2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1989, 1993
5 * The Regents of the University of California. All rights reserved.
6 * (c) UNIX System Laboratories, Inc.
7 * All or some portions of this file are derived from material licensed
8 * to the University of California by American Telephone and Telegraph
9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10 * the permission of UNIX System Laboratories, Inc.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95
40 * External virtual filesystem routines
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
47 #include "opt_watchdog.h"
49 #include <sys/param.h>
50 #include <sys/systm.h>
53 #include <sys/capsicum.h>
54 #include <sys/condvar.h>
56 #include <sys/counter.h>
57 #include <sys/dirent.h>
58 #include <sys/event.h>
59 #include <sys/eventhandler.h>
60 #include <sys/extattr.h>
62 #include <sys/fcntl.h>
65 #include <sys/kernel.h>
66 #include <sys/kthread.h>
68 #include <sys/lockf.h>
69 #include <sys/malloc.h>
70 #include <sys/mount.h>
71 #include <sys/namei.h>
72 #include <sys/pctrie.h>
74 #include <sys/reboot.h>
75 #include <sys/refcount.h>
76 #include <sys/rwlock.h>
77 #include <sys/sched.h>
78 #include <sys/sleepqueue.h>
82 #include <sys/sysctl.h>
83 #include <sys/syslog.h>
84 #include <sys/vmmeter.h>
85 #include <sys/vnode.h>
86 #include <sys/watchdog.h>
88 #include <machine/stdarg.h>
90 #include <security/mac/mac_framework.h>
93 #include <vm/vm_object.h>
94 #include <vm/vm_extern.h>
96 #include <vm/vm_map.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_kern.h>
105 static void delmntque(struct vnode *vp);
106 static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
107 int slpflag, int slptimeo);
108 static void syncer_shutdown(void *arg, int howto);
109 static int vtryrecycle(struct vnode *vp);
110 static void v_init_counters(struct vnode *);
111 static void vn_seqc_init(struct vnode *);
112 static void vn_seqc_write_end_free(struct vnode *vp);
113 static void vgonel(struct vnode *);
114 static bool vhold_recycle_free(struct vnode *);
115 static void vfs_knllock(void *arg);
116 static void vfs_knlunlock(void *arg);
117 static void vfs_knl_assert_lock(void *arg, int what);
118 static void destroy_vpollinfo(struct vpollinfo *vi);
119 static int v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
120 daddr_t startlbn, daddr_t endlbn);
121 static void vnlru_recalc(void);
124 * These fences are intended for cases where some synchronization is
125 * needed between access of v_iflags and lockless vnode refcount (v_holdcnt
126 * and v_usecount) updates. Access to v_iflags is generally synchronized
127 * by the interlock, but we have some internal assertions that check vnode
128 * flags without acquiring the lock. Thus, these fences are INVARIANTS-only
132 #define VNODE_REFCOUNT_FENCE_ACQ() atomic_thread_fence_acq()
133 #define VNODE_REFCOUNT_FENCE_REL() atomic_thread_fence_rel()
135 #define VNODE_REFCOUNT_FENCE_ACQ()
136 #define VNODE_REFCOUNT_FENCE_REL()
140 * Number of vnodes in existence. Increased whenever getnewvnode()
141 * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
143 static u_long __exclusive_cache_line numvnodes;
145 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
146 "Number of vnodes in existence");
148 static counter_u64_t vnodes_created;
149 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
150 "Number of vnodes created by getnewvnode");
153 * Conversion tables for conversion from vnode types to inode formats
156 enum vtype iftovt_tab[16] = {
157 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
158 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
160 int vttoif_tab[10] = {
161 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
162 S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
166 * List of allocates vnodes in the system.
168 static TAILQ_HEAD(freelst, vnode) vnode_list;
169 static struct vnode *vnode_list_free_marker;
170 static struct vnode *vnode_list_reclaim_marker;
173 * "Free" vnode target. Free vnodes are rarely completely free, but are
174 * just ones that are cheap to recycle. Usually they are for files which
175 * have been stat'd but not read; these usually have inode and namecache
176 * data attached to them. This target is the preferred minimum size of a
177 * sub-cache consisting mostly of such files. The system balances the size
178 * of this sub-cache with its complement to try to prevent either from
179 * thrashing while the other is relatively inactive. The targets express
180 * a preference for the best balance.
182 * "Above" this target there are 2 further targets (watermarks) related
183 * to recyling of free vnodes. In the best-operating case, the cache is
184 * exactly full, the free list has size between vlowat and vhiwat above the
185 * free target, and recycling from it and normal use maintains this state.
186 * Sometimes the free list is below vlowat or even empty, but this state
187 * is even better for immediate use provided the cache is not full.
188 * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
189 * ones) to reach one of these states. The watermarks are currently hard-
190 * coded as 4% and 9% of the available space higher. These and the default
191 * of 25% for wantfreevnodes are too large if the memory size is large.
192 * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
193 * whenever vnlru_proc() becomes active.
195 static long wantfreevnodes;
196 static long __exclusive_cache_line freevnodes;
197 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
198 &freevnodes, 0, "Number of \"free\" vnodes");
199 static long freevnodes_old;
201 static counter_u64_t recycles_count;
202 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
203 "Number of vnodes recycled to meet vnode cache targets");
205 static counter_u64_t recycles_free_count;
206 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count,
207 "Number of free vnodes recycled to meet vnode cache targets");
209 static counter_u64_t deferred_inact;
210 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact,
211 "Number of times inactive processing was deferred");
213 /* To keep more than one thread at a time from running vfs_getnewfsid */
214 static struct mtx mntid_mtx;
217 * Lock for any access to the following:
222 static struct mtx __exclusive_cache_line vnode_list_mtx;
224 /* Publicly exported FS */
225 struct nfs_public nfs_pub;
227 static uma_zone_t buf_trie_zone;
228 static smr_t buf_trie_smr;
230 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
231 static uma_zone_t vnode_zone;
232 MALLOC_DEFINE(M_VNODEPOLL, "VN POLL", "vnode poll");
234 __read_frequently smr_t vfs_smr;
237 * The workitem queue.
239 * It is useful to delay writes of file data and filesystem metadata
240 * for tens of seconds so that quickly created and deleted files need
241 * not waste disk bandwidth being created and removed. To realize this,
242 * we append vnodes to a "workitem" queue. When running with a soft
243 * updates implementation, most pending metadata dependencies should
244 * not wait for more than a few seconds. Thus, mounted on block devices
245 * are delayed only about a half the time that file data is delayed.
246 * Similarly, directory updates are more critical, so are only delayed
247 * about a third the time that file data is delayed. Thus, there are
248 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
249 * one each second (driven off the filesystem syncer process). The
250 * syncer_delayno variable indicates the next queue that is to be processed.
251 * Items that need to be processed soon are placed in this queue:
253 * syncer_workitem_pending[syncer_delayno]
255 * A delay of fifteen seconds is done by placing the request fifteen
256 * entries later in the queue:
258 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
261 static int syncer_delayno;
262 static long syncer_mask;
263 LIST_HEAD(synclist, bufobj);
264 static struct synclist *syncer_workitem_pending;
266 * The sync_mtx protects:
271 * syncer_workitem_pending
272 * syncer_worklist_len
275 static struct mtx sync_mtx;
276 static struct cv sync_wakeup;
278 #define SYNCER_MAXDELAY 32
279 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */
280 static int syncdelay = 30; /* max time to delay syncing data */
281 static int filedelay = 30; /* time to delay syncing files */
282 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
283 "Time to delay syncing files (in seconds)");
284 static int dirdelay = 29; /* time to delay syncing directories */
285 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
286 "Time to delay syncing directories (in seconds)");
287 static int metadelay = 28; /* time to delay syncing metadata */
288 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
289 "Time to delay syncing metadata (in seconds)");
290 static int rushjob; /* number of slots to run ASAP */
291 static int stat_rush_requests; /* number of times I/O speeded up */
292 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
293 "Number of times I/O speeded up (rush requests)");
295 #define VDBATCH_SIZE 8
300 struct vnode *tab[VDBATCH_SIZE];
302 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
304 static void vdbatch_dequeue(struct vnode *vp);
307 * When shutting down the syncer, run it at four times normal speed.
309 #define SYNCER_SHUTDOWN_SPEEDUP 4
310 static int sync_vnode_count;
311 static int syncer_worklist_len;
312 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
315 /* Target for maximum number of vnodes. */
316 u_long desiredvnodes;
317 static u_long gapvnodes; /* gap between wanted and desired */
318 static u_long vhiwat; /* enough extras after expansion */
319 static u_long vlowat; /* minimal extras before expansion */
320 static u_long vstir; /* nonzero to stir non-free vnodes */
321 static volatile int vsmalltrigger = 8; /* pref to keep if > this many pages */
323 static u_long vnlru_read_freevnodes(void);
326 * Note that no attempt is made to sanitize these parameters.
329 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
335 error = sysctl_handle_long(oidp, &val, 0, req);
336 if (error != 0 || req->newptr == NULL)
339 if (val == desiredvnodes)
341 mtx_lock(&vnode_list_mtx);
343 wantfreevnodes = desiredvnodes / 4;
345 mtx_unlock(&vnode_list_mtx);
347 * XXX There is no protection against multiple threads changing
348 * desiredvnodes at the same time. Locking above only helps vnlru and
351 vfs_hash_changesize(desiredvnodes);
352 cache_changesize(desiredvnodes);
356 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
357 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
358 "LU", "Target for maximum number of vnodes");
361 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
366 val = wantfreevnodes;
367 error = sysctl_handle_long(oidp, &val, 0, req);
368 if (error != 0 || req->newptr == NULL)
371 if (val == wantfreevnodes)
373 mtx_lock(&vnode_list_mtx);
374 wantfreevnodes = val;
376 mtx_unlock(&vnode_list_mtx);
380 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
381 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
382 "LU", "Target for minimum number of \"free\" vnodes");
384 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
385 &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
386 static int vnlru_nowhere;
387 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
388 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
391 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
396 unsigned long ndflags;
399 if (req->newptr == NULL)
401 if (req->newlen >= PATH_MAX)
404 buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
405 error = SYSCTL_IN(req, buf, req->newlen);
409 buf[req->newlen] = '\0';
411 ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | SAVENAME;
412 NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread);
413 if ((error = namei(&nd)) != 0)
417 if (VN_IS_DOOMED(vp)) {
419 * This vnode is being recycled. Return != 0 to let the caller
420 * know that the sysctl had no effect. Return EAGAIN because a
421 * subsequent call will likely succeed (since namei will create
422 * a new vnode if necessary)
428 counter_u64_add(recycles_count, 1);
438 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
440 struct thread *td = curthread;
446 if (req->newptr == NULL)
449 error = sysctl_handle_int(oidp, &fd, 0, req);
452 error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
457 error = vn_lock(vp, LK_EXCLUSIVE);
461 counter_u64_add(recycles_count, 1);
469 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
470 CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
471 sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
472 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
473 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
474 sysctl_ftry_reclaim_vnode, "I",
475 "Try to reclaim a vnode by its file descriptor");
477 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
481 * Support for the bufobj clean & dirty pctrie.
484 buf_trie_alloc(struct pctrie *ptree)
486 return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT));
490 buf_trie_free(struct pctrie *ptree, void *node)
492 uma_zfree_smr(buf_trie_zone, node);
494 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free,
498 * Initialize the vnode management data structures.
500 * Reevaluate the following cap on the number of vnodes after the physical
501 * memory size exceeds 512GB. In the limit, as the physical memory size
502 * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
504 #ifndef MAXVNODES_MAX
505 #define MAXVNODES_MAX (512UL * 1024 * 1024 / 64) /* 8M */
508 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
510 static struct vnode *
511 vn_alloc_marker(struct mount *mp)
515 vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
516 vp->v_type = VMARKER;
523 vn_free_marker(struct vnode *vp)
526 MPASS(vp->v_type == VMARKER);
527 free(vp, M_VNODE_MARKER);
531 * Initialize a vnode as it first enters the zone.
534 vnode_init(void *mem, int size, int flags)
543 vp->v_vnlock = &vp->v_lock;
544 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
546 * By default, don't allow shared locks unless filesystems opt-in.
548 lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
549 LK_NOSHARE | LK_IS_VNODE);
553 bufobj_init(&vp->v_bufobj, vp);
555 * Initialize namecache.
557 cache_vnode_init(vp);
559 * Initialize rangelocks.
561 rangelock_init(&vp->v_rl);
563 vp->v_dbatchcpu = NOCPU;
566 * Check vhold_recycle_free for an explanation.
568 vp->v_holdcnt = VHOLD_NO_SMR;
570 mtx_lock(&vnode_list_mtx);
571 TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
572 mtx_unlock(&vnode_list_mtx);
577 * Free a vnode when it is cleared from the zone.
580 vnode_fini(void *mem, int size)
587 mtx_lock(&vnode_list_mtx);
588 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
589 mtx_unlock(&vnode_list_mtx);
590 rangelock_destroy(&vp->v_rl);
591 lockdestroy(vp->v_vnlock);
592 mtx_destroy(&vp->v_interlock);
594 rw_destroy(BO_LOCKPTR(bo));
598 * Provide the size of NFS nclnode and NFS fh for calculation of the
599 * vnode memory consumption. The size is specified directly to
600 * eliminate dependency on NFS-private header.
602 * Other filesystems may use bigger or smaller (like UFS and ZFS)
603 * private inode data, but the NFS-based estimation is ample enough.
604 * Still, we care about differences in the size between 64- and 32-bit
607 * Namecache structure size is heuristically
608 * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
611 #define NFS_NCLNODE_SZ (528 + 64)
614 #define NFS_NCLNODE_SZ (360 + 32)
619 vntblinit(void *dummy __unused)
622 int cpu, physvnodes, virtvnodes;
626 * Desiredvnodes is a function of the physical memory size and the
627 * kernel's heap size. Generally speaking, it scales with the
628 * physical memory size. The ratio of desiredvnodes to the physical
629 * memory size is 1:16 until desiredvnodes exceeds 98,304.
631 * marginal ratio of desiredvnodes to the physical memory size is
632 * 1:64. However, desiredvnodes is limited by the kernel's heap
633 * size. The memory required by desiredvnodes vnodes and vm objects
634 * must not exceed 1/10th of the kernel's heap size.
636 physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
637 3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
638 virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
639 sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
640 desiredvnodes = min(physvnodes, virtvnodes);
641 if (desiredvnodes > MAXVNODES_MAX) {
643 printf("Reducing kern.maxvnodes %lu -> %lu\n",
644 desiredvnodes, MAXVNODES_MAX);
645 desiredvnodes = MAXVNODES_MAX;
647 wantfreevnodes = desiredvnodes / 4;
648 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
649 TAILQ_INIT(&vnode_list);
650 mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
652 * The lock is taken to appease WITNESS.
654 mtx_lock(&vnode_list_mtx);
656 mtx_unlock(&vnode_list_mtx);
657 vnode_list_free_marker = vn_alloc_marker(NULL);
658 TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
659 vnode_list_reclaim_marker = vn_alloc_marker(NULL);
660 TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
661 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
662 vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
663 uma_zone_set_smr(vnode_zone, vfs_smr);
665 * Preallocate enough nodes to support one-per buf so that
666 * we can not fail an insert. reassignbuf() callers can not
667 * tolerate the insertion failure.
669 buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
670 NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
671 UMA_ZONE_NOFREE | UMA_ZONE_SMR);
672 buf_trie_smr = uma_zone_get_smr(buf_trie_zone);
673 uma_prealloc(buf_trie_zone, nbuf);
675 vnodes_created = counter_u64_alloc(M_WAITOK);
676 recycles_count = counter_u64_alloc(M_WAITOK);
677 recycles_free_count = counter_u64_alloc(M_WAITOK);
678 deferred_inact = counter_u64_alloc(M_WAITOK);
681 * Initialize the filesystem syncer.
683 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
685 syncer_maxdelay = syncer_mask + 1;
686 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
687 cv_init(&sync_wakeup, "syncer");
688 for (i = 1; i <= sizeof(struct vnode); i <<= 1)
693 vd = DPCPU_ID_PTR((cpu), vd);
694 bzero(vd, sizeof(*vd));
695 mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
698 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
701 * Mark a mount point as busy. Used to synchronize access and to delay
702 * unmounting. Eventually, mountlist_mtx is not released on failure.
704 * vfs_busy() is a custom lock, it can block the caller.
705 * vfs_busy() only sleeps if the unmount is active on the mount point.
706 * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
707 * vnode belonging to mp.
709 * Lookup uses vfs_busy() to traverse mount points.
