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>
54 #include <sys/capsicum.h>
55 #include <sys/condvar.h>
57 #include <sys/counter.h>
58 #include <sys/dirent.h>
59 #include <sys/event.h>
60 #include <sys/eventhandler.h>
61 #include <sys/extattr.h>
63 #include <sys/fcntl.h>
66 #include <sys/kernel.h>
67 #include <sys/kthread.h>
69 #include <sys/limits.h>
70 #include <sys/lockf.h>
71 #include <sys/malloc.h>
72 #include <sys/mount.h>
73 #include <sys/namei.h>
74 #include <sys/pctrie.h>
76 #include <sys/reboot.h>
77 #include <sys/refcount.h>
78 #include <sys/rwlock.h>
79 #include <sys/sched.h>
80 #include <sys/sleepqueue.h>
84 #include <sys/sysctl.h>
85 #include <sys/syslog.h>
86 #include <sys/vmmeter.h>
87 #include <sys/vnode.h>
88 #include <sys/watchdog.h>
90 #include <machine/stdarg.h>
92 #include <security/mac/mac_framework.h>
95 #include <vm/vm_object.h>
96 #include <vm/vm_extern.h>
98 #include <vm/vm_map.h>
99 #include <vm/vm_page.h>
100 #include <vm/vm_kern.h>
103 #if defined(DEBUG_VFS_LOCKS) && (!defined(INVARIANTS) || !defined(WITNESS))
104 #error DEBUG_VFS_LOCKS requires INVARIANTS and WITNESS
111 static void delmntque(struct vnode *vp);
112 static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
113 int slpflag, int slptimeo);
114 static void syncer_shutdown(void *arg, int howto);
115 static int vtryrecycle(struct vnode *vp);
116 static void v_init_counters(struct vnode *);
117 static void vn_seqc_init(struct vnode *);
118 static void vn_seqc_write_end_free(struct vnode *vp);
119 static void vgonel(struct vnode *);
120 static bool vhold_recycle_free(struct vnode *);
121 static void vdropl_recycle(struct vnode *vp);
122 static void vdrop_recycle(struct vnode *vp);
123 static void vfs_knllock(void *arg);
124 static void vfs_knlunlock(void *arg);
125 static void vfs_knl_assert_lock(void *arg, int what);
126 static void destroy_vpollinfo(struct vpollinfo *vi);
127 static int v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
128 daddr_t startlbn, daddr_t endlbn);
129 static void vnlru_recalc(void);
132 * Number of vnodes in existence. Increased whenever getnewvnode()
133 * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
135 static u_long __exclusive_cache_line numvnodes;
137 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
138 "Number of vnodes in existence");
140 static counter_u64_t vnodes_created;
141 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
142 "Number of vnodes created by getnewvnode");
145 * Conversion tables for conversion from vnode types to inode formats
148 enum vtype iftovt_tab[16] = {
149 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
150 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
152 int vttoif_tab[10] = {
153 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
154 S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
158 * List of allocates vnodes in the system.
160 static TAILQ_HEAD(freelst, vnode) vnode_list;
161 static struct vnode *vnode_list_free_marker;
162 static struct vnode *vnode_list_reclaim_marker;
165 * "Free" vnode target. Free vnodes are rarely completely free, but are
166 * just ones that are cheap to recycle. Usually they are for files which
167 * have been stat'd but not read; these usually have inode and namecache
168 * data attached to them. This target is the preferred minimum size of a
169 * sub-cache consisting mostly of such files. The system balances the size
170 * of this sub-cache with its complement to try to prevent either from
171 * thrashing while the other is relatively inactive. The targets express
172 * a preference for the best balance.
174 * "Above" this target there are 2 further targets (watermarks) related
175 * to recyling of free vnodes. In the best-operating case, the cache is
176 * exactly full, the free list has size between vlowat and vhiwat above the
177 * free target, and recycling from it and normal use maintains this state.
178 * Sometimes the free list is below vlowat or even empty, but this state
179 * is even better for immediate use provided the cache is not full.
180 * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
181 * ones) to reach one of these states. The watermarks are currently hard-
182 * coded as 4% and 9% of the available space higher. These and the default
183 * of 25% for wantfreevnodes are too large if the memory size is large.
184 * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
185 * whenever vnlru_proc() becomes active.
187 static long wantfreevnodes;
188 static long __exclusive_cache_line freevnodes;
189 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
190 &freevnodes, 0, "Number of \"free\" vnodes");
191 static long freevnodes_old;
193 static counter_u64_t recycles_count;
194 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
195 "Number of vnodes recycled to meet vnode cache targets");
197 static counter_u64_t recycles_free_count;
198 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count,
199 "Number of free vnodes recycled to meet vnode cache targets");
201 static u_long deferred_inact;
202 SYSCTL_ULONG(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD,
203 &deferred_inact, 0, "Number of times inactive processing was deferred");
205 /* To keep more than one thread at a time from running vfs_getnewfsid */
206 static struct mtx mntid_mtx;
209 * Lock for any access to the following:
214 static struct mtx __exclusive_cache_line vnode_list_mtx;
216 /* Publicly exported FS */
217 struct nfs_public nfs_pub;
219 static uma_zone_t buf_trie_zone;
220 static smr_t buf_trie_smr;
222 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
223 static uma_zone_t vnode_zone;
224 MALLOC_DEFINE(M_VNODEPOLL, "VN POLL", "vnode poll");
226 __read_frequently smr_t vfs_smr;
229 * The workitem queue.
231 * It is useful to delay writes of file data and filesystem metadata
232 * for tens of seconds so that quickly created and deleted files need
233 * not waste disk bandwidth being created and removed. To realize this,
234 * we append vnodes to a "workitem" queue. When running with a soft
235 * updates implementation, most pending metadata dependencies should
236 * not wait for more than a few seconds. Thus, mounted on block devices
237 * are delayed only about a half the time that file data is delayed.
238 * Similarly, directory updates are more critical, so are only delayed
239 * about a third the time that file data is delayed. Thus, there are
240 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
241 * one each second (driven off the filesystem syncer process). The
242 * syncer_delayno variable indicates the next queue that is to be processed.
243 * Items that need to be processed soon are placed in this queue:
245 * syncer_workitem_pending[syncer_delayno]
247 * A delay of fifteen seconds is done by placing the request fifteen
248 * entries later in the queue:
250 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
253 static int syncer_delayno;
254 static long syncer_mask;
255 LIST_HEAD(synclist, bufobj);
256 static struct synclist *syncer_workitem_pending;
258 * The sync_mtx protects:
263 * syncer_workitem_pending
264 * syncer_worklist_len
267 static struct mtx sync_mtx;
268 static struct cv sync_wakeup;
270 #define SYNCER_MAXDELAY 32
271 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */
272 static int syncdelay = 30; /* max time to delay syncing data */
273 static int filedelay = 30; /* time to delay syncing files */
274 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
275 "Time to delay syncing files (in seconds)");
276 static int dirdelay = 29; /* time to delay syncing directories */
277 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
278 "Time to delay syncing directories (in seconds)");
279 static int metadelay = 28; /* time to delay syncing metadata */
280 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
281 "Time to delay syncing metadata (in seconds)");
282 static int rushjob; /* number of slots to run ASAP */
283 static int stat_rush_requests; /* number of times I/O speeded up */
284 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
285 "Number of times I/O speeded up (rush requests)");
287 #define VDBATCH_SIZE 8
291 struct vnode *tab[VDBATCH_SIZE];
293 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
295 static void vdbatch_dequeue(struct vnode *vp);
298 * When shutting down the syncer, run it at four times normal speed.
300 #define SYNCER_SHUTDOWN_SPEEDUP 4
301 static int sync_vnode_count;
302 static int syncer_worklist_len;
303 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
306 /* Target for maximum number of vnodes. */
307 u_long desiredvnodes;
308 static u_long gapvnodes; /* gap between wanted and desired */
309 static u_long vhiwat; /* enough extras after expansion */
310 static u_long vlowat; /* minimal extras before expansion */
311 static u_long vstir; /* nonzero to stir non-free vnodes */
312 static volatile int vsmalltrigger = 8; /* pref to keep if > this many pages */
314 static u_long vnlru_read_freevnodes(void);
317 * Note that no attempt is made to sanitize these parameters.
320 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
326 error = sysctl_handle_long(oidp, &val, 0, req);
327 if (error != 0 || req->newptr == NULL)
330 if (val == desiredvnodes)
332 mtx_lock(&vnode_list_mtx);
334 wantfreevnodes = desiredvnodes / 4;
336 mtx_unlock(&vnode_list_mtx);
338 * XXX There is no protection against multiple threads changing
339 * desiredvnodes at the same time. Locking above only helps vnlru and
342 vfs_hash_changesize(desiredvnodes);
343 cache_changesize(desiredvnodes);
347 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
348 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
349 "LU", "Target for maximum number of vnodes");
352 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
357 val = wantfreevnodes;
358 error = sysctl_handle_long(oidp, &val, 0, req);
359 if (error != 0 || req->newptr == NULL)
362 if (val == wantfreevnodes)
364 mtx_lock(&vnode_list_mtx);
365 wantfreevnodes = val;
367 mtx_unlock(&vnode_list_mtx);
371 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
372 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
373 "LU", "Target for minimum number of \"free\" vnodes");
375 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
376 &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
377 static int vnlru_nowhere;
378 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW | CTLFLAG_STATS,
379 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
382 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
387 unsigned long ndflags;
390 if (req->newptr == NULL)
392 if (req->newlen >= PATH_MAX)
395 buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
396 error = SYSCTL_IN(req, buf, req->newlen);
400 buf[req->newlen] = '\0';
402 ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1;
403 NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf);
404 if ((error = namei(&nd)) != 0)
408 if (VN_IS_DOOMED(vp)) {
410 * This vnode is being recycled. Return != 0 to let the caller
411 * know that the sysctl had no effect. Return EAGAIN because a
412 * subsequent call will likely succeed (since namei will create
413 * a new vnode if necessary)
419 counter_u64_add(recycles_count, 1);
430 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
432 struct thread *td = curthread;
438 if (req->newptr == NULL)
441 error = sysctl_handle_int(oidp, &fd, 0, req);
444 error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
449 error = vn_lock(vp, LK_EXCLUSIVE);
453 counter_u64_add(recycles_count, 1);
461 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
462 CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
463 sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
464 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
465 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
466 sysctl_ftry_reclaim_vnode, "I",
467 "Try to reclaim a vnode by its file descriptor");
469 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
472 _Static_assert(sizeof(struct vnode) >= 1UL << vnsz2log &&
473 sizeof(struct vnode) < 1UL << (vnsz2log + 1),
474 "vnsz2log needs to be updated");
478 * Support for the bufobj clean & dirty pctrie.
481 buf_trie_alloc(struct pctrie *ptree)
483 return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT));
487 buf_trie_free(struct pctrie *ptree, void *node)
489 uma_zfree_smr(buf_trie_zone, node);
491 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free,
495 * Initialize the vnode management data structures.
497 * Reevaluate the following cap on the number of vnodes after the physical
498 * memory size exceeds 512GB. In the limit, as the physical memory size
499 * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
501 #ifndef MAXVNODES_MAX
502 #define MAXVNODES_MAX (512UL * 1024 * 1024 / 64) /* 8M */
505 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
507 static struct vnode *
508 vn_alloc_marker(struct mount *mp)
512 vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
513 vp->v_type = VMARKER;
520 vn_free_marker(struct vnode *vp)
523 MPASS(vp->v_type == VMARKER);
524 free(vp, M_VNODE_MARKER);
529 vnode_ctor(void *mem, int size, void *arg __unused, int flags __unused)
531 kasan_mark(mem, size, roundup2(size, UMA_ALIGN_PTR + 1), 0);
536 vnode_dtor(void *mem, int size, void *arg __unused)
538 size_t end1, end2, off1, off2;
540 _Static_assert(offsetof(struct vnode, v_vnodelist) <
541 offsetof(struct vnode, v_dbatchcpu),
542 "KASAN marks require updating");
544 off1 = offsetof(struct vnode, v_vnodelist);
545 off2 = offsetof(struct vnode, v_dbatchcpu);
546 end1 = off1 + sizeof(((struct vnode *)NULL)->v_vnodelist);
547 end2 = off2 + sizeof(((struct vnode *)NULL)->v_dbatchcpu);
550 * Access to the v_vnodelist and v_dbatchcpu fields are permitted even
551 * after the vnode has been freed. Try to get some KASAN coverage by
552 * marking everything except those two fields as invalid. Because
553 * KASAN's tracking is not byte-granular, any preceding fields sharing
554 * the same 8-byte aligned word must also be marked valid.
557 /* Handle the area from the start until v_vnodelist... */
558 off1 = rounddown2(off1, KASAN_SHADOW_SCALE);
559 kasan_mark(mem, off1, off1, KASAN_UMA_FREED);
561 /* ... then the area between v_vnodelist and v_dbatchcpu ... */
562 off1 = roundup2(end1, KASAN_SHADOW_SCALE);
563 off2 = rounddown2(off2, KASAN_SHADOW_SCALE);
565 kasan_mark((void *)((char *)mem + off1), off2 - off1,
566 off2 - off1, KASAN_UMA_FREED);
568 /* ... and finally the area from v_dbatchcpu to the end. */
569 off2 = roundup2(end2, KASAN_SHADOW_SCALE);
570 kasan_mark((void *)((char *)mem + off2), size - off2, size - off2,
576 * Initialize a vnode as it first enters the zone.
579 vnode_init(void *mem, int size, int flags)
588 vp->v_vnlock = &vp->v_lock;
589 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
591 * By default, don't allow shared locks unless filesystems opt-in.
593 lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
594 LK_NOSHARE | LK_IS_VNODE);
598 bufobj_init(&vp->v_bufobj, vp);
600 * Initialize namecache.
602 cache_vnode_init(vp);
604 * Initialize rangelocks.
606 rangelock_init(&vp->v_rl);
608 vp->v_dbatchcpu = NOCPU;
610 vp->v_state = VSTATE_DEAD;
613 * Check vhold_recycle_free for an explanation.
615 vp->v_holdcnt = VHOLD_NO_SMR;
617 mtx_lock(&vnode_list_mtx);
618 TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
619 mtx_unlock(&vnode_list_mtx);
624 * Free a vnode when it is cleared from the zone.
627 vnode_fini(void *mem, int size)
634 mtx_lock(&vnode_list_mtx);
635 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
636 mtx_unlock(&vnode_list_mtx);
637 rangelock_destroy(&vp->v_rl);
638 lockdestroy(vp->v_vnlock);
639 mtx_destroy(&vp->v_interlock);
641 rw_destroy(BO_LOCKPTR(bo));
643 kasan_mark(mem, size, size, 0);
647 * Provide the size of NFS nclnode and NFS fh for calculation of the
648 * vnode memory consumption. The size is specified directly to
649 * eliminate dependency on NFS-private header.
651 * Other filesystems may use bigger or smaller (like UFS and ZFS)
652 * private inode data, but the NFS-based estimation is ample enough.
653 * Still, we care about differences in the size between 64- and 32-bit
656 * Namecache structure size is heuristically
657 * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
660 #define NFS_NCLNODE_SZ (528 + 64)
663 #define NFS_NCLNODE_SZ (360 + 32)
668 vntblinit(void *dummy __unused)
673 int cpu, physvnodes, virtvnodes;
676 * Desiredvnodes is a function of the physical memory size and the
677 * kernel's heap size. Generally speaking, it scales with the
678 * physical memory size. The ratio of desiredvnodes to the physical
679 * memory size is 1:16 until desiredvnodes exceeds 98,304.
681 * marginal ratio of desiredvnodes to the physical memory size is
682 * 1:64. However, desiredvnodes is limited by the kernel's heap
683 * size. The memory required by desiredvnodes vnodes and vm objects
684 * must not exceed 1/10th of the kernel's heap size.
686 physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
687 3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
688 virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
689 sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
690 desiredvnodes = min(physvnodes, virtvnodes);
691 if (desiredvnodes > MAXVNODES_MAX) {
693 printf("Reducing kern.maxvnodes %lu -> %lu\n",
694 desiredvnodes, MAXVNODES_MAX);
695 desiredvnodes = MAXVNODES_MAX;
697 wantfreevnodes = desiredvnodes / 4;
698 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
699 TAILQ_INIT(&vnode_list);
700 mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
702 * The lock is taken to appease WITNESS.
704 mtx_lock(&vnode_list_mtx);
706 mtx_unlock(&vnode_list_mtx);
707 vnode_list_free_marker = vn_alloc_marker(NULL);
708 TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
709 vnode_list_reclaim_marker = vn_alloc_marker(NULL);
710 TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
719 vnode_zone = uma_zcreate("VNODE", sizeof(struct vnode), ctor, dtor,
720 vnode_init, vnode_fini, UMA_ALIGN_PTR, UMA_ZONE_NOKASAN);
721 uma_zone_set_smr(vnode_zone, vfs_smr);
724 * Preallocate enough nodes to support one-per buf so that
725 * we can not fail an insert. reassignbuf() callers can not
726 * tolerate the insertion failure.
728 buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
729 NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
730 UMA_ZONE_NOFREE | UMA_ZONE_SMR);
731 buf_trie_smr = uma_zone_get_smr(buf_trie_zone);
732 uma_prealloc(buf_trie_zone, nbuf);
734 vnodes_created = counter_u64_alloc(M_WAITOK);
735 recycles_count = counter_u64_alloc(M_WAITOK);
736 recycles_free_count = counter_u64_alloc(M_WAITOK);
739 * Initialize the filesystem syncer.
741 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
743 syncer_maxdelay = syncer_mask + 1;
744 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
745 cv_init(&sync_wakeup, "syncer");
748 vd = DPCPU_ID_PTR((cpu), vd);
749 bzero(vd, sizeof(*vd));
750 mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
753 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
756 * Mark a mount point as busy. Used to synchronize access and to delay
757 * unmounting. Eventually, mountlist_mtx is not released on failure.
759 * vfs_busy() is a custom lock, it can block the caller.
760 * vfs_busy() only sleeps if the unmount is active on the mount point.
761 * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
762 * vnode belonging to mp.
764 * Lookup uses vfs_busy() to traverse mount points.
766 * / vnode lock A / vnode lock (/var) D
767 * /var vnode lock B /log vnode lock(/var/log) E
768 * vfs_busy lock C vfs_busy lock F
770 * Within each file system, the lock order is C->A->B and F->D->E.
772 * When traversing across mounts, the system follows that lock order:
778 * The lookup() process for namei("/var") illustrates the process:
779 * 1. VOP_LOOKUP() obtains B while A is held
780 * 2. vfs_busy() obtains a shared lock on F while A and B are held
781 * 3. vput() releases lock on B
782 * 4. vput() releases lock on A
783 * 5. VFS_ROOT() obtains lock on D while shared lock on F is held
784 * 6. vfs_unbusy() releases shared lock on F
785 * 7. vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
786 * Attempt to lock A (instead of vp_crossmp) while D is held would
787 * violate the global order, causing deadlocks.
