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;
1817 if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1818 vp->v_vflag |= VV_NOKNOTE;
1822 * For the filesystems which do not use vfs_hash_insert(),
1823 * still initialize v_hash to have vfs_hash_index() useful.
1824 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1827 vp->v_hash = (uintptr_t)vp >> vnsz2log;
1834 getnewvnode_reserve(void)
1839 MPASS(td->td_vp_reserved == NULL);
1840 td->td_vp_reserved = vn_alloc(NULL);
1844 getnewvnode_drop_reserve(void)
1849 if (td->td_vp_reserved != NULL) {
1850 vn_free(td->td_vp_reserved);
1851 td->td_vp_reserved = NULL;
1855 static void __noinline
1856 freevnode(struct vnode *vp)
1861 * The vnode has been marked for destruction, so free it.
1863 * The vnode will be returned to the zone where it will
1864 * normally remain until it is needed for another vnode. We
1865 * need to cleanup (or verify that the cleanup has already
1866 * been done) any residual data left from its current use
1867 * so as not to contaminate the freshly allocated vnode.
1869 CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1871 * Paired with vgone.
1873 vn_seqc_write_end_free(vp);
1876 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1877 VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp);
1878 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1879 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1880 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1881 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1882 VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1883 ("clean blk trie not empty"));
1884 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1885 VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1886 ("dirty blk trie not empty"));
1887 VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1888 VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1889 VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1890 VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1891 ("Dangling rangelock waiters"));
1892 VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp,
1893 ("Leaked inactivation"));
1896 mac_vnode_destroy(vp);
1898 if (vp->v_pollinfo != NULL) {
1899 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1900 destroy_vpollinfo(vp->v_pollinfo);
1902 vp->v_pollinfo = NULL;
1904 vp->v_mountedhere = NULL;
1907 vp->v_fifoinfo = NULL;
1915 * Delete from old mount point vnode list, if on one.
1918 delmntque(struct vnode *vp)
1922 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
1931 VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1932 ("bad mount point vnode list size"));
1933 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1934 mp->mnt_nvnodelistsize--;
1940 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1944 vp->v_op = &dead_vnodeops;
1950 * Insert into list of vnodes for the new mount point, if available.
1953 insmntque1(struct vnode *vp, struct mount *mp,
1954 void (*dtr)(struct vnode *, void *), void *dtr_arg)
1957 KASSERT(vp->v_mount == NULL,
1958 ("insmntque: vnode already on per mount vnode list"));
1959 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1960 ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1963 * We acquire the vnode interlock early to ensure that the
1964 * vnode cannot be recycled by another process releasing a
1965 * holdcnt on it before we get it on both the vnode list
1966 * and the active vnode list. The mount mutex protects only
1967 * manipulation of the vnode list and the vnode freelist
1968 * mutex protects only manipulation of the active vnode list.
1969 * Hence the need to hold the vnode interlock throughout.
1973 if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1974 ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1975 mp->mnt_nvnodelistsize == 0)) &&
1976 (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1985 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1986 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1987 ("neg mount point vnode list size"));
1988 mp->mnt_nvnodelistsize++;
1995 insmntque(struct vnode *vp, struct mount *mp)
1998 return (insmntque1(vp, mp, insmntque_stddtr, NULL));
2002 * Flush out and invalidate all buffers associated with a bufobj
2003 * Called with the underlying object locked.
2006 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
2011 if (flags & V_SAVE) {
2012 error = bufobj_wwait(bo, slpflag, slptimeo);
2017 if (bo->bo_dirty.bv_cnt > 0) {
2020 error = BO_SYNC(bo, MNT_WAIT);
2021 } while (error == ERELOOKUP);
2025 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) {
2032 * If you alter this loop please notice that interlock is dropped and
2033 * reacquired in flushbuflist. Special care is needed to ensure that
2034 * no race conditions occur from this.
2037 error = flushbuflist(&bo->bo_clean,
2038 flags, bo, slpflag, slptimeo);
2039 if (error == 0 && !(flags & V_CLEANONLY))
2040 error = flushbuflist(&bo->bo_dirty,
2041 flags, bo, slpflag, slptimeo);
2042 if (error != 0 && error != EAGAIN) {
2046 } while (error != 0);
2049 * Wait for I/O to complete. XXX needs cleaning up. The vnode can
2050 * have write I/O in-progress but if there is a VM object then the
2051 * VM object can also have read-I/O in-progress.
2054 bufobj_wwait(bo, 0, 0);
2055 if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
2057 vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
2060 } while (bo->bo_numoutput > 0);
2064 * Destroy the copy in the VM cache, too.
2066 if (bo->bo_object != NULL &&
2067 (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
2068 VM_OBJECT_WLOCK(bo->bo_object);
2069 vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
2070 OBJPR_CLEANONLY : 0);
2071 VM_OBJECT_WUNLOCK(bo->bo_object);
2076 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
2077 V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
2078 bo->bo_clean.bv_cnt > 0))
2079 panic("vinvalbuf: flush failed");
2080 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
2081 bo->bo_dirty.bv_cnt > 0)
2082 panic("vinvalbuf: flush dirty failed");
2089 * Flush out and invalidate all buffers associated with a vnode.
2090 * Called with the underlying object locked.
2093 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
2096 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2097 ASSERT_VOP_LOCKED(vp, "vinvalbuf");
2098 if (vp->v_object != NULL && vp->v_object->handle != vp)
2100 return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
2104 * Flush out buffers on the specified list.
2108 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2111 struct buf *bp, *nbp;
2116 ASSERT_BO_WLOCKED(bo);
2119 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2121 * If we are flushing both V_NORMAL and V_ALT buffers then
2122 * do not skip any buffers. If we are flushing only V_NORMAL
2123 * buffers then skip buffers marked as BX_ALTDATA. If we are
2124 * flushing only V_ALT buffers then skip buffers not marked
2127 if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2128 (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2129 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2133 lblkno = nbp->b_lblkno;
2134 xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2137 error = BUF_TIMELOCK(bp,
2138 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2139 "flushbuf", slpflag, slptimeo);
2142 return (error != ENOLCK ? error : EAGAIN);
2144 KASSERT(bp->b_bufobj == bo,
2145 ("bp %p wrong b_bufobj %p should be %p",
2146 bp, bp->b_bufobj, bo));
2148 * XXX Since there are no node locks for NFS, I
2149 * believe there is a slight chance that a delayed
2150 * write will occur while sleeping just above, so
2153 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2156 bp->b_flags |= B_ASYNC;
2159 return (EAGAIN); /* XXX: why not loop ? */
2162 bp->b_flags |= (B_INVAL | B_RELBUF);
2163 bp->b_flags &= ~B_ASYNC;
2168 nbp = gbincore(bo, lblkno);
2169 if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2171 break; /* nbp invalid */
2177 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2183 ASSERT_BO_LOCKED(bo);
2185 for (lblkno = startn;;) {
2187 bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2188 if (bp == NULL || bp->b_lblkno >= endn ||
2189 bp->b_lblkno < startn)
2191 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2192 LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2195 if (error == ENOLCK)
2199 KASSERT(bp->b_bufobj == bo,
2200 ("bp %p wrong b_bufobj %p should be %p",
2201 bp, bp->b_bufobj, bo));
2202 lblkno = bp->b_lblkno + 1;
2203 if ((bp->b_flags & B_MANAGED) == 0)
2205 bp->b_flags |= B_RELBUF;
2207 * In the VMIO case, use the B_NOREUSE flag to hint that the
2208 * pages backing each buffer in the range are unlikely to be
2209 * reused. Dirty buffers will have the hint applied once
2210 * they've been written.
2212 if ((bp->b_flags & B_VMIO) != 0)
2213 bp->b_flags |= B_NOREUSE;
2221 * Truncate a file's buffer and pages to a specified length. This
2222 * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2226 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2228 struct buf *bp, *nbp;
2232 CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2233 vp, blksize, (uintmax_t)length);
2236 * Round up to the *next* lbn.
2238 startlbn = howmany(length, blksize);
2240 ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2246 while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2251 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2252 if (bp->b_lblkno > 0)
2255 * Since we hold the vnode lock this should only
2256 * fail if we're racing with the buf daemon.
2259 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2260 BO_LOCKPTR(bo)) == ENOLCK)
2261 goto restart_unlocked;
2263 VNASSERT((bp->b_flags & B_DELWRI), vp,
2264 ("buf(%p) on dirty queue without DELWRI", bp));
2273 bufobj_wwait(bo, 0, 0);
2275 vnode_pager_setsize(vp, length);
2281 * Invalidate the cached pages of a file's buffer within the range of block
2282 * numbers [startlbn, endlbn).
2285 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2291 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2293 start = blksize * startlbn;
2294 end = blksize * endlbn;
2298 MPASS(blksize == bo->bo_bsize);
2300 while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2304 vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2308 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2309 daddr_t startlbn, daddr_t endlbn)
2311 struct buf *bp, *nbp;
2314 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2315 ASSERT_BO_LOCKED(bo);
2319 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2320 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2323 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2324 BO_LOCKPTR(bo)) == ENOLCK) {
2330 bp->b_flags |= B_INVAL | B_RELBUF;
2331 bp->b_flags &= ~B_ASYNC;
2337 (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2339 (nbp->b_flags & B_DELWRI) != 0))
2343 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2344 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2347 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2348 BO_LOCKPTR(bo)) == ENOLCK) {
2353 bp->b_flags |= B_INVAL | B_RELBUF;
2354 bp->b_flags &= ~B_ASYNC;
2360 (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2361 (nbp->b_vp != vp) ||
2362 (nbp->b_flags & B_DELWRI) == 0))
2370 buf_vlist_remove(struct buf *bp)
2375 flags = bp->b_xflags;
2377 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2378 ASSERT_BO_WLOCKED(bp->b_bufobj);
2379 KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 &&
2380 (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN),
2381 ("%s: buffer %p has invalid queue state", __func__, bp));
2383 if ((flags & BX_VNDIRTY) != 0)
2384 bv = &bp->b_bufobj->bo_dirty;
2386 bv = &bp->b_bufobj->bo_clean;
2387 BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2388 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2390 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2394 * Add the buffer to the sorted clean or dirty block list.
2396 * NOTE: xflags is passed as a constant, optimizing this inline function!
2399 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2405 ASSERT_BO_WLOCKED(bo);
2406 KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2407 ("buf_vlist_add: bo %p does not allow bufs", bo));
2408 KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2409 ("dead bo %p", bo));
2410 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2411 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2412 bp->b_xflags |= xflags;
2413 if (xflags & BX_VNDIRTY)
2419 * Keep the list ordered. Optimize empty list insertion. Assume
2420 * we tend to grow at the tail so lookup_le should usually be cheaper
2423 if (bv->bv_cnt == 0 ||
2424 bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2425 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2426 else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2427 TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2429 TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2430 error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2432 panic("buf_vlist_add: Preallocated nodes insufficient.");
2437 * Look up a buffer using the buffer tries.
2440 gbincore(struct bufobj *bo, daddr_t lblkno)
2444 ASSERT_BO_LOCKED(bo);
2445 bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2448 return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno));
2452 * Look up a buf using the buffer tries, without the bufobj lock. This relies
2453 * on SMR for safe lookup, and bufs being in a no-free zone to provide type
2454 * stability of the result. Like other lockless lookups, the found buf may
2455 * already be invalid by the time this function returns.
2458 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno)
2462 ASSERT_BO_UNLOCKED(bo);
2463 bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno);
2466 return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno));
2470 * Associate a buffer with a vnode.
2473 bgetvp(struct vnode *vp, struct buf *bp)
2478 ASSERT_BO_WLOCKED(bo);
2479 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2481 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2482 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2483 ("bgetvp: bp already attached! %p", bp));
2489 * Insert onto list for new vnode.
2491 buf_vlist_add(bp, bo, BX_VNCLEAN);
2495 * Disassociate a buffer from a vnode.
2498 brelvp(struct buf *bp)
2503 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2504 KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2507 * Delete from old vnode list, if on one.
2509 vp = bp->b_vp; /* XXX */
2512 buf_vlist_remove(bp);
2513 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2514 bo->bo_flag &= ~BO_ONWORKLST;
2515 mtx_lock(&sync_mtx);
2516 LIST_REMOVE(bo, bo_synclist);
2517 syncer_worklist_len--;
2518 mtx_unlock(&sync_mtx);
2521 bp->b_bufobj = NULL;
2527 * Add an item to the syncer work queue.
