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/lockf.h>
70 #include <sys/malloc.h>
71 #include <sys/mount.h>
72 #include <sys/namei.h>
73 #include <sys/pctrie.h>
75 #include <sys/reboot.h>
76 #include <sys/refcount.h>
77 #include <sys/rwlock.h>
78 #include <sys/sched.h>
79 #include <sys/sleepqueue.h>
83 #include <sys/sysctl.h>
84 #include <sys/syslog.h>
85 #include <sys/vmmeter.h>
86 #include <sys/vnode.h>
87 #include <sys/watchdog.h>
89 #include <machine/stdarg.h>
91 #include <security/mac/mac_framework.h>
94 #include <vm/vm_object.h>
95 #include <vm/vm_extern.h>
97 #include <vm/vm_map.h>
98 #include <vm/vm_page.h>
99 #include <vm/vm_kern.h>
106 static void delmntque(struct vnode *vp);
107 static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
108 int slpflag, int slptimeo);
109 static void syncer_shutdown(void *arg, int howto);
110 static int vtryrecycle(struct vnode *vp);
111 static void v_init_counters(struct vnode *);
112 static void vn_seqc_init(struct vnode *);
113 static void vn_seqc_write_end_free(struct vnode *vp);
114 static void vgonel(struct vnode *);
115 static bool vhold_recycle_free(struct vnode *);
116 static void vfs_knllock(void *arg);
117 static void vfs_knlunlock(void *arg);
118 static void vfs_knl_assert_lock(void *arg, int what);
119 static void destroy_vpollinfo(struct vpollinfo *vi);
120 static int v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
121 daddr_t startlbn, daddr_t endlbn);
122 static void vnlru_recalc(void);
125 * Number of vnodes in existence. Increased whenever getnewvnode()
126 * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
128 static u_long __exclusive_cache_line numvnodes;
130 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
131 "Number of vnodes in existence");
133 static counter_u64_t vnodes_created;
134 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
135 "Number of vnodes created by getnewvnode");
138 * Conversion tables for conversion from vnode types to inode formats
141 enum vtype iftovt_tab[16] = {
142 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
143 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
145 int vttoif_tab[10] = {
146 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
147 S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
151 * List of allocates vnodes in the system.
153 static TAILQ_HEAD(freelst, vnode) vnode_list;
154 static struct vnode *vnode_list_free_marker;
155 static struct vnode *vnode_list_reclaim_marker;
158 * "Free" vnode target. Free vnodes are rarely completely free, but are
159 * just ones that are cheap to recycle. Usually they are for files which
160 * have been stat'd but not read; these usually have inode and namecache
161 * data attached to them. This target is the preferred minimum size of a
162 * sub-cache consisting mostly of such files. The system balances the size
163 * of this sub-cache with its complement to try to prevent either from
164 * thrashing while the other is relatively inactive. The targets express
165 * a preference for the best balance.
167 * "Above" this target there are 2 further targets (watermarks) related
168 * to recyling of free vnodes. In the best-operating case, the cache is
169 * exactly full, the free list has size between vlowat and vhiwat above the
170 * free target, and recycling from it and normal use maintains this state.
171 * Sometimes the free list is below vlowat or even empty, but this state
172 * is even better for immediate use provided the cache is not full.
173 * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
174 * ones) to reach one of these states. The watermarks are currently hard-
175 * coded as 4% and 9% of the available space higher. These and the default
176 * of 25% for wantfreevnodes are too large if the memory size is large.
177 * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
178 * whenever vnlru_proc() becomes active.
180 static long wantfreevnodes;
181 static long __exclusive_cache_line freevnodes;
182 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
183 &freevnodes, 0, "Number of \"free\" vnodes");
184 static long freevnodes_old;
186 static counter_u64_t recycles_count;
187 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
188 "Number of vnodes recycled to meet vnode cache targets");
190 static counter_u64_t recycles_free_count;
191 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count,
192 "Number of free vnodes recycled to meet vnode cache targets");
194 static counter_u64_t deferred_inact;
195 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact,
196 "Number of times inactive processing was deferred");
198 /* To keep more than one thread at a time from running vfs_getnewfsid */
199 static struct mtx mntid_mtx;
202 * Lock for any access to the following:
207 static struct mtx __exclusive_cache_line vnode_list_mtx;
209 /* Publicly exported FS */
210 struct nfs_public nfs_pub;
212 static uma_zone_t buf_trie_zone;
213 static smr_t buf_trie_smr;
215 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
216 static uma_zone_t vnode_zone;
217 MALLOC_DEFINE(M_VNODEPOLL, "VN POLL", "vnode poll");
219 __read_frequently smr_t vfs_smr;
222 * The workitem queue.
224 * It is useful to delay writes of file data and filesystem metadata
225 * for tens of seconds so that quickly created and deleted files need
226 * not waste disk bandwidth being created and removed. To realize this,
227 * we append vnodes to a "workitem" queue. When running with a soft
228 * updates implementation, most pending metadata dependencies should
229 * not wait for more than a few seconds. Thus, mounted on block devices
230 * are delayed only about a half the time that file data is delayed.
231 * Similarly, directory updates are more critical, so are only delayed
232 * about a third the time that file data is delayed. Thus, there are
233 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
234 * one each second (driven off the filesystem syncer process). The
235 * syncer_delayno variable indicates the next queue that is to be processed.
236 * Items that need to be processed soon are placed in this queue:
238 * syncer_workitem_pending[syncer_delayno]
240 * A delay of fifteen seconds is done by placing the request fifteen
241 * entries later in the queue:
243 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
246 static int syncer_delayno;
247 static long syncer_mask;
248 LIST_HEAD(synclist, bufobj);
249 static struct synclist *syncer_workitem_pending;
251 * The sync_mtx protects:
256 * syncer_workitem_pending
257 * syncer_worklist_len
260 static struct mtx sync_mtx;
261 static struct cv sync_wakeup;
263 #define SYNCER_MAXDELAY 32
264 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */
265 static int syncdelay = 30; /* max time to delay syncing data */
266 static int filedelay = 30; /* time to delay syncing files */
267 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
268 "Time to delay syncing files (in seconds)");
269 static int dirdelay = 29; /* time to delay syncing directories */
270 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
271 "Time to delay syncing directories (in seconds)");
272 static int metadelay = 28; /* time to delay syncing metadata */
273 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
274 "Time to delay syncing metadata (in seconds)");
275 static int rushjob; /* number of slots to run ASAP */
276 static int stat_rush_requests; /* number of times I/O speeded up */
277 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
278 "Number of times I/O speeded up (rush requests)");
280 #define VDBATCH_SIZE 8
285 struct vnode *tab[VDBATCH_SIZE];
287 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
289 static void vdbatch_dequeue(struct vnode *vp);
292 * When shutting down the syncer, run it at four times normal speed.
294 #define SYNCER_SHUTDOWN_SPEEDUP 4
295 static int sync_vnode_count;
296 static int syncer_worklist_len;
297 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
300 /* Target for maximum number of vnodes. */
301 u_long desiredvnodes;
302 static u_long gapvnodes; /* gap between wanted and desired */
303 static u_long vhiwat; /* enough extras after expansion */
304 static u_long vlowat; /* minimal extras before expansion */
305 static u_long vstir; /* nonzero to stir non-free vnodes */
306 static volatile int vsmalltrigger = 8; /* pref to keep if > this many pages */
308 static u_long vnlru_read_freevnodes(void);
311 * Note that no attempt is made to sanitize these parameters.
314 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
320 error = sysctl_handle_long(oidp, &val, 0, req);
321 if (error != 0 || req->newptr == NULL)
324 if (val == desiredvnodes)
326 mtx_lock(&vnode_list_mtx);
328 wantfreevnodes = desiredvnodes / 4;
330 mtx_unlock(&vnode_list_mtx);
332 * XXX There is no protection against multiple threads changing
333 * desiredvnodes at the same time. Locking above only helps vnlru and
336 vfs_hash_changesize(desiredvnodes);
337 cache_changesize(desiredvnodes);
341 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
342 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
343 "LU", "Target for maximum number of vnodes");
346 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
351 val = wantfreevnodes;
352 error = sysctl_handle_long(oidp, &val, 0, req);
353 if (error != 0 || req->newptr == NULL)
356 if (val == wantfreevnodes)
358 mtx_lock(&vnode_list_mtx);
359 wantfreevnodes = val;
361 mtx_unlock(&vnode_list_mtx);
365 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
366 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
367 "LU", "Target for minimum number of \"free\" vnodes");
369 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
370 &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
371 static int vnlru_nowhere;
372 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
373 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
376 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
381 unsigned long ndflags;
384 if (req->newptr == NULL)
386 if (req->newlen >= PATH_MAX)
389 buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
390 error = SYSCTL_IN(req, buf, req->newlen);
394 buf[req->newlen] = '\0';
396 ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | SAVENAME;
397 NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf);
398 if ((error = namei(&nd)) != 0)
402 if (VN_IS_DOOMED(vp)) {
404 * This vnode is being recycled. Return != 0 to let the caller
405 * know that the sysctl had no effect. Return EAGAIN because a
406 * subsequent call will likely succeed (since namei will create
407 * a new vnode if necessary)
413 counter_u64_add(recycles_count, 1);
423 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
425 struct thread *td = curthread;
431 if (req->newptr == NULL)
434 error = sysctl_handle_int(oidp, &fd, 0, req);
437 error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
442 error = vn_lock(vp, LK_EXCLUSIVE);
446 counter_u64_add(recycles_count, 1);
454 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
455 CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
456 sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
457 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
458 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
459 sysctl_ftry_reclaim_vnode, "I",
460 "Try to reclaim a vnode by its file descriptor");
462 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
466 * Support for the bufobj clean & dirty pctrie.
469 buf_trie_alloc(struct pctrie *ptree)
471 return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT));
475 buf_trie_free(struct pctrie *ptree, void *node)
477 uma_zfree_smr(buf_trie_zone, node);
479 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free,
483 * Initialize the vnode management data structures.
485 * Reevaluate the following cap on the number of vnodes after the physical
486 * memory size exceeds 512GB. In the limit, as the physical memory size
487 * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
489 #ifndef MAXVNODES_MAX
490 #define MAXVNODES_MAX (512UL * 1024 * 1024 / 64) /* 8M */
493 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
495 static struct vnode *
496 vn_alloc_marker(struct mount *mp)
500 vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
501 vp->v_type = VMARKER;
508 vn_free_marker(struct vnode *vp)
511 MPASS(vp->v_type == VMARKER);
512 free(vp, M_VNODE_MARKER);
517 vnode_ctor(void *mem, int size, void *arg __unused, int flags __unused)
519 kasan_mark(mem, size, roundup2(size, UMA_ALIGN_PTR + 1), 0);
524 vnode_dtor(void *mem, int size, void *arg __unused)
526 size_t end1, end2, off1, off2;
528 _Static_assert(offsetof(struct vnode, v_vnodelist) <
529 offsetof(struct vnode, v_dbatchcpu),
530 "KASAN marks require updating");
532 off1 = offsetof(struct vnode, v_vnodelist);
533 off2 = offsetof(struct vnode, v_dbatchcpu);
534 end1 = off1 + sizeof(((struct vnode *)NULL)->v_vnodelist);
535 end2 = off2 + sizeof(((struct vnode *)NULL)->v_dbatchcpu);
538 * Access to the v_vnodelist and v_dbatchcpu fields are permitted even
539 * after the vnode has been freed. Try to get some KASAN coverage by
540 * marking everything except those two fields as invalid. Because
541 * KASAN's tracking is not byte-granular, any preceding fields sharing
542 * the same 8-byte aligned word must also be marked valid.
545 /* Handle the area from the start until v_vnodelist... */
546 off1 = rounddown2(off1, KASAN_SHADOW_SCALE);
547 kasan_mark(mem, off1, off1, KASAN_UMA_FREED);
549 /* ... then the area between v_vnodelist and v_dbatchcpu ... */
550 off1 = roundup2(end1, KASAN_SHADOW_SCALE);
551 off2 = rounddown2(off2, KASAN_SHADOW_SCALE);
553 kasan_mark((void *)((char *)mem + off1), off2 - off1,
554 off2 - off1, KASAN_UMA_FREED);
556 /* ... and finally the area from v_dbatchcpu to the end. */
557 off2 = roundup2(end2, KASAN_SHADOW_SCALE);
558 kasan_mark((void *)((char *)mem + off2), size - off2, size - off2,
564 * Initialize a vnode as it first enters the zone.
567 vnode_init(void *mem, int size, int flags)
576 vp->v_vnlock = &vp->v_lock;
577 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
579 * By default, don't allow shared locks unless filesystems opt-in.
581 lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
582 LK_NOSHARE | LK_IS_VNODE);
586 bufobj_init(&vp->v_bufobj, vp);
588 * Initialize namecache.
590 cache_vnode_init(vp);
592 * Initialize rangelocks.
594 rangelock_init(&vp->v_rl);
596 vp->v_dbatchcpu = NOCPU;
599 * Check vhold_recycle_free for an explanation.
601 vp->v_holdcnt = VHOLD_NO_SMR;
603 mtx_lock(&vnode_list_mtx);
604 TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
605 mtx_unlock(&vnode_list_mtx);
610 * Free a vnode when it is cleared from the zone.
613 vnode_fini(void *mem, int size)
620 mtx_lock(&vnode_list_mtx);
621 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
622 mtx_unlock(&vnode_list_mtx);
623 rangelock_destroy(&vp->v_rl);
624 lockdestroy(vp->v_vnlock);
625 mtx_destroy(&vp->v_interlock);
627 rw_destroy(BO_LOCKPTR(bo));
629 kasan_mark(mem, size, size, 0);
633 * Provide the size of NFS nclnode and NFS fh for calculation of the
634 * vnode memory consumption. The size is specified directly to
635 * eliminate dependency on NFS-private header.
637 * Other filesystems may use bigger or smaller (like UFS and ZFS)
638 * private inode data, but the NFS-based estimation is ample enough.
639 * Still, we care about differences in the size between 64- and 32-bit
642 * Namecache structure size is heuristically
643 * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
646 #define NFS_NCLNODE_SZ (528 + 64)
649 #define NFS_NCLNODE_SZ (360 + 32)
654 vntblinit(void *dummy __unused)
659 int cpu, physvnodes, virtvnodes;
663 * Desiredvnodes is a function of the physical memory size and the
664 * kernel's heap size. Generally speaking, it scales with the
665 * physical memory size. The ratio of desiredvnodes to the physical
666 * memory size is 1:16 until desiredvnodes exceeds 98,304.
668 * marginal ratio of desiredvnodes to the physical memory size is
669 * 1:64. However, desiredvnodes is limited by the kernel's heap
670 * size. The memory required by desiredvnodes vnodes and vm objects
671 * must not exceed 1/10th of the kernel's heap size.
673 physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
674 3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
675 virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
676 sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
677 desiredvnodes = min(physvnodes, virtvnodes);
678 if (desiredvnodes > MAXVNODES_MAX) {
680 printf("Reducing kern.maxvnodes %lu -> %lu\n",
681 desiredvnodes, MAXVNODES_MAX);
682 desiredvnodes = MAXVNODES_MAX;
684 wantfreevnodes = desiredvnodes / 4;
685 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
686 TAILQ_INIT(&vnode_list);
687 mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
689 * The lock is taken to appease WITNESS.
691 mtx_lock(&vnode_list_mtx);
693 mtx_unlock(&vnode_list_mtx);
694 vnode_list_free_marker = vn_alloc_marker(NULL);
695 TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
696 vnode_list_reclaim_marker = vn_alloc_marker(NULL);
697 TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
706 vnode_zone = uma_zcreate("VNODE", sizeof(struct vnode), ctor, dtor,
707 vnode_init, vnode_fini, UMA_ALIGN_PTR, UMA_ZONE_NOKASAN);
708 uma_zone_set_smr(vnode_zone, vfs_smr);
711 * Preallocate enough nodes to support one-per buf so that
712 * we can not fail an insert. reassignbuf() callers can not
713 * tolerate the insertion failure.
715 buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
716 NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
717 UMA_ZONE_NOFREE | UMA_ZONE_SMR);
718 buf_trie_smr = uma_zone_get_smr(buf_trie_zone);
719 uma_prealloc(buf_trie_zone, nbuf);
721 vnodes_created = counter_u64_alloc(M_WAITOK);
722 recycles_count = counter_u64_alloc(M_WAITOK);
723 recycles_free_count = counter_u64_alloc(M_WAITOK);
724 deferred_inact = counter_u64_alloc(M_WAITOK);
727 * Initialize the filesystem syncer.
729 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
731 syncer_maxdelay = syncer_mask + 1;
732 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
733 cv_init(&sync_wakeup, "syncer");
734 for (i = 1; i <= sizeof(struct vnode); i <<= 1)
739 vd = DPCPU_ID_PTR((cpu), vd);
740 bzero(vd, sizeof(*vd));
741 mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
744 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
747 * Mark a mount point as busy. Used to synchronize access and to delay
748 * unmounting. Eventually, mountlist_mtx is not released on failure.
750 * vfs_busy() is a custom lock, it can block the caller.
751 * vfs_busy() only sleeps if the unmount is active on the mount point.
752 * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
753 * vnode belonging to mp.
755 * Lookup uses vfs_busy() to traverse mount points.
757 * / vnode lock A / vnode lock (/var) D
758 * /var vnode lock B /log vnode lock(/var/log) E
759 * vfs_busy lock C vfs_busy lock F
761 * Within each file system, the lock order is C->A->B and F->D->E.
763 * When traversing across mounts, the system follows that lock order:
769 * The lookup() process for namei("/var") illustrates the process:
770 * VOP_LOOKUP() obtains B while A is held
771 * vfs_busy() obtains a shared lock on F while A and B are held
772 * vput() releases lock on B
773 * vput() releases lock on A
774 * VFS_ROOT() obtains lock on D while shared lock on F is held
775 * vfs_unbusy() releases shared lock on F
776 * vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
777 * Attempt to lock A (instead of vp_crossmp) while D is held would
778 * violate the global order, causing deadlocks.
780 * dounmount() locks B while F is drained.