789 * dounmount() locks B while F is drained. Note that for stacked
790 * filesystems, D and B in the example above may be the same lock,
791 * which introdues potential lock order reversal deadlock between
792 * dounmount() and step 5 above. These filesystems may avoid the LOR
793 * by setting VV_CROSSLOCK on the covered vnode so that lock B will
794 * remain held until after step 5.
797 vfs_busy(struct mount *mp, int flags)
799 struct mount_pcpu *mpcpu;
801 MPASS((flags & ~MBF_MASK) == 0);
802 CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
804 if (vfs_op_thread_enter(mp, mpcpu)) {
805 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
806 MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
807 MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
808 vfs_mp_count_add_pcpu(mpcpu, ref, 1);
809 vfs_mp_count_add_pcpu(mpcpu, lockref, 1);
810 vfs_op_thread_exit(mp, mpcpu);
811 if (flags & MBF_MNTLSTLOCK)
812 mtx_unlock(&mountlist_mtx);
817 vfs_assert_mount_counters(mp);
820 * If mount point is currently being unmounted, sleep until the
821 * mount point fate is decided. If thread doing the unmounting fails,
822 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
823 * that this mount point has survived the unmount attempt and vfs_busy
824 * should retry. Otherwise the unmounter thread will set MNTK_REFEXPIRE
825 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
826 * about to be really destroyed. vfs_busy needs to release its
827 * reference on the mount point in this case and return with ENOENT,
828 * telling the caller the mount it tried to busy is no longer valid.
830 while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
831 KASSERT(TAILQ_EMPTY(&mp->mnt_uppers),
832 ("%s: non-empty upper mount list with pending unmount",
834 if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
837 CTR1(KTR_VFS, "%s: failed busying before sleeping",
841 if (flags & MBF_MNTLSTLOCK)
842 mtx_unlock(&mountlist_mtx);
843 mp->mnt_kern_flag |= MNTK_MWAIT;
844 msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
845 if (flags & MBF_MNTLSTLOCK)
846 mtx_lock(&mountlist_mtx);
849 if (flags & MBF_MNTLSTLOCK)
850 mtx_unlock(&mountlist_mtx);
857 * Free a busy filesystem.
860 vfs_unbusy(struct mount *mp)
862 struct mount_pcpu *mpcpu;
865 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
867 if (vfs_op_thread_enter(mp, mpcpu)) {
868 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
869 vfs_mp_count_sub_pcpu(mpcpu, lockref, 1);
870 vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
871 vfs_op_thread_exit(mp, mpcpu);
876 vfs_assert_mount_counters(mp);
878 c = --mp->mnt_lockref;
879 if (mp->mnt_vfs_ops == 0) {
880 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
885 vfs_dump_mount_counters(mp);
886 if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
887 MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
888 CTR1(KTR_VFS, "%s: waking up waiters", __func__);
889 mp->mnt_kern_flag &= ~MNTK_DRAINING;
890 wakeup(&mp->mnt_lockref);
896 * Lookup a mount point by filesystem identifier.
899 vfs_getvfs(fsid_t *fsid)
903 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
904 mtx_lock(&mountlist_mtx);
905 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
906 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
908 mtx_unlock(&mountlist_mtx);
912 mtx_unlock(&mountlist_mtx);
913 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
914 return ((struct mount *) 0);
918 * Lookup a mount point by filesystem identifier, busying it before
921 * To avoid congestion on mountlist_mtx, implement simple direct-mapped
922 * cache for popular filesystem identifiers. The cache is lockess, using
923 * the fact that struct mount's are never freed. In worst case we may
924 * get pointer to unmounted or even different filesystem, so we have to
925 * check what we got, and go slow way if so.
928 vfs_busyfs(fsid_t *fsid)
930 #define FSID_CACHE_SIZE 256
931 typedef struct mount * volatile vmp_t;
932 static vmp_t cache[FSID_CACHE_SIZE];
937 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
938 hash = fsid->val[0] ^ fsid->val[1];
939 hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
941 if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0)
943 if (vfs_busy(mp, 0) != 0) {
947 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0)
953 mtx_lock(&mountlist_mtx);
954 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
955 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
956 error = vfs_busy(mp, MBF_MNTLSTLOCK);
959 mtx_unlock(&mountlist_mtx);
966 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
967 mtx_unlock(&mountlist_mtx);
968 return ((struct mount *) 0);
972 * Check if a user can access privileged mount options.
975 vfs_suser(struct mount *mp, struct thread *td)
979 if (jailed(td->td_ucred)) {
981 * If the jail of the calling thread lacks permission for
982 * this type of file system, deny immediately.
984 if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
988 * If the file system was mounted outside the jail of the
989 * calling thread, deny immediately.
991 if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
996 * If file system supports delegated administration, we don't check
997 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
998 * by the file system itself.
999 * If this is not the user that did original mount, we check for
1000 * the PRIV_VFS_MOUNT_OWNER privilege.
1002 if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
1003 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
1004 if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
1011 * Get a new unique fsid. Try to make its val[0] unique, since this value
1012 * will be used to create fake device numbers for stat(). Also try (but
1013 * not so hard) make its val[0] unique mod 2^16, since some emulators only
1014 * support 16-bit device numbers. We end up with unique val[0]'s for the
1015 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
1017 * Keep in mind that several mounts may be running in parallel. Starting
1018 * the search one past where the previous search terminated is both a
1019 * micro-optimization and a defense against returning the same fsid to
1023 vfs_getnewfsid(struct mount *mp)
1025 static uint16_t mntid_base;
1030 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
1031 mtx_lock(&mntid_mtx);
1032 mtype = mp->mnt_vfc->vfc_typenum;
1033 tfsid.val[1] = mtype;
1034 mtype = (mtype & 0xFF) << 24;
1036 tfsid.val[0] = makedev(255,
1037 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
1039 if ((nmp = vfs_getvfs(&tfsid)) == NULL)
1043 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
1044 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
1045 mtx_unlock(&mntid_mtx);
1049 * Knob to control the precision of file timestamps:
1051 * 0 = seconds only; nanoseconds zeroed.
1052 * 1 = seconds and nanoseconds, accurate within 1/HZ.
1053 * 2 = seconds and nanoseconds, truncated to microseconds.
1054 * >=3 = seconds and nanoseconds, maximum precision.
1056 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
1058 static int timestamp_precision = TSP_USEC;
1059 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
1060 ×tamp_precision, 0, "File timestamp precision (0: seconds, "
1061 "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1062 "3+: sec + ns (max. precision))");
1065 * Get a current timestamp.
1068 vfs_timestamp(struct timespec *tsp)
1072 switch (timestamp_precision) {
1074 tsp->tv_sec = time_second;
1082 TIMEVAL_TO_TIMESPEC(&tv, tsp);
1092 * Set vnode attributes to VNOVAL
1095 vattr_null(struct vattr *vap)
1098 vap->va_type = VNON;
1099 vap->va_size = VNOVAL;
1100 vap->va_bytes = VNOVAL;
1101 vap->va_mode = VNOVAL;
1102 vap->va_nlink = VNOVAL;
1103 vap->va_uid = VNOVAL;
1104 vap->va_gid = VNOVAL;
1105 vap->va_fsid = VNOVAL;
1106 vap->va_fileid = VNOVAL;
1107 vap->va_blocksize = VNOVAL;
1108 vap->va_rdev = VNOVAL;
1109 vap->va_atime.tv_sec = VNOVAL;
1110 vap->va_atime.tv_nsec = VNOVAL;
1111 vap->va_mtime.tv_sec = VNOVAL;
1112 vap->va_mtime.tv_nsec = VNOVAL;
1113 vap->va_ctime.tv_sec = VNOVAL;
1114 vap->va_ctime.tv_nsec = VNOVAL;
1115 vap->va_birthtime.tv_sec = VNOVAL;
1116 vap->va_birthtime.tv_nsec = VNOVAL;
1117 vap->va_flags = VNOVAL;
1118 vap->va_gen = VNOVAL;
1119 vap->va_vaflags = 0;
1123 * Try to reduce the total number of vnodes.
1125 * This routine (and its user) are buggy in at least the following ways:
1126 * - all parameters were picked years ago when RAM sizes were significantly
1128 * - it can pick vnodes based on pages used by the vm object, but filesystems
1129 * like ZFS don't use it making the pick broken
1130 * - since ZFS has its own aging policy it gets partially combated by this one
1131 * - a dedicated method should be provided for filesystems to let them decide
1132 * whether the vnode should be recycled
1134 * This routine is called when we have too many vnodes. It attempts
1135 * to free <count> vnodes and will potentially free vnodes that still
1136 * have VM backing store (VM backing store is typically the cause
1137 * of a vnode blowout so we want to do this). Therefore, this operation
1138 * is not considered cheap.
1140 * A number of conditions may prevent a vnode from being reclaimed.
1141 * the buffer cache may have references on the vnode, a directory
1142 * vnode may still have references due to the namei cache representing
1143 * underlying files, or the vnode may be in active use. It is not
1144 * desirable to reuse such vnodes. These conditions may cause the
1145 * number of vnodes to reach some minimum value regardless of what
1146 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low.
1148 * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1149 * entries if this argument is strue
1150 * @param trigger Only reclaim vnodes with fewer than this many resident
1152 * @param target How many vnodes to reclaim.
1153 * @return The number of vnodes that were reclaimed.
1156 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1158 struct vnode *vp, *mvp;
1160 struct vm_object *object;
1164 mtx_assert(&vnode_list_mtx, MA_OWNED);
1169 mvp = vnode_list_reclaim_marker;
1172 while (done < target) {
1173 vp = TAILQ_NEXT(vp, v_vnodelist);
1174 if (__predict_false(vp == NULL))
1177 if (__predict_false(vp->v_type == VMARKER))
1181 * If it's been deconstructed already, it's still
1182 * referenced, or it exceeds the trigger, skip it.
1183 * Also skip free vnodes. We are trying to make space
1184 * to expand the free list, not reduce it.
1186 if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1187 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1190 if (vp->v_type == VBAD || vp->v_type == VNON)
1193 object = atomic_load_ptr(&vp->v_object);
1194 if (object == NULL || object->resident_page_count > trigger) {
1199 * Handle races against vnode allocation. Filesystems lock the
1200 * vnode some time after it gets returned from getnewvnode,
1201 * despite type and hold count being manipulated earlier.
1202 * Resorting to checking v_mount restores guarantees present
1203 * before the global list was reworked to contain all vnodes.
1205 if (!VI_TRYLOCK(vp))
1207 if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1211 if (vp->v_mount == NULL) {
1217 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1218 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1219 mtx_unlock(&vnode_list_mtx);
1221 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1223 goto next_iter_unlocked;
1225 if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1227 vn_finished_write(mp);
1228 goto next_iter_unlocked;
1232 if (vp->v_usecount > 0 ||
1233 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1234 (vp->v_object != NULL && vp->v_object->handle == vp &&
1235 vp->v_object->resident_page_count > trigger)) {
1238 vn_finished_write(mp);
1239 goto next_iter_unlocked;
1241 counter_u64_add(recycles_count, 1);
1245 vn_finished_write(mp);
1249 mtx_lock(&vnode_list_mtx);
1252 MPASS(vp->v_type != VMARKER);
1253 if (!should_yield())
1255 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1256 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1257 mtx_unlock(&vnode_list_mtx);
1258 kern_yield(PRI_USER);
1259 mtx_lock(&vnode_list_mtx);
1262 if (done == 0 && !retried) {
1263 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1264 TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1271 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1272 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1274 "limit on vnode free requests per call to the vnlru_free routine");
1277 * Attempt to reduce the free list by the requested amount.
1280 vnlru_free_impl(int count, struct vfsops *mnt_op, struct vnode *mvp)
1286 mtx_assert(&vnode_list_mtx, MA_OWNED);
1287 if (count > max_vnlru_free)
1288 count = max_vnlru_free;
1295 vp = TAILQ_NEXT(vp, v_vnodelist);
1296 if (__predict_false(vp == NULL)) {
1297 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1298 TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1301 if (__predict_false(vp->v_type == VMARKER))
1303 if (vp->v_holdcnt > 0)
1306 * Don't recycle if our vnode is from different type
1307 * of mount point. Note that mp is type-safe, the
1308 * check does not reach unmapped address even if
1309 * vnode is reclaimed.
1311 if (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1312 mp->mnt_op != mnt_op) {
1315 if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1318 if (!vhold_recycle_free(vp))
1320 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1321 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1322 mtx_unlock(&vnode_list_mtx);
1324 * FIXME: ignores the return value, meaning it may be nothing
1325 * got recycled but it claims otherwise to the caller.
1327 * Originally the value started being ignored in 2005 with
1328 * 114a1006a8204aa156e1f9ad6476cdff89cada7f .
1330 * Respecting the value can run into significant stalls if most
1331 * vnodes belong to one file system and it has writes
1332 * suspended. In presence of many threads and millions of
1333 * vnodes they keep contending on the vnode_list_mtx lock only
1334 * to find vnodes they can't recycle.
1336 * The solution would be to pre-check if the vnode is likely to
1337 * be recycle-able, but it needs to happen with the
1338 * vnode_list_mtx lock held. This runs into a problem where
1339 * VOP_GETWRITEMOUNT (currently needed to find out about if
1340 * writes are frozen) can take locks which LOR against it.
1342 * Check nullfs for one example (null_getwritemount).
1346 mtx_lock(&vnode_list_mtx);
1349 return (ocount - count);
1353 vnlru_free_locked(int count)
1356 mtx_assert(&vnode_list_mtx, MA_OWNED);
1357 return (vnlru_free_impl(count, NULL, vnode_list_free_marker));
1361 vnlru_free_vfsops(int count, struct vfsops *mnt_op, struct vnode *mvp)
1364 MPASS(mnt_op != NULL);
1366 VNPASS(mvp->v_type == VMARKER, mvp);
1367 mtx_lock(&vnode_list_mtx);
1368 vnlru_free_impl(count, mnt_op, mvp);
1369 mtx_unlock(&vnode_list_mtx);
1373 vnlru_alloc_marker(void)
1377 mvp = vn_alloc_marker(NULL);
1378 mtx_lock(&vnode_list_mtx);
1379 TAILQ_INSERT_BEFORE(vnode_list_free_marker, mvp, v_vnodelist);
1380 mtx_unlock(&vnode_list_mtx);
1385 vnlru_free_marker(struct vnode *mvp)
1387 mtx_lock(&vnode_list_mtx);
1388 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1389 mtx_unlock(&vnode_list_mtx);
1390 vn_free_marker(mvp);
1397 mtx_assert(&vnode_list_mtx, MA_OWNED);
1398 gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1399 vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1400 vlowat = vhiwat / 2;
1404 * Attempt to recycle vnodes in a context that is always safe to block.
1405 * Calling vlrurecycle() from the bowels of filesystem code has some
1406 * interesting deadlock problems.
1408 static struct proc *vnlruproc;
1409 static int vnlruproc_sig;
1412 * The main freevnodes counter is only updated when threads requeue their vnode
1413 * batches. CPUs are conditionally walked to compute a more accurate total.
1415 * Limit how much of a slop are we willing to tolerate. Note: the actual value
1416 * at any given moment can still exceed slop, but it should not be by significant
1417 * margin in practice.
1419 #define VNLRU_FREEVNODES_SLOP 126
1421 static void __noinline
1422 vfs_freevnodes_rollup(int8_t *lfreevnodes)
1425 atomic_add_long(&freevnodes, *lfreevnodes);
1430 static __inline void
1431 vfs_freevnodes_inc(void)
1433 int8_t *lfreevnodes;
1436 lfreevnodes = PCPU_PTR(vfs_freevnodes);
1438 if (__predict_false(*lfreevnodes == VNLRU_FREEVNODES_SLOP))
1439 vfs_freevnodes_rollup(lfreevnodes);
1444 static __inline void
1445 vfs_freevnodes_dec(void)
1447 int8_t *lfreevnodes;
1450 lfreevnodes = PCPU_PTR(vfs_freevnodes);
1452 if (__predict_false(*lfreevnodes == -VNLRU_FREEVNODES_SLOP))
1453 vfs_freevnodes_rollup(lfreevnodes);
1459 vnlru_read_freevnodes(void)
1461 long slop, rfreevnodes;
1464 rfreevnodes = atomic_load_long(&freevnodes);
1466 if (rfreevnodes > freevnodes_old)
1467 slop = rfreevnodes - freevnodes_old;
1469 slop = freevnodes_old - rfreevnodes;
1470 if (slop < VNLRU_FREEVNODES_SLOP)
1471 return (rfreevnodes >= 0 ? rfreevnodes : 0);
1472 freevnodes_old = rfreevnodes;
1474 freevnodes_old += cpuid_to_pcpu[cpu]->pc_vfs_freevnodes;
1476 return (freevnodes_old >= 0 ? freevnodes_old : 0);
1480 vnlru_under(u_long rnumvnodes, u_long limit)
1482 u_long rfreevnodes, space;
1484 if (__predict_false(rnumvnodes > desiredvnodes))
1487 space = desiredvnodes - rnumvnodes;
1488 if (space < limit) {
1489 rfreevnodes = vnlru_read_freevnodes();
1490 if (rfreevnodes > wantfreevnodes)
1491 space += rfreevnodes - wantfreevnodes;
1493 return (space < limit);
1497 vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1499 long rfreevnodes, space;
1501 if (__predict_false(rnumvnodes > desiredvnodes))
1504 space = desiredvnodes - rnumvnodes;
1505 if (space < limit) {
1506 rfreevnodes = atomic_load_long(&freevnodes);
1507 if (rfreevnodes > wantfreevnodes)
1508 space += rfreevnodes - wantfreevnodes;
1510 return (space < limit);
1517 mtx_assert(&vnode_list_mtx, MA_OWNED);
1518 if (vnlruproc_sig == 0) {
1527 u_long rnumvnodes, rfreevnodes, target;
1528 unsigned long onumvnodes;
1529 int done, force, trigger, usevnodes;
1530 bool reclaim_nc_src, want_reread;
1532 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1533 SHUTDOWN_PRI_FIRST);
1536 want_reread = false;
1538 kproc_suspend_check(vnlruproc);
1539 mtx_lock(&vnode_list_mtx);
1540 rnumvnodes = atomic_load_long(&numvnodes);
1543 force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1544 want_reread = false;
1548 * If numvnodes is too large (due to desiredvnodes being
1549 * adjusted using its sysctl, or emergency growth), first
1550 * try to reduce it by discarding from the free list.
1552 if (rnumvnodes > desiredvnodes) {
1553 vnlru_free_locked(rnumvnodes - desiredvnodes);
1554 rnumvnodes = atomic_load_long(&numvnodes);
1557 * Sleep if the vnode cache is in a good state. This is
1558 * when it is not over-full and has space for about a 4%
1559 * or 9% expansion (by growing its size or inexcessively
1560 * reducing its free list). Otherwise, try to reclaim
1561 * space for a 10% expansion.
