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>
68 #include <sys/lockf.h>
69 #include <sys/malloc.h>
70 #include <sys/mount.h>
71 #include <sys/namei.h>
72 #include <sys/pctrie.h>
74 #include <sys/reboot.h>
75 #include <sys/refcount.h>
76 #include <sys/rwlock.h>
77 #include <sys/sched.h>
78 #include <sys/sleepqueue.h>
81 #include <sys/sysctl.h>
82 #include <sys/syslog.h>
83 #include <sys/vmmeter.h>
84 #include <sys/vnode.h>
85 #include <sys/watchdog.h>
87 #include <machine/stdarg.h>
89 #include <security/mac/mac_framework.h>
92 #include <vm/vm_object.h>
93 #include <vm/vm_extern.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_page.h>
97 #include <vm/vm_kern.h>
104 static void delmntque(struct vnode *vp);
105 static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
106 int slpflag, int slptimeo);
107 static void syncer_shutdown(void *arg, int howto);
108 static int vtryrecycle(struct vnode *vp);
109 static void v_init_counters(struct vnode *);
110 static void v_incr_devcount(struct vnode *);
111 static void v_decr_devcount(struct vnode *);
112 static void vgonel(struct vnode *);
113 static void vfs_knllock(void *arg);
114 static void vfs_knlunlock(void *arg);
115 static void vfs_knl_assert_locked(void *arg);
116 static void vfs_knl_assert_unlocked(void *arg);
117 static void destroy_vpollinfo(struct vpollinfo *vi);
118 static int v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
119 daddr_t startlbn, daddr_t endlbn);
120 static void vnlru_recalc(void);
123 * These fences are intended for cases where some synchronization is
124 * needed between access of v_iflags and lockless vnode refcount (v_holdcnt
125 * and v_usecount) updates. Access to v_iflags is generally synchronized
126 * by the interlock, but we have some internal assertions that check vnode
127 * flags without acquiring the lock. Thus, these fences are INVARIANTS-only
131 #define VNODE_REFCOUNT_FENCE_ACQ() atomic_thread_fence_acq()
132 #define VNODE_REFCOUNT_FENCE_REL() atomic_thread_fence_rel()
134 #define VNODE_REFCOUNT_FENCE_ACQ()
135 #define VNODE_REFCOUNT_FENCE_REL()
139 * Number of vnodes in existence. Increased whenever getnewvnode()
140 * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
142 static u_long __exclusive_cache_line numvnodes;
144 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
145 "Number of vnodes in existence");
147 static counter_u64_t vnodes_created;
148 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
149 "Number of vnodes created by getnewvnode");
152 * Conversion tables for conversion from vnode types to inode formats
155 enum vtype iftovt_tab[16] = {
156 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
157 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
159 int vttoif_tab[10] = {
160 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
161 S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
165 * List of allocates vnodes in the system.
167 static TAILQ_HEAD(freelst, vnode) vnode_list;
168 static struct vnode *vnode_list_free_marker;
169 static struct vnode *vnode_list_reclaim_marker;
172 * "Free" vnode target. Free vnodes are rarely completely free, but are
173 * just ones that are cheap to recycle. Usually they are for files which
174 * have been stat'd but not read; these usually have inode and namecache
175 * data attached to them. This target is the preferred minimum size of a
176 * sub-cache consisting mostly of such files. The system balances the size
177 * of this sub-cache with its complement to try to prevent either from
178 * thrashing while the other is relatively inactive. The targets express
179 * a preference for the best balance.
181 * "Above" this target there are 2 further targets (watermarks) related
182 * to recyling of free vnodes. In the best-operating case, the cache is
183 * exactly full, the free list has size between vlowat and vhiwat above the
184 * free target, and recycling from it and normal use maintains this state.
185 * Sometimes the free list is below vlowat or even empty, but this state
186 * is even better for immediate use provided the cache is not full.
187 * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
188 * ones) to reach one of these states. The watermarks are currently hard-
189 * coded as 4% and 9% of the available space higher. These and the default
190 * of 25% for wantfreevnodes are too large if the memory size is large.
191 * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
192 * whenever vnlru_proc() becomes active.
194 static long wantfreevnodes;
195 static long __exclusive_cache_line freevnodes;
196 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
197 &freevnodes, 0, "Number of \"free\" vnodes");
198 static long freevnodes_old;
200 static counter_u64_t recycles_count;
201 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
202 "Number of vnodes recycled to meet vnode cache targets");
204 static counter_u64_t recycles_free_count;
205 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count,
206 "Number of free 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 | CTLFLAG_STATS,
215 &reassignbufcalls, 0, "Number of calls to reassignbuf");
217 static counter_u64_t deferred_inact;
218 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact,
219 "Number of times inactive processing was deferred");
221 /* To keep more than one thread at a time from running vfs_getnewfsid */
222 static struct mtx mntid_mtx;
225 * Lock for any access to the following:
230 static struct mtx __exclusive_cache_line vnode_list_mtx;
232 /* Publicly exported FS */
233 struct nfs_public nfs_pub;
235 static uma_zone_t buf_trie_zone;
237 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
238 static uma_zone_t vnode_zone;
239 static uma_zone_t vnodepoll_zone;
242 * The workitem queue.
244 * It is useful to delay writes of file data and filesystem metadata
245 * for tens of seconds so that quickly created and deleted files need
246 * not waste disk bandwidth being created and removed. To realize this,
247 * we append vnodes to a "workitem" queue. When running with a soft
248 * updates implementation, most pending metadata dependencies should
249 * not wait for more than a few seconds. Thus, mounted on block devices
250 * are delayed only about a half the time that file data is delayed.
251 * Similarly, directory updates are more critical, so are only delayed
252 * about a third the time that file data is delayed. Thus, there are
253 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
254 * one each second (driven off the filesystem syncer process). The
255 * syncer_delayno variable indicates the next queue that is to be processed.
256 * Items that need to be processed soon are placed in this queue:
258 * syncer_workitem_pending[syncer_delayno]
260 * A delay of fifteen seconds is done by placing the request fifteen
261 * entries later in the queue:
263 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
266 static int syncer_delayno;
267 static long syncer_mask;
268 LIST_HEAD(synclist, bufobj);
269 static struct synclist *syncer_workitem_pending;
271 * The sync_mtx protects:
276 * syncer_workitem_pending
277 * syncer_worklist_len
280 static struct mtx sync_mtx;
281 static struct cv sync_wakeup;
283 #define SYNCER_MAXDELAY 32
284 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */
285 static int syncdelay = 30; /* max time to delay syncing data */
286 static int filedelay = 30; /* time to delay syncing files */
287 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
288 "Time to delay syncing files (in seconds)");
289 static int dirdelay = 29; /* time to delay syncing directories */
290 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
291 "Time to delay syncing directories (in seconds)");
292 static int metadelay = 28; /* time to delay syncing metadata */
293 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
294 "Time to delay syncing metadata (in seconds)");
295 static int rushjob; /* number of slots to run ASAP */
296 static int stat_rush_requests; /* number of times I/O speeded up */
297 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
298 "Number of times I/O speeded up (rush requests)");
300 #define VDBATCH_SIZE 8
305 struct vnode *tab[VDBATCH_SIZE];
307 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
309 static void vdbatch_dequeue(struct vnode *vp);
312 * When shutting down the syncer, run it at four times normal speed.
314 #define SYNCER_SHUTDOWN_SPEEDUP 4
315 static int sync_vnode_count;
316 static int syncer_worklist_len;
317 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
320 /* Target for maximum number of vnodes. */
321 u_long desiredvnodes;
322 static u_long gapvnodes; /* gap between wanted and desired */
323 static u_long vhiwat; /* enough extras after expansion */
324 static u_long vlowat; /* minimal extras before expansion */
325 static u_long vstir; /* nonzero to stir non-free vnodes */
326 static volatile int vsmalltrigger = 8; /* pref to keep if > this many pages */
328 static u_long vnlru_read_freevnodes(void);
331 * Note that no attempt is made to sanitize these parameters.
334 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
340 error = sysctl_handle_long(oidp, &val, 0, req);
341 if (error != 0 || req->newptr == NULL)
344 if (val == desiredvnodes)
346 mtx_lock(&vnode_list_mtx);
348 wantfreevnodes = desiredvnodes / 4;
350 mtx_unlock(&vnode_list_mtx);
352 * XXX There is no protection against multiple threads changing
353 * desiredvnodes at the same time. Locking above only helps vnlru and
356 vfs_hash_changesize(desiredvnodes);
357 cache_changesize(desiredvnodes);
361 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
362 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
363 "LU", "Target for maximum number of vnodes");
366 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
371 val = wantfreevnodes;
372 error = sysctl_handle_long(oidp, &val, 0, req);
373 if (error != 0 || req->newptr == NULL)
376 if (val == wantfreevnodes)
378 mtx_lock(&vnode_list_mtx);
379 wantfreevnodes = val;
381 mtx_unlock(&vnode_list_mtx);
385 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
386 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
387 "LU", "Target for minimum number of \"free\" vnodes");
389 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
390 &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
391 static int vnlru_nowhere;
392 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
393 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
396 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
401 unsigned long ndflags;
404 if (req->newptr == NULL)
406 if (req->newlen >= PATH_MAX)
409 buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
410 error = SYSCTL_IN(req, buf, req->newlen);
414 buf[req->newlen] = '\0';
416 ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | NOCACHE | SAVENAME;
417 NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread);
418 if ((error = namei(&nd)) != 0)
422 if (VN_IS_DOOMED(vp)) {
424 * This vnode is being recycled. Return != 0 to let the caller
425 * know that the sysctl had no effect. Return EAGAIN because a
426 * subsequent call will likely succeed (since namei will create
427 * a new vnode if necessary)
433 counter_u64_add(recycles_count, 1);
443 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
445 struct thread *td = curthread;
451 if (req->newptr == NULL)
454 error = sysctl_handle_int(oidp, &fd, 0, req);
457 error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
462 error = vn_lock(vp, LK_EXCLUSIVE);
466 counter_u64_add(recycles_count, 1);
474 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
475 CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
476 sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
477 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
478 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
479 sysctl_ftry_reclaim_vnode, "I",
480 "Try to reclaim a vnode by its file descriptor");
482 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
486 * Support for the bufobj clean & dirty pctrie.
489 buf_trie_alloc(struct pctrie *ptree)
492 return uma_zalloc(buf_trie_zone, M_NOWAIT);
496 buf_trie_free(struct pctrie *ptree, void *node)
499 uma_zfree(buf_trie_zone, node);
501 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
504 * Initialize the vnode management data structures.
506 * Reevaluate the following cap on the number of vnodes after the physical
507 * memory size exceeds 512GB. In the limit, as the physical memory size
508 * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
510 #ifndef MAXVNODES_MAX
511 #define MAXVNODES_MAX (512UL * 1024 * 1024 / 64) /* 8M */
514 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
516 static struct vnode *
517 vn_alloc_marker(struct mount *mp)
521 vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
522 vp->v_type = VMARKER;
529 vn_free_marker(struct vnode *vp)
532 MPASS(vp->v_type == VMARKER);
533 free(vp, M_VNODE_MARKER);
537 * Initialize a vnode as it first enters the zone.
540 vnode_init(void *mem, int size, int flags)
549 vp->v_vnlock = &vp->v_lock;
550 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
552 * By default, don't allow shared locks unless filesystems opt-in.
554 lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
555 LK_NOSHARE | LK_IS_VNODE);
559 bufobj_init(&vp->v_bufobj, vp);
561 * Initialize namecache.
563 LIST_INIT(&vp->v_cache_src);
564 TAILQ_INIT(&vp->v_cache_dst);
566 * Initialize rangelocks.
568 rangelock_init(&vp->v_rl);
570 vp->v_dbatchcpu = NOCPU;
572 mtx_lock(&vnode_list_mtx);
573 TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
574 mtx_unlock(&vnode_list_mtx);
579 * Free a vnode when it is cleared from the zone.
582 vnode_fini(void *mem, int size)
589 mtx_lock(&vnode_list_mtx);
590 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
591 mtx_unlock(&vnode_list_mtx);
592 rangelock_destroy(&vp->v_rl);
593 lockdestroy(vp->v_vnlock);
594 mtx_destroy(&vp->v_interlock);
596 rw_destroy(BO_LOCKPTR(bo));
600 * Provide the size of NFS nclnode and NFS fh for calculation of the
601 * vnode memory consumption. The size is specified directly to
602 * eliminate dependency on NFS-private header.
604 * Other filesystems may use bigger or smaller (like UFS and ZFS)
605 * private inode data, but the NFS-based estimation is ample enough.
606 * Still, we care about differences in the size between 64- and 32-bit
609 * Namecache structure size is heuristically
610 * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
613 #define NFS_NCLNODE_SZ (528 + 64)
616 #define NFS_NCLNODE_SZ (360 + 32)
621 vntblinit(void *dummy __unused)
624 int cpu, physvnodes, virtvnodes;
628 * Desiredvnodes is a function of the physical memory size and the
629 * kernel's heap size. Generally speaking, it scales with the
630 * physical memory size. The ratio of desiredvnodes to the physical
631 * memory size is 1:16 until desiredvnodes exceeds 98,304.
633 * marginal ratio of desiredvnodes to the physical memory size is
634 * 1:64. However, desiredvnodes is limited by the kernel's heap
635 * size. The memory required by desiredvnodes vnodes and vm objects
636 * must not exceed 1/10th of the kernel's heap size.
638 physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
639 3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
640 virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
641 sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
642 desiredvnodes = min(physvnodes, virtvnodes);
643 if (desiredvnodes > MAXVNODES_MAX) {
645 printf("Reducing kern.maxvnodes %lu -> %lu\n",
646 desiredvnodes, MAXVNODES_MAX);
647 desiredvnodes = MAXVNODES_MAX;
649 wantfreevnodes = desiredvnodes / 4;
650 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
651 TAILQ_INIT(&vnode_list);
652 mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
654 * The lock is taken to appease WITNESS.
656 mtx_lock(&vnode_list_mtx);
658 mtx_unlock(&vnode_list_mtx);
659 vnode_list_free_marker = vn_alloc_marker(NULL);
660 TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
661 vnode_list_reclaim_marker = vn_alloc_marker(NULL);
662 TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
663 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
664 vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
665 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
666 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
668 * Preallocate enough nodes to support one-per buf so that
669 * we can not fail an insert. reassignbuf() callers can not
670 * tolerate the insertion failure.
672 buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
673 NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
675 uma_prealloc(buf_trie_zone, nbuf);
677 vnodes_created = counter_u64_alloc(M_WAITOK);
678 recycles_count = counter_u64_alloc(M_WAITOK);
679 recycles_free_count = counter_u64_alloc(M_WAITOK);
680 deferred_inact = counter_u64_alloc(M_WAITOK);
683 * Initialize the filesystem syncer.
685 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
687 syncer_maxdelay = syncer_mask + 1;
688 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
689 cv_init(&sync_wakeup, "syncer");
690 for (i = 1; i <= sizeof(struct vnode); i <<= 1)
695 vd = DPCPU_ID_PTR((cpu), vd);
696 bzero(vd, sizeof(*vd));
697 mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
700 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
703 * Mark a mount point as busy. Used to synchronize access and to delay
704 * unmounting. Eventually, mountlist_mtx is not released on failure.
706 * vfs_busy() is a custom lock, it can block the caller.
707 * vfs_busy() only sleeps if the unmount is active on the mount point.
708 * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
709 * vnode belonging to mp.
711 * Lookup uses vfs_busy() to traverse mount points.
713 * / vnode lock A / vnode lock (/var) D
714 * /var vnode lock B /log vnode lock(/var/log) E
715 * vfs_busy lock C vfs_busy lock F
717 * Within each file system, the lock order is C->A->B and F->D->E.
719 * When traversing across mounts, the system follows that lock order:
725 * The lookup() process for namei("/var") illustrates the process:
726 * VOP_LOOKUP() obtains B while A is held
727 * vfs_busy() obtains a shared lock on F while A and B are held
728 * vput() releases lock on B
729 * vput() releases lock on A
730 * VFS_ROOT() obtains lock on D while shared lock on F is held
731 * vfs_unbusy() releases shared lock on F
732 * vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
733 * Attempt to lock A (instead of vp_crossmp) while D is held would
734 * violate the global order, causing deadlocks.
736 * dounmount() locks B while F is drained.
