xref: /freebsd/sys/kern/vfs_subr.c (revision 25fb30bd9abc492359ad1f66901a06cb8cd08370)
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/smr.h>
80 #include <sys/smp.h>
81 #include <sys/stat.h>
82 #include <sys/sysctl.h>
83 #include <sys/syslog.h>
84 #include <sys/vmmeter.h>
85 #include <sys/vnode.h>
86 #include <sys/watchdog.h>
87 
88 #include <machine/stdarg.h>
89 
90 #include <security/mac/mac_framework.h>
91 
92 #include <vm/vm.h>
93 #include <vm/vm_object.h>
94 #include <vm/vm_extern.h>
95 #include <vm/pmap.h>
96 #include <vm/vm_map.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_kern.h>
99 #include <vm/uma.h>
100 
101 #ifdef DDB
102 #include <ddb/ddb.h>
103 #endif
104 
105 static void	delmntque(struct vnode *vp);
106 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
107 		    int slpflag, int slptimeo);
108 static void	syncer_shutdown(void *arg, int howto);
109 static int	vtryrecycle(struct vnode *vp);
110 static void	v_init_counters(struct vnode *);
111 static void	vgonel(struct vnode *);
112 static void	vfs_knllock(void *arg);
113 static void	vfs_knlunlock(void *arg);
114 static void	vfs_knl_assert_locked(void *arg);
115 static void	vfs_knl_assert_unlocked(void *arg);
116 static void	destroy_vpollinfo(struct vpollinfo *vi);
117 static int	v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
118 		    daddr_t startlbn, daddr_t endlbn);
119 static void	vnlru_recalc(void);
120 
121 /*
122  * These fences are intended for cases where some synchronization is
123  * needed between access of v_iflags and lockless vnode refcount (v_holdcnt
124  * and v_usecount) updates.  Access to v_iflags is generally synchronized
125  * by the interlock, but we have some internal assertions that check vnode
126  * flags without acquiring the lock.  Thus, these fences are INVARIANTS-only
127  * for now.
128  */
129 #ifdef INVARIANTS
130 #define	VNODE_REFCOUNT_FENCE_ACQ()	atomic_thread_fence_acq()
131 #define	VNODE_REFCOUNT_FENCE_REL()	atomic_thread_fence_rel()
132 #else
133 #define	VNODE_REFCOUNT_FENCE_ACQ()
134 #define	VNODE_REFCOUNT_FENCE_REL()
135 #endif
136 
137 /*
138  * Number of vnodes in existence.  Increased whenever getnewvnode()
139  * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
140  */
141 static u_long __exclusive_cache_line numvnodes;
142 
143 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
144     "Number of vnodes in existence");
145 
146 static counter_u64_t vnodes_created;
147 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
148     "Number of vnodes created by getnewvnode");
149 
150 /*
151  * Conversion tables for conversion from vnode types to inode formats
152  * and back.
153  */
154 enum vtype iftovt_tab[16] = {
155 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
156 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
157 };
158 int vttoif_tab[10] = {
159 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
160 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
161 };
162 
163 /*
164  * List of allocates vnodes in the system.
165  */
166 static TAILQ_HEAD(freelst, vnode) vnode_list;
167 static struct vnode *vnode_list_free_marker;
168 static struct vnode *vnode_list_reclaim_marker;
169 
170 /*
171  * "Free" vnode target.  Free vnodes are rarely completely free, but are
172  * just ones that are cheap to recycle.  Usually they are for files which
173  * have been stat'd but not read; these usually have inode and namecache
174  * data attached to them.  This target is the preferred minimum size of a
175  * sub-cache consisting mostly of such files. The system balances the size
176  * of this sub-cache with its complement to try to prevent either from
177  * thrashing while the other is relatively inactive.  The targets express
178  * a preference for the best balance.
179  *
180  * "Above" this target there are 2 further targets (watermarks) related
181  * to recyling of free vnodes.  In the best-operating case, the cache is
182  * exactly full, the free list has size between vlowat and vhiwat above the
183  * free target, and recycling from it and normal use maintains this state.
