1/*-
2 * SPDX-License-Identifier: BSD-3-Clause
3 *
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.
11 *
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
14 * are met:
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.
23 *
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
34 * SUCH DAMAGE.
35 *
36 *	@(#)vfs_subr.c	8.31 (Berkeley) 5/26/95
37 */
38
39/*
40 * External virtual filesystem routines
41 */
42
43#include <sys/cdefs.h>
44__FBSDID("$FreeBSD$");
45
46#include "opt_ddb.h"
47#include "opt_watchdog.h"
48
49#include <sys/param.h>
50#include <sys/systm.h>
51#include <sys/bio.h>
52#include <sys/buf.h>
53#include <sys/capsicum.h>
54#include <sys/condvar.h>
55#include <sys/conf.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>
61#include <sys/file.h>
62#include <sys/fcntl.h>
63#include <sys/jail.h>
64#include <sys/kdb.h>
65#include <sys/kernel.h>
66#include <sys/kthread.h>
67#include <sys/ktr.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>
73#include <sys/priv.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>
79#include <sys/smp.h>
80#include <sys/stat.h>
81#include <sys/sysctl.h>
82#include <sys/syslog.h>
83#include <sys/vmmeter.h>
84#include <sys/vnode.h>
85#include <sys/watchdog.h>
86
87#include <machine/stdarg.h>
88
89#include <security/mac/mac_framework.h>
90
91#include <vm/vm.h>
92#include <vm/vm_object.h>
93#include <vm/vm_extern.h>
94#include <vm/pmap.h>
95#include <vm/vm_map.h>
96#include <vm/vm_page.h>
97#include <vm/vm_kern.h>
98#include <vm/uma.h>
99
100#ifdef DDB
101#include <ddb/ddb.h>
102#endif
103
104static void	delmntque(struct vnode *vp);
105static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
106		    int slpflag, int slptimeo);
107static void	syncer_shutdown(void *arg, int howto);
108static int	vtryrecycle(struct vnode *vp);
109static void	v_init_counters(struct vnode *);
110static void	v_incr_devcount(struct vnode *);
111static void	v_decr_devcount(struct vnode *);
112static void	vgonel(struct vnode *);
113static void	vfs_knllock(void *arg);
114static void	vfs_knlunlock(void *arg);
115static void	vfs_knl_assert_locked(void *arg);
116static void	vfs_knl_assert_unlocked(void *arg);
117static void	destroy_vpollinfo(struct vpollinfo *vi);
118static int	v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
119		    daddr_t startlbn, daddr_t endlbn);
120static void	vnlru_recalc(void);
121
122/*
123 * These fences are intended for cases where some synchronization is
124 * needed between access of v_iflags and lockless vnode refcount (v_holdcnt
125 * and v_usecount) updates.  Access to v_iflags is generally synchronized
126 * by the interlock, but we have some internal assertions that check vnode
127 * flags without acquiring the lock.  Thus, these fences are INVARIANTS-only
128 * for now.
129 */
130#ifdef INVARIANTS
131#define	VNODE_REFCOUNT_FENCE_ACQ()	atomic_thread_fence_acq()
132#define	VNODE_REFCOUNT_FENCE_REL()	atomic_thread_fence_rel()
133#else
134#define	VNODE_REFCOUNT_FENCE_ACQ()
135#define	VNODE_REFCOUNT_FENCE_REL()
136#endif
137
138/*
139 * Number of vnodes in existence.  Increased whenever getnewvnode()
140 * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
141 */
142static u_long __exclusive_cache_line numvnodes;
143
144SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
145    "Number of vnodes in existence");
146
147static counter_u64_t vnodes_created;
148SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
149    "Number of vnodes created by getnewvnode");
150
151/*
152 * Conversion tables for conversion from vnode types to inode formats
153 * and back.
154 */
155enum vtype iftovt_tab[16] = {
156	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
157	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
158};
159int vttoif_tab[10] = {
160	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
161	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
162};
163
164/*
165 * List of allocates vnodes in the system.
166 */
167static TAILQ_HEAD(freelst, vnode) vnode_list;
168static struct vnode *vnode_list_free_marker;
169static struct vnode *vnode_list_reclaim_marker;
170
171/*
172 * "Free" vnode target.  Free vnodes are rarely completely free, but are
173 * just ones that are cheap to recycle.  Usually they are for files which
174 * have been stat'd but not read; these usually have inode and namecache
175 * data attached to them.  This target is the preferred minimum size of a
176 * sub-cache consisting mostly of such files. The system balances the size
177 * of this sub-cache with its complement to try to prevent either from
178 * thrashing while the other is relatively inactive.  The targets express
179 * a preference for the best balance.
180 *
181 * "Above" this target there are 2 further targets (watermarks) related
182 * to recyling of free vnodes.  In the best-operating case, the cache is
183 * exactly full, the free list has size between vlowat and vhiwat above the
184 * free target, and recycling from it and normal use maintains this state.
185 * Sometimes the free list is below vlowat or even empty, but this state
186 * is even better for immediate use provided the cache is not full.
187 * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
188 * ones) to reach one of these states.  The watermarks are currently hard-
189 * coded as 4% and 9% of the available space higher.  These and the default
190 * of 25% for wantfreevnodes are too large if the memory size is large.
191 * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
192 * whenever vnlru_proc() becomes active.
193 */
194static long wantfreevnodes;
195static long __exclusive_cache_line freevnodes;
196SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
197    &freevnodes, 0, "Number of \"free\" vnodes");
198static long freevnodes_old;
199
200static counter_u64_t recycles_count;
201SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
202    "Number of vnodes recycled to meet vnode cache targets");
203
204static counter_u64_t recycles_free_count;
205SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count,
206    "Number of free vnodes recycled to meet vnode cache targets");
207
208/*
209 * Various variables used for debugging the new implementation of
210 * reassignbuf().
211 * XXX these are probably of (very) limited utility now.
212 */
213static int reassignbufcalls;
214SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW | CTLFLAG_STATS,
215    &reassignbufcalls, 0, "Number of calls to reassignbuf");
216
217static counter_u64_t deferred_inact;
218SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact,
219    "Number of times inactive processing was deferred");
220
221/* To keep more than one thread at a time from running vfs_getnewfsid */
222static struct mtx mntid_mtx;
223
224/*
225 * Lock for any access to the following:
226 *	vnode_list
227 *	numvnodes
228 *	freevnodes
229 */
230static struct mtx __exclusive_cache_line vnode_list_mtx;
231
232/* Publicly exported FS */
233struct nfs_public nfs_pub;
234
235static uma_zone_t buf_trie_zone;
236
237/* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
238static uma_zone_t vnode_zone;
239static uma_zone_t vnodepoll_zone;
240
241/*
242 * The workitem queue.
243 *
244 * It is useful to delay writes of file data and filesystem metadata
245 * for tens of seconds so that quickly created and deleted files need
246 * not waste disk bandwidth being created and removed. To realize this,
247 * we append vnodes to a "workitem" queue. When running with a soft
248 * updates implementation, most pending metadata dependencies should
249 * not wait for more than a few seconds. Thus, mounted on block devices
250 * are delayed only about a half the time that file data is delayed.
251 * Similarly, directory updates are more critical, so are only delayed
252 * about a third the time that file data is delayed. Thus, there are
253 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
254 * one each second (driven off the filesystem syncer process). The
255 * syncer_delayno variable indicates the next queue that is to be processed.
256 * Items that need to be processed soon are placed in this queue:
257 *
258 *	syncer_workitem_pending[syncer_delayno]
259 *
260 * A delay of fifteen seconds is done by placing the request fifteen
261 * entries later in the queue:
262 *
263 *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
264 *
265 */
266static int syncer_delayno;
267static long syncer_mask;
268LIST_HEAD(synclist, bufobj);
269static struct synclist *syncer_workitem_pending;
270/*
271 * The sync_mtx protects:
272 *	bo->bo_synclist
273 *	sync_vnode_count
274 *	syncer_delayno
275 *	syncer_state
276 *	syncer_workitem_pending
277 *	syncer_worklist_len
278 *	rushjob
279 */
280static struct mtx sync_mtx;
281static struct cv sync_wakeup;
282
283#define SYNCER_MAXDELAY		32
284static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
285static int syncdelay = 30;		/* max time to delay syncing data */
286static int filedelay = 30;		/* time to delay syncing files */
287SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
288    "Time to delay syncing files (in seconds)");
289static int dirdelay = 29;		/* time to delay syncing directories */
290SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
291    "Time to delay syncing directories (in seconds)");
292static int metadelay = 28;		/* time to delay syncing metadata */
293SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
294    "Time to delay syncing metadata (in seconds)");
295static int rushjob;		/* number of slots to run ASAP */
296static int stat_rush_requests;	/* number of times I/O speeded up */
297SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
298    "Number of times I/O speeded up (rush requests)");
299
300#define	VDBATCH_SIZE 8
301struct vdbatch {
302	u_int index;
303	long freevnodes;
304	struct mtx lock;
305	struct vnode *tab[VDBATCH_SIZE];
306};
307DPCPU_DEFINE_STATIC(struct vdbatch, vd);
308
309static void	vdbatch_dequeue(struct vnode *vp);
310
311/*
312 * When shutting down the syncer, run it at four times normal speed.
313 */
314#define SYNCER_SHUTDOWN_SPEEDUP		4
315static int sync_vnode_count;
316static int syncer_worklist_len;
317static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
318    syncer_state;
319
320/* Target for maximum number of vnodes. */
321u_long desiredvnodes;
322static u_long gapvnodes;		/* gap between wanted and desired */
323static u_long vhiwat;		/* enough extras after expansion */
324static u_long vlowat;		/* minimal extras before expansion */
325static u_long vstir;		/* nonzero to stir non-free vnodes */
326static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
327
328static u_long vnlru_read_freevnodes(void);
329
330/*
331 * Note that no attempt is made to sanitize these parameters.
332 */
333static int
334sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
335{
336	u_long val;
337	int error;
338
339	val = desiredvnodes;
340	error = sysctl_handle_long(oidp, &val, 0, req);
341	if (error != 0 || req->newptr == NULL)
342		return (error);
343
344	if (val == desiredvnodes)
345		return (0);
346	mtx_lock(&vnode_list_mtx);
347	desiredvnodes = val;
348	wantfreevnodes = desiredvnodes / 4;
349	vnlru_recalc();
350	mtx_unlock(&vnode_list_mtx);
351	/*
352	 * XXX There is no protection against multiple threads changing
353	 * desiredvnodes at the same time. Locking above only helps vnlru and
354	 * getnewvnode.
355	 */
356	vfs_hash_changesize(desiredvnodes);
357	cache_changesize(desiredvnodes);
358	return (0);
359}
360
361SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
362    CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
363    "LU", "Target for maximum number of vnodes");
364
365static int
366sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
367{
368	u_long val;
369	int error;
370
371	val = wantfreevnodes;
372	error = sysctl_handle_long(oidp, &val, 0, req);
373	if (error != 0 || req->newptr == NULL)
374		return (error);
375
376	if (val == wantfreevnodes)
377		return (0);
378	mtx_lock(&vnode_list_mtx);
379	wantfreevnodes = val;
380	vnlru_recalc();
381	mtx_unlock(&vnode_list_mtx);
382	return (0);
383}
384
385SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
386    CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
387    "LU", "Target for minimum number of \"free\" vnodes");
388
389SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
390    &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
391static int vnlru_nowhere;
392SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
393    &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
394
395static int
396sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
397{
398	struct vnode *vp;
399	struct nameidata nd;
400	char *buf;
401	unsigned long ndflags;
402	int error;
403
404	if (req->newptr == NULL)
405		return (EINVAL);
406	if (req->newlen >= PATH_MAX)
407		return (E2BIG);
408
409	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
410	error = SYSCTL_IN(req, buf, req->newlen);
411	if (error != 0)
412		goto out;
413
414	buf[req->newlen] = '\0';
415
416	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | NOCACHE | SAVENAME;
417	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread);
418	if ((error = namei(&nd)) != 0)
419		goto out;
420	vp = nd.ni_vp;
421
422	if (VN_IS_DOOMED(vp)) {
423		/*
424		 * This vnode is being recycled.  Return != 0 to let the caller
425		 * know that the sysctl had no effect.  Return EAGAIN because a
426		 * subsequent call will likely succeed (since namei will create
427		 * a new vnode if necessary)
428		 */
429		error = EAGAIN;
430		goto putvnode;
431	}
432
433	counter_u64_add(recycles_count, 1);
434	vgone(vp);
435putvnode:
436	NDFREE(&nd, 0);
437out:
438	free(buf, M_TEMP);
439	return (error);
440}
441
442static int
443sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
444{
445	struct thread *td = curthread;
446	struct vnode *vp;
447	struct file *fp;
448	int error;
449	int fd;
450
451	if (req->newptr == NULL)
452		return (EBADF);
453
454        error = sysctl_handle_int(oidp, &fd, 0, req);
455        if (error != 0)
456                return (error);
457	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
458	if (error != 0)
459		return (error);
460	vp = fp->f_vnode;
461
462	error = vn_lock(vp, LK_EXCLUSIVE);
463	if (error != 0)
464		goto drop;
465
466	counter_u64_add(recycles_count, 1);
467	vgone(vp);
468	VOP_UNLOCK(vp);
469drop:
470	fdrop(fp, td);
471	return (error);
472}
473
474SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
475    CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
476    sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
477SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
478    CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
479    sysctl_ftry_reclaim_vnode, "I",
480    "Try to reclaim a vnode by its file descriptor");
481
482/* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
483static int vnsz2log;
484
485/*
486 * Support for the bufobj clean & dirty pctrie.
487 */
488static void *
489buf_trie_alloc(struct pctrie *ptree)
490{
491
492	return uma_zalloc(buf_trie_zone, M_NOWAIT);
493}
494
495static void
496buf_trie_free(struct pctrie *ptree, void *node)
497{
498
499	uma_zfree(buf_trie_zone, node);
500}
501PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
502
503/*
504 * Initialize the vnode management data structures.
505 *
506 * Reevaluate the following cap on the number of vnodes after the physical
507 * memory size exceeds 512GB.  In the limit, as the physical memory size
508 * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
509 */
510#ifndef	MAXVNODES_MAX
511#define	MAXVNODES_MAX	(512UL * 1024 * 1024 / 64)	/* 8M */
512#endif
513
514static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
515
516static struct vnode *
517vn_alloc_marker(struct mount *mp)
518{
519	struct vnode *vp;
520
521	vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
522	vp->v_type = VMARKER;
523	vp->v_mount = mp;
524
525	return (vp);
526}
527
528static void
529vn_free_marker(struct vnode *vp)
530{
531
532	MPASS(vp->v_type == VMARKER);
533	free(vp, M_VNODE_MARKER);
534}
535
536/*
537 * Initialize a vnode as it first enters the zone.
538 */
539static int
540vnode_init(void *mem, int size, int flags)
541{
542	struct vnode *vp;
543
544	vp = mem;
545	bzero(vp, size);
546	/*
547	 * Setup locks.
548	 */
549	vp->v_vnlock = &vp->v_lock;
550	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
551	/*
552	 * By default, don't allow shared locks unless filesystems opt-in.
553	 */
554	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
555	    LK_NOSHARE | LK_IS_VNODE);
556	/*
557	 * Initialize bufobj.
558	 */
559	bufobj_init(&vp->v_bufobj, vp);
560	/*
561	 * Initialize namecache.
562	 */
563	LIST_INIT(&vp->v_cache_src);
564	TAILQ_INIT(&vp->v_cache_dst);
565	/*
566	 * Initialize rangelocks.
567	 */
568	rangelock_init(&vp->v_rl);
569
570	vp->v_dbatchcpu = NOCPU;
571
572	mtx_lock(&vnode_list_mtx);
573	TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
574	mtx_unlock(&vnode_list_mtx);
575	return (0);
576}
577
578/*
579 * Free a vnode when it is cleared from the zone.
580 */
581static void
582vnode_fini(void *mem, int size)
583{
584	struct vnode *vp;
585	struct bufobj *bo;
586
587	vp = mem;
588	vdbatch_dequeue(vp);
589	mtx_lock(&vnode_list_mtx);
590	TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
591	mtx_unlock(&vnode_list_mtx);
592	rangelock_destroy(&vp->v_rl);
593	lockdestroy(vp->v_vnlock);
594	mtx_destroy(&vp->v_interlock);
595	bo = &vp->v_bufobj;
596	rw_destroy(BO_LOCKPTR(bo));
597}
598
599/*
600 * Provide the size of NFS nclnode and NFS fh for calculation of the
601 * vnode memory consumption.  The size is specified directly to
602 * eliminate dependency on NFS-private header.
603 *
604 * Other filesystems may use bigger or smaller (like UFS and ZFS)
605 * private inode data, but the NFS-based estimation is ample enough.
606 * Still, we care about differences in the size between 64- and 32-bit
607 * platforms.
608 *
609 * Namecache structure size is heuristically
610 * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
611 */
612#ifdef _LP64
613#define	NFS_NCLNODE_SZ	(528 + 64)
614#define	NC_SZ		148
615#else
616#define	NFS_NCLNODE_SZ	(360 + 32)
617#define	NC_SZ		92
618#endif
619
620static void
621vntblinit(void *dummy __unused)
622{
623	struct vdbatch *vd;
624	int cpu, physvnodes, virtvnodes;
625	u_int i;
626
627	/*
628	 * Desiredvnodes is a function of the physical memory size and the
629	 * kernel's heap size.  Generally speaking, it scales with the
630	 * physical memory size.  The ratio of desiredvnodes to the physical
631	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
632	 * Thereafter, the
633	 * marginal ratio of desiredvnodes to the physical memory size is
634	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
635	 * size.  The memory required by desiredvnodes vnodes and vm objects
636	 * must not exceed 1/10th of the kernel's heap size.
637	 */
638	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
639	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
640	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
641	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
642	desiredvnodes = min(physvnodes, virtvnodes);
643	if (desiredvnodes > MAXVNODES_MAX) {
644		if (bootverbose)
645			printf("Reducing kern.maxvnodes %lu -> %lu\n",
646			    desiredvnodes, MAXVNODES_MAX);
647		desiredvnodes = MAXVNODES_MAX;
648	}
649	wantfreevnodes = desiredvnodes / 4;
650	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
651	TAILQ_INIT(&vnode_list);
652	mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
653	/*
654	 * The lock is taken to appease WITNESS.
655	 */
656	mtx_lock(&vnode_list_mtx);
657	vnlru_recalc();
658	mtx_unlock(&vnode_list_mtx);
659	vnode_list_free_marker = vn_alloc_marker(NULL);
660	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
661	vnode_list_reclaim_marker = vn_alloc_marker(NULL);
662	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
663	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
664	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
665	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
666	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
667	/*
668	 * Preallocate enough nodes to support one-per buf so that
669	 * we can not fail an insert.  reassignbuf() callers can not
670	 * tolerate the insertion failure.
671	 */
672	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
673	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
674	    UMA_ZONE_NOFREE);
675	uma_prealloc(buf_trie_zone, nbuf);
676
677	vnodes_created = counter_u64_alloc(M_WAITOK);
678	recycles_count = counter_u64_alloc(M_WAITOK);
679	recycles_free_count = counter_u64_alloc(M_WAITOK);
680	deferred_inact = counter_u64_alloc(M_WAITOK);
681
682	/*
683	 * Initialize the filesystem syncer.
684	 */
685	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
686	    &syncer_mask);
687	syncer_maxdelay = syncer_mask + 1;
688	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
689	cv_init(&sync_wakeup, "syncer");
690	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
691		vnsz2log++;
692	vnsz2log--;
693
694	CPU_FOREACH(cpu) {
695		vd = DPCPU_ID_PTR((cpu), vd);
696		bzero(vd, sizeof(*vd));
697		mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
698	}
699}
700SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
701
702/*
703 * Mark a mount point as busy. Used to synchronize access and to delay
704 * unmounting. Eventually, mountlist_mtx is not released on failure.
705 *
706 * vfs_busy() is a custom lock, it can block the caller.
707 * vfs_busy() only sleeps if the unmount is active on the mount point.
708 * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
709 * vnode belonging to mp.
710 *
711 * Lookup uses vfs_busy() to traverse mount points.
712 * root fs			var fs
713 * / vnode lock		A	/ vnode lock (/var)		D
714 * /var vnode lock	B	/log vnode lock(/var/log)	E
715 * vfs_busy lock	C	vfs_busy lock			F
716 *
717 * Within each file system, the lock order is C->A->B and F->D->E.
718 *
719 * When traversing across mounts, the system follows that lock order:
720 *
721 *        C->A->B
722 *              |
723 *              +->F->D->E
724 *
725 * The lookup() process for namei("/var") illustrates the process:
726 *  VOP_LOOKUP() obtains B while A is held
727 *  vfs_busy() obtains a shared lock on F while A and B are held
728 *  vput() releases lock on B
729 *  vput() releases lock on A
730 *  VFS_ROOT() obtains lock on D while shared lock on F is held
731 *  vfs_unbusy() releases shared lock on F
732 *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
733 *    Attempt to lock A (instead of vp_crossmp) while D is held would
734 *    violate the global order, causing deadlocks.
735 *
736 * dounmount() locks B while F is drained.
737 */
738int
739vfs_busy(struct mount *mp, int flags)
740{
741
742	MPASS((flags & ~MBF_MASK) == 0);
743	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
744
745	if (vfs_op_thread_enter(mp)) {
746		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
747		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
748		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
749		vfs_mp_count_add_pcpu(mp, ref, 1);
750		vfs_mp_count_add_pcpu(mp, lockref, 1);
751		vfs_op_thread_exit(mp);
752		if (flags & MBF_MNTLSTLOCK)
753			mtx_unlock(&mountlist_mtx);
754		return (0);
755	}
756
757	MNT_ILOCK(mp);
758	vfs_assert_mount_counters(mp);
759	MNT_REF(mp);
760	/*
761	 * If mount point is currently being unmounted, sleep until the
762	 * mount point fate is decided.  If thread doing the unmounting fails,
763	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
764	 * that this mount point has survived the unmount attempt and vfs_busy
765	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
766	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
767	 * about to be really destroyed.  vfs_busy needs to release its
768	 * reference on the mount point in this case and return with ENOENT,
769	 * telling the caller that mount mount it tried to busy is no longer
770	 * valid.
771	 */
772	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
773		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
774			MNT_REL(mp);
775			MNT_IUNLOCK(mp);
776			CTR1(KTR_VFS, "%s: failed busying before sleeping",
777			    __func__);
778			return (ENOENT);
779		}
780		if (flags & MBF_MNTLSTLOCK)
781			mtx_unlock(&mountlist_mtx);
782		mp->mnt_kern_flag |= MNTK_MWAIT;
783		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
784		if (flags & MBF_MNTLSTLOCK)
785			mtx_lock(&mountlist_mtx);
786		MNT_ILOCK(mp);
787	}
788	if (flags & MBF_MNTLSTLOCK)
789		mtx_unlock(&mountlist_mtx);
790	mp->mnt_lockref++;
791	MNT_IUNLOCK(mp);
792	return (0);
793}
794
795/*
796 * Free a busy filesystem.
