2 * SPDX-License-Identifier: BSD-3-Clause
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.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
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.
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
36 * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95
40 * External virtual filesystem routines
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
47 #include "opt_watchdog.h"
49 #include <sys/param.h>
50 #include <sys/systm.h>
53 #include <sys/capsicum.h>
54 #include <sys/condvar.h>
56 #include <sys/counter.h>
57 #include <sys/dirent.h>
58 #include <sys/event.h>
59 #include <sys/eventhandler.h>
60 #include <sys/extattr.h>
62 #include <sys/fcntl.h>
65 #include <sys/kernel.h>
66 #include <sys/kthread.h>
67 #include <sys/lockf.h>
68 #include <sys/malloc.h>
69 #include <sys/mount.h>
70 #include <sys/namei.h>
71 #include <sys/pctrie.h>
73 #include <sys/reboot.h>
74 #include <sys/refcount.h>
75 #include <sys/rwlock.h>
76 #include <sys/sched.h>
77 #include <sys/sleepqueue.h>
80 #include <sys/sysctl.h>
81 #include <sys/syslog.h>
82 #include <sys/vmmeter.h>
83 #include <sys/vnode.h>
84 #include <sys/watchdog.h>
86 #include <machine/stdarg.h>
88 #include <security/mac/mac_framework.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_extern.h>
94 #include <vm/vm_map.h>
95 #include <vm/vm_page.h>
96 #include <vm/vm_kern.h>
103 static void delmntque(struct vnode *vp);
104 static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
105 int slpflag, int slptimeo);
106 static void syncer_shutdown(void *arg, int howto);
107 static int vtryrecycle(struct vnode *vp);
108 static void v_init_counters(struct vnode *);
109 static void v_incr_usecount(struct vnode *);
110 static void v_incr_usecount_locked(struct vnode *);
111 static void v_incr_devcount(struct vnode *);
112 static void v_decr_devcount(struct vnode *);
113 static void vgonel(struct vnode *);
114 static void vfs_knllock(void *arg);
115 static void vfs_knlunlock(void *arg);
116 static void vfs_knl_assert_locked(void *arg);
117 static void vfs_knl_assert_unlocked(void *arg);
118 static void vnlru_return_batches(struct vfsops *mnt_op);
119 static void destroy_vpollinfo(struct vpollinfo *vi);
120 static int v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
121 daddr_t startlbn, daddr_t endlbn);
124 * These fences are intended for cases where some synchronization is
125 * needed between access of v_iflags and lockless vnode refcount (v_holdcnt
126 * and v_usecount) updates. Access to v_iflags is generally synchronized
127 * by the interlock, but we have some internal assertions that check vnode
128 * flags without acquiring the lock. Thus, these fences are INVARIANTS-only
132 #define VNODE_REFCOUNT_FENCE_ACQ() atomic_thread_fence_acq()
133 #define VNODE_REFCOUNT_FENCE_REL() atomic_thread_fence_rel()
135 #define VNODE_REFCOUNT_FENCE_ACQ()
136 #define VNODE_REFCOUNT_FENCE_REL()
140 * Number of vnodes in existence. Increased whenever getnewvnode()
141 * allocates a new vnode, decreased in vdropl() for VI_DOOMED vnode.
143 static unsigned long numvnodes;
145 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
146 "Number of vnodes in existence");
148 static counter_u64_t vnodes_created;
149 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
150 "Number of vnodes created by getnewvnode");
152 static u_long mnt_free_list_batch = 128;
153 SYSCTL_ULONG(_vfs, OID_AUTO, mnt_free_list_batch, CTLFLAG_RW,
154 &mnt_free_list_batch, 0, "Limit of vnodes held on mnt's free list");
157 * Conversion tables for conversion from vnode types to inode formats
160 enum vtype iftovt_tab[16] = {
161 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
162 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
164 int vttoif_tab[10] = {
165 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
166 S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
170 * List of vnodes that are ready for recycling.
172 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
175 * "Free" vnode target. Free vnodes are rarely completely free, but are
176 * just ones that are cheap to recycle. Usually they are for files which
177 * have been stat'd but not read; these usually have inode and namecache
178 * data attached to them. This target is the preferred minimum size of a
179 * sub-cache consisting mostly of such files. The system balances the size
180 * of this sub-cache with its complement to try to prevent either from
181 * thrashing while the other is relatively inactive. The targets express
182 * a preference for the best balance.
184 * "Above" this target there are 2 further targets (watermarks) related
185 * to recyling of free vnodes. In the best-operating case, the cache is
186 * exactly full, the free list has size between vlowat and vhiwat above the
187 * free target, and recycling from it and normal use maintains this state.
188 * Sometimes the free list is below vlowat or even empty, but this state
189 * is even better for immediate use provided the cache is not full.
190 * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
191 * ones) to reach one of these states. The watermarks are currently hard-
192 * coded as 4% and 9% of the available space higher. These and the default
193 * of 25% for wantfreevnodes are too large if the memory size is large.
194 * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
195 * whenever vnlru_proc() becomes active.
197 static u_long wantfreevnodes;
198 SYSCTL_ULONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW,
199 &wantfreevnodes, 0, "Target for minimum number of \"free\" vnodes");
200 static u_long freevnodes;
201 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
202 &freevnodes, 0, "Number of \"free\" vnodes");
204 static counter_u64_t recycles_count;
205 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
206 "Number of vnodes recycled to meet vnode cache targets");
209 * Various variables used for debugging the new implementation of
211 * XXX these are probably of (very) limited utility now.
213 static int reassignbufcalls;
214 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0,
215 "Number of calls to reassignbuf");
217 static counter_u64_t free_owe_inact;
218 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, free_owe_inact, CTLFLAG_RD, &free_owe_inact,
219 "Number of times free vnodes kept on active list due to VFS "
220 "owing inactivation");
222 /* To keep more than one thread at a time from running vfs_getnewfsid */
223 static struct mtx mntid_mtx;
226 * Lock for any access to the following:
231 static struct mtx vnode_free_list_mtx;
233 /* Publicly exported FS */
234 struct nfs_public nfs_pub;
236 static uma_zone_t buf_trie_zone;
238 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
239 static uma_zone_t vnode_zone;
240 static uma_zone_t vnodepoll_zone;
243 * The workitem queue.
245 * It is useful to delay writes of file data and filesystem metadata
246 * for tens of seconds so that quickly created and deleted files need
247 * not waste disk bandwidth being created and removed. To realize this,
248 * we append vnodes to a "workitem" queue. When running with a soft
249 * updates implementation, most pending metadata dependencies should
250 * not wait for more than a few seconds. Thus, mounted on block devices
251 * are delayed only about a half the time that file data is delayed.
252 * Similarly, directory updates are more critical, so are only delayed
253 * about a third the time that file data is delayed. Thus, there are
254 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
255 * one each second (driven off the filesystem syncer process). The
256 * syncer_delayno variable indicates the next queue that is to be processed.
257 * Items that need to be processed soon are placed in this queue:
259 * syncer_workitem_pending[syncer_delayno]
261 * A delay of fifteen seconds is done by placing the request fifteen
262 * entries later in the queue:
264 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
267 static int syncer_delayno;
268 static long syncer_mask;
269 LIST_HEAD(synclist, bufobj);
270 static struct synclist *syncer_workitem_pending;
272 * The sync_mtx protects:
277 * syncer_workitem_pending
278 * syncer_worklist_len
281 static struct mtx sync_mtx;
282 static struct cv sync_wakeup;
284 #define SYNCER_MAXDELAY 32
285 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */
286 static int syncdelay = 30; /* max time to delay syncing data */
287 static int filedelay = 30; /* time to delay syncing files */
288 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
289 "Time to delay syncing files (in seconds)");
290 static int dirdelay = 29; /* time to delay syncing directories */
291 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
292 "Time to delay syncing directories (in seconds)");
293 static int metadelay = 28; /* time to delay syncing metadata */
294 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
295 "Time to delay syncing metadata (in seconds)");
296 static int rushjob; /* number of slots to run ASAP */
297 static int stat_rush_requests; /* number of times I/O speeded up */
298 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
299 "Number of times I/O speeded up (rush requests)");
302 * When shutting down the syncer, run it at four times normal speed.
304 #define SYNCER_SHUTDOWN_SPEEDUP 4
305 static int sync_vnode_count;
306 static int syncer_worklist_len;
307 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
310 /* Target for maximum number of vnodes. */
312 static int gapvnodes; /* gap between wanted and desired */
313 static int vhiwat; /* enough extras after expansion */
314 static int vlowat; /* minimal extras before expansion */
315 static int vstir; /* nonzero to stir non-free vnodes */
316 static volatile int vsmalltrigger = 8; /* pref to keep if > this many pages */
319 sysctl_update_desiredvnodes(SYSCTL_HANDLER_ARGS)
321 int error, old_desiredvnodes;
323 old_desiredvnodes = desiredvnodes;
324 if ((error = sysctl_handle_int(oidp, arg1, arg2, req)) != 0)
326 if (old_desiredvnodes != desiredvnodes) {
327 wantfreevnodes = desiredvnodes / 4;
328 /* XXX locking seems to be incomplete. */
329 vfs_hash_changesize(desiredvnodes);
330 cache_changesize(desiredvnodes);
335 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
336 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, &desiredvnodes, 0,
337 sysctl_update_desiredvnodes, "I", "Target for maximum number of vnodes");
338 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
339 &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
340 static int vnlru_nowhere;
341 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
342 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
345 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
350 unsigned long ndflags;
353 if (req->newptr == NULL)
355 if (req->newlen > PATH_MAX)
358 buf = malloc(PATH_MAX + 1, M_TEMP, M_WAITOK);
359 error = SYSCTL_IN(req, buf, req->newlen);
363 buf[req->newlen] = '\0';
365 ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | NOCACHE | SAVENAME;
366 NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread);
367 if ((error = namei(&nd)) != 0)
371 if ((vp->v_iflag & VI_DOOMED) != 0) {
373 * This vnode is being recycled. Return != 0 to let the caller
374 * know that the sysctl had no effect. Return EAGAIN because a
375 * subsequent call will likely succeed (since namei will create
376 * a new vnode if necessary)
382 counter_u64_add(recycles_count, 1);
392 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
394 struct thread *td = curthread;
400 if (req->newptr == NULL)
403 error = sysctl_handle_int(oidp, &fd, 0, req);
406 error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
411 error = vn_lock(vp, LK_EXCLUSIVE);
415 counter_u64_add(recycles_count, 1);
423 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
424 CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
425 sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
426 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
427 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
428 sysctl_ftry_reclaim_vnode, "I",
429 "Try to reclaim a vnode by its file descriptor");
431 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
435 * Support for the bufobj clean & dirty pctrie.
438 buf_trie_alloc(struct pctrie *ptree)
441 return uma_zalloc(buf_trie_zone, M_NOWAIT);
445 buf_trie_free(struct pctrie *ptree, void *node)
448 uma_zfree(buf_trie_zone, node);
450 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
453 * Initialize the vnode management data structures.
455 * Reevaluate the following cap on the number of vnodes after the physical
456 * memory size exceeds 512GB. In the limit, as the physical memory size
457 * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
459 #ifndef MAXVNODES_MAX
460 #define MAXVNODES_MAX (512 * 1024 * 1024 / 64) /* 8M */
464 * Initialize a vnode as it first enters the zone.
467 vnode_init(void *mem, int size, int flags)
476 vp->v_vnlock = &vp->v_lock;
477 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
479 * By default, don't allow shared locks unless filesystems opt-in.
481 lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
482 LK_NOSHARE | LK_IS_VNODE);
486 bufobj_init(&vp->v_bufobj, vp);
488 * Initialize namecache.
490 LIST_INIT(&vp->v_cache_src);
491 TAILQ_INIT(&vp->v_cache_dst);
493 * Initialize rangelocks.
495 rangelock_init(&vp->v_rl);
500 * Free a vnode when it is cleared from the zone.
503 vnode_fini(void *mem, int size)
509 rangelock_destroy(&vp->v_rl);
510 lockdestroy(vp->v_vnlock);
511 mtx_destroy(&vp->v_interlock);
513 rw_destroy(BO_LOCKPTR(bo));
517 * Provide the size of NFS nclnode and NFS fh for calculation of the
518 * vnode memory consumption. The size is specified directly to
519 * eliminate dependency on NFS-private header.
521 * Other filesystems may use bigger or smaller (like UFS and ZFS)
522 * private inode data, but the NFS-based estimation is ample enough.
523 * Still, we care about differences in the size between 64- and 32-bit
526 * Namecache structure size is heuristically
527 * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
530 #define NFS_NCLNODE_SZ (528 + 64)
533 #define NFS_NCLNODE_SZ (360 + 32)
538 vntblinit(void *dummy __unused)
541 int physvnodes, virtvnodes;
544 * Desiredvnodes is a function of the physical memory size and the
545 * kernel's heap size. Generally speaking, it scales with the
546 * physical memory size. The ratio of desiredvnodes to the physical
547 * memory size is 1:16 until desiredvnodes exceeds 98,304.
549 * marginal ratio of desiredvnodes to the physical memory size is
550 * 1:64. However, desiredvnodes is limited by the kernel's heap
551 * size. The memory required by desiredvnodes vnodes and vm objects
552 * must not exceed 1/10th of the kernel's heap size.
554 physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
555 3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
556 virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
557 sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
558 desiredvnodes = min(physvnodes, virtvnodes);
559 if (desiredvnodes > MAXVNODES_MAX) {
561 printf("Reducing kern.maxvnodes %d -> %d\n",
562 desiredvnodes, MAXVNODES_MAX);
563 desiredvnodes = MAXVNODES_MAX;
565 wantfreevnodes = desiredvnodes / 4;
566 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
567 TAILQ_INIT(&vnode_free_list);
568 mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
569 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
570 vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
571 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
572 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
574 * Preallocate enough nodes to support one-per buf so that
575 * we can not fail an insert. reassignbuf() callers can not
576 * tolerate the insertion failure.
578 buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
579 NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
580 UMA_ZONE_NOFREE | UMA_ZONE_VM);
581 uma_prealloc(buf_trie_zone, nbuf);
583 vnodes_created = counter_u64_alloc(M_WAITOK);
584 recycles_count = counter_u64_alloc(M_WAITOK);
585 free_owe_inact = counter_u64_alloc(M_WAITOK);
588 * Initialize the filesystem syncer.
590 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
592 syncer_maxdelay = syncer_mask + 1;
593 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
594 cv_init(&sync_wakeup, "syncer");
595 for (i = 1; i <= sizeof(struct vnode); i <<= 1)
599 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
603 * Mark a mount point as busy. Used to synchronize access and to delay
604 * unmounting. Eventually, mountlist_mtx is not released on failure.
606 * vfs_busy() is a custom lock, it can block the caller.
607 * vfs_busy() only sleeps if the unmount is active on the mount point.
608 * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
609 * vnode belonging to mp.
611 * Lookup uses vfs_busy() to traverse mount points.
613 * / vnode lock A / vnode lock (/var) D
614 * /var vnode lock B /log vnode lock(/var/log) E
615 * vfs_busy lock C vfs_busy lock F
617 * Within each file system, the lock order is C->A->B and F->D->E.
619 * When traversing across mounts, the system follows that lock order:
625 * The lookup() process for namei("/var") illustrates the process:
626 * VOP_LOOKUP() obtains B while A is held
627 * vfs_busy() obtains a shared lock on F while A and B are held
628 * vput() releases lock on B
629 * vput() releases lock on A
630 * VFS_ROOT() obtains lock on D while shared lock on F is held
631 * vfs_unbusy() releases shared lock on F
632 * vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
633 * Attempt to lock A (instead of vp_crossmp) while D is held would
634 * violate the global order, causing deadlocks.
636 * dounmount() locks B while F is drained.