184  * Sometimes the free list is below vlowat or even empty, but this state
185  * is even better for immediate use provided the cache is not full.
186  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
187  * ones) to reach one of these states.  The watermarks are currently hard-
188  * coded as 4% and 9% of the available space higher.  These and the default
189  * of 25% for wantfreevnodes are too large if the memory size is large.
190  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
191  * whenever vnlru_proc() becomes active.
192  */
193 static long wantfreevnodes;
194 static long __exclusive_cache_line freevnodes;
195 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
196     &freevnodes, 0, "Number of \"free\" vnodes");
197 static long freevnodes_old;
198 
199 static counter_u64_t recycles_count;
200 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
201     "Number of vnodes recycled to meet vnode cache targets");
202 
203 static counter_u64_t recycles_free_count;
204 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count,
205     "Number of free vnodes recycled to meet vnode cache targets");
206 
207 static counter_u64_t deferred_inact;
208 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact,
209     "Number of times inactive processing was deferred");
210 
211 /* To keep more than one thread at a time from running vfs_getnewfsid */
212 static struct mtx mntid_mtx;
213 
214 /*
215  * Lock for any access to the following:
216  *	vnode_list
217  *	numvnodes
218  *	freevnodes
219  */
220 static struct mtx __exclusive_cache_line vnode_list_mtx;
221 
222 /* Publicly exported FS */
223 struct nfs_public nfs_pub;
224 
225 static uma_zone_t buf_trie_zone;
226 static smr_t buf_trie_smr;
227 
228 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
229 static uma_zone_t vnode_zone;
230 static uma_zone_t vnodepoll_zone;
231 
232 __read_frequently smr_t vfs_smr;
233 
234 /*
235  * The workitem queue.
236  *
237  * It is useful to delay writes of file data and filesystem metadata
238  * for tens of seconds so that quickly created and deleted files need
239  * not waste disk bandwidth being created and removed. To realize this,
240  * we append vnodes to a "workitem" queue. When running with a soft
241  * updates implementation, most pending metadata dependencies should
242  * not wait for more than a few seconds. Thus, mounted on block devices
243  * are delayed only about a half the time that file data is delayed.
244  * Similarly, directory updates are more critical, so are only delayed
245  * about a third the time that file data is delayed. Thus, there are
246  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
247  * one each second (driven off the filesystem syncer process). The
248  * syncer_delayno variable indicates the next queue that is to be processed.
249  * Items that need to be processed soon are placed in this queue:
250  *
251  *	syncer_workitem_pending[syncer_delayno]
252  *
253  * A delay of fifteen seconds is done by placing the request fifteen
254  * entries later in the queue:
255  *
256  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
257  *
258  */
259 static int syncer_delayno;
260 static long syncer_mask;
261 LIST_HEAD(synclist, bufobj);
262 static struct synclist *syncer_workitem_pending;
263 /*
264  * The sync_mtx protects:
265  *	bo->bo_synclist
266  *	sync_vnode_count
267  *	syncer_delayno
268  *	syncer_state
269  *	syncer_workitem_pending
270  *	syncer_worklist_len
271  *	rushjob
272  */
273 static struct mtx sync_mtx;
274 static struct cv sync_wakeup;
275 
276 #define SYNCER_MAXDELAY		32
277 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
278 static int syncdelay = 30;		/* max time to delay syncing data */
279 static int filedelay = 30;		/* time to delay syncing files */
280 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
281     "Time to delay syncing files (in seconds)");
282 static int dirdelay = 29;		/* time to delay syncing directories */
283 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
284     "Time to delay syncing directories (in seconds)");
285 static int metadelay = 28;		/* time to delay syncing metadata */
286 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
287     "Time to delay syncing metadata (in seconds)");
288 static int rushjob;		/* number of slots to run ASAP */
289 static int stat_rush_requests;	/* number of times I/O speeded up */
290 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
291     "Number of times I/O speeded up (rush requests)");
292 
293 #define	VDBATCH_SIZE 8
294 struct vdbatch {
295 	u_int index;
296 	long freevnodes;
297 	struct mtx lock;
298 	struct vnode *tab[VDBATCH_SIZE];
299 };
300 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
301 
302 static void	vdbatch_dequeue(struct vnode *vp);
303 
304 /*
305  * When shutting down the syncer, run it at four times normal speed.