797 */
798void
799vfs_unbusy(struct mount *mp)
800{
801	int c;
802
803	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
804
805	if (vfs_op_thread_enter(mp)) {
806		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
807		vfs_mp_count_sub_pcpu(mp, lockref, 1);
808		vfs_mp_count_sub_pcpu(mp, ref, 1);
809		vfs_op_thread_exit(mp);
810		return;
811	}
812
813	MNT_ILOCK(mp);
814	vfs_assert_mount_counters(mp);
815	MNT_REL(mp);
816	c = --mp->mnt_lockref;
817	if (mp->mnt_vfs_ops == 0) {
818		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
819		MNT_IUNLOCK(mp);
820		return;
821	}
822	if (c < 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);
829	}
830	MNT_IUNLOCK(mp);
831}
832
833/*
834 * Lookup a mount point by filesystem identifier.
835 */
836struct mount *
837vfs_getvfs(fsid_t *fsid)
838{
839	struct mount *mp;
840
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) {
845			vfs_ref(mp);
846			mtx_unlock(&mountlist_mtx);
847			return (mp);
848		}
849	}
850	mtx_unlock(&mountlist_mtx);
851	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
852	return ((struct mount *) 0);
853}
854
855/*
856 * Lookup a mount point by filesystem identifier, busying it before
857 * returning.
858 *
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.
864 */
865struct mount *
866vfs_busyfs(fsid_t *fsid)
867{
868#define	FSID_CACHE_SIZE	256
869	typedef struct mount * volatile vmp_t;
870	static vmp_t cache[FSID_CACHE_SIZE];
871	struct mount *mp;
872	int error;
873	uint32_t hash;
874
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);
878	mp = cache[hash];
879	if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0)
880		goto slow;
881	if (vfs_busy(mp, 0) != 0) {
882		cache[hash] = NULL;
883		goto slow;
884	}
885	if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0)
886		return (mp);
887	else
888	    vfs_unbusy(mp);
889
890slow:
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);
895			if (error) {
896				cache[hash] = NULL;
897				mtx_unlock(&mountlist_mtx);
898				return (NULL);
899			}
900			cache[hash] = mp;
901			return (mp);
902		}
903	}
904	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
905	mtx_unlock(&mountlist_mtx);
906	return ((struct mount *) 0);
907}
908
909/*
910 * Check if a user can access privileged mount options.
911 */
912int
913vfs_suser(struct mount *mp, struct thread *td)
914{
915	int error;
916
917	if (jailed(td->td_ucred)) {
918		/*
919		 * If the jail of the calling thread lacks permission for
920		 * this type of file system, deny immediately.
921		 */
922		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
923			return (EPERM);
924
925		/*
926		 * If the file system was mounted outside the jail of the
927		 * calling thread, deny immediately.
928		 */
929		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
930			return (EPERM);
931	}
932
933	/*
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.
939	 */
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)
943			return (error);
944	}
945	return (0);
946}
947
948/*
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.
954 *
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
958 * different mounts.
959 */
960void
961vfs_getnewfsid(struct mount *mp)
962{
963	static uint16_t mntid_base;
964	struct mount *nmp;
965	fsid_t tfsid;
966	int mtype;
967
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;
973	for (;;) {
974		tfsid.val[0] = makedev(255,
975		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
976		mntid_base++;
977		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
978			break;
979		vfs_rel(nmp);
980	}
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);
984}
985
986/*
987 * Knob to control the precision of file timestamps:
988 *
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.
993 */
994enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
995
996static int timestamp_precision = TSP_USEC;
997SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
998    &timestamp_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))");
1001
1002/*
1003 * Get a current timestamp.
1004 */
1005void
1006vfs_timestamp(struct timespec *tsp)
1007{
1008	struct timeval tv;
1009
1010	switch (timestamp_precision) {
1011	case TSP_SEC:
1012		tsp->tv_sec = time_second;
1013		tsp->tv_nsec = 0;
1014		break;
1015	case TSP_HZ:
1016		getnanotime(tsp);
1017		break;
1018	case TSP_USEC:
1019		microtime(&tv);
1020		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1021		break;
1022	case TSP_NSEC:
1023	default:
1024		nanotime(tsp);
1025		break;
1026	}
1027}
1028
1029/*
1030 * Set vnode attributes to VNOVAL
1031 */
1032void
1033vattr_null(struct vattr *vap)
1034{
1035
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;
1058}
1059
1060/*
1061 * Try to reduce the total number of vnodes.
1062 *
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
1065 *   smaller
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
1071 *
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.
1077 *
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.
1085 *
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
1089 *			 pages.
1090 * @param target	 How many vnodes to reclaim.
1091 * @return		 The number of vnodes that were reclaimed.
1092 */
1093static int
1094vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1095{
1096	struct vnode *vp, *mvp;
1097	struct mount *mp;
1098	struct vm_object *object;
1099	u_long done;
1100	bool retried;
1101
1102	mtx_assert(&vnode_list_mtx, MA_OWNED);
1103
1104	retried = false;
1105	done = 0;
1106
1107	mvp = vnode_list_reclaim_marker;
1108restart:
1109	vp = mvp;
1110	while (done < target) {
1111		vp = TAILQ_NEXT(vp, v_vnodelist);
1112		if (__predict_false(vp == NULL))
1113			break;
1114
1115		if (__predict_false(vp->v_type == VMARKER))
1116			continue;
1117
1118		/*
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.
1123		 */
1124		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1125		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1126			goto next_iter;
1127
1128		if (vp->v_type == VBAD || vp->v_type == VNON)
1129			goto next_iter;
1130
1131		if (!VI_TRYLOCK(vp))
1132			goto next_iter;
1133
1134		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1135		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1136		    VN_IS_DOOMED(vp) || vp->v_type == VNON) {
1137			VI_UNLOCK(vp);
1138			goto next_iter;
1139		}
1140
1141		object = atomic_load_ptr(&vp->v_object);
1142		if (object == NULL || object->resident_page_count > trigger) {
1143			VI_UNLOCK(vp);
1144			goto next_iter;
1145		}
1146
1147		vholdl(vp);
1148		VI_UNLOCK(vp);
1149		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1150		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1151		mtx_unlock(&vnode_list_mtx);
1152
1153		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1154			vdrop(vp);
1155			goto next_iter_unlocked;
1156		}
1157		if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1158			vdrop(vp);
1159			vn_finished_write(mp);
1160			goto next_iter_unlocked;
1161		}
1162
1163		VI_LOCK(vp);
1164		if (vp->v_usecount > 0 ||
1165		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1166		    (vp->v_object != NULL &&
1167		    vp->v_object->resident_page_count > trigger)) {
1168			VOP_UNLOCK(vp);
1169			vdropl(vp);
1170			vn_finished_write(mp);
1171			goto next_iter_unlocked;
1172		}
1173		counter_u64_add(recycles_count, 1);
1174		vgonel(vp);
1175		VOP_UNLOCK(vp);
1176		vdropl(vp);
1177		vn_finished_write(mp);
1178		done++;
1179next_iter_unlocked:
1180		if (should_yield())
1181			kern_yield(PRI_USER);
1182		mtx_lock(&vnode_list_mtx);
1183		goto restart;
1184next_iter:
1185		MPASS(vp->v_type != VMARKER);
1186		if (!should_yield())
1187			continue;
1188		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1189		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1190		mtx_unlock(&vnode_list_mtx);
1191		kern_yield(PRI_USER);
1192		mtx_lock(&vnode_list_mtx);
1193		goto restart;
1194	}
1195	if (done == 0 && !retried) {
1196		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1197		TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1198		retried = true;
1199		goto restart;
1200	}
1201	return (done);
1202}
1203
1204static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1205SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1206    0,
1207    "limit on vnode free requests per call to the vnlru_free routine");
1208
1209/*
1210 * Attempt to reduce the free list by the requested amount.
1211 */
1212static int
1213vnlru_free_locked(int count, struct vfsops *mnt_op)
1214{
1215	struct vnode *vp, *mvp;
1216	struct mount *mp;
1217	int ocount;
1218
1219	mtx_assert(&vnode_list_mtx, MA_OWNED);
1220	if (count > max_vnlru_free)
1221		count = max_vnlru_free;
1222	ocount = count;
1223	mvp = vnode_list_free_marker;
1224restart:
1225	vp = mvp;
1226	while (count > 0) {
1227		vp = TAILQ_NEXT(vp, v_vnodelist);
1228		if (__predict_false(vp == NULL)) {
1229			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1230			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1231			break;
1232		}
1233		if (__predict_false(vp->v_type == VMARKER))
1234			continue;
1235
1236		/*
1237		 * Don't recycle if our vnode is from different type
1238		 * of mount point.  Note that mp is type-safe, the
1239		 * check does not reach unmapped address even if
1240		 * vnode is reclaimed.
1241		 * Don't recycle if we can't get the interlock without
1242		 * blocking.
1243		 */
1244		if (vp->v_holdcnt > 0 || (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1245		    mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) {
1246			continue;
1247		}
1248		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1249		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1250		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1251			VI_UNLOCK(vp);
1252			continue;
1253		}
1254		vholdl(vp);
1255		count--;
1256		mtx_unlock(&vnode_list_mtx);
1257		VI_UNLOCK(vp);
1258		vtryrecycle(vp);
1259		vdrop(vp);
1260		mtx_lock(&vnode_list_mtx);
1261		goto restart;
1262	}
1263	return (ocount - count);
1264}
1265
1266void
1267vnlru_free(int count, struct vfsops *mnt_op)
1268{
1269
1270	mtx_lock(&vnode_list_mtx);
1271	vnlru_free_locked(count, mnt_op);
1272	mtx_unlock(&vnode_list_mtx);
1273}
1274
1275static void
1276vnlru_recalc(void)
1277{
1278
1279	mtx_assert(&vnode_list_mtx, MA_OWNED);
1280	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1281	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1282	vlowat = vhiwat / 2;
1283}
1284
1285/*
1286 * Attempt to recycle vnodes in a context that is always safe to block.
1287 * Calling vlrurecycle() from the bowels of filesystem code has some
1288 * interesting deadlock problems.
1289 */
1290static struct proc *vnlruproc;
1291static int vnlruproc_sig;
1292
1293/*
1294 * The main freevnodes counter is only updated when threads requeue their vnode
1295 * batches. CPUs are conditionally walked to compute a more accurate total.
1296 *
1297 * Limit how much of a slop are we willing to tolerate. Note: the actual value
1298 * at any given moment can still exceed slop, but it should not be by significant
1299 * margin in practice.
1300 */
1301#define VNLRU_FREEVNODES_SLOP 128
1302
1303static u_long
1304vnlru_read_freevnodes(void)
1305{
1306	struct vdbatch *vd;
1307	long slop;
1308	int cpu;
1309
1310	mtx_assert(&vnode_list_mtx, MA_OWNED);
1311	if (freevnodes > freevnodes_old)
1312		slop = freevnodes - freevnodes_old;
1313	else
1314		slop = freevnodes_old - freevnodes;
1315	if (slop < VNLRU_FREEVNODES_SLOP)
1316		return (freevnodes >= 0 ? freevnodes : 0);
1317	freevnodes_old = freevnodes;
1318	CPU_FOREACH(cpu) {
1319		vd = DPCPU_ID_PTR((cpu), vd);
1320		freevnodes_old += vd->freevnodes;
1321	}
1322	return (freevnodes_old >= 0 ? freevnodes_old : 0);
1323}
1324
1325static bool
1326vnlru_under(u_long rnumvnodes, u_long limit)
1327{
1328	u_long rfreevnodes, space;
1329
1330	if (__predict_false(rnumvnodes > desiredvnodes))
1331		return (true);
1332
1333	space = desiredvnodes - rnumvnodes;
1334	if (space < limit) {
1335		rfreevnodes = vnlru_read_freevnodes();
1336		if (rfreevnodes > wantfreevnodes)
1337			space += rfreevnodes - wantfreevnodes;
1338	}
1339	return (space < limit);
1340}
1341
1342static bool
1343vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1344{
1345	long rfreevnodes, space;
1346
1347	if (__predict_false(rnumvnodes > desiredvnodes))
1348		return (true);
1349
1350	space = desiredvnodes - rnumvnodes;
1351	if (space < limit) {
1352		rfreevnodes = atomic_load_long(&freevnodes);
1353		if (rfreevnodes > wantfreevnodes)
1354			space += rfreevnodes - wantfreevnodes;
1355	}
1356	return (space < limit);
1357}
1358
1359static void
1360vnlru_kick(void)
1361{
1362
1363	mtx_assert(&vnode_list_mtx, MA_OWNED);
1364	if (vnlruproc_sig == 0) {
1365		vnlruproc_sig = 1;
1366		wakeup(vnlruproc);
1367	}
1368}
1369
1370static void
1371vnlru_proc(void)
1372{
1373	u_long rnumvnodes, rfreevnodes, target;
1374	unsigned long onumvnodes;
1375	int done, force, trigger, usevnodes;
1376	bool reclaim_nc_src, want_reread;
1377
1378	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1379	    SHUTDOWN_PRI_FIRST);
1380
1381	force = 0;
1382	want_reread = false;
1383	for (;;) {
1384		kproc_suspend_check(vnlruproc);
1385		mtx_lock(&vnode_list_mtx);
1386		rnumvnodes = atomic_load_long(&numvnodes);
1387
1388		if (want_reread) {
1389			force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1390			want_reread = false;
1391		}
1392
1393		/*
1394		 * If numvnodes is too large (due to desiredvnodes being
1395		 * adjusted using its sysctl, or emergency growth), first
1396		 * try to reduce it by discarding from the free list.
1397		 */
1398		if (rnumvnodes > desiredvnodes) {
1399			vnlru_free_locked(rnumvnodes - desiredvnodes, NULL);
1400			rnumvnodes = atomic_load_long(&numvnodes);
1401		}
1402		/*
1403		 * Sleep if the vnode cache is in a good state.  This is
1404		 * when it is not over-full and has space for about a 4%
1405		 * or 9% expansion (by growing its size or inexcessively
1406		 * reducing its free list).  Otherwise, try to reclaim
1407		 * space for a 10% expansion.
1408		 */
1409		if (vstir && force == 0) {
1410			force = 1;
1411			vstir = 0;
1412		}
1413		if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1414			vnlruproc_sig = 0;
1415			wakeup(&vnlruproc_sig);
1416			msleep(vnlruproc, &vnode_list_mtx,
1417			    PVFS|PDROP, "vlruwt", hz);
1418			continue;
1419		}
1420		rfreevnodes = vnlru_read_freevnodes();
1421
1422		onumvnodes = rnumvnodes;
1423		/*
1424		 * Calculate parameters for recycling.  These are the same
1425		 * throughout the loop to give some semblance of fairness.
1426		 * The trigger point is to avoid recycling vnodes with lots
1427		 * of resident pages.  We aren't trying to free memory; we
1428		 * are trying to recycle or at least free vnodes.
1429		 */
1430		if (rnumvnodes <= desiredvnodes)
1431			usevnodes = rnumvnodes - rfreevnodes;
1432		else
1433			usevnodes = rnumvnodes;
1434		if (usevnodes <= 0)
1435			usevnodes = 1;
1436		/*
1437		 * The trigger value is is chosen to give a conservatively
1438		 * large value to ensure that it alone doesn't prevent
1439		 * making progress.  The value can easily be so large that
1440		 * it is effectively infinite in some congested and
1441		 * misconfigured cases, and this is necessary.  Normally
1442		 * it is about 8 to 100 (pages), which is quite large.
1443		 */
1444		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1445		if (force < 2)
1446			trigger = vsmalltrigger;
1447		reclaim_nc_src = force >= 3;
1448		target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1449		target = target / 10 + 1;
1450		done = vlrureclaim(reclaim_nc_src, trigger, target);
1451		mtx_unlock(&vnode_list_mtx);
1452		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1453			uma_reclaim(UMA_RECLAIM_DRAIN);
1454		if (done == 0) {
1455			if (force == 0 || force == 1) {
1456				force = 2;
1457				continue;
1458			}
1459			if (force == 2) {
1460				force = 3;
1461				continue;
1462			}
1463			want_reread = true;
1464			force = 0;
1465			vnlru_nowhere++;
1466			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1467		} else {
1468			want_reread = true;
1469			kern_yield(PRI_USER);
1470		}
1471	}
1472}
1473
1474static struct kproc_desc vnlru_kp = {
1475	"vnlru",
1476	vnlru_proc,
1477	&vnlruproc
1478};
1479SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1480    &vnlru_kp);
1481
1482/*
1483 * Routines having to do with the management of the vnode table.
1484 */
1485
1486/*
1487 * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1488 * before we actually vgone().  This function must be called with the vnode
1489 * held to prevent the vnode from being returned to the free list midway
1490 * through vgone().
1491 */
1492static int
1493vtryrecycle(struct vnode *vp)
1494{
1495	struct mount *vnmp;
1496
1497	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1498	VNASSERT(vp->v_holdcnt, vp,
1499	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1500	/*
1501	 * This vnode may found and locked via some other list, if so we
1502	 * can't recycle it yet.
1503	 */
1504	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1505		CTR2(KTR_VFS,
1506		    "%s: impossible to recycle, vp %p lock is already held",
1507		    __func__, vp);
1508		return (EWOULDBLOCK);
1509	}
1510	/*
1511	 * Don't recycle if its filesystem is being suspended.
1512	 */
1513	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1514		VOP_UNLOCK(vp);
1515		CTR2(KTR_VFS,
1516		    "%s: impossible to recycle, cannot start the write for %p",
1517		    __func__, vp);
1518		return (EBUSY);
1519	}
1520	/*
1521	 * If we got this far, we need to acquire the interlock and see if
1522	 * anyone picked up this vnode from another list.  If not, we will
1523	 * mark it with DOOMED via vgonel() so that anyone who does find it
1524	 * will skip over it.
1525	 */
1526	VI_LOCK(vp);
1527	if (vp->v_usecount) {
1528		VOP_UNLOCK(vp);
1529		VI_UNLOCK(vp);
1530		vn_finished_write(vnmp);
1531		CTR2(KTR_VFS,
1532		    "%s: impossible to recycle, %p is already referenced",
1533		    __func__, vp);
1534		return (EBUSY);
1535	}
1536	if (!VN_IS_DOOMED(vp)) {
1537		counter_u64_add(recycles_free_count, 1);
1538		vgonel(vp);
1539	}
1540	VOP_UNLOCK(vp);
1541	VI_UNLOCK(vp);
1542	vn_finished_write(vnmp);
1543	return (0);
1544}
1545
1546/*
1547 * Allocate a new vnode.
1548 *
1549 * The operation never returns an error. Returning an error was disabled
1550 * in r145385 (dated 2005) with the following comment:
1551 *
1552 * XXX Not all VFS_VGET/ffs_vget callers check returns.
1553 *
1554 * Given the age of this commit (almost 15 years at the time of writing this
1555 * comment) restoring the ability to fail requires a significant audit of
1556 * all codepaths.
1557 *
1558 * The routine can try to free a vnode or stall for up to 1 second waiting for
1559 * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1560 */
1561static u_long vn_alloc_cyclecount;
1562
1563static struct vnode * __noinline
1564vn_alloc_hard(struct mount *mp)
1565{
1566	u_long rnumvnodes, rfreevnodes;
1567
1568	mtx_lock(&vnode_list_mtx);
1569	rnumvnodes = atomic_load_long(&numvnodes);
1570	if (rnumvnodes + 1 < desiredvnodes) {
1571		vn_alloc_cyclecount = 0;
1572		goto alloc;
1573	}
1574	rfreevnodes = vnlru_read_freevnodes();
1575	if (vn_alloc_cyclecount++ >= rfreevnodes) {
1576		vn_alloc_cyclecount = 0;
1577		vstir = 1;
1578	}
1579	/*
1580	 * Grow the vnode cache if it will not be above its target max
1581	 * after growing.  Otherwise, if the free list is nonempty, try
1582	 * to reclaim 1 item from it before growing the cache (possibly
1583	 * above its target max if the reclamation failed or is delayed).
1584	 * Otherwise, wait for some space.  In all cases, schedule
1585	 * vnlru_proc() if we are getting short of space.  The watermarks
1586	 * should be chosen so that we never wait or even reclaim from
1587	 * the free list to below its target minimum.
1588	 */
1589	if (vnlru_free_locked(1, NULL) > 0)
1590		goto alloc;
1591	if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1592		/*
1593		 * Wait for space for a new vnode.
1594		 */
1595		vnlru_kick();
1596		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1597		if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1598		    vnlru_read_freevnodes() > 1)
1599			vnlru_free_locked(1, NULL);
1600	}
1601alloc:
1602	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1603	if (vnlru_under(rnumvnodes, vlowat))
1604		vnlru_kick();
1605	mtx_unlock(&vnode_list_mtx);
1606	return (uma_zalloc(vnode_zone, M_WAITOK));
1607}
1608
1609static struct vnode *
1610vn_alloc(struct mount *mp)
1611{
1612	u_long rnumvnodes;
1613
1614	if (__predict_false(vn_alloc_cyclecount != 0))
1615		return (vn_alloc_hard(mp));
1616	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1617	if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1618		atomic_subtract_long(&numvnodes, 1);
1619		return (vn_alloc_hard(mp));
1620	}
1621
1622	return (uma_zalloc(vnode_zone, M_WAITOK));
1623}
1624
1625static void
1626vn_free(struct vnode *vp)
1627{
1628
1629	atomic_subtract_long(&numvnodes, 1);
1630	uma_zfree(vnode_zone, vp);
1631}
1632
1633/*
1634 * Return the next vnode from the free list.
1635 */
1636int
1637getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1638    struct vnode **vpp)
1639{
1640	struct vnode *vp;
1641	struct thread *td;
1642	struct lock_object *lo;
1643
1644	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1645
1646	KASSERT(vops->registered,
1647	    ("%s: not registered vector op %p\n", __func__, vops));
1648
1649	td = curthread;
1650	if (td->td_vp_reserved != NULL) {
1651		vp = td->td_vp_reserved;
1652		td->td_vp_reserved = NULL;
1653	} else {
1654		vp = vn_alloc(mp);
1655	}
1656	counter_u64_add(vnodes_created, 1);
1657	/*
1658	 * Locks are given the generic name "vnode" when created.
1659	 * Follow the historic practice of using the filesystem
1660	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1661	 *
1662	 * Locks live in a witness group keyed on their name. Thus,
1663	 * when a lock is renamed, it must also move from the witness
1664	 * group of its old name to the witness group of its new name.
