARLYFLUSH; 1993 if ((error = ffs_flushfiles(oldmnt, flags | early, td)) != 0) 1994 break; 1995 if ((error = softdep_flushworklist(oldmnt, &depcount, td)) != 0 || 1996 depcount == 0) 1997 break; 1998 } 1999 /* 2000 * If we are unmounting then it is an error to fail. If we 2001 * are simply trying to downgrade to read-only, then filesystem 2002 * activity can keep us busy forever, so we just fail with EBUSY. 2003 */ 2004 if (loopcnt == 0) { 2005 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) 2006 panic("softdep_flushfiles: looping"); 2007 error = EBUSY; 2008 } 2009 if (!error) 2010 error = softdep_waitidle(oldmnt, flags); 2011 if (!error) { 2012 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) { 2013 retry = 0; 2014 MNT_ILOCK(oldmnt); 2015 KASSERT((oldmnt->mnt_kern_flag & MNTK_NOINSMNTQ) != 0, 2016 ("softdep_flushfiles: !MNTK_NOINSMNTQ")); 2017 morework = oldmnt->mnt_nvnodelistsize > 0; 2018 #ifdef QUOTA 2019 ump = VFSTOUFS(oldmnt); 2020 UFS_LOCK(ump); 2021 for (i = 0; i < MAXQUOTAS; i++) { 2022 if (ump->um_quotas[i] != NULLVP) 2023 morework = 1; 2024 } 2025 UFS_UNLOCK(ump); 2026 #endif 2027 if (morework) { 2028 if (--retry_flush_count > 0) { 2029 retry = 1; 2030 loopcnt = 3; 2031 } else 2032 error = EBUSY; 2033 } 2034 MNT_IUNLOCK(oldmnt); 2035 if (retry) 2036 goto retry_flush; 2037 } 2038 } 2039 return (error); 2040 } 2041 2042 /* 2043 * Structure hashing. 2044 * 2045 * There are four types of structures that can be looked up: 2046 * 1) pagedep structures identified by mount point, inode number, 2047 * and logical block. 2048 * 2) inodedep structures identified by mount point and inode number. 2049 * 3) newblk structures identified by mount point and 2050 * physical block number. 2051 * 4) bmsafemap structures identified by mount point and 2052 * cylinder group number. 2053 * 2054 * The "pagedep" and "inodedep" dependency structures are hashed 2055 * separately from the file blocks and inodes to which they correspond. 2056 * This separation helps when the in-memory copy of an inode or 2057 * file block must be replaced. It also obviates the need to access 2058 * an inode or file page when simply updating (or de-allocating) 2059 * dependency structures. Lookup of newblk structures is needed to 2060 * find newly allocated blocks when trying to associate them with 2061 * their allocdirect or allocindir structure. 2062 * 2063 * The lookup routines optionally create and hash a new instance when 2064 * an existing entry is not found. The bmsafemap lookup routine always 2065 * allocates a new structure if an existing one is not found. 2066 */ 2067 #define DEPALLOC 0x0001 /* allocate structure if lookup fails */ 2068 2069 /* 2070 * Structures and routines associated with pagedep caching. 2071 */ 2072 #define PAGEDEP_HASH(ump, inum, lbn) \ 2073 (&(ump)->pagedep_hashtbl[((inum) + (lbn)) & (ump)->pagedep_hash_size]) 2074 2075 static int 2076 pagedep_find(pagedephd, ino, lbn, pagedeppp) 2077 struct pagedep_hashhead *pagedephd; 2078 ino_t ino; 2079 ufs_lbn_t lbn; 2080 struct pagedep **pagedeppp; 2081 { 2082 struct pagedep *pagedep; 2083 2084 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 2085 if (ino == pagedep->pd_ino && lbn == pagedep->pd_lbn) { 2086 *pagedeppp = pagedep; 2087 return (1); 2088 } 2089 } 2090 *pagedeppp = NULL; 2091 return (0); 2092 } 2093 /* 2094 * Look up a pagedep. Return 1 if found, 0 otherwise. 2095 * If not found, allocate if DEPALLOC flag is passed. 2096 * Found or allocated entry is returned in pagedeppp. 2097 * This routine must be called with splbio interrupts blocked. 2098 */ 2099 static int 2100 pagedep_lookup(mp, bp, ino, lbn, flags, pagedeppp) 2101 struct mount *mp; 2102 struct buf *bp; 2103 ino_t ino; 2104 ufs_lbn_t lbn; 2105 int flags; 2106 struct pagedep **pagedeppp; 2107 { 2108 struct pagedep *pagedep; 2109 struct pagedep_hashhead *pagedephd; 2110 struct worklist *wk; 2111 struct ufsmount *ump; 2112 int ret; 2113 int i; 2114 2115 ump = VFSTOUFS(mp); 2116 LOCK_OWNED(ump); 2117 if (bp) { 2118 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 2119 if (wk->wk_type == D_PAGEDEP) { 2120 *pagedeppp = WK_PAGEDEP(wk); 2121 return (1); 2122 } 2123 } 2124 } 2125 pagedephd = PAGEDEP_HASH(ump, ino, lbn); 2126 ret = pagedep_find(pagedephd, ino, lbn, pagedeppp); 2127 if (ret) { 2128 if (((*pagedeppp)->pd_state & ONWORKLIST) == 0 && bp) 2129 WORKLIST_INSERT(&bp->b_dep, &(*pagedeppp)->pd_list); 2130 return (1); 2131 } 2132 if ((flags & DEPALLOC) == 0) 2133 return (0); 2134 FREE_LOCK(ump); 2135 pagedep = malloc(sizeof(struct pagedep), 2136 M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO); 2137 workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp); 2138 ACQUIRE_LOCK(ump); 2139 ret = pagedep_find(pagedephd, ino, lbn, pagedeppp); 2140 if (*pagedeppp) { 2141 /* 2142 * This should never happen since we only create pagedeps 2143 * with the vnode lock held. Could be an assert. 2144 */ 2145 WORKITEM_FREE(pagedep, D_PAGEDEP); 2146 return (ret); 2147 } 2148 pagedep->pd_ino = ino; 2149 pagedep->pd_lbn = lbn; 2150 LIST_INIT(&pagedep->pd_dirremhd); 2151 LIST_INIT(&pagedep->pd_pendinghd); 2152 for (i = 0; i < DAHASHSZ; i++) 2153 LIST_INIT(&pagedep->pd_diraddhd[i]); 2154 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash); 2155 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list); 2156 *pagedeppp = pagedep; 2157 return (0); 2158 } 2159 2160 /* 2161 * Structures and routines associated with inodedep caching. 2162 */ 2163 #define INODEDEP_HASH(ump, inum) \ 2164 (&(ump)->inodedep_hashtbl[(inum) & (ump)->inodedep_hash_size]) 2165 2166 static int 2167 inodedep_find(inodedephd, inum, inodedeppp) 2168 struct inodedep_hashhead *inodedephd; 2169 ino_t inum; 2170 struct inodedep **inodedeppp; 2171 { 2172 struct inodedep *inodedep; 2173 2174 LIST_FOREACH(inodedep, inodedephd, id_hash) 2175 if (inum == inodedep->id_ino) 2176 break; 2177 if (inodedep) { 2178 *inodedeppp = inodedep; 2179 return (1); 2180 } 2181 *inodedeppp = NULL; 2182 2183 return (0); 2184 } 2185 /* 2186 * Look up an inodedep. Return 1 if found, 0 if not found. 2187 * If not found, allocate if DEPALLOC flag is passed. 2188 * Found or allocated entry is returned in inodedeppp. 2189 * This routine must be called with splbio interrupts blocked. 2190 */ 2191 static int 2192 inodedep_lookup(mp, inum, flags, inodedeppp) 2193 struct mount *mp; 2194 ino_t inum; 2195 int flags; 2196 struct inodedep **inodedeppp; 2197 { 2198 struct inodedep *inodedep; 2199 struct inodedep_hashhead *inodedephd; 2200 struct ufsmount *ump; 2201 struct fs *fs; 2202 2203 ump = VFSTOUFS(mp); 2204 LOCK_OWNED(ump); 2205 fs = ump->um_fs; 2206 inodedephd = INODEDEP_HASH(ump, inum); 2207 2208 if (inodedep_find(inodedephd, inum, inodedeppp)) 2209 return (1); 2210 if ((flags & DEPALLOC) == 0) 2211 return (0); 2212 /* 2213 * If the system is over its limit and our filesystem is 2214 * responsible for more than our share of that usage and 2215 * we are not in a rush, request some inodedep cleanup. 2216 */ 2217 if (softdep_excess_items(ump, D_INODEDEP)) 2218 schedule_cleanup(mp); 2219 else 2220 FREE_LOCK(ump); 2221 inodedep = malloc(sizeof(struct inodedep), 2222 M_INODEDEP, M_SOFTDEP_FLAGS); 2223 workitem_alloc(&inodedep->id_list, D_INODEDEP, mp); 2224 ACQUIRE_LOCK(ump); 2225 if (inodedep_find(inodedephd, inum, inodedeppp)) { 2226 WORKITEM_FREE(inodedep, D_INODEDEP); 2227 return (1); 2228 } 2229 inodedep->id_fs = fs; 2230 inodedep->id_ino = inum; 2231 inodedep->id_state = ALLCOMPLETE; 2232 inodedep->id_nlinkdelta = 0; 2233 inodedep->id_savedino1 = NULL; 2234 inodedep->id_savedsize = -1; 2235 inodedep->id_savedextsize = -1; 2236 inodedep->id_savednlink = -1; 2237 inodedep->id_bmsafemap = NULL; 2238 inodedep->id_mkdiradd = NULL; 2239 LIST_INIT(&inodedep->id_dirremhd); 2240 LIST_INIT(&inodedep->id_pendinghd); 2241 LIST_INIT(&inodedep->id_inowait); 2242 LIST_INIT(&inodedep->id_bufwait); 2243 TAILQ_INIT(&inodedep->id_inoreflst); 2244 TAILQ_INIT(&inodedep->id_inoupdt); 2245 TAILQ_INIT(&inodedep->id_newinoupdt); 2246 TAILQ_INIT(&inodedep->id_extupdt); 2247 TAILQ_INIT(&inodedep->id_newextupdt); 2248 TAILQ_INIT(&inodedep->id_freeblklst); 2249 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash); 2250 *inodedeppp = inodedep; 2251 return (0); 2252 } 2253 2254 /* 2255 * Structures and routines associated with newblk caching. 2256 */ 2257 #define NEWBLK_HASH(ump, inum) \ 2258 (&(ump)->newblk_hashtbl[(inum) & (ump)->newblk_hash_size]) 2259 2260 static int 2261 newblk_find(newblkhd, newblkno, flags, newblkpp) 2262 struct newblk_hashhead *newblkhd; 2263 ufs2_daddr_t newblkno; 2264 int flags; 2265 struct newblk **newblkpp; 2266 { 2267 struct newblk *newblk; 2268 2269 LIST_FOREACH(newblk, newblkhd, nb_hash) { 2270 if (newblkno != newblk->nb_newblkno) 2271 continue; 2272 /* 2273 * If we're creating a new dependency don't match those that 2274 * have already been converted to allocdirects. This is for 2275 * a frag extend. 2276 */ 2277 if ((flags & DEPALLOC) && newblk->nb_list.wk_type != D_NEWBLK) 2278 continue; 2279 break; 2280 } 2281 if (newblk) { 2282 *newblkpp = newblk; 2283 return (1); 2284 } 2285 *newblkpp = NULL; 2286 return (0); 2287 } 2288 2289 /* 2290 * Look up a newblk. Return 1 if found, 0 if not found. 2291 * If not found, allocate if DEPALLOC flag is passed. 2292 * Found or allocated entry is returned in newblkpp. 2293 */ 2294 static int 2295 newblk_lookup(mp, newblkno, flags, newblkpp) 2296 struct mount *mp; 2297 ufs2_daddr_t newblkno; 2298 int flags; 2299 struct newblk **newblkpp; 2300 { 2301 struct newblk *newblk; 2302 struct newblk_hashhead *newblkhd; 2303 struct ufsmount *ump; 2304 2305 ump = VFSTOUFS(mp); 2306 LOCK_OWNED(ump); 2307 newblkhd = NEWBLK_HASH(ump, newblkno); 2308 if (newblk_find(newblkhd, newblkno, flags, newblkpp)) 2309 return (1); 2310 if ((flags & DEPALLOC) == 0) 2311 return (0); 2312 if (softdep_excess_items(ump, D_NEWBLK) || 2313 softdep_excess_items(ump, D_ALLOCDIRECT) || 2314 softdep_excess_items(ump, D_ALLOCINDIR)) 2315 schedule_cleanup(mp); 2316 else 2317 FREE_LOCK(ump); 2318 newblk = malloc(sizeof(union allblk), M_NEWBLK, 2319 M_SOFTDEP_FLAGS | M_ZERO); 2320 workitem_alloc(&newblk->nb_list, D_NEWBLK, mp); 2321 ACQUIRE_LOCK(ump); 2322 if (newblk_find(newblkhd, newblkno, flags, newblkpp)) { 2323 WORKITEM_FREE(newblk, D_NEWBLK); 2324 return (1); 2325 } 2326 newblk->nb_freefrag = NULL; 2327 LIST_INIT(&newblk->nb_indirdeps); 2328 LIST_INIT(&newblk->nb_newdirblk); 2329 LIST_INIT(&newblk->nb_jwork); 2330 newblk->nb_state = ATTACHED; 2331 newblk->nb_newblkno = newblkno; 2332 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash); 2333 *newblkpp = newblk; 2334 return (0); 2335 } 2336 2337 /* 2338 * Structures and routines associated with freed indirect block caching. 2339 */ 2340 #define INDIR_HASH(ump, blkno) \ 2341 (&(ump)->indir_hashtbl[(blkno) & (ump)->indir_hash_size]) 2342 2343 /* 2344 * Lookup an indirect block in the indir hash table. The freework is 2345 * removed and potentially freed. The caller must do a blocking journal 2346 * write before writing to the blkno. 2347 */ 2348 static int 2349 indirblk_lookup(mp, blkno) 2350 struct mount *mp; 2351 ufs2_daddr_t blkno; 2352 { 2353 struct freework *freework; 2354 struct indir_hashhead *wkhd; 2355 struct ufsmount *ump; 2356 2357 ump = VFSTOUFS(mp); 2358 wkhd = INDIR_HASH(ump, blkno); 2359 TAILQ_FOREACH(freework, wkhd, fw_next) { 2360 if (freework->fw_blkno != blkno) 2361 continue; 2362 indirblk_remove(freework); 2363 return (1); 2364 } 2365 return (0); 2366 } 2367 2368 /* 2369 * Insert an indirect block represented by freework into the indirblk 2370 * hash table so that it may prevent the block from being re-used prior 2371 * to the journal being written. 2372 */ 2373 static void 2374 indirblk_insert(freework) 2375 struct freework *freework; 2376 { 2377 struct jblocks *jblocks; 2378 struct jseg *jseg; 2379 struct ufsmount *ump; 2380 2381 ump = VFSTOUFS(freework->fw_list.wk_mp); 2382 jblocks = ump->softdep_jblocks; 2383 jseg = TAILQ_LAST(&jblocks->jb_segs, jseglst); 2384 if (jseg == NULL) 2385 return; 2386 2387 LIST_INSERT_HEAD(&jseg->js_indirs, freework, fw_segs); 2388 TAILQ_INSERT_HEAD(INDIR_HASH(ump, freework->fw_blkno), freework, 2389 fw_next); 2390 freework->fw_state &= ~DEPCOMPLETE; 2391 } 2392 2393 static void 2394 indirblk_remove(freework) 2395 struct freework *freework; 2396 { 2397 struct ufsmount *ump; 2398 2399 ump = VFSTOUFS(freework->fw_list.wk_mp); 2400 LIST_REMOVE(freework, fw_segs); 2401 TAILQ_REMOVE(INDIR_HASH(ump, freework->fw_blkno), freework, fw_next); 2402 freework->fw_state |= DEPCOMPLETE; 2403 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE) 2404 WORKITEM_FREE(freework, D_FREEWORK); 2405 } 2406 2407 /* 2408 * Executed during filesystem system initialization before 2409 * mounting any filesystems. 2410 */ 2411 void 2412 softdep_initialize() 2413 { 2414 2415 TAILQ_INIT(&softdepmounts); 2416 #ifdef __LP64__ 2417 max_softdeps = desiredvnodes * 4; 2418 #else 2419 max_softdeps = desiredvnodes * 2; 2420 #endif 2421 2422 /* initialise bioops hack */ 2423 bioops.io_start = softdep_disk_io_initiation; 2424 bioops.io_complete = softdep_disk_write_complete; 2425 bioops.io_deallocate = softdep_deallocate_dependencies; 2426 bioops.io_countdeps = softdep_count_dependencies; 2427 softdep_ast_cleanup = softdep_ast_cleanup_proc; 2428 2429 /* Initialize the callout with an mtx. */ 2430 callout_init_mtx(&softdep_callout, &lk, 0); 2431 } 2432 2433 /* 2434 * Executed after all filesystems have been unmounted during 2435 * filesystem module unload. 2436 */ 2437 void 2438 softdep_uninitialize() 2439 { 2440 2441 /* clear bioops hack */ 2442 bioops.io_start = NULL; 2443 bioops.io_complete = NULL; 2444 bioops.io_deallocate = NULL; 2445 bioops.io_countdeps = NULL; 2446 softdep_ast_cleanup = NULL; 2447 2448 callout_drain(&softdep_callout); 2449 } 2450 2451 /* 2452 * Called at mount time to notify the dependency code that a 2453 * filesystem wishes to use it. 2454 */ 2455 int 2456 softdep_mount(devvp, mp, fs, cred) 2457 struct vnode *devvp; 2458 struct mount *mp; 2459 struct fs *fs; 2460 struct ucred *cred; 2461 { 2462 struct csum_total cstotal; 2463 struct mount_softdeps *sdp; 2464 struct ufsmount *ump; 2465 struct cg *cgp; 2466 struct buf *bp; 2467 int i, error, cyl; 2468 2469 sdp = malloc(sizeof(struct mount_softdeps), M_MOUNTDATA, 2470 M_WAITOK | M_ZERO); 2471 MNT_ILOCK(mp); 2472 mp->mnt_flag = (mp->mnt_flag & ~MNT_ASYNC) | MNT_SOFTDEP; 2473 if ((mp->mnt_kern_flag & MNTK_SOFTDEP) == 0) { 2474 mp->mnt_kern_flag = (mp->mnt_kern_flag & ~MNTK_ASYNC) | 2475 MNTK_SOFTDEP | MNTK_NOASYNC; 2476 } 2477 ump = VFSTOUFS(mp); 2478 ump->um_softdep = sdp; 2479 MNT_IUNLOCK(mp); 2480 rw_init(LOCK_PTR(ump), "Per-Filesystem Softdep Lock"); 2481 sdp->sd_ump = ump; 2482 LIST_INIT(&ump->softdep_workitem_pending); 2483 LIST_INIT(&ump->softdep_journal_pending); 2484 TAILQ_INIT(&ump->softdep_unlinked); 2485 LIST_INIT(&ump->softdep_dirtycg); 2486 ump->softdep_worklist_tail = NULL; 2487 ump->softdep_on_worklist = 0; 2488 ump->softdep_deps = 0; 2489 LIST_INIT(&ump->softdep_mkdirlisthd); 2490 ump->pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, 2491 &ump->pagedep_hash_size); 2492 ump->pagedep_nextclean = 0; 2493 ump->inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, 2494 &ump->inodedep_hash_size); 2495 ump->inodedep_nextclean = 0; 2496 ump->newblk_hashtbl = hashinit(max_softdeps / 2, M_NEWBLK, 2497 &ump->newblk_hash_size); 2498 ump->bmsafemap_hashtbl = hashinit(1024, M_BMSAFEMAP, 2499 &ump->bmsafemap_hash_size); 2500 i = 1 << (ffs(desiredvnodes / 10) - 1); 2501 ump->indir_hashtbl = malloc(i * sizeof(struct indir_hashhead), 2502 M_FREEWORK, M_WAITOK); 2503 ump->indir_hash_size = i - 1; 2504 for (i = 0; i <= ump->indir_hash_size; i++) 2505 TAILQ_INIT(&ump->indir_hashtbl[i]); 2506 ACQUIRE_GBLLOCK(&lk); 2507 TAILQ_INSERT_TAIL(&softdepmounts, sdp, sd_next); 2508 FREE_GBLLOCK(&lk); 2509 if ((fs->fs_flags & FS_SUJ) && 2510 (error = journal_mount(mp, fs, cred)) != 0) { 2511 printf("Failed to start journal: %d\n", error); 2512 softdep_unmount(mp); 2513 return (error); 2514 } 2515 /* 2516 * Start our flushing thread in the bufdaemon process. 2517 */ 2518 ACQUIRE_LOCK(ump); 2519 ump->softdep_flags |= FLUSH_STARTING; 2520 FREE_LOCK(ump); 2521 kproc_kthread_add(&softdep_flush, mp, &bufdaemonproc, 2522 &ump->softdep_flushtd, 0, 0, "softdepflush", "%s worker", 2523 mp->mnt_stat.f_mntonname); 2524 ACQUIRE_LOCK(ump); 2525 while ((ump->softdep_flags & FLUSH_STARTING) != 0) { 2526 msleep(&ump->softdep_flushtd, LOCK_PTR(ump), PVM, "sdstart", 2527 hz / 2); 2528 } 2529 FREE_LOCK(ump); 2530 /* 2531 * When doing soft updates, the counters in the 2532 * superblock may have gotten out of sync. Recomputation 2533 * can take a long time and can be deferred for background 2534 * fsck. However, the old behavior of scanning the cylinder 2535 * groups and recalculating them at mount time is available 2536 * by setting vfs.ffs.compute_summary_at_mount to one. 2537 */ 2538 if (compute_summary_at_mount == 0 || fs->fs_clean != 0) 2539 return (0); 2540 bzero(&cstotal, sizeof cstotal); 2541 for (cyl = 0; cyl < fs->fs_ncg; cyl++) { 2542 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)), 2543 fs->fs_cgsize, cred, &bp)) != 0) { 2544 brelse(bp); 2545 softdep_unmount(mp); 2546 return (error); 2547 } 2548 cgp = (struct cg *)bp->b_data; 2549 cstotal.cs_nffree += cgp->cg_cs.cs_nffree; 2550 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree; 2551 cstotal.cs_nifree += cgp->cg_cs.cs_nifree; 2552 cstotal.cs_ndir += cgp->cg_cs.cs_ndir; 2553 fs->fs_cs(fs, cyl) = cgp->cg_cs; 2554 brelse(bp); 2555 } 2556 #ifdef DEBUG 2557 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal)) 2558 printf("%s: superblock summary recomputed\n", fs->fs_fsmnt); 2559 #endif 2560 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal); 2561 return (0); 2562 } 2563 2564 void 2565 softdep_unmount(mp) 2566 struct mount *mp; 2567 { 2568 struct ufsmount *ump; 2569 #ifdef INVARIANTS 2570 int i; 2571 #endif 2572 2573 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 2574 ("softdep_unmount called on non-softdep filesystem")); 2575 ump = VFSTOUFS(mp); 2576 MNT_ILOCK(mp); 2577 mp->mnt_flag &= ~MNT_SOFTDEP; 2578 if (MOUNTEDSUJ(mp) == 0) { 2579 MNT_IUNLOCK(mp); 2580 } else { 2581 mp->mnt_flag &= ~MNT_SUJ; 2582 MNT_IUNLOCK(mp); 2583 journal_unmount(ump); 2584 } 2585 /* 2586 * Shut down our flushing thread. Check for NULL is if 2587 * softdep_mount errors out before the thread has been created. 2588 */ 2589 if (ump->softdep_flushtd != NULL) { 2590 ACQUIRE_LOCK(ump); 2591 ump->softdep_flags |= FLUSH_EXIT; 2592 wakeup(&ump->softdep_flushtd); 2593 msleep(&ump->softdep_flags, LOCK_PTR(ump), PVM | PDROP, 2594 "sdwait", 0); 2595 KASSERT((ump->softdep_flags & FLUSH_EXIT) == 0, 2596 ("Thread shutdown failed")); 2597 } 2598 /* 2599 * Free up our resources. 2600 */ 2601 ACQUIRE_GBLLOCK(&lk); 2602 TAILQ_REMOVE(&softdepmounts, ump->um_softdep, sd_next); 2603 FREE_GBLLOCK(&lk); 2604 rw_destroy(LOCK_PTR(ump)); 2605 hashdestroy(ump->pagedep_hashtbl, M_PAGEDEP, ump->pagedep_hash_size); 2606 hashdestroy(ump->inodedep_hashtbl, M_INODEDEP, ump->inodedep_hash_size); 2607 hashdestroy(ump->newblk_hashtbl, M_NEWBLK, ump->newblk_hash_size); 2608 hashdestroy(ump->bmsafemap_hashtbl, M_BMSAFEMAP, 2609 ump->bmsafemap_hash_size); 2610 free(ump->indir_hashtbl, M_FREEWORK); 2611 #ifdef INVARIANTS 2612 for (i = 0; i <= D_LAST; i++) 2613 KASSERT(ump->softdep_curdeps[i] == 0, 2614 ("Unmount %s: Dep type %s != 0 (%ld)", ump->um_fs->fs_fsmnt, 2615 TYPENAME(i), ump->softdep_curdeps[i])); 2616 #endif 2617 free(ump->um_softdep, M_MOUNTDATA); 2618 } 2619 2620 static struct jblocks * 2621 jblocks_create(void) 2622 { 2623 struct jblocks *jblocks; 2624 2625 jblocks = malloc(sizeof(*jblocks), M_JBLOCKS, M_WAITOK | M_ZERO); 2626 TAILQ_INIT(&jblocks->jb_segs); 2627 jblocks->jb_avail = 10; 2628 jblocks->jb_extent = malloc(sizeof(struct jextent) * jblocks->jb_avail, 2629 M_JBLOCKS, M_WAITOK | M_ZERO); 2630 2631 return (jblocks); 2632 } 2633 2634 static ufs2_daddr_t 2635 jblocks_alloc(jblocks, bytes, actual) 2636 struct jblocks *jblocks; 2637 int bytes; 2638 int *actual; 2639 { 2640 ufs2_daddr_t daddr; 2641 struct jextent *jext; 2642 int freecnt; 2643 int blocks; 2644 2645 blocks = bytes / DEV_BSIZE; 2646 jext = &jblocks->jb_extent[jblocks->jb_head]; 2647 freecnt = jext->je_blocks - jblocks->jb_off; 2648 if (freecnt == 0) { 2649 jblocks->jb_off = 0; 2650 if (++jblocks->jb_head > jblocks->jb_used) 2651 jblocks->jb_head = 0; 2652 jext = &jblocks->jb_extent[jblocks->jb_head]; 2653 freecnt = jext->je_blocks; 2654 } 2655 if (freecnt > blocks) 2656 freecnt = blocks; 2657 *actual = freecnt * DEV_BSIZE; 2658 daddr = jext->je_daddr + jblocks->jb_off; 2659 jblocks->jb_off += freecnt; 2660 jblocks->jb_free -= freecnt; 2661 2662 return (daddr); 2663 } 2664 2665 static void 2666 jblocks_free(jblocks, mp, bytes) 2667 struct jblocks *jblocks; 2668 struct mount *mp; 2669 int bytes; 2670 { 2671 2672 LOCK_OWNED(VFSTOUFS(mp)); 2673 jblocks->jb_free += bytes / DEV_BSIZE; 2674 if (jblocks->jb_suspended) 2675 worklist_speedup(mp); 2676 wakeup(jblocks); 2677 } 2678 2679 static void 2680 jblocks_destroy(jblocks) 2681 struct jblocks *jblocks; 2682 { 2683 2684 if (jblocks->jb_extent) 2685 free(jblocks->jb_extent, M_JBLOCKS); 2686 free(jblocks, M_JBLOCKS); 2687 } 2688 2689 static void 2690 jblocks_add(jblocks, daddr, blocks) 2691 struct jblocks *jblocks; 2692 ufs2_daddr_t daddr; 2693 int blocks; 2694 { 2695 struct jextent *jext; 2696 2697 jblocks->jb_blocks += blocks; 2698 jblocks->jb_free += blocks; 2699 jext = &jblocks->jb_extent[jblocks->jb_used]; 2700 /* Adding the first block. */ 2701 if (jext->je_daddr == 0) { 2702 jext->je_daddr = daddr; 2703 jext->je_blocks = blocks; 2704 return; 2705 } 2706 /* Extending the last extent. */ 2707 if (jext->je_daddr + jext->je_blocks == daddr) { 2708 jext->je_blocks += blocks; 2709 return; 2710 } 2711 /* Adding a new extent. */ 2712 if (++jblocks->jb_used == jblocks->jb_avail) { 2713 jblocks->jb_avail *= 2; 2714 jext = malloc(sizeof(struct jextent) * jblocks->jb_avail, 2715 M_JBLOCKS, M_WAITOK | M_ZERO); 2716 memcpy(jext, jblocks->jb_extent, 2717 sizeof(struct jextent) * jblocks->jb_used); 2718 free(jblocks->jb_extent, M_JBLOCKS); 2719 jblocks->jb_extent = jext; 2720 } 2721 jext = &jblocks->jb_extent[jblocks->jb_used]; 2722 jext->je_daddr = daddr; 2723 jext->je_blocks = blocks; 2724 return; 2725 } 2726 2727 int 2728 softdep_journal_lookup(mp, vpp) 2729 struct mount *mp; 2730 struct vnode **vpp; 2731 { 2732 struct componentname cnp; 2733 struct vnode *dvp; 2734 ino_t sujournal; 2735 int error; 2736 2737 error = VFS_VGET(mp, ROOTINO, LK_EXCLUSIVE, &dvp); 2738 if (error) 2739 return (error); 2740 bzero(&cnp, sizeof(cnp)); 2741 cnp.cn_nameiop = LOOKUP; 2742 cnp.cn_flags = ISLASTCN; 2743 cnp.cn_thread = curthread; 2744 cnp.cn_cred = curthread->td_ucred; 2745 cnp.cn_pnbuf = SUJ_FILE; 2746 cnp.cn_nameptr = SUJ_FILE; 2747 cnp.cn_namelen = strlen(SUJ_FILE); 2748 error = ufs_lookup_ino(dvp, NULL, &cnp, &sujournal); 2749 vput(dvp); 2750 if (error != 0) 2751 return (error); 2752 error = VFS_VGET(mp, sujournal, LK_EXCLUSIVE, vpp); 2753 return (error); 2754 } 2755 2756 /* 2757 * Open and verify the journal file. 2758 */ 2759 static int 2760 journal_mount(mp, fs, cred) 2761 struct mount *mp; 2762 struct fs *fs; 2763 struct ucred *cred; 2764 { 2765 struct jblocks *jblocks; 2766 struct ufsmount *ump; 2767 struct vnode *vp; 2768 struct inode *ip; 2769 ufs2_daddr_t blkno; 2770 int bcount; 2771 int error; 2772 int i; 2773 2774 ump = VFSTOUFS(mp); 2775 ump->softdep_journal_tail = NULL; 2776 ump->softdep_on_journal = 0; 2777 ump->softdep_accdeps = 0; 2778 ump->softdep_req = 0; 2779 ump->softdep_jblocks = NULL; 2780 error = softdep_journal_lookup(mp, &vp); 2781 if (error != 0) { 2782 printf("Failed to find journal. Use tunefs to create one\n"); 2783 return (error); 2784 } 2785 ip = VTOI(vp); 2786 if (ip->i_size < SUJ_MIN) { 2787 error = ENOSPC; 2788 goto out; 2789 } 2790 bcount = lblkno(fs, ip->i_size); /* Only use whole blocks. */ 2791 jblocks = jblocks_create(); 2792 for (i = 0; i < bcount; i++) { 2793 error = ufs_bmaparray(vp, i, &blkno, NULL, NULL, NULL); 2794 if (error) 2795 break; 2796 jblocks_add(jblocks, blkno, fsbtodb(fs, fs->fs_frag)); 2797 } 2798 if (error) { 2799 jblocks_destroy(jblocks); 2800 goto out; 2801 } 2802 jblocks->jb_low = jblocks->jb_free / 3; /* Reserve 33%. */ 2803 jblocks->jb_min = jblocks->jb_free / 10; /* Suspend at 10%. */ 2804 ump->softdep_jblocks = jblocks; 2805 out: 2806 if (error == 0) { 2807 MNT_ILOCK(mp); 2808 mp->mnt_flag |= MNT_SUJ; 2809 mp->mnt_flag &= ~MNT_SOFTDEP; 2810 MNT_IUNLOCK(mp); 2811 /* 2812 * Only validate the journal contents if the 2813 * filesystem is clean, otherwise we write the logs 2814 * but they'll never be used. If the filesystem was 2815 * still dirty when we mounted it the journal is 2816 * invalid and a new journal can only be valid if it 2817 * starts from a clean mount. 2818 */ 2819 if (fs->fs_clean) { 2820 DIP_SET(ip, i_modrev, fs->fs_mtime); 2821 ip->i_flags |= IN_MODIFIED; 2822 ffs_update(vp, 1); 2823 } 2824 } 2825 vput(vp); 2826 return (error); 2827 } 2828 2829 static void 2830 journal_unmount(ump) 2831 struct ufsmount *ump; 2832 { 2833 2834 if (ump->softdep_jblocks) 2835 jblocks_destroy(ump->softdep_jblocks); 2836 ump->softdep_jblocks = NULL; 2837 } 2838 2839 /* 2840 * Called when a journal record is ready to be written. Space is allocated 2841 * and the journal entry is created when the journal is flushed to stable 2842 * store. 2843 */ 2844 static void 2845 add_to_journal(wk) 2846 struct worklist *wk; 2847 { 2848 struct ufsmount *ump; 2849 2850 ump = VFSTOUFS(wk->wk_mp); 2851 LOCK_OWNED(ump); 2852 if (wk->wk_state & ONWORKLIST) 2853 panic("add_to_journal: %s(0x%X) already on list", 2854 TYPENAME(wk->wk_type), wk->wk_state); 2855 wk->wk_state |= ONWORKLIST | DEPCOMPLETE; 2856 if (LIST_EMPTY(&ump->softdep_journal_pending)) { 2857 ump->softdep_jblocks->jb_age = ticks; 2858 LIST_INSERT_HEAD(&ump->softdep_journal_pending, wk, wk_list); 2859 } else 2860 LIST_INSERT_AFTER(ump->softdep_journal_tail, wk, wk_list); 2861 ump->softdep_journal_tail = wk; 2862 ump->softdep_on_journal += 1; 2863 } 2864 2865 /* 2866 * Remove an arbitrary item for the journal worklist maintain the tail 2867 * pointer. This happens when a new operation obviates the need to 2868 * journal an old operation. 2869 */ 2870 static void 2871 remove_from_journal(wk) 2872 struct worklist *wk; 2873 { 2874 struct ufsmount *ump; 2875 2876 ump = VFSTOUFS(wk->wk_mp); 2877 LOCK_OWNED(ump); 2878 #ifdef SUJ_DEBUG 2879 { 2880 struct worklist *wkn; 2881 2882 LIST_FOREACH(wkn, &ump->softdep_journal_pending, wk_list) 2883 if (wkn == wk) 2884 break; 2885 if (wkn == NULL) 2886 panic("remove_from_journal: %p is not in journal", wk); 2887 } 2888 #endif 2889 /* 2890 * We emulate a TAILQ to save space in most structures which do not 2891 * require TAILQ semantics. Here we must update the tail position 2892 * when removing the tail which is not the final entry. This works 2893 * only if the worklist linkage are at the beginning of the structure. 2894 */ 2895 if (ump->softdep_journal_tail == wk) 2896 ump->softdep_journal_tail = 2897 (struct worklist *)wk->wk_list.le_prev; 2898 WORKLIST_REMOVE(wk); 2899 ump->softdep_on_journal -= 1; 2900 } 2901 2902 /* 2903 * Check for journal space as well as dependency limits so the prelink 2904 * code can throttle both journaled and non-journaled filesystems. 2905 * Threshold is 0 for low and 1 for min. 2906 */ 2907 static int 2908 journal_space(ump, thresh) 2909 struct ufsmount *ump; 2910 int thresh; 2911 { 2912 struct jblocks *jblocks; 2913 int limit, avail; 2914 2915 jblocks = ump->softdep_jblocks; 2916 if (jblocks == NULL) 2917 return (1); 2918 /* 2919 * We use a tighter restriction here to prevent request_cleanup() 2920 * running in threads from running into locks we currently hold. 2921 * We have to be over the limit and our filesystem has to be 2922 * responsible for more than our share of that usage. 2923 */ 2924 limit = (max_softdeps / 10) * 9; 2925 if (dep_current[D_INODEDEP] > limit && 2926 ump->softdep_curdeps[D_INODEDEP] > limit / stat_flush_threads) 2927 return (0); 2928 if (thresh) 2929 thresh = jblocks->jb_min; 2930 else 2931 thresh = jblocks->jb_low; 2932 avail = (ump->softdep_on_journal * JREC_SIZE) / DEV_BSIZE; 2933 avail = jblocks->jb_free - avail; 2934 2935 return (avail > thresh); 2936 } 2937 2938 static void 2939 journal_suspend(ump) 2940 struct ufsmount *ump; 2941 { 2942 struct jblocks *jblocks; 2943 struct mount *mp; 2944 2945 mp = UFSTOVFS(ump); 2946 jblocks = ump->softdep_jblocks; 2947 MNT_ILOCK(mp); 2948 if ((mp->mnt_kern_flag & MNTK_SUSPEND) == 0) { 2949 stat_journal_min++; 2950 mp->mnt_kern_flag |= MNTK_SUSPEND; 2951 mp->mnt_susp_owner = ump->softdep_flushtd; 2952 } 2953 jblocks->jb_suspended = 1; 2954 MNT_IUNLOCK(mp); 2955 } 2956 2957 static int 2958 journal_unsuspend(struct ufsmount *ump) 2959 { 2960 struct jblocks *jblocks; 2961 struct mount *mp; 2962 2963 mp = UFSTOVFS(ump); 2964 jblocks = ump->softdep_jblocks; 2965 2966 if (jblocks != NULL && jblocks->jb_suspended && 2967 journal_space(ump, jblocks->jb_min)) { 2968 jblocks->jb_suspended = 0; 2969 FREE_LOCK(ump); 2970 mp->mnt_susp_owner = curthread; 2971 vfs_write_resume(mp, 0); 2972 ACQUIRE_LOCK(ump); 2973 return (1); 2974 } 2975 return (0); 2976 } 2977 2978 /* 2979 * Called before any allocation function to be certain that there is 2980 * sufficient space in the journal prior to creating any new records. 2981 * Since in the case of block allocation we may have multiple locked 2982 * buffers at the time of the actual allocation we can not block 2983 * when the journal records are created. Doing so would create a deadlock 2984 * if any of these buffers needed to be flushed to reclaim space. Instead 2985 * we require a sufficiently large amount of available space such that 2986 * each thread in the system could have passed this allocation check and 2987 * still have sufficient free space. With 20% of a minimum journal size 2988 * of 1MB we have 6553 records available. 2989 */ 2990 int 2991 softdep_prealloc(vp, waitok) 2992 struct vnode *vp; 2993 int waitok; 2994 { 2995 struct ufsmount *ump; 2996 2997 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 2998 ("softdep_prealloc called on non-softdep filesystem")); 2999 /* 3000 * Nothing to do if we are not running journaled soft updates. 3001 * If we currently hold the snapshot lock, we must avoid 3002 * handling other resources that could cause deadlock. Do not 3003 * touch quotas vnode since it is typically recursed with 3004 * other vnode locks held. 3005 */ 3006 if (DOINGSUJ(vp) == 0 || IS_SNAPSHOT(VTOI(vp)) || 3007 (vp->v_vflag & VV_SYSTEM) != 0) 3008 return (0); 3009 ump = VFSTOUFS(vp->v_mount); 3010 ACQUIRE_LOCK(ump); 3011 if (journal_space(ump, 0)) { 3012 FREE_LOCK(ump); 3013 return (0); 3014 } 3015 stat_journal_low++; 3016 FREE_LOCK(ump); 3017 if (waitok == MNT_NOWAIT) 3018 return (ENOSPC); 3019 /* 3020 * Attempt to sync this vnode once to flush any journal 3021 * work attached to it. 3022 */ 3023 if ((curthread->td_pflags & TDP_COWINPROGRESS) == 0) 3024 ffs_syncvnode(vp, waitok, 0); 3025 ACQUIRE_LOCK(ump); 3026 process_removes(vp); 3027 process_truncates(vp); 3028 if (journal_space(ump, 0) == 0) { 3029 softdep_speedup(ump); 3030 if (journal_space(ump, 1) == 0) 3031 journal_suspend(ump); 3032 } 3033 FREE_LOCK(ump); 3034 3035 return (0); 3036 } 3037 3038 /* 3039 * Before adjusting a link count on a vnode verify that we have sufficient 3040 * journal space. If not, process operations that depend on the currently 3041 * locked pair of vnodes to try to flush space as the syncer, buf daemon, 3042 * and softdep flush threads can not acquire these locks to reclaim space. 3043 */ 3044 static void 3045 softdep_prelink(dvp, vp) 3046 struct vnode *dvp; 3047 struct vnode *vp; 3048 { 3049 struct ufsmount *ump; 3050 3051 ump = VFSTOUFS(dvp->v_mount); 3052 LOCK_OWNED(ump); 3053 /* 3054 * Nothing to do if we have sufficient journal space. 3055 * If we currently hold the snapshot lock, we must avoid 3056 * handling other resources that could cause deadlock. 3057 */ 3058 if (journal_space(ump, 0) || (vp && IS_SNAPSHOT(VTOI(vp)))) 3059 return; 3060 stat_journal_low++; 3061 FREE_LOCK(ump); 3062 if (vp) 3063 ffs_syncvnode(vp, MNT_NOWAIT, 0); 3064 ffs_syncvnode(dvp, MNT_WAIT, 0); 3065 ACQUIRE_LOCK(ump); 3066 /* Process vp before dvp as it may create .. removes. */ 3067 if (vp) { 3068 process_removes(vp); 3069 process_truncates(vp); 3070 } 3071 process_removes(dvp); 3072 process_truncates(dvp); 3073 softdep_speedup(ump); 3074 process_worklist_item(UFSTOVFS(ump), 2, LK_NOWAIT); 3075 if (journal_space(ump, 0) == 0) { 3076 softdep_speedup(ump); 3077 if (journal_space(ump, 1) == 0) 3078 journal_suspend(ump); 3079 } 3080 } 3081 3082 static void 3083 jseg_write(ump, jseg, data) 3084 struct ufsmount *ump; 3085 struct jseg *jseg; 3086 uint8_t *data; 3087 { 3088 struct jsegrec *rec; 3089 3090 rec = (struct jsegrec *)data; 3091 rec->jsr_seq = jseg->js_seq; 3092 rec->jsr_oldest = jseg->js_oldseq; 3093 rec->jsr_cnt = jseg->js_cnt; 3094 rec->jsr_blocks = jseg->js_size / ump->um_devvp->v_bufobj.bo_bsize; 3095 rec->jsr_crc = 0; 3096 rec->jsr_time = ump->um_fs->fs_mtime; 3097 } 3098 3099 static inline void 3100 inoref_write(inoref, jseg, rec) 3101 struct inoref *inoref; 3102 struct jseg *jseg; 3103 struct jrefrec *rec; 3104 { 3105 3106 inoref->if_jsegdep->jd_seg = jseg; 3107 rec->jr_ino = inoref->if_ino; 3108 rec->jr_parent = inoref->if_parent; 3109 rec->jr_nlink = inoref->if_nlink; 3110 rec->jr_mode = inoref->if_mode; 3111 rec->jr_diroff = inoref->if_diroff; 3112 } 3113 3114 static void 3115 jaddref_write(jaddref, jseg, data) 3116 struct jaddref *jaddref; 3117 struct jseg *jseg; 3118 uint8_t *data; 3119 { 3120 struct jrefrec *rec; 3121 3122 rec = (struct jrefrec *)data; 3123 rec->jr_op = JOP_ADDREF; 3124 inoref_write(&jaddref->ja_ref, jseg, rec); 3125 } 3126 3127 static void 3128 jremref_write(jremref, jseg, data) 3129 struct jremref *jremref; 3130 struct jseg *jseg; 3131 uint8_t *data; 3132 { 3133 struct jrefrec *rec; 3134 3135 rec = (struct jrefrec *)data; 3136 rec->jr_op = JOP_REMREF; 3137 inoref_write(&jremref->jr_ref, jseg, rec); 3138 } 3139 3140 static void 3141 jmvref_write(jmvref, jseg, data) 3142 struct jmvref *jmvref; 3143 struct jseg *jseg; 3144 uint8_t *data; 3145 { 3146 struct jmvrec *rec; 3147 3148 rec = (struct jmvrec *)data; 3149 rec->jm_op = JOP_MVREF; 3150 rec->jm_ino = jmvref->jm_ino; 3151 rec->jm_parent = jmvref->jm_parent; 3152 rec->jm_oldoff = jmvref->jm_oldoff; 3153 rec->jm_newoff = jmvref->jm_newoff; 3154 } 3155 3156 static void 3157 jnewblk_write(jnewblk, jseg, data) 3158 struct jnewblk *jnewblk; 3159 struct jseg *jseg; 3160 uint8_t *data; 3161 { 3162 struct jblkrec *rec; 3163 3164 jnewblk->jn_jsegdep->jd_seg = jseg; 3165 rec = (struct jblkrec *)data; 3166 rec->jb_op = JOP_NEWBLK; 3167 rec->jb_ino = jnewblk->jn_ino; 3168 rec->jb_blkno = jnewblk->jn_blkno; 3169 rec->jb_lbn = jnewblk->jn_lbn; 3170 rec->jb_frags = jnewblk->jn_frags; 3171 rec->jb_oldfrags = jnewblk->jn_oldfrags; 3172 } 3173 3174 static void 3175 jfreeblk_write(jfreeblk, jseg, data) 3176 struct jfreeblk *jfreeblk; 3177 struct jseg *jseg; 3178 uint8_t *data; 3179 { 3180 struct jblkrec *rec; 3181 3182 jfreeblk->jf_dep.jb_jsegdep->jd_seg = jseg; 3183 rec = (struct jblkrec *)data; 3184 rec->jb_op = JOP_FREEBLK; 3185 rec->jb_ino = jfreeblk->jf_ino; 3186 rec->jb_blkno = jfreeblk->jf_blkno; 3187 rec->jb_lbn = jfreeblk->jf_lbn; 3188 rec->jb_frags = jfreeblk->jf_frags; 3189 rec->jb_oldfrags = 0; 3190 } 3191 3192 static void 3193 jfreefrag_write(jfreefrag, jseg, data) 3194 struct jfreefrag *jfreefrag; 3195 struct jseg *jseg; 3196 uint8_t *data; 3197 { 3198 struct jblkrec *rec; 3199 3200 jfreefrag->fr_jsegdep->jd_seg = jseg; 3201 rec = (struct jblkrec *)data; 3202 rec->jb_op = JOP_FREEBLK; 3203 rec->jb_ino = jfreefrag->fr_ino; 3204 rec->jb_blkno = jfreefrag->fr_blkno; 3205 rec->jb_lbn = jfreefrag->fr_lbn; 3206 rec->jb_frags = jfreefrag->fr_frags; 3207 rec->jb_oldfrags = 0; 3208 } 3209 3210 static void 3211 jtrunc_write(jtrunc, jseg, data) 3212 struct jtrunc *jtrunc; 3213 struct jseg *jseg; 3214 uint8_t *data; 3215 { 3216 struct jtrncrec *rec; 3217 3218 jtrunc->jt_dep.jb_jsegdep->jd_seg = jseg; 3219 rec = (struct jtrncrec *)data; 3220 rec->jt_op = JOP_TRUNC; 3221 rec->jt_ino = jtrunc->jt_ino; 3222 rec->jt_size = jtrunc->jt_size; 3223 rec->jt_extsize = jtrunc->jt_extsize; 3224 } 3225 3226 static void 3227 jfsync_write(jfsync, jseg, data) 3228 struct jfsync *jfsync; 3229 struct jseg *jseg; 3230 uint8_t *data; 3231 { 3232 struct jtrncrec *rec; 3233 3234 rec = (struct jtrncrec *)data; 3235 rec->jt_op = JOP_SYNC; 3236 rec->jt_ino = jfsync->jfs_ino; 3237 rec->jt_size = jfsync->jfs_size; 3238 rec->jt_extsize = jfsync->jfs_extsize; 3239 } 3240 3241 static void 3242 softdep_flushjournal(mp) 3243 struct mount *mp; 3244 { 3245 struct jblocks *jblocks; 3246 struct ufsmount *ump; 3247 3248 if (MOUNTEDSUJ(mp) == 0) 3249 return; 3250 ump = VFSTOUFS(mp); 3251 jblocks = ump->softdep_jblocks; 3252 ACQUIRE_LOCK(ump); 3253 while (ump->softdep_on_journal) { 3254 jblocks->jb_needseg = 1; 3255 softdep_process_journal(mp, NULL, MNT_WAIT); 3256 } 3257 FREE_LOCK(ump); 3258 } 3259 3260 static void softdep_synchronize_completed(struct bio *); 3261 static void softdep_synchronize(struct bio *, struct ufsmount *, void *); 3262 3263 static void 3264 softdep_synchronize_completed(bp) 3265 struct bio *bp; 3266 { 3267 struct jseg *oldest; 3268 struct jseg *jseg; 3269 struct ufsmount *ump; 3270 3271 /* 3272 * caller1 marks the last segment written before we issued the 3273 * synchronize cache. 3274 */ 3275 jseg = bp->bio_caller1; 3276 if (jseg == NULL) { 3277 g_destroy_bio(bp); 3278 return; 3279 } 3280 ump = VFSTOUFS(jseg->js_list.wk_mp); 3281 ACQUIRE_LOCK(ump); 3282 oldest = NULL; 3283 /* 3284 * Mark all the journal entries waiting on the synchronize cache 3285 * as completed so they may continue on. 3286 */ 3287 while (jseg != NULL && (jseg->js_state & COMPLETE) == 0) { 3288 jseg->js_state |= COMPLETE; 3289 oldest = jseg; 3290 jseg = TAILQ_PREV(jseg, jseglst, js_next); 3291 } 3292 /* 3293 * Restart deferred journal entry processing from the oldest 3294 * completed jseg. 3295 */ 3296 if (oldest) 3297 complete_jsegs(oldest); 3298 3299 FREE_LOCK(ump); 3300 g_destroy_bio(bp); 3301 } 3302 3303 /* 3304 * Send BIO_FLUSH/SYNCHRONIZE CACHE to the device to enforce write ordering 3305 * barriers. The journal must be written prior to any blocks that depend 3306 * on it and the journal can not be released until the blocks have be 3307 * written. This code handles both barriers simultaneously. 3308 */ 3309 static void 3310 softdep_synchronize(bp, ump, caller1) 3311 struct bio *bp; 3312 struct ufsmount *ump; 3313 void *caller1; 3314 { 3315 3316 bp->bio_cmd = BIO_FLUSH; 3317 bp->bio_flags |= BIO_ORDERED; 3318 bp->bio_data = NULL; 3319 bp->bio_offset = ump->um_cp->provider->mediasize; 3320 bp->bio_length = 0; 3321 bp->bio_done = softdep_synchronize_completed; 3322 bp->bio_caller1 = caller1; 3323 g_io_request(bp, 3324 (struct g_consumer *)ump->um_devvp->v_bufobj.bo_private); 3325 } 3326 3327 /* 3328 * Flush some journal records to disk. 3329 */ 3330 static void 3331 softdep_process_journal(mp, needwk, flags) 3332 struct mount *mp; 3333 struct worklist *needwk; 3334 int flags; 3335 { 3336 struct jblocks *jblocks; 3337 struct ufsmount *ump; 3338 struct worklist *wk; 3339 struct jseg *jseg; 3340 struct buf *bp; 3341 struct bio *bio; 3342 uint8_t *data; 3343 struct fs *fs; 3344 int shouldflush; 3345 int segwritten; 3346 int jrecmin; /* Minimum records per block. */ 3347 int jrecmax; /* Maximum records per block. */ 3348 int size; 3349 int cnt; 3350 int off; 3351 int devbsize; 3352 3353 if (MOUNTEDSUJ(mp) == 0) 3354 return; 3355 shouldflush = softdep_flushcache; 3356 bio = NULL; 3357 jseg = NULL; 3358 ump = VFSTOUFS(mp); 3359 LOCK_OWNED(ump); 3360 fs = ump->um_fs; 3361 jblocks = ump->softdep_jblocks; 3362 devbsize = ump->um_devvp->v_bufobj.bo_bsize; 3363 /* 3364 * We write anywhere between a disk block and fs block. The upper 3365 * bound is picked to prevent buffer cache fragmentation and limit 3366 * processing time per I/O. 3367 */ 3368 jrecmin = (devbsize / JREC_SIZE) - 1; /* -1 for seg header */ 3369 jrecmax = (fs->fs_bsize / devbsize) * jrecmin; 3370 segwritten = 0; 3371 for (;;) { 3372 cnt = ump->softdep_on_journal; 3373 /* 3374 * Criteria for writing a segment: 3375 * 1) We have a full block. 3376 * 2) We're called from jwait() and haven't found the 3377 * journal item yet. 3378 * 3) Always write if needseg is set. 3379 * 4) If we are called from process_worklist and have 3380 * not yet written anything we write a partial block 3381 * to enforce a 1 second maximum latency on journal 3382 * entries. 3383 */ 3384 if (cnt < (jrecmax - 1) && needwk == NULL && 3385 jblocks->jb_needseg == 0 && (segwritten || cnt == 0)) 3386 break; 3387 cnt++; 3388 /* 3389 * Verify some free journal space. softdep_prealloc() should 3390 * guarantee that we don't run out so this is indicative of 3391 * a problem with the flow control. Try to recover 3392 * gracefully in any event. 3393 */ 3394 while (jblocks->jb_free == 0) { 3395 if (flags != MNT_WAIT) 3396 break; 3397 printf("softdep: Out of journal space!\n"); 3398 softdep_speedup(ump); 3399 msleep(jblocks, LOCK_PTR(ump), PRIBIO, "jblocks", hz); 3400 } 3401 FREE_LOCK(ump); 3402 jseg = malloc(sizeof(*jseg), M_JSEG, M_SOFTDEP_FLAGS); 3403 workitem_alloc(&jseg->js_list, D_JSEG, mp); 3404 LIST_INIT(&jseg->js_entries); 3405 LIST_INIT(&jseg->js_indirs); 3406 jseg->js_state = ATTACHED; 3407 if (shouldflush == 0) 3408 jseg->js_state |= COMPLETE; 3409 else if (bio == NULL) 3410 bio = g_alloc_bio(); 3411 jseg->js_jblocks = jblocks; 3412 bp = geteblk(fs->fs_bsize, 0); 3413 ACQUIRE_LOCK(ump); 3414 /* 3415 * If there was a race while we were allocating the block 3416 * and jseg the entry we care about was likely written. 3417 * We bail out in both the WAIT and NOWAIT case and assume 3418 * the caller will loop if the entry it cares about is 3419 * not written. 3420 */ 3421 cnt = ump->softdep_on_journal; 3422 if (cnt + jblocks->jb_needseg == 0 || jblocks->jb_free == 0) { 3423 bp->b_flags |= B_INVAL | B_NOCACHE; 3424 WORKITEM_FREE(jseg, D_JSEG); 3425 FREE_LOCK(ump); 3426 brelse(bp); 3427 ACQUIRE_LOCK(ump); 3428 break; 3429 } 3430 /* 3431 * Calculate the disk block size required for the available 3432 * records rounded to the min size. 3433 */ 3434 if (cnt == 0) 3435 size = devbsize; 3436 else if (cnt < jrecmax) 3437 size = howmany(cnt, jrecmin) * devbsize; 3438 else 3439 size = fs->fs_bsize; 3440 /* 3441 * Allocate a disk block for this journal data and account 3442 * for truncation of the requested size if enough contiguous 3443 * space was not available. 3444 */ 3445 bp->b_blkno = jblocks_alloc(jblocks, size, &size); 3446 bp->b_lblkno = bp->b_blkno; 3447 bp->b_offset = bp->b_blkno * DEV_BSIZE; 3448 bp->b_bcount = size; 3449 bp->b_flags &= ~B_INVAL; 3450 bp->b_flags |= B_VALIDSUSPWRT | B_NOCOPY; 3451 /* 3452 * Initialize our jseg with cnt records. Assign the next 3453 * sequence number to it and link it in-order. 3454 */ 3455 cnt = MIN(cnt, (size / devbsize) * jrecmin); 3456 jseg->js_buf = bp; 3457 jseg->js_cnt = cnt; 3458 jseg->js_refs = cnt + 1; /* Self ref. */ 3459 jseg->js_size = size; 3460 jseg->js_seq = jblocks->jb_nextseq++; 3461 if (jblocks->jb_oldestseg == NULL) 3462 jblocks->jb_oldestseg = jseg; 3463 jseg->js_oldseq = jblocks->jb_oldestseg->js_seq; 3464 TAILQ_INSERT_TAIL(&jblocks->jb_segs, jseg, js_next); 3465 if (jblocks->jb_writeseg == NULL) 3466 jblocks->jb_writeseg = jseg; 3467 /* 3468 * Start filling in records from the pending list. 3469 */ 3470 data = bp->b_data; 3471 off = 0; 3472 3473 /* 3474 * Always put a header on the first block. 3475 * XXX As with below, there might not be a chance to get 3476 * into the loop. Ensure that something valid is written. 3477 */ 3478 jseg_write(ump, jseg, data); 3479 off += JREC_SIZE; 3480 data = bp->b_data + off; 3481 3482 /* 3483 * XXX Something is wrong here. There's no work to do, 3484 * but we need to perform and I/O and allow it to complete 3485 * anyways. 3486 */ 3487 if (LIST_EMPTY(&ump->softdep_journal_pending)) 3488 stat_emptyjblocks++; 3489 3490 while ((wk = LIST_FIRST(&ump->softdep_journal_pending)) 3491 != NULL) { 3492 if (cnt == 0) 3493 break; 3494 /* Place a segment header on every device block. */ 3495 if ((off % devbsize) == 0) { 3496 jseg_write(ump, jseg, data); 3497 off += JREC_SIZE; 3498 data = bp->b_data + off; 3499 } 3500 if (wk == needwk) 3501 needwk = NULL; 3502 remove_from_journal(wk); 3503 wk->wk_state |= INPROGRESS; 3504 WORKLIST_INSERT(&jseg->js_entries, wk); 3505 switch (wk->wk_type) { 3506 case D_JADDREF: 3507 jaddref_write(WK_JADDREF(wk), jseg, data); 3508 break; 3509 case D_JREMREF: 3510 jremref_write(WK_JREMREF(wk), jseg, data); 3511 break; 3512 case D_JMVREF: 3513 jmvref_write(WK_JMVREF(wk), jseg, data); 3514 break; 3515 case D_JNEWBLK: 3516 jnewblk_write(WK_JNEWBLK(wk), jseg, data); 3517 break; 3518 case D_JFREEBLK: 3519 jfreeblk_write(WK_JFREEBLK(wk), jseg, data); 3520 break; 3521 case D_JFREEFRAG: 3522 jfreefrag_write(WK_JFREEFRAG(wk), jseg, data); 3523 break; 3524 case D_JTRUNC: 3525 jtrunc_write(WK_JTRUNC(wk), jseg, data); 3526 break; 3527 case D_JFSYNC: 3528 jfsync_write(WK_JFSYNC(wk), jseg, data); 3529 break; 3530 default: 3531 panic("process_journal: Unknown type %s", 3532 TYPENAME(wk->wk_type)); 3533 /* NOTREACHED */ 3534 } 3535 off += JREC_SIZE; 3536 data = bp->b_data + off; 3537 cnt--; 3538 } 3539 3540 /* Clear any remaining space so we don't leak kernel data */ 3541 if (size > off) 3542 bzero(data, size - off); 3543 3544 /* 3545 * Write this one buffer and continue. 3546 */ 3547 segwritten = 1; 3548 jblocks->jb_needseg = 0; 3549 WORKLIST_INSERT(&bp->b_dep, &jseg->js_list); 3550 FREE_LOCK(ump); 3551 pbgetvp(ump->um_devvp, bp); 3552 /* 3553 * We only do the blocking wait once we find the journal 3554 * entry we're looking for. 3555 */ 3556 if (needwk == NULL && flags == MNT_WAIT) 3557 bwrite(bp); 3558 else 3559 bawrite(bp); 3560 ACQUIRE_LOCK(ump); 3561 } 3562 /* 3563 * If we wrote a segment issue a synchronize cache so the journal 3564 * is reflected on disk before the data is written. Since reclaiming 3565 * journal space also requires writing a journal record this 3566 * process also enforces a barrier before reclamation. 3567 */ 3568 if (segwritten && shouldflush) { 3569 softdep_synchronize(bio, ump, 3570 TAILQ_LAST(&jblocks->jb_segs, jseglst)); 3571 } else if (bio) 3572 g_destroy_bio(bio); 3573 /* 3574 * If we've suspended the filesystem because we ran out of journal 3575 * space either try to sync it here to make some progress or 3576 * unsuspend it if we already have. 3577 */ 3578 if (flags == 0 && jblocks->jb_suspended) { 3579 if (journal_unsuspend(ump)) 3580 return; 3581 FREE_LOCK(ump); 3582 VFS_SYNC(mp, MNT_NOWAIT); 3583 ffs_sbupdate(ump, MNT_WAIT, 0); 3584 ACQUIRE_LOCK(ump); 3585 } 3586 } 3587 3588 /* 3589 * Complete a jseg, allowing all dependencies awaiting journal writes 3590 * to proceed. Each journal dependency also attaches a jsegdep to dependent 3591 * structures so that the journal segment can be freed to reclaim space. 3592 */ 3593 static void 3594 complete_jseg(jseg) 3595 struct jseg *jseg; 3596 { 3597 struct worklist *wk; 3598 struct jmvref *jmvref; 3599 #ifdef INVARIANTS 3600 int i = 0; 3601 #endif 3602 3603 while ((wk = LIST_FIRST(&jseg->js_entries)) != NULL) { 3604 WORKLIST_REMOVE(wk); 3605 wk->wk_state &= ~INPROGRESS; 3606 wk->wk_state |= COMPLETE; 3607 KASSERT(i++ < jseg->js_cnt, 3608 ("handle_written_jseg: overflow %d >= %d", 3609 i - 1, jseg->js_cnt)); 3610 switch (wk->wk_type) { 3611 case D_JADDREF: 3612 handle_written_jaddref(WK_JADDREF(wk)); 3613 break; 3614 case D_JREMREF: 3615 handle_written_jremref(WK_JREMREF(wk)); 3616 break; 3617 case D_JMVREF: 3618 rele_jseg(jseg); /* No jsegdep. */ 3619 jmvref = WK_JMVREF(wk); 3620 LIST_REMOVE(jmvref, jm_deps); 3621 if ((jmvref->jm_pagedep->pd_state & ONWORKLIST) == 0) 3622 free_pagedep(jmvref->jm_pagedep); 3623 WORKITEM_FREE(jmvref, D_JMVREF); 3624 break; 3625 case D_JNEWBLK: 3626 handle_written_jnewblk(WK_JNEWBLK(wk)); 3627 break; 3628 case D_JFREEBLK: 3629 handle_written_jblkdep(&WK_JFREEBLK(wk)->jf_dep); 3630 break; 3631 case D_JTRUNC: 3632 handle_written_jblkdep(&WK_JTRUNC(wk)->jt_dep); 3633 break; 3634 case D_JFSYNC: 3635 rele_jseg(jseg); /* No jsegdep. */ 3636 WORKITEM_FREE(wk, D_JFSYNC); 3637 break; 3638 case D_JFREEFRAG: 3639 handle_written_jfreefrag(WK_JFREEFRAG(wk)); 3640 break; 3641 default: 3642 panic("handle_written_jseg: Unknown type %s", 3643 TYPENAME(wk->wk_type)); 3644 /* NOTREACHED */ 3645 } 3646 } 3647 /* Release the self reference so the structure may be freed. */ 3648 rele_jseg(jseg); 3649 } 3650 3651 /* 3652 * Determine which jsegs are ready for completion processing. Waits for 3653 * synchronize cache to complete as well as forcing in-order completion 3654 * of journal entries. 3655 */ 3656 static void 3657 complete_jsegs(jseg) 3658 struct jseg *jseg; 3659 { 3660 struct jblocks *jblocks; 3661 struct jseg *jsegn; 3662 3663 jblocks = jseg->js_jblocks; 3664 /* 3665 * Don't allow out of order completions. If this isn't the first 3666 * block wait for it to write before we're done. 3667 */ 3668 if (jseg != jblocks->jb_writeseg) 3669 return; 3670 /* Iterate through available jsegs processing their entries. */ 3671 while (jseg && (jseg->js_state & ALLCOMPLETE) == ALLCOMPLETE) { 3672 jblocks->jb_oldestwrseq = jseg->js_oldseq; 3673 jsegn = TAILQ_NEXT(jseg, js_next); 3674 complete_jseg(jseg); 3675 jseg = jsegn; 3676 } 3677 jblocks->jb_writeseg = jseg; 3678 /* 3679 * Attempt to free jsegs now that oldestwrseq may have advanced. 3680 */ 3681 free_jsegs(jblocks); 3682 } 3683 3684 /* 3685 * Mark a jseg as DEPCOMPLETE and throw away the buffer. Attempt to handle 3686 * the final completions. 3687 */ 3688 static void 3689 handle_written_jseg(jseg, bp) 3690 struct jseg *jseg; 3691 struct buf *bp; 3692 { 3693 3694 if (jseg->js_refs == 0) 3695 panic("handle_written_jseg: No self-reference on %p", jseg); 3696 jseg->js_state |= DEPCOMPLETE; 3697 /* 3698 * We'll never need this buffer again, set flags so it will be 3699 * discarded. 3700 */ 3701 bp->b_flags |= B_INVAL | B_NOCACHE; 3702 pbrelvp(bp); 3703 complete_jsegs(jseg); 3704 } 3705 3706 static inline struct jsegdep * 3707 inoref_jseg(inoref) 3708 struct inoref *inoref; 3709 { 3710 struct jsegdep *jsegdep; 3711 3712 jsegdep = inoref->if_jsegdep; 3713 inoref->if_jsegdep = NULL; 3714 3715 return (jsegdep); 3716 } 3717 3718 /* 3719 * Called once a jremref has made it to stable store. The jremref is marked 3720 * complete and we attempt to free it. Any pagedeps writes sleeping waiting 3721 * for the jremref to complete will be awoken by free_jremref. 