783 vfs_busy(struct mount *mp, int flags)
785 struct mount_pcpu *mpcpu;
787 MPASS((flags & ~MBF_MASK) == 0);
788 CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
790 if (vfs_op_thread_enter(mp, mpcpu)) {
791 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
792 MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
793 MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
794 vfs_mp_count_add_pcpu(mpcpu, ref, 1);
795 vfs_mp_count_add_pcpu(mpcpu, lockref, 1);
796 vfs_op_thread_exit(mp, mpcpu);
797 if (flags & MBF_MNTLSTLOCK)
798 mtx_unlock(&mountlist_mtx);
803 vfs_assert_mount_counters(mp);
806 * If mount point is currently being unmounted, sleep until the
807 * mount point fate is decided. If thread doing the unmounting fails,
808 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
809 * that this mount point has survived the unmount attempt and vfs_busy
810 * should retry. Otherwise the unmounter thread will set MNTK_REFEXPIRE
811 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
812 * about to be really destroyed. vfs_busy needs to release its
813 * reference on the mount point in this case and return with ENOENT,
814 * telling the caller that mount mount it tried to busy is no longer
817 while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
818 KASSERT(TAILQ_EMPTY(&mp->mnt_uppers),
819 ("%s: non-empty upper mount list with pending unmount",
821 if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
824 CTR1(KTR_VFS, "%s: failed busying before sleeping",
828 if (flags & MBF_MNTLSTLOCK)
829 mtx_unlock(&mountlist_mtx);
830 mp->mnt_kern_flag |= MNTK_MWAIT;
831 msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
832 if (flags & MBF_MNTLSTLOCK)
833 mtx_lock(&mountlist_mtx);
836 if (flags & MBF_MNTLSTLOCK)
837 mtx_unlock(&mountlist_mtx);
844 * Free a busy filesystem.
847 vfs_unbusy(struct mount *mp)
849 struct mount_pcpu *mpcpu;
852 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
854 if (vfs_op_thread_enter(mp, mpcpu)) {
855 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
856 vfs_mp_count_sub_pcpu(mpcpu, lockref, 1);
857 vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
858 vfs_op_thread_exit(mp, mpcpu);
863 vfs_assert_mount_counters(mp);
865 c = --mp->mnt_lockref;
866 if (mp->mnt_vfs_ops == 0) {
867 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
872 vfs_dump_mount_counters(mp);
873 if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
874 MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
875 CTR1(KTR_VFS, "%s: waking up waiters", __func__);
876 mp->mnt_kern_flag &= ~MNTK_DRAINING;
877 wakeup(&mp->mnt_lockref);
883 * Lookup a mount point by filesystem identifier.
886 vfs_getvfs(fsid_t *fsid)
890 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
891 mtx_lock(&mountlist_mtx);
892 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
893 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
895 mtx_unlock(&mountlist_mtx);
899 mtx_unlock(&mountlist_mtx);
900 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
901 return ((struct mount *) 0);
905 * Lookup a mount point by filesystem identifier, busying it before
908 * To avoid congestion on mountlist_mtx, implement simple direct-mapped
909 * cache for popular filesystem identifiers. The cache is lockess, using
910 * the fact that struct mount's are never freed. In worst case we may
911 * get pointer to unmounted or even different filesystem, so we have to
912 * check what we got, and go slow way if so.
915 vfs_busyfs(fsid_t *fsid)
917 #define FSID_CACHE_SIZE 256
918 typedef struct mount * volatile vmp_t;
919 static vmp_t cache[FSID_CACHE_SIZE];
924 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
925 hash = fsid->val[0] ^ fsid->val[1];
926 hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
928 if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0)
930 if (vfs_busy(mp, 0) != 0) {
934 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0)
940 mtx_lock(&mountlist_mtx);
941 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
942 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
943 error = vfs_busy(mp, MBF_MNTLSTLOCK);
946 mtx_unlock(&mountlist_mtx);
953 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
954 mtx_unlock(&mountlist_mtx);
955 return ((struct mount *) 0);
959 * Check if a user can access privileged mount options.
962 vfs_suser(struct mount *mp, struct thread *td)
966 if (jailed(td->td_ucred)) {
968 * If the jail of the calling thread lacks permission for
969 * this type of file system, deny immediately.
971 if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
975 * If the file system was mounted outside the jail of the
976 * calling thread, deny immediately.
978 if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
983 * If file system supports delegated administration, we don't check
984 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
985 * by the file system itself.
986 * If this is not the user that did original mount, we check for
987 * the PRIV_VFS_MOUNT_OWNER privilege.
989 if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
990 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
991 if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
998 * Get a new unique fsid. Try to make its val[0] unique, since this value
999 * will be used to create fake device numbers for stat(). Also try (but
1000 * not so hard) make its val[0] unique mod 2^16, since some emulators only
1001 * support 16-bit device numbers. We end up with unique val[0]'s for the
1002 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
1004 * Keep in mind that several mounts may be running in parallel. Starting
1005 * the search one past where the previous search terminated is both a
1006 * micro-optimization and a defense against returning the same fsid to
1010 vfs_getnewfsid(struct mount *mp)
1012 static uint16_t mntid_base;
1017 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
1018 mtx_lock(&mntid_mtx);
1019 mtype = mp->mnt_vfc->vfc_typenum;
1020 tfsid.val[1] = mtype;
1021 mtype = (mtype & 0xFF) << 24;
1023 tfsid.val[0] = makedev(255,
1024 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
1026 if ((nmp = vfs_getvfs(&tfsid)) == NULL)
1030 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
1031 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
1032 mtx_unlock(&mntid_mtx);
1036 * Knob to control the precision of file timestamps:
1038 * 0 = seconds only; nanoseconds zeroed.
1039 * 1 = seconds and nanoseconds, accurate within 1/HZ.
1040 * 2 = seconds and nanoseconds, truncated to microseconds.
1041 * >=3 = seconds and nanoseconds, maximum precision.
1043 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
1045 static int timestamp_precision = TSP_USEC;
1046 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
1047 ×tamp_precision, 0, "File timestamp precision (0: seconds, "
1048 "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1049 "3+: sec + ns (max. precision))");
1052 * Get a current timestamp.
1055 vfs_timestamp(struct timespec *tsp)
1059 switch (timestamp_precision) {
1061 tsp->tv_sec = time_second;
1069 TIMEVAL_TO_TIMESPEC(&tv, tsp);
1079 * Set vnode attributes to VNOVAL
1082 vattr_null(struct vattr *vap)
1085 vap->va_type = VNON;
1086 vap->va_size = VNOVAL;
1087 vap->va_bytes = VNOVAL;
1088 vap->va_mode = VNOVAL;
1089 vap->va_nlink = VNOVAL;
1090 vap->va_uid = VNOVAL;
1091 vap->va_gid = VNOVAL;
1092 vap->va_fsid = VNOVAL;
1093 vap->va_fileid = VNOVAL;
1094 vap->va_blocksize = VNOVAL;
1095 vap->va_rdev = VNOVAL;
1096 vap->va_atime.tv_sec = VNOVAL;
1097 vap->va_atime.tv_nsec = VNOVAL;
1098 vap->va_mtime.tv_sec = VNOVAL;
1099 vap->va_mtime.tv_nsec = VNOVAL;
1100 vap->va_ctime.tv_sec = VNOVAL;
1101 vap->va_ctime.tv_nsec = VNOVAL;
1102 vap->va_birthtime.tv_sec = VNOVAL;
1103 vap->va_birthtime.tv_nsec = VNOVAL;
1104 vap->va_flags = VNOVAL;
1105 vap->va_gen = VNOVAL;
1106 vap->va_vaflags = 0;
1110 * Try to reduce the total number of vnodes.
1112 * This routine (and its user) are buggy in at least the following ways:
1113 * - all parameters were picked years ago when RAM sizes were significantly
1115 * - it can pick vnodes based on pages used by the vm object, but filesystems
1116 * like ZFS don't use it making the pick broken
1117 * - since ZFS has its own aging policy it gets partially combated by this one
1118 * - a dedicated method should be provided for filesystems to let them decide
1119 * whether the vnode should be recycled
1121 * This routine is called when we have too many vnodes. It attempts
1122 * to free <count> vnodes and will potentially free vnodes that still
1123 * have VM backing store (VM backing store is typically the cause
1124 * of a vnode blowout so we want to do this). Therefore, this operation
1125 * is not considered cheap.
1127 * A number of conditions may prevent a vnode from being reclaimed.
1128 * the buffer cache may have references on the vnode, a directory
1129 * vnode may still have references due to the namei cache representing
1130 * underlying files, or the vnode may be in active use. It is not
1131 * desirable to reuse such vnodes. These conditions may cause the
1132 * number of vnodes to reach some minimum value regardless of what
1133 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low.
1135 * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1136 * entries if this argument is strue
1137 * @param trigger Only reclaim vnodes with fewer than this many resident
1139 * @param target How many vnodes to reclaim.
1140 * @return The number of vnodes that were reclaimed.
1143 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1145 struct vnode *vp, *mvp;
1147 struct vm_object *object;
1151 mtx_assert(&vnode_list_mtx, MA_OWNED);
1156 mvp = vnode_list_reclaim_marker;
1159 while (done < target) {
1160 vp = TAILQ_NEXT(vp, v_vnodelist);
1161 if (__predict_false(vp == NULL))
1164 if (__predict_false(vp->v_type == VMARKER))
1168 * If it's been deconstructed already, it's still
1169 * referenced, or it exceeds the trigger, skip it.
1170 * Also skip free vnodes. We are trying to make space
1171 * to expand the free list, not reduce it.
1173 if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1174 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1177 if (vp->v_type == VBAD || vp->v_type == VNON)
1180 object = atomic_load_ptr(&vp->v_object);
1181 if (object == NULL || object->resident_page_count > trigger) {
1186 * Handle races against vnode allocation. Filesystems lock the
1187 * vnode some time after it gets returned from getnewvnode,
1188 * despite type and hold count being manipulated earlier.
1189 * Resorting to checking v_mount restores guarantees present
1190 * before the global list was reworked to contain all vnodes.
1192 if (!VI_TRYLOCK(vp))
1194 if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1198 if (vp->v_mount == NULL) {
1204 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1205 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1206 mtx_unlock(&vnode_list_mtx);
1208 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1210 goto next_iter_unlocked;
1212 if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1214 vn_finished_write(mp);
1215 goto next_iter_unlocked;
1219 if (vp->v_usecount > 0 ||
1220 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1221 (vp->v_object != NULL && vp->v_object->handle == vp &&
1222 vp->v_object->resident_page_count > trigger)) {
1225 vn_finished_write(mp);
1226 goto next_iter_unlocked;
1228 counter_u64_add(recycles_count, 1);
1232 vn_finished_write(mp);
1236 kern_yield(PRI_USER);
1237 mtx_lock(&vnode_list_mtx);
1240 MPASS(vp->v_type != VMARKER);
1241 if (!should_yield())
1243 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1244 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1245 mtx_unlock(&vnode_list_mtx);
1246 kern_yield(PRI_USER);
1247 mtx_lock(&vnode_list_mtx);
1250 if (done == 0 && !retried) {
1251 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1252 TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1259 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1260 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1262 "limit on vnode free requests per call to the vnlru_free routine");
1265 * Attempt to reduce the free list by the requested amount.
1268 vnlru_free_impl(int count, struct vfsops *mnt_op, struct vnode *mvp)
1274 mtx_assert(&vnode_list_mtx, MA_OWNED);
1275 if (count > max_vnlru_free)
1276 count = max_vnlru_free;
1283 vp = TAILQ_NEXT(vp, v_vnodelist);
1284 if (__predict_false(vp == NULL)) {
1285 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1286 TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1289 if (__predict_false(vp->v_type == VMARKER))
1291 if (vp->v_holdcnt > 0)
1294 * Don't recycle if our vnode is from different type
1295 * of mount point. Note that mp is type-safe, the
1296 * check does not reach unmapped address even if
1297 * vnode is reclaimed.
1299 if (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1300 mp->mnt_op != mnt_op) {
1303 if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1306 if (!vhold_recycle_free(vp))
1308 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1309 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1310 mtx_unlock(&vnode_list_mtx);
1311 if (vtryrecycle(vp) == 0)
1313 mtx_lock(&vnode_list_mtx);
1316 return (ocount - count);
1320 vnlru_free_locked(int count)
1323 mtx_assert(&vnode_list_mtx, MA_OWNED);
1324 return (vnlru_free_impl(count, NULL, vnode_list_free_marker));
1328 vnlru_free_vfsops(int count, struct vfsops *mnt_op, struct vnode *mvp)
1331 MPASS(mnt_op != NULL);
1333 VNPASS(mvp->v_type == VMARKER, mvp);
1334 mtx_lock(&vnode_list_mtx);
1335 vnlru_free_impl(count, mnt_op, mvp);
1336 mtx_unlock(&vnode_list_mtx);
1340 vnlru_alloc_marker(void)
1344 mvp = vn_alloc_marker(NULL);
1345 mtx_lock(&vnode_list_mtx);
1346 TAILQ_INSERT_BEFORE(vnode_list_free_marker, mvp, v_vnodelist);
1347 mtx_unlock(&vnode_list_mtx);
1352 vnlru_free_marker(struct vnode *mvp)
1354 mtx_lock(&vnode_list_mtx);
1355 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1356 mtx_unlock(&vnode_list_mtx);
1357 vn_free_marker(mvp);
1364 mtx_assert(&vnode_list_mtx, MA_OWNED);
1365 gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1366 vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1367 vlowat = vhiwat / 2;
1371 * Attempt to recycle vnodes in a context that is always safe to block.
1372 * Calling vlrurecycle() from the bowels of filesystem code has some
1373 * interesting deadlock problems.
1375 static struct proc *vnlruproc;
1376 static int vnlruproc_sig;
1379 * The main freevnodes counter is only updated when threads requeue their vnode
1380 * batches. CPUs are conditionally walked to compute a more accurate total.
1382 * Limit how much of a slop are we willing to tolerate. Note: the actual value
1383 * at any given moment can still exceed slop, but it should not be by significant
1384 * margin in practice.
1386 #define VNLRU_FREEVNODES_SLOP 128
1388 static __inline void
1389 vfs_freevnodes_inc(void)
1399 static __inline void
1400 vfs_freevnodes_dec(void)
1411 vnlru_read_freevnodes(void)
1417 mtx_assert(&vnode_list_mtx, MA_OWNED);
1418 if (freevnodes > freevnodes_old)
1419 slop = freevnodes - freevnodes_old;
1421 slop = freevnodes_old - freevnodes;
1422 if (slop < VNLRU_FREEVNODES_SLOP)
1423 return (freevnodes >= 0 ? freevnodes : 0);
1424 freevnodes_old = freevnodes;
1426 vd = DPCPU_ID_PTR((cpu), vd);
1427 freevnodes_old += vd->freevnodes;
1429 return (freevnodes_old >= 0 ? freevnodes_old : 0);
1433 vnlru_under(u_long rnumvnodes, u_long limit)
1435 u_long rfreevnodes, space;
1437 if (__predict_false(rnumvnodes > desiredvnodes))
1440 space = desiredvnodes - rnumvnodes;
1441 if (space < limit) {
1442 rfreevnodes = vnlru_read_freevnodes();
1443 if (rfreevnodes > wantfreevnodes)
1444 space += rfreevnodes - wantfreevnodes;
1446 return (space < limit);
1450 vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1452 long rfreevnodes, space;
1454 if (__predict_false(rnumvnodes > desiredvnodes))
1457 space = desiredvnodes - rnumvnodes;
1458 if (space < limit) {
1459 rfreevnodes = atomic_load_long(&freevnodes);
1460 if (rfreevnodes > wantfreevnodes)
1461 space += rfreevnodes - wantfreevnodes;
1463 return (space < limit);
1470 mtx_assert(&vnode_list_mtx, MA_OWNED);
1471 if (vnlruproc_sig == 0) {
1480 u_long rnumvnodes, rfreevnodes, target;
1481 unsigned long onumvnodes;
1482 int done, force, trigger, usevnodes;
1483 bool reclaim_nc_src, want_reread;
1485 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1486 SHUTDOWN_PRI_FIRST);
1489 want_reread = false;
1491 kproc_suspend_check(vnlruproc);
1492 mtx_lock(&vnode_list_mtx);
1493 rnumvnodes = atomic_load_long(&numvnodes);
1496 force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1497 want_reread = false;
1501 * If numvnodes is too large (due to desiredvnodes being
1502 * adjusted using its sysctl, or emergency growth), first
1503 * try to reduce it by discarding from the free list.
1505 if (rnumvnodes > desiredvnodes) {
1506 vnlru_free_locked(rnumvnodes - desiredvnodes);
1507 rnumvnodes = atomic_load_long(&numvnodes);
1510 * Sleep if the vnode cache is in a good state. This is
1511 * when it is not over-full and has space for about a 4%
1512 * or 9% expansion (by growing its size or inexcessively
1513 * reducing its free list). Otherwise, try to reclaim
1514 * space for a 10% expansion.
1516 if (vstir && force == 0) {
1520 if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1522 wakeup(&vnlruproc_sig);
1523 msleep(vnlruproc, &vnode_list_mtx,
1524 PVFS|PDROP, "vlruwt", hz);
1527 rfreevnodes = vnlru_read_freevnodes();
1529 onumvnodes = rnumvnodes;
1531 * Calculate parameters for recycling. These are the same
1532 * throughout the loop to give some semblance of fairness.
1533 * The trigger point is to avoid recycling vnodes with lots
1534 * of resident pages. We aren't trying to free memory; we
1535 * are trying to recycle or at least free vnodes.
1537 if (rnumvnodes <= desiredvnodes)
1538 usevnodes = rnumvnodes - rfreevnodes;
1540 usevnodes = rnumvnodes;
1544 * The trigger value is is chosen to give a conservatively
1545 * large value to ensure that it alone doesn't prevent
1546 * making progress. The value can easily be so large that
1547 * it is effectively infinite in some congested and
1548 * misconfigured cases, and this is necessary. Normally
1549 * it is about 8 to 100 (pages), which is quite large.