1665	 *
1666	 * The change only needs to be made when the vnode moves
1667	 * from one filesystem type to another. We ensure that each
1668	 * filesystem use a single static name pointer for its tag so
1669	 * that we can compare pointers rather than doing a strcmp().
1670	 */
1671	lo = &vp->v_vnlock->lock_object;
1672#ifdef WITNESS
1673	if (lo->lo_name != tag) {
1674#endif
1675		lo->lo_name = tag;
1676#ifdef WITNESS
1677		WITNESS_DESTROY(lo);
1678		WITNESS_INIT(lo, tag);
1679	}
1680#endif
1681	/*
1682	 * By default, don't allow shared locks unless filesystems opt-in.
1683	 */
1684	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1685	/*
1686	 * Finalize various vnode identity bits.
1687	 */
1688	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1689	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1690	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1691	vp->v_type = VNON;
1692	vp->v_op = vops;
1693	v_init_counters(vp);
1694	vp->v_bufobj.bo_ops = &buf_ops_bio;
1695#ifdef DIAGNOSTIC
1696	if (mp == NULL && vops != &dead_vnodeops)
1697		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1698#endif
1699#ifdef MAC
1700	mac_vnode_init(vp);
1701	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1702		mac_vnode_associate_singlelabel(mp, vp);
1703#endif
1704	if (mp != NULL) {
1705		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1706		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1707			vp->v_vflag |= VV_NOKNOTE;
1708	}
1709
1710	/*
1711	 * For the filesystems which do not use vfs_hash_insert(),
1712	 * still initialize v_hash to have vfs_hash_index() useful.
1713	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1714	 * its own hashing.
1715	 */
1716	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1717
1718	*vpp = vp;
1719	return (0);
1720}
1721
1722void
1723getnewvnode_reserve(void)
1724{
1725	struct thread *td;
1726
1727	td = curthread;
1728	MPASS(td->td_vp_reserved == NULL);
1729	td->td_vp_reserved = vn_alloc(NULL);
1730}
1731
1732void
1733getnewvnode_drop_reserve(void)
1734{
1735	struct thread *td;
1736
1737	td = curthread;
1738	if (td->td_vp_reserved != NULL) {
1739		vn_free(td->td_vp_reserved);
1740		td->td_vp_reserved = NULL;
1741	}
1742}
1743
1744static void
1745freevnode(struct vnode *vp)
1746{
1747	struct bufobj *bo;
1748
1749	/*
1750	 * The vnode has been marked for destruction, so free it.
1751	 *
1752	 * The vnode will be returned to the zone where it will
1753	 * normally remain until it is needed for another vnode. We
1754	 * need to cleanup (or verify that the cleanup has already
1755	 * been done) any residual data left from its current use
1756	 * so as not to contaminate the freshly allocated vnode.
1757	 */
1758	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1759	bo = &vp->v_bufobj;
1760	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1761	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
1762	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1763	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1764	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1765	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1766	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1767	    ("clean blk trie not empty"));
1768	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1769	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1770	    ("dirty blk trie not empty"));
1771	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1772	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1773	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1774	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1775	    ("Dangling rangelock waiters"));
1776	VI_UNLOCK(vp);
1777#ifdef MAC
1778	mac_vnode_destroy(vp);
1779#endif
1780	if (vp->v_pollinfo != NULL) {
1781		destroy_vpollinfo(vp->v_pollinfo);
1782		vp->v_pollinfo = NULL;
1783	}
1784#ifdef INVARIANTS
1785	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
1786	vp->v_op = NULL;
1787#endif
1788	vp->v_mountedhere = NULL;
1789	vp->v_unpcb = NULL;
1790	vp->v_rdev = NULL;
1791	vp->v_fifoinfo = NULL;
1792	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
1793	vp->v_irflag = 0;
1794	vp->v_iflag = 0;
1795	vp->v_vflag = 0;
1796	bo->bo_flag = 0;
1797	vn_free(vp);
1798}
1799
1800/*
1801 * Delete from old mount point vnode list, if on one.
1802 */
1803static void
1804delmntque(struct vnode *vp)
1805{
1806	struct mount *mp;
1807
1808	VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
1809
1810	mp = vp->v_mount;
1811	if (mp == NULL)
1812		return;
1813	MNT_ILOCK(mp);
1814	VI_LOCK(vp);
1815	vp->v_mount = NULL;
1816	VI_UNLOCK(vp);
1817	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1818		("bad mount point vnode list size"));
1819	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1820	mp->mnt_nvnodelistsize--;
1821	MNT_REL(mp);
1822	MNT_IUNLOCK(mp);
1823}
1824
1825static void
1826insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1827{
1828
1829	vp->v_data = NULL;
1830	vp->v_op = &dead_vnodeops;
1831	vgone(vp);
1832	vput(vp);
1833}
1834
1835/*
1836 * Insert into list of vnodes for the new mount point, if available.
1837 */
1838int
1839insmntque1(struct vnode *vp, struct mount *mp,
1840	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1841{
1842
1843	KASSERT(vp->v_mount == NULL,
1844		("insmntque: vnode already on per mount vnode list"));
1845	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1846	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1847
1848	/*
1849	 * We acquire the vnode interlock early to ensure that the
1850	 * vnode cannot be recycled by another process releasing a
1851	 * holdcnt on it before we get it on both the vnode list
1852	 * and the active vnode list. The mount mutex protects only
1853	 * manipulation of the vnode list and the vnode freelist
1854	 * mutex protects only manipulation of the active vnode list.
1855	 * Hence the need to hold the vnode interlock throughout.
1856	 */
1857	MNT_ILOCK(mp);
1858	VI_LOCK(vp);
1859	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1860	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1861	    mp->mnt_nvnodelistsize == 0)) &&
1862	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1863		VI_UNLOCK(vp);
1864		MNT_IUNLOCK(mp);
1865		if (dtr != NULL)
1866			dtr(vp, dtr_arg);
1867		return (EBUSY);
1868	}
1869	vp->v_mount = mp;
1870	MNT_REF(mp);
1871	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1872	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1873		("neg mount point vnode list size"));
1874	mp->mnt_nvnodelistsize++;
1875	VI_UNLOCK(vp);
1876	MNT_IUNLOCK(mp);
1877	return (0);
1878}
1879
1880int
1881insmntque(struct vnode *vp, struct mount *mp)
1882{
1883
1884	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1885}
1886
1887/*
1888 * Flush out and invalidate all buffers associated with a bufobj
1889 * Called with the underlying object locked.
1890 */
1891int
1892bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1893{
1894	int error;
1895
1896	BO_LOCK(bo);
1897	if (flags & V_SAVE) {
1898		error = bufobj_wwait(bo, slpflag, slptimeo);
1899		if (error) {
1900			BO_UNLOCK(bo);
1901			return (error);
1902		}
1903		if (bo->bo_dirty.bv_cnt > 0) {
1904			BO_UNLOCK(bo);
1905			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1906				return (error);
1907			/*
1908			 * XXX We could save a lock/unlock if this was only
1909			 * enabled under INVARIANTS
1910			 */
1911			BO_LOCK(bo);
1912			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1913				panic("vinvalbuf: dirty bufs");
1914		}
1915	}
1916	/*
1917	 * If you alter this loop please notice that interlock is dropped and
1918	 * reacquired in flushbuflist.  Special care is needed to ensure that
1919	 * no race conditions occur from this.
1920	 */
1921	do {
1922		error = flushbuflist(&bo->bo_clean,
1923		    flags, bo, slpflag, slptimeo);
1924		if (error == 0 && !(flags & V_CLEANONLY))
1925			error = flushbuflist(&bo->bo_dirty,
1926			    flags, bo, slpflag, slptimeo);
1927		if (error != 0 && error != EAGAIN) {
1928			BO_UNLOCK(bo);
1929			return (error);
1930		}
1931	} while (error != 0);
1932
1933	/*
1934	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1935	 * have write I/O in-progress but if there is a VM object then the
1936	 * VM object can also have read-I/O in-progress.
1937	 */
1938	do {
1939		bufobj_wwait(bo, 0, 0);
1940		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
1941			BO_UNLOCK(bo);
1942			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
1943			BO_LOCK(bo);
1944		}
1945	} while (bo->bo_numoutput > 0);
1946	BO_UNLOCK(bo);
1947
1948	/*
1949	 * Destroy the copy in the VM cache, too.
1950	 */
1951	if (bo->bo_object != NULL &&
1952	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1953		VM_OBJECT_WLOCK(bo->bo_object);
1954		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1955		    OBJPR_CLEANONLY : 0);
1956		VM_OBJECT_WUNLOCK(bo->bo_object);
1957	}
1958
1959#ifdef INVARIANTS
1960	BO_LOCK(bo);
1961	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1962	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1963	    bo->bo_clean.bv_cnt > 0))
1964		panic("vinvalbuf: flush failed");
1965	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1966	    bo->bo_dirty.bv_cnt > 0)
1967		panic("vinvalbuf: flush dirty failed");
1968	BO_UNLOCK(bo);
1969#endif
1970	return (0);
1971}
1972
1973/*
1974 * Flush out and invalidate all buffers associated with a vnode.
1975 * Called with the underlying object locked.
1976 */
1977int
1978vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1979{
1980
1981	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1982	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1983	if (vp->v_object != NULL && vp->v_object->handle != vp)
1984		return (0);
1985	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1986}
1987
1988/*
1989 * Flush out buffers on the specified list.
1990 *
1991 */
1992static int
1993flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1994    int slptimeo)
1995{
1996	struct buf *bp, *nbp;
1997	int retval, error;
1998	daddr_t lblkno;
1999	b_xflags_t xflags;
2000
2001	ASSERT_BO_WLOCKED(bo);
2002
2003	retval = 0;
2004	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2005		/*
2006		 * If we are flushing both V_NORMAL and V_ALT buffers then
2007		 * do not skip any buffers. If we are flushing only V_NORMAL
2008		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2009		 * flushing only V_ALT buffers then skip buffers not marked
2010		 * as BX_ALTDATA.
2011		 */
2012		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2013		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2014		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2015			continue;
2016		}
2017		if (nbp != NULL) {
2018			lblkno = nbp->b_lblkno;
2019			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2020		}
2021		retval = EAGAIN;
2022		error = BUF_TIMELOCK(bp,
2023		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2024		    "flushbuf", slpflag, slptimeo);
2025		if (error) {
2026			BO_LOCK(bo);
2027			return (error != ENOLCK ? error : EAGAIN);
2028		}
2029		KASSERT(bp->b_bufobj == bo,
2030		    ("bp %p wrong b_bufobj %p should be %p",
2031		    bp, bp->b_bufobj, bo));
2032		/*
2033		 * XXX Since there are no node locks for NFS, I
2034		 * believe there is a slight chance that a delayed
2035		 * write will occur while sleeping just above, so
2036		 * check for it.
2037		 */
2038		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2039		    (flags & V_SAVE)) {
2040			bremfree(bp);
2041			bp->b_flags |= B_ASYNC;
2042			bwrite(bp);
2043			BO_LOCK(bo);
2044			return (EAGAIN);	/* XXX: why not loop ? */
2045		}
2046		bremfree(bp);
2047		bp->b_flags |= (B_INVAL | B_RELBUF);
2048		bp->b_flags &= ~B_ASYNC;
2049		brelse(bp);
2050		BO_LOCK(bo);
2051		if (nbp == NULL)
2052			break;
2053		nbp = gbincore(bo, lblkno);
2054		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2055		    != xflags)
2056			break;			/* nbp invalid */
2057	}
2058	return (retval);
2059}
2060
2061int
2062bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2063{
2064	struct buf *bp;
2065	int error;
2066	daddr_t lblkno;
2067
2068	ASSERT_BO_LOCKED(bo);
2069
2070	for (lblkno = startn;;) {
2071again:
2072		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2073		if (bp == NULL || bp->b_lblkno >= endn ||
2074		    bp->b_lblkno < startn)
2075			break;
2076		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2077		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2078		if (error != 0) {
2079			BO_RLOCK(bo);
2080			if (error == ENOLCK)
2081				goto again;
2082			return (error);
2083		}
2084		KASSERT(bp->b_bufobj == bo,
2085		    ("bp %p wrong b_bufobj %p should be %p",
2086		    bp, bp->b_bufobj, bo));
2087		lblkno = bp->b_lblkno + 1;
2088		if ((bp->b_flags & B_MANAGED) == 0)
2089			bremfree(bp);
2090		bp->b_flags |= B_RELBUF;
2091		/*
2092		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2093		 * pages backing each buffer in the range are unlikely to be
2094		 * reused.  Dirty buffers will have the hint applied once
2095		 * they've been written.
2096		 */
2097		if ((bp->b_flags & B_VMIO) != 0)
2098			bp->b_flags |= B_NOREUSE;
2099		brelse(bp);
2100		BO_RLOCK(bo);
2101	}
2102	return (0);
2103}
2104
2105/*
2106 * Truncate a file's buffer and pages to a specified length.  This
2107 * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2108 * sync activity.
2109 */
2110int
2111vtruncbuf(struct vnode *vp, off_t length, int blksize)
2112{
2113	struct buf *bp, *nbp;
2114	struct bufobj *bo;
2115	daddr_t startlbn;
2116
2117	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2118	    vp, blksize, (uintmax_t)length);
2119
2120	/*
2121	 * Round up to the *next* lbn.
2122	 */
2123	startlbn = howmany(length, blksize);
2124
2125	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2126
2127	bo = &vp->v_bufobj;
2128restart_unlocked:
2129	BO_LOCK(bo);
2130
2131	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2132		;
2133
2134	if (length > 0) {
2135restartsync:
2136		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2137			if (bp->b_lblkno > 0)
2138				continue;
2139			/*
2140			 * Since we hold the vnode lock this should only
2141			 * fail if we're racing with the buf daemon.
2142			 */
2143			if (BUF_LOCK(bp,
2144			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2145			    BO_LOCKPTR(bo)) == ENOLCK)
2146				goto restart_unlocked;
2147
2148			VNASSERT((bp->b_flags & B_DELWRI), vp,
2149			    ("buf(%p) on dirty queue without DELWRI", bp));
2150
2151			bremfree(bp);
2152			bawrite(bp);
2153			BO_LOCK(bo);
2154			goto restartsync;
2155		}
2156	}
2157
2158	bufobj_wwait(bo, 0, 0);
2159	BO_UNLOCK(bo);
2160	vnode_pager_setsize(vp, length);
2161
2162	return (0);
2163}
2164
2165/*
2166 * Invalidate the cached pages of a file's buffer within the range of block
2167 * numbers [startlbn, endlbn).
2168 */
2169void
2170v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2171    int blksize)
2172{
2173	struct bufobj *bo;
2174	off_t start, end;
2175
2176	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2177
2178	start = blksize * startlbn;
2179	end = blksize * endlbn;
2180
2181	bo = &vp->v_bufobj;
2182	BO_LOCK(bo);
2183	MPASS(blksize == bo->bo_bsize);
2184
2185	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2186		;
2187
2188	BO_UNLOCK(bo);
2189	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2190}
2191
2192static int
2193v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2194    daddr_t startlbn, daddr_t endlbn)
2195{
2196	struct buf *bp, *nbp;
2197	bool anyfreed;
2198
2199	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2200	ASSERT_BO_LOCKED(bo);
2201
2202	do {
2203		anyfreed = false;
2204		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2205			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2206				continue;
2207			if (BUF_LOCK(bp,
2208			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2209			    BO_LOCKPTR(bo)) == ENOLCK) {
2210				BO_LOCK(bo);
2211				return (EAGAIN);
2212			}
2213
2214			bremfree(bp);
2215			bp->b_flags |= B_INVAL | B_RELBUF;
2216			bp->b_flags &= ~B_ASYNC;
2217			brelse(bp);
2218			anyfreed = true;
2219
2220			BO_LOCK(bo);
2221			if (nbp != NULL &&
2222			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2223			    nbp->b_vp != vp ||
2224			    (nbp->b_flags & B_DELWRI) != 0))
2225				return (EAGAIN);
2226		}
2227
2228		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2229			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2230				continue;
2231			if (BUF_LOCK(bp,
2232			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2233			    BO_LOCKPTR(bo)) == ENOLCK) {
2234				BO_LOCK(bo);
2235				return (EAGAIN);
2236			}
2237			bremfree(bp);
2238			bp->b_flags |= B_INVAL | B_RELBUF;
2239			bp->b_flags &= ~B_ASYNC;
2240			brelse(bp);
2241			anyfreed = true;
2242
2243			BO_LOCK(bo);
2244			if (nbp != NULL &&
2245			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2246			    (nbp->b_vp != vp) ||
2247			    (nbp->b_flags & B_DELWRI) == 0))
2248				return (EAGAIN);
2249		}
2250	} while (anyfreed);
2251	return (0);
2252}
2253
2254static void
2255buf_vlist_remove(struct buf *bp)
2256{
2257	struct bufv *bv;
2258
2259	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2260	ASSERT_BO_WLOCKED(bp->b_bufobj);
2261	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
2262	    (BX_VNDIRTY|BX_VNCLEAN),
2263	    ("buf_vlist_remove: Buf %p is on two lists", bp));
2264	if (bp->b_xflags & BX_VNDIRTY)
2265		bv = &bp->b_bufobj->bo_dirty;
2266	else
2267		bv = &bp->b_bufobj->bo_clean;
2268	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2269	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2270	bv->bv_cnt--;
2271	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2272}
2273
2274/*
2275 * Add the buffer to the sorted clean or dirty block list.
2276 *
2277 * NOTE: xflags is passed as a constant, optimizing this inline function!
2278 */
2279static void
2280buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2281{
2282	struct bufv *bv;
2283	struct buf *n;
2284	int error;
2285
2286	ASSERT_BO_WLOCKED(bo);
2287	KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2288	    ("buf_vlist_add: bo %p does not allow bufs", bo));
2289	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2290	    ("dead bo %p", bo));
2291	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2292	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2293	bp->b_xflags |= xflags;
2294	if (xflags & BX_VNDIRTY)
2295		bv = &bo->bo_dirty;
2296	else
2297		bv = &bo->bo_clean;
2298
2299	/*
2300	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2301	 * we tend to grow at the tail so lookup_le should usually be cheaper
2302	 * than _ge.
2303	 */
2304	if (bv->bv_cnt == 0 ||
2305	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2306		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2307	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2308		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2309	else
2310		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2311	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2312	if (error)
2313		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2314	bv->bv_cnt++;
2315}
2316
2317/*
2318 * Look up a buffer using the buffer tries.
2319 */
2320struct buf *
2321gbincore(struct bufobj *bo, daddr_t lblkno)
2322{
2323	struct buf *bp;
2324
2325	ASSERT_BO_LOCKED(bo);
2326	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2327	if (bp != NULL)
2328		return (bp);
2329	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
2330}
2331
2332/*
2333 * Associate a buffer with a vnode.
2334 */
2335void
2336bgetvp(struct vnode *vp, struct buf *bp)
2337{
2338	struct bufobj *bo;
2339
2340	bo = &vp->v_bufobj;
2341	ASSERT_BO_WLOCKED(bo);
2342	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2343
2344	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2345	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2346	    ("bgetvp: bp already attached! %p", bp));
2347
2348	vhold(vp);
2349	bp->b_vp = vp;
2350	bp->b_bufobj = bo;
2351	/*
2352	 * Insert onto list for new vnode.
2353	 */
2354	buf_vlist_add(bp, bo, BX_VNCLEAN);
2355}
2356
2357/*
2358 * Disassociate a buffer from a vnode.
2359 */
2360void
2361brelvp(struct buf *bp)
2362{
2363	struct bufobj *bo;
2364	struct vnode *vp;
2365
2366	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2367	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2368
2369	/*
2370	 * Delete from old vnode list, if on one.
2371	 */
2372	vp = bp->b_vp;		/* XXX */
2373	bo = bp->b_bufobj;
2374	BO_LOCK(bo);
2375	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2376		buf_vlist_remove(bp);
2377	else
2378		panic("brelvp: Buffer %p not on queue.", bp);
2379	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2380		bo->bo_flag &= ~BO_ONWORKLST;
2381		mtx_lock(&sync_mtx);
2382		LIST_REMOVE(bo, bo_synclist);
2383		syncer_worklist_len--;
2384		mtx_unlock(&sync_mtx);
2385	}
2386	bp->b_vp = NULL;
2387	bp->b_bufobj = NULL;
2388	BO_UNLOCK(bo);
2389	vdrop(vp);
2390}
2391
2392/*
2393 * Add an item to the syncer work queue.
2394 */
2395static void
2396vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2397{
2398	int slot;
2399
2400	ASSERT_BO_WLOCKED(bo);
2401
2402	mtx_lock(&sync_mtx);
2403	if (bo->bo_flag & BO_ONWORKLST)
2404		LIST_REMOVE(bo, bo_synclist);
2405	else {
2406		bo->bo_flag |= BO_ONWORKLST;
2407		syncer_worklist_len++;
2408	}
2409
2410	if (delay > syncer_maxdelay - 2)
2411		delay = syncer_maxdelay - 2;
2412	slot = (syncer_delayno + delay) & syncer_mask;
2413
2414	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2415	mtx_unlock(&sync_mtx);
2416}
2417
2418static int
2419sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2420{
2421	int error, len;
2422
2423	mtx_lock(&sync_mtx);
2424	len = syncer_worklist_len - sync_vnode_count;
2425	mtx_unlock(&sync_mtx);
2426	error = SYSCTL_OUT(req, &len, sizeof(len));
2427	return (error);
2428}
2429
2430SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2431    CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2432    sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2433
2434static struct proc *updateproc;
2435static void sched_sync(void);
2436static struct kproc_desc up_kp = {
2437	"syncer",
2438	sched_sync,
2439	&updateproc
2440};
2441SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2442
2443static int
2444sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2445{
2446	struct vnode *vp;
2447	struct mount *mp;
2448
2449	*bo = LIST_FIRST(slp);
2450	if (*bo == NULL)
2451		return (0);
2452	vp = bo2vnode(*bo);
2453	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2454		return (1);
2455	/*
2456	 * We use vhold in case the vnode does not
2457	 * successfully sync.  vhold prevents the vnode from
2458	 * going away when we unlock the sync_mtx so that
2459	 * we can acquire the vnode interlock.
2460	 */
2461	vholdl(vp);
2462	mtx_unlock(&sync_mtx);
2463	VI_UNLOCK(vp);
2464	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2465		vdrop(vp);
2466		mtx_lock(&sync_mtx);
2467		return (*bo == LIST_FIRST(slp));
2468	}
2469	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2470	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2471	VOP_UNLOCK(vp);
2472	vn_finished_write(mp);
2473	BO_LOCK(*bo);
2474	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2475		/*
2476		 * Put us back on the worklist.  The worklist
2477		 * routine will remove us from our current
2478		 * position and then add us back in at a later
2479		 * position.