711 * / vnode lock A / vnode lock (/var) D
712 * /var vnode lock B /log vnode lock(/var/log) E
713 * vfs_busy lock C vfs_busy lock F
715 * Within each file system, the lock order is C->A->B and F->D->E.
717 * When traversing across mounts, the system follows that lock order:
723 * The lookup() process for namei("/var") illustrates the process:
724 * VOP_LOOKUP() obtains B while A is held
725 * vfs_busy() obtains a shared lock on F while A and B are held
726 * vput() releases lock on B
727 * vput() releases lock on A
728 * VFS_ROOT() obtains lock on D while shared lock on F is held
729 * vfs_unbusy() releases shared lock on F
730 * vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
731 * Attempt to lock A (instead of vp_crossmp) while D is held would
732 * violate the global order, causing deadlocks.
734 * dounmount() locks B while F is drained.
737 vfs_busy(struct mount *mp, int flags)
739 struct mount_pcpu *mpcpu;
741 MPASS((flags & ~MBF_MASK) == 0);
742 CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
744 if (vfs_op_thread_enter(mp, mpcpu)) {
745 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
746 MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
747 MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
748 vfs_mp_count_add_pcpu(mpcpu, ref, 1);
749 vfs_mp_count_add_pcpu(mpcpu, lockref, 1);
750 vfs_op_thread_exit(mp, mpcpu);
751 if (flags & MBF_MNTLSTLOCK)
752 mtx_unlock(&mountlist_mtx);
757 vfs_assert_mount_counters(mp);
760 * If mount point is currently being unmounted, sleep until the
761 * mount point fate is decided. If thread doing the unmounting fails,
762 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
763 * that this mount point has survived the unmount attempt and vfs_busy
764 * should retry. Otherwise the unmounter thread will set MNTK_REFEXPIRE
765 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
766 * about to be really destroyed. vfs_busy needs to release its
767 * reference on the mount point in this case and return with ENOENT,
768 * telling the caller that mount mount it tried to busy is no longer
771 while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
772 if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
775 CTR1(KTR_VFS, "%s: failed busying before sleeping",
779 if (flags & MBF_MNTLSTLOCK)
780 mtx_unlock(&mountlist_mtx);
781 mp->mnt_kern_flag |= MNTK_MWAIT;
782 msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
783 if (flags & MBF_MNTLSTLOCK)
784 mtx_lock(&mountlist_mtx);
787 if (flags & MBF_MNTLSTLOCK)
788 mtx_unlock(&mountlist_mtx);
795 * Free a busy filesystem.
798 vfs_unbusy(struct mount *mp)
800 struct mount_pcpu *mpcpu;
803 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
805 if (vfs_op_thread_enter(mp, mpcpu)) {
806 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
807 vfs_mp_count_sub_pcpu(mpcpu, lockref, 1);
808 vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
809 vfs_op_thread_exit(mp, mpcpu);
814 vfs_assert_mount_counters(mp);
816 c = --mp->mnt_lockref;
817 if (mp->mnt_vfs_ops == 0) {
818 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
823 vfs_dump_mount_counters(mp);
824 if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
825 MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
826 CTR1(KTR_VFS, "%s: waking up waiters", __func__);
827 mp->mnt_kern_flag &= ~MNTK_DRAINING;
828 wakeup(&mp->mnt_lockref);
834 * Lookup a mount point by filesystem identifier.
837 vfs_getvfs(fsid_t *fsid)
841 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
842 mtx_lock(&mountlist_mtx);
843 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
844 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
846 mtx_unlock(&mountlist_mtx);
850 mtx_unlock(&mountlist_mtx);
851 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
852 return ((struct mount *) 0);
856 * Lookup a mount point by filesystem identifier, busying it before
859 * To avoid congestion on mountlist_mtx, implement simple direct-mapped
860 * cache for popular filesystem identifiers. The cache is lockess, using
861 * the fact that struct mount's are never freed. In worst case we may
862 * get pointer to unmounted or even different filesystem, so we have to
863 * check what we got, and go slow way if so.
866 vfs_busyfs(fsid_t *fsid)
868 #define FSID_CACHE_SIZE 256
869 typedef struct mount * volatile vmp_t;
870 static vmp_t cache[FSID_CACHE_SIZE];
875 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
876 hash = fsid->val[0] ^ fsid->val[1];
877 hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
879 if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0)
881 if (vfs_busy(mp, 0) != 0) {
885 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0)
891 mtx_lock(&mountlist_mtx);
892 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
893 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
894 error = vfs_busy(mp, MBF_MNTLSTLOCK);
897 mtx_unlock(&mountlist_mtx);
904 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
905 mtx_unlock(&mountlist_mtx);
906 return ((struct mount *) 0);
910 * Check if a user can access privileged mount options.
913 vfs_suser(struct mount *mp, struct thread *td)
917 if (jailed(td->td_ucred)) {
919 * If the jail of the calling thread lacks permission for
920 * this type of file system, deny immediately.
922 if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
926 * If the file system was mounted outside the jail of the
927 * calling thread, deny immediately.
929 if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
934 * If file system supports delegated administration, we don't check
935 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
936 * by the file system itself.
937 * If this is not the user that did original mount, we check for
938 * the PRIV_VFS_MOUNT_OWNER privilege.
940 if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
941 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
942 if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
949 * Get a new unique fsid. Try to make its val[0] unique, since this value
950 * will be used to create fake device numbers for stat(). Also try (but
951 * not so hard) make its val[0] unique mod 2^16, since some emulators only
952 * support 16-bit device numbers. We end up with unique val[0]'s for the
953 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
955 * Keep in mind that several mounts may be running in parallel. Starting
956 * the search one past where the previous search terminated is both a
957 * micro-optimization and a defense against returning the same fsid to
961 vfs_getnewfsid(struct mount *mp)
963 static uint16_t mntid_base;
968 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
969 mtx_lock(&mntid_mtx);
970 mtype = mp->mnt_vfc->vfc_typenum;
971 tfsid.val[1] = mtype;
972 mtype = (mtype & 0xFF) << 24;
974 tfsid.val[0] = makedev(255,
975 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
977 if ((nmp = vfs_getvfs(&tfsid)) == NULL)
981 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
982 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
983 mtx_unlock(&mntid_mtx);
987 * Knob to control the precision of file timestamps:
989 * 0 = seconds only; nanoseconds zeroed.
990 * 1 = seconds and nanoseconds, accurate within 1/HZ.
991 * 2 = seconds and nanoseconds, truncated to microseconds.
992 * >=3 = seconds and nanoseconds, maximum precision.
994 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
996 static int timestamp_precision = TSP_USEC;
997 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
998 ×tamp_precision, 0, "File timestamp precision (0: seconds, "
999 "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1000 "3+: sec + ns (max. precision))");
1003 * Get a current timestamp.
1006 vfs_timestamp(struct timespec *tsp)
1010 switch (timestamp_precision) {
1012 tsp->tv_sec = time_second;
1020 TIMEVAL_TO_TIMESPEC(&tv, tsp);
1030 * Set vnode attributes to VNOVAL
1033 vattr_null(struct vattr *vap)
1036 vap->va_type = VNON;
1037 vap->va_size = VNOVAL;
1038 vap->va_bytes = VNOVAL;
1039 vap->va_mode = VNOVAL;
1040 vap->va_nlink = VNOVAL;
1041 vap->va_uid = VNOVAL;
1042 vap->va_gid = VNOVAL;
1043 vap->va_fsid = VNOVAL;
1044 vap->va_fileid = VNOVAL;
1045 vap->va_blocksize = VNOVAL;
1046 vap->va_rdev = VNOVAL;
1047 vap->va_atime.tv_sec = VNOVAL;
1048 vap->va_atime.tv_nsec = VNOVAL;
1049 vap->va_mtime.tv_sec = VNOVAL;
1050 vap->va_mtime.tv_nsec = VNOVAL;
1051 vap->va_ctime.tv_sec = VNOVAL;
1052 vap->va_ctime.tv_nsec = VNOVAL;
1053 vap->va_birthtime.tv_sec = VNOVAL;
1054 vap->va_birthtime.tv_nsec = VNOVAL;
1055 vap->va_flags = VNOVAL;
1056 vap->va_gen = VNOVAL;
1057 vap->va_vaflags = 0;
1061 * Try to reduce the total number of vnodes.
1063 * This routine (and its user) are buggy in at least the following ways:
1064 * - all parameters were picked years ago when RAM sizes were significantly
1066 * - it can pick vnodes based on pages used by the vm object, but filesystems
1067 * like ZFS don't use it making the pick broken
1068 * - since ZFS has its own aging policy it gets partially combated by this one
1069 * - a dedicated method should be provided for filesystems to let them decide
1070 * whether the vnode should be recycled
1072 * This routine is called when we have too many vnodes. It attempts
1073 * to free <count> vnodes and will potentially free vnodes that still
1074 * have VM backing store (VM backing store is typically the cause
1075 * of a vnode blowout so we want to do this). Therefore, this operation
1076 * is not considered cheap.
1078 * A number of conditions may prevent a vnode from being reclaimed.
1079 * the buffer cache may have references on the vnode, a directory
1080 * vnode may still have references due to the namei cache representing
1081 * underlying files, or the vnode may be in active use. It is not
1082 * desirable to reuse such vnodes. These conditions may cause the
1083 * number of vnodes to reach some minimum value regardless of what
1084 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low.
1086 * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1087 * entries if this argument is strue
1088 * @param trigger Only reclaim vnodes with fewer than this many resident
1090 * @param target How many vnodes to reclaim.
1091 * @return The number of vnodes that were reclaimed.
1094 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1096 struct vnode *vp, *mvp;
1098 struct vm_object *object;
1102 mtx_assert(&vnode_list_mtx, MA_OWNED);
1107 mvp = vnode_list_reclaim_marker;
1110 while (done < target) {
1111 vp = TAILQ_NEXT(vp, v_vnodelist);
1112 if (__predict_false(vp == NULL))
1115 if (__predict_false(vp->v_type == VMARKER))
1119 * If it's been deconstructed already, it's still
1120 * referenced, or it exceeds the trigger, skip it.
1121 * Also skip free vnodes. We are trying to make space
1122 * to expand the free list, not reduce it.
1124 if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1125 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1128 if (vp->v_type == VBAD || vp->v_type == VNON)
1131 object = atomic_load_ptr(&vp->v_object);
1132 if (object == NULL || object->resident_page_count > trigger) {
1137 * Handle races against vnode allocation. Filesystems lock the
1138 * vnode some time after it gets returned from getnewvnode,
1139 * despite type and hold count being manipulated earlier.
1140 * Resorting to checking v_mount restores guarantees present
1141 * before the global list was reworked to contain all vnodes.
1143 if (!VI_TRYLOCK(vp))
1145 if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1149 if (vp->v_mount == NULL) {
1155 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1156 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1157 mtx_unlock(&vnode_list_mtx);
1159 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1161 goto next_iter_unlocked;
1163 if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1165 vn_finished_write(mp);
1166 goto next_iter_unlocked;
1170 if (vp->v_usecount > 0 ||
1171 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1172 (vp->v_object != NULL && vp->v_object->handle == vp &&
1173 vp->v_object->resident_page_count > trigger)) {
1176 vn_finished_write(mp);
1177 goto next_iter_unlocked;
1179 counter_u64_add(recycles_count, 1);
1183 vn_finished_write(mp);
1187 kern_yield(PRI_USER);
1188 mtx_lock(&vnode_list_mtx);
1191 MPASS(vp->v_type != VMARKER);
1192 if (!should_yield())
1194 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1195 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1196 mtx_unlock(&vnode_list_mtx);
1197 kern_yield(PRI_USER);
1198 mtx_lock(&vnode_list_mtx);
1201 if (done == 0 && !retried) {
1202 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1203 TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1210 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1211 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1213 "limit on vnode free requests per call to the vnlru_free routine");
1216 * Attempt to reduce the free list by the requested amount.
1219 vnlru_free_impl(int count, struct vfsops *mnt_op, struct vnode *mvp)
1225 mtx_assert(&vnode_list_mtx, MA_OWNED);
1226 if (count > max_vnlru_free)
1227 count = max_vnlru_free;
1234 vp = TAILQ_NEXT(vp, v_vnodelist);
1235 if (__predict_false(vp == NULL)) {
1236 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1237 TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1240 if (__predict_false(vp->v_type == VMARKER))
1242 if (vp->v_holdcnt > 0)
1245 * Don't recycle if our vnode is from different type
1246 * of mount point. Note that mp is type-safe, the
1247 * check does not reach unmapped address even if
1248 * vnode is reclaimed.
1250 if (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1251 mp->mnt_op != mnt_op) {
1254 if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1257 if (!vhold_recycle_free(vp))
1259 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1260 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1261 mtx_unlock(&vnode_list_mtx);
1262 if (vtryrecycle(vp) == 0)
1264 mtx_lock(&vnode_list_mtx);
1267 return (ocount - count);
1271 vnlru_free_locked(int count)
1274 mtx_assert(&vnode_list_mtx, MA_OWNED);
1275 return (vnlru_free_impl(count, NULL, vnode_list_free_marker));
1279 vnlru_free_vfsops(int count, struct vfsops *mnt_op, struct vnode *mvp)
1282 MPASS(mnt_op != NULL);
1284 VNPASS(mvp->v_type == VMARKER, mvp);
1285 mtx_lock(&vnode_list_mtx);
1286 vnlru_free_impl(count, mnt_op, mvp);
1287 mtx_unlock(&vnode_list_mtx);
1291 * Temporary binary compat, don't use. Call vnlru_free_vfsops instead.
1294 vnlru_free(int count, struct vfsops *mnt_op)
1300 mtx_lock(&vnode_list_mtx);
1301 mvp = vnode_list_free_marker;
1302 if (vnlru_free_impl(count, mnt_op, mvp) == 0) {
1304 * It is possible the marker was moved over eligible vnodes by
1305 * callers which filtered by different ops. If so, start from
1308 if (vnlru_read_freevnodes() > 0) {
1309 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1310 TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1312 vnlru_free_impl(count, mnt_op, mvp);
1314 mtx_unlock(&vnode_list_mtx);
1318 vnlru_alloc_marker(void)
1322 mvp = vn_alloc_marker(NULL);
1323 mtx_lock(&vnode_list_mtx);
1324 TAILQ_INSERT_BEFORE(vnode_list_free_marker, mvp, v_vnodelist);
1325 mtx_unlock(&vnode_list_mtx);
1330 vnlru_free_marker(struct vnode *mvp)
1332 mtx_lock(&vnode_list_mtx);
1333 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1334 mtx_unlock(&vnode_list_mtx);
1335 vn_free_marker(mvp);
1342 mtx_assert(&vnode_list_mtx, MA_OWNED);
1343 gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1344 vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1345 vlowat = vhiwat / 2;
1349 * Attempt to recycle vnodes in a context that is always safe to block.
1350 * Calling vlrurecycle() from the bowels of filesystem code has some
1351 * interesting deadlock problems.
1353 static struct proc *vnlruproc;
1354 static int vnlruproc_sig;
1357 * The main freevnodes counter is only updated when threads requeue their vnode
1358 * batches. CPUs are conditionally walked to compute a more accurate total.
1360 * Limit how much of a slop are we willing to tolerate. Note: the actual value
1361 * at any given moment can still exceed slop, but it should not be by significant
1362 * margin in practice.