639 vfs_busy(struct mount *mp, int flags)
642 MPASS((flags & ~MBF_MASK) == 0);
643 CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
648 * If mount point is currently being unmounted, sleep until the
649 * mount point fate is decided. If thread doing the unmounting fails,
650 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
651 * that this mount point has survived the unmount attempt and vfs_busy
652 * should retry. Otherwise the unmounter thread will set MNTK_REFEXPIRE
653 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
654 * about to be really destroyed. vfs_busy needs to release its
655 * reference on the mount point in this case and return with ENOENT,
656 * telling the caller that mount mount it tried to busy is no longer
659 while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
660 if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
663 CTR1(KTR_VFS, "%s: failed busying before sleeping",
667 if (flags & MBF_MNTLSTLOCK)
668 mtx_unlock(&mountlist_mtx);
669 mp->mnt_kern_flag |= MNTK_MWAIT;
670 msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
671 if (flags & MBF_MNTLSTLOCK)
672 mtx_lock(&mountlist_mtx);
675 if (flags & MBF_MNTLSTLOCK)
676 mtx_unlock(&mountlist_mtx);
683 * Free a busy filesystem.
686 vfs_unbusy(struct mount *mp)
689 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
692 KASSERT(mp->mnt_lockref > 0, ("negative mnt_lockref"));
694 if (mp->mnt_lockref == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
695 MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
696 CTR1(KTR_VFS, "%s: waking up waiters", __func__);
697 mp->mnt_kern_flag &= ~MNTK_DRAINING;
698 wakeup(&mp->mnt_lockref);
704 * Lookup a mount point by filesystem identifier.
707 vfs_getvfs(fsid_t *fsid)
711 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
712 mtx_lock(&mountlist_mtx);
713 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
714 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
715 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
717 mtx_unlock(&mountlist_mtx);
721 mtx_unlock(&mountlist_mtx);
722 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
723 return ((struct mount *) 0);
727 * Lookup a mount point by filesystem identifier, busying it before
730 * To avoid congestion on mountlist_mtx, implement simple direct-mapped
731 * cache for popular filesystem identifiers. The cache is lockess, using
732 * the fact that struct mount's are never freed. In worst case we may
733 * get pointer to unmounted or even different filesystem, so we have to
734 * check what we got, and go slow way if so.
737 vfs_busyfs(fsid_t *fsid)
739 #define FSID_CACHE_SIZE 256
740 typedef struct mount * volatile vmp_t;
741 static vmp_t cache[FSID_CACHE_SIZE];
746 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
747 hash = fsid->val[0] ^ fsid->val[1];
748 hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
751 mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
752 mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
754 if (vfs_busy(mp, 0) != 0) {
758 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
759 mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
765 mtx_lock(&mountlist_mtx);
766 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
767 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
768 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
769 error = vfs_busy(mp, MBF_MNTLSTLOCK);
772 mtx_unlock(&mountlist_mtx);
779 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
780 mtx_unlock(&mountlist_mtx);
781 return ((struct mount *) 0);
785 * Check if a user can access privileged mount options.
788 vfs_suser(struct mount *mp, struct thread *td)
792 if (jailed(td->td_ucred)) {
794 * If the jail of the calling thread lacks permission for
795 * this type of file system, deny immediately.
797 if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
801 * If the file system was mounted outside the jail of the
802 * calling thread, deny immediately.
804 if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
809 * If file system supports delegated administration, we don't check
810 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
811 * by the file system itself.
812 * If this is not the user that did original mount, we check for
813 * the PRIV_VFS_MOUNT_OWNER privilege.
815 if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
816 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
817 if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
824 * Get a new unique fsid. Try to make its val[0] unique, since this value
825 * will be used to create fake device numbers for stat(). Also try (but
826 * not so hard) make its val[0] unique mod 2^16, since some emulators only
827 * support 16-bit device numbers. We end up with unique val[0]'s for the
828 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
830 * Keep in mind that several mounts may be running in parallel. Starting
831 * the search one past where the previous search terminated is both a
832 * micro-optimization and a defense against returning the same fsid to
836 vfs_getnewfsid(struct mount *mp)
838 static uint16_t mntid_base;
843 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
844 mtx_lock(&mntid_mtx);
845 mtype = mp->mnt_vfc->vfc_typenum;
846 tfsid.val[1] = mtype;
847 mtype = (mtype & 0xFF) << 24;
849 tfsid.val[0] = makedev(255,
850 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
852 if ((nmp = vfs_getvfs(&tfsid)) == NULL)
856 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
857 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
858 mtx_unlock(&mntid_mtx);
862 * Knob to control the precision of file timestamps:
864 * 0 = seconds only; nanoseconds zeroed.
865 * 1 = seconds and nanoseconds, accurate within 1/HZ.
866 * 2 = seconds and nanoseconds, truncated to microseconds.
867 * >=3 = seconds and nanoseconds, maximum precision.
869 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
871 static int timestamp_precision = TSP_USEC;
872 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
873 ×tamp_precision, 0, "File timestamp precision (0: seconds, "
874 "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
875 "3+: sec + ns (max. precision))");
878 * Get a current timestamp.
881 vfs_timestamp(struct timespec *tsp)
885 switch (timestamp_precision) {
887 tsp->tv_sec = time_second;
895 TIMEVAL_TO_TIMESPEC(&tv, tsp);
905 * Set vnode attributes to VNOVAL
908 vattr_null(struct vattr *vap)
912 vap->va_size = VNOVAL;
913 vap->va_bytes = VNOVAL;
914 vap->va_mode = VNOVAL;
915 vap->va_nlink = VNOVAL;
916 vap->va_uid = VNOVAL;
917 vap->va_gid = VNOVAL;
918 vap->va_fsid = VNOVAL;
919 vap->va_fileid = VNOVAL;
920 vap->va_blocksize = VNOVAL;
921 vap->va_rdev = VNOVAL;
922 vap->va_atime.tv_sec = VNOVAL;
923 vap->va_atime.tv_nsec = VNOVAL;
924 vap->va_mtime.tv_sec = VNOVAL;
925 vap->va_mtime.tv_nsec = VNOVAL;
926 vap->va_ctime.tv_sec = VNOVAL;
927 vap->va_ctime.tv_nsec = VNOVAL;
928 vap->va_birthtime.tv_sec = VNOVAL;
929 vap->va_birthtime.tv_nsec = VNOVAL;
930 vap->va_flags = VNOVAL;
931 vap->va_gen = VNOVAL;
936 * This routine is called when we have too many vnodes. It attempts
937 * to free <count> vnodes and will potentially free vnodes that still
938 * have VM backing store (VM backing store is typically the cause
939 * of a vnode blowout so we want to do this). Therefore, this operation
940 * is not considered cheap.
942 * A number of conditions may prevent a vnode from being reclaimed.
943 * the buffer cache may have references on the vnode, a directory
944 * vnode may still have references due to the namei cache representing
945 * underlying files, or the vnode may be in active use. It is not
946 * desirable to reuse such vnodes. These conditions may cause the
947 * number of vnodes to reach some minimum value regardless of what
948 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low.
950 * @param mp Try to reclaim vnodes from this mountpoint
951 * @param reclaim_nc_src Only reclaim directories with outgoing namecache
952 * entries if this argument is strue
953 * @param trigger Only reclaim vnodes with fewer than this many resident
955 * @return The number of vnodes that were reclaimed.
958 vlrureclaim(struct mount *mp, bool reclaim_nc_src, int trigger)
961 int count, done, target;
964 vn_start_write(NULL, &mp, V_WAIT);
966 count = mp->mnt_nvnodelistsize;
967 target = count * (int64_t)gapvnodes / imax(desiredvnodes, 1);
968 target = target / 10 + 1;
969 while (count != 0 && done < target) {
970 vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
971 while (vp != NULL && vp->v_type == VMARKER)
972 vp = TAILQ_NEXT(vp, v_nmntvnodes);
976 * XXX LRU is completely broken for non-free vnodes. First
977 * by calling here in mountpoint order, then by moving
978 * unselected vnodes to the end here, and most grossly by
979 * removing the vlruvp() function that was supposed to
980 * maintain the order. (This function was born broken
981 * since syncer problems prevented it doing anything.) The
982 * order is closer to LRC (C = Created).
984 * LRU reclaiming of vnodes seems to have last worked in
985 * FreeBSD-3 where LRU wasn't mentioned under any spelling.
986 * Then there was no hold count, and inactive vnodes were
987 * simply put on the free list in LRU order. The separate
988 * lists also break LRU. We prefer to reclaim from the
989 * free list for technical reasons. This tends to thrash
990 * the free list to keep very unrecently used held vnodes.
991 * The problem is mitigated by keeping the free list large.
993 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
994 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
999 * If it's been deconstructed already, it's still
1000 * referenced, or it exceeds the trigger, skip it.
1001 * Also skip free vnodes. We are trying to make space
1002 * to expand the free list, not reduce it.
1004 if (vp->v_usecount ||
1005 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1006 ((vp->v_iflag & VI_FREE) != 0) ||
1007 (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
1008 vp->v_object->resident_page_count > trigger)) {
1014 if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) {
1016 goto next_iter_mntunlocked;
1020 * v_usecount may have been bumped after VOP_LOCK() dropped
1021 * the vnode interlock and before it was locked again.
1023 * It is not necessary to recheck VI_DOOMED because it can
1024 * only be set by another thread that holds both the vnode
1025 * lock and vnode interlock. If another thread has the
1026 * vnode lock before we get to VOP_LOCK() and obtains the
1027 * vnode interlock after VOP_LOCK() drops the vnode
1028 * interlock, the other thread will be unable to drop the
1029 * vnode lock before our VOP_LOCK() call fails.
1031 if (vp->v_usecount ||
1032 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1033 (vp->v_iflag & VI_FREE) != 0 ||
1034 (vp->v_object != NULL &&
1035 vp->v_object->resident_page_count > trigger)) {
1036 VOP_UNLOCK(vp, LK_INTERLOCK);
1038 goto next_iter_mntunlocked;
1040 KASSERT((vp->v_iflag & VI_DOOMED) == 0,
1041 ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
1042 counter_u64_add(recycles_count, 1);
1047 next_iter_mntunlocked:
1048 if (!should_yield())
1052 if (!should_yield())
1056 kern_yield(PRI_USER);
1061 vn_finished_write(mp);
1065 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1066 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1068 "limit on vnode free requests per call to the vnlru_free routine");
1071 * Attempt to reduce the free list by the requested amount.
1074 vnlru_free_locked(int count, struct vfsops *mnt_op)
1080 tried_batches = false;
1081 mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1082 if (count > max_vnlru_free)
1083 count = max_vnlru_free;
1084 for (; count > 0; count--) {
1085 vp = TAILQ_FIRST(&vnode_free_list);
1087 * The list can be modified while the free_list_mtx
1088 * has been dropped and vp could be NULL here.
1093 mtx_unlock(&vnode_free_list_mtx);
1094 vnlru_return_batches(mnt_op);
1095 tried_batches = true;
1096 mtx_lock(&vnode_free_list_mtx);
1100 VNASSERT(vp->v_op != NULL, vp,
1101 ("vnlru_free: vnode already reclaimed."));
1102 KASSERT((vp->v_iflag & VI_FREE) != 0,
1103 ("Removing vnode not on freelist"));
1104 KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1105 ("Mangling active vnode"));
1106 TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
1109 * Don't recycle if our vnode is from different type
1110 * of mount point. Note that mp is type-safe, the
1111 * check does not reach unmapped address even if
1112 * vnode is reclaimed.
1113 * Don't recycle if we can't get the interlock without
1116 if ((mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1117 mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) {
1118 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist);
1121 VNASSERT((vp->v_iflag & VI_FREE) != 0 && vp->v_holdcnt == 0,
1122 vp, ("vp inconsistent on freelist"));
1125 * The clear of VI_FREE prevents activation of the
1126 * vnode. There is no sense in putting the vnode on
1127 * the mount point active list, only to remove it
1128 * later during recycling. Inline the relevant part
1129 * of vholdl(), to avoid triggering assertions or
1133 vp->v_iflag &= ~VI_FREE;
1134 VNODE_REFCOUNT_FENCE_REL();
1135 refcount_acquire(&vp->v_holdcnt);
1137 mtx_unlock(&vnode_free_list_mtx);
1141 * If the recycled succeeded this vdrop will actually free
1142 * the vnode. If not it will simply place it back on
1146 mtx_lock(&vnode_free_list_mtx);
1151 vnlru_free(int count, struct vfsops *mnt_op)
1154 mtx_lock(&vnode_free_list_mtx);
1155 vnlru_free_locked(count, mnt_op);
1156 mtx_unlock(&vnode_free_list_mtx);
1160 /* XXX some names and initialization are bad for limits and watermarks. */
1166 gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1167 vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1168 vlowat = vhiwat / 2;
1169 if (numvnodes > desiredvnodes)
1171 space = desiredvnodes - numvnodes;
1172 if (freevnodes > wantfreevnodes)
1173 space += freevnodes - wantfreevnodes;
1178 vnlru_return_batch_locked(struct mount *mp)
1182 mtx_assert(&mp->mnt_listmtx, MA_OWNED);
1184 if (mp->mnt_tmpfreevnodelistsize == 0)
1187 TAILQ_FOREACH(vp, &mp->mnt_tmpfreevnodelist, v_actfreelist) {
1188 VNASSERT((vp->v_mflag & VMP_TMPMNTFREELIST) != 0, vp,
1189 ("vnode without VMP_TMPMNTFREELIST on mnt_tmpfreevnodelist"));
1190 vp->v_mflag &= ~VMP_TMPMNTFREELIST;
1192 mtx_lock(&vnode_free_list_mtx);
1193 TAILQ_CONCAT(&vnode_free_list, &mp->mnt_tmpfreevnodelist, v_actfreelist);
1194 freevnodes += mp->mnt_tmpfreevnodelistsize;
1195 mtx_unlock(&vnode_free_list_mtx);
1196 mp->mnt_tmpfreevnodelistsize = 0;
1200 vnlru_return_batch(struct mount *mp)
1203 mtx_lock(&mp->mnt_listmtx);
1204 vnlru_return_batch_locked(mp);
1205 mtx_unlock(&mp->mnt_listmtx);
1209 vnlru_return_batches(struct vfsops *mnt_op)
1211 struct mount *mp, *nmp;
1214 mtx_lock(&mountlist_mtx);
1215 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1216 need_unbusy = false;
1217 if (mnt_op != NULL && mp->mnt_op != mnt_op)
1219 if (mp->mnt_tmpfreevnodelistsize == 0)
1221 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK) == 0) {
1222 vnlru_return_batch(mp);
1224 mtx_lock(&mountlist_mtx);
1227 nmp = TAILQ_NEXT(mp, mnt_list);
1231 mtx_unlock(&mountlist_mtx);
1235 * Attempt to recycle vnodes in a context that is always safe to block.
1236 * Calling vlrurecycle() from the bowels of filesystem code has some
1237 * interesting deadlock problems.
1239 static struct proc *vnlruproc;
1240 static int vnlruproc_sig;
1245 struct mount *mp, *nmp;
1246 unsigned long onumvnodes;
1247 int done, force, trigger, usevnodes;
1248 bool reclaim_nc_src;
1250 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1251 SHUTDOWN_PRI_FIRST);
1255 kproc_suspend_check(vnlruproc);
1256 mtx_lock(&vnode_free_list_mtx);
1258 * If numvnodes is too large (due to desiredvnodes being
1259 * adjusted using its sysctl, or emergency growth), first
1260 * try to reduce it by discarding from the free list.
1262 if (numvnodes > desiredvnodes)
1263 vnlru_free_locked(numvnodes - desiredvnodes, NULL);
1265 * Sleep if the vnode cache is in a good state. This is
1266 * when it is not over-full and has space for about a 4%
1267 * or 9% expansion (by growing its size or inexcessively
1268 * reducing its free list). Otherwise, try to reclaim
1269 * space for a 10% expansion.