2530 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2534 ASSERT_BO_WLOCKED(bo);
2536 mtx_lock(&sync_mtx);
2537 if (bo->bo_flag & BO_ONWORKLST)
2538 LIST_REMOVE(bo, bo_synclist);
2540 bo->bo_flag |= BO_ONWORKLST;
2541 syncer_worklist_len++;
2544 if (delay > syncer_maxdelay - 2)
2545 delay = syncer_maxdelay - 2;
2546 slot = (syncer_delayno + delay) & syncer_mask;
2548 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2549 mtx_unlock(&sync_mtx);
2553 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2557 mtx_lock(&sync_mtx);
2558 len = syncer_worklist_len - sync_vnode_count;
2559 mtx_unlock(&sync_mtx);
2560 error = SYSCTL_OUT(req, &len, sizeof(len));
2564 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2565 CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2566 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2568 static struct proc *updateproc;
2569 static void sched_sync(void);
2570 static struct kproc_desc up_kp = {
2575 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2578 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2583 *bo = LIST_FIRST(slp);
2587 if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2590 * We use vhold in case the vnode does not
2591 * successfully sync. vhold prevents the vnode from
2592 * going away when we unlock the sync_mtx so that
2593 * we can acquire the vnode interlock.
2596 mtx_unlock(&sync_mtx);
2598 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2600 mtx_lock(&sync_mtx);
2601 return (*bo == LIST_FIRST(slp));
2603 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2604 (void) VOP_FSYNC(vp, MNT_LAZY, td);
2606 vn_finished_write(mp);
2608 if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2610 * Put us back on the worklist. The worklist
2611 * routine will remove us from our current
2612 * position and then add us back in at a later
2615 vn_syncer_add_to_worklist(*bo, syncdelay);
2619 mtx_lock(&sync_mtx);
2623 static int first_printf = 1;
2626 * System filesystem synchronizer daemon.
2631 struct synclist *next, *slp;
2634 struct thread *td = curthread;
2636 int net_worklist_len;
2637 int syncer_final_iter;
2641 syncer_final_iter = 0;
2642 syncer_state = SYNCER_RUNNING;
2643 starttime = time_uptime;
2644 td->td_pflags |= TDP_NORUNNINGBUF;
2646 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2649 mtx_lock(&sync_mtx);
2651 if (syncer_state == SYNCER_FINAL_DELAY &&
2652 syncer_final_iter == 0) {
2653 mtx_unlock(&sync_mtx);
2654 kproc_suspend_check(td->td_proc);
2655 mtx_lock(&sync_mtx);
2657 net_worklist_len = syncer_worklist_len - sync_vnode_count;
2658 if (syncer_state != SYNCER_RUNNING &&
2659 starttime != time_uptime) {
2661 printf("\nSyncing disks, vnodes remaining... ");
2664 printf("%d ", net_worklist_len);
2666 starttime = time_uptime;
2669 * Push files whose dirty time has expired. Be careful
2670 * of interrupt race on slp queue.
2672 * Skip over empty worklist slots when shutting down.
2675 slp = &syncer_workitem_pending[syncer_delayno];
2676 syncer_delayno += 1;
2677 if (syncer_delayno == syncer_maxdelay)
2679 next = &syncer_workitem_pending[syncer_delayno];
2681 * If the worklist has wrapped since the
2682 * it was emptied of all but syncer vnodes,
2683 * switch to the FINAL_DELAY state and run
2684 * for one more second.
2686 if (syncer_state == SYNCER_SHUTTING_DOWN &&
2687 net_worklist_len == 0 &&
2688 last_work_seen == syncer_delayno) {
2689 syncer_state = SYNCER_FINAL_DELAY;
2690 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2692 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2693 syncer_worklist_len > 0);
2696 * Keep track of the last time there was anything
2697 * on the worklist other than syncer vnodes.
2698 * Return to the SHUTTING_DOWN state if any
2701 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2702 last_work_seen = syncer_delayno;
2703 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2704 syncer_state = SYNCER_SHUTTING_DOWN;
2705 while (!LIST_EMPTY(slp)) {
2706 error = sync_vnode(slp, &bo, td);
2708 LIST_REMOVE(bo, bo_synclist);
2709 LIST_INSERT_HEAD(next, bo, bo_synclist);
2713 if (first_printf == 0) {
2715 * Drop the sync mutex, because some watchdog
2716 * drivers need to sleep while patting
2718 mtx_unlock(&sync_mtx);
2719 wdog_kern_pat(WD_LASTVAL);
2720 mtx_lock(&sync_mtx);
2723 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2724 syncer_final_iter--;
2726 * The variable rushjob allows the kernel to speed up the
2727 * processing of the filesystem syncer process. A rushjob
2728 * value of N tells the filesystem syncer to process the next
2729 * N seconds worth of work on its queue ASAP. Currently rushjob
2730 * is used by the soft update code to speed up the filesystem
2731 * syncer process when the incore state is getting so far
2732 * ahead of the disk that the kernel memory pool is being
2733 * threatened with exhaustion.
2740 * Just sleep for a short period of time between
2741 * iterations when shutting down to allow some I/O
2744 * If it has taken us less than a second to process the
2745 * current work, then wait. Otherwise start right over
2746 * again. We can still lose time if any single round
2747 * takes more than two seconds, but it does not really
2748 * matter as we are just trying to generally pace the
2749 * filesystem activity.
2751 if (syncer_state != SYNCER_RUNNING ||
2752 time_uptime == starttime) {
2754 sched_prio(td, PPAUSE);
2757 if (syncer_state != SYNCER_RUNNING)
2758 cv_timedwait(&sync_wakeup, &sync_mtx,
2759 hz / SYNCER_SHUTDOWN_SPEEDUP);
2760 else if (time_uptime == starttime)
2761 cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2766 * Request the syncer daemon to speed up its work.
2767 * We never push it to speed up more than half of its
2768 * normal turn time, otherwise it could take over the cpu.
2771 speedup_syncer(void)
2775 mtx_lock(&sync_mtx);
2776 if (rushjob < syncdelay / 2) {
2778 stat_rush_requests += 1;
2781 mtx_unlock(&sync_mtx);
2782 cv_broadcast(&sync_wakeup);
2787 * Tell the syncer to speed up its work and run though its work
2788 * list several times, then tell it to shut down.
2791 syncer_shutdown(void *arg, int howto)
2794 if (howto & RB_NOSYNC)
2796 mtx_lock(&sync_mtx);
2797 syncer_state = SYNCER_SHUTTING_DOWN;
2799 mtx_unlock(&sync_mtx);
2800 cv_broadcast(&sync_wakeup);
2801 kproc_shutdown(arg, howto);
2805 syncer_suspend(void)
2808 syncer_shutdown(updateproc, 0);
2815 mtx_lock(&sync_mtx);
2817 syncer_state = SYNCER_RUNNING;
2818 mtx_unlock(&sync_mtx);
2819 cv_broadcast(&sync_wakeup);
2820 kproc_resume(updateproc);
2824 * Move the buffer between the clean and dirty lists of its vnode.
2827 reassignbuf(struct buf *bp)
2839 KASSERT((bp->b_flags & B_PAGING) == 0,
2840 ("%s: cannot reassign paging buffer %p", __func__, bp));
2842 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2843 bp, bp->b_vp, bp->b_flags);
2846 buf_vlist_remove(bp);
2849 * If dirty, put on list of dirty buffers; otherwise insert onto list
2852 if (bp->b_flags & B_DELWRI) {
2853 if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2854 switch (vp->v_type) {
2864 vn_syncer_add_to_worklist(bo, delay);
2866 buf_vlist_add(bp, bo, BX_VNDIRTY);
2868 buf_vlist_add(bp, bo, BX_VNCLEAN);
2870 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2871 mtx_lock(&sync_mtx);
2872 LIST_REMOVE(bo, bo_synclist);
2873 syncer_worklist_len--;
2874 mtx_unlock(&sync_mtx);
2875 bo->bo_flag &= ~BO_ONWORKLST;
2880 bp = TAILQ_FIRST(&bv->bv_hd);
2881 KASSERT(bp == NULL || bp->b_bufobj == bo,
2882 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2883 bp = TAILQ_LAST(&bv->bv_hd, buflists);
2884 KASSERT(bp == NULL || bp->b_bufobj == bo,
2885 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2887 bp = TAILQ_FIRST(&bv->bv_hd);
2888 KASSERT(bp == NULL || bp->b_bufobj == bo,
2889 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2890 bp = TAILQ_LAST(&bv->bv_hd, buflists);
2891 KASSERT(bp == NULL || bp->b_bufobj == bo,
2892 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2898 v_init_counters(struct vnode *vp)
2901 VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2902 vp, ("%s called for an initialized vnode", __FUNCTION__));
2903 ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2905 refcount_init(&vp->v_holdcnt, 1);
2906 refcount_init(&vp->v_usecount, 1);
2910 * Grab a particular vnode from the free list, increment its
2911 * reference count and lock it. VIRF_DOOMED is set if the vnode
2912 * is being destroyed. Only callers who specify LK_RETRY will
2913 * see doomed vnodes. If inactive processing was delayed in
2914 * vput try to do it here.
2916 * usecount is manipulated using atomics without holding any locks.
2918 * holdcnt can be manipulated using atomics without holding any locks,
2919 * except when transitioning 1<->0, in which case the interlock is held.
2921 * Consumers which don't guarantee liveness of the vnode can use SMR to
2922 * try to get a reference. Note this operation can fail since the vnode
2923 * may be awaiting getting freed by the time they get to it.
2926 vget_prep_smr(struct vnode *vp)
2930 VFS_SMR_ASSERT_ENTERED();
2932 if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2944 vget_prep(struct vnode *vp)
2948 if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2958 vget_abort(struct vnode *vp, enum vgetstate vs)
2969 __assert_unreachable();
2974 vget(struct vnode *vp, int flags)
2979 return (vget_finish(vp, flags, vs));
2983 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2987 if ((flags & LK_INTERLOCK) != 0)
2988 ASSERT_VI_LOCKED(vp, __func__);
2990 ASSERT_VI_UNLOCKED(vp, __func__);
2991 VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
2992 VNPASS(vp->v_holdcnt > 0, vp);
2993 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
2995 error = vn_lock(vp, flags);
2996 if (__predict_false(error != 0)) {
2998 CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
3003 vget_finish_ref(vp, vs);
3008 vget_finish_ref(struct vnode *vp, enum vgetstate vs)
3012 VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3013 VNPASS(vp->v_holdcnt > 0, vp);
3014 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3016 if (vs == VGET_USECOUNT)
3020 * We hold the vnode. If the usecount is 0 it will be utilized to keep
3021 * the vnode around. Otherwise someone else lended their hold count and
3022 * we have to drop ours.
3024 old = atomic_fetchadd_int(&vp->v_usecount, 1);
3025 VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3028 old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3029 VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3031 refcount_release(&vp->v_holdcnt);
3037 vref(struct vnode *vp)
3041 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3043 vget_finish_ref(vp, vs);
3047 vrefact(struct vnode *vp)
3050 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3052 int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3053 VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3055 refcount_acquire(&vp->v_usecount);
3060 vlazy(struct vnode *vp)
3064 VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3066 if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3069 * We may get here for inactive routines after the vnode got doomed.
3071 if (VN_IS_DOOMED(vp))
3074 mtx_lock(&mp->mnt_listmtx);
3075 if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3076 vp->v_mflag |= VMP_LAZYLIST;
3077 TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3078 mp->mnt_lazyvnodelistsize++;
3080 mtx_unlock(&mp->mnt_listmtx);
3084 vunlazy(struct vnode *vp)
3088 ASSERT_VI_LOCKED(vp, __func__);
3089 VNPASS(!VN_IS_DOOMED(vp), vp);
3092 mtx_lock(&mp->mnt_listmtx);
3093 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3095 * Don't remove the vnode from the lazy list if another thread
3096 * has increased the hold count. It may have re-enqueued the
3097 * vnode to the lazy list and is now responsible for its
3100 if (vp->v_holdcnt == 0) {
3101 vp->v_mflag &= ~VMP_LAZYLIST;
3102 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3103 mp->mnt_lazyvnodelistsize--;
3105 mtx_unlock(&mp->mnt_listmtx);
3109 * This routine is only meant to be called from vgonel prior to dooming
3113 vunlazy_gone(struct vnode *vp)
3117 ASSERT_VOP_ELOCKED(vp, __func__);
3118 ASSERT_VI_LOCKED(vp, __func__);
3119 VNPASS(!VN_IS_DOOMED(vp), vp);
3121 if (vp->v_mflag & VMP_LAZYLIST) {
3123 mtx_lock(&mp->mnt_listmtx);
3124 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3125 vp->v_mflag &= ~VMP_LAZYLIST;
3126 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3127 mp->mnt_lazyvnodelistsize--;
3128 mtx_unlock(&mp->mnt_listmtx);
3133 vdefer_inactive(struct vnode *vp)
3136 ASSERT_VI_LOCKED(vp, __func__);
3137 VNASSERT(vp->v_holdcnt > 0, vp,
3138 ("%s: vnode without hold count", __func__));
3139 if (VN_IS_DOOMED(vp)) {
3143 if (vp->v_iflag & VI_DEFINACT) {
3144 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3148 if (vp->v_usecount > 0) {
3149 vp->v_iflag &= ~VI_OWEINACT;
3154 vp->v_iflag |= VI_DEFINACT;
3156 counter_u64_add(deferred_inact, 1);
3160 vdefer_inactive_unlocked(struct vnode *vp)
3164 if ((vp->v_iflag & VI_OWEINACT) == 0) {
3168 vdefer_inactive(vp);
3171 enum vput_op { VRELE, VPUT, VUNREF };
3174 * Handle ->v_usecount transitioning to 0.