306  */
307 #define SYNCER_SHUTDOWN_SPEEDUP		4
308 static int sync_vnode_count;
309 static int syncer_worklist_len;
310 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
311     syncer_state;
312 
313 /* Target for maximum number of vnodes. */
314 u_long desiredvnodes;
315 static u_long gapvnodes;		/* gap between wanted and desired */
316 static u_long vhiwat;		/* enough extras after expansion */
317 static u_long vlowat;		/* minimal extras before expansion */
318 static u_long vstir;		/* nonzero to stir non-free vnodes */
319 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
320 
321 static u_long vnlru_read_freevnodes(void);
322 
323 /*
324  * Note that no attempt is made to sanitize these parameters.
325  */
326 static int
327 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
328 {
329 	u_long val;
330 	int error;
331 
332 	val = desiredvnodes;
333 	error = sysctl_handle_long(oidp, &val, 0, req);
334 	if (error != 0 || req->newptr == NULL)
335 		return (error);
336 
337 	if (val == desiredvnodes)
338 		return (0);
339 	mtx_lock(&vnode_list_mtx);
340 	desiredvnodes = val;
341 	wantfreevnodes = desiredvnodes / 4;
342 	vnlru_recalc();
343 	mtx_unlock(&vnode_list_mtx);
344 	/*
345 	 * XXX There is no protection against multiple threads changing
346 	 * desiredvnodes at the same time. Locking above only helps vnlru and
347 	 * getnewvnode.
348 	 */
349 	vfs_hash_changesize(desiredvnodes);
350 	cache_changesize(desiredvnodes);
351 	return (0);
352 }
353 
354 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
355     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
356     "LU", "Target for maximum number of vnodes");
357 
358 static int
359 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
360 {
361 	u_long val;
362 	int error;
363 
364 	val = wantfreevnodes;
365 	error = sysctl_handle_long(oidp, &val, 0, req);
366 	if (error != 0 || req->newptr == NULL)
367 		return (error);
368 
369 	if (val == wantfreevnodes)
370 		return (0);
371 	mtx_lock(&vnode_list_mtx);
372 	wantfreevnodes = val;
373 	vnlru_recalc();
374 	mtx_unlock(&vnode_list_mtx);
375 	return (0);
376 }
377 
378 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
379     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
380     "LU", "Target for minimum number of \"free\" vnodes");
381 
382 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
383     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
384 static int vnlru_nowhere;
385 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
386     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
387 
388 static int
389 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
390 {
391 	struct vnode *vp;
392 	struct nameidata nd;
393 	char *buf;
394 	unsigned long ndflags;
395 	int error;
396 
397 	if (req->newptr == NULL)
398 		return (EINVAL);
399 	if (req->newlen >= PATH_MAX)
400 		return (E2BIG);
401 
402 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
403 	error = SYSCTL_IN(req, buf, req->newlen);
404 	if (error != 0)
405 		goto out;
406 
407 	buf[req->newlen] = '\0';
408 
409 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | SAVENAME;
410 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread);
411 	if ((error = namei(&nd)) != 0)
412 		goto out;
413 	vp = nd.ni_vp;
414 
415 	if (VN_IS_DOOMED(vp)) {
416 		/*
417 		 * This vnode is being recycled.  Return != 0 to let the caller
418 		 * know that the sysctl had no effect.  Return EAGAIN because a
419 		 * subsequent call will likely succeed (since namei will create
420 		 * a new vnode if necessary)
421 		 */
422 		error = EAGAIN;
423 		goto putvnode;
424 	}
425 
426 	counter_u64_add(recycles_count, 1);
427 	vgone(vp);
428 putvnode:
429 	NDFREE(&nd, 0);
430 out:
431 	free(buf, M_TEMP);
432 	return (error);
433 }
434 
435 static int
436 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
437 {
438 	struct thread *td = curthread;
439 	struct vnode *vp;
440 	struct file *fp;
441 	int error;
442 	int fd;
443 
444 	if (req->newptr == NULL)
445 		return (EBADF);
446 
447         error = sysctl_handle_int(oidp, &fd, 0, req);
448         if (error != 0)
449                 return (error);
450 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
451 	if (error != 0)
452 		return (error);
453 	vp = fp->f_vnode;
454 
455 	error = vn_lock(vp, LK_EXCLUSIVE);
456 	if (error != 0)
457 		goto drop;
458 
459 	counter_u64_add(recycles_count, 1);
460 	vgone(vp);
461 	VOP_UNLOCK(vp);
462 drop:
463 	fdrop(fp, td);
464 	return (error);
465 }
466 
467 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
468     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
469     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
470 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
471     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
472     sysctl_ftry_reclaim_vnode, "I",
473     "Try to reclaim a vnode by its file descriptor");
474 
475 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
476 static int vnsz2log;
477 
478 /*
479  * Support for the bufobj clean & dirty pctrie.