1563 if (vstir && force == 0) {
1567 if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1569 wakeup(&vnlruproc_sig);
1570 msleep(vnlruproc, &vnode_list_mtx,
1571 PVFS|PDROP, "vlruwt", hz);
1574 rfreevnodes = vnlru_read_freevnodes();
1576 onumvnodes = rnumvnodes;
1578 * Calculate parameters for recycling. These are the same
1579 * throughout the loop to give some semblance of fairness.
1580 * The trigger point is to avoid recycling vnodes with lots
1581 * of resident pages. We aren't trying to free memory; we
1582 * are trying to recycle or at least free vnodes.
1584 if (rnumvnodes <= desiredvnodes)
1585 usevnodes = rnumvnodes - rfreevnodes;
1587 usevnodes = rnumvnodes;
1591 * The trigger value is chosen to give a conservatively
1592 * large value to ensure that it alone doesn't prevent
1593 * making progress. The value can easily be so large that
1594 * it is effectively infinite in some congested and
1595 * misconfigured cases, and this is necessary. Normally
1596 * it is about 8 to 100 (pages), which is quite large.
1598 trigger = vm_cnt.v_page_count * 2 / usevnodes;
1600 trigger = vsmalltrigger;
1601 reclaim_nc_src = force >= 3;
1602 target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1603 target = target / 10 + 1;
1604 done = vlrureclaim(reclaim_nc_src, trigger, target);
1605 mtx_unlock(&vnode_list_mtx);
1606 if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1607 uma_reclaim(UMA_RECLAIM_DRAIN);
1609 if (force == 0 || force == 1) {
1620 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1623 kern_yield(PRI_USER);
1628 static struct kproc_desc vnlru_kp = {
1633 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1637 * Routines having to do with the management of the vnode table.
1641 * Try to recycle a freed vnode. We abort if anyone picks up a reference
1642 * before we actually vgone(). This function must be called with the vnode
1643 * held to prevent the vnode from being returned to the free list midway
1647 vtryrecycle(struct vnode *vp)
1651 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1652 VNPASS(vp->v_holdcnt > 0, vp);
1654 * This vnode may found and locked via some other list, if so we
1655 * can't recycle it yet.
1657 if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1659 "%s: impossible to recycle, vp %p lock is already held",
1662 return (EWOULDBLOCK);
1665 * Don't recycle if its filesystem is being suspended.
1667 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1670 "%s: impossible to recycle, cannot start the write for %p",
1676 * If we got this far, we need to acquire the interlock and see if
1677 * anyone picked up this vnode from another list. If not, we will
1678 * mark it with DOOMED via vgonel() so that anyone who does find it
1679 * will skip over it.
1682 if (vp->v_usecount) {
1685 vn_finished_write(vnmp);
1687 "%s: impossible to recycle, %p is already referenced",
1691 if (!VN_IS_DOOMED(vp)) {
1692 counter_u64_add(recycles_free_count, 1);
1697 vn_finished_write(vnmp);
1702 * Allocate a new vnode.
1704 * The operation never returns an error. Returning an error was disabled
1705 * in r145385 (dated 2005) with the following comment:
1707 * XXX Not all VFS_VGET/ffs_vget callers check returns.
1709 * Given the age of this commit (almost 15 years at the time of writing this
1710 * comment) restoring the ability to fail requires a significant audit of
1713 * The routine can try to free a vnode or stall for up to 1 second waiting for
1714 * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1716 static u_long vn_alloc_cyclecount;
1718 static struct vnode * __noinline
1719 vn_alloc_hard(struct mount *mp)
1721 u_long rnumvnodes, rfreevnodes;
1723 mtx_lock(&vnode_list_mtx);
1724 rnumvnodes = atomic_load_long(&numvnodes);
1725 if (rnumvnodes + 1 < desiredvnodes) {
1726 vn_alloc_cyclecount = 0;
1729 rfreevnodes = vnlru_read_freevnodes();
1730 if (vn_alloc_cyclecount++ >= rfreevnodes) {
1731 vn_alloc_cyclecount = 0;
1735 * Grow the vnode cache if it will not be above its target max
1736 * after growing. Otherwise, if the free list is nonempty, try
1737 * to reclaim 1 item from it before growing the cache (possibly
1738 * above its target max if the reclamation failed or is delayed).
1739 * Otherwise, wait for some space. In all cases, schedule
1740 * vnlru_proc() if we are getting short of space. The watermarks
1741 * should be chosen so that we never wait or even reclaim from
1742 * the free list to below its target minimum.
1744 if (vnlru_free_locked(1) > 0)
1746 if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1748 * Wait for space for a new vnode.
1751 msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1752 if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1753 vnlru_read_freevnodes() > 1)
1754 vnlru_free_locked(1);
1757 rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1758 if (vnlru_under(rnumvnodes, vlowat))
1760 mtx_unlock(&vnode_list_mtx);
1761 return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1764 static struct vnode *
1765 vn_alloc(struct mount *mp)
1769 if (__predict_false(vn_alloc_cyclecount != 0))
1770 return (vn_alloc_hard(mp));
1771 rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1772 if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1773 atomic_subtract_long(&numvnodes, 1);
1774 return (vn_alloc_hard(mp));
1777 return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1781 vn_free(struct vnode *vp)
1784 atomic_subtract_long(&numvnodes, 1);
1785 uma_zfree_smr(vnode_zone, vp);
1789 * Return the next vnode from the free list.
1792 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1797 struct lock_object *lo;
1799 CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1801 KASSERT(vops->registered,
1802 ("%s: not registered vector op %p\n", __func__, vops));
1805 if (td->td_vp_reserved != NULL) {
1806 vp = td->td_vp_reserved;
1807 td->td_vp_reserved = NULL;
1811 counter_u64_add(vnodes_created, 1);
1813 vn_set_state(vp, VSTATE_UNINITIALIZED);
1816 * Locks are given the generic name "vnode" when created.
1817 * Follow the historic practice of using the filesystem
1818 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1820 * Locks live in a witness group keyed on their name. Thus,
1821 * when a lock is renamed, it must also move from the witness
1822 * group of its old name to the witness group of its new name.
1824 * The change only needs to be made when the vnode moves
1825 * from one filesystem type to another. We ensure that each
1826 * filesystem use a single static name pointer for its tag so
1827 * that we can compare pointers rather than doing a strcmp().
1829 lo = &vp->v_vnlock->lock_object;
1831 if (lo->lo_name != tag) {
1835 WITNESS_DESTROY(lo);
1836 WITNESS_INIT(lo, tag);
1840 * By default, don't allow shared locks unless filesystems opt-in.
1842 vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1844 * Finalize various vnode identity bits.
1846 KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1847 KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1848 KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1852 v_init_counters(vp);
1854 vp->v_bufobj.bo_ops = &buf_ops_bio;
1856 if (mp == NULL && vops != &dead_vnodeops)
1857 printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1861 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1862 mac_vnode_associate_singlelabel(mp, vp);
1865 vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1869 * For the filesystems which do not use vfs_hash_insert(),
1870 * still initialize v_hash to have vfs_hash_index() useful.
1871 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1874 vp->v_hash = (uintptr_t)vp >> vnsz2log;
1881 getnewvnode_reserve(void)
1886 MPASS(td->td_vp_reserved == NULL);
1887 td->td_vp_reserved = vn_alloc(NULL);
1891 getnewvnode_drop_reserve(void)
1896 if (td->td_vp_reserved != NULL) {
1897 vn_free(td->td_vp_reserved);
1898 td->td_vp_reserved = NULL;
1902 static void __noinline
1903 freevnode(struct vnode *vp)
1908 * The vnode has been marked for destruction, so free it.
1910 * The vnode will be returned to the zone where it will
1911 * normally remain until it is needed for another vnode. We
1912 * need to cleanup (or verify that the cleanup has already
1913 * been done) any residual data left from its current use
1914 * so as not to contaminate the freshly allocated vnode.
1916 CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1918 * Paired with vgone.
1920 vn_seqc_write_end_free(vp);
1923 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1924 VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp);
1925 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1926 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1927 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1928 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1929 VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1930 ("clean blk trie not empty"));
1931 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1932 VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1933 ("dirty blk trie not empty"));
1934 VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1935 ("Dangling rangelock waiters"));
1936 VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp,
1937 ("Leaked inactivation"));
1939 cache_assert_no_entries(vp);
1942 mac_vnode_destroy(vp);
1944 if (vp->v_pollinfo != NULL) {
1946 * Use LK_NOWAIT to shut up witness about the lock. We may get
1947 * here while having another vnode locked when trying to
1948 * satisfy a lookup and needing to recycle.
1950 VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT);
1951 destroy_vpollinfo(vp->v_pollinfo);
1953 vp->v_pollinfo = NULL;
1955 vp->v_mountedhere = NULL;
1958 vp->v_fifoinfo = NULL;
1966 * Delete from old mount point vnode list, if on one.
1969 delmntque(struct vnode *vp)
1973 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
1979 VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1980 ("bad mount point vnode list size"));
1981 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1982 mp->mnt_nvnodelistsize--;
1986 * The caller expects the interlock to be still held.
1988 ASSERT_VI_LOCKED(vp, __func__);
1992 insmntque1_int(struct vnode *vp, struct mount *mp, bool dtr)
1995 KASSERT(vp->v_mount == NULL,
1996 ("insmntque: vnode already on per mount vnode list"));
1997 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1998 if ((mp->mnt_kern_flag & MNTK_UNLOCKED_INSMNTQUE) == 0) {
1999 ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
2002 ("%s: can't have MNTK_UNLOCKED_INSMNTQUE and cleanup",
2007 * We acquire the vnode interlock early to ensure that the
2008 * vnode cannot be recycled by another process releasing a
2009 * holdcnt on it before we get it on both the vnode list
2010 * and the active vnode list. The mount mutex protects only
2011 * manipulation of the vnode list and the vnode freelist
2012 * mutex protects only manipulation of the active vnode list.
2013 * Hence the need to hold the vnode interlock throughout.
2017 if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
2018 ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
2019 mp->mnt_nvnodelistsize == 0)) &&
2020 (vp->v_vflag & VV_FORCEINSMQ) == 0) {
2025 vp->v_op = &dead_vnodeops;
2033 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2034 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
2035 ("neg mount point vnode list size"));
2036 mp->mnt_nvnodelistsize++;
2043 * Insert into list of vnodes for the new mount point, if available.
2044 * insmntque() reclaims the vnode on insertion failure, insmntque1()
2045 * leaves handling of the vnode to the caller.
2048 insmntque(struct vnode *vp, struct mount *mp)
2050 return (insmntque1_int(vp, mp, true));
2054 insmntque1(struct vnode *vp, struct mount *mp)
2056 return (insmntque1_int(vp, mp, false));
2060 * Flush out and invalidate all buffers associated with a bufobj
2061 * Called with the underlying object locked.
2064 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
2069 if (flags & V_SAVE) {
2070 error = bufobj_wwait(bo, slpflag, slptimeo);
2075 if (bo->bo_dirty.bv_cnt > 0) {
2078 error = BO_SYNC(bo, MNT_WAIT);
2079 } while (error == ERELOOKUP);
2083 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) {
2090 * If you alter this loop please notice that interlock is dropped and
2091 * reacquired in flushbuflist. Special care is needed to ensure that
2092 * no race conditions occur from this.
2095 error = flushbuflist(&bo->bo_clean,
2096 flags, bo, slpflag, slptimeo);
2097 if (error == 0 && !(flags & V_CLEANONLY))
2098 error = flushbuflist(&bo->bo_dirty,
2099 flags, bo, slpflag, slptimeo);
2100 if (error != 0 && error != EAGAIN) {
2104 } while (error != 0);
2107 * Wait for I/O to complete. XXX needs cleaning up. The vnode can
2108 * have write I/O in-progress but if there is a VM object then the
2109 * VM object can also have read-I/O in-progress.
2112 bufobj_wwait(bo, 0, 0);
2113 if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
2115 vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
2118 } while (bo->bo_numoutput > 0);
2122 * Destroy the copy in the VM cache, too.
2124 if (bo->bo_object != NULL &&
2125 (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
2126 VM_OBJECT_WLOCK(bo->bo_object);
2127 vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
2128 OBJPR_CLEANONLY : 0);
2129 VM_OBJECT_WUNLOCK(bo->bo_object);
2134 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
2135 V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
2136 bo->bo_clean.bv_cnt > 0))
2137 panic("vinvalbuf: flush failed");
2138 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
2139 bo->bo_dirty.bv_cnt > 0)
2140 panic("vinvalbuf: flush dirty failed");
2147 * Flush out and invalidate all buffers associated with a vnode.
2148 * Called with the underlying object locked.
2151 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
2154 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2155 ASSERT_VOP_LOCKED(vp, "vinvalbuf");
2156 if (vp->v_object != NULL && vp->v_object->handle != vp)
2158 return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
2162 * Flush out buffers on the specified list.
2166 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2169 struct buf *bp, *nbp;
2174 ASSERT_BO_WLOCKED(bo);
2177 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2179 * If we are flushing both V_NORMAL and V_ALT buffers then
2180 * do not skip any buffers. If we are flushing only V_NORMAL
2181 * buffers then skip buffers marked as BX_ALTDATA. If we are
2182 * flushing only V_ALT buffers then skip buffers not marked
2185 if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2186 (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2187 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2191 lblkno = nbp->b_lblkno;
2192 xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2195 error = BUF_TIMELOCK(bp,
2196 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2197 "flushbuf", slpflag, slptimeo);
2200 return (error != ENOLCK ? error : EAGAIN);
2202 KASSERT(bp->b_bufobj == bo,
2203 ("bp %p wrong b_bufobj %p should be %p",
2204 bp, bp->b_bufobj, bo));
2206 * XXX Since there are no node locks for NFS, I
2207 * believe there is a slight chance that a delayed
2208 * write will occur while sleeping just above, so
2211 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2214 bp->b_flags |= B_ASYNC;
2217 return (EAGAIN); /* XXX: why not loop ? */
2220 bp->b_flags |= (B_INVAL | B_RELBUF);
2221 bp->b_flags &= ~B_ASYNC;
2226 nbp = gbincore(bo, lblkno);
2227 if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2229 break; /* nbp invalid */
2235 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2241 ASSERT_BO_LOCKED(bo);
2243 for (lblkno = startn;;) {
2245 bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2246 if (bp == NULL || bp->b_lblkno >= endn ||
2247 bp->b_lblkno < startn)
2249 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2250 LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2253 if (error == ENOLCK)
2257 KASSERT(bp->b_bufobj == bo,
2258 ("bp %p wrong b_bufobj %p should be %p",
2259 bp, bp->b_bufobj, bo));
2260 lblkno = bp->b_lblkno + 1;
2261 if ((bp->b_flags & B_MANAGED) == 0)
2263 bp->b_flags |= B_RELBUF;
2265 * In the VMIO case, use the B_NOREUSE flag to hint that the
2266 * pages backing each buffer in the range are unlikely to be
2267 * reused. Dirty buffers will have the hint applied once
2268 * they've been written.
2270 if ((bp->b_flags & B_VMIO) != 0)
2271 bp->b_flags |= B_NOREUSE;
2279 * Truncate a file's buffer and pages to a specified length. This
2280 * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2284 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2286 struct buf *bp, *nbp;
2290 CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2291 vp, blksize, (uintmax_t)length);
2294 * Round up to the *next* lbn.
2296 startlbn = howmany(length, blksize);
2298 ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2304 while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2309 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2310 if (bp->b_lblkno > 0)
2313 * Since we hold the vnode lock this should only
2314 * fail if we're racing with the buf daemon.
2317 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2318 BO_LOCKPTR(bo)) == ENOLCK)
2319 goto restart_unlocked;
2321 VNASSERT((bp->b_flags & B_DELWRI), vp,
2322 ("buf(%p) on dirty queue without DELWRI", bp));
2331 bufobj_wwait(bo, 0, 0);
2333 vnode_pager_setsize(vp, length);
2339 * Invalidate the cached pages of a file's buffer within the range of block
2340 * numbers [startlbn, endlbn).
2343 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2349 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2351 start = blksize * startlbn;
2352 end = blksize * endlbn;
2356 MPASS(blksize == bo->bo_bsize);
2358 while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2362 vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2366 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2367 daddr_t startlbn, daddr_t endlbn)
2369 struct buf *bp, *nbp;
2372 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2373 ASSERT_BO_LOCKED(bo);
2377 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2378 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2381 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2382 BO_LOCKPTR(bo)) == ENOLCK) {
2388 bp->b_flags |= B_INVAL | B_RELBUF;
2389 bp->b_flags &= ~B_ASYNC;
2395 (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2397 (nbp->b_flags & B_DELWRI) != 0))
2401 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2402 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2405 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2406 BO_LOCKPTR(bo)) == ENOLCK) {
2411 bp->b_flags |= B_INVAL | B_RELBUF;
2412 bp->b_flags &= ~B_ASYNC;
2418 (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2419 (nbp->b_vp != vp) ||
2420 (nbp->b_flags & B_DELWRI) == 0))
2428 buf_vlist_remove(struct buf *bp)
2433 flags = bp->b_xflags;
2435 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2436 ASSERT_BO_WLOCKED(bp->b_bufobj);
2437 KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 &&
2438 (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN),
2439 ("%s: buffer %p has invalid queue state", __func__, bp));
2441 if ((flags & BX_VNDIRTY) != 0)
2442 bv = &bp->b_bufobj->bo_dirty;
2444 bv = &bp->b_bufobj->bo_clean;
2445 BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2446 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2448 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2452 * Add the buffer to the sorted clean or dirty block list.
2454 * NOTE: xflags is passed as a constant, optimizing this inline function!
2457 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2463 ASSERT_BO_WLOCKED(bo);
2464 KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2465 ("buf_vlist_add: bo %p does not allow bufs", bo));
2466 KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2467 ("dead bo %p", bo));
2468 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2469 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2470 bp->b_xflags |= xflags;
2471 if (xflags & BX_VNDIRTY)
2477 * Keep the list ordered. Optimize empty list insertion. Assume
2478 * we tend to grow at the tail so lookup_le should usually be cheaper
2481 if (bv->bv_cnt == 0 ||
2482 bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2483 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2484 else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2485 TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2487 TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2488 error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2490 panic("buf_vlist_add: Preallocated nodes insufficient.");
2495 * Look up a buffer using the buffer tries.
2498 gbincore(struct bufobj *bo, daddr_t lblkno)
2502 ASSERT_BO_LOCKED(bo);
2503 bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2506 return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno));
2510 * Look up a buf using the buffer tries, without the bufobj lock. This relies
2511 * on SMR for safe lookup, and bufs being in a no-free zone to provide type
2512 * stability of the result. Like other lockless lookups, the found buf may
2513 * already be invalid by the time this function returns.
2516 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno)
2520 ASSERT_BO_UNLOCKED(bo);
2521 bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno);
2524 return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno));
2528 * Associate a buffer with a vnode.
2531 bgetvp(struct vnode *vp, struct buf *bp)
2536 ASSERT_BO_WLOCKED(bo);
2537 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2539 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2540 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2541 ("bgetvp: bp already attached! %p", bp));
2547 * Insert onto list for new vnode.