3176 * By releasing the last usecount we take ownership of the hold count which
3177 * provides liveness of the vnode, meaning we have to vdrop.
3179 * For all vnodes we may need to perform inactive processing. It requires an
3180 * exclusive lock on the vnode, while it is legal to call here with only a
3181 * shared lock (or no locks). If locking the vnode in an expected manner fails,
3182 * inactive processing gets deferred to the syncer.
3184 * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3185 * on the lock being held all the way until VOP_INACTIVE. This in particular
3186 * happens with UFS which adds half-constructed vnodes to the hash, where they
3187 * can be found by other code.
3190 vput_final(struct vnode *vp, enum vput_op func)
3195 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3196 VNPASS(vp->v_holdcnt > 0, vp);
3201 * By the time we got here someone else might have transitioned
3202 * the count back to > 0.
3204 if (vp->v_usecount > 0)
3208 * If the vnode is doomed vgone already performed inactive processing
3211 if (VN_IS_DOOMED(vp))
3214 if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3217 if (vp->v_iflag & VI_DOINGINACT)
3221 * Locking operations here will drop the interlock and possibly the
3222 * vnode lock, opening a window where the vnode can get doomed all the
3223 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3226 vp->v_iflag |= VI_OWEINACT;
3227 want_unlock = false;
3231 switch (VOP_ISLOCKED(vp)) {
3237 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3242 * The lock has at least one sharer, but we have no way
3243 * to conclude whether this is us. Play it safe and
3252 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3253 error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3259 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3260 error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3266 if (func == VUNREF) {
3267 VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp,
3268 ("recursive vunref"));
3269 vp->v_vflag |= VV_UNREF;
3272 error = vinactive(vp);
3275 if (error != ERELOOKUP || !want_unlock)
3277 VOP_LOCK(vp, LK_EXCLUSIVE);
3280 vp->v_vflag &= ~VV_UNREF;
3283 vdefer_inactive(vp);
3293 * Decrement ->v_usecount for a vnode.
3295 * Releasing the last use count requires additional processing, see vput_final
3296 * above for details.
3298 * Comment above each variant denotes lock state on entry and exit.
3303 * out: same as passed in
3306 vrele(struct vnode *vp)
3309 ASSERT_VI_UNLOCKED(vp, __func__);
3310 if (!refcount_release(&vp->v_usecount))
3312 vput_final(vp, VRELE);
3320 vput(struct vnode *vp)
3323 ASSERT_VOP_LOCKED(vp, __func__);
3324 ASSERT_VI_UNLOCKED(vp, __func__);
3325 if (!refcount_release(&vp->v_usecount)) {
3329 vput_final(vp, VPUT);
3337 vunref(struct vnode *vp)
3340 ASSERT_VOP_LOCKED(vp, __func__);
3341 ASSERT_VI_UNLOCKED(vp, __func__);
3342 if (!refcount_release(&vp->v_usecount))
3344 vput_final(vp, VUNREF);
3348 vhold(struct vnode *vp)
3352 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3353 old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3354 VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3355 ("%s: wrong hold count %d", __func__, old));
3357 vfs_freevnodes_dec();
3361 vholdnz(struct vnode *vp)
3364 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3366 int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3367 VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3368 ("%s: wrong hold count %d", __func__, old));
3370 atomic_add_int(&vp->v_holdcnt, 1);
3375 * Grab a hold count unless the vnode is freed.
3377 * Only use this routine if vfs smr is the only protection you have against
3378 * freeing the vnode.
3380 * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag
3381 * is not set. After the flag is set the vnode becomes immutable to anyone but
3382 * the thread which managed to set the flag.
3384 * It may be tempting to replace the loop with:
3385 * count = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3386 * if (count & VHOLD_NO_SMR) {
3387 * backpedal and error out;
3390 * However, while this is more performant, it hinders debugging by eliminating
3391 * the previously mentioned invariant.
3394 vhold_smr(struct vnode *vp)
3398 VFS_SMR_ASSERT_ENTERED();
3400 count = atomic_load_int(&vp->v_holdcnt);
3402 if (count & VHOLD_NO_SMR) {
3403 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3404 ("non-zero hold count with flags %d\n", count));
3407 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3408 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3410 vfs_freevnodes_dec();
3417 * Hold a free vnode for recycling.
3419 * Note: vnode_init references this comment.
3421 * Attempts to recycle only need the global vnode list lock and have no use for
3424 * However, vnodes get inserted into the global list before they get fully
3425 * initialized and stay there until UMA decides to free the memory. This in
3426 * particular means the target can be found before it becomes usable and after
3427 * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to
3430 * Note: the vnode may gain more references after we transition the count 0->1.
3433 vhold_recycle_free(struct vnode *vp)
3437 mtx_assert(&vnode_list_mtx, MA_OWNED);
3439 count = atomic_load_int(&vp->v_holdcnt);
3441 if (count & VHOLD_NO_SMR) {
3442 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3443 ("non-zero hold count with flags %d\n", count));
3446 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3450 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3451 vfs_freevnodes_dec();
3457 static void __noinline
3458 vdbatch_process(struct vdbatch *vd)
3463 mtx_assert(&vd->lock, MA_OWNED);
3464 MPASS(curthread->td_pinned > 0);
3465 MPASS(vd->index == VDBATCH_SIZE);
3467 mtx_lock(&vnode_list_mtx);
3469 freevnodes += vd->freevnodes;
3470 for (i = 0; i < VDBATCH_SIZE; i++) {
3472 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3473 TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3474 MPASS(vp->v_dbatchcpu != NOCPU);
3475 vp->v_dbatchcpu = NOCPU;
3477 mtx_unlock(&vnode_list_mtx);
3479 bzero(vd->tab, sizeof(vd->tab));
3485 vdbatch_enqueue(struct vnode *vp)
3489 ASSERT_VI_LOCKED(vp, __func__);
3490 VNASSERT(!VN_IS_DOOMED(vp), vp,
3491 ("%s: deferring requeue of a doomed vnode", __func__));
3493 if (vp->v_dbatchcpu != NOCPU) {
3500 mtx_lock(&vd->lock);
3501 MPASS(vd->index < VDBATCH_SIZE);
3502 MPASS(vd->tab[vd->index] == NULL);
3504 * A hack: we depend on being pinned so that we know what to put in
3507 vp->v_dbatchcpu = curcpu;
3508 vd->tab[vd->index] = vp;
3511 if (vd->index == VDBATCH_SIZE)
3512 vdbatch_process(vd);
3513 mtx_unlock(&vd->lock);
3518 * This routine must only be called for vnodes which are about to be
3519 * deallocated. Supporting dequeue for arbitrary vndoes would require
3520 * validating that the locked batch matches.
3523 vdbatch_dequeue(struct vnode *vp)
3529 VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3530 ("%s: called for a used vnode\n", __func__));
3532 cpu = vp->v_dbatchcpu;
3536 vd = DPCPU_ID_PTR(cpu, vd);
3537 mtx_lock(&vd->lock);
3538 for (i = 0; i < vd->index; i++) {
3539 if (vd->tab[i] != vp)
3541 vp->v_dbatchcpu = NOCPU;
3543 vd->tab[i] = vd->tab[vd->index];
3544 vd->tab[vd->index] = NULL;
3547 mtx_unlock(&vd->lock);
3549 * Either we dequeued the vnode above or the target CPU beat us to it.
3551 MPASS(vp->v_dbatchcpu == NOCPU);
3555 * Drop the hold count of the vnode. If this is the last reference to
3556 * the vnode we place it on the free list unless it has been vgone'd
3557 * (marked VIRF_DOOMED) in which case we will free it.
3559 * Because the vnode vm object keeps a hold reference on the vnode if
3560 * there is at least one resident non-cached page, the vnode cannot
3561 * leave the active list without the page cleanup done.
3563 static void __noinline
3564 vdropl_final(struct vnode *vp)
3567 ASSERT_VI_LOCKED(vp, __func__);
3568 VNPASS(VN_IS_DOOMED(vp), vp);
3570 * Set the VHOLD_NO_SMR flag.
3572 * We may be racing against vhold_smr. If they win we can just pretend
3573 * we never got this far, they will vdrop later.
3575 if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) {
3576 vfs_freevnodes_inc();
3579 * We lost the aforementioned race. Any subsequent access is
3580 * invalid as they might have managed to vdropl on their own.
3585 * Don't bump freevnodes as this one is going away.
3591 vdrop(struct vnode *vp)
3594 ASSERT_VI_UNLOCKED(vp, __func__);
3595 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3596 if (refcount_release_if_not_last(&vp->v_holdcnt))
3603 vdropl(struct vnode *vp)
3606 ASSERT_VI_LOCKED(vp, __func__);
3607 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3608 if (!refcount_release(&vp->v_holdcnt)) {
3612 VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp);
3613 VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
3614 if (VN_IS_DOOMED(vp)) {
3619 vfs_freevnodes_inc();
3620 if (vp->v_mflag & VMP_LAZYLIST) {
3624 * Also unlocks the interlock. We can't assert on it as we
3625 * released our hold and by now the vnode might have been
3628 vdbatch_enqueue(vp);
3632 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3633 * flags. DOINGINACT prevents us from recursing in calls to vinactive.
3636 vinactivef(struct vnode *vp)
3638 struct vm_object *obj;
3641 ASSERT_VOP_ELOCKED(vp, "vinactive");
3642 ASSERT_VI_LOCKED(vp, "vinactive");
3643 VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3644 ("vinactive: recursed on VI_DOINGINACT"));
3645 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3646 vp->v_iflag |= VI_DOINGINACT;
3647 vp->v_iflag &= ~VI_OWEINACT;
3650 * Before moving off the active list, we must be sure that any
3651 * modified pages are converted into the vnode's dirty
3652 * buffers, since these will no longer be checked once the
3653 * vnode is on the inactive list.
3655 * The write-out of the dirty pages is asynchronous. At the
3656 * point that VOP_INACTIVE() is called, there could still be
3657 * pending I/O and dirty pages in the object.
3659 if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3660 vm_object_mightbedirty(obj)) {
3661 VM_OBJECT_WLOCK(obj);
3662 vm_object_page_clean(obj, 0, 0, 0);
3663 VM_OBJECT_WUNLOCK(obj);
3665 error = VOP_INACTIVE(vp);
3667 VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3668 ("vinactive: lost VI_DOINGINACT"));
3669 vp->v_iflag &= ~VI_DOINGINACT;
3674 vinactive(struct vnode *vp)
3677 ASSERT_VOP_ELOCKED(vp, "vinactive");
3678 ASSERT_VI_LOCKED(vp, "vinactive");
3679 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3681 if ((vp->v_iflag & VI_OWEINACT) == 0)
3683 if (vp->v_iflag & VI_DOINGINACT)
3685 if (vp->v_usecount > 0) {
3686 vp->v_iflag &= ~VI_OWEINACT;
3689 return (vinactivef(vp));
3693 * Remove any vnodes in the vnode table belonging to mount point mp.
3695 * If FORCECLOSE is not specified, there should not be any active ones,
3696 * return error if any are found (nb: this is a user error, not a
3697 * system error). If FORCECLOSE is specified, detach any active vnodes
3700 * If WRITECLOSE is set, only flush out regular file vnodes open for
3703 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3705 * `rootrefs' specifies the base reference count for the root vnode
3706 * of this filesystem. The root vnode is considered busy if its
3707 * v_usecount exceeds this value. On a successful return, vflush(, td)
3708 * will call vrele() on the root vnode exactly rootrefs times.
3709 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3713 static int busyprt = 0; /* print out busy vnodes */
3714 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3718 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3720 struct vnode *vp, *mvp, *rootvp = NULL;
3722 int busy = 0, error;
3724 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3727 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3728 ("vflush: bad args"));
3730 * Get the filesystem root vnode. We can vput() it
3731 * immediately, since with rootrefs > 0, it won't go away.
3733 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3734 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3741 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3743 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3746 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3750 * Skip over a vnodes marked VV_SYSTEM.
3752 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3758 * If WRITECLOSE is set, flush out unlinked but still open
3759 * files (even if open only for reading) and regular file
3760 * vnodes open for writing.
3762 if (flags & WRITECLOSE) {
3763 if (vp->v_object != NULL) {
3764 VM_OBJECT_WLOCK(vp->v_object);
3765 vm_object_page_clean(vp->v_object, 0, 0, 0);
3766 VM_OBJECT_WUNLOCK(vp->v_object);
3769 error = VOP_FSYNC(vp, MNT_WAIT, td);
3770 } while (error == ERELOOKUP);
3774 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3777 error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3780 if ((vp->v_type == VNON ||
3781 (error == 0 && vattr.va_nlink > 0)) &&
3782 (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3790 * With v_usecount == 0, all we need to do is clear out the
3791 * vnode data structures and we are done.
3793 * If FORCECLOSE is set, forcibly close the vnode.