2480		 */
2481		vn_syncer_add_to_worklist(*bo, syncdelay);
2482	}
2483	BO_UNLOCK(*bo);
2484	vdrop(vp);
2485	mtx_lock(&sync_mtx);
2486	return (0);
2487}
2488
2489static int first_printf = 1;
2490
2491/*
2492 * System filesystem synchronizer daemon.
2493 */
2494static void
2495sched_sync(void)
2496{
2497	struct synclist *next, *slp;
2498	struct bufobj *bo;
2499	long starttime;
2500	struct thread *td = curthread;
2501	int last_work_seen;
2502	int net_worklist_len;
2503	int syncer_final_iter;
2504	int error;
2505
2506	last_work_seen = 0;
2507	syncer_final_iter = 0;
2508	syncer_state = SYNCER_RUNNING;
2509	starttime = time_uptime;
2510	td->td_pflags |= TDP_NORUNNINGBUF;
2511
2512	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2513	    SHUTDOWN_PRI_LAST);
2514
2515	mtx_lock(&sync_mtx);
2516	for (;;) {
2517		if (syncer_state == SYNCER_FINAL_DELAY &&
2518		    syncer_final_iter == 0) {
2519			mtx_unlock(&sync_mtx);
2520			kproc_suspend_check(td->td_proc);
2521			mtx_lock(&sync_mtx);
2522		}
2523		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2524		if (syncer_state != SYNCER_RUNNING &&
2525		    starttime != time_uptime) {
2526			if (first_printf) {
2527				printf("\nSyncing disks, vnodes remaining... ");
2528				first_printf = 0;
2529			}
2530			printf("%d ", net_worklist_len);
2531		}
2532		starttime = time_uptime;
2533
2534		/*
2535		 * Push files whose dirty time has expired.  Be careful
2536		 * of interrupt race on slp queue.
2537		 *
2538		 * Skip over empty worklist slots when shutting down.
2539		 */
2540		do {
2541			slp = &syncer_workitem_pending[syncer_delayno];
2542			syncer_delayno += 1;
2543			if (syncer_delayno == syncer_maxdelay)
2544				syncer_delayno = 0;
2545			next = &syncer_workitem_pending[syncer_delayno];
2546			/*
2547			 * If the worklist has wrapped since the
2548			 * it was emptied of all but syncer vnodes,
2549			 * switch to the FINAL_DELAY state and run
2550			 * for one more second.
2551			 */
2552			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2553			    net_worklist_len == 0 &&
2554			    last_work_seen == syncer_delayno) {
2555				syncer_state = SYNCER_FINAL_DELAY;
2556				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2557			}
2558		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2559		    syncer_worklist_len > 0);
2560
2561		/*
2562		 * Keep track of the last time there was anything
2563		 * on the worklist other than syncer vnodes.
2564		 * Return to the SHUTTING_DOWN state if any
2565		 * new work appears.
2566		 */
2567		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2568			last_work_seen = syncer_delayno;
2569		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2570			syncer_state = SYNCER_SHUTTING_DOWN;
2571		while (!LIST_EMPTY(slp)) {
2572			error = sync_vnode(slp, &bo, td);
2573			if (error == 1) {
2574				LIST_REMOVE(bo, bo_synclist);
2575				LIST_INSERT_HEAD(next, bo, bo_synclist);
2576				continue;
2577			}
2578
2579			if (first_printf == 0) {
2580				/*
2581				 * Drop the sync mutex, because some watchdog
2582				 * drivers need to sleep while patting
2583				 */
2584				mtx_unlock(&sync_mtx);
2585				wdog_kern_pat(WD_LASTVAL);
2586				mtx_lock(&sync_mtx);
2587			}
2588
2589		}
2590		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2591			syncer_final_iter--;
2592		/*
2593		 * The variable rushjob allows the kernel to speed up the
2594		 * processing of the filesystem syncer process. A rushjob
2595		 * value of N tells the filesystem syncer to process the next
2596		 * N seconds worth of work on its queue ASAP. Currently rushjob
2597		 * is used by the soft update code to speed up the filesystem
2598		 * syncer process when the incore state is getting so far
2599		 * ahead of the disk that the kernel memory pool is being
2600		 * threatened with exhaustion.
2601		 */
2602		if (rushjob > 0) {
2603			rushjob -= 1;
2604			continue;
2605		}
2606		/*
2607		 * Just sleep for a short period of time between
2608		 * iterations when shutting down to allow some I/O
2609		 * to happen.
2610		 *
2611		 * If it has taken us less than a second to process the
2612		 * current work, then wait. Otherwise start right over
2613		 * again. We can still lose time if any single round
2614		 * takes more than two seconds, but it does not really
2615		 * matter as we are just trying to generally pace the
2616		 * filesystem activity.
2617		 */
2618		if (syncer_state != SYNCER_RUNNING ||
2619		    time_uptime == starttime) {
2620			thread_lock(td);
2621			sched_prio(td, PPAUSE);
2622			thread_unlock(td);
2623		}
2624		if (syncer_state != SYNCER_RUNNING)
2625			cv_timedwait(&sync_wakeup, &sync_mtx,
2626			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2627		else if (time_uptime == starttime)
2628			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2629	}
2630}
2631
2632/*
2633 * Request the syncer daemon to speed up its work.
2634 * We never push it to speed up more than half of its
2635 * normal turn time, otherwise it could take over the cpu.
2636 */
2637int
2638speedup_syncer(void)
2639{
2640	int ret = 0;
2641
2642	mtx_lock(&sync_mtx);
2643	if (rushjob < syncdelay / 2) {
2644		rushjob += 1;
2645		stat_rush_requests += 1;
2646		ret = 1;
2647	}
2648	mtx_unlock(&sync_mtx);
2649	cv_broadcast(&sync_wakeup);
2650	return (ret);
2651}
2652
2653/*
2654 * Tell the syncer to speed up its work and run though its work
2655 * list several times, then tell it to shut down.
2656 */
2657static void
2658syncer_shutdown(void *arg, int howto)
2659{
2660
2661	if (howto & RB_NOSYNC)
2662		return;
2663	mtx_lock(&sync_mtx);
2664	syncer_state = SYNCER_SHUTTING_DOWN;
2665	rushjob = 0;
2666	mtx_unlock(&sync_mtx);
2667	cv_broadcast(&sync_wakeup);
2668	kproc_shutdown(arg, howto);
2669}
2670
2671void
2672syncer_suspend(void)
2673{
2674
2675	syncer_shutdown(updateproc, 0);
2676}
2677
2678void
2679syncer_resume(void)
2680{
2681
2682	mtx_lock(&sync_mtx);
2683	first_printf = 1;
2684	syncer_state = SYNCER_RUNNING;
2685	mtx_unlock(&sync_mtx);
2686	cv_broadcast(&sync_wakeup);
2687	kproc_resume(updateproc);
2688}
2689
2690/*
2691 * Reassign a buffer from one vnode to another.
2692 * Used to assign file specific control information
2693 * (indirect blocks) to the vnode to which they belong.
2694 */
2695void
2696reassignbuf(struct buf *bp)
2697{
2698	struct vnode *vp;
2699	struct bufobj *bo;
2700	int delay;
2701#ifdef INVARIANTS
2702	struct bufv *bv;
2703#endif
2704
2705	vp = bp->b_vp;
2706	bo = bp->b_bufobj;
2707	++reassignbufcalls;
2708
2709	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2710	    bp, bp->b_vp, bp->b_flags);
2711	/*
2712	 * B_PAGING flagged buffers cannot be reassigned because their vp
2713	 * is not fully linked in.
2714	 */
2715	if (bp->b_flags & B_PAGING)
2716		panic("cannot reassign paging buffer");
2717
2718	/*
2719	 * Delete from old vnode list, if on one.
2720	 */
2721	BO_LOCK(bo);
2722	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2723		buf_vlist_remove(bp);
2724	else
2725		panic("reassignbuf: Buffer %p not on queue.", bp);
2726	/*
2727	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2728	 * of clean buffers.
2729	 */
2730	if (bp->b_flags & B_DELWRI) {
2731		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2732			switch (vp->v_type) {
2733			case VDIR:
2734				delay = dirdelay;
2735				break;
2736			case VCHR:
2737				delay = metadelay;
2738				break;
2739			default:
2740				delay = filedelay;
2741			}
2742			vn_syncer_add_to_worklist(bo, delay);
2743		}
2744		buf_vlist_add(bp, bo, BX_VNDIRTY);
2745	} else {
2746		buf_vlist_add(bp, bo, BX_VNCLEAN);
2747
2748		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2749			mtx_lock(&sync_mtx);
2750			LIST_REMOVE(bo, bo_synclist);
2751			syncer_worklist_len--;
2752			mtx_unlock(&sync_mtx);
2753			bo->bo_flag &= ~BO_ONWORKLST;
2754		}
2755	}
2756#ifdef INVARIANTS
2757	bv = &bo->bo_clean;
2758	bp = TAILQ_FIRST(&bv->bv_hd);
2759	KASSERT(bp == NULL || bp->b_bufobj == bo,
2760	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2761	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2762	KASSERT(bp == NULL || bp->b_bufobj == bo,
2763	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2764	bv = &bo->bo_dirty;
2765	bp = TAILQ_FIRST(&bv->bv_hd);
2766	KASSERT(bp == NULL || bp->b_bufobj == bo,
2767	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2768	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2769	KASSERT(bp == NULL || bp->b_bufobj == bo,
2770	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2771#endif
2772	BO_UNLOCK(bo);
2773}
2774
2775static void
2776v_init_counters(struct vnode *vp)
2777{
2778
2779	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2780	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2781	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2782
2783	refcount_init(&vp->v_holdcnt, 1);
2784	refcount_init(&vp->v_usecount, 1);
2785}
2786
2787/*
2788 * Increment si_usecount of the associated device, if any.
2789 */
2790static void
2791v_incr_devcount(struct vnode *vp)
2792{
2793
2794	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2795	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2796		dev_lock();
2797		vp->v_rdev->si_usecount++;
2798		dev_unlock();
2799	}
2800}
2801
2802/*
2803 * Decrement si_usecount of the associated device, if any.
2804 *
2805 * The caller is required to hold the interlock when transitioning a VCHR use
2806 * count to zero. This prevents a race with devfs_reclaim_vchr() that would
2807 * leak a si_usecount reference. The vnode lock will also prevent this race
2808 * if it is held while dropping the last ref.
2809 *
2810 * The race is:
2811 *
2812 * CPU1					CPU2
2813 *				  	devfs_reclaim_vchr
2814 * make v_usecount == 0
2815 * 				    	  VI_LOCK
2816 * 				    	  sees v_usecount == 0, no updates
2817 * 				    	  vp->v_rdev = NULL;
2818 * 				    	  ...
2819 * 				    	  VI_UNLOCK
2820 * VI_LOCK
2821 * v_decr_devcount
2822 *   sees v_rdev == NULL, no updates
2823 *
2824 * In this scenario si_devcount decrement is not performed.
2825 */
2826static void
2827v_decr_devcount(struct vnode *vp)
2828{
2829
2830	ASSERT_VOP_LOCKED(vp, __func__);
2831	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2832	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2833		dev_lock();
2834		VNPASS(vp->v_rdev->si_usecount > 0, vp);
2835		vp->v_rdev->si_usecount--;
2836		dev_unlock();
2837	}
2838}
2839
2840/*
2841 * Grab a particular vnode from the free list, increment its
2842 * reference count and lock it.  VIRF_DOOMED is set if the vnode
2843 * is being destroyed.  Only callers who specify LK_RETRY will
2844 * see doomed vnodes.  If inactive processing was delayed in
2845 * vput try to do it here.
2846 *
2847 * usecount is manipulated using atomics without holding any locks.
2848 *
2849 * holdcnt can be manipulated using atomics without holding any locks,
2850 * except when transitioning 1<->0, in which case the interlock is held.
2851 */
2852enum vgetstate
2853vget_prep(struct vnode *vp)
2854{
2855	enum vgetstate vs;
2856
2857	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2858		vs = VGET_USECOUNT;
2859	} else {
2860		vhold(vp);
2861		vs = VGET_HOLDCNT;
2862	}
2863	return (vs);
2864}
2865
2866int
2867vget(struct vnode *vp, int flags, struct thread *td)
2868{
2869	enum vgetstate vs;
2870
2871	MPASS(td == curthread);
2872
2873	vs = vget_prep(vp);
2874	return (vget_finish(vp, flags, vs));
2875}
2876
2877static int __noinline
2878vget_finish_vchr(struct vnode *vp)
2879{
2880
2881	VNASSERT(vp->v_type == VCHR, vp, ("type != VCHR)"));
2882
2883	/*
2884	 * See the comment in vget_finish before usecount bump.
2885	 */
2886	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2887#ifdef INVARIANTS
2888		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2889		VNASSERT(old > 0, vp, ("%s: wrong hold count %d", __func__, old));
2890#else
2891		refcount_release(&vp->v_holdcnt);
2892#endif
2893		return (0);
2894	}
2895
2896	VI_LOCK(vp);
2897	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2898#ifdef INVARIANTS
2899		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2900		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2901#else
2902		refcount_release(&vp->v_holdcnt);
2903#endif
2904		VI_UNLOCK(vp);
2905		return (0);
2906	}
2907	v_incr_devcount(vp);
2908	refcount_acquire(&vp->v_usecount);
2909	VI_UNLOCK(vp);
2910	return (0);
2911}
2912
2913int
2914vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2915{
2916	int error, old;
2917
2918	if ((flags & LK_INTERLOCK) != 0)
2919		ASSERT_VI_LOCKED(vp, __func__);
2920	else
2921		ASSERT_VI_UNLOCKED(vp, __func__);
2922	VNPASS(vp->v_holdcnt > 0, vp);
2923	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
2924
2925	error = vn_lock(vp, flags);
2926	if (__predict_false(error != 0)) {
2927		if (vs == VGET_USECOUNT)
2928			vrele(vp);
2929		else
2930			vdrop(vp);
2931		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2932		    vp);
2933		return (error);
2934	}
2935
2936	if (vs == VGET_USECOUNT)
2937		return (0);
2938
2939	if (__predict_false(vp->v_type == VCHR))
2940		return (vget_finish_vchr(vp));
2941
2942	/*
2943	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
2944	 * the vnode around. Otherwise someone else lended their hold count and
2945	 * we have to drop ours.
2946	 */
2947	old = atomic_fetchadd_int(&vp->v_usecount, 1);
2948	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
2949	if (old != 0) {
2950#ifdef INVARIANTS
2951		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2952		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2953#else
2954		refcount_release(&vp->v_holdcnt);
2955#endif
2956	}
2957	return (0);
2958}
2959
2960/*
2961 * Increase the reference (use) and hold count of a vnode.
2962 * This will also remove the vnode from the free list if it is presently free.
2963 */
2964static void __noinline
2965vref_vchr(struct vnode *vp, bool interlock)
2966{
2967
2968	/*
2969	 * See the comment in vget_finish before usecount bump.
2970	 */
2971	if (!interlock) {
2972		if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2973			VNODE_REFCOUNT_FENCE_ACQ();
2974			VNASSERT(vp->v_holdcnt > 0, vp,
2975			    ("%s: active vnode not held", __func__));
2976			return;
2977		}
2978		VI_LOCK(vp);
2979		/*
2980		 * By the time we get here the vnode might have been doomed, at
2981		 * which point the 0->1 use count transition is no longer
2982		 * protected by the interlock. Since it can't bounce back to
2983		 * VCHR and requires vref semantics, punt it back
2984		 */
2985		if (__predict_false(vp->v_type == VBAD)) {
2986			VI_UNLOCK(vp);
2987			vref(vp);
2988			return;
2989		}
2990	}
2991	VNASSERT(vp->v_type == VCHR, vp, ("type != VCHR)"));
2992	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2993		VNODE_REFCOUNT_FENCE_ACQ();
2994		VNASSERT(vp->v_holdcnt > 0, vp,
2995		    ("%s: active vnode not held", __func__));
2996		if (!interlock)
2997			VI_UNLOCK(vp);
2998		return;
2999	}
3000	vhold(vp);
3001	v_incr_devcount(vp);
3002	refcount_acquire(&vp->v_usecount);
3003	if (!interlock)
3004		VI_UNLOCK(vp);
3005	return;
3006}
3007
3008void
3009vref(struct vnode *vp)
3010{
3011	int old;
3012
3013	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3014	if (__predict_false(vp->v_type == VCHR)) {
3015		 vref_vchr(vp, false);
3016		 return;
3017	}
3018
3019	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
3020		VNODE_REFCOUNT_FENCE_ACQ();
3021		VNASSERT(vp->v_holdcnt > 0, vp,
3022		    ("%s: active vnode not held", __func__));
3023		return;
3024	}
3025	vhold(vp);
3026	/*
3027	 * See the comment in vget_finish.
3028	 */
3029	old = atomic_fetchadd_int(&vp->v_usecount, 1);
3030	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3031	if (old != 0) {
3032#ifdef INVARIANTS
3033		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3034		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3035#else
3036		refcount_release(&vp->v_holdcnt);
3037#endif
3038	}
3039}
3040
3041void
3042vrefl(struct vnode *vp)
3043{
3044
3045	ASSERT_VI_LOCKED(vp, __func__);
3046	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3047	if (__predict_false(vp->v_type == VCHR)) {
3048		vref_vchr(vp, true);
3049		return;
3050	}
3051	vref(vp);
3052}
3053
3054void
3055vrefact(struct vnode *vp)
3056{
3057
3058	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3059#ifdef INVARIANTS
3060	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3061	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3062#else
3063	refcount_acquire(&vp->v_usecount);
3064#endif
3065}
3066
3067void
3068vrefactn(struct vnode *vp, u_int n)
3069{
3070
3071	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3072#ifdef INVARIANTS
3073	int old = atomic_fetchadd_int(&vp->v_usecount, n);
3074	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3075#else
3076	atomic_add_int(&vp->v_usecount, n);
3077#endif
3078}
3079
3080/*
3081 * Return reference count of a vnode.
3082 *
3083 * The results of this call are only guaranteed when some mechanism is used to
3084 * stop other processes from gaining references to the vnode.  This may be the
3085 * case if the caller holds the only reference.  This is also useful when stale
3086 * data is acceptable as race conditions may be accounted for by some other
3087 * means.
3088 */
3089int
3090vrefcnt(struct vnode *vp)
3091{
3092
3093	return (vp->v_usecount);
3094}
3095
3096void
3097vlazy(struct vnode *vp)
3098{
3099	struct mount *mp;
3100
3101	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3102
3103	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3104		return;
3105	/*
3106	 * We may get here for inactive routines after the vnode got doomed.
3107	 */
3108	if (VN_IS_DOOMED(vp))
3109		return;
3110	mp = vp->v_mount;
3111	mtx_lock(&mp->mnt_listmtx);
3112	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3113		vp->v_mflag |= VMP_LAZYLIST;
3114		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3115		mp->mnt_lazyvnodelistsize++;
3116	}
3117	mtx_unlock(&mp->mnt_listmtx);
3118}
3119
3120/*
3121 * This routine is only meant to be called from vgonel prior to dooming
3122 * the vnode.
3123 */
3124static void
3125vunlazy_gone(struct vnode *vp)
3126{
3127	struct mount *mp;
3128
3129	ASSERT_VOP_ELOCKED(vp, __func__);
3130	ASSERT_VI_LOCKED(vp, __func__);
3131	VNPASS(!VN_IS_DOOMED(vp), vp);
3132
3133	if (vp->v_mflag & VMP_LAZYLIST) {
3134		mp = vp->v_mount;
3135		mtx_lock(&mp->mnt_listmtx);
3136		VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3137		vp->v_mflag &= ~VMP_LAZYLIST;
3138		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3139		mp->mnt_lazyvnodelistsize--;
3140		mtx_unlock(&mp->mnt_listmtx);
3141	}
3142}
3143
3144static void
3145vdefer_inactive(struct vnode *vp)
3146{
3147
3148	ASSERT_VI_LOCKED(vp, __func__);
3149	VNASSERT(vp->v_holdcnt > 0, vp,
3150	    ("%s: vnode without hold count", __func__));
3151	if (VN_IS_DOOMED(vp)) {
3152		vdropl(vp);
3153		return;
3154	}
3155	if (vp->v_iflag & VI_DEFINACT) {
3156		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3157		vdropl(vp);
3158		return;
3159	}
3160	if (vp->v_usecount > 0) {
3161		vp->v_iflag &= ~VI_OWEINACT;
3162		vdropl(vp);
3163		return;
3164	}
3165	vlazy(vp);
3166	vp->v_iflag |= VI_DEFINACT;
3167	VI_UNLOCK(vp);
3168	counter_u64_add(deferred_inact, 1);
3169}
3170
3171static void
3172vdefer_inactive_unlocked(struct vnode *vp)
3173{
3174
3175	VI_LOCK(vp);
3176	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3177		vdropl(vp);
3178		return;
3179	}
3180	vdefer_inactive(vp);
3181}
3182
3183enum vput_op { VRELE, VPUT, VUNREF };
3184
3185/*
3186 * Handle ->v_usecount transitioning to 0.
3187 *
3188 * By releasing the last usecount we take ownership of the hold count which
3189 * provides liveness of the vnode, meaning we have to vdrop.
3190 *
3191 * If the vnode is of type VCHR we may need to decrement si_usecount, see
3192 * v_decr_devcount for details.
3193 *
3194 * For all vnodes we may need to perform inactive processing. It requires an
3195 * exclusive lock on the vnode, while it is legal to call here with only a
3196 * shared lock (or no locks). If locking the vnode in an expected manner fails,
3197 * inactive processing gets deferred to the syncer.
3198 *
3199 * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3200 * on the lock being held all the way until VOP_INACTIVE. This in particular
3201 * happens with UFS which adds half-constructed vnodes to the hash, where they
3202 * can be found by other code.
3203 */
3204static void
3205vput_final(struct vnode *vp, enum vput_op func)
3206{
3207	int error;
3208	bool want_unlock;
3209
3210	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3211	VNPASS(vp->v_holdcnt > 0, vp);
3212
3213	VI_LOCK(vp);
3214	if (__predict_false(vp->v_type == VCHR && func != VRELE))
3215		v_decr_devcount(vp);
3216
3217	/*
3218	 * By the time we got here someone else might have transitioned
3219	 * the count back to > 0.
3220	 */
3221	if (vp->v_usecount > 0)
3222		goto out;
3223
3224	/*
3225	 * If the vnode is doomed vgone already performed inactive processing
3226	 * (if needed).
3227	 */
3228	if (VN_IS_DOOMED(vp))
3229		goto out;
3230
3231	if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3232		goto out;
3233
3234	if (vp->v_iflag & VI_DOINGINACT)
3235		goto out;
3236
3237	/*
3238	 * Locking operations here will drop the interlock and possibly the
3239	 * vnode lock, opening a window where the vnode can get doomed all the
3240	 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3241	 * perform inactive.