2549 buf_vlist_add(bp, bo, BX_VNCLEAN);
2553 * Disassociate a buffer from a vnode.
2556 brelvp(struct buf *bp)
2561 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2562 KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2565 * Delete from old vnode list, if on one.
2567 vp = bp->b_vp; /* XXX */
2570 buf_vlist_remove(bp);
2571 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2572 bo->bo_flag &= ~BO_ONWORKLST;
2573 mtx_lock(&sync_mtx);
2574 LIST_REMOVE(bo, bo_synclist);
2575 syncer_worklist_len--;
2576 mtx_unlock(&sync_mtx);
2579 bp->b_bufobj = NULL;
2585 * Add an item to the syncer work queue.
2588 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2592 ASSERT_BO_WLOCKED(bo);
2594 mtx_lock(&sync_mtx);
2595 if (bo->bo_flag & BO_ONWORKLST)
2596 LIST_REMOVE(bo, bo_synclist);
2598 bo->bo_flag |= BO_ONWORKLST;
2599 syncer_worklist_len++;
2602 if (delay > syncer_maxdelay - 2)
2603 delay = syncer_maxdelay - 2;
2604 slot = (syncer_delayno + delay) & syncer_mask;
2606 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2607 mtx_unlock(&sync_mtx);
2611 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2615 mtx_lock(&sync_mtx);
2616 len = syncer_worklist_len - sync_vnode_count;
2617 mtx_unlock(&sync_mtx);
2618 error = SYSCTL_OUT(req, &len, sizeof(len));
2622 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2623 CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2624 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2626 static struct proc *updateproc;
2627 static void sched_sync(void);
2628 static struct kproc_desc up_kp = {
2633 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2636 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2641 *bo = LIST_FIRST(slp);
2645 if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2648 * We use vhold in case the vnode does not
2649 * successfully sync. vhold prevents the vnode from
2650 * going away when we unlock the sync_mtx so that
2651 * we can acquire the vnode interlock.
2654 mtx_unlock(&sync_mtx);
2656 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2658 mtx_lock(&sync_mtx);
2659 return (*bo == LIST_FIRST(slp));
2661 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2662 (void) VOP_FSYNC(vp, MNT_LAZY, td);
2664 vn_finished_write(mp);
2666 if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2668 * Put us back on the worklist. The worklist
2669 * routine will remove us from our current
2670 * position and then add us back in at a later
2673 vn_syncer_add_to_worklist(*bo, syncdelay);
2677 mtx_lock(&sync_mtx);
2681 static int first_printf = 1;
2684 * System filesystem synchronizer daemon.
2689 struct synclist *next, *slp;
2692 struct thread *td = curthread;
2694 int net_worklist_len;
2695 int syncer_final_iter;
2699 syncer_final_iter = 0;
2700 syncer_state = SYNCER_RUNNING;
2701 starttime = time_uptime;
2702 td->td_pflags |= TDP_NORUNNINGBUF;
2704 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2707 mtx_lock(&sync_mtx);
2709 if (syncer_state == SYNCER_FINAL_DELAY &&
2710 syncer_final_iter == 0) {
2711 mtx_unlock(&sync_mtx);
2712 kproc_suspend_check(td->td_proc);
2713 mtx_lock(&sync_mtx);
2715 net_worklist_len = syncer_worklist_len - sync_vnode_count;
2716 if (syncer_state != SYNCER_RUNNING &&
2717 starttime != time_uptime) {
2719 printf("\nSyncing disks, vnodes remaining... ");
2722 printf("%d ", net_worklist_len);
2724 starttime = time_uptime;
2727 * Push files whose dirty time has expired. Be careful
2728 * of interrupt race on slp queue.
2730 * Skip over empty worklist slots when shutting down.
2733 slp = &syncer_workitem_pending[syncer_delayno];
2734 syncer_delayno += 1;
2735 if (syncer_delayno == syncer_maxdelay)
2737 next = &syncer_workitem_pending[syncer_delayno];
2739 * If the worklist has wrapped since the
2740 * it was emptied of all but syncer vnodes,
2741 * switch to the FINAL_DELAY state and run
2742 * for one more second.
2744 if (syncer_state == SYNCER_SHUTTING_DOWN &&
2745 net_worklist_len == 0 &&
2746 last_work_seen == syncer_delayno) {
2747 syncer_state = SYNCER_FINAL_DELAY;
2748 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2750 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2751 syncer_worklist_len > 0);
2754 * Keep track of the last time there was anything
2755 * on the worklist other than syncer vnodes.
2756 * Return to the SHUTTING_DOWN state if any
2759 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2760 last_work_seen = syncer_delayno;
2761 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2762 syncer_state = SYNCER_SHUTTING_DOWN;
2763 while (!LIST_EMPTY(slp)) {
2764 error = sync_vnode(slp, &bo, td);
2766 LIST_REMOVE(bo, bo_synclist);
2767 LIST_INSERT_HEAD(next, bo, bo_synclist);
2771 if (first_printf == 0) {
2773 * Drop the sync mutex, because some watchdog
2774 * drivers need to sleep while patting
2776 mtx_unlock(&sync_mtx);
2777 wdog_kern_pat(WD_LASTVAL);
2778 mtx_lock(&sync_mtx);
2781 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2782 syncer_final_iter--;
2784 * The variable rushjob allows the kernel to speed up the
2785 * processing of the filesystem syncer process. A rushjob
2786 * value of N tells the filesystem syncer to process the next
2787 * N seconds worth of work on its queue ASAP. Currently rushjob
2788 * is used by the soft update code to speed up the filesystem
2789 * syncer process when the incore state is getting so far
2790 * ahead of the disk that the kernel memory pool is being
2791 * threatened with exhaustion.
2798 * Just sleep for a short period of time between
2799 * iterations when shutting down to allow some I/O
2802 * If it has taken us less than a second to process the
2803 * current work, then wait. Otherwise start right over
2804 * again. We can still lose time if any single round
2805 * takes more than two seconds, but it does not really
2806 * matter as we are just trying to generally pace the
2807 * filesystem activity.
2809 if (syncer_state != SYNCER_RUNNING ||
2810 time_uptime == starttime) {
2812 sched_prio(td, PPAUSE);
2815 if (syncer_state != SYNCER_RUNNING)
2816 cv_timedwait(&sync_wakeup, &sync_mtx,
2817 hz / SYNCER_SHUTDOWN_SPEEDUP);
2818 else if (time_uptime == starttime)
2819 cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2824 * Request the syncer daemon to speed up its work.
2825 * We never push it to speed up more than half of its
2826 * normal turn time, otherwise it could take over the cpu.
2829 speedup_syncer(void)
2833 mtx_lock(&sync_mtx);
2834 if (rushjob < syncdelay / 2) {
2836 stat_rush_requests += 1;
2839 mtx_unlock(&sync_mtx);
2840 cv_broadcast(&sync_wakeup);
2845 * Tell the syncer to speed up its work and run though its work
2846 * list several times, then tell it to shut down.
2849 syncer_shutdown(void *arg, int howto)
2852 if (howto & RB_NOSYNC)
2854 mtx_lock(&sync_mtx);
2855 syncer_state = SYNCER_SHUTTING_DOWN;
2857 mtx_unlock(&sync_mtx);
2858 cv_broadcast(&sync_wakeup);
2859 kproc_shutdown(arg, howto);
2863 syncer_suspend(void)
2866 syncer_shutdown(updateproc, 0);
2873 mtx_lock(&sync_mtx);
2875 syncer_state = SYNCER_RUNNING;
2876 mtx_unlock(&sync_mtx);
2877 cv_broadcast(&sync_wakeup);
2878 kproc_resume(updateproc);
2882 * Move the buffer between the clean and dirty lists of its vnode.
2885 reassignbuf(struct buf *bp)
2897 KASSERT((bp->b_flags & B_PAGING) == 0,
2898 ("%s: cannot reassign paging buffer %p", __func__, bp));
2900 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2901 bp, bp->b_vp, bp->b_flags);
2904 buf_vlist_remove(bp);
2907 * If dirty, put on list of dirty buffers; otherwise insert onto list
2910 if (bp->b_flags & B_DELWRI) {
2911 if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2912 switch (vp->v_type) {
2922 vn_syncer_add_to_worklist(bo, delay);
2924 buf_vlist_add(bp, bo, BX_VNDIRTY);
2926 buf_vlist_add(bp, bo, BX_VNCLEAN);
2928 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2929 mtx_lock(&sync_mtx);
2930 LIST_REMOVE(bo, bo_synclist);
2931 syncer_worklist_len--;
2932 mtx_unlock(&sync_mtx);
2933 bo->bo_flag &= ~BO_ONWORKLST;
2938 bp = TAILQ_FIRST(&bv->bv_hd);
2939 KASSERT(bp == NULL || bp->b_bufobj == bo,
2940 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2941 bp = TAILQ_LAST(&bv->bv_hd, buflists);
2942 KASSERT(bp == NULL || bp->b_bufobj == bo,
2943 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2945 bp = TAILQ_FIRST(&bv->bv_hd);
2946 KASSERT(bp == NULL || bp->b_bufobj == bo,
2947 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2948 bp = TAILQ_LAST(&bv->bv_hd, buflists);
2949 KASSERT(bp == NULL || bp->b_bufobj == bo,
2950 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2956 v_init_counters(struct vnode *vp)
2959 VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2960 vp, ("%s called for an initialized vnode", __FUNCTION__));
2961 ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2963 refcount_init(&vp->v_holdcnt, 1);
2964 refcount_init(&vp->v_usecount, 1);
2968 * Grab a particular vnode from the free list, increment its
2969 * reference count and lock it. VIRF_DOOMED is set if the vnode
2970 * is being destroyed. Only callers who specify LK_RETRY will
2971 * see doomed vnodes. If inactive processing was delayed in
2972 * vput try to do it here.
2974 * usecount is manipulated using atomics without holding any locks.
2976 * holdcnt can be manipulated using atomics without holding any locks,
2977 * except when transitioning 1<->0, in which case the interlock is held.
2979 * Consumers which don't guarantee liveness of the vnode can use SMR to
2980 * try to get a reference. Note this operation can fail since the vnode
2981 * may be awaiting getting freed by the time they get to it.
2984 vget_prep_smr(struct vnode *vp)
2988 VFS_SMR_ASSERT_ENTERED();
2990 if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
3002 vget_prep(struct vnode *vp)
3006 if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
3016 vget_abort(struct vnode *vp, enum vgetstate vs)
3027 __assert_unreachable();
3032 vget(struct vnode *vp, int flags)
3037 return (vget_finish(vp, flags, vs));
3041 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
3045 if ((flags & LK_INTERLOCK) != 0)
3046 ASSERT_VI_LOCKED(vp, __func__);
3048 ASSERT_VI_UNLOCKED(vp, __func__);
3049 VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3050 VNPASS(vp->v_holdcnt > 0, vp);
3051 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3053 error = vn_lock(vp, flags);
3054 if (__predict_false(error != 0)) {
3056 CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
3061 vget_finish_ref(vp, vs);
3066 vget_finish_ref(struct vnode *vp, enum vgetstate vs)
3070 VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3071 VNPASS(vp->v_holdcnt > 0, vp);
3072 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3074 if (vs == VGET_USECOUNT)
3078 * We hold the vnode. If the usecount is 0 it will be utilized to keep
3079 * the vnode around. Otherwise someone else lended their hold count and
3080 * we have to drop ours.
3082 old = atomic_fetchadd_int(&vp->v_usecount, 1);
3083 VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3086 old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3087 VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3089 refcount_release(&vp->v_holdcnt);
3095 vref(struct vnode *vp)
3099 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3101 vget_finish_ref(vp, vs);
3105 vrefact(struct vnode *vp)
3108 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3110 int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3111 VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3113 refcount_acquire(&vp->v_usecount);
3118 vlazy(struct vnode *vp)
3122 VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3124 if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3127 * We may get here for inactive routines after the vnode got doomed.
3129 if (VN_IS_DOOMED(vp))
3132 mtx_lock(&mp->mnt_listmtx);
3133 if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3134 vp->v_mflag |= VMP_LAZYLIST;
3135 TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3136 mp->mnt_lazyvnodelistsize++;
3138 mtx_unlock(&mp->mnt_listmtx);
3142 vunlazy(struct vnode *vp)
3146 ASSERT_VI_LOCKED(vp, __func__);
3147 VNPASS(!VN_IS_DOOMED(vp), vp);
3150 mtx_lock(&mp->mnt_listmtx);
3151 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3153 * Don't remove the vnode from the lazy list if another thread
3154 * has increased the hold count. It may have re-enqueued the
3155 * vnode to the lazy list and is now responsible for its
3158 if (vp->v_holdcnt == 0) {
3159 vp->v_mflag &= ~VMP_LAZYLIST;
3160 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3161 mp->mnt_lazyvnodelistsize--;
3163 mtx_unlock(&mp->mnt_listmtx);
3167 * This routine is only meant to be called from vgonel prior to dooming
3171 vunlazy_gone(struct vnode *vp)
3175 ASSERT_VOP_ELOCKED(vp, __func__);
3176 ASSERT_VI_LOCKED(vp, __func__);
3177 VNPASS(!VN_IS_DOOMED(vp), vp);
3179 if (vp->v_mflag & VMP_LAZYLIST) {
3181 mtx_lock(&mp->mnt_listmtx);
3182 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3183 vp->v_mflag &= ~VMP_LAZYLIST;
3184 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3185 mp->mnt_lazyvnodelistsize--;
3186 mtx_unlock(&mp->mnt_listmtx);
3191 vdefer_inactive(struct vnode *vp)
3194 ASSERT_VI_LOCKED(vp, __func__);
3195 VNPASS(vp->v_holdcnt > 0, vp);
3196 if (VN_IS_DOOMED(vp)) {
3200 if (vp->v_iflag & VI_DEFINACT) {
3201 VNPASS(vp->v_holdcnt > 1, vp);
3205 if (vp->v_usecount > 0) {
3206 vp->v_iflag &= ~VI_OWEINACT;
3211 vp->v_iflag |= VI_DEFINACT;
3213 atomic_add_long(&deferred_inact, 1);
3217 vdefer_inactive_unlocked(struct vnode *vp)
3221 if ((vp->v_iflag & VI_OWEINACT) == 0) {
3225 vdefer_inactive(vp);
3228 enum vput_op { VRELE, VPUT, VUNREF };
3231 * Handle ->v_usecount transitioning to 0.
3233 * By releasing the last usecount we take ownership of the hold count which
3234 * provides liveness of the vnode, meaning we have to vdrop.
3236 * For all vnodes we may need to perform inactive processing. It requires an
3237 * exclusive lock on the vnode, while it is legal to call here with only a
3238 * shared lock (or no locks). If locking the vnode in an expected manner fails,
3239 * inactive processing gets deferred to the syncer.
3241 * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3242 * on the lock being held all the way until VOP_INACTIVE. This in particular
3243 * happens with UFS which adds half-constructed vnodes to the hash, where they
3244 * can be found by other code.
3247 vput_final(struct vnode *vp, enum vput_op func)
3252 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3253 VNPASS(vp->v_holdcnt > 0, vp);
3258 * By the time we got here someone else might have transitioned
3259 * the count back to > 0.
3261 if (vp->v_usecount > 0)
3265 * If the vnode is doomed vgone already performed inactive processing
3268 if (VN_IS_DOOMED(vp))
3271 if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3274 if (vp->v_iflag & VI_DOINGINACT)
3278 * Locking operations here will drop the interlock and possibly the
3279 * vnode lock, opening a window where the vnode can get doomed all the
3280 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3283 vp->v_iflag |= VI_OWEINACT;
3284 want_unlock = false;
3288 switch (VOP_ISLOCKED(vp)) {
3294 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3299 * The lock has at least one sharer, but we have no way
3300 * to conclude whether this is us. Play it safe and
3309 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3310 error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3316 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3317 error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3323 if (func == VUNREF) {
3324 VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp,
3325 ("recursive vunref"));
3326 vp->v_vflag |= VV_UNREF;
3329 error = vinactive(vp);
3332 if (error != ERELOOKUP || !want_unlock)
3334 VOP_LOCK(vp, LK_EXCLUSIVE);
3337 vp->v_vflag &= ~VV_UNREF;
3340 vdefer_inactive(vp);
3350 * Decrement ->v_usecount for a vnode.
3352 * Releasing the last use count requires additional processing, see vput_final
3353 * above for details.
3355 * Comment above each variant denotes lock state on entry and exit.
3360 * out: same as passed in
3363 vrele(struct vnode *vp)
3366 ASSERT_VI_UNLOCKED(vp, __func__);
3367 if (!refcount_release(&vp->v_usecount))
3369 vput_final(vp, VRELE);
3377 vput(struct vnode *vp)
3380 ASSERT_VOP_LOCKED(vp, __func__);
3381 ASSERT_VI_UNLOCKED(vp, __func__);
3382 if (!refcount_release(&vp->v_usecount)) {
3386 vput_final(vp, VPUT);
3394 vunref(struct vnode *vp)
3397 ASSERT_VOP_LOCKED(vp, __func__);
3398 ASSERT_VI_UNLOCKED(vp, __func__);
3399 if (!refcount_release(&vp->v_usecount))
3401 vput_final(vp, VUNREF);
3405 vhold(struct vnode *vp)
3409 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3410 old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3411 VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3412 ("%s: wrong hold count %d", __func__, old));
3414 vfs_freevnodes_dec();
3418 vholdnz(struct vnode *vp)
3421 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3423 int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3424 VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3425 ("%s: wrong hold count %d", __func__, old));
3427 atomic_add_int(&vp->v_holdcnt, 1);
3432 * Grab a hold count unless the vnode is freed.
3434 * Only use this routine if vfs smr is the only protection you have against
3435 * freeing the vnode.
3437 * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag
3438 * is not set. After the flag is set the vnode becomes immutable to anyone but
3439 * the thread which managed to set the flag.
3441 * It may be tempting to replace the loop with:
3442 * count = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3443 * if (count & VHOLD_NO_SMR) {
3444 * backpedal and error out;
3447 * However, while this is more performant, it hinders debugging by eliminating
3448 * the previously mentioned invariant.
3451 vhold_smr(struct vnode *vp)
3455 VFS_SMR_ASSERT_ENTERED();
3457 count = atomic_load_int(&vp->v_holdcnt);
3459 if (count & VHOLD_NO_SMR) {
3460 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3461 ("non-zero hold count with flags %d\n", count));
3464 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3465 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3467 vfs_freevnodes_dec();
3474 * Hold a free vnode for recycling.
3476 * Note: vnode_init references this comment.
3478 * Attempts to recycle only need the global vnode list lock and have no use for
3481 * However, vnodes get inserted into the global list before they get fully
3482 * initialized and stay there until UMA decides to free the memory. This in
3483 * particular means the target can be found before it becomes usable and after
3484 * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to
3487 * Note: the vnode may gain more references after we transition the count 0->1.