480  */
481 static void *
482 buf_trie_alloc(struct pctrie *ptree)
483 {
484 	return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT));
485 }
486 
487 static void
488 buf_trie_free(struct pctrie *ptree, void *node)
489 {
490 	uma_zfree_smr(buf_trie_zone, node);
491 }
492 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free,
493     buf_trie_smr);
494 
495 /*
496  * Initialize the vnode management data structures.
497  *
498  * Reevaluate the following cap on the number of vnodes after the physical
499  * memory size exceeds 512GB.  In the limit, as the physical memory size
500  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
501  */
502 #ifndef	MAXVNODES_MAX
503 #define	MAXVNODES_MAX	(512UL * 1024 * 1024 / 64)	/* 8M */
504 #endif
505 
506 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
507 
508 static struct vnode *
509 vn_alloc_marker(struct mount *mp)
510 {
511 	struct vnode *vp;
512 
513 	vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
514 	vp->v_type = VMARKER;
515 	vp->v_mount = mp;
516 
517 	return (vp);
518 }
519 
520 static void
521 vn_free_marker(struct vnode *vp)
522 {
523 
524 	MPASS(vp->v_type == VMARKER);
525 	free(vp, M_VNODE_MARKER);
526 }
527 
528 /*
529  * Initialize a vnode as it first enters the zone.
530  */
531 static int
532 vnode_init(void *mem, int size, int flags)
533 {
534 	struct vnode *vp;
535 
536 	vp = mem;
537 	bzero(vp, size);
538 	/*
539 	 * Setup locks.
540 	 */
541 	vp->v_vnlock = &vp->v_lock;
542 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
543 	/*
544 	 * By default, don't allow shared locks unless filesystems opt-in.
545 	 */
546 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
547 	    LK_NOSHARE | LK_IS_VNODE);
548 	/*
549 	 * Initialize bufobj.
550 	 */
551 	bufobj_init(&vp->v_bufobj, vp);
552 	/*
553 	 * Initialize namecache.
554 	 */
555 	cache_vnode_init(vp);
556 	/*
557 	 * Initialize rangelocks.
558 	 */
559 	rangelock_init(&vp->v_rl);
560 
561 	vp->v_dbatchcpu = NOCPU;
562 
563 	mtx_lock(&vnode_list_mtx);
564 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
565 	mtx_unlock(&vnode_list_mtx);
566 	return (0);
567 }
568 
569 /*
570  * Free a vnode when it is cleared from the zone.
571  */
572 static void
573 vnode_fini(void *mem, int size)
574 {
575 	struct vnode *vp;
576 	struct bufobj *bo;
577 
578 	vp = mem;
579 	vdbatch_dequeue(vp);
580 	mtx_lock(&vnode_list_mtx);
581 	TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
582 	mtx_unlock(&vnode_list_mtx);
583 	rangelock_destroy(&vp->v_rl);
584 	lockdestroy(vp->v_vnlock);
585 	mtx_destroy(&vp->v_interlock);
586 	bo = &vp->v_bufobj;
587 	rw_destroy(BO_LOCKPTR(bo));
588 }
589 
590 /*
591  * Provide the size of NFS nclnode and NFS fh for calculation of the
592  * vnode memory consumption.  The size is specified directly to
593  * eliminate dependency on NFS-private header.