3722 */ 3723 static void 3724 handle_written_jremref(jremref) 3725 struct jremref *jremref; 3726 { 3727 struct inodedep *inodedep; 3728 struct jsegdep *jsegdep; 3729 struct dirrem *dirrem; 3730 3731 /* Grab the jsegdep. */ 3732 jsegdep = inoref_jseg(&jremref->jr_ref); 3733 /* 3734 * Remove us from the inoref list. 3735 */ 3736 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 3737 0, &inodedep) == 0) 3738 panic("handle_written_jremref: Lost inodedep"); 3739 TAILQ_REMOVE(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps); 3740 /* 3741 * Complete the dirrem. 3742 */ 3743 dirrem = jremref->jr_dirrem; 3744 jremref->jr_dirrem = NULL; 3745 LIST_REMOVE(jremref, jr_deps); 3746 jsegdep->jd_state |= jremref->jr_state & MKDIR_PARENT; 3747 jwork_insert(&dirrem->dm_jwork, jsegdep); 3748 if (LIST_EMPTY(&dirrem->dm_jremrefhd) && 3749 (dirrem->dm_state & COMPLETE) != 0) 3750 add_to_worklist(&dirrem->dm_list, 0); 3751 free_jremref(jremref); 3752 } 3753 3754 /* 3755 * Called once a jaddref has made it to stable store. The dependency is 3756 * marked complete and any dependent structures are added to the inode 3757 * bufwait list to be completed as soon as it is written. If a bitmap write 3758 * depends on this entry we move the inode into the inodedephd of the 3759 * bmsafemap dependency and attempt to remove the jaddref from the bmsafemap. 3760 */ 3761 static void 3762 handle_written_jaddref(jaddref) 3763 struct jaddref *jaddref; 3764 { 3765 struct jsegdep *jsegdep; 3766 struct inodedep *inodedep; 3767 struct diradd *diradd; 3768 struct mkdir *mkdir; 3769 3770 /* Grab the jsegdep. */ 3771 jsegdep = inoref_jseg(&jaddref->ja_ref); 3772 mkdir = NULL; 3773 diradd = NULL; 3774 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino, 3775 0, &inodedep) == 0) 3776 panic("handle_written_jaddref: Lost inodedep."); 3777 if (jaddref->ja_diradd == NULL) 3778 panic("handle_written_jaddref: No dependency"); 3779 if (jaddref->ja_diradd->da_list.wk_type == D_DIRADD) { 3780 diradd = jaddref->ja_diradd; 3781 WORKLIST_INSERT(&inodedep->id_bufwait, &diradd->da_list); 3782 } else if (jaddref->ja_state & MKDIR_PARENT) { 3783 mkdir = jaddref->ja_mkdir; 3784 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir->md_list); 3785 } else if (jaddref->ja_state & MKDIR_BODY) 3786 mkdir = jaddref->ja_mkdir; 3787 else 3788 panic("handle_written_jaddref: Unknown dependency %p", 3789 jaddref->ja_diradd); 3790 jaddref->ja_diradd = NULL; /* also clears ja_mkdir */ 3791 /* 3792 * Remove us from the inode list. 3793 */ 3794 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, if_deps); 3795 /* 3796 * The mkdir may be waiting on the jaddref to clear before freeing. 3797 */ 3798 if (mkdir) { 3799 KASSERT(mkdir->md_list.wk_type == D_MKDIR, 3800 ("handle_written_jaddref: Incorrect type for mkdir %s", 3801 TYPENAME(mkdir->md_list.wk_type))); 3802 mkdir->md_jaddref = NULL; 3803 diradd = mkdir->md_diradd; 3804 mkdir->md_state |= DEPCOMPLETE; 3805 complete_mkdir(mkdir); 3806 } 3807 jwork_insert(&diradd->da_jwork, jsegdep); 3808 if (jaddref->ja_state & NEWBLOCK) { 3809 inodedep->id_state |= ONDEPLIST; 3810 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_inodedephd, 3811 inodedep, id_deps); 3812 } 3813 free_jaddref(jaddref); 3814 } 3815 3816 /* 3817 * Called once a jnewblk journal is written. The allocdirect or allocindir 3818 * is placed in the bmsafemap to await notification of a written bitmap. If 3819 * the operation was canceled we add the segdep to the appropriate 3820 * dependency to free the journal space once the canceling operation 3821 * completes. 3822 */ 3823 static void 3824 handle_written_jnewblk(jnewblk) 3825 struct jnewblk *jnewblk; 3826 { 3827 struct bmsafemap *bmsafemap; 3828 struct freefrag *freefrag; 3829 struct freework *freework; 3830 struct jsegdep *jsegdep; 3831 struct newblk *newblk; 3832 3833 /* Grab the jsegdep. */ 3834 jsegdep = jnewblk->jn_jsegdep; 3835 jnewblk->jn_jsegdep = NULL; 3836 if (jnewblk->jn_dep == NULL) 3837 panic("handle_written_jnewblk: No dependency for the segdep."); 3838 switch (jnewblk->jn_dep->wk_type) { 3839 case D_NEWBLK: 3840 case D_ALLOCDIRECT: 3841 case D_ALLOCINDIR: 3842 /* 3843 * Add the written block to the bmsafemap so it can 3844 * be notified when the bitmap is on disk. 3845 */ 3846 newblk = WK_NEWBLK(jnewblk->jn_dep); 3847 newblk->nb_jnewblk = NULL; 3848 if ((newblk->nb_state & GOINGAWAY) == 0) { 3849 bmsafemap = newblk->nb_bmsafemap; 3850 newblk->nb_state |= ONDEPLIST; 3851 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, 3852 nb_deps); 3853 } 3854 jwork_insert(&newblk->nb_jwork, jsegdep); 3855 break; 3856 case D_FREEFRAG: 3857 /* 3858 * A newblock being removed by a freefrag when replaced by 3859 * frag extension. 3860 */ 3861 freefrag = WK_FREEFRAG(jnewblk->jn_dep); 3862 freefrag->ff_jdep = NULL; 3863 jwork_insert(&freefrag->ff_jwork, jsegdep); 3864 break; 3865 case D_FREEWORK: 3866 /* 3867 * A direct block was removed by truncate. 3868 */ 3869 freework = WK_FREEWORK(jnewblk->jn_dep); 3870 freework->fw_jnewblk = NULL; 3871 jwork_insert(&freework->fw_freeblks->fb_jwork, jsegdep); 3872 break; 3873 default: 3874 panic("handle_written_jnewblk: Unknown type %d.", 3875 jnewblk->jn_dep->wk_type); 3876 } 3877 jnewblk->jn_dep = NULL; 3878 free_jnewblk(jnewblk); 3879 } 3880 3881 /* 3882 * Cancel a jfreefrag that won't be needed, probably due to colliding with 3883 * an in-flight allocation that has not yet been committed. Divorce us 3884 * from the freefrag and mark it DEPCOMPLETE so that it may be added 3885 * to the worklist. 3886 */ 3887 static void 3888 cancel_jfreefrag(jfreefrag) 3889 struct jfreefrag *jfreefrag; 3890 { 3891 struct freefrag *freefrag; 3892 3893 if (jfreefrag->fr_jsegdep) { 3894 free_jsegdep(jfreefrag->fr_jsegdep); 3895 jfreefrag->fr_jsegdep = NULL; 3896 } 3897 freefrag = jfreefrag->fr_freefrag; 3898 jfreefrag->fr_freefrag = NULL; 3899 free_jfreefrag(jfreefrag); 3900 freefrag->ff_state |= DEPCOMPLETE; 3901 CTR1(KTR_SUJ, "cancel_jfreefrag: blkno %jd", freefrag->ff_blkno); 3902 } 3903 3904 /* 3905 * Free a jfreefrag when the parent freefrag is rendered obsolete. 3906 */ 3907 static void 3908 free_jfreefrag(jfreefrag) 3909 struct jfreefrag *jfreefrag; 3910 { 3911 3912 if (jfreefrag->fr_state & INPROGRESS) 3913 WORKLIST_REMOVE(&jfreefrag->fr_list); 3914 else if (jfreefrag->fr_state & ONWORKLIST) 3915 remove_from_journal(&jfreefrag->fr_list); 3916 if (jfreefrag->fr_freefrag != NULL) 3917 panic("free_jfreefrag: Still attached to a freefrag."); 3918 WORKITEM_FREE(jfreefrag, D_JFREEFRAG); 3919 } 3920 3921 /* 3922 * Called when the journal write for a jfreefrag completes. The parent 3923 * freefrag is added to the worklist if this completes its dependencies. 3924 */ 3925 static void 3926 handle_written_jfreefrag(jfreefrag) 3927 struct jfreefrag *jfreefrag; 3928 { 3929 struct jsegdep *jsegdep; 3930 struct freefrag *freefrag; 3931 3932 /* Grab the jsegdep. */ 3933 jsegdep = jfreefrag->fr_jsegdep; 3934 jfreefrag->fr_jsegdep = NULL; 3935 freefrag = jfreefrag->fr_freefrag; 3936 if (freefrag == NULL) 3937 panic("handle_written_jfreefrag: No freefrag."); 3938 freefrag->ff_state |= DEPCOMPLETE; 3939 freefrag->ff_jdep = NULL; 3940 jwork_insert(&freefrag->ff_jwork, jsegdep); 3941 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE) 3942 add_to_worklist(&freefrag->ff_list, 0); 3943 jfreefrag->fr_freefrag = NULL; 3944 free_jfreefrag(jfreefrag); 3945 } 3946 3947 /* 3948 * Called when the journal write for a jfreeblk completes. The jfreeblk 3949 * is removed from the freeblks list of pending journal writes and the 3950 * jsegdep is moved to the freeblks jwork to be completed when all blocks 3951 * have been reclaimed. 3952 */ 3953 static void 3954 handle_written_jblkdep(jblkdep) 3955 struct jblkdep *jblkdep; 3956 { 3957 struct freeblks *freeblks; 3958 struct jsegdep *jsegdep; 3959 3960 /* Grab the jsegdep. */ 3961 jsegdep = jblkdep->jb_jsegdep; 3962 jblkdep->jb_jsegdep = NULL; 3963 freeblks = jblkdep->jb_freeblks; 3964 LIST_REMOVE(jblkdep, jb_deps); 3965 jwork_insert(&freeblks->fb_jwork, jsegdep); 3966 /* 3967 * If the freeblks is all journaled, we can add it to the worklist. 3968 */ 3969 if (LIST_EMPTY(&freeblks->fb_jblkdephd) && 3970 (freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 3971 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 3972 3973 free_jblkdep(jblkdep); 3974 } 3975 3976 static struct jsegdep * 3977 newjsegdep(struct worklist *wk) 3978 { 3979 struct jsegdep *jsegdep; 3980 3981 jsegdep = malloc(sizeof(*jsegdep), M_JSEGDEP, M_SOFTDEP_FLAGS); 3982 workitem_alloc(&jsegdep->jd_list, D_JSEGDEP, wk->wk_mp); 3983 jsegdep->jd_seg = NULL; 3984 3985 return (jsegdep); 3986 } 3987 3988 static struct jmvref * 3989 newjmvref(dp, ino, oldoff, newoff) 3990 struct inode *dp; 3991 ino_t ino; 3992 off_t oldoff; 3993 off_t newoff; 3994 { 3995 struct jmvref *jmvref; 3996 3997 jmvref = malloc(sizeof(*jmvref), M_JMVREF, M_SOFTDEP_FLAGS); 3998 workitem_alloc(&jmvref->jm_list, D_JMVREF, ITOVFS(dp)); 3999 jmvref->jm_list.wk_state = ATTACHED | DEPCOMPLETE; 4000 jmvref->jm_parent = dp->i_number; 4001 jmvref->jm_ino = ino; 4002 jmvref->jm_oldoff = oldoff; 4003 jmvref->jm_newoff = newoff; 4004 4005 return (jmvref); 4006 } 4007 4008 /* 4009 * Allocate a new jremref that tracks the removal of ip from dp with the 4010 * directory entry offset of diroff. Mark the entry as ATTACHED and 4011 * DEPCOMPLETE as we have all the information required for the journal write 4012 * and the directory has already been removed from the buffer. The caller 4013 * is responsible for linking the jremref into the pagedep and adding it 4014 * to the journal to write. The MKDIR_PARENT flag is set if we're doing 4015 * a DOTDOT addition so handle_workitem_remove() can properly assign 4016 * the jsegdep when we're done. 4017 */ 4018 static struct jremref * 4019 newjremref(struct dirrem *dirrem, struct inode *dp, struct inode *ip, 4020 off_t diroff, nlink_t nlink) 4021 { 4022 struct jremref *jremref; 4023 4024 jremref = malloc(sizeof(*jremref), M_JREMREF, M_SOFTDEP_FLAGS); 4025 workitem_alloc(&jremref->jr_list, D_JREMREF, ITOVFS(dp)); 4026 jremref->jr_state = ATTACHED; 4027 newinoref(&jremref->jr_ref, ip->i_number, dp->i_number, diroff, 4028 nlink, ip->i_mode); 4029 jremref->jr_dirrem = dirrem; 4030 4031 return (jremref); 4032 } 4033 4034 static inline void 4035 newinoref(struct inoref *inoref, ino_t ino, ino_t parent, off_t diroff, 4036 nlink_t nlink, uint16_t mode) 4037 { 4038 4039 inoref->if_jsegdep = newjsegdep(&inoref->if_list); 4040 inoref->if_diroff = diroff; 4041 inoref->if_ino = ino; 4042 inoref->if_parent = parent; 4043 inoref->if_nlink = nlink; 4044 inoref->if_mode = mode; 4045 } 4046 4047 /* 4048 * Allocate a new jaddref to track the addition of ino to dp at diroff. The 4049 * directory offset may not be known until later. The caller is responsible 4050 * adding the entry to the journal when this information is available. nlink 4051 * should be the link count prior to the addition and mode is only required 4052 * to have the correct FMT. 4053 */ 4054 static struct jaddref * 4055 newjaddref(struct inode *dp, ino_t ino, off_t diroff, int16_t nlink, 4056 uint16_t mode) 4057 { 4058 struct jaddref *jaddref; 4059 4060 jaddref = malloc(sizeof(*jaddref), M_JADDREF, M_SOFTDEP_FLAGS); 4061 workitem_alloc(&jaddref->ja_list, D_JADDREF, ITOVFS(dp)); 4062 jaddref->ja_state = ATTACHED; 4063 jaddref->ja_mkdir = NULL; 4064 newinoref(&jaddref->ja_ref, ino, dp->i_number, diroff, nlink, mode); 4065 4066 return (jaddref); 4067 } 4068 4069 /* 4070 * Create a new free dependency for a freework. The caller is responsible 4071 * for adjusting the reference count when it has the lock held. The freedep 4072 * will track an outstanding bitmap write that will ultimately clear the 4073 * freework to continue. 4074 */ 4075 static struct freedep * 4076 newfreedep(struct freework *freework) 4077 { 4078 struct freedep *freedep; 4079 4080 freedep = malloc(sizeof(*freedep), M_FREEDEP, M_SOFTDEP_FLAGS); 4081 workitem_alloc(&freedep->fd_list, D_FREEDEP, freework->fw_list.wk_mp); 4082 freedep->fd_freework = freework; 4083 4084 return (freedep); 4085 } 4086 4087 /* 4088 * Free a freedep structure once the buffer it is linked to is written. If 4089 * this is the last reference to the freework schedule it for completion. 4090 */ 4091 static void 4092 free_freedep(freedep) 4093 struct freedep *freedep; 4094 { 4095 struct freework *freework; 4096 4097 freework = freedep->fd_freework; 4098 freework->fw_freeblks->fb_cgwait--; 4099 if (--freework->fw_ref == 0) 4100 freework_enqueue(freework); 4101 WORKITEM_FREE(freedep, D_FREEDEP); 4102 } 4103 4104 /* 4105 * Allocate a new freework structure that may be a level in an indirect 4106 * when parent is not NULL or a top level block when it is. The top level 4107 * freework structures are allocated without the per-filesystem lock held 4108 * and before the freeblks is visible outside of softdep_setup_freeblocks(). 4109 */ 4110 static struct freework * 4111 newfreework(ump, freeblks, parent, lbn, nb, frags, off, journal) 4112 struct ufsmount *ump; 4113 struct freeblks *freeblks; 4114 struct freework *parent; 4115 ufs_lbn_t lbn; 4116 ufs2_daddr_t nb; 4117 int frags; 4118 int off; 4119 int journal; 4120 { 4121 struct freework *freework; 4122 4123 freework = malloc(sizeof(*freework), M_FREEWORK, M_SOFTDEP_FLAGS); 4124 workitem_alloc(&freework->fw_list, D_FREEWORK, freeblks->fb_list.wk_mp); 4125 freework->fw_state = ATTACHED; 4126 freework->fw_jnewblk = NULL; 4127 freework->fw_freeblks = freeblks; 4128 freework->fw_parent = parent; 4129 freework->fw_lbn = lbn; 4130 freework->fw_blkno = nb; 4131 freework->fw_frags = frags; 4132 freework->fw_indir = NULL; 4133 freework->fw_ref = (MOUNTEDSUJ(UFSTOVFS(ump)) == 0 || lbn >= -NXADDR) 4134 ? 0 : NINDIR(ump->um_fs) + 1; 4135 freework->fw_start = freework->fw_off = off; 4136 if (journal) 4137 newjfreeblk(freeblks, lbn, nb, frags); 4138 if (parent == NULL) { 4139 ACQUIRE_LOCK(ump); 4140 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list); 4141 freeblks->fb_ref++; 4142 FREE_LOCK(ump); 4143 } 4144 4145 return (freework); 4146 } 4147 4148 /* 4149 * Eliminate a jfreeblk for a block that does not need journaling. 4150 */ 4151 static void 4152 cancel_jfreeblk(freeblks, blkno) 4153 struct freeblks *freeblks; 4154 ufs2_daddr_t blkno; 4155 { 4156 struct jfreeblk *jfreeblk; 4157 struct jblkdep *jblkdep; 4158 4159 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) { 4160 if (jblkdep->jb_list.wk_type != D_JFREEBLK) 4161 continue; 4162 jfreeblk = WK_JFREEBLK(&jblkdep->jb_list); 4163 if (jfreeblk->jf_blkno == blkno) 4164 break; 4165 } 4166 if (jblkdep == NULL) 4167 return; 4168 CTR1(KTR_SUJ, "cancel_jfreeblk: blkno %jd", blkno); 4169 free_jsegdep(jblkdep->jb_jsegdep); 4170 LIST_REMOVE(jblkdep, jb_deps); 4171 WORKITEM_FREE(jfreeblk, D_JFREEBLK); 4172 } 4173 4174 /* 4175 * Allocate a new jfreeblk to journal top level block pointer when truncating 4176 * a file. The caller must add this to the worklist when the per-filesystem 4177 * lock is held. 4178 */ 4179 static struct jfreeblk * 4180 newjfreeblk(freeblks, lbn, blkno, frags) 4181 struct freeblks *freeblks; 4182 ufs_lbn_t lbn; 4183 ufs2_daddr_t blkno; 4184 int frags; 4185 { 4186 struct jfreeblk *jfreeblk; 4187 4188 jfreeblk = malloc(sizeof(*jfreeblk), M_JFREEBLK, M_SOFTDEP_FLAGS); 4189 workitem_alloc(&jfreeblk->jf_dep.jb_list, D_JFREEBLK, 4190 freeblks->fb_list.wk_mp); 4191 jfreeblk->jf_dep.jb_jsegdep = newjsegdep(&jfreeblk->jf_dep.jb_list); 4192 jfreeblk->jf_dep.jb_freeblks = freeblks; 4193 jfreeblk->jf_ino = freeblks->fb_inum; 4194 jfreeblk->jf_lbn = lbn; 4195 jfreeblk->jf_blkno = blkno; 4196 jfreeblk->jf_frags = frags; 4197 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jfreeblk->jf_dep, jb_deps); 4198 4199 return (jfreeblk); 4200 } 4201 4202 /* 4203 * The journal is only prepared to handle full-size block numbers, so we 4204 * have to adjust the record to reflect the change to a full-size block. 4205 * For example, suppose we have a block made up of fragments 8-15 and 4206 * want to free its last two fragments. We are given a request that says: 4207 * FREEBLK ino=5, blkno=14, lbn=0, frags=2, oldfrags=0 4208 * where frags are the number of fragments to free and oldfrags are the 4209 * number of fragments to keep. To block align it, we have to change it to 4210 * have a valid full-size blkno, so it becomes: 4211 * FREEBLK ino=5, blkno=8, lbn=0, frags=2, oldfrags=6 4212 */ 4213 static void 4214 adjust_newfreework(freeblks, frag_offset) 4215 struct freeblks *freeblks; 4216 int frag_offset; 4217 { 4218 struct jfreeblk *jfreeblk; 4219 4220 KASSERT((LIST_FIRST(&freeblks->fb_jblkdephd) != NULL && 4221 LIST_FIRST(&freeblks->fb_jblkdephd)->jb_list.wk_type == D_JFREEBLK), 4222 ("adjust_newfreework: Missing freeblks dependency")); 4223 4224 jfreeblk = WK_JFREEBLK(LIST_FIRST(&freeblks->fb_jblkdephd)); 4225 jfreeblk->jf_blkno -= frag_offset; 4226 jfreeblk->jf_frags += frag_offset; 4227 } 4228 4229 /* 4230 * Allocate a new jtrunc to track a partial truncation. 4231 */ 4232 static struct jtrunc * 4233 newjtrunc(freeblks, size, extsize) 4234 struct freeblks *freeblks; 4235 off_t size; 4236 int extsize; 4237 { 4238 struct jtrunc *jtrunc; 4239 4240 jtrunc = malloc(sizeof(*jtrunc), M_JTRUNC, M_SOFTDEP_FLAGS); 4241 workitem_alloc(&jtrunc->jt_dep.jb_list, D_JTRUNC, 4242 freeblks->fb_list.wk_mp); 4243 jtrunc->jt_dep.jb_jsegdep = newjsegdep(&jtrunc->jt_dep.jb_list); 4244 jtrunc->jt_dep.jb_freeblks = freeblks; 4245 jtrunc->jt_ino = freeblks->fb_inum; 4246 jtrunc->jt_size = size; 4247 jtrunc->jt_extsize = extsize; 4248 LIST_INSERT_HEAD(&freeblks->fb_jblkdephd, &jtrunc->jt_dep, jb_deps); 4249 4250 return (jtrunc); 4251 } 4252 4253 /* 4254 * If we're canceling a new bitmap we have to search for another ref 4255 * to move into the bmsafemap dep. This might be better expressed 4256 * with another structure. 4257 */ 4258 static void 4259 move_newblock_dep(jaddref, inodedep) 4260 struct jaddref *jaddref; 4261 struct inodedep *inodedep; 4262 { 4263 struct inoref *inoref; 4264 struct jaddref *jaddrefn; 4265 4266 jaddrefn = NULL; 4267 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref; 4268 inoref = TAILQ_NEXT(inoref, if_deps)) { 4269 if ((jaddref->ja_state & NEWBLOCK) && 4270 inoref->if_list.wk_type == D_JADDREF) { 4271 jaddrefn = (struct jaddref *)inoref; 4272 break; 4273 } 4274 } 4275 if (jaddrefn == NULL) 4276 return; 4277 jaddrefn->ja_state &= ~(ATTACHED | UNDONE); 4278 jaddrefn->ja_state |= jaddref->ja_state & 4279 (ATTACHED | UNDONE | NEWBLOCK); 4280 jaddref->ja_state &= ~(ATTACHED | UNDONE | NEWBLOCK); 4281 jaddref->ja_state |= ATTACHED; 4282 LIST_REMOVE(jaddref, ja_bmdeps); 4283 LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_jaddrefhd, jaddrefn, 4284 ja_bmdeps); 4285 } 4286 4287 /* 4288 * Cancel a jaddref either before it has been written or while it is being 4289 * written. This happens when a link is removed before the add reaches 4290 * the disk. The jaddref dependency is kept linked into the bmsafemap 4291 * and inode to prevent the link count or bitmap from reaching the disk 4292 * until handle_workitem_remove() re-adjusts the counts and bitmaps as 4293 * required. 4294 * 4295 * Returns 1 if the canceled addref requires journaling of the remove and 4296 * 0 otherwise. 4297 */ 4298 static int 4299 cancel_jaddref(jaddref, inodedep, wkhd) 4300 struct jaddref *jaddref; 4301 struct inodedep *inodedep; 4302 struct workhead *wkhd; 4303 { 4304 struct inoref *inoref; 4305 struct jsegdep *jsegdep; 4306 int needsj; 4307 4308 KASSERT((jaddref->ja_state & COMPLETE) == 0, 4309 ("cancel_jaddref: Canceling complete jaddref")); 4310 if (jaddref->ja_state & (INPROGRESS | COMPLETE)) 4311 needsj = 1; 4312 else 4313 needsj = 0; 4314 if (inodedep == NULL) 4315 if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino, 4316 0, &inodedep) == 0) 4317 panic("cancel_jaddref: Lost inodedep"); 4318 /* 4319 * We must adjust the nlink of any reference operation that follows 4320 * us so that it is consistent with the in-memory reference. This 4321 * ensures that inode nlink rollbacks always have the correct link. 4322 */ 4323 if (needsj == 0) { 4324 for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref; 4325 inoref = TAILQ_NEXT(inoref, if_deps)) { 4326 if (inoref->if_state & GOINGAWAY) 4327 break; 4328 inoref->if_nlink--; 4329 } 4330 } 4331 jsegdep = inoref_jseg(&jaddref->ja_ref); 4332 if (jaddref->ja_state & NEWBLOCK) 4333 move_newblock_dep(jaddref, inodedep); 4334 wake_worklist(&jaddref->ja_list); 4335 jaddref->ja_mkdir = NULL; 4336 if (jaddref->ja_state & INPROGRESS) { 4337 jaddref->ja_state &= ~INPROGRESS; 4338 WORKLIST_REMOVE(&jaddref->ja_list); 4339 jwork_insert(wkhd, jsegdep); 4340 } else { 4341 free_jsegdep(jsegdep); 4342 if (jaddref->ja_state & DEPCOMPLETE) 4343 remove_from_journal(&jaddref->ja_list); 4344 } 4345 jaddref->ja_state |= (GOINGAWAY | DEPCOMPLETE); 4346 /* 4347 * Leave NEWBLOCK jaddrefs on the inodedep so handle_workitem_remove 4348 * can arrange for them to be freed with the bitmap. Otherwise we 4349 * no longer need this addref attached to the inoreflst and it 4350 * will incorrectly adjust nlink if we leave it. 4351 */ 4352 if ((jaddref->ja_state & NEWBLOCK) == 0) { 4353 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, 4354 if_deps); 4355 jaddref->ja_state |= COMPLETE; 4356 free_jaddref(jaddref); 4357 return (needsj); 4358 } 4359 /* 4360 * Leave the head of the list for jsegdeps for fast merging. 4361 */ 4362 if (LIST_FIRST(wkhd) != NULL) { 4363 jaddref->ja_state |= ONWORKLIST; 4364 LIST_INSERT_AFTER(LIST_FIRST(wkhd), &jaddref->ja_list, wk_list); 4365 } else 4366 WORKLIST_INSERT(wkhd, &jaddref->ja_list); 4367 4368 return (needsj); 4369 } 4370 4371 /* 4372 * Attempt to free a jaddref structure when some work completes. This 4373 * should only succeed once the entry is written and all dependencies have 4374 * been notified. 4375 */ 4376 static void 4377 free_jaddref(jaddref) 4378 struct jaddref *jaddref; 4379 { 4380 4381 if ((jaddref->ja_state & ALLCOMPLETE) != ALLCOMPLETE) 4382 return; 4383 if (jaddref->ja_ref.if_jsegdep) 4384 panic("free_jaddref: segdep attached to jaddref %p(0x%X)\n", 4385 jaddref, jaddref->ja_state); 4386 if (jaddref->ja_state & NEWBLOCK) 4387 LIST_REMOVE(jaddref, ja_bmdeps); 4388 if (jaddref->ja_state & (INPROGRESS | ONWORKLIST)) 4389 panic("free_jaddref: Bad state %p(0x%X)", 4390 jaddref, jaddref->ja_state); 4391 if (jaddref->ja_mkdir != NULL) 4392 panic("free_jaddref: Work pending, 0x%X\n", jaddref->ja_state); 4393 WORKITEM_FREE(jaddref, D_JADDREF); 4394 } 4395 4396 /* 4397 * Free a jremref structure once it has been written or discarded. 4398 */ 4399 static void 4400 free_jremref(jremref) 4401 struct jremref *jremref; 4402 { 4403 4404 if (jremref->jr_ref.if_jsegdep) 4405 free_jsegdep(jremref->jr_ref.if_jsegdep); 4406 if (jremref->jr_state & INPROGRESS) 4407 panic("free_jremref: IO still pending"); 4408 WORKITEM_FREE(jremref, D_JREMREF); 4409 } 4410 4411 /* 4412 * Free a jnewblk structure. 4413 */ 4414 static void 4415 free_jnewblk(jnewblk) 4416 struct jnewblk *jnewblk; 4417 { 4418 4419 if ((jnewblk->jn_state & ALLCOMPLETE) != ALLCOMPLETE) 4420 return; 4421 LIST_REMOVE(jnewblk, jn_deps); 4422 if (jnewblk->jn_dep != NULL) 4423 panic("free_jnewblk: Dependency still attached."); 4424 WORKITEM_FREE(jnewblk, D_JNEWBLK); 4425 } 4426 4427 /* 4428 * Cancel a jnewblk which has been been made redundant by frag extension. 4429 */ 4430 static void 4431 cancel_jnewblk(jnewblk, wkhd) 4432 struct jnewblk *jnewblk; 4433 struct workhead *wkhd; 4434 { 4435 struct jsegdep *jsegdep; 4436 4437 CTR1(KTR_SUJ, "cancel_jnewblk: blkno %jd", jnewblk->jn_blkno); 4438 jsegdep = jnewblk->jn_jsegdep; 4439 if (jnewblk->jn_jsegdep == NULL || jnewblk->jn_dep == NULL) 4440 panic("cancel_jnewblk: Invalid state"); 4441 jnewblk->jn_jsegdep = NULL; 4442 jnewblk->jn_dep = NULL; 4443 jnewblk->jn_state |= GOINGAWAY; 4444 if (jnewblk->jn_state & INPROGRESS) { 4445 jnewblk->jn_state &= ~INPROGRESS; 4446 WORKLIST_REMOVE(&jnewblk->jn_list); 4447 jwork_insert(wkhd, jsegdep); 4448 } else { 4449 free_jsegdep(jsegdep); 4450 remove_from_journal(&jnewblk->jn_list); 4451 } 4452 wake_worklist(&jnewblk->jn_list); 4453 WORKLIST_INSERT(wkhd, &jnewblk->jn_list); 4454 } 4455 4456 static void 4457 free_jblkdep(jblkdep) 4458 struct jblkdep *jblkdep; 4459 { 4460 4461 if (jblkdep->jb_list.wk_type == D_JFREEBLK) 4462 WORKITEM_FREE(jblkdep, D_JFREEBLK); 4463 else if (jblkdep->jb_list.wk_type == D_JTRUNC) 4464 WORKITEM_FREE(jblkdep, D_JTRUNC); 4465 else 4466 panic("free_jblkdep: Unexpected type %s", 4467 TYPENAME(jblkdep->jb_list.wk_type)); 4468 } 4469 4470 /* 4471 * Free a single jseg once it is no longer referenced in memory or on 4472 * disk. Reclaim journal blocks and dependencies waiting for the segment 4473 * to disappear. 4474 */ 4475 static void 4476 free_jseg(jseg, jblocks) 4477 struct jseg *jseg; 4478 struct jblocks *jblocks; 4479 { 4480 struct freework *freework; 4481 4482 /* 4483 * Free freework structures that were lingering to indicate freed 4484 * indirect blocks that forced journal write ordering on reallocate. 4485 */ 4486 while ((freework = LIST_FIRST(&jseg->js_indirs)) != NULL) 4487 indirblk_remove(freework); 4488 if (jblocks->jb_oldestseg == jseg) 4489 jblocks->jb_oldestseg = TAILQ_NEXT(jseg, js_next); 4490 TAILQ_REMOVE(&jblocks->jb_segs, jseg, js_next); 4491 jblocks_free(jblocks, jseg->js_list.wk_mp, jseg->js_size); 4492 KASSERT(LIST_EMPTY(&jseg->js_entries), 4493 ("free_jseg: Freed jseg has valid entries.")); 4494 WORKITEM_FREE(jseg, D_JSEG); 4495 } 4496 4497 /* 4498 * Free all jsegs that meet the criteria for being reclaimed and update 4499 * oldestseg. 4500 */ 4501 static void 4502 free_jsegs(jblocks) 4503 struct jblocks *jblocks; 4504 { 4505 struct jseg *jseg; 4506 4507 /* 4508 * Free only those jsegs which have none allocated before them to 4509 * preserve the journal space ordering. 4510 */ 4511 while ((jseg = TAILQ_FIRST(&jblocks->jb_segs)) != NULL) { 4512 /* 4513 * Only reclaim space when nothing depends on this journal 4514 * set and another set has written that it is no longer 4515 * valid. 4516 */ 4517 if (jseg->js_refs != 0) { 4518 jblocks->jb_oldestseg = jseg; 4519 return; 4520 } 4521 if ((jseg->js_state & ALLCOMPLETE) != ALLCOMPLETE) 4522 break; 4523 if (jseg->js_seq > jblocks->jb_oldestwrseq) 4524 break; 4525 /* 4526 * We can free jsegs that didn't write entries when 4527 * oldestwrseq == js_seq. 4528 */ 4529 if (jseg->js_seq == jblocks->jb_oldestwrseq && 4530 jseg->js_cnt != 0) 4531 break; 4532 free_jseg(jseg, jblocks); 4533 } 4534 /* 4535 * If we exited the loop above we still must discover the 4536 * oldest valid segment. 4537 */ 4538 if (jseg) 4539 for (jseg = jblocks->jb_oldestseg; jseg != NULL; 4540 jseg = TAILQ_NEXT(jseg, js_next)) 4541 if (jseg->js_refs != 0) 4542 break; 4543 jblocks->jb_oldestseg = jseg; 4544 /* 4545 * The journal has no valid records but some jsegs may still be 4546 * waiting on oldestwrseq to advance. We force a small record 4547 * out to permit these lingering records to be reclaimed. 4548 */ 4549 if (jblocks->jb_oldestseg == NULL && !TAILQ_EMPTY(&jblocks->jb_segs)) 4550 jblocks->jb_needseg = 1; 4551 } 4552 4553 /* 4554 * Release one reference to a jseg and free it if the count reaches 0. This 4555 * should eventually reclaim journal space as well. 4556 */ 4557 static void 4558 rele_jseg(jseg) 4559 struct jseg *jseg; 4560 { 4561 4562 KASSERT(jseg->js_refs > 0, 4563 ("free_jseg: Invalid refcnt %d", jseg->js_refs)); 4564 if (--jseg->js_refs != 0) 4565 return; 4566 free_jsegs(jseg->js_jblocks); 4567 } 4568 4569 /* 4570 * Release a jsegdep and decrement the jseg count. 4571 */ 4572 static void 4573 free_jsegdep(jsegdep) 4574 struct jsegdep *jsegdep; 4575 { 4576 4577 if (jsegdep->jd_seg) 4578 rele_jseg(jsegdep->jd_seg); 4579 WORKITEM_FREE(jsegdep, D_JSEGDEP); 4580 } 4581 4582 /* 4583 * Wait for a journal item to make it to disk. Initiate journal processing 4584 * if required. 4585 */ 4586 static int 4587 jwait(wk, waitfor) 4588 struct worklist *wk; 4589 int waitfor; 4590 { 4591 4592 LOCK_OWNED(VFSTOUFS(wk->wk_mp)); 4593 /* 4594 * Blocking journal waits cause slow synchronous behavior. Record 4595 * stats on the frequency of these blocking operations. 4596 */ 4597 if (waitfor == MNT_WAIT) { 4598 stat_journal_wait++; 4599 switch (wk->wk_type) { 4600 case D_JREMREF: 4601 case D_JMVREF: 4602 stat_jwait_filepage++; 4603 break; 4604 case D_JTRUNC: 4605 case D_JFREEBLK: 4606 stat_jwait_freeblks++; 4607 break; 4608 case D_JNEWBLK: 4609 stat_jwait_newblk++; 4610 break; 4611 case D_JADDREF: 4612 stat_jwait_inode++; 4613 break; 4614 default: 4615 break; 4616 } 4617 } 4618 /* 4619 * If IO has not started we process the journal. We can't mark the 4620 * worklist item as IOWAITING because we drop the lock while 4621 * processing the journal and the worklist entry may be freed after 4622 * this point. The caller may call back in and re-issue the request. 4623 */ 4624 if ((wk->wk_state & INPROGRESS) == 0) { 4625 softdep_process_journal(wk->wk_mp, wk, waitfor); 4626 if (waitfor != MNT_WAIT) 4627 return (EBUSY); 4628 return (0); 4629 } 4630 if (waitfor != MNT_WAIT) 4631 return (EBUSY); 4632 wait_worklist(wk, "jwait"); 4633 return (0); 4634 } 4635 4636 /* 4637 * Lookup an inodedep based on an inode pointer and set the nlinkdelta as 4638 * appropriate. This is a convenience function to reduce duplicate code 4639 * for the setup and revert functions below. 4640 */ 4641 static struct inodedep * 4642 inodedep_lookup_ip(ip) 4643 struct inode *ip; 4644 { 4645 struct inodedep *inodedep; 4646 4647 KASSERT(ip->i_nlink >= ip->i_effnlink, 4648 ("inodedep_lookup_ip: bad delta")); 4649 (void) inodedep_lookup(ITOVFS(ip), ip->i_number, DEPALLOC, 4650 &inodedep); 4651 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 4652 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked")); 4653 4654 return (inodedep); 4655 } 4656 4657 /* 4658 * Called prior to creating a new inode and linking it to a directory. The 4659 * jaddref structure must already be allocated by softdep_setup_inomapdep 4660 * and it is discovered here so we can initialize the mode and update 4661 * nlinkdelta. 4662 */ 4663 void 4664 softdep_setup_create(dp, ip) 4665 struct inode *dp; 4666 struct inode *ip; 4667 { 4668 struct inodedep *inodedep; 4669 struct jaddref *jaddref; 4670 struct vnode *dvp; 4671 4672 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4673 ("softdep_setup_create called on non-softdep filesystem")); 4674 KASSERT(ip->i_nlink == 1, 4675 ("softdep_setup_create: Invalid link count.")); 4676 dvp = ITOV(dp); 4677 ACQUIRE_LOCK(ITOUMP(dp)); 4678 inodedep = inodedep_lookup_ip(ip); 4679 if (DOINGSUJ(dvp)) { 4680 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4681 inoreflst); 4682 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 4683 ("softdep_setup_create: No addref structure present.")); 4684 } 4685 softdep_prelink(dvp, NULL); 4686 FREE_LOCK(ITOUMP(dp)); 4687 } 4688 4689 /* 4690 * Create a jaddref structure to track the addition of a DOTDOT link when 4691 * we are reparenting an inode as part of a rename. This jaddref will be 4692 * found by softdep_setup_directory_change. Adjusts nlinkdelta for 4693 * non-journaling softdep. 4694 */ 4695 void 4696 softdep_setup_dotdot_link(dp, ip) 4697 struct inode *dp; 4698 struct inode *ip; 4699 { 4700 struct inodedep *inodedep; 4701 struct jaddref *jaddref; 4702 struct vnode *dvp; 4703 4704 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4705 ("softdep_setup_dotdot_link called on non-softdep filesystem")); 4706 dvp = ITOV(dp); 4707 jaddref = NULL; 4708 /* 4709 * We don't set MKDIR_PARENT as this is not tied to a mkdir and 4710 * is used as a normal link would be. 4711 */ 4712 if (DOINGSUJ(dvp)) 4713 jaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET, 4714 dp->i_effnlink - 1, dp->i_mode); 4715 ACQUIRE_LOCK(ITOUMP(dp)); 4716 inodedep = inodedep_lookup_ip(dp); 4717 if (jaddref) 4718 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 4719 if_deps); 4720 softdep_prelink(dvp, ITOV(ip)); 4721 FREE_LOCK(ITOUMP(dp)); 4722 } 4723 4724 /* 4725 * Create a jaddref structure to track a new link to an inode. The directory 4726 * offset is not known until softdep_setup_directory_add or 4727 * softdep_setup_directory_change. Adjusts nlinkdelta for non-journaling 4728 * softdep. 4729 */ 4730 void 4731 softdep_setup_link(dp, ip) 4732 struct inode *dp; 4733 struct inode *ip; 4734 { 4735 struct inodedep *inodedep; 4736 struct jaddref *jaddref; 4737 struct vnode *dvp; 4738 4739 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4740 ("softdep_setup_link called on non-softdep filesystem")); 4741 dvp = ITOV(dp); 4742 jaddref = NULL; 4743 if (DOINGSUJ(dvp)) 4744 jaddref = newjaddref(dp, ip->i_number, 0, ip->i_effnlink - 1, 4745 ip->i_mode); 4746 ACQUIRE_LOCK(ITOUMP(dp)); 4747 inodedep = inodedep_lookup_ip(ip); 4748 if (jaddref) 4749 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 4750 if_deps); 4751 softdep_prelink(dvp, ITOV(ip)); 4752 FREE_LOCK(ITOUMP(dp)); 4753 } 4754 4755 /* 4756 * Called to create the jaddref structures to track . and .. references as 4757 * well as lookup and further initialize the incomplete jaddref created 4758 * by softdep_setup_inomapdep when the inode was allocated. Adjusts 4759 * nlinkdelta for non-journaling softdep. 4760 */ 4761 void 4762 softdep_setup_mkdir(dp, ip) 4763 struct inode *dp; 4764 struct inode *ip; 4765 { 4766 struct inodedep *inodedep; 4767 struct jaddref *dotdotaddref; 4768 struct jaddref *dotaddref; 4769 struct jaddref *jaddref; 4770 struct vnode *dvp; 4771 4772 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4773 ("softdep_setup_mkdir called on non-softdep filesystem")); 4774 dvp = ITOV(dp); 4775 dotaddref = dotdotaddref = NULL; 4776 if (DOINGSUJ(dvp)) { 4777 dotaddref = newjaddref(ip, ip->i_number, DOT_OFFSET, 1, 4778 ip->i_mode); 4779 dotaddref->ja_state |= MKDIR_BODY; 4780 dotdotaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET, 4781 dp->i_effnlink - 1, dp->i_mode); 4782 dotdotaddref->ja_state |= MKDIR_PARENT; 4783 } 4784 ACQUIRE_LOCK(ITOUMP(dp)); 4785 inodedep = inodedep_lookup_ip(ip); 4786 if (DOINGSUJ(dvp)) { 4787 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4788 inoreflst); 4789 KASSERT(jaddref != NULL, 4790 ("softdep_setup_mkdir: No addref structure present.")); 4791 KASSERT(jaddref->ja_parent == dp->i_number, 4792 ("softdep_setup_mkdir: bad parent %ju", 4793 (uintmax_t)jaddref->ja_parent)); 4794 TAILQ_INSERT_BEFORE(&jaddref->ja_ref, &dotaddref->ja_ref, 4795 if_deps); 4796 } 4797 inodedep = inodedep_lookup_ip(dp); 4798 if (DOINGSUJ(dvp)) 4799 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, 4800 &dotdotaddref->ja_ref, if_deps); 4801 softdep_prelink(ITOV(dp), NULL); 4802 FREE_LOCK(ITOUMP(dp)); 4803 } 4804 4805 /* 4806 * Called to track nlinkdelta of the inode and parent directories prior to 4807 * unlinking a directory. 4808 */ 4809 void 4810 softdep_setup_rmdir(dp, ip) 4811 struct inode *dp; 4812 struct inode *ip; 4813 { 4814 struct vnode *dvp; 4815 4816 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4817 ("softdep_setup_rmdir called on non-softdep filesystem")); 4818 dvp = ITOV(dp); 4819 ACQUIRE_LOCK(ITOUMP(dp)); 4820 (void) inodedep_lookup_ip(ip); 4821 (void) inodedep_lookup_ip(dp); 4822 softdep_prelink(dvp, ITOV(ip)); 4823 FREE_LOCK(ITOUMP(dp)); 4824 } 4825 4826 /* 4827 * Called to track nlinkdelta of the inode and parent directories prior to 4828 * unlink. 4829 */ 4830 void 4831 softdep_setup_unlink(dp, ip) 4832 struct inode *dp; 4833 struct inode *ip; 4834 { 4835 struct vnode *dvp; 4836 4837 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4838 ("softdep_setup_unlink called on non-softdep filesystem")); 4839 dvp = ITOV(dp); 4840 ACQUIRE_LOCK(ITOUMP(dp)); 4841 (void) inodedep_lookup_ip(ip); 4842 (void) inodedep_lookup_ip(dp); 4843 softdep_prelink(dvp, ITOV(ip)); 4844 FREE_LOCK(ITOUMP(dp)); 4845 } 4846 4847 /* 4848 * Called to release the journal structures created by a failed non-directory 4849 * creation. Adjusts nlinkdelta for non-journaling softdep. 4850 */ 4851 void 4852 softdep_revert_create(dp, ip) 4853 struct inode *dp; 4854 struct inode *ip; 4855 { 4856 struct inodedep *inodedep; 4857 struct jaddref *jaddref; 4858 struct vnode *dvp; 4859 4860 KASSERT(MOUNTEDSOFTDEP(ITOVFS((dp))) != 0, 4861 ("softdep_revert_create called on non-softdep filesystem")); 4862 dvp = ITOV(dp); 4863 ACQUIRE_LOCK(ITOUMP(dp)); 4864 inodedep = inodedep_lookup_ip(ip); 4865 if (DOINGSUJ(dvp)) { 4866 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4867 inoreflst); 4868 KASSERT(jaddref->ja_parent == dp->i_number, 4869 ("softdep_revert_create: addref parent mismatch")); 4870 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4871 } 4872 FREE_LOCK(ITOUMP(dp)); 4873 } 4874 4875 /* 4876 * Called to release the journal structures created by a failed link 4877 * addition. Adjusts nlinkdelta for non-journaling softdep. 4878 */ 4879 void 4880 softdep_revert_link(dp, ip) 4881 struct inode *dp; 4882 struct inode *ip; 4883 { 4884 struct inodedep *inodedep; 4885 struct jaddref *jaddref; 4886 struct vnode *dvp; 4887 4888 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4889 ("softdep_revert_link called on non-softdep filesystem")); 4890 dvp = ITOV(dp); 4891 ACQUIRE_LOCK(ITOUMP(dp)); 4892 inodedep = inodedep_lookup_ip(ip); 4893 if (DOINGSUJ(dvp)) { 4894 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4895 inoreflst); 4896 KASSERT(jaddref->ja_parent == dp->i_number, 4897 ("softdep_revert_link: addref parent mismatch")); 4898 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4899 } 4900 FREE_LOCK(ITOUMP(dp)); 4901 } 4902 4903 /* 4904 * Called to release the journal structures created by a failed mkdir 4905 * attempt. Adjusts nlinkdelta for non-journaling softdep. 4906 */ 4907 void 4908 softdep_revert_mkdir(dp, ip) 4909 struct inode *dp; 4910 struct inode *ip; 4911 { 4912 struct inodedep *inodedep; 4913 struct jaddref *jaddref; 4914 struct jaddref *dotaddref; 4915 struct vnode *dvp; 4916 4917 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4918 ("softdep_revert_mkdir called on non-softdep filesystem")); 4919 dvp = ITOV(dp); 4920 4921 ACQUIRE_LOCK(ITOUMP(dp)); 4922 inodedep = inodedep_lookup_ip(dp); 4923 if (DOINGSUJ(dvp)) { 4924 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4925 inoreflst); 4926 KASSERT(jaddref->ja_parent == ip->i_number, 4927 ("softdep_revert_mkdir: dotdot addref parent mismatch")); 4928 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4929 } 4930 inodedep = inodedep_lookup_ip(ip); 4931 if (DOINGSUJ(dvp)) { 4932 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 4933 inoreflst); 4934 KASSERT(jaddref->ja_parent == dp->i_number, 4935 ("softdep_revert_mkdir: addref parent mismatch")); 4936 dotaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref, 4937 inoreflst, if_deps); 4938 cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait); 4939 KASSERT(dotaddref->ja_parent == ip->i_number, 4940 ("softdep_revert_mkdir: dot addref parent mismatch")); 4941 cancel_jaddref(dotaddref, inodedep, &inodedep->id_inowait); 4942 } 4943 FREE_LOCK(ITOUMP(dp)); 4944 } 4945 4946 /* 4947 * Called to correct nlinkdelta after a failed rmdir. 4948 */ 4949 void 4950 softdep_revert_rmdir(dp, ip) 4951 struct inode *dp; 4952 struct inode *ip; 4953 { 4954 4955 KASSERT(MOUNTEDSOFTDEP(ITOVFS(dp)) != 0, 4956 ("softdep_revert_rmdir called on non-softdep filesystem")); 4957 ACQUIRE_LOCK(ITOUMP(dp)); 4958 (void) inodedep_lookup_ip(ip); 4959 (void) inodedep_lookup_ip(dp); 4960 FREE_LOCK(ITOUMP(dp)); 4961 } 4962 4963 /* 4964 * Protecting the freemaps (or bitmaps). 4965 * 4966 * To eliminate the need to execute fsck before mounting a filesystem 4967 * after a power failure, one must (conservatively) guarantee that the 4968 * on-disk copy of the bitmaps never indicate that a live inode or block is 4969 * free. So, when a block or inode is allocated, the bitmap should be 4970 * updated (on disk) before any new pointers. When a block or inode is 4971 * freed, the bitmap should not be updated until all pointers have been 4972 * reset. The latter dependency is handled by the delayed de-allocation 4973 * approach described below for block and inode de-allocation. The former 4974 * dependency is handled by calling the following procedure when a block or 4975 * inode is allocated. When an inode is allocated an "inodedep" is created 4976 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk. 4977 * Each "inodedep" is also inserted into the hash indexing structure so 4978 * that any additional link additions can be made dependent on the inode 4979 * allocation. 4980 * 4981 * The ufs filesystem maintains a number of free block counts (e.g., per 4982 * cylinder group, per cylinder and per <cylinder, rotational position> pair) 4983 * in addition to the bitmaps. These counts are used to improve efficiency 4984 * during allocation and therefore must be consistent with the bitmaps. 4985 * There is no convenient way to guarantee post-crash consistency of these 4986 * counts with simple update ordering, for two main reasons: (1) The counts 4987 * and bitmaps for a single cylinder group block are not in the same disk 4988 * sector. If a disk write is interrupted (e.g., by power failure), one may 4989 * be written and the other not. (2) Some of the counts are located in the 4990 * superblock rather than the cylinder group block. So, we focus our soft 4991 * updates implementation on protecting the bitmaps. When mounting a 4992 * filesystem, we recompute the auxiliary counts from the bitmaps. 4993 */ 4994 4995 /* 4996 * Called just after updating the cylinder group block to allocate an inode. 4997 */ 4998 void 4999 softdep_setup_inomapdep(bp, ip, newinum, mode) 5000 struct buf *bp; /* buffer for cylgroup block with inode map */ 5001 struct inode *ip; /* inode related to allocation */ 5002 ino_t newinum; /* new inode number being allocated */ 5003 int mode; 5004 { 5005 struct inodedep *inodedep; 5006 struct bmsafemap *bmsafemap; 5007 struct jaddref *jaddref; 5008 struct mount *mp; 5009 struct fs *fs; 5010 5011 mp = ITOVFS(ip); 5012 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5013 ("softdep_setup_inomapdep called on non-softdep filesystem")); 5014 fs = VFSTOUFS(mp)->um_fs; 5015 jaddref = NULL; 5016 5017 /* 5018 * Allocate the journal reference add structure so that the bitmap 5019 * can be dependent on it. 5020 */ 5021 if (MOUNTEDSUJ(mp)) { 5022 jaddref = newjaddref(ip, newinum, 0, 0, mode); 5023 jaddref->ja_state |= NEWBLOCK; 5024 } 5025 5026 /* 5027 * Create a dependency for the newly allocated inode. 5028 * Panic if it already exists as something is seriously wrong. 5029 * Otherwise add it to the dependency list for the buffer holding 5030 * the cylinder group map from which it was allocated. 5031 * 5032 * We have to preallocate a bmsafemap entry in case it is needed 5033 * in bmsafemap_lookup since once we allocate the inodedep, we 5034 * have to finish initializing it before we can FREE_LOCK(). 5035 * By preallocating, we avoid FREE_LOCK() while doing a malloc 5036 * in bmsafemap_lookup. We cannot call bmsafemap_lookup before 5037 * creating the inodedep as it can be freed during the time 5038 * that we FREE_LOCK() while allocating the inodedep. We must 5039 * call workitem_alloc() before entering the locked section as 5040 * it also acquires the lock and we must avoid trying doing so 5041 * recursively. 5042 */ 5043 bmsafemap = malloc(sizeof(struct bmsafemap), 5044 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 5045 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 5046 ACQUIRE_LOCK(ITOUMP(ip)); 5047 if ((inodedep_lookup(mp, newinum, DEPALLOC, &inodedep))) 5048 panic("softdep_setup_inomapdep: dependency %p for new" 5049 "inode already exists", inodedep); 5050 bmsafemap = bmsafemap_lookup(mp, bp, ino_to_cg(fs, newinum), bmsafemap); 5051 if (jaddref) { 5052 LIST_INSERT_HEAD(&bmsafemap->sm_jaddrefhd, jaddref, ja_bmdeps); 5053 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref, 5054 if_deps); 5055 } else { 5056 inodedep->id_state |= ONDEPLIST; 5057 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps); 5058 } 5059 inodedep->id_bmsafemap = bmsafemap; 5060 inodedep->id_state &= ~DEPCOMPLETE; 5061 FREE_LOCK(ITOUMP(ip)); 5062 } 5063 5064 /* 5065 * Called just after updating the cylinder group block to 5066 * allocate block or fragment. 5067 */ 5068 void 5069 softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags) 5070 struct buf *bp; /* buffer for cylgroup block with block map */ 5071 struct mount *mp; /* filesystem doing allocation */ 5072 ufs2_daddr_t newblkno; /* number of newly allocated block */ 5073 int frags; /* Number of fragments. */ 5074 int oldfrags; /* Previous number of fragments for extend. */ 5075 { 5076 struct newblk *newblk; 5077 struct bmsafemap *bmsafemap; 5078 struct jnewblk *jnewblk; 5079 struct ufsmount *ump; 5080 struct fs *fs; 5081 5082 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5083 ("softdep_setup_blkmapdep called on non-softdep filesystem")); 5084 ump = VFSTOUFS(mp); 5085 fs = ump->um_fs; 5086 jnewblk = NULL; 5087 /* 5088 * Create a dependency for the newly allocated block. 5089 * Add it to the dependency list for the buffer holding 5090 * the cylinder group map from which it was allocated. 5091 */ 5092 if (MOUNTEDSUJ(mp)) { 5093 jnewblk = malloc(sizeof(*jnewblk), M_JNEWBLK, M_SOFTDEP_FLAGS); 5094 workitem_alloc(&jnewblk->jn_list, D_JNEWBLK, mp); 5095 jnewblk->jn_jsegdep = newjsegdep(&jnewblk->jn_list); 5096 jnewblk->jn_state = ATTACHED; 5097 jnewblk->jn_blkno = newblkno; 5098 jnewblk->jn_frags = frags; 5099 jnewblk->jn_oldfrags = oldfrags; 5100 #ifdef SUJ_DEBUG 5101 { 5102 struct cg *cgp; 5103 uint8_t *blksfree; 5104 long bno; 5105 int i; 5106 5107 cgp = (struct cg *)bp->b_data; 5108 blksfree = cg_blksfree(cgp); 5109 bno = dtogd(fs, jnewblk->jn_blkno); 5110 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; 5111 i++) { 5112 if (isset(blksfree, bno + i)) 5113 panic("softdep_setup_blkmapdep: " 5114 "free fragment %d from %d-%d " 5115 "state 0x%X dep %p", i, 5116 jnewblk->jn_oldfrags, 5117 jnewblk->jn_frags, 5118 jnewblk->jn_state, 5119 jnewblk->jn_dep); 5120 } 5121 } 5122 #endif 5123 } 5124 5125 CTR3(KTR_SUJ, 5126 "softdep_setup_blkmapdep: blkno %jd frags %d oldfrags %d", 5127 newblkno, frags, oldfrags); 5128 ACQUIRE_LOCK(ump); 5129 if (newblk_lookup(mp, newblkno, DEPALLOC, &newblk) != 0) 5130 panic("softdep_setup_blkmapdep: found block"); 5131 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp, 5132 dtog(fs, newblkno), NULL); 5133 if (jnewblk) { 5134 jnewblk->jn_dep = (struct worklist *)newblk; 5135 LIST_INSERT_HEAD(&bmsafemap->sm_jnewblkhd, jnewblk, jn_deps); 5136 } else { 5137 newblk->nb_state |= ONDEPLIST; 5138 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps); 5139 } 5140 newblk->nb_bmsafemap = bmsafemap; 5141 newblk->nb_jnewblk = jnewblk; 5142 FREE_LOCK(ump); 5143 } 5144 5145 #define BMSAFEMAP_HASH(ump, cg) \ 5146 (&(ump)->bmsafemap_hashtbl[(cg) & (ump)->bmsafemap_hash_size]) 5147 5148 static int 5149 bmsafemap_find(bmsafemaphd, cg, bmsafemapp) 5150 struct bmsafemap_hashhead *bmsafemaphd; 5151 int cg; 5152 struct bmsafemap **bmsafemapp; 5153 { 5154 struct bmsafemap *bmsafemap; 5155 5156 LIST_FOREACH(bmsafemap, bmsafemaphd, sm_hash) 5157 if (bmsafemap->sm_cg == cg) 5158 break; 5159 if (bmsafemap) { 5160 *bmsafemapp = bmsafemap; 5161 return (1); 5162 } 5163 *bmsafemapp = NULL; 5164 5165 return (0); 5166 } 5167 5168 /* 5169 * Find the bmsafemap associated with a cylinder group buffer. 5170 * If none exists, create one. The buffer must be locked when 5171 * this routine is called and this routine must be called with 5172 * the softdep lock held. To avoid giving up the lock while 5173 * allocating a new bmsafemap, a preallocated bmsafemap may be 5174 * provided. If it is provided but not needed, it is freed. 5175 */ 5176 static struct bmsafemap * 5177 bmsafemap_lookup(mp, bp, cg, newbmsafemap) 5178 struct mount *mp; 5179 struct buf *bp; 5180 int cg; 5181 struct bmsafemap *newbmsafemap; 5182 { 5183 struct bmsafemap_hashhead *bmsafemaphd; 5184 struct bmsafemap *bmsafemap, *collision; 5185 struct worklist *wk; 5186 struct ufsmount *ump; 5187 5188 ump = VFSTOUFS(mp); 5189 LOCK_OWNED(ump); 5190 KASSERT(bp != NULL, ("bmsafemap_lookup: missing buffer")); 5191 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5192 if (wk->wk_type == D_BMSAFEMAP) { 5193 if (newbmsafemap) 5194 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP); 5195 return (WK_BMSAFEMAP(wk)); 5196 } 5197 } 5198 bmsafemaphd = BMSAFEMAP_HASH(ump, cg); 5199 if (bmsafemap_find(bmsafemaphd, cg, &bmsafemap) == 1) { 5200 if (newbmsafemap) 5201 WORKITEM_FREE(newbmsafemap, D_BMSAFEMAP); 5202 return (bmsafemap); 5203 } 5204 if (newbmsafemap) { 5205 bmsafemap = newbmsafemap; 5206 } else { 5207 FREE_LOCK(ump); 5208 bmsafemap = malloc(sizeof(struct bmsafemap), 5209 M_BMSAFEMAP, M_SOFTDEP_FLAGS); 5210 workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp); 5211 ACQUIRE_LOCK(ump); 5212 } 5213 bmsafemap->sm_buf = bp; 5214 LIST_INIT(&bmsafemap->sm_inodedephd); 5215 LIST_INIT(&bmsafemap->sm_inodedepwr); 5216 LIST_INIT(&bmsafemap->sm_newblkhd); 5217 LIST_INIT(&bmsafemap->sm_newblkwr); 5218 LIST_INIT(&bmsafemap->sm_jaddrefhd); 5219 LIST_INIT(&bmsafemap->sm_jnewblkhd); 5220 LIST_INIT(&bmsafemap->sm_freehd); 5221 LIST_INIT(&bmsafemap->sm_freewr); 5222 if (bmsafemap_find(bmsafemaphd, cg, &collision) == 1) { 5223 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 5224 return (collision); 5225 } 5226 bmsafemap->sm_cg = cg; 5227 LIST_INSERT_HEAD(bmsafemaphd, bmsafemap, sm_hash); 5228 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next); 5229 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list); 5230 return (bmsafemap); 5231 } 5232 5233 /* 5234 * Direct block allocation dependencies. 5235 * 5236 * When a new block is allocated, the corresponding disk locations must be 5237 * initialized (with zeros or new data) before the on-disk inode points to 5238 * them. Also, the freemap from which the block was allocated must be 5239 * updated (on disk) before the inode's pointer. These two dependencies are 5240 * independent of each other and are needed for all file blocks and indirect 5241 * blocks that are pointed to directly by the inode. Just before the 5242 * "in-core" version of the inode is updated with a newly allocated block 5243 * number, a procedure (below) is called to setup allocation dependency 5244 * structures. These structures are removed when the corresponding 5245 * dependencies are satisfied or when the block allocation becomes obsolete 5246 * (i.e., the file is deleted, the block is de-allocated, or the block is a 5247 * fragment that gets upgraded). All of these cases are handled in 5248 * procedures described later. 5249 * 5250 * When a file extension causes a fragment to be upgraded, either to a larger 5251 * fragment or to a full block, the on-disk location may change (if the 5252 * previous fragment could not simply be extended). In this case, the old 5253 * fragment must be de-allocated, but not until after the inode's pointer has 5254 * been updated. In most cases, this is handled by later procedures, which 5255 * will construct a "freefrag" structure to be added to the workitem queue 5256 * when the inode update is complete (or obsolete). The main exception to 5257 * this is when an allocation occurs while a pending allocation dependency 5258 * (for the same block pointer) remains. This case is handled in the main 5259 * allocation dependency setup procedure by immediately freeing the 5260 * unreferenced fragments. 