1271 if (vstir && force == 0) {
1275 if (vspace() >= vlowat && force == 0) {
1277 wakeup(&vnlruproc_sig);
1278 msleep(vnlruproc, &vnode_free_list_mtx,
1279 PVFS|PDROP, "vlruwt", hz);
1282 mtx_unlock(&vnode_free_list_mtx);
1284 onumvnodes = numvnodes;
1286 * Calculate parameters for recycling. These are the same
1287 * throughout the loop to give some semblance of fairness.
1288 * The trigger point is to avoid recycling vnodes with lots
1289 * of resident pages. We aren't trying to free memory; we
1290 * are trying to recycle or at least free vnodes.
1292 if (numvnodes <= desiredvnodes)
1293 usevnodes = numvnodes - freevnodes;
1295 usevnodes = numvnodes;
1299 * The trigger value is is chosen to give a conservatively
1300 * large value to ensure that it alone doesn't prevent
1301 * making progress. The value can easily be so large that
1302 * it is effectively infinite in some congested and
1303 * misconfigured cases, and this is necessary. Normally
1304 * it is about 8 to 100 (pages), which is quite large.
1306 trigger = vm_cnt.v_page_count * 2 / usevnodes;
1308 trigger = vsmalltrigger;
1309 reclaim_nc_src = force >= 3;
1310 mtx_lock(&mountlist_mtx);
1311 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1312 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) {
1313 nmp = TAILQ_NEXT(mp, mnt_list);
1316 done += vlrureclaim(mp, reclaim_nc_src, trigger);
1317 mtx_lock(&mountlist_mtx);
1318 nmp = TAILQ_NEXT(mp, mnt_list);
1321 mtx_unlock(&mountlist_mtx);
1322 if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1325 if (force == 0 || force == 1) {
1335 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1337 kern_yield(PRI_USER);
1339 * After becoming active to expand above low water, keep
1340 * active until above high water.
1342 force = vspace() < vhiwat;
1346 static struct kproc_desc vnlru_kp = {
1351 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1355 * Routines having to do with the management of the vnode table.
1359 * Try to recycle a freed vnode. We abort if anyone picks up a reference
1360 * before we actually vgone(). This function must be called with the vnode
1361 * held to prevent the vnode from being returned to the free list midway
1365 vtryrecycle(struct vnode *vp)
1369 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1370 VNASSERT(vp->v_holdcnt, vp,
1371 ("vtryrecycle: Recycling vp %p without a reference.", vp));
1373 * This vnode may found and locked via some other list, if so we
1374 * can't recycle it yet.
1376 if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1378 "%s: impossible to recycle, vp %p lock is already held",
1380 return (EWOULDBLOCK);
1383 * Don't recycle if its filesystem is being suspended.
1385 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1388 "%s: impossible to recycle, cannot start the write for %p",
1393 * If we got this far, we need to acquire the interlock and see if
1394 * anyone picked up this vnode from another list. If not, we will
1395 * mark it with DOOMED via vgonel() so that anyone who does find it
1396 * will skip over it.
1399 if (vp->v_usecount) {
1400 VOP_UNLOCK(vp, LK_INTERLOCK);
1401 vn_finished_write(vnmp);
1403 "%s: impossible to recycle, %p is already referenced",
1407 if ((vp->v_iflag & VI_DOOMED) == 0) {
1408 counter_u64_add(recycles_count, 1);
1411 VOP_UNLOCK(vp, LK_INTERLOCK);
1412 vn_finished_write(vnmp);
1420 if (vspace() < vlowat && vnlruproc_sig == 0) {
1427 * Wait if necessary for space for a new vnode.
1430 getnewvnode_wait(int suspended)
1433 mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1434 if (numvnodes >= desiredvnodes) {
1437 * The file system is being suspended. We cannot
1438 * risk a deadlock here, so allow allocation of
1439 * another vnode even if this would give too many.
1443 if (vnlruproc_sig == 0) {
1444 vnlruproc_sig = 1; /* avoid unnecessary wakeups */
1447 msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
1450 /* Post-adjust like the pre-adjust in getnewvnode(). */
1451 if (numvnodes + 1 > desiredvnodes && freevnodes > 1)
1452 vnlru_free_locked(1, NULL);
1453 return (numvnodes >= desiredvnodes ? ENFILE : 0);
1457 * This hack is fragile, and probably not needed any more now that the
1458 * watermark handling works.
1461 getnewvnode_reserve(u_int count)
1465 /* Pre-adjust like the pre-adjust in getnewvnode(), with any count. */
1466 /* XXX no longer so quick, but this part is not racy. */
1467 mtx_lock(&vnode_free_list_mtx);
1468 if (numvnodes + count > desiredvnodes && freevnodes > wantfreevnodes)
1469 vnlru_free_locked(ulmin(numvnodes + count - desiredvnodes,
1470 freevnodes - wantfreevnodes), NULL);
1471 mtx_unlock(&vnode_free_list_mtx);
1474 /* First try to be quick and racy. */
1475 if (atomic_fetchadd_long(&numvnodes, count) + count <= desiredvnodes) {
1476 td->td_vp_reserv += count;
1477 vcheckspace(); /* XXX no longer so quick, but more racy */
1480 atomic_subtract_long(&numvnodes, count);
1482 mtx_lock(&vnode_free_list_mtx);
1484 if (getnewvnode_wait(0) == 0) {
1487 atomic_add_long(&numvnodes, 1);
1491 mtx_unlock(&vnode_free_list_mtx);
1495 * This hack is fragile, especially if desiredvnodes or wantvnodes are
1496 * misconfgured or changed significantly. Reducing desiredvnodes below
1497 * the reserved amount should cause bizarre behaviour like reducing it
1498 * below the number of active vnodes -- the system will try to reduce
1499 * numvnodes to match, but should fail, so the subtraction below should
1503 getnewvnode_drop_reserve(void)
1508 atomic_subtract_long(&numvnodes, td->td_vp_reserv);
1509 td->td_vp_reserv = 0;
1513 * Return the next vnode from the free list.
1516 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1521 struct lock_object *lo;
1522 static int cyclecount;
1525 CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1528 if (td->td_vp_reserv > 0) {
1529 td->td_vp_reserv -= 1;
1532 mtx_lock(&vnode_free_list_mtx);
1533 if (numvnodes < desiredvnodes)
1535 else if (cyclecount++ >= freevnodes) {
1540 * Grow the vnode cache if it will not be above its target max
1541 * after growing. Otherwise, if the free list is nonempty, try
1542 * to reclaim 1 item from it before growing the cache (possibly
1543 * above its target max if the reclamation failed or is delayed).
1544 * Otherwise, wait for some space. In all cases, schedule
1545 * vnlru_proc() if we are getting short of space. The watermarks
1546 * should be chosen so that we never wait or even reclaim from
1547 * the free list to below its target minimum.
1549 if (numvnodes + 1 <= desiredvnodes)
1551 else if (freevnodes > 0)
1552 vnlru_free_locked(1, NULL);
1554 error = getnewvnode_wait(mp != NULL && (mp->mnt_kern_flag &
1556 #if 0 /* XXX Not all VFS_VGET/ffs_vget callers check returns. */
1558 mtx_unlock(&vnode_free_list_mtx);
1564 atomic_add_long(&numvnodes, 1);
1565 mtx_unlock(&vnode_free_list_mtx);
1567 counter_u64_add(vnodes_created, 1);
1568 vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK);
1570 * Locks are given the generic name "vnode" when created.
1571 * Follow the historic practice of using the filesystem
1572 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1574 * Locks live in a witness group keyed on their name. Thus,
1575 * when a lock is renamed, it must also move from the witness
1576 * group of its old name to the witness group of its new name.
1578 * The change only needs to be made when the vnode moves
1579 * from one filesystem type to another. We ensure that each
1580 * filesystem use a single static name pointer for its tag so
1581 * that we can compare pointers rather than doing a strcmp().
1583 lo = &vp->v_vnlock->lock_object;
1584 if (lo->lo_name != tag) {
1586 WITNESS_DESTROY(lo);
1587 WITNESS_INIT(lo, tag);
1590 * By default, don't allow shared locks unless filesystems opt-in.
1592 vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1594 * Finalize various vnode identity bits.
1596 KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1597 KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1598 KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1602 v_init_counters(vp);
1603 vp->v_bufobj.bo_ops = &buf_ops_bio;
1605 if (mp == NULL && vops != &dead_vnodeops)
1606 printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1610 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1611 mac_vnode_associate_singlelabel(mp, vp);
1614 vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1615 if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1616 vp->v_vflag |= VV_NOKNOTE;
1620 * For the filesystems which do not use vfs_hash_insert(),
1621 * still initialize v_hash to have vfs_hash_index() useful.
1622 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1625 vp->v_hash = (uintptr_t)vp >> vnsz2log;
1632 * Delete from old mount point vnode list, if on one.
1635 delmntque(struct vnode *vp)
1645 KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize,
1646 ("Active vnode list size %d > Vnode list size %d",
1647 mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize));
1648 active = vp->v_iflag & VI_ACTIVE;
1649 vp->v_iflag &= ~VI_ACTIVE;
1651 mtx_lock(&mp->mnt_listmtx);
1652 TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist);
1653 mp->mnt_activevnodelistsize--;
1654 mtx_unlock(&mp->mnt_listmtx);
1658 VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1659 ("bad mount point vnode list size"));
1660 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1661 mp->mnt_nvnodelistsize--;
1667 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1671 vp->v_op = &dead_vnodeops;
1677 * Insert into list of vnodes for the new mount point, if available.
1680 insmntque1(struct vnode *vp, struct mount *mp,
1681 void (*dtr)(struct vnode *, void *), void *dtr_arg)
1684 KASSERT(vp->v_mount == NULL,
1685 ("insmntque: vnode already on per mount vnode list"));
1686 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1687 ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1690 * We acquire the vnode interlock early to ensure that the
1691 * vnode cannot be recycled by another process releasing a
1692 * holdcnt on it before we get it on both the vnode list
1693 * and the active vnode list. The mount mutex protects only
1694 * manipulation of the vnode list and the vnode freelist
1695 * mutex protects only manipulation of the active vnode list.
1696 * Hence the need to hold the vnode interlock throughout.
1700 if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1701 ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1702 mp->mnt_nvnodelistsize == 0)) &&
1703 (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1712 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1713 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1714 ("neg mount point vnode list size"));
1715 mp->mnt_nvnodelistsize++;
1716 KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1717 ("Activating already active vnode"));
1718 vp->v_iflag |= VI_ACTIVE;
1719 mtx_lock(&mp->mnt_listmtx);
1720 TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
1721 mp->mnt_activevnodelistsize++;
1722 mtx_unlock(&mp->mnt_listmtx);
1729 insmntque(struct vnode *vp, struct mount *mp)
1732 return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1736 * Flush out and invalidate all buffers associated with a bufobj
1737 * Called with the underlying object locked.
1740 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1745 if (flags & V_SAVE) {
1746 error = bufobj_wwait(bo, slpflag, slptimeo);
1751 if (bo->bo_dirty.bv_cnt > 0) {
1753 if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1756 * XXX We could save a lock/unlock if this was only
1757 * enabled under INVARIANTS
1760 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1761 panic("vinvalbuf: dirty bufs");
1765 * If you alter this loop please notice that interlock is dropped and
1766 * reacquired in flushbuflist. Special care is needed to ensure that
1767 * no race conditions occur from this.
1770 error = flushbuflist(&bo->bo_clean,
1771 flags, bo, slpflag, slptimeo);
1772 if (error == 0 && !(flags & V_CLEANONLY))
1773 error = flushbuflist(&bo->bo_dirty,
1774 flags, bo, slpflag, slptimeo);
1775 if (error != 0 && error != EAGAIN) {
1779 } while (error != 0);
1782 * Wait for I/O to complete. XXX needs cleaning up. The vnode can
1783 * have write I/O in-progress but if there is a VM object then the
1784 * VM object can also have read-I/O in-progress.
1787 bufobj_wwait(bo, 0, 0);
1788 if ((flags & V_VMIO) == 0) {
1790 if (bo->bo_object != NULL) {
1791 VM_OBJECT_WLOCK(bo->bo_object);
1792 vm_object_pip_wait(bo->bo_object, "bovlbx");
1793 VM_OBJECT_WUNLOCK(bo->bo_object);
1797 } while (bo->bo_numoutput > 0);
1801 * Destroy the copy in the VM cache, too.
1803 if (bo->bo_object != NULL &&
1804 (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1805 VM_OBJECT_WLOCK(bo->bo_object);
1806 vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1807 OBJPR_CLEANONLY : 0);
1808 VM_OBJECT_WUNLOCK(bo->bo_object);
1813 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1814 V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1815 bo->bo_clean.bv_cnt > 0))
1816 panic("vinvalbuf: flush failed");
1817 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1818 bo->bo_dirty.bv_cnt > 0)
1819 panic("vinvalbuf: flush dirty failed");
1826 * Flush out and invalidate all buffers associated with a vnode.
1827 * Called with the underlying object locked.
1830 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1833 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1834 ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1835 if (vp->v_object != NULL && vp->v_object->handle != vp)
1837 return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1841 * Flush out buffers on the specified list.
1845 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1848 struct buf *bp, *nbp;
1853 ASSERT_BO_WLOCKED(bo);
1856 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1857 if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1858 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1862 lblkno = nbp->b_lblkno;
1863 xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
1866 error = BUF_TIMELOCK(bp,
1867 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
1868 "flushbuf", slpflag, slptimeo);
1871 return (error != ENOLCK ? error : EAGAIN);
1873 KASSERT(bp->b_bufobj == bo,
1874 ("bp %p wrong b_bufobj %p should be %p",
1875 bp, bp->b_bufobj, bo));
1877 * XXX Since there are no node locks for NFS, I
1878 * believe there is a slight chance that a delayed
1879 * write will occur while sleeping just above, so
1882 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1885 bp->b_flags |= B_ASYNC;
1888 return (EAGAIN); /* XXX: why not loop ? */
1891 bp->b_flags |= (B_INVAL | B_RELBUF);
1892 bp->b_flags &= ~B_ASYNC;
1897 nbp = gbincore(bo, lblkno);
1898 if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1900 break; /* nbp invalid */
1906 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
1912 ASSERT_BO_LOCKED(bo);
1914 for (lblkno = startn;;) {
1916 bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
1917 if (bp == NULL || bp->b_lblkno >= endn ||
1918 bp->b_lblkno < startn)
1920 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
1921 LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
1924 if (error == ENOLCK)
1928 KASSERT(bp->b_bufobj == bo,
1929 ("bp %p wrong b_bufobj %p should be %p",
1930 bp, bp->b_bufobj, bo));
1931 lblkno = bp->b_lblkno + 1;
1932 if ((bp->b_flags & B_MANAGED) == 0)
1934 bp->b_flags |= B_RELBUF;
1936 * In the VMIO case, use the B_NOREUSE flag to hint that the
1937 * pages backing each buffer in the range are unlikely to be
1938 * reused. Dirty buffers will have the hint applied once
1939 * they've been written.
1941 if ((bp->b_flags & B_VMIO) != 0)
1942 bp->b_flags |= B_NOREUSE;
1950 * Truncate a file's buffer and pages to a specified length. This
1951 * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1955 vtruncbuf(struct vnode *vp, off_t length, int blksize)
1957 struct buf *bp, *nbp;
1961 CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
1962 vp, blksize, (uintmax_t)length);
1965 * Round up to the *next* lbn.
1967 startlbn = howmany(length, blksize);
1969 ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1975 while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
1980 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1981 if (bp->b_lblkno > 0)
1984 * Since we hold the vnode lock this should only
1985 * fail if we're racing with the buf daemon.