3242	 */
3243	vp->v_iflag |= VI_OWEINACT;
3244	want_unlock = false;
3245	error = 0;
3246	switch (func) {
3247	case VRELE:
3248		switch (VOP_ISLOCKED(vp)) {
3249		case LK_EXCLUSIVE:
3250			break;
3251		case LK_EXCLOTHER:
3252		case 0:
3253			want_unlock = true;
3254			error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3255			VI_LOCK(vp);
3256			break;
3257		default:
3258			/*
3259			 * The lock has at least one sharer, but we have no way
3260			 * to conclude whether this is us. Play it safe and
3261			 * defer processing.
3262			 */
3263			error = EAGAIN;
3264			break;
3265		}
3266		break;
3267	case VPUT:
3268		want_unlock = true;
3269		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3270			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3271			    LK_NOWAIT);
3272			VI_LOCK(vp);
3273		}
3274		break;
3275	case VUNREF:
3276		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3277			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3278			VI_LOCK(vp);
3279		}
3280		break;
3281	}
3282	if (error == 0) {
3283		vinactive(vp);
3284		if (want_unlock)
3285			VOP_UNLOCK(vp);
3286		vdropl(vp);
3287	} else {
3288		vdefer_inactive(vp);
3289	}
3290	return;
3291out:
3292	if (func == VPUT)
3293		VOP_UNLOCK(vp);
3294	vdropl(vp);
3295}
3296
3297/*
3298 * Decrement ->v_usecount for a vnode.
3299 *
3300 * Releasing the last use count requires additional processing, see vput_final
3301 * above for details.
3302 *
3303 * Note that releasing use count without the vnode lock requires special casing
3304 * for VCHR, see v_decr_devcount for details.
3305 *
3306 * Comment above each variant denotes lock state on entry and exit.
3307 */
3308
3309static void __noinline
3310vrele_vchr(struct vnode *vp)
3311{
3312
3313	if (refcount_release_if_not_last(&vp->v_usecount))
3314		return;
3315	VI_LOCK(vp);
3316	if (!refcount_release(&vp->v_usecount)) {
3317		VI_UNLOCK(vp);
3318		return;
3319	}
3320	v_decr_devcount(vp);
3321	VI_UNLOCK(vp);
3322	vput_final(vp, VRELE);
3323}
3324
3325/*
3326 * in: any
3327 * out: same as passed in
3328 */
3329void
3330vrele(struct vnode *vp)
3331{
3332
3333	ASSERT_VI_UNLOCKED(vp, __func__);
3334	if (__predict_false(vp->v_type == VCHR)) {
3335		vrele_vchr(vp);
3336		return;
3337	}
3338	if (!refcount_release(&vp->v_usecount))
3339		return;
3340	vput_final(vp, VRELE);
3341}
3342
3343/*
3344 * in: locked
3345 * out: unlocked
3346 */
3347void
3348vput(struct vnode *vp)
3349{
3350
3351	ASSERT_VOP_LOCKED(vp, __func__);
3352	ASSERT_VI_UNLOCKED(vp, __func__);
3353	if (!refcount_release(&vp->v_usecount)) {
3354		VOP_UNLOCK(vp);
3355		return;
3356	}
3357	vput_final(vp, VPUT);
3358}
3359
3360/*
3361 * in: locked
3362 * out: locked
3363 */
3364void
3365vunref(struct vnode *vp)
3366{
3367
3368	ASSERT_VOP_LOCKED(vp, __func__);
3369	ASSERT_VI_UNLOCKED(vp, __func__);
3370	if (!refcount_release(&vp->v_usecount))
3371		return;
3372	vput_final(vp, VUNREF);
3373}
3374
3375void
3376vhold(struct vnode *vp)
3377{
3378	struct vdbatch *vd;
3379	int old;
3380
3381	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3382	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3383	VNASSERT(old >= 0, vp, ("%s: wrong hold count %d", __func__, old));
3384	if (old != 0)
3385		return;
3386	critical_enter();
3387	vd = DPCPU_PTR(vd);
3388	vd->freevnodes--;
3389	critical_exit();
3390}
3391
3392void
3393vholdl(struct vnode *vp)
3394{
3395
3396	ASSERT_VI_LOCKED(vp, __func__);
3397	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3398	vhold(vp);
3399}
3400
3401void
3402vholdnz(struct vnode *vp)
3403{
3404
3405	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3406#ifdef INVARIANTS
3407	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3408	VNASSERT(old > 0, vp, ("%s: wrong hold count %d", __func__, old));
3409#else
3410	atomic_add_int(&vp->v_holdcnt, 1);
3411#endif
3412}
3413
3414static void __noinline
3415vdbatch_process(struct vdbatch *vd)
3416{
3417	struct vnode *vp;
3418	int i;
3419
3420	mtx_assert(&vd->lock, MA_OWNED);
3421	MPASS(curthread->td_pinned > 0);
3422	MPASS(vd->index == VDBATCH_SIZE);
3423
3424	mtx_lock(&vnode_list_mtx);
3425	critical_enter();
3426	freevnodes += vd->freevnodes;
3427	for (i = 0; i < VDBATCH_SIZE; i++) {
3428		vp = vd->tab[i];
3429		TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3430		TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3431		MPASS(vp->v_dbatchcpu != NOCPU);
3432		vp->v_dbatchcpu = NOCPU;
3433	}
3434	mtx_unlock(&vnode_list_mtx);
3435	vd->freevnodes = 0;
3436	bzero(vd->tab, sizeof(vd->tab));
3437	vd->index = 0;
3438	critical_exit();
3439}
3440
3441static void
3442vdbatch_enqueue(struct vnode *vp)
3443{
3444	struct vdbatch *vd;
3445
3446	ASSERT_VI_LOCKED(vp, __func__);
3447	VNASSERT(!VN_IS_DOOMED(vp), vp,
3448	    ("%s: deferring requeue of a doomed vnode", __func__));
3449
3450	critical_enter();
3451	vd = DPCPU_PTR(vd);
3452	vd->freevnodes++;
3453	if (vp->v_dbatchcpu != NOCPU) {
3454		VI_UNLOCK(vp);
3455		critical_exit();
3456		return;
3457	}
3458
3459	sched_pin();
3460	critical_exit();
3461	mtx_lock(&vd->lock);
3462	MPASS(vd->index < VDBATCH_SIZE);
3463	MPASS(vd->tab[vd->index] == NULL);
3464	/*
3465	 * A hack: we depend on being pinned so that we know what to put in
3466	 * ->v_dbatchcpu.
3467	 */
3468	vp->v_dbatchcpu = curcpu;
3469	vd->tab[vd->index] = vp;
3470	vd->index++;
3471	VI_UNLOCK(vp);
3472	if (vd->index == VDBATCH_SIZE)
3473		vdbatch_process(vd);
3474	mtx_unlock(&vd->lock);
3475	sched_unpin();
3476}
3477
3478/*
3479 * This routine must only be called for vnodes which are about to be
3480 * deallocated. Supporting dequeue for arbitrary vndoes would require
3481 * validating that the locked batch matches.
3482 */
3483static void
3484vdbatch_dequeue(struct vnode *vp)
3485{
3486	struct vdbatch *vd;
3487	int i;
3488	short cpu;
3489
3490	VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3491	    ("%s: called for a used vnode\n", __func__));
3492
3493	cpu = vp->v_dbatchcpu;
3494	if (cpu == NOCPU)
3495		return;
3496
3497	vd = DPCPU_ID_PTR(cpu, vd);
3498	mtx_lock(&vd->lock);
3499	for (i = 0; i < vd->index; i++) {
3500		if (vd->tab[i] != vp)
3501			continue;
3502		vp->v_dbatchcpu = NOCPU;
3503		vd->index--;
3504		vd->tab[i] = vd->tab[vd->index];
3505		vd->tab[vd->index] = NULL;
3506		break;
3507	}
3508	mtx_unlock(&vd->lock);
3509	/*
3510	 * Either we dequeued the vnode above or the target CPU beat us to it.
3511	 */
3512	MPASS(vp->v_dbatchcpu == NOCPU);
3513}
3514
3515/*
3516 * Drop the hold count of the vnode.  If this is the last reference to
3517 * the vnode we place it on the free list unless it has been vgone'd
3518 * (marked VIRF_DOOMED) in which case we will free it.
3519 *
3520 * Because the vnode vm object keeps a hold reference on the vnode if
3521 * there is at least one resident non-cached page, the vnode cannot
3522 * leave the active list without the page cleanup done.
3523 */
3524static void
3525vdrop_deactivate(struct vnode *vp)
3526{
3527	struct mount *mp;
3528
3529	ASSERT_VI_LOCKED(vp, __func__);
3530	/*
3531	 * Mark a vnode as free: remove it from its active list
3532	 * and put it up for recycling on the freelist.
3533	 */
3534	VNASSERT(!VN_IS_DOOMED(vp), vp,
3535	    ("vdrop: returning doomed vnode"));
3536	VNASSERT(vp->v_op != NULL, vp,
3537	    ("vdrop: vnode already reclaimed."));
3538	VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
3539	    ("vnode with VI_OWEINACT set"));
3540	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp,
3541	    ("vnode with VI_DEFINACT set"));
3542	if (vp->v_mflag & VMP_LAZYLIST) {
3543		mp = vp->v_mount;
3544		mtx_lock(&mp->mnt_listmtx);
3545		VNASSERT(vp->v_mflag & VMP_LAZYLIST, vp, ("lost VMP_LAZYLIST"));
3546		/*
3547		 * Don't remove the vnode from the lazy list if another thread
3548		 * has increased the hold count. It may have re-enqueued the
3549		 * vnode to the lazy list and is now responsible for its
3550		 * removal.
3551		 */
3552		if (vp->v_holdcnt == 0) {
3553			vp->v_mflag &= ~VMP_LAZYLIST;
3554			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3555			mp->mnt_lazyvnodelistsize--;
3556		}
3557		mtx_unlock(&mp->mnt_listmtx);
3558	}
3559	vdbatch_enqueue(vp);
3560}
3561
3562void
3563vdrop(struct vnode *vp)
3564{
3565
3566	ASSERT_VI_UNLOCKED(vp, __func__);
3567	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3568	if (refcount_release_if_not_last(&vp->v_holdcnt))
3569		return;
3570	VI_LOCK(vp);
3571	vdropl(vp);
3572}
3573
3574void
3575vdropl(struct vnode *vp)
3576{
3577
3578	ASSERT_VI_LOCKED(vp, __func__);
3579	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3580	if (!refcount_release(&vp->v_holdcnt)) {
3581		VI_UNLOCK(vp);
3582		return;
3583	}
3584	if (VN_IS_DOOMED(vp)) {
3585		freevnode(vp);
3586		return;
3587	}
3588	vdrop_deactivate(vp);
3589}
3590
3591/*
3592 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3593 * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3594 */
3595static void
3596vinactivef(struct vnode *vp)
3597{
3598	struct vm_object *obj;
3599
3600	ASSERT_VOP_ELOCKED(vp, "vinactive");
3601	ASSERT_VI_LOCKED(vp, "vinactive");
3602	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3603	    ("vinactive: recursed on VI_DOINGINACT"));
3604	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3605	vp->v_iflag |= VI_DOINGINACT;
3606	vp->v_iflag &= ~VI_OWEINACT;
3607	VI_UNLOCK(vp);
3608	/*
3609	 * Before moving off the active list, we must be sure that any
3610	 * modified pages are converted into the vnode's dirty
3611	 * buffers, since these will no longer be checked once the
3612	 * vnode is on the inactive list.
3613	 *
3614	 * The write-out of the dirty pages is asynchronous.  At the
3615	 * point that VOP_INACTIVE() is called, there could still be
3616	 * pending I/O and dirty pages in the object.
3617	 */
3618	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3619	    vm_object_mightbedirty(obj)) {
3620		VM_OBJECT_WLOCK(obj);
3621		vm_object_page_clean(obj, 0, 0, 0);
3622		VM_OBJECT_WUNLOCK(obj);
3623	}
3624	VOP_INACTIVE(vp, curthread);
3625	VI_LOCK(vp);
3626	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3627	    ("vinactive: lost VI_DOINGINACT"));
3628	vp->v_iflag &= ~VI_DOINGINACT;
3629}
3630
3631void
3632vinactive(struct vnode *vp)
3633{
3634
3635	ASSERT_VOP_ELOCKED(vp, "vinactive");
3636	ASSERT_VI_LOCKED(vp, "vinactive");
3637	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3638
3639	if ((vp->v_iflag & VI_OWEINACT) == 0)
3640		return;
3641	if (vp->v_iflag & VI_DOINGINACT)
3642		return;
3643	if (vp->v_usecount > 0) {
3644		vp->v_iflag &= ~VI_OWEINACT;
3645		return;
3646	}
3647	vinactivef(vp);
3648}
3649
3650/*
3651 * Remove any vnodes in the vnode table belonging to mount point mp.
3652 *
3653 * If FORCECLOSE is not specified, there should not be any active ones,
3654 * return error if any are found (nb: this is a user error, not a
3655 * system error). If FORCECLOSE is specified, detach any active vnodes
3656 * that are found.
3657 *
3658 * If WRITECLOSE is set, only flush out regular file vnodes open for
3659 * writing.
3660 *
3661 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3662 *
3663 * `rootrefs' specifies the base reference count for the root vnode
3664 * of this filesystem. The root vnode is considered busy if its
3665 * v_usecount exceeds this value. On a successful return, vflush(, td)
3666 * will call vrele() on the root vnode exactly rootrefs times.
3667 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3668 * be zero.
3669 */
3670#ifdef DIAGNOSTIC
3671static int busyprt = 0;		/* print out busy vnodes */
3672SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3673#endif
3674
3675int
3676vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3677{
3678	struct vnode *vp, *mvp, *rootvp = NULL;
3679	struct vattr vattr;
3680	int busy = 0, error;
3681
3682	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3683	    rootrefs, flags);
3684	if (rootrefs > 0) {
3685		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3686		    ("vflush: bad args"));
3687		/*
3688		 * Get the filesystem root vnode. We can vput() it
3689		 * immediately, since with rootrefs > 0, it won't go away.
3690		 */
3691		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3692			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3693			    __func__, error);
3694			return (error);
3695		}
3696		vput(rootvp);
3697	}
3698loop:
3699	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3700		vholdl(vp);
3701		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3702		if (error) {
3703			vdrop(vp);
3704			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3705			goto loop;
3706		}
3707		/*
3708		 * Skip over a vnodes marked VV_SYSTEM.
3709		 */
3710		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3711			VOP_UNLOCK(vp);
3712			vdrop(vp);
3713			continue;
3714		}
3715		/*
3716		 * If WRITECLOSE is set, flush out unlinked but still open
3717		 * files (even if open only for reading) and regular file
3718		 * vnodes open for writing.
3719		 */
3720		if (flags & WRITECLOSE) {
3721			if (vp->v_object != NULL) {
3722				VM_OBJECT_WLOCK(vp->v_object);
3723				vm_object_page_clean(vp->v_object, 0, 0, 0);
3724				VM_OBJECT_WUNLOCK(vp->v_object);
3725			}
3726			error = VOP_FSYNC(vp, MNT_WAIT, td);
3727			if (error != 0) {
3728				VOP_UNLOCK(vp);
3729				vdrop(vp);
3730				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3731				return (error);
3732			}
3733			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3734			VI_LOCK(vp);
3735
3736			if ((vp->v_type == VNON ||
3737			    (error == 0 && vattr.va_nlink > 0)) &&
3738			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3739				VOP_UNLOCK(vp);
3740				vdropl(vp);
3741				continue;
3742			}
3743		} else
3744			VI_LOCK(vp);
3745		/*
3746		 * With v_usecount == 0, all we need to do is clear out the
3747		 * vnode data structures and we are done.
3748		 *
3749		 * If FORCECLOSE is set, forcibly close the vnode.
3750		 */
3751		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3752			vgonel(vp);
3753		} else {
3754			busy++;
3755#ifdef DIAGNOSTIC
3756			if (busyprt)
3757				vn_printf(vp, "vflush: busy vnode ");
3758#endif
3759		}
3760		VOP_UNLOCK(vp);
3761		vdropl(vp);
3762	}
3763	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3764		/*
3765		 * If just the root vnode is busy, and if its refcount
3766		 * is equal to `rootrefs', then go ahead and kill it.
3767		 */
3768		VI_LOCK(rootvp);
3769		KASSERT(busy > 0, ("vflush: not busy"));
3770		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3771		    ("vflush: usecount %d < rootrefs %d",
3772		     rootvp->v_usecount, rootrefs));
3773		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3774			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3775			vgone(rootvp);
3776			VOP_UNLOCK(rootvp);
3777			busy = 0;
3778		} else
3779			VI_UNLOCK(rootvp);
3780	}
3781	if (busy) {
3782		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3783		    busy);
3784		return (EBUSY);
3785	}
3786	for (; rootrefs > 0; rootrefs--)
3787		vrele(rootvp);
3788	return (0);
3789}
3790
3791/*
3792 * Recycle an unused vnode to the front of the free list.
3793 */
3794int
3795vrecycle(struct vnode *vp)
3796{
3797	int recycled;
3798
3799	VI_LOCK(vp);
3800	recycled = vrecyclel(vp);
3801	VI_UNLOCK(vp);
3802	return (recycled);
3803}
3804
3805/*
3806 * vrecycle, with the vp interlock held.
3807 */
3808int
3809vrecyclel(struct vnode *vp)
3810{
3811	int recycled;
3812
3813	ASSERT_VOP_ELOCKED(vp, __func__);
3814	ASSERT_VI_LOCKED(vp, __func__);
3815	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3816	recycled = 0;
3817	if (vp->v_usecount == 0) {
3818		recycled = 1;
3819		vgonel(vp);
3820	}
3821	return (recycled);
3822}
3823
3824/*
3825 * Eliminate all activity associated with a vnode
3826 * in preparation for reuse.
3827 */
3828void
3829vgone(struct vnode *vp)
3830{
3831	VI_LOCK(vp);
3832	vgonel(vp);
3833	VI_UNLOCK(vp);
3834}
3835
3836static void
3837notify_lowervp_vfs_dummy(struct mount *mp __unused,
3838    struct vnode *lowervp __unused)
3839{
3840}
3841
3842/*
3843 * Notify upper mounts about reclaimed or unlinked vnode.
3844 */
3845void
3846vfs_notify_upper(struct vnode *vp, int event)
3847{
3848	static struct vfsops vgonel_vfsops = {
3849		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3850		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3851	};
3852	struct mount *mp, *ump, *mmp;
3853
3854	mp = vp->v_mount;
3855	if (mp == NULL)
3856		return;
3857	if (TAILQ_EMPTY(&mp->mnt_uppers))
3858		return;
3859
3860	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3861	mmp->mnt_op = &vgonel_vfsops;
3862	mmp->mnt_kern_flag |= MNTK_MARKER;
3863	MNT_ILOCK(mp);
3864	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3865	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3866		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3867			ump = TAILQ_NEXT(ump, mnt_upper_link);
3868			continue;
3869		}
3870		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3871		MNT_IUNLOCK(mp);
3872		switch (event) {
3873		case VFS_NOTIFY_UPPER_RECLAIM:
3874			VFS_RECLAIM_LOWERVP(ump, vp);
3875			break;
3876		case VFS_NOTIFY_UPPER_UNLINK:
3877			VFS_UNLINK_LOWERVP(ump, vp);
3878			break;
3879		default:
3880			KASSERT(0, ("invalid event %d", event));
3881			break;
3882		}
3883		MNT_ILOCK(mp);
3884		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3885		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3886	}
3887	free(mmp, M_TEMP);
3888	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3889	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3890		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3891		wakeup(&mp->mnt_uppers);
3892	}
3893	MNT_IUNLOCK(mp);
3894}
3895
3896/*
3897 * vgone, with the vp interlock held.
3898 */
3899static void
3900vgonel(struct vnode *vp)
3901{
3902	struct thread *td;
3903	struct mount *mp;
3904	vm_object_t object;
3905	bool active, oweinact;
3906
3907	ASSERT_VOP_ELOCKED(vp, "vgonel");
3908	ASSERT_VI_LOCKED(vp, "vgonel");
3909	VNASSERT(vp->v_holdcnt, vp,
3910	    ("vgonel: vp %p has no reference.", vp));
3911	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3912	td = curthread;
3913
3914	/*
3915	 * Don't vgonel if we're already doomed.
3916	 */
3917	if (vp->v_irflag & VIRF_DOOMED)
3918		return;
3919	vunlazy_gone(vp);
3920	vp->v_irflag |= VIRF_DOOMED;
3921
3922	/*
3923	 * Check to see if the vnode is in use.  If so, we have to call
3924	 * VOP_CLOSE() and VOP_INACTIVE().
3925	 */
3926	active = vp->v_usecount > 0;
3927	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3928	/*
3929	 * If we need to do inactive VI_OWEINACT will be set.
3930	 */
3931	if (vp->v_iflag & VI_DEFINACT) {
3932		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3933		vp->v_iflag &= ~VI_DEFINACT;
3934		vdropl(vp);
3935	} else {
3936		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
3937		VI_UNLOCK(vp);
3938	}
3939	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3940
3941	/*
3942	 * If purging an active vnode, it must be closed and
3943	 * deactivated before being reclaimed.
3944	 */
3945	if (active)
3946		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3947	if (oweinact || active) {
3948		VI_LOCK(vp);
3949		vinactivef(vp);
3950		VI_UNLOCK(vp);
3951	}
3952	if (vp->v_type == VSOCK)
3953		vfs_unp_reclaim(vp);
3954
3955	/*
3956	 * Clean out any buffers associated with the vnode.
3957	 * If the flush fails, just toss the buffers.
3958	 */
3959	mp = NULL;
3960	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3961		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3962	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3963		while (vinvalbuf(vp, 0, 0, 0) != 0)
3964			;
3965	}
3966
3967	BO_LOCK(&vp->v_bufobj);
3968	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3969	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3970	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3971	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3972	    ("vp %p bufobj not invalidated", vp));
3973
3974	/*
3975	 * For VMIO bufobj, BO_DEAD is set later, or in
3976	 * vm_object_terminate() after the object's page queue is
3977	 * flushed.
3978	 */
3979	object = vp->v_bufobj.bo_object;
3980	if (object == NULL)
3981		vp->v_bufobj.bo_flag |= BO_DEAD;
3982	BO_UNLOCK(&vp->v_bufobj);
3983
3984	/*
3985	 * Handle the VM part.  Tmpfs handles v_object on its own (the
3986	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
3987	 * should not touch the object borrowed from the lower vnode
3988	 * (the handle check).