739 vfs_busy(struct mount *mp, int flags)
742 MPASS((flags & ~MBF_MASK) == 0);
743 CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
745 if (vfs_op_thread_enter(mp)) {
746 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
747 MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
748 MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
749 vfs_mp_count_add_pcpu(mp, ref, 1);
750 vfs_mp_count_add_pcpu(mp, lockref, 1);
751 vfs_op_thread_exit(mp);
752 if (flags & MBF_MNTLSTLOCK)
753 mtx_unlock(&mountlist_mtx);
758 vfs_assert_mount_counters(mp);
761 * If mount point is currently being unmounted, sleep until the
762 * mount point fate is decided. If thread doing the unmounting fails,
763 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
764 * that this mount point has survived the unmount attempt and vfs_busy
765 * should retry. Otherwise the unmounter thread will set MNTK_REFEXPIRE
766 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
767 * about to be really destroyed. vfs_busy needs to release its
768 * reference on the mount point in this case and return with ENOENT,
769 * telling the caller that mount mount it tried to busy is no longer
772 while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
773 if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
776 CTR1(KTR_VFS, "%s: failed busying before sleeping",
780 if (flags & MBF_MNTLSTLOCK)
781 mtx_unlock(&mountlist_mtx);
782 mp->mnt_kern_flag |= MNTK_MWAIT;
783 msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
784 if (flags & MBF_MNTLSTLOCK)
785 mtx_lock(&mountlist_mtx);
788 if (flags & MBF_MNTLSTLOCK)
789 mtx_unlock(&mountlist_mtx);
796 * Free a busy filesystem.
799 vfs_unbusy(struct mount *mp)
803 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
805 if (vfs_op_thread_enter(mp)) {
806 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
807 vfs_mp_count_sub_pcpu(mp, lockref, 1);
808 vfs_mp_count_sub_pcpu(mp, ref, 1);
809 vfs_op_thread_exit(mp);
814 vfs_assert_mount_counters(mp);
816 c = --mp->mnt_lockref;
817 if (mp->mnt_vfs_ops == 0) {
818 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
823 vfs_dump_mount_counters(mp);
824 if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
825 MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
826 CTR1(KTR_VFS, "%s: waking up waiters", __func__);
827 mp->mnt_kern_flag &= ~MNTK_DRAINING;
828 wakeup(&mp->mnt_lockref);
834 * Lookup a mount point by filesystem identifier.
837 vfs_getvfs(fsid_t *fsid)
841 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
842 mtx_lock(&mountlist_mtx);
843 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
844 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
845 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
847 mtx_unlock(&mountlist_mtx);
851 mtx_unlock(&mountlist_mtx);
852 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
853 return ((struct mount *) 0);
857 * Lookup a mount point by filesystem identifier, busying it before
860 * To avoid congestion on mountlist_mtx, implement simple direct-mapped
861 * cache for popular filesystem identifiers. The cache is lockess, using
862 * the fact that struct mount's are never freed. In worst case we may
863 * get pointer to unmounted or even different filesystem, so we have to
864 * check what we got, and go slow way if so.
867 vfs_busyfs(fsid_t *fsid)
869 #define FSID_CACHE_SIZE 256
870 typedef struct mount * volatile vmp_t;
871 static vmp_t cache[FSID_CACHE_SIZE];
876 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
877 hash = fsid->val[0] ^ fsid->val[1];
878 hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
881 mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
882 mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
884 if (vfs_busy(mp, 0) != 0) {
888 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
889 mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
895 mtx_lock(&mountlist_mtx);
896 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
897 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
898 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
899 error = vfs_busy(mp, MBF_MNTLSTLOCK);
902 mtx_unlock(&mountlist_mtx);
909 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
910 mtx_unlock(&mountlist_mtx);
911 return ((struct mount *) 0);
915 * Check if a user can access privileged mount options.
918 vfs_suser(struct mount *mp, struct thread *td)
922 if (jailed(td->td_ucred)) {
924 * If the jail of the calling thread lacks permission for
925 * this type of file system, deny immediately.
927 if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
931 * If the file system was mounted outside the jail of the
932 * calling thread, deny immediately.
934 if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
939 * If file system supports delegated administration, we don't check
940 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
941 * by the file system itself.
942 * If this is not the user that did original mount, we check for
943 * the PRIV_VFS_MOUNT_OWNER privilege.
945 if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
946 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
947 if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
954 * Get a new unique fsid. Try to make its val[0] unique, since this value
955 * will be used to create fake device numbers for stat(). Also try (but
956 * not so hard) make its val[0] unique mod 2^16, since some emulators only
957 * support 16-bit device numbers. We end up with unique val[0]'s for the
958 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
960 * Keep in mind that several mounts may be running in parallel. Starting
961 * the search one past where the previous search terminated is both a
962 * micro-optimization and a defense against returning the same fsid to
966 vfs_getnewfsid(struct mount *mp)
968 static uint16_t mntid_base;
973 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
974 mtx_lock(&mntid_mtx);
975 mtype = mp->mnt_vfc->vfc_typenum;
976 tfsid.val[1] = mtype;
977 mtype = (mtype & 0xFF) << 24;
979 tfsid.val[0] = makedev(255,
980 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
982 if ((nmp = vfs_getvfs(&tfsid)) == NULL)
986 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
987 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
988 mtx_unlock(&mntid_mtx);
992 * Knob to control the precision of file timestamps:
994 * 0 = seconds only; nanoseconds zeroed.
995 * 1 = seconds and nanoseconds, accurate within 1/HZ.
996 * 2 = seconds and nanoseconds, truncated to microseconds.
997 * >=3 = seconds and nanoseconds, maximum precision.
999 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
1001 static int timestamp_precision = TSP_USEC;
1002 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
1003 ×tamp_precision, 0, "File timestamp precision (0: seconds, "
1004 "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1005 "3+: sec + ns (max. precision))");
1008 * Get a current timestamp.
1011 vfs_timestamp(struct timespec *tsp)
1015 switch (timestamp_precision) {
1017 tsp->tv_sec = time_second;
1025 TIMEVAL_TO_TIMESPEC(&tv, tsp);
1035 * Set vnode attributes to VNOVAL
1038 vattr_null(struct vattr *vap)
1041 vap->va_type = VNON;
1042 vap->va_size = VNOVAL;
1043 vap->va_bytes = VNOVAL;
1044 vap->va_mode = VNOVAL;
1045 vap->va_nlink = VNOVAL;
1046 vap->va_uid = VNOVAL;
1047 vap->va_gid = VNOVAL;
1048 vap->va_fsid = VNOVAL;
1049 vap->va_fileid = VNOVAL;
1050 vap->va_blocksize = VNOVAL;
1051 vap->va_rdev = VNOVAL;
1052 vap->va_atime.tv_sec = VNOVAL;
1053 vap->va_atime.tv_nsec = VNOVAL;
1054 vap->va_mtime.tv_sec = VNOVAL;
1055 vap->va_mtime.tv_nsec = VNOVAL;
1056 vap->va_ctime.tv_sec = VNOVAL;
1057 vap->va_ctime.tv_nsec = VNOVAL;
1058 vap->va_birthtime.tv_sec = VNOVAL;
1059 vap->va_birthtime.tv_nsec = VNOVAL;
1060 vap->va_flags = VNOVAL;
1061 vap->va_gen = VNOVAL;
1062 vap->va_vaflags = 0;
1066 * Try to reduce the total number of vnodes.
1068 * This routine (and its user) are buggy in at least the following ways:
1069 * - all parameters were picked years ago when RAM sizes were significantly
1071 * - it can pick vnodes based on pages used by the vm object, but filesystems
1072 * like ZFS don't use it making the pick broken
1073 * - since ZFS has its own aging policy it gets partially combated by this one
1074 * - a dedicated method should be provided for filesystems to let them decide
1075 * whether the vnode should be recycled
1077 * This routine is called when we have too many vnodes. It attempts
1078 * to free <count> vnodes and will potentially free vnodes that still
1079 * have VM backing store (VM backing store is typically the cause
1080 * of a vnode blowout so we want to do this). Therefore, this operation
1081 * is not considered cheap.
1083 * A number of conditions may prevent a vnode from being reclaimed.
1084 * the buffer cache may have references on the vnode, a directory
1085 * vnode may still have references due to the namei cache representing
1086 * underlying files, or the vnode may be in active use. It is not
1087 * desirable to reuse such vnodes. These conditions may cause the
1088 * number of vnodes to reach some minimum value regardless of what
1089 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low.
1091 * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1092 * entries if this argument is strue
1093 * @param trigger Only reclaim vnodes with fewer than this many resident
1095 * @param target How many vnodes to reclaim.
1096 * @return The number of vnodes that were reclaimed.
1099 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1101 struct vnode *vp, *mvp;
1103 struct vm_object *object;
1107 mtx_assert(&vnode_list_mtx, MA_OWNED);
1112 mvp = vnode_list_reclaim_marker;
1115 while (done < target) {
1116 vp = TAILQ_NEXT(vp, v_vnodelist);
1117 if (__predict_false(vp == NULL))
1120 if (__predict_false(vp->v_type == VMARKER))
1124 * If it's been deconstructed already, it's still
1125 * referenced, or it exceeds the trigger, skip it.
1126 * Also skip free vnodes. We are trying to make space
1127 * to expand the free list, not reduce it.
1129 if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1130 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1133 if (vp->v_type == VBAD || vp->v_type == VNON)
1136 if (!VI_TRYLOCK(vp))
1139 if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1140 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1141 VN_IS_DOOMED(vp) || vp->v_type == VNON) {
1146 object = atomic_load_ptr(&vp->v_object);
1147 if (object == NULL || object->resident_page_count > trigger) {
1154 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1155 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1156 mtx_unlock(&vnode_list_mtx);
1158 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1160 goto next_iter_unlocked;
1162 if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1164 vn_finished_write(mp);
1165 goto next_iter_unlocked;
1169 if (vp->v_usecount > 0 ||
1170 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1171 (vp->v_object != NULL &&
1172 vp->v_object->resident_page_count > trigger)) {
1175 vn_finished_write(mp);
1176 goto next_iter_unlocked;
1178 counter_u64_add(recycles_count, 1);
1182 vn_finished_write(mp);
1186 kern_yield(PRI_USER);
1187 mtx_lock(&vnode_list_mtx);
1190 MPASS(vp->v_type != VMARKER);
1191 if (!should_yield())
1193 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1194 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1195 mtx_unlock(&vnode_list_mtx);
1196 kern_yield(PRI_USER);
1197 mtx_lock(&vnode_list_mtx);
1200 if (done == 0 && !retried) {
1201 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1202 TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1209 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1210 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1212 "limit on vnode free requests per call to the vnlru_free routine");
1215 * Attempt to reduce the free list by the requested amount.
1218 vnlru_free_locked(int count, struct vfsops *mnt_op)
1220 struct vnode *vp, *mvp;
1224 mtx_assert(&vnode_list_mtx, MA_OWNED);
1225 if (count > max_vnlru_free)
1226 count = max_vnlru_free;
1228 mvp = vnode_list_free_marker;
1232 vp = TAILQ_NEXT(vp, v_vnodelist);
1233 if (__predict_false(vp == NULL)) {
1234 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1235 TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1238 if (__predict_false(vp->v_type == VMARKER))
1242 * Don't recycle if our vnode is from different type
1243 * of mount point. Note that mp is type-safe, the
1244 * check does not reach unmapped address even if
1245 * vnode is reclaimed.
1246 * Don't recycle if we can't get the interlock without
1249 if (vp->v_holdcnt > 0 || (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1250 mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) {
1253 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1254 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1255 if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1261 mtx_unlock(&vnode_list_mtx);
1265 mtx_lock(&vnode_list_mtx);
1268 return (ocount - count);
1272 vnlru_free(int count, struct vfsops *mnt_op)
1275 mtx_lock(&vnode_list_mtx);
1276 vnlru_free_locked(count, mnt_op);
1277 mtx_unlock(&vnode_list_mtx);
1284 mtx_assert(&vnode_list_mtx, MA_OWNED);
1285 gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1286 vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1287 vlowat = vhiwat / 2;
1291 * Attempt to recycle vnodes in a context that is always safe to block.
1292 * Calling vlrurecycle() from the bowels of filesystem code has some
1293 * interesting deadlock problems.
1295 static struct proc *vnlruproc;
1296 static int vnlruproc_sig;
1299 * The main freevnodes counter is only updated when threads requeue their vnode
1300 * batches. CPUs are conditionally walked to compute a more accurate total.
1302 * Limit how much of a slop are we willing to tolerate. Note: the actual value
1303 * at any given moment can still exceed slop, but it should not be by significant
1304 * margin in practice.
1306 #define VNLRU_FREEVNODES_SLOP 128
1309 vnlru_read_freevnodes(void)
1315 mtx_assert(&vnode_list_mtx, MA_OWNED);
1316 if (freevnodes > freevnodes_old)
1317 slop = freevnodes - freevnodes_old;
1319 slop = freevnodes_old - freevnodes;
1320 if (slop < VNLRU_FREEVNODES_SLOP)
1321 return (freevnodes >= 0 ? freevnodes : 0);
1322 freevnodes_old = freevnodes;
1324 vd = DPCPU_ID_PTR((cpu), vd);
1325 freevnodes_old += vd->freevnodes;
1327 return (freevnodes_old >= 0 ? freevnodes_old : 0);
1331 vnlru_under(u_long rnumvnodes, u_long limit)
1333 u_long rfreevnodes, space;
1335 if (__predict_false(rnumvnodes > desiredvnodes))
1338 space = desiredvnodes - rnumvnodes;
1339 if (space < limit) {
1340 rfreevnodes = vnlru_read_freevnodes();
1341 if (rfreevnodes > wantfreevnodes)
1342 space += rfreevnodes - wantfreevnodes;
1344 return (space < limit);
1348 vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1350 long rfreevnodes, space;
1352 if (__predict_false(rnumvnodes > desiredvnodes))
1355 space = desiredvnodes - rnumvnodes;
1356 if (space < limit) {
1357 rfreevnodes = atomic_load_long(&freevnodes);
1358 if (rfreevnodes > wantfreevnodes)
1359 space += rfreevnodes - wantfreevnodes;
1361 return (space < limit);
1368 mtx_assert(&vnode_list_mtx, MA_OWNED);
1369 if (vnlruproc_sig == 0) {
1378 u_long rnumvnodes, rfreevnodes, target;
1379 unsigned long onumvnodes;
1380 int done, force, trigger, usevnodes;
1381 bool reclaim_nc_src, want_reread;
1383 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1384 SHUTDOWN_PRI_FIRST);
1387 want_reread = false;
1389 kproc_suspend_check(vnlruproc);
1390 mtx_lock(&vnode_list_mtx);
1391 rnumvnodes = atomic_load_long(&numvnodes);
1394 force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1395 want_reread = false;
1399 * If numvnodes is too large (due to desiredvnodes being
1400 * adjusted using its sysctl, or emergency growth), first
1401 * try to reduce it by discarding from the free list.
1403 if (rnumvnodes > desiredvnodes) {
1404 vnlru_free_locked(rnumvnodes - desiredvnodes, NULL);
1405 rnumvnodes = atomic_load_long(&numvnodes);
1408 * Sleep if the vnode cache is in a good state. This is
1409 * when it is not over-full and has space for about a 4%
1410 * or 9% expansion (by growing its size or inexcessively
1411 * reducing its free list). Otherwise, try to reclaim
1412 * space for a 10% expansion.
1414 if (vstir && force == 0) {
1418 if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1420 wakeup(&vnlruproc_sig);
1421 msleep(vnlruproc, &vnode_list_mtx,
1422 PVFS|PDROP, "vlruwt", hz);
1425 rfreevnodes = vnlru_read_freevnodes();
1427 onumvnodes = rnumvnodes;
1429 * Calculate parameters for recycling. These are the same
1430 * throughout the loop to give some semblance of fairness.
1431 * The trigger point is to avoid recycling vnodes with lots
1432 * of resident pages. We aren't trying to free memory; we
1433 * are trying to recycle or at least free vnodes.
1435 if (rnumvnodes <= desiredvnodes)
1436 usevnodes = rnumvnodes - rfreevnodes;
1438 usevnodes = rnumvnodes;
1442 * The trigger value is is chosen to give a conservatively
1443 * large value to ensure that it alone doesn't prevent
1444 * making progress. The value can easily be so large that
1445 * it is effectively infinite in some congested and
1446 * misconfigured cases, and this is necessary. Normally
1447 * it is about 8 to 100 (pages), which is quite large.
1449 trigger = vm_cnt.v_page_count * 2 / usevnodes;
1451 trigger = vsmalltrigger;
1452 reclaim_nc_src = force >= 3;
1453 target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1454 target = target / 10 + 1;
1455 done = vlrureclaim(reclaim_nc_src, trigger, target);
1456 mtx_unlock(&vnode_list_mtx);
1457 if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1458 uma_reclaim(UMA_RECLAIM_DRAIN);
1460 if (force == 0 || force == 1) {
1471 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1474 kern_yield(PRI_USER);
1479 static struct kproc_desc vnlru_kp = {
1484 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1488 * Routines having to do with the management of the vnode table.