3490 vhold_recycle_free(struct vnode *vp)
3494 mtx_assert(&vnode_list_mtx, MA_OWNED);
3496 count = atomic_load_int(&vp->v_holdcnt);
3498 if (count & VHOLD_NO_SMR) {
3499 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3500 ("non-zero hold count with flags %d\n", count));
3503 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3507 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3508 vfs_freevnodes_dec();
3514 static void __noinline
3515 vdbatch_process(struct vdbatch *vd)
3520 mtx_assert(&vd->lock, MA_OWNED);
3521 MPASS(curthread->td_pinned > 0);
3522 MPASS(vd->index == VDBATCH_SIZE);
3525 if (mtx_trylock(&vnode_list_mtx)) {
3526 for (i = 0; i < VDBATCH_SIZE; i++) {
3529 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3530 TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3531 MPASS(vp->v_dbatchcpu != NOCPU);
3532 vp->v_dbatchcpu = NOCPU;
3534 mtx_unlock(&vnode_list_mtx);
3536 for (i = 0; i < VDBATCH_SIZE; i++) {
3539 MPASS(vp->v_dbatchcpu != NOCPU);
3540 vp->v_dbatchcpu = NOCPU;
3548 vdbatch_enqueue(struct vnode *vp)
3552 ASSERT_VI_LOCKED(vp, __func__);
3553 VNPASS(!VN_IS_DOOMED(vp), vp);
3555 if (vp->v_dbatchcpu != NOCPU) {
3562 mtx_lock(&vd->lock);
3563 MPASS(vd->index < VDBATCH_SIZE);
3564 MPASS(vd->tab[vd->index] == NULL);
3566 * A hack: we depend on being pinned so that we know what to put in
3569 vp->v_dbatchcpu = curcpu;
3570 vd->tab[vd->index] = vp;
3573 if (vd->index == VDBATCH_SIZE)
3574 vdbatch_process(vd);
3575 mtx_unlock(&vd->lock);
3580 * This routine must only be called for vnodes which are about to be
3581 * deallocated. Supporting dequeue for arbitrary vndoes would require
3582 * validating that the locked batch matches.
3585 vdbatch_dequeue(struct vnode *vp)
3591 VNPASS(vp->v_type == VBAD || vp->v_type == VNON, vp);
3593 cpu = vp->v_dbatchcpu;
3597 vd = DPCPU_ID_PTR(cpu, vd);
3598 mtx_lock(&vd->lock);
3599 for (i = 0; i < vd->index; i++) {
3600 if (vd->tab[i] != vp)
3602 vp->v_dbatchcpu = NOCPU;
3604 vd->tab[i] = vd->tab[vd->index];
3605 vd->tab[vd->index] = NULL;
3608 mtx_unlock(&vd->lock);
3610 * Either we dequeued the vnode above or the target CPU beat us to it.
3612 MPASS(vp->v_dbatchcpu == NOCPU);
3616 * Drop the hold count of the vnode. If this is the last reference to
3617 * the vnode we place it on the free list unless it has been vgone'd
3618 * (marked VIRF_DOOMED) in which case we will free it.
3620 * Because the vnode vm object keeps a hold reference on the vnode if
3621 * there is at least one resident non-cached page, the vnode cannot
3622 * leave the active list without the page cleanup done.
3624 static void __noinline
3625 vdropl_final(struct vnode *vp)
3628 ASSERT_VI_LOCKED(vp, __func__);
3629 VNPASS(VN_IS_DOOMED(vp), vp);
3631 * Set the VHOLD_NO_SMR flag.
3633 * We may be racing against vhold_smr. If they win we can just pretend
3634 * we never got this far, they will vdrop later.
3636 if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) {
3637 vfs_freevnodes_inc();
3640 * We lost the aforementioned race. Any subsequent access is
3641 * invalid as they might have managed to vdropl on their own.
3646 * Don't bump freevnodes as this one is going away.
3652 vdrop(struct vnode *vp)
3655 ASSERT_VI_UNLOCKED(vp, __func__);
3656 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3657 if (refcount_release_if_not_last(&vp->v_holdcnt))
3663 static void __always_inline
3664 vdropl_impl(struct vnode *vp, bool enqueue)
3667 ASSERT_VI_LOCKED(vp, __func__);
3668 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3669 if (!refcount_release(&vp->v_holdcnt)) {
3673 VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp);
3674 VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
3675 if (VN_IS_DOOMED(vp)) {
3680 vfs_freevnodes_inc();
3681 if (vp->v_mflag & VMP_LAZYLIST) {
3691 * Also unlocks the interlock. We can't assert on it as we
3692 * released our hold and by now the vnode might have been
3695 vdbatch_enqueue(vp);
3699 vdropl(struct vnode *vp)
3702 vdropl_impl(vp, true);
3706 * vdrop a vnode when recycling
3708 * This is a special case routine only to be used when recycling, differs from
3709 * regular vdrop by not requeieing the vnode on LRU.
3711 * Consider a case where vtryrecycle continuously fails with all vnodes (due to
3712 * e.g., frozen writes on the filesystem), filling the batch and causing it to
3713 * be requeued. Then vnlru will end up revisiting the same vnodes. This is a
3714 * loop which can last for as long as writes are frozen.
3717 vdropl_recycle(struct vnode *vp)
3720 vdropl_impl(vp, false);
3724 vdrop_recycle(struct vnode *vp)
3732 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3733 * flags. DOINGINACT prevents us from recursing in calls to vinactive.
3736 vinactivef(struct vnode *vp)
3738 struct vm_object *obj;
3741 ASSERT_VOP_ELOCKED(vp, "vinactive");
3742 ASSERT_VI_LOCKED(vp, "vinactive");
3743 VNPASS((vp->v_iflag & VI_DOINGINACT) == 0, vp);
3744 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3745 vp->v_iflag |= VI_DOINGINACT;
3746 vp->v_iflag &= ~VI_OWEINACT;
3749 * Before moving off the active list, we must be sure that any
3750 * modified pages are converted into the vnode's dirty
3751 * buffers, since these will no longer be checked once the
3752 * vnode is on the inactive list.
3754 * The write-out of the dirty pages is asynchronous. At the
3755 * point that VOP_INACTIVE() is called, there could still be
3756 * pending I/O and dirty pages in the object.
3758 if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3759 vm_object_mightbedirty(obj)) {
3760 VM_OBJECT_WLOCK(obj);
3761 vm_object_page_clean(obj, 0, 0, 0);
3762 VM_OBJECT_WUNLOCK(obj);
3764 error = VOP_INACTIVE(vp);
3766 VNPASS(vp->v_iflag & VI_DOINGINACT, vp);
3767 vp->v_iflag &= ~VI_DOINGINACT;
3772 vinactive(struct vnode *vp)
3775 ASSERT_VOP_ELOCKED(vp, "vinactive");
3776 ASSERT_VI_LOCKED(vp, "vinactive");
3777 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3779 if ((vp->v_iflag & VI_OWEINACT) == 0)
3781 if (vp->v_iflag & VI_DOINGINACT)
3783 if (vp->v_usecount > 0) {
3784 vp->v_iflag &= ~VI_OWEINACT;
3787 return (vinactivef(vp));
3791 * Remove any vnodes in the vnode table belonging to mount point mp.
3793 * If FORCECLOSE is not specified, there should not be any active ones,
3794 * return error if any are found (nb: this is a user error, not a
3795 * system error). If FORCECLOSE is specified, detach any active vnodes
3798 * If WRITECLOSE is set, only flush out regular file vnodes open for
3801 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3803 * `rootrefs' specifies the base reference count for the root vnode
3804 * of this filesystem. The root vnode is considered busy if its
3805 * v_usecount exceeds this value. On a successful return, vflush(, td)
3806 * will call vrele() on the root vnode exactly rootrefs times.
3807 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3811 static int busyprt = 0; /* print out busy vnodes */
3812 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3816 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3818 struct vnode *vp, *mvp, *rootvp = NULL;
3820 int busy = 0, error;
3822 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3825 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3826 ("vflush: bad args"));
3828 * Get the filesystem root vnode. We can vput() it
3829 * immediately, since with rootrefs > 0, it won't go away.
3831 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3832 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3839 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3841 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3844 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3848 * Skip over a vnodes marked VV_SYSTEM.
3850 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3856 * If WRITECLOSE is set, flush out unlinked but still open
3857 * files (even if open only for reading) and regular file
3858 * vnodes open for writing.
3860 if (flags & WRITECLOSE) {
3861 if (vp->v_object != NULL) {
3862 VM_OBJECT_WLOCK(vp->v_object);
3863 vm_object_page_clean(vp->v_object, 0, 0, 0);
3864 VM_OBJECT_WUNLOCK(vp->v_object);
3867 error = VOP_FSYNC(vp, MNT_WAIT, td);
3868 } while (error == ERELOOKUP);
3872 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3875 error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3878 if ((vp->v_type == VNON ||
3879 (error == 0 && vattr.va_nlink > 0)) &&
3880 (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3888 * With v_usecount == 0, all we need to do is clear out the
3889 * vnode data structures and we are done.
3891 * If FORCECLOSE is set, forcibly close the vnode.
3893 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3899 vn_printf(vp, "vflush: busy vnode ");
3905 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3907 * If just the root vnode is busy, and if its refcount
3908 * is equal to `rootrefs', then go ahead and kill it.
3911 KASSERT(busy > 0, ("vflush: not busy"));
3912 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3913 ("vflush: usecount %d < rootrefs %d",
3914 rootvp->v_usecount, rootrefs));
3915 if (busy == 1 && rootvp->v_usecount == rootrefs) {
3916 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3924 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3928 for (; rootrefs > 0; rootrefs--)
3934 * Recycle an unused vnode to the front of the free list.
3937 vrecycle(struct vnode *vp)
3942 recycled = vrecyclel(vp);
3948 * vrecycle, with the vp interlock held.
3951 vrecyclel(struct vnode *vp)
3955 ASSERT_VOP_ELOCKED(vp, __func__);
3956 ASSERT_VI_LOCKED(vp, __func__);
3957 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3959 if (vp->v_usecount == 0) {
3967 * Eliminate all activity associated with a vnode
3968 * in preparation for reuse.
3971 vgone(struct vnode *vp)
3979 * Notify upper mounts about reclaimed or unlinked vnode.
3982 vfs_notify_upper(struct vnode *vp, enum vfs_notify_upper_type event)
3985 struct mount_upper_node *ump;
3987 mp = atomic_load_ptr(&vp->v_mount);
3990 if (TAILQ_EMPTY(&mp->mnt_notify))
3994 mp->mnt_upper_pending++;
3995 KASSERT(mp->mnt_upper_pending > 0,
3996 ("%s: mnt_upper_pending %d", __func__, mp->mnt_upper_pending));
3997 TAILQ_FOREACH(ump, &mp->mnt_notify, mnt_upper_link) {
4000 case VFS_NOTIFY_UPPER_RECLAIM:
4001 VFS_RECLAIM_LOWERVP(ump->mp, vp);
4003 case VFS_NOTIFY_UPPER_UNLINK:
4004 VFS_UNLINK_LOWERVP(ump->mp, vp);
4009 mp->mnt_upper_pending--;
4010 if ((mp->mnt_kern_flag & MNTK_UPPER_WAITER) != 0 &&
4011 mp->mnt_upper_pending == 0) {
4012 mp->mnt_kern_flag &= ~MNTK_UPPER_WAITER;
4013 wakeup(&mp->mnt_uppers);
4019 * vgone, with the vp interlock held.
4022 vgonel(struct vnode *vp)
4027 bool active, doinginact, oweinact;
4029 ASSERT_VOP_ELOCKED(vp, "vgonel");
4030 ASSERT_VI_LOCKED(vp, "vgonel");
4031 VNASSERT(vp->v_holdcnt, vp,
4032 ("vgonel: vp %p has no reference.", vp));
4033 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4037 * Don't vgonel if we're already doomed.
4039 if (VN_IS_DOOMED(vp)) {
4040 VNPASS(vn_get_state(vp) == VSTATE_DESTROYING || \
4041 vn_get_state(vp) == VSTATE_DEAD, vp);
4045 * Paired with freevnode.
4047 vn_seqc_write_begin_locked(vp);
4049 vn_irflag_set_locked(vp, VIRF_DOOMED);
4050 vn_set_state(vp, VSTATE_DESTROYING);
4053 * Check to see if the vnode is in use. If so, we have to
4054 * call VOP_CLOSE() and VOP_INACTIVE().
4056 * It could be that VOP_INACTIVE() requested reclamation, in
4057 * which case we should avoid recursion, so check
4058 * VI_DOINGINACT. This is not precise but good enough.
4060 active = vp->v_usecount > 0;
4061 oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4062 doinginact = (vp->v_iflag & VI_DOINGINACT) != 0;
4065 * If we need to do inactive VI_OWEINACT will be set.
4067 if (vp->v_iflag & VI_DEFINACT) {
4068 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
4069 vp->v_iflag &= ~VI_DEFINACT;
4072 VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
4075 cache_purge_vgone(vp);
4076 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
4079 * If purging an active vnode, it must be closed and
4080 * deactivated before being reclaimed.
4083 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
4086 if (oweinact || active) {
4089 oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4094 if (vp->v_type == VSOCK)
4095 vfs_unp_reclaim(vp);
4098 * Clean out any buffers associated with the vnode.
4099 * If the flush fails, just toss the buffers.
4102 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
4103 (void) vn_start_secondary_write(vp, &mp, V_WAIT);
4104 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
4105 while (vinvalbuf(vp, 0, 0, 0) != 0)
4109 BO_LOCK(&vp->v_bufobj);
4110 KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
4111 vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
4112 TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
4113 vp->v_bufobj.bo_clean.bv_cnt == 0,
4114 ("vp %p bufobj not invalidated", vp));
4117 * For VMIO bufobj, BO_DEAD is set later, or in
4118 * vm_object_terminate() after the object's page queue is
4121 object = vp->v_bufobj.bo_object;
4123 vp->v_bufobj.bo_flag |= BO_DEAD;
4124 BO_UNLOCK(&vp->v_bufobj);
4127 * Handle the VM part. Tmpfs handles v_object on its own (the
4128 * OBJT_VNODE check). Nullfs or other bypassing filesystems
4129 * should not touch the object borrowed from the lower vnode
4130 * (the handle check).
4132 if (object != NULL && object->type == OBJT_VNODE &&
4133 object->handle == vp)
4134 vnode_destroy_vobject(vp);
4137 * Reclaim the vnode.
4139 if (VOP_RECLAIM(vp))
4140 panic("vgone: cannot reclaim");
4142 vn_finished_secondary_write(mp);
4143 VNASSERT(vp->v_object == NULL, vp,
4144 ("vop_reclaim left v_object vp=%p", vp));
4146 * Clear the advisory locks and wake up waiting threads.
4148 if (vp->v_lockf != NULL) {
4149 (void)VOP_ADVLOCKPURGE(vp);
4153 * Delete from old mount point vnode list.
4155 if (vp->v_mount == NULL) {
4159 ASSERT_VI_LOCKED(vp, "vgonel 2");
4162 * Done with purge, reset to the standard lock and invalidate
4165 vp->v_vnlock = &vp->v_lock;
4166 vp->v_op = &dead_vnodeops;
4168 vn_set_state(vp, VSTATE_DEAD);
4172 * Print out a description of a vnode.
4174 static const char *const vtypename[] = {
4184 [VMARKER] = "VMARKER",
4186 _Static_assert(nitems(vtypename) == VLASTTYPE + 1,
4187 "vnode type name not added to vtypename");
4189 static const char *const vstatename[] = {
4190 [VSTATE_UNINITIALIZED] = "VSTATE_UNINITIALIZED",
4191 [VSTATE_CONSTRUCTED] = "VSTATE_CONSTRUCTED",
4192 [VSTATE_DESTROYING] = "VSTATE_DESTROYING",
4193 [VSTATE_DEAD] = "VSTATE_DEAD",
4195 _Static_assert(nitems(vstatename) == VLASTSTATE + 1,
4196 "vnode state name not added to vstatename");
4198 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0,
4199 "new hold count flag not added to vn_printf");
4202 vn_printf(struct vnode *vp, const char *fmt, ...)
4205 char buf[256], buf2[16];
4213 printf("%p: ", (void *)vp);
4214 printf("type %s state %s\n", vtypename[vp->v_type], vstatename[vp->v_state]);
4215 holdcnt = atomic_load_int(&vp->v_holdcnt);
4216 printf(" usecount %d, writecount %d, refcount %d seqc users %d",
4217 vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS,
4219 switch (vp->v_type) {
4221 printf(" mountedhere %p\n", vp->v_mountedhere);
4224 printf(" rdev %p\n", vp->v_rdev);
4227 printf(" socket %p\n", vp->v_unpcb);
4230 printf(" fifoinfo %p\n", vp->v_fifoinfo);
4238 if (holdcnt & VHOLD_NO_SMR)
4239 strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf));
4240 printf(" hold count flags (%s)\n", buf + 1);
4244 irflag = vn_irflag_read(vp);
4245 if (irflag & VIRF_DOOMED)
4246 strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4247 if (irflag & VIRF_PGREAD)
4248 strlcat(buf, "|VIRF_PGREAD", sizeof(buf));
4249 if (irflag & VIRF_MOUNTPOINT)
4250 strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf));
4251 if (irflag & VIRF_TEXT_REF)
4252 strlcat(buf, "|VIRF_TEXT_REF", sizeof(buf));
4253 flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT | VIRF_TEXT_REF);
4255 snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4256 strlcat(buf, buf2, sizeof(buf));
4258 if (vp->v_vflag & VV_ROOT)
4259 strlcat(buf, "|VV_ROOT", sizeof(buf));
4260 if (vp->v_vflag & VV_ISTTY)
4261 strlcat(buf, "|VV_ISTTY", sizeof(buf));
4262 if (vp->v_vflag & VV_NOSYNC)
4263 strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4264 if (vp->v_vflag & VV_ETERNALDEV)
4265 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4266 if (vp->v_vflag & VV_CACHEDLABEL)
4267 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4268 if (vp->v_vflag & VV_VMSIZEVNLOCK)
4269 strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4270 if (vp->v_vflag & VV_COPYONWRITE)
4271 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4272 if (vp->v_vflag & VV_SYSTEM)
4273 strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4274 if (vp->v_vflag & VV_PROCDEP)
4275 strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4276 if (vp->v_vflag & VV_DELETED)
4277 strlcat(buf, "|VV_DELETED", sizeof(buf));
4278 if (vp->v_vflag & VV_MD)
4279 strlcat(buf, "|VV_MD", sizeof(buf));
4280 if (vp->v_vflag & VV_FORCEINSMQ)
4281 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4282 if (vp->v_vflag & VV_READLINK)
4283 strlcat(buf, "|VV_READLINK", sizeof(buf));
4284 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4285 VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM |
4286 VV_PROCDEP | VV_DELETED | VV_MD | VV_FORCEINSMQ | VV_READLINK);
4288 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4289 strlcat(buf, buf2, sizeof(buf));
4291 if (vp->v_iflag & VI_MOUNT)
4292 strlcat(buf, "|VI_MOUNT", sizeof(buf));
4293 if (vp->v_iflag & VI_DOINGINACT)
4294 strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4295 if (vp->v_iflag & VI_OWEINACT)
4296 strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4297 if (vp->v_iflag & VI_DEFINACT)
4298 strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4299 if (vp->v_iflag & VI_FOPENING)
4300 strlcat(buf, "|VI_FOPENING", sizeof(buf));
4301 flags = vp->v_iflag & ~(VI_MOUNT | VI_DOINGINACT |
4302 VI_OWEINACT | VI_DEFINACT | VI_FOPENING);
4304 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4305 strlcat(buf, buf2, sizeof(buf));
4307 if (vp->v_mflag & VMP_LAZYLIST)
4308 strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4309 flags = vp->v_mflag & ~(VMP_LAZYLIST);
4311 snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4312 strlcat(buf, buf2, sizeof(buf));
4314 printf(" flags (%s)", buf + 1);
4315 if (mtx_owned(VI_MTX(vp)))
4316 printf(" VI_LOCKed");
4318 if (vp->v_object != NULL)
4319 printf(" v_object %p ref %d pages %d "
4320 "cleanbuf %d dirtybuf %d\n",
4321 vp->v_object, vp->v_object->ref_count,
4322 vp->v_object->resident_page_count,
4323 vp->v_bufobj.bo_clean.bv_cnt,
4324 vp->v_bufobj.bo_dirty.bv_cnt);
4326 lockmgr_printinfo(vp->v_vnlock);
4327 if (vp->v_data != NULL)
4333 * List all of the locked vnodes in the system.
