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


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

FreeBSD/Linux Kernel Cross Reference
sys/kern/vfs_subr.c

Version: -  FREEBSD  -  FREEBSD-12-STABLE  -  FREEBSD-12-0  -  FREEBSD-11-STABLE  -  FREEBSD-11-2  -  FREEBSD-11-1  -  FREEBSD-11-0  -  FREEBSD-10-STABLE  -  FREEBSD-10-4  -  FREEBSD-10-3  -  FREEBSD-10-2  -  FREEBSD-10-1  -  FREEBSD-10-0  -  FREEBSD-9-STABLE  -  FREEBSD-9-3  -  FREEBSD-9-2  -  FREEBSD-9-1  -  FREEBSD-9-0  -  FREEBSD-8-STABLE  -  FREEBSD-8-4  -  FREEBSD-8-3  -  FREEBSD-8-2  -  FREEBSD-8-1  -  FREEBSD-8-0  -  FREEBSD-7-STABLE  -  FREEBSD-7-4  -  FREEBSD-7-3  -  FREEBSD-7-2  -  FREEBSD-7-1  -  FREEBSD-7-0  -  FREEBSD-6-STABLE  -  FREEBSD-6-4  -  FREEBSD-6-3  -  FREEBSD-6-2  -  FREEBSD-6-1  -  FREEBSD-6-0  -  FREEBSD-5-STABLE  -  FREEBSD-5-5  -  FREEBSD-5-4  -  FREEBSD-5-3  -  FREEBSD-5-2  -  FREEBSD-5-1  -  FREEBSD-5-0  -  FREEBSD-4-STABLE  -  FREEBSD-3-STABLE  -  FREEBSD22  -  linux-2.6  -  linux-2.4.22  -  MK83  -  MK84  -  PLAN9  -  DFBSD  -  NETBSD  -  NETBSD5  -  NETBSD4  -  NETBSD3  -  NETBSD20  -  OPENBSD  -  xnu-517  -  xnu-792  -  xnu-792.6.70  -  xnu-1228  -  xnu-1456.1.26  -  xnu-1699.24.8  -  xnu-2050.18.24  -  OPENSOLARIS  -  minix-3-1-1 
SearchContext: -  none  -  3  -  10 

    1 /*-
    2  * SPDX-License-Identifier: BSD-3-Clause
    3  *
    4  * Copyright (c) 1989, 1993
    5  *      The Regents of the University of California.  All rights reserved.
    6  * (c) UNIX System Laboratories, Inc.
    7  * All or some portions of this file are derived from material licensed
    8  * to the University of California by American Telephone and Telegraph
    9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
   10  * the permission of UNIX System Laboratories, Inc.
   11  *
   12  * Redistribution and use in source and binary forms, with or without
   13  * modification, are permitted provided that the following conditions
   14  * are met:
   15  * 1. Redistributions of source code must retain the above copyright
   16  *    notice, this list of conditions and the following disclaimer.
   17  * 2. Redistributions in binary form must reproduce the above copyright
   18  *    notice, this list of conditions and the following disclaimer in the
   19  *    documentation and/or other materials provided with the distribution.
   20  * 3. Neither the name of the University nor the names of its contributors
   21  *    may be used to endorse or promote products derived from this software
   22  *    without specific prior written permission.
   23  *
   24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
   25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
   26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
   27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
   28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
   29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
   30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
   31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
   32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
   33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
   34  * SUCH DAMAGE.
   35  *
   36  *      @(#)vfs_subr.c  8.31 (Berkeley) 5/26/95
   37  */
   38 
   39 /*
   40  * External virtual filesystem routines
   41  */
   42 
   43 #include <sys/cdefs.h>
   44 __FBSDID("$FreeBSD: releng/12.0/sys/kern/vfs_subr.c 337977 2018-08-17 16:07:06Z markj $");
   45 
   46 #include "opt_ddb.h"
   47 #include "opt_watchdog.h"
   48 
   49 #include <sys/param.h>
   50 #include <sys/systm.h>
   51 #include <sys/bio.h>
   52 #include <sys/buf.h>
   53 #include <sys/condvar.h>
   54 #include <sys/conf.h>
   55 #include <sys/counter.h>
   56 #include <sys/dirent.h>
   57 #include <sys/event.h>
   58 #include <sys/eventhandler.h>
   59 #include <sys/extattr.h>
   60 #include <sys/file.h>
   61 #include <sys/fcntl.h>
   62 #include <sys/jail.h>
   63 #include <sys/kdb.h>
   64 #include <sys/kernel.h>
   65 #include <sys/kthread.h>
   66 #include <sys/lockf.h>
   67 #include <sys/malloc.h>
   68 #include <sys/mount.h>
   69 #include <sys/namei.h>
   70 #include <sys/pctrie.h>
   71 #include <sys/priv.h>
   72 #include <sys/reboot.h>
   73 #include <sys/refcount.h>
   74 #include <sys/rwlock.h>
   75 #include <sys/sched.h>
   76 #include <sys/sleepqueue.h>
   77 #include <sys/smp.h>
   78 #include <sys/stat.h>
   79 #include <sys/sysctl.h>
   80 #include <sys/syslog.h>
   81 #include <sys/vmmeter.h>
   82 #include <sys/vnode.h>
   83 #include <sys/watchdog.h>
   84 
   85 #include <machine/stdarg.h>
   86 
   87 #include <security/mac/mac_framework.h>
   88 
   89 #include <vm/vm.h>
   90 #include <vm/vm_object.h>
   91 #include <vm/vm_extern.h>
   92 #include <vm/pmap.h>
   93 #include <vm/vm_map.h>
   94 #include <vm/vm_page.h>
   95 #include <vm/vm_kern.h>
   96 #include <vm/uma.h>
   97 
   98 #ifdef DDB
   99 #include <ddb/ddb.h>
  100 #endif
  101 
  102 static void     delmntque(struct vnode *vp);
  103 static int      flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
  104                     int slpflag, int slptimeo);
  105 static void     syncer_shutdown(void *arg, int howto);
  106 static int      vtryrecycle(struct vnode *vp);
  107 static void     v_init_counters(struct vnode *);
  108 static void     v_incr_usecount(struct vnode *);
  109 static void     v_incr_usecount_locked(struct vnode *);
  110 static void     v_incr_devcount(struct vnode *);
  111 static void     v_decr_devcount(struct vnode *);
  112 static void     vgonel(struct vnode *);
  113 static void     vfs_knllock(void *arg);
  114 static void     vfs_knlunlock(void *arg);
  115 static void     vfs_knl_assert_locked(void *arg);
  116 static void     vfs_knl_assert_unlocked(void *arg);
  117 static void     vnlru_return_batches(struct vfsops *mnt_op);
  118 static void     destroy_vpollinfo(struct vpollinfo *vi);
  119 
  120 /*
  121  * These fences are intended for cases where some synchronization is
  122  * needed between access of v_iflags and lockless vnode refcount (v_holdcnt
  123  * and v_usecount) updates.  Access to v_iflags is generally synchronized
  124  * by the interlock, but we have some internal assertions that check vnode
  125  * flags without acquiring the lock.  Thus, these fences are INVARIANTS-only
  126  * for now.
  127  */
  128 #ifdef INVARIANTS
  129 #define VNODE_REFCOUNT_FENCE_ACQ()      atomic_thread_fence_acq()
  130 #define VNODE_REFCOUNT_FENCE_REL()      atomic_thread_fence_rel()
  131 #else
  132 #define VNODE_REFCOUNT_FENCE_ACQ()
  133 #define VNODE_REFCOUNT_FENCE_REL()
  134 #endif
  135 
  136 /*
  137  * Number of vnodes in existence.  Increased whenever getnewvnode()
  138  * allocates a new vnode, decreased in vdropl() for VI_DOOMED vnode.
  139  */
  140 static unsigned long    numvnodes;
  141 
  142 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
  143     "Number of vnodes in existence");
  144 
  145 static counter_u64_t vnodes_created;
  146 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
  147     "Number of vnodes created by getnewvnode");
  148 
  149 static u_long mnt_free_list_batch = 128;
  150 SYSCTL_ULONG(_vfs, OID_AUTO, mnt_free_list_batch, CTLFLAG_RW,
  151     &mnt_free_list_batch, 0, "Limit of vnodes held on mnt's free list");
  152 
  153 /*
  154  * Conversion tables for conversion from vnode types to inode formats
  155  * and back.
  156  */
  157 enum vtype iftovt_tab[16] = {
  158         VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
  159         VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
  160 };
  161 int vttoif_tab[10] = {
  162         0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
  163         S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
  164 };
  165 
  166 /*
  167  * List of vnodes that are ready for recycling.
  168  */
  169 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
  170 
  171 /*
  172  * "Free" vnode target.  Free vnodes are rarely completely free, but are
  173  * just ones that are cheap to recycle.  Usually they are for files which
  174  * have been stat'd but not read; these usually have inode and namecache
  175  * data attached to them.  This target is the preferred minimum size of a
  176  * sub-cache consisting mostly of such files. The system balances the size
  177  * of this sub-cache with its complement to try to prevent either from
  178  * thrashing while the other is relatively inactive.  The targets express
  179  * a preference for the best balance.
  180  *
  181  * "Above" this target there are 2 further targets (watermarks) related
  182  * to recyling of free vnodes.  In the best-operating case, the cache is
  183  * exactly full, the free list has size between vlowat and vhiwat above the
  184  * free target, and recycling from it and normal use maintains this state.
  185  * Sometimes the free list is below vlowat or even empty, but this state
  186  * is even better for immediate use provided the cache is not full.
  187  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
  188  * ones) to reach one of these states.  The watermarks are currently hard-
  189  * coded as 4% and 9% of the available space higher.  These and the default
  190  * of 25% for wantfreevnodes are too large if the memory size is large.
  191  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
  192  * whenever vnlru_proc() becomes active.
  193  */
  194 static u_long wantfreevnodes;
  195 SYSCTL_ULONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW,
  196     &wantfreevnodes, 0, "Target for minimum number of \"free\" vnodes");
  197 static u_long freevnodes;
  198 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
  199     &freevnodes, 0, "Number of \"free\" vnodes");
  200 
  201 static counter_u64_t recycles_count;
  202 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
  203     "Number of vnodes recycled to meet vnode cache targets");
  204 
  205 /*
  206  * Various variables used for debugging the new implementation of
  207  * reassignbuf().
  208  * XXX these are probably of (very) limited utility now.
  209  */
  210 static int reassignbufcalls;
  211 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0,
  212     "Number of calls to reassignbuf");
  213 
  214 static counter_u64_t free_owe_inact;
  215 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, free_owe_inact, CTLFLAG_RD, &free_owe_inact,
  216     "Number of times free vnodes kept on active list due to VFS "
  217     "owing inactivation");
  218 
  219 /* To keep more than one thread at a time from running vfs_getnewfsid */
  220 static struct mtx mntid_mtx;
  221 
  222 /*
  223  * Lock for any access to the following:
  224  *      vnode_free_list
  225  *      numvnodes
  226  *      freevnodes
  227  */
  228 static struct mtx vnode_free_list_mtx;
  229 
  230 /* Publicly exported FS */
  231 struct nfs_public nfs_pub;
  232 
  233 static uma_zone_t buf_trie_zone;
  234 
  235 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
  236 static uma_zone_t vnode_zone;
  237 static uma_zone_t vnodepoll_zone;
  238 
  239 /*
  240  * The workitem queue.
  241  *
  242  * It is useful to delay writes of file data and filesystem metadata
  243  * for tens of seconds so that quickly created and deleted files need
  244  * not waste disk bandwidth being created and removed. To realize this,
  245  * we append vnodes to a "workitem" queue. When running with a soft
  246  * updates implementation, most pending metadata dependencies should
  247  * not wait for more than a few seconds. Thus, mounted on block devices
  248  * are delayed only about a half the time that file data is delayed.
  249  * Similarly, directory updates are more critical, so are only delayed
  250  * about a third the time that file data is delayed. Thus, there are
  251  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
  252  * one each second (driven off the filesystem syncer process). The
  253  * syncer_delayno variable indicates the next queue that is to be processed.
  254  * Items that need to be processed soon are placed in this queue:
  255  *
  256  *      syncer_workitem_pending[syncer_delayno]
  257  *
  258  * A delay of fifteen seconds is done by placing the request fifteen
  259  * entries later in the queue:
  260  *
  261  *      syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
  262  *
  263  */
  264 static int syncer_delayno;
  265 static long syncer_mask;
  266 LIST_HEAD(synclist, bufobj);
  267 static struct synclist *syncer_workitem_pending;
  268 /*
  269  * The sync_mtx protects:
  270  *      bo->bo_synclist
  271  *      sync_vnode_count
  272  *      syncer_delayno
  273  *      syncer_state
  274  *      syncer_workitem_pending
  275  *      syncer_worklist_len
  276  *      rushjob
  277  */
  278 static struct mtx sync_mtx;
  279 static struct cv sync_wakeup;
  280 
  281 #define SYNCER_MAXDELAY         32
  282 static int syncer_maxdelay = SYNCER_MAXDELAY;   /* maximum delay time */
  283 static int syncdelay = 30;              /* max time to delay syncing data */
  284 static int filedelay = 30;              /* time to delay syncing files */
  285 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
  286     "Time to delay syncing files (in seconds)");
  287 static int dirdelay = 29;               /* time to delay syncing directories */
  288 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
  289     "Time to delay syncing directories (in seconds)");
  290 static int metadelay = 28;              /* time to delay syncing metadata */
  291 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
  292     "Time to delay syncing metadata (in seconds)");
  293 static int rushjob;             /* number of slots to run ASAP */
  294 static int stat_rush_requests;  /* number of times I/O speeded up */
  295 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
  296     "Number of times I/O speeded up (rush requests)");
  297 
  298 /*
  299  * When shutting down the syncer, run it at four times normal speed.
  300  */
  301 #define SYNCER_SHUTDOWN_SPEEDUP         4
  302 static int sync_vnode_count;
  303 static int syncer_worklist_len;
  304 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
  305     syncer_state;
  306 
  307 /* Target for maximum number of vnodes. */
  308 int desiredvnodes;
  309 static int gapvnodes;           /* gap between wanted and desired */
  310 static int vhiwat;              /* enough extras after expansion */
  311 static int vlowat;              /* minimal extras before expansion */
  312 static int vstir;               /* nonzero to stir non-free vnodes */
  313 static volatile int vsmalltrigger = 8;  /* pref to keep if > this many pages */
  314 
  315 static int
  316 sysctl_update_desiredvnodes(SYSCTL_HANDLER_ARGS)
  317 {
  318         int error, old_desiredvnodes;
  319 
  320         old_desiredvnodes = desiredvnodes;
  321         if ((error = sysctl_handle_int(oidp, arg1, arg2, req)) != 0)
  322                 return (error);
  323         if (old_desiredvnodes != desiredvnodes) {
  324                 wantfreevnodes = desiredvnodes / 4;
  325                 /* XXX locking seems to be incomplete. */
  326                 vfs_hash_changesize(desiredvnodes);
  327                 cache_changesize(desiredvnodes);
  328         }
  329         return (0);
  330 }
  331 
  332 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
  333     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, &desiredvnodes, 0,
  334     sysctl_update_desiredvnodes, "I", "Target for maximum number of vnodes");
  335 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
  336     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
  337 static int vnlru_nowhere;
  338 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
  339     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
  340 
  341 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
  342 static int vnsz2log;
  343 
  344 /*
  345  * Support for the bufobj clean & dirty pctrie.
  346  */
  347 static void *
  348 buf_trie_alloc(struct pctrie *ptree)
  349 {
  350 
  351         return uma_zalloc(buf_trie_zone, M_NOWAIT);
  352 }
  353 
  354 static void
  355 buf_trie_free(struct pctrie *ptree, void *node)
  356 {
  357 
  358         uma_zfree(buf_trie_zone, node);
  359 }
  360 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
  361 
  362 /*
  363  * Initialize the vnode management data structures.
  364  *
  365  * Reevaluate the following cap on the number of vnodes after the physical
  366  * memory size exceeds 512GB.  In the limit, as the physical memory size
  367  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
  368  */
  369 #ifndef MAXVNODES_MAX
  370 #define MAXVNODES_MAX   (512 * 1024 * 1024 / 64)        /* 8M */
  371 #endif
  372 
  373 /*
  374  * Initialize a vnode as it first enters the zone.
  375  */
  376 static int
  377 vnode_init(void *mem, int size, int flags)
  378 {
  379         struct vnode *vp;
  380 
  381         vp = mem;
  382         bzero(vp, size);
  383         /*
  384          * Setup locks.
  385          */
  386         vp->v_vnlock = &vp->v_lock;
  387         mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
  388         /*
  389          * By default, don't allow shared locks unless filesystems opt-in.
  390          */
  391         lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
  392             LK_NOSHARE | LK_IS_VNODE);
  393         /*
  394          * Initialize bufobj.
  395          */
  396         bufobj_init(&vp->v_bufobj, vp);
  397         /*
  398          * Initialize namecache.
  399          */
  400         LIST_INIT(&vp->v_cache_src);
  401         TAILQ_INIT(&vp->v_cache_dst);
  402         /*
  403          * Initialize rangelocks.
  404          */
  405         rangelock_init(&vp->v_rl);
  406         return (0);
  407 }
  408 
  409 /*
  410  * Free a vnode when it is cleared from the zone.
  411  */
  412 static void
  413 vnode_fini(void *mem, int size)
  414 {
  415         struct vnode *vp;
  416         struct bufobj *bo;
  417 
  418         vp = mem;
  419         rangelock_destroy(&vp->v_rl);
  420         lockdestroy(vp->v_vnlock);
  421         mtx_destroy(&vp->v_interlock);
  422         bo = &vp->v_bufobj;
  423         rw_destroy(BO_LOCKPTR(bo));
  424 }
  425 
  426 /*
  427  * Provide the size of NFS nclnode and NFS fh for calculation of the
  428  * vnode memory consumption.  The size is specified directly to
  429  * eliminate dependency on NFS-private header.
  430  *
  431  * Other filesystems may use bigger or smaller (like UFS and ZFS)
  432  * private inode data, but the NFS-based estimation is ample enough.
  433  * Still, we care about differences in the size between 64- and 32-bit
  434  * platforms.
  435  *
  436  * Namecache structure size is heuristically
  437  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
  438  */
  439 #ifdef _LP64
  440 #define NFS_NCLNODE_SZ  (528 + 64)
  441 #define NC_SZ           148
  442 #else
  443 #define NFS_NCLNODE_SZ  (360 + 32)
  444 #define NC_SZ           92
  445 #endif
  446 
  447 static void
  448 vntblinit(void *dummy __unused)
  449 {
  450         u_int i;
  451         int physvnodes, virtvnodes;
  452 
  453         /*
  454          * Desiredvnodes is a function of the physical memory size and the
  455          * kernel's heap size.  Generally speaking, it scales with the
  456          * physical memory size.  The ratio of desiredvnodes to the physical
  457          * memory size is 1:16 until desiredvnodes exceeds 98,304.
  458          * Thereafter, the
  459          * marginal ratio of desiredvnodes to the physical memory size is
  460          * 1:64.  However, desiredvnodes is limited by the kernel's heap
  461          * size.  The memory required by desiredvnodes vnodes and vm objects
  462          * must not exceed 1/10th of the kernel's heap size.
  463          */
  464         physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
  465             3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
  466         virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
  467             sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
  468         desiredvnodes = min(physvnodes, virtvnodes);
  469         if (desiredvnodes > MAXVNODES_MAX) {
  470                 if (bootverbose)
  471                         printf("Reducing kern.maxvnodes %d -> %d\n",
  472                             desiredvnodes, MAXVNODES_MAX);
  473                 desiredvnodes = MAXVNODES_MAX;
  474         }
  475         wantfreevnodes = desiredvnodes / 4;
  476         mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
  477         TAILQ_INIT(&vnode_free_list);
  478         mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
  479         vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
  480             vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
  481         vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
  482             NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
  483         /*
  484          * Preallocate enough nodes to support one-per buf so that
  485          * we can not fail an insert.  reassignbuf() callers can not
  486          * tolerate the insertion failure.
  487          */
  488         buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
  489             NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR, 
  490             UMA_ZONE_NOFREE | UMA_ZONE_VM);
  491         uma_prealloc(buf_trie_zone, nbuf);
  492 
  493         vnodes_created = counter_u64_alloc(M_WAITOK);
  494         recycles_count = counter_u64_alloc(M_WAITOK);
  495         free_owe_inact = counter_u64_alloc(M_WAITOK);
  496 
  497         /*
  498          * Initialize the filesystem syncer.
  499          */
  500         syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
  501             &syncer_mask);
  502         syncer_maxdelay = syncer_mask + 1;
  503         mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
  504         cv_init(&sync_wakeup, "syncer");
  505         for (i = 1; i <= sizeof(struct vnode); i <<= 1)
  506                 vnsz2log++;
  507         vnsz2log--;
  508 }
  509 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
  510 
  511 
  512 /*
  513  * Mark a mount point as busy. Used to synchronize access and to delay
  514  * unmounting. Eventually, mountlist_mtx is not released on failure.
  515  *
  516  * vfs_busy() is a custom lock, it can block the caller.
  517  * vfs_busy() only sleeps if the unmount is active on the mount point.
  518  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
  519  * vnode belonging to mp.
  520  *
  521  * Lookup uses vfs_busy() to traverse mount points.
  522  * root fs                      var fs
  523  * / vnode lock         A       / vnode lock (/var)             D
  524  * /var vnode lock      B       /log vnode lock(/var/log)       E
  525  * vfs_busy lock        C       vfs_busy lock                   F
  526  *
  527  * Within each file system, the lock order is C->A->B and F->D->E.
  528  *
  529  * When traversing across mounts, the system follows that lock order:
  530  *
  531  *        C->A->B
  532  *              |
  533  *              +->F->D->E
  534  *
  535  * The lookup() process for namei("/var") illustrates the process:
  536  *  VOP_LOOKUP() obtains B while A is held
  537  *  vfs_busy() obtains a shared lock on F while A and B are held
  538  *  vput() releases lock on B
  539  *  vput() releases lock on A
  540  *  VFS_ROOT() obtains lock on D while shared lock on F is held
  541  *  vfs_unbusy() releases shared lock on F
  542  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
  543  *    Attempt to lock A (instead of vp_crossmp) while D is held would
  544  *    violate the global order, causing deadlocks.
  545  *
  546  * dounmount() locks B while F is drained.
  547  */
  548 int
  549 vfs_busy(struct mount *mp, int flags)
  550 {
  551 
  552         MPASS((flags & ~MBF_MASK) == 0);
  553         CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
  554 
  555         MNT_ILOCK(mp);
  556         MNT_REF(mp);
  557         /*
  558          * If mount point is currently being unmounted, sleep until the
  559          * mount point fate is decided.  If thread doing the unmounting fails,
  560          * it will clear MNTK_UNMOUNT flag before waking us up, indicating
  561          * that this mount point has survived the unmount attempt and vfs_busy
  562          * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
  563          * flag in addition to MNTK_UNMOUNT, indicating that mount point is
  564          * about to be really destroyed.  vfs_busy needs to release its
  565          * reference on the mount point in this case and return with ENOENT,
  566          * telling the caller that mount mount it tried to busy is no longer
  567          * valid.
  568          */
  569         while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
  570                 if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
  571                         MNT_REL(mp);
  572                         MNT_IUNLOCK(mp);
  573                         CTR1(KTR_VFS, "%s: failed busying before sleeping",
  574                             __func__);
  575                         return (ENOENT);
  576                 }
  577                 if (flags & MBF_MNTLSTLOCK)
  578                         mtx_unlock(&mountlist_mtx);
  579                 mp->mnt_kern_flag |= MNTK_MWAIT;
  580                 msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
  581                 if (flags & MBF_MNTLSTLOCK)
  582                         mtx_lock(&mountlist_mtx);
  583                 MNT_ILOCK(mp);
  584         }
  585         if (flags & MBF_MNTLSTLOCK)
  586                 mtx_unlock(&mountlist_mtx);
  587         mp->mnt_lockref++;
  588         MNT_IUNLOCK(mp);
  589         return (0);
  590 }
  591 
  592 /*
  593  * Free a busy filesystem.
  594  */
  595 void
  596 vfs_unbusy(struct mount *mp)
  597 {
  598 
  599         CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
  600         MNT_ILOCK(mp);
  601         MNT_REL(mp);
  602         KASSERT(mp->mnt_lockref > 0, ("negative mnt_lockref"));
  603         mp->mnt_lockref--;
  604         if (mp->mnt_lockref == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
  605                 MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
  606                 CTR1(KTR_VFS, "%s: waking up waiters", __func__);
  607                 mp->mnt_kern_flag &= ~MNTK_DRAINING;
  608                 wakeup(&mp->mnt_lockref);
  609         }
  610         MNT_IUNLOCK(mp);
  611 }
  612 
  613 /*
  614  * Lookup a mount point by filesystem identifier.