594  *
595  * Other filesystems may use bigger or smaller (like UFS and ZFS)
596  * private inode data, but the NFS-based estimation is ample enough.
597  * Still, we care about differences in the size between 64- and 32-bit
598  * platforms.
599  *
600  * Namecache structure size is heuristically
601  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
602  */
603 #ifdef _LP64
604 #define	NFS_NCLNODE_SZ	(528 + 64)
605 #define	NC_SZ		148
606 #else
607 #define	NFS_NCLNODE_SZ	(360 + 32)
608 #define	NC_SZ		92
609 #endif
610 
611 static void
612 vntblinit(void *dummy __unused)
613 {
614 	struct vdbatch *vd;
615 	int cpu, physvnodes, virtvnodes;
616 	u_int i;
617 
618 	/*
619 	 * Desiredvnodes is a function of the physical memory size and the
620 	 * kernel's heap size.  Generally speaking, it scales with the
621 	 * physical memory size.  The ratio of desiredvnodes to the physical
622 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
623 	 * Thereafter, the
624 	 * marginal ratio of desiredvnodes to the physical memory size is
625 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
626 	 * size.  The memory required by desiredvnodes vnodes and vm objects
627 	 * must not exceed 1/10th of the kernel's heap size.
628 	 */
629 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
630 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
631 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
632 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
633 	desiredvnodes = min(physvnodes, virtvnodes);
634 	if (desiredvnodes > MAXVNODES_MAX) {
635 		if (bootverbose)
636 			printf("Reducing kern.maxvnodes %lu -> %lu\n",
637 			    desiredvnodes, MAXVNODES_MAX);
638 		desiredvnodes = MAXVNODES_MAX;
639 	}
640 	wantfreevnodes = desiredvnodes / 4;
641 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
642 	TAILQ_INIT(&vnode_list);
643 	mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
644 	/*
645 	 * The lock is taken to appease WITNESS.
646 	 */
647 	mtx_lock(&vnode_list_mtx);
648 	vnlru_recalc();
649 	mtx_unlock(&vnode_list_mtx);
650 	vnode_list_free_marker = vn_alloc_marker(NULL);
651 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
652 	vnode_list_reclaim_marker = vn_alloc_marker(NULL);
653 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
654 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
655 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
656 	uma_zone_set_smr(vnode_zone, vfs_smr);
657 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
658 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
659 	/*
660 	 * Preallocate enough nodes to support one-per buf so that
661 	 * we can not fail an insert.  reassignbuf() callers can not
662 	 * tolerate the insertion failure.
663 	 */
664 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
665 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
666 	    UMA_ZONE_NOFREE | UMA_ZONE_SMR);
667 	buf_trie_smr = uma_zone_get_smr(buf_trie_zone);
668 	uma_prealloc(buf_trie_zone, nbuf);
669 
670 	vnodes_created = counter_u64_alloc(M_WAITOK);
671 	recycles_count = counter_u64_alloc(M_WAITOK);
672 	recycles_free_count = counter_u64_alloc(M_WAITOK);
673 	deferred_inact = counter_u64_alloc(M_WAITOK);
674 
675 	/*
676 	 * Initialize the filesystem syncer.
677 	 */
678 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
679 	    &syncer_mask);
680 	syncer_maxdelay = syncer_mask + 1;
681 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
682 	cv_init(&sync_wakeup, "syncer");
683 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
684 		vnsz2log++;
685 	vnsz2log--;
686 
687 	CPU_FOREACH(cpu) {
688 		vd = DPCPU_ID_PTR((cpu), vd);
689 		bzero(vd, sizeof(*vd));
690 		mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
691 	}
692 }
693 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
694 
695 /*
696  * Mark a mount point as busy. Used to synchronize access and to delay
697  * unmounting. Eventually, mountlist_mtx is not released on failure.