1988 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1989 BO_LOCKPTR(bo)) == ENOLCK)
1990 goto restart_unlocked;
1992 VNASSERT((bp->b_flags & B_DELWRI), vp,
1993 ("buf(%p) on dirty queue without DELWRI", bp));
2002 bufobj_wwait(bo, 0, 0);
2004 vnode_pager_setsize(vp, length);
2010 * Invalidate the cached pages of a file's buffer within the range of block
2011 * numbers [startlbn, endlbn).
2014 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2020 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2022 start = blksize * startlbn;
2023 end = blksize * endlbn;
2027 MPASS(blksize == bo->bo_bsize);
2029 while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2033 vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2037 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2038 daddr_t startlbn, daddr_t endlbn)
2040 struct buf *bp, *nbp;
2043 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2044 ASSERT_BO_LOCKED(bo);
2048 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2049 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2052 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2053 BO_LOCKPTR(bo)) == ENOLCK) {
2059 bp->b_flags |= B_INVAL | B_RELBUF;
2060 bp->b_flags &= ~B_ASYNC;
2066 (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2068 (nbp->b_flags & B_DELWRI) != 0))
2072 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2073 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2076 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2077 BO_LOCKPTR(bo)) == ENOLCK) {
2082 bp->b_flags |= B_INVAL | B_RELBUF;
2083 bp->b_flags &= ~B_ASYNC;
2089 (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2090 (nbp->b_vp != vp) ||
2091 (nbp->b_flags & B_DELWRI) == 0))
2099 buf_vlist_remove(struct buf *bp)
2103 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2104 ASSERT_BO_WLOCKED(bp->b_bufobj);
2105 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
2106 (BX_VNDIRTY|BX_VNCLEAN),
2107 ("buf_vlist_remove: Buf %p is on two lists", bp));
2108 if (bp->b_xflags & BX_VNDIRTY)
2109 bv = &bp->b_bufobj->bo_dirty;
2111 bv = &bp->b_bufobj->bo_clean;
2112 BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2113 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2115 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2119 * Add the buffer to the sorted clean or dirty block list.
2121 * NOTE: xflags is passed as a constant, optimizing this inline function!
2124 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2130 ASSERT_BO_WLOCKED(bo);
2131 KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2132 ("dead bo %p", bo));
2133 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2134 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2135 bp->b_xflags |= xflags;
2136 if (xflags & BX_VNDIRTY)
2142 * Keep the list ordered. Optimize empty list insertion. Assume
2143 * we tend to grow at the tail so lookup_le should usually be cheaper
2146 if (bv->bv_cnt == 0 ||
2147 bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2148 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2149 else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2150 TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2152 TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2153 error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2155 panic("buf_vlist_add: Preallocated nodes insufficient.");
2160 * Look up a buffer using the buffer tries.
2163 gbincore(struct bufobj *bo, daddr_t lblkno)
2167 ASSERT_BO_LOCKED(bo);
2168 bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2171 return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
2175 * Associate a buffer with a vnode.
2178 bgetvp(struct vnode *vp, struct buf *bp)
2183 ASSERT_BO_WLOCKED(bo);
2184 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2186 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2187 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2188 ("bgetvp: bp already attached! %p", bp));
2194 * Insert onto list for new vnode.
2196 buf_vlist_add(bp, bo, BX_VNCLEAN);
2200 * Disassociate a buffer from a vnode.
2203 brelvp(struct buf *bp)
2208 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2209 KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2212 * Delete from old vnode list, if on one.
2214 vp = bp->b_vp; /* XXX */
2217 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2218 buf_vlist_remove(bp);
2220 panic("brelvp: Buffer %p not on queue.", bp);
2221 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2222 bo->bo_flag &= ~BO_ONWORKLST;
2223 mtx_lock(&sync_mtx);
2224 LIST_REMOVE(bo, bo_synclist);
2225 syncer_worklist_len--;
2226 mtx_unlock(&sync_mtx);
2229 bp->b_bufobj = NULL;
2235 * Add an item to the syncer work queue.
2238 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2242 ASSERT_BO_WLOCKED(bo);
2244 mtx_lock(&sync_mtx);
2245 if (bo->bo_flag & BO_ONWORKLST)
2246 LIST_REMOVE(bo, bo_synclist);
2248 bo->bo_flag |= BO_ONWORKLST;
2249 syncer_worklist_len++;
2252 if (delay > syncer_maxdelay - 2)
2253 delay = syncer_maxdelay - 2;
2254 slot = (syncer_delayno + delay) & syncer_mask;
2256 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2257 mtx_unlock(&sync_mtx);
2261 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2265 mtx_lock(&sync_mtx);
2266 len = syncer_worklist_len - sync_vnode_count;
2267 mtx_unlock(&sync_mtx);
2268 error = SYSCTL_OUT(req, &len, sizeof(len));
2272 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
2273 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2275 static struct proc *updateproc;
2276 static void sched_sync(void);
2277 static struct kproc_desc up_kp = {
2282 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2285 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2290 *bo = LIST_FIRST(slp);
2294 if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2297 * We use vhold in case the vnode does not
2298 * successfully sync. vhold prevents the vnode from
2299 * going away when we unlock the sync_mtx so that
2300 * we can acquire the vnode interlock.
2303 mtx_unlock(&sync_mtx);
2305 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2307 mtx_lock(&sync_mtx);
2308 return (*bo == LIST_FIRST(slp));
2310 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2311 (void) VOP_FSYNC(vp, MNT_LAZY, td);
2313 vn_finished_write(mp);
2315 if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2317 * Put us back on the worklist. The worklist
2318 * routine will remove us from our current
2319 * position and then add us back in at a later
2322 vn_syncer_add_to_worklist(*bo, syncdelay);
2326 mtx_lock(&sync_mtx);
2330 static int first_printf = 1;
2333 * System filesystem synchronizer daemon.
2338 struct synclist *next, *slp;
2341 struct thread *td = curthread;
2343 int net_worklist_len;
2344 int syncer_final_iter;
2348 syncer_final_iter = 0;
2349 syncer_state = SYNCER_RUNNING;
2350 starttime = time_uptime;
2351 td->td_pflags |= TDP_NORUNNINGBUF;
2353 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2356 mtx_lock(&sync_mtx);
2358 if (syncer_state == SYNCER_FINAL_DELAY &&
2359 syncer_final_iter == 0) {
2360 mtx_unlock(&sync_mtx);
2361 kproc_suspend_check(td->td_proc);
2362 mtx_lock(&sync_mtx);
2364 net_worklist_len = syncer_worklist_len - sync_vnode_count;
2365 if (syncer_state != SYNCER_RUNNING &&
2366 starttime != time_uptime) {
2368 printf("\nSyncing disks, vnodes remaining... ");
2371 printf("%d ", net_worklist_len);
2373 starttime = time_uptime;
2376 * Push files whose dirty time has expired. Be careful
2377 * of interrupt race on slp queue.
2379 * Skip over empty worklist slots when shutting down.
2382 slp = &syncer_workitem_pending[syncer_delayno];
2383 syncer_delayno += 1;
2384 if (syncer_delayno == syncer_maxdelay)
2386 next = &syncer_workitem_pending[syncer_delayno];
2388 * If the worklist has wrapped since the
2389 * it was emptied of all but syncer vnodes,
2390 * switch to the FINAL_DELAY state and run
2391 * for one more second.
2393 if (syncer_state == SYNCER_SHUTTING_DOWN &&
2394 net_worklist_len == 0 &&
2395 last_work_seen == syncer_delayno) {
2396 syncer_state = SYNCER_FINAL_DELAY;
2397 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2399 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2400 syncer_worklist_len > 0);
2403 * Keep track of the last time there was anything
2404 * on the worklist other than syncer vnodes.
2405 * Return to the SHUTTING_DOWN state if any
2408 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2409 last_work_seen = syncer_delayno;
2410 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2411 syncer_state = SYNCER_SHUTTING_DOWN;
2412 while (!LIST_EMPTY(slp)) {
2413 error = sync_vnode(slp, &bo, td);
2415 LIST_REMOVE(bo, bo_synclist);
2416 LIST_INSERT_HEAD(next, bo, bo_synclist);
2420 if (first_printf == 0) {
2422 * Drop the sync mutex, because some watchdog
2423 * drivers need to sleep while patting
2425 mtx_unlock(&sync_mtx);
2426 wdog_kern_pat(WD_LASTVAL);
2427 mtx_lock(&sync_mtx);
2431 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2432 syncer_final_iter--;
2434 * The variable rushjob allows the kernel to speed up the
2435 * processing of the filesystem syncer process. A rushjob
2436 * value of N tells the filesystem syncer to process the next
2437 * N seconds worth of work on its queue ASAP. Currently rushjob
2438 * is used by the soft update code to speed up the filesystem
2439 * syncer process when the incore state is getting so far
2440 * ahead of the disk that the kernel memory pool is being
2441 * threatened with exhaustion.
2448 * Just sleep for a short period of time between
2449 * iterations when shutting down to allow some I/O
2452 * If it has taken us less than a second to process the
2453 * current work, then wait. Otherwise start right over
2454 * again. We can still lose time if any single round
2455 * takes more than two seconds, but it does not really
2456 * matter as we are just trying to generally pace the
2457 * filesystem activity.
2459 if (syncer_state != SYNCER_RUNNING ||
2460 time_uptime == starttime) {
2462 sched_prio(td, PPAUSE);
2465 if (syncer_state != SYNCER_RUNNING)
2466 cv_timedwait(&sync_wakeup, &sync_mtx,
2467 hz / SYNCER_SHUTDOWN_SPEEDUP);
2468 else if (time_uptime == starttime)
2469 cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2474 * Request the syncer daemon to speed up its work.
2475 * We never push it to speed up more than half of its
2476 * normal turn time, otherwise it could take over the cpu.
2479 speedup_syncer(void)
2483 mtx_lock(&sync_mtx);
2484 if (rushjob < syncdelay / 2) {
2486 stat_rush_requests += 1;
2489 mtx_unlock(&sync_mtx);
2490 cv_broadcast(&sync_wakeup);
2495 * Tell the syncer to speed up its work and run though its work
2496 * list several times, then tell it to shut down.
2499 syncer_shutdown(void *arg, int howto)
2502 if (howto & RB_NOSYNC)
2504 mtx_lock(&sync_mtx);
2505 syncer_state = SYNCER_SHUTTING_DOWN;
2507 mtx_unlock(&sync_mtx);
2508 cv_broadcast(&sync_wakeup);
2509 kproc_shutdown(arg, howto);
2513 syncer_suspend(void)
2516 syncer_shutdown(updateproc, 0);
2523 mtx_lock(&sync_mtx);
2525 syncer_state = SYNCER_RUNNING;
2526 mtx_unlock(&sync_mtx);
2527 cv_broadcast(&sync_wakeup);
2528 kproc_resume(updateproc);
2532 * Reassign a buffer from one vnode to another.
2533 * Used to assign file specific control information
2534 * (indirect blocks) to the vnode to which they belong.
2537 reassignbuf(struct buf *bp)
2550 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2551 bp, bp->b_vp, bp->b_flags);
2553 * B_PAGING flagged buffers cannot be reassigned because their vp
2554 * is not fully linked in.
2556 if (bp->b_flags & B_PAGING)
2557 panic("cannot reassign paging buffer");
2560 * Delete from old vnode list, if on one.
2563 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2564 buf_vlist_remove(bp);
2566 panic("reassignbuf: Buffer %p not on queue.", bp);
2568 * If dirty, put on list of dirty buffers; otherwise insert onto list
2571 if (bp->b_flags & B_DELWRI) {
2572 if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2573 switch (vp->v_type) {
2583 vn_syncer_add_to_worklist(bo, delay);
2585 buf_vlist_add(bp, bo, BX_VNDIRTY);
2587 buf_vlist_add(bp, bo, BX_VNCLEAN);
2589 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2590 mtx_lock(&sync_mtx);
2591 LIST_REMOVE(bo, bo_synclist);
2592 syncer_worklist_len--;
2593 mtx_unlock(&sync_mtx);
2594 bo->bo_flag &= ~BO_ONWORKLST;
2599 bp = TAILQ_FIRST(&bv->bv_hd);
2600 KASSERT(bp == NULL || bp->b_bufobj == bo,
2601 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2602 bp = TAILQ_LAST(&bv->bv_hd, buflists);
2603 KASSERT(bp == NULL || bp->b_bufobj == bo,
2604 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2606 bp = TAILQ_FIRST(&bv->bv_hd);
2607 KASSERT(bp == NULL || bp->b_bufobj == bo,
2608 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2609 bp = TAILQ_LAST(&bv->bv_hd, buflists);
2610 KASSERT(bp == NULL || bp->b_bufobj == bo,
2611 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2617 v_init_counters(struct vnode *vp)
2620 VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2621 vp, ("%s called for an initialized vnode", __FUNCTION__));
2622 ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2624 refcount_init(&vp->v_holdcnt, 1);
2625 refcount_init(&vp->v_usecount, 1);
2629 v_incr_usecount_locked(struct vnode *vp)
2632 ASSERT_VI_LOCKED(vp, __func__);
2633 if ((vp->v_iflag & VI_OWEINACT) != 0) {
2634 VNASSERT(vp->v_usecount == 0, vp,
2635 ("vnode with usecount and VI_OWEINACT set"));
2636 vp->v_iflag &= ~VI_OWEINACT;
2638 refcount_acquire(&vp->v_usecount);
2639 v_incr_devcount(vp);
2643 * Increment the use count on the vnode, taking care to reference
2644 * the driver's usecount if this is a chardev.
2647 v_incr_usecount(struct vnode *vp)
2650 ASSERT_VI_UNLOCKED(vp, __func__);
2651 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2653 if (vp->v_type != VCHR &&
2654 refcount_acquire_if_not_zero(&vp->v_usecount)) {
2655 VNODE_REFCOUNT_FENCE_ACQ();
2656 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2657 ("vnode with usecount and VI_OWEINACT set"));
2660 v_incr_usecount_locked(vp);
2666 * Increment si_usecount of the associated device, if any.
2669 v_incr_devcount(struct vnode *vp)
2672 ASSERT_VI_LOCKED(vp, __FUNCTION__);
2673 if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2675 vp->v_rdev->si_usecount++;
2681 * Decrement si_usecount of the associated device, if any.
2684 v_decr_devcount(struct vnode *vp)
2687 ASSERT_VI_LOCKED(vp, __FUNCTION__);
2688 if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2690 vp->v_rdev->si_usecount--;
2696 * Grab a particular vnode from the free list, increment its
2697 * reference count and lock it. VI_DOOMED is set if the vnode
2698 * is being destroyed. Only callers who specify LK_RETRY will
2699 * see doomed vnodes. If inactive processing was delayed in
2700 * vput try to do it here.
2702 * Notes on lockless counter manipulation:
2703 * _vhold, vputx and other routines make various decisions based
2704 * on either holdcnt or usecount being 0. As long as either counter
2705 * is not transitioning 0->1 nor 1->0, the manipulation can be done
2706 * with atomic operations. Otherwise the interlock is taken covering
2707 * both the atomic and additional actions.
2710 vget(struct vnode *vp, int flags, struct thread *td)
2712 int error, oweinact;
2714 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2715 ("vget: invalid lock operation"));
2717 if ((flags & LK_INTERLOCK) != 0)
2718 ASSERT_VI_LOCKED(vp, __func__);
2720 ASSERT_VI_UNLOCKED(vp, __func__);
2721 if ((flags & LK_VNHELD) != 0)
2722 VNASSERT((vp->v_holdcnt > 0), vp,
2723 ("vget: LK_VNHELD passed but vnode not held"));
2725 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2727 if ((flags & LK_VNHELD) == 0)
2728 _vhold(vp, (flags & LK_INTERLOCK) != 0);
2730 if ((error = vn_lock(vp, flags)) != 0) {
2732 CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2736 if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
2737 panic("vget: vn_lock failed to return ENOENT\n");
2739 * We don't guarantee that any particular close will
2740 * trigger inactive processing so just make a best effort
2741 * here at preventing a reference to a removed file. If
2742 * we don't succeed no harm is done.