  615  */
  616 struct mount *
  617 vfs_getvfs(fsid_t *fsid)
  618 {
  619         struct mount *mp;
  620 
  621         CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
  622         mtx_lock(&mountlist_mtx);
  623         TAILQ_FOREACH(mp, &mountlist, mnt_list) {
  624                 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
  625                     mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
  626                         vfs_ref(mp);
  627                         mtx_unlock(&mountlist_mtx);
  628                         return (mp);
  629                 }
  630         }
  631         mtx_unlock(&mountlist_mtx);
  632         CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
  633         return ((struct mount *) 0);
  634 }
  635 
  636 /*
  637  * Lookup a mount point by filesystem identifier, busying it before
  638  * returning.
  639  *
  640  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
  641  * cache for popular filesystem identifiers.  The cache is lockess, using
  642  * the fact that struct mount's are never freed.  In worst case we may
  643  * get pointer to unmounted or even different filesystem, so we have to
  644  * check what we got, and go slow way if so.
  645  */
  646 struct mount *
  647 vfs_busyfs(fsid_t *fsid)
  648 {
  649 #define FSID_CACHE_SIZE 256
  650         typedef struct mount * volatile vmp_t;
  651         static vmp_t cache[FSID_CACHE_SIZE];
  652         struct mount *mp;
  653         int error;
  654         uint32_t hash;
  655 
  656         CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
  657         hash = fsid->val[0] ^ fsid->val[1];
  658         hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
  659         mp = cache[hash];
  660         if (mp == NULL ||
  661             mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
  662             mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
  663                 goto slow;
  664         if (vfs_busy(mp, 0) != 0) {
  665                 cache[hash] = NULL;
  666                 goto slow;
  667         }
  668         if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
  669             mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
  670                 return (mp);
  671         else
  672             vfs_unbusy(mp);
  673 
  674 slow:
  675         mtx_lock(&mountlist_mtx);
  676         TAILQ_FOREACH(mp, &mountlist, mnt_list) {
  677                 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
  678                     mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
  679                         error = vfs_busy(mp, MBF_MNTLSTLOCK);
  680                         if (error) {
  681                                 cache[hash] = NULL;
  682                                 mtx_unlock(&mountlist_mtx);
  683                                 return (NULL);
  684                         }
  685                         cache[hash] = mp;
  686                         return (mp);
  687                 }
  688         }
  689         CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
  690         mtx_unlock(&mountlist_mtx);
  691         return ((struct mount *) 0);
  692 }
  693 
  694 /*
  695  * Check if a user can access privileged mount options.
  696  */
  697 int
  698 vfs_suser(struct mount *mp, struct thread *td)
  699 {
  700         int error;
  701 
  702         if (jailed(td->td_ucred)) {
  703                 /*
  704                  * If the jail of the calling thread lacks permission for
  705                  * this type of file system, deny immediately.
  706                  */
  707                 if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
  708                         return (EPERM);
  709 
  710                 /*
  711                  * If the file system was mounted outside the jail of the
  712                  * calling thread, deny immediately.
  713                  */
  714                 if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
  715                         return (EPERM);
  716         }
  717 
  718         /*
  719          * If file system supports delegated administration, we don't check
  720          * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
  721          * by the file system itself.
  722          * If this is not the user that did original mount, we check for
  723          * the PRIV_VFS_MOUNT_OWNER privilege.
  724          */
  725         if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
  726             mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
  727                 if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
  728                         return (error);
  729         }
  730         return (0);
  731 }
  732 
  733 /*
  734  * Get a new unique fsid.  Try to make its val[0] unique, since this value
  735  * will be used to create fake device numbers for stat().  Also try (but
  736  * not so hard) make its val[0] unique mod 2^16, since some emulators only
  737  * support 16-bit device numbers.  We end up with unique val[0]'s for the
  738  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
  739  *
  740  * Keep in mind that several mounts may be running in parallel.  Starting
  741  * the search one past where the previous search terminated is both a
  742  * micro-optimization and a defense against returning the same fsid to
  743  * different mounts.
  744  */
  745 void
  746 vfs_getnewfsid(struct mount *mp)
  747 {
  748         static uint16_t mntid_base;
  749         struct mount *nmp;
  750         fsid_t tfsid;
  751         int mtype;
  752 
  753         CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
  754         mtx_lock(&mntid_mtx);
  755         mtype = mp->mnt_vfc->vfc_typenum;
  756         tfsid.val[1] = mtype;
  757         mtype = (mtype & 0xFF) << 24;
  758         for (;;) {
  759                 tfsid.val[0] = makedev(255,
  760                     mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
  761                 mntid_base++;
  762                 if ((nmp = vfs_getvfs(&tfsid)) == NULL)
  763                         break;
  764                 vfs_rel(nmp);
  765         }
  766         mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
  767         mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
  768         mtx_unlock(&mntid_mtx);
  769 }
  770 
  771 /*
  772  * Knob to control the precision of file timestamps:
  773  *
  774  *   0 = seconds only; nanoseconds zeroed.
  775  *   1 = seconds and nanoseconds, accurate within 1/HZ.
  776  *   2 = seconds and nanoseconds, truncated to microseconds.
  777  * >=3 = seconds and nanoseconds, maximum precision.
  778  */
  779 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
  780 
  781 static int timestamp_precision = TSP_USEC;
  782 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
  783     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
  784     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
  785     "3+: sec + ns (max. precision))");
  786 
  787 /*
  788  * Get a current timestamp.
  789  */
  790 void
  791 vfs_timestamp(struct timespec *tsp)
  792 {
  793         struct timeval tv;
  794 
  795         switch (timestamp_precision) {
  796         case TSP_SEC:
  797                 tsp->tv_sec = time_second;
  798                 tsp->tv_nsec = 0;
  799                 break;
  800         case TSP_HZ:
  801                 getnanotime(tsp);
  802                 break;
  803         case TSP_USEC:
  804                 microtime(&tv);
  805                 TIMEVAL_TO_TIMESPEC(&tv, tsp);
  806                 break;
  807         case TSP_NSEC:
  808         default:
  809                 nanotime(tsp);
  810                 break;
  811         }
  812 }
  813 
  814 /*
  815  * Set vnode attributes to VNOVAL
  816  */
  817 void
  818 vattr_null(struct vattr *vap)
  819 {
  820 
  821         vap->va_type = VNON;
  822         vap->va_size = VNOVAL;
  823         vap->va_bytes = VNOVAL;
  824         vap->va_mode = VNOVAL;
  825         vap->va_nlink = VNOVAL;
  826         vap->va_uid = VNOVAL;
  827         vap->va_gid = VNOVAL;
  828         vap->va_fsid = VNOVAL;
  829         vap->va_fileid = VNOVAL;
  830         vap->va_blocksize = VNOVAL;
  831         vap->va_rdev = VNOVAL;
  832         vap->va_atime.tv_sec = VNOVAL;
  833         vap->va_atime.tv_nsec = VNOVAL;
  834         vap->va_mtime.tv_sec = VNOVAL;
  835         vap->va_mtime.tv_nsec = VNOVAL;
  836         vap->va_ctime.tv_sec = VNOVAL;
  837         vap->va_ctime.tv_nsec = VNOVAL;
  838         vap->va_birthtime.tv_sec = VNOVAL;
  839         vap->va_birthtime.tv_nsec = VNOVAL;
  840         vap->va_flags = VNOVAL;
  841         vap->va_gen = VNOVAL;
  842         vap->va_vaflags = 0;
  843 }
  844 
  845 /*
  846  * This routine is called when we have too many vnodes.  It attempts
  847  * to free <count> vnodes and will potentially free vnodes that still
  848  * have VM backing store (VM backing store is typically the cause
  849  * of a vnode blowout so we want to do this).  Therefore, this operation
  850  * is not considered cheap.
  851  *
  852  * A number of conditions may prevent a vnode from being reclaimed.
  853  * the buffer cache may have references on the vnode, a directory
  854  * vnode may still have references due to the namei cache representing
  855  * underlying files, or the vnode may be in active use.   It is not
  856  * desirable to reuse such vnodes.  These conditions may cause the
  857  * number of vnodes to reach some minimum value regardless of what
  858  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
  859  */
  860 static int
  861 vlrureclaim(struct mount *mp, int reclaim_nc_src, int trigger)
  862 {
  863         struct vnode *vp;
  864         int count, done, target;
  865 
  866         done = 0;
  867         vn_start_write(NULL, &mp, V_WAIT);
  868         MNT_ILOCK(mp);
  869         count = mp->mnt_nvnodelistsize;
  870         target = count * (int64_t)gapvnodes / imax(desiredvnodes, 1);
  871         target = target / 10 + 1;
  872         while (count != 0 && done < target) {
  873                 vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
  874                 while (vp != NULL && vp->v_type == VMARKER)
  875                         vp = TAILQ_NEXT(vp, v_nmntvnodes);
  876                 if (vp == NULL)
  877                         break;
  878                 /*
  879                  * XXX LRU is completely broken for non-free vnodes.  First
  880                  * by calling here in mountpoint order, then by moving
  881                  * unselected vnodes to the end here, and most grossly by
  882                  * removing the vlruvp() function that was supposed to
  883                  * maintain the order.  (This function was born broken
  884                  * since syncer problems prevented it doing anything.)  The
  885                  * order is closer to LRC (C = Created).
  886                  *
  887                  * LRU reclaiming of vnodes seems to have last worked in
  888                  * FreeBSD-3 where LRU wasn't mentioned under any spelling.
  889                  * Then there was no hold count, and inactive vnodes were
  890                  * simply put on the free list in LRU order.  The separate
  891                  * lists also break LRU.  We prefer to reclaim from the
  892                  * free list for technical reasons.  This tends to thrash
  893                  * the free list to keep very unrecently used held vnodes.
  894                  * The problem is mitigated by keeping the free list large.
  895                  */
  896                 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
  897                 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
  898                 --count;
  899                 if (!VI_TRYLOCK(vp))
  900                         goto next_iter;
  901                 /*
  902                  * If it's been deconstructed already, it's still
  903                  * referenced, or it exceeds the trigger, skip it.
  904                  * Also skip free vnodes.  We are trying to make space
  905                  * to expand the free list, not reduce it.
  906                  */
  907                 if (vp->v_usecount ||
  908                     (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
  909                     ((vp->v_iflag & VI_FREE) != 0) ||
  910                     (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
  911                     vp->v_object->resident_page_count > trigger)) {
  912                         VI_UNLOCK(vp);
  913                         goto next_iter;
  914                 }
  915                 MNT_IUNLOCK(mp);
  916                 vholdl(vp);
  917                 if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) {
  918                         vdrop(vp);
  919                         goto next_iter_mntunlocked;
  920                 }
  921                 VI_LOCK(vp);
  922                 /*
  923                  * v_usecount may have been bumped after VOP_LOCK() dropped
  924                  * the vnode interlock and before it was locked again.
  925                  *
  926                  * It is not necessary to recheck VI_DOOMED because it can
  927                  * only be set by another thread that holds both the vnode
  928                  * lock and vnode interlock.  If another thread has the
  929                  * vnode lock before we get to VOP_LOCK() and obtains the
  930                  * vnode interlock after VOP_LOCK() drops the vnode
  931                  * interlock, the other thread will be unable to drop the
  932                  * vnode lock before our VOP_LOCK() call fails.
  933                  */
  934                 if (vp->v_usecount ||
  935                     (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
  936                     (vp->v_iflag & VI_FREE) != 0 ||
  937                     (vp->v_object != NULL &&
  938                     vp->v_object->resident_page_count > trigger)) {
  939                         VOP_UNLOCK(vp, LK_INTERLOCK);
  940                         vdrop(vp);
  941                         goto next_iter_mntunlocked;
  942                 }
  943                 KASSERT((vp->v_iflag & VI_DOOMED) == 0,
  944                     ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
  945                 counter_u64_add(recycles_count, 1);
  946                 vgonel(vp);
  947                 VOP_UNLOCK(vp, 0);
  948                 vdropl(vp);
  949                 done++;
  950 next_iter_mntunlocked:
  951                 if (!should_yield())
  952                         goto relock_mnt;
  953                 goto yield;
  954 next_iter:
  955                 if (!should_yield())
  956                         continue;
  957                 MNT_IUNLOCK(mp);
  958 yield:
  959                 kern_yield(PRI_USER);
  960 relock_mnt:
  961                 MNT_ILOCK(mp);
  962         }
  963         MNT_IUNLOCK(mp);
  964         vn_finished_write(mp);
  965         return done;
  966 }
  967 
  968 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
  969 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
  970     0,
  971     "limit on vnode free requests per call to the vnlru_free routine");
  972 
  973 /*
  974  * Attempt to reduce the free list by the requested amount.
  975  */
  976 static void
  977 vnlru_free_locked(int count, struct vfsops *mnt_op)
  978 {
  979         struct vnode *vp;
  980         struct mount *mp;
  981         bool tried_batches;
  982 
  983         tried_batches = false;
  984         mtx_assert(&vnode_free_list_mtx, MA_OWNED);
  985         if (count > max_vnlru_free)
  986                 count = max_vnlru_free;
  987         for (; count > 0; count--) {
  988                 vp = TAILQ_FIRST(&vnode_free_list);
  989                 /*
  990                  * The list can be modified while the free_list_mtx
  991                  * has been dropped and vp could be NULL here.
  992                  */
  993                 if (vp == NULL) {
  994                         if (tried_batches)
  995                                 break;
  996                         mtx_unlock(&vnode_free_list_mtx);
  997                         vnlru_return_batches(mnt_op);
  998                         tried_batches = true;
  999                         mtx_lock(&vnode_free_list_mtx);
 1000                         continue;
 1001                 }
 1002 
 1003                 VNASSERT(vp->v_op != NULL, vp,
 1004                     ("vnlru_free: vnode already reclaimed."));
 1005                 KASSERT((vp->v_iflag & VI_FREE) != 0,
 1006                     ("Removing vnode not on freelist"));
 1007                 KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
 1008                     ("Mangling active vnode"));
 1009                 TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
 1010 
 1011                 /*
 1012                  * Don't recycle if our vnode is from different type
 1013                  * of mount point.  Note that mp is type-safe, the
 1014                  * check does not reach unmapped address even if
 1015                  * vnode is reclaimed.
 1016                  * Don't recycle if we can't get the interlock without
 1017                  * blocking.
 1018                  */
 1019                 if ((mnt_op != NULL && (mp = vp->v_mount) != NULL &&
 1020                     mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) {
 1021                         TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist);
 1022                         continue;
 1023                 }
 1024                 VNASSERT((vp->v_iflag & VI_FREE) != 0 && vp->v_holdcnt == 0,
 1025                     vp, ("vp inconsistent on freelist"));
 1026 
 1027                 /*
 1028                  * The clear of VI_FREE prevents activation of the
 1029                  * vnode.  There is no sense in putting the vnode on
 1030                  * the mount point active list, only to remove it
 1031                  * later during recycling.  Inline the relevant part
 1032                  * of vholdl(), to avoid triggering assertions or
 1033                  * activating.
 1034                  */
 1035                 freevnodes--;
 1036                 vp->v_iflag &= ~VI_FREE;
 1037                 VNODE_REFCOUNT_FENCE_REL();
 1038                 refcount_acquire(&vp->v_holdcnt);
 1039 
 1040                 mtx_unlock(&vnode_free_list_mtx);
 1041                 VI_UNLOCK(vp);
 1042                 vtryrecycle(vp);
 1043                 /*
 1044                  * If the recycled succeeded this vdrop will actually free
 1045                  * the vnode.  If not it will simply place it back on
 1046                  * the free list.
 1047                  */
 1048                 vdrop(vp);
 1049                 mtx_lock(&vnode_free_list_mtx);
 1050         }
 1051 }
 1052 
 1053 void
 1054 vnlru_free(int count, struct vfsops *mnt_op)
 1055 {
 1056 
 1057         mtx_lock(&vnode_free_list_mtx);
 1058         vnlru_free_locked(count, mnt_op);
 1059         mtx_unlock(&vnode_free_list_mtx);
 1060 }
 1061 
 1062 
 1063 /* XXX some names and initialization are bad for limits and watermarks. */
 1064 static int
 1065 vspace(void)
 1066 {
 1067         int space;
 1068 
 1069         gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
 1070         vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
 1071         vlowat = vhiwat / 2;
 1072         if (numvnodes > desiredvnodes)
 1073                 return (0);
 1074         space = desiredvnodes - numvnodes;
 1075         if (freevnodes > wantfreevnodes)
 1076                 space += freevnodes - wantfreevnodes;
 1077         return (space);
 1078 }
 1079 
 1080 static void
 1081 vnlru_return_batch_locked(struct mount *mp)
 1082 {
 1083         struct vnode *vp;
 1084 
 1085         mtx_assert(&mp->mnt_listmtx, MA_OWNED);
 1086 
 1087         if (mp->mnt_tmpfreevnodelistsize == 0)
 1088                 return;
 1089 
 1090         TAILQ_FOREACH(vp, &mp->mnt_tmpfreevnodelist, v_actfreelist) {
 1091                 VNASSERT((vp->v_mflag & VMP_TMPMNTFREELIST) != 0, vp,
 1092                     ("vnode without VMP_TMPMNTFREELIST on mnt_tmpfreevnodelist"));
 1093                 vp->v_mflag &= ~VMP_TMPMNTFREELIST;
 1094         }
 1095         mtx_lock(&vnode_free_list_mtx);
 1096         TAILQ_CONCAT(&vnode_free_list, &mp->mnt_tmpfreevnodelist, v_actfreelist);
 1097         freevnodes += mp->mnt_tmpfreevnodelistsize;
 1098         mtx_unlock(&vnode_free_list_mtx);
 1099         mp->mnt_tmpfreevnodelistsize = 0;
 1100 }
 1101 
 1102 static void
 1103 vnlru_return_batch(struct mount *mp)
 1104 {
 1105 
 1106         mtx_lock(&mp->mnt_listmtx);
 1107         vnlru_return_batch_locked(mp);
 1108         mtx_unlock(&mp->mnt_listmtx);
 1109 }
 1110 
 1111 static void
 1112 vnlru_return_batches(struct vfsops *mnt_op)
 1113 {
 1114         struct mount *mp, *nmp;
 1115         bool need_unbusy;
 1116 
 1117         mtx_lock(&mountlist_mtx);
 1118         for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
 1119                 need_unbusy = false;
 1120                 if (mnt_op != NULL && mp->mnt_op != mnt_op)
 1121                         goto next;
 1122                 if (mp->mnt_tmpfreevnodelistsize == 0)
 1123                         goto next;
 1124                 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK) == 0) {
 1125                         vnlru_return_batch(mp);
 1126                         need_unbusy = true;
 1127                         mtx_lock(&mountlist_mtx);
 1128                 }
 1129 next:
 1130                 nmp = TAILQ_NEXT(mp, mnt_list);
 1131                 if (need_unbusy)
 1132                         vfs_unbusy(mp);
 1133         }
 1134         mtx_unlock(&mountlist_mtx);
 1135 }
 1136 
 1137 /*
 1138  * Attempt to recycle vnodes in a context that is always safe to block.
 1139  * Calling vlrurecycle() from the bowels of filesystem code has some
 1140  * interesting deadlock problems.
 1141  */
 1142 static struct proc *vnlruproc;
 1143 static int vnlruproc_sig;
 1144 
 1145 static void
 1146 vnlru_proc(void)
 1147 {
 1148         struct mount *mp, *nmp;
 1149         unsigned long onumvnodes;
 1150         int done, force, reclaim_nc_src, trigger, usevnodes;
 1151 
 1152         EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
 1153             SHUTDOWN_PRI_FIRST);
 1154 
 1155         force = 0;
 1156         for (;;) {
 1157                 kproc_suspend_check(vnlruproc);
 1158                 mtx_lock(&vnode_free_list_mtx);
 1159                 /*
 1160                  * If numvnodes is too large (due to desiredvnodes being
 1161                  * adjusted using its sysctl, or emergency growth), first
 1162                  * try to reduce it by discarding from the free list.
 1163                  */
 1164                 if (numvnodes > desiredvnodes)
 1165                         vnlru_free_locked(numvnodes - desiredvnodes, NULL);
 1166                 /*
 1167                  * Sleep if the vnode cache is in a good state.  This is
 1168                  * when it is not over-full and has space for about a 4%
 1169                  * or 9% expansion (by growing its size or inexcessively
 1170                  * reducing its free list).  Otherwise, try to reclaim
 1171                  * space for a 10% expansion.
 1172                  */
 1173                 if (vstir && force == 0) {
 1174                         force = 1;
 1175                         vstir = 0;
 1176                 }
 1177                 if (vspace() >= vlowat && force == 0) {
 1178                         vnlruproc_sig = 0;
 1179                         wakeup(&vnlruproc_sig);
 1180                         msleep(vnlruproc, &vnode_free_list_mtx,
 1181                             PVFS|PDROP, "vlruwt", hz);
 1182                         continue;
 1183                 }
 1184                 mtx_unlock(&vnode_free_list_mtx);
 1185                 done = 0;
 1186                 onumvnodes = numvnodes;
 1187                 /*
 1188                  * Calculate parameters for recycling.  These are the same
 1189                  * throughout the loop to give some semblance of fairness.
 1190                  * The trigger point is to avoid recycling vnodes with lots
 1191                  * of resident pages.  We aren't trying to free memory; we
 1192                  * are trying to recycle or at least free vnodes.
 1193                  */
 1194                 if (numvnodes <= desiredvnodes)
 1195                         usevnodes = numvnodes - freevnodes;
 1196                 else
 1197                         usevnodes = numvnodes;
 1198                 if (usevnodes <= 0)
 1199                         usevnodes = 1;
 1200                 /*
 1201                  * The trigger value is is chosen to give a conservatively
 1202                  * large value to ensure that it alone doesn't prevent
 1203                  * making progress.  The value can easily be so large that
 1204                  * it is effectively infinite in some congested and
 1205                  * misconfigured cases, and this is necessary.  Normally
 1206                  * it is about 8 to 100 (pages), which is quite large.
 1207                  */
 1208                 trigger = vm_cnt.v_page_count * 2 / usevnodes;
 1209                 if (force < 2)
 1210                         trigger = vsmalltrigger;
 1211                 reclaim_nc_src = force >= 3;
 1212                 mtx_lock(&mountlist_mtx);
 1213                 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
 1214                         if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) {
 1215                                 nmp = TAILQ_NEXT(mp, mnt_list);
 1216                                 continue;
 1217                         }
 1218                         done += vlrureclaim(mp, reclaim_nc_src, trigger);
 1219                         mtx_lock(&mountlist_mtx);
 1220                         nmp = TAILQ_NEXT(mp, mnt_list);
 1221                         vfs_unbusy(mp);
 1222                 }
 1223                 mtx_unlock(&mountlist_mtx);
 1224                 if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
 1225                         uma_reclaim();
 1226                 if (done == 0) {
 1227                         if (force == 0 || force == 1) {
 1228                                 force = 2;
 1229                                 continue;
 1230                         }
 1231                         if (force == 2) {
 1232                                 force = 3;
 1233                                 continue;
 1234                         }
 1235                         force = 0;
 1236                         vnlru_nowhere++;
 1237                         tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
 1238                 } else
 1239                         kern_yield(PRI_USER);
 1240                 /*
 1241                  * After becoming active to expand above low water, keep
 1242                  * active until above high water.
 1243                  */
 1244                 force = vspace() < vhiwat;
 1245         }
 1246 }
 1247 
 1248 static struct kproc_desc vnlru_kp = {
 1249         "vnlru",
 1250         vnlru_proc,
 1251         &vnlruproc
 1252 };
 1253 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
 1254     &vnlru_kp);
 1255  
 1256 /*
 1257  * Routines having to do with the management of the vnode table.
 1258  */
 1259 
 1260 /*
 1261  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
 1262  * before we actually vgone().  This function must be called with the vnode
 1263  * held to prevent the vnode from being returned to the free list midway
 1264  * through vgone().
 1265  */
 1266 static int
 1267 vtryrecycle(struct vnode *vp)
 1268 {
 1269         struct mount *vnmp;
 1270 
 1271         CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
 1272         VNASSERT(vp->v_holdcnt, vp,
 1273             ("vtryrecycle: Recycling vp %p without a reference.", vp));
 1274         /*
 1275          * This vnode may found and locked via some other list, if so we
 1276          * can't recycle it yet.
 1277          */
 1278         if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
 1279                 CTR2(KTR_VFS,
 1280                     "%s: impossible to recycle, vp %p lock is already held",
 1281                     __func__, vp);
 1282                 return (EWOULDBLOCK);
 1283         }
 1284         /*
 1285          * Don't recycle if its filesystem is being suspended.
 1286          */
 1287         if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
 1288                 VOP_UNLOCK(vp, 0);
 1289                 CTR2(KTR_VFS,
 1290                     "%s: impossible to recycle, cannot start the write for %p",
 1291                     __func__, vp);
 1292                 return (EBUSY);
 1293         }
 1294         /*
 1295          * If we got this far, we need to acquire the interlock and see if
 1296          * anyone picked up this vnode from another list.  If not, we will
 1297          * mark it with DOOMED via vgonel() so that anyone who does find it
 1298          * will skip over it.