698  *
699  * vfs_busy() is a custom lock, it can block the caller.
700  * vfs_busy() only sleeps if the unmount is active on the mount point.
701  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
702  * vnode belonging to mp.
703  *
704  * Lookup uses vfs_busy() to traverse mount points.
705  * root fs			var fs
706  * / vnode lock		A	/ vnode lock (/var)		D
707  * /var vnode lock	B	/log vnode lock(/var/log)	E
708  * vfs_busy lock	C	vfs_busy lock			F
709  *
710  * Within each file system, the lock order is C->A->B and F->D->E.
711  *
712  * When traversing across mounts, the system follows that lock order:
713  *
714  *        C->A->B
715  *              |
716  *              +->F->D->E
717  *
718  * The lookup() process for namei("/var") illustrates the process:
719  *  VOP_LOOKUP() obtains B while A is held
720  *  vfs_busy() obtains a shared lock on F while A and B are held
721  *  vput() releases lock on B
722  *  vput() releases lock on A
723  *  VFS_ROOT() obtains lock on D while shared lock on F is held
724  *  vfs_unbusy() releases shared lock on F
725  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
726  *    Attempt to lock A (instead of vp_crossmp) while D is held would
727  *    violate the global order, causing deadlocks.
728  *
729  * dounmount() locks B while F is drained.
730  */
731 int
732 vfs_busy(struct mount *mp, int flags)
733 {
734 
735 	MPASS((flags & ~MBF_MASK) == 0);
736 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
737 
738 	if (vfs_op_thread_enter(mp)) {
739 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
740 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
741 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
742 		vfs_mp_count_add_pcpu(mp, ref, 1);
743 		vfs_mp_count_add_pcpu(mp, lockref, 1);
744 		vfs_op_thread_exit(mp);
745 		if (flags & MBF_MNTLSTLOCK)
746 			mtx_unlock(&mountlist_mtx);
747 		return (0);
748 	}
749 
750 	MNT_ILOCK(mp);
751 	vfs_assert_mount_counters(mp);
752 	MNT_REF(mp);
753 	/*
754 	 * If mount point is currently being unmounted, sleep until the
755 	 * mount point fate is decided.  If thread doing the unmounting fails,
756 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
757 	 * that this mount point has survived the unmount attempt and vfs_busy
758 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
759 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
760 	 * about to be really destroyed.  vfs_busy needs to release its
761 	 * reference on the mount point in this case and return with ENOENT,
762 	 * telling the caller that mount mount it tried to busy is no longer
763 	 * valid.
764 	 */
765 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
766 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
767 			MNT_REL(mp);
768 			MNT_IUNLOCK(mp);
769 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
770 			    __func__);
771 			return (ENOENT);
772 		}
773 		if (flags & MBF_MNTLSTLOCK)
774 			mtx_unlock(&mountlist_mtx);
775 		mp->mnt_kern_flag |= MNTK_MWAIT;
776 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
777 		if (flags & MBF_MNTLSTLOCK)
778 			mtx_lock(&mountlist_mtx);
779 		MNT_ILOCK(mp);
780 	}
781 	if (flags & MBF_MNTLSTLOCK)
782 		mtx_unlock(&mountlist_mtx);
783 	mp->mnt_lockref++;
784 	MNT_IUNLOCK(mp);
785 	return (0);
786 }
787 
788 /*
789  * Free a busy filesystem.
790  */
791 void
792 vfs_unbusy(struct mount *mp)
793 {
794 	int c;
795 
796 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
797 
798 	if (vfs_op_thread_enter(mp)) {
799 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
800 		vfs_mp_count_sub_pcpu(mp, lockref, 1);
801 		vfs_mp_count_sub_pcpu(mp, ref, 1);
802 		vfs_op_thread_exit(mp);
803 		return;
804 	}
805 
806 	MNT_ILOCK(mp);
807 	vfs_assert_mount_counters(mp);
808 	MNT_REL(mp);
809 	c = --mp->mnt_lockref;
810 	if (mp->mnt_vfs_ops == 0) {
811 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
812 		MNT_IUNLOCK(mp);
813 		return;
814 	}
815 	if (c < 0)
816 		vfs_dump_mount_counters(mp);
817 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
818 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
819 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
820 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
821 		wakeup(&mp->mnt_lockref);
822 	}
823 	MNT_IUNLOCK(mp);
824 }
825 
826 /*
827  * Lookup a mount point by filesystem identifier.