1492 * Try to recycle a freed vnode. We abort if anyone picks up a reference
1493 * before we actually vgone(). This function must be called with the vnode
1494 * held to prevent the vnode from being returned to the free list midway
1498 vtryrecycle(struct vnode *vp)
1502 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1503 VNASSERT(vp->v_holdcnt, vp,
1504 ("vtryrecycle: Recycling vp %p without a reference.", vp));
1506 * This vnode may found and locked via some other list, if so we
1507 * can't recycle it yet.
1509 if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1511 "%s: impossible to recycle, vp %p lock is already held",
1513 return (EWOULDBLOCK);
1516 * Don't recycle if its filesystem is being suspended.
1518 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1521 "%s: impossible to recycle, cannot start the write for %p",
1526 * If we got this far, we need to acquire the interlock and see if
1527 * anyone picked up this vnode from another list. If not, we will
1528 * mark it with DOOMED via vgonel() so that anyone who does find it
1529 * will skip over it.
1532 if (vp->v_usecount) {
1535 vn_finished_write(vnmp);
1537 "%s: impossible to recycle, %p is already referenced",
1541 if (!VN_IS_DOOMED(vp)) {
1542 counter_u64_add(recycles_free_count, 1);
1547 vn_finished_write(vnmp);
1552 * Allocate a new vnode.
1554 * The operation never returns an error. Returning an error was disabled
1555 * in r145385 (dated 2005) with the following comment:
1557 * XXX Not all VFS_VGET/ffs_vget callers check returns.
1559 * Given the age of this commit (almost 15 years at the time of writing this
1560 * comment) restoring the ability to fail requires a significant audit of
1563 * The routine can try to free a vnode or stall for up to 1 second waiting for
1564 * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1566 static u_long vn_alloc_cyclecount;
1568 static struct vnode * __noinline
1569 vn_alloc_hard(struct mount *mp)
1571 u_long rnumvnodes, rfreevnodes;
1573 mtx_lock(&vnode_list_mtx);
1574 rnumvnodes = atomic_load_long(&numvnodes);
1575 if (rnumvnodes + 1 < desiredvnodes) {
1576 vn_alloc_cyclecount = 0;
1579 rfreevnodes = vnlru_read_freevnodes();
1580 if (vn_alloc_cyclecount++ >= rfreevnodes) {
1581 vn_alloc_cyclecount = 0;
1585 * Grow the vnode cache if it will not be above its target max
1586 * after growing. Otherwise, if the free list is nonempty, try
1587 * to reclaim 1 item from it before growing the cache (possibly
1588 * above its target max if the reclamation failed or is delayed).
1589 * Otherwise, wait for some space. In all cases, schedule
1590 * vnlru_proc() if we are getting short of space. The watermarks
1591 * should be chosen so that we never wait or even reclaim from
1592 * the free list to below its target minimum.
1594 if (vnlru_free_locked(1, NULL) > 0)
1596 if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1598 * Wait for space for a new vnode.
1601 msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1602 if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1603 vnlru_read_freevnodes() > 1)
1604 vnlru_free_locked(1, NULL);
1607 rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1608 if (vnlru_under(rnumvnodes, vlowat))
1610 mtx_unlock(&vnode_list_mtx);
1611 return (uma_zalloc(vnode_zone, M_WAITOK));
1614 static struct vnode *
1615 vn_alloc(struct mount *mp)
1619 if (__predict_false(vn_alloc_cyclecount != 0))
1620 return (vn_alloc_hard(mp));
1621 rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1622 if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1623 atomic_subtract_long(&numvnodes, 1);
1624 return (vn_alloc_hard(mp));
1627 return (uma_zalloc(vnode_zone, M_WAITOK));
1631 vn_free(struct vnode *vp)
1634 atomic_subtract_long(&numvnodes, 1);
1635 uma_zfree(vnode_zone, vp);
1639 * Return the next vnode from the free list.
1642 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1647 struct lock_object *lo;
1649 CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1651 KASSERT(vops->registered,
1652 ("%s: not registered vector op %p\n", __func__, vops));
1655 if (td->td_vp_reserved != NULL) {
1656 vp = td->td_vp_reserved;
1657 td->td_vp_reserved = NULL;
1661 counter_u64_add(vnodes_created, 1);
1663 * Locks are given the generic name "vnode" when created.
1664 * Follow the historic practice of using the filesystem
1665 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1667 * Locks live in a witness group keyed on their name. Thus,
1668 * when a lock is renamed, it must also move from the witness
1669 * group of its old name to the witness group of its new name.
1671 * The change only needs to be made when the vnode moves
1672 * from one filesystem type to another. We ensure that each
1673 * filesystem use a single static name pointer for its tag so
1674 * that we can compare pointers rather than doing a strcmp().
1676 lo = &vp->v_vnlock->lock_object;
1678 if (lo->lo_name != tag) {
1682 WITNESS_DESTROY(lo);
1683 WITNESS_INIT(lo, tag);
1687 * By default, don't allow shared locks unless filesystems opt-in.
1689 vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1691 * Finalize various vnode identity bits.
1693 KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1694 KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1695 KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1698 v_init_counters(vp);
1699 vp->v_bufobj.bo_ops = &buf_ops_bio;
1701 if (mp == NULL && vops != &dead_vnodeops)
1702 printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1706 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1707 mac_vnode_associate_singlelabel(mp, vp);
1710 vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1711 if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1712 vp->v_vflag |= VV_NOKNOTE;
1716 * For the filesystems which do not use vfs_hash_insert(),
1717 * still initialize v_hash to have vfs_hash_index() useful.
1718 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1721 vp->v_hash = (uintptr_t)vp >> vnsz2log;
1728 getnewvnode_reserve(void)
1733 MPASS(td->td_vp_reserved == NULL);
1734 td->td_vp_reserved = vn_alloc(NULL);
1738 getnewvnode_drop_reserve(void)
1743 if (td->td_vp_reserved != NULL) {
1744 vn_free(td->td_vp_reserved);
1745 td->td_vp_reserved = NULL;
1750 freevnode(struct vnode *vp)
1755 * The vnode has been marked for destruction, so free it.
1757 * The vnode will be returned to the zone where it will
1758 * normally remain until it is needed for another vnode. We
1759 * need to cleanup (or verify that the cleanup has already
1760 * been done) any residual data left from its current use
1761 * so as not to contaminate the freshly allocated vnode.
1763 CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1765 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1766 VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
1767 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1768 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1769 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1770 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1771 VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1772 ("clean blk trie not empty"));
1773 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1774 VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1775 ("dirty blk trie not empty"));
1776 VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1777 VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1778 VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1779 VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1780 ("Dangling rangelock waiters"));
1783 mac_vnode_destroy(vp);
1785 if (vp->v_pollinfo != NULL) {
1786 destroy_vpollinfo(vp->v_pollinfo);
1787 vp->v_pollinfo = NULL;
1790 /* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
1793 vp->v_mountedhere = NULL;
1796 vp->v_fifoinfo = NULL;
1797 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
1806 * Delete from old mount point vnode list, if on one.
1809 delmntque(struct vnode *vp)
1813 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
1822 VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1823 ("bad mount point vnode list size"));
1824 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1825 mp->mnt_nvnodelistsize--;
1831 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1835 vp->v_op = &dead_vnodeops;
1841 * Insert into list of vnodes for the new mount point, if available.
1844 insmntque1(struct vnode *vp, struct mount *mp,
1845 void (*dtr)(struct vnode *, void *), void *dtr_arg)
1848 KASSERT(vp->v_mount == NULL,
1849 ("insmntque: vnode already on per mount vnode list"));
1850 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1851 ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1854 * We acquire the vnode interlock early to ensure that the
1855 * vnode cannot be recycled by another process releasing a
1856 * holdcnt on it before we get it on both the vnode list
1857 * and the active vnode list. The mount mutex protects only
1858 * manipulation of the vnode list and the vnode freelist
1859 * mutex protects only manipulation of the active vnode list.
1860 * Hence the need to hold the vnode interlock throughout.
1864 if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1865 ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1866 mp->mnt_nvnodelistsize == 0)) &&
1867 (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1876 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1877 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1878 ("neg mount point vnode list size"));
1879 mp->mnt_nvnodelistsize++;
1886 insmntque(struct vnode *vp, struct mount *mp)
1889 return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1893 * Flush out and invalidate all buffers associated with a bufobj
1894 * Called with the underlying object locked.
1897 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1902 if (flags & V_SAVE) {
1903 error = bufobj_wwait(bo, slpflag, slptimeo);
1908 if (bo->bo_dirty.bv_cnt > 0) {
1910 if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1913 * XXX We could save a lock/unlock if this was only
1914 * enabled under INVARIANTS
1917 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1918 panic("vinvalbuf: dirty bufs");
1922 * If you alter this loop please notice that interlock is dropped and
1923 * reacquired in flushbuflist. Special care is needed to ensure that
1924 * no race conditions occur from this.
1927 error = flushbuflist(&bo->bo_clean,
1928 flags, bo, slpflag, slptimeo);
1929 if (error == 0 && !(flags & V_CLEANONLY))
1930 error = flushbuflist(&bo->bo_dirty,
1931 flags, bo, slpflag, slptimeo);
1932 if (error != 0 && error != EAGAIN) {
1936 } while (error != 0);
1939 * Wait for I/O to complete. XXX needs cleaning up. The vnode can
1940 * have write I/O in-progress but if there is a VM object then the
1941 * VM object can also have read-I/O in-progress.
1944 bufobj_wwait(bo, 0, 0);
1945 if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
1947 vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
1950 } while (bo->bo_numoutput > 0);
1954 * Destroy the copy in the VM cache, too.
1956 if (bo->bo_object != NULL &&
1957 (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1958 VM_OBJECT_WLOCK(bo->bo_object);
1959 vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1960 OBJPR_CLEANONLY : 0);
1961 VM_OBJECT_WUNLOCK(bo->bo_object);
1966 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1967 V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1968 bo->bo_clean.bv_cnt > 0))
1969 panic("vinvalbuf: flush failed");
1970 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1971 bo->bo_dirty.bv_cnt > 0)
1972 panic("vinvalbuf: flush dirty failed");
1979 * Flush out and invalidate all buffers associated with a vnode.
1980 * Called with the underlying object locked.
1983 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1986 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1987 ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1988 if (vp->v_object != NULL && vp->v_object->handle != vp)
1990 return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1994 * Flush out buffers on the specified list.
1998 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2001 struct buf *bp, *nbp;
2006 ASSERT_BO_WLOCKED(bo);
2009 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2011 * If we are flushing both V_NORMAL and V_ALT buffers then
2012 * do not skip any buffers. If we are flushing only V_NORMAL
2013 * buffers then skip buffers marked as BX_ALTDATA. If we are
2014 * flushing only V_ALT buffers then skip buffers not marked
2017 if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2018 (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2019 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2023 lblkno = nbp->b_lblkno;
2024 xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2027 error = BUF_TIMELOCK(bp,
2028 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2029 "flushbuf", slpflag, slptimeo);
2032 return (error != ENOLCK ? error : EAGAIN);
2034 KASSERT(bp->b_bufobj == bo,
2035 ("bp %p wrong b_bufobj %p should be %p",
2036 bp, bp->b_bufobj, bo));
2038 * XXX Since there are no node locks for NFS, I
2039 * believe there is a slight chance that a delayed
2040 * write will occur while sleeping just above, so
2043 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2046 bp->b_flags |= B_ASYNC;
2049 return (EAGAIN); /* XXX: why not loop ? */
2052 bp->b_flags |= (B_INVAL | B_RELBUF);
2053 bp->b_flags &= ~B_ASYNC;
2058 nbp = gbincore(bo, lblkno);
2059 if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2061 break; /* nbp invalid */
2067 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2073 ASSERT_BO_LOCKED(bo);
2075 for (lblkno = startn;;) {
2077 bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2078 if (bp == NULL || bp->b_lblkno >= endn ||
2079 bp->b_lblkno < startn)
2081 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2082 LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2085 if (error == ENOLCK)
2089 KASSERT(bp->b_bufobj == bo,
2090 ("bp %p wrong b_bufobj %p should be %p",
2091 bp, bp->b_bufobj, bo));
2092 lblkno = bp->b_lblkno + 1;
2093 if ((bp->b_flags & B_MANAGED) == 0)
2095 bp->b_flags |= B_RELBUF;
2097 * In the VMIO case, use the B_NOREUSE flag to hint that the
2098 * pages backing each buffer in the range are unlikely to be
2099 * reused. Dirty buffers will have the hint applied once
2100 * they've been written.
2102 if ((bp->b_flags & B_VMIO) != 0)
2103 bp->b_flags |= B_NOREUSE;
2111 * Truncate a file's buffer and pages to a specified length. This
2112 * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2116 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2118 struct buf *bp, *nbp;
2122 CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2123 vp, blksize, (uintmax_t)length);
2126 * Round up to the *next* lbn.
2128 startlbn = howmany(length, blksize);
2130 ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2136 while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2141 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2142 if (bp->b_lblkno > 0)
2145 * Since we hold the vnode lock this should only
2146 * fail if we're racing with the buf daemon.
2149 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2150 BO_LOCKPTR(bo)) == ENOLCK)
2151 goto restart_unlocked;
2153 VNASSERT((bp->b_flags & B_DELWRI), vp,
2154 ("buf(%p) on dirty queue without DELWRI", bp));
2163 bufobj_wwait(bo, 0, 0);
2165 vnode_pager_setsize(vp, length);
2171 * Invalidate the cached pages of a file's buffer within the range of block
2172 * numbers [startlbn, endlbn).
2175 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2181 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2183 start = blksize * startlbn;
2184 end = blksize * endlbn;
2188 MPASS(blksize == bo->bo_bsize);
2190 while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2194 vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2198 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2199 daddr_t startlbn, daddr_t endlbn)
2201 struct buf *bp, *nbp;
2204 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2205 ASSERT_BO_LOCKED(bo);
2209 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2210 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2213 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2214 BO_LOCKPTR(bo)) == ENOLCK) {
2220 bp->b_flags |= B_INVAL | B_RELBUF;
2221 bp->b_flags &= ~B_ASYNC;
2227 (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2229 (nbp->b_flags & B_DELWRI) != 0))
2233 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2234 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2237 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2238 BO_LOCKPTR(bo)) == ENOLCK) {
2243 bp->b_flags |= B_INVAL | B_RELBUF;
2244 bp->b_flags &= ~B_ASYNC;
2250 (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2251 (nbp->b_vp != vp) ||
2252 (nbp->b_flags & B_DELWRI) == 0))
2260 buf_vlist_remove(struct buf *bp)
2264 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2265 ASSERT_BO_WLOCKED(bp->b_bufobj);
2266 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
2267 (BX_VNDIRTY|BX_VNCLEAN),
2268 ("buf_vlist_remove: Buf %p is on two lists", bp));
2269 if (bp->b_xflags & BX_VNDIRTY)
2270 bv = &bp->b_bufobj->bo_dirty;
2272 bv = &bp->b_bufobj->bo_clean;
2273 BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2274 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2276 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2280 * Add the buffer to the sorted clean or dirty block list.
2282 * NOTE: xflags is passed as a constant, optimizing this inline function!
2285 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2291 ASSERT_BO_WLOCKED(bo);
2292 KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2293 ("buf_vlist_add: bo %p does not allow bufs", bo));
2294 KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2295 ("dead bo %p", bo));
2296 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2297 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2298 bp->b_xflags |= xflags;
2299 if (xflags & BX_VNDIRTY)
2305 * Keep the list ordered. Optimize empty list insertion. Assume
2306 * we tend to grow at the tail so lookup_le should usually be cheaper
2309 if (bv->bv_cnt == 0 ||
2310 bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2311 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2312 else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2313 TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2315 TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2316 error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2318 panic("buf_vlist_add: Preallocated nodes insufficient.");
2323 * Look up a buffer using the buffer tries.
2326 gbincore(struct bufobj *bo, daddr_t lblkno)
2330 ASSERT_BO_LOCKED(bo);
2331 bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2334 return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
2338 * Associate a buffer with a vnode.
2341 bgetvp(struct vnode *vp, struct buf *bp)
2346 ASSERT_BO_WLOCKED(bo);
2347 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2349 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2350 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2351 ("bgetvp: bp already attached! %p", bp));
2357 * Insert onto list for new vnode.
2359 buf_vlist_add(bp, bo, BX_VNCLEAN);
2363 * Disassociate a buffer from a vnode.
2366 brelvp(struct buf *bp)
2371 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2372 KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2375 * Delete from old vnode list, if on one.