 1299          */
 1300         VI_LOCK(vp);
 1301         if (vp->v_usecount) {
 1302                 VOP_UNLOCK(vp, LK_INTERLOCK);
 1303                 vn_finished_write(vnmp);
 1304                 CTR2(KTR_VFS,
 1305                     "%s: impossible to recycle, %p is already referenced",
 1306                     __func__, vp);
 1307                 return (EBUSY);
 1308         }
 1309         if ((vp->v_iflag & VI_DOOMED) == 0) {
 1310                 counter_u64_add(recycles_count, 1);
 1311                 vgonel(vp);
 1312         }
 1313         VOP_UNLOCK(vp, LK_INTERLOCK);
 1314         vn_finished_write(vnmp);
 1315         return (0);
 1316 }
 1317 
 1318 static void
 1319 vcheckspace(void)
 1320 {
 1321 
 1322         if (vspace() < vlowat && vnlruproc_sig == 0) {
 1323                 vnlruproc_sig = 1;
 1324                 wakeup(vnlruproc);
 1325         }
 1326 }
 1327 
 1328 /*
 1329  * Wait if necessary for space for a new vnode.
 1330  */
 1331 static int
 1332 getnewvnode_wait(int suspended)
 1333 {
 1334 
 1335         mtx_assert(&vnode_free_list_mtx, MA_OWNED);
 1336         if (numvnodes >= desiredvnodes) {
 1337                 if (suspended) {
 1338                         /*
 1339                          * The file system is being suspended.  We cannot
 1340                          * risk a deadlock here, so allow allocation of
 1341                          * another vnode even if this would give too many.
 1342                          */
 1343                         return (0);
 1344                 }
 1345                 if (vnlruproc_sig == 0) {
 1346                         vnlruproc_sig = 1;      /* avoid unnecessary wakeups */
 1347                         wakeup(vnlruproc);
 1348                 }
 1349                 msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
 1350                     "vlruwk", hz);
 1351         }
 1352         /* Post-adjust like the pre-adjust in getnewvnode(). */
 1353         if (numvnodes + 1 > desiredvnodes && freevnodes > 1)
 1354                 vnlru_free_locked(1, NULL);
 1355         return (numvnodes >= desiredvnodes ? ENFILE : 0);
 1356 }
 1357 
 1358 /*
 1359  * This hack is fragile, and probably not needed any more now that the
 1360  * watermark handling works.
 1361  */
 1362 void
 1363 getnewvnode_reserve(u_int count)
 1364 {
 1365         struct thread *td;
 1366 
 1367         /* Pre-adjust like the pre-adjust in getnewvnode(), with any count. */
 1368         /* XXX no longer so quick, but this part is not racy. */
 1369         mtx_lock(&vnode_free_list_mtx);
 1370         if (numvnodes + count > desiredvnodes && freevnodes > wantfreevnodes)
 1371                 vnlru_free_locked(ulmin(numvnodes + count - desiredvnodes,
 1372                     freevnodes - wantfreevnodes), NULL);
 1373         mtx_unlock(&vnode_free_list_mtx);
 1374 
 1375         td = curthread;
 1376         /* First try to be quick and racy. */
 1377         if (atomic_fetchadd_long(&numvnodes, count) + count <= desiredvnodes) {
 1378                 td->td_vp_reserv += count;
 1379                 vcheckspace();  /* XXX no longer so quick, but more racy */
 1380                 return;
 1381         } else
 1382                 atomic_subtract_long(&numvnodes, count);
 1383 
 1384         mtx_lock(&vnode_free_list_mtx);
 1385         while (count > 0) {
 1386                 if (getnewvnode_wait(0) == 0) {
 1387                         count--;
 1388                         td->td_vp_reserv++;
 1389                         atomic_add_long(&numvnodes, 1);
 1390                 }
 1391         }
 1392         vcheckspace();
 1393         mtx_unlock(&vnode_free_list_mtx);
 1394 }
 1395 
 1396 /*
 1397  * This hack is fragile, especially if desiredvnodes or wantvnodes are
 1398  * misconfgured or changed significantly.  Reducing desiredvnodes below
 1399  * the reserved amount should cause bizarre behaviour like reducing it
 1400  * below the number of active vnodes -- the system will try to reduce
 1401  * numvnodes to match, but should fail, so the subtraction below should
 1402  * not overflow.
 1403  */
 1404 void
 1405 getnewvnode_drop_reserve(void)
 1406 {
 1407         struct thread *td;
 1408 
 1409         td = curthread;
 1410         atomic_subtract_long(&numvnodes, td->td_vp_reserv);
 1411         td->td_vp_reserv = 0;
 1412 }
 1413 
 1414 /*
 1415  * Return the next vnode from the free list.
 1416  */
 1417 int
 1418 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
 1419     struct vnode **vpp)
 1420 {
 1421         struct vnode *vp;
 1422         struct thread *td;
 1423         struct lock_object *lo;
 1424         static int cyclecount;
 1425         int error __unused;
 1426 
 1427         CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
 1428         vp = NULL;
 1429         td = curthread;
 1430         if (td->td_vp_reserv > 0) {
 1431                 td->td_vp_reserv -= 1;
 1432                 goto alloc;
 1433         }
 1434         mtx_lock(&vnode_free_list_mtx);
 1435         if (numvnodes < desiredvnodes)
 1436                 cyclecount = 0;
 1437         else if (cyclecount++ >= freevnodes) {
 1438                 cyclecount = 0;
 1439                 vstir = 1;
 1440         }
 1441         /*
 1442          * Grow the vnode cache if it will not be above its target max
 1443          * after growing.  Otherwise, if the free list is nonempty, try
 1444          * to reclaim 1 item from it before growing the cache (possibly
 1445          * above its target max if the reclamation failed or is delayed).
 1446          * Otherwise, wait for some space.  In all cases, schedule
 1447          * vnlru_proc() if we are getting short of space.  The watermarks
 1448          * should be chosen so that we never wait or even reclaim from
 1449          * the free list to below its target minimum.
 1450          */
 1451         if (numvnodes + 1 <= desiredvnodes)
 1452                 ;
 1453         else if (freevnodes > 0)
 1454                 vnlru_free_locked(1, NULL);
 1455         else {
 1456                 error = getnewvnode_wait(mp != NULL && (mp->mnt_kern_flag &
 1457                     MNTK_SUSPEND));
 1458 #if 0   /* XXX Not all VFS_VGET/ffs_vget callers check returns. */
 1459                 if (error != 0) {
 1460                         mtx_unlock(&vnode_free_list_mtx);
 1461                         return (error);
 1462                 }
 1463 #endif
 1464         }
 1465         vcheckspace();
 1466         atomic_add_long(&numvnodes, 1);
 1467         mtx_unlock(&vnode_free_list_mtx);
 1468 alloc:
 1469         counter_u64_add(vnodes_created, 1);
 1470         vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK);
 1471         /*
 1472          * Locks are given the generic name "vnode" when created.
 1473          * Follow the historic practice of using the filesystem
 1474          * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
 1475          *
 1476          * Locks live in a witness group keyed on their name. Thus,
 1477          * when a lock is renamed, it must also move from the witness
 1478          * group of its old name to the witness group of its new name.
 1479          *
 1480          * The change only needs to be made when the vnode moves
 1481          * from one filesystem type to another. We ensure that each
 1482          * filesystem use a single static name pointer for its tag so
 1483          * that we can compare pointers rather than doing a strcmp().
 1484          */
 1485         lo = &vp->v_vnlock->lock_object;
 1486         if (lo->lo_name != tag) {
 1487                 lo->lo_name = tag;
 1488                 WITNESS_DESTROY(lo);
 1489                 WITNESS_INIT(lo, tag);
 1490         }
 1491         /*
 1492          * By default, don't allow shared locks unless filesystems opt-in.
 1493          */
 1494         vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
 1495         /*
 1496          * Finalize various vnode identity bits.
 1497          */
 1498         KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
 1499         KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
 1500         KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
 1501         vp->v_type = VNON;
 1502         vp->v_tag = tag;
 1503         vp->v_op = vops;
 1504         v_init_counters(vp);
 1505         vp->v_bufobj.bo_ops = &buf_ops_bio;
 1506 #ifdef DIAGNOSTIC
 1507         if (mp == NULL && vops != &dead_vnodeops)
 1508                 printf("NULL mp in getnewvnode(9), tag %s\n", tag);
 1509 #endif
 1510 #ifdef MAC
 1511         mac_vnode_init(vp);
 1512         if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
 1513                 mac_vnode_associate_singlelabel(mp, vp);
 1514 #endif
 1515         if (mp != NULL) {
 1516                 vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
 1517                 if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
 1518                         vp->v_vflag |= VV_NOKNOTE;
 1519         }
 1520 
 1521         /*
 1522          * For the filesystems which do not use vfs_hash_insert(),
 1523          * still initialize v_hash to have vfs_hash_index() useful.
 1524          * E.g., nullfs uses vfs_hash_index() on the lower vnode for
 1525          * its own hashing.
 1526          */
 1527         vp->v_hash = (uintptr_t)vp >> vnsz2log;
 1528 
 1529         *vpp = vp;
 1530         return (0);
 1531 }
 1532 
 1533 /*
 1534  * Delete from old mount point vnode list, if on one.
 1535  */
 1536 static void
 1537 delmntque(struct vnode *vp)
 1538 {
 1539         struct mount *mp;
 1540         int active;
 1541 
 1542         mp = vp->v_mount;
 1543         if (mp == NULL)
 1544                 return;
 1545         MNT_ILOCK(mp);
 1546         VI_LOCK(vp);
 1547         KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize,
 1548             ("Active vnode list size %d > Vnode list size %d",
 1549              mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize));
 1550         active = vp->v_iflag & VI_ACTIVE;
 1551         vp->v_iflag &= ~VI_ACTIVE;
 1552         if (active) {
 1553                 mtx_lock(&mp->mnt_listmtx);
 1554                 TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist);
 1555                 mp->mnt_activevnodelistsize--;
 1556                 mtx_unlock(&mp->mnt_listmtx);
 1557         }
 1558         vp->v_mount = NULL;
 1559         VI_UNLOCK(vp);
 1560         VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
 1561                 ("bad mount point vnode list size"));
 1562         TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
 1563         mp->mnt_nvnodelistsize--;
 1564         MNT_REL(mp);
 1565         MNT_IUNLOCK(mp);
 1566 }
 1567 
 1568 static void
 1569 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
 1570 {
 1571 
 1572         vp->v_data = NULL;
 1573         vp->v_op = &dead_vnodeops;
 1574         vgone(vp);
 1575         vput(vp);
 1576 }
 1577 
 1578 /*
 1579  * Insert into list of vnodes for the new mount point, if available.
 1580  */
 1581 int
 1582 insmntque1(struct vnode *vp, struct mount *mp,
 1583         void (*dtr)(struct vnode *, void *), void *dtr_arg)
 1584 {
 1585 
 1586         KASSERT(vp->v_mount == NULL,
 1587                 ("insmntque: vnode already on per mount vnode list"));
 1588         VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
 1589         ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
 1590 
 1591         /*
 1592          * We acquire the vnode interlock early to ensure that the
 1593          * vnode cannot be recycled by another process releasing a
 1594          * holdcnt on it before we get it on both the vnode list
 1595          * and the active vnode list. The mount mutex protects only
 1596          * manipulation of the vnode list and the vnode freelist
 1597          * mutex protects only manipulation of the active vnode list.
 1598          * Hence the need to hold the vnode interlock throughout.
 1599          */
 1600         MNT_ILOCK(mp);
 1601         VI_LOCK(vp);
 1602         if (((mp->mnt_kern_flag & MNTK_NOINSMNTQ) != 0 &&
 1603             ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
 1604             mp->mnt_nvnodelistsize == 0)) &&
 1605             (vp->v_vflag & VV_FORCEINSMQ) == 0) {
 1606                 VI_UNLOCK(vp);
 1607                 MNT_IUNLOCK(mp);
 1608                 if (dtr != NULL)
 1609                         dtr(vp, dtr_arg);
 1610                 return (EBUSY);
 1611         }
 1612         vp->v_mount = mp;
 1613         MNT_REF(mp);
 1614         TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
 1615         VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
 1616                 ("neg mount point vnode list size"));
 1617         mp->mnt_nvnodelistsize++;
 1618         KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
 1619             ("Activating already active vnode"));
 1620         vp->v_iflag |= VI_ACTIVE;
 1621         mtx_lock(&mp->mnt_listmtx);
 1622         TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
 1623         mp->mnt_activevnodelistsize++;
 1624         mtx_unlock(&mp->mnt_listmtx);
 1625         VI_UNLOCK(vp);
 1626         MNT_IUNLOCK(mp);
 1627         return (0);
 1628 }
 1629 
 1630 int
 1631 insmntque(struct vnode *vp, struct mount *mp)
 1632 {
 1633 
 1634         return (insmntque1(vp, mp, insmntque_stddtr, NULL));
 1635 }
 1636 
 1637 /*
 1638  * Flush out and invalidate all buffers associated with a bufobj
 1639  * Called with the underlying object locked.
 1640  */
 1641 int
 1642 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
 1643 {
 1644         int error;
 1645 
 1646         BO_LOCK(bo);
 1647         if (flags & V_SAVE) {
 1648                 error = bufobj_wwait(bo, slpflag, slptimeo);
 1649                 if (error) {
 1650                         BO_UNLOCK(bo);
 1651                         return (error);
 1652                 }
 1653                 if (bo->bo_dirty.bv_cnt > 0) {
 1654                         BO_UNLOCK(bo);
 1655                         if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
 1656                                 return (error);
 1657                         /*
 1658                          * XXX We could save a lock/unlock if this was only
 1659                          * enabled under INVARIANTS
 1660                          */
 1661                         BO_LOCK(bo);
 1662                         if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
 1663                                 panic("vinvalbuf: dirty bufs");
 1664                 }
 1665         }
 1666         /*
 1667          * If you alter this loop please notice that interlock is dropped and
 1668          * reacquired in flushbuflist.  Special care is needed to ensure that
 1669          * no race conditions occur from this.
 1670          */
 1671         do {
 1672                 error = flushbuflist(&bo->bo_clean,
 1673                     flags, bo, slpflag, slptimeo);
 1674                 if (error == 0 && !(flags & V_CLEANONLY))
 1675                         error = flushbuflist(&bo->bo_dirty,
 1676                             flags, bo, slpflag, slptimeo);
 1677                 if (error != 0 && error != EAGAIN) {
 1678                         BO_UNLOCK(bo);
 1679                         return (error);
 1680                 }
 1681         } while (error != 0);
 1682 
 1683         /*
 1684          * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
 1685          * have write I/O in-progress but if there is a VM object then the
 1686          * VM object can also have read-I/O in-progress.
 1687          */
 1688         do {
 1689                 bufobj_wwait(bo, 0, 0);
 1690                 if ((flags & V_VMIO) == 0) {
 1691                         BO_UNLOCK(bo);
 1692                         if (bo->bo_object != NULL) {
 1693                                 VM_OBJECT_WLOCK(bo->bo_object);
 1694                                 vm_object_pip_wait(bo->bo_object, "bovlbx");
 1695                                 VM_OBJECT_WUNLOCK(bo->bo_object);
 1696                         }
 1697                         BO_LOCK(bo);
 1698                 }
 1699         } while (bo->bo_numoutput > 0);
 1700         BO_UNLOCK(bo);
 1701 
 1702         /*
 1703          * Destroy the copy in the VM cache, too.
 1704          */
 1705         if (bo->bo_object != NULL &&
 1706             (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
 1707                 VM_OBJECT_WLOCK(bo->bo_object);
 1708                 vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
 1709                     OBJPR_CLEANONLY : 0);
 1710                 VM_OBJECT_WUNLOCK(bo->bo_object);
 1711         }
 1712 
 1713 #ifdef INVARIANTS
 1714         BO_LOCK(bo);
 1715         if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
 1716             V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
 1717             bo->bo_clean.bv_cnt > 0))
 1718                 panic("vinvalbuf: flush failed");
 1719         if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
 1720             bo->bo_dirty.bv_cnt > 0)
 1721                 panic("vinvalbuf: flush dirty failed");
 1722         BO_UNLOCK(bo);
 1723 #endif
 1724         return (0);
 1725 }
 1726 
 1727 /*
 1728  * Flush out and invalidate all buffers associated with a vnode.
 1729  * Called with the underlying object locked.
 1730  */
 1731 int
 1732 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
 1733 {
 1734 
 1735         CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
 1736         ASSERT_VOP_LOCKED(vp, "vinvalbuf");
 1737         if (vp->v_object != NULL && vp->v_object->handle != vp)
 1738                 return (0);
 1739         return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
 1740 }
 1741 
 1742 /*
 1743  * Flush out buffers on the specified list.
 1744  *
 1745  */
 1746 static int
 1747 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
 1748     int slptimeo)
 1749 {
 1750         struct buf *bp, *nbp;
 1751         int retval, error;
 1752         daddr_t lblkno;
 1753         b_xflags_t xflags;
 1754 
 1755         ASSERT_BO_WLOCKED(bo);
 1756 
 1757         retval = 0;
 1758         TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
 1759                 if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
 1760                     ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
 1761                         continue;
 1762                 }
 1763                 if (nbp != NULL) {
 1764                         lblkno = nbp->b_lblkno;
 1765                         xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
 1766                 }
 1767                 retval = EAGAIN;
 1768                 error = BUF_TIMELOCK(bp,
 1769                     LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
 1770                     "flushbuf", slpflag, slptimeo);
 1771                 if (error) {
 1772                         BO_LOCK(bo);
 1773                         return (error != ENOLCK ? error : EAGAIN);
 1774                 }
 1775                 KASSERT(bp->b_bufobj == bo,
 1776                     ("bp %p wrong b_bufobj %p should be %p",
 1777                     bp, bp->b_bufobj, bo));
 1778                 /*
 1779                  * XXX Since there are no node locks for NFS, I
 1780                  * believe there is a slight chance that a delayed
 1781                  * write will occur while sleeping just above, so
 1782                  * check for it.
 1783                  */
 1784                 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
 1785                     (flags & V_SAVE)) {
 1786                         bremfree(bp);
 1787                         bp->b_flags |= B_ASYNC;
 1788                         bwrite(bp);
 1789                         BO_LOCK(bo);
 1790                         return (EAGAIN);        /* XXX: why not loop ? */
 1791                 }
 1792                 bremfree(bp);
 1793                 bp->b_flags |= (B_INVAL | B_RELBUF);
 1794                 bp->b_flags &= ~B_ASYNC;
 1795                 brelse(bp);
 1796                 BO_LOCK(bo);
 1797                 if (nbp == NULL)
 1798                         break;
 1799                 nbp = gbincore(bo, lblkno);
 1800                 if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
 1801                     != xflags)
 1802                         break;                  /* nbp invalid */
 1803         }
 1804         return (retval);
 1805 }
 1806 
 1807 int
 1808 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
 1809 {
 1810         struct buf *bp;
 1811         int error;
 1812         daddr_t lblkno;
 1813 
 1814         ASSERT_BO_LOCKED(bo);
 1815 
 1816         for (lblkno = startn;;) {
 1817 again:
 1818                 bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
 1819                 if (bp == NULL || bp->b_lblkno >= endn ||
 1820                     bp->b_lblkno < startn)
 1821                         break;
 1822                 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
 1823                     LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
 1824                 if (error != 0) {
 1825                         BO_RLOCK(bo);
 1826                         if (error == ENOLCK)
 1827                                 goto again;
 1828                         return (error);
 1829                 }
 1830                 KASSERT(bp->b_bufobj == bo,
 1831                     ("bp %p wrong b_bufobj %p should be %p",
 1832                     bp, bp->b_bufobj, bo));
 1833                 lblkno = bp->b_lblkno + 1;
 1834                 if ((bp->b_flags & B_MANAGED) == 0)
 1835                         bremfree(bp);
 1836                 bp->b_flags |= B_RELBUF;
 1837                 /*
 1838                  * In the VMIO case, use the B_NOREUSE flag to hint that the
 1839                  * pages backing each buffer in the range are unlikely to be
 1840                  * reused.  Dirty buffers will have the hint applied once
 1841                  * they've been written.
 1842                  */
 1843                 if (bp->b_vp->v_object != NULL)
 1844                         bp->b_flags |= B_NOREUSE;
 1845                 brelse(bp);
 1846                 BO_RLOCK(bo);
 1847         }
 1848         return (0);
 1849 }
 1850 
 1851 /*
 1852  * Truncate a file's buffer and pages to a specified length.  This
 1853  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
 1854  * sync activity.
 1855  */
 1856 int
 1857 vtruncbuf(struct vnode *vp, struct ucred *cred, off_t length, int blksize)
 1858 {
 1859         struct buf *bp, *nbp;
 1860         int anyfreed;
 1861         int trunclbn;
 1862         struct bufobj *bo;
 1863 
 1864         CTR5(KTR_VFS, "%s: vp %p with cred %p and block %d:%ju", __func__,
 1865             vp, cred, blksize, (uintmax_t)length);
 1866 
 1867         /*
 1868          * Round up to the *next* lbn.
 1869          */
 1870         trunclbn = howmany(length, blksize);
 1871 
 1872         ASSERT_VOP_LOCKED(vp, "vtruncbuf");
 1873 restart:
 1874         bo = &vp->v_bufobj;
 1875         BO_LOCK(bo);
 1876         anyfreed = 1;
 1877         for (;anyfreed;) {
 1878                 anyfreed = 0;
 1879                 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
 1880                         if (bp->b_lblkno < trunclbn)
 1881                                 continue;
 1882                         if (BUF_LOCK(bp,
 1883                             LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
 1884                             BO_LOCKPTR(bo)) == ENOLCK)
 1885                                 goto restart;
 1886 
 1887                         bremfree(bp);
 1888                         bp->b_flags |= (B_INVAL | B_RELBUF);
 1889                         bp->b_flags &= ~B_ASYNC;
 1890                         brelse(bp);
 1891                         anyfreed = 1;
 1892 
 1893                         BO_LOCK(bo);
 1894                         if (nbp != NULL &&
 1895                             (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
 1896                             (nbp->b_vp != vp) ||
 1897                             (nbp->b_flags & B_DELWRI))) {
 1898                                 BO_UNLOCK(bo);
 1899                                 goto restart;
 1900                         }
 1901                 }
 1902 
 1903                 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
 1904                         if (bp->b_lblkno < trunclbn)
 1905                                 continue;
 1906                         if (BUF_LOCK(bp,
 1907                             LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
 1908                             BO_LOCKPTR(bo)) == ENOLCK)
 1909                                 goto restart;
 1910                         bremfree(bp);
 1911                         bp->b_flags |= (B_INVAL | B_RELBUF);
 1912                         bp->b_flags &= ~B_ASYNC;
 1913                         brelse(bp);
 1914                         anyfreed = 1;
 1915 
 1916                         BO_LOCK(bo);
 1917                         if (nbp != NULL &&
 1918                             (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
 1919                             (nbp->b_vp != vp) ||
 1920                             (nbp->b_flags & B_DELWRI) == 0)) {
 1921                                 BO_UNLOCK(bo);
 1922                                 goto restart;
 1923                         }
 1924                 }
 1925         }
 1926 
 1927         if (length > 0) {
 1928 restartsync:
 1929                 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
 1930                         if (bp->b_lblkno > 0)
 1931                                 continue;
 1932                         /*
 1933                          * Since we hold the vnode lock this should only
 1934                          * fail if we're racing with the buf daemon.
 1935                          */
 1936                         if (BUF_LOCK(bp,
 1937                             LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
 1938                             BO_LOCKPTR(bo)) == ENOLCK) {
 1939                                 goto restart;
 1940                         }
 1941                         VNASSERT((bp->b_flags & B_DELWRI), vp,
 1942                             ("buf(%p) on dirty queue without DELWRI", bp));
 1943 
 1944                         bremfree(bp);
 1945                         bawrite(bp);
 1946                         BO_LOCK(bo);
 1947                         goto restartsync;
 1948                 }
 1949         }
 1950 
 1951         bufobj_wwait(bo, 0, 0);
 1952         BO_UNLOCK(bo);
 1953         vnode_pager_setsize(vp, length);
 1954 
 1955         return (0);
 1956 }
 1957 
 1958 static void
 1959 buf_vlist_remove(struct buf *bp)
 1960 {
 1961         struct bufv *bv;
 1962 
 1963         KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
 1964         ASSERT_BO_WLOCKED(bp->b_bufobj);
 1965         KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
 1966             (BX_VNDIRTY|BX_VNCLEAN),
 1967             ("buf_vlist_remove: Buf %p is on two lists", bp));
 1968         if (bp->b_xflags & BX_VNDIRTY)
 1969                 bv = &bp->b_bufobj->bo_dirty;
 1970         else
 1971                 bv = &bp->b_bufobj->bo_clean;
 1972         BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
 1973         TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
 1974         bv->bv_cnt--;
 1975         bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
 1976 }
 1977 
 1978 /*
 1979  * Add the buffer to the sorted clean or dirty block list.
 1980  *
 1981  * NOTE: xflags is passed as a constant, optimizing this inline function!