5261 */ 5262 void 5263 softdep_setup_allocdirect(ip, off, newblkno, oldblkno, newsize, oldsize, bp) 5264 struct inode *ip; /* inode to which block is being added */ 5265 ufs_lbn_t off; /* block pointer within inode */ 5266 ufs2_daddr_t newblkno; /* disk block number being added */ 5267 ufs2_daddr_t oldblkno; /* previous block number, 0 unless frag */ 5268 long newsize; /* size of new block */ 5269 long oldsize; /* size of new block */ 5270 struct buf *bp; /* bp for allocated block */ 5271 { 5272 struct allocdirect *adp, *oldadp; 5273 struct allocdirectlst *adphead; 5274 struct freefrag *freefrag; 5275 struct inodedep *inodedep; 5276 struct pagedep *pagedep; 5277 struct jnewblk *jnewblk; 5278 struct newblk *newblk; 5279 struct mount *mp; 5280 ufs_lbn_t lbn; 5281 5282 lbn = bp->b_lblkno; 5283 mp = ITOVFS(ip); 5284 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5285 ("softdep_setup_allocdirect called on non-softdep filesystem")); 5286 if (oldblkno && oldblkno != newblkno) 5287 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn); 5288 else 5289 freefrag = NULL; 5290 5291 CTR6(KTR_SUJ, 5292 "softdep_setup_allocdirect: ino %d blkno %jd oldblkno %jd " 5293 "off %jd newsize %ld oldsize %d", 5294 ip->i_number, newblkno, oldblkno, off, newsize, oldsize); 5295 ACQUIRE_LOCK(ITOUMP(ip)); 5296 if (off >= NDADDR) { 5297 if (lbn > 0) 5298 panic("softdep_setup_allocdirect: bad lbn %jd, off %jd", 5299 lbn, off); 5300 /* allocating an indirect block */ 5301 if (oldblkno != 0) 5302 panic("softdep_setup_allocdirect: non-zero indir"); 5303 } else { 5304 if (off != lbn) 5305 panic("softdep_setup_allocdirect: lbn %jd != off %jd", 5306 lbn, off); 5307 /* 5308 * Allocating a direct block. 5309 * 5310 * If we are allocating a directory block, then we must 5311 * allocate an associated pagedep to track additions and 5312 * deletions. 5313 */ 5314 if ((ip->i_mode & IFMT) == IFDIR) 5315 pagedep_lookup(mp, bp, ip->i_number, off, DEPALLOC, 5316 &pagedep); 5317 } 5318 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0) 5319 panic("softdep_setup_allocdirect: lost block"); 5320 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5321 ("softdep_setup_allocdirect: newblk already initialized")); 5322 /* 5323 * Convert the newblk to an allocdirect. 5324 */ 5325 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT); 5326 adp = (struct allocdirect *)newblk; 5327 newblk->nb_freefrag = freefrag; 5328 adp->ad_offset = off; 5329 adp->ad_oldblkno = oldblkno; 5330 adp->ad_newsize = newsize; 5331 adp->ad_oldsize = oldsize; 5332 5333 /* 5334 * Finish initializing the journal. 5335 */ 5336 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5337 jnewblk->jn_ino = ip->i_number; 5338 jnewblk->jn_lbn = lbn; 5339 add_to_journal(&jnewblk->jn_list); 5340 } 5341 if (freefrag && freefrag->ff_jdep != NULL && 5342 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5343 add_to_journal(freefrag->ff_jdep); 5344 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 5345 adp->ad_inodedep = inodedep; 5346 5347 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list); 5348 /* 5349 * The list of allocdirects must be kept in sorted and ascending 5350 * order so that the rollback routines can quickly determine the 5351 * first uncommitted block (the size of the file stored on disk 5352 * ends at the end of the lowest committed fragment, or if there 5353 * are no fragments, at the end of the highest committed block). 5354 * Since files generally grow, the typical case is that the new 5355 * block is to be added at the end of the list. We speed this 5356 * special case by checking against the last allocdirect in the 5357 * list before laboriously traversing the list looking for the 5358 * insertion point. 5359 */ 5360 adphead = &inodedep->id_newinoupdt; 5361 oldadp = TAILQ_LAST(adphead, allocdirectlst); 5362 if (oldadp == NULL || oldadp->ad_offset <= off) { 5363 /* insert at end of list */ 5364 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 5365 if (oldadp != NULL && oldadp->ad_offset == off) 5366 allocdirect_merge(adphead, adp, oldadp); 5367 FREE_LOCK(ITOUMP(ip)); 5368 return; 5369 } 5370 TAILQ_FOREACH(oldadp, adphead, ad_next) { 5371 if (oldadp->ad_offset >= off) 5372 break; 5373 } 5374 if (oldadp == NULL) 5375 panic("softdep_setup_allocdirect: lost entry"); 5376 /* insert in middle of list */ 5377 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 5378 if (oldadp->ad_offset == off) 5379 allocdirect_merge(adphead, adp, oldadp); 5380 5381 FREE_LOCK(ITOUMP(ip)); 5382 } 5383 5384 /* 5385 * Merge a newer and older journal record to be stored either in a 5386 * newblock or freefrag. This handles aggregating journal records for 5387 * fragment allocation into a second record as well as replacing a 5388 * journal free with an aborted journal allocation. A segment for the 5389 * oldest record will be placed on wkhd if it has been written. If not 5390 * the segment for the newer record will suffice. 5391 */ 5392 static struct worklist * 5393 jnewblk_merge(new, old, wkhd) 5394 struct worklist *new; 5395 struct worklist *old; 5396 struct workhead *wkhd; 5397 { 5398 struct jnewblk *njnewblk; 5399 struct jnewblk *jnewblk; 5400 5401 /* Handle NULLs to simplify callers. */ 5402 if (new == NULL) 5403 return (old); 5404 if (old == NULL) 5405 return (new); 5406 /* Replace a jfreefrag with a jnewblk. */ 5407 if (new->wk_type == D_JFREEFRAG) { 5408 if (WK_JNEWBLK(old)->jn_blkno != WK_JFREEFRAG(new)->fr_blkno) 5409 panic("jnewblk_merge: blkno mismatch: %p, %p", 5410 old, new); 5411 cancel_jfreefrag(WK_JFREEFRAG(new)); 5412 return (old); 5413 } 5414 if (old->wk_type != D_JNEWBLK || new->wk_type != D_JNEWBLK) 5415 panic("jnewblk_merge: Bad type: old %d new %d\n", 5416 old->wk_type, new->wk_type); 5417 /* 5418 * Handle merging of two jnewblk records that describe 5419 * different sets of fragments in the same block. 5420 */ 5421 jnewblk = WK_JNEWBLK(old); 5422 njnewblk = WK_JNEWBLK(new); 5423 if (jnewblk->jn_blkno != njnewblk->jn_blkno) 5424 panic("jnewblk_merge: Merging disparate blocks."); 5425 /* 5426 * The record may be rolled back in the cg. 5427 */ 5428 if (jnewblk->jn_state & UNDONE) { 5429 jnewblk->jn_state &= ~UNDONE; 5430 njnewblk->jn_state |= UNDONE; 5431 njnewblk->jn_state &= ~ATTACHED; 5432 } 5433 /* 5434 * We modify the newer addref and free the older so that if neither 5435 * has been written the most up-to-date copy will be on disk. If 5436 * both have been written but rolled back we only temporarily need 5437 * one of them to fix the bits when the cg write completes. 5438 */ 5439 jnewblk->jn_state |= ATTACHED | COMPLETE; 5440 njnewblk->jn_oldfrags = jnewblk->jn_oldfrags; 5441 cancel_jnewblk(jnewblk, wkhd); 5442 WORKLIST_REMOVE(&jnewblk->jn_list); 5443 free_jnewblk(jnewblk); 5444 return (new); 5445 } 5446 5447 /* 5448 * Replace an old allocdirect dependency with a newer one. 5449 * This routine must be called with splbio interrupts blocked. 5450 */ 5451 static void 5452 allocdirect_merge(adphead, newadp, oldadp) 5453 struct allocdirectlst *adphead; /* head of list holding allocdirects */ 5454 struct allocdirect *newadp; /* allocdirect being added */ 5455 struct allocdirect *oldadp; /* existing allocdirect being checked */ 5456 { 5457 struct worklist *wk; 5458 struct freefrag *freefrag; 5459 5460 freefrag = NULL; 5461 LOCK_OWNED(VFSTOUFS(newadp->ad_list.wk_mp)); 5462 if (newadp->ad_oldblkno != oldadp->ad_newblkno || 5463 newadp->ad_oldsize != oldadp->ad_newsize || 5464 newadp->ad_offset >= NDADDR) 5465 panic("%s %jd != new %jd || old size %ld != new %ld", 5466 "allocdirect_merge: old blkno", 5467 (intmax_t)newadp->ad_oldblkno, 5468 (intmax_t)oldadp->ad_newblkno, 5469 newadp->ad_oldsize, oldadp->ad_newsize); 5470 newadp->ad_oldblkno = oldadp->ad_oldblkno; 5471 newadp->ad_oldsize = oldadp->ad_oldsize; 5472 /* 5473 * If the old dependency had a fragment to free or had never 5474 * previously had a block allocated, then the new dependency 5475 * can immediately post its freefrag and adopt the old freefrag. 5476 * This action is done by swapping the freefrag dependencies. 5477 * The new dependency gains the old one's freefrag, and the 5478 * old one gets the new one and then immediately puts it on 5479 * the worklist when it is freed by free_newblk. It is 5480 * not possible to do this swap when the old dependency had a 5481 * non-zero size but no previous fragment to free. This condition 5482 * arises when the new block is an extension of the old block. 5483 * Here, the first part of the fragment allocated to the new 5484 * dependency is part of the block currently claimed on disk by 5485 * the old dependency, so cannot legitimately be freed until the 5486 * conditions for the new dependency are fulfilled. 5487 */ 5488 freefrag = newadp->ad_freefrag; 5489 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) { 5490 newadp->ad_freefrag = oldadp->ad_freefrag; 5491 oldadp->ad_freefrag = freefrag; 5492 } 5493 /* 5494 * If we are tracking a new directory-block allocation, 5495 * move it from the old allocdirect to the new allocdirect. 5496 */ 5497 if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) { 5498 WORKLIST_REMOVE(wk); 5499 if (!LIST_EMPTY(&oldadp->ad_newdirblk)) 5500 panic("allocdirect_merge: extra newdirblk"); 5501 WORKLIST_INSERT(&newadp->ad_newdirblk, wk); 5502 } 5503 TAILQ_REMOVE(adphead, oldadp, ad_next); 5504 /* 5505 * We need to move any journal dependencies over to the freefrag 5506 * that releases this block if it exists. Otherwise we are 5507 * extending an existing block and we'll wait until that is 5508 * complete to release the journal space and extend the 5509 * new journal to cover this old space as well. 5510 */ 5511 if (freefrag == NULL) { 5512 if (oldadp->ad_newblkno != newadp->ad_newblkno) 5513 panic("allocdirect_merge: %jd != %jd", 5514 oldadp->ad_newblkno, newadp->ad_newblkno); 5515 newadp->ad_block.nb_jnewblk = (struct jnewblk *) 5516 jnewblk_merge(&newadp->ad_block.nb_jnewblk->jn_list, 5517 &oldadp->ad_block.nb_jnewblk->jn_list, 5518 &newadp->ad_block.nb_jwork); 5519 oldadp->ad_block.nb_jnewblk = NULL; 5520 cancel_newblk(&oldadp->ad_block, NULL, 5521 &newadp->ad_block.nb_jwork); 5522 } else { 5523 wk = (struct worklist *) cancel_newblk(&oldadp->ad_block, 5524 &freefrag->ff_list, &freefrag->ff_jwork); 5525 freefrag->ff_jdep = jnewblk_merge(freefrag->ff_jdep, wk, 5526 &freefrag->ff_jwork); 5527 } 5528 free_newblk(&oldadp->ad_block); 5529 } 5530 5531 /* 5532 * Allocate a jfreefrag structure to journal a single block free. 5533 */ 5534 static struct jfreefrag * 5535 newjfreefrag(freefrag, ip, blkno, size, lbn) 5536 struct freefrag *freefrag; 5537 struct inode *ip; 5538 ufs2_daddr_t blkno; 5539 long size; 5540 ufs_lbn_t lbn; 5541 { 5542 struct jfreefrag *jfreefrag; 5543 struct fs *fs; 5544 5545 fs = ITOFS(ip); 5546 jfreefrag = malloc(sizeof(struct jfreefrag), M_JFREEFRAG, 5547 M_SOFTDEP_FLAGS); 5548 workitem_alloc(&jfreefrag->fr_list, D_JFREEFRAG, ITOVFS(ip)); 5549 jfreefrag->fr_jsegdep = newjsegdep(&jfreefrag->fr_list); 5550 jfreefrag->fr_state = ATTACHED | DEPCOMPLETE; 5551 jfreefrag->fr_ino = ip->i_number; 5552 jfreefrag->fr_lbn = lbn; 5553 jfreefrag->fr_blkno = blkno; 5554 jfreefrag->fr_frags = numfrags(fs, size); 5555 jfreefrag->fr_freefrag = freefrag; 5556 5557 return (jfreefrag); 5558 } 5559 5560 /* 5561 * Allocate a new freefrag structure. 5562 */ 5563 static struct freefrag * 5564 newfreefrag(ip, blkno, size, lbn) 5565 struct inode *ip; 5566 ufs2_daddr_t blkno; 5567 long size; 5568 ufs_lbn_t lbn; 5569 { 5570 struct freefrag *freefrag; 5571 struct ufsmount *ump; 5572 struct fs *fs; 5573 5574 CTR4(KTR_SUJ, "newfreefrag: ino %d blkno %jd size %ld lbn %jd", 5575 ip->i_number, blkno, size, lbn); 5576 ump = ITOUMP(ip); 5577 fs = ump->um_fs; 5578 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag) 5579 panic("newfreefrag: frag size"); 5580 freefrag = malloc(sizeof(struct freefrag), 5581 M_FREEFRAG, M_SOFTDEP_FLAGS); 5582 workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ump)); 5583 freefrag->ff_state = ATTACHED; 5584 LIST_INIT(&freefrag->ff_jwork); 5585 freefrag->ff_inum = ip->i_number; 5586 freefrag->ff_vtype = ITOV(ip)->v_type; 5587 freefrag->ff_blkno = blkno; 5588 freefrag->ff_fragsize = size; 5589 5590 if (MOUNTEDSUJ(UFSTOVFS(ump))) { 5591 freefrag->ff_jdep = (struct worklist *) 5592 newjfreefrag(freefrag, ip, blkno, size, lbn); 5593 } else { 5594 freefrag->ff_state |= DEPCOMPLETE; 5595 freefrag->ff_jdep = NULL; 5596 } 5597 5598 return (freefrag); 5599 } 5600 5601 /* 5602 * This workitem de-allocates fragments that were replaced during 5603 * file block allocation. 5604 */ 5605 static void 5606 handle_workitem_freefrag(freefrag) 5607 struct freefrag *freefrag; 5608 { 5609 struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp); 5610 struct workhead wkhd; 5611 5612 CTR3(KTR_SUJ, 5613 "handle_workitem_freefrag: ino %d blkno %jd size %ld", 5614 freefrag->ff_inum, freefrag->ff_blkno, freefrag->ff_fragsize); 5615 /* 5616 * It would be illegal to add new completion items to the 5617 * freefrag after it was schedule to be done so it must be 5618 * safe to modify the list head here. 5619 */ 5620 LIST_INIT(&wkhd); 5621 ACQUIRE_LOCK(ump); 5622 LIST_SWAP(&freefrag->ff_jwork, &wkhd, worklist, wk_list); 5623 /* 5624 * If the journal has not been written we must cancel it here. 5625 */ 5626 if (freefrag->ff_jdep) { 5627 if (freefrag->ff_jdep->wk_type != D_JNEWBLK) 5628 panic("handle_workitem_freefrag: Unexpected type %d\n", 5629 freefrag->ff_jdep->wk_type); 5630 cancel_jnewblk(WK_JNEWBLK(freefrag->ff_jdep), &wkhd); 5631 } 5632 FREE_LOCK(ump); 5633 ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno, 5634 freefrag->ff_fragsize, freefrag->ff_inum, freefrag->ff_vtype, &wkhd); 5635 ACQUIRE_LOCK(ump); 5636 WORKITEM_FREE(freefrag, D_FREEFRAG); 5637 FREE_LOCK(ump); 5638 } 5639 5640 /* 5641 * Set up a dependency structure for an external attributes data block. 5642 * This routine follows much of the structure of softdep_setup_allocdirect. 5643 * See the description of softdep_setup_allocdirect above for details. 5644 */ 5645 void 5646 softdep_setup_allocext(ip, off, newblkno, oldblkno, newsize, oldsize, bp) 5647 struct inode *ip; 5648 ufs_lbn_t off; 5649 ufs2_daddr_t newblkno; 5650 ufs2_daddr_t oldblkno; 5651 long newsize; 5652 long oldsize; 5653 struct buf *bp; 5654 { 5655 struct allocdirect *adp, *oldadp; 5656 struct allocdirectlst *adphead; 5657 struct freefrag *freefrag; 5658 struct inodedep *inodedep; 5659 struct jnewblk *jnewblk; 5660 struct newblk *newblk; 5661 struct mount *mp; 5662 struct ufsmount *ump; 5663 ufs_lbn_t lbn; 5664 5665 mp = ITOVFS(ip); 5666 ump = VFSTOUFS(mp); 5667 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5668 ("softdep_setup_allocext called on non-softdep filesystem")); 5669 KASSERT(off < NXADDR, ("softdep_setup_allocext: lbn %lld > NXADDR", 5670 (long long)off)); 5671 5672 lbn = bp->b_lblkno; 5673 if (oldblkno && oldblkno != newblkno) 5674 freefrag = newfreefrag(ip, oldblkno, oldsize, lbn); 5675 else 5676 freefrag = NULL; 5677 5678 ACQUIRE_LOCK(ump); 5679 if (newblk_lookup(mp, newblkno, 0, &newblk) == 0) 5680 panic("softdep_setup_allocext: lost block"); 5681 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5682 ("softdep_setup_allocext: newblk already initialized")); 5683 /* 5684 * Convert the newblk to an allocdirect. 5685 */ 5686 WORKITEM_REASSIGN(newblk, D_ALLOCDIRECT); 5687 adp = (struct allocdirect *)newblk; 5688 newblk->nb_freefrag = freefrag; 5689 adp->ad_offset = off; 5690 adp->ad_oldblkno = oldblkno; 5691 adp->ad_newsize = newsize; 5692 adp->ad_oldsize = oldsize; 5693 adp->ad_state |= EXTDATA; 5694 5695 /* 5696 * Finish initializing the journal. 5697 */ 5698 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5699 jnewblk->jn_ino = ip->i_number; 5700 jnewblk->jn_lbn = lbn; 5701 add_to_journal(&jnewblk->jn_list); 5702 } 5703 if (freefrag && freefrag->ff_jdep != NULL && 5704 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5705 add_to_journal(freefrag->ff_jdep); 5706 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 5707 adp->ad_inodedep = inodedep; 5708 5709 WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list); 5710 /* 5711 * The list of allocdirects must be kept in sorted and ascending 5712 * order so that the rollback routines can quickly determine the 5713 * first uncommitted block (the size of the file stored on disk 5714 * ends at the end of the lowest committed fragment, or if there 5715 * are no fragments, at the end of the highest committed block). 5716 * Since files generally grow, the typical case is that the new 5717 * block is to be added at the end of the list. We speed this 5718 * special case by checking against the last allocdirect in the 5719 * list before laboriously traversing the list looking for the 5720 * insertion point. 5721 */ 5722 adphead = &inodedep->id_newextupdt; 5723 oldadp = TAILQ_LAST(adphead, allocdirectlst); 5724 if (oldadp == NULL || oldadp->ad_offset <= off) { 5725 /* insert at end of list */ 5726 TAILQ_INSERT_TAIL(adphead, adp, ad_next); 5727 if (oldadp != NULL && oldadp->ad_offset == off) 5728 allocdirect_merge(adphead, adp, oldadp); 5729 FREE_LOCK(ump); 5730 return; 5731 } 5732 TAILQ_FOREACH(oldadp, adphead, ad_next) { 5733 if (oldadp->ad_offset >= off) 5734 break; 5735 } 5736 if (oldadp == NULL) 5737 panic("softdep_setup_allocext: lost entry"); 5738 /* insert in middle of list */ 5739 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next); 5740 if (oldadp->ad_offset == off) 5741 allocdirect_merge(adphead, adp, oldadp); 5742 FREE_LOCK(ump); 5743 } 5744 5745 /* 5746 * Indirect block allocation dependencies. 5747 * 5748 * The same dependencies that exist for a direct block also exist when 5749 * a new block is allocated and pointed to by an entry in a block of 5750 * indirect pointers. The undo/redo states described above are also 5751 * used here. Because an indirect block contains many pointers that 5752 * may have dependencies, a second copy of the entire in-memory indirect 5753 * block is kept. The buffer cache copy is always completely up-to-date. 5754 * The second copy, which is used only as a source for disk writes, 5755 * contains only the safe pointers (i.e., those that have no remaining 5756 * update dependencies). The second copy is freed when all pointers 5757 * are safe. The cache is not allowed to replace indirect blocks with 5758 * pending update dependencies. If a buffer containing an indirect 5759 * block with dependencies is written, these routines will mark it 5760 * dirty again. It can only be successfully written once all the 5761 * dependencies are removed. The ffs_fsync routine in conjunction with 5762 * softdep_sync_metadata work together to get all the dependencies 5763 * removed so that a file can be successfully written to disk. Three 5764 * procedures are used when setting up indirect block pointer 5765 * dependencies. The division is necessary because of the organization 5766 * of the "balloc" routine and because of the distinction between file 5767 * pages and file metadata blocks. 5768 */ 5769 5770 /* 5771 * Allocate a new allocindir structure. 5772 */ 5773 static struct allocindir * 5774 newallocindir(ip, ptrno, newblkno, oldblkno, lbn) 5775 struct inode *ip; /* inode for file being extended */ 5776 int ptrno; /* offset of pointer in indirect block */ 5777 ufs2_daddr_t newblkno; /* disk block number being added */ 5778 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 5779 ufs_lbn_t lbn; 5780 { 5781 struct newblk *newblk; 5782 struct allocindir *aip; 5783 struct freefrag *freefrag; 5784 struct jnewblk *jnewblk; 5785 5786 if (oldblkno) 5787 freefrag = newfreefrag(ip, oldblkno, ITOFS(ip)->fs_bsize, lbn); 5788 else 5789 freefrag = NULL; 5790 ACQUIRE_LOCK(ITOUMP(ip)); 5791 if (newblk_lookup(ITOVFS(ip), newblkno, 0, &newblk) == 0) 5792 panic("new_allocindir: lost block"); 5793 KASSERT(newblk->nb_list.wk_type == D_NEWBLK, 5794 ("newallocindir: newblk already initialized")); 5795 WORKITEM_REASSIGN(newblk, D_ALLOCINDIR); 5796 newblk->nb_freefrag = freefrag; 5797 aip = (struct allocindir *)newblk; 5798 aip->ai_offset = ptrno; 5799 aip->ai_oldblkno = oldblkno; 5800 aip->ai_lbn = lbn; 5801 if ((jnewblk = newblk->nb_jnewblk) != NULL) { 5802 jnewblk->jn_ino = ip->i_number; 5803 jnewblk->jn_lbn = lbn; 5804 add_to_journal(&jnewblk->jn_list); 5805 } 5806 if (freefrag && freefrag->ff_jdep != NULL && 5807 freefrag->ff_jdep->wk_type == D_JFREEFRAG) 5808 add_to_journal(freefrag->ff_jdep); 5809 return (aip); 5810 } 5811 5812 /* 5813 * Called just before setting an indirect block pointer 5814 * to a newly allocated file page. 5815 */ 5816 void 5817 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp) 5818 struct inode *ip; /* inode for file being extended */ 5819 ufs_lbn_t lbn; /* allocated block number within file */ 5820 struct buf *bp; /* buffer with indirect blk referencing page */ 5821 int ptrno; /* offset of pointer in indirect block */ 5822 ufs2_daddr_t newblkno; /* disk block number being added */ 5823 ufs2_daddr_t oldblkno; /* previous block number, 0 if none */ 5824 struct buf *nbp; /* buffer holding allocated page */ 5825 { 5826 struct inodedep *inodedep; 5827 struct freefrag *freefrag; 5828 struct allocindir *aip; 5829 struct pagedep *pagedep; 5830 struct mount *mp; 5831 struct ufsmount *ump; 5832 5833 mp = ITOVFS(ip); 5834 ump = VFSTOUFS(mp); 5835 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 5836 ("softdep_setup_allocindir_page called on non-softdep filesystem")); 5837 KASSERT(lbn == nbp->b_lblkno, 5838 ("softdep_setup_allocindir_page: lbn %jd != lblkno %jd", 5839 lbn, bp->b_lblkno)); 5840 CTR4(KTR_SUJ, 5841 "softdep_setup_allocindir_page: ino %d blkno %jd oldblkno %jd " 5842 "lbn %jd", ip->i_number, newblkno, oldblkno, lbn); 5843 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page"); 5844 aip = newallocindir(ip, ptrno, newblkno, oldblkno, lbn); 5845 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 5846 /* 5847 * If we are allocating a directory page, then we must 5848 * allocate an associated pagedep to track additions and 5849 * deletions. 5850 */ 5851 if ((ip->i_mode & IFMT) == IFDIR) 5852 pagedep_lookup(mp, nbp, ip->i_number, lbn, DEPALLOC, &pagedep); 5853 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list); 5854 freefrag = setup_allocindir_phase2(bp, ip, inodedep, aip, lbn); 5855 FREE_LOCK(ump); 5856 if (freefrag) 5857 handle_workitem_freefrag(freefrag); 5858 } 5859 5860 /* 5861 * Called just before setting an indirect block pointer to a 5862 * newly allocated indirect block. 5863 */ 5864 void 5865 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno) 5866 struct buf *nbp; /* newly allocated indirect block */ 5867 struct inode *ip; /* inode for file being extended */ 5868 struct buf *bp; /* indirect block referencing allocated block */ 5869 int ptrno; /* offset of pointer in indirect block */ 5870 ufs2_daddr_t newblkno; /* disk block number being added */ 5871 { 5872 struct inodedep *inodedep; 5873 struct allocindir *aip; 5874 struct ufsmount *ump; 5875 ufs_lbn_t lbn; 5876 5877 ump = ITOUMP(ip); 5878 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 5879 ("softdep_setup_allocindir_meta called on non-softdep filesystem")); 5880 CTR3(KTR_SUJ, 5881 "softdep_setup_allocindir_meta: ino %d blkno %jd ptrno %d", 5882 ip->i_number, newblkno, ptrno); 5883 lbn = nbp->b_lblkno; 5884 ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta"); 5885 aip = newallocindir(ip, ptrno, newblkno, 0, lbn); 5886 inodedep_lookup(UFSTOVFS(ump), ip->i_number, DEPALLOC, &inodedep); 5887 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list); 5888 if (setup_allocindir_phase2(bp, ip, inodedep, aip, lbn)) 5889 panic("softdep_setup_allocindir_meta: Block already existed"); 5890 FREE_LOCK(ump); 5891 } 5892 5893 static void 5894 indirdep_complete(indirdep) 5895 struct indirdep *indirdep; 5896 { 5897 struct allocindir *aip; 5898 5899 LIST_REMOVE(indirdep, ir_next); 5900 indirdep->ir_state |= DEPCOMPLETE; 5901 5902 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) { 5903 LIST_REMOVE(aip, ai_next); 5904 free_newblk(&aip->ai_block); 5905 } 5906 /* 5907 * If this indirdep is not attached to a buf it was simply waiting 5908 * on completion to clear completehd. free_indirdep() asserts 5909 * that nothing is dangling. 5910 */ 5911 if ((indirdep->ir_state & ONWORKLIST) == 0) 5912 free_indirdep(indirdep); 5913 } 5914 5915 static struct indirdep * 5916 indirdep_lookup(mp, ip, bp) 5917 struct mount *mp; 5918 struct inode *ip; 5919 struct buf *bp; 5920 { 5921 struct indirdep *indirdep, *newindirdep; 5922 struct newblk *newblk; 5923 struct ufsmount *ump; 5924 struct worklist *wk; 5925 struct fs *fs; 5926 ufs2_daddr_t blkno; 5927 5928 ump = VFSTOUFS(mp); 5929 LOCK_OWNED(ump); 5930 indirdep = NULL; 5931 newindirdep = NULL; 5932 fs = ump->um_fs; 5933 for (;;) { 5934 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 5935 if (wk->wk_type != D_INDIRDEP) 5936 continue; 5937 indirdep = WK_INDIRDEP(wk); 5938 break; 5939 } 5940 /* Found on the buffer worklist, no new structure to free. */ 5941 if (indirdep != NULL && newindirdep == NULL) 5942 return (indirdep); 5943 if (indirdep != NULL && newindirdep != NULL) 5944 panic("indirdep_lookup: simultaneous create"); 5945 /* None found on the buffer and a new structure is ready. */ 5946 if (indirdep == NULL && newindirdep != NULL) 5947 break; 5948 /* None found and no new structure available. */ 5949 FREE_LOCK(ump); 5950 newindirdep = malloc(sizeof(struct indirdep), 5951 M_INDIRDEP, M_SOFTDEP_FLAGS); 5952 workitem_alloc(&newindirdep->ir_list, D_INDIRDEP, mp); 5953 newindirdep->ir_state = ATTACHED; 5954 if (I_IS_UFS1(ip)) 5955 newindirdep->ir_state |= UFS1FMT; 5956 TAILQ_INIT(&newindirdep->ir_trunc); 5957 newindirdep->ir_saveddata = NULL; 5958 LIST_INIT(&newindirdep->ir_deplisthd); 5959 LIST_INIT(&newindirdep->ir_donehd); 5960 LIST_INIT(&newindirdep->ir_writehd); 5961 LIST_INIT(&newindirdep->ir_completehd); 5962 if (bp->b_blkno == bp->b_lblkno) { 5963 ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp, 5964 NULL, NULL); 5965 bp->b_blkno = blkno; 5966 } 5967 newindirdep->ir_freeblks = NULL; 5968 newindirdep->ir_savebp = 5969 getblk(ump->um_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0); 5970 newindirdep->ir_bp = bp; 5971 BUF_KERNPROC(newindirdep->ir_savebp); 5972 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount); 5973 ACQUIRE_LOCK(ump); 5974 } 5975 indirdep = newindirdep; 5976 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list); 5977 /* 5978 * If the block is not yet allocated we don't set DEPCOMPLETE so 5979 * that we don't free dependencies until the pointers are valid. 5980 * This could search b_dep for D_ALLOCDIRECT/D_ALLOCINDIR rather 5981 * than using the hash. 5982 */ 5983 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk)) 5984 LIST_INSERT_HEAD(&newblk->nb_indirdeps, indirdep, ir_next); 5985 else 5986 indirdep->ir_state |= DEPCOMPLETE; 5987 return (indirdep); 5988 } 5989 5990 /* 5991 * Called to finish the allocation of the "aip" allocated 5992 * by one of the two routines above. 5993 */ 5994 static struct freefrag * 5995 setup_allocindir_phase2(bp, ip, inodedep, aip, lbn) 5996 struct buf *bp; /* in-memory copy of the indirect block */ 5997 struct inode *ip; /* inode for file being extended */ 5998 struct inodedep *inodedep; /* Inodedep for ip */ 5999 struct allocindir *aip; /* allocindir allocated by the above routines */ 6000 ufs_lbn_t lbn; /* Logical block number for this block. */ 6001 { 6002 struct fs *fs; 6003 struct indirdep *indirdep; 6004 struct allocindir *oldaip; 6005 struct freefrag *freefrag; 6006 struct mount *mp; 6007 struct ufsmount *ump; 6008 6009 mp = ITOVFS(ip); 6010 ump = VFSTOUFS(mp); 6011 LOCK_OWNED(ump); 6012 fs = ump->um_fs; 6013 if (bp->b_lblkno >= 0) 6014 panic("setup_allocindir_phase2: not indir blk"); 6015 KASSERT(aip->ai_offset >= 0 && aip->ai_offset < NINDIR(fs), 6016 ("setup_allocindir_phase2: Bad offset %d", aip->ai_offset)); 6017 indirdep = indirdep_lookup(mp, ip, bp); 6018 KASSERT(indirdep->ir_savebp != NULL, 6019 ("setup_allocindir_phase2 NULL ir_savebp")); 6020 aip->ai_indirdep = indirdep; 6021 /* 6022 * Check for an unwritten dependency for this indirect offset. If 6023 * there is, merge the old dependency into the new one. This happens 6024 * as a result of reallocblk only. 6025 */ 6026 freefrag = NULL; 6027 if (aip->ai_oldblkno != 0) { 6028 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next) { 6029 if (oldaip->ai_offset == aip->ai_offset) { 6030 freefrag = allocindir_merge(aip, oldaip); 6031 goto done; 6032 } 6033 } 6034 LIST_FOREACH(oldaip, &indirdep->ir_donehd, ai_next) { 6035 if (oldaip->ai_offset == aip->ai_offset) { 6036 freefrag = allocindir_merge(aip, oldaip); 6037 goto done; 6038 } 6039 } 6040 } 6041 done: 6042 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next); 6043 return (freefrag); 6044 } 6045 6046 /* 6047 * Merge two allocindirs which refer to the same block. Move newblock 6048 * dependencies and setup the freefrags appropriately. 6049 */ 6050 static struct freefrag * 6051 allocindir_merge(aip, oldaip) 6052 struct allocindir *aip; 6053 struct allocindir *oldaip; 6054 { 6055 struct freefrag *freefrag; 6056 struct worklist *wk; 6057 6058 if (oldaip->ai_newblkno != aip->ai_oldblkno) 6059 panic("allocindir_merge: blkno"); 6060 aip->ai_oldblkno = oldaip->ai_oldblkno; 6061 freefrag = aip->ai_freefrag; 6062 aip->ai_freefrag = oldaip->ai_freefrag; 6063 oldaip->ai_freefrag = NULL; 6064 KASSERT(freefrag != NULL, ("setup_allocindir_phase2: No freefrag")); 6065 /* 6066 * If we are tracking a new directory-block allocation, 6067 * move it from the old allocindir to the new allocindir. 6068 */ 6069 if ((wk = LIST_FIRST(&oldaip->ai_newdirblk)) != NULL) { 6070 WORKLIST_REMOVE(wk); 6071 if (!LIST_EMPTY(&oldaip->ai_newdirblk)) 6072 panic("allocindir_merge: extra newdirblk"); 6073 WORKLIST_INSERT(&aip->ai_newdirblk, wk); 6074 } 6075 /* 6076 * We can skip journaling for this freefrag and just complete 6077 * any pending journal work for the allocindir that is being 6078 * removed after the freefrag completes. 6079 */ 6080 if (freefrag->ff_jdep) 6081 cancel_jfreefrag(WK_JFREEFRAG(freefrag->ff_jdep)); 6082 LIST_REMOVE(oldaip, ai_next); 6083 freefrag->ff_jdep = (struct worklist *)cancel_newblk(&oldaip->ai_block, 6084 &freefrag->ff_list, &freefrag->ff_jwork); 6085 free_newblk(&oldaip->ai_block); 6086 6087 return (freefrag); 6088 } 6089 6090 static inline void 6091 setup_freedirect(freeblks, ip, i, needj) 6092 struct freeblks *freeblks; 6093 struct inode *ip; 6094 int i; 6095 int needj; 6096 { 6097 struct ufsmount *ump; 6098 ufs2_daddr_t blkno; 6099 int frags; 6100 6101 blkno = DIP(ip, i_db[i]); 6102 if (blkno == 0) 6103 return; 6104 DIP_SET(ip, i_db[i], 0); 6105 ump = ITOUMP(ip); 6106 frags = sblksize(ump->um_fs, ip->i_size, i); 6107 frags = numfrags(ump->um_fs, frags); 6108 newfreework(ump, freeblks, NULL, i, blkno, frags, 0, needj); 6109 } 6110 6111 static inline void 6112 setup_freeext(freeblks, ip, i, needj) 6113 struct freeblks *freeblks; 6114 struct inode *ip; 6115 int i; 6116 int needj; 6117 { 6118 struct ufsmount *ump; 6119 ufs2_daddr_t blkno; 6120 int frags; 6121 6122 blkno = ip->i_din2->di_extb[i]; 6123 if (blkno == 0) 6124 return; 6125 ip->i_din2->di_extb[i] = 0; 6126 ump = ITOUMP(ip); 6127 frags = sblksize(ump->um_fs, ip->i_din2->di_extsize, i); 6128 frags = numfrags(ump->um_fs, frags); 6129 newfreework(ump, freeblks, NULL, -1 - i, blkno, frags, 0, needj); 6130 } 6131 6132 static inline void 6133 setup_freeindir(freeblks, ip, i, lbn, needj) 6134 struct freeblks *freeblks; 6135 struct inode *ip; 6136 int i; 6137 ufs_lbn_t lbn; 6138 int needj; 6139 { 6140 struct ufsmount *ump; 6141 ufs2_daddr_t blkno; 6142 6143 blkno = DIP(ip, i_ib[i]); 6144 if (blkno == 0) 6145 return; 6146 DIP_SET(ip, i_ib[i], 0); 6147 ump = ITOUMP(ip); 6148 newfreework(ump, freeblks, NULL, lbn, blkno, ump->um_fs->fs_frag, 6149 0, needj); 6150 } 6151 6152 static inline struct freeblks * 6153 newfreeblks(mp, ip) 6154 struct mount *mp; 6155 struct inode *ip; 6156 { 6157 struct freeblks *freeblks; 6158 6159 freeblks = malloc(sizeof(struct freeblks), 6160 M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO); 6161 workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp); 6162 LIST_INIT(&freeblks->fb_jblkdephd); 6163 LIST_INIT(&freeblks->fb_jwork); 6164 freeblks->fb_ref = 0; 6165 freeblks->fb_cgwait = 0; 6166 freeblks->fb_state = ATTACHED; 6167 freeblks->fb_uid = ip->i_uid; 6168 freeblks->fb_inum = ip->i_number; 6169 freeblks->fb_vtype = ITOV(ip)->v_type; 6170 freeblks->fb_modrev = DIP(ip, i_modrev); 6171 freeblks->fb_devvp = ITODEVVP(ip); 6172 freeblks->fb_chkcnt = 0; 6173 freeblks->fb_len = 0; 6174 6175 return (freeblks); 6176 } 6177 6178 static void 6179 trunc_indirdep(indirdep, freeblks, bp, off) 6180 struct indirdep *indirdep; 6181 struct freeblks *freeblks; 6182 struct buf *bp; 6183 int off; 6184 { 6185 struct allocindir *aip, *aipn; 6186 6187 /* 6188 * The first set of allocindirs won't be in savedbp. 6189 */ 6190 LIST_FOREACH_SAFE(aip, &indirdep->ir_deplisthd, ai_next, aipn) 6191 if (aip->ai_offset > off) 6192 cancel_allocindir(aip, bp, freeblks, 1); 6193 LIST_FOREACH_SAFE(aip, &indirdep->ir_donehd, ai_next, aipn) 6194 if (aip->ai_offset > off) 6195 cancel_allocindir(aip, bp, freeblks, 1); 6196 /* 6197 * These will exist in savedbp. 6198 */ 6199 LIST_FOREACH_SAFE(aip, &indirdep->ir_writehd, ai_next, aipn) 6200 if (aip->ai_offset > off) 6201 cancel_allocindir(aip, NULL, freeblks, 0); 6202 LIST_FOREACH_SAFE(aip, &indirdep->ir_completehd, ai_next, aipn) 6203 if (aip->ai_offset > off) 6204 cancel_allocindir(aip, NULL, freeblks, 0); 6205 } 6206 6207 /* 6208 * Follow the chain of indirects down to lastlbn creating a freework 6209 * structure for each. This will be used to start indir_trunc() at 6210 * the right offset and create the journal records for the parrtial 6211 * truncation. A second step will handle the truncated dependencies. 6212 */ 6213 static int 6214 setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno) 6215 struct freeblks *freeblks; 6216 struct inode *ip; 6217 ufs_lbn_t lbn; 6218 ufs_lbn_t lastlbn; 6219 ufs2_daddr_t blkno; 6220 { 6221 struct indirdep *indirdep; 6222 struct indirdep *indirn; 6223 struct freework *freework; 6224 struct newblk *newblk; 6225 struct mount *mp; 6226 struct ufsmount *ump; 6227 struct buf *bp; 6228 uint8_t *start; 6229 uint8_t *end; 6230 ufs_lbn_t lbnadd; 6231 int level; 6232 int error; 6233 int off; 6234 6235 6236 freework = NULL; 6237 if (blkno == 0) 6238 return (0); 6239 mp = freeblks->fb_list.wk_mp; 6240 ump = VFSTOUFS(mp); 6241 bp = getblk(ITOV(ip), lbn, mp->mnt_stat.f_iosize, 0, 0, 0); 6242 if ((bp->b_flags & B_CACHE) == 0) { 6243 bp->b_blkno = blkptrtodb(VFSTOUFS(mp), blkno); 6244 bp->b_iocmd = BIO_READ; 6245 bp->b_flags &= ~B_INVAL; 6246 bp->b_ioflags &= ~BIO_ERROR; 6247 vfs_busy_pages(bp, 0); 6248 bp->b_iooffset = dbtob(bp->b_blkno); 6249 bstrategy(bp); 6250 #ifdef RACCT 6251 if (racct_enable) { 6252 PROC_LOCK(curproc); 6253 racct_add_buf(curproc, bp, 0); 6254 PROC_UNLOCK(curproc); 6255 } 6256 #endif /* RACCT */ 6257 curthread->td_ru.ru_inblock++; 6258 error = bufwait(bp); 6259 if (error) { 6260 brelse(bp); 6261 return (error); 6262 } 6263 } 6264 level = lbn_level(lbn); 6265 lbnadd = lbn_offset(ump->um_fs, level); 6266 /* 6267 * Compute the offset of the last block we want to keep. Store 6268 * in the freework the first block we want to completely free. 6269 */ 6270 off = (lastlbn - -(lbn + level)) / lbnadd; 6271 if (off + 1 == NINDIR(ump->um_fs)) 6272 goto nowork; 6273 freework = newfreework(ump, freeblks, NULL, lbn, blkno, 0, off + 1, 0); 6274 /* 6275 * Link the freework into the indirdep. This will prevent any new 6276 * allocations from proceeding until we are finished with the 6277 * truncate and the block is written. 6278 */ 6279 ACQUIRE_LOCK(ump); 6280 indirdep = indirdep_lookup(mp, ip, bp); 6281 if (indirdep->ir_freeblks) 6282 panic("setup_trunc_indir: indirdep already truncated."); 6283 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, freework, fw_next); 6284 freework->fw_indir = indirdep; 6285 /* 6286 * Cancel any allocindirs that will not make it to disk. 6287 * We have to do this for all copies of the indirdep that 6288 * live on this newblk. 6289 */ 6290 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 6291 newblk_lookup(mp, dbtofsb(ump->um_fs, bp->b_blkno), 0, &newblk); 6292 LIST_FOREACH(indirn, &newblk->nb_indirdeps, ir_next) 6293 trunc_indirdep(indirn, freeblks, bp, off); 6294 } else 6295 trunc_indirdep(indirdep, freeblks, bp, off); 6296 FREE_LOCK(ump); 6297 /* 6298 * Creation is protected by the buf lock. The saveddata is only 6299 * needed if a full truncation follows a partial truncation but it 6300 * is difficult to allocate in that case so we fetch it anyway. 6301 */ 6302 if (indirdep->ir_saveddata == NULL) 6303 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP, 6304 M_SOFTDEP_FLAGS); 6305 nowork: 6306 /* Fetch the blkno of the child and the zero start offset. */ 6307 if (I_IS_UFS1(ip)) { 6308 blkno = ((ufs1_daddr_t *)bp->b_data)[off]; 6309 start = (uint8_t *)&((ufs1_daddr_t *)bp->b_data)[off+1]; 6310 } else { 6311 blkno = ((ufs2_daddr_t *)bp->b_data)[off]; 6312 start = (uint8_t *)&((ufs2_daddr_t *)bp->b_data)[off+1]; 6313 } 6314 if (freework) { 6315 /* Zero the truncated pointers. */ 6316 end = bp->b_data + bp->b_bcount; 6317 bzero(start, end - start); 6318 bdwrite(bp); 6319 } else 6320 bqrelse(bp); 6321 if (level == 0) 6322 return (0); 6323 lbn++; /* adjust level */ 6324 lbn -= (off * lbnadd); 6325 return setup_trunc_indir(freeblks, ip, lbn, lastlbn, blkno); 6326 } 6327 6328 /* 6329 * Complete the partial truncation of an indirect block setup by 6330 * setup_trunc_indir(). This zeros the truncated pointers in the saved 6331 * copy and writes them to disk before the freeblks is allowed to complete. 6332 */ 6333 static void 6334 complete_trunc_indir(freework) 6335 struct freework *freework; 6336 { 6337 struct freework *fwn; 6338 struct indirdep *indirdep; 6339 struct ufsmount *ump; 6340 struct buf *bp; 6341 uintptr_t start; 6342 int count; 6343 6344 ump = VFSTOUFS(freework->fw_list.wk_mp); 6345 LOCK_OWNED(ump); 6346 indirdep = freework->fw_indir; 6347 for (;;) { 6348 bp = indirdep->ir_bp; 6349 /* See if the block was discarded. */ 6350 if (bp == NULL) 6351 break; 6352 /* Inline part of getdirtybuf(). We dont want bremfree. */ 6353 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) 6354 break; 6355 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 6356 LOCK_PTR(ump)) == 0) 6357 BUF_UNLOCK(bp); 6358 ACQUIRE_LOCK(ump); 6359 } 6360 freework->fw_state |= DEPCOMPLETE; 6361 TAILQ_REMOVE(&indirdep->ir_trunc, freework, fw_next); 6362 /* 6363 * Zero the pointers in the saved copy. 6364 */ 6365 if (indirdep->ir_state & UFS1FMT) 6366 start = sizeof(ufs1_daddr_t); 6367 else 6368 start = sizeof(ufs2_daddr_t); 6369 start *= freework->fw_start; 6370 count = indirdep->ir_savebp->b_bcount - start; 6371 start += (uintptr_t)indirdep->ir_savebp->b_data; 6372 bzero((char *)start, count); 6373 /* 6374 * We need to start the next truncation in the list if it has not 6375 * been started yet. 6376 */ 6377 fwn = TAILQ_FIRST(&indirdep->ir_trunc); 6378 if (fwn != NULL) { 6379 if (fwn->fw_freeblks == indirdep->ir_freeblks) 6380 TAILQ_REMOVE(&indirdep->ir_trunc, fwn, fw_next); 6381 if ((fwn->fw_state & ONWORKLIST) == 0) 6382 freework_enqueue(fwn); 6383 } 6384 /* 6385 * If bp is NULL the block was fully truncated, restore 6386 * the saved block list otherwise free it if it is no 6387 * longer needed. 6388 */ 6389 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 6390 if (bp == NULL) 6391 bcopy(indirdep->ir_saveddata, 6392 indirdep->ir_savebp->b_data, 6393 indirdep->ir_savebp->b_bcount); 6394 free(indirdep->ir_saveddata, M_INDIRDEP); 6395 indirdep->ir_saveddata = NULL; 6396 } 6397 /* 6398 * When bp is NULL there is a full truncation pending. We 6399 * must wait for this full truncation to be journaled before 6400 * we can release this freework because the disk pointers will 6401 * never be written as zero. 6402 */ 6403 if (bp == NULL) { 6404 if (LIST_EMPTY(&indirdep->ir_freeblks->fb_jblkdephd)) 6405 handle_written_freework(freework); 6406 else 6407 WORKLIST_INSERT(&indirdep->ir_freeblks->fb_freeworkhd, 6408 &freework->fw_list); 6409 } else { 6410 /* Complete when the real copy is written. */ 6411 WORKLIST_INSERT(&bp->b_dep, &freework->fw_list); 6412 BUF_UNLOCK(bp); 6413 } 6414 } 6415 6416 /* 6417 * Calculate the number of blocks we are going to release where datablocks 6418 * is the current total and length is the new file size. 6419 */ 6420 static ufs2_daddr_t 6421 blkcount(fs, datablocks, length) 6422 struct fs *fs; 6423 ufs2_daddr_t datablocks; 6424 off_t length; 6425 { 6426 off_t totblks, numblks; 6427 6428 totblks = 0; 6429 numblks = howmany(length, fs->fs_bsize); 6430 if (numblks <= NDADDR) { 6431 totblks = howmany(length, fs->fs_fsize); 6432 goto out; 6433 } 6434 totblks = blkstofrags(fs, numblks); 6435 numblks -= NDADDR; 6436 /* 6437 * Count all single, then double, then triple indirects required. 6438 * Subtracting one indirects worth of blocks for each pass 6439 * acknowledges one of each pointed to by the inode. 6440 */ 6441 for (;;) { 6442 totblks += blkstofrags(fs, howmany(numblks, NINDIR(fs))); 6443 numblks -= NINDIR(fs); 6444 if (numblks <= 0) 6445 break; 6446 numblks = howmany(numblks, NINDIR(fs)); 6447 } 6448 out: 6449 totblks = fsbtodb(fs, totblks); 6450 /* 6451 * Handle sparse files. We can't reclaim more blocks than the inode 6452 * references. We will correct it later in handle_complete_freeblks() 6453 * when we know the real count. 6454 */ 6455 if (totblks > datablocks) 6456 return (0); 6457 return (datablocks - totblks); 6458 } 6459 6460 /* 6461 * Handle freeblocks for journaled softupdate filesystems. 6462 * 6463 * Contrary to normal softupdates, we must preserve the block pointers in 6464 * indirects until their subordinates are free. This is to avoid journaling 6465 * every block that is freed which may consume more space than the journal 6466 * itself. The recovery program will see the free block journals at the 6467 * base of the truncated area and traverse them to reclaim space. The 6468 * pointers in the inode may be cleared immediately after the journal 6469 * records are written because each direct and indirect pointer in the 6470 * inode is recorded in a journal. This permits full truncation to proceed 6471 * asynchronously. The write order is journal -> inode -> cgs -> indirects. 6472 * 6473 * The algorithm is as follows: 6474 * 1) Traverse the in-memory state and create journal entries to release 6475 * the relevant blocks and full indirect trees. 6476 * 2) Traverse the indirect block chain adding partial truncation freework 6477 * records to indirects in the path to lastlbn. The freework will 6478 * prevent new allocation dependencies from being satisfied in this 6479 * indirect until the truncation completes. 6480 * 3) Read and lock the inode block, performing an update with the new size 6481 * and pointers. This prevents truncated data from becoming valid on 6482 * disk through step 4. 6483 * 4) Reap unsatisfied dependencies that are beyond the truncated area, 6484 * eliminate journal work for those records that do not require it. 6485 * 5) Schedule the journal records to be written followed by the inode block. 6486 * 6) Allocate any necessary frags for the end of file. 6487 * 7) Zero any partially truncated blocks. 6488 * 6489 * From this truncation proceeds asynchronously using the freework and 6490 * indir_trunc machinery. The file will not be extended again into a 6491 * partially truncated indirect block until all work is completed but 6492 * the normal dependency mechanism ensures that it is rolled back/forward 6493 * as appropriate. Further truncation may occur without delay and is 6494 * serialized in indir_trunc(). 6495 */ 6496 void 6497 softdep_journal_freeblocks(ip, cred, length, flags) 6498 struct inode *ip; /* The inode whose length is to be reduced */ 6499 struct ucred *cred; 6500 off_t length; /* The new length for the file */ 6501 int flags; /* IO_EXT and/or IO_NORMAL */ 6502 { 6503 struct freeblks *freeblks, *fbn; 6504 struct worklist *wk, *wkn; 6505 struct inodedep *inodedep; 6506 struct jblkdep *jblkdep; 6507 struct allocdirect *adp, *adpn; 6508 struct ufsmount *ump; 6509 struct fs *fs; 6510 struct buf *bp; 6511 struct vnode *vp; 6512 struct mount *mp; 6513 ufs2_daddr_t extblocks, datablocks; 6514 ufs_lbn_t tmpval, lbn, lastlbn; 6515 int frags, lastoff, iboff, allocblock, needj, error, i; 6516 6517 ump = ITOUMP(ip); 6518 mp = UFSTOVFS(ump); 6519 fs = ump->um_fs; 6520 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 6521 ("softdep_journal_freeblocks called on non-softdep filesystem")); 6522 vp = ITOV(ip); 6523 needj = 1; 6524 iboff = -1; 6525 allocblock = 0; 6526 extblocks = 0; 6527 datablocks = 0; 6528 frags = 0; 6529 freeblks = newfreeblks(mp, ip); 6530 ACQUIRE_LOCK(ump); 6531 /* 6532 * If we're truncating a removed file that will never be written 6533 * we don't need to journal the block frees. The canceled journals 6534 * for the allocations will suffice. 6535 */ 6536 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6537 if ((inodedep->id_state & (UNLINKED | DEPCOMPLETE)) == UNLINKED && 6538 length == 0) 6539 needj = 0; 6540 CTR3(KTR_SUJ, "softdep_journal_freeblks: ip %d length %ld needj %d", 6541 ip->i_number, length, needj); 6542 FREE_LOCK(ump); 6543 /* 6544 * Calculate the lbn that we are truncating to. This results in -1 6545 * if we're truncating the 0 bytes. So it is the last lbn we want 6546 * to keep, not the first lbn we want to truncate. 6547 */ 6548 lastlbn = lblkno(fs, length + fs->fs_bsize - 1) - 1; 6549 lastoff = blkoff(fs, length); 6550 /* 6551 * Compute frags we are keeping in lastlbn. 0 means all. 6552 */ 6553 if (lastlbn >= 0 && lastlbn < NDADDR) { 6554 frags = fragroundup(fs, lastoff); 6555 /* adp offset of last valid allocdirect. */ 6556 iboff = lastlbn; 6557 } else if (lastlbn > 0) 6558 iboff = NDADDR; 6559 if (fs->fs_magic == FS_UFS2_MAGIC) 6560 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 6561 /* 6562 * Handle normal data blocks and indirects. This section saves 6563 * values used after the inode update to complete frag and indirect 6564 * truncation. 6565 */ 6566 if ((flags & IO_NORMAL) != 0) { 6567 /* 6568 * Handle truncation of whole direct and indirect blocks. 6569 */ 6570 for (i = iboff + 1; i < NDADDR; i++) 6571 setup_freedirect(freeblks, ip, i, needj); 6572 for (i = 0, tmpval = NINDIR(fs), lbn = NDADDR; i < NIADDR; 6573 i++, lbn += tmpval, tmpval *= NINDIR(fs)) { 6574 /* Release a whole indirect tree. */ 6575 if (lbn > lastlbn) { 6576 setup_freeindir(freeblks, ip, i, -lbn -i, 6577 needj); 6578 continue; 6579 } 6580 iboff = i + NDADDR; 6581 /* 6582 * Traverse partially truncated indirect tree. 6583 */ 6584 if (lbn <= lastlbn && lbn + tmpval - 1 > lastlbn) 6585 setup_trunc_indir(freeblks, ip, -lbn - i, 6586 lastlbn, DIP(ip, i_ib[i])); 6587 } 6588 /* 6589 * Handle partial truncation to a frag boundary. 6590 */ 6591 if (frags) { 6592 ufs2_daddr_t blkno; 6593 long oldfrags; 6594 6595 oldfrags = blksize(fs, ip, lastlbn); 6596 blkno = DIP(ip, i_db[lastlbn]); 6597 if (blkno && oldfrags != frags) { 6598 oldfrags -= frags; 6599 oldfrags = numfrags(fs, oldfrags); 6600 blkno += numfrags(fs, frags); 6601 newfreework(ump, freeblks, NULL, lastlbn, 6602 blkno, oldfrags, 0, needj); 6603 if (needj) 6604 adjust_newfreework(freeblks, 6605 numfrags(fs, frags)); 6606 } else if (blkno == 0) 6607 allocblock = 1; 6608 } 6609 /* 6610 * Add a journal record for partial truncate if we are 6611 * handling indirect blocks. Non-indirects need no extra 6612 * journaling. 6613 */ 6614 if (length != 0 && lastlbn >= NDADDR) { 6615 ip->i_flag |= IN_TRUNCATED; 6616 newjtrunc(freeblks, length, 0); 6617 } 6618 ip->i_size = length; 6619 DIP_SET(ip, i_size, ip->i_size); 6620 ip->i_flag |= IN_SIZEMOD | IN_CHANGE; 6621 datablocks = DIP(ip, i_blocks) - extblocks; 6622 if (length != 0) 6623 datablocks = blkcount(fs, datablocks, length); 6624 freeblks->fb_len = length; 6625 } 6626 if ((flags & IO_EXT) != 0) { 6627 for (i = 0; i < NXADDR; i++) 6628 setup_freeext(freeblks, ip, i, needj); 6629 ip->i_din2->di_extsize = 0; 6630 datablocks += extblocks; 6631 ip->i_flag |= IN_SIZEMOD | IN_CHANGE; 6632 } 6633 #ifdef QUOTA 6634 /* Reference the quotas in case the block count is wrong in the end. */ 6635 quotaref(vp, freeblks->fb_quota); 6636 (void) chkdq(ip, -datablocks, NOCRED, 0); 6637 #endif 6638 freeblks->fb_chkcnt = -datablocks; 6639 UFS_LOCK(ump); 6640 fs->fs_pendingblocks += datablocks; 6641 UFS_UNLOCK(ump); 6642 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks); 6643 /* 6644 * Handle truncation of incomplete alloc direct dependencies. We 6645 * hold the inode block locked to prevent incomplete dependencies 6646 * from reaching the disk while we are eliminating those that 6647 * have been truncated. This is a partially inlined ffs_update(). 6648 */ 6649 ufs_itimes(vp); 6650 ip->i_flag &= ~(IN_LAZYACCESS | IN_LAZYMOD | IN_MODIFIED); 6651 error = bread(ump->um_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 6652 (int)fs->fs_bsize, cred, &bp); 6653 if (error) { 6654 brelse(bp); 6655 softdep_error("softdep_journal_freeblocks", error); 6656 return; 6657 } 6658 if (bp->b_bufsize == fs->fs_bsize) 6659 bp->b_flags |= B_CLUSTEROK; 6660 softdep_update_inodeblock(ip, bp, 0); 6661 if (ump->um_fstype == UFS1) 6662 *((struct ufs1_dinode *)bp->b_data + 6663 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1; 6664 else 6665 *((struct ufs2_dinode *)bp->b_data + 6666 ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2; 6667 ACQUIRE_LOCK(ump); 6668 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6669 if ((inodedep->id_state & IOSTARTED) != 0) 6670 panic("softdep_setup_freeblocks: inode busy"); 6671 /* 6672 * Add the freeblks structure to the list of operations that 6673 * must await the zero'ed inode being written to disk. If we 6674 * still have a bitmap dependency (needj), then the inode 6675 * has never been written to disk, so we can process the 6676 * freeblks below once we have deleted the dependencies. 