3989	 */
3990	if (object != NULL && object->type == OBJT_VNODE &&
3991	    object->handle == vp)
3992		vnode_destroy_vobject(vp);
3993
3994	/*
3995	 * Reclaim the vnode.
3996	 */
3997	if (VOP_RECLAIM(vp, td))
3998		panic("vgone: cannot reclaim");
3999	if (mp != NULL)
4000		vn_finished_secondary_write(mp);
4001	VNASSERT(vp->v_object == NULL, vp,
4002	    ("vop_reclaim left v_object vp=%p", vp));
4003	/*
4004	 * Clear the advisory locks and wake up waiting threads.
4005	 */
4006	(void)VOP_ADVLOCKPURGE(vp);
4007	vp->v_lockf = NULL;
4008	/*
4009	 * Delete from old mount point vnode list.
4010	 */
4011	delmntque(vp);
4012	cache_purge(vp);
4013	/*
4014	 * Done with purge, reset to the standard lock and invalidate
4015	 * the vnode.
4016	 */
4017	VI_LOCK(vp);
4018	vp->v_vnlock = &vp->v_lock;
4019	vp->v_op = &dead_vnodeops;
4020	vp->v_type = VBAD;
4021}
4022
4023/*
4024 * Calculate the total number of references to a special device.
4025 */
4026int
4027vcount(struct vnode *vp)
4028{
4029	int count;
4030
4031	dev_lock();
4032	count = vp->v_rdev->si_usecount;
4033	dev_unlock();
4034	return (count);
4035}
4036
4037/*
4038 * Print out a description of a vnode.
4039 */
4040static const char * const typename[] =
4041{"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
4042 "VMARKER"};
4043
4044void
4045vn_printf(struct vnode *vp, const char *fmt, ...)
4046{
4047	va_list ap;
4048	char buf[256], buf2[16];
4049	u_long flags;
4050
4051	va_start(ap, fmt);
4052	vprintf(fmt, ap);
4053	va_end(ap);
4054	printf("%p: ", (void *)vp);
4055	printf("type %s\n", typename[vp->v_type]);
4056	printf("    usecount %d, writecount %d, refcount %d",
4057	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
4058	switch (vp->v_type) {
4059	case VDIR:
4060		printf(" mountedhere %p\n", vp->v_mountedhere);
4061		break;
4062	case VCHR:
4063		printf(" rdev %p\n", vp->v_rdev);
4064		break;
4065	case VSOCK:
4066		printf(" socket %p\n", vp->v_unpcb);
4067		break;
4068	case VFIFO:
4069		printf(" fifoinfo %p\n", vp->v_fifoinfo);
4070		break;
4071	default:
4072		printf("\n");
4073		break;
4074	}
4075	buf[0] = '\0';
4076	buf[1] = '\0';
4077	if (vp->v_irflag & VIRF_DOOMED)
4078		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4079	flags = vp->v_irflag & ~(VIRF_DOOMED);
4080	if (flags != 0) {
4081		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4082		strlcat(buf, buf2, sizeof(buf));
4083	}
4084	if (vp->v_vflag & VV_ROOT)
4085		strlcat(buf, "|VV_ROOT", sizeof(buf));
4086	if (vp->v_vflag & VV_ISTTY)
4087		strlcat(buf, "|VV_ISTTY", sizeof(buf));
4088	if (vp->v_vflag & VV_NOSYNC)
4089		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4090	if (vp->v_vflag & VV_ETERNALDEV)
4091		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4092	if (vp->v_vflag & VV_CACHEDLABEL)
4093		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4094	if (vp->v_vflag & VV_VMSIZEVNLOCK)
4095		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4096	if (vp->v_vflag & VV_COPYONWRITE)
4097		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4098	if (vp->v_vflag & VV_SYSTEM)
4099		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4100	if (vp->v_vflag & VV_PROCDEP)
4101		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4102	if (vp->v_vflag & VV_NOKNOTE)
4103		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
4104	if (vp->v_vflag & VV_DELETED)
4105		strlcat(buf, "|VV_DELETED", sizeof(buf));
4106	if (vp->v_vflag & VV_MD)
4107		strlcat(buf, "|VV_MD", sizeof(buf));
4108	if (vp->v_vflag & VV_FORCEINSMQ)
4109		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4110	if (vp->v_vflag & VV_READLINK)
4111		strlcat(buf, "|VV_READLINK", sizeof(buf));
4112	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4113	    VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
4114	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
4115	if (flags != 0) {
4116		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4117		strlcat(buf, buf2, sizeof(buf));
4118	}
4119	if (vp->v_iflag & VI_TEXT_REF)
4120		strlcat(buf, "|VI_TEXT_REF", sizeof(buf));
4121	if (vp->v_iflag & VI_MOUNT)
4122		strlcat(buf, "|VI_MOUNT", sizeof(buf));
4123	if (vp->v_iflag & VI_DOINGINACT)
4124		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4125	if (vp->v_iflag & VI_OWEINACT)
4126		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4127	if (vp->v_iflag & VI_DEFINACT)
4128		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4129	flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_DOINGINACT |
4130	    VI_OWEINACT | VI_DEFINACT);
4131	if (flags != 0) {
4132		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4133		strlcat(buf, buf2, sizeof(buf));
4134	}
4135	if (vp->v_mflag & VMP_LAZYLIST)
4136		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4137	flags = vp->v_mflag & ~(VMP_LAZYLIST);
4138	if (flags != 0) {
4139		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4140		strlcat(buf, buf2, sizeof(buf));
4141	}
4142	printf("    flags (%s)\n", buf + 1);
4143	if (mtx_owned(VI_MTX(vp)))
4144		printf(" VI_LOCKed");
4145	if (vp->v_object != NULL)
4146		printf("    v_object %p ref %d pages %d "
4147		    "cleanbuf %d dirtybuf %d\n",
4148		    vp->v_object, vp->v_object->ref_count,
4149		    vp->v_object->resident_page_count,
4150		    vp->v_bufobj.bo_clean.bv_cnt,
4151		    vp->v_bufobj.bo_dirty.bv_cnt);
4152	printf("    ");
4153	lockmgr_printinfo(vp->v_vnlock);
4154	if (vp->v_data != NULL)
4155		VOP_PRINT(vp);
4156}
4157
4158#ifdef DDB
4159/*
4160 * List all of the locked vnodes in the system.
4161 * Called when debugging the kernel.
4162 */
4163DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
4164{
4165	struct mount *mp;
4166	struct vnode *vp;
4167
4168	/*
4169	 * Note: because this is DDB, we can't obey the locking semantics
4170	 * for these structures, which means we could catch an inconsistent
4171	 * state and dereference a nasty pointer.  Not much to be done
4172	 * about that.
4173	 */
4174	db_printf("Locked vnodes\n");
4175	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4176		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4177			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4178				vn_printf(vp, "vnode ");
4179		}
4180	}
4181}
4182
4183/*
4184 * Show details about the given vnode.
4185 */
4186DB_SHOW_COMMAND(vnode, db_show_vnode)
4187{
4188	struct vnode *vp;
4189
4190	if (!have_addr)
4191		return;
4192	vp = (struct vnode *)addr;
4193	vn_printf(vp, "vnode ");
4194}
4195
4196/*
4197 * Show details about the given mount point.
4198 */
4199DB_SHOW_COMMAND(mount, db_show_mount)
4200{
4201	struct mount *mp;
4202	struct vfsopt *opt;
4203	struct statfs *sp;
4204	struct vnode *vp;
4205	char buf[512];
4206	uint64_t mflags;
4207	u_int flags;
4208
4209	if (!have_addr) {
4210		/* No address given, print short info about all mount points. */
4211		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4212			db_printf("%p %s on %s (%s)\n", mp,
4213			    mp->mnt_stat.f_mntfromname,
4214			    mp->mnt_stat.f_mntonname,
4215			    mp->mnt_stat.f_fstypename);
4216			if (db_pager_quit)
4217				break;
4218		}
4219		db_printf("\nMore info: show mount <addr>\n");
4220		return;
4221	}
4222
4223	mp = (struct mount *)addr;
4224	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4225	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4226
4227	buf[0] = '\0';
4228	mflags = mp->mnt_flag;
4229#define	MNT_FLAG(flag)	do {						\
4230	if (mflags & (flag)) {						\
4231		if (buf[0] != '\0')					\
4232			strlcat(buf, ", ", sizeof(buf));		\
4233		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4234		mflags &= ~(flag);					\
4235	}								\
4236} while (0)
4237	MNT_FLAG(MNT_RDONLY);
4238	MNT_FLAG(MNT_SYNCHRONOUS);
4239	MNT_FLAG(MNT_NOEXEC);
4240	MNT_FLAG(MNT_NOSUID);
4241	MNT_FLAG(MNT_NFS4ACLS);
4242	MNT_FLAG(MNT_UNION);
4243	MNT_FLAG(MNT_ASYNC);
4244	MNT_FLAG(MNT_SUIDDIR);
4245	MNT_FLAG(MNT_SOFTDEP);
4246	MNT_FLAG(MNT_NOSYMFOLLOW);
4247	MNT_FLAG(MNT_GJOURNAL);
4248	MNT_FLAG(MNT_MULTILABEL);
4249	MNT_FLAG(MNT_ACLS);
4250	MNT_FLAG(MNT_NOATIME);
4251	MNT_FLAG(MNT_NOCLUSTERR);
4252	MNT_FLAG(MNT_NOCLUSTERW);
4253	MNT_FLAG(MNT_SUJ);
4254	MNT_FLAG(MNT_EXRDONLY);
4255	MNT_FLAG(MNT_EXPORTED);
4256	MNT_FLAG(MNT_DEFEXPORTED);
4257	MNT_FLAG(MNT_EXPORTANON);
4258	MNT_FLAG(MNT_EXKERB);
4259	MNT_FLAG(MNT_EXPUBLIC);
4260	MNT_FLAG(MNT_LOCAL);
4261	MNT_FLAG(MNT_QUOTA);
4262	MNT_FLAG(MNT_ROOTFS);
4263	MNT_FLAG(MNT_USER);
4264	MNT_FLAG(MNT_IGNORE);
4265	MNT_FLAG(MNT_UPDATE);
4266	MNT_FLAG(MNT_DELEXPORT);
4267	MNT_FLAG(MNT_RELOAD);
4268	MNT_FLAG(MNT_FORCE);
4269	MNT_FLAG(MNT_SNAPSHOT);
4270	MNT_FLAG(MNT_BYFSID);
4271#undef MNT_FLAG
4272	if (mflags != 0) {
4273		if (buf[0] != '\0')
4274			strlcat(buf, ", ", sizeof(buf));
4275		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4276		    "0x%016jx", mflags);
4277	}
4278	db_printf("    mnt_flag = %s\n", buf);
4279
4280	buf[0] = '\0';
4281	flags = mp->mnt_kern_flag;
4282#define	MNT_KERN_FLAG(flag)	do {					\
4283	if (flags & (flag)) {						\
4284		if (buf[0] != '\0')					\
4285			strlcat(buf, ", ", sizeof(buf));		\
4286		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4287		flags &= ~(flag);					\
4288	}								\
4289} while (0)
4290	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4291	MNT_KERN_FLAG(MNTK_ASYNC);
4292	MNT_KERN_FLAG(MNTK_SOFTDEP);
4293	MNT_KERN_FLAG(MNTK_DRAINING);
4294	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4295	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4296	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4297	MNT_KERN_FLAG(MNTK_NO_IOPF);
4298	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
4299	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
4300	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
4301	MNT_KERN_FLAG(MNTK_MARKER);
4302	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4303	MNT_KERN_FLAG(MNTK_NOASYNC);
4304	MNT_KERN_FLAG(MNTK_UNMOUNT);
4305	MNT_KERN_FLAG(MNTK_MWAIT);
4306	MNT_KERN_FLAG(MNTK_SUSPEND);
4307	MNT_KERN_FLAG(MNTK_SUSPEND2);
4308	MNT_KERN_FLAG(MNTK_SUSPENDED);
4309	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4310	MNT_KERN_FLAG(MNTK_NOKNOTE);
4311#undef MNT_KERN_FLAG
4312	if (flags != 0) {
4313		if (buf[0] != '\0')
4314			strlcat(buf, ", ", sizeof(buf));
4315		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4316		    "0x%08x", flags);
4317	}
4318	db_printf("    mnt_kern_flag = %s\n", buf);
4319
4320	db_printf("    mnt_opt = ");
4321	opt = TAILQ_FIRST(mp->mnt_opt);
4322	if (opt != NULL) {
4323		db_printf("%s", opt->name);
4324		opt = TAILQ_NEXT(opt, link);
4325		while (opt != NULL) {
4326			db_printf(", %s", opt->name);
4327			opt = TAILQ_NEXT(opt, link);
4328		}
4329	}
4330	db_printf("\n");
4331
4332	sp = &mp->mnt_stat;
4333	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4334	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4335	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4336	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4337	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4338	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4339	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4340	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4341	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4342	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4343	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4344	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4345
4346	db_printf("    mnt_cred = { uid=%u ruid=%u",
4347	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4348	if (jailed(mp->mnt_cred))
4349		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4350	db_printf(" }\n");
4351	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4352	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4353	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4354	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4355	db_printf("    mnt_lazyvnodelistsize = %d\n",
4356	    mp->mnt_lazyvnodelistsize);
4357	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4358	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4359	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
4360	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4361	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4362	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4363	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4364	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4365	db_printf("    mnt_secondary_accwrites = %d\n",
4366	    mp->mnt_secondary_accwrites);
4367	db_printf("    mnt_gjprovider = %s\n",
4368	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4369	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4370
4371	db_printf("\n\nList of active vnodes\n");
4372	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4373		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4374			vn_printf(vp, "vnode ");
4375			if (db_pager_quit)
4376				break;
4377		}
4378	}
4379	db_printf("\n\nList of inactive vnodes\n");
4380	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4381		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4382			vn_printf(vp, "vnode ");
4383			if (db_pager_quit)
4384				break;
4385		}
4386	}
4387}
4388#endif	/* DDB */
4389
4390/*
4391 * Fill in a struct xvfsconf based on a struct vfsconf.
4392 */
4393static int
4394vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4395{
4396	struct xvfsconf xvfsp;
4397
4398	bzero(&xvfsp, sizeof(xvfsp));
4399	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4400	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4401	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4402	xvfsp.vfc_flags = vfsp->vfc_flags;
4403	/*
4404	 * These are unused in userland, we keep them
4405	 * to not break binary compatibility.
4406	 */
4407	xvfsp.vfc_vfsops = NULL;
4408	xvfsp.vfc_next = NULL;
4409	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4410}
4411
4412#ifdef COMPAT_FREEBSD32
4413struct xvfsconf32 {
4414	uint32_t	vfc_vfsops;
4415	char		vfc_name[MFSNAMELEN];
4416	int32_t		vfc_typenum;
4417	int32_t		vfc_refcount;
4418	int32_t		vfc_flags;
4419	uint32_t	vfc_next;
4420};
4421
4422static int
4423vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4424{
4425	struct xvfsconf32 xvfsp;
4426
4427	bzero(&xvfsp, sizeof(xvfsp));
4428	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4429	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4430	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4431	xvfsp.vfc_flags = vfsp->vfc_flags;
4432	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4433}
4434#endif
4435
4436/*
4437 * Top level filesystem related information gathering.
4438 */
4439static int
4440sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4441{
4442	struct vfsconf *vfsp;
4443	int error;
4444
4445	error = 0;
4446	vfsconf_slock();
4447	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4448#ifdef COMPAT_FREEBSD32
4449		if (req->flags & SCTL_MASK32)
4450			error = vfsconf2x32(req, vfsp);
4451		else
4452#endif
4453			error = vfsconf2x(req, vfsp);
4454		if (error)
4455			break;
4456	}
4457	vfsconf_sunlock();
4458	return (error);
4459}
4460
4461SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4462    CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4463    "S,xvfsconf", "List of all configured filesystems");
4464
4465#ifndef BURN_BRIDGES
4466static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4467
4468static int
4469vfs_sysctl(SYSCTL_HANDLER_ARGS)
4470{
4471	int *name = (int *)arg1 - 1;	/* XXX */
4472	u_int namelen = arg2 + 1;	/* XXX */
4473	struct vfsconf *vfsp;
4474
4475	log(LOG_WARNING, "userland calling deprecated sysctl, "
4476	    "please rebuild world\n");
4477
4478#if 1 || defined(COMPAT_PRELITE2)
4479	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4480	if (namelen == 1)
4481		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4482#endif
4483
4484	switch (name[1]) {
4485	case VFS_MAXTYPENUM:
4486		if (namelen != 2)
4487			return (ENOTDIR);
4488		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4489	case VFS_CONF:
4490		if (namelen != 3)
4491			return (ENOTDIR);	/* overloaded */
4492		vfsconf_slock();
4493		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4494			if (vfsp->vfc_typenum == name[2])
4495				break;
4496		}
4497		vfsconf_sunlock();
4498		if (vfsp == NULL)
4499			return (EOPNOTSUPP);
4500#ifdef COMPAT_FREEBSD32
4501		if (req->flags & SCTL_MASK32)
4502			return (vfsconf2x32(req, vfsp));
4503		else
4504#endif
4505			return (vfsconf2x(req, vfsp));
4506	}
4507	return (EOPNOTSUPP);
4508}
4509
4510static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4511    CTLFLAG_MPSAFE, vfs_sysctl,
4512    "Generic filesystem");
4513
4514#if 1 || defined(COMPAT_PRELITE2)
4515
4516static int
4517sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4518{
4519	int error;
4520	struct vfsconf *vfsp;
4521	struct ovfsconf ovfs;
4522
4523	vfsconf_slock();
4524	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4525		bzero(&ovfs, sizeof(ovfs));
4526		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4527		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4528		ovfs.vfc_index = vfsp->vfc_typenum;
4529		ovfs.vfc_refcount = vfsp->vfc_refcount;
4530		ovfs.vfc_flags = vfsp->vfc_flags;
4531		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4532		if (error != 0) {
4533			vfsconf_sunlock();
4534			return (error);
4535		}
4536	}
4537	vfsconf_sunlock();
4538	return (0);
4539}
4540
4541#endif /* 1 || COMPAT_PRELITE2 */
4542#endif /* !BURN_BRIDGES */
4543
4544#define KINFO_VNODESLOP		10
4545#ifdef notyet
4546/*
4547 * Dump vnode list (via sysctl).
4548 */
4549/* ARGSUSED */
4550static int
4551sysctl_vnode(SYSCTL_HANDLER_ARGS)
4552{
4553	struct xvnode *xvn;
4554	struct mount *mp;
4555	struct vnode *vp;
4556	int error, len, n;
4557
4558	/*
4559	 * Stale numvnodes access is not fatal here.
4560	 */
4561	req->lock = 0;
4562	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4563	if (!req->oldptr)
4564		/* Make an estimate */
4565		return (SYSCTL_OUT(req, 0, len));
4566
4567	error = sysctl_wire_old_buffer(req, 0);
4568	if (error != 0)
4569		return (error);
4570	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4571	n = 0;
4572	mtx_lock(&mountlist_mtx);
4573	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4574		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4575			continue;
4576		MNT_ILOCK(mp);
4577		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4578			if (n == len)
4579				break;
4580			vref(vp);
4581			xvn[n].xv_size = sizeof *xvn;
4582			xvn[n].xv_vnode = vp;
4583			xvn[n].xv_id = 0;	/* XXX compat */
4584#define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4585			XV_COPY(usecount);
4586			XV_COPY(writecount);
4587			XV_COPY(holdcnt);
4588			XV_COPY(mount);
4589			XV_COPY(numoutput);
4590			XV_COPY(type);
4591#undef XV_COPY
4592			xvn[n].xv_flag = vp->v_vflag;
4593
4594			switch (vp->v_type) {
4595			case VREG:
4596			case VDIR:
4597			case VLNK:
4598				break;
4599			case VBLK:
4600			case VCHR:
4601				if (vp->v_rdev == NULL) {
4602					vrele(vp);
4603					continue;
4604				}
4605				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4606				break;
4607			case VSOCK:
4608				xvn[n].xv_socket = vp->v_socket;
4609				break;
4610			case VFIFO:
4611				xvn[n].xv_fifo = vp->v_fifoinfo;
4612				break;
4613			case VNON:
4614			case VBAD:
4615			default:
4616				/* shouldn't happen? */
4617				vrele(vp);
4618				continue;
4619			}
4620			vrele(vp);
4621			++n;
4622		}
4623		MNT_IUNLOCK(mp);
4624		mtx_lock(&mountlist_mtx);
4625		vfs_unbusy(mp);
4626		if (n == len)
4627			break;
4628	}
4629	mtx_unlock(&mountlist_mtx);
4630
4631	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4632	free(xvn, M_TEMP);
4633	return (error);
4634}
4635
4636SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4637    CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4638    "");
4639#endif
4640
4641static void
4642unmount_or_warn(struct mount *mp)
4643{
4644	int error;
4645
4646	error = dounmount(mp, MNT_FORCE, curthread);
4647	if (error != 0) {
4648		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4649		if (error == EBUSY)
4650			printf("BUSY)\n");
4651		else
4652			printf("%d)\n", error);
4653	}
4654}
4655
4656/*
4657 * Unmount all filesystems. The list is traversed in reverse order
4658 * of mounting to avoid dependencies.
4659 */
4660void
4661vfs_unmountall(void)
4662{
4663	struct mount *mp, *tmp;
4664
4665	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4666
4667	/*
4668	 * Since this only runs when rebooting, it is not interlocked.
4669	 */
4670	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4671		vfs_ref(mp);
4672
4673		/*
4674		 * Forcibly unmounting "/dev" before "/" would prevent clean
4675		 * unmount of the latter.
4676		 */
4677		if (mp == rootdevmp)
4678			continue;
4679
4680		unmount_or_warn(mp);
4681	}
4682
4683	if (rootdevmp != NULL)
4684		unmount_or_warn(rootdevmp);
4685}
4686
4687static void
4688vfs_deferred_inactive(struct vnode *vp, int lkflags)
4689{
4690
4691	ASSERT_VI_LOCKED(vp, __func__);
4692	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4693	if ((vp->v_iflag & VI_OWEINACT) == 0) {
4694		vdropl(vp);
4695		return;
4696	}
4697	if (vn_lock(vp, lkflags) == 0) {
4698		VI_LOCK(vp);
4699		vinactive(vp);
4700		VOP_UNLOCK(vp);
4701		vdropl(vp);
4702		return;
4703	}
4704	vdefer_inactive_unlocked(vp);
4705}
4706
4707static int
4708vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4709{
4710
4711	return (vp->v_iflag & VI_DEFINACT);
4712}
4713
4714static void __noinline
4715vfs_periodic_inactive(struct mount *mp, int flags)
4716{
4717	struct vnode *vp, *mvp;
4718	int lkflags;
4719
4720	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4721	if (flags != MNT_WAIT)
4722		lkflags |= LK_NOWAIT;
4723
4724	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4725		if ((vp->v_iflag & VI_DEFINACT) == 0) {
4726			VI_UNLOCK(vp);
4727			continue;
4728		}
4729		vp->v_iflag &= ~VI_DEFINACT;
4730		vfs_deferred_inactive(vp, lkflags);
4731	}
4732}
4733
4734static inline bool
4735vfs_want_msync(struct vnode *vp)
4736{
4737	struct vm_object *obj;
4738
4739	/*
4740	 * This test may be performed without any locks held.