 1982  */
 1983 static void
 1984 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
 1985 {
 1986         struct bufv *bv;
 1987         struct buf *n;
 1988         int error;
 1989 
 1990         ASSERT_BO_WLOCKED(bo);
 1991         KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
 1992             ("dead bo %p", bo));
 1993         KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
 1994             ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
 1995         bp->b_xflags |= xflags;
 1996         if (xflags & BX_VNDIRTY)
 1997                 bv = &bo->bo_dirty;
 1998         else
 1999                 bv = &bo->bo_clean;
 2000 
 2001         /*
 2002          * Keep the list ordered.  Optimize empty list insertion.  Assume
 2003          * we tend to grow at the tail so lookup_le should usually be cheaper
 2004          * than _ge. 
 2005          */
 2006         if (bv->bv_cnt == 0 ||
 2007             bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
 2008                 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
 2009         else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
 2010                 TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
 2011         else
 2012                 TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
 2013         error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
 2014         if (error)
 2015                 panic("buf_vlist_add:  Preallocated nodes insufficient.");
 2016         bv->bv_cnt++;
 2017 }
 2018 
 2019 /*
 2020  * Look up a buffer using the buffer tries.
 2021  */
 2022 struct buf *
 2023 gbincore(struct bufobj *bo, daddr_t lblkno)
 2024 {
 2025         struct buf *bp;
 2026 
 2027         ASSERT_BO_LOCKED(bo);
 2028         bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
 2029         if (bp != NULL)
 2030                 return (bp);
 2031         return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
 2032 }
 2033 
 2034 /*
 2035  * Associate a buffer with a vnode.
 2036  */
 2037 void
 2038 bgetvp(struct vnode *vp, struct buf *bp)
 2039 {
 2040         struct bufobj *bo;
 2041 
 2042         bo = &vp->v_bufobj;
 2043         ASSERT_BO_WLOCKED(bo);
 2044         VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
 2045 
 2046         CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
 2047         VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
 2048             ("bgetvp: bp already attached! %p", bp));
 2049 
 2050         vhold(vp);
 2051         bp->b_vp = vp;
 2052         bp->b_bufobj = bo;
 2053         /*
 2054          * Insert onto list for new vnode.
 2055          */
 2056         buf_vlist_add(bp, bo, BX_VNCLEAN);
 2057 }
 2058 
 2059 /*
 2060  * Disassociate a buffer from a vnode.
 2061  */
 2062 void
 2063 brelvp(struct buf *bp)
 2064 {
 2065         struct bufobj *bo;
 2066         struct vnode *vp;
 2067 
 2068         CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
 2069         KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
 2070 
 2071         /*
 2072          * Delete from old vnode list, if on one.
 2073          */
 2074         vp = bp->b_vp;          /* XXX */
 2075         bo = bp->b_bufobj;
 2076         BO_LOCK(bo);
 2077         if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
 2078                 buf_vlist_remove(bp);
 2079         else
 2080                 panic("brelvp: Buffer %p not on queue.", bp);
 2081         if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
 2082                 bo->bo_flag &= ~BO_ONWORKLST;
 2083                 mtx_lock(&sync_mtx);
 2084                 LIST_REMOVE(bo, bo_synclist);
 2085                 syncer_worklist_len--;
 2086                 mtx_unlock(&sync_mtx);
 2087         }
 2088         bp->b_vp = NULL;
 2089         bp->b_bufobj = NULL;
 2090         BO_UNLOCK(bo);
 2091         vdrop(vp);
 2092 }
 2093 
 2094 /*
 2095  * Add an item to the syncer work queue.
 2096  */
 2097 static void
 2098 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
 2099 {
 2100         int slot;
 2101 
 2102         ASSERT_BO_WLOCKED(bo);
 2103 
 2104         mtx_lock(&sync_mtx);
 2105         if (bo->bo_flag & BO_ONWORKLST)
 2106                 LIST_REMOVE(bo, bo_synclist);
 2107         else {
 2108                 bo->bo_flag |= BO_ONWORKLST;
 2109                 syncer_worklist_len++;
 2110         }
 2111 
 2112         if (delay > syncer_maxdelay - 2)
 2113                 delay = syncer_maxdelay - 2;
 2114         slot = (syncer_delayno + delay) & syncer_mask;
 2115 
 2116         LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
 2117         mtx_unlock(&sync_mtx);
 2118 }
 2119 
 2120 static int
 2121 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
 2122 {
 2123         int error, len;
 2124 
 2125         mtx_lock(&sync_mtx);
 2126         len = syncer_worklist_len - sync_vnode_count;
 2127         mtx_unlock(&sync_mtx);
 2128         error = SYSCTL_OUT(req, &len, sizeof(len));
 2129         return (error);
 2130 }
 2131 
 2132 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
 2133     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
 2134 
 2135 static struct proc *updateproc;
 2136 static void sched_sync(void);
 2137 static struct kproc_desc up_kp = {
 2138         "syncer",
 2139         sched_sync,
 2140         &updateproc
 2141 };
 2142 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
 2143 
 2144 static int
 2145 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
 2146 {
 2147         struct vnode *vp;
 2148         struct mount *mp;
 2149 
 2150         *bo = LIST_FIRST(slp);
 2151         if (*bo == NULL)
 2152                 return (0);
 2153         vp = bo2vnode(*bo);
 2154         if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
 2155                 return (1);
 2156         /*
 2157          * We use vhold in case the vnode does not
 2158          * successfully sync.  vhold prevents the vnode from
 2159          * going away when we unlock the sync_mtx so that
 2160          * we can acquire the vnode interlock.
 2161          */
 2162         vholdl(vp);
 2163         mtx_unlock(&sync_mtx);
 2164         VI_UNLOCK(vp);
 2165         if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
 2166                 vdrop(vp);
 2167                 mtx_lock(&sync_mtx);
 2168                 return (*bo == LIST_FIRST(slp));
 2169         }
 2170         vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 2171         (void) VOP_FSYNC(vp, MNT_LAZY, td);
 2172         VOP_UNLOCK(vp, 0);
 2173         vn_finished_write(mp);
 2174         BO_LOCK(*bo);
 2175         if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
 2176                 /*
 2177                  * Put us back on the worklist.  The worklist
 2178                  * routine will remove us from our current
 2179                  * position and then add us back in at a later
 2180                  * position.
 2181                  */
 2182                 vn_syncer_add_to_worklist(*bo, syncdelay);
 2183         }
 2184         BO_UNLOCK(*bo);
 2185         vdrop(vp);
 2186         mtx_lock(&sync_mtx);
 2187         return (0);
 2188 }
 2189 
 2190 static int first_printf = 1;
 2191 
 2192 /*
 2193  * System filesystem synchronizer daemon.
 2194  */
 2195 static void
 2196 sched_sync(void)
 2197 {
 2198         struct synclist *next, *slp;
 2199         struct bufobj *bo;
 2200         long starttime;
 2201         struct thread *td = curthread;
 2202         int last_work_seen;
 2203         int net_worklist_len;
 2204         int syncer_final_iter;
 2205         int error;
 2206 
 2207         last_work_seen = 0;
 2208         syncer_final_iter = 0;
 2209         syncer_state = SYNCER_RUNNING;
 2210         starttime = time_uptime;
 2211         td->td_pflags |= TDP_NORUNNINGBUF;
 2212 
 2213         EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
 2214             SHUTDOWN_PRI_LAST);
 2215 
 2216         mtx_lock(&sync_mtx);
 2217         for (;;) {
 2218                 if (syncer_state == SYNCER_FINAL_DELAY &&
 2219                     syncer_final_iter == 0) {
 2220                         mtx_unlock(&sync_mtx);
 2221                         kproc_suspend_check(td->td_proc);
 2222                         mtx_lock(&sync_mtx);
 2223                 }
 2224                 net_worklist_len = syncer_worklist_len - sync_vnode_count;
 2225                 if (syncer_state != SYNCER_RUNNING &&
 2226                     starttime != time_uptime) {
 2227                         if (first_printf) {
 2228                                 printf("\nSyncing disks, vnodes remaining... ");
 2229                                 first_printf = 0;
 2230                         }
 2231                         printf("%d ", net_worklist_len);
 2232                 }
 2233                 starttime = time_uptime;
 2234 
 2235                 /*
 2236                  * Push files whose dirty time has expired.  Be careful
 2237                  * of interrupt race on slp queue.
 2238                  *
 2239                  * Skip over empty worklist slots when shutting down.
 2240                  */
 2241                 do {
 2242                         slp = &syncer_workitem_pending[syncer_delayno];
 2243                         syncer_delayno += 1;
 2244                         if (syncer_delayno == syncer_maxdelay)
 2245                                 syncer_delayno = 0;
 2246                         next = &syncer_workitem_pending[syncer_delayno];
 2247                         /*
 2248                          * If the worklist has wrapped since the
 2249                          * it was emptied of all but syncer vnodes,
 2250                          * switch to the FINAL_DELAY state and run
 2251                          * for one more second.
 2252                          */
 2253                         if (syncer_state == SYNCER_SHUTTING_DOWN &&
 2254                             net_worklist_len == 0 &&
 2255                             last_work_seen == syncer_delayno) {
 2256                                 syncer_state = SYNCER_FINAL_DELAY;
 2257                                 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
 2258                         }
 2259                 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
 2260                     syncer_worklist_len > 0);
 2261 
 2262                 /*
 2263                  * Keep track of the last time there was anything
 2264                  * on the worklist other than syncer vnodes.
 2265                  * Return to the SHUTTING_DOWN state if any
 2266                  * new work appears.
 2267                  */
 2268                 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
 2269                         last_work_seen = syncer_delayno;
 2270                 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
 2271                         syncer_state = SYNCER_SHUTTING_DOWN;
 2272                 while (!LIST_EMPTY(slp)) {
 2273                         error = sync_vnode(slp, &bo, td);
 2274                         if (error == 1) {
 2275                                 LIST_REMOVE(bo, bo_synclist);
 2276                                 LIST_INSERT_HEAD(next, bo, bo_synclist);
 2277                                 continue;
 2278                         }
 2279 
 2280                         if (first_printf == 0) {
 2281                                 /*
 2282                                  * Drop the sync mutex, because some watchdog
 2283                                  * drivers need to sleep while patting
 2284                                  */
 2285                                 mtx_unlock(&sync_mtx);
 2286                                 wdog_kern_pat(WD_LASTVAL);
 2287                                 mtx_lock(&sync_mtx);
 2288                         }
 2289 
 2290                 }
 2291                 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
 2292                         syncer_final_iter--;
 2293                 /*
 2294                  * The variable rushjob allows the kernel to speed up the
 2295                  * processing of the filesystem syncer process. A rushjob
 2296                  * value of N tells the filesystem syncer to process the next
 2297                  * N seconds worth of work on its queue ASAP. Currently rushjob
 2298                  * is used by the soft update code to speed up the filesystem
 2299                  * syncer process when the incore state is getting so far
 2300                  * ahead of the disk that the kernel memory pool is being
 2301                  * threatened with exhaustion.
 2302                  */
 2303                 if (rushjob > 0) {
 2304                         rushjob -= 1;
 2305                         continue;
 2306                 }
 2307                 /*
 2308                  * Just sleep for a short period of time between
 2309                  * iterations when shutting down to allow some I/O
 2310                  * to happen.
 2311                  *
 2312                  * If it has taken us less than a second to process the
 2313                  * current work, then wait. Otherwise start right over
 2314                  * again. We can still lose time if any single round
 2315                  * takes more than two seconds, but it does not really
 2316                  * matter as we are just trying to generally pace the
 2317                  * filesystem activity.
 2318                  */
 2319                 if (syncer_state != SYNCER_RUNNING ||
 2320                     time_uptime == starttime) {
 2321                         thread_lock(td);
 2322                         sched_prio(td, PPAUSE);
 2323                         thread_unlock(td);
 2324                 }
 2325                 if (syncer_state != SYNCER_RUNNING)
 2326                         cv_timedwait(&sync_wakeup, &sync_mtx,
 2327                             hz / SYNCER_SHUTDOWN_SPEEDUP);
 2328                 else if (time_uptime == starttime)
 2329                         cv_timedwait(&sync_wakeup, &sync_mtx, hz);
 2330         }
 2331 }
 2332 
 2333 /*
 2334  * Request the syncer daemon to speed up its work.
 2335  * We never push it to speed up more than half of its
 2336  * normal turn time, otherwise it could take over the cpu.
 2337  */
 2338 int
 2339 speedup_syncer(void)
 2340 {
 2341         int ret = 0;
 2342 
 2343         mtx_lock(&sync_mtx);
 2344         if (rushjob < syncdelay / 2) {
 2345                 rushjob += 1;
 2346                 stat_rush_requests += 1;
 2347                 ret = 1;
 2348         }
 2349         mtx_unlock(&sync_mtx);
 2350         cv_broadcast(&sync_wakeup);
 2351         return (ret);
 2352 }
 2353 
 2354 /*
 2355  * Tell the syncer to speed up its work and run though its work
 2356  * list several times, then tell it to shut down.
 2357  */
 2358 static void
 2359 syncer_shutdown(void *arg, int howto)
 2360 {
 2361 
 2362         if (howto & RB_NOSYNC)
 2363                 return;
 2364         mtx_lock(&sync_mtx);
 2365         syncer_state = SYNCER_SHUTTING_DOWN;
 2366         rushjob = 0;
 2367         mtx_unlock(&sync_mtx);
 2368         cv_broadcast(&sync_wakeup);
 2369         kproc_shutdown(arg, howto);
 2370 }
 2371 
 2372 void
 2373 syncer_suspend(void)
 2374 {
 2375 
 2376         syncer_shutdown(updateproc, 0);
 2377 }
 2378 
 2379 void
 2380 syncer_resume(void)
 2381 {
 2382 
 2383         mtx_lock(&sync_mtx);
 2384         first_printf = 1;
 2385         syncer_state = SYNCER_RUNNING;
 2386         mtx_unlock(&sync_mtx);
 2387         cv_broadcast(&sync_wakeup);
 2388         kproc_resume(updateproc);
 2389 }
 2390 
 2391 /*
 2392  * Reassign a buffer from one vnode to another.
 2393  * Used to assign file specific control information
 2394  * (indirect blocks) to the vnode to which they belong.
 2395  */
 2396 void
 2397 reassignbuf(struct buf *bp)
 2398 {
 2399         struct vnode *vp;
 2400         struct bufobj *bo;
 2401         int delay;
 2402 #ifdef INVARIANTS
 2403         struct bufv *bv;
 2404 #endif
 2405 
 2406         vp = bp->b_vp;
 2407         bo = bp->b_bufobj;
 2408         ++reassignbufcalls;
 2409 
 2410         CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
 2411             bp, bp->b_vp, bp->b_flags);
 2412         /*
 2413          * B_PAGING flagged buffers cannot be reassigned because their vp
 2414          * is not fully linked in.
 2415          */
 2416         if (bp->b_flags & B_PAGING)
 2417                 panic("cannot reassign paging buffer");
 2418 
 2419         /*
 2420          * Delete from old vnode list, if on one.
 2421          */
 2422         BO_LOCK(bo);
 2423         if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
 2424                 buf_vlist_remove(bp);
 2425         else
 2426                 panic("reassignbuf: Buffer %p not on queue.", bp);
 2427         /*
 2428          * If dirty, put on list of dirty buffers; otherwise insert onto list
 2429          * of clean buffers.
 2430          */
 2431         if (bp->b_flags & B_DELWRI) {
 2432                 if ((bo->bo_flag & BO_ONWORKLST) == 0) {
 2433                         switch (vp->v_type) {
 2434                         case VDIR:
 2435                                 delay = dirdelay;
 2436                                 break;
 2437                         case VCHR:
 2438                                 delay = metadelay;
 2439                                 break;
 2440                         default:
 2441                                 delay = filedelay;
 2442                         }
 2443                         vn_syncer_add_to_worklist(bo, delay);
 2444                 }
 2445                 buf_vlist_add(bp, bo, BX_VNDIRTY);
 2446         } else {
 2447                 buf_vlist_add(bp, bo, BX_VNCLEAN);
 2448 
 2449                 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
 2450                         mtx_lock(&sync_mtx);
 2451                         LIST_REMOVE(bo, bo_synclist);
 2452                         syncer_worklist_len--;
 2453                         mtx_unlock(&sync_mtx);
 2454                         bo->bo_flag &= ~BO_ONWORKLST;
 2455                 }
 2456         }
 2457 #ifdef INVARIANTS
 2458         bv = &bo->bo_clean;
 2459         bp = TAILQ_FIRST(&bv->bv_hd);
 2460         KASSERT(bp == NULL || bp->b_bufobj == bo,
 2461             ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
 2462         bp = TAILQ_LAST(&bv->bv_hd, buflists);
 2463         KASSERT(bp == NULL || bp->b_bufobj == bo,
 2464             ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
 2465         bv = &bo->bo_dirty;
 2466         bp = TAILQ_FIRST(&bv->bv_hd);
 2467         KASSERT(bp == NULL || bp->b_bufobj == bo,
 2468             ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
 2469         bp = TAILQ_LAST(&bv->bv_hd, buflists);
 2470         KASSERT(bp == NULL || bp->b_bufobj == bo,
 2471             ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
 2472 #endif
 2473         BO_UNLOCK(bo);
 2474 }
 2475 
 2476 static void
 2477 v_init_counters(struct vnode *vp)
 2478 {
 2479 
 2480         VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
 2481             vp, ("%s called for an initialized vnode", __FUNCTION__));
 2482         ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
 2483 
 2484         refcount_init(&vp->v_holdcnt, 1);
 2485         refcount_init(&vp->v_usecount, 1);
 2486 }
 2487 
 2488 static void
 2489 v_incr_usecount_locked(struct vnode *vp)
 2490 {
 2491 
 2492         ASSERT_VI_LOCKED(vp, __func__);
 2493         if ((vp->v_iflag & VI_OWEINACT) != 0) {
 2494                 VNASSERT(vp->v_usecount == 0, vp,
 2495                     ("vnode with usecount and VI_OWEINACT set"));
 2496                 vp->v_iflag &= ~VI_OWEINACT;
 2497         }
 2498         refcount_acquire(&vp->v_usecount);
 2499         v_incr_devcount(vp);
 2500 }
 2501 
 2502 /*
 2503  * Increment the use count on the vnode, taking care to reference
 2504  * the driver's usecount if this is a chardev.
 2505  */
 2506 static void
 2507 v_incr_usecount(struct vnode *vp)
 2508 {
 2509 
 2510         ASSERT_VI_UNLOCKED(vp, __func__);
 2511         CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
 2512 
 2513         if (vp->v_type != VCHR &&
 2514             refcount_acquire_if_not_zero(&vp->v_usecount)) {
 2515                 VNODE_REFCOUNT_FENCE_ACQ();
 2516                 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
 2517                     ("vnode with usecount and VI_OWEINACT set"));
 2518         } else {
 2519                 VI_LOCK(vp);
 2520                 v_incr_usecount_locked(vp);
 2521                 VI_UNLOCK(vp);
 2522         }
 2523 }
 2524 
 2525 /*
 2526  * Increment si_usecount of the associated device, if any.
 2527  */
 2528 static void
 2529 v_incr_devcount(struct vnode *vp)
 2530 {
 2531 
 2532         ASSERT_VI_LOCKED(vp, __FUNCTION__);
 2533         if (vp->v_type == VCHR && vp->v_rdev != NULL) {
 2534                 dev_lock();
 2535                 vp->v_rdev->si_usecount++;
 2536                 dev_unlock();
 2537         }
 2538 }
 2539 
 2540 /*
 2541  * Decrement si_usecount of the associated device, if any.
 2542  */
 2543 static void
 2544 v_decr_devcount(struct vnode *vp)
 2545 {
 2546 
 2547         ASSERT_VI_LOCKED(vp, __FUNCTION__);
 2548         if (vp->v_type == VCHR && vp->v_rdev != NULL) {
 2549                 dev_lock();
 2550                 vp->v_rdev->si_usecount--;
 2551                 dev_unlock();
 2552         }
 2553 }
 2554 
 2555 /*
 2556  * Grab a particular vnode from the free list, increment its
 2557  * reference count and lock it.  VI_DOOMED is set if the vnode
 2558  * is being destroyed.  Only callers who specify LK_RETRY will
 2559  * see doomed vnodes.  If inactive processing was delayed in
 2560  * vput try to do it here.
 2561  *
 2562  * Notes on lockless counter manipulation:
 2563  * _vhold, vputx and other routines make various decisions based
 2564  * on either holdcnt or usecount being 0. As long as either counter
 2565  * is not transitioning 0->1 nor 1->0, the manipulation can be done
 2566  * with atomic operations. Otherwise the interlock is taken covering
 2567  * both the atomic and additional actions.
 2568  */
 2569 int
 2570 vget(struct vnode *vp, int flags, struct thread *td)
 2571 {
 2572         int error, oweinact;
 2573 
 2574         VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
 2575             ("vget: invalid lock operation"));
 2576 
 2577         if ((flags & LK_INTERLOCK) != 0)
 2578                 ASSERT_VI_LOCKED(vp, __func__);
 2579         else
 2580                 ASSERT_VI_UNLOCKED(vp, __func__);
 2581         if ((flags & LK_VNHELD) != 0)
 2582                 VNASSERT((vp->v_holdcnt > 0), vp,
 2583                     ("vget: LK_VNHELD passed but vnode not held"));
 2584 
 2585         CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
 2586 
 2587         if ((flags & LK_VNHELD) == 0)
 2588                 _vhold(vp, (flags & LK_INTERLOCK) != 0);
 2589 
 2590         if ((error = vn_lock(vp, flags)) != 0) {
 2591                 vdrop(vp);
 2592                 CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
 2593                     vp);
 2594                 return (error);
 2595         }
 2596         if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
 2597                 panic("vget: vn_lock failed to return ENOENT\n");
 2598         /*
 2599          * We don't guarantee that any particular close will
 2600          * trigger inactive processing so just make a best effort
 2601          * here at preventing a reference to a removed file.  If
 2602          * we don't succeed no harm is done.
 2603          *
 2604          * Upgrade our holdcnt to a usecount.
 2605          */
 2606         if (vp->v_type == VCHR ||
 2607             !refcount_acquire_if_not_zero(&vp->v_usecount)) {
 2608                 VI_LOCK(vp);
 2609                 if ((vp->v_iflag & VI_OWEINACT) == 0) {
 2610                         oweinact = 0;
 2611                 } else {
 2612                         oweinact = 1;
 2613                         vp->v_iflag &= ~VI_OWEINACT;
 2614                         VNODE_REFCOUNT_FENCE_REL();
 2615                 }
 2616                 refcount_acquire(&vp->v_usecount);
 2617                 v_incr_devcount(vp);
 2618                 if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
 2619                     (flags & LK_NOWAIT) == 0)
 2620                         vinactive(vp, td);
 2621                 VI_UNLOCK(vp);
 2622         }
 2623         return (0);
 2624 }
 2625 
 2626 /*
 2627  * Increase the reference (use) and hold count of a vnode.
 2628  * This will also remove the vnode from the free list if it is presently free.
 2629  */
 2630 void
 2631 vref(struct vnode *vp)
 2632 {
 2633 
 2634         CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
 2635         _vhold(vp, false);
 2636         v_incr_usecount(vp);
 2637 }
 2638 
 2639 void
 2640 vrefl(struct vnode *vp)
 2641 {
 2642 
 2643         ASSERT_VI_LOCKED(vp, __func__);
 2644         CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
 2645         _vhold(vp, true);
 2646         v_incr_usecount_locked(vp);
 2647 }
 2648 
 2649 void
 2650 vrefact(struct vnode *vp)
 2651 {
 2652 
 2653         CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
 2654         if (__predict_false(vp->v_type == VCHR)) {
 2655                 VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp,
 2656                     ("%s: wrong ref counts", __func__));
 2657                 vref(vp);
 2658                 return;
 2659         }
 2660 #ifdef INVARIANTS
 2661         int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
 2662         VNASSERT(old > 0, vp, ("%s: wrong hold count", __func__));
 2663         old = atomic_fetchadd_int(&vp->v_usecount, 1);
 2664         VNASSERT(old > 0, vp, ("%s: wrong use count", __func__));
 2665 #else
 2666         refcount_acquire(&vp->v_holdcnt);
 2667         refcount_acquire(&vp->v_usecount);
 2668 #endif
 2669 }
 2670 
 2671 /*
 2672  * Return reference count of a vnode.
 2673  *
 2674  * The results of this call are only guaranteed when some mechanism is used to
 2675  * stop other processes from gaining references to the vnode.  This may be the
 2676  * case if the caller holds the only reference.  This is also useful when stale
 2677  * data is acceptable as race conditions may be accounted for by some other
 2678  * means.
 2679  */
 2680 int
 2681 vrefcnt(struct vnode *vp)
 2682 {
 2683 
 2684         return (vp->v_usecount);
 2685 }
 2686 
 2687 #define VPUTX_VRELE     1
 2688 #define VPUTX_VPUT      2
 2689 #define VPUTX_VUNREF    3
 2690 
 2691 /*
 2692  * Decrement the use and hold counts for a vnode.
 2693  *
 2694  * See an explanation near vget() as to why atomic operation is safe.