2377 vp = bp->b_vp; /* XXX */
2380 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2381 buf_vlist_remove(bp);
2383 panic("brelvp: Buffer %p not on queue.", bp);
2384 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2385 bo->bo_flag &= ~BO_ONWORKLST;
2386 mtx_lock(&sync_mtx);
2387 LIST_REMOVE(bo, bo_synclist);
2388 syncer_worklist_len--;
2389 mtx_unlock(&sync_mtx);
2392 bp->b_bufobj = NULL;
2398 * Add an item to the syncer work queue.
2401 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2405 ASSERT_BO_WLOCKED(bo);
2407 mtx_lock(&sync_mtx);
2408 if (bo->bo_flag & BO_ONWORKLST)
2409 LIST_REMOVE(bo, bo_synclist);
2411 bo->bo_flag |= BO_ONWORKLST;
2412 syncer_worklist_len++;
2415 if (delay > syncer_maxdelay - 2)
2416 delay = syncer_maxdelay - 2;
2417 slot = (syncer_delayno + delay) & syncer_mask;
2419 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2420 mtx_unlock(&sync_mtx);
2424 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2428 mtx_lock(&sync_mtx);
2429 len = syncer_worklist_len - sync_vnode_count;
2430 mtx_unlock(&sync_mtx);
2431 error = SYSCTL_OUT(req, &len, sizeof(len));
2435 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2436 CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2437 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2439 static struct proc *updateproc;
2440 static void sched_sync(void);
2441 static struct kproc_desc up_kp = {
2446 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2449 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2454 *bo = LIST_FIRST(slp);
2458 if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2461 * We use vhold in case the vnode does not
2462 * successfully sync. vhold prevents the vnode from
2463 * going away when we unlock the sync_mtx so that
2464 * we can acquire the vnode interlock.
2467 mtx_unlock(&sync_mtx);
2469 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2471 mtx_lock(&sync_mtx);
2472 return (*bo == LIST_FIRST(slp));
2474 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2475 (void) VOP_FSYNC(vp, MNT_LAZY, td);
2477 vn_finished_write(mp);
2479 if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2481 * Put us back on the worklist. The worklist
2482 * routine will remove us from our current
2483 * position and then add us back in at a later
2486 vn_syncer_add_to_worklist(*bo, syncdelay);
2490 mtx_lock(&sync_mtx);
2494 static int first_printf = 1;
2497 * System filesystem synchronizer daemon.
2502 struct synclist *next, *slp;
2505 struct thread *td = curthread;
2507 int net_worklist_len;
2508 int syncer_final_iter;
2512 syncer_final_iter = 0;
2513 syncer_state = SYNCER_RUNNING;
2514 starttime = time_uptime;
2515 td->td_pflags |= TDP_NORUNNINGBUF;
2517 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2520 mtx_lock(&sync_mtx);
2522 if (syncer_state == SYNCER_FINAL_DELAY &&
2523 syncer_final_iter == 0) {
2524 mtx_unlock(&sync_mtx);
2525 kproc_suspend_check(td->td_proc);
2526 mtx_lock(&sync_mtx);
2528 net_worklist_len = syncer_worklist_len - sync_vnode_count;
2529 if (syncer_state != SYNCER_RUNNING &&
2530 starttime != time_uptime) {
2532 printf("\nSyncing disks, vnodes remaining... ");
2535 printf("%d ", net_worklist_len);
2537 starttime = time_uptime;
2540 * Push files whose dirty time has expired. Be careful
2541 * of interrupt race on slp queue.
2543 * Skip over empty worklist slots when shutting down.
2546 slp = &syncer_workitem_pending[syncer_delayno];
2547 syncer_delayno += 1;
2548 if (syncer_delayno == syncer_maxdelay)
2550 next = &syncer_workitem_pending[syncer_delayno];
2552 * If the worklist has wrapped since the
2553 * it was emptied of all but syncer vnodes,
2554 * switch to the FINAL_DELAY state and run
2555 * for one more second.
2557 if (syncer_state == SYNCER_SHUTTING_DOWN &&
2558 net_worklist_len == 0 &&
2559 last_work_seen == syncer_delayno) {
2560 syncer_state = SYNCER_FINAL_DELAY;
2561 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2563 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2564 syncer_worklist_len > 0);
2567 * Keep track of the last time there was anything
2568 * on the worklist other than syncer vnodes.
2569 * Return to the SHUTTING_DOWN state if any
2572 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2573 last_work_seen = syncer_delayno;
2574 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2575 syncer_state = SYNCER_SHUTTING_DOWN;
2576 while (!LIST_EMPTY(slp)) {
2577 error = sync_vnode(slp, &bo, td);
2579 LIST_REMOVE(bo, bo_synclist);
2580 LIST_INSERT_HEAD(next, bo, bo_synclist);
2584 if (first_printf == 0) {
2586 * Drop the sync mutex, because some watchdog
2587 * drivers need to sleep while patting
2589 mtx_unlock(&sync_mtx);
2590 wdog_kern_pat(WD_LASTVAL);
2591 mtx_lock(&sync_mtx);
2595 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2596 syncer_final_iter--;
2598 * The variable rushjob allows the kernel to speed up the
2599 * processing of the filesystem syncer process. A rushjob
2600 * value of N tells the filesystem syncer to process the next
2601 * N seconds worth of work on its queue ASAP. Currently rushjob
2602 * is used by the soft update code to speed up the filesystem
2603 * syncer process when the incore state is getting so far
2604 * ahead of the disk that the kernel memory pool is being
2605 * threatened with exhaustion.
2612 * Just sleep for a short period of time between
2613 * iterations when shutting down to allow some I/O
2616 * If it has taken us less than a second to process the
2617 * current work, then wait. Otherwise start right over
2618 * again. We can still lose time if any single round
2619 * takes more than two seconds, but it does not really
2620 * matter as we are just trying to generally pace the
2621 * filesystem activity.
2623 if (syncer_state != SYNCER_RUNNING ||
2624 time_uptime == starttime) {
2626 sched_prio(td, PPAUSE);
2629 if (syncer_state != SYNCER_RUNNING)
2630 cv_timedwait(&sync_wakeup, &sync_mtx,
2631 hz / SYNCER_SHUTDOWN_SPEEDUP);
2632 else if (time_uptime == starttime)
2633 cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2638 * Request the syncer daemon to speed up its work.
2639 * We never push it to speed up more than half of its
2640 * normal turn time, otherwise it could take over the cpu.
2643 speedup_syncer(void)
2647 mtx_lock(&sync_mtx);
2648 if (rushjob < syncdelay / 2) {
2650 stat_rush_requests += 1;
2653 mtx_unlock(&sync_mtx);
2654 cv_broadcast(&sync_wakeup);
2659 * Tell the syncer to speed up its work and run though its work
2660 * list several times, then tell it to shut down.
2663 syncer_shutdown(void *arg, int howto)
2666 if (howto & RB_NOSYNC)
2668 mtx_lock(&sync_mtx);
2669 syncer_state = SYNCER_SHUTTING_DOWN;
2671 mtx_unlock(&sync_mtx);
2672 cv_broadcast(&sync_wakeup);
2673 kproc_shutdown(arg, howto);
2677 syncer_suspend(void)
2680 syncer_shutdown(updateproc, 0);
2687 mtx_lock(&sync_mtx);
2689 syncer_state = SYNCER_RUNNING;
2690 mtx_unlock(&sync_mtx);
2691 cv_broadcast(&sync_wakeup);
2692 kproc_resume(updateproc);
2696 * Reassign a buffer from one vnode to another.
2697 * Used to assign file specific control information
2698 * (indirect blocks) to the vnode to which they belong.
2701 reassignbuf(struct buf *bp)
2714 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2715 bp, bp->b_vp, bp->b_flags);
2717 * B_PAGING flagged buffers cannot be reassigned because their vp
2718 * is not fully linked in.
2720 if (bp->b_flags & B_PAGING)
2721 panic("cannot reassign paging buffer");
2724 * Delete from old vnode list, if on one.
2727 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2728 buf_vlist_remove(bp);
2730 panic("reassignbuf: Buffer %p not on queue.", bp);
2732 * If dirty, put on list of dirty buffers; otherwise insert onto list
2735 if (bp->b_flags & B_DELWRI) {
2736 if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2737 switch (vp->v_type) {
2747 vn_syncer_add_to_worklist(bo, delay);
2749 buf_vlist_add(bp, bo, BX_VNDIRTY);
2751 buf_vlist_add(bp, bo, BX_VNCLEAN);
2753 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2754 mtx_lock(&sync_mtx);
2755 LIST_REMOVE(bo, bo_synclist);
2756 syncer_worklist_len--;
2757 mtx_unlock(&sync_mtx);
2758 bo->bo_flag &= ~BO_ONWORKLST;
2763 bp = TAILQ_FIRST(&bv->bv_hd);
2764 KASSERT(bp == NULL || bp->b_bufobj == bo,
2765 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2766 bp = TAILQ_LAST(&bv->bv_hd, buflists);
2767 KASSERT(bp == NULL || bp->b_bufobj == bo,
2768 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2770 bp = TAILQ_FIRST(&bv->bv_hd);
2771 KASSERT(bp == NULL || bp->b_bufobj == bo,
2772 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2773 bp = TAILQ_LAST(&bv->bv_hd, buflists);
2774 KASSERT(bp == NULL || bp->b_bufobj == bo,
2775 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2781 v_init_counters(struct vnode *vp)
2784 VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2785 vp, ("%s called for an initialized vnode", __FUNCTION__));
2786 ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2788 refcount_init(&vp->v_holdcnt, 1);
2789 refcount_init(&vp->v_usecount, 1);
2793 * Increment si_usecount of the associated device, if any.
2796 v_incr_devcount(struct vnode *vp)
2799 ASSERT_VI_LOCKED(vp, __FUNCTION__);
2800 if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2802 vp->v_rdev->si_usecount++;
2808 * Decrement si_usecount of the associated device, if any.
2810 * The caller is required to hold the interlock when transitioning a VCHR use
2811 * count to zero. This prevents a race with devfs_reclaim_vchr() that would
2812 * leak a si_usecount reference. The vnode lock will also prevent this race
2813 * if it is held while dropping the last ref.
2818 * devfs_reclaim_vchr
2819 * make v_usecount == 0
2821 * sees v_usecount == 0, no updates
2822 * vp->v_rdev = NULL;
2827 * sees v_rdev == NULL, no updates
2829 * In this scenario si_devcount decrement is not performed.
2832 v_decr_devcount(struct vnode *vp)
2835 ASSERT_VOP_LOCKED(vp, __func__);
2836 ASSERT_VI_LOCKED(vp, __FUNCTION__);
2837 if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2839 VNPASS(vp->v_rdev->si_usecount > 0, vp);
2840 vp->v_rdev->si_usecount--;
2846 * Grab a particular vnode from the free list, increment its
2847 * reference count and lock it. VIRF_DOOMED is set if the vnode
2848 * is being destroyed. Only callers who specify LK_RETRY will
2849 * see doomed vnodes. If inactive processing was delayed in
2850 * vput try to do it here.
2852 * usecount is manipulated using atomics without holding any locks.
2854 * holdcnt can be manipulated using atomics without holding any locks,
2855 * except when transitioning 1<->0, in which case the interlock is held.
2858 vget_prep(struct vnode *vp)
2862 if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2872 vget(struct vnode *vp, int flags, struct thread *td)
2876 MPASS(td == curthread);
2879 return (vget_finish(vp, flags, vs));
2882 static int __noinline
2883 vget_finish_vchr(struct vnode *vp)
2886 VNASSERT(vp->v_type == VCHR, vp, ("type != VCHR)"));
2889 * See the comment in vget_finish before usecount bump.
2891 if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2893 int old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2894 VNASSERT(old > 0, vp, ("%s: wrong hold count %d", __func__, old));
2896 refcount_release(&vp->v_holdcnt);
2902 if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2904 int old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2905 VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2907 refcount_release(&vp->v_holdcnt);
2912 v_incr_devcount(vp);
2913 refcount_acquire(&vp->v_usecount);
2919 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2923 if ((flags & LK_INTERLOCK) != 0)
2924 ASSERT_VI_LOCKED(vp, __func__);
2926 ASSERT_VI_UNLOCKED(vp, __func__);
2927 VNPASS(vp->v_holdcnt > 0, vp);
2928 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
2930 error = vn_lock(vp, flags);
2931 if (__predict_false(error != 0)) {
2932 if (vs == VGET_USECOUNT)
2936 CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2941 if (vs == VGET_USECOUNT)
2944 if (__predict_false(vp->v_type == VCHR))
2945 return (vget_finish_vchr(vp));
2948 * We hold the vnode. If the usecount is 0 it will be utilized to keep
2949 * the vnode around. Otherwise someone else lended their hold count and
2950 * we have to drop ours.
2952 old = atomic_fetchadd_int(&vp->v_usecount, 1);
2953 VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
2956 old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2957 VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2959 refcount_release(&vp->v_holdcnt);
2966 * Increase the reference (use) and hold count of a vnode.
2967 * This will also remove the vnode from the free list if it is presently free.
2969 static void __noinline
2970 vref_vchr(struct vnode *vp, bool interlock)
2974 * See the comment in vget_finish before usecount bump.
2977 if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2978 VNODE_REFCOUNT_FENCE_ACQ();
2979 VNASSERT(vp->v_holdcnt > 0, vp,
2980 ("%s: active vnode not held", __func__));
2985 * By the time we get here the vnode might have been doomed, at
2986 * which point the 0->1 use count transition is no longer
2987 * protected by the interlock. Since it can't bounce back to
2988 * VCHR and requires vref semantics, punt it back
2990 if (__predict_false(vp->v_type == VBAD)) {
2996 VNASSERT(vp->v_type == VCHR, vp, ("type != VCHR)"));
2997 if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2998 VNODE_REFCOUNT_FENCE_ACQ();
2999 VNASSERT(vp->v_holdcnt > 0, vp,
3000 ("%s: active vnode not held", __func__));
3006 v_incr_devcount(vp);
3007 refcount_acquire(&vp->v_usecount);
3014 vref(struct vnode *vp)
3018 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3019 if (__predict_false(vp->v_type == VCHR)) {
3020 vref_vchr(vp, false);
3024 if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
3025 VNODE_REFCOUNT_FENCE_ACQ();
3026 VNASSERT(vp->v_holdcnt > 0, vp,
3027 ("%s: active vnode not held", __func__));
3032 * See the comment in vget_finish.
3034 old = atomic_fetchadd_int(&vp->v_usecount, 1);
3035 VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3038 old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3039 VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3041 refcount_release(&vp->v_holdcnt);
3047 vrefl(struct vnode *vp)
3050 ASSERT_VI_LOCKED(vp, __func__);
3051 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3052 if (__predict_false(vp->v_type == VCHR)) {
3053 vref_vchr(vp, true);
3060 vrefact(struct vnode *vp)
3063 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3065 int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3066 VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3068 refcount_acquire(&vp->v_usecount);
3073 vrefactn(struct vnode *vp, u_int n)
3076 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3078 int old = atomic_fetchadd_int(&vp->v_usecount, n);
3079 VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3081 atomic_add_int(&vp->v_usecount, n);
3086 * Return reference count of a vnode.
3088 * The results of this call are only guaranteed when some mechanism is used to
3089 * stop other processes from gaining references to the vnode. This may be the
3090 * case if the caller holds the only reference. This is also useful when stale
3091 * data is acceptable as race conditions may be accounted for by some other
3095 vrefcnt(struct vnode *vp)
3098 return (vp->v_usecount);
3102 vlazy(struct vnode *vp)
3106 VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3108 if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3111 * We may get here for inactive routines after the vnode got doomed.
3113 if (VN_IS_DOOMED(vp))
3116 mtx_lock(&mp->mnt_listmtx);
3117 if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3118 vp->v_mflag |= VMP_LAZYLIST;
3119 TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3120 mp->mnt_lazyvnodelistsize++;
3122 mtx_unlock(&mp->mnt_listmtx);
3126 * This routine is only meant to be called from vgonel prior to dooming
3130 vunlazy_gone(struct vnode *vp)
3134 ASSERT_VOP_ELOCKED(vp, __func__);
3135 ASSERT_VI_LOCKED(vp, __func__);
3136 VNPASS(!VN_IS_DOOMED(vp), vp);
3138 if (vp->v_mflag & VMP_LAZYLIST) {
3140 mtx_lock(&mp->mnt_listmtx);
3141 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3142 vp->v_mflag &= ~VMP_LAZYLIST;
3143 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3144 mp->mnt_lazyvnodelistsize--;
3145 mtx_unlock(&mp->mnt_listmtx);
3150 vdefer_inactive(struct vnode *vp)
3153 ASSERT_VI_LOCKED(vp, __func__);
3154 VNASSERT(vp->v_holdcnt > 0, vp,
3155 ("%s: vnode without hold count", __func__));
3156 if (VN_IS_DOOMED(vp)) {
3160 if (vp->v_iflag & VI_DEFINACT) {
3161 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3165 if (vp->v_usecount > 0) {
3166 vp->v_iflag &= ~VI_OWEINACT;
3171 vp->v_iflag |= VI_DEFINACT;
3173 counter_u64_add(deferred_inact, 1);
3177 vdefer_inactive_unlocked(struct vnode *vp)
3181 if ((vp->v_iflag & VI_OWEINACT) == 0) {
3185 vdefer_inactive(vp);
3188 enum vput_op { VRELE, VPUT, VUNREF };
3191 * Handle ->v_usecount transitioning to 0.