1551 trigger = vm_cnt.v_page_count * 2 / usevnodes;
1553 trigger = vsmalltrigger;
1554 reclaim_nc_src = force >= 3;
1555 target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1556 target = target / 10 + 1;
1557 done = vlrureclaim(reclaim_nc_src, trigger, target);
1558 mtx_unlock(&vnode_list_mtx);
1559 if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1560 uma_reclaim(UMA_RECLAIM_DRAIN);
1562 if (force == 0 || force == 1) {
1573 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1576 kern_yield(PRI_USER);
1581 static struct kproc_desc vnlru_kp = {
1586 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1590 * Routines having to do with the management of the vnode table.
1594 * Try to recycle a freed vnode. We abort if anyone picks up a reference
1595 * before we actually vgone(). This function must be called with the vnode
1596 * held to prevent the vnode from being returned to the free list midway
1600 vtryrecycle(struct vnode *vp)
1604 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1605 VNASSERT(vp->v_holdcnt, vp,
1606 ("vtryrecycle: Recycling vp %p without a reference.", vp));
1608 * This vnode may found and locked via some other list, if so we
1609 * can't recycle it yet.
1611 if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1613 "%s: impossible to recycle, vp %p lock is already held",
1616 return (EWOULDBLOCK);
1619 * Don't recycle if its filesystem is being suspended.
1621 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1624 "%s: impossible to recycle, cannot start the write for %p",
1630 * If we got this far, we need to acquire the interlock and see if
1631 * anyone picked up this vnode from another list. If not, we will
1632 * mark it with DOOMED via vgonel() so that anyone who does find it
1633 * will skip over it.
1636 if (vp->v_usecount) {
1639 vn_finished_write(vnmp);
1641 "%s: impossible to recycle, %p is already referenced",
1645 if (!VN_IS_DOOMED(vp)) {
1646 counter_u64_add(recycles_free_count, 1);
1651 vn_finished_write(vnmp);
1656 * Allocate a new vnode.
1658 * The operation never returns an error. Returning an error was disabled
1659 * in r145385 (dated 2005) with the following comment:
1661 * XXX Not all VFS_VGET/ffs_vget callers check returns.
1663 * Given the age of this commit (almost 15 years at the time of writing this
1664 * comment) restoring the ability to fail requires a significant audit of
1667 * The routine can try to free a vnode or stall for up to 1 second waiting for
1668 * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1670 static u_long vn_alloc_cyclecount;
1672 static struct vnode * __noinline
1673 vn_alloc_hard(struct mount *mp)
1675 u_long rnumvnodes, rfreevnodes;
1677 mtx_lock(&vnode_list_mtx);
1678 rnumvnodes = atomic_load_long(&numvnodes);
1679 if (rnumvnodes + 1 < desiredvnodes) {
1680 vn_alloc_cyclecount = 0;
1683 rfreevnodes = vnlru_read_freevnodes();
1684 if (vn_alloc_cyclecount++ >= rfreevnodes) {
1685 vn_alloc_cyclecount = 0;
1689 * Grow the vnode cache if it will not be above its target max
1690 * after growing. Otherwise, if the free list is nonempty, try
1691 * to reclaim 1 item from it before growing the cache (possibly
1692 * above its target max if the reclamation failed or is delayed).
1693 * Otherwise, wait for some space. In all cases, schedule
1694 * vnlru_proc() if we are getting short of space. The watermarks
1695 * should be chosen so that we never wait or even reclaim from
1696 * the free list to below its target minimum.
1698 if (vnlru_free_locked(1) > 0)
1700 if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1702 * Wait for space for a new vnode.
1705 msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1706 if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1707 vnlru_read_freevnodes() > 1)
1708 vnlru_free_locked(1);
1711 rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1712 if (vnlru_under(rnumvnodes, vlowat))
1714 mtx_unlock(&vnode_list_mtx);
1715 return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1718 static struct vnode *
1719 vn_alloc(struct mount *mp)
1723 if (__predict_false(vn_alloc_cyclecount != 0))
1724 return (vn_alloc_hard(mp));
1725 rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1726 if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1727 atomic_subtract_long(&numvnodes, 1);
1728 return (vn_alloc_hard(mp));
1731 return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1735 vn_free(struct vnode *vp)
1738 atomic_subtract_long(&numvnodes, 1);
1739 uma_zfree_smr(vnode_zone, vp);
1743 * Return the next vnode from the free list.
1746 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1751 struct lock_object *lo;
1753 CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1755 KASSERT(vops->registered,
1756 ("%s: not registered vector op %p\n", __func__, vops));
1759 if (td->td_vp_reserved != NULL) {
1760 vp = td->td_vp_reserved;
1761 td->td_vp_reserved = NULL;
1765 counter_u64_add(vnodes_created, 1);
1767 * Locks are given the generic name "vnode" when created.
1768 * Follow the historic practice of using the filesystem
1769 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1771 * Locks live in a witness group keyed on their name. Thus,
1772 * when a lock is renamed, it must also move from the witness
1773 * group of its old name to the witness group of its new name.
1775 * The change only needs to be made when the vnode moves
1776 * from one filesystem type to another. We ensure that each
1777 * filesystem use a single static name pointer for its tag so
1778 * that we can compare pointers rather than doing a strcmp().
1780 lo = &vp->v_vnlock->lock_object;
1782 if (lo->lo_name != tag) {
1786 WITNESS_DESTROY(lo);
1787 WITNESS_INIT(lo, tag);
1791 * By default, don't allow shared locks unless filesystems opt-in.
1793 vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1795 * Finalize various vnode identity bits.
1797 KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1798 KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1799 KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1803 v_init_counters(vp);
1805 vp->v_bufobj.bo_ops = &buf_ops_bio;
1807 if (mp == NULL && vops != &dead_vnodeops)
1808 printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1812 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1813 mac_vnode_associate_singlelabel(mp, vp);
1816 vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1820 * For the filesystems which do not use vfs_hash_insert(),
1821 * still initialize v_hash to have vfs_hash_index() useful.
1822 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1825 vp->v_hash = (uintptr_t)vp >> vnsz2log;
1832 getnewvnode_reserve(void)
1837 MPASS(td->td_vp_reserved == NULL);
1838 td->td_vp_reserved = vn_alloc(NULL);
1842 getnewvnode_drop_reserve(void)
1847 if (td->td_vp_reserved != NULL) {
1848 vn_free(td->td_vp_reserved);
1849 td->td_vp_reserved = NULL;
1853 static void __noinline
1854 freevnode(struct vnode *vp)
1859 * The vnode has been marked for destruction, so free it.
1861 * The vnode will be returned to the zone where it will
1862 * normally remain until it is needed for another vnode. We
1863 * need to cleanup (or verify that the cleanup has already
1864 * been done) any residual data left from its current use
1865 * so as not to contaminate the freshly allocated vnode.
1867 CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1869 * Paired with vgone.
1871 vn_seqc_write_end_free(vp);
1874 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1875 VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp);
1876 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1877 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1878 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1879 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1880 VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1881 ("clean blk trie not empty"));
1882 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1883 VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1884 ("dirty blk trie not empty"));
1885 VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1886 VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1887 VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1888 VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1889 ("Dangling rangelock waiters"));
1890 VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp,
1891 ("Leaked inactivation"));
1894 mac_vnode_destroy(vp);
1896 if (vp->v_pollinfo != NULL) {
1897 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1898 destroy_vpollinfo(vp->v_pollinfo);
1900 vp->v_pollinfo = NULL;
1902 vp->v_mountedhere = NULL;
1905 vp->v_fifoinfo = NULL;
1913 * Delete from old mount point vnode list, if on one.
1916 delmntque(struct vnode *vp)
1920 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
1929 VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1930 ("bad mount point vnode list size"));
1931 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1932 mp->mnt_nvnodelistsize--;
1938 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1942 vp->v_op = &dead_vnodeops;
1948 * Insert into list of vnodes for the new mount point, if available.
1951 insmntque1(struct vnode *vp, struct mount *mp,
1952 void (*dtr)(struct vnode *, void *), void *dtr_arg)
1955 KASSERT(vp->v_mount == NULL,
1956 ("insmntque: vnode already on per mount vnode list"));
1957 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1958 ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1961 * We acquire the vnode interlock early to ensure that the
1962 * vnode cannot be recycled by another process releasing a
1963 * holdcnt on it before we get it on both the vnode list
1964 * and the active vnode list. The mount mutex protects only
1965 * manipulation of the vnode list and the vnode freelist
1966 * mutex protects only manipulation of the active vnode list.
1967 * Hence the need to hold the vnode interlock throughout.
1971 if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1972 ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1973 mp->mnt_nvnodelistsize == 0)) &&
1974 (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1983 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1984 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1985 ("neg mount point vnode list size"));
1986 mp->mnt_nvnodelistsize++;
1993 insmntque(struct vnode *vp, struct mount *mp)
1996 return (insmntque1(vp, mp, insmntque_stddtr, NULL));
2000 * Flush out and invalidate all buffers associated with a bufobj
2001 * Called with the underlying object locked.
2004 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
2009 if (flags & V_SAVE) {
2010 error = bufobj_wwait(bo, slpflag, slptimeo);
2015 if (bo->bo_dirty.bv_cnt > 0) {
2018 error = BO_SYNC(bo, MNT_WAIT);
2019 } while (error == ERELOOKUP);
2023 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) {
2030 * If you alter this loop please notice that interlock is dropped and
2031 * reacquired in flushbuflist. Special care is needed to ensure that
2032 * no race conditions occur from this.
2035 error = flushbuflist(&bo->bo_clean,
2036 flags, bo, slpflag, slptimeo);
2037 if (error == 0 && !(flags & V_CLEANONLY))
2038 error = flushbuflist(&bo->bo_dirty,
2039 flags, bo, slpflag, slptimeo);
2040 if (error != 0 && error != EAGAIN) {
2044 } while (error != 0);
2047 * Wait for I/O to complete. XXX needs cleaning up. The vnode can
2048 * have write I/O in-progress but if there is a VM object then the
2049 * VM object can also have read-I/O in-progress.
2052 bufobj_wwait(bo, 0, 0);
2053 if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
2055 vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
2058 } while (bo->bo_numoutput > 0);
2062 * Destroy the copy in the VM cache, too.
2064 if (bo->bo_object != NULL &&
2065 (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
2066 VM_OBJECT_WLOCK(bo->bo_object);
2067 vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
2068 OBJPR_CLEANONLY : 0);
2069 VM_OBJECT_WUNLOCK(bo->bo_object);
2074 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
2075 V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
2076 bo->bo_clean.bv_cnt > 0))
2077 panic("vinvalbuf: flush failed");
2078 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
2079 bo->bo_dirty.bv_cnt > 0)
2080 panic("vinvalbuf: flush dirty failed");
2087 * Flush out and invalidate all buffers associated with a vnode.
2088 * Called with the underlying object locked.
2091 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
2094 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2095 ASSERT_VOP_LOCKED(vp, "vinvalbuf");
2096 if (vp->v_object != NULL && vp->v_object->handle != vp)
2098 return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
2102 * Flush out buffers on the specified list.
2106 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2109 struct buf *bp, *nbp;
2114 ASSERT_BO_WLOCKED(bo);
2117 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2119 * If we are flushing both V_NORMAL and V_ALT buffers then
2120 * do not skip any buffers. If we are flushing only V_NORMAL
2121 * buffers then skip buffers marked as BX_ALTDATA. If we are
2122 * flushing only V_ALT buffers then skip buffers not marked
2125 if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2126 (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2127 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2131 lblkno = nbp->b_lblkno;
2132 xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2135 error = BUF_TIMELOCK(bp,
2136 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2137 "flushbuf", slpflag, slptimeo);
2140 return (error != ENOLCK ? error : EAGAIN);
2142 KASSERT(bp->b_bufobj == bo,
2143 ("bp %p wrong b_bufobj %p should be %p",
2144 bp, bp->b_bufobj, bo));
2146 * XXX Since there are no node locks for NFS, I
2147 * believe there is a slight chance that a delayed
2148 * write will occur while sleeping just above, so
2151 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2154 bp->b_flags |= B_ASYNC;
2157 return (EAGAIN); /* XXX: why not loop ? */
2160 bp->b_flags |= (B_INVAL | B_RELBUF);
2161 bp->b_flags &= ~B_ASYNC;
2166 nbp = gbincore(bo, lblkno);
2167 if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2169 break; /* nbp invalid */
2175 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2181 ASSERT_BO_LOCKED(bo);
2183 for (lblkno = startn;;) {
2185 bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2186 if (bp == NULL || bp->b_lblkno >= endn ||
2187 bp->b_lblkno < startn)
2189 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2190 LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2193 if (error == ENOLCK)
2197 KASSERT(bp->b_bufobj == bo,
2198 ("bp %p wrong b_bufobj %p should be %p",
2199 bp, bp->b_bufobj, bo));
2200 lblkno = bp->b_lblkno + 1;
2201 if ((bp->b_flags & B_MANAGED) == 0)
2203 bp->b_flags |= B_RELBUF;
2205 * In the VMIO case, use the B_NOREUSE flag to hint that the
2206 * pages backing each buffer in the range are unlikely to be
2207 * reused. Dirty buffers will have the hint applied once
2208 * they've been written.
2210 if ((bp->b_flags & B_VMIO) != 0)
2211 bp->b_flags |= B_NOREUSE;
2219 * Truncate a file's buffer and pages to a specified length. This
2220 * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2224 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2226 struct buf *bp, *nbp;
2230 CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2231 vp, blksize, (uintmax_t)length);
2234 * Round up to the *next* lbn.
2236 startlbn = howmany(length, blksize);
2238 ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2244 while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2249 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2250 if (bp->b_lblkno > 0)
2253 * Since we hold the vnode lock this should only
2254 * fail if we're racing with the buf daemon.
2257 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2258 BO_LOCKPTR(bo)) == ENOLCK)
2259 goto restart_unlocked;
2261 VNASSERT((bp->b_flags & B_DELWRI), vp,
2262 ("buf(%p) on dirty queue without DELWRI", bp));
2271 bufobj_wwait(bo, 0, 0);
2273 vnode_pager_setsize(vp, length);
2279 * Invalidate the cached pages of a file's buffer within the range of block
2280 * numbers [startlbn, endlbn).
2283 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2289 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2291 start = blksize * startlbn;
2292 end = blksize * endlbn;
2296 MPASS(blksize == bo->bo_bsize);
2298 while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2302 vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2306 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2307 daddr_t startlbn, daddr_t endlbn)
2309 struct buf *bp, *nbp;
2312 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2313 ASSERT_BO_LOCKED(bo);
2317 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2318 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2321 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2322 BO_LOCKPTR(bo)) == ENOLCK) {
2328 bp->b_flags |= B_INVAL | B_RELBUF;
2329 bp->b_flags &= ~B_ASYNC;
2335 (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2337 (nbp->b_flags & B_DELWRI) != 0))
2341 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2342 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2345 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2346 BO_LOCKPTR(bo)) == ENOLCK) {
2351 bp->b_flags |= B_INVAL | B_RELBUF;
2352 bp->b_flags &= ~B_ASYNC;
2358 (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2359 (nbp->b_vp != vp) ||
2360 (nbp->b_flags & B_DELWRI) == 0))
2368 buf_vlist_remove(struct buf *bp)
2373 flags = bp->b_xflags;
2375 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2376 ASSERT_BO_WLOCKED(bp->b_bufobj);
2377 KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 &&
2378 (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN),
2379 ("%s: buffer %p has invalid queue state", __func__, bp));
2381 if ((flags & BX_VNDIRTY) != 0)
2382 bv = &bp->b_bufobj->bo_dirty;
2384 bv = &bp->b_bufobj->bo_clean;
2385 BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2386 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2388 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2392 * Add the buffer to the sorted clean or dirty block list.
2394 * NOTE: xflags is passed as a constant, optimizing this inline function!
2397 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2403 ASSERT_BO_WLOCKED(bo);
2404 KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2405 ("buf_vlist_add: bo %p does not allow bufs", bo));
2406 KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2407 ("dead bo %p", bo));
2408 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2409 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2410 bp->b_xflags |= xflags;
2411 if (xflags & BX_VNDIRTY)
2417 * Keep the list ordered. Optimize empty list insertion. Assume
2418 * we tend to grow at the tail so lookup_le should usually be cheaper
2421 if (bv->bv_cnt == 0 ||
2422 bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2423 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2424 else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2425 TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2427 TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2428 error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2430 panic("buf_vlist_add: Preallocated nodes insufficient.");
2435 * Look up a buffer using the buffer tries.
2438 gbincore(struct bufobj *bo, daddr_t lblkno)
2442 ASSERT_BO_LOCKED(bo);
2443 bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2446 return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno));
2450 * Look up a buf using the buffer tries, without the bufobj lock. This relies
2451 * on SMR for safe lookup, and bufs being in a no-free zone to provide type
2452 * stability of the result. Like other lockless lookups, the found buf may
2453 * already be invalid by the time this function returns.
2456 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno)
2460 ASSERT_BO_UNLOCKED(bo);
2461 bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno);
2464 return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno));
2468 * Associate a buffer with a vnode.
2471 bgetvp(struct vnode *vp, struct buf *bp)
2476 ASSERT_BO_WLOCKED(bo);
2477 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2479 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2480 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2481 ("bgetvp: bp already attached! %p", bp));
2487 * Insert onto list for new vnode.
2489 buf_vlist_add(bp, bo, BX_VNCLEAN);
2493 * Disassociate a buffer from a vnode.
2496 brelvp(struct buf *bp)
2501 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2502 KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2505 * Delete from old vnode list, if on one.
2507 vp = bp->b_vp; /* XXX */
2510 buf_vlist_remove(bp);
2511 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2512 bo->bo_flag &= ~BO_ONWORKLST;
2513 mtx_lock(&sync_mtx);
2514 LIST_REMOVE(bo, bo_synclist);
2515 syncer_worklist_len--;
2516 mtx_unlock(&sync_mtx);
2519 bp->b_bufobj = NULL;
2525 * Add an item to the syncer work queue.