 2695  */
 2696 static void
 2697 vputx(struct vnode *vp, int func)
 2698 {
 2699         int error;
 2700 
 2701         KASSERT(vp != NULL, ("vputx: null vp"));
 2702         if (func == VPUTX_VUNREF)
 2703                 ASSERT_VOP_LOCKED(vp, "vunref");
 2704         else if (func == VPUTX_VPUT)
 2705                 ASSERT_VOP_LOCKED(vp, "vput");
 2706         else
 2707                 KASSERT(func == VPUTX_VRELE, ("vputx: wrong func"));
 2708         ASSERT_VI_UNLOCKED(vp, __func__);
 2709         CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
 2710 
 2711         if (vp->v_type != VCHR &&
 2712             refcount_release_if_not_last(&vp->v_usecount)) {
 2713                 if (func == VPUTX_VPUT)
 2714                         VOP_UNLOCK(vp, 0);
 2715                 vdrop(vp);
 2716                 return;
 2717         }
 2718 
 2719         VI_LOCK(vp);
 2720 
 2721         /*
 2722          * We want to hold the vnode until the inactive finishes to
 2723          * prevent vgone() races.  We drop the use count here and the
 2724          * hold count below when we're done.
 2725          */
 2726         if (!refcount_release(&vp->v_usecount) ||
 2727             (vp->v_iflag & VI_DOINGINACT)) {
 2728                 if (func == VPUTX_VPUT)
 2729                         VOP_UNLOCK(vp, 0);
 2730                 v_decr_devcount(vp);
 2731                 vdropl(vp);
 2732                 return;
 2733         }
 2734 
 2735         v_decr_devcount(vp);
 2736 
 2737         error = 0;
 2738 
 2739         if (vp->v_usecount != 0) {
 2740                 vn_printf(vp, "vputx: usecount not zero for vnode ");
 2741                 panic("vputx: usecount not zero");
 2742         }
 2743 
 2744         CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp);
 2745 
 2746         /*
 2747          * We must call VOP_INACTIVE with the node locked. Mark
 2748          * as VI_DOINGINACT to avoid recursion.
 2749          */
 2750         vp->v_iflag |= VI_OWEINACT;
 2751         switch (func) {
 2752         case VPUTX_VRELE:
 2753                 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
 2754                 VI_LOCK(vp);
 2755                 break;
 2756         case VPUTX_VPUT:
 2757                 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
 2758                         error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
 2759                             LK_NOWAIT);
 2760                         VI_LOCK(vp);
 2761                 }
 2762                 break;
 2763         case VPUTX_VUNREF:
 2764                 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
 2765                         error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
 2766                         VI_LOCK(vp);
 2767                 }
 2768                 break;
 2769         }
 2770         VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
 2771             ("vnode with usecount and VI_OWEINACT set"));
 2772         if (error == 0) {
 2773                 if (vp->v_iflag & VI_OWEINACT)
 2774                         vinactive(vp, curthread);
 2775                 if (func != VPUTX_VUNREF)
 2776                         VOP_UNLOCK(vp, 0);
 2777         }
 2778         vdropl(vp);
 2779 }
 2780 
 2781 /*
 2782  * Vnode put/release.
 2783  * If count drops to zero, call inactive routine and return to freelist.
 2784  */
 2785 void
 2786 vrele(struct vnode *vp)
 2787 {
 2788 
 2789         vputx(vp, VPUTX_VRELE);
 2790 }
 2791 
 2792 /*
 2793  * Release an already locked vnode.  This give the same effects as
 2794  * unlock+vrele(), but takes less time and avoids releasing and
 2795  * re-aquiring the lock (as vrele() acquires the lock internally.)
 2796  */
 2797 void
 2798 vput(struct vnode *vp)
 2799 {
 2800 
 2801         vputx(vp, VPUTX_VPUT);
 2802 }
 2803 
 2804 /*
 2805  * Release an exclusively locked vnode. Do not unlock the vnode lock.
 2806  */
 2807 void
 2808 vunref(struct vnode *vp)
 2809 {
 2810 
 2811         vputx(vp, VPUTX_VUNREF);
 2812 }
 2813 
 2814 /*
 2815  * Increase the hold count and activate if this is the first reference.
 2816  */
 2817 void
 2818 _vhold(struct vnode *vp, bool locked)
 2819 {
 2820         struct mount *mp;
 2821 
 2822         if (locked)
 2823                 ASSERT_VI_LOCKED(vp, __func__);
 2824         else
 2825                 ASSERT_VI_UNLOCKED(vp, __func__);
 2826         CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
 2827         if (!locked) {
 2828                 if (refcount_acquire_if_not_zero(&vp->v_holdcnt)) {
 2829                         VNODE_REFCOUNT_FENCE_ACQ();
 2830                         VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
 2831                             ("_vhold: vnode with holdcnt is free"));
 2832                         return;
 2833                 }
 2834                 VI_LOCK(vp);
 2835         }
 2836         if ((vp->v_iflag & VI_FREE) == 0) {
 2837                 refcount_acquire(&vp->v_holdcnt);
 2838                 if (!locked)
 2839                         VI_UNLOCK(vp);
 2840                 return;
 2841         }
 2842         VNASSERT(vp->v_holdcnt == 0, vp,
 2843             ("%s: wrong hold count", __func__));
 2844         VNASSERT(vp->v_op != NULL, vp,
 2845             ("%s: vnode already reclaimed.", __func__));
 2846         /*
 2847          * Remove a vnode from the free list, mark it as in use,
 2848          * and put it on the active list.
 2849          */
 2850         VNASSERT(vp->v_mount != NULL, vp,
 2851             ("_vhold: vnode not on per mount vnode list"));
 2852         mp = vp->v_mount;
 2853         mtx_lock(&mp->mnt_listmtx);
 2854         if ((vp->v_mflag & VMP_TMPMNTFREELIST) != 0) {
 2855                 TAILQ_REMOVE(&mp->mnt_tmpfreevnodelist, vp, v_actfreelist);
 2856                 mp->mnt_tmpfreevnodelistsize--;
 2857                 vp->v_mflag &= ~VMP_TMPMNTFREELIST;
 2858         } else {
 2859                 mtx_lock(&vnode_free_list_mtx);
 2860                 TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
 2861                 freevnodes--;
 2862                 mtx_unlock(&vnode_free_list_mtx);
 2863         }
 2864         KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
 2865             ("Activating already active vnode"));
 2866         vp->v_iflag &= ~VI_FREE;
 2867         vp->v_iflag |= VI_ACTIVE;
 2868         TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
 2869         mp->mnt_activevnodelistsize++;
 2870         mtx_unlock(&mp->mnt_listmtx);
 2871         refcount_acquire(&vp->v_holdcnt);
 2872         if (!locked)
 2873                 VI_UNLOCK(vp);
 2874 }
 2875 
 2876 /*
 2877  * Drop the hold count of the vnode.  If this is the last reference to
 2878  * the vnode we place it on the free list unless it has been vgone'd
 2879  * (marked VI_DOOMED) in which case we will free it.
 2880  *
 2881  * Because the vnode vm object keeps a hold reference on the vnode if
 2882  * there is at least one resident non-cached page, the vnode cannot
 2883  * leave the active list without the page cleanup done.
 2884  */
 2885 void
 2886 _vdrop(struct vnode *vp, bool locked)
 2887 {
 2888         struct bufobj *bo;
 2889         struct mount *mp;
 2890         int active;
 2891 
 2892         if (locked)
 2893                 ASSERT_VI_LOCKED(vp, __func__);
 2894         else
 2895                 ASSERT_VI_UNLOCKED(vp, __func__);
 2896         CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
 2897         if ((int)vp->v_holdcnt <= 0)
 2898                 panic("vdrop: holdcnt %d", vp->v_holdcnt);
 2899         if (!locked) {
 2900                 if (refcount_release_if_not_last(&vp->v_holdcnt))
 2901                         return;
 2902                 VI_LOCK(vp);
 2903         }
 2904         if (refcount_release(&vp->v_holdcnt) == 0) {
 2905                 VI_UNLOCK(vp);
 2906                 return;
 2907         }
 2908         if ((vp->v_iflag & VI_DOOMED) == 0) {
 2909                 /*
 2910                  * Mark a vnode as free: remove it from its active list
 2911                  * and put it up for recycling on the freelist.
 2912                  */
 2913                 VNASSERT(vp->v_op != NULL, vp,
 2914                     ("vdropl: vnode already reclaimed."));
 2915                 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
 2916                     ("vnode already free"));
 2917                 VNASSERT(vp->v_holdcnt == 0, vp,
 2918                     ("vdropl: freeing when we shouldn't"));
 2919                 active = vp->v_iflag & VI_ACTIVE;
 2920                 if ((vp->v_iflag & VI_OWEINACT) == 0) {
 2921                         vp->v_iflag &= ~VI_ACTIVE;
 2922                         mp = vp->v_mount;
 2923                         if (mp != NULL) {
 2924                                 mtx_lock(&mp->mnt_listmtx);
 2925                                 if (active) {
 2926                                         TAILQ_REMOVE(&mp->mnt_activevnodelist,
 2927                                             vp, v_actfreelist);
 2928                                         mp->mnt_activevnodelistsize--;
 2929                                 }
 2930                                 TAILQ_INSERT_TAIL(&mp->mnt_tmpfreevnodelist,
 2931                                     vp, v_actfreelist);
 2932                                 mp->mnt_tmpfreevnodelistsize++;
 2933                                 vp->v_iflag |= VI_FREE;
 2934                                 vp->v_mflag |= VMP_TMPMNTFREELIST;
 2935                                 VI_UNLOCK(vp);
 2936                                 if (mp->mnt_tmpfreevnodelistsize >=
 2937                                     mnt_free_list_batch)
 2938                                         vnlru_return_batch_locked(mp);
 2939                                 mtx_unlock(&mp->mnt_listmtx);
 2940                         } else {
 2941                                 VNASSERT(active == 0, vp,
 2942                                     ("vdropl: active vnode not on per mount "
 2943                                     "vnode list"));
 2944                                 mtx_lock(&vnode_free_list_mtx);
 2945                                 TAILQ_INSERT_TAIL(&vnode_free_list, vp,
 2946                                     v_actfreelist);
 2947                                 freevnodes++;
 2948                                 vp->v_iflag |= VI_FREE;
 2949                                 VI_UNLOCK(vp);
 2950                                 mtx_unlock(&vnode_free_list_mtx);
 2951                         }
 2952                 } else {
 2953                         VI_UNLOCK(vp);
 2954                         counter_u64_add(free_owe_inact, 1);
 2955                 }
 2956                 return;
 2957         }
 2958         /*
 2959          * The vnode has been marked for destruction, so free it.
 2960          *
 2961          * The vnode will be returned to the zone where it will
 2962          * normally remain until it is needed for another vnode. We
 2963          * need to cleanup (or verify that the cleanup has already
 2964          * been done) any residual data left from its current use
 2965          * so as not to contaminate the freshly allocated vnode.
 2966          */
 2967         CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
 2968         atomic_subtract_long(&numvnodes, 1);
 2969         bo = &vp->v_bufobj;
 2970         VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
 2971             ("cleaned vnode still on the free list."));
 2972         VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
 2973         VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
 2974         VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
 2975         VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
 2976         VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
 2977         VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
 2978         VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
 2979             ("clean blk trie not empty"));
 2980         VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
 2981         VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
 2982             ("dirty blk trie not empty"));
 2983         VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
 2984         VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
 2985         VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
 2986         VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
 2987             ("Dangling rangelock waiters"));
 2988         VI_UNLOCK(vp);
 2989 #ifdef MAC
 2990         mac_vnode_destroy(vp);
 2991 #endif
 2992         if (vp->v_pollinfo != NULL) {
 2993                 destroy_vpollinfo(vp->v_pollinfo);
 2994                 vp->v_pollinfo = NULL;
 2995         }
 2996 #ifdef INVARIANTS
 2997         /* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
 2998         vp->v_op = NULL;
 2999 #endif
 3000         vp->v_mountedhere = NULL;
 3001         vp->v_unpcb = NULL;
 3002         vp->v_rdev = NULL;
 3003         vp->v_fifoinfo = NULL;
 3004         vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
 3005         vp->v_iflag = 0;
 3006         vp->v_vflag = 0;
 3007         bo->bo_flag = 0;
 3008         uma_zfree(vnode_zone, vp);
 3009 }
 3010 
 3011 /*
 3012  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
 3013  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
 3014  * OWEINACT tracks whether a vnode missed a call to inactive due to a
 3015  * failed lock upgrade.
 3016  */
 3017 void
 3018 vinactive(struct vnode *vp, struct thread *td)
 3019 {
 3020         struct vm_object *obj;
 3021 
 3022         ASSERT_VOP_ELOCKED(vp, "vinactive");
 3023         ASSERT_VI_LOCKED(vp, "vinactive");
 3024         VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
 3025             ("vinactive: recursed on VI_DOINGINACT"));
 3026         CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
 3027         vp->v_iflag |= VI_DOINGINACT;
 3028         vp->v_iflag &= ~VI_OWEINACT;
 3029         VI_UNLOCK(vp);
 3030         /*
 3031          * Before moving off the active list, we must be sure that any
 3032          * modified pages are converted into the vnode's dirty
 3033          * buffers, since these will no longer be checked once the
 3034          * vnode is on the inactive list.
 3035          *
 3036          * The write-out of the dirty pages is asynchronous.  At the
 3037          * point that VOP_INACTIVE() is called, there could still be
 3038          * pending I/O and dirty pages in the object.
 3039          */
 3040         if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
 3041             (obj->flags & OBJ_MIGHTBEDIRTY) != 0) {
 3042                 VM_OBJECT_WLOCK(obj);
 3043                 vm_object_page_clean(obj, 0, 0, 0);
 3044                 VM_OBJECT_WUNLOCK(obj);
 3045         }
 3046         VOP_INACTIVE(vp, td);
 3047         VI_LOCK(vp);
 3048         VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
 3049             ("vinactive: lost VI_DOINGINACT"));
 3050         vp->v_iflag &= ~VI_DOINGINACT;
 3051 }
 3052 
 3053 /*
 3054  * Remove any vnodes in the vnode table belonging to mount point mp.
 3055  *
 3056  * If FORCECLOSE is not specified, there should not be any active ones,
 3057  * return error if any are found (nb: this is a user error, not a
 3058  * system error). If FORCECLOSE is specified, detach any active vnodes
 3059  * that are found.
 3060  *
 3061  * If WRITECLOSE is set, only flush out regular file vnodes open for
 3062  * writing.
 3063  *
 3064  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
 3065  *
 3066  * `rootrefs' specifies the base reference count for the root vnode
 3067  * of this filesystem. The root vnode is considered busy if its
 3068  * v_usecount exceeds this value. On a successful return, vflush(, td)
 3069  * will call vrele() on the root vnode exactly rootrefs times.
 3070  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
 3071  * be zero.
 3072  */
 3073 #ifdef DIAGNOSTIC
 3074 static int busyprt = 0;         /* print out busy vnodes */
 3075 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
 3076 #endif
 3077 
 3078 int
 3079 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
 3080 {
 3081         struct vnode *vp, *mvp, *rootvp = NULL;
 3082         struct vattr vattr;
 3083         int busy = 0, error;
 3084 
 3085         CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
 3086             rootrefs, flags);
 3087         if (rootrefs > 0) {
 3088                 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
 3089                     ("vflush: bad args"));
 3090                 /*
 3091                  * Get the filesystem root vnode. We can vput() it
 3092                  * immediately, since with rootrefs > 0, it won't go away.
 3093                  */
 3094                 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
 3095                         CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
 3096                             __func__, error);
 3097                         return (error);
 3098                 }
 3099                 vput(rootvp);
 3100         }
 3101 loop:
 3102         MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
 3103                 vholdl(vp);
 3104                 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
 3105                 if (error) {
 3106                         vdrop(vp);
 3107                         MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
 3108                         goto loop;
 3109                 }
 3110                 /*
 3111                  * Skip over a vnodes marked VV_SYSTEM.
 3112                  */
 3113                 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
 3114                         VOP_UNLOCK(vp, 0);
 3115                         vdrop(vp);
 3116                         continue;
 3117                 }
 3118                 /*
 3119                  * If WRITECLOSE is set, flush out unlinked but still open
 3120                  * files (even if open only for reading) and regular file
 3121                  * vnodes open for writing.
 3122                  */
 3123                 if (flags & WRITECLOSE) {
 3124                         if (vp->v_object != NULL) {
 3125                                 VM_OBJECT_WLOCK(vp->v_object);
 3126                                 vm_object_page_clean(vp->v_object, 0, 0, 0);
 3127                                 VM_OBJECT_WUNLOCK(vp->v_object);
 3128                         }
 3129                         error = VOP_FSYNC(vp, MNT_WAIT, td);
 3130                         if (error != 0) {
 3131                                 VOP_UNLOCK(vp, 0);
 3132                                 vdrop(vp);
 3133                                 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
 3134                                 return (error);
 3135                         }
 3136                         error = VOP_GETATTR(vp, &vattr, td->td_ucred);
 3137                         VI_LOCK(vp);
 3138 
 3139                         if ((vp->v_type == VNON ||
 3140                             (error == 0 && vattr.va_nlink > 0)) &&
 3141                             (vp->v_writecount == 0 || vp->v_type != VREG)) {
 3142                                 VOP_UNLOCK(vp, 0);
 3143                                 vdropl(vp);
 3144                                 continue;
 3145                         }
 3146                 } else
 3147                         VI_LOCK(vp);
 3148                 /*
 3149                  * With v_usecount == 0, all we need to do is clear out the
 3150                  * vnode data structures and we are done.
 3151                  *
 3152                  * If FORCECLOSE is set, forcibly close the vnode.
 3153                  */
 3154                 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
 3155                         vgonel(vp);
 3156                 } else {
 3157                         busy++;
 3158 #ifdef DIAGNOSTIC
 3159                         if (busyprt)
 3160                                 vn_printf(vp, "vflush: busy vnode ");
 3161 #endif
 3162                 }
 3163                 VOP_UNLOCK(vp, 0);
 3164                 vdropl(vp);
 3165         }
 3166         if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
 3167                 /*
 3168                  * If just the root vnode is busy, and if its refcount
 3169                  * is equal to `rootrefs', then go ahead and kill it.
 3170                  */
 3171                 VI_LOCK(rootvp);
 3172                 KASSERT(busy > 0, ("vflush: not busy"));
 3173                 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
 3174                     ("vflush: usecount %d < rootrefs %d",
 3175                      rootvp->v_usecount, rootrefs));
 3176                 if (busy == 1 && rootvp->v_usecount == rootrefs) {
 3177                         VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
 3178                         vgone(rootvp);
 3179                         VOP_UNLOCK(rootvp, 0);
 3180                         busy = 0;
 3181                 } else
 3182                         VI_UNLOCK(rootvp);
 3183         }
 3184         if (busy) {
 3185                 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
 3186                     busy);
 3187                 return (EBUSY);
 3188         }
 3189         for (; rootrefs > 0; rootrefs--)
 3190                 vrele(rootvp);
 3191         return (0);
 3192 }
 3193 
 3194 /*
 3195  * Recycle an unused vnode to the front of the free list.
 3196  */
 3197 int
 3198 vrecycle(struct vnode *vp)
 3199 {
 3200         int recycled;
 3201 
 3202         VI_LOCK(vp);
 3203         recycled = vrecyclel(vp);
 3204         VI_UNLOCK(vp);
 3205         return (recycled);
 3206 }
 3207 
 3208 /*
 3209  * vrecycle, with the vp interlock held.
 3210  */
 3211 int
 3212 vrecyclel(struct vnode *vp)
 3213 {
 3214         int recycled;
 3215 
 3216         ASSERT_VOP_ELOCKED(vp, __func__);
 3217         ASSERT_VI_LOCKED(vp, __func__);
 3218         CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
 3219         recycled = 0;
 3220         if (vp->v_usecount == 0) {
 3221                 recycled = 1;
 3222                 vgonel(vp);
 3223         }
 3224         return (recycled);
 3225 }
 3226 
 3227 /*
 3228  * Eliminate all activity associated with a vnode
 3229  * in preparation for reuse.
 3230  */
 3231 void
 3232 vgone(struct vnode *vp)
 3233 {
 3234         VI_LOCK(vp);
 3235         vgonel(vp);
 3236         VI_UNLOCK(vp);
 3237 }
 3238 
 3239 static void
 3240 notify_lowervp_vfs_dummy(struct mount *mp __unused,
 3241     struct vnode *lowervp __unused)
 3242 {
 3243 }
 3244 
 3245 /*
 3246  * Notify upper mounts about reclaimed or unlinked vnode.
 3247  */
 3248 void
 3249 vfs_notify_upper(struct vnode *vp, int event)
 3250 {
 3251         static struct vfsops vgonel_vfsops = {
 3252                 .vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
 3253                 .vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
 3254         };
 3255         struct mount *mp, *ump, *mmp;
 3256 
 3257         mp = vp->v_mount;
 3258         if (mp == NULL)
 3259                 return;
 3260 
 3261         MNT_ILOCK(mp);
 3262         if (TAILQ_EMPTY(&mp->mnt_uppers))
 3263                 goto unlock;
 3264         MNT_IUNLOCK(mp);
 3265         mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
 3266         mmp->mnt_op = &vgonel_vfsops;
 3267         mmp->mnt_kern_flag |= MNTK_MARKER;
 3268         MNT_ILOCK(mp);
 3269         mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
 3270         for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
 3271                 if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
 3272                         ump = TAILQ_NEXT(ump, mnt_upper_link);
 3273                         continue;
 3274                 }
 3275                 TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
 3276                 MNT_IUNLOCK(mp);
 3277                 switch (event) {
 3278                 case VFS_NOTIFY_UPPER_RECLAIM:
 3279                         VFS_RECLAIM_LOWERVP(ump, vp);
 3280                         break;
 3281                 case VFS_NOTIFY_UPPER_UNLINK:
 3282                         VFS_UNLINK_LOWERVP(ump, vp);
 3283                         break;
 3284                 default:
 3285                         KASSERT(0, ("invalid event %d", event));
 3286                         break;
 3287                 }
 3288                 MNT_ILOCK(mp);
 3289                 ump = TAILQ_NEXT(mmp, mnt_upper_link);
 3290                 TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
 3291         }
 3292         free(mmp, M_TEMP);
 3293         mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
 3294         if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
 3295                 mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
 3296                 wakeup(&mp->mnt_uppers);
 3297         }
 3298 unlock:
 3299         MNT_IUNLOCK(mp);
 3300 }
 3301 
 3302 /*
 3303  * vgone, with the vp interlock held.
 3304  */
 3305 static void
 3306 vgonel(struct vnode *vp)
 3307 {
 3308         struct thread *td;
 3309         int oweinact;
 3310         int active;
 3311         struct mount *mp;
 3312 
 3313         ASSERT_VOP_ELOCKED(vp, "vgonel");
 3314         ASSERT_VI_LOCKED(vp, "vgonel");
 3315         VNASSERT(vp->v_holdcnt, vp,
 3316             ("vgonel: vp %p has no reference.", vp));
 3317         CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
 3318         td = curthread;
 3319 
 3320         /*
 3321          * Don't vgonel if we're already doomed.
 3322          */
 3323         if (vp->v_iflag & VI_DOOMED)
 3324                 return;
 3325         vp->v_iflag |= VI_DOOMED;
 3326 
 3327         /*
 3328          * Check to see if the vnode is in use.  If so, we have to call
 3329          * VOP_CLOSE() and VOP_INACTIVE().
 3330          */
 3331         active = vp->v_usecount;
 3332         oweinact = (vp->v_iflag & VI_OWEINACT);
 3333         VI_UNLOCK(vp);
 3334         vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
 3335 
 3336         /*
 3337          * If purging an active vnode, it must be closed and
 3338          * deactivated before being reclaimed.
 3339          */
 3340         if (active)
 3341                 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
 3342         if (oweinact || active) {
 3343                 VI_LOCK(vp);
 3344                 if ((vp->v_iflag & VI_DOINGINACT) == 0)
 3345                         vinactive(vp, td);
 3346                 VI_UNLOCK(vp);
 3347         }
 3348         if (vp->v_type == VSOCK)
 3349                 vfs_unp_reclaim(vp);
 3350 
 3351         /*
 3352          * Clean out any buffers associated with the vnode.
 3353          * If the flush fails, just toss the buffers.
 3354          */
 3355         mp = NULL;
 3356         if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
 3357                 (void) vn_start_secondary_write(vp, &mp, V_WAIT);
 3358         if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
 3359                 while (vinvalbuf(vp, 0, 0, 0) != 0)
 3360                         ;
 3361         }
 3362 
 3363         BO_LOCK(&vp->v_bufobj);
 3364         KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
 3365             vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
 3366             TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
 3367             vp->v_bufobj.bo_clean.bv_cnt == 0,
 3368             ("vp %p bufobj not invalidated", vp));
 3369 
 3370         /*
 3371          * For VMIO bufobj, BO_DEAD is set in vm_object_terminate()
 3372          * after the object's page queue is flushed.
 3373          */
 3374         if (vp->v_bufobj.bo_object == NULL)
 3375                 vp->v_bufobj.bo_flag |= BO_DEAD;
 3376         BO_UNLOCK(&vp->v_bufobj);
 3377 
 3378         /*
 3379          * Reclaim the vnode.