2744 * Upgrade our holdcnt to a usecount.
2746 if (vp->v_type == VCHR ||
2747 !refcount_acquire_if_not_zero(&vp->v_usecount)) {
2749 if ((vp->v_iflag & VI_OWEINACT) == 0) {
2753 vp->v_iflag &= ~VI_OWEINACT;
2754 VNODE_REFCOUNT_FENCE_REL();
2756 refcount_acquire(&vp->v_usecount);
2757 v_incr_devcount(vp);
2758 if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2759 (flags & LK_NOWAIT) == 0)
2767 * Increase the reference (use) and hold count of a vnode.
2768 * This will also remove the vnode from the free list if it is presently free.
2771 vref(struct vnode *vp)
2774 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2776 v_incr_usecount(vp);
2780 vrefl(struct vnode *vp)
2783 ASSERT_VI_LOCKED(vp, __func__);
2784 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2786 v_incr_usecount_locked(vp);
2790 vrefact(struct vnode *vp)
2793 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2794 if (__predict_false(vp->v_type == VCHR)) {
2795 VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp,
2796 ("%s: wrong ref counts", __func__));
2801 int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
2802 VNASSERT(old > 0, vp, ("%s: wrong hold count", __func__));
2803 old = atomic_fetchadd_int(&vp->v_usecount, 1);
2804 VNASSERT(old > 0, vp, ("%s: wrong use count", __func__));
2806 refcount_acquire(&vp->v_holdcnt);
2807 refcount_acquire(&vp->v_usecount);
2812 * Return reference count of a vnode.
2814 * The results of this call are only guaranteed when some mechanism is used to
2815 * stop other processes from gaining references to the vnode. This may be the
2816 * case if the caller holds the only reference. This is also useful when stale
2817 * data is acceptable as race conditions may be accounted for by some other
2821 vrefcnt(struct vnode *vp)
2824 return (vp->v_usecount);
2827 #define VPUTX_VRELE 1
2828 #define VPUTX_VPUT 2
2829 #define VPUTX_VUNREF 3
2832 * Decrement the use and hold counts for a vnode.
2834 * See an explanation near vget() as to why atomic operation is safe.
2837 vputx(struct vnode *vp, int func)
2841 KASSERT(vp != NULL, ("vputx: null vp"));
2842 if (func == VPUTX_VUNREF)
2843 ASSERT_VOP_LOCKED(vp, "vunref");
2844 else if (func == VPUTX_VPUT)
2845 ASSERT_VOP_LOCKED(vp, "vput");
2847 KASSERT(func == VPUTX_VRELE, ("vputx: wrong func"));
2848 ASSERT_VI_UNLOCKED(vp, __func__);
2849 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2851 if (vp->v_type != VCHR &&
2852 refcount_release_if_not_last(&vp->v_usecount)) {
2853 if (func == VPUTX_VPUT)
2862 * We want to hold the vnode until the inactive finishes to
2863 * prevent vgone() races. We drop the use count here and the
2864 * hold count below when we're done.
2866 if (!refcount_release(&vp->v_usecount) ||
2867 (vp->v_iflag & VI_DOINGINACT)) {
2868 if (func == VPUTX_VPUT)
2870 v_decr_devcount(vp);
2875 v_decr_devcount(vp);
2879 if (vp->v_usecount != 0) {
2880 vn_printf(vp, "vputx: usecount not zero for vnode ");
2881 panic("vputx: usecount not zero");
2884 CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp);
2887 * We must call VOP_INACTIVE with the node locked. Mark
2888 * as VI_DOINGINACT to avoid recursion.
2890 vp->v_iflag |= VI_OWEINACT;
2893 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
2897 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2898 error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
2904 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2905 error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
2910 VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
2911 ("vnode with usecount and VI_OWEINACT set"));
2913 if (vp->v_iflag & VI_OWEINACT)
2914 vinactive(vp, curthread);
2915 if (func != VPUTX_VUNREF)
2922 * Vnode put/release.
2923 * If count drops to zero, call inactive routine and return to freelist.
2926 vrele(struct vnode *vp)
2929 vputx(vp, VPUTX_VRELE);
2933 * Release an already locked vnode. This give the same effects as
2934 * unlock+vrele(), but takes less time and avoids releasing and
2935 * re-aquiring the lock (as vrele() acquires the lock internally.)
2938 vput(struct vnode *vp)
2941 vputx(vp, VPUTX_VPUT);
2945 * Release an exclusively locked vnode. Do not unlock the vnode lock.
2948 vunref(struct vnode *vp)
2951 vputx(vp, VPUTX_VUNREF);
2955 * Increase the hold count and activate if this is the first reference.
2958 _vhold(struct vnode *vp, bool locked)
2963 ASSERT_VI_LOCKED(vp, __func__);
2965 ASSERT_VI_UNLOCKED(vp, __func__);
2966 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2968 if (refcount_acquire_if_not_zero(&vp->v_holdcnt)) {
2969 VNODE_REFCOUNT_FENCE_ACQ();
2970 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2971 ("_vhold: vnode with holdcnt is free"));
2976 if ((vp->v_iflag & VI_FREE) == 0) {
2977 refcount_acquire(&vp->v_holdcnt);
2982 VNASSERT(vp->v_holdcnt == 0, vp,
2983 ("%s: wrong hold count", __func__));
2984 VNASSERT(vp->v_op != NULL, vp,
2985 ("%s: vnode already reclaimed.", __func__));
2987 * Remove a vnode from the free list, mark it as in use,
2988 * and put it on the active list.
2990 VNASSERT(vp->v_mount != NULL, vp,
2991 ("_vhold: vnode not on per mount vnode list"));
2993 mtx_lock(&mp->mnt_listmtx);
2994 if ((vp->v_mflag & VMP_TMPMNTFREELIST) != 0) {
2995 TAILQ_REMOVE(&mp->mnt_tmpfreevnodelist, vp, v_actfreelist);
2996 mp->mnt_tmpfreevnodelistsize--;
2997 vp->v_mflag &= ~VMP_TMPMNTFREELIST;
2999 mtx_lock(&vnode_free_list_mtx);
3000 TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
3002 mtx_unlock(&vnode_free_list_mtx);
3004 KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
3005 ("Activating already active vnode"));
3006 vp->v_iflag &= ~VI_FREE;
3007 vp->v_iflag |= VI_ACTIVE;
3008 TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
3009 mp->mnt_activevnodelistsize++;
3010 mtx_unlock(&mp->mnt_listmtx);
3011 refcount_acquire(&vp->v_holdcnt);
3017 * Drop the hold count of the vnode. If this is the last reference to
3018 * the vnode we place it on the free list unless it has been vgone'd
3019 * (marked VI_DOOMED) in which case we will free it.
3021 * Because the vnode vm object keeps a hold reference on the vnode if
3022 * there is at least one resident non-cached page, the vnode cannot
3023 * leave the active list without the page cleanup done.
3026 _vdrop(struct vnode *vp, bool locked)
3033 ASSERT_VI_LOCKED(vp, __func__);
3035 ASSERT_VI_UNLOCKED(vp, __func__);
3036 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3037 if ((int)vp->v_holdcnt <= 0)
3038 panic("vdrop: holdcnt %d", vp->v_holdcnt);
3040 if (refcount_release_if_not_last(&vp->v_holdcnt))
3044 if (refcount_release(&vp->v_holdcnt) == 0) {
3048 if ((vp->v_iflag & VI_DOOMED) == 0) {
3050 * Mark a vnode as free: remove it from its active list
3051 * and put it up for recycling on the freelist.
3053 VNASSERT(vp->v_op != NULL, vp,
3054 ("vdropl: vnode already reclaimed."));
3055 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
3056 ("vnode already free"));
3057 VNASSERT(vp->v_holdcnt == 0, vp,
3058 ("vdropl: freeing when we shouldn't"));
3059 active = vp->v_iflag & VI_ACTIVE;
3060 if ((vp->v_iflag & VI_OWEINACT) == 0) {
3061 vp->v_iflag &= ~VI_ACTIVE;
3064 mtx_lock(&mp->mnt_listmtx);
3066 TAILQ_REMOVE(&mp->mnt_activevnodelist,
3068 mp->mnt_activevnodelistsize--;
3070 TAILQ_INSERT_TAIL(&mp->mnt_tmpfreevnodelist,
3072 mp->mnt_tmpfreevnodelistsize++;
3073 vp->v_iflag |= VI_FREE;
3074 vp->v_mflag |= VMP_TMPMNTFREELIST;
3076 if (mp->mnt_tmpfreevnodelistsize >=
3077 mnt_free_list_batch)
3078 vnlru_return_batch_locked(mp);
3079 mtx_unlock(&mp->mnt_listmtx);
3081 VNASSERT(active == 0, vp,
3082 ("vdropl: active vnode not on per mount "
3084 mtx_lock(&vnode_free_list_mtx);
3085 TAILQ_INSERT_TAIL(&vnode_free_list, vp,
3088 vp->v_iflag |= VI_FREE;
3090 mtx_unlock(&vnode_free_list_mtx);
3094 counter_u64_add(free_owe_inact, 1);
3099 * The vnode has been marked for destruction, so free it.
3101 * The vnode will be returned to the zone where it will
3102 * normally remain until it is needed for another vnode. We
3103 * need to cleanup (or verify that the cleanup has already
3104 * been done) any residual data left from its current use
3105 * so as not to contaminate the freshly allocated vnode.
3107 CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
3108 atomic_subtract_long(&numvnodes, 1);
3110 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
3111 ("cleaned vnode still on the free list."));
3112 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
3113 VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
3114 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
3115 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
3116 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
3117 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
3118 VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
3119 ("clean blk trie not empty"));
3120 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
3121 VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
3122 ("dirty blk trie not empty"));
3123 VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
3124 VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
3125 VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
3126 VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
3127 ("Dangling rangelock waiters"));
3130 mac_vnode_destroy(vp);
3132 if (vp->v_pollinfo != NULL) {
3133 destroy_vpollinfo(vp->v_pollinfo);
3134 vp->v_pollinfo = NULL;
3137 /* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
3140 vp->v_mountedhere = NULL;
3143 vp->v_fifoinfo = NULL;
3144 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
3148 uma_zfree(vnode_zone, vp);
3152 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3153 * flags. DOINGINACT prevents us from recursing in calls to vinactive.
3154 * OWEINACT tracks whether a vnode missed a call to inactive due to a
3155 * failed lock upgrade.
3158 vinactive(struct vnode *vp, struct thread *td)
3160 struct vm_object *obj;
3162 ASSERT_VOP_ELOCKED(vp, "vinactive");
3163 ASSERT_VI_LOCKED(vp, "vinactive");
3164 VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3165 ("vinactive: recursed on VI_DOINGINACT"));
3166 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3167 vp->v_iflag |= VI_DOINGINACT;
3168 vp->v_iflag &= ~VI_OWEINACT;
3171 * Before moving off the active list, we must be sure that any
3172 * modified pages are converted into the vnode's dirty
3173 * buffers, since these will no longer be checked once the
3174 * vnode is on the inactive list.
3176 * The write-out of the dirty pages is asynchronous. At the
3177 * point that VOP_INACTIVE() is called, there could still be
3178 * pending I/O and dirty pages in the object.
3180 if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3181 (obj->flags & OBJ_MIGHTBEDIRTY) != 0) {
3182 VM_OBJECT_WLOCK(obj);
3183 vm_object_page_clean(obj, 0, 0, 0);
3184 VM_OBJECT_WUNLOCK(obj);
3186 VOP_INACTIVE(vp, td);
3188 VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3189 ("vinactive: lost VI_DOINGINACT"));
3190 vp->v_iflag &= ~VI_DOINGINACT;
3194 * Remove any vnodes in the vnode table belonging to mount point mp.
3196 * If FORCECLOSE is not specified, there should not be any active ones,
3197 * return error if any are found (nb: this is a user error, not a
3198 * system error). If FORCECLOSE is specified, detach any active vnodes
3201 * If WRITECLOSE is set, only flush out regular file vnodes open for
3204 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3206 * `rootrefs' specifies the base reference count for the root vnode
3207 * of this filesystem. The root vnode is considered busy if its
3208 * v_usecount exceeds this value. On a successful return, vflush(, td)
3209 * will call vrele() on the root vnode exactly rootrefs times.
3210 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3214 static int busyprt = 0; /* print out busy vnodes */
3215 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3219 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3221 struct vnode *vp, *mvp, *rootvp = NULL;
3223 int busy = 0, error;
3225 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3228 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3229 ("vflush: bad args"));
3231 * Get the filesystem root vnode. We can vput() it
3232 * immediately, since with rootrefs > 0, it won't go away.
3234 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3235 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3242 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3244 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3247 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3251 * Skip over a vnodes marked VV_SYSTEM.
3253 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3259 * If WRITECLOSE is set, flush out unlinked but still open
3260 * files (even if open only for reading) and regular file
3261 * vnodes open for writing.
3263 if (flags & WRITECLOSE) {
3264 if (vp->v_object != NULL) {
3265 VM_OBJECT_WLOCK(vp->v_object);
3266 vm_object_page_clean(vp->v_object, 0, 0, 0);
3267 VM_OBJECT_WUNLOCK(vp->v_object);
3269 error = VOP_FSYNC(vp, MNT_WAIT, td);
3273 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3276 error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3279 if ((vp->v_type == VNON ||
3280 (error == 0 && vattr.va_nlink > 0)) &&
3281 (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3289 * With v_usecount == 0, all we need to do is clear out the
3290 * vnode data structures and we are done.
3292 * If FORCECLOSE is set, forcibly close the vnode.
3294 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3300 vn_printf(vp, "vflush: busy vnode ");
3306 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3308 * If just the root vnode is busy, and if its refcount
3309 * is equal to `rootrefs', then go ahead and kill it.
3312 KASSERT(busy > 0, ("vflush: not busy"));
3313 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3314 ("vflush: usecount %d < rootrefs %d",
3315 rootvp->v_usecount, rootrefs));
3316 if (busy == 1 && rootvp->v_usecount == rootrefs) {
3317 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3319 VOP_UNLOCK(rootvp, 0);
3325 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3329 for (; rootrefs > 0; rootrefs--)
3335 * Recycle an unused vnode to the front of the free list.
3338 vrecycle(struct vnode *vp)
3343 recycled = vrecyclel(vp);
3349 * vrecycle, with the vp interlock held.
3352 vrecyclel(struct vnode *vp)
3356 ASSERT_VOP_ELOCKED(vp, __func__);
3357 ASSERT_VI_LOCKED(vp, __func__);
3358 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3360 if (vp->v_usecount == 0) {
3368 * Eliminate all activity associated with a vnode
3369 * in preparation for reuse.
3372 vgone(struct vnode *vp)
3380 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3381 struct vnode *lowervp __unused)
3386 * Notify upper mounts about reclaimed or unlinked vnode.
3389 vfs_notify_upper(struct vnode *vp, int event)
3391 static struct vfsops vgonel_vfsops = {
3392 .vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3393 .vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3395 struct mount *mp, *ump, *mmp;
3402 if (TAILQ_EMPTY(&mp->mnt_uppers))
3405 mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3406 mmp->mnt_op = &vgonel_vfsops;
3407 mmp->mnt_kern_flag |= MNTK_MARKER;
3409 mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3410 for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3411 if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3412 ump = TAILQ_NEXT(ump, mnt_upper_link);
3415 TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3418 case VFS_NOTIFY_UPPER_RECLAIM:
3419 VFS_RECLAIM_LOWERVP(ump, vp);
3421 case VFS_NOTIFY_UPPER_UNLINK:
3422 VFS_UNLINK_LOWERVP(ump, vp);
3425 KASSERT(0, ("invalid event %d", event));
3429 ump = TAILQ_NEXT(mmp, mnt_upper_link);
3430 TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3433 mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3434 if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3435 mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3436 wakeup(&mp->mnt_uppers);
3443 * vgone, with the vp interlock held.