2528 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2532 ASSERT_BO_WLOCKED(bo);
2534 mtx_lock(&sync_mtx);
2535 if (bo->bo_flag & BO_ONWORKLST)
2536 LIST_REMOVE(bo, bo_synclist);
2538 bo->bo_flag |= BO_ONWORKLST;
2539 syncer_worklist_len++;
2542 if (delay > syncer_maxdelay - 2)
2543 delay = syncer_maxdelay - 2;
2544 slot = (syncer_delayno + delay) & syncer_mask;
2546 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2547 mtx_unlock(&sync_mtx);
2551 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2555 mtx_lock(&sync_mtx);
2556 len = syncer_worklist_len - sync_vnode_count;
2557 mtx_unlock(&sync_mtx);
2558 error = SYSCTL_OUT(req, &len, sizeof(len));
2562 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2563 CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2564 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2566 static struct proc *updateproc;
2567 static void sched_sync(void);
2568 static struct kproc_desc up_kp = {
2573 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2576 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2581 *bo = LIST_FIRST(slp);
2585 if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2588 * We use vhold in case the vnode does not
2589 * successfully sync. vhold prevents the vnode from
2590 * going away when we unlock the sync_mtx so that
2591 * we can acquire the vnode interlock.
2594 mtx_unlock(&sync_mtx);
2596 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2598 mtx_lock(&sync_mtx);
2599 return (*bo == LIST_FIRST(slp));
2601 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2602 (void) VOP_FSYNC(vp, MNT_LAZY, td);
2604 vn_finished_write(mp);
2606 if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2608 * Put us back on the worklist. The worklist
2609 * routine will remove us from our current
2610 * position and then add us back in at a later
2613 vn_syncer_add_to_worklist(*bo, syncdelay);
2617 mtx_lock(&sync_mtx);
2621 static int first_printf = 1;
2624 * System filesystem synchronizer daemon.
2629 struct synclist *next, *slp;
2632 struct thread *td = curthread;
2634 int net_worklist_len;
2635 int syncer_final_iter;
2639 syncer_final_iter = 0;
2640 syncer_state = SYNCER_RUNNING;
2641 starttime = time_uptime;
2642 td->td_pflags |= TDP_NORUNNINGBUF;
2644 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2647 mtx_lock(&sync_mtx);
2649 if (syncer_state == SYNCER_FINAL_DELAY &&
2650 syncer_final_iter == 0) {
2651 mtx_unlock(&sync_mtx);
2652 kproc_suspend_check(td->td_proc);
2653 mtx_lock(&sync_mtx);
2655 net_worklist_len = syncer_worklist_len - sync_vnode_count;
2656 if (syncer_state != SYNCER_RUNNING &&
2657 starttime != time_uptime) {
2659 printf("\nSyncing disks, vnodes remaining... ");
2662 printf("%d ", net_worklist_len);
2664 starttime = time_uptime;
2667 * Push files whose dirty time has expired. Be careful
2668 * of interrupt race on slp queue.
2670 * Skip over empty worklist slots when shutting down.
2673 slp = &syncer_workitem_pending[syncer_delayno];
2674 syncer_delayno += 1;
2675 if (syncer_delayno == syncer_maxdelay)
2677 next = &syncer_workitem_pending[syncer_delayno];
2679 * If the worklist has wrapped since the
2680 * it was emptied of all but syncer vnodes,
2681 * switch to the FINAL_DELAY state and run
2682 * for one more second.
2684 if (syncer_state == SYNCER_SHUTTING_DOWN &&
2685 net_worklist_len == 0 &&
2686 last_work_seen == syncer_delayno) {
2687 syncer_state = SYNCER_FINAL_DELAY;
2688 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2690 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2691 syncer_worklist_len > 0);
2694 * Keep track of the last time there was anything
2695 * on the worklist other than syncer vnodes.
2696 * Return to the SHUTTING_DOWN state if any
2699 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2700 last_work_seen = syncer_delayno;
2701 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2702 syncer_state = SYNCER_SHUTTING_DOWN;
2703 while (!LIST_EMPTY(slp)) {
2704 error = sync_vnode(slp, &bo, td);
2706 LIST_REMOVE(bo, bo_synclist);
2707 LIST_INSERT_HEAD(next, bo, bo_synclist);
2711 if (first_printf == 0) {
2713 * Drop the sync mutex, because some watchdog
2714 * drivers need to sleep while patting
2716 mtx_unlock(&sync_mtx);
2717 wdog_kern_pat(WD_LASTVAL);
2718 mtx_lock(&sync_mtx);
2721 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2722 syncer_final_iter--;
2724 * The variable rushjob allows the kernel to speed up the
2725 * processing of the filesystem syncer process. A rushjob
2726 * value of N tells the filesystem syncer to process the next
2727 * N seconds worth of work on its queue ASAP. Currently rushjob
2728 * is used by the soft update code to speed up the filesystem
2729 * syncer process when the incore state is getting so far
2730 * ahead of the disk that the kernel memory pool is being
2731 * threatened with exhaustion.
2738 * Just sleep for a short period of time between
2739 * iterations when shutting down to allow some I/O
2742 * If it has taken us less than a second to process the
2743 * current work, then wait. Otherwise start right over
2744 * again. We can still lose time if any single round
2745 * takes more than two seconds, but it does not really
2746 * matter as we are just trying to generally pace the
2747 * filesystem activity.
2749 if (syncer_state != SYNCER_RUNNING ||
2750 time_uptime == starttime) {
2752 sched_prio(td, PPAUSE);
2755 if (syncer_state != SYNCER_RUNNING)
2756 cv_timedwait(&sync_wakeup, &sync_mtx,
2757 hz / SYNCER_SHUTDOWN_SPEEDUP);
2758 else if (time_uptime == starttime)
2759 cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2764 * Request the syncer daemon to speed up its work.
2765 * We never push it to speed up more than half of its
2766 * normal turn time, otherwise it could take over the cpu.
2769 speedup_syncer(void)
2773 mtx_lock(&sync_mtx);
2774 if (rushjob < syncdelay / 2) {
2776 stat_rush_requests += 1;
2779 mtx_unlock(&sync_mtx);
2780 cv_broadcast(&sync_wakeup);
2785 * Tell the syncer to speed up its work and run though its work
2786 * list several times, then tell it to shut down.
2789 syncer_shutdown(void *arg, int howto)
2792 if (howto & RB_NOSYNC)
2794 mtx_lock(&sync_mtx);
2795 syncer_state = SYNCER_SHUTTING_DOWN;
2797 mtx_unlock(&sync_mtx);
2798 cv_broadcast(&sync_wakeup);
2799 kproc_shutdown(arg, howto);
2803 syncer_suspend(void)
2806 syncer_shutdown(updateproc, 0);
2813 mtx_lock(&sync_mtx);
2815 syncer_state = SYNCER_RUNNING;
2816 mtx_unlock(&sync_mtx);
2817 cv_broadcast(&sync_wakeup);
2818 kproc_resume(updateproc);
2822 * Move the buffer between the clean and dirty lists of its vnode.
2825 reassignbuf(struct buf *bp)
2837 KASSERT((bp->b_flags & B_PAGING) == 0,
2838 ("%s: cannot reassign paging buffer %p", __func__, bp));
2840 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2841 bp, bp->b_vp, bp->b_flags);
2844 buf_vlist_remove(bp);
2847 * If dirty, put on list of dirty buffers; otherwise insert onto list
2850 if (bp->b_flags & B_DELWRI) {
2851 if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2852 switch (vp->v_type) {
2862 vn_syncer_add_to_worklist(bo, delay);
2864 buf_vlist_add(bp, bo, BX_VNDIRTY);
2866 buf_vlist_add(bp, bo, BX_VNCLEAN);
2868 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2869 mtx_lock(&sync_mtx);
2870 LIST_REMOVE(bo, bo_synclist);
2871 syncer_worklist_len--;
2872 mtx_unlock(&sync_mtx);
2873 bo->bo_flag &= ~BO_ONWORKLST;
2878 bp = TAILQ_FIRST(&bv->bv_hd);
2879 KASSERT(bp == NULL || bp->b_bufobj == bo,
2880 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2881 bp = TAILQ_LAST(&bv->bv_hd, buflists);
2882 KASSERT(bp == NULL || bp->b_bufobj == bo,
2883 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2885 bp = TAILQ_FIRST(&bv->bv_hd);
2886 KASSERT(bp == NULL || bp->b_bufobj == bo,
2887 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2888 bp = TAILQ_LAST(&bv->bv_hd, buflists);
2889 KASSERT(bp == NULL || bp->b_bufobj == bo,
2890 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2896 v_init_counters(struct vnode *vp)
2899 VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2900 vp, ("%s called for an initialized vnode", __FUNCTION__));
2901 ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2903 refcount_init(&vp->v_holdcnt, 1);
2904 refcount_init(&vp->v_usecount, 1);
2908 * Grab a particular vnode from the free list, increment its
2909 * reference count and lock it. VIRF_DOOMED is set if the vnode
2910 * is being destroyed. Only callers who specify LK_RETRY will
2911 * see doomed vnodes. If inactive processing was delayed in
2912 * vput try to do it here.
2914 * usecount is manipulated using atomics without holding any locks.
2916 * holdcnt can be manipulated using atomics without holding any locks,
2917 * except when transitioning 1<->0, in which case the interlock is held.
2919 * Consumers which don't guarantee liveness of the vnode can use SMR to
2920 * try to get a reference. Note this operation can fail since the vnode
2921 * may be awaiting getting freed by the time they get to it.
2924 vget_prep_smr(struct vnode *vp)
2928 VFS_SMR_ASSERT_ENTERED();
2930 if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2942 vget_prep(struct vnode *vp)
2946 if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2956 vget_abort(struct vnode *vp, enum vgetstate vs)
2967 __assert_unreachable();
2972 vget(struct vnode *vp, int flags)
2977 return (vget_finish(vp, flags, vs));
2981 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2985 if ((flags & LK_INTERLOCK) != 0)
2986 ASSERT_VI_LOCKED(vp, __func__);
2988 ASSERT_VI_UNLOCKED(vp, __func__);
2989 VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
2990 VNPASS(vp->v_holdcnt > 0, vp);
2991 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
2993 error = vn_lock(vp, flags);
2994 if (__predict_false(error != 0)) {
2996 CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
3001 vget_finish_ref(vp, vs);
3006 vget_finish_ref(struct vnode *vp, enum vgetstate vs)
3010 VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3011 VNPASS(vp->v_holdcnt > 0, vp);
3012 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3014 if (vs == VGET_USECOUNT)
3018 * We hold the vnode. If the usecount is 0 it will be utilized to keep
3019 * the vnode around. Otherwise someone else lended their hold count and
3020 * we have to drop ours.
3022 old = atomic_fetchadd_int(&vp->v_usecount, 1);
3023 VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3026 old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3027 VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3029 refcount_release(&vp->v_holdcnt);
3035 vref(struct vnode *vp)
3039 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3041 vget_finish_ref(vp, vs);
3045 vrefact(struct vnode *vp)
3048 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3050 int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3051 VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3053 refcount_acquire(&vp->v_usecount);
3058 vlazy(struct vnode *vp)
3062 VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3064 if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3067 * We may get here for inactive routines after the vnode got doomed.
3069 if (VN_IS_DOOMED(vp))
3072 mtx_lock(&mp->mnt_listmtx);
3073 if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3074 vp->v_mflag |= VMP_LAZYLIST;
3075 TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3076 mp->mnt_lazyvnodelistsize++;
3078 mtx_unlock(&mp->mnt_listmtx);
3082 vunlazy(struct vnode *vp)
3086 ASSERT_VI_LOCKED(vp, __func__);
3087 VNPASS(!VN_IS_DOOMED(vp), vp);
3090 mtx_lock(&mp->mnt_listmtx);
3091 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3093 * Don't remove the vnode from the lazy list if another thread
3094 * has increased the hold count. It may have re-enqueued the
3095 * vnode to the lazy list and is now responsible for its
3098 if (vp->v_holdcnt == 0) {
3099 vp->v_mflag &= ~VMP_LAZYLIST;
3100 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3101 mp->mnt_lazyvnodelistsize--;
3103 mtx_unlock(&mp->mnt_listmtx);
3107 * This routine is only meant to be called from vgonel prior to dooming
3111 vunlazy_gone(struct vnode *vp)
3115 ASSERT_VOP_ELOCKED(vp, __func__);
3116 ASSERT_VI_LOCKED(vp, __func__);
3117 VNPASS(!VN_IS_DOOMED(vp), vp);
3119 if (vp->v_mflag & VMP_LAZYLIST) {
3121 mtx_lock(&mp->mnt_listmtx);
3122 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3123 vp->v_mflag &= ~VMP_LAZYLIST;
3124 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3125 mp->mnt_lazyvnodelistsize--;
3126 mtx_unlock(&mp->mnt_listmtx);
3131 vdefer_inactive(struct vnode *vp)
3134 ASSERT_VI_LOCKED(vp, __func__);
3135 VNASSERT(vp->v_holdcnt > 0, vp,
3136 ("%s: vnode without hold count", __func__));
3137 if (VN_IS_DOOMED(vp)) {
3141 if (vp->v_iflag & VI_DEFINACT) {
3142 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3146 if (vp->v_usecount > 0) {
3147 vp->v_iflag &= ~VI_OWEINACT;
3152 vp->v_iflag |= VI_DEFINACT;
3154 counter_u64_add(deferred_inact, 1);
3158 vdefer_inactive_unlocked(struct vnode *vp)
3162 if ((vp->v_iflag & VI_OWEINACT) == 0) {
3166 vdefer_inactive(vp);
3169 enum vput_op { VRELE, VPUT, VUNREF };
3172 * Handle ->v_usecount transitioning to 0.
3174 * By releasing the last usecount we take ownership of the hold count which
3175 * provides liveness of the vnode, meaning we have to vdrop.
3177 * For all vnodes we may need to perform inactive processing. It requires an
3178 * exclusive lock on the vnode, while it is legal to call here with only a
3179 * shared lock (or no locks). If locking the vnode in an expected manner fails,
3180 * inactive processing gets deferred to the syncer.
3182 * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3183 * on the lock being held all the way until VOP_INACTIVE. This in particular
3184 * happens with UFS which adds half-constructed vnodes to the hash, where they
3185 * can be found by other code.
3188 vput_final(struct vnode *vp, enum vput_op func)
3193 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3194 VNPASS(vp->v_holdcnt > 0, vp);
3199 * By the time we got here someone else might have transitioned
3200 * the count back to > 0.
3202 if (vp->v_usecount > 0)
3206 * If the vnode is doomed vgone already performed inactive processing
3209 if (VN_IS_DOOMED(vp))
3212 if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3215 if (vp->v_iflag & VI_DOINGINACT)
3219 * Locking operations here will drop the interlock and possibly the
3220 * vnode lock, opening a window where the vnode can get doomed all the
3221 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3224 vp->v_iflag |= VI_OWEINACT;
3225 want_unlock = false;
3229 switch (VOP_ISLOCKED(vp)) {
3235 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3240 * The lock has at least one sharer, but we have no way
3241 * to conclude whether this is us. Play it safe and
3250 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3251 error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3257 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3258 error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3264 if (func == VUNREF) {
3265 VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp,
3266 ("recursive vunref"));
3267 vp->v_vflag |= VV_UNREF;
3270 error = vinactive(vp);
3273 if (error != ERELOOKUP || !want_unlock)
3275 VOP_LOCK(vp, LK_EXCLUSIVE);
3278 vp->v_vflag &= ~VV_UNREF;
3281 vdefer_inactive(vp);
3291 * Decrement ->v_usecount for a vnode.
3293 * Releasing the last use count requires additional processing, see vput_final
3294 * above for details.
3296 * Comment above each variant denotes lock state on entry and exit.
3301 * out: same as passed in
3304 vrele(struct vnode *vp)
3307 ASSERT_VI_UNLOCKED(vp, __func__);
3308 if (!refcount_release(&vp->v_usecount))
3310 vput_final(vp, VRELE);
3318 vput(struct vnode *vp)
3321 ASSERT_VOP_LOCKED(vp, __func__);
3322 ASSERT_VI_UNLOCKED(vp, __func__);
3323 if (!refcount_release(&vp->v_usecount)) {
3327 vput_final(vp, VPUT);
3335 vunref(struct vnode *vp)
3338 ASSERT_VOP_LOCKED(vp, __func__);
3339 ASSERT_VI_UNLOCKED(vp, __func__);
3340 if (!refcount_release(&vp->v_usecount))
3342 vput_final(vp, VUNREF);
3346 vhold(struct vnode *vp)
3350 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3351 old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3352 VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3353 ("%s: wrong hold count %d", __func__, old));
3355 vfs_freevnodes_dec();
3359 vholdnz(struct vnode *vp)
3362 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3364 int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3365 VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3366 ("%s: wrong hold count %d", __func__, old));
3368 atomic_add_int(&vp->v_holdcnt, 1);
3373 * Grab a hold count unless the vnode is freed.
3375 * Only use this routine if vfs smr is the only protection you have against
3376 * freeing the vnode.
3378 * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag
3379 * is not set. After the flag is set the vnode becomes immutable to anyone but
3380 * the thread which managed to set the flag.
3382 * It may be tempting to replace the loop with:
3383 * count = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3384 * if (count & VHOLD_NO_SMR) {
3385 * backpedal and error out;
3388 * However, while this is more performant, it hinders debugging by eliminating
3389 * the previously mentioned invariant.
3392 vhold_smr(struct vnode *vp)
3396 VFS_SMR_ASSERT_ENTERED();
3398 count = atomic_load_int(&vp->v_holdcnt);
3400 if (count & VHOLD_NO_SMR) {
3401 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3402 ("non-zero hold count with flags %d\n", count));
3405 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3406 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3408 vfs_freevnodes_dec();
3415 * Hold a free vnode for recycling.
3417 * Note: vnode_init references this comment.
3419 * Attempts to recycle only need the global vnode list lock and have no use for
3422 * However, vnodes get inserted into the global list before they get fully
3423 * initialized and stay there until UMA decides to free the memory. This in
3424 * particular means the target can be found before it becomes usable and after
3425 * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to
3428 * Note: the vnode may gain more references after we transition the count 0->1.