6677 */ 6678 if (needj) 6679 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list); 6680 else 6681 freeblks->fb_state |= COMPLETE; 6682 if ((flags & IO_NORMAL) != 0) { 6683 TAILQ_FOREACH_SAFE(adp, &inodedep->id_inoupdt, ad_next, adpn) { 6684 if (adp->ad_offset > iboff) 6685 cancel_allocdirect(&inodedep->id_inoupdt, adp, 6686 freeblks); 6687 /* 6688 * Truncate the allocdirect. We could eliminate 6689 * or modify journal records as well. 6690 */ 6691 else if (adp->ad_offset == iboff && frags) 6692 adp->ad_newsize = frags; 6693 } 6694 } 6695 if ((flags & IO_EXT) != 0) 6696 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 6697 cancel_allocdirect(&inodedep->id_extupdt, adp, 6698 freeblks); 6699 /* 6700 * Scan the bufwait list for newblock dependencies that will never 6701 * make it to disk. 6702 */ 6703 LIST_FOREACH_SAFE(wk, &inodedep->id_bufwait, wk_list, wkn) { 6704 if (wk->wk_type != D_ALLOCDIRECT) 6705 continue; 6706 adp = WK_ALLOCDIRECT(wk); 6707 if (((flags & IO_NORMAL) != 0 && (adp->ad_offset > iboff)) || 6708 ((flags & IO_EXT) != 0 && (adp->ad_state & EXTDATA))) { 6709 cancel_jfreeblk(freeblks, adp->ad_newblkno); 6710 cancel_newblk(WK_NEWBLK(wk), NULL, &freeblks->fb_jwork); 6711 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk); 6712 } 6713 } 6714 /* 6715 * Add journal work. 6716 */ 6717 LIST_FOREACH(jblkdep, &freeblks->fb_jblkdephd, jb_deps) 6718 add_to_journal(&jblkdep->jb_list); 6719 FREE_LOCK(ump); 6720 bdwrite(bp); 6721 /* 6722 * Truncate dependency structures beyond length. 6723 */ 6724 trunc_dependencies(ip, freeblks, lastlbn, frags, flags); 6725 /* 6726 * This is only set when we need to allocate a fragment because 6727 * none existed at the end of a frag-sized file. It handles only 6728 * allocating a new, zero filled block. 6729 */ 6730 if (allocblock) { 6731 ip->i_size = length - lastoff; 6732 DIP_SET(ip, i_size, ip->i_size); 6733 error = UFS_BALLOC(vp, length - 1, 1, cred, BA_CLRBUF, &bp); 6734 if (error != 0) { 6735 softdep_error("softdep_journal_freeblks", error); 6736 return; 6737 } 6738 ip->i_size = length; 6739 DIP_SET(ip, i_size, length); 6740 ip->i_flag |= IN_SIZEMOD | IN_CHANGE | IN_UPDATE; 6741 allocbuf(bp, frags); 6742 ffs_update(vp, 0); 6743 bawrite(bp); 6744 } else if (lastoff != 0 && vp->v_type != VDIR) { 6745 int size; 6746 6747 /* 6748 * Zero the end of a truncated frag or block. 6749 */ 6750 size = sblksize(fs, length, lastlbn); 6751 error = bread(vp, lastlbn, size, cred, &bp); 6752 if (error) { 6753 softdep_error("softdep_journal_freeblks", error); 6754 return; 6755 } 6756 bzero((char *)bp->b_data + lastoff, size - lastoff); 6757 bawrite(bp); 6758 6759 } 6760 ACQUIRE_LOCK(ump); 6761 inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6762 TAILQ_INSERT_TAIL(&inodedep->id_freeblklst, freeblks, fb_next); 6763 freeblks->fb_state |= DEPCOMPLETE | ONDEPLIST; 6764 /* 6765 * We zero earlier truncations so they don't erroneously 6766 * update i_blocks. 6767 */ 6768 if (freeblks->fb_len == 0 && (flags & IO_NORMAL) != 0) 6769 TAILQ_FOREACH(fbn, &inodedep->id_freeblklst, fb_next) 6770 fbn->fb_len = 0; 6771 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE && 6772 LIST_EMPTY(&freeblks->fb_jblkdephd)) 6773 freeblks->fb_state |= INPROGRESS; 6774 else 6775 freeblks = NULL; 6776 FREE_LOCK(ump); 6777 if (freeblks) 6778 handle_workitem_freeblocks(freeblks, 0); 6779 trunc_pages(ip, length, extblocks, flags); 6780 6781 } 6782 6783 /* 6784 * Flush a JOP_SYNC to the journal. 6785 */ 6786 void 6787 softdep_journal_fsync(ip) 6788 struct inode *ip; 6789 { 6790 struct jfsync *jfsync; 6791 struct ufsmount *ump; 6792 6793 ump = ITOUMP(ip); 6794 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 6795 ("softdep_journal_fsync called on non-softdep filesystem")); 6796 if ((ip->i_flag & IN_TRUNCATED) == 0) 6797 return; 6798 ip->i_flag &= ~IN_TRUNCATED; 6799 jfsync = malloc(sizeof(*jfsync), M_JFSYNC, M_SOFTDEP_FLAGS | M_ZERO); 6800 workitem_alloc(&jfsync->jfs_list, D_JFSYNC, UFSTOVFS(ump)); 6801 jfsync->jfs_size = ip->i_size; 6802 jfsync->jfs_ino = ip->i_number; 6803 ACQUIRE_LOCK(ump); 6804 add_to_journal(&jfsync->jfs_list); 6805 jwait(&jfsync->jfs_list, MNT_WAIT); 6806 FREE_LOCK(ump); 6807 } 6808 6809 /* 6810 * Block de-allocation dependencies. 6811 * 6812 * When blocks are de-allocated, the on-disk pointers must be nullified before 6813 * the blocks are made available for use by other files. (The true 6814 * requirement is that old pointers must be nullified before new on-disk 6815 * pointers are set. We chose this slightly more stringent requirement to 6816 * reduce complexity.) Our implementation handles this dependency by updating 6817 * the inode (or indirect block) appropriately but delaying the actual block 6818 * de-allocation (i.e., freemap and free space count manipulation) until 6819 * after the updated versions reach stable storage. After the disk is 6820 * updated, the blocks can be safely de-allocated whenever it is convenient. 6821 * This implementation handles only the common case of reducing a file's 6822 * length to zero. Other cases are handled by the conventional synchronous 6823 * write approach. 6824 * 6825 * The ffs implementation with which we worked double-checks 6826 * the state of the block pointers and file size as it reduces 6827 * a file's length. Some of this code is replicated here in our 6828 * soft updates implementation. The freeblks->fb_chkcnt field is 6829 * used to transfer a part of this information to the procedure 6830 * that eventually de-allocates the blocks. 6831 * 6832 * This routine should be called from the routine that shortens 6833 * a file's length, before the inode's size or block pointers 6834 * are modified. It will save the block pointer information for 6835 * later release and zero the inode so that the calling routine 6836 * can release it. 6837 */ 6838 void 6839 softdep_setup_freeblocks(ip, length, flags) 6840 struct inode *ip; /* The inode whose length is to be reduced */ 6841 off_t length; /* The new length for the file */ 6842 int flags; /* IO_EXT and/or IO_NORMAL */ 6843 { 6844 struct ufs1_dinode *dp1; 6845 struct ufs2_dinode *dp2; 6846 struct freeblks *freeblks; 6847 struct inodedep *inodedep; 6848 struct allocdirect *adp; 6849 struct ufsmount *ump; 6850 struct buf *bp; 6851 struct fs *fs; 6852 ufs2_daddr_t extblocks, datablocks; 6853 struct mount *mp; 6854 int i, delay, error; 6855 ufs_lbn_t tmpval; 6856 ufs_lbn_t lbn; 6857 6858 ump = ITOUMP(ip); 6859 mp = UFSTOVFS(ump); 6860 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 6861 ("softdep_setup_freeblocks called on non-softdep filesystem")); 6862 CTR2(KTR_SUJ, "softdep_setup_freeblks: ip %d length %ld", 6863 ip->i_number, length); 6864 KASSERT(length == 0, ("softdep_setup_freeblocks: non-zero length")); 6865 fs = ump->um_fs; 6866 if ((error = bread(ump->um_devvp, 6867 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 6868 (int)fs->fs_bsize, NOCRED, &bp)) != 0) { 6869 brelse(bp); 6870 softdep_error("softdep_setup_freeblocks", error); 6871 return; 6872 } 6873 freeblks = newfreeblks(mp, ip); 6874 extblocks = 0; 6875 datablocks = 0; 6876 if (fs->fs_magic == FS_UFS2_MAGIC) 6877 extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize)); 6878 if ((flags & IO_NORMAL) != 0) { 6879 for (i = 0; i < NDADDR; i++) 6880 setup_freedirect(freeblks, ip, i, 0); 6881 for (i = 0, tmpval = NINDIR(fs), lbn = NDADDR; i < NIADDR; 6882 i++, lbn += tmpval, tmpval *= NINDIR(fs)) 6883 setup_freeindir(freeblks, ip, i, -lbn -i, 0); 6884 ip->i_size = 0; 6885 DIP_SET(ip, i_size, 0); 6886 ip->i_flag |= IN_SIZEMOD | IN_CHANGE; 6887 datablocks = DIP(ip, i_blocks) - extblocks; 6888 } 6889 if ((flags & IO_EXT) != 0) { 6890 for (i = 0; i < NXADDR; i++) 6891 setup_freeext(freeblks, ip, i, 0); 6892 ip->i_din2->di_extsize = 0; 6893 datablocks += extblocks; 6894 ip->i_flag |= IN_SIZEMOD | IN_CHANGE; 6895 } 6896 #ifdef QUOTA 6897 /* Reference the quotas in case the block count is wrong in the end. */ 6898 quotaref(ITOV(ip), freeblks->fb_quota); 6899 (void) chkdq(ip, -datablocks, NOCRED, 0); 6900 #endif 6901 freeblks->fb_chkcnt = -datablocks; 6902 UFS_LOCK(ump); 6903 fs->fs_pendingblocks += datablocks; 6904 UFS_UNLOCK(ump); 6905 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - datablocks); 6906 /* 6907 * Push the zero'ed inode to its disk buffer so that we are free 6908 * to delete its dependencies below. Once the dependencies are gone 6909 * the buffer can be safely released. 6910 */ 6911 if (ump->um_fstype == UFS1) { 6912 dp1 = ((struct ufs1_dinode *)bp->b_data + 6913 ino_to_fsbo(fs, ip->i_number)); 6914 ip->i_din1->di_freelink = dp1->di_freelink; 6915 *dp1 = *ip->i_din1; 6916 } else { 6917 dp2 = ((struct ufs2_dinode *)bp->b_data + 6918 ino_to_fsbo(fs, ip->i_number)); 6919 ip->i_din2->di_freelink = dp2->di_freelink; 6920 *dp2 = *ip->i_din2; 6921 } 6922 /* 6923 * Find and eliminate any inode dependencies. 6924 */ 6925 ACQUIRE_LOCK(ump); 6926 (void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep); 6927 if ((inodedep->id_state & IOSTARTED) != 0) 6928 panic("softdep_setup_freeblocks: inode busy"); 6929 /* 6930 * Add the freeblks structure to the list of operations that 6931 * must await the zero'ed inode being written to disk. If we 6932 * still have a bitmap dependency (delay == 0), then the inode 6933 * has never been written to disk, so we can process the 6934 * freeblks below once we have deleted the dependencies. 6935 */ 6936 delay = (inodedep->id_state & DEPCOMPLETE); 6937 if (delay) 6938 WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list); 6939 else 6940 freeblks->fb_state |= COMPLETE; 6941 /* 6942 * Because the file length has been truncated to zero, any 6943 * pending block allocation dependency structures associated 6944 * with this inode are obsolete and can simply be de-allocated. 6945 * We must first merge the two dependency lists to get rid of 6946 * any duplicate freefrag structures, then purge the merged list. 6947 * If we still have a bitmap dependency, then the inode has never 6948 * been written to disk, so we can free any fragments without delay. 6949 */ 6950 if (flags & IO_NORMAL) { 6951 merge_inode_lists(&inodedep->id_newinoupdt, 6952 &inodedep->id_inoupdt); 6953 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 6954 cancel_allocdirect(&inodedep->id_inoupdt, adp, 6955 freeblks); 6956 } 6957 if (flags & IO_EXT) { 6958 merge_inode_lists(&inodedep->id_newextupdt, 6959 &inodedep->id_extupdt); 6960 while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 6961 cancel_allocdirect(&inodedep->id_extupdt, adp, 6962 freeblks); 6963 } 6964 FREE_LOCK(ump); 6965 bdwrite(bp); 6966 trunc_dependencies(ip, freeblks, -1, 0, flags); 6967 ACQUIRE_LOCK(ump); 6968 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 6969 (void) free_inodedep(inodedep); 6970 freeblks->fb_state |= DEPCOMPLETE; 6971 /* 6972 * If the inode with zeroed block pointers is now on disk 6973 * we can start freeing blocks. 6974 */ 6975 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) 6976 freeblks->fb_state |= INPROGRESS; 6977 else 6978 freeblks = NULL; 6979 FREE_LOCK(ump); 6980 if (freeblks) 6981 handle_workitem_freeblocks(freeblks, 0); 6982 trunc_pages(ip, length, extblocks, flags); 6983 } 6984 6985 /* 6986 * Eliminate pages from the page cache that back parts of this inode and 6987 * adjust the vnode pager's idea of our size. This prevents stale data 6988 * from hanging around in the page cache. 6989 */ 6990 static void 6991 trunc_pages(ip, length, extblocks, flags) 6992 struct inode *ip; 6993 off_t length; 6994 ufs2_daddr_t extblocks; 6995 int flags; 6996 { 6997 struct vnode *vp; 6998 struct fs *fs; 6999 ufs_lbn_t lbn; 7000 off_t end, extend; 7001 7002 vp = ITOV(ip); 7003 fs = ITOFS(ip); 7004 extend = OFF_TO_IDX(lblktosize(fs, -extblocks)); 7005 if ((flags & IO_EXT) != 0) 7006 vn_pages_remove(vp, extend, 0); 7007 if ((flags & IO_NORMAL) == 0) 7008 return; 7009 BO_LOCK(&vp->v_bufobj); 7010 drain_output(vp); 7011 BO_UNLOCK(&vp->v_bufobj); 7012 /* 7013 * The vnode pager eliminates file pages we eliminate indirects 7014 * below. 7015 */ 7016 vnode_pager_setsize(vp, length); 7017 /* 7018 * Calculate the end based on the last indirect we want to keep. If 7019 * the block extends into indirects we can just use the negative of 7020 * its lbn. Doubles and triples exist at lower numbers so we must 7021 * be careful not to remove those, if they exist. double and triple 7022 * indirect lbns do not overlap with others so it is not important 7023 * to verify how many levels are required. 7024 */ 7025 lbn = lblkno(fs, length); 7026 if (lbn >= NDADDR) { 7027 /* Calculate the virtual lbn of the triple indirect. */ 7028 lbn = -lbn - (NIADDR - 1); 7029 end = OFF_TO_IDX(lblktosize(fs, lbn)); 7030 } else 7031 end = extend; 7032 vn_pages_remove(vp, OFF_TO_IDX(OFF_MAX), end); 7033 } 7034 7035 /* 7036 * See if the buf bp is in the range eliminated by truncation. 7037 */ 7038 static int 7039 trunc_check_buf(bp, blkoffp, lastlbn, lastoff, flags) 7040 struct buf *bp; 7041 int *blkoffp; 7042 ufs_lbn_t lastlbn; 7043 int lastoff; 7044 int flags; 7045 { 7046 ufs_lbn_t lbn; 7047 7048 *blkoffp = 0; 7049 /* Only match ext/normal blocks as appropriate. */ 7050 if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) || 7051 ((flags & IO_NORMAL) == 0 && (bp->b_xflags & BX_ALTDATA) == 0)) 7052 return (0); 7053 /* ALTDATA is always a full truncation. */ 7054 if ((bp->b_xflags & BX_ALTDATA) != 0) 7055 return (1); 7056 /* -1 is full truncation. */ 7057 if (lastlbn == -1) 7058 return (1); 7059 /* 7060 * If this is a partial truncate we only want those 7061 * blocks and indirect blocks that cover the range 7062 * we're after. 7063 */ 7064 lbn = bp->b_lblkno; 7065 if (lbn < 0) 7066 lbn = -(lbn + lbn_level(lbn)); 7067 if (lbn < lastlbn) 7068 return (0); 7069 /* Here we only truncate lblkno if it's partial. */ 7070 if (lbn == lastlbn) { 7071 if (lastoff == 0) 7072 return (0); 7073 *blkoffp = lastoff; 7074 } 7075 return (1); 7076 } 7077 7078 /* 7079 * Eliminate any dependencies that exist in memory beyond lblkno:off 7080 */ 7081 static void 7082 trunc_dependencies(ip, freeblks, lastlbn, lastoff, flags) 7083 struct inode *ip; 7084 struct freeblks *freeblks; 7085 ufs_lbn_t lastlbn; 7086 int lastoff; 7087 int flags; 7088 { 7089 struct bufobj *bo; 7090 struct vnode *vp; 7091 struct buf *bp; 7092 int blkoff; 7093 7094 /* 7095 * We must wait for any I/O in progress to finish so that 7096 * all potential buffers on the dirty list will be visible. 7097 * Once they are all there, walk the list and get rid of 7098 * any dependencies. 7099 */ 7100 vp = ITOV(ip); 7101 bo = &vp->v_bufobj; 7102 BO_LOCK(bo); 7103 drain_output(vp); 7104 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) 7105 bp->b_vflags &= ~BV_SCANNED; 7106 restart: 7107 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { 7108 if (bp->b_vflags & BV_SCANNED) 7109 continue; 7110 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) { 7111 bp->b_vflags |= BV_SCANNED; 7112 continue; 7113 } 7114 KASSERT(bp->b_bufobj == bo, ("Wrong object in buffer")); 7115 if ((bp = getdirtybuf(bp, BO_LOCKPTR(bo), MNT_WAIT)) == NULL) 7116 goto restart; 7117 BO_UNLOCK(bo); 7118 if (deallocate_dependencies(bp, freeblks, blkoff)) 7119 bqrelse(bp); 7120 else 7121 brelse(bp); 7122 BO_LOCK(bo); 7123 goto restart; 7124 } 7125 /* 7126 * Now do the work of vtruncbuf while also matching indirect blocks. 7127 */ 7128 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) 7129 bp->b_vflags &= ~BV_SCANNED; 7130 cleanrestart: 7131 TAILQ_FOREACH(bp, &bo->bo_clean.bv_hd, b_bobufs) { 7132 if (bp->b_vflags & BV_SCANNED) 7133 continue; 7134 if (!trunc_check_buf(bp, &blkoff, lastlbn, lastoff, flags)) { 7135 bp->b_vflags |= BV_SCANNED; 7136 continue; 7137 } 7138 if (BUF_LOCK(bp, 7139 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 7140 BO_LOCKPTR(bo)) == ENOLCK) { 7141 BO_LOCK(bo); 7142 goto cleanrestart; 7143 } 7144 bp->b_vflags |= BV_SCANNED; 7145 bremfree(bp); 7146 if (blkoff != 0) { 7147 allocbuf(bp, blkoff); 7148 bqrelse(bp); 7149 } else { 7150 bp->b_flags |= B_INVAL | B_NOCACHE | B_RELBUF; 7151 brelse(bp); 7152 } 7153 BO_LOCK(bo); 7154 goto cleanrestart; 7155 } 7156 drain_output(vp); 7157 BO_UNLOCK(bo); 7158 } 7159 7160 static int 7161 cancel_pagedep(pagedep, freeblks, blkoff) 7162 struct pagedep *pagedep; 7163 struct freeblks *freeblks; 7164 int blkoff; 7165 { 7166 struct jremref *jremref; 7167 struct jmvref *jmvref; 7168 struct dirrem *dirrem, *tmp; 7169 int i; 7170 7171 /* 7172 * Copy any directory remove dependencies to the list 7173 * to be processed after the freeblks proceeds. If 7174 * directory entry never made it to disk they 7175 * can be dumped directly onto the work list. 7176 */ 7177 LIST_FOREACH_SAFE(dirrem, &pagedep->pd_dirremhd, dm_next, tmp) { 7178 /* Skip this directory removal if it is intended to remain. */ 7179 if (dirrem->dm_offset < blkoff) 7180 continue; 7181 /* 7182 * If there are any dirrems we wait for the journal write 7183 * to complete and then restart the buf scan as the lock 7184 * has been dropped. 7185 */ 7186 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) { 7187 jwait(&jremref->jr_list, MNT_WAIT); 7188 return (ERESTART); 7189 } 7190 LIST_REMOVE(dirrem, dm_next); 7191 dirrem->dm_dirinum = pagedep->pd_ino; 7192 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &dirrem->dm_list); 7193 } 7194 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) { 7195 jwait(&jmvref->jm_list, MNT_WAIT); 7196 return (ERESTART); 7197 } 7198 /* 7199 * When we're partially truncating a pagedep we just want to flush 7200 * journal entries and return. There can not be any adds in the 7201 * truncated portion of the directory and newblk must remain if 7202 * part of the block remains. 7203 */ 7204 if (blkoff != 0) { 7205 struct diradd *dap; 7206 7207 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 7208 if (dap->da_offset > blkoff) 7209 panic("cancel_pagedep: diradd %p off %d > %d", 7210 dap, dap->da_offset, blkoff); 7211 for (i = 0; i < DAHASHSZ; i++) 7212 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) 7213 if (dap->da_offset > blkoff) 7214 panic("cancel_pagedep: diradd %p off %d > %d", 7215 dap, dap->da_offset, blkoff); 7216 return (0); 7217 } 7218 /* 7219 * There should be no directory add dependencies present 7220 * as the directory could not be truncated until all 7221 * children were removed. 7222 */ 7223 KASSERT(LIST_FIRST(&pagedep->pd_pendinghd) == NULL, 7224 ("deallocate_dependencies: pendinghd != NULL")); 7225 for (i = 0; i < DAHASHSZ; i++) 7226 KASSERT(LIST_FIRST(&pagedep->pd_diraddhd[i]) == NULL, 7227 ("deallocate_dependencies: diraddhd != NULL")); 7228 if ((pagedep->pd_state & NEWBLOCK) != 0) 7229 free_newdirblk(pagedep->pd_newdirblk); 7230 if (free_pagedep(pagedep) == 0) 7231 panic("Failed to free pagedep %p", pagedep); 7232 return (0); 7233 } 7234 7235 /* 7236 * Reclaim any dependency structures from a buffer that is about to 7237 * be reallocated to a new vnode. The buffer must be locked, thus, 7238 * no I/O completion operations can occur while we are manipulating 7239 * its associated dependencies. The mutex is held so that other I/O's 7240 * associated with related dependencies do not occur. 7241 */ 7242 static int 7243 deallocate_dependencies(bp, freeblks, off) 7244 struct buf *bp; 7245 struct freeblks *freeblks; 7246 int off; 7247 { 7248 struct indirdep *indirdep; 7249 struct pagedep *pagedep; 7250 struct worklist *wk, *wkn; 7251 struct ufsmount *ump; 7252 7253 ump = softdep_bp_to_mp(bp); 7254 if (ump == NULL) 7255 goto done; 7256 ACQUIRE_LOCK(ump); 7257 LIST_FOREACH_SAFE(wk, &bp->b_dep, wk_list, wkn) { 7258 switch (wk->wk_type) { 7259 case D_INDIRDEP: 7260 indirdep = WK_INDIRDEP(wk); 7261 if (bp->b_lblkno >= 0 || 7262 bp->b_blkno != indirdep->ir_savebp->b_lblkno) 7263 panic("deallocate_dependencies: not indir"); 7264 cancel_indirdep(indirdep, bp, freeblks); 7265 continue; 7266 7267 case D_PAGEDEP: 7268 pagedep = WK_PAGEDEP(wk); 7269 if (cancel_pagedep(pagedep, freeblks, off)) { 7270 FREE_LOCK(ump); 7271 return (ERESTART); 7272 } 7273 continue; 7274 7275 case D_ALLOCINDIR: 7276 /* 7277 * Simply remove the allocindir, we'll find it via 7278 * the indirdep where we can clear pointers if 7279 * needed. 7280 */ 7281 WORKLIST_REMOVE(wk); 7282 continue; 7283 7284 case D_FREEWORK: 7285 /* 7286 * A truncation is waiting for the zero'd pointers 7287 * to be written. It can be freed when the freeblks 7288 * is journaled. 7289 */ 7290 WORKLIST_REMOVE(wk); 7291 wk->wk_state |= ONDEPLIST; 7292 WORKLIST_INSERT(&freeblks->fb_freeworkhd, wk); 7293 break; 7294 7295 case D_ALLOCDIRECT: 7296 if (off != 0) 7297 continue; 7298 /* FALLTHROUGH */ 7299 default: 7300 panic("deallocate_dependencies: Unexpected type %s", 7301 TYPENAME(wk->wk_type)); 7302 /* NOTREACHED */ 7303 } 7304 } 7305 FREE_LOCK(ump); 7306 done: 7307 /* 7308 * Don't throw away this buf, we were partially truncating and 7309 * some deps may always remain. 7310 */ 7311 if (off) { 7312 allocbuf(bp, off); 7313 bp->b_vflags |= BV_SCANNED; 7314 return (EBUSY); 7315 } 7316 bp->b_flags |= B_INVAL | B_NOCACHE; 7317 7318 return (0); 7319 } 7320 7321 /* 7322 * An allocdirect is being canceled due to a truncate. We must make sure 7323 * the journal entry is released in concert with the blkfree that releases 7324 * the storage. Completed journal entries must not be released until the 7325 * space is no longer pointed to by the inode or in the bitmap. 7326 */ 7327 static void 7328 cancel_allocdirect(adphead, adp, freeblks) 7329 struct allocdirectlst *adphead; 7330 struct allocdirect *adp; 7331 struct freeblks *freeblks; 7332 { 7333 struct freework *freework; 7334 struct newblk *newblk; 7335 struct worklist *wk; 7336 7337 TAILQ_REMOVE(adphead, adp, ad_next); 7338 newblk = (struct newblk *)adp; 7339 freework = NULL; 7340 /* 7341 * Find the correct freework structure. 7342 */ 7343 LIST_FOREACH(wk, &freeblks->fb_freeworkhd, wk_list) { 7344 if (wk->wk_type != D_FREEWORK) 7345 continue; 7346 freework = WK_FREEWORK(wk); 7347 if (freework->fw_blkno == newblk->nb_newblkno) 7348 break; 7349 } 7350 if (freework == NULL) 7351 panic("cancel_allocdirect: Freework not found"); 7352 /* 7353 * If a newblk exists at all we still have the journal entry that 7354 * initiated the allocation so we do not need to journal the free. 7355 */ 7356 cancel_jfreeblk(freeblks, freework->fw_blkno); 7357 /* 7358 * If the journal hasn't been written the jnewblk must be passed 7359 * to the call to ffs_blkfree that reclaims the space. We accomplish 7360 * this by linking the journal dependency into the freework to be 7361 * freed when freework_freeblock() is called. If the journal has 7362 * been written we can simply reclaim the journal space when the 7363 * freeblks work is complete. 7364 */ 7365 freework->fw_jnewblk = cancel_newblk(newblk, &freework->fw_list, 7366 &freeblks->fb_jwork); 7367 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list); 7368 } 7369 7370 7371 /* 7372 * Cancel a new block allocation. May be an indirect or direct block. We 7373 * remove it from various lists and return any journal record that needs to 7374 * be resolved by the caller. 7375 * 7376 * A special consideration is made for indirects which were never pointed 7377 * at on disk and will never be found once this block is released. 7378 */ 7379 static struct jnewblk * 7380 cancel_newblk(newblk, wk, wkhd) 7381 struct newblk *newblk; 7382 struct worklist *wk; 7383 struct workhead *wkhd; 7384 { 7385 struct jnewblk *jnewblk; 7386 7387 CTR1(KTR_SUJ, "cancel_newblk: blkno %jd", newblk->nb_newblkno); 7388 7389 newblk->nb_state |= GOINGAWAY; 7390 /* 7391 * Previously we traversed the completedhd on each indirdep 7392 * attached to this newblk to cancel them and gather journal 7393 * work. Since we need only the oldest journal segment and 7394 * the lowest point on the tree will always have the oldest 7395 * journal segment we are free to release the segments 7396 * of any subordinates and may leave the indirdep list to 7397 * indirdep_complete() when this newblk is freed. 7398 */ 7399 if (newblk->nb_state & ONDEPLIST) { 7400 newblk->nb_state &= ~ONDEPLIST; 7401 LIST_REMOVE(newblk, nb_deps); 7402 } 7403 if (newblk->nb_state & ONWORKLIST) 7404 WORKLIST_REMOVE(&newblk->nb_list); 7405 /* 7406 * If the journal entry hasn't been written we save a pointer to 7407 * the dependency that frees it until it is written or the 7408 * superseding operation completes. 7409 */ 7410 jnewblk = newblk->nb_jnewblk; 7411 if (jnewblk != NULL && wk != NULL) { 7412 newblk->nb_jnewblk = NULL; 7413 jnewblk->jn_dep = wk; 7414 } 7415 if (!LIST_EMPTY(&newblk->nb_jwork)) 7416 jwork_move(wkhd, &newblk->nb_jwork); 7417 /* 7418 * When truncating we must free the newdirblk early to remove 7419 * the pagedep from the hash before returning. 7420 */ 7421 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL) 7422 free_newdirblk(WK_NEWDIRBLK(wk)); 7423 if (!LIST_EMPTY(&newblk->nb_newdirblk)) 7424 panic("cancel_newblk: extra newdirblk"); 7425 7426 return (jnewblk); 7427 } 7428 7429 /* 7430 * Schedule the freefrag associated with a newblk to be released once 7431 * the pointers are written and the previous block is no longer needed. 7432 */ 7433 static void 7434 newblk_freefrag(newblk) 7435 struct newblk *newblk; 7436 { 7437 struct freefrag *freefrag; 7438 7439 if (newblk->nb_freefrag == NULL) 7440 return; 7441 freefrag = newblk->nb_freefrag; 7442 newblk->nb_freefrag = NULL; 7443 freefrag->ff_state |= COMPLETE; 7444 if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE) 7445 add_to_worklist(&freefrag->ff_list, 0); 7446 } 7447 7448 /* 7449 * Free a newblk. Generate a new freefrag work request if appropriate. 7450 * This must be called after the inode pointer and any direct block pointers 7451 * are valid or fully removed via truncate or frag extension. 7452 */ 7453 static void 7454 free_newblk(newblk) 7455 struct newblk *newblk; 7456 { 7457 struct indirdep *indirdep; 7458 struct worklist *wk; 7459 7460 KASSERT(newblk->nb_jnewblk == NULL, 7461 ("free_newblk: jnewblk %p still attached", newblk->nb_jnewblk)); 7462 KASSERT(newblk->nb_list.wk_type != D_NEWBLK, 7463 ("free_newblk: unclaimed newblk")); 7464 LOCK_OWNED(VFSTOUFS(newblk->nb_list.wk_mp)); 7465 newblk_freefrag(newblk); 7466 if (newblk->nb_state & ONDEPLIST) 7467 LIST_REMOVE(newblk, nb_deps); 7468 if (newblk->nb_state & ONWORKLIST) 7469 WORKLIST_REMOVE(&newblk->nb_list); 7470 LIST_REMOVE(newblk, nb_hash); 7471 if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL) 7472 free_newdirblk(WK_NEWDIRBLK(wk)); 7473 if (!LIST_EMPTY(&newblk->nb_newdirblk)) 7474 panic("free_newblk: extra newdirblk"); 7475 while ((indirdep = LIST_FIRST(&newblk->nb_indirdeps)) != NULL) 7476 indirdep_complete(indirdep); 7477 handle_jwork(&newblk->nb_jwork); 7478 WORKITEM_FREE(newblk, D_NEWBLK); 7479 } 7480 7481 /* 7482 * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep. 7483 * This routine must be called with splbio interrupts blocked. 7484 */ 7485 static void 7486 free_newdirblk(newdirblk) 7487 struct newdirblk *newdirblk; 7488 { 7489 struct pagedep *pagedep; 7490 struct diradd *dap; 7491 struct worklist *wk; 7492 7493 LOCK_OWNED(VFSTOUFS(newdirblk->db_list.wk_mp)); 7494 WORKLIST_REMOVE(&newdirblk->db_list); 7495 /* 7496 * If the pagedep is still linked onto the directory buffer 7497 * dependency chain, then some of the entries on the 7498 * pd_pendinghd list may not be committed to disk yet. In 7499 * this case, we will simply clear the NEWBLOCK flag and 7500 * let the pd_pendinghd list be processed when the pagedep 7501 * is next written. If the pagedep is no longer on the buffer 7502 * dependency chain, then all the entries on the pd_pending 7503 * list are committed to disk and we can free them here. 7504 */ 7505 pagedep = newdirblk->db_pagedep; 7506 pagedep->pd_state &= ~NEWBLOCK; 7507 if ((pagedep->pd_state & ONWORKLIST) == 0) { 7508 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 7509 free_diradd(dap, NULL); 7510 /* 7511 * If no dependencies remain, the pagedep will be freed. 7512 */ 7513 free_pagedep(pagedep); 7514 } 7515 /* Should only ever be one item in the list. */ 7516 while ((wk = LIST_FIRST(&newdirblk->db_mkdir)) != NULL) { 7517 WORKLIST_REMOVE(wk); 7518 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 7519 } 7520 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 7521 } 7522 7523 /* 7524 * Prepare an inode to be freed. The actual free operation is not 7525 * done until the zero'ed inode has been written to disk. 7526 */ 7527 void 7528 softdep_freefile(pvp, ino, mode) 7529 struct vnode *pvp; 7530 ino_t ino; 7531 int mode; 7532 { 7533 struct inode *ip = VTOI(pvp); 7534 struct inodedep *inodedep; 7535 struct freefile *freefile; 7536 struct freeblks *freeblks; 7537 struct ufsmount *ump; 7538 7539 ump = ITOUMP(ip); 7540 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 7541 ("softdep_freefile called on non-softdep filesystem")); 7542 /* 7543 * This sets up the inode de-allocation dependency. 7544 */ 7545 freefile = malloc(sizeof(struct freefile), 7546 M_FREEFILE, M_SOFTDEP_FLAGS); 7547 workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount); 7548 freefile->fx_mode = mode; 7549 freefile->fx_oldinum = ino; 7550 freefile->fx_devvp = ump->um_devvp; 7551 LIST_INIT(&freefile->fx_jwork); 7552 UFS_LOCK(ump); 7553 ump->um_fs->fs_pendinginodes += 1; 7554 UFS_UNLOCK(ump); 7555 7556 /* 7557 * If the inodedep does not exist, then the zero'ed inode has 7558 * been written to disk. If the allocated inode has never been 7559 * written to disk, then the on-disk inode is zero'ed. In either 7560 * case we can free the file immediately. If the journal was 7561 * canceled before being written the inode will never make it to 7562 * disk and we must send the canceled journal entrys to 7563 * ffs_freefile() to be cleared in conjunction with the bitmap. 7564 * Any blocks waiting on the inode to write can be safely freed 7565 * here as it will never been written. 7566 */ 7567 ACQUIRE_LOCK(ump); 7568 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep); 7569 if (inodedep) { 7570 /* 7571 * Clear out freeblks that no longer need to reference 7572 * this inode. 7573 */ 7574 while ((freeblks = 7575 TAILQ_FIRST(&inodedep->id_freeblklst)) != NULL) { 7576 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, 7577 fb_next); 7578 freeblks->fb_state &= ~ONDEPLIST; 7579 } 7580 /* 7581 * Remove this inode from the unlinked list. 7582 */ 7583 if (inodedep->id_state & UNLINKED) { 7584 /* 7585 * Save the journal work to be freed with the bitmap 7586 * before we clear UNLINKED. Otherwise it can be lost 7587 * if the inode block is written. 7588 */ 7589 handle_bufwait(inodedep, &freefile->fx_jwork); 7590 clear_unlinked_inodedep(inodedep); 7591 /* 7592 * Re-acquire inodedep as we've dropped the 7593 * per-filesystem lock in clear_unlinked_inodedep(). 7594 */ 7595 inodedep_lookup(pvp->v_mount, ino, 0, &inodedep); 7596 } 7597 } 7598 if (inodedep == NULL || check_inode_unwritten(inodedep)) { 7599 FREE_LOCK(ump); 7600 handle_workitem_freefile(freefile); 7601 return; 7602 } 7603 if ((inodedep->id_state & DEPCOMPLETE) == 0) 7604 inodedep->id_state |= GOINGAWAY; 7605 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list); 7606 FREE_LOCK(ump); 7607 if (ip->i_number == ino) 7608 ip->i_flag |= IN_MODIFIED; 7609 } 7610 7611 /* 7612 * Check to see if an inode has never been written to disk. If 7613 * so free the inodedep and return success, otherwise return failure. 7614 * This routine must be called with splbio interrupts blocked. 7615 * 7616 * If we still have a bitmap dependency, then the inode has never 7617 * been written to disk. Drop the dependency as it is no longer 7618 * necessary since the inode is being deallocated. We set the 7619 * ALLCOMPLETE flags since the bitmap now properly shows that the 7620 * inode is not allocated. Even if the inode is actively being 7621 * written, it has been rolled back to its zero'ed state, so we 7622 * are ensured that a zero inode is what is on the disk. For short 7623 * lived files, this change will usually result in removing all the 7624 * dependencies from the inode so that it can be freed immediately. 7625 */ 7626 static int 7627 check_inode_unwritten(inodedep) 7628 struct inodedep *inodedep; 7629 { 7630 7631 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp)); 7632 7633 if ((inodedep->id_state & (DEPCOMPLETE | UNLINKED)) != 0 || 7634 !LIST_EMPTY(&inodedep->id_dirremhd) || 7635 !LIST_EMPTY(&inodedep->id_pendinghd) || 7636 !LIST_EMPTY(&inodedep->id_bufwait) || 7637 !LIST_EMPTY(&inodedep->id_inowait) || 7638 !TAILQ_EMPTY(&inodedep->id_inoreflst) || 7639 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 7640 !TAILQ_EMPTY(&inodedep->id_newinoupdt) || 7641 !TAILQ_EMPTY(&inodedep->id_extupdt) || 7642 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 7643 !TAILQ_EMPTY(&inodedep->id_freeblklst) || 7644 inodedep->id_mkdiradd != NULL || 7645 inodedep->id_nlinkdelta != 0) 7646 return (0); 7647 /* 7648 * Another process might be in initiate_write_inodeblock_ufs[12] 7649 * trying to allocate memory without holding "Softdep Lock". 7650 */ 7651 if ((inodedep->id_state & IOSTARTED) != 0 && 7652 inodedep->id_savedino1 == NULL) 7653 return (0); 7654 7655 if (inodedep->id_state & ONDEPLIST) 7656 LIST_REMOVE(inodedep, id_deps); 7657 inodedep->id_state &= ~ONDEPLIST; 7658 inodedep->id_state |= ALLCOMPLETE; 7659 inodedep->id_bmsafemap = NULL; 7660 if (inodedep->id_state & ONWORKLIST) 7661 WORKLIST_REMOVE(&inodedep->id_list); 7662 if (inodedep->id_savedino1 != NULL) { 7663 free(inodedep->id_savedino1, M_SAVEDINO); 7664 inodedep->id_savedino1 = NULL; 7665 } 7666 if (free_inodedep(inodedep) == 0) 7667 panic("check_inode_unwritten: busy inode"); 7668 return (1); 7669 } 7670 7671 static int 7672 check_inodedep_free(inodedep) 7673 struct inodedep *inodedep; 7674 { 7675 7676 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp)); 7677 if ((inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE || 7678 !LIST_EMPTY(&inodedep->id_dirremhd) || 7679 !LIST_EMPTY(&inodedep->id_pendinghd) || 7680 !LIST_EMPTY(&inodedep->id_bufwait) || 7681 !LIST_EMPTY(&inodedep->id_inowait) || 7682 !TAILQ_EMPTY(&inodedep->id_inoreflst) || 7683 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 7684 !TAILQ_EMPTY(&inodedep->id_newinoupdt) || 7685 !TAILQ_EMPTY(&inodedep->id_extupdt) || 7686 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 7687 !TAILQ_EMPTY(&inodedep->id_freeblklst) || 7688 inodedep->id_mkdiradd != NULL || 7689 inodedep->id_nlinkdelta != 0 || 7690 inodedep->id_savedino1 != NULL) 7691 return (0); 7692 return (1); 7693 } 7694 7695 /* 7696 * Try to free an inodedep structure. Return 1 if it could be freed. 7697 */ 7698 static int 7699 free_inodedep(inodedep) 7700 struct inodedep *inodedep; 7701 { 7702 7703 LOCK_OWNED(VFSTOUFS(inodedep->id_list.wk_mp)); 7704 if ((inodedep->id_state & (ONWORKLIST | UNLINKED)) != 0 || 7705 !check_inodedep_free(inodedep)) 7706 return (0); 7707 if (inodedep->id_state & ONDEPLIST) 7708 LIST_REMOVE(inodedep, id_deps); 7709 LIST_REMOVE(inodedep, id_hash); 7710 WORKITEM_FREE(inodedep, D_INODEDEP); 7711 return (1); 7712 } 7713 7714 /* 7715 * Free the block referenced by a freework structure. The parent freeblks 7716 * structure is released and completed when the final cg bitmap reaches 7717 * the disk. This routine may be freeing a jnewblk which never made it to 7718 * disk in which case we do not have to wait as the operation is undone 7719 * in memory immediately. 7720 */ 7721 static void 7722 freework_freeblock(freework) 7723 struct freework *freework; 7724 { 7725 struct freeblks *freeblks; 7726 struct jnewblk *jnewblk; 7727 struct ufsmount *ump; 7728 struct workhead wkhd; 7729 struct fs *fs; 7730 int bsize; 7731 int needj; 7732 7733 ump = VFSTOUFS(freework->fw_list.wk_mp); 7734 LOCK_OWNED(ump); 7735 /* 7736 * Handle partial truncate separately. 7737 */ 7738 if (freework->fw_indir) { 7739 complete_trunc_indir(freework); 7740 return; 7741 } 7742 freeblks = freework->fw_freeblks; 7743 fs = ump->um_fs; 7744 needj = MOUNTEDSUJ(freeblks->fb_list.wk_mp) != 0; 7745 bsize = lfragtosize(fs, freework->fw_frags); 7746 LIST_INIT(&wkhd); 7747 /* 7748 * DEPCOMPLETE is cleared in indirblk_insert() if the block lives 7749 * on the indirblk hashtable and prevents premature freeing. 7750 */ 7751 freework->fw_state |= DEPCOMPLETE; 7752 /* 7753 * SUJ needs to wait for the segment referencing freed indirect 7754 * blocks to expire so that we know the checker will not confuse 7755 * a re-allocated indirect block with its old contents. 7756 */ 7757 if (needj && freework->fw_lbn <= -NDADDR) 7758 indirblk_insert(freework); 7759 /* 7760 * If we are canceling an existing jnewblk pass it to the free 7761 * routine, otherwise pass the freeblk which will ultimately 7762 * release the freeblks. If we're not journaling, we can just 7763 * free the freeblks immediately. 7764 */ 7765 jnewblk = freework->fw_jnewblk; 7766 if (jnewblk != NULL) { 7767 cancel_jnewblk(jnewblk, &wkhd); 7768 needj = 0; 7769 } else if (needj) { 7770 freework->fw_state |= DELAYEDFREE; 7771 freeblks->fb_cgwait++; 7772 WORKLIST_INSERT(&wkhd, &freework->fw_list); 7773 } 7774 FREE_LOCK(ump); 7775 freeblks_free(ump, freeblks, btodb(bsize)); 7776 CTR4(KTR_SUJ, 7777 "freework_freeblock: ino %d blkno %jd lbn %jd size %ld", 7778 freeblks->fb_inum, freework->fw_blkno, freework->fw_lbn, bsize); 7779 ffs_blkfree(ump, fs, freeblks->fb_devvp, freework->fw_blkno, bsize, 7780 freeblks->fb_inum, freeblks->fb_vtype, &wkhd); 7781 ACQUIRE_LOCK(ump); 7782 /* 7783 * The jnewblk will be discarded and the bits in the map never 7784 * made it to disk. We can immediately free the freeblk. 7785 */ 7786 if (needj == 0) 7787 handle_written_freework(freework); 7788 } 7789 7790 /* 7791 * We enqueue freework items that need processing back on the freeblks and 7792 * add the freeblks to the worklist. This makes it easier to find all work 7793 * required to flush a truncation in process_truncates(). 7794 */ 7795 static void 7796 freework_enqueue(freework) 7797 struct freework *freework; 7798 { 7799 struct freeblks *freeblks; 7800 7801 freeblks = freework->fw_freeblks; 7802 if ((freework->fw_state & INPROGRESS) == 0) 7803 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &freework->fw_list); 7804 if ((freeblks->fb_state & 7805 (ONWORKLIST | INPROGRESS | ALLCOMPLETE)) == ALLCOMPLETE && 7806 LIST_EMPTY(&freeblks->fb_jblkdephd)) 7807 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 7808 } 7809 7810 /* 7811 * Start, continue, or finish the process of freeing an indirect block tree. 7812 * The free operation may be paused at any point with fw_off containing the 7813 * offset to restart from. This enables us to implement some flow control 7814 * for large truncates which may fan out and generate a huge number of 7815 * dependencies. 7816 */ 7817 static void 7818 handle_workitem_indirblk(freework) 7819 struct freework *freework; 7820 { 7821 struct freeblks *freeblks; 7822 struct ufsmount *ump; 7823 struct fs *fs; 7824 7825 freeblks = freework->fw_freeblks; 7826 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7827 fs = ump->um_fs; 7828 if (freework->fw_state & DEPCOMPLETE) { 7829 handle_written_freework(freework); 7830 return; 7831 } 7832 if (freework->fw_off == NINDIR(fs)) { 7833 freework_freeblock(freework); 7834 return; 7835 } 7836 freework->fw_state |= INPROGRESS; 7837 FREE_LOCK(ump); 7838 indir_trunc(freework, fsbtodb(fs, freework->fw_blkno), 7839 freework->fw_lbn); 7840 ACQUIRE_LOCK(ump); 7841 } 7842 7843 /* 7844 * Called when a freework structure attached to a cg buf is written. The 7845 * ref on either the parent or the freeblks structure is released and 7846 * the freeblks is added back to the worklist if there is more work to do. 7847 */ 7848 static void 7849 handle_written_freework(freework) 7850 struct freework *freework; 7851 { 7852 struct freeblks *freeblks; 7853 struct freework *parent; 7854 7855 freeblks = freework->fw_freeblks; 7856 parent = freework->fw_parent; 7857 if (freework->fw_state & DELAYEDFREE) 7858 freeblks->fb_cgwait--; 7859 freework->fw_state |= COMPLETE; 7860 if ((freework->fw_state & ALLCOMPLETE) == ALLCOMPLETE) 7861 WORKITEM_FREE(freework, D_FREEWORK); 7862 if (parent) { 7863 if (--parent->fw_ref == 0) 7864 freework_enqueue(parent); 7865 return; 7866 } 7867 if (--freeblks->fb_ref != 0) 7868 return; 7869 if ((freeblks->fb_state & (ALLCOMPLETE | ONWORKLIST | INPROGRESS)) == 7870 ALLCOMPLETE && LIST_EMPTY(&freeblks->fb_jblkdephd)) 7871 add_to_worklist(&freeblks->fb_list, WK_NODELAY); 7872 } 7873 7874 /* 7875 * This workitem routine performs the block de-allocation. 7876 * The workitem is added to the pending list after the updated 7877 * inode block has been written to disk. As mentioned above, 7878 * checks regarding the number of blocks de-allocated (compared 7879 * to the number of blocks allocated for the file) are also 7880 * performed in this function. 7881 */ 7882 static int 7883 handle_workitem_freeblocks(freeblks, flags) 7884 struct freeblks *freeblks; 7885 int flags; 7886 { 7887 struct freework *freework; 7888 struct newblk *newblk; 7889 struct allocindir *aip; 7890 struct ufsmount *ump; 7891 struct worklist *wk; 7892 7893 KASSERT(LIST_EMPTY(&freeblks->fb_jblkdephd), 7894 ("handle_workitem_freeblocks: Journal entries not written.")); 7895 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7896 ACQUIRE_LOCK(ump); 7897 while ((wk = LIST_FIRST(&freeblks->fb_freeworkhd)) != NULL) { 7898 WORKLIST_REMOVE(wk); 7899 switch (wk->wk_type) { 7900 case D_DIRREM: 7901 wk->wk_state |= COMPLETE; 7902 add_to_worklist(wk, 0); 7903 continue; 7904 7905 case D_ALLOCDIRECT: 7906 free_newblk(WK_NEWBLK(wk)); 7907 continue; 7908 7909 case D_ALLOCINDIR: 7910 aip = WK_ALLOCINDIR(wk); 7911 freework = NULL; 7912 if (aip->ai_state & DELAYEDFREE) { 7913 FREE_LOCK(ump); 7914 freework = newfreework(ump, freeblks, NULL, 7915 aip->ai_lbn, aip->ai_newblkno, 7916 ump->um_fs->fs_frag, 0, 0); 7917 ACQUIRE_LOCK(ump); 7918 } 7919 newblk = WK_NEWBLK(wk); 7920 if (newblk->nb_jnewblk) { 7921 freework->fw_jnewblk = newblk->nb_jnewblk; 7922 newblk->nb_jnewblk->jn_dep = &freework->fw_list; 7923 newblk->nb_jnewblk = NULL; 7924 } 7925 free_newblk(newblk); 7926 continue; 7927 7928 case D_FREEWORK: 7929 freework = WK_FREEWORK(wk); 7930 if (freework->fw_lbn <= -NDADDR) 7931 handle_workitem_indirblk(freework); 7932 else 7933 freework_freeblock(freework); 7934 continue; 7935 default: 7936 panic("handle_workitem_freeblocks: Unknown type %s", 7937 TYPENAME(wk->wk_type)); 7938 } 7939 } 7940 if (freeblks->fb_ref != 0) { 7941 freeblks->fb_state &= ~INPROGRESS; 7942 wake_worklist(&freeblks->fb_list); 7943 freeblks = NULL; 7944 } 7945 FREE_LOCK(ump); 7946 if (freeblks) 7947 return handle_complete_freeblocks(freeblks, flags); 7948 return (0); 7949 } 7950 7951 /* 7952 * Handle completion of block free via truncate. This allows fs_pending 7953 * to track the actual free block count more closely than if we only updated 7954 * it at the end. We must be careful to handle cases where the block count 7955 * on free was incorrect. 7956 */ 7957 static void 7958 freeblks_free(ump, freeblks, blocks) 7959 struct ufsmount *ump; 7960 struct freeblks *freeblks; 7961 int blocks; 7962 { 7963 struct fs *fs; 7964 ufs2_daddr_t remain; 7965 7966 UFS_LOCK(ump); 7967 remain = -freeblks->fb_chkcnt; 7968 freeblks->fb_chkcnt += blocks; 7969 if (remain > 0) { 7970 if (remain < blocks) 7971 blocks = remain; 7972 fs = ump->um_fs; 7973 fs->fs_pendingblocks -= blocks; 7974 } 7975 UFS_UNLOCK(ump); 7976 } 7977 7978 /* 7979 * Once all of the freework workitems are complete we can retire the 7980 * freeblocks dependency and any journal work awaiting completion. This 7981 * can not be called until all other dependencies are stable on disk. 7982 */ 7983 static int 7984 handle_complete_freeblocks(freeblks, flags) 7985 struct freeblks *freeblks; 7986 int flags; 7987 { 7988 struct inodedep *inodedep; 7989 struct inode *ip; 7990 struct vnode *vp; 7991 struct fs *fs; 7992 struct ufsmount *ump; 7993 ufs2_daddr_t spare; 7994 7995 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 7996 fs = ump->um_fs; 7997 flags = LK_EXCLUSIVE | flags; 7998 spare = freeblks->fb_chkcnt; 7999 8000 /* 8001 * If we did not release the expected number of blocks we may have 8002 * to adjust the inode block count here. Only do so if it wasn't 8003 * a truncation to zero and the modrev still matches. 8004 */ 8005 if (spare && freeblks->fb_len != 0) { 8006 if (ffs_vgetf(freeblks->fb_list.wk_mp, freeblks->fb_inum, 8007 flags, &vp, FFSV_FORCEINSMQ) != 0) 8008 return (EBUSY); 8009 ip = VTOI(vp); 8010 if (DIP(ip, i_modrev) == freeblks->fb_modrev) { 8011 DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - spare); 8012 ip->i_flag |= IN_CHANGE; 8013 /* 8014 * We must wait so this happens before the 8015 * journal is reclaimed. 8016 */ 8017 ffs_update(vp, 1); 8018 } 8019 vput(vp); 8020 } 8021 if (spare < 0) { 8022 UFS_LOCK(ump); 8023 fs->fs_pendingblocks += spare; 8024 UFS_UNLOCK(ump); 8025 } 8026 #ifdef QUOTA 8027 /* Handle spare. */ 8028 if (spare) 8029 quotaadj(freeblks->fb_quota, ump, -spare); 8030 quotarele(freeblks->fb_quota); 8031 #endif 8032 ACQUIRE_LOCK(ump); 8033 if (freeblks->fb_state & ONDEPLIST) { 8034 inodedep_lookup(freeblks->fb_list.wk_mp, freeblks->fb_inum, 8035 0, &inodedep); 8036 TAILQ_REMOVE(&inodedep->id_freeblklst, freeblks, fb_next); 8037 freeblks->fb_state &= ~ONDEPLIST; 8038 if (TAILQ_EMPTY(&inodedep->id_freeblklst)) 8039 free_inodedep(inodedep); 8040 } 8041 /* 8042 * All of the freeblock deps must be complete prior to this call 8043 * so it's now safe to complete earlier outstanding journal entries. 8044 */ 8045 handle_jwork(&freeblks->fb_jwork); 8046 WORKITEM_FREE(freeblks, D_FREEBLKS); 8047 FREE_LOCK(ump); 8048 return (0); 8049 } 8050 8051 /* 8052 * Release blocks associated with the freeblks and stored in the indirect 8053 * block dbn. If level is greater than SINGLE, the block is an indirect block 8054 * and recursive calls to indirtrunc must be used to cleanse other indirect 8055 * blocks. 8056 * 8057 * This handles partial and complete truncation of blocks. Partial is noted 8058 * with goingaway == 0. In this case the freework is completed after the 8059 * zero'd indirects are written to disk. For full truncation the freework 8060 * is completed after the block is freed. 8061 */ 8062 static void 8063 indir_trunc(freework, dbn, lbn) 8064 struct freework *freework; 8065 ufs2_daddr_t dbn; 8066 ufs_lbn_t lbn; 8067 { 8068 struct freework *nfreework; 8069 struct workhead wkhd; 8070 struct freeblks *freeblks; 8071 struct buf *bp; 8072 struct fs *fs; 8073 struct indirdep *indirdep; 8074 struct ufsmount *ump; 8075 ufs1_daddr_t *bap1; 8076 ufs2_daddr_t nb, nnb, *bap2; 8077 ufs_lbn_t lbnadd, nlbn; 8078 int i, nblocks, ufs1fmt; 8079 int freedblocks; 8080 int goingaway; 8081 int freedeps; 8082 int needj; 8083 int level; 8084 int cnt; 8085 8086 freeblks = freework->fw_freeblks; 8087 ump = VFSTOUFS(freeblks->fb_list.wk_mp); 8088 fs = ump->um_fs; 8089 /* 8090 * Get buffer of block pointers to be freed. There are three cases: 8091 * 8092 * 1) Partial truncate caches the indirdep pointer in the freework 8093 * which provides us a back copy to the save bp which holds the 8094 * pointers we want to clear. When this completes the zero 8095 * pointers are written to the real copy. 8096 * 2) The indirect is being completely truncated, cancel_indirdep() 8097 * eliminated the real copy and placed the indirdep on the saved 8098 * copy. The indirdep and buf are discarded when this completes. 8099 * 3) The indirect was not in memory, we read a copy off of the disk 8100 * using the devvp and drop and invalidate the buffer when we're 8101 * done. 8102 */ 8103 goingaway = 1; 8104 indirdep = NULL; 8105 if (freework->fw_indir != NULL) { 8106 goingaway = 0; 8107 indirdep = freework->fw_indir; 8108 bp = indirdep->ir_savebp; 8109 if (bp == NULL || bp->b_blkno != dbn) 8110 panic("indir_trunc: Bad saved buf %p blkno %jd", 8111 bp, (intmax_t)dbn); 8112 } else if ((bp = incore(&freeblks->fb_devvp->v_bufobj, dbn)) != NULL) { 8113 /* 8114 * The lock prevents the buf dep list from changing and 8115 * indirects on devvp should only ever have one dependency. 8116 */ 8117 indirdep = WK_INDIRDEP(LIST_FIRST(&bp->b_dep)); 8118 if (indirdep == NULL || (indirdep->ir_state & GOINGAWAY) == 0) 8119 panic("indir_trunc: Bad indirdep %p from buf %p", 8120 indirdep, bp); 8121 } else if (bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 8122 NOCRED, &bp) != 0) { 8123 brelse(bp); 8124 return; 8125 } 8126 ACQUIRE_LOCK(ump); 8127 /* Protects against a race with complete_trunc_indir(). */ 8128 freework->fw_state &= ~INPROGRESS; 8129 /* 8130 * If we have an indirdep we need to enforce the truncation order 8131 * and discard it when it is complete. 8132 */ 8133 if (indirdep) { 8134 if (freework != TAILQ_FIRST(&indirdep->ir_trunc) && 8135 !TAILQ_EMPTY(&indirdep->ir_trunc)) { 8136 /* 8137 * Add the complete truncate to the list on the 8138 * indirdep to enforce in-order processing. 8139 */ 8140 if (freework->fw_indir == NULL) 8141 TAILQ_INSERT_TAIL(&indirdep->ir_trunc, 8142 freework, fw_next); 8143 FREE_LOCK(ump); 8144 return; 8145 } 8146 /* 8147 * If we're goingaway, free the indirdep. Otherwise it will 8148 * linger until the write completes. 8149 */ 8150 if (goingaway) 8151 free_indirdep(indirdep); 8152 } 8153 FREE_LOCK(ump); 8154 /* Initialize pointers depending on block size. */ 8155 if (ump->um_fstype == UFS1) { 8156 bap1 = (ufs1_daddr_t *)bp->b_data; 8157 nb = bap1[freework->fw_off]; 8158 ufs1fmt = 1; 8159 bap2 = NULL; 8160 } else { 8161 bap2 = (ufs2_daddr_t *)bp->b_data; 8162 nb = bap2[freework->fw_off]; 8163 ufs1fmt = 0; 8164 bap1 = NULL; 8165 } 8166 level = lbn_level(lbn); 8167 needj = MOUNTEDSUJ(UFSTOVFS(ump)) != 0; 8168 lbnadd = lbn_offset(fs, level); 8169 nblocks = btodb(fs->fs_bsize); 8170 nfreework = freework; 8171 freedeps = 0; 8172 cnt = 0; 8173 /* 8174 * Reclaim blocks. Traverses into nested indirect levels and 8175 * arranges for the current level to be freed when subordinates 8176 * are free when journaling. 