3446 vgonel(struct vnode *vp)
3453 ASSERT_VOP_ELOCKED(vp, "vgonel");
3454 ASSERT_VI_LOCKED(vp, "vgonel");
3455 VNASSERT(vp->v_holdcnt, vp,
3456 ("vgonel: vp %p has no reference.", vp));
3457 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3461 * Don't vgonel if we're already doomed.
3463 if (vp->v_iflag & VI_DOOMED)
3465 vp->v_iflag |= VI_DOOMED;
3468 * Check to see if the vnode is in use. If so, we have to call
3469 * VOP_CLOSE() and VOP_INACTIVE().
3471 active = vp->v_usecount;
3472 oweinact = (vp->v_iflag & VI_OWEINACT);
3474 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3477 * If purging an active vnode, it must be closed and
3478 * deactivated before being reclaimed.
3481 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3482 if (oweinact || active) {
3484 if ((vp->v_iflag & VI_DOINGINACT) == 0)
3488 if (vp->v_type == VSOCK)
3489 vfs_unp_reclaim(vp);
3492 * Clean out any buffers associated with the vnode.
3493 * If the flush fails, just toss the buffers.
3496 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3497 (void) vn_start_secondary_write(vp, &mp, V_WAIT);
3498 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3499 while (vinvalbuf(vp, 0, 0, 0) != 0)
3503 BO_LOCK(&vp->v_bufobj);
3504 KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3505 vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3506 TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3507 vp->v_bufobj.bo_clean.bv_cnt == 0,
3508 ("vp %p bufobj not invalidated", vp));
3511 * For VMIO bufobj, BO_DEAD is set in vm_object_terminate()
3512 * after the object's page queue is flushed.
3514 if (vp->v_bufobj.bo_object == NULL)
3515 vp->v_bufobj.bo_flag |= BO_DEAD;
3516 BO_UNLOCK(&vp->v_bufobj);
3519 * Reclaim the vnode.
3521 if (VOP_RECLAIM(vp, td))
3522 panic("vgone: cannot reclaim");
3524 vn_finished_secondary_write(mp);
3525 VNASSERT(vp->v_object == NULL, vp,
3526 ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
3528 * Clear the advisory locks and wake up waiting threads.
3530 (void)VOP_ADVLOCKPURGE(vp);
3533 * Delete from old mount point vnode list.
3538 * Done with purge, reset to the standard lock and invalidate
3542 vp->v_vnlock = &vp->v_lock;
3543 vp->v_op = &dead_vnodeops;
3549 * Calculate the total number of references to a special device.
3552 vcount(struct vnode *vp)
3557 count = vp->v_rdev->si_usecount;
3563 * Same as above, but using the struct cdev *as argument
3566 count_dev(struct cdev *dev)
3571 count = dev->si_usecount;
3577 * Print out a description of a vnode.
3579 static char *typename[] =
3580 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3584 vn_printf(struct vnode *vp, const char *fmt, ...)
3587 char buf[256], buf2[16];
3593 printf("%p: ", (void *)vp);
3594 printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
3595 printf(" usecount %d, writecount %d, refcount %d",
3596 vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
3597 switch (vp->v_type) {
3599 printf(" mountedhere %p\n", vp->v_mountedhere);
3602 printf(" rdev %p\n", vp->v_rdev);
3605 printf(" socket %p\n", vp->v_unpcb);
3608 printf(" fifoinfo %p\n", vp->v_fifoinfo);
3616 if (vp->v_vflag & VV_ROOT)
3617 strlcat(buf, "|VV_ROOT", sizeof(buf));
3618 if (vp->v_vflag & VV_ISTTY)
3619 strlcat(buf, "|VV_ISTTY", sizeof(buf));
3620 if (vp->v_vflag & VV_NOSYNC)
3621 strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3622 if (vp->v_vflag & VV_ETERNALDEV)
3623 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3624 if (vp->v_vflag & VV_CACHEDLABEL)
3625 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3626 if (vp->v_vflag & VV_COPYONWRITE)
3627 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3628 if (vp->v_vflag & VV_SYSTEM)
3629 strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3630 if (vp->v_vflag & VV_PROCDEP)
3631 strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3632 if (vp->v_vflag & VV_NOKNOTE)
3633 strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3634 if (vp->v_vflag & VV_DELETED)
3635 strlcat(buf, "|VV_DELETED", sizeof(buf));
3636 if (vp->v_vflag & VV_MD)
3637 strlcat(buf, "|VV_MD", sizeof(buf));
3638 if (vp->v_vflag & VV_FORCEINSMQ)
3639 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
3640 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
3641 VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
3642 VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
3644 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
3645 strlcat(buf, buf2, sizeof(buf));
3647 if (vp->v_iflag & VI_MOUNT)
3648 strlcat(buf, "|VI_MOUNT", sizeof(buf));
3649 if (vp->v_iflag & VI_DOOMED)
3650 strlcat(buf, "|VI_DOOMED", sizeof(buf));
3651 if (vp->v_iflag & VI_FREE)
3652 strlcat(buf, "|VI_FREE", sizeof(buf));
3653 if (vp->v_iflag & VI_ACTIVE)
3654 strlcat(buf, "|VI_ACTIVE", sizeof(buf));
3655 if (vp->v_iflag & VI_DOINGINACT)
3656 strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
3657 if (vp->v_iflag & VI_OWEINACT)
3658 strlcat(buf, "|VI_OWEINACT", sizeof(buf));
3659 flags = vp->v_iflag & ~(VI_MOUNT | VI_DOOMED | VI_FREE |
3660 VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT);
3662 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
3663 strlcat(buf, buf2, sizeof(buf));
3665 printf(" flags (%s)\n", buf + 1);
3666 if (mtx_owned(VI_MTX(vp)))
3667 printf(" VI_LOCKed");
3668 if (vp->v_object != NULL)
3669 printf(" v_object %p ref %d pages %d "
3670 "cleanbuf %d dirtybuf %d\n",
3671 vp->v_object, vp->v_object->ref_count,
3672 vp->v_object->resident_page_count,
3673 vp->v_bufobj.bo_clean.bv_cnt,
3674 vp->v_bufobj.bo_dirty.bv_cnt);
3676 lockmgr_printinfo(vp->v_vnlock);
3677 if (vp->v_data != NULL)
3683 * List all of the locked vnodes in the system.
3684 * Called when debugging the kernel.
3686 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
3692 * Note: because this is DDB, we can't obey the locking semantics
3693 * for these structures, which means we could catch an inconsistent
3694 * state and dereference a nasty pointer. Not much to be done
3697 db_printf("Locked vnodes\n");
3698 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3699 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3700 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
3701 vn_printf(vp, "vnode ");
3707 * Show details about the given vnode.
3709 DB_SHOW_COMMAND(vnode, db_show_vnode)
3715 vp = (struct vnode *)addr;
3716 vn_printf(vp, "vnode ");
3720 * Show details about the given mount point.
3722 DB_SHOW_COMMAND(mount, db_show_mount)
3733 /* No address given, print short info about all mount points. */
3734 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3735 db_printf("%p %s on %s (%s)\n", mp,
3736 mp->mnt_stat.f_mntfromname,
3737 mp->mnt_stat.f_mntonname,
3738 mp->mnt_stat.f_fstypename);
3742 db_printf("\nMore info: show mount <addr>\n");
3746 mp = (struct mount *)addr;
3747 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
3748 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
3751 mflags = mp->mnt_flag;
3752 #define MNT_FLAG(flag) do { \
3753 if (mflags & (flag)) { \
3754 if (buf[0] != '\0') \
3755 strlcat(buf, ", ", sizeof(buf)); \
3756 strlcat(buf, (#flag) + 4, sizeof(buf)); \
3757 mflags &= ~(flag); \
3760 MNT_FLAG(MNT_RDONLY);
3761 MNT_FLAG(MNT_SYNCHRONOUS);
3762 MNT_FLAG(MNT_NOEXEC);
3763 MNT_FLAG(MNT_NOSUID);
3764 MNT_FLAG(MNT_NFS4ACLS);
3765 MNT_FLAG(MNT_UNION);
3766 MNT_FLAG(MNT_ASYNC);
3767 MNT_FLAG(MNT_SUIDDIR);
3768 MNT_FLAG(MNT_SOFTDEP);
3769 MNT_FLAG(MNT_NOSYMFOLLOW);
3770 MNT_FLAG(MNT_GJOURNAL);
3771 MNT_FLAG(MNT_MULTILABEL);
3773 MNT_FLAG(MNT_NOATIME);
3774 MNT_FLAG(MNT_NOCLUSTERR);
3775 MNT_FLAG(MNT_NOCLUSTERW);
3777 MNT_FLAG(MNT_EXRDONLY);
3778 MNT_FLAG(MNT_EXPORTED);
3779 MNT_FLAG(MNT_DEFEXPORTED);
3780 MNT_FLAG(MNT_EXPORTANON);
3781 MNT_FLAG(MNT_EXKERB);
3782 MNT_FLAG(MNT_EXPUBLIC);
3783 MNT_FLAG(MNT_LOCAL);
3784 MNT_FLAG(MNT_QUOTA);
3785 MNT_FLAG(MNT_ROOTFS);
3787 MNT_FLAG(MNT_IGNORE);
3788 MNT_FLAG(MNT_UPDATE);
3789 MNT_FLAG(MNT_DELEXPORT);
3790 MNT_FLAG(MNT_RELOAD);
3791 MNT_FLAG(MNT_FORCE);
3792 MNT_FLAG(MNT_SNAPSHOT);
3793 MNT_FLAG(MNT_BYFSID);
3797 strlcat(buf, ", ", sizeof(buf));
3798 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3799 "0x%016jx", mflags);
3801 db_printf(" mnt_flag = %s\n", buf);
3804 flags = mp->mnt_kern_flag;
3805 #define MNT_KERN_FLAG(flag) do { \
3806 if (flags & (flag)) { \
3807 if (buf[0] != '\0') \
3808 strlcat(buf, ", ", sizeof(buf)); \
3809 strlcat(buf, (#flag) + 5, sizeof(buf)); \
3813 MNT_KERN_FLAG(MNTK_UNMOUNTF);
3814 MNT_KERN_FLAG(MNTK_ASYNC);
3815 MNT_KERN_FLAG(MNTK_SOFTDEP);
3816 MNT_KERN_FLAG(MNTK_DRAINING);
3817 MNT_KERN_FLAG(MNTK_REFEXPIRE);
3818 MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
3819 MNT_KERN_FLAG(MNTK_SHARED_WRITES);
3820 MNT_KERN_FLAG(MNTK_NO_IOPF);
3821 MNT_KERN_FLAG(MNTK_VGONE_UPPER);
3822 MNT_KERN_FLAG(MNTK_VGONE_WAITER);
3823 MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
3824 MNT_KERN_FLAG(MNTK_MARKER);
3825 MNT_KERN_FLAG(MNTK_USES_BCACHE);
3826 MNT_KERN_FLAG(MNTK_NOASYNC);
3827 MNT_KERN_FLAG(MNTK_UNMOUNT);
3828 MNT_KERN_FLAG(MNTK_MWAIT);
3829 MNT_KERN_FLAG(MNTK_SUSPEND);
3830 MNT_KERN_FLAG(MNTK_SUSPEND2);
3831 MNT_KERN_FLAG(MNTK_SUSPENDED);
3832 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
3833 MNT_KERN_FLAG(MNTK_NOKNOTE);
3834 #undef MNT_KERN_FLAG
3837 strlcat(buf, ", ", sizeof(buf));
3838 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3841 db_printf(" mnt_kern_flag = %s\n", buf);
3843 db_printf(" mnt_opt = ");
3844 opt = TAILQ_FIRST(mp->mnt_opt);
3846 db_printf("%s", opt->name);
3847 opt = TAILQ_NEXT(opt, link);
3848 while (opt != NULL) {
3849 db_printf(", %s", opt->name);
3850 opt = TAILQ_NEXT(opt, link);
3856 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx "
3857 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
3858 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
3859 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
3860 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
3861 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
3862 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
3863 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
3864 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
3865 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
3866 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
3867 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
3869 db_printf(" mnt_cred = { uid=%u ruid=%u",
3870 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
3871 if (jailed(mp->mnt_cred))
3872 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
3874 db_printf(" mnt_ref = %d\n", mp->mnt_ref);
3875 db_printf(" mnt_gen = %d\n", mp->mnt_gen);
3876 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
3877 db_printf(" mnt_activevnodelistsize = %d\n",
3878 mp->mnt_activevnodelistsize);
3879 db_printf(" mnt_writeopcount = %d\n", mp->mnt_writeopcount);
3880 db_printf(" mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
3881 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max);
3882 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed);
3883 db_printf(" mnt_lockref = %d\n", mp->mnt_lockref);
3884 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
3885 db_printf(" mnt_secondary_accwrites = %d\n",
3886 mp->mnt_secondary_accwrites);
3887 db_printf(" mnt_gjprovider = %s\n",
3888 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
3890 db_printf("\n\nList of active vnodes\n");
3891 TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) {
3892 if (vp->v_type != VMARKER) {
3893 vn_printf(vp, "vnode ");
3898 db_printf("\n\nList of inactive vnodes\n");
3899 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3900 if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) {
3901 vn_printf(vp, "vnode ");
3910 * Fill in a struct xvfsconf based on a struct vfsconf.
3913 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
3915 struct xvfsconf xvfsp;
3917 bzero(&xvfsp, sizeof(xvfsp));
3918 strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3919 xvfsp.vfc_typenum = vfsp->vfc_typenum;
3920 xvfsp.vfc_refcount = vfsp->vfc_refcount;
3921 xvfsp.vfc_flags = vfsp->vfc_flags;
3923 * These are unused in userland, we keep them
3924 * to not break binary compatibility.
3926 xvfsp.vfc_vfsops = NULL;
3927 xvfsp.vfc_next = NULL;
3928 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3931 #ifdef COMPAT_FREEBSD32
3933 uint32_t vfc_vfsops;
3934 char vfc_name[MFSNAMELEN];
3935 int32_t vfc_typenum;
3936 int32_t vfc_refcount;
3942 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
3944 struct xvfsconf32 xvfsp;
3946 bzero(&xvfsp, sizeof(xvfsp));
3947 strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3948 xvfsp.vfc_typenum = vfsp->vfc_typenum;
3949 xvfsp.vfc_refcount = vfsp->vfc_refcount;
3950 xvfsp.vfc_flags = vfsp->vfc_flags;
3951 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3956 * Top level filesystem related information gathering.
3959 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
3961 struct vfsconf *vfsp;
3966 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3967 #ifdef COMPAT_FREEBSD32
3968 if (req->flags & SCTL_MASK32)
3969 error = vfsconf2x32(req, vfsp);
3972 error = vfsconf2x(req, vfsp);
3980 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
3981 CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
3982 "S,xvfsconf", "List of all configured filesystems");
3984 #ifndef BURN_BRIDGES
3985 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
3988 vfs_sysctl(SYSCTL_HANDLER_ARGS)
3990 int *name = (int *)arg1 - 1; /* XXX */
3991 u_int namelen = arg2 + 1; /* XXX */
3992 struct vfsconf *vfsp;
3994 log(LOG_WARNING, "userland calling deprecated sysctl, "
3995 "please rebuild world\n");
3997 #if 1 || defined(COMPAT_PRELITE2)
3998 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4000 return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4004 case VFS_MAXTYPENUM:
4007 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4010 return (ENOTDIR); /* overloaded */
4012 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4013 if (vfsp->vfc_typenum == name[2])
4018 return (EOPNOTSUPP);
4019 #ifdef COMPAT_FREEBSD32
4020 if (req->flags & SCTL_MASK32)
4021 return (vfsconf2x32(req, vfsp));
4024 return (vfsconf2x(req, vfsp));
4026 return (EOPNOTSUPP);
4029 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4030 CTLFLAG_MPSAFE, vfs_sysctl,
4031 "Generic filesystem");
4033 #if 1 || defined(COMPAT_PRELITE2)
4036 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4039 struct vfsconf *vfsp;
4040 struct ovfsconf ovfs;
4043 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4044 bzero(&ovfs, sizeof(ovfs));
4045 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */
4046 strcpy(ovfs.vfc_name, vfsp->vfc_name);
4047 ovfs.vfc_index = vfsp->vfc_typenum;
4048 ovfs.vfc_refcount = vfsp->vfc_refcount;
4049 ovfs.vfc_flags = vfsp->vfc_flags;
4050 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4060 #endif /* 1 || COMPAT_PRELITE2 */
4061 #endif /* !BURN_BRIDGES */
4063 #define KINFO_VNODESLOP 10
4066 * Dump vnode list (via sysctl).