3193 * By releasing the last usecount we take ownership of the hold count which
3194 * provides liveness of the vnode, meaning we have to vdrop.
3196 * If the vnode is of type VCHR we may need to decrement si_usecount, see
3197 * v_decr_devcount for details.
3199 * For all vnodes we may need to perform inactive processing. It requires an
3200 * exclusive lock on the vnode, while it is legal to call here with only a
3201 * shared lock (or no locks). If locking the vnode in an expected manner fails,
3202 * inactive processing gets deferred to the syncer.
3204 * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3205 * on the lock being held all the way until VOP_INACTIVE. This in particular
3206 * happens with UFS which adds half-constructed vnodes to the hash, where they
3207 * can be found by other code.
3210 vput_final(struct vnode *vp, enum vput_op func)
3215 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3216 VNPASS(vp->v_holdcnt > 0, vp);
3219 if (__predict_false(vp->v_type == VCHR && func != VRELE))
3220 v_decr_devcount(vp);
3223 * By the time we got here someone else might have transitioned
3224 * the count back to > 0.
3226 if (vp->v_usecount > 0)
3230 * If the vnode is doomed vgone already performed inactive processing
3233 if (VN_IS_DOOMED(vp))
3236 if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3239 if (vp->v_iflag & VI_DOINGINACT)
3243 * Locking operations here will drop the interlock and possibly the
3244 * vnode lock, opening a window where the vnode can get doomed all the
3245 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3248 vp->v_iflag |= VI_OWEINACT;
3249 want_unlock = false;
3253 switch (VOP_ISLOCKED(vp)) {
3259 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3264 * The lock has at least one sharer, but we have no way
3265 * to conclude whether this is us. Play it safe and
3274 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3275 error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3281 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3282 error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3293 vdefer_inactive(vp);
3303 * Decrement ->v_usecount for a vnode.
3305 * Releasing the last use count requires additional processing, see vput_final
3306 * above for details.
3308 * Note that releasing use count without the vnode lock requires special casing
3309 * for VCHR, see v_decr_devcount for details.
3311 * Comment above each variant denotes lock state on entry and exit.
3314 static void __noinline
3315 vrele_vchr(struct vnode *vp)
3318 if (refcount_release_if_not_last(&vp->v_usecount))
3321 if (!refcount_release(&vp->v_usecount)) {
3325 v_decr_devcount(vp);
3327 vput_final(vp, VRELE);
3332 * out: same as passed in
3335 vrele(struct vnode *vp)
3338 ASSERT_VI_UNLOCKED(vp, __func__);
3339 if (__predict_false(vp->v_type == VCHR)) {
3343 if (!refcount_release(&vp->v_usecount))
3345 vput_final(vp, VRELE);
3353 vput(struct vnode *vp)
3356 ASSERT_VOP_LOCKED(vp, __func__);
3357 ASSERT_VI_UNLOCKED(vp, __func__);
3358 if (!refcount_release(&vp->v_usecount)) {
3362 vput_final(vp, VPUT);
3370 vunref(struct vnode *vp)
3373 ASSERT_VOP_LOCKED(vp, __func__);
3374 ASSERT_VI_UNLOCKED(vp, __func__);
3375 if (!refcount_release(&vp->v_usecount))
3377 vput_final(vp, VUNREF);
3381 vhold(struct vnode *vp)
3386 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3387 old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3388 VNASSERT(old >= 0, vp, ("%s: wrong hold count %d", __func__, old));
3398 vholdl(struct vnode *vp)
3401 ASSERT_VI_LOCKED(vp, __func__);
3402 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3407 vholdnz(struct vnode *vp)
3410 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3412 int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3413 VNASSERT(old > 0, vp, ("%s: wrong hold count %d", __func__, old));
3415 atomic_add_int(&vp->v_holdcnt, 1);
3419 static void __noinline
3420 vdbatch_process(struct vdbatch *vd)
3425 mtx_assert(&vd->lock, MA_OWNED);
3426 MPASS(curthread->td_pinned > 0);
3427 MPASS(vd->index == VDBATCH_SIZE);
3429 mtx_lock(&vnode_list_mtx);
3431 freevnodes += vd->freevnodes;
3432 for (i = 0; i < VDBATCH_SIZE; i++) {
3434 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3435 TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3436 MPASS(vp->v_dbatchcpu != NOCPU);
3437 vp->v_dbatchcpu = NOCPU;
3439 mtx_unlock(&vnode_list_mtx);
3441 bzero(vd->tab, sizeof(vd->tab));
3447 vdbatch_enqueue(struct vnode *vp)
3451 ASSERT_VI_LOCKED(vp, __func__);
3452 VNASSERT(!VN_IS_DOOMED(vp), vp,
3453 ("%s: deferring requeue of a doomed vnode", __func__));
3458 if (vp->v_dbatchcpu != NOCPU) {
3466 mtx_lock(&vd->lock);
3467 MPASS(vd->index < VDBATCH_SIZE);
3468 MPASS(vd->tab[vd->index] == NULL);
3470 * A hack: we depend on being pinned so that we know what to put in
3473 vp->v_dbatchcpu = curcpu;
3474 vd->tab[vd->index] = vp;
3477 if (vd->index == VDBATCH_SIZE)
3478 vdbatch_process(vd);
3479 mtx_unlock(&vd->lock);
3484 * This routine must only be called for vnodes which are about to be
3485 * deallocated. Supporting dequeue for arbitrary vndoes would require
3486 * validating that the locked batch matches.
3489 vdbatch_dequeue(struct vnode *vp)
3495 VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3496 ("%s: called for a used vnode\n", __func__));
3498 cpu = vp->v_dbatchcpu;
3502 vd = DPCPU_ID_PTR(cpu, vd);
3503 mtx_lock(&vd->lock);
3504 for (i = 0; i < vd->index; i++) {
3505 if (vd->tab[i] != vp)
3507 vp->v_dbatchcpu = NOCPU;
3509 vd->tab[i] = vd->tab[vd->index];
3510 vd->tab[vd->index] = NULL;
3513 mtx_unlock(&vd->lock);
3515 * Either we dequeued the vnode above or the target CPU beat us to it.
3517 MPASS(vp->v_dbatchcpu == NOCPU);
3521 * Drop the hold count of the vnode. If this is the last reference to
3522 * the vnode we place it on the free list unless it has been vgone'd
3523 * (marked VIRF_DOOMED) in which case we will free it.
3525 * Because the vnode vm object keeps a hold reference on the vnode if
3526 * there is at least one resident non-cached page, the vnode cannot
3527 * leave the active list without the page cleanup done.
3530 vdrop_deactivate(struct vnode *vp)
3534 ASSERT_VI_LOCKED(vp, __func__);
3536 * Mark a vnode as free: remove it from its active list
3537 * and put it up for recycling on the freelist.
3539 VNASSERT(!VN_IS_DOOMED(vp), vp,
3540 ("vdrop: returning doomed vnode"));
3541 VNASSERT(vp->v_op != NULL, vp,
3542 ("vdrop: vnode already reclaimed."));
3543 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
3544 ("vnode with VI_OWEINACT set"));
3545 VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp,
3546 ("vnode with VI_DEFINACT set"));
3547 if (vp->v_mflag & VMP_LAZYLIST) {
3549 mtx_lock(&mp->mnt_listmtx);
3550 VNASSERT(vp->v_mflag & VMP_LAZYLIST, vp, ("lost VMP_LAZYLIST"));
3552 * Don't remove the vnode from the lazy list if another thread
3553 * has increased the hold count. It may have re-enqueued the
3554 * vnode to the lazy list and is now responsible for its
3557 if (vp->v_holdcnt == 0) {
3558 vp->v_mflag &= ~VMP_LAZYLIST;
3559 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3560 mp->mnt_lazyvnodelistsize--;
3562 mtx_unlock(&mp->mnt_listmtx);
3564 vdbatch_enqueue(vp);
3568 vdrop(struct vnode *vp)
3571 ASSERT_VI_UNLOCKED(vp, __func__);
3572 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3573 if (refcount_release_if_not_last(&vp->v_holdcnt))
3580 vdropl(struct vnode *vp)
3583 ASSERT_VI_LOCKED(vp, __func__);
3584 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3585 if (!refcount_release(&vp->v_holdcnt)) {
3589 if (VN_IS_DOOMED(vp)) {
3593 vdrop_deactivate(vp);
3597 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3598 * flags. DOINGINACT prevents us from recursing in calls to vinactive.
3601 vinactivef(struct vnode *vp)
3603 struct vm_object *obj;
3605 ASSERT_VOP_ELOCKED(vp, "vinactive");
3606 ASSERT_VI_LOCKED(vp, "vinactive");
3607 VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3608 ("vinactive: recursed on VI_DOINGINACT"));
3609 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3610 vp->v_iflag |= VI_DOINGINACT;
3611 vp->v_iflag &= ~VI_OWEINACT;
3614 * Before moving off the active list, we must be sure that any
3615 * modified pages are converted into the vnode's dirty
3616 * buffers, since these will no longer be checked once the
3617 * vnode is on the inactive list.
3619 * The write-out of the dirty pages is asynchronous. At the
3620 * point that VOP_INACTIVE() is called, there could still be
3621 * pending I/O and dirty pages in the object.
3623 if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3624 vm_object_mightbedirty(obj)) {
3625 VM_OBJECT_WLOCK(obj);
3626 vm_object_page_clean(obj, 0, 0, 0);
3627 VM_OBJECT_WUNLOCK(obj);
3629 VOP_INACTIVE(vp, curthread);
3631 VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3632 ("vinactive: lost VI_DOINGINACT"));
3633 vp->v_iflag &= ~VI_DOINGINACT;
3637 vinactive(struct vnode *vp)
3640 ASSERT_VOP_ELOCKED(vp, "vinactive");
3641 ASSERT_VI_LOCKED(vp, "vinactive");
3642 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3644 if ((vp->v_iflag & VI_OWEINACT) == 0)
3646 if (vp->v_iflag & VI_DOINGINACT)
3648 if (vp->v_usecount > 0) {
3649 vp->v_iflag &= ~VI_OWEINACT;
3656 * Remove any vnodes in the vnode table belonging to mount point mp.
3658 * If FORCECLOSE is not specified, there should not be any active ones,
3659 * return error if any are found (nb: this is a user error, not a
3660 * system error). If FORCECLOSE is specified, detach any active vnodes
3663 * If WRITECLOSE is set, only flush out regular file vnodes open for
3666 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3668 * `rootrefs' specifies the base reference count for the root vnode
3669 * of this filesystem. The root vnode is considered busy if its
3670 * v_usecount exceeds this value. On a successful return, vflush(, td)
3671 * will call vrele() on the root vnode exactly rootrefs times.
3672 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3676 static int busyprt = 0; /* print out busy vnodes */
3677 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3681 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3683 struct vnode *vp, *mvp, *rootvp = NULL;
3685 int busy = 0, error;
3687 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3690 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3691 ("vflush: bad args"));
3693 * Get the filesystem root vnode. We can vput() it
3694 * immediately, since with rootrefs > 0, it won't go away.
3696 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3697 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3704 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3706 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3709 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3713 * Skip over a vnodes marked VV_SYSTEM.
3715 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3721 * If WRITECLOSE is set, flush out unlinked but still open
3722 * files (even if open only for reading) and regular file
3723 * vnodes open for writing.
3725 if (flags & WRITECLOSE) {
3726 if (vp->v_object != NULL) {
3727 VM_OBJECT_WLOCK(vp->v_object);
3728 vm_object_page_clean(vp->v_object, 0, 0, 0);
3729 VM_OBJECT_WUNLOCK(vp->v_object);
3731 error = VOP_FSYNC(vp, MNT_WAIT, td);
3735 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3738 error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3741 if ((vp->v_type == VNON ||
3742 (error == 0 && vattr.va_nlink > 0)) &&
3743 (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3751 * With v_usecount == 0, all we need to do is clear out the
3752 * vnode data structures and we are done.
3754 * If FORCECLOSE is set, forcibly close the vnode.
3756 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3762 vn_printf(vp, "vflush: busy vnode ");
3768 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3770 * If just the root vnode is busy, and if its refcount
3771 * is equal to `rootrefs', then go ahead and kill it.
3774 KASSERT(busy > 0, ("vflush: not busy"));
3775 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3776 ("vflush: usecount %d < rootrefs %d",
3777 rootvp->v_usecount, rootrefs));
3778 if (busy == 1 && rootvp->v_usecount == rootrefs) {
3779 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3787 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3791 for (; rootrefs > 0; rootrefs--)
3797 * Recycle an unused vnode to the front of the free list.
3800 vrecycle(struct vnode *vp)
3805 recycled = vrecyclel(vp);
3811 * vrecycle, with the vp interlock held.
3814 vrecyclel(struct vnode *vp)
3818 ASSERT_VOP_ELOCKED(vp, __func__);
3819 ASSERT_VI_LOCKED(vp, __func__);
3820 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3822 if (vp->v_usecount == 0) {
3830 * Eliminate all activity associated with a vnode
3831 * in preparation for reuse.
3834 vgone(struct vnode *vp)
3842 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3843 struct vnode *lowervp __unused)
3848 * Notify upper mounts about reclaimed or unlinked vnode.
3851 vfs_notify_upper(struct vnode *vp, int event)
3853 static struct vfsops vgonel_vfsops = {
3854 .vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3855 .vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3857 struct mount *mp, *ump, *mmp;
3862 if (TAILQ_EMPTY(&mp->mnt_uppers))
3865 mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3866 mmp->mnt_op = &vgonel_vfsops;
3867 mmp->mnt_kern_flag |= MNTK_MARKER;
3869 mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3870 for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3871 if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3872 ump = TAILQ_NEXT(ump, mnt_upper_link);
3875 TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3878 case VFS_NOTIFY_UPPER_RECLAIM:
3879 VFS_RECLAIM_LOWERVP(ump, vp);
3881 case VFS_NOTIFY_UPPER_UNLINK:
3882 VFS_UNLINK_LOWERVP(ump, vp);
3885 KASSERT(0, ("invalid event %d", event));
3889 ump = TAILQ_NEXT(mmp, mnt_upper_link);
3890 TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3893 mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3894 if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3895 mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3896 wakeup(&mp->mnt_uppers);
3902 * vgone, with the vp interlock held.
3905 vgonel(struct vnode *vp)
3910 bool active, oweinact;
3912 ASSERT_VOP_ELOCKED(vp, "vgonel");
3913 ASSERT_VI_LOCKED(vp, "vgonel");
3914 VNASSERT(vp->v_holdcnt, vp,
3915 ("vgonel: vp %p has no reference.", vp));
3916 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3920 * Don't vgonel if we're already doomed.
3922 if (vp->v_irflag & VIRF_DOOMED)
3925 vp->v_irflag |= VIRF_DOOMED;
3928 * Check to see if the vnode is in use. If so, we have to call
3929 * VOP_CLOSE() and VOP_INACTIVE().
3931 active = vp->v_usecount > 0;
3932 oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3934 * If we need to do inactive VI_OWEINACT will be set.
3936 if (vp->v_iflag & VI_DEFINACT) {
3937 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3938 vp->v_iflag &= ~VI_DEFINACT;
3941 VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
3944 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3947 * If purging an active vnode, it must be closed and
3948 * deactivated before being reclaimed.
3951 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3952 if (oweinact || active) {
3957 if (vp->v_type == VSOCK)
3958 vfs_unp_reclaim(vp);
3961 * Clean out any buffers associated with the vnode.
3962 * If the flush fails, just toss the buffers.
3965 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3966 (void) vn_start_secondary_write(vp, &mp, V_WAIT);
3967 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3968 while (vinvalbuf(vp, 0, 0, 0) != 0)
3972 BO_LOCK(&vp->v_bufobj);
3973 KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3974 vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3975 TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3976 vp->v_bufobj.bo_clean.bv_cnt == 0,
3977 ("vp %p bufobj not invalidated", vp));
3980 * For VMIO bufobj, BO_DEAD is set later, or in
3981 * vm_object_terminate() after the object's page queue is
3984 object = vp->v_bufobj.bo_object;
3986 vp->v_bufobj.bo_flag |= BO_DEAD;
3987 BO_UNLOCK(&vp->v_bufobj);
3990 * Handle the VM part. Tmpfs handles v_object on its own (the
3991 * OBJT_VNODE check). Nullfs or other bypassing filesystems
3992 * should not touch the object borrowed from the lower vnode
3993 * (the handle check).