3431 vhold_recycle_free(struct vnode *vp)
3435 mtx_assert(&vnode_list_mtx, MA_OWNED);
3437 count = atomic_load_int(&vp->v_holdcnt);
3439 if (count & VHOLD_NO_SMR) {
3440 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3441 ("non-zero hold count with flags %d\n", count));
3444 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3448 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3449 vfs_freevnodes_dec();
3455 static void __noinline
3456 vdbatch_process(struct vdbatch *vd)
3461 mtx_assert(&vd->lock, MA_OWNED);
3462 MPASS(curthread->td_pinned > 0);
3463 MPASS(vd->index == VDBATCH_SIZE);
3465 mtx_lock(&vnode_list_mtx);
3467 freevnodes += vd->freevnodes;
3468 for (i = 0; i < VDBATCH_SIZE; i++) {
3470 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3471 TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3472 MPASS(vp->v_dbatchcpu != NOCPU);
3473 vp->v_dbatchcpu = NOCPU;
3475 mtx_unlock(&vnode_list_mtx);
3477 bzero(vd->tab, sizeof(vd->tab));
3483 vdbatch_enqueue(struct vnode *vp)
3487 ASSERT_VI_LOCKED(vp, __func__);
3488 VNASSERT(!VN_IS_DOOMED(vp), vp,
3489 ("%s: deferring requeue of a doomed vnode", __func__));
3491 if (vp->v_dbatchcpu != NOCPU) {
3498 mtx_lock(&vd->lock);
3499 MPASS(vd->index < VDBATCH_SIZE);
3500 MPASS(vd->tab[vd->index] == NULL);
3502 * A hack: we depend on being pinned so that we know what to put in
3505 vp->v_dbatchcpu = curcpu;
3506 vd->tab[vd->index] = vp;
3509 if (vd->index == VDBATCH_SIZE)
3510 vdbatch_process(vd);
3511 mtx_unlock(&vd->lock);
3516 * This routine must only be called for vnodes which are about to be
3517 * deallocated. Supporting dequeue for arbitrary vndoes would require
3518 * validating that the locked batch matches.
3521 vdbatch_dequeue(struct vnode *vp)
3527 VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3528 ("%s: called for a used vnode\n", __func__));
3530 cpu = vp->v_dbatchcpu;
3534 vd = DPCPU_ID_PTR(cpu, vd);
3535 mtx_lock(&vd->lock);
3536 for (i = 0; i < vd->index; i++) {
3537 if (vd->tab[i] != vp)
3539 vp->v_dbatchcpu = NOCPU;
3541 vd->tab[i] = vd->tab[vd->index];
3542 vd->tab[vd->index] = NULL;
3545 mtx_unlock(&vd->lock);
3547 * Either we dequeued the vnode above or the target CPU beat us to it.
3549 MPASS(vp->v_dbatchcpu == NOCPU);
3553 * Drop the hold count of the vnode. If this is the last reference to
3554 * the vnode we place it on the free list unless it has been vgone'd
3555 * (marked VIRF_DOOMED) in which case we will free it.
3557 * Because the vnode vm object keeps a hold reference on the vnode if
3558 * there is at least one resident non-cached page, the vnode cannot
3559 * leave the active list without the page cleanup done.
3561 static void __noinline
3562 vdropl_final(struct vnode *vp)
3565 ASSERT_VI_LOCKED(vp, __func__);
3566 VNPASS(VN_IS_DOOMED(vp), vp);
3568 * Set the VHOLD_NO_SMR flag.
3570 * We may be racing against vhold_smr. If they win we can just pretend
3571 * we never got this far, they will vdrop later.
3573 if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) {
3574 vfs_freevnodes_inc();
3577 * We lost the aforementioned race. Any subsequent access is
3578 * invalid as they might have managed to vdropl on their own.
3583 * Don't bump freevnodes as this one is going away.
3589 vdrop(struct vnode *vp)
3592 ASSERT_VI_UNLOCKED(vp, __func__);
3593 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3594 if (refcount_release_if_not_last(&vp->v_holdcnt))
3601 vdropl(struct vnode *vp)
3604 ASSERT_VI_LOCKED(vp, __func__);
3605 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3606 if (!refcount_release(&vp->v_holdcnt)) {
3610 VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp);
3611 VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
3612 if (VN_IS_DOOMED(vp)) {
3617 vfs_freevnodes_inc();
3618 if (vp->v_mflag & VMP_LAZYLIST) {
3622 * Also unlocks the interlock. We can't assert on it as we
3623 * released our hold and by now the vnode might have been
3626 vdbatch_enqueue(vp);
3630 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3631 * flags. DOINGINACT prevents us from recursing in calls to vinactive.
3634 vinactivef(struct vnode *vp)
3636 struct vm_object *obj;
3639 ASSERT_VOP_ELOCKED(vp, "vinactive");
3640 ASSERT_VI_LOCKED(vp, "vinactive");
3641 VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3642 ("vinactive: recursed on VI_DOINGINACT"));
3643 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3644 vp->v_iflag |= VI_DOINGINACT;
3645 vp->v_iflag &= ~VI_OWEINACT;
3648 * Before moving off the active list, we must be sure that any
3649 * modified pages are converted into the vnode's dirty
3650 * buffers, since these will no longer be checked once the
3651 * vnode is on the inactive list.
3653 * The write-out of the dirty pages is asynchronous. At the
3654 * point that VOP_INACTIVE() is called, there could still be
3655 * pending I/O and dirty pages in the object.
3657 if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3658 vm_object_mightbedirty(obj)) {
3659 VM_OBJECT_WLOCK(obj);
3660 vm_object_page_clean(obj, 0, 0, 0);
3661 VM_OBJECT_WUNLOCK(obj);
3663 error = VOP_INACTIVE(vp);
3665 VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3666 ("vinactive: lost VI_DOINGINACT"));
3667 vp->v_iflag &= ~VI_DOINGINACT;
3672 vinactive(struct vnode *vp)
3675 ASSERT_VOP_ELOCKED(vp, "vinactive");
3676 ASSERT_VI_LOCKED(vp, "vinactive");
3677 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3679 if ((vp->v_iflag & VI_OWEINACT) == 0)
3681 if (vp->v_iflag & VI_DOINGINACT)
3683 if (vp->v_usecount > 0) {
3684 vp->v_iflag &= ~VI_OWEINACT;
3687 return (vinactivef(vp));
3691 * Remove any vnodes in the vnode table belonging to mount point mp.
3693 * If FORCECLOSE is not specified, there should not be any active ones,
3694 * return error if any are found (nb: this is a user error, not a
3695 * system error). If FORCECLOSE is specified, detach any active vnodes
3698 * If WRITECLOSE is set, only flush out regular file vnodes open for
3701 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3703 * `rootrefs' specifies the base reference count for the root vnode
3704 * of this filesystem. The root vnode is considered busy if its
3705 * v_usecount exceeds this value. On a successful return, vflush(, td)
3706 * will call vrele() on the root vnode exactly rootrefs times.
3707 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3711 static int busyprt = 0; /* print out busy vnodes */
3712 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3716 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3718 struct vnode *vp, *mvp, *rootvp = NULL;
3720 int busy = 0, error;
3722 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3725 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3726 ("vflush: bad args"));
3728 * Get the filesystem root vnode. We can vput() it
3729 * immediately, since with rootrefs > 0, it won't go away.
3731 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3732 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3739 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3741 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3744 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3748 * Skip over a vnodes marked VV_SYSTEM.
3750 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3756 * If WRITECLOSE is set, flush out unlinked but still open
3757 * files (even if open only for reading) and regular file
3758 * vnodes open for writing.
3760 if (flags & WRITECLOSE) {
3761 if (vp->v_object != NULL) {
3762 VM_OBJECT_WLOCK(vp->v_object);
3763 vm_object_page_clean(vp->v_object, 0, 0, 0);
3764 VM_OBJECT_WUNLOCK(vp->v_object);
3767 error = VOP_FSYNC(vp, MNT_WAIT, td);
3768 } while (error == ERELOOKUP);
3772 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3775 error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3778 if ((vp->v_type == VNON ||
3779 (error == 0 && vattr.va_nlink > 0)) &&
3780 (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3788 * With v_usecount == 0, all we need to do is clear out the
3789 * vnode data structures and we are done.
3791 * If FORCECLOSE is set, forcibly close the vnode.
3793 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3799 vn_printf(vp, "vflush: busy vnode ");
3805 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3807 * If just the root vnode is busy, and if its refcount
3808 * is equal to `rootrefs', then go ahead and kill it.
3811 KASSERT(busy > 0, ("vflush: not busy"));
3812 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3813 ("vflush: usecount %d < rootrefs %d",
3814 rootvp->v_usecount, rootrefs));
3815 if (busy == 1 && rootvp->v_usecount == rootrefs) {
3816 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3824 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3828 for (; rootrefs > 0; rootrefs--)
3834 * Recycle an unused vnode to the front of the free list.
3837 vrecycle(struct vnode *vp)
3842 recycled = vrecyclel(vp);
3848 * vrecycle, with the vp interlock held.
3851 vrecyclel(struct vnode *vp)
3855 ASSERT_VOP_ELOCKED(vp, __func__);
3856 ASSERT_VI_LOCKED(vp, __func__);
3857 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3859 if (vp->v_usecount == 0) {
3867 * Eliminate all activity associated with a vnode
3868 * in preparation for reuse.
3871 vgone(struct vnode *vp)
3879 * Notify upper mounts about reclaimed or unlinked vnode.
3882 vfs_notify_upper(struct vnode *vp, int event)
3885 struct mount_upper_node *ump;
3887 mp = atomic_load_ptr(&vp->v_mount);
3890 if (TAILQ_EMPTY(&mp->mnt_notify))
3894 mp->mnt_upper_pending++;
3895 KASSERT(mp->mnt_upper_pending > 0,
3896 ("%s: mnt_upper_pending %d", __func__, mp->mnt_upper_pending));
3897 TAILQ_FOREACH(ump, &mp->mnt_notify, mnt_upper_link) {
3900 case VFS_NOTIFY_UPPER_RECLAIM:
3901 VFS_RECLAIM_LOWERVP(ump->mp, vp);
3903 case VFS_NOTIFY_UPPER_UNLINK:
3904 VFS_UNLINK_LOWERVP(ump->mp, vp);
3907 KASSERT(0, ("invalid event %d", event));
3912 mp->mnt_upper_pending--;
3913 if ((mp->mnt_kern_flag & MNTK_UPPER_WAITER) != 0 &&
3914 mp->mnt_upper_pending == 0) {
3915 mp->mnt_kern_flag &= ~MNTK_UPPER_WAITER;
3916 wakeup(&mp->mnt_uppers);
3922 * vgone, with the vp interlock held.
3925 vgonel(struct vnode *vp)
3930 bool active, doinginact, oweinact;
3932 ASSERT_VOP_ELOCKED(vp, "vgonel");
3933 ASSERT_VI_LOCKED(vp, "vgonel");
3934 VNASSERT(vp->v_holdcnt, vp,
3935 ("vgonel: vp %p has no reference.", vp));
3936 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3940 * Don't vgonel if we're already doomed.
3942 if (VN_IS_DOOMED(vp))
3945 * Paired with freevnode.
3947 vn_seqc_write_begin_locked(vp);
3949 vn_irflag_set_locked(vp, VIRF_DOOMED);
3952 * Check to see if the vnode is in use. If so, we have to
3953 * call VOP_CLOSE() and VOP_INACTIVE().
3955 * It could be that VOP_INACTIVE() requested reclamation, in
3956 * which case we should avoid recursion, so check
3957 * VI_DOINGINACT. This is not precise but good enough.
3959 active = vp->v_usecount > 0;
3960 oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3961 doinginact = (vp->v_iflag & VI_DOINGINACT) != 0;
3964 * If we need to do inactive VI_OWEINACT will be set.
3966 if (vp->v_iflag & VI_DEFINACT) {
3967 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3968 vp->v_iflag &= ~VI_DEFINACT;
3971 VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
3974 cache_purge_vgone(vp);
3975 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3978 * If purging an active vnode, it must be closed and
3979 * deactivated before being reclaimed.
3982 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3985 if (oweinact || active) {
3988 oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3993 if (vp->v_type == VSOCK)
3994 vfs_unp_reclaim(vp);
3997 * Clean out any buffers associated with the vnode.
3998 * If the flush fails, just toss the buffers.
4001 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
4002 (void) vn_start_secondary_write(vp, &mp, V_WAIT);
4003 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
4004 while (vinvalbuf(vp, 0, 0, 0) != 0)
4008 BO_LOCK(&vp->v_bufobj);
4009 KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
4010 vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
4011 TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
4012 vp->v_bufobj.bo_clean.bv_cnt == 0,
4013 ("vp %p bufobj not invalidated", vp));
4016 * For VMIO bufobj, BO_DEAD is set later, or in
4017 * vm_object_terminate() after the object's page queue is
4020 object = vp->v_bufobj.bo_object;
4022 vp->v_bufobj.bo_flag |= BO_DEAD;
4023 BO_UNLOCK(&vp->v_bufobj);
4026 * Handle the VM part. Tmpfs handles v_object on its own (the
4027 * OBJT_VNODE check). Nullfs or other bypassing filesystems
4028 * should not touch the object borrowed from the lower vnode
4029 * (the handle check).
4031 if (object != NULL && object->type == OBJT_VNODE &&
4032 object->handle == vp)
4033 vnode_destroy_vobject(vp);
4036 * Reclaim the vnode.
4038 if (VOP_RECLAIM(vp))
4039 panic("vgone: cannot reclaim");
4041 vn_finished_secondary_write(mp);
4042 VNASSERT(vp->v_object == NULL, vp,
4043 ("vop_reclaim left v_object vp=%p", vp));
4045 * Clear the advisory locks and wake up waiting threads.
4047 (void)VOP_ADVLOCKPURGE(vp);
4050 * Delete from old mount point vnode list.
4054 * Done with purge, reset to the standard lock and invalidate
4058 vp->v_vnlock = &vp->v_lock;
4059 vp->v_op = &dead_vnodeops;
4064 * Print out a description of a vnode.
4066 static const char * const typename[] =
4067 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
4070 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0,
4071 "new hold count flag not added to vn_printf");
4074 vn_printf(struct vnode *vp, const char *fmt, ...)
4077 char buf[256], buf2[16];
4085 printf("%p: ", (void *)vp);
4086 printf("type %s\n", typename[vp->v_type]);
4087 holdcnt = atomic_load_int(&vp->v_holdcnt);
4088 printf(" usecount %d, writecount %d, refcount %d seqc users %d",
4089 vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS,
4091 switch (vp->v_type) {
4093 printf(" mountedhere %p\n", vp->v_mountedhere);
4096 printf(" rdev %p\n", vp->v_rdev);
4099 printf(" socket %p\n", vp->v_unpcb);
4102 printf(" fifoinfo %p\n", vp->v_fifoinfo);
4110 if (holdcnt & VHOLD_NO_SMR)
4111 strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf));
4112 printf(" hold count flags (%s)\n", buf + 1);
4116 irflag = vn_irflag_read(vp);
4117 if (irflag & VIRF_DOOMED)
4118 strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4119 if (irflag & VIRF_PGREAD)
4120 strlcat(buf, "|VIRF_PGREAD", sizeof(buf));
4121 if (irflag & VIRF_MOUNTPOINT)
4122 strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf));
4123 if (irflag & VIRF_TEXT_REF)
4124 strlcat(buf, "|VIRF_TEXT_REF", sizeof(buf));
4125 flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT | VIRF_TEXT_REF);
4127 snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4128 strlcat(buf, buf2, sizeof(buf));
4130 if (vp->v_vflag & VV_ROOT)
4131 strlcat(buf, "|VV_ROOT", sizeof(buf));
4132 if (vp->v_vflag & VV_ISTTY)
4133 strlcat(buf, "|VV_ISTTY", sizeof(buf));
4134 if (vp->v_vflag & VV_NOSYNC)
4135 strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4136 if (vp->v_vflag & VV_ETERNALDEV)
4137 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4138 if (vp->v_vflag & VV_CACHEDLABEL)
4139 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4140 if (vp->v_vflag & VV_VMSIZEVNLOCK)
4141 strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4142 if (vp->v_vflag & VV_COPYONWRITE)
4143 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4144 if (vp->v_vflag & VV_SYSTEM)
4145 strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4146 if (vp->v_vflag & VV_PROCDEP)
4147 strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4148 if (vp->v_vflag & VV_DELETED)
4149 strlcat(buf, "|VV_DELETED", sizeof(buf));
4150 if (vp->v_vflag & VV_MD)
4151 strlcat(buf, "|VV_MD", sizeof(buf));
4152 if (vp->v_vflag & VV_FORCEINSMQ)
4153 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4154 if (vp->v_vflag & VV_READLINK)
4155 strlcat(buf, "|VV_READLINK", sizeof(buf));
4156 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4157 VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM |
4158 VV_PROCDEP | VV_DELETED | VV_MD | VV_FORCEINSMQ | VV_READLINK);
4160 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4161 strlcat(buf, buf2, sizeof(buf));
4163 if (vp->v_iflag & VI_MOUNT)
4164 strlcat(buf, "|VI_MOUNT", sizeof(buf));
4165 if (vp->v_iflag & VI_DOINGINACT)
4166 strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4167 if (vp->v_iflag & VI_OWEINACT)
4168 strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4169 if (vp->v_iflag & VI_DEFINACT)
4170 strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4171 if (vp->v_iflag & VI_FOPENING)
4172 strlcat(buf, "|VI_FOPENING", sizeof(buf));
4173 flags = vp->v_iflag & ~(VI_MOUNT | VI_DOINGINACT |
4174 VI_OWEINACT | VI_DEFINACT | VI_FOPENING);
4176 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4177 strlcat(buf, buf2, sizeof(buf));
4179 if (vp->v_mflag & VMP_LAZYLIST)
4180 strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4181 flags = vp->v_mflag & ~(VMP_LAZYLIST);
4183 snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4184 strlcat(buf, buf2, sizeof(buf));
4186 printf(" flags (%s)", buf + 1);
4187 if (mtx_owned(VI_MTX(vp)))
4188 printf(" VI_LOCKed");
4190 if (vp->v_object != NULL)
4191 printf(" v_object %p ref %d pages %d "
4192 "cleanbuf %d dirtybuf %d\n",
4193 vp->v_object, vp->v_object->ref_count,
4194 vp->v_object->resident_page_count,
4195 vp->v_bufobj.bo_clean.bv_cnt,
4196 vp->v_bufobj.bo_dirty.bv_cnt);
4198 lockmgr_printinfo(vp->v_vnlock);
4199 if (vp->v_data != NULL)
4205 * List all of the locked vnodes in the system.