4070 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4078 * Stale numvnodes access is not fatal here.
4081 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4083 /* Make an estimate */
4084 return (SYSCTL_OUT(req, 0, len));
4086 error = sysctl_wire_old_buffer(req, 0);
4089 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4091 mtx_lock(&mountlist_mtx);
4092 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4093 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4096 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4100 xvn[n].xv_size = sizeof *xvn;
4101 xvn[n].xv_vnode = vp;
4102 xvn[n].xv_id = 0; /* XXX compat */
4103 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4105 XV_COPY(writecount);
4111 xvn[n].xv_flag = vp->v_vflag;
4113 switch (vp->v_type) {
4120 if (vp->v_rdev == NULL) {
4124 xvn[n].xv_dev = dev2udev(vp->v_rdev);
4127 xvn[n].xv_socket = vp->v_socket;
4130 xvn[n].xv_fifo = vp->v_fifoinfo;
4135 /* shouldn't happen? */
4143 mtx_lock(&mountlist_mtx);
4148 mtx_unlock(&mountlist_mtx);
4150 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4155 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4156 CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4161 unmount_or_warn(struct mount *mp)
4165 error = dounmount(mp, MNT_FORCE, curthread);
4167 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4171 printf("%d)\n", error);
4176 * Unmount all filesystems. The list is traversed in reverse order
4177 * of mounting to avoid dependencies.
4180 vfs_unmountall(void)
4182 struct mount *mp, *tmp;
4184 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4187 * Since this only runs when rebooting, it is not interlocked.
4189 TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4193 * Forcibly unmounting "/dev" before "/" would prevent clean
4194 * unmount of the latter.
4196 if (mp == rootdevmp)
4199 unmount_or_warn(mp);
4202 if (rootdevmp != NULL)
4203 unmount_or_warn(rootdevmp);
4207 * perform msync on all vnodes under a mount point
4208 * the mount point must be locked.
4211 vfs_msync(struct mount *mp, int flags)
4213 struct vnode *vp, *mvp;
4214 struct vm_object *obj;
4216 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4218 vnlru_return_batch(mp);
4220 MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
4222 if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 &&
4223 (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
4225 LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
4227 if (vp->v_vflag & VV_NOSYNC) { /* unlinked */
4234 VM_OBJECT_WLOCK(obj);
4235 vm_object_page_clean(obj, 0, 0,
4237 OBJPC_SYNC : OBJPC_NOSYNC);
4238 VM_OBJECT_WUNLOCK(obj);
4248 destroy_vpollinfo_free(struct vpollinfo *vi)
4251 knlist_destroy(&vi->vpi_selinfo.si_note);
4252 mtx_destroy(&vi->vpi_lock);
4253 uma_zfree(vnodepoll_zone, vi);
4257 destroy_vpollinfo(struct vpollinfo *vi)
4260 knlist_clear(&vi->vpi_selinfo.si_note, 1);
4261 seldrain(&vi->vpi_selinfo);
4262 destroy_vpollinfo_free(vi);
4266 * Initialize per-vnode helper structure to hold poll-related state.
4269 v_addpollinfo(struct vnode *vp)
4271 struct vpollinfo *vi;
4273 if (vp->v_pollinfo != NULL)
4275 vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4276 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4277 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4278 vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4280 if (vp->v_pollinfo != NULL) {
4282 destroy_vpollinfo_free(vi);
4285 vp->v_pollinfo = vi;
4290 * Record a process's interest in events which might happen to
4291 * a vnode. Because poll uses the historic select-style interface
4292 * internally, this routine serves as both the ``check for any
4293 * pending events'' and the ``record my interest in future events''
4294 * functions. (These are done together, while the lock is held,
4295 * to avoid race conditions.)
4298 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4302 mtx_lock(&vp->v_pollinfo->vpi_lock);
4303 if (vp->v_pollinfo->vpi_revents & events) {
4305 * This leaves events we are not interested
4306 * in available for the other process which
4307 * which presumably had requested them
4308 * (otherwise they would never have been
4311 events &= vp->v_pollinfo->vpi_revents;
4312 vp->v_pollinfo->vpi_revents &= ~events;
4314 mtx_unlock(&vp->v_pollinfo->vpi_lock);
4317 vp->v_pollinfo->vpi_events |= events;
4318 selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4319 mtx_unlock(&vp->v_pollinfo->vpi_lock);
4324 * Routine to create and manage a filesystem syncer vnode.
4326 #define sync_close ((int (*)(struct vop_close_args *))nullop)
4327 static int sync_fsync(struct vop_fsync_args *);
4328 static int sync_inactive(struct vop_inactive_args *);
4329 static int sync_reclaim(struct vop_reclaim_args *);
4331 static struct vop_vector sync_vnodeops = {
4332 .vop_bypass = VOP_EOPNOTSUPP,
4333 .vop_close = sync_close, /* close */
4334 .vop_fsync = sync_fsync, /* fsync */
4335 .vop_inactive = sync_inactive, /* inactive */
4336 .vop_reclaim = sync_reclaim, /* reclaim */
4337 .vop_lock1 = vop_stdlock, /* lock */
4338 .vop_unlock = vop_stdunlock, /* unlock */
4339 .vop_islocked = vop_stdislocked, /* islocked */
4343 * Create a new filesystem syncer vnode for the specified mount point.
4346 vfs_allocate_syncvnode(struct mount *mp)
4350 static long start, incr, next;
4353 /* Allocate a new vnode */
4354 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4356 panic("vfs_allocate_syncvnode: getnewvnode() failed");
4358 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4359 vp->v_vflag |= VV_FORCEINSMQ;
4360 error = insmntque(vp, mp);
4362 panic("vfs_allocate_syncvnode: insmntque() failed");
4363 vp->v_vflag &= ~VV_FORCEINSMQ;
4366 * Place the vnode onto the syncer worklist. We attempt to
4367 * scatter them about on the list so that they will go off
4368 * at evenly distributed times even if all the filesystems
4369 * are mounted at once.
4372 if (next == 0 || next > syncer_maxdelay) {
4376 start = syncer_maxdelay / 2;
4377 incr = syncer_maxdelay;
4383 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4384 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4385 mtx_lock(&sync_mtx);
4387 if (mp->mnt_syncer == NULL) {
4388 mp->mnt_syncer = vp;
4391 mtx_unlock(&sync_mtx);
4394 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4401 vfs_deallocate_syncvnode(struct mount *mp)
4405 mtx_lock(&sync_mtx);
4406 vp = mp->mnt_syncer;
4408 mp->mnt_syncer = NULL;
4409 mtx_unlock(&sync_mtx);
4415 * Do a lazy sync of the filesystem.
4418 sync_fsync(struct vop_fsync_args *ap)
4420 struct vnode *syncvp = ap->a_vp;
4421 struct mount *mp = syncvp->v_mount;
4426 * We only need to do something if this is a lazy evaluation.
4428 if (ap->a_waitfor != MNT_LAZY)
4432 * Move ourselves to the back of the sync list.
4434 bo = &syncvp->v_bufobj;
4436 vn_syncer_add_to_worklist(bo, syncdelay);
4440 * Walk the list of vnodes pushing all that are dirty and
4441 * not already on the sync list.
4443 if (vfs_busy(mp, MBF_NOWAIT) != 0)
4445 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4449 save = curthread_pflags_set(TDP_SYNCIO);
4450 vfs_msync(mp, MNT_NOWAIT);
4451 error = VFS_SYNC(mp, MNT_LAZY);
4452 curthread_pflags_restore(save);
4453 vn_finished_write(mp);
4459 * The syncer vnode is no referenced.
4462 sync_inactive(struct vop_inactive_args *ap)
4470 * The syncer vnode is no longer needed and is being decommissioned.
4472 * Modifications to the worklist must be protected by sync_mtx.
4475 sync_reclaim(struct vop_reclaim_args *ap)
4477 struct vnode *vp = ap->a_vp;
4482 mtx_lock(&sync_mtx);
4483 if (vp->v_mount->mnt_syncer == vp)
4484 vp->v_mount->mnt_syncer = NULL;
4485 if (bo->bo_flag & BO_ONWORKLST) {
4486 LIST_REMOVE(bo, bo_synclist);
4487 syncer_worklist_len--;
4489 bo->bo_flag &= ~BO_ONWORKLST;
4491 mtx_unlock(&sync_mtx);
4498 * Check if vnode represents a disk device
4501 vn_isdisk(struct vnode *vp, int *errp)
4505 if (vp->v_type != VCHR) {
4511 if (vp->v_rdev == NULL)
4513 else if (vp->v_rdev->si_devsw == NULL)
4515 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
4521 return (error == 0);
4525 * Common filesystem object access control check routine. Accepts a
4526 * vnode's type, "mode", uid and gid, requested access mode, credentials,
4527 * and optional call-by-reference privused argument allowing vaccess()
4528 * to indicate to the caller whether privilege was used to satisfy the
4529 * request (obsoleted). Returns 0 on success, or an errno on failure.
4532 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
4533 accmode_t accmode, struct ucred *cred, int *privused)
4535 accmode_t dac_granted;
4536 accmode_t priv_granted;
4538 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
4539 ("invalid bit in accmode"));
4540 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
4541 ("VAPPEND without VWRITE"));
4544 * Look for a normal, non-privileged way to access the file/directory
4545 * as requested. If it exists, go with that.
4548 if (privused != NULL)
4553 /* Check the owner. */
4554 if (cred->cr_uid == file_uid) {
4555 dac_granted |= VADMIN;
4556 if (file_mode & S_IXUSR)
4557 dac_granted |= VEXEC;
4558 if (file_mode & S_IRUSR)
4559 dac_granted |= VREAD;
4560 if (file_mode & S_IWUSR)
4561 dac_granted |= (VWRITE | VAPPEND);
4563 if ((accmode & dac_granted) == accmode)
4569 /* Otherwise, check the groups (first match) */
4570 if (groupmember(file_gid, cred)) {
4571 if (file_mode & S_IXGRP)
4572 dac_granted |= VEXEC;
4573 if (file_mode & S_IRGRP)
4574 dac_granted |= VREAD;
4575 if (file_mode & S_IWGRP)
4576 dac_granted |= (VWRITE | VAPPEND);
4578 if ((accmode & dac_granted) == accmode)
4584 /* Otherwise, check everyone else. */
4585 if (file_mode & S_IXOTH)
4586 dac_granted |= VEXEC;
4587 if (file_mode & S_IROTH)
4588 dac_granted |= VREAD;
4589 if (file_mode & S_IWOTH)
4590 dac_granted |= (VWRITE | VAPPEND);
4591 if ((accmode & dac_granted) == accmode)
4596 * Build a privilege mask to determine if the set of privileges
4597 * satisfies the requirements when combined with the granted mask
4598 * from above. For each privilege, if the privilege is required,
4599 * bitwise or the request type onto the priv_granted mask.
4605 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
4606 * requests, instead of PRIV_VFS_EXEC.
4608 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4609 !priv_check_cred(cred, PRIV_VFS_LOOKUP, 0))
4610 priv_granted |= VEXEC;
4613 * Ensure that at least one execute bit is on. Otherwise,
4614 * a privileged user will always succeed, and we don't want
4615 * this to happen unless the file really is executable.
4617 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4618 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
4619 !priv_check_cred(cred, PRIV_VFS_EXEC, 0))
4620 priv_granted |= VEXEC;
4623 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
4624 !priv_check_cred(cred, PRIV_VFS_READ, 0))
4625 priv_granted |= VREAD;
4627 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
4628 !priv_check_cred(cred, PRIV_VFS_WRITE, 0))
4629 priv_granted |= (VWRITE | VAPPEND);
4631 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
4632 !priv_check_cred(cred, PRIV_VFS_ADMIN, 0))
4633 priv_granted |= VADMIN;
4635 if ((accmode & (priv_granted | dac_granted)) == accmode) {
4636 /* XXX audit: privilege used */
4637 if (privused != NULL)
4642 return ((accmode & VADMIN) ? EPERM : EACCES);
4646 * Credential check based on process requesting service, and per-attribute
4650 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
4651 struct thread *td, accmode_t accmode)
4655 * Kernel-invoked always succeeds.
4661 * Do not allow privileged processes in jail to directly manipulate
4662 * system attributes.
4664 switch (attrnamespace) {
4665 case EXTATTR_NAMESPACE_SYSTEM:
4666 /* Potentially should be: return (EPERM); */
4667 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0));
4668 case EXTATTR_NAMESPACE_USER:
4669 return (VOP_ACCESS(vp, accmode, cred, td));
4675 #ifdef DEBUG_VFS_LOCKS
4677 * This only exists to suppress warnings from unlocked specfs accesses. It is
4678 * no longer ok to have an unlocked VFS.
4680 #define IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL || \
4681 (vp)->v_type == VCHR || (vp)->v_type == VBAD)
4683 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */
4684 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
4685 "Drop into debugger on lock violation");
4687 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */
4688 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
4689 0, "Check for interlock across VOPs");
4691 int vfs_badlock_print = 1; /* Print lock violations. */
4692 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
4693 0, "Print lock violations");
4695 int vfs_badlock_vnode = 1; /* Print vnode details on lock violations. */
4696 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
4697 0, "Print vnode details on lock violations");
4700 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */
4701 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
4702 &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
4706 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
4710 if (vfs_badlock_backtrace)
4713 if (vfs_badlock_vnode)
4714 vn_printf(vp, "vnode ");
4715 if (vfs_badlock_print)
4716 printf("%s: %p %s\n", str, (void *)vp, msg);
4717 if (vfs_badlock_ddb)
4718 kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4722 assert_vi_locked(struct vnode *vp, const char *str)
4725 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
4726 vfs_badlock("interlock is not locked but should be", str, vp);
4730 assert_vi_unlocked(struct vnode *vp, const char *str)
4733 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
4734 vfs_badlock("interlock is locked but should not be", str, vp);
4738 assert_vop_locked(struct vnode *vp, const char *str)
4742 if (!IGNORE_LOCK(vp)) {
4743 locked = VOP_ISLOCKED(vp);
4744 if (locked == 0 || locked == LK_EXCLOTHER)
4745 vfs_badlock("is not locked but should be", str, vp);
4750 assert_vop_unlocked(struct vnode *vp, const char *str)
4753 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
4754 vfs_badlock("is locked but should not be", str, vp);
4758 assert_vop_elocked(struct vnode *vp, const char *str)
4761 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
4762 vfs_badlock("is not exclusive locked but should be", str, vp);
4764 #endif /* DEBUG_VFS_LOCKS */
4767 vop_rename_fail(struct vop_rename_args *ap)
4770 if (ap->a_tvp != NULL)
4772 if (ap->a_tdvp == ap->a_tvp)
4781 vop_rename_pre(void *ap)
4783 struct vop_rename_args *a = ap;
4785 #ifdef DEBUG_VFS_LOCKS
4787 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
4788 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
4789 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
4790 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
4792 /* Check the source (from). */
4793 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
4794 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
4795 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
4796 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
4797 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
4799 /* Check the target. */
4801 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
4802 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
4804 if (a->a_tdvp != a->a_fdvp)
4806 if (a->a_tvp != a->a_fvp)
4813 #ifdef DEBUG_VFS_LOCKS
4815 vop_strategy_pre(void *ap)
4817 struct vop_strategy_args *a;
4824 * Cluster ops lock their component buffers but not the IO container.