1364 #define VNLRU_FREEVNODES_SLOP 128
1366 static __inline void
1367 vn_freevnodes_inc(void)
1377 static __inline void
1378 vn_freevnodes_dec(void)
1389 vnlru_read_freevnodes(void)
1395 mtx_assert(&vnode_list_mtx, MA_OWNED);
1396 if (freevnodes > freevnodes_old)
1397 slop = freevnodes - freevnodes_old;
1399 slop = freevnodes_old - freevnodes;
1400 if (slop < VNLRU_FREEVNODES_SLOP)
1401 return (freevnodes >= 0 ? freevnodes : 0);
1402 freevnodes_old = freevnodes;
1404 vd = DPCPU_ID_PTR((cpu), vd);
1405 freevnodes_old += vd->freevnodes;
1407 return (freevnodes_old >= 0 ? freevnodes_old : 0);
1411 vnlru_under(u_long rnumvnodes, u_long limit)
1413 u_long rfreevnodes, space;
1415 if (__predict_false(rnumvnodes > desiredvnodes))
1418 space = desiredvnodes - rnumvnodes;
1419 if (space < limit) {
1420 rfreevnodes = vnlru_read_freevnodes();
1421 if (rfreevnodes > wantfreevnodes)
1422 space += rfreevnodes - wantfreevnodes;
1424 return (space < limit);
1428 vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1430 long rfreevnodes, space;
1432 if (__predict_false(rnumvnodes > desiredvnodes))
1435 space = desiredvnodes - rnumvnodes;
1436 if (space < limit) {
1437 rfreevnodes = atomic_load_long(&freevnodes);
1438 if (rfreevnodes > wantfreevnodes)
1439 space += rfreevnodes - wantfreevnodes;
1441 return (space < limit);
1448 mtx_assert(&vnode_list_mtx, MA_OWNED);
1449 if (vnlruproc_sig == 0) {
1458 u_long rnumvnodes, rfreevnodes, target;
1459 unsigned long onumvnodes;
1460 int done, force, trigger, usevnodes;
1461 bool reclaim_nc_src, want_reread;
1463 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1464 SHUTDOWN_PRI_FIRST);
1467 want_reread = false;
1469 kproc_suspend_check(vnlruproc);
1470 mtx_lock(&vnode_list_mtx);
1471 rnumvnodes = atomic_load_long(&numvnodes);
1474 force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1475 want_reread = false;
1479 * If numvnodes is too large (due to desiredvnodes being
1480 * adjusted using its sysctl, or emergency growth), first
1481 * try to reduce it by discarding from the free list.
1483 if (rnumvnodes > desiredvnodes) {
1484 vnlru_free_locked(rnumvnodes - desiredvnodes);
1485 rnumvnodes = atomic_load_long(&numvnodes);
1488 * Sleep if the vnode cache is in a good state. This is
1489 * when it is not over-full and has space for about a 4%
1490 * or 9% expansion (by growing its size or inexcessively
1491 * reducing its free list). Otherwise, try to reclaim
1492 * space for a 10% expansion.
1494 if (vstir && force == 0) {
1498 if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1500 wakeup(&vnlruproc_sig);
1501 msleep(vnlruproc, &vnode_list_mtx,
1502 PVFS|PDROP, "vlruwt", hz);
1505 rfreevnodes = vnlru_read_freevnodes();
1507 onumvnodes = rnumvnodes;
1509 * Calculate parameters for recycling. These are the same
1510 * throughout the loop to give some semblance of fairness.
1511 * The trigger point is to avoid recycling vnodes with lots
1512 * of resident pages. We aren't trying to free memory; we
1513 * are trying to recycle or at least free vnodes.
1515 if (rnumvnodes <= desiredvnodes)
1516 usevnodes = rnumvnodes - rfreevnodes;
1518 usevnodes = rnumvnodes;
1522 * The trigger value is is chosen to give a conservatively
1523 * large value to ensure that it alone doesn't prevent
1524 * making progress. The value can easily be so large that
1525 * it is effectively infinite in some congested and
1526 * misconfigured cases, and this is necessary. Normally
1527 * it is about 8 to 100 (pages), which is quite large.
1529 trigger = vm_cnt.v_page_count * 2 / usevnodes;
1531 trigger = vsmalltrigger;
1532 reclaim_nc_src = force >= 3;
1533 target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1534 target = target / 10 + 1;
1535 done = vlrureclaim(reclaim_nc_src, trigger, target);
1536 mtx_unlock(&vnode_list_mtx);
1537 if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1538 uma_reclaim(UMA_RECLAIM_DRAIN);
1540 if (force == 0 || force == 1) {
1551 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1554 kern_yield(PRI_USER);
1559 static struct kproc_desc vnlru_kp = {
1564 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1568 * Routines having to do with the management of the vnode table.
1572 * Try to recycle a freed vnode. We abort if anyone picks up a reference
1573 * before we actually vgone(). This function must be called with the vnode
1574 * held to prevent the vnode from being returned to the free list midway
1578 vtryrecycle(struct vnode *vp)
1582 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1583 VNASSERT(vp->v_holdcnt, vp,
1584 ("vtryrecycle: Recycling vp %p without a reference.", vp));
1586 * This vnode may found and locked via some other list, if so we
1587 * can't recycle it yet.
1589 if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1591 "%s: impossible to recycle, vp %p lock is already held",
1594 return (EWOULDBLOCK);
1597 * Don't recycle if its filesystem is being suspended.
1599 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1602 "%s: impossible to recycle, cannot start the write for %p",
1608 * If we got this far, we need to acquire the interlock and see if
1609 * anyone picked up this vnode from another list. If not, we will
1610 * mark it with DOOMED via vgonel() so that anyone who does find it
1611 * will skip over it.
1614 if (vp->v_usecount) {
1617 vn_finished_write(vnmp);
1619 "%s: impossible to recycle, %p is already referenced",
1623 if (!VN_IS_DOOMED(vp)) {
1624 counter_u64_add(recycles_free_count, 1);
1629 vn_finished_write(vnmp);
1634 * Allocate a new vnode.
1636 * The operation never returns an error. Returning an error was disabled
1637 * in r145385 (dated 2005) with the following comment:
1639 * XXX Not all VFS_VGET/ffs_vget callers check returns.
1641 * Given the age of this commit (almost 15 years at the time of writing this
1642 * comment) restoring the ability to fail requires a significant audit of
1645 * The routine can try to free a vnode or stall for up to 1 second waiting for
1646 * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1648 static u_long vn_alloc_cyclecount;
1650 static struct vnode * __noinline
1651 vn_alloc_hard(struct mount *mp)
1653 u_long rnumvnodes, rfreevnodes;
1655 mtx_lock(&vnode_list_mtx);
1656 rnumvnodes = atomic_load_long(&numvnodes);
1657 if (rnumvnodes + 1 < desiredvnodes) {
1658 vn_alloc_cyclecount = 0;
1661 rfreevnodes = vnlru_read_freevnodes();
1662 if (vn_alloc_cyclecount++ >= rfreevnodes) {
1663 vn_alloc_cyclecount = 0;
1667 * Grow the vnode cache if it will not be above its target max
1668 * after growing. Otherwise, if the free list is nonempty, try
1669 * to reclaim 1 item from it before growing the cache (possibly
1670 * above its target max if the reclamation failed or is delayed).
1671 * Otherwise, wait for some space. In all cases, schedule
1672 * vnlru_proc() if we are getting short of space. The watermarks
1673 * should be chosen so that we never wait or even reclaim from
1674 * the free list to below its target minimum.
1676 if (vnlru_free_locked(1) > 0)
1678 if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1680 * Wait for space for a new vnode.
1683 msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1684 if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1685 vnlru_read_freevnodes() > 1)
1686 vnlru_free_locked(1);
1689 rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1690 if (vnlru_under(rnumvnodes, vlowat))
1692 mtx_unlock(&vnode_list_mtx);
1693 return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1696 static struct vnode *
1697 vn_alloc(struct mount *mp)
1701 if (__predict_false(vn_alloc_cyclecount != 0))
1702 return (vn_alloc_hard(mp));
1703 rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1704 if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1705 atomic_subtract_long(&numvnodes, 1);
1706 return (vn_alloc_hard(mp));
1709 return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1713 vn_free(struct vnode *vp)
1716 atomic_subtract_long(&numvnodes, 1);
1717 uma_zfree_smr(vnode_zone, vp);
1721 * Return the next vnode from the free list.
1724 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1729 struct lock_object *lo;
1731 CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1733 KASSERT(vops->registered,
1734 ("%s: not registered vector op %p\n", __func__, vops));
1737 if (td->td_vp_reserved != NULL) {
1738 vp = td->td_vp_reserved;
1739 td->td_vp_reserved = NULL;
1743 counter_u64_add(vnodes_created, 1);
1745 * Locks are given the generic name "vnode" when created.
1746 * Follow the historic practice of using the filesystem
1747 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1749 * Locks live in a witness group keyed on their name. Thus,
1750 * when a lock is renamed, it must also move from the witness
1751 * group of its old name to the witness group of its new name.
1753 * The change only needs to be made when the vnode moves
1754 * from one filesystem type to another. We ensure that each
1755 * filesystem use a single static name pointer for its tag so
1756 * that we can compare pointers rather than doing a strcmp().
1758 lo = &vp->v_vnlock->lock_object;
1760 if (lo->lo_name != tag) {
1764 WITNESS_DESTROY(lo);
1765 WITNESS_INIT(lo, tag);
1769 * By default, don't allow shared locks unless filesystems opt-in.
1771 vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1773 * Finalize various vnode identity bits.
1775 KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1776 KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1777 KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1781 v_init_counters(vp);
1783 vp->v_bufobj.bo_ops = &buf_ops_bio;
1785 if (mp == NULL && vops != &dead_vnodeops)
1786 printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1790 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1791 mac_vnode_associate_singlelabel(mp, vp);
1794 vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1795 if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1796 vp->v_vflag |= VV_NOKNOTE;
1800 * For the filesystems which do not use vfs_hash_insert(),
1801 * still initialize v_hash to have vfs_hash_index() useful.
1802 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1805 vp->v_hash = (uintptr_t)vp >> vnsz2log;
1812 getnewvnode_reserve(void)
1817 MPASS(td->td_vp_reserved == NULL);
1818 td->td_vp_reserved = vn_alloc(NULL);
1822 getnewvnode_drop_reserve(void)
1827 if (td->td_vp_reserved != NULL) {
1828 vn_free(td->td_vp_reserved);
1829 td->td_vp_reserved = NULL;
1833 static void __noinline
1834 freevnode(struct vnode *vp)
1839 * The vnode has been marked for destruction, so free it.
1841 * The vnode will be returned to the zone where it will
1842 * normally remain until it is needed for another vnode. We
1843 * need to cleanup (or verify that the cleanup has already
1844 * been done) any residual data left from its current use
1845 * so as not to contaminate the freshly allocated vnode.
1847 CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1849 * Paired with vgone.
1851 vn_seqc_write_end_free(vp);
1854 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1855 VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp);
1856 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1857 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1858 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1859 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1860 VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1861 ("clean blk trie not empty"));
1862 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1863 VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1864 ("dirty blk trie not empty"));
1865 VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1866 VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1867 VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1868 VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1869 ("Dangling rangelock waiters"));
1870 VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp,
1871 ("Leaked inactivation"));
1874 mac_vnode_destroy(vp);
1876 if (vp->v_pollinfo != NULL) {
1877 destroy_vpollinfo(vp->v_pollinfo);
1878 vp->v_pollinfo = NULL;
1880 vp->v_mountedhere = NULL;
1883 vp->v_fifoinfo = NULL;
1891 * Delete from old mount point vnode list, if on one.
1894 delmntque(struct vnode *vp)
1898 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
1907 VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1908 ("bad mount point vnode list size"));
1909 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1910 mp->mnt_nvnodelistsize--;
1916 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1920 vp->v_op = &dead_vnodeops;
1926 * Insert into list of vnodes for the new mount point, if available.
1929 insmntque1(struct vnode *vp, struct mount *mp,
1930 void (*dtr)(struct vnode *, void *), void *dtr_arg)
1933 KASSERT(vp->v_mount == NULL,
1934 ("insmntque: vnode already on per mount vnode list"));
1935 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1936 ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1939 * We acquire the vnode interlock early to ensure that the
1940 * vnode cannot be recycled by another process releasing a
1941 * holdcnt on it before we get it on both the vnode list
1942 * and the active vnode list. The mount mutex protects only
1943 * manipulation of the vnode list and the vnode freelist
1944 * mutex protects only manipulation of the active vnode list.
1945 * Hence the need to hold the vnode interlock throughout.
1949 if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1950 ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1951 mp->mnt_nvnodelistsize == 0)) &&
1952 (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1961 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1962 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1963 ("neg mount point vnode list size"));
1964 mp->mnt_nvnodelistsize++;
1971 insmntque(struct vnode *vp, struct mount *mp)
1974 return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1978 * Flush out and invalidate all buffers associated with a bufobj
1979 * Called with the underlying object locked.
1982 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1987 if (flags & V_SAVE) {
1988 error = bufobj_wwait(bo, slpflag, slptimeo);
1993 if (bo->bo_dirty.bv_cnt > 0) {
1996 error = BO_SYNC(bo, MNT_WAIT);
1997 } while (error == ERELOOKUP);
2001 * XXX We could save a lock/unlock if this was only
2002 * enabled under INVARIANTS
2005 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
2006 panic("vinvalbuf: dirty bufs");
2010 * If you alter this loop please notice that interlock is dropped and
2011 * reacquired in flushbuflist. Special care is needed to ensure that
2012 * no race conditions occur from this.
2015 error = flushbuflist(&bo->bo_clean,
2016 flags, bo, slpflag, slptimeo);
2017 if (error == 0 && !(flags & V_CLEANONLY))
2018 error = flushbuflist(&bo->bo_dirty,
2019 flags, bo, slpflag, slptimeo);
2020 if (error != 0 && error != EAGAIN) {
2024 } while (error != 0);
2027 * Wait for I/O to complete. XXX needs cleaning up. The vnode can
2028 * have write I/O in-progress but if there is a VM object then the
2029 * VM object can also have read-I/O in-progress.
2032 bufobj_wwait(bo, 0, 0);
2033 if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
2035 vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
2038 } while (bo->bo_numoutput > 0);
2042 * Destroy the copy in the VM cache, too.
2044 if (bo->bo_object != NULL &&
2045 (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
2046 VM_OBJECT_WLOCK(bo->bo_object);
2047 vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
2048 OBJPR_CLEANONLY : 0);
2049 VM_OBJECT_WUNLOCK(bo->bo_object);
2054 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
2055 V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
2056 bo->bo_clean.bv_cnt > 0))
2057 panic("vinvalbuf: flush failed");
2058 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
2059 bo->bo_dirty.bv_cnt > 0)
2060 panic("vinvalbuf: flush dirty failed");
2067 * Flush out and invalidate all buffers associated with a vnode.
2068 * Called with the underlying object locked.
2071 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
2074 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2075 ASSERT_VOP_LOCKED(vp, "vinvalbuf");
2076 if (vp->v_object != NULL && vp->v_object->handle != vp)
2078 return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
2082 * Flush out buffers on the specified list.
2086 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2089 struct buf *bp, *nbp;
2094 ASSERT_BO_WLOCKED(bo);
2097 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2099 * If we are flushing both V_NORMAL and V_ALT buffers then
2100 * do not skip any buffers. If we are flushing only V_NORMAL
2101 * buffers then skip buffers marked as BX_ALTDATA. If we are
2102 * flushing only V_ALT buffers then skip buffers not marked
2105 if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2106 (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2107 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2111 lblkno = nbp->b_lblkno;
2112 xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2115 error = BUF_TIMELOCK(bp,
2116 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2117 "flushbuf", slpflag, slptimeo);
2120 return (error != ENOLCK ? error : EAGAIN);
2122 KASSERT(bp->b_bufobj == bo,
2123 ("bp %p wrong b_bufobj %p should be %p",
2124 bp, bp->b_bufobj, bo));
2126 * XXX Since there are no node locks for NFS, I
2127 * believe there is a slight chance that a delayed
2128 * write will occur while sleeping just above, so
2131 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2134 bp->b_flags |= B_ASYNC;
2137 return (EAGAIN); /* XXX: why not loop ? */
2140 bp->b_flags |= (B_INVAL | B_RELBUF);
2141 bp->b_flags &= ~B_ASYNC;
2146 nbp = gbincore(bo, lblkno);
2147 if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2149 break; /* nbp invalid */
2155 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2161 ASSERT_BO_LOCKED(bo);
2163 for (lblkno = startn;;) {
2165 bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2166 if (bp == NULL || bp->b_lblkno >= endn ||
2167 bp->b_lblkno < startn)
2169 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2170 LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2173 if (error == ENOLCK)
2177 KASSERT(bp->b_bufobj == bo,
2178 ("bp %p wrong b_bufobj %p should be %p",
2179 bp, bp->b_bufobj, bo));
2180 lblkno = bp->b_lblkno + 1;
2181 if ((bp->b_flags & B_MANAGED) == 0)
2183 bp->b_flags |= B_RELBUF;
2185 * In the VMIO case, use the B_NOREUSE flag to hint that the
2186 * pages backing each buffer in the range are unlikely to be
2187 * reused. Dirty buffers will have the hint applied once
2188 * they've been written.
2190 if ((bp->b_flags & B_VMIO) != 0)
2191 bp->b_flags |= B_NOREUSE;
2199 * Truncate a file's buffer and pages to a specified length. This
2200 * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2204 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2206 struct buf *bp, *nbp;
2210 CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2211 vp, blksize, (uintmax_t)length);
2214 * Round up to the *next* lbn.
2216 startlbn = howmany(length, blksize);
2218 ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2224 while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2229 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2230 if (bp->b_lblkno > 0)
2233 * Since we hold the vnode lock this should only
2234 * fail if we're racing with the buf daemon.