3795 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3801 vn_printf(vp, "vflush: busy vnode ");
3807 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3809 * If just the root vnode is busy, and if its refcount
3810 * is equal to `rootrefs', then go ahead and kill it.
3813 KASSERT(busy > 0, ("vflush: not busy"));
3814 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3815 ("vflush: usecount %d < rootrefs %d",
3816 rootvp->v_usecount, rootrefs));
3817 if (busy == 1 && rootvp->v_usecount == rootrefs) {
3818 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3826 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3830 for (; rootrefs > 0; rootrefs--)
3836 * Recycle an unused vnode to the front of the free list.
3839 vrecycle(struct vnode *vp)
3844 recycled = vrecyclel(vp);
3850 * vrecycle, with the vp interlock held.
3853 vrecyclel(struct vnode *vp)
3857 ASSERT_VOP_ELOCKED(vp, __func__);
3858 ASSERT_VI_LOCKED(vp, __func__);
3859 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3861 if (vp->v_usecount == 0) {
3869 * Eliminate all activity associated with a vnode
3870 * in preparation for reuse.
3873 vgone(struct vnode *vp)
3881 * Notify upper mounts about reclaimed or unlinked vnode.
3884 vfs_notify_upper(struct vnode *vp, int event)
3887 struct mount_upper_node *ump;
3889 mp = atomic_load_ptr(&vp->v_mount);
3892 if (TAILQ_EMPTY(&mp->mnt_notify))
3896 mp->mnt_upper_pending++;
3897 KASSERT(mp->mnt_upper_pending > 0,
3898 ("%s: mnt_upper_pending %d", __func__, mp->mnt_upper_pending));
3899 TAILQ_FOREACH(ump, &mp->mnt_notify, mnt_upper_link) {
3902 case VFS_NOTIFY_UPPER_RECLAIM:
3903 VFS_RECLAIM_LOWERVP(ump->mp, vp);
3905 case VFS_NOTIFY_UPPER_UNLINK:
3906 VFS_UNLINK_LOWERVP(ump->mp, vp);
3909 KASSERT(0, ("invalid event %d", event));
3914 mp->mnt_upper_pending--;
3915 if ((mp->mnt_kern_flag & MNTK_UPPER_WAITER) != 0 &&
3916 mp->mnt_upper_pending == 0) {
3917 mp->mnt_kern_flag &= ~MNTK_UPPER_WAITER;
3918 wakeup(&mp->mnt_uppers);
3924 * vgone, with the vp interlock held.
3927 vgonel(struct vnode *vp)
3932 bool active, doinginact, oweinact;
3934 ASSERT_VOP_ELOCKED(vp, "vgonel");
3935 ASSERT_VI_LOCKED(vp, "vgonel");
3936 VNASSERT(vp->v_holdcnt, vp,
3937 ("vgonel: vp %p has no reference.", vp));
3938 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3942 * Don't vgonel if we're already doomed.
3944 if (VN_IS_DOOMED(vp))
3947 * Paired with freevnode.
3949 vn_seqc_write_begin_locked(vp);
3951 vn_irflag_set_locked(vp, VIRF_DOOMED);
3954 * Check to see if the vnode is in use. If so, we have to
3955 * call VOP_CLOSE() and VOP_INACTIVE().
3957 * It could be that VOP_INACTIVE() requested reclamation, in
3958 * which case we should avoid recursion, so check
3959 * VI_DOINGINACT. This is not precise but good enough.
3961 active = vp->v_usecount > 0;
3962 oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3963 doinginact = (vp->v_iflag & VI_DOINGINACT) != 0;
3966 * If we need to do inactive VI_OWEINACT will be set.
3968 if (vp->v_iflag & VI_DEFINACT) {
3969 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3970 vp->v_iflag &= ~VI_DEFINACT;
3973 VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
3976 cache_purge_vgone(vp);
3977 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3980 * If purging an active vnode, it must be closed and
3981 * deactivated before being reclaimed.
3984 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3987 if (oweinact || active) {
3990 oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3995 if (vp->v_type == VSOCK)
3996 vfs_unp_reclaim(vp);
3999 * Clean out any buffers associated with the vnode.
4000 * If the flush fails, just toss the buffers.
4003 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
4004 (void) vn_start_secondary_write(vp, &mp, V_WAIT);
4005 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
4006 while (vinvalbuf(vp, 0, 0, 0) != 0)
4010 BO_LOCK(&vp->v_bufobj);
4011 KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
4012 vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
4013 TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
4014 vp->v_bufobj.bo_clean.bv_cnt == 0,
4015 ("vp %p bufobj not invalidated", vp));
4018 * For VMIO bufobj, BO_DEAD is set later, or in
4019 * vm_object_terminate() after the object's page queue is
4022 object = vp->v_bufobj.bo_object;
4024 vp->v_bufobj.bo_flag |= BO_DEAD;
4025 BO_UNLOCK(&vp->v_bufobj);
4028 * Handle the VM part. Tmpfs handles v_object on its own (the
4029 * OBJT_VNODE check). Nullfs or other bypassing filesystems
4030 * should not touch the object borrowed from the lower vnode
4031 * (the handle check).
4033 if (object != NULL && object->type == OBJT_VNODE &&
4034 object->handle == vp)
4035 vnode_destroy_vobject(vp);
4038 * Reclaim the vnode.
4040 if (VOP_RECLAIM(vp))
4041 panic("vgone: cannot reclaim");
4043 vn_finished_secondary_write(mp);
4044 VNASSERT(vp->v_object == NULL, vp,
4045 ("vop_reclaim left v_object vp=%p", vp));
4047 * Clear the advisory locks and wake up waiting threads.
4049 (void)VOP_ADVLOCKPURGE(vp);
4052 * Delete from old mount point vnode list.
4056 * Done with purge, reset to the standard lock and invalidate
4060 vp->v_vnlock = &vp->v_lock;
4061 vp->v_op = &dead_vnodeops;
4066 * Print out a description of a vnode.
4068 static const char * const typename[] =
4069 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
4072 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0,
4073 "new hold count flag not added to vn_printf");
4076 vn_printf(struct vnode *vp, const char *fmt, ...)
4079 char buf[256], buf2[16];
4087 printf("%p: ", (void *)vp);
4088 printf("type %s\n", typename[vp->v_type]);
4089 holdcnt = atomic_load_int(&vp->v_holdcnt);
4090 printf(" usecount %d, writecount %d, refcount %d seqc users %d",
4091 vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS,
4093 switch (vp->v_type) {
4095 printf(" mountedhere %p\n", vp->v_mountedhere);
4098 printf(" rdev %p\n", vp->v_rdev);
4101 printf(" socket %p\n", vp->v_unpcb);
4104 printf(" fifoinfo %p\n", vp->v_fifoinfo);
4112 if (holdcnt & VHOLD_NO_SMR)
4113 strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf));
4114 printf(" hold count flags (%s)\n", buf + 1);
4118 irflag = vn_irflag_read(vp);
4119 if (irflag & VIRF_DOOMED)
4120 strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4121 if (irflag & VIRF_PGREAD)
4122 strlcat(buf, "|VIRF_PGREAD", sizeof(buf));
4123 if (irflag & VIRF_MOUNTPOINT)
4124 strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf));
4125 if (irflag & VIRF_TEXT_REF)
4126 strlcat(buf, "|VIRF_TEXT_REF", sizeof(buf));
4127 flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT | VIRF_TEXT_REF);
4129 snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4130 strlcat(buf, buf2, sizeof(buf));
4132 if (vp->v_vflag & VV_ROOT)
4133 strlcat(buf, "|VV_ROOT", sizeof(buf));
4134 if (vp->v_vflag & VV_ISTTY)
4135 strlcat(buf, "|VV_ISTTY", sizeof(buf));
4136 if (vp->v_vflag & VV_NOSYNC)
4137 strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4138 if (vp->v_vflag & VV_ETERNALDEV)
4139 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4140 if (vp->v_vflag & VV_CACHEDLABEL)
4141 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4142 if (vp->v_vflag & VV_VMSIZEVNLOCK)
4143 strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4144 if (vp->v_vflag & VV_COPYONWRITE)
4145 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4146 if (vp->v_vflag & VV_SYSTEM)
4147 strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4148 if (vp->v_vflag & VV_PROCDEP)
4149 strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4150 if (vp->v_vflag & VV_NOKNOTE)
4151 strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
4152 if (vp->v_vflag & VV_DELETED)
4153 strlcat(buf, "|VV_DELETED", sizeof(buf));
4154 if (vp->v_vflag & VV_MD)
4155 strlcat(buf, "|VV_MD", sizeof(buf));
4156 if (vp->v_vflag & VV_FORCEINSMQ)
4157 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4158 if (vp->v_vflag & VV_READLINK)
4159 strlcat(buf, "|VV_READLINK", sizeof(buf));
4160 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4161 VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM |
4162 VV_PROCDEP | VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ |
4165 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4166 strlcat(buf, buf2, sizeof(buf));
4168 if (vp->v_iflag & VI_MOUNT)
4169 strlcat(buf, "|VI_MOUNT", sizeof(buf));
4170 if (vp->v_iflag & VI_DOINGINACT)
4171 strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4172 if (vp->v_iflag & VI_OWEINACT)
4173 strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4174 if (vp->v_iflag & VI_DEFINACT)
4175 strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4176 if (vp->v_iflag & VI_FOPENING)
4177 strlcat(buf, "|VI_FOPENING", sizeof(buf));
4178 flags = vp->v_iflag & ~(VI_MOUNT | VI_DOINGINACT |
4179 VI_OWEINACT | VI_DEFINACT | VI_FOPENING);
4181 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4182 strlcat(buf, buf2, sizeof(buf));
4184 if (vp->v_mflag & VMP_LAZYLIST)
4185 strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4186 flags = vp->v_mflag & ~(VMP_LAZYLIST);
4188 snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4189 strlcat(buf, buf2, sizeof(buf));
4191 printf(" flags (%s)", buf + 1);
4192 if (mtx_owned(VI_MTX(vp)))
4193 printf(" VI_LOCKed");
4195 if (vp->v_object != NULL)
4196 printf(" v_object %p ref %d pages %d "
4197 "cleanbuf %d dirtybuf %d\n",
4198 vp->v_object, vp->v_object->ref_count,
4199 vp->v_object->resident_page_count,
4200 vp->v_bufobj.bo_clean.bv_cnt,
4201 vp->v_bufobj.bo_dirty.bv_cnt);
4203 lockmgr_printinfo(vp->v_vnlock);
4204 if (vp->v_data != NULL)
4210 * List all of the locked vnodes in the system.
4211 * Called when debugging the kernel.
4213 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
4219 * Note: because this is DDB, we can't obey the locking semantics
4220 * for these structures, which means we could catch an inconsistent
4221 * state and dereference a nasty pointer. Not much to be done
4224 db_printf("Locked vnodes\n");
4225 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4226 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4227 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4228 vn_printf(vp, "vnode ");
4234 * Show details about the given vnode.
4236 DB_SHOW_COMMAND(vnode, db_show_vnode)
4242 vp = (struct vnode *)addr;
4243 vn_printf(vp, "vnode ");
4247 * Show details about the given mount point.