4826 if ((bp->b_flags & B_CLUSTER) != 0)
4829 if (panicstr == NULL && !BUF_ISLOCKED(bp)) {
4830 if (vfs_badlock_print)
4832 "VOP_STRATEGY: bp is not locked but should be\n");
4833 if (vfs_badlock_ddb)
4834 kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4839 vop_lock_pre(void *ap)
4841 struct vop_lock1_args *a = ap;
4843 if ((a->a_flags & LK_INTERLOCK) == 0)
4844 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4846 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
4850 vop_lock_post(void *ap, int rc)
4852 struct vop_lock1_args *a = ap;
4854 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4855 if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
4856 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
4860 vop_unlock_pre(void *ap)
4862 struct vop_unlock_args *a = ap;
4864 if (a->a_flags & LK_INTERLOCK)
4865 ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
4866 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
4870 vop_unlock_post(void *ap, int rc)
4872 struct vop_unlock_args *a = ap;
4874 if (a->a_flags & LK_INTERLOCK)
4875 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
4880 vop_create_post(void *ap, int rc)
4882 struct vop_create_args *a = ap;
4885 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4889 vop_deleteextattr_post(void *ap, int rc)
4891 struct vop_deleteextattr_args *a = ap;
4894 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4898 vop_link_post(void *ap, int rc)
4900 struct vop_link_args *a = ap;
4903 VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
4904 VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
4909 vop_mkdir_post(void *ap, int rc)
4911 struct vop_mkdir_args *a = ap;
4914 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4918 vop_mknod_post(void *ap, int rc)
4920 struct vop_mknod_args *a = ap;
4923 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4927 vop_reclaim_post(void *ap, int rc)
4929 struct vop_reclaim_args *a = ap;
4932 VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
4936 vop_remove_post(void *ap, int rc)
4938 struct vop_remove_args *a = ap;
4941 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4942 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4947 vop_rename_post(void *ap, int rc)
4949 struct vop_rename_args *a = ap;
4954 if (a->a_fdvp == a->a_tdvp) {
4955 if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
4957 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
4958 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
4960 hint |= NOTE_EXTEND;
4961 if (a->a_fvp->v_type == VDIR)
4963 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
4965 if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
4966 a->a_tvp->v_type == VDIR)
4968 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
4971 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
4973 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
4975 if (a->a_tdvp != a->a_fdvp)
4977 if (a->a_tvp != a->a_fvp)
4985 vop_rmdir_post(void *ap, int rc)
4987 struct vop_rmdir_args *a = ap;
4990 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4991 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4996 vop_setattr_post(void *ap, int rc)
4998 struct vop_setattr_args *a = ap;
5001 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5005 vop_setextattr_post(void *ap, int rc)
5007 struct vop_setextattr_args *a = ap;
5010 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5014 vop_symlink_post(void *ap, int rc)
5016 struct vop_symlink_args *a = ap;
5019 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5023 vop_open_post(void *ap, int rc)
5025 struct vop_open_args *a = ap;
5028 VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5032 vop_close_post(void *ap, int rc)
5034 struct vop_close_args *a = ap;
5036 if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5037 (a->a_vp->v_iflag & VI_DOOMED) == 0)) {
5038 VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5039 NOTE_CLOSE_WRITE : NOTE_CLOSE);
5044 vop_read_post(void *ap, int rc)
5046 struct vop_read_args *a = ap;
5049 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5053 vop_readdir_post(void *ap, int rc)
5055 struct vop_readdir_args *a = ap;
5058 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5061 static struct knlist fs_knlist;
5064 vfs_event_init(void *arg)
5066 knlist_init_mtx(&fs_knlist, NULL);
5068 /* XXX - correct order? */
5069 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5072 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5075 KNOTE_UNLOCKED(&fs_knlist, event);
5078 static int filt_fsattach(struct knote *kn);
5079 static void filt_fsdetach(struct knote *kn);
5080 static int filt_fsevent(struct knote *kn, long hint);
5082 struct filterops fs_filtops = {
5084 .f_attach = filt_fsattach,
5085 .f_detach = filt_fsdetach,
5086 .f_event = filt_fsevent
5090 filt_fsattach(struct knote *kn)
5093 kn->kn_flags |= EV_CLEAR;
5094 knlist_add(&fs_knlist, kn, 0);
5099 filt_fsdetach(struct knote *kn)
5102 knlist_remove(&fs_knlist, kn, 0);
5106 filt_fsevent(struct knote *kn, long hint)
5109 kn->kn_fflags |= hint;
5110 return (kn->kn_fflags != 0);
5114 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5120 error = SYSCTL_IN(req, &vc, sizeof(vc));
5123 if (vc.vc_vers != VFS_CTL_VERS1)
5125 mp = vfs_getvfs(&vc.vc_fsid);
5128 /* ensure that a specific sysctl goes to the right filesystem. */
5129 if (strcmp(vc.vc_fstypename, "*") != 0 &&
5130 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
5134 VCTLTOREQ(&vc, req);
5135 error = VFS_SYSCTL(mp, vc.vc_op, req);
5140 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR,
5141 NULL, 0, sysctl_vfs_ctl, "",
5145 * Function to initialize a va_filerev field sensibly.
5146 * XXX: Wouldn't a random number make a lot more sense ??
5149 init_va_filerev(void)
5154 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5157 static int filt_vfsread(struct knote *kn, long hint);
5158 static int filt_vfswrite(struct knote *kn, long hint);
5159 static int filt_vfsvnode(struct knote *kn, long hint);
5160 static void filt_vfsdetach(struct knote *kn);
5161 static struct filterops vfsread_filtops = {
5163 .f_detach = filt_vfsdetach,
5164 .f_event = filt_vfsread
5166 static struct filterops vfswrite_filtops = {
5168 .f_detach = filt_vfsdetach,
5169 .f_event = filt_vfswrite
5171 static struct filterops vfsvnode_filtops = {
5173 .f_detach = filt_vfsdetach,
5174 .f_event = filt_vfsvnode
5178 vfs_knllock(void *arg)
5180 struct vnode *vp = arg;
5182 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5186 vfs_knlunlock(void *arg)
5188 struct vnode *vp = arg;
5194 vfs_knl_assert_locked(void *arg)
5196 #ifdef DEBUG_VFS_LOCKS
5197 struct vnode *vp = arg;
5199 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
5204 vfs_knl_assert_unlocked(void *arg)
5206 #ifdef DEBUG_VFS_LOCKS
5207 struct vnode *vp = arg;
5209 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
5214 vfs_kqfilter(struct vop_kqfilter_args *ap)
5216 struct vnode *vp = ap->a_vp;
5217 struct knote *kn = ap->a_kn;
5220 switch (kn->kn_filter) {
5222 kn->kn_fop = &vfsread_filtops;
5225 kn->kn_fop = &vfswrite_filtops;
5228 kn->kn_fop = &vfsvnode_filtops;
5234 kn->kn_hook = (caddr_t)vp;
5237 if (vp->v_pollinfo == NULL)
5239 knl = &vp->v_pollinfo->vpi_selinfo.si_note;
5241 knlist_add(knl, kn, 0);
5247 * Detach knote from vnode
5250 filt_vfsdetach(struct knote *kn)
5252 struct vnode *vp = (struct vnode *)kn->kn_hook;
5254 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5255 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5261 filt_vfsread(struct knote *kn, long hint)
5263 struct vnode *vp = (struct vnode *)kn->kn_hook;
5268 * filesystem is gone, so set the EOF flag and schedule
5269 * the knote for deletion.
5271 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5273 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5278 if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5282 kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5283 res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5290 filt_vfswrite(struct knote *kn, long hint)
5292 struct vnode *vp = (struct vnode *)kn->kn_hook;
5297 * filesystem is gone, so set the EOF flag and schedule
5298 * the knote for deletion.
5300 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5301 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5309 filt_vfsvnode(struct knote *kn, long hint)
5311 struct vnode *vp = (struct vnode *)kn->kn_hook;
5315 if (kn->kn_sfflags & hint)
5316 kn->kn_fflags |= hint;
5317 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5318 kn->kn_flags |= EV_EOF;
5322 res = (kn->kn_fflags != 0);
5328 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5332 if (dp->d_reclen > ap->a_uio->uio_resid)
5333 return (ENAMETOOLONG);
5334 error = uiomove(dp, dp->d_reclen, ap->a_uio);
5336 if (ap->a_ncookies != NULL) {
5337 if (ap->a_cookies != NULL)
5338 free(ap->a_cookies, M_TEMP);
5339 ap->a_cookies = NULL;
5340 *ap->a_ncookies = 0;
5344 if (ap->a_ncookies == NULL)
5347 KASSERT(ap->a_cookies,
5348 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
5350 *ap->a_cookies = realloc(*ap->a_cookies,
5351 (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
5352 (*ap->a_cookies)[*ap->a_ncookies] = off;
5353 *ap->a_ncookies += 1;
5358 * Mark for update the access time of the file if the filesystem
5359 * supports VOP_MARKATIME. This functionality is used by execve and
5360 * mmap, so we want to avoid the I/O implied by directly setting
5361 * va_atime for the sake of efficiency.
5364 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
5369 ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
5370 if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
5371 (void)VOP_MARKATIME(vp);
5375 * The purpose of this routine is to remove granularity from accmode_t,
5376 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
5377 * VADMIN and VAPPEND.
5379 * If it returns 0, the caller is supposed to continue with the usual
5380 * access checks using 'accmode' as modified by this routine. If it
5381 * returns nonzero value, the caller is supposed to return that value
5384 * Note that after this routine runs, accmode may be zero.
5387 vfs_unixify_accmode(accmode_t *accmode)
5390 * There is no way to specify explicit "deny" rule using
5391 * file mode or POSIX.1e ACLs.
5393 if (*accmode & VEXPLICIT_DENY) {
5399 * None of these can be translated into usual access bits.
5400 * Also, the common case for NFSv4 ACLs is to not contain
5401 * either of these bits. Caller should check for VWRITE
5402 * on the containing directory instead.
5404 if (*accmode & (VDELETE_CHILD | VDELETE))
5407 if (*accmode & VADMIN_PERMS) {
5408 *accmode &= ~VADMIN_PERMS;
5413 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5414 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5416 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5422 * These are helper functions for filesystems to traverse all
5423 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
5425 * This interface replaces MNT_VNODE_FOREACH.
5428 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
5431 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
5436 kern_yield(PRI_USER);
5438 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5439 for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
5440 vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
5441 /* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
5442 if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
5445 if ((vp->v_iflag & VI_DOOMED) != 0) {
5452 __mnt_vnode_markerfree_all(mvp, mp);
5453 /* MNT_IUNLOCK(mp); -- done in above function */
5454 mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
5457 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5458 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5464 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
5468 *mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5471 (*mvp)->v_mount = mp;
5472 (*mvp)->v_type = VMARKER;
5474 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
5475 /* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
5476 if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
5479 if ((vp->v_iflag & VI_DOOMED) != 0) {
5488 free(*mvp, M_VNODE_MARKER);
5492 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5498 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
5506 mtx_assert(MNT_MTX(mp), MA_OWNED);
5508 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5509 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5512 free(*mvp, M_VNODE_MARKER);
5517 * These are helper functions for filesystems to traverse their
5518 * active vnodes. See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h
5521 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5524 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5529 free(*mvp, M_VNODE_MARKER);
5534 * Relock the mp mount vnode list lock with the vp vnode interlock in the
5535 * conventional lock order during mnt_vnode_next_active iteration.
5537 * On entry, the mount vnode list lock is held and the vnode interlock is not.
5538 * The list lock is dropped and reacquired. On success, both locks are held.
5539 * On failure, the mount vnode list lock is held but the vnode interlock is
5540 * not, and the procedure may have yielded.
5543 mnt_vnode_next_active_relock(struct vnode *mvp, struct mount *mp,
5546 const struct vnode *tmp;
5549 VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
5550 TAILQ_NEXT(mvp, v_actfreelist) != NULL, mvp,
5551 ("%s: bad marker", __func__));
5552 VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
5553 ("%s: inappropriate vnode", __func__));
5554 ASSERT_VI_UNLOCKED(vp, __func__);
5555 mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5559 TAILQ_REMOVE(&mp->mnt_activevnodelist, mvp, v_actfreelist);
5560 TAILQ_INSERT_BEFORE(vp, mvp, v_actfreelist);
5563 * Use a hold to prevent vp from disappearing while the mount vnode
5564 * list lock is dropped and reacquired. Normally a hold would be
5565 * acquired with vhold(), but that might try to acquire the vnode
5566 * interlock, which would be a LOR with the mount vnode list lock.
5568 held = refcount_acquire_if_not_zero(&vp->v_holdcnt);
5569 mtx_unlock(&mp->mnt_listmtx);
5573 if (!refcount_release_if_not_last(&vp->v_holdcnt)) {
5577 mtx_lock(&mp->mnt_listmtx);
5580 * Determine whether the vnode is still the next one after the marker,
5581 * excepting any other markers. If the vnode has not been doomed by
5582 * vgone() then the hold should have ensured that it remained on the
5583 * active list. If it has been doomed but is still on the active list,
5584 * don't abort, but rather skip over it (avoid spinning on doomed
5589 tmp = TAILQ_NEXT(tmp, v_actfreelist);
5590 } while (tmp != NULL && tmp->v_type == VMARKER);
5592 mtx_unlock(&mp->mnt_listmtx);
5601 mtx_lock(&mp->mnt_listmtx);
5604 ASSERT_VI_LOCKED(vp, __func__);
5606 ASSERT_VI_UNLOCKED(vp, __func__);
5607 mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5611 static struct vnode *
5612 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5614 struct vnode *vp, *nvp;
5616 mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5617 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5619 vp = TAILQ_NEXT(*mvp, v_actfreelist);
5620 while (vp != NULL) {
5621 if (vp->v_type == VMARKER) {
5622 vp = TAILQ_NEXT(vp, v_actfreelist);
5626 * Try-lock because this is the wrong lock order. If that does
5627 * not succeed, drop the mount vnode list lock and try to
5628 * reacquire it and the vnode interlock in the right order.
5630 if (!VI_TRYLOCK(vp) &&
5631 !mnt_vnode_next_active_relock(*mvp, mp, vp))
5633 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
5634 KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
5635 ("alien vnode on the active list %p %p", vp, mp));
5636 if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0)
5638 nvp = TAILQ_NEXT(vp, v_actfreelist);
5642 TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5644 /* Check if we are done */
5646 mtx_unlock(&mp->mnt_listmtx);
5647 mnt_vnode_markerfree_active(mvp, mp);
5650 TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist);
5651 mtx_unlock(&mp->mnt_listmtx);
5652 ASSERT_VI_LOCKED(vp, "active iter");
5653 KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp));
5658 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5662 kern_yield(PRI_USER);
5663 mtx_lock(&mp->mnt_listmtx);
5664 return (mnt_vnode_next_active(mvp, mp));
5668 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp)
5672 *mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5676 (*mvp)->v_type = VMARKER;
5677 (*mvp)->v_mount = mp;
5679 mtx_lock(&mp->mnt_listmtx);
5680 vp = TAILQ_FIRST(&mp->mnt_activevnodelist);
5682 mtx_unlock(&mp->mnt_listmtx);
5683 mnt_vnode_markerfree_active(mvp, mp);
5686 TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5687 return (mnt_vnode_next_active(mvp, mp));
5691 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5697 mtx_lock(&mp->mnt_listmtx);
5698 TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5699 mtx_unlock(&mp->mnt_listmtx);
5700 mnt_vnode_markerfree_active(mvp, mp);