2237 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2238 BO_LOCKPTR(bo)) == ENOLCK)
2239 goto restart_unlocked;
2241 VNASSERT((bp->b_flags & B_DELWRI), vp,
2242 ("buf(%p) on dirty queue without DELWRI", bp));
2251 bufobj_wwait(bo, 0, 0);
2253 vnode_pager_setsize(vp, length);
2259 * Invalidate the cached pages of a file's buffer within the range of block
2260 * numbers [startlbn, endlbn).
2263 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2269 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2271 start = blksize * startlbn;
2272 end = blksize * endlbn;
2276 MPASS(blksize == bo->bo_bsize);
2278 while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2282 vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2286 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2287 daddr_t startlbn, daddr_t endlbn)
2289 struct buf *bp, *nbp;
2292 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2293 ASSERT_BO_LOCKED(bo);
2297 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2298 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2301 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2302 BO_LOCKPTR(bo)) == ENOLCK) {
2308 bp->b_flags |= B_INVAL | B_RELBUF;
2309 bp->b_flags &= ~B_ASYNC;
2315 (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2317 (nbp->b_flags & B_DELWRI) != 0))
2321 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2322 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2325 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2326 BO_LOCKPTR(bo)) == ENOLCK) {
2331 bp->b_flags |= B_INVAL | B_RELBUF;
2332 bp->b_flags &= ~B_ASYNC;
2338 (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2339 (nbp->b_vp != vp) ||
2340 (nbp->b_flags & B_DELWRI) == 0))
2348 buf_vlist_remove(struct buf *bp)
2353 flags = bp->b_xflags;
2355 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2356 ASSERT_BO_WLOCKED(bp->b_bufobj);
2357 KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 &&
2358 (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN),
2359 ("%s: buffer %p has invalid queue state", __func__, bp));
2361 if ((flags & BX_VNDIRTY) != 0)
2362 bv = &bp->b_bufobj->bo_dirty;
2364 bv = &bp->b_bufobj->bo_clean;
2365 BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2366 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2368 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2372 * Add the buffer to the sorted clean or dirty block list.
2374 * NOTE: xflags is passed as a constant, optimizing this inline function!
2377 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2383 ASSERT_BO_WLOCKED(bo);
2384 KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2385 ("buf_vlist_add: bo %p does not allow bufs", bo));
2386 KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2387 ("dead bo %p", bo));
2388 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2389 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2390 bp->b_xflags |= xflags;
2391 if (xflags & BX_VNDIRTY)
2397 * Keep the list ordered. Optimize empty list insertion. Assume
2398 * we tend to grow at the tail so lookup_le should usually be cheaper
2401 if (bv->bv_cnt == 0 ||
2402 bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2403 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2404 else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2405 TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2407 TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2408 error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2410 panic("buf_vlist_add: Preallocated nodes insufficient.");
2415 * Look up a buffer using the buffer tries.
2418 gbincore(struct bufobj *bo, daddr_t lblkno)
2422 ASSERT_BO_LOCKED(bo);
2423 bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2426 return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno));
2430 * Look up a buf using the buffer tries, without the bufobj lock. This relies
2431 * on SMR for safe lookup, and bufs being in a no-free zone to provide type
2432 * stability of the result. Like other lockless lookups, the found buf may
2433 * already be invalid by the time this function returns.
2436 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno)
2440 ASSERT_BO_UNLOCKED(bo);
2441 bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno);
2444 return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno));
2448 * Associate a buffer with a vnode.
2451 bgetvp(struct vnode *vp, struct buf *bp)
2456 ASSERT_BO_WLOCKED(bo);
2457 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2459 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2460 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2461 ("bgetvp: bp already attached! %p", bp));
2467 * Insert onto list for new vnode.
2469 buf_vlist_add(bp, bo, BX_VNCLEAN);
2473 * Disassociate a buffer from a vnode.
2476 brelvp(struct buf *bp)
2481 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2482 KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2485 * Delete from old vnode list, if on one.
2487 vp = bp->b_vp; /* XXX */
2490 buf_vlist_remove(bp);
2491 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2492 bo->bo_flag &= ~BO_ONWORKLST;
2493 mtx_lock(&sync_mtx);
2494 LIST_REMOVE(bo, bo_synclist);
2495 syncer_worklist_len--;
2496 mtx_unlock(&sync_mtx);
2499 bp->b_bufobj = NULL;
2505 * Add an item to the syncer work queue.
2508 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2512 ASSERT_BO_WLOCKED(bo);
2514 mtx_lock(&sync_mtx);
2515 if (bo->bo_flag & BO_ONWORKLST)
2516 LIST_REMOVE(bo, bo_synclist);
2518 bo->bo_flag |= BO_ONWORKLST;
2519 syncer_worklist_len++;
2522 if (delay > syncer_maxdelay - 2)
2523 delay = syncer_maxdelay - 2;
2524 slot = (syncer_delayno + delay) & syncer_mask;
2526 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2527 mtx_unlock(&sync_mtx);
2531 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2535 mtx_lock(&sync_mtx);
2536 len = syncer_worklist_len - sync_vnode_count;
2537 mtx_unlock(&sync_mtx);
2538 error = SYSCTL_OUT(req, &len, sizeof(len));
2542 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2543 CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2544 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2546 static struct proc *updateproc;
2547 static void sched_sync(void);
2548 static struct kproc_desc up_kp = {
2553 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2556 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2561 *bo = LIST_FIRST(slp);
2565 if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2568 * We use vhold in case the vnode does not
2569 * successfully sync. vhold prevents the vnode from
2570 * going away when we unlock the sync_mtx so that
2571 * we can acquire the vnode interlock.
2574 mtx_unlock(&sync_mtx);
2576 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2578 mtx_lock(&sync_mtx);
2579 return (*bo == LIST_FIRST(slp));
2581 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2582 (void) VOP_FSYNC(vp, MNT_LAZY, td);
2584 vn_finished_write(mp);
2586 if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2588 * Put us back on the worklist. The worklist
2589 * routine will remove us from our current
2590 * position and then add us back in at a later
2593 vn_syncer_add_to_worklist(*bo, syncdelay);
2597 mtx_lock(&sync_mtx);
2601 static int first_printf = 1;
2604 * System filesystem synchronizer daemon.
2609 struct synclist *next, *slp;
2612 struct thread *td = curthread;
2614 int net_worklist_len;
2615 int syncer_final_iter;
2619 syncer_final_iter = 0;
2620 syncer_state = SYNCER_RUNNING;
2621 starttime = time_uptime;
2622 td->td_pflags |= TDP_NORUNNINGBUF;
2624 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2627 mtx_lock(&sync_mtx);
2629 if (syncer_state == SYNCER_FINAL_DELAY &&
2630 syncer_final_iter == 0) {
2631 mtx_unlock(&sync_mtx);
2632 kproc_suspend_check(td->td_proc);
2633 mtx_lock(&sync_mtx);
2635 net_worklist_len = syncer_worklist_len - sync_vnode_count;
2636 if (syncer_state != SYNCER_RUNNING &&
2637 starttime != time_uptime) {
2639 printf("\nSyncing disks, vnodes remaining... ");
2642 printf("%d ", net_worklist_len);
2644 starttime = time_uptime;
2647 * Push files whose dirty time has expired. Be careful
2648 * of interrupt race on slp queue.
2650 * Skip over empty worklist slots when shutting down.
2653 slp = &syncer_workitem_pending[syncer_delayno];
2654 syncer_delayno += 1;
2655 if (syncer_delayno == syncer_maxdelay)
2657 next = &syncer_workitem_pending[syncer_delayno];
2659 * If the worklist has wrapped since the
2660 * it was emptied of all but syncer vnodes,
2661 * switch to the FINAL_DELAY state and run
2662 * for one more second.
2664 if (syncer_state == SYNCER_SHUTTING_DOWN &&
2665 net_worklist_len == 0 &&
2666 last_work_seen == syncer_delayno) {
2667 syncer_state = SYNCER_FINAL_DELAY;
2668 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2670 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2671 syncer_worklist_len > 0);
2674 * Keep track of the last time there was anything
2675 * on the worklist other than syncer vnodes.
2676 * Return to the SHUTTING_DOWN state if any
2679 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2680 last_work_seen = syncer_delayno;
2681 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2682 syncer_state = SYNCER_SHUTTING_DOWN;
2683 while (!LIST_EMPTY(slp)) {
2684 error = sync_vnode(slp, &bo, td);
2686 LIST_REMOVE(bo, bo_synclist);
2687 LIST_INSERT_HEAD(next, bo, bo_synclist);
2691 if (first_printf == 0) {
2693 * Drop the sync mutex, because some watchdog
2694 * drivers need to sleep while patting
2696 mtx_unlock(&sync_mtx);
2697 wdog_kern_pat(WD_LASTVAL);
2698 mtx_lock(&sync_mtx);
2701 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2702 syncer_final_iter--;
2704 * The variable rushjob allows the kernel to speed up the
2705 * processing of the filesystem syncer process. A rushjob
2706 * value of N tells the filesystem syncer to process the next
2707 * N seconds worth of work on its queue ASAP. Currently rushjob
2708 * is used by the soft update code to speed up the filesystem
2709 * syncer process when the incore state is getting so far
2710 * ahead of the disk that the kernel memory pool is being
2711 * threatened with exhaustion.
2718 * Just sleep for a short period of time between
2719 * iterations when shutting down to allow some I/O
2722 * If it has taken us less than a second to process the
2723 * current work, then wait. Otherwise start right over
2724 * again. We can still lose time if any single round
2725 * takes more than two seconds, but it does not really
2726 * matter as we are just trying to generally pace the
2727 * filesystem activity.
2729 if (syncer_state != SYNCER_RUNNING ||
2730 time_uptime == starttime) {
2732 sched_prio(td, PPAUSE);
2735 if (syncer_state != SYNCER_RUNNING)
2736 cv_timedwait(&sync_wakeup, &sync_mtx,
2737 hz / SYNCER_SHUTDOWN_SPEEDUP);
2738 else if (time_uptime == starttime)
2739 cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2744 * Request the syncer daemon to speed up its work.
2745 * We never push it to speed up more than half of its
2746 * normal turn time, otherwise it could take over the cpu.
2749 speedup_syncer(void)
2753 mtx_lock(&sync_mtx);
2754 if (rushjob < syncdelay / 2) {
2756 stat_rush_requests += 1;
2759 mtx_unlock(&sync_mtx);
2760 cv_broadcast(&sync_wakeup);
2765 * Tell the syncer to speed up its work and run though its work
2766 * list several times, then tell it to shut down.
2769 syncer_shutdown(void *arg, int howto)
2772 if (howto & RB_NOSYNC)
2774 mtx_lock(&sync_mtx);
2775 syncer_state = SYNCER_SHUTTING_DOWN;
2777 mtx_unlock(&sync_mtx);
2778 cv_broadcast(&sync_wakeup);
2779 kproc_shutdown(arg, howto);
2783 syncer_suspend(void)
2786 syncer_shutdown(updateproc, 0);
2793 mtx_lock(&sync_mtx);
2795 syncer_state = SYNCER_RUNNING;
2796 mtx_unlock(&sync_mtx);
2797 cv_broadcast(&sync_wakeup);
2798 kproc_resume(updateproc);
2802 * Move the buffer between the clean and dirty lists of its vnode.
2805 reassignbuf(struct buf *bp)
2817 KASSERT((bp->b_flags & B_PAGING) == 0,
2818 ("%s: cannot reassign paging buffer %p", __func__, bp));
2820 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2821 bp, bp->b_vp, bp->b_flags);
2824 buf_vlist_remove(bp);
2827 * If dirty, put on list of dirty buffers; otherwise insert onto list
2830 if (bp->b_flags & B_DELWRI) {
2831 if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2832 switch (vp->v_type) {
2842 vn_syncer_add_to_worklist(bo, delay);
2844 buf_vlist_add(bp, bo, BX_VNDIRTY);
2846 buf_vlist_add(bp, bo, BX_VNCLEAN);
2848 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2849 mtx_lock(&sync_mtx);
2850 LIST_REMOVE(bo, bo_synclist);
2851 syncer_worklist_len--;
2852 mtx_unlock(&sync_mtx);
2853 bo->bo_flag &= ~BO_ONWORKLST;
2858 bp = TAILQ_FIRST(&bv->bv_hd);
2859 KASSERT(bp == NULL || bp->b_bufobj == bo,
2860 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2861 bp = TAILQ_LAST(&bv->bv_hd, buflists);
2862 KASSERT(bp == NULL || bp->b_bufobj == bo,
2863 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2865 bp = TAILQ_FIRST(&bv->bv_hd);
2866 KASSERT(bp == NULL || bp->b_bufobj == bo,
2867 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2868 bp = TAILQ_LAST(&bv->bv_hd, buflists);
2869 KASSERT(bp == NULL || bp->b_bufobj == bo,
2870 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2876 v_init_counters(struct vnode *vp)
2879 VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2880 vp, ("%s called for an initialized vnode", __FUNCTION__));
2881 ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2883 refcount_init(&vp->v_holdcnt, 1);
2884 refcount_init(&vp->v_usecount, 1);
2888 * Grab a particular vnode from the free list, increment its
2889 * reference count and lock it. VIRF_DOOMED is set if the vnode
2890 * is being destroyed. Only callers who specify LK_RETRY will
2891 * see doomed vnodes. If inactive processing was delayed in
2892 * vput try to do it here.
2894 * usecount is manipulated using atomics without holding any locks.
2896 * holdcnt can be manipulated using atomics without holding any locks,
2897 * except when transitioning 1<->0, in which case the interlock is held.
2899 * Consumers which don't guarantee liveness of the vnode can use SMR to
2900 * try to get a reference. Note this operation can fail since the vnode
2901 * may be awaiting getting freed by the time they get to it.
2904 vget_prep_smr(struct vnode *vp)
2908 VFS_SMR_ASSERT_ENTERED();
2910 if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2922 vget_prep(struct vnode *vp)
2926 if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2936 vget_abort(struct vnode *vp, enum vgetstate vs)
2947 __assert_unreachable();
2952 vget(struct vnode *vp, int flags)
2957 return (vget_finish(vp, flags, vs));
2961 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2965 if ((flags & LK_INTERLOCK) != 0)
2966 ASSERT_VI_LOCKED(vp, __func__);
2968 ASSERT_VI_UNLOCKED(vp, __func__);
2969 VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
2970 VNPASS(vp->v_holdcnt > 0, vp);
2971 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
2973 error = vn_lock(vp, flags);
2974 if (__predict_false(error != 0)) {
2976 CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2981 vget_finish_ref(vp, vs);
2986 vget_finish_ref(struct vnode *vp, enum vgetstate vs)
2990 VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
2991 VNPASS(vp->v_holdcnt > 0, vp);
2992 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
2994 if (vs == VGET_USECOUNT)
2998 * We hold the vnode. If the usecount is 0 it will be utilized to keep
2999 * the vnode around. Otherwise someone else lended their hold count and
3000 * we have to drop ours.
3002 old = atomic_fetchadd_int(&vp->v_usecount, 1);
3003 VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3006 old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3007 VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3009 refcount_release(&vp->v_holdcnt);
3015 vref(struct vnode *vp)
3019 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3021 vget_finish_ref(vp, vs);
3025 vrefact(struct vnode *vp)
3028 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3030 int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3031 VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3033 refcount_acquire(&vp->v_usecount);
3038 vlazy(struct vnode *vp)
3042 VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3044 if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3047 * We may get here for inactive routines after the vnode got doomed.
3049 if (VN_IS_DOOMED(vp))
3052 mtx_lock(&mp->mnt_listmtx);
3053 if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3054 vp->v_mflag |= VMP_LAZYLIST;
3055 TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3056 mp->mnt_lazyvnodelistsize++;
3058 mtx_unlock(&mp->mnt_listmtx);
3062 * This routine is only meant to be called from vgonel prior to dooming
3066 vunlazy_gone(struct vnode *vp)
3070 ASSERT_VOP_ELOCKED(vp, __func__);
3071 ASSERT_VI_LOCKED(vp, __func__);
3072 VNPASS(!VN_IS_DOOMED(vp), vp);
3074 if (vp->v_mflag & VMP_LAZYLIST) {
3076 mtx_lock(&mp->mnt_listmtx);
3077 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3078 vp->v_mflag &= ~VMP_LAZYLIST;
3079 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3080 mp->mnt_lazyvnodelistsize--;
3081 mtx_unlock(&mp->mnt_listmtx);
3086 vdefer_inactive(struct vnode *vp)
3089 ASSERT_VI_LOCKED(vp, __func__);
3090 VNASSERT(vp->v_holdcnt > 0, vp,
3091 ("%s: vnode without hold count", __func__));
3092 if (VN_IS_DOOMED(vp)) {
3096 if (vp->v_iflag & VI_DEFINACT) {
3097 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3101 if (vp->v_usecount > 0) {
3102 vp->v_iflag &= ~VI_OWEINACT;
3107 vp->v_iflag |= VI_DEFINACT;
3109 counter_u64_add(deferred_inact, 1);
3113 vdefer_inactive_unlocked(struct vnode *vp)
3117 if ((vp->v_iflag & VI_OWEINACT) == 0) {
3121 vdefer_inactive(vp);
3124 enum vput_op { VRELE, VPUT, VUNREF };
3127 * Handle ->v_usecount transitioning to 0.