4334 * Called when debugging the kernel.
4336 DB_SHOW_COMMAND_FLAGS(lockedvnods, lockedvnodes, DB_CMD_MEMSAFE)
4342 * Note: because this is DDB, we can't obey the locking semantics
4343 * for these structures, which means we could catch an inconsistent
4344 * state and dereference a nasty pointer. Not much to be done
4347 db_printf("Locked vnodes\n");
4348 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4349 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4350 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4351 vn_printf(vp, "vnode ");
4357 * Show details about the given vnode.
4359 DB_SHOW_COMMAND(vnode, db_show_vnode)
4365 vp = (struct vnode *)addr;
4366 vn_printf(vp, "vnode ");
4370 * Show details about the given mount point.
4372 DB_SHOW_COMMAND(mount, db_show_mount)
4383 /* No address given, print short info about all mount points. */
4384 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4385 db_printf("%p %s on %s (%s)\n", mp,
4386 mp->mnt_stat.f_mntfromname,
4387 mp->mnt_stat.f_mntonname,
4388 mp->mnt_stat.f_fstypename);
4392 db_printf("\nMore info: show mount <addr>\n");
4396 mp = (struct mount *)addr;
4397 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4398 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4401 mflags = mp->mnt_flag;
4402 #define MNT_FLAG(flag) do { \
4403 if (mflags & (flag)) { \
4404 if (buf[0] != '\0') \
4405 strlcat(buf, ", ", sizeof(buf)); \
4406 strlcat(buf, (#flag) + 4, sizeof(buf)); \
4407 mflags &= ~(flag); \
4410 MNT_FLAG(MNT_RDONLY);
4411 MNT_FLAG(MNT_SYNCHRONOUS);
4412 MNT_FLAG(MNT_NOEXEC);
4413 MNT_FLAG(MNT_NOSUID);
4414 MNT_FLAG(MNT_NFS4ACLS);
4415 MNT_FLAG(MNT_UNION);
4416 MNT_FLAG(MNT_ASYNC);
4417 MNT_FLAG(MNT_SUIDDIR);
4418 MNT_FLAG(MNT_SOFTDEP);
4419 MNT_FLAG(MNT_NOSYMFOLLOW);
4420 MNT_FLAG(MNT_GJOURNAL);
4421 MNT_FLAG(MNT_MULTILABEL);
4423 MNT_FLAG(MNT_NOATIME);
4424 MNT_FLAG(MNT_NOCLUSTERR);
4425 MNT_FLAG(MNT_NOCLUSTERW);
4427 MNT_FLAG(MNT_EXRDONLY);
4428 MNT_FLAG(MNT_EXPORTED);
4429 MNT_FLAG(MNT_DEFEXPORTED);
4430 MNT_FLAG(MNT_EXPORTANON);
4431 MNT_FLAG(MNT_EXKERB);
4432 MNT_FLAG(MNT_EXPUBLIC);
4433 MNT_FLAG(MNT_LOCAL);
4434 MNT_FLAG(MNT_QUOTA);
4435 MNT_FLAG(MNT_ROOTFS);
4437 MNT_FLAG(MNT_IGNORE);
4438 MNT_FLAG(MNT_UPDATE);
4439 MNT_FLAG(MNT_DELEXPORT);
4440 MNT_FLAG(MNT_RELOAD);
4441 MNT_FLAG(MNT_FORCE);
4442 MNT_FLAG(MNT_SNAPSHOT);
4443 MNT_FLAG(MNT_BYFSID);
4447 strlcat(buf, ", ", sizeof(buf));
4448 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4449 "0x%016jx", mflags);
4451 db_printf(" mnt_flag = %s\n", buf);
4454 flags = mp->mnt_kern_flag;
4455 #define MNT_KERN_FLAG(flag) do { \
4456 if (flags & (flag)) { \
4457 if (buf[0] != '\0') \
4458 strlcat(buf, ", ", sizeof(buf)); \
4459 strlcat(buf, (#flag) + 5, sizeof(buf)); \
4463 MNT_KERN_FLAG(MNTK_UNMOUNTF);
4464 MNT_KERN_FLAG(MNTK_ASYNC);
4465 MNT_KERN_FLAG(MNTK_SOFTDEP);
4466 MNT_KERN_FLAG(MNTK_NOMSYNC);
4467 MNT_KERN_FLAG(MNTK_DRAINING);
4468 MNT_KERN_FLAG(MNTK_REFEXPIRE);
4469 MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4470 MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4471 MNT_KERN_FLAG(MNTK_NO_IOPF);
4472 MNT_KERN_FLAG(MNTK_RECURSE);
4473 MNT_KERN_FLAG(MNTK_UPPER_WAITER);
4474 MNT_KERN_FLAG(MNTK_UNLOCKED_INSMNTQUE);
4475 MNT_KERN_FLAG(MNTK_USES_BCACHE);
4476 MNT_KERN_FLAG(MNTK_VMSETSIZE_BUG);
4477 MNT_KERN_FLAG(MNTK_FPLOOKUP);
4478 MNT_KERN_FLAG(MNTK_TASKQUEUE_WAITER);
4479 MNT_KERN_FLAG(MNTK_NOASYNC);
4480 MNT_KERN_FLAG(MNTK_UNMOUNT);
4481 MNT_KERN_FLAG(MNTK_MWAIT);
4482 MNT_KERN_FLAG(MNTK_SUSPEND);
4483 MNT_KERN_FLAG(MNTK_SUSPEND2);
4484 MNT_KERN_FLAG(MNTK_SUSPENDED);
4485 MNT_KERN_FLAG(MNTK_NULL_NOCACHE);
4486 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4487 #undef MNT_KERN_FLAG
4490 strlcat(buf, ", ", sizeof(buf));
4491 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4494 db_printf(" mnt_kern_flag = %s\n", buf);
4496 db_printf(" mnt_opt = ");
4497 opt = TAILQ_FIRST(mp->mnt_opt);
4499 db_printf("%s", opt->name);
4500 opt = TAILQ_NEXT(opt, link);
4501 while (opt != NULL) {
4502 db_printf(", %s", opt->name);
4503 opt = TAILQ_NEXT(opt, link);
4509 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx "
4510 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4511 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4512 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4513 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4514 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4515 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4516 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4517 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4518 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4519 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4520 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4522 db_printf(" mnt_cred = { uid=%u ruid=%u",
4523 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4524 if (jailed(mp->mnt_cred))
4525 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4527 db_printf(" mnt_ref = %d (with %d in the struct)\n",
4528 vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4529 db_printf(" mnt_gen = %d\n", mp->mnt_gen);
4530 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4531 db_printf(" mnt_lazyvnodelistsize = %d\n",
4532 mp->mnt_lazyvnodelistsize);
4533 db_printf(" mnt_writeopcount = %d (with %d in the struct)\n",
4534 vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4535 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4536 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed);
4537 db_printf(" mnt_lockref = %d (with %d in the struct)\n",
4538 vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4539 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4540 db_printf(" mnt_secondary_accwrites = %d\n",
4541 mp->mnt_secondary_accwrites);
4542 db_printf(" mnt_gjprovider = %s\n",
4543 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4544 db_printf(" mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4546 db_printf("\n\nList of active vnodes\n");
4547 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4548 if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4549 vn_printf(vp, "vnode ");
4554 db_printf("\n\nList of inactive vnodes\n");
4555 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4556 if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4557 vn_printf(vp, "vnode ");
4566 * Fill in a struct xvfsconf based on a struct vfsconf.
4569 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4571 struct xvfsconf xvfsp;
4573 bzero(&xvfsp, sizeof(xvfsp));
4574 strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4575 xvfsp.vfc_typenum = vfsp->vfc_typenum;
4576 xvfsp.vfc_refcount = vfsp->vfc_refcount;
4577 xvfsp.vfc_flags = vfsp->vfc_flags;
4579 * These are unused in userland, we keep them
4580 * to not break binary compatibility.
4582 xvfsp.vfc_vfsops = NULL;
4583 xvfsp.vfc_next = NULL;
4584 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4587 #ifdef COMPAT_FREEBSD32
4589 uint32_t vfc_vfsops;
4590 char vfc_name[MFSNAMELEN];
4591 int32_t vfc_typenum;
4592 int32_t vfc_refcount;
4598 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4600 struct xvfsconf32 xvfsp;
4602 bzero(&xvfsp, sizeof(xvfsp));
4603 strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4604 xvfsp.vfc_typenum = vfsp->vfc_typenum;
4605 xvfsp.vfc_refcount = vfsp->vfc_refcount;
4606 xvfsp.vfc_flags = vfsp->vfc_flags;
4607 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4612 * Top level filesystem related information gathering.
4615 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4617 struct vfsconf *vfsp;
4622 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4623 #ifdef COMPAT_FREEBSD32
4624 if (req->flags & SCTL_MASK32)
4625 error = vfsconf2x32(req, vfsp);
4628 error = vfsconf2x(req, vfsp);
4636 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4637 CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4638 "S,xvfsconf", "List of all configured filesystems");
4640 #ifndef BURN_BRIDGES
4641 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4644 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4646 int *name = (int *)arg1 - 1; /* XXX */
4647 u_int namelen = arg2 + 1; /* XXX */
4648 struct vfsconf *vfsp;
4650 log(LOG_WARNING, "userland calling deprecated sysctl, "
4651 "please rebuild world\n");
4653 #if 1 || defined(COMPAT_PRELITE2)
4654 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4656 return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4660 case VFS_MAXTYPENUM:
4663 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4666 return (ENOTDIR); /* overloaded */
4668 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4669 if (vfsp->vfc_typenum == name[2])
4674 return (EOPNOTSUPP);
4675 #ifdef COMPAT_FREEBSD32
4676 if (req->flags & SCTL_MASK32)
4677 return (vfsconf2x32(req, vfsp));
4680 return (vfsconf2x(req, vfsp));
4682 return (EOPNOTSUPP);
4685 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4686 CTLFLAG_MPSAFE, vfs_sysctl,
4687 "Generic filesystem");
4689 #if 1 || defined(COMPAT_PRELITE2)
4692 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4695 struct vfsconf *vfsp;
4696 struct ovfsconf ovfs;
4699 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4700 bzero(&ovfs, sizeof(ovfs));
4701 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */
4702 strcpy(ovfs.vfc_name, vfsp->vfc_name);
4703 ovfs.vfc_index = vfsp->vfc_typenum;
4704 ovfs.vfc_refcount = vfsp->vfc_refcount;
4705 ovfs.vfc_flags = vfsp->vfc_flags;
4706 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4716 #endif /* 1 || COMPAT_PRELITE2 */
4717 #endif /* !BURN_BRIDGES */
4720 unmount_or_warn(struct mount *mp)
4724 error = dounmount(mp, MNT_FORCE, curthread);
4726 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4730 printf("%d)\n", error);
4735 * Unmount all filesystems. The list is traversed in reverse order
4736 * of mounting to avoid dependencies.
4739 vfs_unmountall(void)
4741 struct mount *mp, *tmp;
4743 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4746 * Since this only runs when rebooting, it is not interlocked.
4748 TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4752 * Forcibly unmounting "/dev" before "/" would prevent clean
4753 * unmount of the latter.
4755 if (mp == rootdevmp)
4758 unmount_or_warn(mp);
4761 if (rootdevmp != NULL)
4762 unmount_or_warn(rootdevmp);
4766 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4769 ASSERT_VI_LOCKED(vp, __func__);
4770 VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
4771 if ((vp->v_iflag & VI_OWEINACT) == 0) {
4775 if (vn_lock(vp, lkflags) == 0) {
4782 vdefer_inactive_unlocked(vp);
4786 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4789 return (vp->v_iflag & VI_DEFINACT);
4792 static void __noinline
4793 vfs_periodic_inactive(struct mount *mp, int flags)
4795 struct vnode *vp, *mvp;
4798 lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4799 if (flags != MNT_WAIT)
4800 lkflags |= LK_NOWAIT;
4802 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4803 if ((vp->v_iflag & VI_DEFINACT) == 0) {
4807 vp->v_iflag &= ~VI_DEFINACT;
4808 vfs_deferred_inactive(vp, lkflags);
4813 vfs_want_msync(struct vnode *vp)
4815 struct vm_object *obj;
4818 * This test may be performed without any locks held.
4819 * We rely on vm_object's type stability.
4821 if (vp->v_vflag & VV_NOSYNC)
4824 return (obj != NULL && vm_object_mightbedirty(obj));
4828 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4831 if (vp->v_vflag & VV_NOSYNC)
4833 if (vp->v_iflag & VI_DEFINACT)
4835 return (vfs_want_msync(vp));
4838 static void __noinline
4839 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4841 struct vnode *vp, *mvp;
4842 struct vm_object *obj;
4843 int lkflags, objflags;
4846 lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4847 if (flags != MNT_WAIT) {
4848 lkflags |= LK_NOWAIT;
4849 objflags = OBJPC_NOSYNC;
4851 objflags = OBJPC_SYNC;
4854 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4856 if (vp->v_iflag & VI_DEFINACT) {
4857 vp->v_iflag &= ~VI_DEFINACT;
4860 if (!vfs_want_msync(vp)) {
4862 vfs_deferred_inactive(vp, lkflags);
4867 if (vget(vp, lkflags) == 0) {
4869 if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4870 VM_OBJECT_WLOCK(obj);
4871 vm_object_page_clean(obj, 0, 0, objflags);
4872 VM_OBJECT_WUNLOCK(obj);
4879 vdefer_inactive_unlocked(vp);
4885 vfs_periodic(struct mount *mp, int flags)
4888 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4890 if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4891 vfs_periodic_inactive(mp, flags);
4893 vfs_periodic_msync_inactive(mp, flags);
4897 destroy_vpollinfo_free(struct vpollinfo *vi)
4900 knlist_destroy(&vi->vpi_selinfo.si_note);
4901 mtx_destroy(&vi->vpi_lock);
4902 free(vi, M_VNODEPOLL);
4906 destroy_vpollinfo(struct vpollinfo *vi)
4909 knlist_clear(&vi->vpi_selinfo.si_note, 1);
4910 seldrain(&vi->vpi_selinfo);
4911 destroy_vpollinfo_free(vi);
4915 * Initialize per-vnode helper structure to hold poll-related state.
4918 v_addpollinfo(struct vnode *vp)
4920 struct vpollinfo *vi;
4922 if (vp->v_pollinfo != NULL)
4924 vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO);
4925 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4926 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4927 vfs_knlunlock, vfs_knl_assert_lock);
4929 if (vp->v_pollinfo != NULL) {
4931 destroy_vpollinfo_free(vi);
4934 vp->v_pollinfo = vi;
4939 * Record a process's interest in events which might happen to
4940 * a vnode. Because poll uses the historic select-style interface
4941 * internally, this routine serves as both the ``check for any
4942 * pending events'' and the ``record my interest in future events''
4943 * functions. (These are done together, while the lock is held,
4944 * to avoid race conditions.)
4947 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4951 mtx_lock(&vp->v_pollinfo->vpi_lock);
4952 if (vp->v_pollinfo->vpi_revents & events) {
4954 * This leaves events we are not interested
4955 * in available for the other process which
4956 * which presumably had requested them
4957 * (otherwise they would never have been
4960 events &= vp->v_pollinfo->vpi_revents;
4961 vp->v_pollinfo->vpi_revents &= ~events;
4963 mtx_unlock(&vp->v_pollinfo->vpi_lock);
4966 vp->v_pollinfo->vpi_events |= events;
4967 selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4968 mtx_unlock(&vp->v_pollinfo->vpi_lock);
4973 * Routine to create and manage a filesystem syncer vnode.
4975 #define sync_close ((int (*)(struct vop_close_args *))nullop)
4976 static int sync_fsync(struct vop_fsync_args *);
4977 static int sync_inactive(struct vop_inactive_args *);
4978 static int sync_reclaim(struct vop_reclaim_args *);
4980 static struct vop_vector sync_vnodeops = {
4981 .vop_bypass = VOP_EOPNOTSUPP,
4982 .vop_close = sync_close, /* close */
4983 .vop_fsync = sync_fsync, /* fsync */
4984 .vop_inactive = sync_inactive, /* inactive */
4985 .vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4986 .vop_reclaim = sync_reclaim, /* reclaim */
4987 .vop_lock1 = vop_stdlock, /* lock */
4988 .vop_unlock = vop_stdunlock, /* unlock */
4989 .vop_islocked = vop_stdislocked, /* islocked */
4991 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4994 * Create a new filesystem syncer vnode for the specified mount point.
4997 vfs_allocate_syncvnode(struct mount *mp)
5001 static long start, incr, next;
5004 /* Allocate a new vnode */
5005 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
5007 panic("vfs_allocate_syncvnode: getnewvnode() failed");
5009 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5010 vp->v_vflag |= VV_FORCEINSMQ;
5011 error = insmntque1(vp, mp);
5013 panic("vfs_allocate_syncvnode: insmntque() failed");
5014 vp->v_vflag &= ~VV_FORCEINSMQ;
5015 vn_set_state(vp, VSTATE_CONSTRUCTED);
5018 * Place the vnode onto the syncer worklist. We attempt to
5019 * scatter them about on the list so that they will go off
5020 * at evenly distributed times even if all the filesystems
5021 * are mounted at once.