4249 DB_SHOW_COMMAND(mount, db_show_mount)
4260 /* No address given, print short info about all mount points. */
4261 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4262 db_printf("%p %s on %s (%s)\n", mp,
4263 mp->mnt_stat.f_mntfromname,
4264 mp->mnt_stat.f_mntonname,
4265 mp->mnt_stat.f_fstypename);
4269 db_printf("\nMore info: show mount <addr>\n");
4273 mp = (struct mount *)addr;
4274 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4275 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4278 mflags = mp->mnt_flag;
4279 #define MNT_FLAG(flag) do { \
4280 if (mflags & (flag)) { \
4281 if (buf[0] != '\0') \
4282 strlcat(buf, ", ", sizeof(buf)); \
4283 strlcat(buf, (#flag) + 4, sizeof(buf)); \
4284 mflags &= ~(flag); \
4287 MNT_FLAG(MNT_RDONLY);
4288 MNT_FLAG(MNT_SYNCHRONOUS);
4289 MNT_FLAG(MNT_NOEXEC);
4290 MNT_FLAG(MNT_NOSUID);
4291 MNT_FLAG(MNT_NFS4ACLS);
4292 MNT_FLAG(MNT_UNION);
4293 MNT_FLAG(MNT_ASYNC);
4294 MNT_FLAG(MNT_SUIDDIR);
4295 MNT_FLAG(MNT_SOFTDEP);
4296 MNT_FLAG(MNT_NOSYMFOLLOW);
4297 MNT_FLAG(MNT_GJOURNAL);
4298 MNT_FLAG(MNT_MULTILABEL);
4300 MNT_FLAG(MNT_NOATIME);
4301 MNT_FLAG(MNT_NOCLUSTERR);
4302 MNT_FLAG(MNT_NOCLUSTERW);
4304 MNT_FLAG(MNT_EXRDONLY);
4305 MNT_FLAG(MNT_EXPORTED);
4306 MNT_FLAG(MNT_DEFEXPORTED);
4307 MNT_FLAG(MNT_EXPORTANON);
4308 MNT_FLAG(MNT_EXKERB);
4309 MNT_FLAG(MNT_EXPUBLIC);
4310 MNT_FLAG(MNT_LOCAL);
4311 MNT_FLAG(MNT_QUOTA);
4312 MNT_FLAG(MNT_ROOTFS);
4314 MNT_FLAG(MNT_IGNORE);
4315 MNT_FLAG(MNT_UPDATE);
4316 MNT_FLAG(MNT_DELEXPORT);
4317 MNT_FLAG(MNT_RELOAD);
4318 MNT_FLAG(MNT_FORCE);
4319 MNT_FLAG(MNT_SNAPSHOT);
4320 MNT_FLAG(MNT_BYFSID);
4324 strlcat(buf, ", ", sizeof(buf));
4325 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4326 "0x%016jx", mflags);
4328 db_printf(" mnt_flag = %s\n", buf);
4331 flags = mp->mnt_kern_flag;
4332 #define MNT_KERN_FLAG(flag) do { \
4333 if (flags & (flag)) { \
4334 if (buf[0] != '\0') \
4335 strlcat(buf, ", ", sizeof(buf)); \
4336 strlcat(buf, (#flag) + 5, sizeof(buf)); \
4340 MNT_KERN_FLAG(MNTK_UNMOUNTF);
4341 MNT_KERN_FLAG(MNTK_ASYNC);
4342 MNT_KERN_FLAG(MNTK_SOFTDEP);
4343 MNT_KERN_FLAG(MNTK_DRAINING);
4344 MNT_KERN_FLAG(MNTK_REFEXPIRE);
4345 MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4346 MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4347 MNT_KERN_FLAG(MNTK_NO_IOPF);
4348 MNT_KERN_FLAG(MNTK_RECURSE);
4349 MNT_KERN_FLAG(MNTK_UPPER_WAITER);
4350 MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
4351 MNT_KERN_FLAG(MNTK_USES_BCACHE);
4352 MNT_KERN_FLAG(MNTK_FPLOOKUP);
4353 MNT_KERN_FLAG(MNTK_TASKQUEUE_WAITER);
4354 MNT_KERN_FLAG(MNTK_NOASYNC);
4355 MNT_KERN_FLAG(MNTK_UNMOUNT);
4356 MNT_KERN_FLAG(MNTK_MWAIT);
4357 MNT_KERN_FLAG(MNTK_SUSPEND);
4358 MNT_KERN_FLAG(MNTK_SUSPEND2);
4359 MNT_KERN_FLAG(MNTK_SUSPENDED);
4360 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4361 MNT_KERN_FLAG(MNTK_NOKNOTE);
4362 #undef MNT_KERN_FLAG
4365 strlcat(buf, ", ", sizeof(buf));
4366 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4369 db_printf(" mnt_kern_flag = %s\n", buf);
4371 db_printf(" mnt_opt = ");
4372 opt = TAILQ_FIRST(mp->mnt_opt);
4374 db_printf("%s", opt->name);
4375 opt = TAILQ_NEXT(opt, link);
4376 while (opt != NULL) {
4377 db_printf(", %s", opt->name);
4378 opt = TAILQ_NEXT(opt, link);
4384 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx "
4385 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4386 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4387 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4388 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4389 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4390 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4391 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4392 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4393 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4394 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4395 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4397 db_printf(" mnt_cred = { uid=%u ruid=%u",
4398 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4399 if (jailed(mp->mnt_cred))
4400 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4402 db_printf(" mnt_ref = %d (with %d in the struct)\n",
4403 vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4404 db_printf(" mnt_gen = %d\n", mp->mnt_gen);
4405 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4406 db_printf(" mnt_lazyvnodelistsize = %d\n",
4407 mp->mnt_lazyvnodelistsize);
4408 db_printf(" mnt_writeopcount = %d (with %d in the struct)\n",
4409 vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4410 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4411 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed);
4412 db_printf(" mnt_lockref = %d (with %d in the struct)\n",
4413 vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4414 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4415 db_printf(" mnt_secondary_accwrites = %d\n",
4416 mp->mnt_secondary_accwrites);
4417 db_printf(" mnt_gjprovider = %s\n",
4418 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4419 db_printf(" mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4421 db_printf("\n\nList of active vnodes\n");
4422 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4423 if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4424 vn_printf(vp, "vnode ");
4429 db_printf("\n\nList of inactive vnodes\n");
4430 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4431 if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4432 vn_printf(vp, "vnode ");
4441 * Fill in a struct xvfsconf based on a struct vfsconf.
4444 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4446 struct xvfsconf xvfsp;
4448 bzero(&xvfsp, sizeof(xvfsp));
4449 strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4450 xvfsp.vfc_typenum = vfsp->vfc_typenum;
4451 xvfsp.vfc_refcount = vfsp->vfc_refcount;
4452 xvfsp.vfc_flags = vfsp->vfc_flags;
4454 * These are unused in userland, we keep them
4455 * to not break binary compatibility.
4457 xvfsp.vfc_vfsops = NULL;
4458 xvfsp.vfc_next = NULL;
4459 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4462 #ifdef COMPAT_FREEBSD32
4464 uint32_t vfc_vfsops;
4465 char vfc_name[MFSNAMELEN];
4466 int32_t vfc_typenum;
4467 int32_t vfc_refcount;
4473 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4475 struct xvfsconf32 xvfsp;
4477 bzero(&xvfsp, sizeof(xvfsp));
4478 strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4479 xvfsp.vfc_typenum = vfsp->vfc_typenum;
4480 xvfsp.vfc_refcount = vfsp->vfc_refcount;
4481 xvfsp.vfc_flags = vfsp->vfc_flags;
4482 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4487 * Top level filesystem related information gathering.
4490 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4492 struct vfsconf *vfsp;
4497 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4498 #ifdef COMPAT_FREEBSD32
4499 if (req->flags & SCTL_MASK32)
4500 error = vfsconf2x32(req, vfsp);
4503 error = vfsconf2x(req, vfsp);
4511 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4512 CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4513 "S,xvfsconf", "List of all configured filesystems");
4515 #ifndef BURN_BRIDGES
4516 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4519 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4521 int *name = (int *)arg1 - 1; /* XXX */
4522 u_int namelen = arg2 + 1; /* XXX */
4523 struct vfsconf *vfsp;
4525 log(LOG_WARNING, "userland calling deprecated sysctl, "
4526 "please rebuild world\n");
4528 #if 1 || defined(COMPAT_PRELITE2)
4529 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4531 return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4535 case VFS_MAXTYPENUM:
4538 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4541 return (ENOTDIR); /* overloaded */
4543 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4544 if (vfsp->vfc_typenum == name[2])
4549 return (EOPNOTSUPP);
4550 #ifdef COMPAT_FREEBSD32
4551 if (req->flags & SCTL_MASK32)
4552 return (vfsconf2x32(req, vfsp));
4555 return (vfsconf2x(req, vfsp));
4557 return (EOPNOTSUPP);
4560 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4561 CTLFLAG_MPSAFE, vfs_sysctl,
4562 "Generic filesystem");
4564 #if 1 || defined(COMPAT_PRELITE2)
4567 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4570 struct vfsconf *vfsp;
4571 struct ovfsconf ovfs;
4574 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4575 bzero(&ovfs, sizeof(ovfs));
4576 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */
4577 strcpy(ovfs.vfc_name, vfsp->vfc_name);
4578 ovfs.vfc_index = vfsp->vfc_typenum;
4579 ovfs.vfc_refcount = vfsp->vfc_refcount;
4580 ovfs.vfc_flags = vfsp->vfc_flags;
4581 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4591 #endif /* 1 || COMPAT_PRELITE2 */
4592 #endif /* !BURN_BRIDGES */
4594 #define KINFO_VNODESLOP 10
4597 * Dump vnode list (via sysctl).
4601 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4609 * Stale numvnodes access is not fatal here.
4612 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4614 /* Make an estimate */
4615 return (SYSCTL_OUT(req, 0, len));
4617 error = sysctl_wire_old_buffer(req, 0);
4620 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4622 mtx_lock(&mountlist_mtx);
4623 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4624 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4627 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4631 xvn[n].xv_size = sizeof *xvn;
4632 xvn[n].xv_vnode = vp;
4633 xvn[n].xv_id = 0; /* XXX compat */
4634 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4636 XV_COPY(writecount);
4642 xvn[n].xv_flag = vp->v_vflag;
4644 switch (vp->v_type) {
4651 if (vp->v_rdev == NULL) {
4655 xvn[n].xv_dev = dev2udev(vp->v_rdev);
4658 xvn[n].xv_socket = vp->v_socket;
4661 xvn[n].xv_fifo = vp->v_fifoinfo;
4666 /* shouldn't happen? */
4674 mtx_lock(&mountlist_mtx);
4679 mtx_unlock(&mountlist_mtx);
4681 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4686 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4687 CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4692 unmount_or_warn(struct mount *mp)
4696 error = dounmount(mp, MNT_FORCE, curthread);
4698 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4702 printf("%d)\n", error);
4707 * Unmount all filesystems. The list is traversed in reverse order
4708 * of mounting to avoid dependencies.
4711 vfs_unmountall(void)
4713 struct mount *mp, *tmp;
4715 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4718 * Since this only runs when rebooting, it is not interlocked.
4720 TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4724 * Forcibly unmounting "/dev" before "/" would prevent clean
4725 * unmount of the latter.
4727 if (mp == rootdevmp)
4730 unmount_or_warn(mp);
4733 if (rootdevmp != NULL)
4734 unmount_or_warn(rootdevmp);
4738 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4741 ASSERT_VI_LOCKED(vp, __func__);
4742 VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4743 if ((vp->v_iflag & VI_OWEINACT) == 0) {
4747 if (vn_lock(vp, lkflags) == 0) {
4754 vdefer_inactive_unlocked(vp);
4758 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4761 return (vp->v_iflag & VI_DEFINACT);
4764 static void __noinline
4765 vfs_periodic_inactive(struct mount *mp, int flags)
4767 struct vnode *vp, *mvp;
4770 lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4771 if (flags != MNT_WAIT)
4772 lkflags |= LK_NOWAIT;
4774 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4775 if ((vp->v_iflag & VI_DEFINACT) == 0) {
4779 vp->v_iflag &= ~VI_DEFINACT;
4780 vfs_deferred_inactive(vp, lkflags);
4785 vfs_want_msync(struct vnode *vp)
4787 struct vm_object *obj;
4790 * This test may be performed without any locks held.
4791 * We rely on vm_object's type stability.
4793 if (vp->v_vflag & VV_NOSYNC)
4796 return (obj != NULL && vm_object_mightbedirty(obj));
4800 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4803 if (vp->v_vflag & VV_NOSYNC)
4805 if (vp->v_iflag & VI_DEFINACT)
4807 return (vfs_want_msync(vp));
4810 static void __noinline
4811 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4813 struct vnode *vp, *mvp;
4814 struct vm_object *obj;
4815 int lkflags, objflags;
4818 lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4819 if (flags != MNT_WAIT) {
4820 lkflags |= LK_NOWAIT;
4821 objflags = OBJPC_NOSYNC;
4823 objflags = OBJPC_SYNC;
4826 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4828 if (vp->v_iflag & VI_DEFINACT) {
4829 vp->v_iflag &= ~VI_DEFINACT;
4832 if (!vfs_want_msync(vp)) {
4834 vfs_deferred_inactive(vp, lkflags);
4839 if (vget(vp, lkflags) == 0) {
4841 if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4842 VM_OBJECT_WLOCK(obj);
4843 vm_object_page_clean(obj, 0, 0, objflags);
4844 VM_OBJECT_WUNLOCK(obj);
4851 vdefer_inactive_unlocked(vp);
4857 vfs_periodic(struct mount *mp, int flags)
4860 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4862 if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4863 vfs_periodic_inactive(mp, flags);
4865 vfs_periodic_msync_inactive(mp, flags);
4869 destroy_vpollinfo_free(struct vpollinfo *vi)
4872 knlist_destroy(&vi->vpi_selinfo.si_note);
4873 mtx_destroy(&vi->vpi_lock);
4874 free(vi, M_VNODEPOLL);
4878 destroy_vpollinfo(struct vpollinfo *vi)
4881 knlist_clear(&vi->vpi_selinfo.si_note, 1);
4882 seldrain(&vi->vpi_selinfo);
4883 destroy_vpollinfo_free(vi);
4887 * Initialize per-vnode helper structure to hold poll-related state.
4890 v_addpollinfo(struct vnode *vp)
4892 struct vpollinfo *vi;
4894 if (vp->v_pollinfo != NULL)
4896 vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO);
4897 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4898 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4899 vfs_knlunlock, vfs_knl_assert_lock);
4901 if (vp->v_pollinfo != NULL) {
4903 destroy_vpollinfo_free(vi);
4906 vp->v_pollinfo = vi;
4911 * Record a process's interest in events which might happen to
4912 * a vnode. Because poll uses the historic select-style interface
4913 * internally, this routine serves as both the ``check for any
4914 * pending events'' and the ``record my interest in future events''
4915 * functions. (These are done together, while the lock is held,
4916 * to avoid race conditions.)