3129 * By releasing the last usecount we take ownership of the hold count which
3130 * provides liveness of the vnode, meaning we have to vdrop.
3132 * For all vnodes we may need to perform inactive processing. It requires an
3133 * exclusive lock on the vnode, while it is legal to call here with only a
3134 * shared lock (or no locks). If locking the vnode in an expected manner fails,
3135 * inactive processing gets deferred to the syncer.
3137 * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3138 * on the lock being held all the way until VOP_INACTIVE. This in particular
3139 * happens with UFS which adds half-constructed vnodes to the hash, where they
3140 * can be found by other code.
3143 vput_final(struct vnode *vp, enum vput_op func)
3148 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3149 VNPASS(vp->v_holdcnt > 0, vp);
3154 * By the time we got here someone else might have transitioned
3155 * the count back to > 0.
3157 if (vp->v_usecount > 0)
3161 * If the vnode is doomed vgone already performed inactive processing
3164 if (VN_IS_DOOMED(vp))
3167 if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3170 if (vp->v_iflag & VI_DOINGINACT)
3174 * Locking operations here will drop the interlock and possibly the
3175 * vnode lock, opening a window where the vnode can get doomed all the
3176 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3179 vp->v_iflag |= VI_OWEINACT;
3180 want_unlock = false;
3184 switch (VOP_ISLOCKED(vp)) {
3190 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3195 * The lock has at least one sharer, but we have no way
3196 * to conclude whether this is us. Play it safe and
3205 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3206 error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3212 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3213 error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3219 if (func == VUNREF) {
3220 VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp,
3221 ("recursive vunref"));
3222 vp->v_vflag |= VV_UNREF;
3225 error = vinactive(vp);
3228 if (error != ERELOOKUP || !want_unlock)
3230 VOP_LOCK(vp, LK_EXCLUSIVE);
3233 vp->v_vflag &= ~VV_UNREF;
3236 vdefer_inactive(vp);
3246 * Decrement ->v_usecount for a vnode.
3248 * Releasing the last use count requires additional processing, see vput_final
3249 * above for details.
3251 * Comment above each variant denotes lock state on entry and exit.
3256 * out: same as passed in
3259 vrele(struct vnode *vp)
3262 ASSERT_VI_UNLOCKED(vp, __func__);
3263 if (!refcount_release(&vp->v_usecount))
3265 vput_final(vp, VRELE);
3273 vput(struct vnode *vp)
3276 ASSERT_VOP_LOCKED(vp, __func__);
3277 ASSERT_VI_UNLOCKED(vp, __func__);
3278 if (!refcount_release(&vp->v_usecount)) {
3282 vput_final(vp, VPUT);
3290 vunref(struct vnode *vp)
3293 ASSERT_VOP_LOCKED(vp, __func__);
3294 ASSERT_VI_UNLOCKED(vp, __func__);
3295 if (!refcount_release(&vp->v_usecount))
3297 vput_final(vp, VUNREF);
3301 vhold(struct vnode *vp)
3305 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3306 old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3307 VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3308 ("%s: wrong hold count %d", __func__, old));
3310 vn_freevnodes_dec();
3314 vholdnz(struct vnode *vp)
3317 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3319 int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3320 VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3321 ("%s: wrong hold count %d", __func__, old));
3323 atomic_add_int(&vp->v_holdcnt, 1);
3328 * Grab a hold count unless the vnode is freed.
3330 * Only use this routine if vfs smr is the only protection you have against
3331 * freeing the vnode.
3333 * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag
3334 * is not set. After the flag is set the vnode becomes immutable to anyone but
3335 * the thread which managed to set the flag.
3337 * It may be tempting to replace the loop with:
3338 * count = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3339 * if (count & VHOLD_NO_SMR) {
3340 * backpedal and error out;
3343 * However, while this is more performant, it hinders debugging by eliminating
3344 * the previously mentioned invariant.
3347 vhold_smr(struct vnode *vp)
3351 VFS_SMR_ASSERT_ENTERED();
3353 count = atomic_load_int(&vp->v_holdcnt);
3355 if (count & VHOLD_NO_SMR) {
3356 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3357 ("non-zero hold count with flags %d\n", count));
3360 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3361 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3363 vn_freevnodes_dec();
3370 * Hold a free vnode for recycling.
3372 * Note: vnode_init references this comment.
3374 * Attempts to recycle only need the global vnode list lock and have no use for
3377 * However, vnodes get inserted into the global list before they get fully
3378 * initialized and stay there until UMA decides to free the memory. This in
3379 * particular means the target can be found before it becomes usable and after
3380 * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to
3383 * Note: the vnode may gain more references after we transition the count 0->1.
3386 vhold_recycle_free(struct vnode *vp)
3390 mtx_assert(&vnode_list_mtx, MA_OWNED);
3392 count = atomic_load_int(&vp->v_holdcnt);
3394 if (count & VHOLD_NO_SMR) {
3395 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3396 ("non-zero hold count with flags %d\n", count));
3399 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3403 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3404 vn_freevnodes_dec();
3410 static void __noinline
3411 vdbatch_process(struct vdbatch *vd)
3416 mtx_assert(&vd->lock, MA_OWNED);
3417 MPASS(curthread->td_pinned > 0);
3418 MPASS(vd->index == VDBATCH_SIZE);
3420 mtx_lock(&vnode_list_mtx);
3422 freevnodes += vd->freevnodes;
3423 for (i = 0; i < VDBATCH_SIZE; i++) {
3425 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3426 TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3427 MPASS(vp->v_dbatchcpu != NOCPU);
3428 vp->v_dbatchcpu = NOCPU;
3430 mtx_unlock(&vnode_list_mtx);
3432 bzero(vd->tab, sizeof(vd->tab));
3438 vdbatch_enqueue(struct vnode *vp)
3442 ASSERT_VI_LOCKED(vp, __func__);
3443 VNASSERT(!VN_IS_DOOMED(vp), vp,
3444 ("%s: deferring requeue of a doomed vnode", __func__));
3446 if (vp->v_dbatchcpu != NOCPU) {
3453 mtx_lock(&vd->lock);
3454 MPASS(vd->index < VDBATCH_SIZE);
3455 MPASS(vd->tab[vd->index] == NULL);
3457 * A hack: we depend on being pinned so that we know what to put in
3460 vp->v_dbatchcpu = curcpu;
3461 vd->tab[vd->index] = vp;
3464 if (vd->index == VDBATCH_SIZE)
3465 vdbatch_process(vd);
3466 mtx_unlock(&vd->lock);
3471 * This routine must only be called for vnodes which are about to be
3472 * deallocated. Supporting dequeue for arbitrary vndoes would require
3473 * validating that the locked batch matches.
3476 vdbatch_dequeue(struct vnode *vp)
3482 VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3483 ("%s: called for a used vnode\n", __func__));
3485 cpu = vp->v_dbatchcpu;
3489 vd = DPCPU_ID_PTR(cpu, vd);
3490 mtx_lock(&vd->lock);
3491 for (i = 0; i < vd->index; i++) {
3492 if (vd->tab[i] != vp)
3494 vp->v_dbatchcpu = NOCPU;
3496 vd->tab[i] = vd->tab[vd->index];
3497 vd->tab[vd->index] = NULL;
3500 mtx_unlock(&vd->lock);
3502 * Either we dequeued the vnode above or the target CPU beat us to it.
3504 MPASS(vp->v_dbatchcpu == NOCPU);
3508 * Drop the hold count of the vnode. If this is the last reference to
3509 * the vnode we place it on the free list unless it has been vgone'd
3510 * (marked VIRF_DOOMED) in which case we will free it.
3512 * Because the vnode vm object keeps a hold reference on the vnode if
3513 * there is at least one resident non-cached page, the vnode cannot
3514 * leave the active list without the page cleanup done.
3517 vdrop_deactivate(struct vnode *vp)
3521 ASSERT_VI_LOCKED(vp, __func__);
3523 * Mark a vnode as free: remove it from its active list
3524 * and put it up for recycling on the freelist.
3526 VNASSERT(!VN_IS_DOOMED(vp), vp,
3527 ("vdrop: returning doomed vnode"));
3528 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
3529 ("vnode with VI_OWEINACT set"));
3530 VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp,
3531 ("vnode with VI_DEFINACT set"));
3532 if (vp->v_mflag & VMP_LAZYLIST) {
3534 mtx_lock(&mp->mnt_listmtx);
3535 VNASSERT(vp->v_mflag & VMP_LAZYLIST, vp, ("lost VMP_LAZYLIST"));
3537 * Don't remove the vnode from the lazy list if another thread
3538 * has increased the hold count. It may have re-enqueued the
3539 * vnode to the lazy list and is now responsible for its
3542 if (vp->v_holdcnt == 0) {
3543 vp->v_mflag &= ~VMP_LAZYLIST;
3544 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3545 mp->mnt_lazyvnodelistsize--;
3547 mtx_unlock(&mp->mnt_listmtx);
3549 vdbatch_enqueue(vp);
3552 static void __noinline
3553 vdropl_final(struct vnode *vp)
3556 ASSERT_VI_LOCKED(vp, __func__);
3557 VNPASS(VN_IS_DOOMED(vp), vp);
3559 * Set the VHOLD_NO_SMR flag.
3561 * We may be racing against vhold_smr. If they win we can just pretend
3562 * we never got this far, they will vdrop later.
3564 if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) {
3565 vn_freevnodes_inc();
3568 * We lost the aforementioned race. Any subsequent access is
3569 * invalid as they might have managed to vdropl on their own.
3574 * Don't bump freevnodes as this one is going away.
3580 vdrop(struct vnode *vp)
3583 ASSERT_VI_UNLOCKED(vp, __func__);
3584 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3585 if (refcount_release_if_not_last(&vp->v_holdcnt))
3592 vdropl(struct vnode *vp)
3595 ASSERT_VI_LOCKED(vp, __func__);
3596 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3597 if (!refcount_release(&vp->v_holdcnt)) {
3601 if (!VN_IS_DOOMED(vp)) {
3602 vn_freevnodes_inc();
3603 vdrop_deactivate(vp);
3605 * Also unlocks the interlock. We can't assert on it as we
3606 * released our hold and by now the vnode might have been
3615 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3616 * flags. DOINGINACT prevents us from recursing in calls to vinactive.
3619 vinactivef(struct vnode *vp)
3621 struct vm_object *obj;
3624 ASSERT_VOP_ELOCKED(vp, "vinactive");
3625 ASSERT_VI_LOCKED(vp, "vinactive");
3626 VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3627 ("vinactive: recursed on VI_DOINGINACT"));
3628 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3629 vp->v_iflag |= VI_DOINGINACT;
3630 vp->v_iflag &= ~VI_OWEINACT;
3633 * Before moving off the active list, we must be sure that any
3634 * modified pages are converted into the vnode's dirty
3635 * buffers, since these will no longer be checked once the
3636 * vnode is on the inactive list.
3638 * The write-out of the dirty pages is asynchronous. At the
3639 * point that VOP_INACTIVE() is called, there could still be
3640 * pending I/O and dirty pages in the object.
3642 if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3643 vm_object_mightbedirty(obj)) {
3644 VM_OBJECT_WLOCK(obj);
3645 vm_object_page_clean(obj, 0, 0, 0);
3646 VM_OBJECT_WUNLOCK(obj);
3648 error = VOP_INACTIVE(vp);
3650 VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3651 ("vinactive: lost VI_DOINGINACT"));
3652 vp->v_iflag &= ~VI_DOINGINACT;
3657 vinactive(struct vnode *vp)
3660 ASSERT_VOP_ELOCKED(vp, "vinactive");
3661 ASSERT_VI_LOCKED(vp, "vinactive");
3662 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3664 if ((vp->v_iflag & VI_OWEINACT) == 0)
3666 if (vp->v_iflag & VI_DOINGINACT)
3668 if (vp->v_usecount > 0) {
3669 vp->v_iflag &= ~VI_OWEINACT;
3672 return (vinactivef(vp));
3676 * Remove any vnodes in the vnode table belonging to mount point mp.
3678 * If FORCECLOSE is not specified, there should not be any active ones,
3679 * return error if any are found (nb: this is a user error, not a
3680 * system error). If FORCECLOSE is specified, detach any active vnodes
3683 * If WRITECLOSE is set, only flush out regular file vnodes open for
3686 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3688 * `rootrefs' specifies the base reference count for the root vnode
3689 * of this filesystem. The root vnode is considered busy if its
3690 * v_usecount exceeds this value. On a successful return, vflush(, td)
3691 * will call vrele() on the root vnode exactly rootrefs times.
3692 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3696 static int busyprt = 0; /* print out busy vnodes */
3697 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3701 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3703 struct vnode *vp, *mvp, *rootvp = NULL;
3705 int busy = 0, error;
3707 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3710 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3711 ("vflush: bad args"));
3713 * Get the filesystem root vnode. We can vput() it
3714 * immediately, since with rootrefs > 0, it won't go away.
3716 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3717 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3724 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3726 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3729 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3733 * Skip over a vnodes marked VV_SYSTEM.
3735 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3741 * If WRITECLOSE is set, flush out unlinked but still open
3742 * files (even if open only for reading) and regular file
3743 * vnodes open for writing.
3745 if (flags & WRITECLOSE) {
3746 if (vp->v_object != NULL) {
3747 VM_OBJECT_WLOCK(vp->v_object);
3748 vm_object_page_clean(vp->v_object, 0, 0, 0);
3749 VM_OBJECT_WUNLOCK(vp->v_object);
3752 error = VOP_FSYNC(vp, MNT_WAIT, td);
3753 } while (error == ERELOOKUP);
3757 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3760 error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3763 if ((vp->v_type == VNON ||
3764 (error == 0 && vattr.va_nlink > 0)) &&
3765 (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3773 * With v_usecount == 0, all we need to do is clear out the
3774 * vnode data structures and we are done.
3776 * If FORCECLOSE is set, forcibly close the vnode.
3778 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3784 vn_printf(vp, "vflush: busy vnode ");
3790 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3792 * If just the root vnode is busy, and if its refcount
3793 * is equal to `rootrefs', then go ahead and kill it.
3796 KASSERT(busy > 0, ("vflush: not busy"));
3797 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3798 ("vflush: usecount %d < rootrefs %d",
3799 rootvp->v_usecount, rootrefs));
3800 if (busy == 1 && rootvp->v_usecount == rootrefs) {
3801 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3809 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3813 for (; rootrefs > 0; rootrefs--)
3819 * Recycle an unused vnode to the front of the free list.
3822 vrecycle(struct vnode *vp)
3827 recycled = vrecyclel(vp);
3833 * vrecycle, with the vp interlock held.
3836 vrecyclel(struct vnode *vp)
3840 ASSERT_VOP_ELOCKED(vp, __func__);
3841 ASSERT_VI_LOCKED(vp, __func__);
3842 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3844 if (vp->v_usecount == 0) {
3852 * Eliminate all activity associated with a vnode
3853 * in preparation for reuse.
3856 vgone(struct vnode *vp)
3864 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3865 struct vnode *lowervp __unused)
3870 * Notify upper mounts about reclaimed or unlinked vnode.