 3380          */
 3381         if (VOP_RECLAIM(vp, td))
 3382                 panic("vgone: cannot reclaim");
 3383         if (mp != NULL)
 3384                 vn_finished_secondary_write(mp);
 3385         VNASSERT(vp->v_object == NULL, vp,
 3386             ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
 3387         /*
 3388          * Clear the advisory locks and wake up waiting threads.
 3389          */
 3390         (void)VOP_ADVLOCKPURGE(vp);
 3391         vp->v_lockf = NULL;
 3392         /*
 3393          * Delete from old mount point vnode list.
 3394          */
 3395         delmntque(vp);
 3396         cache_purge(vp);
 3397         /*
 3398          * Done with purge, reset to the standard lock and invalidate
 3399          * the vnode.
 3400          */
 3401         VI_LOCK(vp);
 3402         vp->v_vnlock = &vp->v_lock;
 3403         vp->v_op = &dead_vnodeops;
 3404         vp->v_tag = "none";
 3405         vp->v_type = VBAD;
 3406 }
 3407 
 3408 /*
 3409  * Calculate the total number of references to a special device.
 3410  */
 3411 int
 3412 vcount(struct vnode *vp)
 3413 {
 3414         int count;
 3415 
 3416         dev_lock();
 3417         count = vp->v_rdev->si_usecount;
 3418         dev_unlock();
 3419         return (count);
 3420 }
 3421 
 3422 /*
 3423  * Same as above, but using the struct cdev *as argument
 3424  */
 3425 int
 3426 count_dev(struct cdev *dev)
 3427 {
 3428         int count;
 3429 
 3430         dev_lock();
 3431         count = dev->si_usecount;
 3432         dev_unlock();
 3433         return(count);
 3434 }
 3435 
 3436 /*
 3437  * Print out a description of a vnode.
 3438  */
 3439 static char *typename[] =
 3440 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
 3441  "VMARKER"};
 3442 
 3443 void
 3444 vn_printf(struct vnode *vp, const char *fmt, ...)
 3445 {
 3446         va_list ap;
 3447         char buf[256], buf2[16];
 3448         u_long flags;
 3449 
 3450         va_start(ap, fmt);
 3451         vprintf(fmt, ap);
 3452         va_end(ap);
 3453         printf("%p: ", (void *)vp);
 3454         printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
 3455         printf("    usecount %d, writecount %d, refcount %d",
 3456             vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
 3457         switch (vp->v_type) {
 3458         case VDIR:
 3459                 printf(" mountedhere %p\n", vp->v_mountedhere);
 3460                 break;
 3461         case VCHR:
 3462                 printf(" rdev %p\n", vp->v_rdev);
 3463                 break;
 3464         case VSOCK:
 3465                 printf(" socket %p\n", vp->v_unpcb);
 3466                 break;
 3467         case VFIFO:
 3468                 printf(" fifoinfo %p\n", vp->v_fifoinfo);
 3469                 break;
 3470         default:
 3471                 printf("\n");
 3472                 break;
 3473         }
 3474         buf[0] = '\0';
 3475         buf[1] = '\0';
 3476         if (vp->v_vflag & VV_ROOT)
 3477                 strlcat(buf, "|VV_ROOT", sizeof(buf));
 3478         if (vp->v_vflag & VV_ISTTY)
 3479                 strlcat(buf, "|VV_ISTTY", sizeof(buf));
 3480         if (vp->v_vflag & VV_NOSYNC)
 3481                 strlcat(buf, "|VV_NOSYNC", sizeof(buf));
 3482         if (vp->v_vflag & VV_ETERNALDEV)
 3483                 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
 3484         if (vp->v_vflag & VV_CACHEDLABEL)
 3485                 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
 3486         if (vp->v_vflag & VV_TEXT)
 3487                 strlcat(buf, "|VV_TEXT", sizeof(buf));
 3488         if (vp->v_vflag & VV_COPYONWRITE)
 3489                 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
 3490         if (vp->v_vflag & VV_SYSTEM)
 3491                 strlcat(buf, "|VV_SYSTEM", sizeof(buf));
 3492         if (vp->v_vflag & VV_PROCDEP)
 3493                 strlcat(buf, "|VV_PROCDEP", sizeof(buf));
 3494         if (vp->v_vflag & VV_NOKNOTE)
 3495                 strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
 3496         if (vp->v_vflag & VV_DELETED)
 3497                 strlcat(buf, "|VV_DELETED", sizeof(buf));
 3498         if (vp->v_vflag & VV_MD)
 3499                 strlcat(buf, "|VV_MD", sizeof(buf));
 3500         if (vp->v_vflag & VV_FORCEINSMQ)
 3501                 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
 3502         flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
 3503             VV_CACHEDLABEL | VV_TEXT | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
 3504             VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
 3505         if (flags != 0) {
 3506                 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
 3507                 strlcat(buf, buf2, sizeof(buf));
 3508         }
 3509         if (vp->v_iflag & VI_MOUNT)
 3510                 strlcat(buf, "|VI_MOUNT", sizeof(buf));
 3511         if (vp->v_iflag & VI_DOOMED)
 3512                 strlcat(buf, "|VI_DOOMED", sizeof(buf));
 3513         if (vp->v_iflag & VI_FREE)
 3514                 strlcat(buf, "|VI_FREE", sizeof(buf));
 3515         if (vp->v_iflag & VI_ACTIVE)
 3516                 strlcat(buf, "|VI_ACTIVE", sizeof(buf));
 3517         if (vp->v_iflag & VI_DOINGINACT)
 3518                 strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
 3519         if (vp->v_iflag & VI_OWEINACT)
 3520                 strlcat(buf, "|VI_OWEINACT", sizeof(buf));
 3521         flags = vp->v_iflag & ~(VI_MOUNT | VI_DOOMED | VI_FREE |
 3522             VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT);
 3523         if (flags != 0) {
 3524                 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
 3525                 strlcat(buf, buf2, sizeof(buf));
 3526         }
 3527         printf("    flags (%s)\n", buf + 1);
 3528         if (mtx_owned(VI_MTX(vp)))
 3529                 printf(" VI_LOCKed");
 3530         if (vp->v_object != NULL)
 3531                 printf("    v_object %p ref %d pages %d "
 3532                     "cleanbuf %d dirtybuf %d\n",
 3533                     vp->v_object, vp->v_object->ref_count,
 3534                     vp->v_object->resident_page_count,
 3535                     vp->v_bufobj.bo_clean.bv_cnt,
 3536                     vp->v_bufobj.bo_dirty.bv_cnt);
 3537         printf("    ");
 3538         lockmgr_printinfo(vp->v_vnlock);
 3539         if (vp->v_data != NULL)
 3540                 VOP_PRINT(vp);
 3541 }
 3542 
 3543 #ifdef DDB
 3544 /*
 3545  * List all of the locked vnodes in the system.
 3546  * Called when debugging the kernel.
 3547  */
 3548 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
 3549 {
 3550         struct mount *mp;
 3551         struct vnode *vp;
 3552 
 3553         /*
 3554          * Note: because this is DDB, we can't obey the locking semantics
 3555          * for these structures, which means we could catch an inconsistent
 3556          * state and dereference a nasty pointer.  Not much to be done
 3557          * about that.
 3558          */
 3559         db_printf("Locked vnodes\n");
 3560         TAILQ_FOREACH(mp, &mountlist, mnt_list) {
 3561                 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
 3562                         if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
 3563                                 vn_printf(vp, "vnode ");
 3564                 }
 3565         }
 3566 }
 3567 
 3568 /*
 3569  * Show details about the given vnode.
 3570  */
 3571 DB_SHOW_COMMAND(vnode, db_show_vnode)
 3572 {
 3573         struct vnode *vp;
 3574 
 3575         if (!have_addr)
 3576                 return;
 3577         vp = (struct vnode *)addr;
 3578         vn_printf(vp, "vnode ");
 3579 }
 3580 
 3581 /*
 3582  * Show details about the given mount point.
 3583  */
 3584 DB_SHOW_COMMAND(mount, db_show_mount)
 3585 {
 3586         struct mount *mp;
 3587         struct vfsopt *opt;
 3588         struct statfs *sp;
 3589         struct vnode *vp;
 3590         char buf[512];
 3591         uint64_t mflags;
 3592         u_int flags;
 3593 
 3594         if (!have_addr) {
 3595                 /* No address given, print short info about all mount points. */
 3596                 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
 3597                         db_printf("%p %s on %s (%s)\n", mp,
 3598                             mp->mnt_stat.f_mntfromname,
 3599                             mp->mnt_stat.f_mntonname,
 3600                             mp->mnt_stat.f_fstypename);
 3601                         if (db_pager_quit)
 3602                                 break;
 3603                 }
 3604                 db_printf("\nMore info: show mount <addr>\n");
 3605                 return;
 3606         }
 3607 
 3608         mp = (struct mount *)addr;
 3609         db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
 3610             mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
 3611 
 3612         buf[0] = '\0';
 3613         mflags = mp->mnt_flag;
 3614 #define MNT_FLAG(flag)  do {                                            \
 3615         if (mflags & (flag)) {                                          \
 3616                 if (buf[0] != '\0')                                     \
 3617                         strlcat(buf, ", ", sizeof(buf));                \
 3618                 strlcat(buf, (#flag) + 4, sizeof(buf));                 \
 3619                 mflags &= ~(flag);                                      \
 3620         }                                                               \
 3621 } while (0)
 3622         MNT_FLAG(MNT_RDONLY);
 3623         MNT_FLAG(MNT_SYNCHRONOUS);
 3624         MNT_FLAG(MNT_NOEXEC);
 3625         MNT_FLAG(MNT_NOSUID);
 3626         MNT_FLAG(MNT_NFS4ACLS);
 3627         MNT_FLAG(MNT_UNION);
 3628         MNT_FLAG(MNT_ASYNC);
 3629         MNT_FLAG(MNT_SUIDDIR);
 3630         MNT_FLAG(MNT_SOFTDEP);
 3631         MNT_FLAG(MNT_NOSYMFOLLOW);
 3632         MNT_FLAG(MNT_GJOURNAL);
 3633         MNT_FLAG(MNT_MULTILABEL);
 3634         MNT_FLAG(MNT_ACLS);
 3635         MNT_FLAG(MNT_NOATIME);
 3636         MNT_FLAG(MNT_NOCLUSTERR);
 3637         MNT_FLAG(MNT_NOCLUSTERW);
 3638         MNT_FLAG(MNT_SUJ);
 3639         MNT_FLAG(MNT_EXRDONLY);
 3640         MNT_FLAG(MNT_EXPORTED);
 3641         MNT_FLAG(MNT_DEFEXPORTED);
 3642         MNT_FLAG(MNT_EXPORTANON);
 3643         MNT_FLAG(MNT_EXKERB);
 3644         MNT_FLAG(MNT_EXPUBLIC);
 3645         MNT_FLAG(MNT_LOCAL);
 3646         MNT_FLAG(MNT_QUOTA);
 3647         MNT_FLAG(MNT_ROOTFS);
 3648         MNT_FLAG(MNT_USER);
 3649         MNT_FLAG(MNT_IGNORE);
 3650         MNT_FLAG(MNT_UPDATE);
 3651         MNT_FLAG(MNT_DELEXPORT);
 3652         MNT_FLAG(MNT_RELOAD);
 3653         MNT_FLAG(MNT_FORCE);
 3654         MNT_FLAG(MNT_SNAPSHOT);
 3655         MNT_FLAG(MNT_BYFSID);
 3656 #undef MNT_FLAG
 3657         if (mflags != 0) {
 3658                 if (buf[0] != '\0')
 3659                         strlcat(buf, ", ", sizeof(buf));
 3660                 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
 3661                     "0x%016jx", mflags);
 3662         }
 3663         db_printf("    mnt_flag = %s\n", buf);
 3664 
 3665         buf[0] = '\0';
 3666         flags = mp->mnt_kern_flag;
 3667 #define MNT_KERN_FLAG(flag)     do {                                    \
 3668         if (flags & (flag)) {                                           \
 3669                 if (buf[0] != '\0')                                     \
 3670                         strlcat(buf, ", ", sizeof(buf));                \
 3671                 strlcat(buf, (#flag) + 5, sizeof(buf));                 \
 3672                 flags &= ~(flag);                                       \
 3673         }                                                               \
 3674 } while (0)
 3675         MNT_KERN_FLAG(MNTK_UNMOUNTF);
 3676         MNT_KERN_FLAG(MNTK_ASYNC);
 3677         MNT_KERN_FLAG(MNTK_SOFTDEP);
 3678         MNT_KERN_FLAG(MNTK_NOINSMNTQ);
 3679         MNT_KERN_FLAG(MNTK_DRAINING);
 3680         MNT_KERN_FLAG(MNTK_REFEXPIRE);
 3681         MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
 3682         MNT_KERN_FLAG(MNTK_SHARED_WRITES);
 3683         MNT_KERN_FLAG(MNTK_NO_IOPF);
 3684         MNT_KERN_FLAG(MNTK_VGONE_UPPER);
 3685         MNT_KERN_FLAG(MNTK_VGONE_WAITER);
 3686         MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
 3687         MNT_KERN_FLAG(MNTK_MARKER);
 3688         MNT_KERN_FLAG(MNTK_USES_BCACHE);
 3689         MNT_KERN_FLAG(MNTK_NOASYNC);
 3690         MNT_KERN_FLAG(MNTK_UNMOUNT);
 3691         MNT_KERN_FLAG(MNTK_MWAIT);
 3692         MNT_KERN_FLAG(MNTK_SUSPEND);
 3693         MNT_KERN_FLAG(MNTK_SUSPEND2);
 3694         MNT_KERN_FLAG(MNTK_SUSPENDED);
 3695         MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
 3696         MNT_KERN_FLAG(MNTK_NOKNOTE);
 3697 #undef MNT_KERN_FLAG
 3698         if (flags != 0) {
 3699                 if (buf[0] != '\0')
 3700                         strlcat(buf, ", ", sizeof(buf));
 3701                 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
 3702                     "0x%08x", flags);
 3703         }
 3704         db_printf("    mnt_kern_flag = %s\n", buf);
 3705 
 3706         db_printf("    mnt_opt = ");
 3707         opt = TAILQ_FIRST(mp->mnt_opt);
 3708         if (opt != NULL) {
 3709                 db_printf("%s", opt->name);
 3710                 opt = TAILQ_NEXT(opt, link);
 3711                 while (opt != NULL) {
 3712                         db_printf(", %s", opt->name);
 3713                         opt = TAILQ_NEXT(opt, link);
 3714                 }
 3715         }
 3716         db_printf("\n");
 3717 
 3718         sp = &mp->mnt_stat;
 3719         db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
 3720             "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
 3721             "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
 3722             "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
 3723             (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
 3724             (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
 3725             (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
 3726             (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
 3727             (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
 3728             (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
 3729             (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
 3730             (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
 3731 
 3732         db_printf("    mnt_cred = { uid=%u ruid=%u",
 3733             (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
 3734         if (jailed(mp->mnt_cred))
 3735                 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
 3736         db_printf(" }\n");
 3737         db_printf("    mnt_ref = %d\n", mp->mnt_ref);
 3738         db_printf("    mnt_gen = %d\n", mp->mnt_gen);
 3739         db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
 3740         db_printf("    mnt_activevnodelistsize = %d\n",
 3741             mp->mnt_activevnodelistsize);
 3742         db_printf("    mnt_writeopcount = %d\n", mp->mnt_writeopcount);
 3743         db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
 3744         db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
 3745         db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
 3746         db_printf("    mnt_lockref = %d\n", mp->mnt_lockref);
 3747         db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
 3748         db_printf("    mnt_secondary_accwrites = %d\n",
 3749             mp->mnt_secondary_accwrites);
 3750         db_printf("    mnt_gjprovider = %s\n",
 3751             mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
 3752 
 3753         db_printf("\n\nList of active vnodes\n");
 3754         TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) {
 3755                 if (vp->v_type != VMARKER) {
 3756                         vn_printf(vp, "vnode ");
 3757                         if (db_pager_quit)
 3758                                 break;
 3759                 }
 3760         }
 3761         db_printf("\n\nList of inactive vnodes\n");
 3762         TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
 3763                 if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) {
 3764                         vn_printf(vp, "vnode ");
 3765                         if (db_pager_quit)
 3766                                 break;
 3767                 }
 3768         }
 3769 }
 3770 #endif  /* DDB */
 3771 
 3772 /*
 3773  * Fill in a struct xvfsconf based on a struct vfsconf.
 3774  */
 3775 static int
 3776 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
 3777 {
 3778         struct xvfsconf xvfsp;
 3779 
 3780         bzero(&xvfsp, sizeof(xvfsp));
 3781         strcpy(xvfsp.vfc_name, vfsp->vfc_name);
 3782         xvfsp.vfc_typenum = vfsp->vfc_typenum;
 3783         xvfsp.vfc_refcount = vfsp->vfc_refcount;
 3784         xvfsp.vfc_flags = vfsp->vfc_flags;
 3785         /*
 3786          * These are unused in userland, we keep them
 3787          * to not break binary compatibility.
 3788          */
 3789         xvfsp.vfc_vfsops = NULL;
 3790         xvfsp.vfc_next = NULL;
 3791         return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
 3792 }
 3793 
 3794 #ifdef COMPAT_FREEBSD32
 3795 struct xvfsconf32 {
 3796         uint32_t        vfc_vfsops;
 3797         char            vfc_name[MFSNAMELEN];
 3798         int32_t         vfc_typenum;
 3799         int32_t         vfc_refcount;
 3800         int32_t         vfc_flags;
 3801         uint32_t        vfc_next;
 3802 };
 3803 
 3804 static int
 3805 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
 3806 {
 3807         struct xvfsconf32 xvfsp;
 3808 
 3809         bzero(&xvfsp, sizeof(xvfsp));
 3810         strcpy(xvfsp.vfc_name, vfsp->vfc_name);
 3811         xvfsp.vfc_typenum = vfsp->vfc_typenum;
 3812         xvfsp.vfc_refcount = vfsp->vfc_refcount;
 3813         xvfsp.vfc_flags = vfsp->vfc_flags;
 3814         return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
 3815 }
 3816 #endif
 3817 
 3818 /*
 3819  * Top level filesystem related information gathering.
 3820  */
 3821 static int
 3822 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
 3823 {
 3824         struct vfsconf *vfsp;
 3825         int error;
 3826 
 3827         error = 0;
 3828         vfsconf_slock();
 3829         TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
 3830 #ifdef COMPAT_FREEBSD32
 3831                 if (req->flags & SCTL_MASK32)
 3832                         error = vfsconf2x32(req, vfsp);
 3833                 else
 3834 #endif
 3835                         error = vfsconf2x(req, vfsp);
 3836                 if (error)
 3837                         break;
 3838         }
 3839         vfsconf_sunlock();
 3840         return (error);
 3841 }
 3842 
 3843 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
 3844     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
 3845     "S,xvfsconf", "List of all configured filesystems");
 3846 
 3847 #ifndef BURN_BRIDGES
 3848 static int      sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
 3849 
 3850 static int
 3851 vfs_sysctl(SYSCTL_HANDLER_ARGS)
 3852 {
 3853         int *name = (int *)arg1 - 1;    /* XXX */
 3854         u_int namelen = arg2 + 1;       /* XXX */
 3855         struct vfsconf *vfsp;
 3856 
 3857         log(LOG_WARNING, "userland calling deprecated sysctl, "
 3858             "please rebuild world\n");
 3859 
 3860 #if 1 || defined(COMPAT_PRELITE2)
 3861         /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
 3862         if (namelen == 1)
 3863                 return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
 3864 #endif
 3865 
 3866         switch (name[1]) {
 3867         case VFS_MAXTYPENUM:
 3868                 if (namelen != 2)
 3869                         return (ENOTDIR);
 3870                 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
 3871         case VFS_CONF:
 3872                 if (namelen != 3)
 3873                         return (ENOTDIR);       /* overloaded */
 3874                 vfsconf_slock();
 3875                 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
 3876                         if (vfsp->vfc_typenum == name[2])
 3877                                 break;
 3878                 }
 3879                 vfsconf_sunlock();
 3880                 if (vfsp == NULL)
 3881                         return (EOPNOTSUPP);
 3882 #ifdef COMPAT_FREEBSD32
 3883                 if (req->flags & SCTL_MASK32)
 3884                         return (vfsconf2x32(req, vfsp));
 3885                 else
 3886 #endif
 3887                         return (vfsconf2x(req, vfsp));
 3888         }
 3889         return (EOPNOTSUPP);
 3890 }
 3891 
 3892 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
 3893     CTLFLAG_MPSAFE, vfs_sysctl,
 3894     "Generic filesystem");
 3895 
 3896 #if 1 || defined(COMPAT_PRELITE2)
 3897 
 3898 static int
 3899 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
 3900 {
 3901         int error;
 3902         struct vfsconf *vfsp;
 3903         struct ovfsconf ovfs;
 3904 
 3905         vfsconf_slock();
 3906         TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
 3907                 bzero(&ovfs, sizeof(ovfs));
 3908                 ovfs.vfc_vfsops = vfsp->vfc_vfsops;     /* XXX used as flag */
 3909                 strcpy(ovfs.vfc_name, vfsp->vfc_name);
 3910                 ovfs.vfc_index = vfsp->vfc_typenum;
 3911                 ovfs.vfc_refcount = vfsp->vfc_refcount;
 3912                 ovfs.vfc_flags = vfsp->vfc_flags;
 3913                 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
 3914                 if (error != 0) {
 3915                         vfsconf_sunlock();
 3916                         return (error);
 3917                 }
 3918         }
 3919         vfsconf_sunlock();
 3920         return (0);
 3921 }
 3922 
 3923 #endif /* 1 || COMPAT_PRELITE2 */
 3924 #endif /* !BURN_BRIDGES */
 3925 
 3926 #define KINFO_VNODESLOP         10
 3927 #ifdef notyet
 3928 /*
 3929  * Dump vnode list (via sysctl).
 3930  */
 3931 /* ARGSUSED */
 3932 static int
 3933 sysctl_vnode(SYSCTL_HANDLER_ARGS)
 3934 {
 3935         struct xvnode *xvn;
 3936         struct mount *mp;
 3937         struct vnode *vp;
 3938         int error, len, n;
 3939 
 3940         /*
 3941          * Stale numvnodes access is not fatal here.
 3942          */
 3943         req->lock = 0;
 3944         len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
 3945         if (!req->oldptr)
 3946                 /* Make an estimate */
 3947                 return (SYSCTL_OUT(req, 0, len));
 3948 
 3949         error = sysctl_wire_old_buffer(req, 0);
 3950         if (error != 0)
 3951                 return (error);
 3952         xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
 3953         n = 0;
 3954         mtx_lock(&mountlist_mtx);
 3955         TAILQ_FOREACH(mp, &mountlist, mnt_list) {
 3956                 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
 3957                         continue;
 3958                 MNT_ILOCK(mp);
 3959                 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
 3960                         if (n == len)
 3961                                 break;
 3962                         vref(vp);
 3963                         xvn[n].xv_size = sizeof *xvn;
 3964                         xvn[n].xv_vnode = vp;
 3965                         xvn[n].xv_id = 0;       /* XXX compat */
 3966 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
 3967                         XV_COPY(usecount);
 3968                         XV_COPY(writecount);
 3969                         XV_COPY(holdcnt);
 3970                         XV_COPY(mount);
 3971                         XV_COPY(numoutput);
 3972                         XV_COPY(type);
 3973 #undef XV_COPY
 3974                         xvn[n].xv_flag = vp->v_vflag;
 3975 
 3976                         switch (vp->v_type) {
 3977                         case VREG:
 3978                         case VDIR:
 3979                         case VLNK:
 3980                                 break;
 3981                         case VBLK:
 3982                         case VCHR:
 3983                                 if (vp->v_rdev == NULL) {
 3984                                         vrele(vp);
 3985                                         continue;
 3986                                 }
 3987                                 xvn[n].xv_dev = dev2udev(vp->v_rdev);
 3988                                 break;
 3989                         case VSOCK:
 3990                                 xvn[n].xv_socket = vp->v_socket;
 3991                                 break;
 3992                         case VFIFO:
 3993                                 xvn[n].xv_fifo = vp->v_fifoinfo;
 3994                                 break;
 3995                         case VNON:
 3996                         case VBAD:
 3997                         default:
 3998                                 /* shouldn't happen? */
 3999                                 vrele(vp);
 4000                                 continue;
 4001                         }
 4002                         vrele(vp);
 4003                         ++n;
 4004                 }
 4005                 MNT_IUNLOCK(mp);
 4006                 mtx_lock(&mountlist_mtx);
 4007                 vfs_unbusy(mp);
 4008                 if (n == len)
 4009                         break;
 4010         }
 4011         mtx_unlock(&mountlist_mtx);
 4012 
 4013         error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
 4014         free(xvn, M_TEMP);
 4015         return (error);
 4016 }
 4017 
 4018 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
 4019     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
 4020     "");
 4021 #endif
 4022 
 4023 static void
 4024 unmount_or_warn(struct mount *mp)
 4025 {
 4026         int error;
 4027 
 4028         error = dounmount(mp, MNT_FORCE, curthread);
 4029         if (error != 0) {
 4030                 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
 4031                 if (error == EBUSY)
 4032                         printf("BUSY)\n");
 4033                 else
 4034                         printf("%d)\n", error);
 4035         }
 4036 }
 4037 
 4038 /*
 4039  * Unmount all filesystems. The list is traversed in reverse order
 4040  * of mounting to avoid dependencies.