4206 * Called when debugging the kernel.
4208 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
4214 * Note: because this is DDB, we can't obey the locking semantics
4215 * for these structures, which means we could catch an inconsistent
4216 * state and dereference a nasty pointer. Not much to be done
4219 db_printf("Locked vnodes\n");
4220 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4221 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4222 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4223 vn_printf(vp, "vnode ");
4229 * Show details about the given vnode.
4231 DB_SHOW_COMMAND(vnode, db_show_vnode)
4237 vp = (struct vnode *)addr;
4238 vn_printf(vp, "vnode ");
4242 * Show details about the given mount point.
4244 DB_SHOW_COMMAND(mount, db_show_mount)
4255 /* No address given, print short info about all mount points. */
4256 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4257 db_printf("%p %s on %s (%s)\n", mp,
4258 mp->mnt_stat.f_mntfromname,
4259 mp->mnt_stat.f_mntonname,
4260 mp->mnt_stat.f_fstypename);
4264 db_printf("\nMore info: show mount <addr>\n");
4268 mp = (struct mount *)addr;
4269 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4270 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4273 mflags = mp->mnt_flag;
4274 #define MNT_FLAG(flag) do { \
4275 if (mflags & (flag)) { \
4276 if (buf[0] != '\0') \
4277 strlcat(buf, ", ", sizeof(buf)); \
4278 strlcat(buf, (#flag) + 4, sizeof(buf)); \
4279 mflags &= ~(flag); \
4282 MNT_FLAG(MNT_RDONLY);
4283 MNT_FLAG(MNT_SYNCHRONOUS);
4284 MNT_FLAG(MNT_NOEXEC);
4285 MNT_FLAG(MNT_NOSUID);
4286 MNT_FLAG(MNT_NFS4ACLS);
4287 MNT_FLAG(MNT_UNION);
4288 MNT_FLAG(MNT_ASYNC);
4289 MNT_FLAG(MNT_SUIDDIR);
4290 MNT_FLAG(MNT_SOFTDEP);
4291 MNT_FLAG(MNT_NOSYMFOLLOW);
4292 MNT_FLAG(MNT_GJOURNAL);
4293 MNT_FLAG(MNT_MULTILABEL);
4295 MNT_FLAG(MNT_NOATIME);
4296 MNT_FLAG(MNT_NOCLUSTERR);
4297 MNT_FLAG(MNT_NOCLUSTERW);
4299 MNT_FLAG(MNT_EXRDONLY);
4300 MNT_FLAG(MNT_EXPORTED);
4301 MNT_FLAG(MNT_DEFEXPORTED);
4302 MNT_FLAG(MNT_EXPORTANON);
4303 MNT_FLAG(MNT_EXKERB);
4304 MNT_FLAG(MNT_EXPUBLIC);
4305 MNT_FLAG(MNT_LOCAL);
4306 MNT_FLAG(MNT_QUOTA);
4307 MNT_FLAG(MNT_ROOTFS);
4309 MNT_FLAG(MNT_IGNORE);
4310 MNT_FLAG(MNT_UPDATE);
4311 MNT_FLAG(MNT_DELEXPORT);
4312 MNT_FLAG(MNT_RELOAD);
4313 MNT_FLAG(MNT_FORCE);
4314 MNT_FLAG(MNT_SNAPSHOT);
4315 MNT_FLAG(MNT_BYFSID);
4319 strlcat(buf, ", ", sizeof(buf));
4320 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4321 "0x%016jx", mflags);
4323 db_printf(" mnt_flag = %s\n", buf);
4326 flags = mp->mnt_kern_flag;
4327 #define MNT_KERN_FLAG(flag) do { \
4328 if (flags & (flag)) { \
4329 if (buf[0] != '\0') \
4330 strlcat(buf, ", ", sizeof(buf)); \
4331 strlcat(buf, (#flag) + 5, sizeof(buf)); \
4335 MNT_KERN_FLAG(MNTK_UNMOUNTF);
4336 MNT_KERN_FLAG(MNTK_ASYNC);
4337 MNT_KERN_FLAG(MNTK_SOFTDEP);
4338 MNT_KERN_FLAG(MNTK_DRAINING);
4339 MNT_KERN_FLAG(MNTK_REFEXPIRE);
4340 MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4341 MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4342 MNT_KERN_FLAG(MNTK_NO_IOPF);
4343 MNT_KERN_FLAG(MNTK_RECURSE);
4344 MNT_KERN_FLAG(MNTK_UPPER_WAITER);
4345 MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
4346 MNT_KERN_FLAG(MNTK_USES_BCACHE);
4347 MNT_KERN_FLAG(MNTK_FPLOOKUP);
4348 MNT_KERN_FLAG(MNTK_TASKQUEUE_WAITER);
4349 MNT_KERN_FLAG(MNTK_NOASYNC);
4350 MNT_KERN_FLAG(MNTK_UNMOUNT);
4351 MNT_KERN_FLAG(MNTK_MWAIT);
4352 MNT_KERN_FLAG(MNTK_SUSPEND);
4353 MNT_KERN_FLAG(MNTK_SUSPEND2);
4354 MNT_KERN_FLAG(MNTK_SUSPENDED);
4355 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4356 #undef MNT_KERN_FLAG
4359 strlcat(buf, ", ", sizeof(buf));
4360 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4363 db_printf(" mnt_kern_flag = %s\n", buf);
4365 db_printf(" mnt_opt = ");
4366 opt = TAILQ_FIRST(mp->mnt_opt);
4368 db_printf("%s", opt->name);
4369 opt = TAILQ_NEXT(opt, link);
4370 while (opt != NULL) {
4371 db_printf(", %s", opt->name);
4372 opt = TAILQ_NEXT(opt, link);
4378 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx "
4379 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4380 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4381 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4382 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4383 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4384 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4385 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4386 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4387 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4388 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4389 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4391 db_printf(" mnt_cred = { uid=%u ruid=%u",
4392 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4393 if (jailed(mp->mnt_cred))
4394 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4396 db_printf(" mnt_ref = %d (with %d in the struct)\n",
4397 vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4398 db_printf(" mnt_gen = %d\n", mp->mnt_gen);
4399 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4400 db_printf(" mnt_lazyvnodelistsize = %d\n",
4401 mp->mnt_lazyvnodelistsize);
4402 db_printf(" mnt_writeopcount = %d (with %d in the struct)\n",
4403 vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4404 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4405 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed);
4406 db_printf(" mnt_lockref = %d (with %d in the struct)\n",
4407 vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4408 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4409 db_printf(" mnt_secondary_accwrites = %d\n",
4410 mp->mnt_secondary_accwrites);
4411 db_printf(" mnt_gjprovider = %s\n",
4412 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4413 db_printf(" mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4415 db_printf("\n\nList of active vnodes\n");
4416 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4417 if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4418 vn_printf(vp, "vnode ");
4423 db_printf("\n\nList of inactive vnodes\n");
4424 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4425 if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4426 vn_printf(vp, "vnode ");
4435 * Fill in a struct xvfsconf based on a struct vfsconf.
4438 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4440 struct xvfsconf xvfsp;
4442 bzero(&xvfsp, sizeof(xvfsp));
4443 strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4444 xvfsp.vfc_typenum = vfsp->vfc_typenum;
4445 xvfsp.vfc_refcount = vfsp->vfc_refcount;
4446 xvfsp.vfc_flags = vfsp->vfc_flags;
4448 * These are unused in userland, we keep them
4449 * to not break binary compatibility.
4451 xvfsp.vfc_vfsops = NULL;
4452 xvfsp.vfc_next = NULL;
4453 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4456 #ifdef COMPAT_FREEBSD32
4458 uint32_t vfc_vfsops;
4459 char vfc_name[MFSNAMELEN];
4460 int32_t vfc_typenum;
4461 int32_t vfc_refcount;
4467 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4469 struct xvfsconf32 xvfsp;
4471 bzero(&xvfsp, sizeof(xvfsp));
4472 strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4473 xvfsp.vfc_typenum = vfsp->vfc_typenum;
4474 xvfsp.vfc_refcount = vfsp->vfc_refcount;
4475 xvfsp.vfc_flags = vfsp->vfc_flags;
4476 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4481 * Top level filesystem related information gathering.
4484 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4486 struct vfsconf *vfsp;
4491 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4492 #ifdef COMPAT_FREEBSD32
4493 if (req->flags & SCTL_MASK32)
4494 error = vfsconf2x32(req, vfsp);
4497 error = vfsconf2x(req, vfsp);
4505 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4506 CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4507 "S,xvfsconf", "List of all configured filesystems");
4509 #ifndef BURN_BRIDGES
4510 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4513 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4515 int *name = (int *)arg1 - 1; /* XXX */
4516 u_int namelen = arg2 + 1; /* XXX */
4517 struct vfsconf *vfsp;
4519 log(LOG_WARNING, "userland calling deprecated sysctl, "
4520 "please rebuild world\n");
4522 #if 1 || defined(COMPAT_PRELITE2)
4523 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4525 return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4529 case VFS_MAXTYPENUM:
4532 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4535 return (ENOTDIR); /* overloaded */
4537 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4538 if (vfsp->vfc_typenum == name[2])
4543 return (EOPNOTSUPP);
4544 #ifdef COMPAT_FREEBSD32
4545 if (req->flags & SCTL_MASK32)
4546 return (vfsconf2x32(req, vfsp));
4549 return (vfsconf2x(req, vfsp));
4551 return (EOPNOTSUPP);
4554 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4555 CTLFLAG_MPSAFE, vfs_sysctl,
4556 "Generic filesystem");
4558 #if 1 || defined(COMPAT_PRELITE2)
4561 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4564 struct vfsconf *vfsp;
4565 struct ovfsconf ovfs;
4568 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4569 bzero(&ovfs, sizeof(ovfs));
4570 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */
4571 strcpy(ovfs.vfc_name, vfsp->vfc_name);
4572 ovfs.vfc_index = vfsp->vfc_typenum;
4573 ovfs.vfc_refcount = vfsp->vfc_refcount;
4574 ovfs.vfc_flags = vfsp->vfc_flags;
4575 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4585 #endif /* 1 || COMPAT_PRELITE2 */
4586 #endif /* !BURN_BRIDGES */
4588 #define KINFO_VNODESLOP 10
4591 * Dump vnode list (via sysctl).
4595 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4603 * Stale numvnodes access is not fatal here.
4606 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4608 /* Make an estimate */
4609 return (SYSCTL_OUT(req, 0, len));
4611 error = sysctl_wire_old_buffer(req, 0);
4614 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4616 mtx_lock(&mountlist_mtx);
4617 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4618 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4621 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4625 xvn[n].xv_size = sizeof *xvn;
4626 xvn[n].xv_vnode = vp;
4627 xvn[n].xv_id = 0; /* XXX compat */
4628 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4630 XV_COPY(writecount);
4636 xvn[n].xv_flag = vp->v_vflag;
4638 switch (vp->v_type) {
4645 if (vp->v_rdev == NULL) {
4649 xvn[n].xv_dev = dev2udev(vp->v_rdev);
4652 xvn[n].xv_socket = vp->v_socket;
4655 xvn[n].xv_fifo = vp->v_fifoinfo;
4660 /* shouldn't happen? */
4668 mtx_lock(&mountlist_mtx);
4673 mtx_unlock(&mountlist_mtx);
4675 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4680 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4681 CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4686 unmount_or_warn(struct mount *mp)
4690 error = dounmount(mp, MNT_FORCE, curthread);
4692 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4696 printf("%d)\n", error);
4701 * Unmount all filesystems. The list is traversed in reverse order
4702 * of mounting to avoid dependencies.
4705 vfs_unmountall(void)
4707 struct mount *mp, *tmp;
4709 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4712 * Since this only runs when rebooting, it is not interlocked.
4714 TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4718 * Forcibly unmounting "/dev" before "/" would prevent clean
4719 * unmount of the latter.
4721 if (mp == rootdevmp)
4724 unmount_or_warn(mp);
4727 if (rootdevmp != NULL)
4728 unmount_or_warn(rootdevmp);
4732 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4735 ASSERT_VI_LOCKED(vp, __func__);
4736 VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4737 if ((vp->v_iflag & VI_OWEINACT) == 0) {
4741 if (vn_lock(vp, lkflags) == 0) {
4748 vdefer_inactive_unlocked(vp);
4752 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4755 return (vp->v_iflag & VI_DEFINACT);
4758 static void __noinline
4759 vfs_periodic_inactive(struct mount *mp, int flags)
4761 struct vnode *vp, *mvp;
4764 lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4765 if (flags != MNT_WAIT)
4766 lkflags |= LK_NOWAIT;
4768 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4769 if ((vp->v_iflag & VI_DEFINACT) == 0) {
4773 vp->v_iflag &= ~VI_DEFINACT;
4774 vfs_deferred_inactive(vp, lkflags);
4779 vfs_want_msync(struct vnode *vp)
4781 struct vm_object *obj;
4784 * This test may be performed without any locks held.
4785 * We rely on vm_object's type stability.
4787 if (vp->v_vflag & VV_NOSYNC)
4790 return (obj != NULL && vm_object_mightbedirty(obj));
4794 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4797 if (vp->v_vflag & VV_NOSYNC)
4799 if (vp->v_iflag & VI_DEFINACT)
4801 return (vfs_want_msync(vp));
4804 static void __noinline
4805 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4807 struct vnode *vp, *mvp;
4808 struct vm_object *obj;
4809 int lkflags, objflags;
4812 lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4813 if (flags != MNT_WAIT) {
4814 lkflags |= LK_NOWAIT;
4815 objflags = OBJPC_NOSYNC;
4817 objflags = OBJPC_SYNC;
4820 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4822 if (vp->v_iflag & VI_DEFINACT) {
4823 vp->v_iflag &= ~VI_DEFINACT;
4826 if (!vfs_want_msync(vp)) {
4828 vfs_deferred_inactive(vp, lkflags);
4833 if (vget(vp, lkflags) == 0) {
4835 if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4836 VM_OBJECT_WLOCK(obj);
4837 vm_object_page_clean(obj, 0, 0, objflags);
4838 VM_OBJECT_WUNLOCK(obj);
4845 vdefer_inactive_unlocked(vp);
4851 vfs_periodic(struct mount *mp, int flags)
4854 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4856 if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4857 vfs_periodic_inactive(mp, flags);
4859 vfs_periodic_msync_inactive(mp, flags);
4863 destroy_vpollinfo_free(struct vpollinfo *vi)
4866 knlist_destroy(&vi->vpi_selinfo.si_note);
4867 mtx_destroy(&vi->vpi_lock);
4868 free(vi, M_VNODEPOLL);
4872 destroy_vpollinfo(struct vpollinfo *vi)
4875 knlist_clear(&vi->vpi_selinfo.si_note, 1);
4876 seldrain(&vi->vpi_selinfo);
4877 destroy_vpollinfo_free(vi);
4881 * Initialize per-vnode helper structure to hold poll-related state.
4884 v_addpollinfo(struct vnode *vp)
4886 struct vpollinfo *vi;
4888 if (vp->v_pollinfo != NULL)
4890 vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO);
4891 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4892 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4893 vfs_knlunlock, vfs_knl_assert_lock);
4895 if (vp->v_pollinfo != NULL) {
4897 destroy_vpollinfo_free(vi);
4900 vp->v_pollinfo = vi;
4905 * Record a process's interest in events which might happen to
4906 * a vnode. Because poll uses the historic select-style interface
4907 * internally, this routine serves as both the ``check for any
4908 * pending events'' and the ``record my interest in future events''
4909 * functions. (These are done together, while the lock is held,
4910 * to avoid race conditions.)
4913 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4917 mtx_lock(&vp->v_pollinfo->vpi_lock);
4918 if (vp->v_pollinfo->vpi_revents & events) {
4920 * This leaves events we are not interested
4921 * in available for the other process which
4922 * which presumably had requested them
4923 * (otherwise they would never have been
4926 events &= vp->v_pollinfo->vpi_revents;
4927 vp->v_pollinfo->vpi_revents &= ~events;
4929 mtx_unlock(&vp->v_pollinfo->vpi_lock);
4932 vp->v_pollinfo->vpi_events |= events;
4933 selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4934 mtx_unlock(&vp->v_pollinfo->vpi_lock);
4939 * Routine to create and manage a filesystem syncer vnode.
4941 #define sync_close ((int (*)(struct vop_close_args *))nullop)
4942 static int sync_fsync(struct vop_fsync_args *);
4943 static int sync_inactive(struct vop_inactive_args *);
4944 static int sync_reclaim(struct vop_reclaim_args *);
4946 static struct vop_vector sync_vnodeops = {
4947 .vop_bypass = VOP_EOPNOTSUPP,
4948 .vop_close = sync_close, /* close */
4949 .vop_fsync = sync_fsync, /* fsync */
4950 .vop_inactive = sync_inactive, /* inactive */
4951 .vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4952 .vop_reclaim = sync_reclaim, /* reclaim */
4953 .vop_lock1 = vop_stdlock, /* lock */
4954 .vop_unlock = vop_stdunlock, /* unlock */
4955 .vop_islocked = vop_stdislocked, /* islocked */
4957 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4960 * Create a new filesystem syncer vnode for the specified mount point.
4963 vfs_allocate_syncvnode(struct mount *mp)
4967 static long start, incr, next;
4970 /* Allocate a new vnode */
4971 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4973 panic("vfs_allocate_syncvnode: getnewvnode() failed");
4975 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4976 vp->v_vflag |= VV_FORCEINSMQ;
4977 error = insmntque(vp, mp);
4979 panic("vfs_allocate_syncvnode: insmntque() failed");
4980 vp->v_vflag &= ~VV_FORCEINSMQ;
4983 * Place the vnode onto the syncer worklist. We attempt to
4984 * scatter them about on the list so that they will go off
4985 * at evenly distributed times even if all the filesystems
4986 * are mounted at once.