4741	 * We rely on vm_object's type stability.
4742	 */
4743	if (vp->v_vflag & VV_NOSYNC)
4744		return (false);
4745	obj = vp->v_object;
4746	return (obj != NULL && vm_object_mightbedirty(obj));
4747}
4748
4749static int
4750vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4751{
4752
4753	if (vp->v_vflag & VV_NOSYNC)
4754		return (false);
4755	if (vp->v_iflag & VI_DEFINACT)
4756		return (true);
4757	return (vfs_want_msync(vp));
4758}
4759
4760static void __noinline
4761vfs_periodic_msync_inactive(struct mount *mp, int flags)
4762{
4763	struct vnode *vp, *mvp;
4764	struct vm_object *obj;
4765	struct thread *td;
4766	int lkflags, objflags;
4767	bool seen_defer;
4768
4769	td = curthread;
4770
4771	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4772	if (flags != MNT_WAIT) {
4773		lkflags |= LK_NOWAIT;
4774		objflags = OBJPC_NOSYNC;
4775	} else {
4776		objflags = OBJPC_SYNC;
4777	}
4778
4779	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4780		seen_defer = false;
4781		if (vp->v_iflag & VI_DEFINACT) {
4782			vp->v_iflag &= ~VI_DEFINACT;
4783			seen_defer = true;
4784		}
4785		if (!vfs_want_msync(vp)) {
4786			if (seen_defer)
4787				vfs_deferred_inactive(vp, lkflags);
4788			else
4789				VI_UNLOCK(vp);
4790			continue;
4791		}
4792		if (vget(vp, lkflags, td) == 0) {
4793			obj = vp->v_object;
4794			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4795				VM_OBJECT_WLOCK(obj);
4796				vm_object_page_clean(obj, 0, 0, objflags);
4797				VM_OBJECT_WUNLOCK(obj);
4798			}
4799			vput(vp);
4800			if (seen_defer)
4801				vdrop(vp);
4802		} else {
4803			if (seen_defer)
4804				vdefer_inactive_unlocked(vp);
4805		}
4806	}
4807}
4808
4809void
4810vfs_periodic(struct mount *mp, int flags)
4811{
4812
4813	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4814
4815	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4816		vfs_periodic_inactive(mp, flags);
4817	else
4818		vfs_periodic_msync_inactive(mp, flags);
4819}
4820
4821static void
4822destroy_vpollinfo_free(struct vpollinfo *vi)
4823{
4824
4825	knlist_destroy(&vi->vpi_selinfo.si_note);
4826	mtx_destroy(&vi->vpi_lock);
4827	uma_zfree(vnodepoll_zone, vi);
4828}
4829
4830static void
4831destroy_vpollinfo(struct vpollinfo *vi)
4832{
4833
4834	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4835	seldrain(&vi->vpi_selinfo);
4836	destroy_vpollinfo_free(vi);
4837}
4838
4839/*
4840 * Initialize per-vnode helper structure to hold poll-related state.
4841 */
4842void
4843v_addpollinfo(struct vnode *vp)
4844{
4845	struct vpollinfo *vi;
4846
4847	if (vp->v_pollinfo != NULL)
4848		return;
4849	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4850	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4851	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4852	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4853	VI_LOCK(vp);
4854	if (vp->v_pollinfo != NULL) {
4855		VI_UNLOCK(vp);
4856		destroy_vpollinfo_free(vi);
4857		return;
4858	}
4859	vp->v_pollinfo = vi;
4860	VI_UNLOCK(vp);
4861}
4862
4863/*
4864 * Record a process's interest in events which might happen to
4865 * a vnode.  Because poll uses the historic select-style interface
4866 * internally, this routine serves as both the ``check for any
4867 * pending events'' and the ``record my interest in future events''
4868 * functions.  (These are done together, while the lock is held,
4869 * to avoid race conditions.)
4870 */
4871int
4872vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4873{
4874
4875	v_addpollinfo(vp);
4876	mtx_lock(&vp->v_pollinfo->vpi_lock);
4877	if (vp->v_pollinfo->vpi_revents & events) {
4878		/*
4879		 * This leaves events we are not interested
4880		 * in available for the other process which
4881		 * which presumably had requested them
4882		 * (otherwise they would never have been
4883		 * recorded).
4884		 */
4885		events &= vp->v_pollinfo->vpi_revents;
4886		vp->v_pollinfo->vpi_revents &= ~events;
4887
4888		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4889		return (events);
4890	}
4891	vp->v_pollinfo->vpi_events |= events;
4892	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4893	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4894	return (0);
4895}
4896
4897/*
4898 * Routine to create and manage a filesystem syncer vnode.
4899 */
4900#define sync_close ((int (*)(struct  vop_close_args *))nullop)
4901static int	sync_fsync(struct  vop_fsync_args *);
4902static int	sync_inactive(struct  vop_inactive_args *);
4903static int	sync_reclaim(struct  vop_reclaim_args *);
4904
4905static struct vop_vector sync_vnodeops = {
4906	.vop_bypass =	VOP_EOPNOTSUPP,
4907	.vop_close =	sync_close,		/* close */
4908	.vop_fsync =	sync_fsync,		/* fsync */
4909	.vop_inactive =	sync_inactive,	/* inactive */
4910	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4911	.vop_reclaim =	sync_reclaim,	/* reclaim */
4912	.vop_lock1 =	vop_stdlock,	/* lock */
4913	.vop_unlock =	vop_stdunlock,	/* unlock */
4914	.vop_islocked =	vop_stdislocked,	/* islocked */
4915};
4916VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4917
4918/*
4919 * Create a new filesystem syncer vnode for the specified mount point.
4920 */
4921void
4922vfs_allocate_syncvnode(struct mount *mp)
4923{
4924	struct vnode *vp;
4925	struct bufobj *bo;
4926	static long start, incr, next;
4927	int error;
4928
4929	/* Allocate a new vnode */
4930	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4931	if (error != 0)
4932		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4933	vp->v_type = VNON;
4934	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4935	vp->v_vflag |= VV_FORCEINSMQ;
4936	error = insmntque(vp, mp);
4937	if (error != 0)
4938		panic("vfs_allocate_syncvnode: insmntque() failed");
4939	vp->v_vflag &= ~VV_FORCEINSMQ;
4940	VOP_UNLOCK(vp);
4941	/*
4942	 * Place the vnode onto the syncer worklist. We attempt to
4943	 * scatter them about on the list so that they will go off
4944	 * at evenly distributed times even if all the filesystems
4945	 * are mounted at once.
4946	 */
4947	next += incr;
4948	if (next == 0 || next > syncer_maxdelay) {
4949		start /= 2;
4950		incr /= 2;
4951		if (start == 0) {
4952			start = syncer_maxdelay / 2;
4953			incr = syncer_maxdelay;
4954		}
4955		next = start;
4956	}
4957	bo = &vp->v_bufobj;
4958	BO_LOCK(bo);
4959	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4960	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4961	mtx_lock(&sync_mtx);
4962	sync_vnode_count++;
4963	if (mp->mnt_syncer == NULL) {
4964		mp->mnt_syncer = vp;
4965		vp = NULL;
4966	}
4967	mtx_unlock(&sync_mtx);
4968	BO_UNLOCK(bo);
4969	if (vp != NULL) {
4970		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4971		vgone(vp);
4972		vput(vp);
4973	}
4974}
4975
4976void
4977vfs_deallocate_syncvnode(struct mount *mp)
4978{
4979	struct vnode *vp;
4980
4981	mtx_lock(&sync_mtx);
4982	vp = mp->mnt_syncer;
4983	if (vp != NULL)
4984		mp->mnt_syncer = NULL;
4985	mtx_unlock(&sync_mtx);
4986	if (vp != NULL)
4987		vrele(vp);
4988}
4989
4990/*
4991 * Do a lazy sync of the filesystem.
4992 */
4993static int
4994sync_fsync(struct vop_fsync_args *ap)
4995{
4996	struct vnode *syncvp = ap->a_vp;
4997	struct mount *mp = syncvp->v_mount;
4998	int error, save;
4999	struct bufobj *bo;
5000
5001	/*
5002	 * We only need to do something if this is a lazy evaluation.
5003	 */
5004	if (ap->a_waitfor != MNT_LAZY)
5005		return (0);
5006
5007	/*
5008	 * Move ourselves to the back of the sync list.
5009	 */
5010	bo = &syncvp->v_bufobj;
5011	BO_LOCK(bo);
5012	vn_syncer_add_to_worklist(bo, syncdelay);
5013	BO_UNLOCK(bo);
5014
5015	/*
5016	 * Walk the list of vnodes pushing all that are dirty and
5017	 * not already on the sync list.
5018	 */
5019	if (vfs_busy(mp, MBF_NOWAIT) != 0)
5020		return (0);
5021	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
5022		vfs_unbusy(mp);
5023		return (0);
5024	}
5025	save = curthread_pflags_set(TDP_SYNCIO);
5026	/*
5027	 * The filesystem at hand may be idle with free vnodes stored in the
5028	 * batch.  Return them instead of letting them stay there indefinitely.
5029	 */
5030	vfs_periodic(mp, MNT_NOWAIT);
5031	error = VFS_SYNC(mp, MNT_LAZY);
5032	curthread_pflags_restore(save);
5033	vn_finished_write(mp);
5034	vfs_unbusy(mp);
5035	return (error);
5036}
5037
5038/*
5039 * The syncer vnode is no referenced.
5040 */
5041static int
5042sync_inactive(struct vop_inactive_args *ap)
5043{
5044
5045	vgone(ap->a_vp);
5046	return (0);
5047}
5048
5049/*
5050 * The syncer vnode is no longer needed and is being decommissioned.
5051 *
5052 * Modifications to the worklist must be protected by sync_mtx.
5053 */
5054static int
5055sync_reclaim(struct vop_reclaim_args *ap)
5056{
5057	struct vnode *vp = ap->a_vp;
5058	struct bufobj *bo;
5059
5060	bo = &vp->v_bufobj;
5061	BO_LOCK(bo);
5062	mtx_lock(&sync_mtx);
5063	if (vp->v_mount->mnt_syncer == vp)
5064		vp->v_mount->mnt_syncer = NULL;
5065	if (bo->bo_flag & BO_ONWORKLST) {
5066		LIST_REMOVE(bo, bo_synclist);
5067		syncer_worklist_len--;
5068		sync_vnode_count--;
5069		bo->bo_flag &= ~BO_ONWORKLST;
5070	}
5071	mtx_unlock(&sync_mtx);
5072	BO_UNLOCK(bo);
5073
5074	return (0);
5075}
5076
5077int
5078vn_need_pageq_flush(struct vnode *vp)
5079{
5080	struct vm_object *obj;
5081	int need;
5082
5083	MPASS(mtx_owned(VI_MTX(vp)));
5084	need = 0;
5085	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5086	    vm_object_mightbedirty(obj))
5087		need = 1;
5088	return (need);
5089}
5090
5091/*
5092 * Check if vnode represents a disk device
5093 */
5094int
5095vn_isdisk(struct vnode *vp, int *errp)
5096{
5097	int error;
5098
5099	if (vp->v_type != VCHR) {
5100		error = ENOTBLK;
5101		goto out;
5102	}
5103	error = 0;
5104	dev_lock();
5105	if (vp->v_rdev == NULL)
5106		error = ENXIO;
5107	else if (vp->v_rdev->si_devsw == NULL)
5108		error = ENXIO;
5109	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5110		error = ENOTBLK;
5111	dev_unlock();
5112out:
5113	if (errp != NULL)
5114		*errp = error;
5115	return (error == 0);
5116}
5117
5118/*
5119 * Common filesystem object access control check routine.  Accepts a
5120 * vnode's type, "mode", uid and gid, requested access mode, credentials,
5121 * and optional call-by-reference privused argument allowing vaccess()
5122 * to indicate to the caller whether privilege was used to satisfy the
5123 * request (obsoleted).  Returns 0 on success, or an errno on failure.
5124 */
5125int
5126vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5127    accmode_t accmode, struct ucred *cred, int *privused)
5128{
5129	accmode_t dac_granted;
5130	accmode_t priv_granted;
5131
5132	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5133	    ("invalid bit in accmode"));
5134	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5135	    ("VAPPEND without VWRITE"));
5136
5137	/*
5138	 * Look for a normal, non-privileged way to access the file/directory
5139	 * as requested.  If it exists, go with that.
5140	 */
5141
5142	if (privused != NULL)
5143		*privused = 0;
5144
5145	dac_granted = 0;
5146
5147	/* Check the owner. */
5148	if (cred->cr_uid == file_uid) {
5149		dac_granted |= VADMIN;
5150		if (file_mode & S_IXUSR)
5151			dac_granted |= VEXEC;
5152		if (file_mode & S_IRUSR)
5153			dac_granted |= VREAD;
5154		if (file_mode & S_IWUSR)
5155			dac_granted |= (VWRITE | VAPPEND);
5156
5157		if ((accmode & dac_granted) == accmode)
5158			return (0);
5159
5160		goto privcheck;
5161	}
5162
5163	/* Otherwise, check the groups (first match) */
5164	if (groupmember(file_gid, cred)) {
5165		if (file_mode & S_IXGRP)
5166			dac_granted |= VEXEC;
5167		if (file_mode & S_IRGRP)
5168			dac_granted |= VREAD;
5169		if (file_mode & S_IWGRP)
5170			dac_granted |= (VWRITE | VAPPEND);
5171
5172		if ((accmode & dac_granted) == accmode)
5173			return (0);
5174
5175		goto privcheck;
5176	}
5177
5178	/* Otherwise, check everyone else. */
5179	if (file_mode & S_IXOTH)
5180		dac_granted |= VEXEC;
5181	if (file_mode & S_IROTH)
5182		dac_granted |= VREAD;
5183	if (file_mode & S_IWOTH)
5184		dac_granted |= (VWRITE | VAPPEND);
5185	if ((accmode & dac_granted) == accmode)
5186		return (0);
5187
5188privcheck:
5189	/*
5190	 * Build a privilege mask to determine if the set of privileges
5191	 * satisfies the requirements when combined with the granted mask
5192	 * from above.  For each privilege, if the privilege is required,
5193	 * bitwise or the request type onto the priv_granted mask.
5194	 */
5195	priv_granted = 0;
5196
5197	if (type == VDIR) {
5198		/*
5199		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5200		 * requests, instead of PRIV_VFS_EXEC.
5201		 */
5202		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5203		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5204			priv_granted |= VEXEC;
5205	} else {
5206		/*
5207		 * Ensure that at least one execute bit is on. Otherwise,
5208		 * a privileged user will always succeed, and we don't want
5209		 * this to happen unless the file really is executable.
5210		 */
5211		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5212		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5213		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5214			priv_granted |= VEXEC;
5215	}
5216
5217	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5218	    !priv_check_cred(cred, PRIV_VFS_READ))
5219		priv_granted |= VREAD;
5220
5221	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5222	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5223		priv_granted |= (VWRITE | VAPPEND);
5224
5225	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5226	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5227		priv_granted |= VADMIN;
5228
5229	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5230		/* XXX audit: privilege used */
5231		if (privused != NULL)
5232			*privused = 1;
5233		return (0);
5234	}
5235
5236	return ((accmode & VADMIN) ? EPERM : EACCES);
5237}
5238
5239/*
5240 * Credential check based on process requesting service, and per-attribute
5241 * permissions.
5242 */
5243int
5244extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5245    struct thread *td, accmode_t accmode)
5246{
5247
5248	/*
5249	 * Kernel-invoked always succeeds.
5250	 */
5251	if (cred == NOCRED)
5252		return (0);
5253
5254	/*
5255	 * Do not allow privileged processes in jail to directly manipulate
5256	 * system attributes.
5257	 */
5258	switch (attrnamespace) {
5259	case EXTATTR_NAMESPACE_SYSTEM:
5260		/* Potentially should be: return (EPERM); */
5261		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5262	case EXTATTR_NAMESPACE_USER:
5263		return (VOP_ACCESS(vp, accmode, cred, td));
5264	default:
5265		return (EPERM);
5266	}
5267}
5268
5269#ifdef DEBUG_VFS_LOCKS
5270/*
5271 * This only exists to suppress warnings from unlocked specfs accesses.  It is
5272 * no longer ok to have an unlocked VFS.
5273 */
5274#define	IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL ||		\
5275	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
5276
5277int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5278SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5279    "Drop into debugger on lock violation");
5280
5281int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5282SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5283    0, "Check for interlock across VOPs");
5284
5285int vfs_badlock_print = 1;	/* Print lock violations. */
5286SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5287    0, "Print lock violations");
5288
5289int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5290SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5291    0, "Print vnode details on lock violations");
5292
5293#ifdef KDB
5294int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5295SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5296    &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5297#endif
5298
5299static void
5300vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5301{
5302
5303#ifdef KDB
5304	if (vfs_badlock_backtrace)
5305		kdb_backtrace();
5306#endif
5307	if (vfs_badlock_vnode)
5308		vn_printf(vp, "vnode ");
5309	if (vfs_badlock_print)
5310		printf("%s: %p %s\n", str, (void *)vp, msg);
5311	if (vfs_badlock_ddb)
5312		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5313}
5314
5315void
5316assert_vi_locked(struct vnode *vp, const char *str)
5317{
5318
5319	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5320		vfs_badlock("interlock is not locked but should be", str, vp);
5321}
5322
5323void
5324assert_vi_unlocked(struct vnode *vp, const char *str)
5325{
5326
5327	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5328		vfs_badlock("interlock is locked but should not be", str, vp);
5329}
5330
5331void
5332assert_vop_locked(struct vnode *vp, const char *str)
5333{
5334	int locked;
5335
5336	if (!IGNORE_LOCK(vp)) {
5337		locked = VOP_ISLOCKED(vp);
5338		if (locked == 0 || locked == LK_EXCLOTHER)
5339			vfs_badlock("is not locked but should be", str, vp);
5340	}
5341}
5342
5343void
5344assert_vop_unlocked(struct vnode *vp, const char *str)
5345{
5346
5347	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5348		vfs_badlock("is locked but should not be", str, vp);
5349}
5350
5351void
5352assert_vop_elocked(struct vnode *vp, const char *str)
5353{
5354
5355	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5356		vfs_badlock("is not exclusive locked but should be", str, vp);
5357}
5358#endif /* DEBUG_VFS_LOCKS */
5359
5360void
5361vop_rename_fail(struct vop_rename_args *ap)
5362{
5363
5364	if (ap->a_tvp != NULL)
5365		vput(ap->a_tvp);
5366	if (ap->a_tdvp == ap->a_tvp)
5367		vrele(ap->a_tdvp);
5368	else
5369		vput(ap->a_tdvp);
5370	vrele(ap->a_fdvp);
5371	vrele(ap->a_fvp);
5372}
5373
5374void
5375vop_rename_pre(void *ap)
5376{
5377	struct vop_rename_args *a = ap;
5378
5379#ifdef DEBUG_VFS_LOCKS
5380	if (a->a_tvp)
5381		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5382	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5383	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5384	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5385
5386	/* Check the source (from). */
5387	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5388	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5389		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5390	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5391		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5392
5393	/* Check the target. */
5394	if (a->a_tvp)
5395		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5396	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5397#endif
5398	if (a->a_tdvp != a->a_fdvp)
5399		vhold(a->a_fdvp);
5400	if (a->a_tvp != a->a_fvp)
5401		vhold(a->a_fvp);
5402	vhold(a->a_tdvp);
5403	if (a->a_tvp)
5404		vhold(a->a_tvp);
5405}
5406
5407#ifdef DEBUG_VFS_LOCKS
5408void
5409vop_strategy_pre(void *ap)
5410{
5411	struct vop_strategy_args *a;
5412	struct buf *bp;
5413
5414	a = ap;
5415	bp = a->a_bp;
5416
5417	/*
5418	 * Cluster ops lock their component buffers but not the IO container.