3873 vfs_notify_upper(struct vnode *vp, int event)
3875 static struct vfsops vgonel_vfsops = {
3876 .vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3877 .vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3879 struct mount *mp, *ump, *mmp;
3884 if (TAILQ_EMPTY(&mp->mnt_uppers))
3887 mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3888 mmp->mnt_op = &vgonel_vfsops;
3889 mmp->mnt_kern_flag |= MNTK_MARKER;
3891 mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3892 for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3893 if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3894 ump = TAILQ_NEXT(ump, mnt_upper_link);
3897 TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3900 case VFS_NOTIFY_UPPER_RECLAIM:
3901 VFS_RECLAIM_LOWERVP(ump, vp);
3903 case VFS_NOTIFY_UPPER_UNLINK:
3904 VFS_UNLINK_LOWERVP(ump, vp);
3907 KASSERT(0, ("invalid event %d", event));
3911 ump = TAILQ_NEXT(mmp, mnt_upper_link);
3912 TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3915 mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3916 if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3917 mp->mnt_kern_flag &= ~MNTK_VGONE_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 flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT);
4127 snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4128 strlcat(buf, buf2, sizeof(buf));
4130 if (vp->v_vflag & VV_ROOT)
4131 strlcat(buf, "|VV_ROOT", sizeof(buf));
4132 if (vp->v_vflag & VV_ISTTY)
4133 strlcat(buf, "|VV_ISTTY", sizeof(buf));
4134 if (vp->v_vflag & VV_NOSYNC)
4135 strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4136 if (vp->v_vflag & VV_ETERNALDEV)
4137 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4138 if (vp->v_vflag & VV_CACHEDLABEL)
4139 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4140 if (vp->v_vflag & VV_VMSIZEVNLOCK)
4141 strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4142 if (vp->v_vflag & VV_COPYONWRITE)
4143 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4144 if (vp->v_vflag & VV_SYSTEM)
4145 strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4146 if (vp->v_vflag & VV_PROCDEP)
4147 strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4148 if (vp->v_vflag & VV_NOKNOTE)
4149 strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
4150 if (vp->v_vflag & VV_DELETED)
4151 strlcat(buf, "|VV_DELETED", sizeof(buf));
4152 if (vp->v_vflag & VV_MD)
4153 strlcat(buf, "|VV_MD", sizeof(buf));
4154 if (vp->v_vflag & VV_FORCEINSMQ)
4155 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4156 if (vp->v_vflag & VV_READLINK)
4157 strlcat(buf, "|VV_READLINK", sizeof(buf));
4158 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4159 VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM |
4160 VV_PROCDEP | VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ |
4163 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4164 strlcat(buf, buf2, sizeof(buf));
4166 if (vp->v_iflag & VI_TEXT_REF)
4167 strlcat(buf, "|VI_TEXT_REF", 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_TEXT_REF | 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_VGONE_UPPER);
4349 MNT_KERN_FLAG(MNTK_VGONE_WAITER);
4350 MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
4351 MNT_KERN_FLAG(MNTK_MARKER);
4352 MNT_KERN_FLAG(MNTK_USES_BCACHE);
4353 MNT_KERN_FLAG(MNTK_FPLOOKUP);
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_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
4411 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4412 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed);
4413 db_printf(" mnt_lockref = %d (with %d in the struct)\n",
4414 vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4415 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4416 db_printf(" mnt_secondary_accwrites = %d\n",
4417 mp->mnt_secondary_accwrites);
4418 db_printf(" mnt_gjprovider = %s\n",
4419 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4420 db_printf(" mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4422 db_printf("\n\nList of active vnodes\n");
4423 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4424 if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4425 vn_printf(vp, "vnode ");
4430 db_printf("\n\nList of inactive vnodes\n");
4431 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4432 if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4433 vn_printf(vp, "vnode ");
4442 * Fill in a struct xvfsconf based on a struct vfsconf.
4445 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4447 struct xvfsconf xvfsp;
4449 bzero(&xvfsp, sizeof(xvfsp));
4450 strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4451 xvfsp.vfc_typenum = vfsp->vfc_typenum;
4452 xvfsp.vfc_refcount = vfsp->vfc_refcount;
4453 xvfsp.vfc_flags = vfsp->vfc_flags;
4455 * These are unused in userland, we keep them
4456 * to not break binary compatibility.
4458 xvfsp.vfc_vfsops = NULL;
4459 xvfsp.vfc_next = NULL;
4460 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4463 #ifdef COMPAT_FREEBSD32
4465 uint32_t vfc_vfsops;
4466 char vfc_name[MFSNAMELEN];
4467 int32_t vfc_typenum;
4468 int32_t vfc_refcount;
4474 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4476 struct xvfsconf32 xvfsp;
4478 bzero(&xvfsp, sizeof(xvfsp));
4479 strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4480 xvfsp.vfc_typenum = vfsp->vfc_typenum;
4481 xvfsp.vfc_refcount = vfsp->vfc_refcount;
4482 xvfsp.vfc_flags = vfsp->vfc_flags;
4483 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4488 * Top level filesystem related information gathering.
4491 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4493 struct vfsconf *vfsp;
4498 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4499 #ifdef COMPAT_FREEBSD32
4500 if (req->flags & SCTL_MASK32)
4501 error = vfsconf2x32(req, vfsp);
4504 error = vfsconf2x(req, vfsp);
4512 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4513 CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4514 "S,xvfsconf", "List of all configured filesystems");
4516 #ifndef BURN_BRIDGES
4517 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4520 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4522 int *name = (int *)arg1 - 1; /* XXX */
4523 u_int namelen = arg2 + 1; /* XXX */
4524 struct vfsconf *vfsp;
4526 log(LOG_WARNING, "userland calling deprecated sysctl, "
4527 "please rebuild world\n");
4529 #if 1 || defined(COMPAT_PRELITE2)
4530 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4532 return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4536 case VFS_MAXTYPENUM:
4539 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4542 return (ENOTDIR); /* overloaded */
4544 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4545 if (vfsp->vfc_typenum == name[2])
4550 return (EOPNOTSUPP);
4551 #ifdef COMPAT_FREEBSD32
4552 if (req->flags & SCTL_MASK32)
4553 return (vfsconf2x32(req, vfsp));
4556 return (vfsconf2x(req, vfsp));
4558 return (EOPNOTSUPP);
4561 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4562 CTLFLAG_MPSAFE, vfs_sysctl,
4563 "Generic filesystem");
4565 #if 1 || defined(COMPAT_PRELITE2)
4568 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4571 struct vfsconf *vfsp;
4572 struct ovfsconf ovfs;
4575 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4576 bzero(&ovfs, sizeof(ovfs));
4577 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */
4578 strcpy(ovfs.vfc_name, vfsp->vfc_name);
4579 ovfs.vfc_index = vfsp->vfc_typenum;
4580 ovfs.vfc_refcount = vfsp->vfc_refcount;
4581 ovfs.vfc_flags = vfsp->vfc_flags;
4582 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4592 #endif /* 1 || COMPAT_PRELITE2 */
4593 #endif /* !BURN_BRIDGES */
4595 #define KINFO_VNODESLOP 10
4598 * Dump vnode list (via sysctl).
4602 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4610 * Stale numvnodes access is not fatal here.
4613 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4615 /* Make an estimate */
4616 return (SYSCTL_OUT(req, 0, len));
4618 error = sysctl_wire_old_buffer(req, 0);
4621 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4623 mtx_lock(&mountlist_mtx);
4624 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4625 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4628 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4632 xvn[n].xv_size = sizeof *xvn;
4633 xvn[n].xv_vnode = vp;
4634 xvn[n].xv_id = 0; /* XXX compat */
4635 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4637 XV_COPY(writecount);
4643 xvn[n].xv_flag = vp->v_vflag;
4645 switch (vp->v_type) {
4652 if (vp->v_rdev == NULL) {
4656 xvn[n].xv_dev = dev2udev(vp->v_rdev);
4659 xvn[n].xv_socket = vp->v_socket;
4662 xvn[n].xv_fifo = vp->v_fifoinfo;
4667 /* shouldn't happen? */
4675 mtx_lock(&mountlist_mtx);
4680 mtx_unlock(&mountlist_mtx);
4682 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4687 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4688 CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4693 unmount_or_warn(struct mount *mp)
4697 error = dounmount(mp, MNT_FORCE, curthread);
4699 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4703 printf("%d)\n", error);
4708 * Unmount all filesystems. The list is traversed in reverse order
4709 * of mounting to avoid dependencies.
4712 vfs_unmountall(void)
4714 struct mount *mp, *tmp;
4716 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4719 * Since this only runs when rebooting, it is not interlocked.
4721 TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4725 * Forcibly unmounting "/dev" before "/" would prevent clean
4726 * unmount of the latter.
4728 if (mp == rootdevmp)
4731 unmount_or_warn(mp);
4734 if (rootdevmp != NULL)
4735 unmount_or_warn(rootdevmp);
4739 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4742 ASSERT_VI_LOCKED(vp, __func__);
4743 VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4744 if ((vp->v_iflag & VI_OWEINACT) == 0) {
4748 if (vn_lock(vp, lkflags) == 0) {
4755 vdefer_inactive_unlocked(vp);
4759 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4762 return (vp->v_iflag & VI_DEFINACT);
4765 static void __noinline
4766 vfs_periodic_inactive(struct mount *mp, int flags)
4768 struct vnode *vp, *mvp;
4771 lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4772 if (flags != MNT_WAIT)
4773 lkflags |= LK_NOWAIT;
4775 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4776 if ((vp->v_iflag & VI_DEFINACT) == 0) {
4780 vp->v_iflag &= ~VI_DEFINACT;
4781 vfs_deferred_inactive(vp, lkflags);
4786 vfs_want_msync(struct vnode *vp)
4788 struct vm_object *obj;
4791 * This test may be performed without any locks held.
4792 * We rely on vm_object's type stability.
4794 if (vp->v_vflag & VV_NOSYNC)
4797 return (obj != NULL && vm_object_mightbedirty(obj));
4801 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4804 if (vp->v_vflag & VV_NOSYNC)
4806 if (vp->v_iflag & VI_DEFINACT)
4808 return (vfs_want_msync(vp));
4811 static void __noinline
4812 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4814 struct vnode *vp, *mvp;
4815 struct vm_object *obj;
4816 int lkflags, objflags;
4819 lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4820 if (flags != MNT_WAIT) {
4821 lkflags |= LK_NOWAIT;
4822 objflags = OBJPC_NOSYNC;
4824 objflags = OBJPC_SYNC;
4827 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4829 if (vp->v_iflag & VI_DEFINACT) {
4830 vp->v_iflag &= ~VI_DEFINACT;
4833 if (!vfs_want_msync(vp)) {
4835 vfs_deferred_inactive(vp, lkflags);
4840 if (vget(vp, lkflags) == 0) {
4842 if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4843 VM_OBJECT_WLOCK(obj);
4844 vm_object_page_clean(obj, 0, 0, objflags);
4845 VM_OBJECT_WUNLOCK(obj);
4852 vdefer_inactive_unlocked(vp);
4858 vfs_periodic(struct mount *mp, int flags)
4861 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4863 if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4864 vfs_periodic_inactive(mp, flags);
4866 vfs_periodic_msync_inactive(mp, flags);
4870 destroy_vpollinfo_free(struct vpollinfo *vi)
4873 knlist_destroy(&vi->vpi_selinfo.si_note);
4874 mtx_destroy(&vi->vpi_lock);
4875 free(vi, M_VNODEPOLL);
4879 destroy_vpollinfo(struct vpollinfo *vi)
4882 knlist_clear(&vi->vpi_selinfo.si_note, 1);
4883 seldrain(&vi->vpi_selinfo);
4884 destroy_vpollinfo_free(vi);
4888 * Initialize per-vnode helper structure to hold poll-related state.
4891 v_addpollinfo(struct vnode *vp)
4893 struct vpollinfo *vi;
4895 if (vp->v_pollinfo != NULL)
4897 vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO);
4898 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4899 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4900 vfs_knlunlock, vfs_knl_assert_lock);
4902 if (vp->v_pollinfo != NULL) {
4904 destroy_vpollinfo_free(vi);
4907 vp->v_pollinfo = vi;
4912 * Record a process's interest in events which might happen to
4913 * a vnode. Because poll uses the historic select-style interface
4914 * internally, this routine serves as both the ``check for any
4915 * pending events'' and the ``record my interest in future events''
4916 * functions. (These are done together, while the lock is held,
4917 * to avoid race conditions.)
4920 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4924 mtx_lock(&vp->v_pollinfo->vpi_lock);
4925 if (vp->v_pollinfo->vpi_revents & events) {
4927 * This leaves events we are not interested
4928 * in available for the other process which
4929 * which presumably had requested them
4930 * (otherwise they would never have been
4933 events &= vp->v_pollinfo->vpi_revents;
4934 vp->v_pollinfo->vpi_revents &= ~events;
4936 mtx_unlock(&vp->v_pollinfo->vpi_lock);
4939 vp->v_pollinfo->vpi_events |= events;
4940 selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4941 mtx_unlock(&vp->v_pollinfo->vpi_lock);
4946 * Routine to create and manage a filesystem syncer vnode.
4948 #define sync_close ((int (*)(struct vop_close_args *))nullop)
4949 static int sync_fsync(struct vop_fsync_args *);
4950 static int sync_inactive(struct vop_inactive_args *);
4951 static int sync_reclaim(struct vop_reclaim_args *);
4953 static struct vop_vector sync_vnodeops = {
4954 .vop_bypass = VOP_EOPNOTSUPP,
4955 .vop_close = sync_close, /* close */
4956 .vop_fsync = sync_fsync, /* fsync */
4957 .vop_inactive = sync_inactive, /* inactive */
4958 .vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4959 .vop_reclaim = sync_reclaim, /* reclaim */
4960 .vop_lock1 = vop_stdlock, /* lock */
4961 .vop_unlock = vop_stdunlock, /* unlock */
4962 .vop_islocked = vop_stdislocked, /* islocked */
4964 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4967 * Create a new filesystem syncer vnode for the specified mount point.
4970 vfs_allocate_syncvnode(struct mount *mp)
4974 static long start, incr, next;
4977 /* Allocate a new vnode */
4978 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4980 panic("vfs_allocate_syncvnode: getnewvnode() failed");
4982 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4983 vp->v_vflag |= VV_FORCEINSMQ;
4984 error = insmntque(vp, mp);
4986 panic("vfs_allocate_syncvnode: insmntque() failed");
4987 vp->v_vflag &= ~VV_FORCEINSMQ;
4990 * Place the vnode onto the syncer worklist. We attempt to
4991 * scatter them about on the list so that they will go off
4992 * at evenly distributed times even if all the filesystems
4993 * are mounted at once.
4996 if (next == 0 || next > syncer_maxdelay) {
5000 start = syncer_maxdelay / 2;
5001 incr = syncer_maxdelay;
5007 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
5008 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
5009 mtx_lock(&sync_mtx);
5011 if (mp->mnt_syncer == NULL) {
5012 mp->mnt_syncer = vp;
5015 mtx_unlock(&sync_mtx);
5018 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5025 vfs_deallocate_syncvnode(struct mount *mp)
5029 mtx_lock(&sync_mtx);
5030 vp = mp->mnt_syncer;
5032 mp->mnt_syncer = NULL;
5033 mtx_unlock(&sync_mtx);
5039 * Do a lazy sync of the filesystem.
5042 sync_fsync(struct vop_fsync_args *ap)
5044 struct vnode *syncvp = ap->a_vp;
5045 struct mount *mp = syncvp->v_mount;
5050 * We only need to do something if this is a lazy evaluation.
5052 if (ap->a_waitfor != MNT_LAZY)
5056 * Move ourselves to the back of the sync list.
5058 bo = &syncvp->v_bufobj;
5060 vn_syncer_add_to_worklist(bo, syncdelay);
5064 * Walk the list of vnodes pushing all that are dirty and
5065 * not already on the sync list.
5067 if (vfs_busy(mp, MBF_NOWAIT) != 0)
5069 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
5073 save = curthread_pflags_set(TDP_SYNCIO);
5075 * The filesystem at hand may be idle with free vnodes stored in the
5076 * batch. Return them instead of letting them stay there indefinitely.
5078 vfs_periodic(mp, MNT_NOWAIT);
5079 error = VFS_SYNC(mp, MNT_LAZY);
5080 curthread_pflags_restore(save);
5081 vn_finished_write(mp);
5087 * The syncer vnode is no referenced.
5090 sync_inactive(struct vop_inactive_args *ap)
5098 * The syncer vnode is no longer needed and is being decommissioned.
5100 * Modifications to the worklist must be protected by sync_mtx.