8177 */ 8178 for (i = freework->fw_off; i < NINDIR(fs); i++, nb = nnb) { 8179 if (i != NINDIR(fs) - 1) { 8180 if (ufs1fmt) 8181 nnb = bap1[i+1]; 8182 else 8183 nnb = bap2[i+1]; 8184 } else 8185 nnb = 0; 8186 if (nb == 0) 8187 continue; 8188 cnt++; 8189 if (level != 0) { 8190 nlbn = (lbn + 1) - (i * lbnadd); 8191 if (needj != 0) { 8192 nfreework = newfreework(ump, freeblks, freework, 8193 nlbn, nb, fs->fs_frag, 0, 0); 8194 freedeps++; 8195 } 8196 indir_trunc(nfreework, fsbtodb(fs, nb), nlbn); 8197 } else { 8198 struct freedep *freedep; 8199 8200 /* 8201 * Attempt to aggregate freedep dependencies for 8202 * all blocks being released to the same CG. 8203 */ 8204 LIST_INIT(&wkhd); 8205 if (needj != 0 && 8206 (nnb == 0 || (dtog(fs, nb) != dtog(fs, nnb)))) { 8207 freedep = newfreedep(freework); 8208 WORKLIST_INSERT_UNLOCKED(&wkhd, 8209 &freedep->fd_list); 8210 freedeps++; 8211 } 8212 CTR3(KTR_SUJ, 8213 "indir_trunc: ino %d blkno %jd size %ld", 8214 freeblks->fb_inum, nb, fs->fs_bsize); 8215 ffs_blkfree(ump, fs, freeblks->fb_devvp, nb, 8216 fs->fs_bsize, freeblks->fb_inum, 8217 freeblks->fb_vtype, &wkhd); 8218 } 8219 } 8220 if (goingaway) { 8221 bp->b_flags |= B_INVAL | B_NOCACHE; 8222 brelse(bp); 8223 } 8224 freedblocks = 0; 8225 if (level == 0) 8226 freedblocks = (nblocks * cnt); 8227 if (needj == 0) 8228 freedblocks += nblocks; 8229 freeblks_free(ump, freeblks, freedblocks); 8230 /* 8231 * If we are journaling set up the ref counts and offset so this 8232 * indirect can be completed when its children are free. 8233 */ 8234 if (needj) { 8235 ACQUIRE_LOCK(ump); 8236 freework->fw_off = i; 8237 freework->fw_ref += freedeps; 8238 freework->fw_ref -= NINDIR(fs) + 1; 8239 if (level == 0) 8240 freeblks->fb_cgwait += freedeps; 8241 if (freework->fw_ref == 0) 8242 freework_freeblock(freework); 8243 FREE_LOCK(ump); 8244 return; 8245 } 8246 /* 8247 * If we're not journaling we can free the indirect now. 8248 */ 8249 dbn = dbtofsb(fs, dbn); 8250 CTR3(KTR_SUJ, 8251 "indir_trunc 2: ino %d blkno %jd size %ld", 8252 freeblks->fb_inum, dbn, fs->fs_bsize); 8253 ffs_blkfree(ump, fs, freeblks->fb_devvp, dbn, fs->fs_bsize, 8254 freeblks->fb_inum, freeblks->fb_vtype, NULL); 8255 /* Non SUJ softdep does single-threaded truncations. */ 8256 if (freework->fw_blkno == dbn) { 8257 freework->fw_state |= ALLCOMPLETE; 8258 ACQUIRE_LOCK(ump); 8259 handle_written_freework(freework); 8260 FREE_LOCK(ump); 8261 } 8262 return; 8263 } 8264 8265 /* 8266 * Cancel an allocindir when it is removed via truncation. When bp is not 8267 * NULL the indirect never appeared on disk and is scheduled to be freed 8268 * independently of the indir so we can more easily track journal work. 8269 */ 8270 static void 8271 cancel_allocindir(aip, bp, freeblks, trunc) 8272 struct allocindir *aip; 8273 struct buf *bp; 8274 struct freeblks *freeblks; 8275 int trunc; 8276 { 8277 struct indirdep *indirdep; 8278 struct freefrag *freefrag; 8279 struct newblk *newblk; 8280 8281 newblk = (struct newblk *)aip; 8282 LIST_REMOVE(aip, ai_next); 8283 /* 8284 * We must eliminate the pointer in bp if it must be freed on its 8285 * own due to partial truncate or pending journal work. 8286 */ 8287 if (bp && (trunc || newblk->nb_jnewblk)) { 8288 /* 8289 * Clear the pointer and mark the aip to be freed 8290 * directly if it never existed on disk. 8291 */ 8292 aip->ai_state |= DELAYEDFREE; 8293 indirdep = aip->ai_indirdep; 8294 if (indirdep->ir_state & UFS1FMT) 8295 ((ufs1_daddr_t *)bp->b_data)[aip->ai_offset] = 0; 8296 else 8297 ((ufs2_daddr_t *)bp->b_data)[aip->ai_offset] = 0; 8298 } 8299 /* 8300 * When truncating the previous pointer will be freed via 8301 * savedbp. Eliminate the freefrag which would dup free. 8302 */ 8303 if (trunc && (freefrag = newblk->nb_freefrag) != NULL) { 8304 newblk->nb_freefrag = NULL; 8305 if (freefrag->ff_jdep) 8306 cancel_jfreefrag( 8307 WK_JFREEFRAG(freefrag->ff_jdep)); 8308 jwork_move(&freeblks->fb_jwork, &freefrag->ff_jwork); 8309 WORKITEM_FREE(freefrag, D_FREEFRAG); 8310 } 8311 /* 8312 * If the journal hasn't been written the jnewblk must be passed 8313 * to the call to ffs_blkfree that reclaims the space. We accomplish 8314 * this by leaving the journal dependency on the newblk to be freed 8315 * when a freework is created in handle_workitem_freeblocks(). 8316 */ 8317 cancel_newblk(newblk, NULL, &freeblks->fb_jwork); 8318 WORKLIST_INSERT(&freeblks->fb_freeworkhd, &newblk->nb_list); 8319 } 8320 8321 /* 8322 * Create the mkdir dependencies for . and .. in a new directory. Link them 8323 * in to a newdirblk so any subsequent additions are tracked properly. The 8324 * caller is responsible for adding the mkdir1 dependency to the journal 8325 * and updating id_mkdiradd. This function returns with the per-filesystem 8326 * lock held. 8327 */ 8328 static struct mkdir * 8329 setup_newdir(dap, newinum, dinum, newdirbp, mkdirp) 8330 struct diradd *dap; 8331 ino_t newinum; 8332 ino_t dinum; 8333 struct buf *newdirbp; 8334 struct mkdir **mkdirp; 8335 { 8336 struct newblk *newblk; 8337 struct pagedep *pagedep; 8338 struct inodedep *inodedep; 8339 struct newdirblk *newdirblk; 8340 struct mkdir *mkdir1, *mkdir2; 8341 struct worklist *wk; 8342 struct jaddref *jaddref; 8343 struct ufsmount *ump; 8344 struct mount *mp; 8345 8346 mp = dap->da_list.wk_mp; 8347 ump = VFSTOUFS(mp); 8348 newdirblk = malloc(sizeof(struct newdirblk), M_NEWDIRBLK, 8349 M_SOFTDEP_FLAGS); 8350 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 8351 LIST_INIT(&newdirblk->db_mkdir); 8352 mkdir1 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); 8353 workitem_alloc(&mkdir1->md_list, D_MKDIR, mp); 8354 mkdir1->md_state = ATTACHED | MKDIR_BODY; 8355 mkdir1->md_diradd = dap; 8356 mkdir1->md_jaddref = NULL; 8357 mkdir2 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS); 8358 workitem_alloc(&mkdir2->md_list, D_MKDIR, mp); 8359 mkdir2->md_state = ATTACHED | MKDIR_PARENT; 8360 mkdir2->md_diradd = dap; 8361 mkdir2->md_jaddref = NULL; 8362 if (MOUNTEDSUJ(mp) == 0) { 8363 mkdir1->md_state |= DEPCOMPLETE; 8364 mkdir2->md_state |= DEPCOMPLETE; 8365 } 8366 /* 8367 * Dependency on "." and ".." being written to disk. 8368 */ 8369 mkdir1->md_buf = newdirbp; 8370 ACQUIRE_LOCK(VFSTOUFS(mp)); 8371 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir1, md_mkdirs); 8372 /* 8373 * We must link the pagedep, allocdirect, and newdirblk for 8374 * the initial file page so the pointer to the new directory 8375 * is not written until the directory contents are live and 8376 * any subsequent additions are not marked live until the 8377 * block is reachable via the inode. 8378 */ 8379 if (pagedep_lookup(mp, newdirbp, newinum, 0, 0, &pagedep) == 0) 8380 panic("setup_newdir: lost pagedep"); 8381 LIST_FOREACH(wk, &newdirbp->b_dep, wk_list) 8382 if (wk->wk_type == D_ALLOCDIRECT) 8383 break; 8384 if (wk == NULL) 8385 panic("setup_newdir: lost allocdirect"); 8386 if (pagedep->pd_state & NEWBLOCK) 8387 panic("setup_newdir: NEWBLOCK already set"); 8388 newblk = WK_NEWBLK(wk); 8389 pagedep->pd_state |= NEWBLOCK; 8390 pagedep->pd_newdirblk = newdirblk; 8391 newdirblk->db_pagedep = pagedep; 8392 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list); 8393 WORKLIST_INSERT(&newdirblk->db_mkdir, &mkdir1->md_list); 8394 /* 8395 * Look up the inodedep for the parent directory so that we 8396 * can link mkdir2 into the pending dotdot jaddref or 8397 * the inode write if there is none. If the inode is 8398 * ALLCOMPLETE and no jaddref is present all dependencies have 8399 * been satisfied and mkdir2 can be freed. 8400 */ 8401 inodedep_lookup(mp, dinum, 0, &inodedep); 8402 if (MOUNTEDSUJ(mp)) { 8403 if (inodedep == NULL) 8404 panic("setup_newdir: Lost parent."); 8405 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 8406 inoreflst); 8407 KASSERT(jaddref != NULL && jaddref->ja_parent == newinum && 8408 (jaddref->ja_state & MKDIR_PARENT), 8409 ("setup_newdir: bad dotdot jaddref %p", jaddref)); 8410 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs); 8411 mkdir2->md_jaddref = jaddref; 8412 jaddref->ja_mkdir = mkdir2; 8413 } else if (inodedep == NULL || 8414 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 8415 dap->da_state &= ~MKDIR_PARENT; 8416 WORKITEM_FREE(mkdir2, D_MKDIR); 8417 mkdir2 = NULL; 8418 } else { 8419 LIST_INSERT_HEAD(&ump->softdep_mkdirlisthd, mkdir2, md_mkdirs); 8420 WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir2->md_list); 8421 } 8422 *mkdirp = mkdir2; 8423 8424 return (mkdir1); 8425 } 8426 8427 /* 8428 * Directory entry addition dependencies. 8429 * 8430 * When adding a new directory entry, the inode (with its incremented link 8431 * count) must be written to disk before the directory entry's pointer to it. 8432 * Also, if the inode is newly allocated, the corresponding freemap must be 8433 * updated (on disk) before the directory entry's pointer. These requirements 8434 * are met via undo/redo on the directory entry's pointer, which consists 8435 * simply of the inode number. 8436 * 8437 * As directory entries are added and deleted, the free space within a 8438 * directory block can become fragmented. The ufs filesystem will compact 8439 * a fragmented directory block to make space for a new entry. When this 8440 * occurs, the offsets of previously added entries change. Any "diradd" 8441 * dependency structures corresponding to these entries must be updated with 8442 * the new offsets. 8443 */ 8444 8445 /* 8446 * This routine is called after the in-memory inode's link 8447 * count has been incremented, but before the directory entry's 8448 * pointer to the inode has been set. 8449 */ 8450 int 8451 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk) 8452 struct buf *bp; /* buffer containing directory block */ 8453 struct inode *dp; /* inode for directory */ 8454 off_t diroffset; /* offset of new entry in directory */ 8455 ino_t newinum; /* inode referenced by new directory entry */ 8456 struct buf *newdirbp; /* non-NULL => contents of new mkdir */ 8457 int isnewblk; /* entry is in a newly allocated block */ 8458 { 8459 int offset; /* offset of new entry within directory block */ 8460 ufs_lbn_t lbn; /* block in directory containing new entry */ 8461 struct fs *fs; 8462 struct diradd *dap; 8463 struct newblk *newblk; 8464 struct pagedep *pagedep; 8465 struct inodedep *inodedep; 8466 struct newdirblk *newdirblk; 8467 struct mkdir *mkdir1, *mkdir2; 8468 struct jaddref *jaddref; 8469 struct ufsmount *ump; 8470 struct mount *mp; 8471 int isindir; 8472 8473 mp = ITOVFS(dp); 8474 ump = VFSTOUFS(mp); 8475 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 8476 ("softdep_setup_directory_add called on non-softdep filesystem")); 8477 /* 8478 * Whiteouts have no dependencies. 8479 */ 8480 if (newinum == WINO) { 8481 if (newdirbp != NULL) 8482 bdwrite(newdirbp); 8483 return (0); 8484 } 8485 jaddref = NULL; 8486 mkdir1 = mkdir2 = NULL; 8487 fs = ump->um_fs; 8488 lbn = lblkno(fs, diroffset); 8489 offset = blkoff(fs, diroffset); 8490 dap = malloc(sizeof(struct diradd), M_DIRADD, 8491 M_SOFTDEP_FLAGS|M_ZERO); 8492 workitem_alloc(&dap->da_list, D_DIRADD, mp); 8493 dap->da_offset = offset; 8494 dap->da_newinum = newinum; 8495 dap->da_state = ATTACHED; 8496 LIST_INIT(&dap->da_jwork); 8497 isindir = bp->b_lblkno >= NDADDR; 8498 newdirblk = NULL; 8499 if (isnewblk && 8500 (isindir ? blkoff(fs, diroffset) : fragoff(fs, diroffset)) == 0) { 8501 newdirblk = malloc(sizeof(struct newdirblk), 8502 M_NEWDIRBLK, M_SOFTDEP_FLAGS); 8503 workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp); 8504 LIST_INIT(&newdirblk->db_mkdir); 8505 } 8506 /* 8507 * If we're creating a new directory setup the dependencies and set 8508 * the dap state to wait for them. Otherwise it's COMPLETE and 8509 * we can move on. 8510 */ 8511 if (newdirbp == NULL) { 8512 dap->da_state |= DEPCOMPLETE; 8513 ACQUIRE_LOCK(ump); 8514 } else { 8515 dap->da_state |= MKDIR_BODY | MKDIR_PARENT; 8516 mkdir1 = setup_newdir(dap, newinum, dp->i_number, newdirbp, 8517 &mkdir2); 8518 } 8519 /* 8520 * Link into parent directory pagedep to await its being written. 8521 */ 8522 pagedep_lookup(mp, bp, dp->i_number, lbn, DEPALLOC, &pagedep); 8523 #ifdef DEBUG 8524 if (diradd_lookup(pagedep, offset) != NULL) 8525 panic("softdep_setup_directory_add: %p already at off %d\n", 8526 diradd_lookup(pagedep, offset), offset); 8527 #endif 8528 dap->da_pagedep = pagedep; 8529 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap, 8530 da_pdlist); 8531 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep); 8532 /* 8533 * If we're journaling, link the diradd into the jaddref so it 8534 * may be completed after the journal entry is written. Otherwise, 8535 * link the diradd into its inodedep. If the inode is not yet 8536 * written place it on the bufwait list, otherwise do the post-inode 8537 * write processing to put it on the id_pendinghd list. 8538 */ 8539 if (MOUNTEDSUJ(mp)) { 8540 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 8541 inoreflst); 8542 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 8543 ("softdep_setup_directory_add: bad jaddref %p", jaddref)); 8544 jaddref->ja_diroff = diroffset; 8545 jaddref->ja_diradd = dap; 8546 add_to_journal(&jaddref->ja_list); 8547 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) 8548 diradd_inode_written(dap, inodedep); 8549 else 8550 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 8551 /* 8552 * Add the journal entries for . and .. links now that the primary 8553 * link is written. 8554 */ 8555 if (mkdir1 != NULL && MOUNTEDSUJ(mp)) { 8556 jaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref, 8557 inoreflst, if_deps); 8558 KASSERT(jaddref != NULL && 8559 jaddref->ja_ino == jaddref->ja_parent && 8560 (jaddref->ja_state & MKDIR_BODY), 8561 ("softdep_setup_directory_add: bad dot jaddref %p", 8562 jaddref)); 8563 mkdir1->md_jaddref = jaddref; 8564 jaddref->ja_mkdir = mkdir1; 8565 /* 8566 * It is important that the dotdot journal entry 8567 * is added prior to the dot entry since dot writes 8568 * both the dot and dotdot links. These both must 8569 * be added after the primary link for the journal 8570 * to remain consistent. 8571 */ 8572 add_to_journal(&mkdir2->md_jaddref->ja_list); 8573 add_to_journal(&jaddref->ja_list); 8574 } 8575 /* 8576 * If we are adding a new directory remember this diradd so that if 8577 * we rename it we can keep the dot and dotdot dependencies. If 8578 * we are adding a new name for an inode that has a mkdiradd we 8579 * must be in rename and we have to move the dot and dotdot 8580 * dependencies to this new name. The old name is being orphaned 8581 * soon. 8582 */ 8583 if (mkdir1 != NULL) { 8584 if (inodedep->id_mkdiradd != NULL) 8585 panic("softdep_setup_directory_add: Existing mkdir"); 8586 inodedep->id_mkdiradd = dap; 8587 } else if (inodedep->id_mkdiradd) 8588 merge_diradd(inodedep, dap); 8589 if (newdirblk != NULL) { 8590 /* 8591 * There is nothing to do if we are already tracking 8592 * this block. 8593 */ 8594 if ((pagedep->pd_state & NEWBLOCK) != 0) { 8595 WORKITEM_FREE(newdirblk, D_NEWDIRBLK); 8596 FREE_LOCK(ump); 8597 return (0); 8598 } 8599 if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk) 8600 == 0) 8601 panic("softdep_setup_directory_add: lost entry"); 8602 WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list); 8603 pagedep->pd_state |= NEWBLOCK; 8604 pagedep->pd_newdirblk = newdirblk; 8605 newdirblk->db_pagedep = pagedep; 8606 FREE_LOCK(ump); 8607 /* 8608 * If we extended into an indirect signal direnter to sync. 8609 */ 8610 if (isindir) 8611 return (1); 8612 return (0); 8613 } 8614 FREE_LOCK(ump); 8615 return (0); 8616 } 8617 8618 /* 8619 * This procedure is called to change the offset of a directory 8620 * entry when compacting a directory block which must be owned 8621 * exclusively by the caller. Note that the actual entry movement 8622 * must be done in this procedure to ensure that no I/O completions 8623 * occur while the move is in progress. 8624 */ 8625 void 8626 softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize) 8627 struct buf *bp; /* Buffer holding directory block. */ 8628 struct inode *dp; /* inode for directory */ 8629 caddr_t base; /* address of dp->i_offset */ 8630 caddr_t oldloc; /* address of old directory location */ 8631 caddr_t newloc; /* address of new directory location */ 8632 int entrysize; /* size of directory entry */ 8633 { 8634 int offset, oldoffset, newoffset; 8635 struct pagedep *pagedep; 8636 struct jmvref *jmvref; 8637 struct diradd *dap; 8638 struct direct *de; 8639 struct mount *mp; 8640 struct ufsmount *ump; 8641 ufs_lbn_t lbn; 8642 int flags; 8643 8644 mp = ITOVFS(dp); 8645 ump = VFSTOUFS(mp); 8646 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 8647 ("softdep_change_directoryentry_offset called on " 8648 "non-softdep filesystem")); 8649 de = (struct direct *)oldloc; 8650 jmvref = NULL; 8651 flags = 0; 8652 /* 8653 * Moves are always journaled as it would be too complex to 8654 * determine if any affected adds or removes are present in the 8655 * journal. 8656 */ 8657 if (MOUNTEDSUJ(mp)) { 8658 flags = DEPALLOC; 8659 jmvref = newjmvref(dp, de->d_ino, 8660 dp->i_offset + (oldloc - base), 8661 dp->i_offset + (newloc - base)); 8662 } 8663 lbn = lblkno(ump->um_fs, dp->i_offset); 8664 offset = blkoff(ump->um_fs, dp->i_offset); 8665 oldoffset = offset + (oldloc - base); 8666 newoffset = offset + (newloc - base); 8667 ACQUIRE_LOCK(ump); 8668 if (pagedep_lookup(mp, bp, dp->i_number, lbn, flags, &pagedep) == 0) 8669 goto done; 8670 dap = diradd_lookup(pagedep, oldoffset); 8671 if (dap) { 8672 dap->da_offset = newoffset; 8673 newoffset = DIRADDHASH(newoffset); 8674 oldoffset = DIRADDHASH(oldoffset); 8675 if ((dap->da_state & ALLCOMPLETE) != ALLCOMPLETE && 8676 newoffset != oldoffset) { 8677 LIST_REMOVE(dap, da_pdlist); 8678 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[newoffset], 8679 dap, da_pdlist); 8680 } 8681 } 8682 done: 8683 if (jmvref) { 8684 jmvref->jm_pagedep = pagedep; 8685 LIST_INSERT_HEAD(&pagedep->pd_jmvrefhd, jmvref, jm_deps); 8686 add_to_journal(&jmvref->jm_list); 8687 } 8688 bcopy(oldloc, newloc, entrysize); 8689 FREE_LOCK(ump); 8690 } 8691 8692 /* 8693 * Move the mkdir dependencies and journal work from one diradd to another 8694 * when renaming a directory. The new name must depend on the mkdir deps 8695 * completing as the old name did. Directories can only have one valid link 8696 * at a time so one must be canonical. 8697 */ 8698 static void 8699 merge_diradd(inodedep, newdap) 8700 struct inodedep *inodedep; 8701 struct diradd *newdap; 8702 { 8703 struct diradd *olddap; 8704 struct mkdir *mkdir, *nextmd; 8705 struct ufsmount *ump; 8706 short state; 8707 8708 olddap = inodedep->id_mkdiradd; 8709 inodedep->id_mkdiradd = newdap; 8710 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8711 newdap->da_state &= ~DEPCOMPLETE; 8712 ump = VFSTOUFS(inodedep->id_list.wk_mp); 8713 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 8714 mkdir = nextmd) { 8715 nextmd = LIST_NEXT(mkdir, md_mkdirs); 8716 if (mkdir->md_diradd != olddap) 8717 continue; 8718 mkdir->md_diradd = newdap; 8719 state = mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY); 8720 newdap->da_state |= state; 8721 olddap->da_state &= ~state; 8722 if ((olddap->da_state & 8723 (MKDIR_PARENT | MKDIR_BODY)) == 0) 8724 break; 8725 } 8726 if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 8727 panic("merge_diradd: unfound ref"); 8728 } 8729 /* 8730 * Any mkdir related journal items are not safe to be freed until 8731 * the new name is stable. 8732 */ 8733 jwork_move(&newdap->da_jwork, &olddap->da_jwork); 8734 olddap->da_state |= DEPCOMPLETE; 8735 complete_diradd(olddap); 8736 } 8737 8738 /* 8739 * Move the diradd to the pending list when all diradd dependencies are 8740 * complete. 8741 */ 8742 static void 8743 complete_diradd(dap) 8744 struct diradd *dap; 8745 { 8746 struct pagedep *pagedep; 8747 8748 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 8749 if (dap->da_state & DIRCHG) 8750 pagedep = dap->da_previous->dm_pagedep; 8751 else 8752 pagedep = dap->da_pagedep; 8753 LIST_REMOVE(dap, da_pdlist); 8754 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 8755 } 8756 } 8757 8758 /* 8759 * Cancel a diradd when a dirrem overlaps with it. We must cancel the journal 8760 * add entries and conditonally journal the remove. 8761 */ 8762 static void 8763 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref) 8764 struct diradd *dap; 8765 struct dirrem *dirrem; 8766 struct jremref *jremref; 8767 struct jremref *dotremref; 8768 struct jremref *dotdotremref; 8769 { 8770 struct inodedep *inodedep; 8771 struct jaddref *jaddref; 8772 struct inoref *inoref; 8773 struct ufsmount *ump; 8774 struct mkdir *mkdir; 8775 8776 /* 8777 * If no remove references were allocated we're on a non-journaled 8778 * filesystem and can skip the cancel step. 8779 */ 8780 if (jremref == NULL) { 8781 free_diradd(dap, NULL); 8782 return; 8783 } 8784 /* 8785 * Cancel the primary name an free it if it does not require 8786 * journaling. 8787 */ 8788 if (inodedep_lookup(dap->da_list.wk_mp, dap->da_newinum, 8789 0, &inodedep) != 0) { 8790 /* Abort the addref that reference this diradd. */ 8791 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 8792 if (inoref->if_list.wk_type != D_JADDREF) 8793 continue; 8794 jaddref = (struct jaddref *)inoref; 8795 if (jaddref->ja_diradd != dap) 8796 continue; 8797 if (cancel_jaddref(jaddref, inodedep, 8798 &dirrem->dm_jwork) == 0) { 8799 free_jremref(jremref); 8800 jremref = NULL; 8801 } 8802 break; 8803 } 8804 } 8805 /* 8806 * Cancel subordinate names and free them if they do not require 8807 * journaling. 8808 */ 8809 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8810 ump = VFSTOUFS(dap->da_list.wk_mp); 8811 LIST_FOREACH(mkdir, &ump->softdep_mkdirlisthd, md_mkdirs) { 8812 if (mkdir->md_diradd != dap) 8813 continue; 8814 if ((jaddref = mkdir->md_jaddref) == NULL) 8815 continue; 8816 mkdir->md_jaddref = NULL; 8817 if (mkdir->md_state & MKDIR_PARENT) { 8818 if (cancel_jaddref(jaddref, NULL, 8819 &dirrem->dm_jwork) == 0) { 8820 free_jremref(dotdotremref); 8821 dotdotremref = NULL; 8822 } 8823 } else { 8824 if (cancel_jaddref(jaddref, inodedep, 8825 &dirrem->dm_jwork) == 0) { 8826 free_jremref(dotremref); 8827 dotremref = NULL; 8828 } 8829 } 8830 } 8831 } 8832 8833 if (jremref) 8834 journal_jremref(dirrem, jremref, inodedep); 8835 if (dotremref) 8836 journal_jremref(dirrem, dotremref, inodedep); 8837 if (dotdotremref) 8838 journal_jremref(dirrem, dotdotremref, NULL); 8839 jwork_move(&dirrem->dm_jwork, &dap->da_jwork); 8840 free_diradd(dap, &dirrem->dm_jwork); 8841 } 8842 8843 /* 8844 * Free a diradd dependency structure. This routine must be called 8845 * with splbio interrupts blocked. 8846 */ 8847 static void 8848 free_diradd(dap, wkhd) 8849 struct diradd *dap; 8850 struct workhead *wkhd; 8851 { 8852 struct dirrem *dirrem; 8853 struct pagedep *pagedep; 8854 struct inodedep *inodedep; 8855 struct mkdir *mkdir, *nextmd; 8856 struct ufsmount *ump; 8857 8858 ump = VFSTOUFS(dap->da_list.wk_mp); 8859 LOCK_OWNED(ump); 8860 LIST_REMOVE(dap, da_pdlist); 8861 if (dap->da_state & ONWORKLIST) 8862 WORKLIST_REMOVE(&dap->da_list); 8863 if ((dap->da_state & DIRCHG) == 0) { 8864 pagedep = dap->da_pagedep; 8865 } else { 8866 dirrem = dap->da_previous; 8867 pagedep = dirrem->dm_pagedep; 8868 dirrem->dm_dirinum = pagedep->pd_ino; 8869 dirrem->dm_state |= COMPLETE; 8870 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 8871 add_to_worklist(&dirrem->dm_list, 0); 8872 } 8873 if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum, 8874 0, &inodedep) != 0) 8875 if (inodedep->id_mkdiradd == dap) 8876 inodedep->id_mkdiradd = NULL; 8877 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) { 8878 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 8879 mkdir = nextmd) { 8880 nextmd = LIST_NEXT(mkdir, md_mkdirs); 8881 if (mkdir->md_diradd != dap) 8882 continue; 8883 dap->da_state &= 8884 ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)); 8885 LIST_REMOVE(mkdir, md_mkdirs); 8886 if (mkdir->md_state & ONWORKLIST) 8887 WORKLIST_REMOVE(&mkdir->md_list); 8888 if (mkdir->md_jaddref != NULL) 8889 panic("free_diradd: Unexpected jaddref"); 8890 WORKITEM_FREE(mkdir, D_MKDIR); 8891 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) 8892 break; 8893 } 8894 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) 8895 panic("free_diradd: unfound ref"); 8896 } 8897 if (inodedep) 8898 free_inodedep(inodedep); 8899 /* 8900 * Free any journal segments waiting for the directory write. 8901 */ 8902 handle_jwork(&dap->da_jwork); 8903 WORKITEM_FREE(dap, D_DIRADD); 8904 } 8905 8906 /* 8907 * Directory entry removal dependencies. 8908 * 8909 * When removing a directory entry, the entry's inode pointer must be 8910 * zero'ed on disk before the corresponding inode's link count is decremented 8911 * (possibly freeing the inode for re-use). This dependency is handled by 8912 * updating the directory entry but delaying the inode count reduction until 8913 * after the directory block has been written to disk. After this point, the 8914 * inode count can be decremented whenever it is convenient. 8915 */ 8916 8917 /* 8918 * This routine should be called immediately after removing 8919 * a directory entry. The inode's link count should not be 8920 * decremented by the calling procedure -- the soft updates 8921 * code will do this task when it is safe. 8922 */ 8923 void 8924 softdep_setup_remove(bp, dp, ip, isrmdir) 8925 struct buf *bp; /* buffer containing directory block */ 8926 struct inode *dp; /* inode for the directory being modified */ 8927 struct inode *ip; /* inode for directory entry being removed */ 8928 int isrmdir; /* indicates if doing RMDIR */ 8929 { 8930 struct dirrem *dirrem, *prevdirrem; 8931 struct inodedep *inodedep; 8932 struct ufsmount *ump; 8933 int direct; 8934 8935 ump = ITOUMP(ip); 8936 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 8937 ("softdep_setup_remove called on non-softdep filesystem")); 8938 /* 8939 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK. We want 8940 * newdirrem() to setup the full directory remove which requires 8941 * isrmdir > 1. 8942 */ 8943 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 8944 /* 8945 * Add the dirrem to the inodedep's pending remove list for quick 8946 * discovery later. 8947 */ 8948 if (inodedep_lookup(UFSTOVFS(ump), ip->i_number, 0, &inodedep) == 0) 8949 panic("softdep_setup_remove: Lost inodedep."); 8950 KASSERT((inodedep->id_state & UNLINKED) == 0, ("inode unlinked")); 8951 dirrem->dm_state |= ONDEPLIST; 8952 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 8953 8954 /* 8955 * If the COMPLETE flag is clear, then there were no active 8956 * entries and we want to roll back to a zeroed entry until 8957 * the new inode is committed to disk. If the COMPLETE flag is 8958 * set then we have deleted an entry that never made it to 8959 * disk. If the entry we deleted resulted from a name change, 8960 * then the old name still resides on disk. We cannot delete 8961 * its inode (returned to us in prevdirrem) until the zeroed 8962 * directory entry gets to disk. The new inode has never been 8963 * referenced on the disk, so can be deleted immediately. 8964 */ 8965 if ((dirrem->dm_state & COMPLETE) == 0) { 8966 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem, 8967 dm_next); 8968 FREE_LOCK(ump); 8969 } else { 8970 if (prevdirrem != NULL) 8971 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, 8972 prevdirrem, dm_next); 8973 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino; 8974 direct = LIST_EMPTY(&dirrem->dm_jremrefhd); 8975 FREE_LOCK(ump); 8976 if (direct) 8977 handle_workitem_remove(dirrem, 0); 8978 } 8979 } 8980 8981 /* 8982 * Check for an entry matching 'offset' on both the pd_dirraddhd list and the 8983 * pd_pendinghd list of a pagedep. 8984 */ 8985 static struct diradd * 8986 diradd_lookup(pagedep, offset) 8987 struct pagedep *pagedep; 8988 int offset; 8989 { 8990 struct diradd *dap; 8991 8992 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist) 8993 if (dap->da_offset == offset) 8994 return (dap); 8995 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) 8996 if (dap->da_offset == offset) 8997 return (dap); 8998 return (NULL); 8999 } 9000 9001 /* 9002 * Search for a .. diradd dependency in a directory that is being removed. 9003 * If the directory was renamed to a new parent we have a diradd rather 9004 * than a mkdir for the .. entry. We need to cancel it now before 9005 * it is found in truncate(). 9006 */ 9007 static struct jremref * 9008 cancel_diradd_dotdot(ip, dirrem, jremref) 9009 struct inode *ip; 9010 struct dirrem *dirrem; 9011 struct jremref *jremref; 9012 { 9013 struct pagedep *pagedep; 9014 struct diradd *dap; 9015 struct worklist *wk; 9016 9017 if (pagedep_lookup(ITOVFS(ip), NULL, ip->i_number, 0, 0, &pagedep) == 0) 9018 return (jremref); 9019 dap = diradd_lookup(pagedep, DOTDOT_OFFSET); 9020 if (dap == NULL) 9021 return (jremref); 9022 cancel_diradd(dap, dirrem, jremref, NULL, NULL); 9023 /* 9024 * Mark any journal work as belonging to the parent so it is freed 9025 * with the .. reference. 9026 */ 9027 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list) 9028 wk->wk_state |= MKDIR_PARENT; 9029 return (NULL); 9030 } 9031 9032 /* 9033 * Cancel the MKDIR_PARENT mkdir component of a diradd when we're going to 9034 * replace it with a dirrem/diradd pair as a result of re-parenting a 9035 * directory. This ensures that we don't simultaneously have a mkdir and 9036 * a diradd for the same .. entry. 9037 */ 9038 static struct jremref * 9039 cancel_mkdir_dotdot(ip, dirrem, jremref) 9040 struct inode *ip; 9041 struct dirrem *dirrem; 9042 struct jremref *jremref; 9043 { 9044 struct inodedep *inodedep; 9045 struct jaddref *jaddref; 9046 struct ufsmount *ump; 9047 struct mkdir *mkdir; 9048 struct diradd *dap; 9049 struct mount *mp; 9050 9051 mp = ITOVFS(ip); 9052 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 9053 return (jremref); 9054 dap = inodedep->id_mkdiradd; 9055 if (dap == NULL || (dap->da_state & MKDIR_PARENT) == 0) 9056 return (jremref); 9057 ump = VFSTOUFS(inodedep->id_list.wk_mp); 9058 for (mkdir = LIST_FIRST(&ump->softdep_mkdirlisthd); mkdir; 9059 mkdir = LIST_NEXT(mkdir, md_mkdirs)) 9060 if (mkdir->md_diradd == dap && mkdir->md_state & MKDIR_PARENT) 9061 break; 9062 if (mkdir == NULL) 9063 panic("cancel_mkdir_dotdot: Unable to find mkdir\n"); 9064 if ((jaddref = mkdir->md_jaddref) != NULL) { 9065 mkdir->md_jaddref = NULL; 9066 jaddref->ja_state &= ~MKDIR_PARENT; 9067 if (inodedep_lookup(mp, jaddref->ja_ino, 0, &inodedep) == 0) 9068 panic("cancel_mkdir_dotdot: Lost parent inodedep"); 9069 if (cancel_jaddref(jaddref, inodedep, &dirrem->dm_jwork)) { 9070 journal_jremref(dirrem, jremref, inodedep); 9071 jremref = NULL; 9072 } 9073 } 9074 if (mkdir->md_state & ONWORKLIST) 9075 WORKLIST_REMOVE(&mkdir->md_list); 9076 mkdir->md_state |= ALLCOMPLETE; 9077 complete_mkdir(mkdir); 9078 return (jremref); 9079 } 9080 9081 static void 9082 journal_jremref(dirrem, jremref, inodedep) 9083 struct dirrem *dirrem; 9084 struct jremref *jremref; 9085 struct inodedep *inodedep; 9086 { 9087 9088 if (inodedep == NULL) 9089 if (inodedep_lookup(jremref->jr_list.wk_mp, 9090 jremref->jr_ref.if_ino, 0, &inodedep) == 0) 9091 panic("journal_jremref: Lost inodedep"); 9092 LIST_INSERT_HEAD(&dirrem->dm_jremrefhd, jremref, jr_deps); 9093 TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps); 9094 add_to_journal(&jremref->jr_list); 9095 } 9096 9097 static void 9098 dirrem_journal(dirrem, jremref, dotremref, dotdotremref) 9099 struct dirrem *dirrem; 9100 struct jremref *jremref; 9101 struct jremref *dotremref; 9102 struct jremref *dotdotremref; 9103 { 9104 struct inodedep *inodedep; 9105 9106 9107 if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 0, 9108 &inodedep) == 0) 9109 panic("dirrem_journal: Lost inodedep"); 9110 journal_jremref(dirrem, jremref, inodedep); 9111 if (dotremref) 9112 journal_jremref(dirrem, dotremref, inodedep); 9113 if (dotdotremref) 9114 journal_jremref(dirrem, dotdotremref, NULL); 9115 } 9116 9117 /* 9118 * Allocate a new dirrem if appropriate and return it along with 9119 * its associated pagedep. Called without a lock, returns with lock. 9120 */ 9121 static struct dirrem * 9122 newdirrem(bp, dp, ip, isrmdir, prevdirremp) 9123 struct buf *bp; /* buffer containing directory block */ 9124 struct inode *dp; /* inode for the directory being modified */ 9125 struct inode *ip; /* inode for directory entry being removed */ 9126 int isrmdir; /* indicates if doing RMDIR */ 9127 struct dirrem **prevdirremp; /* previously referenced inode, if any */ 9128 { 9129 int offset; 9130 ufs_lbn_t lbn; 9131 struct diradd *dap; 9132 struct dirrem *dirrem; 9133 struct pagedep *pagedep; 9134 struct jremref *jremref; 9135 struct jremref *dotremref; 9136 struct jremref *dotdotremref; 9137 struct vnode *dvp; 9138 struct ufsmount *ump; 9139 9140 /* 9141 * Whiteouts have no deletion dependencies. 9142 */ 9143 if (ip == NULL) 9144 panic("newdirrem: whiteout"); 9145 dvp = ITOV(dp); 9146 ump = ITOUMP(dp); 9147 9148 /* 9149 * If the system is over its limit and our filesystem is 9150 * responsible for more than our share of that usage and 9151 * we are not a snapshot, request some inodedep cleanup. 9152 * Limiting the number of dirrem structures will also limit 9153 * the number of freefile and freeblks structures. 9154 */ 9155 ACQUIRE_LOCK(ump); 9156 if (!IS_SNAPSHOT(ip) && softdep_excess_items(ump, D_DIRREM)) 9157 schedule_cleanup(UFSTOVFS(ump)); 9158 else 9159 FREE_LOCK(ump); 9160 dirrem = malloc(sizeof(struct dirrem), M_DIRREM, M_SOFTDEP_FLAGS | 9161 M_ZERO); 9162 workitem_alloc(&dirrem->dm_list, D_DIRREM, dvp->v_mount); 9163 LIST_INIT(&dirrem->dm_jremrefhd); 9164 LIST_INIT(&dirrem->dm_jwork); 9165 dirrem->dm_state = isrmdir ? RMDIR : 0; 9166 dirrem->dm_oldinum = ip->i_number; 9167 *prevdirremp = NULL; 9168 /* 9169 * Allocate remove reference structures to track journal write 9170 * dependencies. We will always have one for the link and 9171 * when doing directories we will always have one more for dot. 9172 * When renaming a directory we skip the dotdot link change so 9173 * this is not needed. 9174 */ 9175 jremref = dotremref = dotdotremref = NULL; 9176 if (DOINGSUJ(dvp)) { 9177 if (isrmdir) { 9178 jremref = newjremref(dirrem, dp, ip, dp->i_offset, 9179 ip->i_effnlink + 2); 9180 dotremref = newjremref(dirrem, ip, ip, DOT_OFFSET, 9181 ip->i_effnlink + 1); 9182 dotdotremref = newjremref(dirrem, ip, dp, DOTDOT_OFFSET, 9183 dp->i_effnlink + 1); 9184 dotdotremref->jr_state |= MKDIR_PARENT; 9185 } else 9186 jremref = newjremref(dirrem, dp, ip, dp->i_offset, 9187 ip->i_effnlink + 1); 9188 } 9189 ACQUIRE_LOCK(ump); 9190 lbn = lblkno(ump->um_fs, dp->i_offset); 9191 offset = blkoff(ump->um_fs, dp->i_offset); 9192 pagedep_lookup(UFSTOVFS(ump), bp, dp->i_number, lbn, DEPALLOC, 9193 &pagedep); 9194 dirrem->dm_pagedep = pagedep; 9195 dirrem->dm_offset = offset; 9196 /* 9197 * If we're renaming a .. link to a new directory, cancel any 9198 * existing MKDIR_PARENT mkdir. If it has already been canceled 9199 * the jremref is preserved for any potential diradd in this 9200 * location. This can not coincide with a rmdir. 9201 */ 9202 if (dp->i_offset == DOTDOT_OFFSET) { 9203 if (isrmdir) 9204 panic("newdirrem: .. directory change during remove?"); 9205 jremref = cancel_mkdir_dotdot(dp, dirrem, jremref); 9206 } 9207 /* 9208 * If we're removing a directory search for the .. dependency now and 9209 * cancel it. Any pending journal work will be added to the dirrem 9210 * to be completed when the workitem remove completes. 9211 */ 9212 if (isrmdir) 9213 dotdotremref = cancel_diradd_dotdot(ip, dirrem, dotdotremref); 9214 /* 9215 * Check for a diradd dependency for the same directory entry. 9216 * If present, then both dependencies become obsolete and can 9217 * be de-allocated. 9218 */ 9219 dap = diradd_lookup(pagedep, offset); 9220 if (dap == NULL) { 9221 /* 9222 * Link the jremref structures into the dirrem so they are 9223 * written prior to the pagedep. 9224 */ 9225 if (jremref) 9226 dirrem_journal(dirrem, jremref, dotremref, 9227 dotdotremref); 9228 return (dirrem); 9229 } 9230 /* 9231 * Must be ATTACHED at this point. 9232 */ 9233 if ((dap->da_state & ATTACHED) == 0) 9234 panic("newdirrem: not ATTACHED"); 9235 if (dap->da_newinum != ip->i_number) 9236 panic("newdirrem: inum %ju should be %ju", 9237 (uintmax_t)ip->i_number, (uintmax_t)dap->da_newinum); 9238 /* 9239 * If we are deleting a changed name that never made it to disk, 9240 * then return the dirrem describing the previous inode (which 9241 * represents the inode currently referenced from this entry on disk). 9242 */ 9243 if ((dap->da_state & DIRCHG) != 0) { 9244 *prevdirremp = dap->da_previous; 9245 dap->da_state &= ~DIRCHG; 9246 dap->da_pagedep = pagedep; 9247 } 9248 /* 9249 * We are deleting an entry that never made it to disk. 9250 * Mark it COMPLETE so we can delete its inode immediately. 9251 */ 9252 dirrem->dm_state |= COMPLETE; 9253 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref); 9254 #ifdef SUJ_DEBUG 9255 if (isrmdir == 0) { 9256 struct worklist *wk; 9257 9258 LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list) 9259 if (wk->wk_state & (MKDIR_BODY | MKDIR_PARENT)) 9260 panic("bad wk %p (0x%X)\n", wk, wk->wk_state); 9261 } 9262 #endif 9263 9264 return (dirrem); 9265 } 9266 9267 /* 9268 * Directory entry change dependencies. 9269 * 9270 * Changing an existing directory entry requires that an add operation 9271 * be completed first followed by a deletion. The semantics for the addition 9272 * are identical to the description of adding a new entry above except 9273 * that the rollback is to the old inode number rather than zero. Once 9274 * the addition dependency is completed, the removal is done as described 9275 * in the removal routine above. 9276 */ 9277 9278 /* 9279 * This routine should be called immediately after changing 9280 * a directory entry. The inode's link count should not be 9281 * decremented by the calling procedure -- the soft updates 9282 * code will perform this task when it is safe. 9283 */ 9284 void 9285 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir) 9286 struct buf *bp; /* buffer containing directory block */ 9287 struct inode *dp; /* inode for the directory being modified */ 9288 struct inode *ip; /* inode for directory entry being removed */ 9289 ino_t newinum; /* new inode number for changed entry */ 9290 int isrmdir; /* indicates if doing RMDIR */ 9291 { 9292 int offset; 9293 struct diradd *dap = NULL; 9294 struct dirrem *dirrem, *prevdirrem; 9295 struct pagedep *pagedep; 9296 struct inodedep *inodedep; 9297 struct jaddref *jaddref; 9298 struct mount *mp; 9299 struct ufsmount *ump; 9300 9301 mp = ITOVFS(dp); 9302 ump = VFSTOUFS(mp); 9303 offset = blkoff(ump->um_fs, dp->i_offset); 9304 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 9305 ("softdep_setup_directory_change called on non-softdep filesystem")); 9306 9307 /* 9308 * Whiteouts do not need diradd dependencies. 9309 */ 9310 if (newinum != WINO) { 9311 dap = malloc(sizeof(struct diradd), 9312 M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO); 9313 workitem_alloc(&dap->da_list, D_DIRADD, mp); 9314 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE; 9315 dap->da_offset = offset; 9316 dap->da_newinum = newinum; 9317 LIST_INIT(&dap->da_jwork); 9318 } 9319 9320 /* 9321 * Allocate a new dirrem and ACQUIRE_LOCK. 9322 */ 9323 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem); 9324 pagedep = dirrem->dm_pagedep; 9325 /* 9326 * The possible values for isrmdir: 9327 * 0 - non-directory file rename 9328 * 1 - directory rename within same directory 9329 * inum - directory rename to new directory of given inode number 9330 * When renaming to a new directory, we are both deleting and 9331 * creating a new directory entry, so the link count on the new 9332 * directory should not change. Thus we do not need the followup 9333 * dirrem which is usually done in handle_workitem_remove. We set 9334 * the DIRCHG flag to tell handle_workitem_remove to skip the 9335 * followup dirrem. 9336 */ 9337 if (isrmdir > 1) 9338 dirrem->dm_state |= DIRCHG; 9339 9340 /* 9341 * Whiteouts have no additional dependencies, 9342 * so just put the dirrem on the correct list. 9343 */ 9344 if (newinum == WINO) { 9345 if ((dirrem->dm_state & COMPLETE) == 0) { 9346 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem, 9347 dm_next); 9348 } else { 9349 dirrem->dm_dirinum = pagedep->pd_ino; 9350 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 9351 add_to_worklist(&dirrem->dm_list, 0); 9352 } 9353 FREE_LOCK(ump); 9354 return; 9355 } 9356 /* 9357 * Add the dirrem to the inodedep's pending remove list for quick 9358 * discovery later. A valid nlinkdelta ensures that this lookup 9359 * will not fail. 9360 */ 9361 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 9362 panic("softdep_setup_directory_change: Lost inodedep."); 9363 dirrem->dm_state |= ONDEPLIST; 9364 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 9365 9366 /* 9367 * If the COMPLETE flag is clear, then there were no active 9368 * entries and we want to roll back to the previous inode until 9369 * the new inode is committed to disk. If the COMPLETE flag is 9370 * set, then we have deleted an entry that never made it to disk. 9371 * If the entry we deleted resulted from a name change, then the old 9372 * inode reference still resides on disk. Any rollback that we do 9373 * needs to be to that old inode (returned to us in prevdirrem). If 9374 * the entry we deleted resulted from a create, then there is 9375 * no entry on the disk, so we want to roll back to zero rather 9376 * than the uncommitted inode. In either of the COMPLETE cases we 9377 * want to immediately free the unwritten and unreferenced inode. 9378 */ 9379 if ((dirrem->dm_state & COMPLETE) == 0) { 9380 dap->da_previous = dirrem; 9381 } else { 9382 if (prevdirrem != NULL) { 9383 dap->da_previous = prevdirrem; 9384 } else { 9385 dap->da_state &= ~DIRCHG; 9386 dap->da_pagedep = pagedep; 9387 } 9388 dirrem->dm_dirinum = pagedep->pd_ino; 9389 if (LIST_EMPTY(&dirrem->dm_jremrefhd)) 9390 add_to_worklist(&dirrem->dm_list, 0); 9391 } 9392 /* 9393 * Lookup the jaddref for this journal entry. We must finish 9394 * initializing it and make the diradd write dependent on it. 9395 * If we're not journaling, put it on the id_bufwait list if the 9396 * inode is not yet written. If it is written, do the post-inode 9397 * write processing to put it on the id_pendinghd list. 9398 */ 9399 inodedep_lookup(mp, newinum, DEPALLOC, &inodedep); 9400 if (MOUNTEDSUJ(mp)) { 9401 jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst, 9402 inoreflst); 9403 KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number, 9404 ("softdep_setup_directory_change: bad jaddref %p", 9405 jaddref)); 9406 jaddref->ja_diroff = dp->i_offset; 9407 jaddref->ja_diradd = dap; 9408 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 9409 dap, da_pdlist); 9410 add_to_journal(&jaddref->ja_list); 9411 } else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) { 9412 dap->da_state |= COMPLETE; 9413 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist); 9414 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 9415 } else { 9416 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], 9417 dap, da_pdlist); 9418 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list); 9419 } 9420 /* 9421 * If we're making a new name for a directory that has not been 9422 * committed when need to move the dot and dotdot references to 9423 * this new name. 9424 */ 9425 if (inodedep->id_mkdiradd && dp->i_offset != DOTDOT_OFFSET) 9426 merge_diradd(inodedep, dap); 9427 FREE_LOCK(ump); 9428 } 9429 9430 /* 9431 * Called whenever the link count on an inode is changed. 9432 * It creates an inode dependency so that the new reference(s) 9433 * to the inode cannot be committed to disk until the updated 9434 * inode has been written. 9435 */ 9436 void 9437 softdep_change_linkcnt(ip) 9438 struct inode *ip; /* the inode with the increased link count */ 9439 { 9440 struct inodedep *inodedep; 9441 struct ufsmount *ump; 9442 9443 ump = ITOUMP(ip); 9444 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 9445 ("softdep_change_linkcnt called on non-softdep filesystem")); 9446 ACQUIRE_LOCK(ump); 9447 inodedep_lookup(UFSTOVFS(ump), ip->i_number, DEPALLOC, &inodedep); 9448 if (ip->i_nlink < ip->i_effnlink) 9449 panic("softdep_change_linkcnt: bad delta"); 9450 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9451 FREE_LOCK(ump); 9452 } 9453 9454 /* 9455 * Attach a sbdep dependency to the superblock buf so that we can keep 9456 * track of the head of the linked list of referenced but unlinked inodes. 9457 */ 9458 void 9459 softdep_setup_sbupdate(ump, fs, bp) 9460 struct ufsmount *ump; 9461 struct fs *fs; 9462 struct buf *bp; 9463 { 9464 struct sbdep *sbdep; 9465 struct worklist *wk; 9466 9467 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 9468 ("softdep_setup_sbupdate called on non-softdep filesystem")); 9469 LIST_FOREACH(wk, &bp->b_dep, wk_list) 9470 if (wk->wk_type == D_SBDEP) 9471 break; 9472 if (wk != NULL) 9473 return; 9474 sbdep = malloc(sizeof(struct sbdep), M_SBDEP, M_SOFTDEP_FLAGS); 9475 workitem_alloc(&sbdep->sb_list, D_SBDEP, UFSTOVFS(ump)); 9476 sbdep->sb_fs = fs; 9477 sbdep->sb_ump = ump; 9478 ACQUIRE_LOCK(ump); 9479 WORKLIST_INSERT(&bp->b_dep, &sbdep->sb_list); 9480 FREE_LOCK(ump); 9481 } 9482 9483 /* 9484 * Return the first unlinked inodedep which is ready to be the head of the 9485 * list. The inodedep and all those after it must have valid next pointers. 9486 */ 9487 static struct inodedep * 9488 first_unlinked_inodedep(ump) 9489 struct ufsmount *ump; 9490 { 9491 struct inodedep *inodedep; 9492 struct inodedep *idp; 9493 9494 LOCK_OWNED(ump); 9495 for (inodedep = TAILQ_LAST(&ump->softdep_unlinked, inodedeplst); 9496 inodedep; inodedep = idp) { 9497 if ((inodedep->id_state & UNLINKNEXT) == 0) 9498 return (NULL); 9499 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9500 if (idp == NULL || (idp->id_state & UNLINKNEXT) == 0) 9501 break; 9502 if ((inodedep->id_state & UNLINKPREV) == 0) 9503 break; 9504 } 9505 return (inodedep); 9506 } 9507 9508 /* 9509 * Set the sujfree unlinked head pointer prior to writing a superblock. 9510 */ 9511 static void 9512 initiate_write_sbdep(sbdep) 9513 struct sbdep *sbdep; 9514 { 9515 struct inodedep *inodedep; 9516 struct fs *bpfs; 9517 struct fs *fs; 9518 9519 bpfs = sbdep->sb_fs; 9520 fs = sbdep->sb_ump->um_fs; 9521 inodedep = first_unlinked_inodedep(sbdep->sb_ump); 9522 if (inodedep) { 9523 fs->fs_sujfree = inodedep->id_ino; 9524 inodedep->id_state |= UNLINKPREV; 9525 } else 9526 fs->fs_sujfree = 0; 9527 bpfs->fs_sujfree = fs->fs_sujfree; 9528 } 9529 9530 /* 9531 * After a superblock is written determine whether it must be written again 9532 * due to a changing unlinked list head. 9533 */ 9534 static int 9535 handle_written_sbdep(sbdep, bp) 9536 struct sbdep *sbdep; 9537 struct buf *bp; 9538 { 9539 struct inodedep *inodedep; 9540 struct fs *fs; 9541 9542 LOCK_OWNED(sbdep->sb_ump); 9543 fs = sbdep->sb_fs; 9544 /* 9545 * If the superblock doesn't match the in-memory list start over. 9546 */ 9547 inodedep = first_unlinked_inodedep(sbdep->sb_ump); 9548 if ((inodedep && fs->fs_sujfree != inodedep->id_ino) || 9549 (inodedep == NULL && fs->fs_sujfree != 0)) { 9550 bdirty(bp); 9551 return (1); 9552 } 9553 WORKITEM_FREE(sbdep, D_SBDEP); 9554 if (fs->fs_sujfree == 0) 9555 return (0); 9556 /* 9557 * Now that we have a record of this inode in stable store allow it 9558 * to be written to free up pending work. Inodes may see a lot of 9559 * write activity after they are unlinked which we must not hold up. 9560 */ 9561 for (; inodedep != NULL; inodedep = TAILQ_NEXT(inodedep, id_unlinked)) { 9562 if ((inodedep->id_state & UNLINKLINKS) != UNLINKLINKS) 9563 panic("handle_written_sbdep: Bad inodedep %p (0x%X)", 9564 inodedep, inodedep->id_state); 9565 if (inodedep->id_state & UNLINKONLIST) 9566 break; 9567 inodedep->id_state |= DEPCOMPLETE | UNLINKONLIST; 9568 } 9569 9570 return (0); 9571 } 9572 9573 /* 9574 * Mark an inodedep as unlinked and insert it into the in-memory unlinked list. 9575 */ 9576 static void 9577 unlinked_inodedep(mp, inodedep) 9578 struct mount *mp; 9579 struct inodedep *inodedep; 9580 { 9581 struct ufsmount *ump; 9582 9583 ump = VFSTOUFS(mp); 9584 LOCK_OWNED(ump); 9585 if (MOUNTEDSUJ(mp) == 0) 9586 return; 9587 ump->um_fs->fs_fmod = 1; 9588 if (inodedep->id_state & UNLINKED) 9589 panic("unlinked_inodedep: %p already unlinked\n", inodedep); 9590 inodedep->id_state |= UNLINKED; 9591 TAILQ_INSERT_HEAD(&ump->softdep_unlinked, inodedep, id_unlinked); 9592 } 9593 9594 /* 9595 * Remove an inodedep from the unlinked inodedep list. This may require 9596 * disk writes if the inode has made it that far. 9597 */ 9598 static void 9599 clear_unlinked_inodedep(inodedep) 9600 struct inodedep *inodedep; 9601 { 9602 struct ufsmount *ump; 9603 struct inodedep *idp; 9604 struct inodedep *idn; 9605 struct fs *fs; 9606 struct buf *bp; 9607 ino_t ino; 9608 ino_t nino; 9609 ino_t pino; 9610 int error; 9611 9612 ump = VFSTOUFS(inodedep->id_list.wk_mp); 9613 fs = ump->um_fs; 9614 ino = inodedep->id_ino; 9615 error = 0; 9616 for (;;) { 9617 LOCK_OWNED(ump); 9618 KASSERT((inodedep->id_state & UNLINKED) != 0, 9619 ("clear_unlinked_inodedep: inodedep %p not unlinked", 9620 inodedep)); 9621 /* 9622 * If nothing has yet been written simply remove us from 9623 * the in memory list and return. This is the most common 9624 * case where handle_workitem_remove() loses the final 9625 * reference. 9626 */ 9627 if ((inodedep->id_state & UNLINKLINKS) == 0) 9628 break; 9629 /* 9630 * If we have a NEXT pointer and no PREV pointer we can simply 9631 * clear NEXT's PREV and remove ourselves from the list. Be 9632 * careful not to clear PREV if the superblock points at 9633 * next as well. 9634 */ 9635 idn = TAILQ_NEXT(inodedep, id_unlinked); 9636 if ((inodedep->id_state & UNLINKLINKS) == UNLINKNEXT) { 9637 if (idn && fs->fs_sujfree != idn->id_ino) 9638 idn->id_state &= ~UNLINKPREV; 9639 break; 9640 } 9641 /* 9642 * Here we have an inodedep which is actually linked into 9643 * the list. We must remove it by forcing a write to the 9644 * link before us, whether it be the superblock or an inode. 9645 * Unfortunately the list may change while we're waiting 9646 * on the buf lock for either resource so we must loop until 9647 * we lock the right one. If both the superblock and an 9648 * inode point to this inode we must clear the inode first 9649 * followed by the superblock. 9650 */ 9651 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9652 pino = 0; 9653 if (idp && (idp->id_state & UNLINKNEXT)) 9654 pino = idp->id_ino; 9655 FREE_LOCK(ump); 9656 if (pino == 0) { 9657 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc), 9658 (int)fs->fs_sbsize, 0, 0, 0); 9659 } else { 9660 error = bread(ump->um_devvp, 9661 fsbtodb(fs, ino_to_fsba(fs, pino)), 9662 (int)fs->fs_bsize, NOCRED, &bp); 9663 if (error) 9664 brelse(bp); 9665 } 9666 ACQUIRE_LOCK(ump); 9667 if (error) 9668 break; 9669 /* If the list has changed restart the loop. */ 9670 idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked); 9671 nino = 0; 9672 if (idp && (idp->id_state & UNLINKNEXT)) 9673 nino = idp->id_ino; 9674 if (nino != pino || 9675 (inodedep->id_state & UNLINKPREV) != UNLINKPREV) { 9676 FREE_LOCK(ump); 9677 brelse(bp); 9678 ACQUIRE_LOCK(ump); 9679 continue; 9680 } 9681 nino = 0; 9682 idn = TAILQ_NEXT(inodedep, id_unlinked); 9683 if (idn) 9684 nino = idn->id_ino; 9685 /* 9686 * Remove us from the in memory list. After this we cannot 9687 * access the inodedep. 9688 */ 9689 KASSERT((inodedep->id_state & UNLINKED) != 0, 9690 ("clear_unlinked_inodedep: inodedep %p not unlinked", 9691 inodedep)); 9692 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST); 9693 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked); 9694 FREE_LOCK(ump); 9695 /* 9696 * The predecessor's next pointer is manually updated here 9697 * so that the NEXT flag is never cleared for an element 9698 * that is in the list. 9699 */ 9700 if (pino == 0) { 9701 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize); 9702 ffs_oldfscompat_write((struct fs *)bp->b_data, ump); 9703 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data, 9704 bp); 9705 } else if (fs->fs_magic == FS_UFS1_MAGIC) 9706 ((struct ufs1_dinode *)bp->b_data + 9707 ino_to_fsbo(fs, pino))->di_freelink = nino; 9708 else 9709 ((struct ufs2_dinode *)bp->b_data + 9710 ino_to_fsbo(fs, pino))->di_freelink = nino; 9711 /* 9712 * If the bwrite fails we have no recourse to recover. The 9713 * filesystem is corrupted already. 