4919 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4923 mtx_lock(&vp->v_pollinfo->vpi_lock);
4924 if (vp->v_pollinfo->vpi_revents & events) {
4926 * This leaves events we are not interested
4927 * in available for the other process which
4928 * which presumably had requested them
4929 * (otherwise they would never have been
4932 events &= vp->v_pollinfo->vpi_revents;
4933 vp->v_pollinfo->vpi_revents &= ~events;
4935 mtx_unlock(&vp->v_pollinfo->vpi_lock);
4938 vp->v_pollinfo->vpi_events |= events;
4939 selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4940 mtx_unlock(&vp->v_pollinfo->vpi_lock);
4945 * Routine to create and manage a filesystem syncer vnode.
4947 #define sync_close ((int (*)(struct vop_close_args *))nullop)
4948 static int sync_fsync(struct vop_fsync_args *);
4949 static int sync_inactive(struct vop_inactive_args *);
4950 static int sync_reclaim(struct vop_reclaim_args *);
4952 static struct vop_vector sync_vnodeops = {
4953 .vop_bypass = VOP_EOPNOTSUPP,
4954 .vop_close = sync_close, /* close */
4955 .vop_fsync = sync_fsync, /* fsync */
4956 .vop_inactive = sync_inactive, /* inactive */
4957 .vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4958 .vop_reclaim = sync_reclaim, /* reclaim */
4959 .vop_lock1 = vop_stdlock, /* lock */
4960 .vop_unlock = vop_stdunlock, /* unlock */
4961 .vop_islocked = vop_stdislocked, /* islocked */
4963 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4966 * Create a new filesystem syncer vnode for the specified mount point.
4969 vfs_allocate_syncvnode(struct mount *mp)
4973 static long start, incr, next;
4976 /* Allocate a new vnode */
4977 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4979 panic("vfs_allocate_syncvnode: getnewvnode() failed");
4981 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4982 vp->v_vflag |= VV_FORCEINSMQ;
4983 error = insmntque(vp, mp);
4985 panic("vfs_allocate_syncvnode: insmntque() failed");
4986 vp->v_vflag &= ~VV_FORCEINSMQ;
4989 * Place the vnode onto the syncer worklist. We attempt to
4990 * scatter them about on the list so that they will go off
4991 * at evenly distributed times even if all the filesystems
4992 * are mounted at once.
4995 if (next == 0 || next > syncer_maxdelay) {
4999 start = syncer_maxdelay / 2;
5000 incr = syncer_maxdelay;
5006 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
5007 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
5008 mtx_lock(&sync_mtx);
5010 if (mp->mnt_syncer == NULL) {
5011 mp->mnt_syncer = vp;
5014 mtx_unlock(&sync_mtx);
5017 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5024 vfs_deallocate_syncvnode(struct mount *mp)
5028 mtx_lock(&sync_mtx);
5029 vp = mp->mnt_syncer;
5031 mp->mnt_syncer = NULL;
5032 mtx_unlock(&sync_mtx);
5038 * Do a lazy sync of the filesystem.
5041 sync_fsync(struct vop_fsync_args *ap)
5043 struct vnode *syncvp = ap->a_vp;
5044 struct mount *mp = syncvp->v_mount;
5049 * We only need to do something if this is a lazy evaluation.
5051 if (ap->a_waitfor != MNT_LAZY)
5055 * Move ourselves to the back of the sync list.
5057 bo = &syncvp->v_bufobj;
5059 vn_syncer_add_to_worklist(bo, syncdelay);
5063 * Walk the list of vnodes pushing all that are dirty and
5064 * not already on the sync list.
5066 if (vfs_busy(mp, MBF_NOWAIT) != 0)
5068 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
5072 save = curthread_pflags_set(TDP_SYNCIO);
5074 * The filesystem at hand may be idle with free vnodes stored in the
5075 * batch. Return them instead of letting them stay there indefinitely.
5077 vfs_periodic(mp, MNT_NOWAIT);
5078 error = VFS_SYNC(mp, MNT_LAZY);
5079 curthread_pflags_restore(save);
5080 vn_finished_write(mp);
5086 * The syncer vnode is no referenced.
5089 sync_inactive(struct vop_inactive_args *ap)
5097 * The syncer vnode is no longer needed and is being decommissioned.
5099 * Modifications to the worklist must be protected by sync_mtx.
5102 sync_reclaim(struct vop_reclaim_args *ap)
5104 struct vnode *vp = ap->a_vp;
5109 mtx_lock(&sync_mtx);
5110 if (vp->v_mount->mnt_syncer == vp)
5111 vp->v_mount->mnt_syncer = NULL;
5112 if (bo->bo_flag & BO_ONWORKLST) {
5113 LIST_REMOVE(bo, bo_synclist);
5114 syncer_worklist_len--;
5116 bo->bo_flag &= ~BO_ONWORKLST;
5118 mtx_unlock(&sync_mtx);
5125 vn_need_pageq_flush(struct vnode *vp)
5127 struct vm_object *obj;
5130 return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5131 vm_object_mightbedirty(obj));
5135 * Check if vnode represents a disk device
5138 vn_isdisk_error(struct vnode *vp, int *errp)
5142 if (vp->v_type != VCHR) {
5148 if (vp->v_rdev == NULL)
5150 else if (vp->v_rdev->si_devsw == NULL)
5152 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5157 return (error == 0);
5161 vn_isdisk(struct vnode *vp)
5165 return (vn_isdisk_error(vp, &error));
5169 * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
5170 * the comment above cache_fplookup for details.
5173 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred)
5177 VFS_SMR_ASSERT_ENTERED();
5179 /* Check the owner. */
5180 if (cred->cr_uid == file_uid) {
5181 if (file_mode & S_IXUSR)
5186 /* Otherwise, check the groups (first match) */
5187 if (groupmember(file_gid, cred)) {
5188 if (file_mode & S_IXGRP)
5193 /* Otherwise, check everyone else. */
5194 if (file_mode & S_IXOTH)
5198 * Permission check failed, but it is possible denial will get overwritten
5199 * (e.g., when root is traversing through a 700 directory owned by someone
5202 * vaccess() calls priv_check_cred which in turn can descent into MAC
5203 * modules overriding this result. It's quite unclear what semantics
5204 * are allowed for them to operate, thus for safety we don't call them
5205 * from within the SMR section. This also means if any such modules
5206 * are present, we have to let the regular lookup decide.
5208 error = priv_check_cred_vfs_lookup_nomac(cred);
5214 * MAC modules present.
5225 * Common filesystem object access control check routine. Accepts a
5226 * vnode's type, "mode", uid and gid, requested access mode, and credentials.
5227 * Returns 0 on success, or an errno on failure.
5230 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5231 accmode_t accmode, struct ucred *cred)
5233 accmode_t dac_granted;
5234 accmode_t priv_granted;
5236 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5237 ("invalid bit in accmode"));
5238 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5239 ("VAPPEND without VWRITE"));
5242 * Look for a normal, non-privileged way to access the file/directory
5243 * as requested. If it exists, go with that.
5248 /* Check the owner. */
5249 if (cred->cr_uid == file_uid) {
5250 dac_granted |= VADMIN;
5251 if (file_mode & S_IXUSR)
5252 dac_granted |= VEXEC;
5253 if (file_mode & S_IRUSR)
5254 dac_granted |= VREAD;
5255 if (file_mode & S_IWUSR)
5256 dac_granted |= (VWRITE | VAPPEND);
5258 if ((accmode & dac_granted) == accmode)
5264 /* Otherwise, check the groups (first match) */
5265 if (groupmember(file_gid, cred)) {
5266 if (file_mode & S_IXGRP)
5267 dac_granted |= VEXEC;
5268 if (file_mode & S_IRGRP)
5269 dac_granted |= VREAD;
5270 if (file_mode & S_IWGRP)
5271 dac_granted |= (VWRITE | VAPPEND);
5273 if ((accmode & dac_granted) == accmode)
5279 /* Otherwise, check everyone else. */
5280 if (file_mode & S_IXOTH)
5281 dac_granted |= VEXEC;
5282 if (file_mode & S_IROTH)
5283 dac_granted |= VREAD;
5284 if (file_mode & S_IWOTH)
5285 dac_granted |= (VWRITE | VAPPEND);
5286 if ((accmode & dac_granted) == accmode)
5291 * Build a privilege mask to determine if the set of privileges
5292 * satisfies the requirements when combined with the granted mask
5293 * from above. For each privilege, if the privilege is required,
5294 * bitwise or the request type onto the priv_granted mask.
5300 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5301 * requests, instead of PRIV_VFS_EXEC.
5303 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5304 !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5305 priv_granted |= VEXEC;
5308 * Ensure that at least one execute bit is on. Otherwise,
5309 * a privileged user will always succeed, and we don't want
5310 * this to happen unless the file really is executable.
5312 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5313 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5314 !priv_check_cred(cred, PRIV_VFS_EXEC))
5315 priv_granted |= VEXEC;
5318 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5319 !priv_check_cred(cred, PRIV_VFS_READ))
5320 priv_granted |= VREAD;
5322 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5323 !priv_check_cred(cred, PRIV_VFS_WRITE))
5324 priv_granted |= (VWRITE | VAPPEND);
5326 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5327 !priv_check_cred(cred, PRIV_VFS_ADMIN))
5328 priv_granted |= VADMIN;
5330 if ((accmode & (priv_granted | dac_granted)) == accmode) {
5334 return ((accmode & VADMIN) ? EPERM : EACCES);
5338 * Credential check based on process requesting service, and per-attribute
5342 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5343 struct thread *td, accmode_t accmode)
5347 * Kernel-invoked always succeeds.
5353 * Do not allow privileged processes in jail to directly manipulate
5354 * system attributes.
5356 switch (attrnamespace) {
5357 case EXTATTR_NAMESPACE_SYSTEM:
5358 /* Potentially should be: return (EPERM); */
5359 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5360 case EXTATTR_NAMESPACE_USER:
5361 return (VOP_ACCESS(vp, accmode, cred, td));
5367 #ifdef DEBUG_VFS_LOCKS
5368 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */
5369 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5370 "Drop into debugger on lock violation");
5372 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */
5373 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5374 0, "Check for interlock across VOPs");
5376 int vfs_badlock_print = 1; /* Print lock violations. */
5377 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5378 0, "Print lock violations");
5380 int vfs_badlock_vnode = 1; /* Print vnode details on lock violations. */
5381 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5382 0, "Print vnode details on lock violations");
5385 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */
5386 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5387 &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5391 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5395 if (vfs_badlock_backtrace)
5398 if (vfs_badlock_vnode)
5399 vn_printf(vp, "vnode ");
5400 if (vfs_badlock_print)
5401 printf("%s: %p %s\n", str, (void *)vp, msg);
5402 if (vfs_badlock_ddb)
5403 kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5407 assert_vi_locked(struct vnode *vp, const char *str)
5410 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5411 vfs_badlock("interlock is not locked but should be", str, vp);
5415 assert_vi_unlocked(struct vnode *vp, const char *str)
5418 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5419 vfs_badlock("interlock is locked but should not be", str, vp);
5423 assert_vop_locked(struct vnode *vp, const char *str)
5427 if (KERNEL_PANICKED() || vp == NULL)
5430 locked = VOP_ISLOCKED(vp);
5431 if (locked == 0 || locked == LK_EXCLOTHER)
5432 vfs_badlock("is not locked but should be", str, vp);
5436 assert_vop_unlocked(struct vnode *vp, const char *str)
5438 if (KERNEL_PANICKED() || vp == NULL)
5441 if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5442 vfs_badlock("is locked but should not be", str, vp);
5446 assert_vop_elocked(struct vnode *vp, const char *str)
5448 if (KERNEL_PANICKED() || vp == NULL)
5451 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5452 vfs_badlock("is not exclusive locked but should be", str, vp);
5454 #endif /* DEBUG_VFS_LOCKS */
5457 vop_rename_fail(struct vop_rename_args *ap)
5460 if (ap->a_tvp != NULL)
5462 if (ap->a_tdvp == ap->a_tvp)
5471 vop_rename_pre(void *ap)
5473 struct vop_rename_args *a = ap;
5475 #ifdef DEBUG_VFS_LOCKS
5477 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5478 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5479 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5480 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5482 /* Check the source (from). */
5483 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5484 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5485 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5486 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5487 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5489 /* Check the target. */
5491 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5492 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5495 * It may be tempting to add vn_seqc_write_begin/end calls here and
5496 * in vop_rename_post but that's not going to work out since some
5497 * filesystems relookup vnodes mid-rename. This is probably a bug.
5499 * For now filesystems are expected to do the relevant calls after they
5500 * decide what vnodes to operate on.