4989 if (next == 0 || next > syncer_maxdelay) {
4993 start = syncer_maxdelay / 2;
4994 incr = syncer_maxdelay;
5000 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
5001 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
5002 mtx_lock(&sync_mtx);
5004 if (mp->mnt_syncer == NULL) {
5005 mp->mnt_syncer = vp;
5008 mtx_unlock(&sync_mtx);
5011 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5018 vfs_deallocate_syncvnode(struct mount *mp)
5022 mtx_lock(&sync_mtx);
5023 vp = mp->mnt_syncer;
5025 mp->mnt_syncer = NULL;
5026 mtx_unlock(&sync_mtx);
5032 * Do a lazy sync of the filesystem.
5035 sync_fsync(struct vop_fsync_args *ap)
5037 struct vnode *syncvp = ap->a_vp;
5038 struct mount *mp = syncvp->v_mount;
5043 * We only need to do something if this is a lazy evaluation.
5045 if (ap->a_waitfor != MNT_LAZY)
5049 * Move ourselves to the back of the sync list.
5051 bo = &syncvp->v_bufobj;
5053 vn_syncer_add_to_worklist(bo, syncdelay);
5057 * Walk the list of vnodes pushing all that are dirty and
5058 * not already on the sync list.
5060 if (vfs_busy(mp, MBF_NOWAIT) != 0)
5062 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
5066 save = curthread_pflags_set(TDP_SYNCIO);
5068 * The filesystem at hand may be idle with free vnodes stored in the
5069 * batch. Return them instead of letting them stay there indefinitely.
5071 vfs_periodic(mp, MNT_NOWAIT);
5072 error = VFS_SYNC(mp, MNT_LAZY);
5073 curthread_pflags_restore(save);
5074 vn_finished_write(mp);
5080 * The syncer vnode is no referenced.
5083 sync_inactive(struct vop_inactive_args *ap)
5091 * The syncer vnode is no longer needed and is being decommissioned.
5093 * Modifications to the worklist must be protected by sync_mtx.
5096 sync_reclaim(struct vop_reclaim_args *ap)
5098 struct vnode *vp = ap->a_vp;
5103 mtx_lock(&sync_mtx);
5104 if (vp->v_mount->mnt_syncer == vp)
5105 vp->v_mount->mnt_syncer = NULL;
5106 if (bo->bo_flag & BO_ONWORKLST) {
5107 LIST_REMOVE(bo, bo_synclist);
5108 syncer_worklist_len--;
5110 bo->bo_flag &= ~BO_ONWORKLST;
5112 mtx_unlock(&sync_mtx);
5119 vn_need_pageq_flush(struct vnode *vp)
5121 struct vm_object *obj;
5124 return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5125 vm_object_mightbedirty(obj));
5129 * Check if vnode represents a disk device
5132 vn_isdisk_error(struct vnode *vp, int *errp)
5136 if (vp->v_type != VCHR) {
5142 if (vp->v_rdev == NULL)
5144 else if (vp->v_rdev->si_devsw == NULL)
5146 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5151 return (error == 0);
5155 vn_isdisk(struct vnode *vp)
5159 return (vn_isdisk_error(vp, &error));
5163 * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
5164 * the comment above cache_fplookup for details.
5167 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred)
5171 VFS_SMR_ASSERT_ENTERED();
5173 /* Check the owner. */
5174 if (cred->cr_uid == file_uid) {
5175 if (file_mode & S_IXUSR)
5180 /* Otherwise, check the groups (first match) */
5181 if (groupmember(file_gid, cred)) {
5182 if (file_mode & S_IXGRP)
5187 /* Otherwise, check everyone else. */
5188 if (file_mode & S_IXOTH)
5192 * Permission check failed, but it is possible denial will get overwritten
5193 * (e.g., when root is traversing through a 700 directory owned by someone
5196 * vaccess() calls priv_check_cred which in turn can descent into MAC
5197 * modules overriding this result. It's quite unclear what semantics
5198 * are allowed for them to operate, thus for safety we don't call them
5199 * from within the SMR section. This also means if any such modules
5200 * are present, we have to let the regular lookup decide.
5202 error = priv_check_cred_vfs_lookup_nomac(cred);
5208 * MAC modules present.
5219 * Common filesystem object access control check routine. Accepts a
5220 * vnode's type, "mode", uid and gid, requested access mode, and credentials.
5221 * Returns 0 on success, or an errno on failure.
5224 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5225 accmode_t accmode, struct ucred *cred)
5227 accmode_t dac_granted;
5228 accmode_t priv_granted;
5230 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5231 ("invalid bit in accmode"));
5232 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5233 ("VAPPEND without VWRITE"));
5236 * Look for a normal, non-privileged way to access the file/directory
5237 * as requested. If it exists, go with that.
5242 /* Check the owner. */
5243 if (cred->cr_uid == file_uid) {
5244 dac_granted |= VADMIN;
5245 if (file_mode & S_IXUSR)
5246 dac_granted |= VEXEC;
5247 if (file_mode & S_IRUSR)
5248 dac_granted |= VREAD;
5249 if (file_mode & S_IWUSR)
5250 dac_granted |= (VWRITE | VAPPEND);
5252 if ((accmode & dac_granted) == accmode)
5258 /* Otherwise, check the groups (first match) */
5259 if (groupmember(file_gid, cred)) {
5260 if (file_mode & S_IXGRP)
5261 dac_granted |= VEXEC;
5262 if (file_mode & S_IRGRP)
5263 dac_granted |= VREAD;
5264 if (file_mode & S_IWGRP)
5265 dac_granted |= (VWRITE | VAPPEND);
5267 if ((accmode & dac_granted) == accmode)
5273 /* Otherwise, check everyone else. */
5274 if (file_mode & S_IXOTH)
5275 dac_granted |= VEXEC;
5276 if (file_mode & S_IROTH)
5277 dac_granted |= VREAD;
5278 if (file_mode & S_IWOTH)
5279 dac_granted |= (VWRITE | VAPPEND);
5280 if ((accmode & dac_granted) == accmode)
5285 * Build a privilege mask to determine if the set of privileges
5286 * satisfies the requirements when combined with the granted mask
5287 * from above. For each privilege, if the privilege is required,
5288 * bitwise or the request type onto the priv_granted mask.
5294 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5295 * requests, instead of PRIV_VFS_EXEC.
5297 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5298 !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5299 priv_granted |= VEXEC;
5302 * Ensure that at least one execute bit is on. Otherwise,
5303 * a privileged user will always succeed, and we don't want
5304 * this to happen unless the file really is executable.
5306 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5307 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5308 !priv_check_cred(cred, PRIV_VFS_EXEC))
5309 priv_granted |= VEXEC;
5312 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5313 !priv_check_cred(cred, PRIV_VFS_READ))
5314 priv_granted |= VREAD;
5316 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5317 !priv_check_cred(cred, PRIV_VFS_WRITE))
5318 priv_granted |= (VWRITE | VAPPEND);
5320 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5321 !priv_check_cred(cred, PRIV_VFS_ADMIN))
5322 priv_granted |= VADMIN;
5324 if ((accmode & (priv_granted | dac_granted)) == accmode) {
5328 return ((accmode & VADMIN) ? EPERM : EACCES);
5332 * Credential check based on process requesting service, and per-attribute
5336 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5337 struct thread *td, accmode_t accmode)
5341 * Kernel-invoked always succeeds.
5347 * Do not allow privileged processes in jail to directly manipulate
5348 * system attributes.
5350 switch (attrnamespace) {
5351 case EXTATTR_NAMESPACE_SYSTEM:
5352 /* Potentially should be: return (EPERM); */
5353 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5354 case EXTATTR_NAMESPACE_USER:
5355 return (VOP_ACCESS(vp, accmode, cred, td));
5361 #ifdef DEBUG_VFS_LOCKS
5362 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */
5363 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5364 "Drop into debugger on lock violation");
5366 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */
5367 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5368 0, "Check for interlock across VOPs");
5370 int vfs_badlock_print = 1; /* Print lock violations. */
5371 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5372 0, "Print lock violations");
5374 int vfs_badlock_vnode = 1; /* Print vnode details on lock violations. */
5375 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5376 0, "Print vnode details on lock violations");
5379 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */
5380 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5381 &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5385 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5389 if (vfs_badlock_backtrace)
5392 if (vfs_badlock_vnode)
5393 vn_printf(vp, "vnode ");
5394 if (vfs_badlock_print)
5395 printf("%s: %p %s\n", str, (void *)vp, msg);
5396 if (vfs_badlock_ddb)
5397 kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5401 assert_vi_locked(struct vnode *vp, const char *str)
5404 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5405 vfs_badlock("interlock is not locked but should be", str, vp);
5409 assert_vi_unlocked(struct vnode *vp, const char *str)
5412 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5413 vfs_badlock("interlock is locked but should not be", str, vp);
5417 assert_vop_locked(struct vnode *vp, const char *str)
5421 if (KERNEL_PANICKED() || vp == NULL)
5424 locked = VOP_ISLOCKED(vp);
5425 if (locked == 0 || locked == LK_EXCLOTHER)
5426 vfs_badlock("is not locked but should be", str, vp);
5430 assert_vop_unlocked(struct vnode *vp, const char *str)
5432 if (KERNEL_PANICKED() || vp == NULL)
5435 if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5436 vfs_badlock("is locked but should not be", str, vp);
5440 assert_vop_elocked(struct vnode *vp, const char *str)
5442 if (KERNEL_PANICKED() || vp == NULL)
5445 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5446 vfs_badlock("is not exclusive locked but should be", str, vp);
5448 #endif /* DEBUG_VFS_LOCKS */
5451 vop_rename_fail(struct vop_rename_args *ap)
5454 if (ap->a_tvp != NULL)
5456 if (ap->a_tdvp == ap->a_tvp)
5465 vop_rename_pre(void *ap)
5467 struct vop_rename_args *a = ap;
5469 #ifdef DEBUG_VFS_LOCKS
5471 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5472 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5473 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5474 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5476 /* Check the source (from). */
5477 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5478 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5479 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5480 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5481 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5483 /* Check the target. */
5485 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5486 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5489 * It may be tempting to add vn_seqc_write_begin/end calls here and
5490 * in vop_rename_post but that's not going to work out since some
5491 * filesystems relookup vnodes mid-rename. This is probably a bug.
5493 * For now filesystems are expected to do the relevant calls after they
5494 * decide what vnodes to operate on.
5496 if (a->a_tdvp != a->a_fdvp)
5498 if (a->a_tvp != a->a_fvp)
5505 #ifdef DEBUG_VFS_LOCKS
5507 vop_fplookup_vexec_debugpre(void *ap __unused)
5510 VFS_SMR_ASSERT_ENTERED();
5514 vop_fplookup_vexec_debugpost(void *ap __unused, int rc __unused)
5517 VFS_SMR_ASSERT_ENTERED();
5521 vop_fplookup_symlink_debugpre(void *ap __unused)
5524 VFS_SMR_ASSERT_ENTERED();
5528 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused)
5531 VFS_SMR_ASSERT_ENTERED();
5535 vop_fsync_debugprepost(struct vnode *vp, const char *name)
5537 if (vp->v_type == VCHR)
5539 else if (MNT_EXTENDED_SHARED(vp->v_mount))
5540 ASSERT_VOP_LOCKED(vp, name);
5542 ASSERT_VOP_ELOCKED(vp, name);
5546 vop_fsync_debugpre(void *a)
5548 struct vop_fsync_args *ap;
5551 vop_fsync_debugprepost(ap->a_vp, "fsync");
5555 vop_fsync_debugpost(void *a, int rc __unused)
5557 struct vop_fsync_args *ap;
5560 vop_fsync_debugprepost(ap->a_vp, "fsync");
5564 vop_fdatasync_debugpre(void *a)
5566 struct vop_fdatasync_args *ap;
5569 vop_fsync_debugprepost(ap->a_vp, "fsync");
5573 vop_fdatasync_debugpost(void *a, int rc __unused)
5575 struct vop_fdatasync_args *ap;
5578 vop_fsync_debugprepost(ap->a_vp, "fsync");
5582 vop_strategy_debugpre(void *ap)
5584 struct vop_strategy_args *a;
5591 * Cluster ops lock their component buffers but not the IO container.
5593 if ((bp->b_flags & B_CLUSTER) != 0)
5596 if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5597 if (vfs_badlock_print)
5599 "VOP_STRATEGY: bp is not locked but should be\n");
5600 if (vfs_badlock_ddb)
5601 kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5606 vop_lock_debugpre(void *ap)
5608 struct vop_lock1_args *a = ap;
5610 if ((a->a_flags & LK_INTERLOCK) == 0)
5611 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5613 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5617 vop_lock_debugpost(void *ap, int rc)
5619 struct vop_lock1_args *a = ap;
5621 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5622 if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5623 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5627 vop_unlock_debugpre(void *ap)
5629 struct vop_unlock_args *a = ap;
5631 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5635 vop_need_inactive_debugpre(void *ap)
5637 struct vop_need_inactive_args *a = ap;
5639 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5643 vop_need_inactive_debugpost(void *ap, int rc)
5645 struct vop_need_inactive_args *a = ap;
5647 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5652 vop_create_pre(void *ap)
5654 struct vop_create_args *a;
5659 vn_seqc_write_begin(dvp);
5663 vop_create_post(void *ap, int rc)
5665 struct vop_create_args *a;
5670 vn_seqc_write_end(dvp);
5672 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5676 vop_whiteout_pre(void *ap)
5678 struct vop_whiteout_args *a;
5683 vn_seqc_write_begin(dvp);
5687 vop_whiteout_post(void *ap, int rc)
5689 struct vop_whiteout_args *a;
5694 vn_seqc_write_end(dvp);
5698 vop_deleteextattr_pre(void *ap)
5700 struct vop_deleteextattr_args *a;
5705 vn_seqc_write_begin(vp);
5709 vop_deleteextattr_post(void *ap, int rc)
5711 struct vop_deleteextattr_args *a;
5716 vn_seqc_write_end(vp);
5718 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5722 vop_link_pre(void *ap)
5724 struct vop_link_args *a;
5725 struct vnode *vp, *tdvp;
5730 vn_seqc_write_begin(vp);
5731 vn_seqc_write_begin(tdvp);
5735 vop_link_post(void *ap, int rc)
5737 struct vop_link_args *a;
5738 struct vnode *vp, *tdvp;
5743 vn_seqc_write_end(vp);
5744 vn_seqc_write_end(tdvp);
5746 VFS_KNOTE_LOCKED(vp, NOTE_LINK);
5747 VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE);
5752 vop_mkdir_pre(void *ap)
5754 struct vop_mkdir_args *a;
5759 vn_seqc_write_begin(dvp);
5763 vop_mkdir_post(void *ap, int rc)
5765 struct vop_mkdir_args *a;
5770 vn_seqc_write_end(dvp);
5772 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5775 #ifdef DEBUG_VFS_LOCKS
5777 vop_mkdir_debugpost(void *ap, int rc)
5779 struct vop_mkdir_args *a;
5783 cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp);
5788 vop_mknod_pre(void *ap)
5790 struct vop_mknod_args *a;
5795 vn_seqc_write_begin(dvp);
5799 vop_mknod_post(void *ap, int rc)
5801 struct vop_mknod_args *a;
5806 vn_seqc_write_end(dvp);
5808 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5812 vop_reclaim_post(void *ap, int rc)
5814 struct vop_reclaim_args *a;
5819 ASSERT_VOP_IN_SEQC(vp);
5821 VFS_KNOTE_LOCKED(vp, NOTE_REVOKE);
5825 vop_remove_pre(void *ap)
5827 struct vop_remove_args *a;
5828 struct vnode *dvp, *vp;
5833 vn_seqc_write_begin(dvp);
5834 vn_seqc_write_begin(vp);
5838 vop_remove_post(void *ap, int rc)
5840 struct vop_remove_args *a;
5841 struct vnode *dvp, *vp;
5846 vn_seqc_write_end(dvp);
5847 vn_seqc_write_end(vp);
5849 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5850 VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5855 vop_rename_post(void *ap, int rc)
5857 struct vop_rename_args *a = ap;
5862 if (a->a_fdvp == a->a_tdvp) {
5863 if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5865 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5866 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5868 hint |= NOTE_EXTEND;
5869 if (a->a_fvp->v_type == VDIR)
5871 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5873 if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5874 a->a_tvp->v_type == VDIR)
5876 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5879 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5881 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5883 if (a->a_tdvp != a->a_fdvp)
5885 if (a->a_tvp != a->a_fvp)
5893 vop_rmdir_pre(void *ap)
5895 struct vop_rmdir_args *a;
5896 struct vnode *dvp, *vp;
5901 vn_seqc_write_begin(dvp);
5902 vn_seqc_write_begin(vp);
5906 vop_rmdir_post(void *ap, int rc)
5908 struct vop_rmdir_args *a;
5909 struct vnode *dvp, *vp;
5914 vn_seqc_write_end(dvp);
5915 vn_seqc_write_end(vp);
5917 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5918 VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5923 vop_setattr_pre(void *ap)
5925 struct vop_setattr_args *a;
5930 vn_seqc_write_begin(vp);
5934 vop_setattr_post(void *ap, int rc)
5936 struct vop_setattr_args *a;
5941 vn_seqc_write_end(vp);
5943 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5947 vop_setacl_pre(void *ap)
5949 struct vop_setacl_args *a;
5954 vn_seqc_write_begin(vp);
5958 vop_setacl_post(void *ap, int rc __unused)
5960 struct vop_setacl_args *a;
5965 vn_seqc_write_end(vp);
5969 vop_setextattr_pre(void *ap)
5971 struct vop_setextattr_args *a;
5976 vn_seqc_write_begin(vp);
5980 vop_setextattr_post(void *ap, int rc)
5982 struct vop_setextattr_args *a;
5987 vn_seqc_write_end(vp);
5989 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5993 vop_symlink_pre(void *ap)
5995 struct vop_symlink_args *a;
6000 vn_seqc_write_begin(dvp);
6004 vop_symlink_post(void *ap, int rc)
6006 struct vop_symlink_args *a;
6011 vn_seqc_write_end(dvp);
6013 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6017 vop_open_post(void *ap, int rc)
6019 struct vop_open_args *a = ap;
6022 VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
6026 vop_close_post(void *ap, int rc)
6028 struct vop_close_args *a = ap;
6030 if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
6031 !VN_IS_DOOMED(a->a_vp))) {
6032 VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
6033 NOTE_CLOSE_WRITE : NOTE_CLOSE);
6038 vop_read_post(void *ap, int rc)
6040 struct vop_read_args *a = ap;
6043 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6047 vop_read_pgcache_post(void *ap, int rc)
6049 struct vop_read_pgcache_args *a = ap;
6052 VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ);
6056 vop_readdir_post(void *ap, int rc)
6058 struct vop_readdir_args *a = ap;
6061 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6064 static struct knlist fs_knlist;
6067 vfs_event_init(void *arg)
6069 knlist_init_mtx(&fs_knlist, NULL);
6071 /* XXX - correct order? */
6072 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
6075 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
6078 KNOTE_UNLOCKED(&fs_knlist, event);
6081 static int filt_fsattach(struct knote *kn);
6082 static void filt_fsdetach(struct knote *kn);
6083 static int filt_fsevent(struct knote *kn, long hint);
6085 struct filterops fs_filtops = {
6087 .f_attach = filt_fsattach,
6088 .f_detach = filt_fsdetach,
6089 .f_event = filt_fsevent
6093 filt_fsattach(struct knote *kn)
6096 kn->kn_flags |= EV_CLEAR;
6097 knlist_add(&fs_knlist, kn, 0);
6102 filt_fsdetach(struct knote *kn)
6105 knlist_remove(&fs_knlist, kn, 0);
6109 filt_fsevent(struct knote *kn, long hint)
6112 kn->kn_fflags |= kn->kn_sfflags & hint;
6114 return (kn->kn_fflags != 0);
6118 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
6124 error = SYSCTL_IN(req, &vc, sizeof(vc));
6127 if (vc.vc_vers != VFS_CTL_VERS1)
6129 mp = vfs_getvfs(&vc.vc_fsid);
6132 /* ensure that a specific sysctl goes to the right filesystem. */
6133 if (strcmp(vc.vc_fstypename, "*") != 0 &&
6134 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
6138 VCTLTOREQ(&vc, req);
6139 error = VFS_SYSCTL(mp, vc.vc_op, req);
6144 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
6145 NULL, 0, sysctl_vfs_ctl, "",
6149 * Function to initialize a va_filerev field sensibly.