3995 if (object != NULL && object->type == OBJT_VNODE &&
3996 object->handle == vp)
3997 vnode_destroy_vobject(vp);
4000 * Reclaim the vnode.
4002 if (VOP_RECLAIM(vp, td))
4003 panic("vgone: cannot reclaim");
4005 vn_finished_secondary_write(mp);
4006 VNASSERT(vp->v_object == NULL, vp,
4007 ("vop_reclaim left v_object vp=%p", vp));
4009 * Clear the advisory locks and wake up waiting threads.
4011 (void)VOP_ADVLOCKPURGE(vp);
4014 * Delete from old mount point vnode list.
4019 * Done with purge, reset to the standard lock and invalidate
4023 vp->v_vnlock = &vp->v_lock;
4024 vp->v_op = &dead_vnodeops;
4029 * Calculate the total number of references to a special device.
4032 vcount(struct vnode *vp)
4037 count = vp->v_rdev->si_usecount;
4043 * Print out a description of a vnode.
4045 static const char * const typename[] =
4046 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
4050 vn_printf(struct vnode *vp, const char *fmt, ...)
4053 char buf[256], buf2[16];
4059 printf("%p: ", (void *)vp);
4060 printf("type %s\n", typename[vp->v_type]);
4061 printf(" usecount %d, writecount %d, refcount %d",
4062 vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
4063 switch (vp->v_type) {
4065 printf(" mountedhere %p\n", vp->v_mountedhere);
4068 printf(" rdev %p\n", vp->v_rdev);
4071 printf(" socket %p\n", vp->v_unpcb);
4074 printf(" fifoinfo %p\n", vp->v_fifoinfo);
4082 if (vp->v_irflag & VIRF_DOOMED)
4083 strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4084 flags = vp->v_irflag & ~(VIRF_DOOMED);
4086 snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4087 strlcat(buf, buf2, sizeof(buf));
4089 if (vp->v_vflag & VV_ROOT)
4090 strlcat(buf, "|VV_ROOT", sizeof(buf));
4091 if (vp->v_vflag & VV_ISTTY)
4092 strlcat(buf, "|VV_ISTTY", sizeof(buf));
4093 if (vp->v_vflag & VV_NOSYNC)
4094 strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4095 if (vp->v_vflag & VV_ETERNALDEV)
4096 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4097 if (vp->v_vflag & VV_CACHEDLABEL)
4098 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4099 if (vp->v_vflag & VV_VMSIZEVNLOCK)
4100 strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4101 if (vp->v_vflag & VV_COPYONWRITE)
4102 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4103 if (vp->v_vflag & VV_SYSTEM)
4104 strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4105 if (vp->v_vflag & VV_PROCDEP)
4106 strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4107 if (vp->v_vflag & VV_NOKNOTE)
4108 strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
4109 if (vp->v_vflag & VV_DELETED)
4110 strlcat(buf, "|VV_DELETED", sizeof(buf));
4111 if (vp->v_vflag & VV_MD)
4112 strlcat(buf, "|VV_MD", sizeof(buf));
4113 if (vp->v_vflag & VV_FORCEINSMQ)
4114 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4115 if (vp->v_vflag & VV_READLINK)
4116 strlcat(buf, "|VV_READLINK", sizeof(buf));
4117 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4118 VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
4119 VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
4121 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4122 strlcat(buf, buf2, sizeof(buf));
4124 if (vp->v_iflag & VI_TEXT_REF)
4125 strlcat(buf, "|VI_TEXT_REF", sizeof(buf));
4126 if (vp->v_iflag & VI_MOUNT)
4127 strlcat(buf, "|VI_MOUNT", sizeof(buf));
4128 if (vp->v_iflag & VI_DOINGINACT)
4129 strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4130 if (vp->v_iflag & VI_OWEINACT)
4131 strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4132 if (vp->v_iflag & VI_DEFINACT)
4133 strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4134 flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_DOINGINACT |
4135 VI_OWEINACT | VI_DEFINACT);
4137 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4138 strlcat(buf, buf2, sizeof(buf));
4140 if (vp->v_mflag & VMP_LAZYLIST)
4141 strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4142 flags = vp->v_mflag & ~(VMP_LAZYLIST);
4144 snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4145 strlcat(buf, buf2, sizeof(buf));
4147 printf(" flags (%s)\n", buf + 1);
4148 if (mtx_owned(VI_MTX(vp)))
4149 printf(" VI_LOCKed");
4150 if (vp->v_object != NULL)
4151 printf(" v_object %p ref %d pages %d "
4152 "cleanbuf %d dirtybuf %d\n",
4153 vp->v_object, vp->v_object->ref_count,
4154 vp->v_object->resident_page_count,
4155 vp->v_bufobj.bo_clean.bv_cnt,
4156 vp->v_bufobj.bo_dirty.bv_cnt);
4158 lockmgr_printinfo(vp->v_vnlock);
4159 if (vp->v_data != NULL)
4165 * List all of the locked vnodes in the system.
4166 * Called when debugging the kernel.
4168 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
4174 * Note: because this is DDB, we can't obey the locking semantics
4175 * for these structures, which means we could catch an inconsistent
4176 * state and dereference a nasty pointer. Not much to be done
4179 db_printf("Locked vnodes\n");
4180 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4181 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4182 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4183 vn_printf(vp, "vnode ");
4189 * Show details about the given vnode.
4191 DB_SHOW_COMMAND(vnode, db_show_vnode)
4197 vp = (struct vnode *)addr;
4198 vn_printf(vp, "vnode ");
4202 * Show details about the given mount point.
4204 DB_SHOW_COMMAND(mount, db_show_mount)
4215 /* No address given, print short info about all mount points. */
4216 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4217 db_printf("%p %s on %s (%s)\n", mp,
4218 mp->mnt_stat.f_mntfromname,
4219 mp->mnt_stat.f_mntonname,
4220 mp->mnt_stat.f_fstypename);
4224 db_printf("\nMore info: show mount <addr>\n");
4228 mp = (struct mount *)addr;
4229 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4230 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4233 mflags = mp->mnt_flag;
4234 #define MNT_FLAG(flag) do { \
4235 if (mflags & (flag)) { \
4236 if (buf[0] != '\0') \
4237 strlcat(buf, ", ", sizeof(buf)); \
4238 strlcat(buf, (#flag) + 4, sizeof(buf)); \
4239 mflags &= ~(flag); \
4242 MNT_FLAG(MNT_RDONLY);
4243 MNT_FLAG(MNT_SYNCHRONOUS);
4244 MNT_FLAG(MNT_NOEXEC);
4245 MNT_FLAG(MNT_NOSUID);
4246 MNT_FLAG(MNT_NFS4ACLS);
4247 MNT_FLAG(MNT_UNION);
4248 MNT_FLAG(MNT_ASYNC);
4249 MNT_FLAG(MNT_SUIDDIR);
4250 MNT_FLAG(MNT_SOFTDEP);
4251 MNT_FLAG(MNT_NOSYMFOLLOW);
4252 MNT_FLAG(MNT_GJOURNAL);
4253 MNT_FLAG(MNT_MULTILABEL);
4255 MNT_FLAG(MNT_NOATIME);
4256 MNT_FLAG(MNT_NOCLUSTERR);
4257 MNT_FLAG(MNT_NOCLUSTERW);
4259 MNT_FLAG(MNT_EXRDONLY);
4260 MNT_FLAG(MNT_EXPORTED);
4261 MNT_FLAG(MNT_DEFEXPORTED);
4262 MNT_FLAG(MNT_EXPORTANON);
4263 MNT_FLAG(MNT_EXKERB);
4264 MNT_FLAG(MNT_EXPUBLIC);
4265 MNT_FLAG(MNT_LOCAL);
4266 MNT_FLAG(MNT_QUOTA);
4267 MNT_FLAG(MNT_ROOTFS);
4269 MNT_FLAG(MNT_IGNORE);
4270 MNT_FLAG(MNT_UPDATE);
4271 MNT_FLAG(MNT_DELEXPORT);
4272 MNT_FLAG(MNT_RELOAD);
4273 MNT_FLAG(MNT_FORCE);
4274 MNT_FLAG(MNT_SNAPSHOT);
4275 MNT_FLAG(MNT_BYFSID);
4279 strlcat(buf, ", ", sizeof(buf));
4280 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4281 "0x%016jx", mflags);
4283 db_printf(" mnt_flag = %s\n", buf);
4286 flags = mp->mnt_kern_flag;
4287 #define MNT_KERN_FLAG(flag) do { \
4288 if (flags & (flag)) { \
4289 if (buf[0] != '\0') \
4290 strlcat(buf, ", ", sizeof(buf)); \
4291 strlcat(buf, (#flag) + 5, sizeof(buf)); \
4295 MNT_KERN_FLAG(MNTK_UNMOUNTF);
4296 MNT_KERN_FLAG(MNTK_ASYNC);
4297 MNT_KERN_FLAG(MNTK_SOFTDEP);
4298 MNT_KERN_FLAG(MNTK_DRAINING);
4299 MNT_KERN_FLAG(MNTK_REFEXPIRE);
4300 MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4301 MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4302 MNT_KERN_FLAG(MNTK_NO_IOPF);
4303 MNT_KERN_FLAG(MNTK_VGONE_UPPER);
4304 MNT_KERN_FLAG(MNTK_VGONE_WAITER);
4305 MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
4306 MNT_KERN_FLAG(MNTK_MARKER);
4307 MNT_KERN_FLAG(MNTK_USES_BCACHE);
4308 MNT_KERN_FLAG(MNTK_NOASYNC);
4309 MNT_KERN_FLAG(MNTK_UNMOUNT);
4310 MNT_KERN_FLAG(MNTK_MWAIT);
4311 MNT_KERN_FLAG(MNTK_SUSPEND);
4312 MNT_KERN_FLAG(MNTK_SUSPEND2);
4313 MNT_KERN_FLAG(MNTK_SUSPENDED);
4314 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4315 MNT_KERN_FLAG(MNTK_NOKNOTE);
4316 #undef MNT_KERN_FLAG
4319 strlcat(buf, ", ", sizeof(buf));
4320 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4323 db_printf(" mnt_kern_flag = %s\n", buf);
4325 db_printf(" mnt_opt = ");
4326 opt = TAILQ_FIRST(mp->mnt_opt);
4328 db_printf("%s", opt->name);
4329 opt = TAILQ_NEXT(opt, link);
4330 while (opt != NULL) {
4331 db_printf(", %s", opt->name);
4332 opt = TAILQ_NEXT(opt, link);
4338 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx "
4339 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4340 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4341 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4342 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4343 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4344 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4345 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4346 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4347 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4348 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4349 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4351 db_printf(" mnt_cred = { uid=%u ruid=%u",
4352 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4353 if (jailed(mp->mnt_cred))
4354 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4356 db_printf(" mnt_ref = %d (with %d in the struct)\n",
4357 vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4358 db_printf(" mnt_gen = %d\n", mp->mnt_gen);
4359 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4360 db_printf(" mnt_lazyvnodelistsize = %d\n",
4361 mp->mnt_lazyvnodelistsize);
4362 db_printf(" mnt_writeopcount = %d (with %d in the struct)\n",
4363 vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4364 db_printf(" mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
4365 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4366 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed);
4367 db_printf(" mnt_lockref = %d (with %d in the struct)\n",
4368 vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4369 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4370 db_printf(" mnt_secondary_accwrites = %d\n",
4371 mp->mnt_secondary_accwrites);
4372 db_printf(" mnt_gjprovider = %s\n",
4373 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4374 db_printf(" mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4376 db_printf("\n\nList of active vnodes\n");
4377 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4378 if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4379 vn_printf(vp, "vnode ");
4384 db_printf("\n\nList of inactive vnodes\n");
4385 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4386 if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4387 vn_printf(vp, "vnode ");
4396 * Fill in a struct xvfsconf based on a struct vfsconf.
4399 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4401 struct xvfsconf xvfsp;
4403 bzero(&xvfsp, sizeof(xvfsp));
4404 strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4405 xvfsp.vfc_typenum = vfsp->vfc_typenum;
4406 xvfsp.vfc_refcount = vfsp->vfc_refcount;
4407 xvfsp.vfc_flags = vfsp->vfc_flags;
4409 * These are unused in userland, we keep them
4410 * to not break binary compatibility.
4412 xvfsp.vfc_vfsops = NULL;
4413 xvfsp.vfc_next = NULL;
4414 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4417 #ifdef COMPAT_FREEBSD32
4419 uint32_t vfc_vfsops;
4420 char vfc_name[MFSNAMELEN];
4421 int32_t vfc_typenum;
4422 int32_t vfc_refcount;
4428 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4430 struct xvfsconf32 xvfsp;
4432 bzero(&xvfsp, sizeof(xvfsp));
4433 strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4434 xvfsp.vfc_typenum = vfsp->vfc_typenum;
4435 xvfsp.vfc_refcount = vfsp->vfc_refcount;
4436 xvfsp.vfc_flags = vfsp->vfc_flags;
4437 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4442 * Top level filesystem related information gathering.
4445 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4447 struct vfsconf *vfsp;
4452 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4453 #ifdef COMPAT_FREEBSD32
4454 if (req->flags & SCTL_MASK32)
4455 error = vfsconf2x32(req, vfsp);
4458 error = vfsconf2x(req, vfsp);
4466 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4467 CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4468 "S,xvfsconf", "List of all configured filesystems");
4470 #ifndef BURN_BRIDGES
4471 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4474 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4476 int *name = (int *)arg1 - 1; /* XXX */
4477 u_int namelen = arg2 + 1; /* XXX */
4478 struct vfsconf *vfsp;
4480 log(LOG_WARNING, "userland calling deprecated sysctl, "
4481 "please rebuild world\n");
4483 #if 1 || defined(COMPAT_PRELITE2)
4484 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4486 return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4490 case VFS_MAXTYPENUM:
4493 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4496 return (ENOTDIR); /* overloaded */
4498 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4499 if (vfsp->vfc_typenum == name[2])
4504 return (EOPNOTSUPP);
4505 #ifdef COMPAT_FREEBSD32
4506 if (req->flags & SCTL_MASK32)
4507 return (vfsconf2x32(req, vfsp));
4510 return (vfsconf2x(req, vfsp));
4512 return (EOPNOTSUPP);
4515 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4516 CTLFLAG_MPSAFE, vfs_sysctl,
4517 "Generic filesystem");
4519 #if 1 || defined(COMPAT_PRELITE2)
4522 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4525 struct vfsconf *vfsp;
4526 struct ovfsconf ovfs;
4529 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4530 bzero(&ovfs, sizeof(ovfs));
4531 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */
4532 strcpy(ovfs.vfc_name, vfsp->vfc_name);
4533 ovfs.vfc_index = vfsp->vfc_typenum;
4534 ovfs.vfc_refcount = vfsp->vfc_refcount;
4535 ovfs.vfc_flags = vfsp->vfc_flags;
4536 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4546 #endif /* 1 || COMPAT_PRELITE2 */
4547 #endif /* !BURN_BRIDGES */
4549 #define KINFO_VNODESLOP 10
4552 * Dump vnode list (via sysctl).
4556 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4564 * Stale numvnodes access is not fatal here.
4567 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4569 /* Make an estimate */
4570 return (SYSCTL_OUT(req, 0, len));
4572 error = sysctl_wire_old_buffer(req, 0);
4575 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4577 mtx_lock(&mountlist_mtx);
4578 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4579 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4582 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4586 xvn[n].xv_size = sizeof *xvn;
4587 xvn[n].xv_vnode = vp;
4588 xvn[n].xv_id = 0; /* XXX compat */
4589 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4591 XV_COPY(writecount);
4597 xvn[n].xv_flag = vp->v_vflag;
4599 switch (vp->v_type) {
4606 if (vp->v_rdev == NULL) {
4610 xvn[n].xv_dev = dev2udev(vp->v_rdev);
4613 xvn[n].xv_socket = vp->v_socket;
4616 xvn[n].xv_fifo = vp->v_fifoinfo;
4621 /* shouldn't happen? */
4629 mtx_lock(&mountlist_mtx);
4634 mtx_unlock(&mountlist_mtx);
4636 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4641 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4642 CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4647 unmount_or_warn(struct mount *mp)
4651 error = dounmount(mp, MNT_FORCE, curthread);
4653 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4657 printf("%d)\n", error);
4662 * Unmount all filesystems. The list is traversed in reverse order
4663 * of mounting to avoid dependencies.