5103 sync_reclaim(struct vop_reclaim_args *ap)
5105 struct vnode *vp = ap->a_vp;
5110 mtx_lock(&sync_mtx);
5111 if (vp->v_mount->mnt_syncer == vp)
5112 vp->v_mount->mnt_syncer = NULL;
5113 if (bo->bo_flag & BO_ONWORKLST) {
5114 LIST_REMOVE(bo, bo_synclist);
5115 syncer_worklist_len--;
5117 bo->bo_flag &= ~BO_ONWORKLST;
5119 mtx_unlock(&sync_mtx);
5126 vn_need_pageq_flush(struct vnode *vp)
5128 struct vm_object *obj;
5131 MPASS(mtx_owned(VI_MTX(vp)));
5133 if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5134 vm_object_mightbedirty(obj))
5140 * Check if vnode represents a disk device
5143 vn_isdisk_error(struct vnode *vp, int *errp)
5147 if (vp->v_type != VCHR) {
5153 if (vp->v_rdev == NULL)
5155 else if (vp->v_rdev->si_devsw == NULL)
5157 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5162 return (error == 0);
5166 vn_isdisk(struct vnode *vp)
5170 return (vn_isdisk_error(vp, &error));
5174 * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
5175 * the comment above cache_fplookup for details.
5178 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred)
5182 VFS_SMR_ASSERT_ENTERED();
5184 /* Check the owner. */
5185 if (cred->cr_uid == file_uid) {
5186 if (file_mode & S_IXUSR)
5191 /* Otherwise, check the groups (first match) */
5192 if (groupmember(file_gid, cred)) {
5193 if (file_mode & S_IXGRP)
5198 /* Otherwise, check everyone else. */
5199 if (file_mode & S_IXOTH)
5203 * Permission check failed, but it is possible denial will get overwritten
5204 * (e.g., when root is traversing through a 700 directory owned by someone
5207 * vaccess() calls priv_check_cred which in turn can descent into MAC
5208 * modules overriding this result. It's quite unclear what semantics
5209 * are allowed for them to operate, thus for safety we don't call them
5210 * from within the SMR section. This also means if any such modules
5211 * are present, we have to let the regular lookup decide.
5213 error = priv_check_cred_vfs_lookup_nomac(cred);
5219 * MAC modules present.
5230 * Common filesystem object access control check routine. Accepts a
5231 * vnode's type, "mode", uid and gid, requested access mode, and credentials.
5232 * Returns 0 on success, or an errno on failure.
5235 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5236 accmode_t accmode, struct ucred *cred)
5238 accmode_t dac_granted;
5239 accmode_t priv_granted;
5241 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5242 ("invalid bit in accmode"));
5243 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5244 ("VAPPEND without VWRITE"));
5247 * Look for a normal, non-privileged way to access the file/directory
5248 * as requested. If it exists, go with that.
5253 /* Check the owner. */
5254 if (cred->cr_uid == file_uid) {
5255 dac_granted |= VADMIN;
5256 if (file_mode & S_IXUSR)
5257 dac_granted |= VEXEC;
5258 if (file_mode & S_IRUSR)
5259 dac_granted |= VREAD;
5260 if (file_mode & S_IWUSR)
5261 dac_granted |= (VWRITE | VAPPEND);
5263 if ((accmode & dac_granted) == accmode)
5269 /* Otherwise, check the groups (first match) */
5270 if (groupmember(file_gid, cred)) {
5271 if (file_mode & S_IXGRP)
5272 dac_granted |= VEXEC;
5273 if (file_mode & S_IRGRP)
5274 dac_granted |= VREAD;
5275 if (file_mode & S_IWGRP)
5276 dac_granted |= (VWRITE | VAPPEND);
5278 if ((accmode & dac_granted) == accmode)
5284 /* Otherwise, check everyone else. */
5285 if (file_mode & S_IXOTH)
5286 dac_granted |= VEXEC;
5287 if (file_mode & S_IROTH)
5288 dac_granted |= VREAD;
5289 if (file_mode & S_IWOTH)
5290 dac_granted |= (VWRITE | VAPPEND);
5291 if ((accmode & dac_granted) == accmode)
5296 * Build a privilege mask to determine if the set of privileges
5297 * satisfies the requirements when combined with the granted mask
5298 * from above. For each privilege, if the privilege is required,
5299 * bitwise or the request type onto the priv_granted mask.
5305 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5306 * requests, instead of PRIV_VFS_EXEC.
5308 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5309 !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5310 priv_granted |= VEXEC;
5313 * Ensure that at least one execute bit is on. Otherwise,
5314 * a privileged user will always succeed, and we don't want
5315 * this to happen unless the file really is executable.
5317 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5318 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5319 !priv_check_cred(cred, PRIV_VFS_EXEC))
5320 priv_granted |= VEXEC;
5323 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5324 !priv_check_cred(cred, PRIV_VFS_READ))
5325 priv_granted |= VREAD;
5327 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5328 !priv_check_cred(cred, PRIV_VFS_WRITE))
5329 priv_granted |= (VWRITE | VAPPEND);
5331 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5332 !priv_check_cred(cred, PRIV_VFS_ADMIN))
5333 priv_granted |= VADMIN;
5335 if ((accmode & (priv_granted | dac_granted)) == accmode) {
5339 return ((accmode & VADMIN) ? EPERM : EACCES);
5343 * Credential check based on process requesting service, and per-attribute
5347 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5348 struct thread *td, accmode_t accmode)
5352 * Kernel-invoked always succeeds.
5358 * Do not allow privileged processes in jail to directly manipulate
5359 * system attributes.
5361 switch (attrnamespace) {
5362 case EXTATTR_NAMESPACE_SYSTEM:
5363 /* Potentially should be: return (EPERM); */
5364 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5365 case EXTATTR_NAMESPACE_USER:
5366 return (VOP_ACCESS(vp, accmode, cred, td));
5372 #ifdef DEBUG_VFS_LOCKS
5374 * This only exists to suppress warnings from unlocked specfs accesses. It is
5375 * no longer ok to have an unlocked VFS.
5377 #define IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL || \
5378 (vp)->v_type == VCHR || (vp)->v_type == VBAD)
5380 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */
5381 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5382 "Drop into debugger on lock violation");
5384 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */
5385 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5386 0, "Check for interlock across VOPs");
5388 int vfs_badlock_print = 1; /* Print lock violations. */
5389 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5390 0, "Print lock violations");
5392 int vfs_badlock_vnode = 1; /* Print vnode details on lock violations. */
5393 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5394 0, "Print vnode details on lock violations");
5397 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */
5398 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5399 &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5403 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5407 if (vfs_badlock_backtrace)
5410 if (vfs_badlock_vnode)
5411 vn_printf(vp, "vnode ");
5412 if (vfs_badlock_print)
5413 printf("%s: %p %s\n", str, (void *)vp, msg);
5414 if (vfs_badlock_ddb)
5415 kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5419 assert_vi_locked(struct vnode *vp, const char *str)
5422 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5423 vfs_badlock("interlock is not locked but should be", str, vp);
5427 assert_vi_unlocked(struct vnode *vp, const char *str)
5430 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5431 vfs_badlock("interlock is locked but should not be", str, vp);
5435 assert_vop_locked(struct vnode *vp, const char *str)
5439 if (!IGNORE_LOCK(vp)) {
5440 locked = VOP_ISLOCKED(vp);
5441 if (locked == 0 || locked == LK_EXCLOTHER)
5442 vfs_badlock("is not locked but should be", str, vp);
5447 assert_vop_unlocked(struct vnode *vp, const char *str)
5450 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5451 vfs_badlock("is locked but should not be", str, vp);
5455 assert_vop_elocked(struct vnode *vp, const char *str)
5458 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5459 vfs_badlock("is not exclusive locked but should be", str, vp);
5461 #endif /* DEBUG_VFS_LOCKS */
5464 vop_rename_fail(struct vop_rename_args *ap)
5467 if (ap->a_tvp != NULL)
5469 if (ap->a_tdvp == ap->a_tvp)
5478 vop_rename_pre(void *ap)
5480 struct vop_rename_args *a = ap;
5482 #ifdef DEBUG_VFS_LOCKS
5484 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5485 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5486 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5487 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5489 /* Check the source (from). */
5490 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5491 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5492 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5493 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5494 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5496 /* Check the target. */
5498 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5499 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5502 * It may be tempting to add vn_seqc_write_begin/end calls here and
5503 * in vop_rename_post but that's not going to work out since some
5504 * filesystems relookup vnodes mid-rename. This is probably a bug.
5506 * For now filesystems are expected to do the relevant calls after they
5507 * decide what vnodes to operate on.
5509 if (a->a_tdvp != a->a_fdvp)
5511 if (a->a_tvp != a->a_fvp)
5518 #ifdef DEBUG_VFS_LOCKS
5520 vop_fplookup_vexec_debugpre(void *ap __unused)
5523 VFS_SMR_ASSERT_ENTERED();
5527 vop_fplookup_vexec_debugpost(void *ap __unused, int rc __unused)
5530 VFS_SMR_ASSERT_ENTERED();
5534 vop_fplookup_symlink_debugpre(void *ap __unused)
5537 VFS_SMR_ASSERT_ENTERED();
5541 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused)
5544 VFS_SMR_ASSERT_ENTERED();
5547 vop_strategy_debugpre(void *ap)
5549 struct vop_strategy_args *a;
5556 * Cluster ops lock their component buffers but not the IO container.
5558 if ((bp->b_flags & B_CLUSTER) != 0)
5561 if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5562 if (vfs_badlock_print)
5564 "VOP_STRATEGY: bp is not locked but should be\n");
5565 if (vfs_badlock_ddb)
5566 kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5571 vop_lock_debugpre(void *ap)
5573 struct vop_lock1_args *a = ap;
5575 if ((a->a_flags & LK_INTERLOCK) == 0)
5576 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5578 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5582 vop_lock_debugpost(void *ap, int rc)
5584 struct vop_lock1_args *a = ap;
5586 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5587 if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5588 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5592 vop_unlock_debugpre(void *ap)
5594 struct vop_unlock_args *a = ap;
5596 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5600 vop_need_inactive_debugpre(void *ap)
5602 struct vop_need_inactive_args *a = ap;
5604 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5608 vop_need_inactive_debugpost(void *ap, int rc)
5610 struct vop_need_inactive_args *a = ap;
5612 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5617 vop_create_pre(void *ap)
5619 struct vop_create_args *a;
5624 vn_seqc_write_begin(dvp);
5628 vop_create_post(void *ap, int rc)
5630 struct vop_create_args *a;
5635 vn_seqc_write_end(dvp);
5637 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5641 vop_whiteout_pre(void *ap)
5643 struct vop_whiteout_args *a;
5648 vn_seqc_write_begin(dvp);
5652 vop_whiteout_post(void *ap, int rc)
5654 struct vop_whiteout_args *a;
5659 vn_seqc_write_end(dvp);
5663 vop_deleteextattr_pre(void *ap)
5665 struct vop_deleteextattr_args *a;
5670 vn_seqc_write_begin(vp);
5674 vop_deleteextattr_post(void *ap, int rc)
5676 struct vop_deleteextattr_args *a;
5681 vn_seqc_write_end(vp);
5683 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5687 vop_link_pre(void *ap)
5689 struct vop_link_args *a;
5690 struct vnode *vp, *tdvp;
5695 vn_seqc_write_begin(vp);
5696 vn_seqc_write_begin(tdvp);
5700 vop_link_post(void *ap, int rc)
5702 struct vop_link_args *a;
5703 struct vnode *vp, *tdvp;
5708 vn_seqc_write_end(vp);
5709 vn_seqc_write_end(tdvp);
5711 VFS_KNOTE_LOCKED(vp, NOTE_LINK);
5712 VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE);
5717 vop_mkdir_pre(void *ap)
5719 struct vop_mkdir_args *a;
5724 vn_seqc_write_begin(dvp);
5728 vop_mkdir_post(void *ap, int rc)
5730 struct vop_mkdir_args *a;
5735 vn_seqc_write_end(dvp);
5737 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5740 #ifdef DEBUG_VFS_LOCKS
5742 vop_mkdir_debugpost(void *ap, int rc)
5744 struct vop_mkdir_args *a;
5748 cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp);
5753 vop_mknod_pre(void *ap)
5755 struct vop_mknod_args *a;
5760 vn_seqc_write_begin(dvp);
5764 vop_mknod_post(void *ap, int rc)
5766 struct vop_mknod_args *a;
5771 vn_seqc_write_end(dvp);
5773 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5777 vop_reclaim_post(void *ap, int rc)
5779 struct vop_reclaim_args *a;
5784 ASSERT_VOP_IN_SEQC(vp);
5786 VFS_KNOTE_LOCKED(vp, NOTE_REVOKE);
5790 vop_remove_pre(void *ap)
5792 struct vop_remove_args *a;
5793 struct vnode *dvp, *vp;
5798 vn_seqc_write_begin(dvp);
5799 vn_seqc_write_begin(vp);
5803 vop_remove_post(void *ap, int rc)
5805 struct vop_remove_args *a;
5806 struct vnode *dvp, *vp;
5811 vn_seqc_write_end(dvp);
5812 vn_seqc_write_end(vp);
5814 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5815 VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5820 vop_rename_post(void *ap, int rc)
5822 struct vop_rename_args *a = ap;
5827 if (a->a_fdvp == a->a_tdvp) {
5828 if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5830 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5831 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5833 hint |= NOTE_EXTEND;
5834 if (a->a_fvp->v_type == VDIR)
5836 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5838 if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5839 a->a_tvp->v_type == VDIR)
5841 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5844 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5846 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5848 if (a->a_tdvp != a->a_fdvp)
5850 if (a->a_tvp != a->a_fvp)
5858 vop_rmdir_pre(void *ap)
5860 struct vop_rmdir_args *a;
5861 struct vnode *dvp, *vp;
5866 vn_seqc_write_begin(dvp);
5867 vn_seqc_write_begin(vp);
5871 vop_rmdir_post(void *ap, int rc)
5873 struct vop_rmdir_args *a;
5874 struct vnode *dvp, *vp;
5879 vn_seqc_write_end(dvp);
5880 vn_seqc_write_end(vp);
5882 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5883 VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5888 vop_setattr_pre(void *ap)
5890 struct vop_setattr_args *a;
5895 vn_seqc_write_begin(vp);
5899 vop_setattr_post(void *ap, int rc)
5901 struct vop_setattr_args *a;
5906 vn_seqc_write_end(vp);
5908 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5912 vop_setacl_pre(void *ap)
5914 struct vop_setacl_args *a;
5919 vn_seqc_write_begin(vp);
5923 vop_setacl_post(void *ap, int rc __unused)
5925 struct vop_setacl_args *a;
5930 vn_seqc_write_end(vp);
5934 vop_setextattr_pre(void *ap)
5936 struct vop_setextattr_args *a;
5941 vn_seqc_write_begin(vp);
5945 vop_setextattr_post(void *ap, int rc)
5947 struct vop_setextattr_args *a;
5952 vn_seqc_write_end(vp);
5954 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5958 vop_symlink_pre(void *ap)
5960 struct vop_symlink_args *a;
5965 vn_seqc_write_begin(dvp);
5969 vop_symlink_post(void *ap, int rc)
5971 struct vop_symlink_args *a;
5976 vn_seqc_write_end(dvp);
5978 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5982 vop_open_post(void *ap, int rc)
5984 struct vop_open_args *a = ap;
5987 VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5991 vop_close_post(void *ap, int rc)
5993 struct vop_close_args *a = ap;
5995 if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5996 !VN_IS_DOOMED(a->a_vp))) {
5997 VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5998 NOTE_CLOSE_WRITE : NOTE_CLOSE);
6003 vop_read_post(void *ap, int rc)
6005 struct vop_read_args *a = ap;
6008 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6012 vop_read_pgcache_post(void *ap, int rc)
6014 struct vop_read_pgcache_args *a = ap;
6017 VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ);
6021 vop_readdir_post(void *ap, int rc)
6023 struct vop_readdir_args *a = ap;
6026 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6029 static struct knlist fs_knlist;
6032 vfs_event_init(void *arg)
6034 knlist_init_mtx(&fs_knlist, NULL);
6036 /* XXX - correct order? */
6037 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
6040 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
6043 KNOTE_UNLOCKED(&fs_knlist, event);
6046 static int filt_fsattach(struct knote *kn);
6047 static void filt_fsdetach(struct knote *kn);
6048 static int filt_fsevent(struct knote *kn, long hint);
6050 struct filterops fs_filtops = {
6052 .f_attach = filt_fsattach,
6053 .f_detach = filt_fsdetach,
6054 .f_event = filt_fsevent
6058 filt_fsattach(struct knote *kn)
6061 kn->kn_flags |= EV_CLEAR;
6062 knlist_add(&fs_knlist, kn, 0);
6067 filt_fsdetach(struct knote *kn)
6070 knlist_remove(&fs_knlist, kn, 0);
6074 filt_fsevent(struct knote *kn, long hint)
6077 kn->kn_fflags |= kn->kn_sfflags & hint;
6079 return (kn->kn_fflags != 0);
6083 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
6089 error = SYSCTL_IN(req, &vc, sizeof(vc));
6092 if (vc.vc_vers != VFS_CTL_VERS1)
6094 mp = vfs_getvfs(&vc.vc_fsid);
6097 /* ensure that a specific sysctl goes to the right filesystem. */
6098 if (strcmp(vc.vc_fstypename, "*") != 0 &&
6099 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
6103 VCTLTOREQ(&vc, req);
6104 error = VFS_SYSCTL(mp, vc.vc_op, req);
6109 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
6110 NULL, 0, sysctl_vfs_ctl, "",
6114 * Function to initialize a va_filerev field sensibly.