828  */
829 struct mount *
830 vfs_getvfs(fsid_t *fsid)
831 {
832 	struct mount *mp;
833 
834 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
835 	mtx_lock(&mountlist_mtx);
836 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
837 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
838 			vfs_ref(mp);
839 			mtx_unlock(&mountlist_mtx);
840 			return (mp);
841 		}
842 	}
843 	mtx_unlock(&mountlist_mtx);
844 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
845 	return ((struct mount *) 0);
846 }
847 
848 /*
849  * Lookup a mount point by filesystem identifier, busying it before
850  * returning.
851  *
852  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
853  * cache for popular filesystem identifiers.  The cache is lockess, using
854  * the fact that struct mount's are never freed.  In worst case we may
855  * get pointer to unmounted or even different filesystem, so we have to
856  * check what we got, and go slow way if so.
857  */
858 struct mount *
859 vfs_busyfs(fsid_t *fsid)
860 {
861 #define	FSID_CACHE_SIZE	256
862 	typedef struct mount * volatile vmp_t;
863 	static vmp_t cache[FSID_CACHE_SIZE];
864 	struct mount *mp;
865 	int error;
866 	uint32_t hash;
867 
868 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
869 	hash = fsid->val[0] ^ fsid->val[1];
870 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
871 	mp = cache[hash];
872 	if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0)
873 		goto slow;
874 	if (vfs_busy(mp, 0) != 0) {
875 		cache[hash] = NULL;
876 		goto slow;
877 	}
878 	if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0)
879 		return (mp);
880 	else
881 	    vfs_unbusy(mp);
882 
883 slow:
884 	mtx_lock(&mountlist_mtx);
885 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
886 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
887 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
888 			if (error) {
889 				cache[hash] = NULL;
890 				mtx_unlock(&mountlist_mtx);
891 				return (NULL);
892 			}
893 			cache[hash] = mp;
894 			return (mp);
895 		}
896 	}
897 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
898 	mtx_unlock(&mountlist_mtx);
899 	return ((struct mount *) 0);
900 }
901 
902 /*
903  * Check if a user can access privileged mount options.
904  */
905 int
906 vfs_suser(struct mount *mp, struct thread *td)
907 {
908 	int error;
909 
910 	if (jailed(td->td_ucred)) {
911 		/*
912 		 * If the jail of the calling thread lacks permission for
913 		 * this type of file system, deny immediately.
914 		 */
915 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
916 			return (EPERM);
917 
918 		/*
919 		 * If the file system was mounted outside the jail of the
920 		 * calling thread, deny immediately.
921 		 */
922 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
923 			return (EPERM);
924 	}
925 
926 	/*
927 	 * If file system supports delegated administration, we don't check
928 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
929 	 * by the file system itself.
930 	 * If this is not the user that did original mount, we check for
931 	 * the PRIV_VFS_MOUNT_OWNER privilege.
932 	 */
933 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
934 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
935 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
936 			return (error);
937 	}
938 	return (0);
939 }
940 
941 /*
942  * Get a new unique fsid.  Try to make its val[0] unique, since this value
943  * will be used to create fake device numbers for stat().  Also try (but
944  * not so hard) make its val[0] unique mod 2^16, since some emulators only
945  * support 16-bit device numbers.  We end up with unique val[0]'s for the
946  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
947  *
948  * Keep in mind that several mounts may be running in parallel.  Starting
949  * the search one past where the previous search terminated is both a
950  * micro-optimization and a defense against returning the same fsid to
951  * different mounts.