6150 * XXX: Wouldn't a random number make a lot more sense ??
6153 init_va_filerev(void)
6158 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
6161 static int filt_vfsread(struct knote *kn, long hint);
6162 static int filt_vfswrite(struct knote *kn, long hint);
6163 static int filt_vfsvnode(struct knote *kn, long hint);
6164 static void filt_vfsdetach(struct knote *kn);
6165 static struct filterops vfsread_filtops = {
6167 .f_detach = filt_vfsdetach,
6168 .f_event = filt_vfsread
6170 static struct filterops vfswrite_filtops = {
6172 .f_detach = filt_vfsdetach,
6173 .f_event = filt_vfswrite
6175 static struct filterops vfsvnode_filtops = {
6177 .f_detach = filt_vfsdetach,
6178 .f_event = filt_vfsvnode
6182 vfs_knllock(void *arg)
6184 struct vnode *vp = arg;
6186 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
6190 vfs_knlunlock(void *arg)
6192 struct vnode *vp = arg;
6198 vfs_knl_assert_lock(void *arg, int what)
6200 #ifdef DEBUG_VFS_LOCKS
6201 struct vnode *vp = arg;
6203 if (what == LA_LOCKED)
6204 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
6206 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
6211 vfs_kqfilter(struct vop_kqfilter_args *ap)
6213 struct vnode *vp = ap->a_vp;
6214 struct knote *kn = ap->a_kn;
6217 KASSERT(vp->v_type != VFIFO || (kn->kn_filter != EVFILT_READ &&
6218 kn->kn_filter != EVFILT_WRITE),
6219 ("READ/WRITE filter on a FIFO leaked through"));
6220 switch (kn->kn_filter) {
6222 kn->kn_fop = &vfsread_filtops;
6225 kn->kn_fop = &vfswrite_filtops;
6228 kn->kn_fop = &vfsvnode_filtops;
6234 kn->kn_hook = (caddr_t)vp;
6237 if (vp->v_pollinfo == NULL)
6239 knl = &vp->v_pollinfo->vpi_selinfo.si_note;
6241 knlist_add(knl, kn, 0);
6247 * Detach knote from vnode
6250 filt_vfsdetach(struct knote *kn)
6252 struct vnode *vp = (struct vnode *)kn->kn_hook;
6254 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
6255 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
6261 filt_vfsread(struct knote *kn, long hint)
6263 struct vnode *vp = (struct vnode *)kn->kn_hook;
6268 * filesystem is gone, so set the EOF flag and schedule
6269 * the knote for deletion.
6271 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6273 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6278 if (VOP_GETATTR(vp, &va, curthread->td_ucred))
6282 kn->kn_data = va.va_size - kn->kn_fp->f_offset;
6283 res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
6290 filt_vfswrite(struct knote *kn, long hint)
6292 struct vnode *vp = (struct vnode *)kn->kn_hook;
6297 * filesystem is gone, so set the EOF flag and schedule
6298 * the knote for deletion.
6300 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
6301 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6309 filt_vfsvnode(struct knote *kn, long hint)
6311 struct vnode *vp = (struct vnode *)kn->kn_hook;
6315 if (kn->kn_sfflags & hint)
6316 kn->kn_fflags |= hint;
6317 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6318 kn->kn_flags |= EV_EOF;
6322 res = (kn->kn_fflags != 0);
6328 * Returns whether the directory is empty or not.
6329 * If it is empty, the return value is 0; otherwise
6330 * the return value is an error value (which may
6334 vfs_emptydir(struct vnode *vp)
6338 struct dirent *dirent, *dp, *endp;
6344 ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
6345 VNASSERT(vp->v_type == VDIR, vp, ("vp is not a directory"));
6347 dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
6348 iov.iov_base = dirent;
6349 iov.iov_len = sizeof(struct dirent);
6354 uio.uio_resid = sizeof(struct dirent);
6355 uio.uio_segflg = UIO_SYSSPACE;
6356 uio.uio_rw = UIO_READ;
6357 uio.uio_td = curthread;
6359 while (eof == 0 && error == 0) {
6360 error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
6364 endp = (void *)((uint8_t *)dirent +
6365 sizeof(struct dirent) - uio.uio_resid);
6366 for (dp = dirent; dp < endp;
6367 dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
6368 if (dp->d_type == DT_WHT)
6370 if (dp->d_namlen == 0)
6372 if (dp->d_type != DT_DIR &&
6373 dp->d_type != DT_UNKNOWN) {
6377 if (dp->d_namlen > 2) {
6381 if (dp->d_namlen == 1 &&
6382 dp->d_name[0] != '.') {
6386 if (dp->d_namlen == 2 &&
6387 dp->d_name[1] != '.') {
6391 uio.uio_resid = sizeof(struct dirent);
6394 free(dirent, M_TEMP);
6399 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6403 if (dp->d_reclen > ap->a_uio->uio_resid)
6404 return (ENAMETOOLONG);
6405 error = uiomove(dp, dp->d_reclen, ap->a_uio);
6407 if (ap->a_ncookies != NULL) {
6408 if (ap->a_cookies != NULL)
6409 free(ap->a_cookies, M_TEMP);
6410 ap->a_cookies = NULL;
6411 *ap->a_ncookies = 0;
6415 if (ap->a_ncookies == NULL)
6418 KASSERT(ap->a_cookies,
6419 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6421 *ap->a_cookies = realloc(*ap->a_cookies,
6422 (*ap->a_ncookies + 1) * sizeof(uint64_t), M_TEMP, M_WAITOK | M_ZERO);
6423 (*ap->a_cookies)[*ap->a_ncookies] = off;
6424 *ap->a_ncookies += 1;
6429 * The purpose of this routine is to remove granularity from accmode_t,
6430 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6431 * VADMIN and VAPPEND.
6433 * If it returns 0, the caller is supposed to continue with the usual
6434 * access checks using 'accmode' as modified by this routine. If it
6435 * returns nonzero value, the caller is supposed to return that value
6438 * Note that after this routine runs, accmode may be zero.
6441 vfs_unixify_accmode(accmode_t *accmode)
6444 * There is no way to specify explicit "deny" rule using
6445 * file mode or POSIX.1e ACLs.
6447 if (*accmode & VEXPLICIT_DENY) {
6453 * None of these can be translated into usual access bits.
6454 * Also, the common case for NFSv4 ACLs is to not contain
6455 * either of these bits. Caller should check for VWRITE
6456 * on the containing directory instead.
6458 if (*accmode & (VDELETE_CHILD | VDELETE))
6461 if (*accmode & VADMIN_PERMS) {
6462 *accmode &= ~VADMIN_PERMS;
6467 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6468 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6470 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6476 * Clear out a doomed vnode (if any) and replace it with a new one as long
6477 * as the fs is not being unmounted. Return the root vnode to the caller.
6479 static int __noinline
6480 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6486 if (mp->mnt_rootvnode != NULL) {
6488 vp = mp->mnt_rootvnode;
6490 if (!VN_IS_DOOMED(vp)) {
6493 error = vn_lock(vp, flags);
6502 * Clear the old one.
6504 mp->mnt_rootvnode = NULL;
6508 vfs_op_barrier_wait(mp);
6512 error = VFS_CACHEDROOT(mp, flags, vpp);
6515 if (mp->mnt_vfs_ops == 0) {
6517 if (mp->mnt_vfs_ops != 0) {
6521 if (mp->mnt_rootvnode == NULL) {
6523 mp->mnt_rootvnode = *vpp;
6525 if (mp->mnt_rootvnode != *vpp) {
6526 if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6527 panic("%s: mismatch between vnode returned "
6528 " by VFS_CACHEDROOT and the one cached "
6530 __func__, *vpp, mp->mnt_rootvnode);
6540 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6542 struct mount_pcpu *mpcpu;
6546 if (!vfs_op_thread_enter(mp, mpcpu))
6547 return (vfs_cache_root_fallback(mp, flags, vpp));
6548 vp = atomic_load_ptr(&mp->mnt_rootvnode);
6549 if (vp == NULL || VN_IS_DOOMED(vp)) {
6550 vfs_op_thread_exit(mp, mpcpu);
6551 return (vfs_cache_root_fallback(mp, flags, vpp));
6554 vfs_op_thread_exit(mp, mpcpu);
6555 error = vn_lock(vp, flags);
6558 return (vfs_cache_root_fallback(mp, flags, vpp));
6565 vfs_cache_root_clear(struct mount *mp)
6570 * ops > 0 guarantees there is nobody who can see this vnode
6572 MPASS(mp->mnt_vfs_ops > 0);
6573 vp = mp->mnt_rootvnode;
6575 vn_seqc_write_begin(vp);
6576 mp->mnt_rootvnode = NULL;
6581 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6584 MPASS(mp->mnt_vfs_ops > 0);
6586 mp->mnt_rootvnode = vp;
6590 * These are helper functions for filesystems to traverse all
6591 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6593 * This interface replaces MNT_VNODE_FOREACH.
6597 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6602 kern_yield(PRI_USER);
6604 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6605 for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6606 vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6607 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6608 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6611 if (VN_IS_DOOMED(vp)) {
6618 __mnt_vnode_markerfree_all(mvp, mp);
6619 /* MNT_IUNLOCK(mp); -- done in above function */
6620 mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6623 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6624 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6630 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6634 *mvp = vn_alloc_marker(mp);
6638 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6639 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6640 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6643 if (VN_IS_DOOMED(vp)) {
6652 vn_free_marker(*mvp);
6656 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6662 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6670 mtx_assert(MNT_MTX(mp), MA_OWNED);
6672 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6673 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6676 vn_free_marker(*mvp);
6681 * These are helper functions for filesystems to traverse their
6682 * lazy vnodes. See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6685 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6688 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6693 vn_free_marker(*mvp);
6698 * Relock the mp mount vnode list lock with the vp vnode interlock in the
6699 * conventional lock order during mnt_vnode_next_lazy iteration.
6701 * On entry, the mount vnode list lock is held and the vnode interlock is not.
6702 * The list lock is dropped and reacquired. On success, both locks are held.
6703 * On failure, the mount vnode list lock is held but the vnode interlock is
6704 * not, and the procedure may have yielded.
6707 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6711 VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6712 TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6713 ("%s: bad marker", __func__));
6714 VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6715 ("%s: inappropriate vnode", __func__));
6716 ASSERT_VI_UNLOCKED(vp, __func__);
6717 mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6719 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6720 TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6723 * Note we may be racing against vdrop which transitioned the hold
6724 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6725 * if we are the only user after we get the interlock we will just
6729 mtx_unlock(&mp->mnt_listmtx);
6731 if (VN_IS_DOOMED(vp)) {
6732 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6735 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6737 * There is nothing to do if we are the last user.
6739 if (!refcount_release_if_not_last(&vp->v_holdcnt))
6741 mtx_lock(&mp->mnt_listmtx);
6746 mtx_lock(&mp->mnt_listmtx);
6750 static struct vnode *
6751 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6756 mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6757 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6759 vp = TAILQ_NEXT(*mvp, v_lazylist);
6760 while (vp != NULL) {
6761 if (vp->v_type == VMARKER) {
6762 vp = TAILQ_NEXT(vp, v_lazylist);
6766 * See if we want to process the vnode. Note we may encounter a
6767 * long string of vnodes we don't care about and hog the list
6768 * as a result. Check for it and requeue the marker.
6770 VNPASS(!VN_IS_DOOMED(vp), vp);
6771 if (!cb(vp, cbarg)) {
6772 if (!should_yield()) {
6773 vp = TAILQ_NEXT(vp, v_lazylist);
6776 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6778 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6780 mtx_unlock(&mp->mnt_listmtx);
6781 kern_yield(PRI_USER);
6782 mtx_lock(&mp->mnt_listmtx);
6786 * Try-lock because this is the wrong lock order.
6788 if (!VI_TRYLOCK(vp) &&
6789 !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6791 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6792 KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6793 ("alien vnode on the lazy list %p %p", vp, mp));
6794 VNPASS(vp->v_mount == mp, vp);
6795 VNPASS(!VN_IS_DOOMED(vp), vp);
6798 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6800 /* Check if we are done */
6802 mtx_unlock(&mp->mnt_listmtx);
6803 mnt_vnode_markerfree_lazy(mvp, mp);
6806 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6807 mtx_unlock(&mp->mnt_listmtx);
6808 ASSERT_VI_LOCKED(vp, "lazy iter");
6813 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6818 kern_yield(PRI_USER);
6819 mtx_lock(&mp->mnt_listmtx);
6820 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6824 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6829 if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6832 *mvp = vn_alloc_marker(mp);
6837 mtx_lock(&mp->mnt_listmtx);
6838 vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6840 mtx_unlock(&mp->mnt_listmtx);
6841 mnt_vnode_markerfree_lazy(mvp, mp);
6844 TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6845 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6849 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6855 mtx_lock(&mp->mnt_listmtx);
6856 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6857 mtx_unlock(&mp->mnt_listmtx);
6858 mnt_vnode_markerfree_lazy(mvp, mp);
6862 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6865 if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6866 cnp->cn_flags &= ~NOEXECCHECK;
6870 return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, curthread));
6874 * Do not use this variant unless you have means other than the hold count
6875 * to prevent the vnode from getting freed.
6878 vn_seqc_write_begin_locked(struct vnode *vp)
6881 ASSERT_VI_LOCKED(vp, __func__);
6882 VNPASS(vp->v_holdcnt > 0, vp);
6883 VNPASS(vp->v_seqc_users >= 0, vp);
6885 if (vp->v_seqc_users == 1)
6886 seqc_sleepable_write_begin(&vp->v_seqc);
6890 vn_seqc_write_begin(struct vnode *vp)
6894 vn_seqc_write_begin_locked(vp);
6899 vn_seqc_write_end_locked(struct vnode *vp)
6902 ASSERT_VI_LOCKED(vp, __func__);
6903 VNPASS(vp->v_seqc_users > 0, vp);
6905 if (vp->v_seqc_users == 0)
6906 seqc_sleepable_write_end(&vp->v_seqc);
6910 vn_seqc_write_end(struct vnode *vp)
6914 vn_seqc_write_end_locked(vp);
6919 * Special case handling for allocating and freeing vnodes.
6921 * The counter remains unchanged on free so that a doomed vnode will
6922 * keep testing as in modify as long as it is accessible with SMR.
6925 vn_seqc_init(struct vnode *vp)
6929 vp->v_seqc_users = 0;
6933 vn_seqc_write_end_free(struct vnode *vp)
6936 VNPASS(seqc_in_modify(vp->v_seqc), vp);
6937 VNPASS(vp->v_seqc_users == 1, vp);
6941 vn_irflag_set_locked(struct vnode *vp, short toset)
6945 ASSERT_VI_LOCKED(vp, __func__);
6946 flags = vn_irflag_read(vp);
6947 VNASSERT((flags & toset) == 0, vp,
6948 ("%s: some of the passed flags already set (have %d, passed %d)\n",
6949 __func__, flags, toset));
6950 atomic_store_short(&vp->v_irflag, flags | toset);
6954 vn_irflag_set(struct vnode *vp, short toset)
6958 vn_irflag_set_locked(vp, toset);
6963 vn_irflag_set_cond_locked(struct vnode *vp, short toset)
6967 ASSERT_VI_LOCKED(vp, __func__);
6968 flags = vn_irflag_read(vp);
6969 atomic_store_short(&vp->v_irflag, flags | toset);
6973 vn_irflag_set_cond(struct vnode *vp, short toset)
6977 vn_irflag_set_cond_locked(vp, toset);
6982 vn_irflag_unset_locked(struct vnode *vp, short tounset)
6986 ASSERT_VI_LOCKED(vp, __func__);
6987 flags = vn_irflag_read(vp);
6988 VNASSERT((flags & tounset) == tounset, vp,
6989 ("%s: some of the passed flags not set (have %d, passed %d)\n",
6990 __func__, flags, tounset));
6991 atomic_store_short(&vp->v_irflag, flags & ~tounset);
6995 vn_irflag_unset(struct vnode *vp, short tounset)
6999 vn_irflag_unset_locked(vp, tounset);