6115 * XXX: Wouldn't a random number make a lot more sense ??
6118 init_va_filerev(void)
6123 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
6126 static int filt_vfsread(struct knote *kn, long hint);
6127 static int filt_vfswrite(struct knote *kn, long hint);
6128 static int filt_vfsvnode(struct knote *kn, long hint);
6129 static void filt_vfsdetach(struct knote *kn);
6130 static struct filterops vfsread_filtops = {
6132 .f_detach = filt_vfsdetach,
6133 .f_event = filt_vfsread
6135 static struct filterops vfswrite_filtops = {
6137 .f_detach = filt_vfsdetach,
6138 .f_event = filt_vfswrite
6140 static struct filterops vfsvnode_filtops = {
6142 .f_detach = filt_vfsdetach,
6143 .f_event = filt_vfsvnode
6147 vfs_knllock(void *arg)
6149 struct vnode *vp = arg;
6151 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
6155 vfs_knlunlock(void *arg)
6157 struct vnode *vp = arg;
6163 vfs_knl_assert_lock(void *arg, int what)
6165 #ifdef DEBUG_VFS_LOCKS
6166 struct vnode *vp = arg;
6168 if (what == LA_LOCKED)
6169 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
6171 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
6176 vfs_kqfilter(struct vop_kqfilter_args *ap)
6178 struct vnode *vp = ap->a_vp;
6179 struct knote *kn = ap->a_kn;
6182 switch (kn->kn_filter) {
6184 kn->kn_fop = &vfsread_filtops;
6187 kn->kn_fop = &vfswrite_filtops;
6190 kn->kn_fop = &vfsvnode_filtops;
6196 kn->kn_hook = (caddr_t)vp;
6199 if (vp->v_pollinfo == NULL)
6201 knl = &vp->v_pollinfo->vpi_selinfo.si_note;
6203 knlist_add(knl, kn, 0);
6209 * Detach knote from vnode
6212 filt_vfsdetach(struct knote *kn)
6214 struct vnode *vp = (struct vnode *)kn->kn_hook;
6216 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
6217 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
6223 filt_vfsread(struct knote *kn, long hint)
6225 struct vnode *vp = (struct vnode *)kn->kn_hook;
6230 * filesystem is gone, so set the EOF flag and schedule
6231 * the knote for deletion.
6233 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6235 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6240 if (VOP_GETATTR(vp, &va, curthread->td_ucred))
6244 kn->kn_data = va.va_size - kn->kn_fp->f_offset;
6245 res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
6252 filt_vfswrite(struct knote *kn, long hint)
6254 struct vnode *vp = (struct vnode *)kn->kn_hook;
6259 * filesystem is gone, so set the EOF flag and schedule
6260 * the knote for deletion.
6262 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
6263 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6271 filt_vfsvnode(struct knote *kn, long hint)
6273 struct vnode *vp = (struct vnode *)kn->kn_hook;
6277 if (kn->kn_sfflags & hint)
6278 kn->kn_fflags |= hint;
6279 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6280 kn->kn_flags |= EV_EOF;
6284 res = (kn->kn_fflags != 0);
6290 * Returns whether the directory is empty or not.
6291 * If it is empty, the return value is 0; otherwise
6292 * the return value is an error value (which may
6296 vfs_emptydir(struct vnode *vp)
6300 struct dirent *dirent, *dp, *endp;
6306 ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
6308 dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
6309 iov.iov_base = dirent;
6310 iov.iov_len = sizeof(struct dirent);
6315 uio.uio_resid = sizeof(struct dirent);
6316 uio.uio_segflg = UIO_SYSSPACE;
6317 uio.uio_rw = UIO_READ;
6318 uio.uio_td = curthread;
6320 while (eof == 0 && error == 0) {
6321 error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
6325 endp = (void *)((uint8_t *)dirent +
6326 sizeof(struct dirent) - uio.uio_resid);
6327 for (dp = dirent; dp < endp;
6328 dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
6329 if (dp->d_type == DT_WHT)
6331 if (dp->d_namlen == 0)
6333 if (dp->d_type != DT_DIR &&
6334 dp->d_type != DT_UNKNOWN) {
6338 if (dp->d_namlen > 2) {
6342 if (dp->d_namlen == 1 &&
6343 dp->d_name[0] != '.') {
6347 if (dp->d_namlen == 2 &&
6348 dp->d_name[1] != '.') {
6352 uio.uio_resid = sizeof(struct dirent);
6355 free(dirent, M_TEMP);
6360 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6364 if (dp->d_reclen > ap->a_uio->uio_resid)
6365 return (ENAMETOOLONG);
6366 error = uiomove(dp, dp->d_reclen, ap->a_uio);
6368 if (ap->a_ncookies != NULL) {
6369 if (ap->a_cookies != NULL)
6370 free(ap->a_cookies, M_TEMP);
6371 ap->a_cookies = NULL;
6372 *ap->a_ncookies = 0;
6376 if (ap->a_ncookies == NULL)
6379 KASSERT(ap->a_cookies,
6380 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6382 *ap->a_cookies = realloc(*ap->a_cookies,
6383 (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
6384 (*ap->a_cookies)[*ap->a_ncookies] = off;
6385 *ap->a_ncookies += 1;
6390 * The purpose of this routine is to remove granularity from accmode_t,
6391 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6392 * VADMIN and VAPPEND.
6394 * If it returns 0, the caller is supposed to continue with the usual
6395 * access checks using 'accmode' as modified by this routine. If it
6396 * returns nonzero value, the caller is supposed to return that value
6399 * Note that after this routine runs, accmode may be zero.
6402 vfs_unixify_accmode(accmode_t *accmode)
6405 * There is no way to specify explicit "deny" rule using
6406 * file mode or POSIX.1e ACLs.
6408 if (*accmode & VEXPLICIT_DENY) {
6414 * None of these can be translated into usual access bits.
6415 * Also, the common case for NFSv4 ACLs is to not contain
6416 * either of these bits. Caller should check for VWRITE
6417 * on the containing directory instead.
6419 if (*accmode & (VDELETE_CHILD | VDELETE))
6422 if (*accmode & VADMIN_PERMS) {
6423 *accmode &= ~VADMIN_PERMS;
6428 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6429 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6431 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6437 * Clear out a doomed vnode (if any) and replace it with a new one as long
6438 * as the fs is not being unmounted. Return the root vnode to the caller.
6440 static int __noinline
6441 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6447 if (mp->mnt_rootvnode != NULL) {
6449 vp = mp->mnt_rootvnode;
6451 if (!VN_IS_DOOMED(vp)) {
6454 error = vn_lock(vp, flags);
6463 * Clear the old one.
6465 mp->mnt_rootvnode = NULL;
6469 vfs_op_barrier_wait(mp);
6473 error = VFS_CACHEDROOT(mp, flags, vpp);
6476 if (mp->mnt_vfs_ops == 0) {
6478 if (mp->mnt_vfs_ops != 0) {
6482 if (mp->mnt_rootvnode == NULL) {
6484 mp->mnt_rootvnode = *vpp;
6486 if (mp->mnt_rootvnode != *vpp) {
6487 if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6488 panic("%s: mismatch between vnode returned "
6489 " by VFS_CACHEDROOT and the one cached "
6491 __func__, *vpp, mp->mnt_rootvnode);
6501 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6503 struct mount_pcpu *mpcpu;
6507 if (!vfs_op_thread_enter(mp, mpcpu))
6508 return (vfs_cache_root_fallback(mp, flags, vpp));
6509 vp = atomic_load_ptr(&mp->mnt_rootvnode);
6510 if (vp == NULL || VN_IS_DOOMED(vp)) {
6511 vfs_op_thread_exit(mp, mpcpu);
6512 return (vfs_cache_root_fallback(mp, flags, vpp));
6515 vfs_op_thread_exit(mp, mpcpu);
6516 error = vn_lock(vp, flags);
6519 return (vfs_cache_root_fallback(mp, flags, vpp));
6526 vfs_cache_root_clear(struct mount *mp)
6531 * ops > 0 guarantees there is nobody who can see this vnode
6533 MPASS(mp->mnt_vfs_ops > 0);
6534 vp = mp->mnt_rootvnode;
6536 vn_seqc_write_begin(vp);
6537 mp->mnt_rootvnode = NULL;
6542 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6545 MPASS(mp->mnt_vfs_ops > 0);
6547 mp->mnt_rootvnode = vp;
6551 * These are helper functions for filesystems to traverse all
6552 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6554 * This interface replaces MNT_VNODE_FOREACH.
6558 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6563 kern_yield(PRI_USER);
6565 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6566 for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6567 vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6568 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6569 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6572 if (VN_IS_DOOMED(vp)) {
6579 __mnt_vnode_markerfree_all(mvp, mp);
6580 /* MNT_IUNLOCK(mp); -- done in above function */
6581 mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6584 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6585 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6591 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6595 *mvp = vn_alloc_marker(mp);
6599 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6600 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6601 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6604 if (VN_IS_DOOMED(vp)) {
6613 vn_free_marker(*mvp);
6617 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6623 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6631 mtx_assert(MNT_MTX(mp), MA_OWNED);
6633 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6634 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6637 vn_free_marker(*mvp);
6642 * These are helper functions for filesystems to traverse their
6643 * lazy vnodes. See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6646 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6649 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6654 vn_free_marker(*mvp);
6659 * Relock the mp mount vnode list lock with the vp vnode interlock in the
6660 * conventional lock order during mnt_vnode_next_lazy iteration.
6662 * On entry, the mount vnode list lock is held and the vnode interlock is not.
6663 * The list lock is dropped and reacquired. On success, both locks are held.
6664 * On failure, the mount vnode list lock is held but the vnode interlock is
6665 * not, and the procedure may have yielded.
6668 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6672 VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6673 TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6674 ("%s: bad marker", __func__));
6675 VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6676 ("%s: inappropriate vnode", __func__));
6677 ASSERT_VI_UNLOCKED(vp, __func__);
6678 mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6680 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6681 TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6684 * Note we may be racing against vdrop which transitioned the hold
6685 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6686 * if we are the only user after we get the interlock we will just
6690 mtx_unlock(&mp->mnt_listmtx);
6692 if (VN_IS_DOOMED(vp)) {
6693 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6696 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6698 * There is nothing to do if we are the last user.
6700 if (!refcount_release_if_not_last(&vp->v_holdcnt))
6702 mtx_lock(&mp->mnt_listmtx);
6707 mtx_lock(&mp->mnt_listmtx);
6711 static struct vnode *
6712 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6717 mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6718 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6720 vp = TAILQ_NEXT(*mvp, v_lazylist);
6721 while (vp != NULL) {
6722 if (vp->v_type == VMARKER) {
6723 vp = TAILQ_NEXT(vp, v_lazylist);
6727 * See if we want to process the vnode. Note we may encounter a
6728 * long string of vnodes we don't care about and hog the list
6729 * as a result. Check for it and requeue the marker.
6731 VNPASS(!VN_IS_DOOMED(vp), vp);
6732 if (!cb(vp, cbarg)) {
6733 if (!should_yield()) {
6734 vp = TAILQ_NEXT(vp, v_lazylist);
6737 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6739 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6741 mtx_unlock(&mp->mnt_listmtx);
6742 kern_yield(PRI_USER);
6743 mtx_lock(&mp->mnt_listmtx);
6747 * Try-lock because this is the wrong lock order.
6749 if (!VI_TRYLOCK(vp) &&
6750 !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6752 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6753 KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6754 ("alien vnode on the lazy list %p %p", vp, mp));
6755 VNPASS(vp->v_mount == mp, vp);
6756 VNPASS(!VN_IS_DOOMED(vp), vp);
6759 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6761 /* Check if we are done */
6763 mtx_unlock(&mp->mnt_listmtx);
6764 mnt_vnode_markerfree_lazy(mvp, mp);
6767 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6768 mtx_unlock(&mp->mnt_listmtx);
6769 ASSERT_VI_LOCKED(vp, "lazy iter");
6774 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6779 kern_yield(PRI_USER);
6780 mtx_lock(&mp->mnt_listmtx);
6781 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6785 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6790 if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6793 *mvp = vn_alloc_marker(mp);
6798 mtx_lock(&mp->mnt_listmtx);
6799 vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6801 mtx_unlock(&mp->mnt_listmtx);
6802 mnt_vnode_markerfree_lazy(mvp, mp);
6805 TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6806 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6810 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6816 mtx_lock(&mp->mnt_listmtx);
6817 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6818 mtx_unlock(&mp->mnt_listmtx);
6819 mnt_vnode_markerfree_lazy(mvp, mp);
6823 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6826 if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6827 cnp->cn_flags &= ~NOEXECCHECK;
6831 return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, cnp->cn_thread));
6835 * Do not use this variant unless you have means other than the hold count
6836 * to prevent the vnode from getting freed.
6839 vn_seqc_write_begin_unheld_locked(struct vnode *vp)
6842 ASSERT_VI_LOCKED(vp, __func__);
6843 VNPASS(vp->v_seqc_users >= 0, vp);
6845 if (vp->v_seqc_users == 1)
6846 seqc_sleepable_write_begin(&vp->v_seqc);
6850 vn_seqc_write_begin_locked(struct vnode *vp)
6853 ASSERT_VI_LOCKED(vp, __func__);
6854 VNPASS(vp->v_holdcnt > 0, vp);
6855 vn_seqc_write_begin_unheld_locked(vp);
6859 vn_seqc_write_begin(struct vnode *vp)
6863 vn_seqc_write_begin_locked(vp);
6868 vn_seqc_write_begin_unheld(struct vnode *vp)
6872 vn_seqc_write_begin_unheld_locked(vp);
6877 vn_seqc_write_end_locked(struct vnode *vp)
6880 ASSERT_VI_LOCKED(vp, __func__);
6881 VNPASS(vp->v_seqc_users > 0, vp);
6883 if (vp->v_seqc_users == 0)
6884 seqc_sleepable_write_end(&vp->v_seqc);
6888 vn_seqc_write_end(struct vnode *vp)
6892 vn_seqc_write_end_locked(vp);
6897 * Special case handling for allocating and freeing vnodes.
6899 * The counter remains unchanged on free so that a doomed vnode will
6900 * keep testing as in modify as long as it is accessible with SMR.
6903 vn_seqc_init(struct vnode *vp)
6907 vp->v_seqc_users = 0;
6911 vn_seqc_write_end_free(struct vnode *vp)
6914 VNPASS(seqc_in_modify(vp->v_seqc), vp);
6915 VNPASS(vp->v_seqc_users == 1, vp);
6919 vn_irflag_set_locked(struct vnode *vp, short toset)
6923 ASSERT_VI_LOCKED(vp, __func__);
6924 flags = vn_irflag_read(vp);
6925 VNASSERT((flags & toset) == 0, vp,
6926 ("%s: some of the passed flags already set (have %d, passed %d)\n",
6927 __func__, flags, toset));
6928 atomic_store_short(&vp->v_irflag, flags | toset);
6932 vn_irflag_set(struct vnode *vp, short toset)
6936 vn_irflag_set_locked(vp, toset);
6941 vn_irflag_set_cond_locked(struct vnode *vp, short toset)
6945 ASSERT_VI_LOCKED(vp, __func__);
6946 flags = vn_irflag_read(vp);
6947 atomic_store_short(&vp->v_irflag, flags | toset);
6951 vn_irflag_set_cond(struct vnode *vp, short toset)
6955 vn_irflag_set_cond_locked(vp, toset);
6960 vn_irflag_unset_locked(struct vnode *vp, short tounset)
6964 ASSERT_VI_LOCKED(vp, __func__);
6965 flags = vn_irflag_read(vp);
6966 VNASSERT((flags & tounset) == tounset, vp,
6967 ("%s: some of the passed flags not set (have %d, passed %d)\n",
6968 __func__, flags, tounset));
6969 atomic_store_short(&vp->v_irflag, flags & ~tounset);
6973 vn_irflag_unset(struct vnode *vp, short tounset)
6977 vn_irflag_unset_locked(vp, tounset);