5024 if (next == 0 || next > syncer_maxdelay) {
5028 start = syncer_maxdelay / 2;
5029 incr = syncer_maxdelay;
5035 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
5036 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
5037 mtx_lock(&sync_mtx);
5039 if (mp->mnt_syncer == NULL) {
5040 mp->mnt_syncer = vp;
5043 mtx_unlock(&sync_mtx);
5046 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5053 vfs_deallocate_syncvnode(struct mount *mp)
5057 mtx_lock(&sync_mtx);
5058 vp = mp->mnt_syncer;
5060 mp->mnt_syncer = NULL;
5061 mtx_unlock(&sync_mtx);
5067 * Do a lazy sync of the filesystem.
5070 sync_fsync(struct vop_fsync_args *ap)
5072 struct vnode *syncvp = ap->a_vp;
5073 struct mount *mp = syncvp->v_mount;
5078 * We only need to do something if this is a lazy evaluation.
5080 if (ap->a_waitfor != MNT_LAZY)
5084 * Move ourselves to the back of the sync list.
5086 bo = &syncvp->v_bufobj;
5088 vn_syncer_add_to_worklist(bo, syncdelay);
5092 * Walk the list of vnodes pushing all that are dirty and
5093 * not already on the sync list.
5095 if (vfs_busy(mp, MBF_NOWAIT) != 0)
5098 save = curthread_pflags_set(TDP_SYNCIO);
5100 * The filesystem at hand may be idle with free vnodes stored in the
5101 * batch. Return them instead of letting them stay there indefinitely.
5103 vfs_periodic(mp, MNT_NOWAIT);
5104 error = VFS_SYNC(mp, MNT_LAZY);
5105 curthread_pflags_restore(save);
5106 vn_lock(syncvp, LK_EXCLUSIVE | LK_RETRY);
5112 * The syncer vnode is no referenced.
5115 sync_inactive(struct vop_inactive_args *ap)
5123 * The syncer vnode is no longer needed and is being decommissioned.
5125 * Modifications to the worklist must be protected by sync_mtx.
5128 sync_reclaim(struct vop_reclaim_args *ap)
5130 struct vnode *vp = ap->a_vp;
5135 mtx_lock(&sync_mtx);
5136 if (vp->v_mount->mnt_syncer == vp)
5137 vp->v_mount->mnt_syncer = NULL;
5138 if (bo->bo_flag & BO_ONWORKLST) {
5139 LIST_REMOVE(bo, bo_synclist);
5140 syncer_worklist_len--;
5142 bo->bo_flag &= ~BO_ONWORKLST;
5144 mtx_unlock(&sync_mtx);
5151 vn_need_pageq_flush(struct vnode *vp)
5153 struct vm_object *obj;
5156 return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5157 vm_object_mightbedirty(obj));
5161 * Check if vnode represents a disk device
5164 vn_isdisk_error(struct vnode *vp, int *errp)
5168 if (vp->v_type != VCHR) {
5174 if (vp->v_rdev == NULL)
5176 else if (vp->v_rdev->si_devsw == NULL)
5178 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5183 return (error == 0);
5187 vn_isdisk(struct vnode *vp)
5191 return (vn_isdisk_error(vp, &error));
5195 * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
5196 * the comment above cache_fplookup for details.
5199 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred)
5203 VFS_SMR_ASSERT_ENTERED();
5205 /* Check the owner. */
5206 if (cred->cr_uid == file_uid) {
5207 if (file_mode & S_IXUSR)
5212 /* Otherwise, check the groups (first match) */
5213 if (groupmember(file_gid, cred)) {
5214 if (file_mode & S_IXGRP)
5219 /* Otherwise, check everyone else. */
5220 if (file_mode & S_IXOTH)
5224 * Permission check failed, but it is possible denial will get overwritten
5225 * (e.g., when root is traversing through a 700 directory owned by someone
5228 * vaccess() calls priv_check_cred which in turn can descent into MAC
5229 * modules overriding this result. It's quite unclear what semantics
5230 * are allowed for them to operate, thus for safety we don't call them
5231 * from within the SMR section. This also means if any such modules
5232 * are present, we have to let the regular lookup decide.
5234 error = priv_check_cred_vfs_lookup_nomac(cred);
5240 * MAC modules present.
5251 * Common filesystem object access control check routine. Accepts a
5252 * vnode's type, "mode", uid and gid, requested access mode, and credentials.
5253 * Returns 0 on success, or an errno on failure.
5256 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5257 accmode_t accmode, struct ucred *cred)
5259 accmode_t dac_granted;
5260 accmode_t priv_granted;
5262 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5263 ("invalid bit in accmode"));
5264 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5265 ("VAPPEND without VWRITE"));
5268 * Look for a normal, non-privileged way to access the file/directory
5269 * as requested. If it exists, go with that.
5274 /* Check the owner. */
5275 if (cred->cr_uid == file_uid) {
5276 dac_granted |= VADMIN;
5277 if (file_mode & S_IXUSR)
5278 dac_granted |= VEXEC;
5279 if (file_mode & S_IRUSR)
5280 dac_granted |= VREAD;
5281 if (file_mode & S_IWUSR)
5282 dac_granted |= (VWRITE | VAPPEND);
5284 if ((accmode & dac_granted) == accmode)
5290 /* Otherwise, check the groups (first match) */
5291 if (groupmember(file_gid, cred)) {
5292 if (file_mode & S_IXGRP)
5293 dac_granted |= VEXEC;
5294 if (file_mode & S_IRGRP)
5295 dac_granted |= VREAD;
5296 if (file_mode & S_IWGRP)
5297 dac_granted |= (VWRITE | VAPPEND);
5299 if ((accmode & dac_granted) == accmode)
5305 /* Otherwise, check everyone else. */
5306 if (file_mode & S_IXOTH)
5307 dac_granted |= VEXEC;
5308 if (file_mode & S_IROTH)
5309 dac_granted |= VREAD;
5310 if (file_mode & S_IWOTH)
5311 dac_granted |= (VWRITE | VAPPEND);
5312 if ((accmode & dac_granted) == accmode)
5317 * Build a privilege mask to determine if the set of privileges
5318 * satisfies the requirements when combined with the granted mask
5319 * from above. For each privilege, if the privilege is required,
5320 * bitwise or the request type onto the priv_granted mask.
5326 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5327 * requests, instead of PRIV_VFS_EXEC.
5329 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5330 !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5331 priv_granted |= VEXEC;
5334 * Ensure that at least one execute bit is on. Otherwise,
5335 * a privileged user will always succeed, and we don't want
5336 * this to happen unless the file really is executable.
5338 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5339 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5340 !priv_check_cred(cred, PRIV_VFS_EXEC))
5341 priv_granted |= VEXEC;
5344 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5345 !priv_check_cred(cred, PRIV_VFS_READ))
5346 priv_granted |= VREAD;
5348 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5349 !priv_check_cred(cred, PRIV_VFS_WRITE))
5350 priv_granted |= (VWRITE | VAPPEND);
5352 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5353 !priv_check_cred(cred, PRIV_VFS_ADMIN))
5354 priv_granted |= VADMIN;
5356 if ((accmode & (priv_granted | dac_granted)) == accmode) {
5360 return ((accmode & VADMIN) ? EPERM : EACCES);
5364 * Credential check based on process requesting service, and per-attribute
5368 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5369 struct thread *td, accmode_t accmode)
5373 * Kernel-invoked always succeeds.
5379 * Do not allow privileged processes in jail to directly manipulate
5380 * system attributes.
5382 switch (attrnamespace) {
5383 case EXTATTR_NAMESPACE_SYSTEM:
5384 /* Potentially should be: return (EPERM); */
5385 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5386 case EXTATTR_NAMESPACE_USER:
5387 return (VOP_ACCESS(vp, accmode, cred, td));
5393 #ifdef DEBUG_VFS_LOCKS
5394 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */
5395 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5396 "Drop into debugger on lock violation");
5398 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */
5399 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5400 0, "Check for interlock across VOPs");
5402 int vfs_badlock_print = 1; /* Print lock violations. */
5403 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5404 0, "Print lock violations");
5406 int vfs_badlock_vnode = 1; /* Print vnode details on lock violations. */
5407 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5408 0, "Print vnode details on lock violations");
5411 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */
5412 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5413 &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5417 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5421 if (vfs_badlock_backtrace)
5424 if (vfs_badlock_vnode)
5425 vn_printf(vp, "vnode ");
5426 if (vfs_badlock_print)
5427 printf("%s: %p %s\n", str, (void *)vp, msg);
5428 if (vfs_badlock_ddb)
5429 kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5433 assert_vi_locked(struct vnode *vp, const char *str)
5436 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5437 vfs_badlock("interlock is not locked but should be", str, vp);
5441 assert_vi_unlocked(struct vnode *vp, const char *str)
5444 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5445 vfs_badlock("interlock is locked but should not be", str, vp);
5449 assert_vop_locked(struct vnode *vp, const char *str)
5453 if (KERNEL_PANICKED() || vp == NULL)
5456 locked = VOP_ISLOCKED(vp);
5457 if (locked == 0 || locked == LK_EXCLOTHER)
5458 vfs_badlock("is not locked but should be", str, vp);
5462 assert_vop_unlocked(struct vnode *vp, const char *str)
5464 if (KERNEL_PANICKED() || vp == NULL)
5467 if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5468 vfs_badlock("is locked but should not be", str, vp);
5472 assert_vop_elocked(struct vnode *vp, const char *str)
5474 if (KERNEL_PANICKED() || vp == NULL)
5477 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5478 vfs_badlock("is not exclusive locked but should be", str, vp);
5480 #endif /* DEBUG_VFS_LOCKS */
5483 vop_rename_fail(struct vop_rename_args *ap)
5486 if (ap->a_tvp != NULL)
5488 if (ap->a_tdvp == ap->a_tvp)
5497 vop_rename_pre(void *ap)
5499 struct vop_rename_args *a = ap;
5501 #ifdef DEBUG_VFS_LOCKS
5503 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5504 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5505 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5506 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5508 /* Check the source (from). */
5509 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5510 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5511 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5512 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5513 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5515 /* Check the target. */
5517 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5518 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5521 * It may be tempting to add vn_seqc_write_begin/end calls here and
5522 * in vop_rename_post but that's not going to work out since some
5523 * filesystems relookup vnodes mid-rename. This is probably a bug.
5525 * For now filesystems are expected to do the relevant calls after they
5526 * decide what vnodes to operate on.
5528 if (a->a_tdvp != a->a_fdvp)
5530 if (a->a_tvp != a->a_fvp)
5537 #ifdef DEBUG_VFS_LOCKS
5539 vop_fplookup_vexec_debugpre(void *ap __unused)
5542 VFS_SMR_ASSERT_ENTERED();
5546 vop_fplookup_vexec_debugpost(void *ap __unused, int rc __unused)
5549 VFS_SMR_ASSERT_ENTERED();
5553 vop_fplookup_symlink_debugpre(void *ap __unused)
5556 VFS_SMR_ASSERT_ENTERED();
5560 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused)
5563 VFS_SMR_ASSERT_ENTERED();
5567 vop_fsync_debugprepost(struct vnode *vp, const char *name)
5569 if (vp->v_type == VCHR)
5571 else if (MNT_EXTENDED_SHARED(vp->v_mount))
5572 ASSERT_VOP_LOCKED(vp, name);
5574 ASSERT_VOP_ELOCKED(vp, name);
5578 vop_fsync_debugpre(void *a)
5580 struct vop_fsync_args *ap;
5583 vop_fsync_debugprepost(ap->a_vp, "fsync");
5587 vop_fsync_debugpost(void *a, int rc __unused)
5589 struct vop_fsync_args *ap;
5592 vop_fsync_debugprepost(ap->a_vp, "fsync");
5596 vop_fdatasync_debugpre(void *a)
5598 struct vop_fdatasync_args *ap;
5601 vop_fsync_debugprepost(ap->a_vp, "fsync");
5605 vop_fdatasync_debugpost(void *a, int rc __unused)
5607 struct vop_fdatasync_args *ap;
5610 vop_fsync_debugprepost(ap->a_vp, "fsync");
5614 vop_strategy_debugpre(void *ap)
5616 struct vop_strategy_args *a;
5623 * Cluster ops lock their component buffers but not the IO container.
5625 if ((bp->b_flags & B_CLUSTER) != 0)
5628 if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5629 if (vfs_badlock_print)
5631 "VOP_STRATEGY: bp is not locked but should be\n");
5632 if (vfs_badlock_ddb)
5633 kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5638 vop_lock_debugpre(void *ap)
5640 struct vop_lock1_args *a = ap;
5642 if ((a->a_flags & LK_INTERLOCK) == 0)
5643 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5645 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5649 vop_lock_debugpost(void *ap, int rc)
5651 struct vop_lock1_args *a = ap;
5653 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5654 if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5655 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5659 vop_unlock_debugpre(void *ap)
5661 struct vop_unlock_args *a = ap;
5662 struct vnode *vp = a->a_vp;
5664 VNPASS(vn_get_state(vp) != VSTATE_UNINITIALIZED, vp);
5665 ASSERT_VOP_LOCKED(vp, "VOP_UNLOCK");
5669 vop_need_inactive_debugpre(void *ap)
5671 struct vop_need_inactive_args *a = ap;
5673 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5677 vop_need_inactive_debugpost(void *ap, int rc)
5679 struct vop_need_inactive_args *a = ap;
5681 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5686 vop_create_pre(void *ap)
5688 struct vop_create_args *a;
5693 vn_seqc_write_begin(dvp);
5697 vop_create_post(void *ap, int rc)
5699 struct vop_create_args *a;
5704 vn_seqc_write_end(dvp);
5706 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5710 vop_whiteout_pre(void *ap)
5712 struct vop_whiteout_args *a;
5717 vn_seqc_write_begin(dvp);
5721 vop_whiteout_post(void *ap, int rc)
5723 struct vop_whiteout_args *a;
5728 vn_seqc_write_end(dvp);
5732 vop_deleteextattr_pre(void *ap)
5734 struct vop_deleteextattr_args *a;
5739 vn_seqc_write_begin(vp);
5743 vop_deleteextattr_post(void *ap, int rc)
5745 struct vop_deleteextattr_args *a;
5750 vn_seqc_write_end(vp);
5752 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5756 vop_link_pre(void *ap)
5758 struct vop_link_args *a;
5759 struct vnode *vp, *tdvp;
5764 vn_seqc_write_begin(vp);
5765 vn_seqc_write_begin(tdvp);
5769 vop_link_post(void *ap, int rc)
5771 struct vop_link_args *a;
5772 struct vnode *vp, *tdvp;
5777 vn_seqc_write_end(vp);
5778 vn_seqc_write_end(tdvp);
5780 VFS_KNOTE_LOCKED(vp, NOTE_LINK);
5781 VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE);
5786 vop_mkdir_pre(void *ap)
5788 struct vop_mkdir_args *a;
5793 vn_seqc_write_begin(dvp);
5797 vop_mkdir_post(void *ap, int rc)
5799 struct vop_mkdir_args *a;
5804 vn_seqc_write_end(dvp);
5806 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5809 #ifdef DEBUG_VFS_LOCKS
5811 vop_mkdir_debugpost(void *ap, int rc)
5813 struct vop_mkdir_args *a;
5817 cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp);
5822 vop_mknod_pre(void *ap)
5824 struct vop_mknod_args *a;
5829 vn_seqc_write_begin(dvp);
5833 vop_mknod_post(void *ap, int rc)
5835 struct vop_mknod_args *a;
5840 vn_seqc_write_end(dvp);
5842 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5846 vop_reclaim_post(void *ap, int rc)
5848 struct vop_reclaim_args *a;
5853 ASSERT_VOP_IN_SEQC(vp);
5855 VFS_KNOTE_LOCKED(vp, NOTE_REVOKE);
5859 vop_remove_pre(void *ap)
5861 struct vop_remove_args *a;
5862 struct vnode *dvp, *vp;
5867 vn_seqc_write_begin(dvp);
5868 vn_seqc_write_begin(vp);
5872 vop_remove_post(void *ap, int rc)
5874 struct vop_remove_args *a;
5875 struct vnode *dvp, *vp;
5880 vn_seqc_write_end(dvp);
5881 vn_seqc_write_end(vp);
5883 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5884 VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5889 vop_rename_post(void *ap, int rc)
5891 struct vop_rename_args *a = ap;
5896 if (a->a_fdvp == a->a_tdvp) {
5897 if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5899 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5900 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5902 hint |= NOTE_EXTEND;
5903 if (a->a_fvp->v_type == VDIR)
5905 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5907 if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5908 a->a_tvp->v_type == VDIR)
5910 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5913 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5915 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5917 if (a->a_tdvp != a->a_fdvp)
5919 if (a->a_tvp != a->a_fvp)
5927 vop_rmdir_pre(void *ap)
5929 struct vop_rmdir_args *a;
5930 struct vnode *dvp, *vp;
5935 vn_seqc_write_begin(dvp);
5936 vn_seqc_write_begin(vp);
5940 vop_rmdir_post(void *ap, int rc)
5942 struct vop_rmdir_args *a;
5943 struct vnode *dvp, *vp;
5948 vn_seqc_write_end(dvp);
5949 vn_seqc_write_end(vp);
5951 vp->v_vflag |= VV_UNLINKED;
5952 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5953 VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5958 vop_setattr_pre(void *ap)
5960 struct vop_setattr_args *a;
5965 vn_seqc_write_begin(vp);
5969 vop_setattr_post(void *ap, int rc)
5971 struct vop_setattr_args *a;
5976 vn_seqc_write_end(vp);
5978 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5982 vop_setacl_pre(void *ap)
5984 struct vop_setacl_args *a;
5989 vn_seqc_write_begin(vp);
5993 vop_setacl_post(void *ap, int rc __unused)
5995 struct vop_setacl_args *a;
6000 vn_seqc_write_end(vp);
6004 vop_setextattr_pre(void *ap)
6006 struct vop_setextattr_args *a;
6011 vn_seqc_write_begin(vp);
6015 vop_setextattr_post(void *ap, int rc)
6017 struct vop_setextattr_args *a;
6022 vn_seqc_write_end(vp);
6024 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
6028 vop_symlink_pre(void *ap)
6030 struct vop_symlink_args *a;
6035 vn_seqc_write_begin(dvp);
6039 vop_symlink_post(void *ap, int rc)
6041 struct vop_symlink_args *a;
6046 vn_seqc_write_end(dvp);
6048 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6052 vop_open_post(void *ap, int rc)
6054 struct vop_open_args *a = ap;
6057 VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
6061 vop_close_post(void *ap, int rc)
6063 struct vop_close_args *a = ap;
6065 if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
6066 !VN_IS_DOOMED(a->a_vp))) {
6067 VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
6068 NOTE_CLOSE_WRITE : NOTE_CLOSE);
6073 vop_read_post(void *ap, int rc)
6075 struct vop_read_args *a = ap;
6078 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6082 vop_read_pgcache_post(void *ap, int rc)
6084 struct vop_read_pgcache_args *a = ap;
6087 VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ);
6091 vop_readdir_post(void *ap, int rc)
6093 struct vop_readdir_args *a = ap;
6096 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6099 static struct knlist fs_knlist;
6102 vfs_event_init(void *arg)
6104 knlist_init_mtx(&fs_knlist, NULL);
6106 /* XXX - correct order? */
6107 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
6110 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
6113 KNOTE_UNLOCKED(&fs_knlist, event);
6116 static int filt_fsattach(struct knote *kn);
6117 static void filt_fsdetach(struct knote *kn);
6118 static int filt_fsevent(struct knote *kn, long hint);
6120 struct filterops fs_filtops = {
6122 .f_attach = filt_fsattach,
6123 .f_detach = filt_fsdetach,
6124 .f_event = filt_fsevent
6128 filt_fsattach(struct knote *kn)
6131 kn->kn_flags |= EV_CLEAR;
6132 knlist_add(&fs_knlist, kn, 0);
6137 filt_fsdetach(struct knote *kn)
6140 knlist_remove(&fs_knlist, kn, 0);
6144 filt_fsevent(struct knote *kn, long hint)
6147 kn->kn_fflags |= kn->kn_sfflags & hint;
6149 return (kn->kn_fflags != 0);
6153 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
6159 error = SYSCTL_IN(req, &vc, sizeof(vc));
6162 if (vc.vc_vers != VFS_CTL_VERS1)
6164 mp = vfs_getvfs(&vc.vc_fsid);
6167 /* ensure that a specific sysctl goes to the right filesystem. */
6168 if (strcmp(vc.vc_fstypename, "*") != 0 &&
6169 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
6173 VCTLTOREQ(&vc, req);
6174 error = VFS_SYSCTL(mp, vc.vc_op, req);
6179 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
6180 NULL, 0, sysctl_vfs_ctl, "",
6184 * Function to initialize a va_filerev field sensibly.