 4041  */
 4042 void
 4043 vfs_unmountall(void)
 4044 {
 4045         struct mount *mp, *tmp;
 4046 
 4047         CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
 4048 
 4049         /*
 4050          * Since this only runs when rebooting, it is not interlocked.
 4051          */
 4052         TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
 4053                 vfs_ref(mp);
 4054 
 4055                 /*
 4056                  * Forcibly unmounting "/dev" before "/" would prevent clean
 4057                  * unmount of the latter.
 4058                  */
 4059                 if (mp == rootdevmp)
 4060                         continue;
 4061 
 4062                 unmount_or_warn(mp);
 4063         }
 4064 
 4065         if (rootdevmp != NULL)
 4066                 unmount_or_warn(rootdevmp);
 4067 }
 4068 
 4069 /*
 4070  * perform msync on all vnodes under a mount point
 4071  * the mount point must be locked.
 4072  */
 4073 void
 4074 vfs_msync(struct mount *mp, int flags)
 4075 {
 4076         struct vnode *vp, *mvp;
 4077         struct vm_object *obj;
 4078 
 4079         CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
 4080 
 4081         vnlru_return_batch(mp);
 4082 
 4083         MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
 4084                 obj = vp->v_object;
 4085                 if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 &&
 4086                     (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
 4087                         if (!vget(vp,
 4088                             LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
 4089                             curthread)) {
 4090                                 if (vp->v_vflag & VV_NOSYNC) {  /* unlinked */
 4091                                         vput(vp);
 4092                                         continue;
 4093                                 }
 4094 
 4095                                 obj = vp->v_object;
 4096                                 if (obj != NULL) {
 4097                                         VM_OBJECT_WLOCK(obj);
 4098                                         vm_object_page_clean(obj, 0, 0,
 4099                                             flags == MNT_WAIT ?
 4100                                             OBJPC_SYNC : OBJPC_NOSYNC);
 4101                                         VM_OBJECT_WUNLOCK(obj);
 4102                                 }
 4103                                 vput(vp);
 4104                         }
 4105                 } else
 4106                         VI_UNLOCK(vp);
 4107         }
 4108 }
 4109 
 4110 static void
 4111 destroy_vpollinfo_free(struct vpollinfo *vi)
 4112 {
 4113 
 4114         knlist_destroy(&vi->vpi_selinfo.si_note);
 4115         mtx_destroy(&vi->vpi_lock);
 4116         uma_zfree(vnodepoll_zone, vi);
 4117 }
 4118 
 4119 static void
 4120 destroy_vpollinfo(struct vpollinfo *vi)
 4121 {
 4122 
 4123         knlist_clear(&vi->vpi_selinfo.si_note, 1);
 4124         seldrain(&vi->vpi_selinfo);
 4125         destroy_vpollinfo_free(vi);
 4126 }
 4127 
 4128 /*
 4129  * Initialize per-vnode helper structure to hold poll-related state.
 4130  */
 4131 void
 4132 v_addpollinfo(struct vnode *vp)
 4133 {
 4134         struct vpollinfo *vi;
 4135 
 4136         if (vp->v_pollinfo != NULL)
 4137                 return;
 4138         vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
 4139         mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
 4140         knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
 4141             vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
 4142         VI_LOCK(vp);
 4143         if (vp->v_pollinfo != NULL) {
 4144                 VI_UNLOCK(vp);
 4145                 destroy_vpollinfo_free(vi);
 4146                 return;
 4147         }
 4148         vp->v_pollinfo = vi;
 4149         VI_UNLOCK(vp);
 4150 }
 4151 
 4152 /*
 4153  * Record a process's interest in events which might happen to
 4154  * a vnode.  Because poll uses the historic select-style interface
 4155  * internally, this routine serves as both the ``check for any
 4156  * pending events'' and the ``record my interest in future events''
 4157  * functions.  (These are done together, while the lock is held,
 4158  * to avoid race conditions.)
 4159  */
 4160 int
 4161 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
 4162 {
 4163 
 4164         v_addpollinfo(vp);
 4165         mtx_lock(&vp->v_pollinfo->vpi_lock);
 4166         if (vp->v_pollinfo->vpi_revents & events) {
 4167                 /*
 4168                  * This leaves events we are not interested
 4169                  * in available for the other process which
 4170                  * which presumably had requested them
 4171                  * (otherwise they would never have been
 4172                  * recorded).
 4173                  */
 4174                 events &= vp->v_pollinfo->vpi_revents;
 4175                 vp->v_pollinfo->vpi_revents &= ~events;
 4176 
 4177                 mtx_unlock(&vp->v_pollinfo->vpi_lock);
 4178                 return (events);
 4179         }
 4180         vp->v_pollinfo->vpi_events |= events;
 4181         selrecord(td, &vp->v_pollinfo->vpi_selinfo);
 4182         mtx_unlock(&vp->v_pollinfo->vpi_lock);
 4183         return (0);
 4184 }
 4185 
 4186 /*
 4187  * Routine to create and manage a filesystem syncer vnode.
 4188  */
 4189 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
 4190 static int      sync_fsync(struct  vop_fsync_args *);
 4191 static int      sync_inactive(struct  vop_inactive_args *);
 4192 static int      sync_reclaim(struct  vop_reclaim_args *);
 4193 
 4194 static struct vop_vector sync_vnodeops = {
 4195         .vop_bypass =   VOP_EOPNOTSUPP,
 4196         .vop_close =    sync_close,             /* close */
 4197         .vop_fsync =    sync_fsync,             /* fsync */
 4198         .vop_inactive = sync_inactive,  /* inactive */
 4199         .vop_reclaim =  sync_reclaim,   /* reclaim */
 4200         .vop_lock1 =    vop_stdlock,    /* lock */
 4201         .vop_unlock =   vop_stdunlock,  /* unlock */
 4202         .vop_islocked = vop_stdislocked,        /* islocked */
 4203 };
 4204 
 4205 /*
 4206  * Create a new filesystem syncer vnode for the specified mount point.
 4207  */
 4208 void
 4209 vfs_allocate_syncvnode(struct mount *mp)
 4210 {
 4211         struct vnode *vp;
 4212         struct bufobj *bo;
 4213         static long start, incr, next;
 4214         int error;
 4215 
 4216         /* Allocate a new vnode */
 4217         error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
 4218         if (error != 0)
 4219                 panic("vfs_allocate_syncvnode: getnewvnode() failed");
 4220         vp->v_type = VNON;
 4221         vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 4222         vp->v_vflag |= VV_FORCEINSMQ;
 4223         error = insmntque(vp, mp);
 4224         if (error != 0)
 4225                 panic("vfs_allocate_syncvnode: insmntque() failed");
 4226         vp->v_vflag &= ~VV_FORCEINSMQ;
 4227         VOP_UNLOCK(vp, 0);
 4228         /*
 4229          * Place the vnode onto the syncer worklist. We attempt to
 4230          * scatter them about on the list so that they will go off
 4231          * at evenly distributed times even if all the filesystems
 4232          * are mounted at once.
 4233          */
 4234         next += incr;
 4235         if (next == 0 || next > syncer_maxdelay) {
 4236                 start /= 2;
 4237                 incr /= 2;
 4238                 if (start == 0) {
 4239                         start = syncer_maxdelay / 2;
 4240                         incr = syncer_maxdelay;
 4241                 }
 4242                 next = start;
 4243         }
 4244         bo = &vp->v_bufobj;
 4245         BO_LOCK(bo);
 4246         vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
 4247         /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
 4248         mtx_lock(&sync_mtx);
 4249         sync_vnode_count++;
 4250         if (mp->mnt_syncer == NULL) {
 4251                 mp->mnt_syncer = vp;
 4252                 vp = NULL;
 4253         }
 4254         mtx_unlock(&sync_mtx);
 4255         BO_UNLOCK(bo);
 4256         if (vp != NULL) {
 4257                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 4258                 vgone(vp);
 4259                 vput(vp);
 4260         }
 4261 }
 4262 
 4263 void
 4264 vfs_deallocate_syncvnode(struct mount *mp)
 4265 {
 4266         struct vnode *vp;
 4267 
 4268         mtx_lock(&sync_mtx);
 4269         vp = mp->mnt_syncer;
 4270         if (vp != NULL)
 4271                 mp->mnt_syncer = NULL;
 4272         mtx_unlock(&sync_mtx);
 4273         if (vp != NULL)
 4274                 vrele(vp);
 4275 }
 4276 
 4277 /*
 4278  * Do a lazy sync of the filesystem.
 4279  */
 4280 static int
 4281 sync_fsync(struct vop_fsync_args *ap)
 4282 {
 4283         struct vnode *syncvp = ap->a_vp;
 4284         struct mount *mp = syncvp->v_mount;
 4285         int error, save;
 4286         struct bufobj *bo;
 4287 
 4288         /*
 4289          * We only need to do something if this is a lazy evaluation.
 4290          */
 4291         if (ap->a_waitfor != MNT_LAZY)
 4292                 return (0);
 4293 
 4294         /*
 4295          * Move ourselves to the back of the sync list.
 4296          */
 4297         bo = &syncvp->v_bufobj;
 4298         BO_LOCK(bo);
 4299         vn_syncer_add_to_worklist(bo, syncdelay);
 4300         BO_UNLOCK(bo);
 4301 
 4302         /*
 4303          * Walk the list of vnodes pushing all that are dirty and
 4304          * not already on the sync list.
 4305          */
 4306         if (vfs_busy(mp, MBF_NOWAIT) != 0)
 4307                 return (0);
 4308         if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
 4309                 vfs_unbusy(mp);
 4310                 return (0);
 4311         }
 4312         save = curthread_pflags_set(TDP_SYNCIO);
 4313         vfs_msync(mp, MNT_NOWAIT);
 4314         error = VFS_SYNC(mp, MNT_LAZY);
 4315         curthread_pflags_restore(save);
 4316         vn_finished_write(mp);
 4317         vfs_unbusy(mp);
 4318         return (error);
 4319 }
 4320 
 4321 /*
 4322  * The syncer vnode is no referenced.
 4323  */
 4324 static int
 4325 sync_inactive(struct vop_inactive_args *ap)
 4326 {
 4327 
 4328         vgone(ap->a_vp);
 4329         return (0);
 4330 }
 4331 
 4332 /*
 4333  * The syncer vnode is no longer needed and is being decommissioned.
 4334  *
 4335  * Modifications to the worklist must be protected by sync_mtx.
 4336  */
 4337 static int
 4338 sync_reclaim(struct vop_reclaim_args *ap)
 4339 {
 4340         struct vnode *vp = ap->a_vp;
 4341         struct bufobj *bo;
 4342 
 4343         bo = &vp->v_bufobj;
 4344         BO_LOCK(bo);
 4345         mtx_lock(&sync_mtx);
 4346         if (vp->v_mount->mnt_syncer == vp)
 4347                 vp->v_mount->mnt_syncer = NULL;
 4348         if (bo->bo_flag & BO_ONWORKLST) {
 4349                 LIST_REMOVE(bo, bo_synclist);
 4350                 syncer_worklist_len--;
 4351                 sync_vnode_count--;
 4352                 bo->bo_flag &= ~BO_ONWORKLST;
 4353         }
 4354         mtx_unlock(&sync_mtx);
 4355         BO_UNLOCK(bo);
 4356 
 4357         return (0);
 4358 }
 4359 
 4360 /*
 4361  * Check if vnode represents a disk device
 4362  */
 4363 int
 4364 vn_isdisk(struct vnode *vp, int *errp)
 4365 {
 4366         int error;
 4367 
 4368         if (vp->v_type != VCHR) {
 4369                 error = ENOTBLK;
 4370                 goto out;
 4371         }
 4372         error = 0;
 4373         dev_lock();
 4374         if (vp->v_rdev == NULL)
 4375                 error = ENXIO;
 4376         else if (vp->v_rdev->si_devsw == NULL)
 4377                 error = ENXIO;
 4378         else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
 4379                 error = ENOTBLK;
 4380         dev_unlock();
 4381 out:
 4382         if (errp != NULL)
 4383                 *errp = error;
 4384         return (error == 0);
 4385 }
 4386 
 4387 /*
 4388  * Common filesystem object access control check routine.  Accepts a
 4389  * vnode's type, "mode", uid and gid, requested access mode, credentials,
 4390  * and optional call-by-reference privused argument allowing vaccess()
 4391  * to indicate to the caller whether privilege was used to satisfy the
 4392  * request (obsoleted).  Returns 0 on success, or an errno on failure.
 4393  */
 4394 int
 4395 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
 4396     accmode_t accmode, struct ucred *cred, int *privused)
 4397 {
 4398         accmode_t dac_granted;
 4399         accmode_t priv_granted;
 4400 
 4401         KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
 4402             ("invalid bit in accmode"));
 4403         KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
 4404             ("VAPPEND without VWRITE"));
 4405 
 4406         /*
 4407          * Look for a normal, non-privileged way to access the file/directory
 4408          * as requested.  If it exists, go with that.
 4409          */
 4410 
 4411         if (privused != NULL)
 4412                 *privused = 0;
 4413 
 4414         dac_granted = 0;
 4415 
 4416         /* Check the owner. */
 4417         if (cred->cr_uid == file_uid) {
 4418                 dac_granted |= VADMIN;
 4419                 if (file_mode & S_IXUSR)
 4420                         dac_granted |= VEXEC;
 4421                 if (file_mode & S_IRUSR)
 4422                         dac_granted |= VREAD;
 4423                 if (file_mode & S_IWUSR)
 4424                         dac_granted |= (VWRITE | VAPPEND);
 4425 
 4426                 if ((accmode & dac_granted) == accmode)
 4427                         return (0);
 4428 
 4429                 goto privcheck;
 4430         }
 4431 
 4432         /* Otherwise, check the groups (first match) */
 4433         if (groupmember(file_gid, cred)) {
 4434                 if (file_mode & S_IXGRP)
 4435                         dac_granted |= VEXEC;
 4436                 if (file_mode & S_IRGRP)
 4437                         dac_granted |= VREAD;
 4438                 if (file_mode & S_IWGRP)
 4439                         dac_granted |= (VWRITE | VAPPEND);
 4440 
 4441                 if ((accmode & dac_granted) == accmode)
 4442                         return (0);
 4443 
 4444                 goto privcheck;
 4445         }
 4446 
 4447         /* Otherwise, check everyone else. */
 4448         if (file_mode & S_IXOTH)
 4449                 dac_granted |= VEXEC;
 4450         if (file_mode & S_IROTH)
 4451                 dac_granted |= VREAD;
 4452         if (file_mode & S_IWOTH)
 4453                 dac_granted |= (VWRITE | VAPPEND);
 4454         if ((accmode & dac_granted) == accmode)
 4455                 return (0);
 4456 
 4457 privcheck:
 4458         /*
 4459          * Build a privilege mask to determine if the set of privileges
 4460          * satisfies the requirements when combined with the granted mask
 4461          * from above.  For each privilege, if the privilege is required,
 4462          * bitwise or the request type onto the priv_granted mask.
 4463          */
 4464         priv_granted = 0;
 4465 
 4466         if (type == VDIR) {
 4467                 /*
 4468                  * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
 4469                  * requests, instead of PRIV_VFS_EXEC.
 4470                  */
 4471                 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
 4472                     !priv_check_cred(cred, PRIV_VFS_LOOKUP, 0))
 4473                         priv_granted |= VEXEC;
 4474         } else {
 4475                 /*
 4476                  * Ensure that at least one execute bit is on. Otherwise,
 4477                  * a privileged user will always succeed, and we don't want
 4478                  * this to happen unless the file really is executable.
 4479                  */
 4480                 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
 4481                     (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
 4482                     !priv_check_cred(cred, PRIV_VFS_EXEC, 0))
 4483                         priv_granted |= VEXEC;
 4484         }
 4485 
 4486         if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
 4487             !priv_check_cred(cred, PRIV_VFS_READ, 0))
 4488                 priv_granted |= VREAD;
 4489 
 4490         if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
 4491             !priv_check_cred(cred, PRIV_VFS_WRITE, 0))
 4492                 priv_granted |= (VWRITE | VAPPEND);
 4493 
 4494         if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
 4495             !priv_check_cred(cred, PRIV_VFS_ADMIN, 0))
 4496                 priv_granted |= VADMIN;
 4497 
 4498         if ((accmode & (priv_granted | dac_granted)) == accmode) {
 4499                 /* XXX audit: privilege used */
 4500                 if (privused != NULL)
 4501                         *privused = 1;
 4502                 return (0);
 4503         }
 4504 
 4505         return ((accmode & VADMIN) ? EPERM : EACCES);
 4506 }
 4507 
 4508 /*
 4509  * Credential check based on process requesting service, and per-attribute
 4510  * permissions.
 4511  */
 4512 int
 4513 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
 4514     struct thread *td, accmode_t accmode)
 4515 {
 4516 
 4517         /*
 4518          * Kernel-invoked always succeeds.
 4519          */
 4520         if (cred == NOCRED)
 4521                 return (0);
 4522 
 4523         /*
 4524          * Do not allow privileged processes in jail to directly manipulate
 4525          * system attributes.
 4526          */
 4527         switch (attrnamespace) {
 4528         case EXTATTR_NAMESPACE_SYSTEM:
 4529                 /* Potentially should be: return (EPERM); */
 4530                 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0));
 4531         case EXTATTR_NAMESPACE_USER:
 4532                 return (VOP_ACCESS(vp, accmode, cred, td));
 4533         default:
 4534                 return (EPERM);
 4535         }
 4536 }
 4537 
 4538 #ifdef DEBUG_VFS_LOCKS
 4539 /*
 4540  * This only exists to suppress warnings from unlocked specfs accesses.  It is
 4541  * no longer ok to have an unlocked VFS.
 4542  */
 4543 #define IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL ||            \
 4544         (vp)->v_type == VCHR || (vp)->v_type == VBAD)
 4545 
 4546 int vfs_badlock_ddb = 1;        /* Drop into debugger on violation. */
 4547 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
 4548     "Drop into debugger on lock violation");
 4549 
 4550 int vfs_badlock_mutex = 1;      /* Check for interlock across VOPs. */
 4551 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
 4552     0, "Check for interlock across VOPs");
 4553 
 4554 int vfs_badlock_print = 1;      /* Print lock violations. */
 4555 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
 4556     0, "Print lock violations");
 4557 
 4558 int vfs_badlock_vnode = 1;      /* Print vnode details on lock violations. */
 4559 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
 4560     0, "Print vnode details on lock violations");
 4561 
 4562 #ifdef KDB
 4563 int vfs_badlock_backtrace = 1;  /* Print backtrace at lock violations. */
 4564 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
 4565     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
 4566 #endif
 4567 
 4568 static void
 4569 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
 4570 {
 4571 
 4572 #ifdef KDB
 4573         if (vfs_badlock_backtrace)
 4574                 kdb_backtrace();
 4575 #endif
 4576         if (vfs_badlock_vnode)
 4577                 vn_printf(vp, "vnode ");
 4578         if (vfs_badlock_print)
 4579                 printf("%s: %p %s\n", str, (void *)vp, msg);
 4580         if (vfs_badlock_ddb)
 4581                 kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
 4582 }
 4583 
 4584 void
 4585 assert_vi_locked(struct vnode *vp, const char *str)
 4586 {
 4587 
 4588         if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
 4589                 vfs_badlock("interlock is not locked but should be", str, vp);
 4590 }
 4591 
 4592 void
 4593 assert_vi_unlocked(struct vnode *vp, const char *str)
 4594 {
 4595 
 4596         if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
 4597                 vfs_badlock("interlock is locked but should not be", str, vp);
 4598 }
 4599 
 4600 void
 4601 assert_vop_locked(struct vnode *vp, const char *str)
 4602 {
 4603         int locked;
 4604 
 4605         if (!IGNORE_LOCK(vp)) {
 4606                 locked = VOP_ISLOCKED(vp);
 4607                 if (locked == 0 || locked == LK_EXCLOTHER)
 4608                         vfs_badlock("is not locked but should be", str, vp);
 4609         }
 4610 }
 4611 
 4612 void
 4613 assert_vop_unlocked(struct vnode *vp, const char *str)
 4614 {
 4615 
 4616         if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
 4617                 vfs_badlock("is locked but should not be", str, vp);
 4618 }
 4619 
 4620 void
 4621 assert_vop_elocked(struct vnode *vp, const char *str)
 4622 {
 4623 
 4624         if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
 4625                 vfs_badlock("is not exclusive locked but should be", str, vp);
 4626 }
 4627 #endif /* DEBUG_VFS_LOCKS */
 4628 
 4629 void
 4630 vop_rename_fail(struct vop_rename_args *ap)
 4631 {
 4632 
 4633         if (ap->a_tvp != NULL)
 4634                 vput(ap->a_tvp);
 4635         if (ap->a_tdvp == ap->a_tvp)
 4636                 vrele(ap->a_tdvp);
 4637         else
 4638                 vput(ap->a_tdvp);
 4639         vrele(ap->a_fdvp);
 4640         vrele(ap->a_fvp);
 4641 }
 4642 
 4643 void
 4644 vop_rename_pre(void *ap)
 4645 {
 4646         struct vop_rename_args *a = ap;
 4647 
 4648 #ifdef DEBUG_VFS_LOCKS
 4649         if (a->a_tvp)
 4650                 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
 4651         ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
 4652         ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
 4653         ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
 4654 
 4655         /* Check the source (from). */
 4656         if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
 4657             (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
 4658                 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
 4659         if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
 4660                 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
 4661 
 4662         /* Check the target. */
 4663         if (a->a_tvp)
 4664                 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
 4665         ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
 4666 #endif
 4667         if (a->a_tdvp != a->a_fdvp)
 4668                 vhold(a->a_fdvp);
 4669         if (a->a_tvp != a->a_fvp)
 4670                 vhold(a->a_fvp);
 4671         vhold(a->a_tdvp);
 4672         if (a->a_tvp)
 4673                 vhold(a->a_tvp);
 4674 }
 4675 
 4676 #ifdef DEBUG_VFS_LOCKS
 4677 void
 4678 vop_strategy_pre(void *ap)
 4679 {
 4680         struct vop_strategy_args *a;
 4681         struct buf *bp;
 4682 
 4683         a = ap;
 4684         bp = a->a_bp;
 4685 
 4686         /*
 4687          * Cluster ops lock their component buffers but not the IO container.