5419	 */
5420	if ((bp->b_flags & B_CLUSTER) != 0)
5421		return;
5422
5423	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5424		if (vfs_badlock_print)
5425			printf(
5426			    "VOP_STRATEGY: bp is not locked but should be\n");
5427		if (vfs_badlock_ddb)
5428			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5429	}
5430}
5431
5432void
5433vop_lock_pre(void *ap)
5434{
5435	struct vop_lock1_args *a = ap;
5436
5437	if ((a->a_flags & LK_INTERLOCK) == 0)
5438		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5439	else
5440		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5441}
5442
5443void
5444vop_lock_post(void *ap, int rc)
5445{
5446	struct vop_lock1_args *a = ap;
5447
5448	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5449	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5450		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5451}
5452
5453void
5454vop_unlock_pre(void *ap)
5455{
5456	struct vop_unlock_args *a = ap;
5457
5458	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5459}
5460
5461void
5462vop_need_inactive_pre(void *ap)
5463{
5464	struct vop_need_inactive_args *a = ap;
5465
5466	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5467}
5468
5469void
5470vop_need_inactive_post(void *ap, int rc)
5471{
5472	struct vop_need_inactive_args *a = ap;
5473
5474	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5475}
5476#endif
5477
5478void
5479vop_create_post(void *ap, int rc)
5480{
5481	struct vop_create_args *a = ap;
5482
5483	if (!rc)
5484		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5485}
5486
5487void
5488vop_deleteextattr_post(void *ap, int rc)
5489{
5490	struct vop_deleteextattr_args *a = ap;
5491
5492	if (!rc)
5493		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5494}
5495
5496void
5497vop_link_post(void *ap, int rc)
5498{
5499	struct vop_link_args *a = ap;
5500
5501	if (!rc) {
5502		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
5503		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
5504	}
5505}
5506
5507void
5508vop_mkdir_post(void *ap, int rc)
5509{
5510	struct vop_mkdir_args *a = ap;
5511
5512	if (!rc)
5513		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5514}
5515
5516void
5517vop_mknod_post(void *ap, int rc)
5518{
5519	struct vop_mknod_args *a = ap;
5520
5521	if (!rc)
5522		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5523}
5524
5525void
5526vop_reclaim_post(void *ap, int rc)
5527{
5528	struct vop_reclaim_args *a = ap;
5529
5530	if (!rc)
5531		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
5532}
5533
5534void
5535vop_remove_post(void *ap, int rc)
5536{
5537	struct vop_remove_args *a = ap;
5538
5539	if (!rc) {
5540		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5541		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5542	}
5543}
5544
5545void
5546vop_rename_post(void *ap, int rc)
5547{
5548	struct vop_rename_args *a = ap;
5549	long hint;
5550
5551	if (!rc) {
5552		hint = NOTE_WRITE;
5553		if (a->a_fdvp == a->a_tdvp) {
5554			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5555				hint |= NOTE_LINK;
5556			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5557			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5558		} else {
5559			hint |= NOTE_EXTEND;
5560			if (a->a_fvp->v_type == VDIR)
5561				hint |= NOTE_LINK;
5562			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5563
5564			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5565			    a->a_tvp->v_type == VDIR)
5566				hint &= ~NOTE_LINK;
5567			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5568		}
5569
5570		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5571		if (a->a_tvp)
5572			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5573	}
5574	if (a->a_tdvp != a->a_fdvp)
5575		vdrop(a->a_fdvp);
5576	if (a->a_tvp != a->a_fvp)
5577		vdrop(a->a_fvp);
5578	vdrop(a->a_tdvp);
5579	if (a->a_tvp)
5580		vdrop(a->a_tvp);
5581}
5582
5583void
5584vop_rmdir_post(void *ap, int rc)
5585{
5586	struct vop_rmdir_args *a = ap;
5587
5588	if (!rc) {
5589		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5590		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5591	}
5592}
5593
5594void
5595vop_setattr_post(void *ap, int rc)
5596{
5597	struct vop_setattr_args *a = ap;
5598
5599	if (!rc)
5600		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5601}
5602
5603void
5604vop_setextattr_post(void *ap, int rc)
5605{
5606	struct vop_setextattr_args *a = ap;
5607
5608	if (!rc)
5609		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5610}
5611
5612void
5613vop_symlink_post(void *ap, int rc)
5614{
5615	struct vop_symlink_args *a = ap;
5616
5617	if (!rc)
5618		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5619}
5620
5621void
5622vop_open_post(void *ap, int rc)
5623{
5624	struct vop_open_args *a = ap;
5625
5626	if (!rc)
5627		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5628}
5629
5630void
5631vop_close_post(void *ap, int rc)
5632{
5633	struct vop_close_args *a = ap;
5634
5635	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5636	    !VN_IS_DOOMED(a->a_vp))) {
5637		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5638		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
5639	}
5640}
5641
5642void
5643vop_read_post(void *ap, int rc)
5644{
5645	struct vop_read_args *a = ap;
5646
5647	if (!rc)
5648		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5649}
5650
5651void
5652vop_readdir_post(void *ap, int rc)
5653{
5654	struct vop_readdir_args *a = ap;
5655
5656	if (!rc)
5657		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5658}
5659
5660static struct knlist fs_knlist;
5661
5662static void
5663vfs_event_init(void *arg)
5664{
5665	knlist_init_mtx(&fs_knlist, NULL);
5666}
5667/* XXX - correct order? */
5668SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5669
5670void
5671vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5672{
5673
5674	KNOTE_UNLOCKED(&fs_knlist, event);
5675}
5676
5677static int	filt_fsattach(struct knote *kn);
5678static void	filt_fsdetach(struct knote *kn);
5679static int	filt_fsevent(struct knote *kn, long hint);
5680
5681struct filterops fs_filtops = {
5682	.f_isfd = 0,
5683	.f_attach = filt_fsattach,
5684	.f_detach = filt_fsdetach,
5685	.f_event = filt_fsevent
5686};
5687
5688static int
5689filt_fsattach(struct knote *kn)
5690{
5691
5692	kn->kn_flags |= EV_CLEAR;
5693	knlist_add(&fs_knlist, kn, 0);
5694	return (0);
5695}
5696
5697static void
5698filt_fsdetach(struct knote *kn)
5699{
5700
5701	knlist_remove(&fs_knlist, kn, 0);
5702}
5703
5704static int
5705filt_fsevent(struct knote *kn, long hint)
5706{
5707
5708	kn->kn_fflags |= hint;
5709	return (kn->kn_fflags != 0);
5710}
5711
5712static int
5713sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5714{
5715	struct vfsidctl vc;
5716	int error;
5717	struct mount *mp;
5718
5719	error = SYSCTL_IN(req, &vc, sizeof(vc));
5720	if (error)
5721		return (error);
5722	if (vc.vc_vers != VFS_CTL_VERS1)
5723		return (EINVAL);
5724	mp = vfs_getvfs(&vc.vc_fsid);
5725	if (mp == NULL)
5726		return (ENOENT);
5727	/* ensure that a specific sysctl goes to the right filesystem. */
5728	if (strcmp(vc.vc_fstypename, "*") != 0 &&
5729	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
5730		vfs_rel(mp);
5731		return (EINVAL);
5732	}
5733	VCTLTOREQ(&vc, req);
5734	error = VFS_SYSCTL(mp, vc.vc_op, req);
5735	vfs_rel(mp);
5736	return (error);
5737}
5738
5739SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
5740    NULL, 0, sysctl_vfs_ctl, "",
5741    "Sysctl by fsid");
5742
5743/*
5744 * Function to initialize a va_filerev field sensibly.
5745 * XXX: Wouldn't a random number make a lot more sense ??
5746 */
5747u_quad_t
5748init_va_filerev(void)
5749{
5750	struct bintime bt;
5751
5752	getbinuptime(&bt);
5753	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5754}
5755
5756static int	filt_vfsread(struct knote *kn, long hint);
5757static int	filt_vfswrite(struct knote *kn, long hint);
5758static int	filt_vfsvnode(struct knote *kn, long hint);
5759static void	filt_vfsdetach(struct knote *kn);
5760static struct filterops vfsread_filtops = {
5761	.f_isfd = 1,
5762	.f_detach = filt_vfsdetach,
5763	.f_event = filt_vfsread
5764};
5765static struct filterops vfswrite_filtops = {
5766	.f_isfd = 1,
5767	.f_detach = filt_vfsdetach,
5768	.f_event = filt_vfswrite
5769};
5770static struct filterops vfsvnode_filtops = {
5771	.f_isfd = 1,
5772	.f_detach = filt_vfsdetach,
5773	.f_event = filt_vfsvnode
5774};
5775
5776static void
5777vfs_knllock(void *arg)
5778{
5779	struct vnode *vp = arg;
5780
5781	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5782}
5783
5784static void
5785vfs_knlunlock(void *arg)
5786{
5787	struct vnode *vp = arg;
5788
5789	VOP_UNLOCK(vp);
5790}
5791
5792static void
5793vfs_knl_assert_locked(void *arg)
5794{
5795#ifdef DEBUG_VFS_LOCKS
5796	struct vnode *vp = arg;
5797
5798	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
5799#endif
5800}
5801
5802static void
5803vfs_knl_assert_unlocked(void *arg)
5804{
5805#ifdef DEBUG_VFS_LOCKS
5806	struct vnode *vp = arg;
5807
5808	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
5809#endif
5810}
5811
5812int
5813vfs_kqfilter(struct vop_kqfilter_args *ap)
5814{
5815	struct vnode *vp = ap->a_vp;
5816	struct knote *kn = ap->a_kn;
5817	struct knlist *knl;
5818
5819	switch (kn->kn_filter) {
5820	case EVFILT_READ:
5821		kn->kn_fop = &vfsread_filtops;
5822		break;
5823	case EVFILT_WRITE:
5824		kn->kn_fop = &vfswrite_filtops;
5825		break;
5826	case EVFILT_VNODE:
5827		kn->kn_fop = &vfsvnode_filtops;
5828		break;
5829	default:
5830		return (EINVAL);
5831	}
5832
5833	kn->kn_hook = (caddr_t)vp;
5834
5835	v_addpollinfo(vp);
5836	if (vp->v_pollinfo == NULL)
5837		return (ENOMEM);
5838	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
5839	vhold(vp);
5840	knlist_add(knl, kn, 0);
5841
5842	return (0);
5843}
5844
5845/*
5846 * Detach knote from vnode
5847 */
5848static void
5849filt_vfsdetach(struct knote *kn)
5850{
5851	struct vnode *vp = (struct vnode *)kn->kn_hook;
5852
5853	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5854	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5855	vdrop(vp);
5856}
5857
5858/*ARGSUSED*/
5859static int
5860filt_vfsread(struct knote *kn, long hint)
5861{
5862	struct vnode *vp = (struct vnode *)kn->kn_hook;
5863	struct vattr va;
5864	int res;
5865
5866	/*
5867	 * filesystem is gone, so set the EOF flag and schedule
5868	 * the knote for deletion.
5869	 */
5870	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5871		VI_LOCK(vp);
5872		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5873		VI_UNLOCK(vp);
5874		return (1);
5875	}
5876
5877	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5878		return (0);
5879
5880	VI_LOCK(vp);
5881	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5882	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5883	VI_UNLOCK(vp);
5884	return (res);
5885}
5886
5887/*ARGSUSED*/
5888static int
5889filt_vfswrite(struct knote *kn, long hint)
5890{
5891	struct vnode *vp = (struct vnode *)kn->kn_hook;
5892
5893	VI_LOCK(vp);
5894
5895	/*
5896	 * filesystem is gone, so set the EOF flag and schedule
5897	 * the knote for deletion.
5898	 */
5899	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5900		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5901
5902	kn->kn_data = 0;
5903	VI_UNLOCK(vp);
5904	return (1);
5905}
5906
5907static int
5908filt_vfsvnode(struct knote *kn, long hint)
5909{
5910	struct vnode *vp = (struct vnode *)kn->kn_hook;
5911	int res;
5912
5913	VI_LOCK(vp);
5914	if (kn->kn_sfflags & hint)
5915		kn->kn_fflags |= hint;
5916	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5917		kn->kn_flags |= EV_EOF;
5918		VI_UNLOCK(vp);
5919		return (1);
5920	}
5921	res = (kn->kn_fflags != 0);
5922	VI_UNLOCK(vp);
5923	return (res);
5924}
5925
5926/*
5927 * Returns whether the directory is empty or not.
5928 * If it is empty, the return value is 0; otherwise
5929 * the return value is an error value (which may
5930 * be ENOTEMPTY).
5931 */
5932int
5933vfs_emptydir(struct vnode *vp)
5934{
5935	struct uio uio;
5936	struct iovec iov;
5937	struct dirent *dirent, *dp, *endp;
5938	int error, eof;
5939
5940	error = 0;
5941	eof = 0;
5942
5943	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
5944
5945	dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
5946	iov.iov_base = dirent;
5947	iov.iov_len = sizeof(struct dirent);
5948
5949	uio.uio_iov = &iov;
5950	uio.uio_iovcnt = 1;
5951	uio.uio_offset = 0;
5952	uio.uio_resid = sizeof(struct dirent);
5953	uio.uio_segflg = UIO_SYSSPACE;
5954	uio.uio_rw = UIO_READ;
5955	uio.uio_td = curthread;
5956
5957	while (eof == 0 && error == 0) {
5958		error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
5959		    NULL, NULL);
5960		if (error != 0)
5961			break;
5962		endp = (void *)((uint8_t *)dirent +
5963		    sizeof(struct dirent) - uio.uio_resid);
5964		for (dp = dirent; dp < endp;
5965		     dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
5966			if (dp->d_type == DT_WHT)
5967				continue;
5968			if (dp->d_namlen == 0)
5969				continue;
5970			if (dp->d_type != DT_DIR &&
5971			    dp->d_type != DT_UNKNOWN) {
5972				error = ENOTEMPTY;
5973				break;
5974			}
5975			if (dp->d_namlen > 2) {
5976				error = ENOTEMPTY;
5977				break;
5978			}
5979			if (dp->d_namlen == 1 &&
5980			    dp->d_name[0] != '.') {
5981				error = ENOTEMPTY;
5982				break;
5983			}
5984			if (dp->d_namlen == 2 &&
5985			    dp->d_name[1] != '.') {
5986				error = ENOTEMPTY;
5987				break;
5988			}
5989			uio.uio_resid = sizeof(struct dirent);
5990		}
5991	}
5992	free(dirent, M_TEMP);
5993	return (error);
5994}
5995
5996int
5997vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5998{
5999	int error;
6000
6001	if (dp->d_reclen > ap->a_uio->uio_resid)
6002		return (ENAMETOOLONG);
6003	error = uiomove(dp, dp->d_reclen, ap->a_uio);
6004	if (error) {
6005		if (ap->a_ncookies != NULL) {
6006			if (ap->a_cookies != NULL)
6007				free(ap->a_cookies, M_TEMP);
6008			ap->a_cookies = NULL;
6009			*ap->a_ncookies = 0;
6010		}
6011		return (error);
6012	}
6013	if (ap->a_ncookies == NULL)
6014		return (0);
6015
6016	KASSERT(ap->a_cookies,
6017	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6018
6019	*ap->a_cookies = realloc(*ap->a_cookies,
6020	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
6021	(*ap->a_cookies)[*ap->a_ncookies] = off;
6022	*ap->a_ncookies += 1;
6023	return (0);
6024}
6025
6026/*
6027 * The purpose of this routine is to remove granularity from accmode_t,
6028 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6029 * VADMIN and VAPPEND.
6030 *
6031 * If it returns 0, the caller is supposed to continue with the usual
6032 * access checks using 'accmode' as modified by this routine.  If it
6033 * returns nonzero value, the caller is supposed to return that value
6034 * as errno.
6035 *
6036 * Note that after this routine runs, accmode may be zero.
6037 */
6038int
6039vfs_unixify_accmode(accmode_t *accmode)
6040{
6041	/*
6042	 * There is no way to specify explicit "deny" rule using
6043	 * file mode or POSIX.1e ACLs.
6044	 */
6045	if (*accmode & VEXPLICIT_DENY) {
6046		*accmode = 0;
6047		return (0);
6048	}
6049
6050	/*
6051	 * None of these can be translated into usual access bits.
6052	 * Also, the common case for NFSv4 ACLs is to not contain
6053	 * either of these bits. Caller should check for VWRITE
6054	 * on the containing directory instead.
6055	 */
6056	if (*accmode & (VDELETE_CHILD | VDELETE))
6057		return (EPERM);
6058
6059	if (*accmode & VADMIN_PERMS) {
6060		*accmode &= ~VADMIN_PERMS;
6061		*accmode |= VADMIN;
6062	}
6063
6064	/*
6065	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6066	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6067	 */
6068	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6069
6070	return (0);
6071}
6072
6073/*
6074 * Clear out a doomed vnode (if any) and replace it with a new one as long
6075 * as the fs is not being unmounted. Return the root vnode to the caller.
6076 */
6077static int __noinline
6078vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6079{
6080	struct vnode *vp;
6081	int error;
6082
6083restart:
6084	if (mp->mnt_rootvnode != NULL) {
6085		MNT_ILOCK(mp);
6086		vp = mp->mnt_rootvnode;
6087		if (vp != NULL) {
6088			if (!VN_IS_DOOMED(vp)) {
6089				vrefact(vp);
6090				MNT_IUNLOCK(mp);
6091				error = vn_lock(vp, flags);
6092				if (error == 0) {
6093					*vpp = vp;
6094					return (0);
6095				}
6096				vrele(vp);
6097				goto restart;
6098			}
6099			/*
6100			 * Clear the old one.
6101			 */
6102			mp->mnt_rootvnode = NULL;
6103		}
6104		MNT_IUNLOCK(mp);
6105		if (vp != NULL) {
6106			vfs_op_barrier_wait(mp);
6107			vrele(vp);
6108		}
6109	}
6110	error = VFS_CACHEDROOT(mp, flags, vpp);
6111	if (error != 0)
6112		return (error);
6113	if (mp->mnt_vfs_ops == 0) {
6114		MNT_ILOCK(mp);
6115		if (mp->mnt_vfs_ops != 0) {
6116			MNT_IUNLOCK(mp);
6117			return (0);
6118		}
6119		if (mp->mnt_rootvnode == NULL) {
6120			vrefact(*vpp);
6121			mp->mnt_rootvnode = *vpp;
6122		} else {
6123			if (mp->mnt_rootvnode != *vpp) {
6124				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6125					panic("%s: mismatch between vnode returned "
6126					    " by VFS_CACHEDROOT and the one cached "
6127					    " (%p != %p)",
6128					    __func__, *vpp, mp->mnt_rootvnode);
6129				}
6130			}
6131		}
6132		MNT_IUNLOCK(mp);
6133	}
6134	return (0);
6135}
6136
6137int
6138vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6139{
6140	struct vnode *vp;
6141	int error;
6142
6143	if (!vfs_op_thread_enter(mp))
6144		return (vfs_cache_root_fallback(mp, flags, vpp));
6145	vp = atomic_load_ptr(&mp->mnt_rootvnode);
6146	if (vp == NULL || VN_IS_DOOMED(vp)) {
6147		vfs_op_thread_exit(mp);
6148		return (vfs_cache_root_fallback(mp, flags, vpp));
6149	}
6150	vrefact(vp);
6151	vfs_op_thread_exit(mp);
6152	error = vn_lock(vp, flags);
6153	if (error != 0) {
6154		vrele(vp);
6155		return (vfs_cache_root_fallback(mp, flags, vpp));
6156	}
6157	*vpp = vp;
6158	return (0);
6159}
6160
6161struct vnode *
6162vfs_cache_root_clear(struct mount *mp)
6163{
6164	struct vnode *vp;
6165
6166	/*
6167	 * ops > 0 guarantees there is nobody who can see this vnode
6168	 */
6169	MPASS(mp->mnt_vfs_ops > 0);
6170	vp = mp->mnt_rootvnode;
6171	mp->mnt_rootvnode = NULL;
6172	return (vp);
6173}
6174
6175void
6176vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6177{
6178
6179	MPASS(mp->mnt_vfs_ops > 0);
6180	vrefact(vp);
6181	mp->mnt_rootvnode = vp;
6182}
6183
6184/*
6185 * These are helper functions for filesystems to traverse all
6186 * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6187 *
6188 * This interface replaces MNT_VNODE_FOREACH.
6189 */
6190
6191struct vnode *
6192__mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6193{
6194	struct vnode *vp;
6195
6196	if (should_yield())
6197		kern_yield(PRI_USER);
6198	MNT_ILOCK(mp);
6199	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6200	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6201	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6202		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6203		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6204			continue;
6205		VI_LOCK(vp);
6206		if (VN_IS_DOOMED(vp)) {
6207			VI_UNLOCK(vp);
6208			continue;
6209		}
6210		break;
6211	}
6212	if (vp == NULL) {
6213		__mnt_vnode_markerfree_all(mvp, mp);
6214		/* MNT_IUNLOCK(mp); -- done in above function */
6215		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6216		return (NULL);
6217	}
6218	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6219	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6220	MNT_IUNLOCK(mp);
6221	return (vp);
6222}
6223
6224struct vnode *
6225__mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6226{
6227	struct vnode *vp;
6228
6229	*mvp = vn_alloc_marker(mp);
6230	MNT_ILOCK(mp);
6231	MNT_REF(mp);
6232
6233	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6234		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6235		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6236			continue;
6237		VI_LOCK(vp);
6238		if (VN_IS_DOOMED(vp)) {
6239			VI_UNLOCK(vp);
6240			continue;
6241		}
6242		break;
6243	}
6244	if (vp == NULL) {
6245		MNT_REL(mp);
6246		MNT_IUNLOCK(mp);
6247		vn_free_marker(*mvp);
6248		*mvp = NULL;
6249		return (NULL);
6250	}
6251	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6252	MNT_IUNLOCK(mp);
6253	return (vp);
6254}
6255
6256void
6257__mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6258{
6259
6260	if (*mvp == NULL) {
6261		MNT_IUNLOCK(mp);
6262		return;
6263	}
6264
6265	mtx_assert(MNT_MTX(mp), MA_OWNED);
6266
6267	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6268	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6269	MNT_REL(mp);
6270	MNT_IUNLOCK(mp);
6271	vn_free_marker(*mvp);
6272	*mvp = NULL;
6273}
6274
6275/*
6276 * These are helper functions for filesystems to traverse their
6277 * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6278 */
6279static void
6280mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6281{
6282
6283	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6284
6285	MNT_ILOCK(mp);
6286	MNT_REL(mp);
6287	MNT_IUNLOCK(mp);
6288	vn_free_marker(*mvp);
6289	*mvp = NULL;
6290}
6291
6292/*
6293 * Relock the mp mount vnode list lock with the vp vnode interlock in the
6294 * conventional lock order during mnt_vnode_next_lazy iteration.
6295 *
6296 * On entry, the mount vnode list lock is held and the vnode interlock is not.
6297 * The list lock is dropped and reacquired.  On success, both locks are held.
6298 * On failure, the mount vnode list lock is held but the vnode interlock is
6299 * not, and the procedure may have yielded.
6300 */
6301static bool
6302mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6303    struct vnode *vp)
6304{
6305
6306	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6307	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6308	    ("%s: bad marker", __func__));
6309	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6310	    ("%s: inappropriate vnode", __func__));
6311	ASSERT_VI_UNLOCKED(vp, __func__);
6312	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6313
6314	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6315	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6316
6317	/*
6318	 * Note we may be racing against vdrop which transitioned the hold
6319	 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6320	 * if we are the only user after we get the interlock we will just
6321	 * vdrop.
6322	 */
6323	vhold(vp);
6324	mtx_unlock(&mp->mnt_listmtx);
6325	VI_LOCK(vp);
6326	if (VN_IS_DOOMED(vp)) {
6327		VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6328		goto out_lost;
6329	}
6330	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6331	/*
6332	 * There is nothing to do if we are the last user.
6333	 */
6334	if (!refcount_release_if_not_last(&vp->v_holdcnt))
6335		goto out_lost;
6336	mtx_lock(&mp->mnt_listmtx);
6337	return (true);
6338out_lost:
6339	vdropl(vp);
6340	maybe_yield();
6341	mtx_lock(&mp->mnt_listmtx);
6342	return (false);
6343}
6344
6345static struct vnode *
6346mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6347    void *cbarg)
6348{
6349	struct vnode *vp;
6350
6351	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6352	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6353restart:
6354	vp = TAILQ_NEXT(*mvp, v_lazylist);
6355	while (vp != NULL) {
6356		if (vp->v_type == VMARKER) {
6357			vp = TAILQ_NEXT(vp, v_lazylist);
6358			continue;
6359		}
6360		/*
6361		 * See if we want to process the vnode. Note we may encounter a
6362		 * long string of vnodes we don't care about and hog the list
6363		 * as a result. Check for it and requeue the marker.
6364		 */
6365		VNPASS(!VN_IS_DOOMED(vp), vp);
6366		if (!cb(vp, cbarg)) {
6367			if (!should_yield()) {
6368				vp = TAILQ_NEXT(vp, v_lazylist);
6369				continue;
6370			}
6371			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6372			    v_lazylist);
6373			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6374			    v_lazylist);
6375			mtx_unlock(&mp<