9714 */ 9715 bwrite(bp); 9716 ACQUIRE_LOCK(ump); 9717 /* 9718 * If the superblock pointer still needs to be cleared force 9719 * a write here. 9720 */ 9721 if (fs->fs_sujfree == ino) { 9722 FREE_LOCK(ump); 9723 bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc), 9724 (int)fs->fs_sbsize, 0, 0, 0); 9725 bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize); 9726 ffs_oldfscompat_write((struct fs *)bp->b_data, ump); 9727 softdep_setup_sbupdate(ump, (struct fs *)bp->b_data, 9728 bp); 9729 bwrite(bp); 9730 ACQUIRE_LOCK(ump); 9731 } 9732 9733 if (fs->fs_sujfree != ino) 9734 return; 9735 panic("clear_unlinked_inodedep: Failed to clear free head"); 9736 } 9737 if (inodedep->id_ino == fs->fs_sujfree) 9738 panic("clear_unlinked_inodedep: Freeing head of free list"); 9739 inodedep->id_state &= ~(UNLINKED | UNLINKLINKS | UNLINKONLIST); 9740 TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked); 9741 return; 9742 } 9743 9744 /* 9745 * This workitem decrements the inode's link count. 9746 * If the link count reaches zero, the file is removed. 9747 */ 9748 static int 9749 handle_workitem_remove(dirrem, flags) 9750 struct dirrem *dirrem; 9751 int flags; 9752 { 9753 struct inodedep *inodedep; 9754 struct workhead dotdotwk; 9755 struct worklist *wk; 9756 struct ufsmount *ump; 9757 struct mount *mp; 9758 struct vnode *vp; 9759 struct inode *ip; 9760 ino_t oldinum; 9761 9762 if (dirrem->dm_state & ONWORKLIST) 9763 panic("handle_workitem_remove: dirrem %p still on worklist", 9764 dirrem); 9765 oldinum = dirrem->dm_oldinum; 9766 mp = dirrem->dm_list.wk_mp; 9767 ump = VFSTOUFS(mp); 9768 flags |= LK_EXCLUSIVE; 9769 if (ffs_vgetf(mp, oldinum, flags, &vp, FFSV_FORCEINSMQ) != 0) 9770 return (EBUSY); 9771 ip = VTOI(vp); 9772 ACQUIRE_LOCK(ump); 9773 if ((inodedep_lookup(mp, oldinum, 0, &inodedep)) == 0) 9774 panic("handle_workitem_remove: lost inodedep"); 9775 if (dirrem->dm_state & ONDEPLIST) 9776 LIST_REMOVE(dirrem, dm_inonext); 9777 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd), 9778 ("handle_workitem_remove: Journal entries not written.")); 9779 9780 /* 9781 * Move all dependencies waiting on the remove to complete 9782 * from the dirrem to the inode inowait list to be completed 9783 * after the inode has been updated and written to disk. 9784 * 9785 * Any marked MKDIR_PARENT are saved to be completed when the 9786 * dotdot ref is removed unless DIRCHG is specified. For 9787 * directory change operations there will be no further 9788 * directory writes and the jsegdeps need to be moved along 9789 * with the rest to be completed when the inode is free or 9790 * stable in the inode free list. 9791 */ 9792 LIST_INIT(&dotdotwk); 9793 while ((wk = LIST_FIRST(&dirrem->dm_jwork)) != NULL) { 9794 WORKLIST_REMOVE(wk); 9795 if ((dirrem->dm_state & DIRCHG) == 0 && 9796 wk->wk_state & MKDIR_PARENT) { 9797 wk->wk_state &= ~MKDIR_PARENT; 9798 WORKLIST_INSERT(&dotdotwk, wk); 9799 continue; 9800 } 9801 WORKLIST_INSERT(&inodedep->id_inowait, wk); 9802 } 9803 LIST_SWAP(&dirrem->dm_jwork, &dotdotwk, worklist, wk_list); 9804 /* 9805 * Normal file deletion. 9806 */ 9807 if ((dirrem->dm_state & RMDIR) == 0) { 9808 ip->i_nlink--; 9809 DIP_SET(ip, i_nlink, ip->i_nlink); 9810 ip->i_flag |= IN_CHANGE; 9811 if (ip->i_nlink < ip->i_effnlink) 9812 panic("handle_workitem_remove: bad file delta"); 9813 if (ip->i_nlink == 0) 9814 unlinked_inodedep(mp, inodedep); 9815 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9816 KASSERT(LIST_EMPTY(&dirrem->dm_jwork), 9817 ("handle_workitem_remove: worklist not empty. %s", 9818 TYPENAME(LIST_FIRST(&dirrem->dm_jwork)->wk_type))); 9819 WORKITEM_FREE(dirrem, D_DIRREM); 9820 FREE_LOCK(ump); 9821 goto out; 9822 } 9823 /* 9824 * Directory deletion. Decrement reference count for both the 9825 * just deleted parent directory entry and the reference for ".". 9826 * Arrange to have the reference count on the parent decremented 9827 * to account for the loss of "..". 9828 */ 9829 ip->i_nlink -= 2; 9830 DIP_SET(ip, i_nlink, ip->i_nlink); 9831 ip->i_flag |= IN_CHANGE; 9832 if (ip->i_nlink < ip->i_effnlink) 9833 panic("handle_workitem_remove: bad dir delta"); 9834 if (ip->i_nlink == 0) 9835 unlinked_inodedep(mp, inodedep); 9836 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink; 9837 /* 9838 * Rename a directory to a new parent. Since, we are both deleting 9839 * and creating a new directory entry, the link count on the new 9840 * directory should not change. Thus we skip the followup dirrem. 9841 */ 9842 if (dirrem->dm_state & DIRCHG) { 9843 KASSERT(LIST_EMPTY(&dirrem->dm_jwork), 9844 ("handle_workitem_remove: DIRCHG and worklist not empty.")); 9845 WORKITEM_FREE(dirrem, D_DIRREM); 9846 FREE_LOCK(ump); 9847 goto out; 9848 } 9849 dirrem->dm_state = ONDEPLIST; 9850 dirrem->dm_oldinum = dirrem->dm_dirinum; 9851 /* 9852 * Place the dirrem on the parent's diremhd list. 9853 */ 9854 if (inodedep_lookup(mp, dirrem->dm_oldinum, 0, &inodedep) == 0) 9855 panic("handle_workitem_remove: lost dir inodedep"); 9856 LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext); 9857 /* 9858 * If the allocated inode has never been written to disk, then 9859 * the on-disk inode is zero'ed and we can remove the file 9860 * immediately. When journaling if the inode has been marked 9861 * unlinked and not DEPCOMPLETE we know it can never be written. 9862 */ 9863 inodedep_lookup(mp, oldinum, 0, &inodedep); 9864 if (inodedep == NULL || 9865 (inodedep->id_state & (DEPCOMPLETE | UNLINKED)) == UNLINKED || 9866 check_inode_unwritten(inodedep)) { 9867 FREE_LOCK(ump); 9868 vput(vp); 9869 return handle_workitem_remove(dirrem, flags); 9870 } 9871 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list); 9872 FREE_LOCK(ump); 9873 ip->i_flag |= IN_CHANGE; 9874 out: 9875 ffs_update(vp, 0); 9876 vput(vp); 9877 return (0); 9878 } 9879 9880 /* 9881 * Inode de-allocation dependencies. 9882 * 9883 * When an inode's link count is reduced to zero, it can be de-allocated. We 9884 * found it convenient to postpone de-allocation until after the inode is 9885 * written to disk with its new link count (zero). At this point, all of the 9886 * on-disk inode's block pointers are nullified and, with careful dependency 9887 * list ordering, all dependencies related to the inode will be satisfied and 9888 * the corresponding dependency structures de-allocated. So, if/when the 9889 * inode is reused, there will be no mixing of old dependencies with new 9890 * ones. This artificial dependency is set up by the block de-allocation 9891 * procedure above (softdep_setup_freeblocks) and completed by the 9892 * following procedure. 9893 */ 9894 static void 9895 handle_workitem_freefile(freefile) 9896 struct freefile *freefile; 9897 { 9898 struct workhead wkhd; 9899 struct fs *fs; 9900 struct inodedep *idp; 9901 struct ufsmount *ump; 9902 int error; 9903 9904 ump = VFSTOUFS(freefile->fx_list.wk_mp); 9905 fs = ump->um_fs; 9906 #ifdef DEBUG 9907 ACQUIRE_LOCK(ump); 9908 error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp); 9909 FREE_LOCK(ump); 9910 if (error) 9911 panic("handle_workitem_freefile: inodedep %p survived", idp); 9912 #endif 9913 UFS_LOCK(ump); 9914 fs->fs_pendinginodes -= 1; 9915 UFS_UNLOCK(ump); 9916 LIST_INIT(&wkhd); 9917 LIST_SWAP(&freefile->fx_jwork, &wkhd, worklist, wk_list); 9918 if ((error = ffs_freefile(ump, fs, freefile->fx_devvp, 9919 freefile->fx_oldinum, freefile->fx_mode, &wkhd)) != 0) 9920 softdep_error("handle_workitem_freefile", error); 9921 ACQUIRE_LOCK(ump); 9922 WORKITEM_FREE(freefile, D_FREEFILE); 9923 FREE_LOCK(ump); 9924 } 9925 9926 9927 /* 9928 * Helper function which unlinks marker element from work list and returns 9929 * the next element on the list. 9930 */ 9931 static __inline struct worklist * 9932 markernext(struct worklist *marker) 9933 { 9934 struct worklist *next; 9935 9936 next = LIST_NEXT(marker, wk_list); 9937 LIST_REMOVE(marker, wk_list); 9938 return next; 9939 } 9940 9941 /* 9942 * Disk writes. 9943 * 9944 * The dependency structures constructed above are most actively used when file 9945 * system blocks are written to disk. No constraints are placed on when a 9946 * block can be written, but unsatisfied update dependencies are made safe by 9947 * modifying (or replacing) the source memory for the duration of the disk 9948 * write. When the disk write completes, the memory block is again brought 9949 * up-to-date. 9950 * 9951 * In-core inode structure reclamation. 9952 * 9953 * Because there are a finite number of "in-core" inode structures, they are 9954 * reused regularly. By transferring all inode-related dependencies to the 9955 * in-memory inode block and indexing them separately (via "inodedep"s), we 9956 * can allow "in-core" inode structures to be reused at any time and avoid 9957 * any increase in contention. 9958 * 9959 * Called just before entering the device driver to initiate a new disk I/O. 9960 * The buffer must be locked, thus, no I/O completion operations can occur 9961 * while we are manipulating its associated dependencies. 9962 */ 9963 static void 9964 softdep_disk_io_initiation(bp) 9965 struct buf *bp; /* structure describing disk write to occur */ 9966 { 9967 struct worklist *wk; 9968 struct worklist marker; 9969 struct inodedep *inodedep; 9970 struct freeblks *freeblks; 9971 struct jblkdep *jblkdep; 9972 struct newblk *newblk; 9973 struct ufsmount *ump; 9974 9975 /* 9976 * We only care about write operations. There should never 9977 * be dependencies for reads. 9978 */ 9979 if (bp->b_iocmd != BIO_WRITE) 9980 panic("softdep_disk_io_initiation: not write"); 9981 9982 if (bp->b_vflags & BV_BKGRDINPROG) 9983 panic("softdep_disk_io_initiation: Writing buffer with " 9984 "background write in progress: %p", bp); 9985 9986 ump = softdep_bp_to_mp(bp); 9987 if (ump == NULL) 9988 return; 9989 9990 marker.wk_type = D_LAST + 1; /* Not a normal workitem */ 9991 PHOLD(curproc); /* Don't swap out kernel stack */ 9992 ACQUIRE_LOCK(ump); 9993 /* 9994 * Do any necessary pre-I/O processing. 9995 */ 9996 for (wk = LIST_FIRST(&bp->b_dep); wk != NULL; 9997 wk = markernext(&marker)) { 9998 LIST_INSERT_AFTER(wk, &marker, wk_list); 9999 switch (wk->wk_type) { 10000 10001 case D_PAGEDEP: 10002 initiate_write_filepage(WK_PAGEDEP(wk), bp); 10003 continue; 10004 10005 case D_INODEDEP: 10006 inodedep = WK_INODEDEP(wk); 10007 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) 10008 initiate_write_inodeblock_ufs1(inodedep, bp); 10009 else 10010 initiate_write_inodeblock_ufs2(inodedep, bp); 10011 continue; 10012 10013 case D_INDIRDEP: 10014 initiate_write_indirdep(WK_INDIRDEP(wk), bp); 10015 continue; 10016 10017 case D_BMSAFEMAP: 10018 initiate_write_bmsafemap(WK_BMSAFEMAP(wk), bp); 10019 continue; 10020 10021 case D_JSEG: 10022 WK_JSEG(wk)->js_buf = NULL; 10023 continue; 10024 10025 case D_FREEBLKS: 10026 freeblks = WK_FREEBLKS(wk); 10027 jblkdep = LIST_FIRST(&freeblks->fb_jblkdephd); 10028 /* 10029 * We have to wait for the freeblks to be journaled 10030 * before we can write an inodeblock with updated 10031 * pointers. Be careful to arrange the marker so 10032 * we revisit the freeblks if it's not removed by 10033 * the first jwait(). 10034 */ 10035 if (jblkdep != NULL) { 10036 LIST_REMOVE(&marker, wk_list); 10037 LIST_INSERT_BEFORE(wk, &marker, wk_list); 10038 jwait(&jblkdep->jb_list, MNT_WAIT); 10039 } 10040 continue; 10041 case D_ALLOCDIRECT: 10042 case D_ALLOCINDIR: 10043 /* 10044 * We have to wait for the jnewblk to be journaled 10045 * before we can write to a block if the contents 10046 * may be confused with an earlier file's indirect 10047 * at recovery time. Handle the marker as described 10048 * above. 10049 */ 10050 newblk = WK_NEWBLK(wk); 10051 if (newblk->nb_jnewblk != NULL && 10052 indirblk_lookup(newblk->nb_list.wk_mp, 10053 newblk->nb_newblkno)) { 10054 LIST_REMOVE(&marker, wk_list); 10055 LIST_INSERT_BEFORE(wk, &marker, wk_list); 10056 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 10057 } 10058 continue; 10059 10060 case D_SBDEP: 10061 initiate_write_sbdep(WK_SBDEP(wk)); 10062 continue; 10063 10064 case D_MKDIR: 10065 case D_FREEWORK: 10066 case D_FREEDEP: 10067 case D_JSEGDEP: 10068 continue; 10069 10070 default: 10071 panic("handle_disk_io_initiation: Unexpected type %s", 10072 TYPENAME(wk->wk_type)); 10073 /* NOTREACHED */ 10074 } 10075 } 10076 FREE_LOCK(ump); 10077 PRELE(curproc); /* Allow swapout of kernel stack */ 10078 } 10079 10080 /* 10081 * Called from within the procedure above to deal with unsatisfied 10082 * allocation dependencies in a directory. The buffer must be locked, 10083 * thus, no I/O completion operations can occur while we are 10084 * manipulating its associated dependencies. 10085 */ 10086 static void 10087 initiate_write_filepage(pagedep, bp) 10088 struct pagedep *pagedep; 10089 struct buf *bp; 10090 { 10091 struct jremref *jremref; 10092 struct jmvref *jmvref; 10093 struct dirrem *dirrem; 10094 struct diradd *dap; 10095 struct direct *ep; 10096 int i; 10097 10098 if (pagedep->pd_state & IOSTARTED) { 10099 /* 10100 * This can only happen if there is a driver that does not 10101 * understand chaining. Here biodone will reissue the call 10102 * to strategy for the incomplete buffers. 10103 */ 10104 printf("initiate_write_filepage: already started\n"); 10105 return; 10106 } 10107 pagedep->pd_state |= IOSTARTED; 10108 /* 10109 * Wait for all journal remove dependencies to hit the disk. 10110 * We can not allow any potentially conflicting directory adds 10111 * to be visible before removes and rollback is too difficult. 10112 * The per-filesystem lock may be dropped and re-acquired, however 10113 * we hold the buf locked so the dependency can not go away. 10114 */ 10115 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) 10116 while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) 10117 jwait(&jremref->jr_list, MNT_WAIT); 10118 while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) 10119 jwait(&jmvref->jm_list, MNT_WAIT); 10120 for (i = 0; i < DAHASHSZ; i++) { 10121 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 10122 ep = (struct direct *) 10123 ((char *)bp->b_data + dap->da_offset); 10124 if (ep->d_ino != dap->da_newinum) 10125 panic("%s: dir inum %ju != new %ju", 10126 "initiate_write_filepage", 10127 (uintmax_t)ep->d_ino, 10128 (uintmax_t)dap->da_newinum); 10129 if (dap->da_state & DIRCHG) 10130 ep->d_ino = dap->da_previous->dm_oldinum; 10131 else 10132 ep->d_ino = 0; 10133 dap->da_state &= ~ATTACHED; 10134 dap->da_state |= UNDONE; 10135 } 10136 } 10137 } 10138 10139 /* 10140 * Version of initiate_write_inodeblock that handles UFS1 dinodes. 10141 * Note that any bug fixes made to this routine must be done in the 10142 * version found below. 10143 * 10144 * Called from within the procedure above to deal with unsatisfied 10145 * allocation dependencies in an inodeblock. The buffer must be 10146 * locked, thus, no I/O completion operations can occur while we 10147 * are manipulating its associated dependencies. 10148 */ 10149 static void 10150 initiate_write_inodeblock_ufs1(inodedep, bp) 10151 struct inodedep *inodedep; 10152 struct buf *bp; /* The inode block */ 10153 { 10154 struct allocdirect *adp, *lastadp; 10155 struct ufs1_dinode *dp; 10156 struct ufs1_dinode *sip; 10157 struct inoref *inoref; 10158 struct ufsmount *ump; 10159 struct fs *fs; 10160 ufs_lbn_t i; 10161 #ifdef INVARIANTS 10162 ufs_lbn_t prevlbn = 0; 10163 #endif 10164 int deplist; 10165 10166 if (inodedep->id_state & IOSTARTED) 10167 panic("initiate_write_inodeblock_ufs1: already started"); 10168 inodedep->id_state |= IOSTARTED; 10169 fs = inodedep->id_fs; 10170 ump = VFSTOUFS(inodedep->id_list.wk_mp); 10171 LOCK_OWNED(ump); 10172 dp = (struct ufs1_dinode *)bp->b_data + 10173 ino_to_fsbo(fs, inodedep->id_ino); 10174 10175 /* 10176 * If we're on the unlinked list but have not yet written our 10177 * next pointer initialize it here. 10178 */ 10179 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 10180 struct inodedep *inon; 10181 10182 inon = TAILQ_NEXT(inodedep, id_unlinked); 10183 dp->di_freelink = inon ? inon->id_ino : 0; 10184 } 10185 /* 10186 * If the bitmap is not yet written, then the allocated 10187 * inode cannot be written to disk. 10188 */ 10189 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 10190 if (inodedep->id_savedino1 != NULL) 10191 panic("initiate_write_inodeblock_ufs1: I/O underway"); 10192 FREE_LOCK(ump); 10193 sip = malloc(sizeof(struct ufs1_dinode), 10194 M_SAVEDINO, M_SOFTDEP_FLAGS); 10195 ACQUIRE_LOCK(ump); 10196 inodedep->id_savedino1 = sip; 10197 *inodedep->id_savedino1 = *dp; 10198 bzero((caddr_t)dp, sizeof(struct ufs1_dinode)); 10199 dp->di_gen = inodedep->id_savedino1->di_gen; 10200 dp->di_freelink = inodedep->id_savedino1->di_freelink; 10201 return; 10202 } 10203 /* 10204 * If no dependencies, then there is nothing to roll back. 10205 */ 10206 inodedep->id_savedsize = dp->di_size; 10207 inodedep->id_savedextsize = 0; 10208 inodedep->id_savednlink = dp->di_nlink; 10209 if (TAILQ_EMPTY(&inodedep->id_inoupdt) && 10210 TAILQ_EMPTY(&inodedep->id_inoreflst)) 10211 return; 10212 /* 10213 * Revert the link count to that of the first unwritten journal entry. 10214 */ 10215 inoref = TAILQ_FIRST(&inodedep->id_inoreflst); 10216 if (inoref) 10217 dp->di_nlink = inoref->if_nlink; 10218 /* 10219 * Set the dependencies to busy. 10220 */ 10221 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10222 adp = TAILQ_NEXT(adp, ad_next)) { 10223 #ifdef INVARIANTS 10224 if (deplist != 0 && prevlbn >= adp->ad_offset) 10225 panic("softdep_write_inodeblock: lbn order"); 10226 prevlbn = adp->ad_offset; 10227 if (adp->ad_offset < NDADDR && 10228 dp->di_db[adp->ad_offset] != adp->ad_newblkno) 10229 panic("initiate_write_inodeblock_ufs1: " 10230 "direct pointer #%jd mismatch %d != %jd", 10231 (intmax_t)adp->ad_offset, 10232 dp->di_db[adp->ad_offset], 10233 (intmax_t)adp->ad_newblkno); 10234 if (adp->ad_offset >= NDADDR && 10235 dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno) 10236 panic("initiate_write_inodeblock_ufs1: " 10237 "indirect pointer #%jd mismatch %d != %jd", 10238 (intmax_t)adp->ad_offset - NDADDR, 10239 dp->di_ib[adp->ad_offset - NDADDR], 10240 (intmax_t)adp->ad_newblkno); 10241 deplist |= 1 << adp->ad_offset; 10242 if ((adp->ad_state & ATTACHED) == 0) 10243 panic("initiate_write_inodeblock_ufs1: " 10244 "Unknown state 0x%x", adp->ad_state); 10245 #endif /* INVARIANTS */ 10246 adp->ad_state &= ~ATTACHED; 10247 adp->ad_state |= UNDONE; 10248 } 10249 /* 10250 * The on-disk inode cannot claim to be any larger than the last 10251 * fragment that has been written. Otherwise, the on-disk inode 10252 * might have fragments that were not the last block in the file 10253 * which would corrupt the filesystem. 10254 */ 10255 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10256 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10257 if (adp->ad_offset >= NDADDR) 10258 break; 10259 dp->di_db[adp->ad_offset] = adp->ad_oldblkno; 10260 /* keep going until hitting a rollback to a frag */ 10261 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10262 continue; 10263 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10264 for (i = adp->ad_offset + 1; i < NDADDR; i++) { 10265 #ifdef INVARIANTS 10266 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 10267 panic("initiate_write_inodeblock_ufs1: " 10268 "lost dep1"); 10269 #endif /* INVARIANTS */ 10270 dp->di_db[i] = 0; 10271 } 10272 for (i = 0; i < NIADDR; i++) { 10273 #ifdef INVARIANTS 10274 if (dp->di_ib[i] != 0 && 10275 (deplist & ((1 << NDADDR) << i)) == 0) 10276 panic("initiate_write_inodeblock_ufs1: " 10277 "lost dep2"); 10278 #endif /* INVARIANTS */ 10279 dp->di_ib[i] = 0; 10280 } 10281 return; 10282 } 10283 /* 10284 * If we have zero'ed out the last allocated block of the file, 10285 * roll back the size to the last currently allocated block. 10286 * We know that this last allocated block is a full-sized as 10287 * we already checked for fragments in the loop above. 10288 */ 10289 if (lastadp != NULL && 10290 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10291 for (i = lastadp->ad_offset; i >= 0; i--) 10292 if (dp->di_db[i] != 0) 10293 break; 10294 dp->di_size = (i + 1) * fs->fs_bsize; 10295 } 10296 /* 10297 * The only dependencies are for indirect blocks. 10298 * 10299 * The file size for indirect block additions is not guaranteed. 10300 * Such a guarantee would be non-trivial to achieve. The conventional 10301 * synchronous write implementation also does not make this guarantee. 10302 * Fsck should catch and fix discrepancies. Arguably, the file size 10303 * can be over-estimated without destroying integrity when the file 10304 * moves into the indirect blocks (i.e., is large). If we want to 10305 * postpone fsck, we are stuck with this argument. 10306 */ 10307 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 10308 dp->di_ib[adp->ad_offset - NDADDR] = 0; 10309 } 10310 10311 /* 10312 * Version of initiate_write_inodeblock that handles UFS2 dinodes. 10313 * Note that any bug fixes made to this routine must be done in the 10314 * version found above. 10315 * 10316 * Called from within the procedure above to deal with unsatisfied 10317 * allocation dependencies in an inodeblock. The buffer must be 10318 * locked, thus, no I/O completion operations can occur while we 10319 * are manipulating its associated dependencies. 10320 */ 10321 static void 10322 initiate_write_inodeblock_ufs2(inodedep, bp) 10323 struct inodedep *inodedep; 10324 struct buf *bp; /* The inode block */ 10325 { 10326 struct allocdirect *adp, *lastadp; 10327 struct ufs2_dinode *dp; 10328 struct ufs2_dinode *sip; 10329 struct inoref *inoref; 10330 struct ufsmount *ump; 10331 struct fs *fs; 10332 ufs_lbn_t i; 10333 #ifdef INVARIANTS 10334 ufs_lbn_t prevlbn = 0; 10335 #endif 10336 int deplist; 10337 10338 if (inodedep->id_state & IOSTARTED) 10339 panic("initiate_write_inodeblock_ufs2: already started"); 10340 inodedep->id_state |= IOSTARTED; 10341 fs = inodedep->id_fs; 10342 ump = VFSTOUFS(inodedep->id_list.wk_mp); 10343 LOCK_OWNED(ump); 10344 dp = (struct ufs2_dinode *)bp->b_data + 10345 ino_to_fsbo(fs, inodedep->id_ino); 10346 10347 /* 10348 * If we're on the unlinked list but have not yet written our 10349 * next pointer initialize it here. 10350 */ 10351 if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) { 10352 struct inodedep *inon; 10353 10354 inon = TAILQ_NEXT(inodedep, id_unlinked); 10355 dp->di_freelink = inon ? inon->id_ino : 0; 10356 } 10357 /* 10358 * If the bitmap is not yet written, then the allocated 10359 * inode cannot be written to disk. 10360 */ 10361 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 10362 if (inodedep->id_savedino2 != NULL) 10363 panic("initiate_write_inodeblock_ufs2: I/O underway"); 10364 FREE_LOCK(ump); 10365 sip = malloc(sizeof(struct ufs2_dinode), 10366 M_SAVEDINO, M_SOFTDEP_FLAGS); 10367 ACQUIRE_LOCK(ump); 10368 inodedep->id_savedino2 = sip; 10369 *inodedep->id_savedino2 = *dp; 10370 bzero((caddr_t)dp, sizeof(struct ufs2_dinode)); 10371 dp->di_gen = inodedep->id_savedino2->di_gen; 10372 dp->di_freelink = inodedep->id_savedino2->di_freelink; 10373 return; 10374 } 10375 /* 10376 * If no dependencies, then there is nothing to roll back. 10377 */ 10378 inodedep->id_savedsize = dp->di_size; 10379 inodedep->id_savedextsize = dp->di_extsize; 10380 inodedep->id_savednlink = dp->di_nlink; 10381 if (TAILQ_EMPTY(&inodedep->id_inoupdt) && 10382 TAILQ_EMPTY(&inodedep->id_extupdt) && 10383 TAILQ_EMPTY(&inodedep->id_inoreflst)) 10384 return; 10385 /* 10386 * Revert the link count to that of the first unwritten journal entry. 10387 */ 10388 inoref = TAILQ_FIRST(&inodedep->id_inoreflst); 10389 if (inoref) 10390 dp->di_nlink = inoref->if_nlink; 10391 10392 /* 10393 * Set the ext data dependencies to busy. 10394 */ 10395 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 10396 adp = TAILQ_NEXT(adp, ad_next)) { 10397 #ifdef INVARIANTS 10398 if (deplist != 0 && prevlbn >= adp->ad_offset) 10399 panic("initiate_write_inodeblock_ufs2: lbn order"); 10400 prevlbn = adp->ad_offset; 10401 if (dp->di_extb[adp->ad_offset] != adp->ad_newblkno) 10402 panic("initiate_write_inodeblock_ufs2: " 10403 "ext pointer #%jd mismatch %jd != %jd", 10404 (intmax_t)adp->ad_offset, 10405 (intmax_t)dp->di_extb[adp->ad_offset], 10406 (intmax_t)adp->ad_newblkno); 10407 deplist |= 1 << adp->ad_offset; 10408 if ((adp->ad_state & ATTACHED) == 0) 10409 panic("initiate_write_inodeblock_ufs2: Unknown " 10410 "state 0x%x", adp->ad_state); 10411 #endif /* INVARIANTS */ 10412 adp->ad_state &= ~ATTACHED; 10413 adp->ad_state |= UNDONE; 10414 } 10415 /* 10416 * The on-disk inode cannot claim to be any larger than the last 10417 * fragment that has been written. Otherwise, the on-disk inode 10418 * might have fragments that were not the last block in the ext 10419 * data which would corrupt the filesystem. 10420 */ 10421 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; 10422 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10423 dp->di_extb[adp->ad_offset] = adp->ad_oldblkno; 10424 /* keep going until hitting a rollback to a frag */ 10425 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10426 continue; 10427 dp->di_extsize = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10428 for (i = adp->ad_offset + 1; i < NXADDR; i++) { 10429 #ifdef INVARIANTS 10430 if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) 10431 panic("initiate_write_inodeblock_ufs2: " 10432 "lost dep1"); 10433 #endif /* INVARIANTS */ 10434 dp->di_extb[i] = 0; 10435 } 10436 lastadp = NULL; 10437 break; 10438 } 10439 /* 10440 * If we have zero'ed out the last allocated block of the ext 10441 * data, roll back the size to the last currently allocated block. 10442 * We know that this last allocated block is a full-sized as 10443 * we already checked for fragments in the loop above. 10444 */ 10445 if (lastadp != NULL && 10446 dp->di_extsize <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10447 for (i = lastadp->ad_offset; i >= 0; i--) 10448 if (dp->di_extb[i] != 0) 10449 break; 10450 dp->di_extsize = (i + 1) * fs->fs_bsize; 10451 } 10452 /* 10453 * Set the file data dependencies to busy. 10454 */ 10455 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10456 adp = TAILQ_NEXT(adp, ad_next)) { 10457 #ifdef INVARIANTS 10458 if (deplist != 0 && prevlbn >= adp->ad_offset) 10459 panic("softdep_write_inodeblock: lbn order"); 10460 if ((adp->ad_state & ATTACHED) == 0) 10461 panic("inodedep %p and adp %p not attached", inodedep, adp); 10462 prevlbn = adp->ad_offset; 10463 if (adp->ad_offset < NDADDR && 10464 dp->di_db[adp->ad_offset] != adp->ad_newblkno) 10465 panic("initiate_write_inodeblock_ufs2: " 10466 "direct pointer #%jd mismatch %jd != %jd", 10467 (intmax_t)adp->ad_offset, 10468 (intmax_t)dp->di_db[adp->ad_offset], 10469 (intmax_t)adp->ad_newblkno); 10470 if (adp->ad_offset >= NDADDR && 10471 dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno) 10472 panic("initiate_write_inodeblock_ufs2: " 10473 "indirect pointer #%jd mismatch %jd != %jd", 10474 (intmax_t)adp->ad_offset - NDADDR, 10475 (intmax_t)dp->di_ib[adp->ad_offset - NDADDR], 10476 (intmax_t)adp->ad_newblkno); 10477 deplist |= 1 << adp->ad_offset; 10478 if ((adp->ad_state & ATTACHED) == 0) 10479 panic("initiate_write_inodeblock_ufs2: Unknown " 10480 "state 0x%x", adp->ad_state); 10481 #endif /* INVARIANTS */ 10482 adp->ad_state &= ~ATTACHED; 10483 adp->ad_state |= UNDONE; 10484 } 10485 /* 10486 * The on-disk inode cannot claim to be any larger than the last 10487 * fragment that has been written. Otherwise, the on-disk inode 10488 * might have fragments that were not the last block in the file 10489 * which would corrupt the filesystem. 10490 */ 10491 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; 10492 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) { 10493 if (adp->ad_offset >= NDADDR) 10494 break; 10495 dp->di_db[adp->ad_offset] = adp->ad_oldblkno; 10496 /* keep going until hitting a rollback to a frag */ 10497 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize) 10498 continue; 10499 dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize; 10500 for (i = adp->ad_offset + 1; i < NDADDR; i++) { 10501 #ifdef INVARIANTS 10502 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) 10503 panic("initiate_write_inodeblock_ufs2: " 10504 "lost dep2"); 10505 #endif /* INVARIANTS */ 10506 dp->di_db[i] = 0; 10507 } 10508 for (i = 0; i < NIADDR; i++) { 10509 #ifdef INVARIANTS 10510 if (dp->di_ib[i] != 0 && 10511 (deplist & ((1 << NDADDR) << i)) == 0) 10512 panic("initiate_write_inodeblock_ufs2: " 10513 "lost dep3"); 10514 #endif /* INVARIANTS */ 10515 dp->di_ib[i] = 0; 10516 } 10517 return; 10518 } 10519 /* 10520 * If we have zero'ed out the last allocated block of the file, 10521 * roll back the size to the last currently allocated block. 10522 * We know that this last allocated block is a full-sized as 10523 * we already checked for fragments in the loop above. 10524 */ 10525 if (lastadp != NULL && 10526 dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) { 10527 for (i = lastadp->ad_offset; i >= 0; i--) 10528 if (dp->di_db[i] != 0) 10529 break; 10530 dp->di_size = (i + 1) * fs->fs_bsize; 10531 } 10532 /* 10533 * The only dependencies are for indirect blocks. 10534 * 10535 * The file size for indirect block additions is not guaranteed. 10536 * Such a guarantee would be non-trivial to achieve. The conventional 10537 * synchronous write implementation also does not make this guarantee. 10538 * Fsck should catch and fix discrepancies. Arguably, the file size 10539 * can be over-estimated without destroying integrity when the file 10540 * moves into the indirect blocks (i.e., is large). If we want to 10541 * postpone fsck, we are stuck with this argument. 10542 */ 10543 for (; adp; adp = TAILQ_NEXT(adp, ad_next)) 10544 dp->di_ib[adp->ad_offset - NDADDR] = 0; 10545 } 10546 10547 /* 10548 * Cancel an indirdep as a result of truncation. Release all of the 10549 * children allocindirs and place their journal work on the appropriate 10550 * list. 10551 */ 10552 static void 10553 cancel_indirdep(indirdep, bp, freeblks) 10554 struct indirdep *indirdep; 10555 struct buf *bp; 10556 struct freeblks *freeblks; 10557 { 10558 struct allocindir *aip; 10559 10560 /* 10561 * None of the indirect pointers will ever be visible, 10562 * so they can simply be tossed. GOINGAWAY ensures 10563 * that allocated pointers will be saved in the buffer 10564 * cache until they are freed. Note that they will 10565 * only be able to be found by their physical address 10566 * since the inode mapping the logical address will 10567 * be gone. The save buffer used for the safe copy 10568 * was allocated in setup_allocindir_phase2 using 10569 * the physical address so it could be used for this 10570 * purpose. Hence we swap the safe copy with the real 10571 * copy, allowing the safe copy to be freed and holding 10572 * on to the real copy for later use in indir_trunc. 10573 */ 10574 if (indirdep->ir_state & GOINGAWAY) 10575 panic("cancel_indirdep: already gone"); 10576 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 10577 indirdep->ir_state |= DEPCOMPLETE; 10578 LIST_REMOVE(indirdep, ir_next); 10579 } 10580 indirdep->ir_state |= GOINGAWAY; 10581 /* 10582 * Pass in bp for blocks still have journal writes 10583 * pending so we can cancel them on their own. 10584 */ 10585 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != NULL) 10586 cancel_allocindir(aip, bp, freeblks, 0); 10587 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != NULL) 10588 cancel_allocindir(aip, NULL, freeblks, 0); 10589 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != NULL) 10590 cancel_allocindir(aip, NULL, freeblks, 0); 10591 while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) 10592 cancel_allocindir(aip, NULL, freeblks, 0); 10593 /* 10594 * If there are pending partial truncations we need to keep the 10595 * old block copy around until they complete. This is because 10596 * the current b_data is not a perfect superset of the available 10597 * blocks. 10598 */ 10599 if (TAILQ_EMPTY(&indirdep->ir_trunc)) 10600 bcopy(bp->b_data, indirdep->ir_savebp->b_data, bp->b_bcount); 10601 else 10602 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 10603 WORKLIST_REMOVE(&indirdep->ir_list); 10604 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, &indirdep->ir_list); 10605 indirdep->ir_bp = NULL; 10606 indirdep->ir_freeblks = freeblks; 10607 } 10608 10609 /* 10610 * Free an indirdep once it no longer has new pointers to track. 10611 */ 10612 static void 10613 free_indirdep(indirdep) 10614 struct indirdep *indirdep; 10615 { 10616 10617 KASSERT(TAILQ_EMPTY(&indirdep->ir_trunc), 10618 ("free_indirdep: Indir trunc list not empty.")); 10619 KASSERT(LIST_EMPTY(&indirdep->ir_completehd), 10620 ("free_indirdep: Complete head not empty.")); 10621 KASSERT(LIST_EMPTY(&indirdep->ir_writehd), 10622 ("free_indirdep: write head not empty.")); 10623 KASSERT(LIST_EMPTY(&indirdep->ir_donehd), 10624 ("free_indirdep: done head not empty.")); 10625 KASSERT(LIST_EMPTY(&indirdep->ir_deplisthd), 10626 ("free_indirdep: deplist head not empty.")); 10627 KASSERT((indirdep->ir_state & DEPCOMPLETE), 10628 ("free_indirdep: %p still on newblk list.", indirdep)); 10629 KASSERT(indirdep->ir_saveddata == NULL, 10630 ("free_indirdep: %p still has saved data.", indirdep)); 10631 if (indirdep->ir_state & ONWORKLIST) 10632 WORKLIST_REMOVE(&indirdep->ir_list); 10633 WORKITEM_FREE(indirdep, D_INDIRDEP); 10634 } 10635 10636 /* 10637 * Called before a write to an indirdep. This routine is responsible for 10638 * rolling back pointers to a safe state which includes only those 10639 * allocindirs which have been completed. 10640 */ 10641 static void 10642 initiate_write_indirdep(indirdep, bp) 10643 struct indirdep *indirdep; 10644 struct buf *bp; 10645 { 10646 struct ufsmount *ump; 10647 10648 indirdep->ir_state |= IOSTARTED; 10649 if (indirdep->ir_state & GOINGAWAY) 10650 panic("disk_io_initiation: indirdep gone"); 10651 /* 10652 * If there are no remaining dependencies, this will be writing 10653 * the real pointers. 10654 */ 10655 if (LIST_EMPTY(&indirdep->ir_deplisthd) && 10656 TAILQ_EMPTY(&indirdep->ir_trunc)) 10657 return; 10658 /* 10659 * Replace up-to-date version with safe version. 10660 */ 10661 if (indirdep->ir_saveddata == NULL) { 10662 ump = VFSTOUFS(indirdep->ir_list.wk_mp); 10663 LOCK_OWNED(ump); 10664 FREE_LOCK(ump); 10665 indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP, 10666 M_SOFTDEP_FLAGS); 10667 ACQUIRE_LOCK(ump); 10668 } 10669 indirdep->ir_state &= ~ATTACHED; 10670 indirdep->ir_state |= UNDONE; 10671 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount); 10672 bcopy(indirdep->ir_savebp->b_data, bp->b_data, 10673 bp->b_bcount); 10674 } 10675 10676 /* 10677 * Called when an inode has been cleared in a cg bitmap. This finally 10678 * eliminates any canceled jaddrefs 10679 */ 10680 void 10681 softdep_setup_inofree(mp, bp, ino, wkhd) 10682 struct mount *mp; 10683 struct buf *bp; 10684 ino_t ino; 10685 struct workhead *wkhd; 10686 { 10687 struct worklist *wk, *wkn; 10688 struct inodedep *inodedep; 10689 struct ufsmount *ump; 10690 uint8_t *inosused; 10691 struct cg *cgp; 10692 struct fs *fs; 10693 10694 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 10695 ("softdep_setup_inofree called on non-softdep filesystem")); 10696 ump = VFSTOUFS(mp); 10697 ACQUIRE_LOCK(ump); 10698 fs = ump->um_fs; 10699 cgp = (struct cg *)bp->b_data; 10700 inosused = cg_inosused(cgp); 10701 if (isset(inosused, ino % fs->fs_ipg)) 10702 panic("softdep_setup_inofree: inode %ju not freed.", 10703 (uintmax_t)ino); 10704 if (inodedep_lookup(mp, ino, 0, &inodedep)) 10705 panic("softdep_setup_inofree: ino %ju has existing inodedep %p", 10706 (uintmax_t)ino, inodedep); 10707 if (wkhd) { 10708 LIST_FOREACH_SAFE(wk, wkhd, wk_list, wkn) { 10709 if (wk->wk_type != D_JADDREF) 10710 continue; 10711 WORKLIST_REMOVE(wk); 10712 /* 10713 * We can free immediately even if the jaddref 10714 * isn't attached in a background write as now 10715 * the bitmaps are reconciled. 10716 */ 10717 wk->wk_state |= COMPLETE | ATTACHED; 10718 free_jaddref(WK_JADDREF(wk)); 10719 } 10720 jwork_move(&bp->b_dep, wkhd); 10721 } 10722 FREE_LOCK(ump); 10723 } 10724 10725 10726 /* 10727 * Called via ffs_blkfree() after a set of frags has been cleared from a cg 10728 * map. Any dependencies waiting for the write to clear are added to the 10729 * buf's list and any jnewblks that are being canceled are discarded 10730 * immediately. 10731 */ 10732 void 10733 softdep_setup_blkfree(mp, bp, blkno, frags, wkhd) 10734 struct mount *mp; 10735 struct buf *bp; 10736 ufs2_daddr_t blkno; 10737 int frags; 10738 struct workhead *wkhd; 10739 { 10740 struct bmsafemap *bmsafemap; 10741 struct jnewblk *jnewblk; 10742 struct ufsmount *ump; 10743 struct worklist *wk; 10744 struct fs *fs; 10745 #ifdef SUJ_DEBUG 10746 uint8_t *blksfree; 10747 struct cg *cgp; 10748 ufs2_daddr_t jstart; 10749 ufs2_daddr_t jend; 10750 ufs2_daddr_t end; 10751 long bno; 10752 int i; 10753 #endif 10754 10755 CTR3(KTR_SUJ, 10756 "softdep_setup_blkfree: blkno %jd frags %d wk head %p", 10757 blkno, frags, wkhd); 10758 10759 ump = VFSTOUFS(mp); 10760 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 10761 ("softdep_setup_blkfree called on non-softdep filesystem")); 10762 ACQUIRE_LOCK(ump); 10763 /* Lookup the bmsafemap so we track when it is dirty. */ 10764 fs = ump->um_fs; 10765 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL); 10766 /* 10767 * Detach any jnewblks which have been canceled. They must linger 10768 * until the bitmap is cleared again by ffs_blkfree() to prevent 10769 * an unjournaled allocation from hitting the disk. 10770 */ 10771 if (wkhd) { 10772 while ((wk = LIST_FIRST(wkhd)) != NULL) { 10773 CTR2(KTR_SUJ, 10774 "softdep_setup_blkfree: blkno %jd wk type %d", 10775 blkno, wk->wk_type); 10776 WORKLIST_REMOVE(wk); 10777 if (wk->wk_type != D_JNEWBLK) { 10778 WORKLIST_INSERT(&bmsafemap->sm_freehd, wk); 10779 continue; 10780 } 10781 jnewblk = WK_JNEWBLK(wk); 10782 KASSERT(jnewblk->jn_state & GOINGAWAY, 10783 ("softdep_setup_blkfree: jnewblk not canceled.")); 10784 #ifdef SUJ_DEBUG 10785 /* 10786 * Assert that this block is free in the bitmap 10787 * before we discard the jnewblk. 10788 */ 10789 cgp = (struct cg *)bp->b_data; 10790 blksfree = cg_blksfree(cgp); 10791 bno = dtogd(fs, jnewblk->jn_blkno); 10792 for (i = jnewblk->jn_oldfrags; 10793 i < jnewblk->jn_frags; i++) { 10794 if (isset(blksfree, bno + i)) 10795 continue; 10796 panic("softdep_setup_blkfree: not free"); 10797 } 10798 #endif 10799 /* 10800 * Even if it's not attached we can free immediately 10801 * as the new bitmap is correct. 10802 */ 10803 wk->wk_state |= COMPLETE | ATTACHED; 10804 free_jnewblk(jnewblk); 10805 } 10806 } 10807 10808 #ifdef SUJ_DEBUG 10809 /* 10810 * Assert that we are not freeing a block which has an outstanding 10811 * allocation dependency. 10812 */ 10813 fs = VFSTOUFS(mp)->um_fs; 10814 bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno), NULL); 10815 end = blkno + frags; 10816 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) { 10817 /* 10818 * Don't match against blocks that will be freed when the 10819 * background write is done. 10820 */ 10821 if ((jnewblk->jn_state & (ATTACHED | COMPLETE | DEPCOMPLETE)) == 10822 (COMPLETE | DEPCOMPLETE)) 10823 continue; 10824 jstart = jnewblk->jn_blkno + jnewblk->jn_oldfrags; 10825 jend = jnewblk->jn_blkno + jnewblk->jn_frags; 10826 if ((blkno >= jstart && blkno < jend) || 10827 (end > jstart && end <= jend)) { 10828 printf("state 0x%X %jd - %d %d dep %p\n", 10829 jnewblk->jn_state, jnewblk->jn_blkno, 10830 jnewblk->jn_oldfrags, jnewblk->jn_frags, 10831 jnewblk->jn_dep); 10832 panic("softdep_setup_blkfree: " 10833 "%jd-%jd(%d) overlaps with %jd-%jd", 10834 blkno, end, frags, jstart, jend); 10835 } 10836 } 10837 #endif 10838 FREE_LOCK(ump); 10839 } 10840 10841 /* 10842 * Revert a block allocation when the journal record that describes it 10843 * is not yet written. 10844 */ 10845 static int 10846 jnewblk_rollback(jnewblk, fs, cgp, blksfree) 10847 struct jnewblk *jnewblk; 10848 struct fs *fs; 10849 struct cg *cgp; 10850 uint8_t *blksfree; 10851 { 10852 ufs1_daddr_t fragno; 10853 long cgbno, bbase; 10854 int frags, blk; 10855 int i; 10856 10857 frags = 0; 10858 cgbno = dtogd(fs, jnewblk->jn_blkno); 10859 /* 10860 * We have to test which frags need to be rolled back. We may 10861 * be operating on a stale copy when doing background writes. 10862 */ 10863 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) 10864 if (isclr(blksfree, cgbno + i)) 10865 frags++; 10866 if (frags == 0) 10867 return (0); 10868 /* 10869 * This is mostly ffs_blkfree() sans some validation and 10870 * superblock updates. 10871 */ 10872 if (frags == fs->fs_frag) { 10873 fragno = fragstoblks(fs, cgbno); 10874 ffs_setblock(fs, blksfree, fragno); 10875 ffs_clusteracct(fs, cgp, fragno, 1); 10876 cgp->cg_cs.cs_nbfree++; 10877 } else { 10878 cgbno += jnewblk->jn_oldfrags; 10879 bbase = cgbno - fragnum(fs, cgbno); 10880 /* Decrement the old frags. */ 10881 blk = blkmap(fs, blksfree, bbase); 10882 ffs_fragacct(fs, blk, cgp->cg_frsum, -1); 10883 /* Deallocate the fragment */ 10884 for (i = 0; i < frags; i++) 10885 setbit(blksfree, cgbno + i); 10886 cgp->cg_cs.cs_nffree += frags; 10887 /* Add back in counts associated with the new frags */ 10888 blk = blkmap(fs, blksfree, bbase); 10889 ffs_fragacct(fs, blk, cgp->cg_frsum, 1); 10890 /* If a complete block has been reassembled, account for it. */ 10891 fragno = fragstoblks(fs, bbase); 10892 if (ffs_isblock(fs, blksfree, fragno)) { 10893 cgp->cg_cs.cs_nffree -= fs->fs_frag; 10894 ffs_clusteracct(fs, cgp, fragno, 1); 10895 cgp->cg_cs.cs_nbfree++; 10896 } 10897 } 10898 stat_jnewblk++; 10899 jnewblk->jn_state &= ~ATTACHED; 10900 jnewblk->jn_state |= UNDONE; 10901 10902 return (frags); 10903 } 10904 10905 static void 10906 initiate_write_bmsafemap(bmsafemap, bp) 10907 struct bmsafemap *bmsafemap; 10908 struct buf *bp; /* The cg block. */ 10909 { 10910 struct jaddref *jaddref; 10911 struct jnewblk *jnewblk; 10912 uint8_t *inosused; 10913 uint8_t *blksfree; 10914 struct cg *cgp; 10915 struct fs *fs; 10916 ino_t ino; 10917 10918 /* 10919 * If this is a background write, we did this at the time that 10920 * the copy was made, so do not need to do it again. 10921 */ 10922 if (bmsafemap->sm_state & IOSTARTED) 10923 return; 10924 bmsafemap->sm_state |= IOSTARTED; 10925 /* 10926 * Clear any inode allocations which are pending journal writes. 10927 */ 10928 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd) != NULL) { 10929 cgp = (struct cg *)bp->b_data; 10930 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 10931 inosused = cg_inosused(cgp); 10932 LIST_FOREACH(jaddref, &bmsafemap->sm_jaddrefhd, ja_bmdeps) { 10933 ino = jaddref->ja_ino % fs->fs_ipg; 10934 if (isset(inosused, ino)) { 10935 if ((jaddref->ja_mode & IFMT) == IFDIR) 10936 cgp->cg_cs.cs_ndir--; 10937 cgp->cg_cs.cs_nifree++; 10938 clrbit(inosused, ino); 10939 jaddref->ja_state &= ~ATTACHED; 10940 jaddref->ja_state |= UNDONE; 10941 stat_jaddref++; 10942 } else 10943 panic("initiate_write_bmsafemap: inode %ju " 10944 "marked free", (uintmax_t)jaddref->ja_ino); 10945 } 10946 } 10947 /* 10948 * Clear any block allocations which are pending journal writes. 10949 */ 10950 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) { 10951 cgp = (struct cg *)bp->b_data; 10952 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 10953 blksfree = cg_blksfree(cgp); 10954 LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) { 10955 if (jnewblk_rollback(jnewblk, fs, cgp, blksfree)) 10956 continue; 10957 panic("initiate_write_bmsafemap: block %jd " 10958 "marked free", jnewblk->jn_blkno); 10959 } 10960 } 10961 /* 10962 * Move allocation lists to the written lists so they can be 10963 * cleared once the block write is complete. 10964 */ 10965 LIST_SWAP(&bmsafemap->sm_inodedephd, &bmsafemap->sm_inodedepwr, 10966 inodedep, id_deps); 10967 LIST_SWAP(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr, 10968 newblk, nb_deps); 10969 LIST_SWAP(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, worklist, 10970 wk_list); 10971 } 10972 10973 /* 10974 * This routine is called during the completion interrupt 10975 * service routine for a disk write (from the procedure called 10976 * by the device driver to inform the filesystem caches of 10977 * a request completion). It should be called early in this 10978 * procedure, before the block is made available to other 10979 * processes or other routines are called. 10980 * 10981 */ 10982 static void 10983 softdep_disk_write_complete(bp) 10984 struct buf *bp; /* describes the completed disk write */ 10985 { 10986 struct worklist *wk; 10987 struct worklist *owk; 10988 struct ufsmount *ump; 10989 struct workhead reattach; 10990 struct freeblks *freeblks; 10991 struct buf *sbp; 10992 10993 ump = softdep_bp_to_mp(bp); 10994 if (ump == NULL) 10995 return; 10996 10997 /* 10998 * If an error occurred while doing the write, then the data 10999 * has not hit the disk and the dependencies cannot be processed. 11000 * But we do have to go through and roll forward any dependencies 11001 * that were rolled back before the disk write. 11002 */ 11003 ACQUIRE_LOCK(ump); 11004 if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0) { 11005 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 11006 switch (wk->wk_type) { 11007 11008 case D_PAGEDEP: 11009 handle_written_filepage(WK_PAGEDEP(wk), bp, 0); 11010 continue; 11011 11012 case D_INODEDEP: 11013 handle_written_inodeblock(WK_INODEDEP(wk), 11014 bp, 0); 11015 continue; 11016 11017 case D_BMSAFEMAP: 11018 handle_written_bmsafemap(WK_BMSAFEMAP(wk), 11019 bp, 0); 11020 continue; 11021 11022 case D_INDIRDEP: 11023 handle_written_indirdep(WK_INDIRDEP(wk), 11024 bp, &sbp, 0); 11025 continue; 11026 default: 11027 /* nothing to roll forward */ 11028 continue; 11029 } 11030 } 11031 FREE_LOCK(ump); 11032 return; 11033 } 11034 LIST_INIT(&reattach); 11035 11036 /* 11037 * Ump SU lock must not be released anywhere in this code segment. 11038 */ 11039 sbp = NULL; 11040 owk = NULL; 11041 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) { 11042 WORKLIST_REMOVE(wk); 11043 atomic_add_long(&dep_write[wk->wk_type], 1); 11044 if (wk == owk) 11045 panic("duplicate worklist: %p\n", wk); 11046 owk = wk; 11047 switch (wk->wk_type) { 11048 11049 case D_PAGEDEP: 11050 if (handle_written_filepage(WK_PAGEDEP(wk), bp, 11051 WRITESUCCEEDED)) 11052 WORKLIST_INSERT(&reattach, wk); 11053 continue; 11054 11055 case D_INODEDEP: 11056 if (handle_written_inodeblock(WK_INODEDEP(wk), bp, 11057 WRITESUCCEEDED)) 11058 WORKLIST_INSERT(&reattach, wk); 11059 continue; 11060 11061 case D_BMSAFEMAP: 11062 if (handle_written_bmsafemap(WK_BMSAFEMAP(wk), bp, 11063 WRITESUCCEEDED)) 11064 WORKLIST_INSERT(&reattach, wk); 11065 continue; 11066 11067 case D_MKDIR: 11068 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY); 11069 continue; 11070 11071 case D_ALLOCDIRECT: 11072 wk->wk_state |= COMPLETE; 11073 handle_allocdirect_partdone(WK_ALLOCDIRECT(wk), NULL); 11074 continue; 11075 11076 case D_ALLOCINDIR: 11077 wk->wk_state |= COMPLETE; 11078 handle_allocindir_partdone(WK_ALLOCINDIR(wk)); 11079 continue; 11080 11081 case D_INDIRDEP: 11082 if (handle_written_indirdep(WK_INDIRDEP(wk), bp, &sbp, 11083 WRITESUCCEEDED)) 11084 WORKLIST_INSERT(&reattach, wk); 11085 continue; 11086 11087 case D_FREEBLKS: 11088 wk->wk_state |= COMPLETE; 11089 freeblks = WK_FREEBLKS(wk); 11090 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE && 11091 LIST_EMPTY(&freeblks->fb_jblkdephd)) 11092 add_to_worklist(wk, WK_NODELAY); 11093 continue; 11094 11095 case D_FREEWORK: 11096 handle_written_freework(WK_FREEWORK(wk)); 11097 break; 11098 11099 case D_JSEGDEP: 11100 free_jsegdep(WK_JSEGDEP(wk)); 11101 continue; 11102 11103 case D_JSEG: 11104 handle_written_jseg(WK_JSEG(wk), bp); 11105 continue; 11106 11107 case D_SBDEP: 11108 if (handle_written_sbdep(WK_SBDEP(wk), bp)) 11109 WORKLIST_INSERT(&reattach, wk); 11110 continue; 11111 11112 case D_FREEDEP: 11113 free_freedep(WK_FREEDEP(wk)); 11114 continue; 11115 11116 default: 11117 panic("handle_disk_write_complete: Unknown type %s", 11118 TYPENAME(wk->wk_type)); 11119 /* NOTREACHED */ 11120 } 11121 } 11122 /* 11123 * Reattach any requests that must be redone. 11124 */ 11125 while ((wk = LIST_FIRST(&reattach)) != NULL) { 11126 WORKLIST_REMOVE(wk); 11127 WORKLIST_INSERT(&bp->b_dep, wk); 11128 } 11129 FREE_LOCK(ump); 11130 if (sbp) 11131 brelse(sbp); 11132 } 11133 11134 /* 11135 * Called from within softdep_disk_write_complete above. Note that 11136 * this routine is always called from interrupt level with further 11137 * splbio interrupts blocked. 11138 */ 11139 static void 11140 handle_allocdirect_partdone(adp, wkhd) 11141 struct allocdirect *adp; /* the completed allocdirect */ 11142 struct workhead *wkhd; /* Work to do when inode is writtne. */ 11143 { 11144 struct allocdirectlst *listhead; 11145 struct allocdirect *listadp; 11146 struct inodedep *inodedep; 11147 long bsize; 11148 11149 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 11150 return; 11151 /* 11152 * The on-disk inode cannot claim to be any larger than the last 11153 * fragment that has been written. Otherwise, the on-disk inode 11154 * might have fragments that were not the last block in the file 11155 * which would corrupt the filesystem. Thus, we cannot free any 11156 * allocdirects after one whose ad_oldblkno claims a fragment as 11157 * these blocks must be rolled back to zero before writing the inode. 11158 * We check the currently active set of allocdirects in id_inoupdt 11159 * or id_extupdt as appropriate. 11160 */ 11161 inodedep = adp->ad_inodedep; 11162 bsize = inodedep->id_fs->fs_bsize; 11163 if (adp->ad_state & EXTDATA) 11164 listhead = &inodedep->id_extupdt; 11165 else 11166 listhead = &inodedep->id_inoupdt; 11167 TAILQ_FOREACH(listadp, listhead, ad_next) { 11168 /* found our block */ 11169 if (listadp == adp) 11170 break; 11171 /* continue if ad_oldlbn is not a fragment */ 11172 if (listadp->ad_oldsize == 0 || 11173 listadp->ad_oldsize == bsize) 11174 continue; 11175 /* hit a fragment */ 11176 return; 11177 } 11178 /* 11179 * If we have reached the end of the current list without 11180 * finding the just finished dependency, then it must be 11181 * on the future dependency list. Future dependencies cannot 11182 * be freed until they are moved to the current list. 11183 */ 11184 if (listadp == NULL) { 11185 #ifdef DEBUG 11186 if (adp->ad_state & EXTDATA) 11187 listhead = &inodedep->id_newextupdt; 11188 else 11189 listhead = &inodedep->id_newinoupdt; 11190 TAILQ_FOREACH(listadp, listhead, ad_next) 11191 /* found our block */ 11192 if (listadp == adp) 11193 break; 11194 if (listadp == NULL) 11195 panic("handle_allocdirect_partdone: lost dep"); 11196 #endif /* DEBUG */ 11197 return; 11198 } 11199 /* 11200 * If we have found the just finished dependency, then queue 11201 * it along with anything that follows it that is complete. 11202 * Since the pointer has not yet been written in the inode 11203 * as the dependency prevents it, place the allocdirect on the 11204 * bufwait list where it will be freed once the pointer is 11205 * valid. 