6185 * XXX: Wouldn't a random number make a lot more sense ??
6188 init_va_filerev(void)
6193 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
6196 static int filt_vfsread(struct knote *kn, long hint);
6197 static int filt_vfswrite(struct knote *kn, long hint);
6198 static int filt_vfsvnode(struct knote *kn, long hint);
6199 static void filt_vfsdetach(struct knote *kn);
6200 static struct filterops vfsread_filtops = {
6202 .f_detach = filt_vfsdetach,
6203 .f_event = filt_vfsread
6205 static struct filterops vfswrite_filtops = {
6207 .f_detach = filt_vfsdetach,
6208 .f_event = filt_vfswrite
6210 static struct filterops vfsvnode_filtops = {
6212 .f_detach = filt_vfsdetach,
6213 .f_event = filt_vfsvnode
6217 vfs_knllock(void *arg)
6219 struct vnode *vp = arg;
6221 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
6225 vfs_knlunlock(void *arg)
6227 struct vnode *vp = arg;
6233 vfs_knl_assert_lock(void *arg, int what)
6235 #ifdef DEBUG_VFS_LOCKS
6236 struct vnode *vp = arg;
6238 if (what == LA_LOCKED)
6239 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
6241 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
6246 vfs_kqfilter(struct vop_kqfilter_args *ap)
6248 struct vnode *vp = ap->a_vp;
6249 struct knote *kn = ap->a_kn;
6252 KASSERT(vp->v_type != VFIFO || (kn->kn_filter != EVFILT_READ &&
6253 kn->kn_filter != EVFILT_WRITE),
6254 ("READ/WRITE filter on a FIFO leaked through"));
6255 switch (kn->kn_filter) {
6257 kn->kn_fop = &vfsread_filtops;
6260 kn->kn_fop = &vfswrite_filtops;
6263 kn->kn_fop = &vfsvnode_filtops;
6269 kn->kn_hook = (caddr_t)vp;
6272 if (vp->v_pollinfo == NULL)
6274 knl = &vp->v_pollinfo->vpi_selinfo.si_note;
6276 knlist_add(knl, kn, 0);
6282 * Detach knote from vnode
6285 filt_vfsdetach(struct knote *kn)
6287 struct vnode *vp = (struct vnode *)kn->kn_hook;
6289 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
6290 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
6296 filt_vfsread(struct knote *kn, long hint)
6298 struct vnode *vp = (struct vnode *)kn->kn_hook;
6303 * filesystem is gone, so set the EOF flag and schedule
6304 * the knote for deletion.
6306 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6308 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6313 if (vn_getsize_locked(vp, &size, curthread->td_ucred) != 0)
6317 kn->kn_data = size - kn->kn_fp->f_offset;
6318 res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
6325 filt_vfswrite(struct knote *kn, long hint)
6327 struct vnode *vp = (struct vnode *)kn->kn_hook;
6332 * filesystem is gone, so set the EOF flag and schedule
6333 * the knote for deletion.
6335 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
6336 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6344 filt_vfsvnode(struct knote *kn, long hint)
6346 struct vnode *vp = (struct vnode *)kn->kn_hook;
6350 if (kn->kn_sfflags & hint)
6351 kn->kn_fflags |= hint;
6352 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6353 kn->kn_flags |= EV_EOF;
6357 res = (kn->kn_fflags != 0);
6363 * Returns whether the directory is empty or not.
6364 * If it is empty, the return value is 0; otherwise
6365 * the return value is an error value (which may
6369 vfs_emptydir(struct vnode *vp)
6373 struct dirent *dirent, *dp, *endp;
6379 ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
6380 VNPASS(vp->v_type == VDIR, vp);
6382 dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
6383 iov.iov_base = dirent;
6384 iov.iov_len = sizeof(struct dirent);
6389 uio.uio_resid = sizeof(struct dirent);
6390 uio.uio_segflg = UIO_SYSSPACE;
6391 uio.uio_rw = UIO_READ;
6392 uio.uio_td = curthread;
6394 while (eof == 0 && error == 0) {
6395 error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
6399 endp = (void *)((uint8_t *)dirent +
6400 sizeof(struct dirent) - uio.uio_resid);
6401 for (dp = dirent; dp < endp;
6402 dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
6403 if (dp->d_type == DT_WHT)
6405 if (dp->d_namlen == 0)
6407 if (dp->d_type != DT_DIR &&
6408 dp->d_type != DT_UNKNOWN) {
6412 if (dp->d_namlen > 2) {
6416 if (dp->d_namlen == 1 &&
6417 dp->d_name[0] != '.') {
6421 if (dp->d_namlen == 2 &&
6422 dp->d_name[1] != '.') {
6426 uio.uio_resid = sizeof(struct dirent);
6429 free(dirent, M_TEMP);
6434 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6438 if (dp->d_reclen > ap->a_uio->uio_resid)
6439 return (ENAMETOOLONG);
6440 error = uiomove(dp, dp->d_reclen, ap->a_uio);
6442 if (ap->a_ncookies != NULL) {
6443 if (ap->a_cookies != NULL)
6444 free(ap->a_cookies, M_TEMP);
6445 ap->a_cookies = NULL;
6446 *ap->a_ncookies = 0;
6450 if (ap->a_ncookies == NULL)
6453 KASSERT(ap->a_cookies,
6454 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6456 *ap->a_cookies = realloc(*ap->a_cookies,
6457 (*ap->a_ncookies + 1) * sizeof(uint64_t), M_TEMP, M_WAITOK | M_ZERO);
6458 (*ap->a_cookies)[*ap->a_ncookies] = off;
6459 *ap->a_ncookies += 1;
6464 * The purpose of this routine is to remove granularity from accmode_t,
6465 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6466 * VADMIN and VAPPEND.
6468 * If it returns 0, the caller is supposed to continue with the usual
6469 * access checks using 'accmode' as modified by this routine. If it
6470 * returns nonzero value, the caller is supposed to return that value
6473 * Note that after this routine runs, accmode may be zero.
6476 vfs_unixify_accmode(accmode_t *accmode)
6479 * There is no way to specify explicit "deny" rule using
6480 * file mode or POSIX.1e ACLs.
6482 if (*accmode & VEXPLICIT_DENY) {
6488 * None of these can be translated into usual access bits.
6489 * Also, the common case for NFSv4 ACLs is to not contain
6490 * either of these bits. Caller should check for VWRITE
6491 * on the containing directory instead.
6493 if (*accmode & (VDELETE_CHILD | VDELETE))
6496 if (*accmode & VADMIN_PERMS) {
6497 *accmode &= ~VADMIN_PERMS;
6502 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6503 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6505 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6511 * Clear out a doomed vnode (if any) and replace it with a new one as long
6512 * as the fs is not being unmounted. Return the root vnode to the caller.
6514 static int __noinline
6515 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6521 if (mp->mnt_rootvnode != NULL) {
6523 vp = mp->mnt_rootvnode;
6525 if (!VN_IS_DOOMED(vp)) {
6528 error = vn_lock(vp, flags);
6537 * Clear the old one.
6539 mp->mnt_rootvnode = NULL;
6543 vfs_op_barrier_wait(mp);
6547 error = VFS_CACHEDROOT(mp, flags, vpp);
6550 if (mp->mnt_vfs_ops == 0) {
6552 if (mp->mnt_vfs_ops != 0) {
6556 if (mp->mnt_rootvnode == NULL) {
6558 mp->mnt_rootvnode = *vpp;
6560 if (mp->mnt_rootvnode != *vpp) {
6561 if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6562 panic("%s: mismatch between vnode returned "
6563 " by VFS_CACHEDROOT and the one cached "
6565 __func__, *vpp, mp->mnt_rootvnode);
6575 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6577 struct mount_pcpu *mpcpu;
6581 if (!vfs_op_thread_enter(mp, mpcpu))
6582 return (vfs_cache_root_fallback(mp, flags, vpp));
6583 vp = atomic_load_ptr(&mp->mnt_rootvnode);
6584 if (vp == NULL || VN_IS_DOOMED(vp)) {
6585 vfs_op_thread_exit(mp, mpcpu);
6586 return (vfs_cache_root_fallback(mp, flags, vpp));
6589 vfs_op_thread_exit(mp, mpcpu);
6590 error = vn_lock(vp, flags);
6593 return (vfs_cache_root_fallback(mp, flags, vpp));
6600 vfs_cache_root_clear(struct mount *mp)
6605 * ops > 0 guarantees there is nobody who can see this vnode
6607 MPASS(mp->mnt_vfs_ops > 0);
6608 vp = mp->mnt_rootvnode;
6610 vn_seqc_write_begin(vp);
6611 mp->mnt_rootvnode = NULL;
6616 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6619 MPASS(mp->mnt_vfs_ops > 0);
6621 mp->mnt_rootvnode = vp;
6625 * These are helper functions for filesystems to traverse all
6626 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6628 * This interface replaces MNT_VNODE_FOREACH.
6632 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6638 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6639 for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6640 vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6641 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6642 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6645 if (VN_IS_DOOMED(vp)) {
6652 __mnt_vnode_markerfree_all(mvp, mp);
6653 /* MNT_IUNLOCK(mp); -- done in above function */
6654 mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6657 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6658 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6664 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6668 *mvp = vn_alloc_marker(mp);
6672 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6673 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6674 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6677 if (VN_IS_DOOMED(vp)) {
6686 vn_free_marker(*mvp);
6690 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6696 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6704 mtx_assert(MNT_MTX(mp), MA_OWNED);
6706 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6707 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6710 vn_free_marker(*mvp);
6715 * These are helper functions for filesystems to traverse their
6716 * lazy vnodes. See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6719 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6722 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6727 vn_free_marker(*mvp);
6732 * Relock the mp mount vnode list lock with the vp vnode interlock in the
6733 * conventional lock order during mnt_vnode_next_lazy iteration.
6735 * On entry, the mount vnode list lock is held and the vnode interlock is not.
6736 * The list lock is dropped and reacquired. On success, both locks are held.
6737 * On failure, the mount vnode list lock is held but the vnode interlock is
6738 * not, and the procedure may have yielded.
6741 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6745 VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6746 TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6747 ("%s: bad marker", __func__));
6748 VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6749 ("%s: inappropriate vnode", __func__));
6750 ASSERT_VI_UNLOCKED(vp, __func__);
6751 mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6753 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6754 TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6757 * Note we may be racing against vdrop which transitioned the hold
6758 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6759 * if we are the only user after we get the interlock we will just
6763 mtx_unlock(&mp->mnt_listmtx);
6765 if (VN_IS_DOOMED(vp)) {
6766 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6769 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6771 * There is nothing to do if we are the last user.
6773 if (!refcount_release_if_not_last(&vp->v_holdcnt))
6775 mtx_lock(&mp->mnt_listmtx);
6780 mtx_lock(&mp->mnt_listmtx);
6784 static struct vnode *
6785 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6790 mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6791 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6793 vp = TAILQ_NEXT(*mvp, v_lazylist);
6794 while (vp != NULL) {
6795 if (vp->v_type == VMARKER) {
6796 vp = TAILQ_NEXT(vp, v_lazylist);
6800 * See if we want to process the vnode. Note we may encounter a
6801 * long string of vnodes we don't care about and hog the list
6802 * as a result. Check for it and requeue the marker.
6804 VNPASS(!VN_IS_DOOMED(vp), vp);
6805 if (!cb(vp, cbarg)) {
6806 if (!should_yield()) {
6807 vp = TAILQ_NEXT(vp, v_lazylist);
6810 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6812 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6814 mtx_unlock(&mp->mnt_listmtx);
6815 kern_yield(PRI_USER);
6816 mtx_lock(&mp->mnt_listmtx);
6820 * Try-lock because this is the wrong lock order.
6822 if (!VI_TRYLOCK(vp) &&
6823 !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6825 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6826 KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6827 ("alien vnode on the lazy list %p %p", vp, mp));
6828 VNPASS(vp->v_mount == mp, vp);
6829 VNPASS(!VN_IS_DOOMED(vp), vp);
6832 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6834 /* Check if we are done */
6836 mtx_unlock(&mp->mnt_listmtx);
6837 mnt_vnode_markerfree_lazy(mvp, mp);
6840 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6841 mtx_unlock(&mp->mnt_listmtx);
6842 ASSERT_VI_LOCKED(vp, "lazy iter");
6847 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6852 mtx_lock(&mp->mnt_listmtx);
6853 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6857 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6862 if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6865 *mvp = vn_alloc_marker(mp);
6870 mtx_lock(&mp->mnt_listmtx);
6871 vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6873 mtx_unlock(&mp->mnt_listmtx);
6874 mnt_vnode_markerfree_lazy(mvp, mp);
6877 TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6878 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6882 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6888 mtx_lock(&mp->mnt_listmtx);
6889 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6890 mtx_unlock(&mp->mnt_listmtx);
6891 mnt_vnode_markerfree_lazy(mvp, mp);
6895 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6898 if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6899 cnp->cn_flags &= ~NOEXECCHECK;
6903 return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, curthread));
6907 * Do not use this variant unless you have means other than the hold count
6908 * to prevent the vnode from getting freed.
6911 vn_seqc_write_begin_locked(struct vnode *vp)
6914 ASSERT_VI_LOCKED(vp, __func__);
6915 VNPASS(vp->v_holdcnt > 0, vp);
6916 VNPASS(vp->v_seqc_users >= 0, vp);
6918 if (vp->v_seqc_users == 1)
6919 seqc_sleepable_write_begin(&vp->v_seqc);
6923 vn_seqc_write_begin(struct vnode *vp)
6927 vn_seqc_write_begin_locked(vp);
6932 vn_seqc_write_end_locked(struct vnode *vp)
6935 ASSERT_VI_LOCKED(vp, __func__);
6936 VNPASS(vp->v_seqc_users > 0, vp);
6938 if (vp->v_seqc_users == 0)
6939 seqc_sleepable_write_end(&vp->v_seqc);
6943 vn_seqc_write_end(struct vnode *vp)
6947 vn_seqc_write_end_locked(vp);
6952 * Special case handling for allocating and freeing vnodes.
6954 * The counter remains unchanged on free so that a doomed vnode will
6955 * keep testing as in modify as long as it is accessible with SMR.
6958 vn_seqc_init(struct vnode *vp)
6962 vp->v_seqc_users = 0;
6966 vn_seqc_write_end_free(struct vnode *vp)
6969 VNPASS(seqc_in_modify(vp->v_seqc), vp);
6970 VNPASS(vp->v_seqc_users == 1, vp);
6974 vn_irflag_set_locked(struct vnode *vp, short toset)
6978 ASSERT_VI_LOCKED(vp, __func__);
6979 flags = vn_irflag_read(vp);
6980 VNASSERT((flags & toset) == 0, vp,
6981 ("%s: some of the passed flags already set (have %d, passed %d)\n",
6982 __func__, flags, toset));
6983 atomic_store_short(&vp->v_irflag, flags | toset);
6987 vn_irflag_set(struct vnode *vp, short toset)
6991 vn_irflag_set_locked(vp, toset);
6996 vn_irflag_set_cond_locked(struct vnode *vp, short toset)
7000 ASSERT_VI_LOCKED(vp, __func__);
7001 flags = vn_irflag_read(vp);
7002 atomic_store_short(&vp->v_irflag, flags | toset);
7006 vn_irflag_set_cond(struct vnode *vp, short toset)
7010 vn_irflag_set_cond_locked(vp, toset);
7015 vn_irflag_unset_locked(struct vnode *vp, short tounset)
7019 ASSERT_VI_LOCKED(vp, __func__);
7020 flags = vn_irflag_read(vp);
7021 VNASSERT((flags & tounset) == tounset, vp,
7022 ("%s: some of the passed flags not set (have %d, passed %d)\n",
7023 __func__, flags, tounset));
7024 atomic_store_short(&vp->v_irflag, flags & ~tounset);
7028 vn_irflag_unset(struct vnode *vp, short tounset)
7032 vn_irflag_unset_locked(vp, tounset);
7037 vn_getsize_locked(struct vnode *vp, off_t *size, struct ucred *cred)
7042 ASSERT_VOP_LOCKED(vp, __func__);
7043 error = VOP_GETATTR(vp, &vattr, cred);
7044 if (__predict_true(error == 0)) {
7045 if (vattr.va_size <= OFF_MAX)
7046 *size = vattr.va_size;
7054 vn_getsize(struct vnode *vp, off_t *size, struct ucred *cred)
7058 VOP_LOCK(vp, LK_SHARED);
7059 error = vn_getsize_locked(vp, size, cred);
7066 vn_set_state_validate(struct vnode *vp, enum vstate state)
7069 switch (vp->v_state) {
7070 case VSTATE_UNINITIALIZED:
7072 case VSTATE_CONSTRUCTED:
7073 case VSTATE_DESTROYING:
7079 case VSTATE_CONSTRUCTED:
7080 ASSERT_VOP_ELOCKED(vp, __func__);
7082 case VSTATE_DESTROYING:
7088 case VSTATE_DESTROYING:
7089 ASSERT_VOP_ELOCKED(vp, __func__);
7099 case VSTATE_UNINITIALIZED:
7107 vn_printf(vp, "invalid state transition %d -> %d\n", vp->v_state, state);
7108 panic("invalid state transition %d -> %d\n", vp->v_state, state);