 4688          */
 4689         if ((bp->b_flags & B_CLUSTER) != 0)
 4690                 return;
 4691 
 4692         if (panicstr == NULL && !BUF_ISLOCKED(bp)) {
 4693                 if (vfs_badlock_print)
 4694                         printf(
 4695                             "VOP_STRATEGY: bp is not locked but should be\n");
 4696                 if (vfs_badlock_ddb)
 4697                         kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
 4698         }
 4699 }
 4700 
 4701 void
 4702 vop_lock_pre(void *ap)
 4703 {
 4704         struct vop_lock1_args *a = ap;
 4705 
 4706         if ((a->a_flags & LK_INTERLOCK) == 0)
 4707                 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
 4708         else
 4709                 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
 4710 }
 4711 
 4712 void
 4713 vop_lock_post(void *ap, int rc)
 4714 {
 4715         struct vop_lock1_args *a = ap;
 4716 
 4717         ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
 4718         if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
 4719                 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
 4720 }
 4721 
 4722 void
 4723 vop_unlock_pre(void *ap)
 4724 {
 4725         struct vop_unlock_args *a = ap;
 4726 
 4727         if (a->a_flags & LK_INTERLOCK)
 4728                 ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
 4729         ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
 4730 }
 4731 
 4732 void
 4733 vop_unlock_post(void *ap, int rc)
 4734 {
 4735         struct vop_unlock_args *a = ap;
 4736 
 4737         if (a->a_flags & LK_INTERLOCK)
 4738                 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
 4739 }
 4740 #endif
 4741 
 4742 void
 4743 vop_create_post(void *ap, int rc)
 4744 {
 4745         struct vop_create_args *a = ap;
 4746 
 4747         if (!rc)
 4748                 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
 4749 }
 4750 
 4751 void
 4752 vop_deleteextattr_post(void *ap, int rc)
 4753 {
 4754         struct vop_deleteextattr_args *a = ap;
 4755 
 4756         if (!rc)
 4757                 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
 4758 }
 4759 
 4760 void
 4761 vop_link_post(void *ap, int rc)
 4762 {
 4763         struct vop_link_args *a = ap;
 4764 
 4765         if (!rc) {
 4766                 VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
 4767                 VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
 4768         }
 4769 }
 4770 
 4771 void
 4772 vop_mkdir_post(void *ap, int rc)
 4773 {
 4774         struct vop_mkdir_args *a = ap;
 4775 
 4776         if (!rc)
 4777                 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
 4778 }
 4779 
 4780 void
 4781 vop_mknod_post(void *ap, int rc)
 4782 {
 4783         struct vop_mknod_args *a = ap;
 4784 
 4785         if (!rc)
 4786                 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
 4787 }
 4788 
 4789 void
 4790 vop_reclaim_post(void *ap, int rc)
 4791 {
 4792         struct vop_reclaim_args *a = ap;
 4793 
 4794         if (!rc)
 4795                 VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
 4796 }
 4797 
 4798 void
 4799 vop_remove_post(void *ap, int rc)
 4800 {
 4801         struct vop_remove_args *a = ap;
 4802 
 4803         if (!rc) {
 4804                 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
 4805                 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
 4806         }
 4807 }
 4808 
 4809 void
 4810 vop_rename_post(void *ap, int rc)
 4811 {
 4812         struct vop_rename_args *a = ap;
 4813         long hint;
 4814 
 4815         if (!rc) {
 4816                 hint = NOTE_WRITE;
 4817                 if (a->a_fdvp == a->a_tdvp) {
 4818                         if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
 4819                                 hint |= NOTE_LINK;
 4820                         VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
 4821                         VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
 4822                 } else {
 4823                         hint |= NOTE_EXTEND;
 4824                         if (a->a_fvp->v_type == VDIR)
 4825                                 hint |= NOTE_LINK;
 4826                         VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
 4827 
 4828                         if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
 4829                             a->a_tvp->v_type == VDIR)
 4830                                 hint &= ~NOTE_LINK;
 4831                         VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
 4832                 }
 4833 
 4834                 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
 4835                 if (a->a_tvp)
 4836                         VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
 4837         }
 4838         if (a->a_tdvp != a->a_fdvp)
 4839                 vdrop(a->a_fdvp);
 4840         if (a->a_tvp != a->a_fvp)
 4841                 vdrop(a->a_fvp);
 4842         vdrop(a->a_tdvp);
 4843         if (a->a_tvp)
 4844                 vdrop(a->a_tvp);
 4845 }
 4846 
 4847 void
 4848 vop_rmdir_post(void *ap, int rc)
 4849 {
 4850         struct vop_rmdir_args *a = ap;
 4851 
 4852         if (!rc) {
 4853                 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
 4854                 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
 4855         }
 4856 }
 4857 
 4858 void
 4859 vop_setattr_post(void *ap, int rc)
 4860 {
 4861         struct vop_setattr_args *a = ap;
 4862 
 4863         if (!rc)
 4864                 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
 4865 }
 4866 
 4867 void
 4868 vop_setextattr_post(void *ap, int rc)
 4869 {
 4870         struct vop_setextattr_args *a = ap;
 4871 
 4872         if (!rc)
 4873                 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
 4874 }
 4875 
 4876 void
 4877 vop_symlink_post(void *ap, int rc)
 4878 {
 4879         struct vop_symlink_args *a = ap;
 4880 
 4881         if (!rc)
 4882                 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
 4883 }
 4884 
 4885 void
 4886 vop_open_post(void *ap, int rc)
 4887 {
 4888         struct vop_open_args *a = ap;
 4889 
 4890         if (!rc)
 4891                 VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
 4892 }
 4893 
 4894 void
 4895 vop_close_post(void *ap, int rc)
 4896 {
 4897         struct vop_close_args *a = ap;
 4898 
 4899         if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
 4900             (a->a_vp->v_iflag & VI_DOOMED) == 0)) {
 4901                 VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
 4902                     NOTE_CLOSE_WRITE : NOTE_CLOSE);
 4903         }
 4904 }
 4905 
 4906 void
 4907 vop_read_post(void *ap, int rc)
 4908 {
 4909         struct vop_read_args *a = ap;
 4910 
 4911         if (!rc)
 4912                 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
 4913 }
 4914 
 4915 void
 4916 vop_readdir_post(void *ap, int rc)
 4917 {
 4918         struct vop_readdir_args *a = ap;
 4919 
 4920         if (!rc)
 4921                 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
 4922 }
 4923 
 4924 static struct knlist fs_knlist;
 4925 
 4926 static void
 4927 vfs_event_init(void *arg)
 4928 {
 4929         knlist_init_mtx(&fs_knlist, NULL);
 4930 }
 4931 /* XXX - correct order? */
 4932 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
 4933 
 4934 void
 4935 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
 4936 {
 4937 
 4938         KNOTE_UNLOCKED(&fs_knlist, event);
 4939 }
 4940 
 4941 static int      filt_fsattach(struct knote *kn);
 4942 static void     filt_fsdetach(struct knote *kn);
 4943 static int      filt_fsevent(struct knote *kn, long hint);
 4944 
 4945 struct filterops fs_filtops = {
 4946         .f_isfd = 0,
 4947         .f_attach = filt_fsattach,
 4948         .f_detach = filt_fsdetach,
 4949         .f_event = filt_fsevent
 4950 };
 4951 
 4952 static int
 4953 filt_fsattach(struct knote *kn)
 4954 {
 4955 
 4956         kn->kn_flags |= EV_CLEAR;
 4957         knlist_add(&fs_knlist, kn, 0);
 4958         return (0);
 4959 }
 4960 
 4961 static void
 4962 filt_fsdetach(struct knote *kn)
 4963 {
 4964 
 4965         knlist_remove(&fs_knlist, kn, 0);
 4966 }
 4967 
 4968 static int
 4969 filt_fsevent(struct knote *kn, long hint)
 4970 {
 4971 
 4972         kn->kn_fflags |= hint;
 4973         return (kn->kn_fflags != 0);
 4974 }
 4975 
 4976 static int
 4977 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
 4978 {
 4979         struct vfsidctl vc;
 4980         int error;
 4981         struct mount *mp;
 4982 
 4983         error = SYSCTL_IN(req, &vc, sizeof(vc));
 4984         if (error)
 4985                 return (error);
 4986         if (vc.vc_vers != VFS_CTL_VERS1)
 4987                 return (EINVAL);
 4988         mp = vfs_getvfs(&vc.vc_fsid);
 4989         if (mp == NULL)
 4990                 return (ENOENT);
 4991         /* ensure that a specific sysctl goes to the right filesystem. */
 4992         if (strcmp(vc.vc_fstypename, "*") != 0 &&
 4993             strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
 4994                 vfs_rel(mp);
 4995                 return (EINVAL);
 4996         }
 4997         VCTLTOREQ(&vc, req);
 4998         error = VFS_SYSCTL(mp, vc.vc_op, req);
 4999         vfs_rel(mp);
 5000         return (error);
 5001 }
 5002 
 5003 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR,
 5004     NULL, 0, sysctl_vfs_ctl, "",
 5005     "Sysctl by fsid");
 5006 
 5007 /*
 5008  * Function to initialize a va_filerev field sensibly.
 5009  * XXX: Wouldn't a random number make a lot more sense ??
 5010  */
 5011 u_quad_t
 5012 init_va_filerev(void)
 5013 {
 5014         struct bintime bt;
 5015 
 5016         getbinuptime(&bt);
 5017         return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
 5018 }
 5019 
 5020 static int      filt_vfsread(struct knote *kn, long hint);
 5021 static int      filt_vfswrite(struct knote *kn, long hint);
 5022 static int      filt_vfsvnode(struct knote *kn, long hint);
 5023 static void     filt_vfsdetach(struct knote *kn);
 5024 static struct filterops vfsread_filtops = {
 5025         .f_isfd = 1,
 5026         .f_detach = filt_vfsdetach,
 5027         .f_event = filt_vfsread
 5028 };
 5029 static struct filterops vfswrite_filtops = {
 5030         .f_isfd = 1,
 5031         .f_detach = filt_vfsdetach,
 5032         .f_event = filt_vfswrite
 5033 };
 5034 static struct filterops vfsvnode_filtops = {
 5035         .f_isfd = 1,
 5036         .f_detach = filt_vfsdetach,
 5037         .f_event = filt_vfsvnode
 5038 };
 5039 
 5040 static void
 5041 vfs_knllock(void *arg)
 5042 {
 5043         struct vnode *vp = arg;
 5044 
 5045         vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
 5046 }
 5047 
 5048 static void
 5049 vfs_knlunlock(void *arg)
 5050 {
 5051         struct vnode *vp = arg;
 5052 
 5053         VOP_UNLOCK(vp, 0);
 5054 }
 5055 
 5056 static void
 5057 vfs_knl_assert_locked(void *arg)
 5058 {
 5059 #ifdef DEBUG_VFS_LOCKS
 5060         struct vnode *vp = arg;
 5061 
 5062         ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
 5063 #endif
 5064 }
 5065 
 5066 static void
 5067 vfs_knl_assert_unlocked(void *arg)
 5068 {
 5069 #ifdef DEBUG_VFS_LOCKS
 5070         struct vnode *vp = arg;
 5071 
 5072         ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
 5073 #endif
 5074 }
 5075 
 5076 int
 5077 vfs_kqfilter(struct vop_kqfilter_args *ap)
 5078 {
 5079         struct vnode *vp = ap->a_vp;
 5080         struct knote *kn = ap->a_kn;
 5081         struct knlist *knl;
 5082 
 5083         switch (kn->kn_filter) {
 5084         case EVFILT_READ:
 5085                 kn->kn_fop = &vfsread_filtops;
 5086                 break;
 5087         case EVFILT_WRITE:
 5088                 kn->kn_fop = &vfswrite_filtops;
 5089                 break;
 5090         case EVFILT_VNODE:
 5091                 kn->kn_fop = &vfsvnode_filtops;
 5092                 break;
 5093         default:
 5094                 return (EINVAL);
 5095         }
 5096 
 5097         kn->kn_hook = (caddr_t)vp;
 5098 
 5099         v_addpollinfo(vp);
 5100         if (vp->v_pollinfo == NULL)
 5101                 return (ENOMEM);
 5102         knl = &vp->v_pollinfo->vpi_selinfo.si_note;
 5103         vhold(vp);
 5104         knlist_add(knl, kn, 0);
 5105 
 5106         return (0);
 5107 }
 5108 
 5109 /*
 5110  * Detach knote from vnode
 5111  */
 5112 static void
 5113 filt_vfsdetach(struct knote *kn)
 5114 {
 5115         struct vnode *vp = (struct vnode *)kn->kn_hook;
 5116 
 5117         KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
 5118         knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
 5119         vdrop(vp);
 5120 }
 5121 
 5122 /*ARGSUSED*/
 5123 static int
 5124 filt_vfsread(struct knote *kn, long hint)
 5125 {
 5126         struct vnode *vp = (struct vnode *)kn->kn_hook;
 5127         struct vattr va;
 5128         int res;
 5129 
 5130         /*
 5131          * filesystem is gone, so set the EOF flag and schedule
 5132          * the knote for deletion.
 5133          */
 5134         if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
 5135                 VI_LOCK(vp);
 5136                 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
 5137                 VI_UNLOCK(vp);
 5138                 return (1);
 5139         }
 5140 
 5141         if (VOP_GETATTR(vp, &va, curthread->td_ucred))
 5142                 return (0);
 5143 
 5144         VI_LOCK(vp);
 5145         kn->kn_data = va.va_size - kn->kn_fp->f_offset;
 5146         res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
 5147         VI_UNLOCK(vp);
 5148         return (res);
 5149 }
 5150 
 5151 /*ARGSUSED*/
 5152 static int
 5153 filt_vfswrite(struct knote *kn, long hint)
 5154 {
 5155         struct vnode *vp = (struct vnode *)kn->kn_hook;
 5156 
 5157         VI_LOCK(vp);
 5158 
 5159         /*
 5160          * filesystem is gone, so set the EOF flag and schedule
 5161          * the knote for deletion.
 5162          */
 5163         if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
 5164                 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
 5165 
 5166         kn->kn_data = 0;
 5167         VI_UNLOCK(vp);
 5168         return (1);
 5169 }
 5170 
 5171 static int
 5172 filt_vfsvnode(struct knote *kn, long hint)
 5173 {
 5174         struct vnode *vp = (struct vnode *)kn->kn_hook;
 5175         int res;
 5176 
 5177         VI_LOCK(vp);
 5178         if (kn->kn_sfflags & hint)
 5179                 kn->kn_fflags |= hint;
 5180         if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
 5181                 kn->kn_flags |= EV_EOF;
 5182                 VI_UNLOCK(vp);
 5183                 return (1);
 5184         }
 5185         res = (kn->kn_fflags != 0);
 5186         VI_UNLOCK(vp);
 5187         return (res);
 5188 }
 5189 
 5190 int
 5191 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
 5192 {
 5193         int error;
 5194 
 5195         if (dp->d_reclen > ap->a_uio->uio_resid)
 5196                 return (ENAMETOOLONG);
 5197         error = uiomove(dp, dp->d_reclen, ap->a_uio);
 5198         if (error) {
 5199                 if (ap->a_ncookies != NULL) {
 5200                         if (ap->a_cookies != NULL)
 5201                                 free(ap->a_cookies, M_TEMP);
 5202                         ap->a_cookies = NULL;
 5203                         *ap->a_ncookies = 0;
 5204                 }
 5205                 return (error);
 5206         }
 5207         if (ap->a_ncookies == NULL)
 5208                 return (0);
 5209 
 5210         KASSERT(ap->a_cookies,
 5211             ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
 5212 
 5213         *ap->a_cookies = realloc(*ap->a_cookies,
 5214             (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
 5215         (*ap->a_cookies)[*ap->a_ncookies] = off;
 5216         *ap->a_ncookies += 1;
 5217         return (0);
 5218 }
 5219 
 5220 /*
 5221  * Mark for update the access time of the file if the filesystem
 5222  * supports VOP_MARKATIME.  This functionality is used by execve and
 5223  * mmap, so we want to avoid the I/O implied by directly setting
 5224  * va_atime for the sake of efficiency.
 5225  */
 5226 void
 5227 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
 5228 {
 5229         struct mount *mp;
 5230 
 5231         mp = vp->v_mount;
 5232         ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
 5233         if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
 5234                 (void)VOP_MARKATIME(vp);
 5235 }
 5236 
 5237 /*
 5238  * The purpose of this routine is to remove granularity from accmode_t,
 5239  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
 5240  * VADMIN and VAPPEND.
 5241  *
 5242  * If it returns 0, the caller is supposed to continue with the usual
 5243  * access checks using 'accmode' as modified by this routine.  If it
 5244  * returns nonzero value, the caller is supposed to return that value
 5245  * as errno.
 5246  *
 5247  * Note that after this routine runs, accmode may be zero.
 5248  */
 5249 int
 5250 vfs_unixify_accmode(accmode_t *accmode)
 5251 {
 5252         /*
 5253          * There is no way to specify explicit "deny" rule using
 5254          * file mode or POSIX.1e ACLs.
 5255          */
 5256         if (*accmode & VEXPLICIT_DENY) {
 5257                 *accmode = 0;
 5258                 return (0);
 5259         }
 5260 
 5261         /*
 5262          * None of these can be translated into usual access bits.
 5263          * Also, the common case for NFSv4 ACLs is to not contain
 5264          * either of these bits. Caller should check for VWRITE
 5265          * on the containing directory instead.
 5266          */
 5267         if (*accmode & (VDELETE_CHILD | VDELETE))
 5268                 return (EPERM);
 5269 
 5270         if (*accmode & VADMIN_PERMS) {
 5271                 *accmode &= ~VADMIN_PERMS;
 5272                 *accmode |= VADMIN;
 5273         }
 5274 
 5275         /*
 5276          * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
 5277          * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
 5278          */
 5279         *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
 5280 
 5281         return (0);
 5282 }
 5283 
 5284 /*
 5285  * These are helper functions for filesystems to traverse all
 5286  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
 5287  *
 5288  * This interface replaces MNT_VNODE_FOREACH.
 5289  */
 5290 
 5291 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
 5292 
 5293 struct vnode *
 5294 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
 5295 {
 5296         struct vnode *vp;
 5297 
 5298         if (should_yield())
 5299                 kern_yield(PRI_USER);
 5300         MNT_ILOCK(mp);
 5301         KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
 5302         for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
 5303             vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
 5304                 /* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
 5305                 if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
 5306                         continue;
 5307                 VI_LOCK(vp);
 5308                 if ((vp->v_iflag & VI_DOOMED) != 0) {
 5309                         VI_UNLOCK(vp);
 5310                         continue;
 5311                 }
 5312                 break;
 5313         }
 5314         if (vp == NULL) {
 5315                 __mnt_vnode_markerfree_all(mvp, mp);
 5316                 /* MNT_IUNLOCK(mp); -- done in above function */
 5317                 mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
 5318                 return (NULL);
 5319         }
 5320         TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
 5321         TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
 5322         MNT_IUNLOCK(mp);
 5323         return (vp);
 5324 }
 5325 
 5326 struct vnode *
 5327 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
 5328 {
 5329         struct vnode *vp;
 5330 
 5331         *mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
 5332         MNT_ILOCK(mp);
 5333         MNT_REF(mp);
 5334         (*mvp)->v_mount = mp;
 5335         (*mvp)->v_type = VMARKER;
 5336 
 5337         TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
 5338                 /* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
 5339                 if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
 5340                         continue;
 5341                 VI_LOCK(vp);
 5342                 if ((vp->v_iflag & VI_DOOMED) != 0) {
 5343                         VI_UNLOCK(vp);
 5344                         continue;
 5345                 }
 5346                 break;
 5347         }
 5348         if (vp == NULL) {
 5349                 MNT_REL(mp);
 5350                 MNT_IUNLOCK(mp);
 5351                 free(*mvp, M_VNODE_MARKER);
 5352                 *mvp = NULL;
 5353                 return (NULL);
 5354         }
 5355         TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
 5356         MNT_IUNLOCK(mp);
 5357         return (vp);
 5358 }
 5359 
 5360 void
 5361 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
 5362 {
 5363 
 5364         if (*mvp == NULL) {
 5365                 MNT_IUNLOCK(mp);
 5366                 return;
 5367         }
 5368 
 5369         mtx_assert(MNT_MTX(mp), MA_OWNED);
 5370 
 5371         KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
 5372         TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
 5373         MNT_REL(mp);
 5374         MNT_IUNLOCK(mp);
 5375         free(*mvp, M_VNODE_MARKER);
 5376         *mvp = NULL;
 5377 }
 5378 
 5379 /*
 5380  * These are helper functions for filesystems to traverse their
 5381  * active vnodes.  See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h
 5382  */
 5383 static void
 5384 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
 5385 {
 5386 
 5387         KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
 5388 
 5389         MNT_ILOCK(mp);
 5390         MNT_REL(mp);
 5391         MNT_IUNLOCK(mp);
 5392         free(*mvp, M_VNODE_MARKER);
 5393         *mvp = NULL;
 5394 }
 5395 
 5396 /*
 5397  * Relock the mp mount vnode list lock with the vp vnode interlock in the
 5398  * conventional lock order during mnt_vnode_next_active iteration.
 5399  *
 5400  * On entry, the mount vnode list lock is held and the vnode interlock is not.
 5401  * The list lock is dropped and reacquired.  On success, both locks are held.
 5402  * On failure, the mount vnode list lock is held but the vnode interlock is
 5403  * not, and the procedure may have yielded.
 5404  */
 5405 static bool
 5406 mnt_vnode_next_active_relock(struct vnode *mvp, struct mount *mp,
 5407     struct vnode *vp)
 5408 {
 5409         const struct vnode *tmp;
 5410         bool held, ret;
 5411 
 5412         VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
 5413             TAILQ_NEXT(mvp, v_actfreelist) != NULL, mvp,
 5414             ("%s: bad marker", __func__));
 5415         VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
 5416             ("%s: inappropriate vnode", __func__));
 5417         ASSERT_VI_UNLOCKED(vp, __func__);
 5418         mtx_assert(&mp->mnt_listmtx, MA_OWNED);
 5419 
 5420         ret = false;
 5421 
 5422         TAILQ_REMOVE(&mp->mnt_activevnodelist, mvp, v_actfreelist);
 5423         TAILQ_INSERT_BEFORE(vp, mvp, v_actfreelist);
 5424 
 5425         /*
 5426          * Use a hold to prevent vp from disappearing while the mount vnode
 5427          * list lock is dropped and reacquired.  Normally a hold would be
 5428          * acquired with vhold(), but that might try to acquire the vnode
 5429          * interlock, which would be a LOR with the mount vnode list lock.
 5430          */
 5431         held = refcount_acquire_if_not_zero(&vp->v_holdcnt);
 5432         mtx_unlock(&mp->mnt_listmtx);
 5433         if (!held)
 5434                 goto abort;
 5435         VI_LOCK(vp);
 5436         if (!refcount_release_if_not_last(&vp->v_holdcnt)) {
 5437                 vdropl(vp);
 5438                 goto abort;
 5439         }
 5440         mtx_lock(&mp->mnt_listmtx);
 5441 
 5442         /*
 5443          * Determine whether the vnode is still the next one after the marker,
 5444          * excepting any other markers.  If the vnode has not been doomed by
 5445          * vgone() then the hold should have ensured that it remained on the
 5446          * active list.  If it has been doomed but is still on the active list,
 5447          * don't abort, but rather skip over it (avoid spinning on doomed
 5448          * vnodes).
 5449          */
 5450         tmp = mvp;
 5451         do {
 5452                 tmp = TAILQ_NEXT(tmp, v_actfreelist);
 5453         } while (tmp != NULL && tmp->v_type == VMARKER);
 5454         if (tmp != vp) {
 5455                 mtx_unlock(&mp->mnt_listmtx);
 5456                 VI_UNLOCK(vp);
 5457                 goto abort;
 5458         }
 5459 
 5460         ret = true;
 5461         goto out;
 5462 abort:
 5463         maybe_yield();
 5464         mtx_lock(&mp->mnt_listmtx);
 5465 out:
 5466         if (ret)
 5467                 ASSERT_VI_LOCKED(vp, __func__);
 5468         else
 5469                 ASSERT_VI_UNLOCKED(vp, __func__);
 5470         mtx_assert(&mp->mnt_listmtx, MA_OWNED);
 5471         return (ret);
 5472 }
 5473 
 5474 static struct vnode *
 5475 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
 5476 {
 5477         struct vnode *vp, *nvp;
 5478 
 5479         mtx_assert(&mp->mnt_listmtx, MA_OWNED);
 5480         KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
 5481 restart:
 5482         vp = TAILQ_NEXT(*mvp, v_actfreelist);
 5483         while (vp != NULL) {
 5484                 if (vp->v_type == VMARKER) {
 5485                         vp = TAILQ_NEXT(vp, v_actfreelist);
 5486                         continue;
 5487                 }
 5488                 /*
 5489                  * Try-lock because this is the wrong lock order.  If that does
 5490                  * not succeed, drop the mount vnode list lock and try to
 5491                  * reacquire it and the vnode interlock in the right order.
 5492                  */
 5493                 if (!VI_TRYLOCK(vp) &&
 5494                     !mnt_vnode_next_active_relock(*mvp, mp, vp))
 5495                         goto restart;
 5496                 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
 5497                 KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
 5498                     ("alien vnode on the active list %p %p", vp, mp));
 5499                 if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0)
 5500                         break;
 5501                 nvp = TAILQ_NEXT(vp, v_actfreelist);
 5502                 VI_UNLOCK(vp);
 5503                 vp = nvp;
 5504         }
 5505         TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
 5506 
 5507         /* Check if we are done */
 5508         if (vp == NULL) {
 5509                 mtx_unlock(&mp->mnt_listmtx);
 5510                 mnt_vnode_markerfree_active(mvp, mp);
 5511                 return (NULL);
 5512         }
 5513         TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist);
 5514         mtx_unlock(&mp->mnt_listmtx);
 5515         ASSERT_VI_LOCKED(vp, "active iter");
 5516         KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp));
 5517         return (vp);
 5518 }
 5519 
 5520 struct vnode *
 5521 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
 5522 {
 5523 
 5524         if (should_yield())
 5525                 kern_yield(PRI_USER);
 5526         mtx_lock(&mp->mnt_listmtx);
 5527         return (mnt_vnode_next_active(mvp, mp));
 5528 }
 5529 
 5530 struct vnode *
 5531 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp)
 5532 {
 5533         struct vnode *vp;
 5534 
 5535         *mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
 5536         MNT_ILOCK(mp);
 5537         MNT_REF(mp);
 5538         MNT_IUNLOCK(mp);
 5539         (*mvp)->v_type = VMARKER;
 5540         (*mvp)->v_mount = mp;
 5541 
 5542         mtx_lock(&mp->mnt_listmtx);
 5543         vp = TAILQ_FIRST(&mp->mnt_activevnodelist);
 5544         if (vp == NULL) {
 5545                 mtx_unlock(&mp->mnt_listmtx);
 5546                 mnt_vnode_markerfree_active(mvp, mp);
 5547                 return (NULL);
 5548         }
 5549         TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
 5550         return (mnt_vnode_next_active(mvp, mp));
 5551 }
 5552 
 5553 void
 5554 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
 5555 {
 5556 
 5557         if (*mvp == NULL)
 5558                 return;
 5559 
 5560         mtx_lock(&mp->mnt_listmtx);
 5561         TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
 5562         mtx_unlock(&mp->mnt_listmtx);
 5563         mnt_vnode_markerfree_active(mvp, mp);
 5564 }

Cache object: f8bd1e3a692fbec170579da76ee5e697


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


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