11206 */ 11207 if (wkhd == NULL) 11208 wkhd = &inodedep->id_bufwait; 11209 for (; adp; adp = listadp) { 11210 listadp = TAILQ_NEXT(adp, ad_next); 11211 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE) 11212 return; 11213 TAILQ_REMOVE(listhead, adp, ad_next); 11214 WORKLIST_INSERT(wkhd, &adp->ad_block.nb_list); 11215 } 11216 } 11217 11218 /* 11219 * Called from within softdep_disk_write_complete above. This routine 11220 * completes successfully written allocindirs. 11221 */ 11222 static void 11223 handle_allocindir_partdone(aip) 11224 struct allocindir *aip; /* the completed allocindir */ 11225 { 11226 struct indirdep *indirdep; 11227 11228 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE) 11229 return; 11230 indirdep = aip->ai_indirdep; 11231 LIST_REMOVE(aip, ai_next); 11232 /* 11233 * Don't set a pointer while the buffer is undergoing IO or while 11234 * we have active truncations. 11235 */ 11236 if (indirdep->ir_state & UNDONE || !TAILQ_EMPTY(&indirdep->ir_trunc)) { 11237 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next); 11238 return; 11239 } 11240 if (indirdep->ir_state & UFS1FMT) 11241 ((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 11242 aip->ai_newblkno; 11243 else 11244 ((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] = 11245 aip->ai_newblkno; 11246 /* 11247 * Await the pointer write before freeing the allocindir. 11248 */ 11249 LIST_INSERT_HEAD(&indirdep->ir_writehd, aip, ai_next); 11250 } 11251 11252 /* 11253 * Release segments held on a jwork list. 11254 */ 11255 static void 11256 handle_jwork(wkhd) 11257 struct workhead *wkhd; 11258 { 11259 struct worklist *wk; 11260 11261 while ((wk = LIST_FIRST(wkhd)) != NULL) { 11262 WORKLIST_REMOVE(wk); 11263 switch (wk->wk_type) { 11264 case D_JSEGDEP: 11265 free_jsegdep(WK_JSEGDEP(wk)); 11266 continue; 11267 case D_FREEDEP: 11268 free_freedep(WK_FREEDEP(wk)); 11269 continue; 11270 case D_FREEFRAG: 11271 rele_jseg(WK_JSEG(WK_FREEFRAG(wk)->ff_jdep)); 11272 WORKITEM_FREE(wk, D_FREEFRAG); 11273 continue; 11274 case D_FREEWORK: 11275 handle_written_freework(WK_FREEWORK(wk)); 11276 continue; 11277 default: 11278 panic("handle_jwork: Unknown type %s\n", 11279 TYPENAME(wk->wk_type)); 11280 } 11281 } 11282 } 11283 11284 /* 11285 * Handle the bufwait list on an inode when it is safe to release items 11286 * held there. This normally happens after an inode block is written but 11287 * may be delayed and handled later if there are pending journal items that 11288 * are not yet safe to be released. 11289 */ 11290 static struct freefile * 11291 handle_bufwait(inodedep, refhd) 11292 struct inodedep *inodedep; 11293 struct workhead *refhd; 11294 { 11295 struct jaddref *jaddref; 11296 struct freefile *freefile; 11297 struct worklist *wk; 11298 11299 freefile = NULL; 11300 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) { 11301 WORKLIST_REMOVE(wk); 11302 switch (wk->wk_type) { 11303 case D_FREEFILE: 11304 /* 11305 * We defer adding freefile to the worklist 11306 * until all other additions have been made to 11307 * ensure that it will be done after all the 11308 * old blocks have been freed. 11309 */ 11310 if (freefile != NULL) 11311 panic("handle_bufwait: freefile"); 11312 freefile = WK_FREEFILE(wk); 11313 continue; 11314 11315 case D_MKDIR: 11316 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT); 11317 continue; 11318 11319 case D_DIRADD: 11320 diradd_inode_written(WK_DIRADD(wk), inodedep); 11321 continue; 11322 11323 case D_FREEFRAG: 11324 wk->wk_state |= COMPLETE; 11325 if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE) 11326 add_to_worklist(wk, 0); 11327 continue; 11328 11329 case D_DIRREM: 11330 wk->wk_state |= COMPLETE; 11331 add_to_worklist(wk, 0); 11332 continue; 11333 11334 case D_ALLOCDIRECT: 11335 case D_ALLOCINDIR: 11336 free_newblk(WK_NEWBLK(wk)); 11337 continue; 11338 11339 case D_JNEWBLK: 11340 wk->wk_state |= COMPLETE; 11341 free_jnewblk(WK_JNEWBLK(wk)); 11342 continue; 11343 11344 /* 11345 * Save freed journal segments and add references on 11346 * the supplied list which will delay their release 11347 * until the cg bitmap is cleared on disk. 11348 */ 11349 case D_JSEGDEP: 11350 if (refhd == NULL) 11351 free_jsegdep(WK_JSEGDEP(wk)); 11352 else 11353 WORKLIST_INSERT(refhd, wk); 11354 continue; 11355 11356 case D_JADDREF: 11357 jaddref = WK_JADDREF(wk); 11358 TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, 11359 if_deps); 11360 /* 11361 * Transfer any jaddrefs to the list to be freed with 11362 * the bitmap if we're handling a removed file. 11363 */ 11364 if (refhd == NULL) { 11365 wk->wk_state |= COMPLETE; 11366 free_jaddref(jaddref); 11367 } else 11368 WORKLIST_INSERT(refhd, wk); 11369 continue; 11370 11371 default: 11372 panic("handle_bufwait: Unknown type %p(%s)", 11373 wk, TYPENAME(wk->wk_type)); 11374 /* NOTREACHED */ 11375 } 11376 } 11377 return (freefile); 11378 } 11379 /* 11380 * Called from within softdep_disk_write_complete above to restore 11381 * in-memory inode block contents to their most up-to-date state. Note 11382 * that this routine is always called from interrupt level with further 11383 * interrupts from this device blocked. 11384 * 11385 * If the write did not succeed, we will do all the roll-forward 11386 * operations, but we will not take the actions that will allow its 11387 * dependencies to be processed. 11388 */ 11389 static int 11390 handle_written_inodeblock(inodedep, bp, flags) 11391 struct inodedep *inodedep; 11392 struct buf *bp; /* buffer containing the inode block */ 11393 int flags; 11394 { 11395 struct freefile *freefile; 11396 struct allocdirect *adp, *nextadp; 11397 struct ufs1_dinode *dp1 = NULL; 11398 struct ufs2_dinode *dp2 = NULL; 11399 struct workhead wkhd; 11400 int hadchanges, fstype; 11401 ino_t freelink; 11402 11403 LIST_INIT(&wkhd); 11404 hadchanges = 0; 11405 freefile = NULL; 11406 if ((inodedep->id_state & IOSTARTED) == 0) 11407 panic("handle_written_inodeblock: not started"); 11408 inodedep->id_state &= ~IOSTARTED; 11409 if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) { 11410 fstype = UFS1; 11411 dp1 = (struct ufs1_dinode *)bp->b_data + 11412 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 11413 freelink = dp1->di_freelink; 11414 } else { 11415 fstype = UFS2; 11416 dp2 = (struct ufs2_dinode *)bp->b_data + 11417 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino); 11418 freelink = dp2->di_freelink; 11419 } 11420 /* 11421 * Leave this inodeblock dirty until it's in the list. 11422 */ 11423 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) == UNLINKED && 11424 (flags & WRITESUCCEEDED)) { 11425 struct inodedep *inon; 11426 11427 inon = TAILQ_NEXT(inodedep, id_unlinked); 11428 if ((inon == NULL && freelink == 0) || 11429 (inon && inon->id_ino == freelink)) { 11430 if (inon) 11431 inon->id_state |= UNLINKPREV; 11432 inodedep->id_state |= UNLINKNEXT; 11433 } 11434 hadchanges = 1; 11435 } 11436 /* 11437 * If we had to rollback the inode allocation because of 11438 * bitmaps being incomplete, then simply restore it. 11439 * Keep the block dirty so that it will not be reclaimed until 11440 * all associated dependencies have been cleared and the 11441 * corresponding updates written to disk. 11442 */ 11443 if (inodedep->id_savedino1 != NULL) { 11444 hadchanges = 1; 11445 if (fstype == UFS1) 11446 *dp1 = *inodedep->id_savedino1; 11447 else 11448 *dp2 = *inodedep->id_savedino2; 11449 free(inodedep->id_savedino1, M_SAVEDINO); 11450 inodedep->id_savedino1 = NULL; 11451 if ((bp->b_flags & B_DELWRI) == 0) 11452 stat_inode_bitmap++; 11453 bdirty(bp); 11454 /* 11455 * If the inode is clear here and GOINGAWAY it will never 11456 * be written. Process the bufwait and clear any pending 11457 * work which may include the freefile. 11458 */ 11459 if (inodedep->id_state & GOINGAWAY) 11460 goto bufwait; 11461 return (1); 11462 } 11463 if (flags & WRITESUCCEEDED) 11464 inodedep->id_state |= COMPLETE; 11465 /* 11466 * Roll forward anything that had to be rolled back before 11467 * the inode could be updated. 11468 */ 11469 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) { 11470 nextadp = TAILQ_NEXT(adp, ad_next); 11471 if (adp->ad_state & ATTACHED) 11472 panic("handle_written_inodeblock: new entry"); 11473 if (fstype == UFS1) { 11474 if (adp->ad_offset < NDADDR) { 11475 if (dp1->di_db[adp->ad_offset]!=adp->ad_oldblkno) 11476 panic("%s %s #%jd mismatch %d != %jd", 11477 "handle_written_inodeblock:", 11478 "direct pointer", 11479 (intmax_t)adp->ad_offset, 11480 dp1->di_db[adp->ad_offset], 11481 (intmax_t)adp->ad_oldblkno); 11482 dp1->di_db[adp->ad_offset] = adp->ad_newblkno; 11483 } else { 11484 if (dp1->di_ib[adp->ad_offset - NDADDR] != 0) 11485 panic("%s: %s #%jd allocated as %d", 11486 "handle_written_inodeblock", 11487 "indirect pointer", 11488 (intmax_t)adp->ad_offset - NDADDR, 11489 dp1->di_ib[adp->ad_offset - NDADDR]); 11490 dp1->di_ib[adp->ad_offset - NDADDR] = 11491 adp->ad_newblkno; 11492 } 11493 } else { 11494 if (adp->ad_offset < NDADDR) { 11495 if (dp2->di_db[adp->ad_offset]!=adp->ad_oldblkno) 11496 panic("%s: %s #%jd %s %jd != %jd", 11497 "handle_written_inodeblock", 11498 "direct pointer", 11499 (intmax_t)adp->ad_offset, "mismatch", 11500 (intmax_t)dp2->di_db[adp->ad_offset], 11501 (intmax_t)adp->ad_oldblkno); 11502 dp2->di_db[adp->ad_offset] = adp->ad_newblkno; 11503 } else { 11504 if (dp2->di_ib[adp->ad_offset - NDADDR] != 0) 11505 panic("%s: %s #%jd allocated as %jd", 11506 "handle_written_inodeblock", 11507 "indirect pointer", 11508 (intmax_t)adp->ad_offset - NDADDR, 11509 (intmax_t) 11510 dp2->di_ib[adp->ad_offset - NDADDR]); 11511 dp2->di_ib[adp->ad_offset - NDADDR] = 11512 adp->ad_newblkno; 11513 } 11514 } 11515 adp->ad_state &= ~UNDONE; 11516 adp->ad_state |= ATTACHED; 11517 hadchanges = 1; 11518 } 11519 for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) { 11520 nextadp = TAILQ_NEXT(adp, ad_next); 11521 if (adp->ad_state & ATTACHED) 11522 panic("handle_written_inodeblock: new entry"); 11523 if (dp2->di_extb[adp->ad_offset] != adp->ad_oldblkno) 11524 panic("%s: direct pointers #%jd %s %jd != %jd", 11525 "handle_written_inodeblock", 11526 (intmax_t)adp->ad_offset, "mismatch", 11527 (intmax_t)dp2->di_extb[adp->ad_offset], 11528 (intmax_t)adp->ad_oldblkno); 11529 dp2->di_extb[adp->ad_offset] = adp->ad_newblkno; 11530 adp->ad_state &= ~UNDONE; 11531 adp->ad_state |= ATTACHED; 11532 hadchanges = 1; 11533 } 11534 if (hadchanges && (bp->b_flags & B_DELWRI) == 0) 11535 stat_direct_blk_ptrs++; 11536 /* 11537 * Reset the file size to its most up-to-date value. 11538 */ 11539 if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) 11540 panic("handle_written_inodeblock: bad size"); 11541 if (inodedep->id_savednlink > LINK_MAX) 11542 panic("handle_written_inodeblock: Invalid link count " 11543 "%jd for inodedep %p", (uintmax_t)inodedep->id_savednlink, 11544 inodedep); 11545 if (fstype == UFS1) { 11546 if (dp1->di_nlink != inodedep->id_savednlink) { 11547 dp1->di_nlink = inodedep->id_savednlink; 11548 hadchanges = 1; 11549 } 11550 if (dp1->di_size != inodedep->id_savedsize) { 11551 dp1->di_size = inodedep->id_savedsize; 11552 hadchanges = 1; 11553 } 11554 } else { 11555 if (dp2->di_nlink != inodedep->id_savednlink) { 11556 dp2->di_nlink = inodedep->id_savednlink; 11557 hadchanges = 1; 11558 } 11559 if (dp2->di_size != inodedep->id_savedsize) { 11560 dp2->di_size = inodedep->id_savedsize; 11561 hadchanges = 1; 11562 } 11563 if (dp2->di_extsize != inodedep->id_savedextsize) { 11564 dp2->di_extsize = inodedep->id_savedextsize; 11565 hadchanges = 1; 11566 } 11567 } 11568 inodedep->id_savedsize = -1; 11569 inodedep->id_savedextsize = -1; 11570 inodedep->id_savednlink = -1; 11571 /* 11572 * If there were any rollbacks in the inode block, then it must be 11573 * marked dirty so that its will eventually get written back in 11574 * its correct form. 11575 */ 11576 if (hadchanges) 11577 bdirty(bp); 11578 bufwait: 11579 /* 11580 * If the write did not succeed, we have done all the roll-forward 11581 * operations, but we cannot take the actions that will allow its 11582 * dependencies to be processed. 11583 */ 11584 if ((flags & WRITESUCCEEDED) == 0) 11585 return (hadchanges); 11586 /* 11587 * Process any allocdirects that completed during the update. 11588 */ 11589 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL) 11590 handle_allocdirect_partdone(adp, &wkhd); 11591 if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL) 11592 handle_allocdirect_partdone(adp, &wkhd); 11593 /* 11594 * Process deallocations that were held pending until the 11595 * inode had been written to disk. Freeing of the inode 11596 * is delayed until after all blocks have been freed to 11597 * avoid creation of new <vfsid, inum, lbn> triples 11598 * before the old ones have been deleted. Completely 11599 * unlinked inodes are not processed until the unlinked 11600 * inode list is written or the last reference is removed. 11601 */ 11602 if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) != UNLINKED) { 11603 freefile = handle_bufwait(inodedep, NULL); 11604 if (freefile && !LIST_EMPTY(&wkhd)) { 11605 WORKLIST_INSERT(&wkhd, &freefile->fx_list); 11606 freefile = NULL; 11607 } 11608 } 11609 /* 11610 * Move rolled forward dependency completions to the bufwait list 11611 * now that those that were already written have been processed. 11612 */ 11613 if (!LIST_EMPTY(&wkhd) && hadchanges == 0) 11614 panic("handle_written_inodeblock: bufwait but no changes"); 11615 jwork_move(&inodedep->id_bufwait, &wkhd); 11616 11617 if (freefile != NULL) { 11618 /* 11619 * If the inode is goingaway it was never written. Fake up 11620 * the state here so free_inodedep() can succeed. 11621 */ 11622 if (inodedep->id_state & GOINGAWAY) 11623 inodedep->id_state |= COMPLETE | DEPCOMPLETE; 11624 if (free_inodedep(inodedep) == 0) 11625 panic("handle_written_inodeblock: live inodedep %p", 11626 inodedep); 11627 add_to_worklist(&freefile->fx_list, 0); 11628 return (0); 11629 } 11630 11631 /* 11632 * If no outstanding dependencies, free it. 11633 */ 11634 if (free_inodedep(inodedep) || 11635 (TAILQ_FIRST(&inodedep->id_inoreflst) == 0 && 11636 TAILQ_FIRST(&inodedep->id_inoupdt) == 0 && 11637 TAILQ_FIRST(&inodedep->id_extupdt) == 0 && 11638 LIST_FIRST(&inodedep->id_bufwait) == 0)) 11639 return (0); 11640 return (hadchanges); 11641 } 11642 11643 /* 11644 * Perform needed roll-forwards and kick off any dependencies that 11645 * can now be processed. 11646 * 11647 * If the write did not succeed, we will do all the roll-forward 11648 * operations, but we will not take the actions that will allow its 11649 * dependencies to be processed. 11650 */ 11651 static int 11652 handle_written_indirdep(indirdep, bp, bpp, flags) 11653 struct indirdep *indirdep; 11654 struct buf *bp; 11655 struct buf **bpp; 11656 int flags; 11657 { 11658 struct allocindir *aip; 11659 struct buf *sbp; 11660 int chgs; 11661 11662 if (indirdep->ir_state & GOINGAWAY) 11663 panic("handle_written_indirdep: indirdep gone"); 11664 if ((indirdep->ir_state & IOSTARTED) == 0) 11665 panic("handle_written_indirdep: IO not started"); 11666 chgs = 0; 11667 /* 11668 * If there were rollbacks revert them here. 11669 */ 11670 if (indirdep->ir_saveddata) { 11671 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount); 11672 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 11673 free(indirdep->ir_saveddata, M_INDIRDEP); 11674 indirdep->ir_saveddata = NULL; 11675 } 11676 chgs = 1; 11677 } 11678 indirdep->ir_state &= ~(UNDONE | IOSTARTED); 11679 indirdep->ir_state |= ATTACHED; 11680 /* 11681 * If the write did not succeed, we have done all the roll-forward 11682 * operations, but we cannot take the actions that will allow its 11683 * dependencies to be processed. 11684 */ 11685 if ((flags & WRITESUCCEEDED) == 0) { 11686 stat_indir_blk_ptrs++; 11687 bdirty(bp); 11688 return (1); 11689 } 11690 /* 11691 * Move allocindirs with written pointers to the completehd if 11692 * the indirdep's pointer is not yet written. Otherwise 11693 * free them here. 11694 */ 11695 while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != NULL) { 11696 LIST_REMOVE(aip, ai_next); 11697 if ((indirdep->ir_state & DEPCOMPLETE) == 0) { 11698 LIST_INSERT_HEAD(&indirdep->ir_completehd, aip, 11699 ai_next); 11700 newblk_freefrag(&aip->ai_block); 11701 continue; 11702 } 11703 free_newblk(&aip->ai_block); 11704 } 11705 /* 11706 * Move allocindirs that have finished dependency processing from 11707 * the done list to the write list after updating the pointers. 11708 */ 11709 if (TAILQ_EMPTY(&indirdep->ir_trunc)) { 11710 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != NULL) { 11711 handle_allocindir_partdone(aip); 11712 if (aip == LIST_FIRST(&indirdep->ir_donehd)) 11713 panic("disk_write_complete: not gone"); 11714 chgs = 1; 11715 } 11716 } 11717 /* 11718 * Preserve the indirdep if there were any changes or if it is not 11719 * yet valid on disk. 11720 */ 11721 if (chgs) { 11722 stat_indir_blk_ptrs++; 11723 bdirty(bp); 11724 return (1); 11725 } 11726 /* 11727 * If there were no changes we can discard the savedbp and detach 11728 * ourselves from the buf. We are only carrying completed pointers 11729 * in this case. 11730 */ 11731 sbp = indirdep->ir_savebp; 11732 sbp->b_flags |= B_INVAL | B_NOCACHE; 11733 indirdep->ir_savebp = NULL; 11734 indirdep->ir_bp = NULL; 11735 if (*bpp != NULL) 11736 panic("handle_written_indirdep: bp already exists."); 11737 *bpp = sbp; 11738 /* 11739 * The indirdep may not be freed until its parent points at it. 11740 */ 11741 if (indirdep->ir_state & DEPCOMPLETE) 11742 free_indirdep(indirdep); 11743 11744 return (0); 11745 } 11746 11747 /* 11748 * Process a diradd entry after its dependent inode has been written. 11749 * This routine must be called with splbio interrupts blocked. 11750 */ 11751 static void 11752 diradd_inode_written(dap, inodedep) 11753 struct diradd *dap; 11754 struct inodedep *inodedep; 11755 { 11756 11757 dap->da_state |= COMPLETE; 11758 complete_diradd(dap); 11759 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list); 11760 } 11761 11762 /* 11763 * Returns true if the bmsafemap will have rollbacks when written. Must only 11764 * be called with the per-filesystem lock and the buf lock on the cg held. 11765 */ 11766 static int 11767 bmsafemap_backgroundwrite(bmsafemap, bp) 11768 struct bmsafemap *bmsafemap; 11769 struct buf *bp; 11770 { 11771 int dirty; 11772 11773 LOCK_OWNED(VFSTOUFS(bmsafemap->sm_list.wk_mp)); 11774 dirty = !LIST_EMPTY(&bmsafemap->sm_jaddrefhd) | 11775 !LIST_EMPTY(&bmsafemap->sm_jnewblkhd); 11776 /* 11777 * If we're initiating a background write we need to process the 11778 * rollbacks as they exist now, not as they exist when IO starts. 11779 * No other consumers will look at the contents of the shadowed 11780 * buf so this is safe to do here. 11781 */ 11782 if (bp->b_xflags & BX_BKGRDMARKER) 11783 initiate_write_bmsafemap(bmsafemap, bp); 11784 11785 return (dirty); 11786 } 11787 11788 /* 11789 * Re-apply an allocation when a cg write is complete. 11790 */ 11791 static int 11792 jnewblk_rollforward(jnewblk, fs, cgp, blksfree) 11793 struct jnewblk *jnewblk; 11794 struct fs *fs; 11795 struct cg *cgp; 11796 uint8_t *blksfree; 11797 { 11798 ufs1_daddr_t fragno; 11799 ufs2_daddr_t blkno; 11800 long cgbno, bbase; 11801 int frags, blk; 11802 int i; 11803 11804 frags = 0; 11805 cgbno = dtogd(fs, jnewblk->jn_blkno); 11806 for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags; i++) { 11807 if (isclr(blksfree, cgbno + i)) 11808 panic("jnewblk_rollforward: re-allocated fragment"); 11809 frags++; 11810 } 11811 if (frags == fs->fs_frag) { 11812 blkno = fragstoblks(fs, cgbno); 11813 ffs_clrblock(fs, blksfree, (long)blkno); 11814 ffs_clusteracct(fs, cgp, blkno, -1); 11815 cgp->cg_cs.cs_nbfree--; 11816 } else { 11817 bbase = cgbno - fragnum(fs, cgbno); 11818 cgbno += jnewblk->jn_oldfrags; 11819 /* If a complete block had been reassembled, account for it. */ 11820 fragno = fragstoblks(fs, bbase); 11821 if (ffs_isblock(fs, blksfree, fragno)) { 11822 cgp->cg_cs.cs_nffree += fs->fs_frag; 11823 ffs_clusteracct(fs, cgp, fragno, -1); 11824 cgp->cg_cs.cs_nbfree--; 11825 } 11826 /* Decrement the old frags. */ 11827 blk = blkmap(fs, blksfree, bbase); 11828 ffs_fragacct(fs, blk, cgp->cg_frsum, -1); 11829 /* Allocate the fragment */ 11830 for (i = 0; i < frags; i++) 11831 clrbit(blksfree, cgbno + i); 11832 cgp->cg_cs.cs_nffree -= frags; 11833 /* Add back in counts associated with the new frags */ 11834 blk = blkmap(fs, blksfree, bbase); 11835 ffs_fragacct(fs, blk, cgp->cg_frsum, 1); 11836 } 11837 return (frags); 11838 } 11839 11840 /* 11841 * Complete a write to a bmsafemap structure. Roll forward any bitmap 11842 * changes if it's not a background write. Set all written dependencies 11843 * to DEPCOMPLETE and free the structure if possible. 11844 * 11845 * If the write did not succeed, we will do all the roll-forward 11846 * operations, but we will not take the actions that will allow its 11847 * dependencies to be processed. 11848 */ 11849 static int 11850 handle_written_bmsafemap(bmsafemap, bp, flags) 11851 struct bmsafemap *bmsafemap; 11852 struct buf *bp; 11853 int flags; 11854 { 11855 struct newblk *newblk; 11856 struct inodedep *inodedep; 11857 struct jaddref *jaddref, *jatmp; 11858 struct jnewblk *jnewblk, *jntmp; 11859 struct ufsmount *ump; 11860 uint8_t *inosused; 11861 uint8_t *blksfree; 11862 struct cg *cgp; 11863 struct fs *fs; 11864 ino_t ino; 11865 int foreground; 11866 int chgs; 11867 11868 if ((bmsafemap->sm_state & IOSTARTED) == 0) 11869 panic("handle_written_bmsafemap: Not started\n"); 11870 ump = VFSTOUFS(bmsafemap->sm_list.wk_mp); 11871 chgs = 0; 11872 bmsafemap->sm_state &= ~IOSTARTED; 11873 foreground = (bp->b_xflags & BX_BKGRDMARKER) == 0; 11874 /* 11875 * If write was successful, release journal work that was waiting 11876 * on the write. Otherwise move the work back. 11877 */ 11878 if (flags & WRITESUCCEEDED) 11879 handle_jwork(&bmsafemap->sm_freewr); 11880 else 11881 LIST_CONCAT(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, 11882 worklist, wk_list); 11883 11884 /* 11885 * Restore unwritten inode allocation pending jaddref writes. 11886 */ 11887 if (!LIST_EMPTY(&bmsafemap->sm_jaddrefhd)) { 11888 cgp = (struct cg *)bp->b_data; 11889 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 11890 inosused = cg_inosused(cgp); 11891 LIST_FOREACH_SAFE(jaddref, &bmsafemap->sm_jaddrefhd, 11892 ja_bmdeps, jatmp) { 11893 if ((jaddref->ja_state & UNDONE) == 0) 11894 continue; 11895 ino = jaddref->ja_ino % fs->fs_ipg; 11896 if (isset(inosused, ino)) 11897 panic("handle_written_bmsafemap: " 11898 "re-allocated inode"); 11899 /* Do the roll-forward only if it's a real copy. */ 11900 if (foreground) { 11901 if ((jaddref->ja_mode & IFMT) == IFDIR) 11902 cgp->cg_cs.cs_ndir++; 11903 cgp->cg_cs.cs_nifree--; 11904 setbit(inosused, ino); 11905 chgs = 1; 11906 } 11907 jaddref->ja_state &= ~UNDONE; 11908 jaddref->ja_state |= ATTACHED; 11909 free_jaddref(jaddref); 11910 } 11911 } 11912 /* 11913 * Restore any block allocations which are pending journal writes. 11914 */ 11915 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) { 11916 cgp = (struct cg *)bp->b_data; 11917 fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs; 11918 blksfree = cg_blksfree(cgp); 11919 LIST_FOREACH_SAFE(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps, 11920 jntmp) { 11921 if ((jnewblk->jn_state & UNDONE) == 0) 11922 continue; 11923 /* Do the roll-forward only if it's a real copy. */ 11924 if (foreground && 11925 jnewblk_rollforward(jnewblk, fs, cgp, blksfree)) 11926 chgs = 1; 11927 jnewblk->jn_state &= ~(UNDONE | NEWBLOCK); 11928 jnewblk->jn_state |= ATTACHED; 11929 free_jnewblk(jnewblk); 11930 } 11931 } 11932 /* 11933 * If the write did not succeed, we have done all the roll-forward 11934 * operations, but we cannot take the actions that will allow its 11935 * dependencies to be processed. 11936 */ 11937 if ((flags & WRITESUCCEEDED) == 0) { 11938 LIST_CONCAT(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr, 11939 newblk, nb_deps); 11940 LIST_CONCAT(&bmsafemap->sm_freehd, &bmsafemap->sm_freewr, 11941 worklist, wk_list); 11942 if (foreground) 11943 bdirty(bp); 11944 return (1); 11945 } 11946 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkwr))) { 11947 newblk->nb_state |= DEPCOMPLETE; 11948 newblk->nb_state &= ~ONDEPLIST; 11949 newblk->nb_bmsafemap = NULL; 11950 LIST_REMOVE(newblk, nb_deps); 11951 if (newblk->nb_list.wk_type == D_ALLOCDIRECT) 11952 handle_allocdirect_partdone( 11953 WK_ALLOCDIRECT(&newblk->nb_list), NULL); 11954 else if (newblk->nb_list.wk_type == D_ALLOCINDIR) 11955 handle_allocindir_partdone( 11956 WK_ALLOCINDIR(&newblk->nb_list)); 11957 else if (newblk->nb_list.wk_type != D_NEWBLK) 11958 panic("handle_written_bmsafemap: Unexpected type: %s", 11959 TYPENAME(newblk->nb_list.wk_type)); 11960 } 11961 while ((inodedep = LIST_FIRST(&bmsafemap->sm_inodedepwr)) != NULL) { 11962 inodedep->id_state |= DEPCOMPLETE; 11963 inodedep->id_state &= ~ONDEPLIST; 11964 LIST_REMOVE(inodedep, id_deps); 11965 inodedep->id_bmsafemap = NULL; 11966 } 11967 LIST_REMOVE(bmsafemap, sm_next); 11968 if (chgs == 0 && LIST_EMPTY(&bmsafemap->sm_jaddrefhd) && 11969 LIST_EMPTY(&bmsafemap->sm_jnewblkhd) && 11970 LIST_EMPTY(&bmsafemap->sm_newblkhd) && 11971 LIST_EMPTY(&bmsafemap->sm_inodedephd) && 11972 LIST_EMPTY(&bmsafemap->sm_freehd)) { 11973 LIST_REMOVE(bmsafemap, sm_hash); 11974 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP); 11975 return (0); 11976 } 11977 LIST_INSERT_HEAD(&ump->softdep_dirtycg, bmsafemap, sm_next); 11978 if (foreground) 11979 bdirty(bp); 11980 return (1); 11981 } 11982 11983 /* 11984 * Try to free a mkdir dependency. 11985 */ 11986 static void 11987 complete_mkdir(mkdir) 11988 struct mkdir *mkdir; 11989 { 11990 struct diradd *dap; 11991 11992 if ((mkdir->md_state & ALLCOMPLETE) != ALLCOMPLETE) 11993 return; 11994 LIST_REMOVE(mkdir, md_mkdirs); 11995 dap = mkdir->md_diradd; 11996 dap->da_state &= ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)); 11997 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) { 11998 dap->da_state |= DEPCOMPLETE; 11999 complete_diradd(dap); 12000 } 12001 WORKITEM_FREE(mkdir, D_MKDIR); 12002 } 12003 12004 /* 12005 * Handle the completion of a mkdir dependency. 12006 */ 12007 static void 12008 handle_written_mkdir(mkdir, type) 12009 struct mkdir *mkdir; 12010 int type; 12011 { 12012 12013 if ((mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)) != type) 12014 panic("handle_written_mkdir: bad type"); 12015 mkdir->md_state |= COMPLETE; 12016 complete_mkdir(mkdir); 12017 } 12018 12019 static int 12020 free_pagedep(pagedep) 12021 struct pagedep *pagedep; 12022 { 12023 int i; 12024 12025 if (pagedep->pd_state & NEWBLOCK) 12026 return (0); 12027 if (!LIST_EMPTY(&pagedep->pd_dirremhd)) 12028 return (0); 12029 for (i = 0; i < DAHASHSZ; i++) 12030 if (!LIST_EMPTY(&pagedep->pd_diraddhd[i])) 12031 return (0); 12032 if (!LIST_EMPTY(&pagedep->pd_pendinghd)) 12033 return (0); 12034 if (!LIST_EMPTY(&pagedep->pd_jmvrefhd)) 12035 return (0); 12036 if (pagedep->pd_state & ONWORKLIST) 12037 WORKLIST_REMOVE(&pagedep->pd_list); 12038 LIST_REMOVE(pagedep, pd_hash); 12039 WORKITEM_FREE(pagedep, D_PAGEDEP); 12040 12041 return (1); 12042 } 12043 12044 /* 12045 * Called from within softdep_disk_write_complete above. 12046 * A write operation was just completed. Removed inodes can 12047 * now be freed and associated block pointers may be committed. 12048 * Note that this routine is always called from interrupt level 12049 * with further interrupts from this device blocked. 12050 * 12051 * If the write did not succeed, we will do all the roll-forward 12052 * operations, but we will not take the actions that will allow its 12053 * dependencies to be processed. 12054 */ 12055 static int 12056 handle_written_filepage(pagedep, bp, flags) 12057 struct pagedep *pagedep; 12058 struct buf *bp; /* buffer containing the written page */ 12059 int flags; 12060 { 12061 struct dirrem *dirrem; 12062 struct diradd *dap, *nextdap; 12063 struct direct *ep; 12064 int i, chgs; 12065 12066 if ((pagedep->pd_state & IOSTARTED) == 0) 12067 panic("handle_written_filepage: not started"); 12068 pagedep->pd_state &= ~IOSTARTED; 12069 if ((flags & WRITESUCCEEDED) == 0) 12070 goto rollforward; 12071 /* 12072 * Process any directory removals that have been committed. 12073 */ 12074 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) { 12075 LIST_REMOVE(dirrem, dm_next); 12076 dirrem->dm_state |= COMPLETE; 12077 dirrem->dm_dirinum = pagedep->pd_ino; 12078 KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd), 12079 ("handle_written_filepage: Journal entries not written.")); 12080 add_to_worklist(&dirrem->dm_list, 0); 12081 } 12082 /* 12083 * Free any directory additions that have been committed. 12084 * If it is a newly allocated block, we have to wait until 12085 * the on-disk directory inode claims the new block. 12086 */ 12087 if ((pagedep->pd_state & NEWBLOCK) == 0) 12088 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL) 12089 free_diradd(dap, NULL); 12090 rollforward: 12091 /* 12092 * Uncommitted directory entries must be restored. 12093 */ 12094 for (chgs = 0, i = 0; i < DAHASHSZ; i++) { 12095 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap; 12096 dap = nextdap) { 12097 nextdap = LIST_NEXT(dap, da_pdlist); 12098 if (dap->da_state & ATTACHED) 12099 panic("handle_written_filepage: attached"); 12100 ep = (struct direct *) 12101 ((char *)bp->b_data + dap->da_offset); 12102 ep->d_ino = dap->da_newinum; 12103 dap->da_state &= ~UNDONE; 12104 dap->da_state |= ATTACHED; 12105 chgs = 1; 12106 /* 12107 * If the inode referenced by the directory has 12108 * been written out, then the dependency can be 12109 * moved to the pending list. 12110 */ 12111 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) { 12112 LIST_REMOVE(dap, da_pdlist); 12113 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, 12114 da_pdlist); 12115 } 12116 } 12117 } 12118 /* 12119 * If there were any rollbacks in the directory, then it must be 12120 * marked dirty so that its will eventually get written back in 12121 * its correct form. 12122 */ 12123 if (chgs || (flags & WRITESUCCEEDED) == 0) { 12124 if ((bp->b_flags & B_DELWRI) == 0) 12125 stat_dir_entry++; 12126 bdirty(bp); 12127 return (1); 12128 } 12129 /* 12130 * If we are not waiting for a new directory block to be 12131 * claimed by its inode, then the pagedep will be freed. 12132 * Otherwise it will remain to track any new entries on 12133 * the page in case they are fsync'ed. 12134 */ 12135 free_pagedep(pagedep); 12136 return (0); 12137 } 12138 12139 /* 12140 * Writing back in-core inode structures. 12141 * 12142 * The filesystem only accesses an inode's contents when it occupies an 12143 * "in-core" inode structure. These "in-core" structures are separate from 12144 * the page frames used to cache inode blocks. Only the latter are 12145 * transferred to/from the disk. So, when the updated contents of the 12146 * "in-core" inode structure are copied to the corresponding in-memory inode 12147 * block, the dependencies are also transferred. The following procedure is 12148 * called when copying a dirty "in-core" inode to a cached inode block. 12149 */ 12150 12151 /* 12152 * Called when an inode is loaded from disk. If the effective link count 12153 * differed from the actual link count when it was last flushed, then we 12154 * need to ensure that the correct effective link count is put back. 12155 */ 12156 void 12157 softdep_load_inodeblock(ip) 12158 struct inode *ip; /* the "in_core" copy of the inode */ 12159 { 12160 struct inodedep *inodedep; 12161 struct ufsmount *ump; 12162 12163 ump = ITOUMP(ip); 12164 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 12165 ("softdep_load_inodeblock called on non-softdep filesystem")); 12166 /* 12167 * Check for alternate nlink count. 12168 */ 12169 ip->i_effnlink = ip->i_nlink; 12170 ACQUIRE_LOCK(ump); 12171 if (inodedep_lookup(UFSTOVFS(ump), ip->i_number, 0, &inodedep) == 0) { 12172 FREE_LOCK(ump); 12173 return; 12174 } 12175 ip->i_effnlink -= inodedep->id_nlinkdelta; 12176 FREE_LOCK(ump); 12177 } 12178 12179 /* 12180 * This routine is called just before the "in-core" inode 12181 * information is to be copied to the in-memory inode block. 12182 * Recall that an inode block contains several inodes. If 12183 * the force flag is set, then the dependencies will be 12184 * cleared so that the update can always be made. Note that 12185 * the buffer is locked when this routine is called, so we 12186 * will never be in the middle of writing the inode block 12187 * to disk. 12188 */ 12189 void 12190 softdep_update_inodeblock(ip, bp, waitfor) 12191 struct inode *ip; /* the "in_core" copy of the inode */ 12192 struct buf *bp; /* the buffer containing the inode block */ 12193 int waitfor; /* nonzero => update must be allowed */ 12194 { 12195 struct inodedep *inodedep; 12196 struct inoref *inoref; 12197 struct ufsmount *ump; 12198 struct worklist *wk; 12199 struct mount *mp; 12200 struct buf *ibp; 12201 struct fs *fs; 12202 int error; 12203 12204 ump = ITOUMP(ip); 12205 mp = UFSTOVFS(ump); 12206 KASSERT(MOUNTEDSOFTDEP(mp) != 0, 12207 ("softdep_update_inodeblock called on non-softdep filesystem")); 12208 fs = ump->um_fs; 12209 /* 12210 * Preserve the freelink that is on disk. clear_unlinked_inodedep() 12211 * does not have access to the in-core ip so must write directly into 12212 * the inode block buffer when setting freelink. 12213 */ 12214 if (fs->fs_magic == FS_UFS1_MAGIC) 12215 DIP_SET(ip, i_freelink, ((struct ufs1_dinode *)bp->b_data + 12216 ino_to_fsbo(fs, ip->i_number))->di_freelink); 12217 else 12218 DIP_SET(ip, i_freelink, ((struct ufs2_dinode *)bp->b_data + 12219 ino_to_fsbo(fs, ip->i_number))->di_freelink); 12220 /* 12221 * If the effective link count is not equal to the actual link 12222 * count, then we must track the difference in an inodedep while 12223 * the inode is (potentially) tossed out of the cache. Otherwise, 12224 * if there is no existing inodedep, then there are no dependencies 12225 * to track. 12226 */ 12227 ACQUIRE_LOCK(ump); 12228 again: 12229 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 12230 FREE_LOCK(ump); 12231 if (ip->i_effnlink != ip->i_nlink) 12232 panic("softdep_update_inodeblock: bad link count"); 12233 return; 12234 } 12235 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) 12236 panic("softdep_update_inodeblock: bad delta"); 12237 /* 12238 * If we're flushing all dependencies we must also move any waiting 12239 * for journal writes onto the bufwait list prior to I/O. 12240 */ 12241 if (waitfor) { 12242 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12243 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12244 == DEPCOMPLETE) { 12245 jwait(&inoref->if_list, MNT_WAIT); 12246 goto again; 12247 } 12248 } 12249 } 12250 /* 12251 * Changes have been initiated. Anything depending on these 12252 * changes cannot occur until this inode has been written. 12253 */ 12254 inodedep->id_state &= ~COMPLETE; 12255 if ((inodedep->id_state & ONWORKLIST) == 0) 12256 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list); 12257 /* 12258 * Any new dependencies associated with the incore inode must 12259 * now be moved to the list associated with the buffer holding 12260 * the in-memory copy of the inode. Once merged process any 12261 * allocdirects that are completed by the merger. 12262 */ 12263 merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt); 12264 if (!TAILQ_EMPTY(&inodedep->id_inoupdt)) 12265 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt), 12266 NULL); 12267 merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt); 12268 if (!TAILQ_EMPTY(&inodedep->id_extupdt)) 12269 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt), 12270 NULL); 12271 /* 12272 * Now that the inode has been pushed into the buffer, the 12273 * operations dependent on the inode being written to disk 12274 * can be moved to the id_bufwait so that they will be 12275 * processed when the buffer I/O completes. 12276 */ 12277 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) { 12278 WORKLIST_REMOVE(wk); 12279 WORKLIST_INSERT(&inodedep->id_bufwait, wk); 12280 } 12281 /* 12282 * Newly allocated inodes cannot be written until the bitmap 12283 * that allocates them have been written (indicated by 12284 * DEPCOMPLETE being set in id_state). If we are doing a 12285 * forced sync (e.g., an fsync on a file), we force the bitmap 12286 * to be written so that the update can be done. 12287 */ 12288 if (waitfor == 0) { 12289 FREE_LOCK(ump); 12290 return; 12291 } 12292 retry: 12293 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) != 0) { 12294 FREE_LOCK(ump); 12295 return; 12296 } 12297 ibp = inodedep->id_bmsafemap->sm_buf; 12298 ibp = getdirtybuf(ibp, LOCK_PTR(ump), MNT_WAIT); 12299 if (ibp == NULL) { 12300 /* 12301 * If ibp came back as NULL, the dependency could have been 12302 * freed while we slept. Look it up again, and check to see 12303 * that it has completed. 12304 */ 12305 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) 12306 goto retry; 12307 FREE_LOCK(ump); 12308 return; 12309 } 12310 FREE_LOCK(ump); 12311 if ((error = bwrite(ibp)) != 0) 12312 softdep_error("softdep_update_inodeblock: bwrite", error); 12313 } 12314 12315 /* 12316 * Merge the a new inode dependency list (such as id_newinoupdt) into an 12317 * old inode dependency list (such as id_inoupdt). This routine must be 12318 * called with splbio interrupts blocked. 12319 */ 12320 static void 12321 merge_inode_lists(newlisthead, oldlisthead) 12322 struct allocdirectlst *newlisthead; 12323 struct allocdirectlst *oldlisthead; 12324 { 12325 struct allocdirect *listadp, *newadp; 12326 12327 newadp = TAILQ_FIRST(newlisthead); 12328 for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) { 12329 if (listadp->ad_offset < newadp->ad_offset) { 12330 listadp = TAILQ_NEXT(listadp, ad_next); 12331 continue; 12332 } 12333 TAILQ_REMOVE(newlisthead, newadp, ad_next); 12334 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next); 12335 if (listadp->ad_offset == newadp->ad_offset) { 12336 allocdirect_merge(oldlisthead, newadp, 12337 listadp); 12338 listadp = newadp; 12339 } 12340 newadp = TAILQ_FIRST(newlisthead); 12341 } 12342 while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) { 12343 TAILQ_REMOVE(newlisthead, newadp, ad_next); 12344 TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next); 12345 } 12346 } 12347 12348 /* 12349 * If we are doing an fsync, then we must ensure that any directory 12350 * entries for the inode have been written after the inode gets to disk. 12351 */ 12352 int 12353 softdep_fsync(vp) 12354 struct vnode *vp; /* the "in_core" copy of the inode */ 12355 { 12356 struct inodedep *inodedep; 12357 struct pagedep *pagedep; 12358 struct inoref *inoref; 12359 struct ufsmount *ump; 12360 struct worklist *wk; 12361 struct diradd *dap; 12362 struct mount *mp; 12363 struct vnode *pvp; 12364 struct inode *ip; 12365 struct buf *bp; 12366 struct fs *fs; 12367 struct thread *td = curthread; 12368 int error, flushparent, pagedep_new_block; 12369 ino_t parentino; 12370 ufs_lbn_t lbn; 12371 12372 ip = VTOI(vp); 12373 mp = vp->v_mount; 12374 ump = VFSTOUFS(mp); 12375 fs = ump->um_fs; 12376 if (MOUNTEDSOFTDEP(mp) == 0) 12377 return (0); 12378 ACQUIRE_LOCK(ump); 12379 restart: 12380 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) { 12381 FREE_LOCK(ump); 12382 return (0); 12383 } 12384 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12385 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12386 == DEPCOMPLETE) { 12387 jwait(&inoref->if_list, MNT_WAIT); 12388 goto restart; 12389 } 12390 } 12391 if (!LIST_EMPTY(&inodedep->id_inowait) || 12392 !TAILQ_EMPTY(&inodedep->id_extupdt) || 12393 !TAILQ_EMPTY(&inodedep->id_newextupdt) || 12394 !TAILQ_EMPTY(&inodedep->id_inoupdt) || 12395 !TAILQ_EMPTY(&inodedep->id_newinoupdt)) 12396 panic("softdep_fsync: pending ops %p", inodedep); 12397 for (error = 0, flushparent = 0; ; ) { 12398 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL) 12399 break; 12400 if (wk->wk_type != D_DIRADD) 12401 panic("softdep_fsync: Unexpected type %s", 12402 TYPENAME(wk->wk_type)); 12403 dap = WK_DIRADD(wk); 12404 /* 12405 * Flush our parent if this directory entry has a MKDIR_PARENT 12406 * dependency or is contained in a newly allocated block. 12407 */ 12408 if (dap->da_state & DIRCHG) 12409 pagedep = dap->da_previous->dm_pagedep; 12410 else 12411 pagedep = dap->da_pagedep; 12412 parentino = pagedep->pd_ino; 12413 lbn = pagedep->pd_lbn; 12414 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) 12415 panic("softdep_fsync: dirty"); 12416 if ((dap->da_state & MKDIR_PARENT) || 12417 (pagedep->pd_state & NEWBLOCK)) 12418 flushparent = 1; 12419 else 12420 flushparent = 0; 12421 /* 12422 * If we are being fsync'ed as part of vgone'ing this vnode, 12423 * then we will not be able to release and recover the 12424 * vnode below, so we just have to give up on writing its 12425 * directory entry out. It will eventually be written, just 12426 * not now, but then the user was not asking to have it 12427 * written, so we are not breaking any promises. 12428 */ 12429 if (vp->v_iflag & VI_DOOMED) 12430 break; 12431 /* 12432 * We prevent deadlock by always fetching inodes from the 12433 * root, moving down the directory tree. Thus, when fetching 12434 * our parent directory, we first try to get the lock. If 12435 * that fails, we must unlock ourselves before requesting 12436 * the lock on our parent. See the comment in ufs_lookup 12437 * for details on possible races. 12438 */ 12439 FREE_LOCK(ump); 12440 if (ffs_vgetf(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp, 12441 FFSV_FORCEINSMQ)) { 12442 /* 12443 * Unmount cannot proceed after unlock because 12444 * caller must have called vn_start_write(). 12445 */ 12446 VOP_UNLOCK(vp, 0); 12447 error = ffs_vgetf(mp, parentino, LK_EXCLUSIVE, 12448 &pvp, FFSV_FORCEINSMQ); 12449 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 12450 if (vp->v_iflag & VI_DOOMED) { 12451 if (error == 0) 12452 vput(pvp); 12453 error = ENOENT; 12454 } 12455 if (error != 0) 12456 return (error); 12457 } 12458 /* 12459 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps 12460 * that are contained in direct blocks will be resolved by 12461 * doing a ffs_update. Pagedeps contained in indirect blocks 12462 * may require a complete sync'ing of the directory. So, we 12463 * try the cheap and fast ffs_update first, and if that fails, 12464 * then we do the slower ffs_syncvnode of the directory. 12465 */ 12466 if (flushparent) { 12467 int locked; 12468 12469 if ((error = ffs_update(pvp, 1)) != 0) { 12470 vput(pvp); 12471 return (error); 12472 } 12473 ACQUIRE_LOCK(ump); 12474 locked = 1; 12475 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) { 12476 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) { 12477 if (wk->wk_type != D_DIRADD) 12478 panic("softdep_fsync: Unexpected type %s", 12479 TYPENAME(wk->wk_type)); 12480 dap = WK_DIRADD(wk); 12481 if (dap->da_state & DIRCHG) 12482 pagedep = dap->da_previous->dm_pagedep; 12483 else 12484 pagedep = dap->da_pagedep; 12485 pagedep_new_block = pagedep->pd_state & NEWBLOCK; 12486 FREE_LOCK(ump); 12487 locked = 0; 12488 if (pagedep_new_block && (error = 12489 ffs_syncvnode(pvp, MNT_WAIT, 0))) { 12490 vput(pvp); 12491 return (error); 12492 } 12493 } 12494 } 12495 if (locked) 12496 FREE_LOCK(ump); 12497 } 12498 /* 12499 * Flush directory page containing the inode's name. 12500 */ 12501 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred, 12502 &bp); 12503 if (error == 0) 12504 error = bwrite(bp); 12505 else 12506 brelse(bp); 12507 vput(pvp); 12508 if (error != 0) 12509 return (error); 12510 ACQUIRE_LOCK(ump); 12511 if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) 12512 break; 12513 } 12514 FREE_LOCK(ump); 12515 return (0); 12516 } 12517 12518 /* 12519 * Flush all the dirty bitmaps associated with the block device 12520 * before flushing the rest of the dirty blocks so as to reduce 12521 * the number of dependencies that will have to be rolled back. 12522 * 12523 * XXX Unused? 12524 */ 12525 void 12526 softdep_fsync_mountdev(vp) 12527 struct vnode *vp; 12528 { 12529 struct buf *bp, *nbp; 12530 struct worklist *wk; 12531 struct bufobj *bo; 12532 12533 if (!vn_isdisk(vp, NULL)) 12534 panic("softdep_fsync_mountdev: vnode not a disk"); 12535 bo = &vp->v_bufobj; 12536 restart: 12537 BO_LOCK(bo); 12538 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 12539 /* 12540 * If it is already scheduled, skip to the next buffer. 12541 */ 12542 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) 12543 continue; 12544 12545 if ((bp->b_flags & B_DELWRI) == 0) 12546 panic("softdep_fsync_mountdev: not dirty"); 12547 /* 12548 * We are only interested in bitmaps with outstanding 12549 * dependencies. 12550 */ 12551 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL || 12552 wk->wk_type != D_BMSAFEMAP || 12553 (bp->b_vflags & BV_BKGRDINPROG)) { 12554 BUF_UNLOCK(bp); 12555 continue; 12556 } 12557 BO_UNLOCK(bo); 12558 bremfree(bp); 12559 (void) bawrite(bp); 12560 goto restart; 12561 } 12562 drain_output(vp); 12563 BO_UNLOCK(bo); 12564 } 12565 12566 /* 12567 * Sync all cylinder groups that were dirty at the time this function is 12568 * called. Newly dirtied cgs will be inserted before the sentinel. This 12569 * is used to flush freedep activity that may be holding up writes to a 12570 * indirect block. 12571 */ 12572 static int 12573 sync_cgs(mp, waitfor) 12574 struct mount *mp; 12575 int waitfor; 12576 { 12577 struct bmsafemap *bmsafemap; 12578 struct bmsafemap *sentinel; 12579 struct ufsmount *ump; 12580 struct buf *bp; 12581 int error; 12582 12583 sentinel = malloc(sizeof(*sentinel), M_BMSAFEMAP, M_ZERO | M_WAITOK); 12584 sentinel->sm_cg = -1; 12585 ump = VFSTOUFS(mp); 12586 error = 0; 12587 ACQUIRE_LOCK(ump); 12588 LIST_INSERT_HEAD(&ump->softdep_dirtycg, sentinel, sm_next); 12589 for (bmsafemap = LIST_NEXT(sentinel, sm_next); bmsafemap != NULL; 12590 bmsafemap = LIST_NEXT(sentinel, sm_next)) { 12591 /* Skip sentinels and cgs with no work to release. */ 12592 if (bmsafemap->sm_cg == -1 || 12593 (LIST_EMPTY(&bmsafemap->sm_freehd) && 12594 LIST_EMPTY(&bmsafemap->sm_freewr))) { 12595 LIST_REMOVE(sentinel, sm_next); 12596 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next); 12597 continue; 12598 } 12599 /* 12600 * If we don't get the lock and we're waiting try again, if 12601 * not move on to the next buf and try to sync it. 12602 */ 12603 bp = getdirtybuf(bmsafemap->sm_buf, LOCK_PTR(ump), waitfor); 12604 if (bp == NULL && waitfor == MNT_WAIT) 12605 continue; 12606 LIST_REMOVE(sentinel, sm_next); 12607 LIST_INSERT_AFTER(bmsafemap, sentinel, sm_next); 12608 if (bp == NULL) 12609 continue; 12610 FREE_LOCK(ump); 12611 if (waitfor == MNT_NOWAIT) 12612 bawrite(bp); 12613 else 12614 error = bwrite(bp); 12615 ACQUIRE_LOCK(ump); 12616 if (error) 12617 break; 12618 } 12619 LIST_REMOVE(sentinel, sm_next); 12620 FREE_LOCK(ump); 12621 free(sentinel, M_BMSAFEMAP); 12622 return (error); 12623 } 12624 12625 /* 12626 * This routine is called when we are trying to synchronously flush a 12627 * file. This routine must eliminate any filesystem metadata dependencies 12628 * so that the syncing routine can succeed. 12629 */ 12630 int 12631 softdep_sync_metadata(struct vnode *vp) 12632 { 12633 struct inode *ip; 12634 int error; 12635 12636 ip = VTOI(vp); 12637 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 12638 ("softdep_sync_metadata called on non-softdep filesystem")); 12639 /* 12640 * Ensure that any direct block dependencies have been cleared, 12641 * truncations are started, and inode references are journaled. 12642 */ 12643 ACQUIRE_LOCK(VFSTOUFS(vp->v_mount)); 12644 /* 12645 * Write all journal records to prevent rollbacks on devvp. 12646 */ 12647 if (vp->v_type == VCHR) 12648 softdep_flushjournal(vp->v_mount); 12649 error = flush_inodedep_deps(vp, vp->v_mount, ip->i_number); 12650 /* 12651 * Ensure that all truncates are written so we won't find deps on 12652 * indirect blocks. 12653 */ 12654 process_truncates(vp); 12655 FREE_LOCK(VFSTOUFS(vp->v_mount)); 12656 12657 return (error); 12658 } 12659 12660 /* 12661 * This routine is called when we are attempting to sync a buf with 12662 * dependencies. If waitfor is MNT_NOWAIT it attempts to schedule any 12663 * other IO it can but returns EBUSY if the buffer is not yet able to 12664 * be written. Dependencies which will not cause rollbacks will always 12665 * return 0. 12666 */ 12667 int 12668 softdep_sync_buf(struct vnode *vp, struct buf *bp, int waitfor) 12669 { 12670 struct indirdep *indirdep; 12671 struct pagedep *pagedep; 12672 struct allocindir *aip; 12673 struct newblk *newblk; 12674 struct ufsmount *ump; 12675 struct buf *nbp; 12676 struct worklist *wk; 12677 int i, error; 12678 12679 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 12680 ("softdep_sync_buf called on non-softdep filesystem")); 12681 /* 12682 * For VCHR we just don't want to force flush any dependencies that 12683 * will cause rollbacks. 12684 */ 12685 if (vp->v_type == VCHR) { 12686 if (waitfor == MNT_NOWAIT && softdep_count_dependencies(bp, 0)) 12687 return (EBUSY); 12688 return (0); 12689 } 12690 ump = VFSTOUFS(vp->v_mount); 12691 ACQUIRE_LOCK(ump); 12692 /* 12693 * As we hold the buffer locked, none of its dependencies 12694 * will disappear. 12695 */ 12696 error = 0; 12697 top: 12698 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 12699 switch (wk->wk_type) { 12700 12701 case D_ALLOCDIRECT: 12702 case D_ALLOCINDIR: 12703 newblk = WK_NEWBLK(wk); 12704 if (newblk->nb_jnewblk != NULL) { 12705 if (waitfor == MNT_NOWAIT) { 12706 error = EBUSY; 12707 goto out_unlock; 12708 } 12709 jwait(&newblk->nb_jnewblk->jn_list, waitfor); 12710 goto top; 12711 } 12712 if (newblk->nb_state & DEPCOMPLETE || 12713 waitfor == MNT_NOWAIT) 12714 continue; 12715 nbp = newblk->nb_bmsafemap->sm_buf; 12716 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor); 12717 if (nbp == NULL) 12718 goto top; 12719 FREE_LOCK(ump); 12720 if ((error = bwrite(nbp)) != 0) 12721 goto out; 12722 ACQUIRE_LOCK(ump); 12723 continue; 12724 12725 case D_INDIRDEP: 12726 indirdep = WK_INDIRDEP(wk); 12727 if (waitfor == MNT_NOWAIT) { 12728 if (!TAILQ_EMPTY(&indirdep->ir_trunc) || 12729 !LIST_EMPTY(&indirdep->ir_deplisthd)) { 12730 error = EBUSY; 12731 goto out_unlock; 12732 } 12733 } 12734 if (!TAILQ_EMPTY(&indirdep->ir_trunc)) 12735 panic("softdep_sync_buf: truncation pending."); 12736 restart: 12737 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 12738 newblk = (struct newblk *)aip; 12739 if (newblk->nb_jnewblk != NULL) { 12740 jwait(&newblk->nb_jnewblk->jn_list, 12741 waitfor); 12742 goto restart; 12743 } 12744 if (newblk->nb_state & DEPCOMPLETE) 12745 continue; 12746 nbp = newblk->nb_bmsafemap->sm_buf; 12747 nbp = getdirtybuf(nbp, LOCK_PTR(ump), waitfor); 12748 if (nbp == NULL) 12749 goto restart; 12750 FREE_LOCK(ump); 12751 if ((error = bwrite(nbp)) != 0) 12752 goto out; 12753 ACQUIRE_LOCK(ump); 12754 goto restart; 12755 } 12756 continue; 12757 12758 case D_PAGEDEP: 12759 /* 12760 * Only flush directory entries in synchronous passes. 