5502 if (a->a_tdvp != a->a_fdvp)
5504 if (a->a_tvp != a->a_fvp)
5511 #ifdef DEBUG_VFS_LOCKS
5513 vop_fplookup_vexec_debugpre(void *ap __unused)
5516 VFS_SMR_ASSERT_ENTERED();
5520 vop_fplookup_vexec_debugpost(void *ap __unused, int rc __unused)
5523 VFS_SMR_ASSERT_ENTERED();
5527 vop_fplookup_symlink_debugpre(void *ap __unused)
5530 VFS_SMR_ASSERT_ENTERED();
5534 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused)
5537 VFS_SMR_ASSERT_ENTERED();
5541 vop_fsync_debugprepost(struct vnode *vp, const char *name)
5543 if (vp->v_type == VCHR)
5545 else if (MNT_EXTENDED_SHARED(vp->v_mount))
5546 ASSERT_VOP_LOCKED(vp, name);
5548 ASSERT_VOP_ELOCKED(vp, name);
5552 vop_fsync_debugpre(void *a)
5554 struct vop_fsync_args *ap;
5557 vop_fsync_debugprepost(ap->a_vp, "fsync");
5561 vop_fsync_debugpost(void *a, int rc __unused)
5563 struct vop_fsync_args *ap;
5566 vop_fsync_debugprepost(ap->a_vp, "fsync");
5570 vop_fdatasync_debugpre(void *a)
5572 struct vop_fdatasync_args *ap;
5575 vop_fsync_debugprepost(ap->a_vp, "fsync");
5579 vop_fdatasync_debugpost(void *a, int rc __unused)
5581 struct vop_fdatasync_args *ap;
5584 vop_fsync_debugprepost(ap->a_vp, "fsync");
5588 vop_strategy_debugpre(void *ap)
5590 struct vop_strategy_args *a;
5597 * Cluster ops lock their component buffers but not the IO container.
5599 if ((bp->b_flags & B_CLUSTER) != 0)
5602 if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5603 if (vfs_badlock_print)
5605 "VOP_STRATEGY: bp is not locked but should be\n");
5606 if (vfs_badlock_ddb)
5607 kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5612 vop_lock_debugpre(void *ap)
5614 struct vop_lock1_args *a = ap;
5616 if ((a->a_flags & LK_INTERLOCK) == 0)
5617 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5619 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5623 vop_lock_debugpost(void *ap, int rc)
5625 struct vop_lock1_args *a = ap;
5627 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5628 if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5629 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5633 vop_unlock_debugpre(void *ap)
5635 struct vop_unlock_args *a = ap;
5637 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5641 vop_need_inactive_debugpre(void *ap)
5643 struct vop_need_inactive_args *a = ap;
5645 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5649 vop_need_inactive_debugpost(void *ap, int rc)
5651 struct vop_need_inactive_args *a = ap;
5653 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5658 vop_create_pre(void *ap)
5660 struct vop_create_args *a;
5665 vn_seqc_write_begin(dvp);
5669 vop_create_post(void *ap, int rc)
5671 struct vop_create_args *a;
5676 vn_seqc_write_end(dvp);
5678 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5682 vop_whiteout_pre(void *ap)
5684 struct vop_whiteout_args *a;
5689 vn_seqc_write_begin(dvp);
5693 vop_whiteout_post(void *ap, int rc)
5695 struct vop_whiteout_args *a;
5700 vn_seqc_write_end(dvp);
5704 vop_deleteextattr_pre(void *ap)
5706 struct vop_deleteextattr_args *a;
5711 vn_seqc_write_begin(vp);
5715 vop_deleteextattr_post(void *ap, int rc)
5717 struct vop_deleteextattr_args *a;
5722 vn_seqc_write_end(vp);
5724 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5728 vop_link_pre(void *ap)
5730 struct vop_link_args *a;
5731 struct vnode *vp, *tdvp;
5736 vn_seqc_write_begin(vp);
5737 vn_seqc_write_begin(tdvp);
5741 vop_link_post(void *ap, int rc)
5743 struct vop_link_args *a;
5744 struct vnode *vp, *tdvp;
5749 vn_seqc_write_end(vp);
5750 vn_seqc_write_end(tdvp);
5752 VFS_KNOTE_LOCKED(vp, NOTE_LINK);
5753 VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE);
5758 vop_mkdir_pre(void *ap)
5760 struct vop_mkdir_args *a;
5765 vn_seqc_write_begin(dvp);
5769 vop_mkdir_post(void *ap, int rc)
5771 struct vop_mkdir_args *a;
5776 vn_seqc_write_end(dvp);
5778 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5781 #ifdef DEBUG_VFS_LOCKS
5783 vop_mkdir_debugpost(void *ap, int rc)
5785 struct vop_mkdir_args *a;
5789 cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp);
5794 vop_mknod_pre(void *ap)
5796 struct vop_mknod_args *a;
5801 vn_seqc_write_begin(dvp);
5805 vop_mknod_post(void *ap, int rc)
5807 struct vop_mknod_args *a;
5812 vn_seqc_write_end(dvp);
5814 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5818 vop_reclaim_post(void *ap, int rc)
5820 struct vop_reclaim_args *a;
5825 ASSERT_VOP_IN_SEQC(vp);
5827 VFS_KNOTE_LOCKED(vp, NOTE_REVOKE);
5831 vop_remove_pre(void *ap)
5833 struct vop_remove_args *a;
5834 struct vnode *dvp, *vp;
5839 vn_seqc_write_begin(dvp);
5840 vn_seqc_write_begin(vp);
5844 vop_remove_post(void *ap, int rc)
5846 struct vop_remove_args *a;
5847 struct vnode *dvp, *vp;
5852 vn_seqc_write_end(dvp);
5853 vn_seqc_write_end(vp);
5855 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5856 VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5861 vop_rename_post(void *ap, int rc)
5863 struct vop_rename_args *a = ap;
5868 if (a->a_fdvp == a->a_tdvp) {
5869 if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5871 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5872 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5874 hint |= NOTE_EXTEND;
5875 if (a->a_fvp->v_type == VDIR)
5877 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5879 if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5880 a->a_tvp->v_type == VDIR)
5882 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5885 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5887 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5889 if (a->a_tdvp != a->a_fdvp)
5891 if (a->a_tvp != a->a_fvp)
5899 vop_rmdir_pre(void *ap)
5901 struct vop_rmdir_args *a;
5902 struct vnode *dvp, *vp;
5907 vn_seqc_write_begin(dvp);
5908 vn_seqc_write_begin(vp);
5912 vop_rmdir_post(void *ap, int rc)
5914 struct vop_rmdir_args *a;
5915 struct vnode *dvp, *vp;
5920 vn_seqc_write_end(dvp);
5921 vn_seqc_write_end(vp);
5923 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5924 VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5929 vop_setattr_pre(void *ap)
5931 struct vop_setattr_args *a;
5936 vn_seqc_write_begin(vp);
5940 vop_setattr_post(void *ap, int rc)
5942 struct vop_setattr_args *a;
5947 vn_seqc_write_end(vp);
5949 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5953 vop_setacl_pre(void *ap)
5955 struct vop_setacl_args *a;
5960 vn_seqc_write_begin(vp);
5964 vop_setacl_post(void *ap, int rc __unused)
5966 struct vop_setacl_args *a;
5971 vn_seqc_write_end(vp);
5975 vop_setextattr_pre(void *ap)
5977 struct vop_setextattr_args *a;
5982 vn_seqc_write_begin(vp);
5986 vop_setextattr_post(void *ap, int rc)
5988 struct vop_setextattr_args *a;
5993 vn_seqc_write_end(vp);
5995 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5999 vop_symlink_pre(void *ap)
6001 struct vop_symlink_args *a;
6006 vn_seqc_write_begin(dvp);
6010 vop_symlink_post(void *ap, int rc)
6012 struct vop_symlink_args *a;
6017 vn_seqc_write_end(dvp);
6019 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6023 vop_open_post(void *ap, int rc)
6025 struct vop_open_args *a = ap;
6028 VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
6032 vop_close_post(void *ap, int rc)
6034 struct vop_close_args *a = ap;
6036 if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
6037 !VN_IS_DOOMED(a->a_vp))) {
6038 VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
6039 NOTE_CLOSE_WRITE : NOTE_CLOSE);
6044 vop_read_post(void *ap, int rc)
6046 struct vop_read_args *a = ap;
6049 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6053 vop_read_pgcache_post(void *ap, int rc)
6055 struct vop_read_pgcache_args *a = ap;
6058 VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ);
6062 vop_readdir_post(void *ap, int rc)
6064 struct vop_readdir_args *a = ap;
6067 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6070 static struct knlist fs_knlist;
6073 vfs_event_init(void *arg)
6075 knlist_init_mtx(&fs_knlist, NULL);
6077 /* XXX - correct order? */
6078 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
6081 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
6084 KNOTE_UNLOCKED(&fs_knlist, event);
6087 static int filt_fsattach(struct knote *kn);
6088 static void filt_fsdetach(struct knote *kn);
6089 static int filt_fsevent(struct knote *kn, long hint);
6091 struct filterops fs_filtops = {
6093 .f_attach = filt_fsattach,
6094 .f_detach = filt_fsdetach,
6095 .f_event = filt_fsevent
6099 filt_fsattach(struct knote *kn)
6102 kn->kn_flags |= EV_CLEAR;
6103 knlist_add(&fs_knlist, kn, 0);
6108 filt_fsdetach(struct knote *kn)
6111 knlist_remove(&fs_knlist, kn, 0);
6115 filt_fsevent(struct knote *kn, long hint)
6118 kn->kn_fflags |= kn->kn_sfflags & hint;
6120 return (kn->kn_fflags != 0);
6124 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
6130 error = SYSCTL_IN(req, &vc, sizeof(vc));
6133 if (vc.vc_vers != VFS_CTL_VERS1)
6135 mp = vfs_getvfs(&vc.vc_fsid);
6138 /* ensure that a specific sysctl goes to the right filesystem. */
6139 if (strcmp(vc.vc_fstypename, "*") != 0 &&
6140 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
6144 VCTLTOREQ(&vc, req);
6145 error = VFS_SYSCTL(mp, vc.vc_op, req);
6150 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
6151 NULL, 0, sysctl_vfs_ctl, "",
6155 * Function to initialize a va_filerev field sensibly.
6156 * XXX: Wouldn't a random number make a lot more sense ??
6159 init_va_filerev(void)
6164 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
6167 static int filt_vfsread(struct knote *kn, long hint);
6168 static int filt_vfswrite(struct knote *kn, long hint);
6169 static int filt_vfsvnode(struct knote *kn, long hint);
6170 static void filt_vfsdetach(struct knote *kn);
6171 static struct filterops vfsread_filtops = {
6173 .f_detach = filt_vfsdetach,
6174 .f_event = filt_vfsread
6176 static struct filterops vfswrite_filtops = {
6178 .f_detach = filt_vfsdetach,
6179 .f_event = filt_vfswrite
6181 static struct filterops vfsvnode_filtops = {
6183 .f_detach = filt_vfsdetach,
6184 .f_event = filt_vfsvnode
6188 vfs_knllock(void *arg)
6190 struct vnode *vp = arg;
6192 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
6196 vfs_knlunlock(void *arg)
6198 struct vnode *vp = arg;
6204 vfs_knl_assert_lock(void *arg, int what)
6206 #ifdef DEBUG_VFS_LOCKS
6207 struct vnode *vp = arg;
6209 if (what == LA_LOCKED)
6210 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
6212 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
6217 vfs_kqfilter(struct vop_kqfilter_args *ap)
6219 struct vnode *vp = ap->a_vp;
6220 struct knote *kn = ap->a_kn;
6223 KASSERT(vp->v_type != VFIFO || (kn->kn_filter != EVFILT_READ &&
6224 kn->kn_filter != EVFILT_WRITE),
6225 ("READ/WRITE filter on a FIFO leaked through"));
6226 switch (kn->kn_filter) {
6228 kn->kn_fop = &vfsread_filtops;
6231 kn->kn_fop = &vfswrite_filtops;
6234 kn->kn_fop = &vfsvnode_filtops;
6240 kn->kn_hook = (caddr_t)vp;
6243 if (vp->v_pollinfo == NULL)
6245 knl = &vp->v_pollinfo->vpi_selinfo.si_note;
6247 knlist_add(knl, kn, 0);
6253 * Detach knote from vnode
6256 filt_vfsdetach(struct knote *kn)
6258 struct vnode *vp = (struct vnode *)kn->kn_hook;
6260 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
6261 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
6267 filt_vfsread(struct knote *kn, long hint)
6269 struct vnode *vp = (struct vnode *)kn->kn_hook;
6274 * filesystem is gone, so set the EOF flag and schedule
6275 * the knote for deletion.
6277 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6279 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6284 if (VOP_GETATTR(vp, &va, curthread->td_ucred))
6288 kn->kn_data = va.va_size - kn->kn_fp->f_offset;
6289 res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
6296 filt_vfswrite(struct knote *kn, long hint)
6298 struct vnode *vp = (struct vnode *)kn->kn_hook;
6303 * filesystem is gone, so set the EOF flag and schedule
6304 * the knote for deletion.