4666 vfs_unmountall(void)
4668 struct mount *mp, *tmp;
4670 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4673 * Since this only runs when rebooting, it is not interlocked.
4675 TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4679 * Forcibly unmounting "/dev" before "/" would prevent clean
4680 * unmount of the latter.
4682 if (mp == rootdevmp)
4685 unmount_or_warn(mp);
4688 if (rootdevmp != NULL)
4689 unmount_or_warn(rootdevmp);
4693 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4696 ASSERT_VI_LOCKED(vp, __func__);
4697 VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4698 if ((vp->v_iflag & VI_OWEINACT) == 0) {
4702 if (vn_lock(vp, lkflags) == 0) {
4709 vdefer_inactive_unlocked(vp);
4713 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4716 return (vp->v_iflag & VI_DEFINACT);
4719 static void __noinline
4720 vfs_periodic_inactive(struct mount *mp, int flags)
4722 struct vnode *vp, *mvp;
4725 lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4726 if (flags != MNT_WAIT)
4727 lkflags |= LK_NOWAIT;
4729 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4730 if ((vp->v_iflag & VI_DEFINACT) == 0) {
4734 vp->v_iflag &= ~VI_DEFINACT;
4735 vfs_deferred_inactive(vp, lkflags);
4740 vfs_want_msync(struct vnode *vp)
4742 struct vm_object *obj;
4745 * This test may be performed without any locks held.
4746 * We rely on vm_object's type stability.
4748 if (vp->v_vflag & VV_NOSYNC)
4751 return (obj != NULL && vm_object_mightbedirty(obj));
4755 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4758 if (vp->v_vflag & VV_NOSYNC)
4760 if (vp->v_iflag & VI_DEFINACT)
4762 return (vfs_want_msync(vp));
4765 static void __noinline
4766 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4768 struct vnode *vp, *mvp;
4769 struct vm_object *obj;
4771 int lkflags, objflags;
4776 lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4777 if (flags != MNT_WAIT) {
4778 lkflags |= LK_NOWAIT;
4779 objflags = OBJPC_NOSYNC;
4781 objflags = OBJPC_SYNC;
4784 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4786 if (vp->v_iflag & VI_DEFINACT) {
4787 vp->v_iflag &= ~VI_DEFINACT;
4790 if (!vfs_want_msync(vp)) {
4792 vfs_deferred_inactive(vp, lkflags);
4797 if (vget(vp, lkflags, td) == 0) {
4799 if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4800 VM_OBJECT_WLOCK(obj);
4801 vm_object_page_clean(obj, 0, 0, objflags);
4802 VM_OBJECT_WUNLOCK(obj);
4809 vdefer_inactive_unlocked(vp);
4815 vfs_periodic(struct mount *mp, int flags)
4818 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4820 if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4821 vfs_periodic_inactive(mp, flags);
4823 vfs_periodic_msync_inactive(mp, flags);
4827 destroy_vpollinfo_free(struct vpollinfo *vi)
4830 knlist_destroy(&vi->vpi_selinfo.si_note);
4831 mtx_destroy(&vi->vpi_lock);
4832 uma_zfree(vnodepoll_zone, vi);
4836 destroy_vpollinfo(struct vpollinfo *vi)
4839 knlist_clear(&vi->vpi_selinfo.si_note, 1);
4840 seldrain(&vi->vpi_selinfo);
4841 destroy_vpollinfo_free(vi);
4845 * Initialize per-vnode helper structure to hold poll-related state.
4848 v_addpollinfo(struct vnode *vp)
4850 struct vpollinfo *vi;
4852 if (vp->v_pollinfo != NULL)
4854 vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4855 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4856 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4857 vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4859 if (vp->v_pollinfo != NULL) {
4861 destroy_vpollinfo_free(vi);
4864 vp->v_pollinfo = vi;
4869 * Record a process's interest in events which might happen to
4870 * a vnode. Because poll uses the historic select-style interface
4871 * internally, this routine serves as both the ``check for any
4872 * pending events'' and the ``record my interest in future events''
4873 * functions. (These are done together, while the lock is held,
4874 * to avoid race conditions.)
4877 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4881 mtx_lock(&vp->v_pollinfo->vpi_lock);
4882 if (vp->v_pollinfo->vpi_revents & events) {
4884 * This leaves events we are not interested
4885 * in available for the other process which
4886 * which presumably had requested them
4887 * (otherwise they would never have been
4890 events &= vp->v_pollinfo->vpi_revents;
4891 vp->v_pollinfo->vpi_revents &= ~events;
4893 mtx_unlock(&vp->v_pollinfo->vpi_lock);
4896 vp->v_pollinfo->vpi_events |= events;
4897 selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4898 mtx_unlock(&vp->v_pollinfo->vpi_lock);
4903 * Routine to create and manage a filesystem syncer vnode.
4905 #define sync_close ((int (*)(struct vop_close_args *))nullop)
4906 static int sync_fsync(struct vop_fsync_args *);
4907 static int sync_inactive(struct vop_inactive_args *);
4908 static int sync_reclaim(struct vop_reclaim_args *);
4910 static struct vop_vector sync_vnodeops = {
4911 .vop_bypass = VOP_EOPNOTSUPP,
4912 .vop_close = sync_close, /* close */
4913 .vop_fsync = sync_fsync, /* fsync */
4914 .vop_inactive = sync_inactive, /* inactive */
4915 .vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4916 .vop_reclaim = sync_reclaim, /* reclaim */
4917 .vop_lock1 = vop_stdlock, /* lock */
4918 .vop_unlock = vop_stdunlock, /* unlock */
4919 .vop_islocked = vop_stdislocked, /* islocked */
4921 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4924 * Create a new filesystem syncer vnode for the specified mount point.
4927 vfs_allocate_syncvnode(struct mount *mp)
4931 static long start, incr, next;
4934 /* Allocate a new vnode */
4935 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4937 panic("vfs_allocate_syncvnode: getnewvnode() failed");
4939 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4940 vp->v_vflag |= VV_FORCEINSMQ;
4941 error = insmntque(vp, mp);
4943 panic("vfs_allocate_syncvnode: insmntque() failed");
4944 vp->v_vflag &= ~VV_FORCEINSMQ;
4947 * Place the vnode onto the syncer worklist. We attempt to
4948 * scatter them about on the list so that they will go off
4949 * at evenly distributed times even if all the filesystems
4950 * are mounted at once.
4953 if (next == 0 || next > syncer_maxdelay) {
4957 start = syncer_maxdelay / 2;
4958 incr = syncer_maxdelay;
4964 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4965 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4966 mtx_lock(&sync_mtx);
4968 if (mp->mnt_syncer == NULL) {
4969 mp->mnt_syncer = vp;
4972 mtx_unlock(&sync_mtx);
4975 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4982 vfs_deallocate_syncvnode(struct mount *mp)
4986 mtx_lock(&sync_mtx);
4987 vp = mp->mnt_syncer;
4989 mp->mnt_syncer = NULL;
4990 mtx_unlock(&sync_mtx);
4996 * Do a lazy sync of the filesystem.
4999 sync_fsync(struct vop_fsync_args *ap)
5001 struct vnode *syncvp = ap->a_vp;
5002 struct mount *mp = syncvp->v_mount;
5007 * We only need to do something if this is a lazy evaluation.
5009 if (ap->a_waitfor != MNT_LAZY)
5013 * Move ourselves to the back of the sync list.
5015 bo = &syncvp->v_bufobj;
5017 vn_syncer_add_to_worklist(bo, syncdelay);
5021 * Walk the list of vnodes pushing all that are dirty and
5022 * not already on the sync list.
5024 if (vfs_busy(mp, MBF_NOWAIT) != 0)
5026 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
5030 save = curthread_pflags_set(TDP_SYNCIO);
5032 * The filesystem at hand may be idle with free vnodes stored in the
5033 * batch. Return them instead of letting them stay there indefinitely.
5035 vfs_periodic(mp, MNT_NOWAIT);
5036 error = VFS_SYNC(mp, MNT_LAZY);
5037 curthread_pflags_restore(save);
5038 vn_finished_write(mp);
5044 * The syncer vnode is no referenced.
5047 sync_inactive(struct vop_inactive_args *ap)
5055 * The syncer vnode is no longer needed and is being decommissioned.
5057 * Modifications to the worklist must be protected by sync_mtx.
5060 sync_reclaim(struct vop_reclaim_args *ap)
5062 struct vnode *vp = ap->a_vp;
5067 mtx_lock(&sync_mtx);
5068 if (vp->v_mount->mnt_syncer == vp)
5069 vp->v_mount->mnt_syncer = NULL;
5070 if (bo->bo_flag & BO_ONWORKLST) {
5071 LIST_REMOVE(bo, bo_synclist);
5072 syncer_worklist_len--;
5074 bo->bo_flag &= ~BO_ONWORKLST;
5076 mtx_unlock(&sync_mtx);
5083 vn_need_pageq_flush(struct vnode *vp)
5085 struct vm_object *obj;
5088 MPASS(mtx_owned(VI_MTX(vp)));
5090 if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5091 vm_object_mightbedirty(obj))
5097 * Check if vnode represents a disk device
5100 vn_isdisk(struct vnode *vp, int *errp)
5104 if (vp->v_type != VCHR) {
5110 if (vp->v_rdev == NULL)
5112 else if (vp->v_rdev->si_devsw == NULL)
5114 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5120 return (error == 0);
5124 * Common filesystem object access control check routine. Accepts a
5125 * vnode's type, "mode", uid and gid, requested access mode, credentials,
5126 * and optional call-by-reference privused argument allowing vaccess()
5127 * to indicate to the caller whether privilege was used to satisfy the
5128 * request (obsoleted). Returns 0 on success, or an errno on failure.
5131 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5132 accmode_t accmode, struct ucred *cred, int *privused)
5134 accmode_t dac_granted;
5135 accmode_t priv_granted;
5137 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5138 ("invalid bit in accmode"));
5139 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5140 ("VAPPEND without VWRITE"));
5143 * Look for a normal, non-privileged way to access the file/directory
5144 * as requested. If it exists, go with that.
5147 if (privused != NULL)
5152 /* Check the owner. */
5153 if (cred->cr_uid == file_uid) {
5154 dac_granted |= VADMIN;
5155 if (file_mode & S_IXUSR)
5156 dac_granted |= VEXEC;
5157 if (file_mode & S_IRUSR)
5158 dac_granted |= VREAD;
5159 if (file_mode & S_IWUSR)
5160 dac_granted |= (VWRITE | VAPPEND);
5162 if ((accmode & dac_granted) == accmode)
5168 /* Otherwise, check the groups (first match) */
5169 if (groupmember(file_gid, cred)) {
5170 if (file_mode & S_IXGRP)
5171 dac_granted |= VEXEC;
5172 if (file_mode & S_IRGRP)
5173 dac_granted |= VREAD;
5174 if (file_mode & S_IWGRP)
5175 dac_granted |= (VWRITE | VAPPEND);
5177 if ((accmode & dac_granted) == accmode)
5183 /* Otherwise, check everyone else. */
5184 if (file_mode & S_IXOTH)
5185 dac_granted |= VEXEC;
5186 if (file_mode & S_IROTH)
5187 dac_granted |= VREAD;
5188 if (file_mode & S_IWOTH)
5189 dac_granted |= (VWRITE | VAPPEND);
5190 if ((accmode & dac_granted) == accmode)
5195 * Build a privilege mask to determine if the set of privileges
5196 * satisfies the requirements when combined with the granted mask
5197 * from above. For each privilege, if the privilege is required,
5198 * bitwise or the request type onto the priv_granted mask.
5204 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5205 * requests, instead of PRIV_VFS_EXEC.
5207 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5208 !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5209 priv_granted |= VEXEC;
5212 * Ensure that at least one execute bit is on. Otherwise,
5213 * a privileged user will always succeed, and we don't want
5214 * this to happen unless the file really is executable.
5216 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5217 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5218 !priv_check_cred(cred, PRIV_VFS_EXEC))
5219 priv_granted |= VEXEC;
5222 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5223 !priv_check_cred(cred, PRIV_VFS_READ))
5224 priv_granted |= VREAD;
5226 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5227 !priv_check_cred(cred, PRIV_VFS_WRITE))
5228 priv_granted |= (VWRITE | VAPPEND);
5230 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5231 !priv_check_cred(cred, PRIV_VFS_ADMIN))
5232 priv_granted |= VADMIN;
5234 if ((accmode & (priv_granted | dac_granted)) == accmode) {
5235 /* XXX audit: privilege used */
5236 if (privused != NULL)
5241 return ((accmode & VADMIN) ? EPERM : EACCES);
5245 * Credential check based on process requesting service, and per-attribute
5249 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5250 struct thread *td, accmode_t accmode)
5254 * Kernel-invoked always succeeds.
5260 * Do not allow privileged processes in jail to directly manipulate
5261 * system attributes.
5263 switch (attrnamespace) {
5264 case EXTATTR_NAMESPACE_SYSTEM:
5265 /* Potentially should be: return (EPERM); */
5266 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5267 case EXTATTR_NAMESPACE_USER:
5268 return (VOP_ACCESS(vp, accmode, cred, td));
5274 #ifdef DEBUG_VFS_LOCKS
5276 * This only exists to suppress warnings from unlocked specfs accesses. It is
5277 * no longer ok to have an unlocked VFS.
5279 #define IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL || \
5280 (vp)->v_type == VCHR || (vp)->v_type == VBAD)
5282 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */
5283 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5284 "Drop into debugger on lock violation");
5286 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */
5287 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5288 0, "Check for interlock across VOPs");
5290 int vfs_badlock_print = 1; /* Print lock violations. */
5291 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5292 0, "Print lock violations");
5294 int vfs_badlock_vnode = 1; /* Print vnode details on lock violations. */
5295 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5296 0, "Print vnode details on lock violations");
5299 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */
5300 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5301 &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5305 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5309 if (vfs_badlock_backtrace)
5312 if (vfs_badlock_vnode)
5313 vn_printf(vp, "vnode ");
5314 if (vfs_badlock_print)
5315 printf("%s: %p %s\n", str, (void *)vp, msg);
5316 if (vfs_badlock_ddb)
5317 kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5321 assert_vi_locked(struct vnode *vp, const char *str)
5324 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5325 vfs_badlock("interlock is not locked but should be", str, vp);
5329 assert_vi_unlocked(struct vnode *vp, const char *str)
5332 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5333 vfs_badlock("interlock is locked but should not be", str, vp);
5337 assert_vop_locked(struct vnode *vp, const char *str)
5341 if (!IGNORE_LOCK(vp)) {
5342 locked = VOP_ISLOCKED(vp);
5343 if (locked == 0 || locked == LK_EXCLOTHER)
5344 vfs_badlock("is not locked but should be", str, vp);
5349 assert_vop_unlocked(struct vnode *vp, const char *str)
5352 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5353 vfs_badlock("is locked but should not be", str, vp);
5357 assert_vop_elocked(struct vnode *vp, const char *str)
5360 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5361 vfs_badlock("is not exclusive locked but should be", str, vp);
5363 #endif /* DEBUG_VFS_LOCKS */
5366 vop_rename_fail(struct vop_rename_args *ap)
5369 if (ap->a_tvp != NULL)
5371 if (ap->a_tdvp == ap->a_tvp)
5380 vop_rename_pre(void *ap)
5382 struct vop_rename_args *a = ap;
5384 #ifdef DEBUG_VFS_LOCKS
5386 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5387 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5388 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5389 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5391 /* Check the source (from). */
5392 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5393 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5394 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5395 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5396 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5398 /* Check the target. */
5400 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5401 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5403 if (a->a_tdvp != a->a_fdvp)
5405 if (a->a_tvp != a->a_fvp)
5412 #ifdef DEBUG_VFS_LOCKS
5414 vop_strategy_pre(void *ap)
5416 struct vop_strategy_args *a;
5423 * Cluster ops lock their component buffers but not the IO container.