12761 */ 12762 if (waitfor != MNT_WAIT) { 12763 error = EBUSY; 12764 goto out_unlock; 12765 } 12766 /* 12767 * While syncing snapshots, we must allow recursive 12768 * lookups. 12769 */ 12770 BUF_AREC(bp); 12771 /* 12772 * We are trying to sync a directory that may 12773 * have dependencies on both its own metadata 12774 * and/or dependencies on the inodes of any 12775 * recently allocated files. We walk its diradd 12776 * lists pushing out the associated inode. 12777 */ 12778 pagedep = WK_PAGEDEP(wk); 12779 for (i = 0; i < DAHASHSZ; i++) { 12780 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0) 12781 continue; 12782 if ((error = flush_pagedep_deps(vp, wk->wk_mp, 12783 &pagedep->pd_diraddhd[i]))) { 12784 BUF_NOREC(bp); 12785 goto out_unlock; 12786 } 12787 } 12788 BUF_NOREC(bp); 12789 continue; 12790 12791 case D_FREEWORK: 12792 case D_FREEDEP: 12793 case D_JSEGDEP: 12794 case D_JNEWBLK: 12795 continue; 12796 12797 default: 12798 panic("softdep_sync_buf: Unknown type %s", 12799 TYPENAME(wk->wk_type)); 12800 /* NOTREACHED */ 12801 } 12802 } 12803 out_unlock: 12804 FREE_LOCK(ump); 12805 out: 12806 return (error); 12807 } 12808 12809 /* 12810 * Flush the dependencies associated with an inodedep. 12811 * Called with splbio blocked. 12812 */ 12813 static int 12814 flush_inodedep_deps(vp, mp, ino) 12815 struct vnode *vp; 12816 struct mount *mp; 12817 ino_t ino; 12818 { 12819 struct inodedep *inodedep; 12820 struct inoref *inoref; 12821 struct ufsmount *ump; 12822 int error, waitfor; 12823 12824 /* 12825 * This work is done in two passes. The first pass grabs most 12826 * of the buffers and begins asynchronously writing them. The 12827 * only way to wait for these asynchronous writes is to sleep 12828 * on the filesystem vnode which may stay busy for a long time 12829 * if the filesystem is active. So, instead, we make a second 12830 * pass over the dependencies blocking on each write. In the 12831 * usual case we will be blocking against a write that we 12832 * initiated, so when it is done the dependency will have been 12833 * resolved. Thus the second pass is expected to end quickly. 12834 * We give a brief window at the top of the loop to allow 12835 * any pending I/O to complete. 12836 */ 12837 ump = VFSTOUFS(mp); 12838 LOCK_OWNED(ump); 12839 for (error = 0, waitfor = MNT_NOWAIT; ; ) { 12840 if (error) 12841 return (error); 12842 FREE_LOCK(ump); 12843 ACQUIRE_LOCK(ump); 12844 restart: 12845 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 12846 return (0); 12847 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 12848 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 12849 == DEPCOMPLETE) { 12850 jwait(&inoref->if_list, MNT_WAIT); 12851 goto restart; 12852 } 12853 } 12854 if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) || 12855 flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) || 12856 flush_deplist(&inodedep->id_extupdt, waitfor, &error) || 12857 flush_deplist(&inodedep->id_newextupdt, waitfor, &error)) 12858 continue; 12859 /* 12860 * If pass2, we are done, otherwise do pass 2. 12861 */ 12862 if (waitfor == MNT_WAIT) 12863 break; 12864 waitfor = MNT_WAIT; 12865 } 12866 /* 12867 * Try freeing inodedep in case all dependencies have been removed. 12868 */ 12869 if (inodedep_lookup(mp, ino, 0, &inodedep) != 0) 12870 (void) free_inodedep(inodedep); 12871 return (0); 12872 } 12873 12874 /* 12875 * Flush an inode dependency list. 12876 * Called with splbio blocked. 12877 */ 12878 static int 12879 flush_deplist(listhead, waitfor, errorp) 12880 struct allocdirectlst *listhead; 12881 int waitfor; 12882 int *errorp; 12883 { 12884 struct allocdirect *adp; 12885 struct newblk *newblk; 12886 struct ufsmount *ump; 12887 struct buf *bp; 12888 12889 if ((adp = TAILQ_FIRST(listhead)) == NULL) 12890 return (0); 12891 ump = VFSTOUFS(adp->ad_list.wk_mp); 12892 LOCK_OWNED(ump); 12893 TAILQ_FOREACH(adp, listhead, ad_next) { 12894 newblk = (struct newblk *)adp; 12895 if (newblk->nb_jnewblk != NULL) { 12896 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 12897 return (1); 12898 } 12899 if (newblk->nb_state & DEPCOMPLETE) 12900 continue; 12901 bp = newblk->nb_bmsafemap->sm_buf; 12902 bp = getdirtybuf(bp, LOCK_PTR(ump), waitfor); 12903 if (bp == NULL) { 12904 if (waitfor == MNT_NOWAIT) 12905 continue; 12906 return (1); 12907 } 12908 FREE_LOCK(ump); 12909 if (waitfor == MNT_NOWAIT) 12910 bawrite(bp); 12911 else 12912 *errorp = bwrite(bp); 12913 ACQUIRE_LOCK(ump); 12914 return (1); 12915 } 12916 return (0); 12917 } 12918 12919 /* 12920 * Flush dependencies associated with an allocdirect block. 12921 */ 12922 static int 12923 flush_newblk_dep(vp, mp, lbn) 12924 struct vnode *vp; 12925 struct mount *mp; 12926 ufs_lbn_t lbn; 12927 { 12928 struct newblk *newblk; 12929 struct ufsmount *ump; 12930 struct bufobj *bo; 12931 struct inode *ip; 12932 struct buf *bp; 12933 ufs2_daddr_t blkno; 12934 int error; 12935 12936 error = 0; 12937 bo = &vp->v_bufobj; 12938 ip = VTOI(vp); 12939 blkno = DIP(ip, i_db[lbn]); 12940 if (blkno == 0) 12941 panic("flush_newblk_dep: Missing block"); 12942 ump = VFSTOUFS(mp); 12943 ACQUIRE_LOCK(ump); 12944 /* 12945 * Loop until all dependencies related to this block are satisfied. 12946 * We must be careful to restart after each sleep in case a write 12947 * completes some part of this process for us. 12948 */ 12949 for (;;) { 12950 if (newblk_lookup(mp, blkno, 0, &newblk) == 0) { 12951 FREE_LOCK(ump); 12952 break; 12953 } 12954 if (newblk->nb_list.wk_type != D_ALLOCDIRECT) 12955 panic("flush_newblk_deps: Bad newblk %p", newblk); 12956 /* 12957 * Flush the journal. 12958 */ 12959 if (newblk->nb_jnewblk != NULL) { 12960 jwait(&newblk->nb_jnewblk->jn_list, MNT_WAIT); 12961 continue; 12962 } 12963 /* 12964 * Write the bitmap dependency. 12965 */ 12966 if ((newblk->nb_state & DEPCOMPLETE) == 0) { 12967 bp = newblk->nb_bmsafemap->sm_buf; 12968 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT); 12969 if (bp == NULL) 12970 continue; 12971 FREE_LOCK(ump); 12972 error = bwrite(bp); 12973 if (error) 12974 break; 12975 ACQUIRE_LOCK(ump); 12976 continue; 12977 } 12978 /* 12979 * Write the buffer. 12980 */ 12981 FREE_LOCK(ump); 12982 BO_LOCK(bo); 12983 bp = gbincore(bo, lbn); 12984 if (bp != NULL) { 12985 error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | 12986 LK_INTERLOCK, BO_LOCKPTR(bo)); 12987 if (error == ENOLCK) { 12988 ACQUIRE_LOCK(ump); 12989 error = 0; 12990 continue; /* Slept, retry */ 12991 } 12992 if (error != 0) 12993 break; /* Failed */ 12994 if (bp->b_flags & B_DELWRI) { 12995 bremfree(bp); 12996 error = bwrite(bp); 12997 if (error) 12998 break; 12999 } else 13000 BUF_UNLOCK(bp); 13001 } else 13002 BO_UNLOCK(bo); 13003 /* 13004 * We have to wait for the direct pointers to 13005 * point at the newdirblk before the dependency 13006 * will go away. 13007 */ 13008 error = ffs_update(vp, 1); 13009 if (error) 13010 break; 13011 ACQUIRE_LOCK(ump); 13012 } 13013 return (error); 13014 } 13015 13016 /* 13017 * Eliminate a pagedep dependency by flushing out all its diradd dependencies. 13018 * Called with splbio blocked. 13019 */ 13020 static int 13021 flush_pagedep_deps(pvp, mp, diraddhdp) 13022 struct vnode *pvp; 13023 struct mount *mp; 13024 struct diraddhd *diraddhdp; 13025 { 13026 struct inodedep *inodedep; 13027 struct inoref *inoref; 13028 struct ufsmount *ump; 13029 struct diradd *dap; 13030 struct vnode *vp; 13031 int error = 0; 13032 struct buf *bp; 13033 ino_t inum; 13034 struct diraddhd unfinished; 13035 13036 LIST_INIT(&unfinished); 13037 ump = VFSTOUFS(mp); 13038 LOCK_OWNED(ump); 13039 restart: 13040 while ((dap = LIST_FIRST(diraddhdp)) != NULL) { 13041 /* 13042 * Flush ourselves if this directory entry 13043 * has a MKDIR_PARENT dependency. 13044 */ 13045 if (dap->da_state & MKDIR_PARENT) { 13046 FREE_LOCK(ump); 13047 if ((error = ffs_update(pvp, 1)) != 0) 13048 break; 13049 ACQUIRE_LOCK(ump); 13050 /* 13051 * If that cleared dependencies, go on to next. 13052 */ 13053 if (dap != LIST_FIRST(diraddhdp)) 13054 continue; 13055 /* 13056 * All MKDIR_PARENT dependencies and all the 13057 * NEWBLOCK pagedeps that are contained in direct 13058 * blocks were resolved by doing above ffs_update. 13059 * Pagedeps contained in indirect blocks may 13060 * require a complete sync'ing of the directory. 13061 * We are in the midst of doing a complete sync, 13062 * so if they are not resolved in this pass we 13063 * defer them for now as they will be sync'ed by 13064 * our caller shortly. 13065 */ 13066 LIST_REMOVE(dap, da_pdlist); 13067 LIST_INSERT_HEAD(&unfinished, dap, da_pdlist); 13068 continue; 13069 } 13070 /* 13071 * A newly allocated directory must have its "." and 13072 * ".." entries written out before its name can be 13073 * committed in its parent. 13074 */ 13075 inum = dap->da_newinum; 13076 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 13077 panic("flush_pagedep_deps: lost inode1"); 13078 /* 13079 * Wait for any pending journal adds to complete so we don't 13080 * cause rollbacks while syncing. 13081 */ 13082 TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) { 13083 if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY)) 13084 == DEPCOMPLETE) { 13085 jwait(&inoref->if_list, MNT_WAIT); 13086 goto restart; 13087 } 13088 } 13089 if (dap->da_state & MKDIR_BODY) { 13090 FREE_LOCK(ump); 13091 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp, 13092 FFSV_FORCEINSMQ))) 13093 break; 13094 error = flush_newblk_dep(vp, mp, 0); 13095 /* 13096 * If we still have the dependency we might need to 13097 * update the vnode to sync the new link count to 13098 * disk. 13099 */ 13100 if (error == 0 && dap == LIST_FIRST(diraddhdp)) 13101 error = ffs_update(vp, 1); 13102 vput(vp); 13103 if (error != 0) 13104 break; 13105 ACQUIRE_LOCK(ump); 13106 /* 13107 * If that cleared dependencies, go on to next. 13108 */ 13109 if (dap != LIST_FIRST(diraddhdp)) 13110 continue; 13111 if (dap->da_state & MKDIR_BODY) { 13112 inodedep_lookup(UFSTOVFS(ump), inum, 0, 13113 &inodedep); 13114 panic("flush_pagedep_deps: MKDIR_BODY " 13115 "inodedep %p dap %p vp %p", 13116 inodedep, dap, vp); 13117 } 13118 } 13119 /* 13120 * Flush the inode on which the directory entry depends. 13121 * Having accounted for MKDIR_PARENT and MKDIR_BODY above, 13122 * the only remaining dependency is that the updated inode 13123 * count must get pushed to disk. The inode has already 13124 * been pushed into its inode buffer (via VOP_UPDATE) at 13125 * the time of the reference count change. So we need only 13126 * locate that buffer, ensure that there will be no rollback 13127 * caused by a bitmap dependency, then write the inode buffer. 13128 */ 13129 retry: 13130 if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0) 13131 panic("flush_pagedep_deps: lost inode"); 13132 /* 13133 * If the inode still has bitmap dependencies, 13134 * push them to disk. 13135 */ 13136 if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) == 0) { 13137 bp = inodedep->id_bmsafemap->sm_buf; 13138 bp = getdirtybuf(bp, LOCK_PTR(ump), MNT_WAIT); 13139 if (bp == NULL) 13140 goto retry; 13141 FREE_LOCK(ump); 13142 if ((error = bwrite(bp)) != 0) 13143 break; 13144 ACQUIRE_LOCK(ump); 13145 if (dap != LIST_FIRST(diraddhdp)) 13146 continue; 13147 } 13148 /* 13149 * If the inode is still sitting in a buffer waiting 13150 * to be written or waiting for the link count to be 13151 * adjusted update it here to flush it to disk. 13152 */ 13153 if (dap == LIST_FIRST(diraddhdp)) { 13154 FREE_LOCK(ump); 13155 if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp, 13156 FFSV_FORCEINSMQ))) 13157 break; 13158 error = ffs_update(vp, 1); 13159 vput(vp); 13160 if (error) 13161 break; 13162 ACQUIRE_LOCK(ump); 13163 } 13164 /* 13165 * If we have failed to get rid of all the dependencies 13166 * then something is seriously wrong. 13167 */ 13168 if (dap == LIST_FIRST(diraddhdp)) { 13169 inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep); 13170 panic("flush_pagedep_deps: failed to flush " 13171 "inodedep %p ino %ju dap %p", 13172 inodedep, (uintmax_t)inum, dap); 13173 } 13174 } 13175 if (error) 13176 ACQUIRE_LOCK(ump); 13177 while ((dap = LIST_FIRST(&unfinished)) != NULL) { 13178 LIST_REMOVE(dap, da_pdlist); 13179 LIST_INSERT_HEAD(diraddhdp, dap, da_pdlist); 13180 } 13181 return (error); 13182 } 13183 13184 /* 13185 * A large burst of file addition or deletion activity can drive the 13186 * memory load excessively high. First attempt to slow things down 13187 * using the techniques below. If that fails, this routine requests 13188 * the offending operations to fall back to running synchronously 13189 * until the memory load returns to a reasonable level. 13190 */ 13191 int 13192 softdep_slowdown(vp) 13193 struct vnode *vp; 13194 { 13195 struct ufsmount *ump; 13196 int jlow; 13197 int max_softdeps_hard; 13198 13199 KASSERT(MOUNTEDSOFTDEP(vp->v_mount) != 0, 13200 ("softdep_slowdown called on non-softdep filesystem")); 13201 ump = VFSTOUFS(vp->v_mount); 13202 ACQUIRE_LOCK(ump); 13203 jlow = 0; 13204 /* 13205 * Check for journal space if needed. 13206 */ 13207 if (DOINGSUJ(vp)) { 13208 if (journal_space(ump, 0) == 0) 13209 jlow = 1; 13210 } 13211 /* 13212 * If the system is under its limits and our filesystem is 13213 * not responsible for more than our share of the usage and 13214 * we are not low on journal space, then no need to slow down. 13215 */ 13216 max_softdeps_hard = max_softdeps * 11 / 10; 13217 if (dep_current[D_DIRREM] < max_softdeps_hard / 2 && 13218 dep_current[D_INODEDEP] < max_softdeps_hard && 13219 dep_current[D_INDIRDEP] < max_softdeps_hard / 1000 && 13220 dep_current[D_FREEBLKS] < max_softdeps_hard && jlow == 0 && 13221 ump->softdep_curdeps[D_DIRREM] < 13222 (max_softdeps_hard / 2) / stat_flush_threads && 13223 ump->softdep_curdeps[D_INODEDEP] < 13224 max_softdeps_hard / stat_flush_threads && 13225 ump->softdep_curdeps[D_INDIRDEP] < 13226 (max_softdeps_hard / 1000) / stat_flush_threads && 13227 ump->softdep_curdeps[D_FREEBLKS] < 13228 max_softdeps_hard / stat_flush_threads) { 13229 FREE_LOCK(ump); 13230 return (0); 13231 } 13232 /* 13233 * If the journal is low or our filesystem is over its limit 13234 * then speedup the cleanup. 13235 */ 13236 if (ump->softdep_curdeps[D_INDIRDEP] < 13237 (max_softdeps_hard / 1000) / stat_flush_threads || jlow) 13238 softdep_speedup(ump); 13239 stat_sync_limit_hit += 1; 13240 FREE_LOCK(ump); 13241 /* 13242 * We only slow down the rate at which new dependencies are 13243 * generated if we are not using journaling. With journaling, 13244 * the cleanup should always be sufficient to keep things 13245 * under control. 13246 */ 13247 if (DOINGSUJ(vp)) 13248 return (0); 13249 return (1); 13250 } 13251 13252 /* 13253 * Called by the allocation routines when they are about to fail 13254 * in the hope that we can free up the requested resource (inodes 13255 * or disk space). 13256 * 13257 * First check to see if the work list has anything on it. If it has, 13258 * clean up entries until we successfully free the requested resource. 13259 * Because this process holds inodes locked, we cannot handle any remove 13260 * requests that might block on a locked inode as that could lead to 13261 * deadlock. If the worklist yields none of the requested resource, 13262 * start syncing out vnodes to free up the needed space. 13263 */ 13264 int 13265 softdep_request_cleanup(fs, vp, cred, resource) 13266 struct fs *fs; 13267 struct vnode *vp; 13268 struct ucred *cred; 13269 int resource; 13270 { 13271 struct ufsmount *ump; 13272 struct mount *mp; 13273 long starttime; 13274 ufs2_daddr_t needed; 13275 int error, failed_vnode; 13276 13277 /* 13278 * If we are being called because of a process doing a 13279 * copy-on-write, then it is not safe to process any 13280 * worklist items as we will recurse into the copyonwrite 13281 * routine. This will result in an incoherent snapshot. 13282 * If the vnode that we hold is a snapshot, we must avoid 13283 * handling other resources that could cause deadlock. 13284 */ 13285 if ((curthread->td_pflags & TDP_COWINPROGRESS) || IS_SNAPSHOT(VTOI(vp))) 13286 return (0); 13287 13288 if (resource == FLUSH_BLOCKS_WAIT) 13289 stat_cleanup_blkrequests += 1; 13290 else 13291 stat_cleanup_inorequests += 1; 13292 13293 mp = vp->v_mount; 13294 ump = VFSTOUFS(mp); 13295 mtx_assert(UFS_MTX(ump), MA_OWNED); 13296 UFS_UNLOCK(ump); 13297 error = ffs_update(vp, 1); 13298 if (error != 0 || MOUNTEDSOFTDEP(mp) == 0) { 13299 UFS_LOCK(ump); 13300 return (0); 13301 } 13302 /* 13303 * If we are in need of resources, start by cleaning up 13304 * any block removals associated with our inode. 13305 */ 13306 ACQUIRE_LOCK(ump); 13307 process_removes(vp); 13308 process_truncates(vp); 13309 FREE_LOCK(ump); 13310 /* 13311 * Now clean up at least as many resources as we will need. 13312 * 13313 * When requested to clean up inodes, the number that are needed 13314 * is set by the number of simultaneous writers (mnt_writeopcount) 13315 * plus a bit of slop (2) in case some more writers show up while 13316 * we are cleaning. 13317 * 13318 * When requested to free up space, the amount of space that 13319 * we need is enough blocks to allocate a full-sized segment 13320 * (fs_contigsumsize). The number of such segments that will 13321 * be needed is set by the number of simultaneous writers 13322 * (mnt_writeopcount) plus a bit of slop (2) in case some more 13323 * writers show up while we are cleaning. 13324 * 13325 * Additionally, if we are unpriviledged and allocating space, 13326 * we need to ensure that we clean up enough blocks to get the 13327 * needed number of blocks over the threshold of the minimum 13328 * number of blocks required to be kept free by the filesystem 13329 * (fs_minfree). 13330 */ 13331 if (resource == FLUSH_INODES_WAIT) { 13332 needed = vp->v_mount->mnt_writeopcount + 2; 13333 } else if (resource == FLUSH_BLOCKS_WAIT) { 13334 needed = (vp->v_mount->mnt_writeopcount + 2) * 13335 fs->fs_contigsumsize; 13336 if (priv_check_cred(cred, PRIV_VFS_BLOCKRESERVE, 0)) 13337 needed += fragstoblks(fs, 13338 roundup((fs->fs_dsize * fs->fs_minfree / 100) - 13339 fs->fs_cstotal.cs_nffree, fs->fs_frag)); 13340 } else { 13341 UFS_LOCK(ump); 13342 printf("softdep_request_cleanup: Unknown resource type %d\n", 13343 resource); 13344 return (0); 13345 } 13346 starttime = time_second; 13347 retry: 13348 if ((resource == FLUSH_BLOCKS_WAIT && ump->softdep_on_worklist > 0 && 13349 fs->fs_cstotal.cs_nbfree <= needed) || 13350 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 && 13351 fs->fs_cstotal.cs_nifree <= needed)) { 13352 ACQUIRE_LOCK(ump); 13353 if (ump->softdep_on_worklist > 0 && 13354 process_worklist_item(UFSTOVFS(ump), 13355 ump->softdep_on_worklist, LK_NOWAIT) != 0) 13356 stat_worklist_push += 1; 13357 FREE_LOCK(ump); 13358 } 13359 /* 13360 * If we still need resources and there are no more worklist 13361 * entries to process to obtain them, we have to start flushing 13362 * the dirty vnodes to force the release of additional requests 13363 * to the worklist that we can then process to reap addition 13364 * resources. We walk the vnodes associated with the mount point 13365 * until we get the needed worklist requests that we can reap. 13366 * 13367 * If there are several threads all needing to clean the same 13368 * mount point, only one is allowed to walk the mount list. 13369 * When several threads all try to walk the same mount list, 13370 * they end up competing with each other and often end up in 13371 * livelock. This approach ensures that forward progress is 13372 * made at the cost of occational ENOSPC errors being returned 13373 * that might otherwise have been avoided. 13374 */ 13375 error = 1; 13376 if ((resource == FLUSH_BLOCKS_WAIT && 13377 fs->fs_cstotal.cs_nbfree <= needed) || 13378 (resource == FLUSH_INODES_WAIT && fs->fs_pendinginodes > 0 && 13379 fs->fs_cstotal.cs_nifree <= needed)) { 13380 ACQUIRE_LOCK(ump); 13381 if ((ump->um_softdep->sd_flags & FLUSH_RC_ACTIVE) == 0) { 13382 ump->um_softdep->sd_flags |= FLUSH_RC_ACTIVE; 13383 FREE_LOCK(ump); 13384 failed_vnode = softdep_request_cleanup_flush(mp, ump); 13385 ACQUIRE_LOCK(ump); 13386 ump->um_softdep->sd_flags &= ~FLUSH_RC_ACTIVE; 13387 FREE_LOCK(ump); 13388 if (ump->softdep_on_worklist > 0) { 13389 stat_cleanup_retries += 1; 13390 if (!failed_vnode) 13391 goto retry; 13392 } 13393 } else { 13394 FREE_LOCK(ump); 13395 error = 0; 13396 } 13397 stat_cleanup_failures += 1; 13398 } 13399 if (time_second - starttime > stat_cleanup_high_delay) 13400 stat_cleanup_high_delay = time_second - starttime; 13401 UFS_LOCK(ump); 13402 return (error); 13403 } 13404 13405 /* 13406 * Scan the vnodes for the specified mount point flushing out any 13407 * vnodes that can be locked without waiting. Finally, try to flush 13408 * the device associated with the mount point if it can be locked 13409 * without waiting. 13410 * 13411 * We return 0 if we were able to lock every vnode in our scan. 13412 * If we had to skip one or more vnodes, we return 1. 13413 */ 13414 static int 13415 softdep_request_cleanup_flush(mp, ump) 13416 struct mount *mp; 13417 struct ufsmount *ump; 13418 { 13419 struct thread *td; 13420 struct vnode *lvp, *mvp; 13421 int failed_vnode; 13422 13423 failed_vnode = 0; 13424 td = curthread; 13425 MNT_VNODE_FOREACH_ALL(lvp, mp, mvp) { 13426 if (TAILQ_FIRST(&lvp->v_bufobj.bo_dirty.bv_hd) == 0) { 13427 VI_UNLOCK(lvp); 13428 continue; 13429 } 13430 if (vget(lvp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT, 13431 td) != 0) { 13432 failed_vnode = 1; 13433 continue; 13434 } 13435 if (lvp->v_vflag & VV_NOSYNC) { /* unlinked */ 13436 vput(lvp); 13437 continue; 13438 } 13439 (void) ffs_syncvnode(lvp, MNT_NOWAIT, 0); 13440 vput(lvp); 13441 } 13442 lvp = ump->um_devvp; 13443 if (vn_lock(lvp, LK_EXCLUSIVE | LK_NOWAIT) == 0) { 13444 VOP_FSYNC(lvp, MNT_NOWAIT, td); 13445 VOP_UNLOCK(lvp, 0); 13446 } 13447 return (failed_vnode); 13448 } 13449 13450 static bool 13451 softdep_excess_items(struct ufsmount *ump, int item) 13452 { 13453 13454 KASSERT(item >= 0 && item < D_LAST, ("item %d", item)); 13455 return (dep_current[item] > max_softdeps && 13456 ump->softdep_curdeps[item] > max_softdeps / 13457 stat_flush_threads); 13458 } 13459 13460 static void 13461 schedule_cleanup(struct mount *mp) 13462 { 13463 struct ufsmount *ump; 13464 struct thread *td; 13465 13466 ump = VFSTOUFS(mp); 13467 LOCK_OWNED(ump); 13468 FREE_LOCK(ump); 13469 td = curthread; 13470 if ((td->td_pflags & TDP_KTHREAD) != 0 && 13471 (td->td_proc->p_flag2 & P2_AST_SU) == 0) { 13472 /* 13473 * No ast is delivered to kernel threads, so nobody 13474 * would deref the mp. Some kernel threads 13475 * explicitely check for AST, e.g. NFS daemon does 13476 * this in the serving loop. 13477 */ 13478 return; 13479 } 13480 if (td->td_su != NULL) 13481 vfs_rel(td->td_su); 13482 vfs_ref(mp); 13483 td->td_su = mp; 13484 thread_lock(td); 13485 td->td_flags |= TDF_ASTPENDING; 13486 thread_unlock(td); 13487 } 13488 13489 static void 13490 softdep_ast_cleanup_proc(struct thread *td) 13491 { 13492 struct mount *mp; 13493 struct ufsmount *ump; 13494 int error; 13495 bool req; 13496 13497 while ((mp = td->td_su) != NULL) { 13498 td->td_su = NULL; 13499 error = vfs_busy(mp, MBF_NOWAIT); 13500 vfs_rel(mp); 13501 if (error != 0) 13502 return; 13503 if (ffs_own_mount(mp) && MOUNTEDSOFTDEP(mp)) { 13504 ump = VFSTOUFS(mp); 13505 for (;;) { 13506 req = false; 13507 ACQUIRE_LOCK(ump); 13508 if (softdep_excess_items(ump, D_INODEDEP)) { 13509 req = true; 13510 request_cleanup(mp, FLUSH_INODES); 13511 } 13512 if (softdep_excess_items(ump, D_DIRREM)) { 13513 req = true; 13514 request_cleanup(mp, FLUSH_BLOCKS); 13515 } 13516 FREE_LOCK(ump); 13517 if (softdep_excess_items(ump, D_NEWBLK) || 13518 softdep_excess_items(ump, D_ALLOCDIRECT) || 13519 softdep_excess_items(ump, D_ALLOCINDIR)) { 13520 error = vn_start_write(NULL, &mp, 13521 V_WAIT); 13522 if (error == 0) { 13523 req = true; 13524 VFS_SYNC(mp, MNT_WAIT); 13525 vn_finished_write(mp); 13526 } 13527 } 13528 if ((td->td_pflags & TDP_KTHREAD) != 0 || !req) 13529 break; 13530 } 13531 } 13532 vfs_unbusy(mp); 13533 } 13534 if ((mp = td->td_su) != NULL) { 13535 td->td_su = NULL; 13536 vfs_rel(mp); 13537 } 13538 } 13539 13540 /* 13541 * If memory utilization has gotten too high, deliberately slow things 13542 * down and speed up the I/O processing. 13543 */ 13544 static int 13545 request_cleanup(mp, resource) 13546 struct mount *mp; 13547 int resource; 13548 { 13549 struct thread *td = curthread; 13550 struct ufsmount *ump; 13551 13552 ump = VFSTOUFS(mp); 13553 LOCK_OWNED(ump); 13554 /* 13555 * We never hold up the filesystem syncer or buf daemon. 13556 */ 13557 if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF)) 13558 return (0); 13559 /* 13560 * First check to see if the work list has gotten backlogged. 13561 * If it has, co-opt this process to help clean up two entries. 13562 * Because this process may hold inodes locked, we cannot 13563 * handle any remove requests that might block on a locked 13564 * inode as that could lead to deadlock. We set TDP_SOFTDEP 13565 * to avoid recursively processing the worklist. 13566 */ 13567 if (ump->softdep_on_worklist > max_softdeps / 10) { 13568 td->td_pflags |= TDP_SOFTDEP; 13569 process_worklist_item(mp, 2, LK_NOWAIT); 13570 td->td_pflags &= ~TDP_SOFTDEP; 13571 stat_worklist_push += 2; 13572 return(1); 13573 } 13574 /* 13575 * Next, we attempt to speed up the syncer process. If that 13576 * is successful, then we allow the process to continue. 13577 */ 13578 if (softdep_speedup(ump) && 13579 resource != FLUSH_BLOCKS_WAIT && 13580 resource != FLUSH_INODES_WAIT) 13581 return(0); 13582 /* 13583 * If we are resource constrained on inode dependencies, try 13584 * flushing some dirty inodes. Otherwise, we are constrained 13585 * by file deletions, so try accelerating flushes of directories 13586 * with removal dependencies. We would like to do the cleanup 13587 * here, but we probably hold an inode locked at this point and 13588 * that might deadlock against one that we try to clean. So, 13589 * the best that we can do is request the syncer daemon to do 13590 * the cleanup for us. 13591 */ 13592 switch (resource) { 13593 13594 case FLUSH_INODES: 13595 case FLUSH_INODES_WAIT: 13596 ACQUIRE_GBLLOCK(&lk); 13597 stat_ino_limit_push += 1; 13598 req_clear_inodedeps += 1; 13599 FREE_GBLLOCK(&lk); 13600 stat_countp = &stat_ino_limit_hit; 13601 break; 13602 13603 case FLUSH_BLOCKS: 13604 case FLUSH_BLOCKS_WAIT: 13605 ACQUIRE_GBLLOCK(&lk); 13606 stat_blk_limit_push += 1; 13607 req_clear_remove += 1; 13608 FREE_GBLLOCK(&lk); 13609 stat_countp = &stat_blk_limit_hit; 13610 break; 13611 13612 default: 13613 panic("request_cleanup: unknown type"); 13614 } 13615 /* 13616 * Hopefully the syncer daemon will catch up and awaken us. 13617 * We wait at most tickdelay before proceeding in any case. 13618 */ 13619 ACQUIRE_GBLLOCK(&lk); 13620 FREE_LOCK(ump); 13621 proc_waiting += 1; 13622 if (callout_pending(&softdep_callout) == FALSE) 13623 callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2, 13624 pause_timer, 0); 13625 13626 if ((td->td_pflags & TDP_KTHREAD) == 0) 13627 msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0); 13628 proc_waiting -= 1; 13629 FREE_GBLLOCK(&lk); 13630 ACQUIRE_LOCK(ump); 13631 return (1); 13632 } 13633 13634 /* 13635 * Awaken processes pausing in request_cleanup and clear proc_waiting 13636 * to indicate that there is no longer a timer running. Pause_timer 13637 * will be called with the global softdep mutex (&lk) locked. 13638 */ 13639 static void 13640 pause_timer(arg) 13641 void *arg; 13642 { 13643 13644 GBLLOCK_OWNED(&lk); 13645 /* 13646 * The callout_ API has acquired mtx and will hold it around this 13647 * function call. 13648 */ 13649 *stat_countp += proc_waiting; 13650 wakeup(&proc_waiting); 13651 } 13652 13653 /* 13654 * If requested, try removing inode or removal dependencies. 13655 */ 13656 static void 13657 check_clear_deps(mp) 13658 struct mount *mp; 13659 { 13660 13661 /* 13662 * If we are suspended, it may be because of our using 13663 * too many inodedeps, so help clear them out. 13664 */ 13665 if (MOUNTEDSUJ(mp) && VFSTOUFS(mp)->softdep_jblocks->jb_suspended) 13666 clear_inodedeps(mp); 13667 /* 13668 * General requests for cleanup of backed up dependencies 13669 */ 13670 ACQUIRE_GBLLOCK(&lk); 13671 if (req_clear_inodedeps) { 13672 req_clear_inodedeps -= 1; 13673 FREE_GBLLOCK(&lk); 13674 clear_inodedeps(mp); 13675 ACQUIRE_GBLLOCK(&lk); 13676 wakeup(&proc_waiting); 13677 } 13678 if (req_clear_remove) { 13679 req_clear_remove -= 1; 13680 FREE_GBLLOCK(&lk); 13681 clear_remove(mp); 13682 ACQUIRE_GBLLOCK(&lk); 13683 wakeup(&proc_waiting); 13684 } 13685 FREE_GBLLOCK(&lk); 13686 } 13687 13688 /* 13689 * Flush out a directory with at least one removal dependency in an effort to 13690 * reduce the number of dirrem, freefile, and freeblks dependency structures. 13691 */ 13692 static void 13693 clear_remove(mp) 13694 struct mount *mp; 13695 { 13696 struct pagedep_hashhead *pagedephd; 13697 struct pagedep *pagedep; 13698 struct ufsmount *ump; 13699 struct vnode *vp; 13700 struct bufobj *bo; 13701 int error, cnt; 13702 ino_t ino; 13703 13704 ump = VFSTOUFS(mp); 13705 LOCK_OWNED(ump); 13706 13707 for (cnt = 0; cnt <= ump->pagedep_hash_size; cnt++) { 13708 pagedephd = &ump->pagedep_hashtbl[ump->pagedep_nextclean++]; 13709 if (ump->pagedep_nextclean > ump->pagedep_hash_size) 13710 ump->pagedep_nextclean = 0; 13711 LIST_FOREACH(pagedep, pagedephd, pd_hash) { 13712 if (LIST_EMPTY(&pagedep->pd_dirremhd)) 13713 continue; 13714 ino = pagedep->pd_ino; 13715 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 13716 continue; 13717 FREE_LOCK(ump); 13718 13719 /* 13720 * Let unmount clear deps 13721 */ 13722 error = vfs_busy(mp, MBF_NOWAIT); 13723 if (error != 0) 13724 goto finish_write; 13725 error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, 13726 FFSV_FORCEINSMQ); 13727 vfs_unbusy(mp); 13728 if (error != 0) { 13729 softdep_error("clear_remove: vget", error); 13730 goto finish_write; 13731 } 13732 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0))) 13733 softdep_error("clear_remove: fsync", error); 13734 bo = &vp->v_bufobj; 13735 BO_LOCK(bo); 13736 drain_output(vp); 13737 BO_UNLOCK(bo); 13738 vput(vp); 13739 finish_write: 13740 vn_finished_write(mp); 13741 ACQUIRE_LOCK(ump); 13742 return; 13743 } 13744 } 13745 } 13746 13747 /* 13748 * Clear out a block of dirty inodes in an effort to reduce 13749 * the number of inodedep dependency structures. 13750 */ 13751 static void 13752 clear_inodedeps(mp) 13753 struct mount *mp; 13754 { 13755 struct inodedep_hashhead *inodedephd; 13756 struct inodedep *inodedep; 13757 struct ufsmount *ump; 13758 struct vnode *vp; 13759 struct fs *fs; 13760 int error, cnt; 13761 ino_t firstino, lastino, ino; 13762 13763 ump = VFSTOUFS(mp); 13764 fs = ump->um_fs; 13765 LOCK_OWNED(ump); 13766 /* 13767 * Pick a random inode dependency to be cleared. 13768 * We will then gather up all the inodes in its block 13769 * that have dependencies and flush them out. 13770 */ 13771 for (cnt = 0; cnt <= ump->inodedep_hash_size; cnt++) { 13772 inodedephd = &ump->inodedep_hashtbl[ump->inodedep_nextclean++]; 13773 if (ump->inodedep_nextclean > ump->inodedep_hash_size) 13774 ump->inodedep_nextclean = 0; 13775 if ((inodedep = LIST_FIRST(inodedephd)) != NULL) 13776 break; 13777 } 13778 if (inodedep == NULL) 13779 return; 13780 /* 13781 * Find the last inode in the block with dependencies. 13782 */ 13783 firstino = rounddown2(inodedep->id_ino, INOPB(fs)); 13784 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--) 13785 if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0) 13786 break; 13787 /* 13788 * Asynchronously push all but the last inode with dependencies. 13789 * Synchronously push the last inode with dependencies to ensure 13790 * that the inode block gets written to free up the inodedeps. 13791 */ 13792 for (ino = firstino; ino <= lastino; ino++) { 13793 if (inodedep_lookup(mp, ino, 0, &inodedep) == 0) 13794 continue; 13795 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) 13796 continue; 13797 FREE_LOCK(ump); 13798 error = vfs_busy(mp, MBF_NOWAIT); /* Let unmount clear deps */ 13799 if (error != 0) { 13800 vn_finished_write(mp); 13801 ACQUIRE_LOCK(ump); 13802 return; 13803 } 13804 if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp, 13805 FFSV_FORCEINSMQ)) != 0) { 13806 softdep_error("clear_inodedeps: vget", error); 13807 vfs_unbusy(mp); 13808 vn_finished_write(mp); 13809 ACQUIRE_LOCK(ump); 13810 return; 13811 } 13812 vfs_unbusy(mp); 13813 if (ino == lastino) { 13814 if ((error = ffs_syncvnode(vp, MNT_WAIT, 0))) 13815 softdep_error("clear_inodedeps: fsync1", error); 13816 } else { 13817 if ((error = ffs_syncvnode(vp, MNT_NOWAIT, 0))) 13818 softdep_error("clear_inodedeps: fsync2", error); 13819 BO_LOCK(&vp->v_bufobj); 13820 drain_output(vp); 13821 BO_UNLOCK(&vp->v_bufobj); 13822 } 13823 vput(vp); 13824 vn_finished_write(mp); 13825 ACQUIRE_LOCK(ump); 13826 } 13827 } 13828 13829 void 13830 softdep_buf_append(bp, wkhd) 13831 struct buf *bp; 13832 struct workhead *wkhd; 13833 { 13834 struct worklist *wk; 13835 struct ufsmount *ump; 13836 13837 if ((wk = LIST_FIRST(wkhd)) == NULL) 13838 return; 13839 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 13840 ("softdep_buf_append called on non-softdep filesystem")); 13841 ump = VFSTOUFS(wk->wk_mp); 13842 ACQUIRE_LOCK(ump); 13843 while ((wk = LIST_FIRST(wkhd)) != NULL) { 13844 WORKLIST_REMOVE(wk); 13845 WORKLIST_INSERT(&bp->b_dep, wk); 13846 } 13847 FREE_LOCK(ump); 13848 13849 } 13850 13851 void 13852 softdep_inode_append(ip, cred, wkhd) 13853 struct inode *ip; 13854 struct ucred *cred; 13855 struct workhead *wkhd; 13856 { 13857 struct buf *bp; 13858 struct fs *fs; 13859 struct ufsmount *ump; 13860 int error; 13861 13862 ump = ITOUMP(ip); 13863 KASSERT(MOUNTEDSOFTDEP(UFSTOVFS(ump)) != 0, 13864 ("softdep_inode_append called on non-softdep filesystem")); 13865 fs = ump->um_fs; 13866 error = bread(ump->um_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)), 13867 (int)fs->fs_bsize, cred, &bp); 13868 if (error) { 13869 bqrelse(bp); 13870 softdep_freework(wkhd); 13871 return; 13872 } 13873 softdep_buf_append(bp, wkhd); 13874 bqrelse(bp); 13875 } 13876 13877 void 13878 softdep_freework(wkhd) 13879 struct workhead *wkhd; 13880 { 13881 struct worklist *wk; 13882 struct ufsmount *ump; 13883 13884 if ((wk = LIST_FIRST(wkhd)) == NULL) 13885 return; 13886 KASSERT(MOUNTEDSOFTDEP(wk->wk_mp) != 0, 13887 ("softdep_freework called on non-softdep filesystem")); 13888 ump = VFSTOUFS(wk->wk_mp); 13889 ACQUIRE_LOCK(ump); 13890 handle_jwork(wkhd); 13891 FREE_LOCK(ump); 13892 } 13893 13894 static struct ufsmount * 13895 softdep_bp_to_mp(bp) 13896 struct buf *bp; 13897 { 13898 struct mount *mp; 13899 struct vnode *vp; 13900 13901 if (LIST_EMPTY(&bp->b_dep)) 13902 return (NULL); 13903 vp = bp->b_vp; 13904 KASSERT(vp != NULL, 13905 ("%s, buffer with dependencies lacks vnode", __func__)); 13906 13907 /* 13908 * The ump mount point is stable after we get a correct 13909 * pointer, since bp is locked and this prevents unmount from 13910 * proceeding. But to get to it, we cannot dereference bp->b_dep 13911 * head wk_mp, because we do not yet own SU ump lock and 13912 * workitem might be freed while dereferenced. 13913 */ 13914 retry: 13915 switch (vp->v_type) { 13916 case VCHR: 13917 VI_LOCK(vp); 13918 mp = vp->v_type == VCHR ? vp->v_rdev->si_mountpt : NULL; 13919 VI_UNLOCK(vp); 13920 if (mp == NULL) 13921 goto retry; 13922 break; 13923 case VREG: 13924 case VDIR: 13925 case VLNK: 13926 case VFIFO: 13927 case VSOCK: 13928 mp = vp->v_mount; 13929 break; 13930 case VBLK: 13931 vn_printf(vp, "softdep_bp_to_mp: unexpected block device\n"); 13932 /* FALLTHROUGH */ 13933 case VNON: 13934 case VBAD: 13935 case VMARKER: 13936 mp = NULL; 13937 break; 13938 default: 13939 vn_printf(vp, "unknown vnode type"); 13940 mp = NULL; 13941 break; 13942 } 13943 return (VFSTOUFS(mp)); 13944 } 13945 13946 /* 13947 * Function to determine if the buffer has outstanding dependencies 13948 * that will cause a roll-back if the buffer is written. If wantcount 13949 * is set, return number of dependencies, otherwise just yes or no. 13950 */ 13951 static int 13952 softdep_count_dependencies(bp, wantcount) 13953 struct buf *bp; 13954 int wantcount; 13955 { 13956 struct worklist *wk; 13957 struct ufsmount *ump; 13958 struct bmsafemap *bmsafemap; 13959 struct freework *freework; 13960 struct inodedep *inodedep; 13961 struct indirdep *indirdep; 13962 struct freeblks *freeblks; 13963 struct allocindir *aip; 13964 struct pagedep *pagedep; 13965 struct dirrem *dirrem; 13966 struct newblk *newblk; 13967 struct mkdir *mkdir; 13968 struct diradd *dap; 13969 int i, retval; 13970 13971 ump = softdep_bp_to_mp(bp); 13972 if (ump == NULL) 13973 return (0); 13974 retval = 0; 13975 ACQUIRE_LOCK(ump); 13976 LIST_FOREACH(wk, &bp->b_dep, wk_list) { 13977 switch (wk->wk_type) { 13978 13979 case D_INODEDEP: 13980 inodedep = WK_INODEDEP(wk); 13981 if ((inodedep->id_state & DEPCOMPLETE) == 0) { 13982 /* bitmap allocation dependency */ 13983 retval += 1; 13984 if (!wantcount) 13985 goto out; 13986 } 13987 if (TAILQ_FIRST(&inodedep->id_inoupdt)) { 13988 /* direct block pointer dependency */ 13989 retval += 1; 13990 if (!wantcount) 13991 goto out; 13992 } 13993 if (TAILQ_FIRST(&inodedep->id_extupdt)) { 13994 /* direct block pointer dependency */ 13995 retval += 1; 13996 if (!wantcount) 13997 goto out; 13998 } 13999 if (TAILQ_FIRST(&inodedep->id_inoreflst)) { 14000 /* Add reference dependency. */ 14001 retval += 1; 14002 if (!wantcount) 14003 goto out; 14004 } 14005 continue; 14006 14007 case D_INDIRDEP: 14008 indirdep = WK_INDIRDEP(wk); 14009 14010 TAILQ_FOREACH(freework, &indirdep->ir_trunc, fw_next) { 14011 /* indirect truncation dependency */ 14012 retval += 1; 14013 if (!wantcount) 14014 goto out; 14015 } 14016 14017 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) { 14018 /* indirect block pointer dependency */ 14019 retval += 1; 14020 if (!wantcount) 14021 goto out; 14022 } 14023 continue; 14024 14025 case D_PAGEDEP: 14026 pagedep = WK_PAGEDEP(wk); 14027 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) { 14028 if (LIST_FIRST(&dirrem->dm_jremrefhd)) { 14029 /* Journal remove ref dependency. */ 14030 retval += 1; 14031 if (!wantcount) 14032 goto out; 14033 } 14034 } 14035 for (i = 0; i < DAHASHSZ; i++) { 14036 14037 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) { 14038 /* directory entry dependency */ 14039 retval += 1; 14040 if (!wantcount) 14041 goto out; 14042 } 14043 } 14044 continue; 14045 14046 case D_BMSAFEMAP: 14047 bmsafemap = WK_BMSAFEMAP(wk); 14048 if (LIST_FIRST(&bmsafemap->sm_jaddrefhd)) { 14049 /* Add reference dependency. */ 14050 retval += 1; 14051 if (!wantcount) 14052 goto out; 14053 } 14054 if (LIST_FIRST(&bmsafemap->sm_jnewblkhd)) { 14055 /* Allocate block dependency. */ 14056 retval += 1; 14057 if (!wantcount) 14058 goto out; 14059 } 14060 continue; 14061 14062 case D_FREEBLKS: 14063 freeblks = WK_FREEBLKS(wk); 14064 if (LIST_FIRST(&freeblks->fb_jblkdephd)) { 14065 /* Freeblk journal dependency. */ 14066 retval += 1; 14067 if (!wantcount) 14068 goto out; 14069 } 14070 continue; 14071 14072 case D_ALLOCDIRECT: 14073 case D_ALLOCINDIR: 14074 newblk = WK_NEWBLK(wk); 14075 if (newblk->nb_jnewblk) { 14076 /* Journal allocate dependency. */ 14077 retval += 1; 14078 if (!wantcount) 14079 goto out; 14080 } 14081 continue; 14082 14083 case D_MKDIR: 14084 mkdir = WK_MKDIR(wk); 14085 if (mkdir->md_jaddref) { 14086 /* Journal reference dependency. */ 14087 retval += 1; 14088 if (!wantcount) 14089 goto out; 14090 } 14091 continue; 14092 14093 case D_FREEWORK: 14094 case D_FREEDEP: 14095 case D_JSEGDEP: 14096 case D_JSEG: 14097 case D_SBDEP: 14098 /* never a dependency on these blocks */ 14099 continue; 14100 14101 default: 14102 panic("softdep_count_dependencies: Unexpected type %s", 14103 TYPENAME(wk->wk_type)); 14104 /* NOTREACHED */ 14105 } 14106 } 14107 out: 14108 FREE_LOCK(ump); 14109 return (retval); 14110 } 14111 14112 /* 14113 * Acquire exclusive access to a buffer. 14114 * Must be called with a locked mtx parameter. 14115 * Return acquired buffer or NULL on failure. 14116 */ 14117 static struct buf * 14118 getdirtybuf(bp, lock, waitfor) 14119 struct buf *bp; 14120 struct rwlock *lock; 14121 int waitfor; 14122 { 14123 int error; 14124 14125 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) { 14126 if (waitfor != MNT_WAIT) 14127 return (NULL); 14128 error = BUF_LOCK(bp, 14129 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, lock); 14130 /* 14131 * Even if we successfully acquire bp here, we have dropped 14132 * lock, which may violates our guarantee. 14133 */ 14134 if (error == 0) 14135 BUF_UNLOCK(bp); 14136 else if (error != ENOLCK) 14137 panic("getdirtybuf: inconsistent lock: %d", error); 14138 rw_wlock(lock); 14139 return (NULL); 14140 } 14141 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 14142 if (lock != BO_LOCKPTR(bp->b_bufobj) && waitfor == MNT_WAIT) { 14143 rw_wunlock(lock); 14144 BO_LOCK(bp->b_bufobj); 14145 BUF_UNLOCK(bp); 14146 if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { 14147 bp->b_vflags |= BV_BKGRDWAIT; 14148 msleep(&bp->b_xflags, BO_LOCKPTR(bp->b_bufobj), 14149 PRIBIO | PDROP, "getbuf", 0); 14150 } else 14151 BO_UNLOCK(bp->b_bufobj); 14152 rw_wlock(lock); 14153 return (NULL); 14154 } 14155 BUF_UNLOCK(bp); 14156 if (waitfor != MNT_WAIT) 14157 return (NULL); 14158 #ifdef DEBUG_VFS_LOCKS 14159 if (bp->b_vp->v_type != VCHR) 14160 ASSERT_BO_WLOCKED(bp->b_bufobj); 14161 #endif 14162 bp->b_vflags |= BV_BKGRDWAIT; 14163 rw_sleep(&bp->b_xflags, lock, PRIBIO, "getbuf", 0); 14164 return (NULL); 14165 } 14166 if ((bp->b_flags & B_DELWRI) == 0) { 14167 BUF_UNLOCK(bp); 14168 return (NULL); 14169 } 14170 bremfree(bp); 14171 return (bp); 14172 } 14173 14174 14175 /* 14176 * Check if it is safe to suspend the file system now. On entry, 14177 * the vnode interlock for devvp should be held. Return 0 with 14178 * the mount interlock held if the file system can be suspended now, 14179 * otherwise return EAGAIN with the mount interlock held. 14180 */ 14181 int 14182 softdep_check_suspend(struct mount *mp, 14183 struct vnode *devvp, 14184 int softdep_depcnt, 14185 int softdep_accdepcnt, 14186 int secondary_writes, 14187 int secondary_accwrites) 14188 { 14189 struct bufobj *bo; 14190 struct ufsmount *ump; 14191 struct inodedep *inodedep; 14192 int error, unlinked; 14193 14194 bo = &devvp->v_bufobj; 14195 ASSERT_BO_WLOCKED(bo); 14196 14197 /* 14198 * If we are not running with soft updates, then we need only 14199 * deal with secondary writes as we try to suspend. 14200 */ 14201 if (MOUNTEDSOFTDEP(mp) == 0) { 14202 MNT_ILOCK(mp); 14203 while (mp->mnt_secondary_writes != 0) { 14204 BO_UNLOCK(bo); 14205 msleep(&mp->mnt_secondary_writes, MNT_MTX(mp), 14206 (PUSER - 1) | PDROP, "secwr", 0); 14207 BO_LOCK(bo); 14208 MNT_ILOCK(mp); 14209 } 14210 14211 /* 14212 * Reasons for needing more work before suspend: 14213 * - Dirty buffers on devvp. 14214 * - Secondary writes occurred after start of vnode sync loop 14215 */ 14216 error = 0; 14217 if (bo->bo_numoutput > 0 || 14218 bo->bo_dirty.bv_cnt > 0 || 14219 secondary_writes != 0 || 14220 mp->mnt_secondary_writes != 0 || 14221 secondary_accwrites != mp->mnt_secondary_accwrites) 14222 error = EAGAIN; 14223 BO_UNLOCK(bo); 14224 return (error); 14225 } 14226 14227 /* 14228 * If we are running with soft updates, then we need to coordinate 14229 * with them as we try to suspend. 14230 */ 14231 ump = VFSTOUFS(mp); 14232 for (;;) { 14233 if (!TRY_ACQUIRE_LOCK(ump)) { 14234 BO_UNLOCK(bo); 14235 ACQUIRE_LOCK(ump); 14236 FREE_LOCK(ump); 14237 BO_LOCK(bo); 14238 continue; 14239 } 14240 MNT_ILOCK(mp); 14241 if (mp->mnt_secondary_writes != 0) { 14242 FREE_LOCK(ump); 14243 BO_UNLOCK(bo); 14244 msleep(&mp->mnt_secondary_writes, 14245 MNT_MTX(mp), 14246 (PUSER - 1) | PDROP, "secwr", 0); 14247 BO_LOCK(bo); 14248 continue; 14249 } 14250 break; 14251 } 14252 14253 unlinked = 0; 14254 if (MOUNTEDSUJ(mp)) { 14255 for (inodedep = TAILQ_FIRST(&ump->softdep_unlinked); 14256 inodedep != NULL; 14257 inodedep = TAILQ_NEXT(inodedep, id_unlinked)) { 14258 if ((inodedep->id_state & (UNLINKED | UNLINKLINKS | 14259 UNLINKONLIST)) != (UNLINKED | UNLINKLINKS | 14260 UNLINKONLIST) || 14261 !check_inodedep_free(inodedep)) 14262 continue; 14263 unlinked++; 14264 } 14265 } 14266 14267 /* 14268 * Reasons for needing more work before suspend: 14269 * - Dirty buffers on devvp. 14270 * - Softdep activity occurred after start of vnode sync loop 14271 * - Secondary writes occurred after start of vnode sync loop 14272 */ 14273 error = 0; 14274 if (bo->bo_numoutput > 0 || 14275 bo->bo_dirty.bv_cnt > 0 || 14276 softdep_depcnt != unlinked || 14277 ump->softdep_deps != unlinked || 14278 softdep_accdepcnt != ump->softdep_accdeps || 14279 secondary_writes != 0 || 14280 mp->mnt_secondary_writes != 0 || 14281 secondary_accwrites != mp->mnt_secondary_accwrites) 14282 error = EAGAIN; 14283 FREE_LOCK(ump); 14284 BO_UNLOCK(bo); 14285 return (error); 14286 } 14287 14288 14289 /* 14290 * Get the number of dependency structures for the file system, both 14291 * the current number and the total number allocated. These will 14292 * later be used to detect that softdep processing has occurred. 14293 */ 14294 void 14295 softdep_get_depcounts(struct mount *mp, 14296 int *softdep_depsp, 14297 int *softdep_accdepsp) 14298 { 14299 struct ufsmount *ump; 14300 14301 if (MOUNTEDSOFTDEP(mp) == 0) { 14302 *softdep_depsp = 0; 14303 *softdep_accdepsp = 0; 14304 return; 14305 } 14306 ump = VFSTOUFS(mp); 14307 ACQUIRE_LOCK(ump); 14308 *softdep_depsp = ump->softdep_deps; 14309 *softdep_accdepsp = ump->softdep_accdeps; 14310 FREE_LOCK(ump); 14311 } 14312 14313 /* 14314 * Wait for pending output on a vnode to complete. 14315 */ 14316 static void 14317 drain_output(vp) 14318 struct vnode *vp; 14319 { 14320 14321 ASSERT_VOP_LOCKED(vp, "drain_output"); 14322 (void)bufobj_wwait(&vp->v_bufobj, 0, 0); 14323 } 14324 14325 /* 14326 * Called whenever a buffer that is being invalidated or reallocated 14327 * contains dependencies. This should only happen if an I/O error has 14328 * occurred. The routine is called with the buffer locked. 14329 */ 14330 static void 14331 softdep_deallocate_dependencies(bp) 14332 struct buf *bp; 14333 { 14334 14335 if ((bp->b_ioflags & BIO_ERROR) == 0) 14336 panic("softdep_deallocate_dependencies: dangling deps"); 14337 if (bp->b_vp != NULL && bp->b_vp->v_mount != NULL) 14338 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error); 14339 else 14340 printf("softdep_deallocate_dependencies: " 14341 "got error %d while accessing filesystem\n", bp->b_error); 14342 if (bp->b_error != ENXIO) 14343 panic("softdep_deallocate_dependencies: unrecovered I/O error"); 14344 } 14345 14346 /* 14347 * Function to handle asynchronous write errors in the filesystem. 14348 */ 14349 static void 14350 softdep_error(func, error) 14351 char *func; 14352 int error; 14353 { 14354 14355 /* XXX should do something better! */ 14356 printf("%s: got error %d while accessing filesystem\n", func, error); 14357 } 14358 14359 #ifdef DDB 14360 14361 static void 14362 inodedep_print(struct inodedep *inodedep, int verbose) 14363 { 14364 db_printf("%p fs %p st %x ino %jd inoblk %jd delta %jd nlink %jd" 14365 " saveino %p\n", 14366 inodedep, inodedep->id_fs, inodedep->id_state, 14367 (intmax_t)inodedep->id_ino, 14368 (intmax_t)fsbtodb(inodedep->id_fs, 14369 ino_to_fsba(inodedep->id_fs, inodedep->id_ino)), 14370 (intmax_t)inodedep->id_nlinkdelta, 14371 (intmax_t)inodedep->id_savednlink, 14372 inodedep->id_savedino1); 14373 14374 if (verbose == 0) 14375 return; 14376 14377 db_printf("\tpendinghd %p, bufwait %p, inowait %p, inoreflst %p, " 14378 "mkdiradd %p\n", 14379 LIST_FIRST(&inodedep->id_pendinghd), 14380 LIST_FIRST(&inodedep->id_bufwait), 14381 LIST_FIRST(&inodedep->id_inowait), 14382 TAILQ_FIRST(&inodedep->id_inoreflst), 14383 inodedep->id_mkdiradd); 14384 db_printf("\tinoupdt %p, newinoupdt %p, extupdt %p, newextupdt %p\n", 14385 TAILQ_FIRST(&inodedep->id_inoupdt), 14386 TAILQ_FIRST(&inodedep->id_newinoupdt), 14387 TAILQ_FIRST(&inodedep->id_extupdt), 14388 TAILQ_FIRST(&inodedep->id_newextupdt)); 14389 } 14390 14391 DB_SHOW_COMMAND(inodedep, db_show_inodedep) 14392 { 14393 14394 if (have_addr == 0) { 14395 db_printf("Address required\n"); 14396 return; 14397 } 14398 inodedep_print((struct inodedep*)addr, 1); 14399 } 14400 14401 DB_SHOW_COMMAND(inodedeps, db_show_inodedeps) 14402 { 14403 struct inodedep_hashhead *inodedephd; 14404 struct inodedep *inodedep; 14405 struct ufsmount *ump; 14406 int cnt; 14407 14408 if (have_addr == 0) { 14409 db_printf("Address required\n"); 14410 return; 14411 } 14412 ump = (struct ufsmount *)addr; 14413 for (cnt = 0; cnt < ump->inodedep_hash_size; cnt++) { 14414 inodedephd = &ump->inodedep_hashtbl[cnt]; 14415 LIST_FOREACH(inodedep, inodedephd, id_hash) { 14416 inodedep_print(inodedep, 0); 14417 } 14418 } 14419 } 14420 14421 DB_SHOW_COMMAND(worklist, db_show_worklist) 14422 { 14423 struct worklist *wk; 14424 14425 if (have_addr == 0) { 14426 db_printf("Address required\n"); 14427 return; 14428 } 14429 wk = (struct worklist *)addr; 14430 printf("worklist: %p type %s state 0x%X\n", 14431 wk, TYPENAME(wk->wk_type), wk->wk_state); 14432 } 14433 14434 DB_SHOW_COMMAND(workhead, db_show_workhead) 14435 { 14436 struct workhead *wkhd; 14437 struct worklist *wk; 14438 int i; 14439 14440 if (have_addr == 0) { 14441 db_printf("Address required\n"); 14442 return; 14443 } 14444 wkhd = (struct workhead *)addr; 14445 wk = LIST_FIRST(wkhd); 14446 for (i = 0; i < 100 && wk != NULL; i++, wk = LIST_NEXT(wk, wk_list)) 14447 db_printf("worklist: %p type %s state 0x%X", 14448 wk, TYPENAME(wk->wk_type), wk->wk_state); 14449 if (i == 100) 14450 db_printf("workhead overflow"); 14451 printf("\n"); 14452 } 14453 14454 14455 DB_SHOW_COMMAND(mkdirs, db_show_mkdirs) 14456 { 14457 struct mkdirlist *mkdirlisthd; 14458 struct jaddref *jaddref; 14459 struct diradd *diradd; 14460 struct mkdir *mkdir; 14461 14462 if (have_addr == 0) { 14463 db_printf("Address required\n"); 14464 return; 14465 } 14466 mkdirlisthd = (struct mkdirlist *)addr; 14467 LIST_FOREACH(mkdir, mkdirlisthd, md_mkdirs) { 14468 diradd = mkdir->md_diradd; 14469 db_printf("mkdir: %p state 0x%X dap %p state 0x%X", 14470 mkdir, mkdir->md_state, diradd, diradd->da_state); 14471 if ((jaddref = mkdir->md_jaddref) != NULL) 14472 db_printf(" jaddref %p jaddref state 0x%X", 14473 jaddref, jaddref->ja_state); 14474 db_printf("\n"); 14475 } 14476 } 14477 14478 /* exported to ffs_vfsops.c */ 14479 extern void db_print_ffs(struct ufsmount *ump); 14480 void 14481 db_print_ffs(struct ufsmount *ump) 14482 { 14483 db_printf("mp %p %s devvp %p fs %p su_wl %d su_deps %d su_req %d\n", 14484 ump->um_mountp, ump->um_mountp->mnt_stat.f_mntonname, 14485 ump->um_devvp, ump->um_fs, ump->softdep_on_worklist, 14486 ump->softdep_deps, ump->softdep_req); 14487 } 14488 14489 #endif /* DDB */ 14490 14491 #endif /* SOFTUPDATES */
Cache object: c1e527735145d97a7d36eda6347fa279
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]