6306 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
6307 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6315 filt_vfsvnode(struct knote *kn, long hint)
6317 struct vnode *vp = (struct vnode *)kn->kn_hook;
6321 if (kn->kn_sfflags & hint)
6322 kn->kn_fflags |= hint;
6323 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6324 kn->kn_flags |= EV_EOF;
6328 res = (kn->kn_fflags != 0);
6334 * Returns whether the directory is empty or not.
6335 * If it is empty, the return value is 0; otherwise
6336 * the return value is an error value (which may
6340 vfs_emptydir(struct vnode *vp)
6344 struct dirent *dirent, *dp, *endp;
6350 ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
6351 VNASSERT(vp->v_type == VDIR, vp, ("vp is not a directory"));
6353 dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
6354 iov.iov_base = dirent;
6355 iov.iov_len = sizeof(struct dirent);
6360 uio.uio_resid = sizeof(struct dirent);
6361 uio.uio_segflg = UIO_SYSSPACE;
6362 uio.uio_rw = UIO_READ;
6363 uio.uio_td = curthread;
6365 while (eof == 0 && error == 0) {
6366 error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
6370 endp = (void *)((uint8_t *)dirent +
6371 sizeof(struct dirent) - uio.uio_resid);
6372 for (dp = dirent; dp < endp;
6373 dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
6374 if (dp->d_type == DT_WHT)
6376 if (dp->d_namlen == 0)
6378 if (dp->d_type != DT_DIR &&
6379 dp->d_type != DT_UNKNOWN) {
6383 if (dp->d_namlen > 2) {
6387 if (dp->d_namlen == 1 &&
6388 dp->d_name[0] != '.') {
6392 if (dp->d_namlen == 2 &&
6393 dp->d_name[1] != '.') {
6397 uio.uio_resid = sizeof(struct dirent);
6400 free(dirent, M_TEMP);
6405 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6409 if (dp->d_reclen > ap->a_uio->uio_resid)
6410 return (ENAMETOOLONG);
6411 error = uiomove(dp, dp->d_reclen, ap->a_uio);
6413 if (ap->a_ncookies != NULL) {
6414 if (ap->a_cookies != NULL)
6415 free(ap->a_cookies, M_TEMP);
6416 ap->a_cookies = NULL;
6417 *ap->a_ncookies = 0;
6421 if (ap->a_ncookies == NULL)
6424 KASSERT(ap->a_cookies,
6425 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6427 *ap->a_cookies = realloc(*ap->a_cookies,
6428 (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
6429 (*ap->a_cookies)[*ap->a_ncookies] = off;
6430 *ap->a_ncookies += 1;
6435 * The purpose of this routine is to remove granularity from accmode_t,
6436 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6437 * VADMIN and VAPPEND.
6439 * If it returns 0, the caller is supposed to continue with the usual
6440 * access checks using 'accmode' as modified by this routine. If it
6441 * returns nonzero value, the caller is supposed to return that value
6444 * Note that after this routine runs, accmode may be zero.
6447 vfs_unixify_accmode(accmode_t *accmode)
6450 * There is no way to specify explicit "deny" rule using
6451 * file mode or POSIX.1e ACLs.
6453 if (*accmode & VEXPLICIT_DENY) {
6459 * None of these can be translated into usual access bits.
6460 * Also, the common case for NFSv4 ACLs is to not contain
6461 * either of these bits. Caller should check for VWRITE
6462 * on the containing directory instead.
6464 if (*accmode & (VDELETE_CHILD | VDELETE))
6467 if (*accmode & VADMIN_PERMS) {
6468 *accmode &= ~VADMIN_PERMS;
6473 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6474 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6476 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6482 * Clear out a doomed vnode (if any) and replace it with a new one as long
6483 * as the fs is not being unmounted. Return the root vnode to the caller.
6485 static int __noinline
6486 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6492 if (mp->mnt_rootvnode != NULL) {
6494 vp = mp->mnt_rootvnode;
6496 if (!VN_IS_DOOMED(vp)) {
6499 error = vn_lock(vp, flags);
6508 * Clear the old one.
6510 mp->mnt_rootvnode = NULL;
6514 vfs_op_barrier_wait(mp);
6518 error = VFS_CACHEDROOT(mp, flags, vpp);
6521 if (mp->mnt_vfs_ops == 0) {
6523 if (mp->mnt_vfs_ops != 0) {
6527 if (mp->mnt_rootvnode == NULL) {
6529 mp->mnt_rootvnode = *vpp;
6531 if (mp->mnt_rootvnode != *vpp) {
6532 if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6533 panic("%s: mismatch between vnode returned "
6534 " by VFS_CACHEDROOT and the one cached "
6536 __func__, *vpp, mp->mnt_rootvnode);
6546 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6548 struct mount_pcpu *mpcpu;
6552 if (!vfs_op_thread_enter(mp, mpcpu))
6553 return (vfs_cache_root_fallback(mp, flags, vpp));
6554 vp = atomic_load_ptr(&mp->mnt_rootvnode);
6555 if (vp == NULL || VN_IS_DOOMED(vp)) {
6556 vfs_op_thread_exit(mp, mpcpu);
6557 return (vfs_cache_root_fallback(mp, flags, vpp));
6560 vfs_op_thread_exit(mp, mpcpu);
6561 error = vn_lock(vp, flags);
6564 return (vfs_cache_root_fallback(mp, flags, vpp));
6571 vfs_cache_root_clear(struct mount *mp)
6576 * ops > 0 guarantees there is nobody who can see this vnode
6578 MPASS(mp->mnt_vfs_ops > 0);
6579 vp = mp->mnt_rootvnode;
6581 vn_seqc_write_begin(vp);
6582 mp->mnt_rootvnode = NULL;
6587 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6590 MPASS(mp->mnt_vfs_ops > 0);
6592 mp->mnt_rootvnode = vp;
6596 * These are helper functions for filesystems to traverse all
6597 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6599 * This interface replaces MNT_VNODE_FOREACH.
6603 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6608 kern_yield(PRI_USER);
6610 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6611 for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6612 vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6613 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6614 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6617 if (VN_IS_DOOMED(vp)) {
6624 __mnt_vnode_markerfree_all(mvp, mp);
6625 /* MNT_IUNLOCK(mp); -- done in above function */
6626 mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6629 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6630 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6636 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6640 *mvp = vn_alloc_marker(mp);
6644 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6645 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6646 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6649 if (VN_IS_DOOMED(vp)) {
6658 vn_free_marker(*mvp);
6662 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6668 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6676 mtx_assert(MNT_MTX(mp), MA_OWNED);
6678 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6679 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6682 vn_free_marker(*mvp);
6687 * These are helper functions for filesystems to traverse their
6688 * lazy vnodes. See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6691 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6694 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6699 vn_free_marker(*mvp);
6704 * Relock the mp mount vnode list lock with the vp vnode interlock in the
6705 * conventional lock order during mnt_vnode_next_lazy iteration.
6707 * On entry, the mount vnode list lock is held and the vnode interlock is not.
6708 * The list lock is dropped and reacquired. On success, both locks are held.
6709 * On failure, the mount vnode list lock is held but the vnode interlock is
6710 * not, and the procedure may have yielded.
6713 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6717 VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6718 TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6719 ("%s: bad marker", __func__));
6720 VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6721 ("%s: inappropriate vnode", __func__));
6722 ASSERT_VI_UNLOCKED(vp, __func__);
6723 mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6725 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6726 TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6729 * Note we may be racing against vdrop which transitioned the hold
6730 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6731 * if we are the only user after we get the interlock we will just
6735 mtx_unlock(&mp->mnt_listmtx);
6737 if (VN_IS_DOOMED(vp)) {
6738 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6741 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6743 * There is nothing to do if we are the last user.
6745 if (!refcount_release_if_not_last(&vp->v_holdcnt))
6747 mtx_lock(&mp->mnt_listmtx);
6752 mtx_lock(&mp->mnt_listmtx);
6756 static struct vnode *
6757 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6762 mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6763 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6765 vp = TAILQ_NEXT(*mvp, v_lazylist);
6766 while (vp != NULL) {
6767 if (vp->v_type == VMARKER) {
6768 vp = TAILQ_NEXT(vp, v_lazylist);
6772 * See if we want to process the vnode. Note we may encounter a
6773 * long string of vnodes we don't care about and hog the list
6774 * as a result. Check for it and requeue the marker.
6776 VNPASS(!VN_IS_DOOMED(vp), vp);
6777 if (!cb(vp, cbarg)) {
6778 if (!should_yield()) {
6779 vp = TAILQ_NEXT(vp, v_lazylist);
6782 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6784 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6786 mtx_unlock(&mp->mnt_listmtx);
6787 kern_yield(PRI_USER);
6788 mtx_lock(&mp->mnt_listmtx);
6792 * Try-lock because this is the wrong lock order.
6794 if (!VI_TRYLOCK(vp) &&
6795 !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6797 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6798 KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6799 ("alien vnode on the lazy list %p %p", vp, mp));
6800 VNPASS(vp->v_mount == mp, vp);
6801 VNPASS(!VN_IS_DOOMED(vp), vp);
6804 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6806 /* Check if we are done */
6808 mtx_unlock(&mp->mnt_listmtx);
6809 mnt_vnode_markerfree_lazy(mvp, mp);
6812 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6813 mtx_unlock(&mp->mnt_listmtx);
6814 ASSERT_VI_LOCKED(vp, "lazy iter");
6819 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6824 kern_yield(PRI_USER);
6825 mtx_lock(&mp->mnt_listmtx);
6826 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6830 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6835 if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6838 *mvp = vn_alloc_marker(mp);
6843 mtx_lock(&mp->mnt_listmtx);
6844 vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6846 mtx_unlock(&mp->mnt_listmtx);
6847 mnt_vnode_markerfree_lazy(mvp, mp);
6850 TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6851 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6855 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6861 mtx_lock(&mp->mnt_listmtx);
6862 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6863 mtx_unlock(&mp->mnt_listmtx);
6864 mnt_vnode_markerfree_lazy(mvp, mp);
6868 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6871 if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6872 cnp->cn_flags &= ~NOEXECCHECK;
6876 return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, curthread));
6880 * Do not use this variant unless you have means other than the hold count
6881 * to prevent the vnode from getting freed.
6884 vn_seqc_write_begin_locked(struct vnode *vp)
6887 ASSERT_VI_LOCKED(vp, __func__);
6888 VNPASS(vp->v_holdcnt > 0, vp);
6889 VNPASS(vp->v_seqc_users >= 0, vp);
6891 if (vp->v_seqc_users == 1)
6892 seqc_sleepable_write_begin(&vp->v_seqc);
6896 vn_seqc_write_begin(struct vnode *vp)
6900 vn_seqc_write_begin_locked(vp);
6905 vn_seqc_write_end_locked(struct vnode *vp)
6908 ASSERT_VI_LOCKED(vp, __func__);
6909 VNPASS(vp->v_seqc_users > 0, vp);
6911 if (vp->v_seqc_users == 0)
6912 seqc_sleepable_write_end(&vp->v_seqc);
6916 vn_seqc_write_end(struct vnode *vp)
6920 vn_seqc_write_end_locked(vp);
6925 * Special case handling for allocating and freeing vnodes.
6927 * The counter remains unchanged on free so that a doomed vnode will
6928 * keep testing as in modify as long as it is accessible with SMR.
6931 vn_seqc_init(struct vnode *vp)
6935 vp->v_seqc_users = 0;
6939 vn_seqc_write_end_free(struct vnode *vp)
6942 VNPASS(seqc_in_modify(vp->v_seqc), vp);
6943 VNPASS(vp->v_seqc_users == 1, vp);
6947 vn_irflag_set_locked(struct vnode *vp, short toset)
6951 ASSERT_VI_LOCKED(vp, __func__);
6952 flags = vn_irflag_read(vp);
6953 VNASSERT((flags & toset) == 0, vp,
6954 ("%s: some of the passed flags already set (have %d, passed %d)\n",
6955 __func__, flags, toset));
6956 atomic_store_short(&vp->v_irflag, flags | toset);
6960 vn_irflag_set(struct vnode *vp, short toset)
6964 vn_irflag_set_locked(vp, toset);
6969 vn_irflag_set_cond_locked(struct vnode *vp, short toset)
6973 ASSERT_VI_LOCKED(vp, __func__);
6974 flags = vn_irflag_read(vp);
6975 atomic_store_short(&vp->v_irflag, flags | toset);
6979 vn_irflag_set_cond(struct vnode *vp, short toset)
6983 vn_irflag_set_cond_locked(vp, toset);
6988 vn_irflag_unset_locked(struct vnode *vp, short tounset)
6992 ASSERT_VI_LOCKED(vp, __func__);
6993 flags = vn_irflag_read(vp);
6994 VNASSERT((flags & tounset) == tounset, vp,
6995 ("%s: some of the passed flags not set (have %d, passed %d)\n",
6996 __func__, flags, tounset));
6997 atomic_store_short(&vp->v_irflag, flags & ~tounset);
7001 vn_irflag_unset(struct vnode *vp, short tounset)
7005 vn_irflag_unset_locked(vp, tounset);