5425 if ((bp->b_flags & B_CLUSTER) != 0)
5428 if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5429 if (vfs_badlock_print)
5431 "VOP_STRATEGY: bp is not locked but should be\n");
5432 if (vfs_badlock_ddb)
5433 kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5438 vop_lock_pre(void *ap)
5440 struct vop_lock1_args *a = ap;
5442 if ((a->a_flags & LK_INTERLOCK) == 0)
5443 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5445 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5449 vop_lock_post(void *ap, int rc)
5451 struct vop_lock1_args *a = ap;
5453 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5454 if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5455 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5459 vop_unlock_pre(void *ap)
5461 struct vop_unlock_args *a = ap;
5463 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5467 vop_need_inactive_pre(void *ap)
5469 struct vop_need_inactive_args *a = ap;
5471 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5475 vop_need_inactive_post(void *ap, int rc)
5477 struct vop_need_inactive_args *a = ap;
5479 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5484 vop_create_post(void *ap, int rc)
5486 struct vop_create_args *a = ap;
5489 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5493 vop_deleteextattr_post(void *ap, int rc)
5495 struct vop_deleteextattr_args *a = ap;
5498 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5502 vop_link_post(void *ap, int rc)
5504 struct vop_link_args *a = ap;
5507 VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
5508 VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
5513 vop_mkdir_post(void *ap, int rc)
5515 struct vop_mkdir_args *a = ap;
5518 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5522 vop_mknod_post(void *ap, int rc)
5524 struct vop_mknod_args *a = ap;
5527 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5531 vop_reclaim_post(void *ap, int rc)
5533 struct vop_reclaim_args *a = ap;
5536 VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
5540 vop_remove_post(void *ap, int rc)
5542 struct vop_remove_args *a = ap;
5545 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5546 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5551 vop_rename_post(void *ap, int rc)
5553 struct vop_rename_args *a = ap;
5558 if (a->a_fdvp == a->a_tdvp) {
5559 if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5561 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5562 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5564 hint |= NOTE_EXTEND;
5565 if (a->a_fvp->v_type == VDIR)
5567 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5569 if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5570 a->a_tvp->v_type == VDIR)
5572 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5575 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5577 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5579 if (a->a_tdvp != a->a_fdvp)
5581 if (a->a_tvp != a->a_fvp)
5589 vop_rmdir_post(void *ap, int rc)
5591 struct vop_rmdir_args *a = ap;
5594 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5595 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5600 vop_setattr_post(void *ap, int rc)
5602 struct vop_setattr_args *a = ap;
5605 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5609 vop_setextattr_post(void *ap, int rc)
5611 struct vop_setextattr_args *a = ap;
5614 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5618 vop_symlink_post(void *ap, int rc)
5620 struct vop_symlink_args *a = ap;
5623 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5627 vop_open_post(void *ap, int rc)
5629 struct vop_open_args *a = ap;
5632 VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5636 vop_close_post(void *ap, int rc)
5638 struct vop_close_args *a = ap;
5640 if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5641 !VN_IS_DOOMED(a->a_vp))) {
5642 VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5643 NOTE_CLOSE_WRITE : NOTE_CLOSE);
5648 vop_read_post(void *ap, int rc)
5650 struct vop_read_args *a = ap;
5653 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5657 vop_readdir_post(void *ap, int rc)
5659 struct vop_readdir_args *a = ap;
5662 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5665 static struct knlist fs_knlist;
5668 vfs_event_init(void *arg)
5670 knlist_init_mtx(&fs_knlist, NULL);
5672 /* XXX - correct order? */
5673 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5676 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5679 KNOTE_UNLOCKED(&fs_knlist, event);
5682 static int filt_fsattach(struct knote *kn);
5683 static void filt_fsdetach(struct knote *kn);
5684 static int filt_fsevent(struct knote *kn, long hint);
5686 struct filterops fs_filtops = {
5688 .f_attach = filt_fsattach,
5689 .f_detach = filt_fsdetach,
5690 .f_event = filt_fsevent
5694 filt_fsattach(struct knote *kn)
5697 kn->kn_flags |= EV_CLEAR;
5698 knlist_add(&fs_knlist, kn, 0);
5703 filt_fsdetach(struct knote *kn)
5706 knlist_remove(&fs_knlist, kn, 0);
5710 filt_fsevent(struct knote *kn, long hint)
5713 kn->kn_fflags |= hint;
5714 return (kn->kn_fflags != 0);
5718 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5724 error = SYSCTL_IN(req, &vc, sizeof(vc));
5727 if (vc.vc_vers != VFS_CTL_VERS1)
5729 mp = vfs_getvfs(&vc.vc_fsid);
5732 /* ensure that a specific sysctl goes to the right filesystem. */
5733 if (strcmp(vc.vc_fstypename, "*") != 0 &&
5734 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
5738 VCTLTOREQ(&vc, req);
5739 error = VFS_SYSCTL(mp, vc.vc_op, req);
5744 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
5745 NULL, 0, sysctl_vfs_ctl, "",
5749 * Function to initialize a va_filerev field sensibly.
5750 * XXX: Wouldn't a random number make a lot more sense ??
5753 init_va_filerev(void)
5758 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5761 static int filt_vfsread(struct knote *kn, long hint);
5762 static int filt_vfswrite(struct knote *kn, long hint);
5763 static int filt_vfsvnode(struct knote *kn, long hint);
5764 static void filt_vfsdetach(struct knote *kn);
5765 static struct filterops vfsread_filtops = {
5767 .f_detach = filt_vfsdetach,
5768 .f_event = filt_vfsread
5770 static struct filterops vfswrite_filtops = {
5772 .f_detach = filt_vfsdetach,
5773 .f_event = filt_vfswrite
5775 static struct filterops vfsvnode_filtops = {
5777 .f_detach = filt_vfsdetach,
5778 .f_event = filt_vfsvnode
5782 vfs_knllock(void *arg)
5784 struct vnode *vp = arg;
5786 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5790 vfs_knlunlock(void *arg)
5792 struct vnode *vp = arg;
5798 vfs_knl_assert_locked(void *arg)
5800 #ifdef DEBUG_VFS_LOCKS
5801 struct vnode *vp = arg;
5803 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
5808 vfs_knl_assert_unlocked(void *arg)
5810 #ifdef DEBUG_VFS_LOCKS
5811 struct vnode *vp = arg;
5813 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
5818 vfs_kqfilter(struct vop_kqfilter_args *ap)
5820 struct vnode *vp = ap->a_vp;
5821 struct knote *kn = ap->a_kn;
5824 switch (kn->kn_filter) {
5826 kn->kn_fop = &vfsread_filtops;
5829 kn->kn_fop = &vfswrite_filtops;
5832 kn->kn_fop = &vfsvnode_filtops;
5838 kn->kn_hook = (caddr_t)vp;
5841 if (vp->v_pollinfo == NULL)
5843 knl = &vp->v_pollinfo->vpi_selinfo.si_note;
5845 knlist_add(knl, kn, 0);
5851 * Detach knote from vnode
5854 filt_vfsdetach(struct knote *kn)
5856 struct vnode *vp = (struct vnode *)kn->kn_hook;
5858 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5859 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5865 filt_vfsread(struct knote *kn, long hint)
5867 struct vnode *vp = (struct vnode *)kn->kn_hook;
5872 * filesystem is gone, so set the EOF flag and schedule
5873 * the knote for deletion.
5875 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5877 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5882 if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5886 kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5887 res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5894 filt_vfswrite(struct knote *kn, long hint)
5896 struct vnode *vp = (struct vnode *)kn->kn_hook;
5901 * filesystem is gone, so set the EOF flag and schedule
5902 * the knote for deletion.
5904 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5905 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5913 filt_vfsvnode(struct knote *kn, long hint)
5915 struct vnode *vp = (struct vnode *)kn->kn_hook;
5919 if (kn->kn_sfflags & hint)
5920 kn->kn_fflags |= hint;
5921 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5922 kn->kn_flags |= EV_EOF;
5926 res = (kn->kn_fflags != 0);
5932 * Returns whether the directory is empty or not.
5933 * If it is empty, the return value is 0; otherwise
5934 * the return value is an error value (which may
5938 vfs_emptydir(struct vnode *vp)
5942 struct dirent *dirent, *dp, *endp;
5948 ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
5950 dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
5951 iov.iov_base = dirent;
5952 iov.iov_len = sizeof(struct dirent);
5957 uio.uio_resid = sizeof(struct dirent);
5958 uio.uio_segflg = UIO_SYSSPACE;
5959 uio.uio_rw = UIO_READ;
5960 uio.uio_td = curthread;
5962 while (eof == 0 && error == 0) {
5963 error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
5967 endp = (void *)((uint8_t *)dirent +
5968 sizeof(struct dirent) - uio.uio_resid);
5969 for (dp = dirent; dp < endp;
5970 dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
5971 if (dp->d_type == DT_WHT)
5973 if (dp->d_namlen == 0)
5975 if (dp->d_type != DT_DIR &&
5976 dp->d_type != DT_UNKNOWN) {
5980 if (dp->d_namlen > 2) {
5984 if (dp->d_namlen == 1 &&
5985 dp->d_name[0] != '.') {
5989 if (dp->d_namlen == 2 &&
5990 dp->d_name[1] != '.') {
5994 uio.uio_resid = sizeof(struct dirent);
5997 free(dirent, M_TEMP);
6002 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6006 if (dp->d_reclen > ap->a_uio->uio_resid)
6007 return (ENAMETOOLONG);
6008 error = uiomove(dp, dp->d_reclen, ap->a_uio);
6010 if (ap->a_ncookies != NULL) {
6011 if (ap->a_cookies != NULL)
6012 free(ap->a_cookies, M_TEMP);
6013 ap->a_cookies = NULL;
6014 *ap->a_ncookies = 0;
6018 if (ap->a_ncookies == NULL)
6021 KASSERT(ap->a_cookies,
6022 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6024 *ap->a_cookies = realloc(*ap->a_cookies,
6025 (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
6026 (*ap->a_cookies)[*ap->a_ncookies] = off;
6027 *ap->a_ncookies += 1;
6032 * The purpose of this routine is to remove granularity from accmode_t,
6033 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6034 * VADMIN and VAPPEND.
6036 * If it returns 0, the caller is supposed to continue with the usual
6037 * access checks using 'accmode' as modified by this routine. If it
6038 * returns nonzero value, the caller is supposed to return that value
6041 * Note that after this routine runs, accmode may be zero.
6044 vfs_unixify_accmode(accmode_t *accmode)
6047 * There is no way to specify explicit "deny" rule using
6048 * file mode or POSIX.1e ACLs.
6050 if (*accmode & VEXPLICIT_DENY) {
6056 * None of these can be translated into usual access bits.
6057 * Also, the common case for NFSv4 ACLs is to not contain
6058 * either of these bits. Caller should check for VWRITE
6059 * on the containing directory instead.
6061 if (*accmode & (VDELETE_CHILD | VDELETE))
6064 if (*accmode & VADMIN_PERMS) {
6065 *accmode &= ~VADMIN_PERMS;
6070 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6071 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6073 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6079 * Clear out a doomed vnode (if any) and replace it with a new one as long
6080 * as the fs is not being unmounted. Return the root vnode to the caller.
6082 static int __noinline
6083 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6089 if (mp->mnt_rootvnode != NULL) {
6091 vp = mp->mnt_rootvnode;
6093 if (!VN_IS_DOOMED(vp)) {
6096 error = vn_lock(vp, flags);
6105 * Clear the old one.
6107 mp->mnt_rootvnode = NULL;
6111 vfs_op_barrier_wait(mp);
6115 error = VFS_CACHEDROOT(mp, flags, vpp);
6118 if (mp->mnt_vfs_ops == 0) {
6120 if (mp->mnt_vfs_ops != 0) {
6124 if (mp->mnt_rootvnode == NULL) {
6126 mp->mnt_rootvnode = *vpp;
6128 if (mp->mnt_rootvnode != *vpp) {
6129 if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6130 panic("%s: mismatch between vnode returned "
6131 " by VFS_CACHEDROOT and the one cached "
6133 __func__, *vpp, mp->mnt_rootvnode);
6143 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6148 if (!vfs_op_thread_enter(mp))
6149 return (vfs_cache_root_fallback(mp, flags, vpp));
6150 vp = atomic_load_ptr(&mp->mnt_rootvnode);
6151 if (vp == NULL || VN_IS_DOOMED(vp)) {
6152 vfs_op_thread_exit(mp);
6153 return (vfs_cache_root_fallback(mp, flags, vpp));
6156 vfs_op_thread_exit(mp);
6157 error = vn_lock(vp, flags);
6160 return (vfs_cache_root_fallback(mp, flags, vpp));
6167 vfs_cache_root_clear(struct mount *mp)
6172 * ops > 0 guarantees there is nobody who can see this vnode
6174 MPASS(mp->mnt_vfs_ops > 0);
6175 vp = mp->mnt_rootvnode;
6176 mp->mnt_rootvnode = NULL;
6181 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6184 MPASS(mp->mnt_vfs_ops > 0);
6186 mp->mnt_rootvnode = vp;
6190 * These are helper functions for filesystems to traverse all
6191 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6193 * This interface replaces MNT_VNODE_FOREACH.
6197 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6202 kern_yield(PRI_USER);
6204 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6205 for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6206 vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6207 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6208 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6211 if (VN_IS_DOOMED(vp)) {
6218 __mnt_vnode_markerfree_all(mvp, mp);
6219 /* MNT_IUNLOCK(mp); -- done in above function */
6220 mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6223 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6224 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6230 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6234 *mvp = vn_alloc_marker(mp);
6238 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6239 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6240 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6243 if (VN_IS_DOOMED(vp)) {
6252 vn_free_marker(*mvp);
6256 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6262 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6270 mtx_assert(MNT_MTX(mp), MA_OWNED);
6272 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6273 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6276 vn_free_marker(*mvp);
6281 * These are helper functions for filesystems to traverse their
6282 * lazy vnodes. See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6285 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6288 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6293 vn_free_marker(*mvp);
6298 * Relock the mp mount vnode list lock with the vp vnode interlock in the
6299 * conventional lock order during mnt_vnode_next_lazy iteration.
6301 * On entry, the mount vnode list lock is held and the vnode interlock is not.
6302 * The list lock is dropped and reacquired. On success, both locks are held.
6303 * On failure, the mount vnode list lock is held but the vnode interlock is
6304 * not, and the procedure may have yielded.
6307 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6311 VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6312 TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6313 ("%s: bad marker", __func__));
6314 VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6315 ("%s: inappropriate vnode", __func__));
6316 ASSERT_VI_UNLOCKED(vp, __func__);
6317 mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6319 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6320 TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6323 * Note we may be racing against vdrop which transitioned the hold
6324 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6325 * if we are the only user after we get the interlock we will just
6329 mtx_unlock(&mp->mnt_listmtx);
6331 if (VN_IS_DOOMED(vp)) {
6332 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6335 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6337 * There is nothing to do if we are the last user.
6339 if (!refcount_release_if_not_last(&vp->v_holdcnt))
6341 mtx_lock(&mp->mnt_listmtx);
6346 mtx_lock(&mp->mnt_listmtx);
6350 static struct vnode *
6351 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6356 mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6357 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6359 vp = TAILQ_NEXT(*mvp, v_lazylist);
6360 while (vp != NULL) {
6361 if (vp->v_type == VMARKER) {
6362 vp = TAILQ_NEXT(vp, v_lazylist);
6366 * See if we want to process the vnode. Note we may encounter a
6367 * long string of vnodes we don't care about and hog the list
6368 * as a result. Check for it and requeue the marker.
6370 VNPASS(!VN_IS_DOOMED(vp), vp);
6371 if (!cb(vp, cbarg)) {
6372 if (!should_yield()) {
6373 vp = TAILQ_NEXT(vp, v_lazylist);
6376 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6378 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6380 mtx_unlock(&mp->mnt_listmtx);
6381 kern_yield(PRI_USER);
6382 mtx_lock(&mp->mnt_listmtx);
6386 * Try-lock because this is the wrong lock order.
6388 if (!VI_TRYLOCK(vp) &&
6389 !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6391 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6392 KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6393 ("alien vnode on the lazy list %p %p", vp, mp));
6394 VNPASS(vp->v_mount == mp, vp);
6395 VNPASS(!VN_IS_DOOMED(vp), vp);
6398 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6400 /* Check if we are done */
6402 mtx_unlock(&mp->mnt_listmtx);
6403 mnt_vnode_markerfree_lazy(mvp, mp);
6406 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6407 mtx_unlock(&mp->mnt_listmtx);
6408 ASSERT_VI_LOCKED(vp, "lazy iter");
6413 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6418 kern_yield(PRI_USER);
6419 mtx_lock(&mp->mnt_listmtx);
6420 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6424 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6429 if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6432 *mvp = vn_alloc_marker(mp);
6437 mtx_lock(&mp->mnt_listmtx);
6438 vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6440 mtx_unlock(&mp->mnt_listmtx);
6441 mnt_vnode_markerfree_lazy(mvp, mp);
6444 TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6445 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6449 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6455 mtx_lock(&mp->mnt_listmtx);
6456 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6457 mtx_unlock(&mp->mnt_listmtx);
6458 mnt_vnode_markerfree_lazy(mvp, mp);
6462 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6465 if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6466 cnp->cn_flags &= ~NOEXECCHECK;
6470 return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, cnp->cn_thread));