952  */
953 void
954 vfs_getnewfsid(struct mount *mp)
955 {
956 	static uint16_t mntid_base;
957 	struct mount *nmp;
958 	fsid_t tfsid;
959 	int mtype;
960 
961 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
962 	mtx_lock(&mntid_mtx);
963 	mtype = mp->mnt_vfc->vfc_typenum;
964 	tfsid.val[1] = mtype;
965 	mtype = (mtype & 0xFF) << 24;
966 	for (;;) {
967 		tfsid.val[0] = makedev(255,
968 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
969 		mntid_base++;
970 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
971 			break;
972 		vfs_rel(nmp);
973 	}
974 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
975 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
976 	mtx_unlock(&mntid_mtx);
977 }
978 
979 /*
980  * Knob to control the precision of file timestamps:
981  *
982  *   0 = seconds only; nanoseconds zeroed.
983  *   1 = seconds and nanoseconds, accurate within 1/HZ.
984  *   2 = seconds and nanoseconds, truncated to microseconds.
985  * >=3 = seconds and nanoseconds, maximum precision.
986  */
987 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
988 
989 static int timestamp_precision = TSP_USEC;
990 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
991     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
992     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
993     "3+: sec + ns (max. precision))");
994 
995 /*
996  * Get a current timestamp.
997  */
998 void
999 vfs_timestamp(struct timespec *tsp)
1000 {
1001 	struct timeval tv;
1002 
1003 	switch (timestamp_precision) {
1004 	case TSP_SEC:
1005 		tsp->tv_sec = time_second;
1006 		tsp->tv_nsec = 0;
1007 		break;
1008 	case TSP_HZ:
1009 		getnanotime(tsp);
1010 		break;
1011 	case TSP_USEC:
1012 		microtime(&tv);
1013 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1014 		break;
1015 	case TSP_NSEC:
1016 	default:
1017 		nanotime(tsp);
1018 		break;
1019 	}
1020 }
1021 
1022 /*
1023  * Set vnode attributes to VNOVAL
1024  */
1025 void
1026 vattr_null(struct vattr *vap)
1027 {
1028 
1029 	vap->va_type = VNON;
1030 	vap->va_size = VNOVAL;
1031 	vap->va_bytes = VNOVAL;
1032 	vap->va_mode = VNOVAL;
1033 	vap->va_nlink = VNOVAL;
1034 	vap->va_uid = VNOVAL;
1035 	vap->va_gid = VNOVAL;
1036 	vap->va_fsid = VNOVAL;
1037 	vap->va_fileid = VNOVAL;
1038 	vap->va_blocksize = VNOVAL;
1039 	vap->va_rdev = VNOVAL;
1040 	vap->va_atime.tv_sec = VNOVAL;
1041 	vap->va_atime.tv_nsec = VNOVAL;
1042 	vap->va_mtime.tv_sec = VNOVAL;
1043 	vap->va_mtime.tv_nsec = VNOVAL;
1044 	vap->va_ctime.tv_sec = VNOVAL;
1045 	vap->va_ctime.tv_nsec = VNOVAL;
1046 	vap->va_birthtime.tv_sec = VNOVAL;
1047 	vap->va_birthtime.tv_nsec = VNOVAL;
1048 	vap->va_flags = VNOVAL;
1049 	vap->va_gen = VNOVAL;
1050 	vap->va_vaflags = 0;
1051 }
1052 
1053 /*
1054  * Try to reduce the total number of vnodes.
1055  *
1056  * This routine (and its user) are buggy in at least the following ways:
1057  * - all parameters were picked years ago when RAM sizes were significantly
1058  *   smaller
1059  * - it can pick vnodes based on pages used by the vm object, but filesystems
1060  *   like ZFS don't use it making the pick broken
1061  * - since ZFS has its own aging policy it gets partially combated by this one
1062  * - a dedicated method should be provided for filesystems to let them decide
1063  *   whether the vnode should be recycled
1064  *
1065  * This routine is called when we have too many vnodes.  It attempts
1066  * to free <count> vnodes and will potentially free vnodes that still
1067  * have VM backing store (VM backing store is typically the cause
1068  * of a vnode blowout so we want to do this).  Therefore, this operation
1069  * is not considered cheap.
1070  *
1071  * A number of conditions may prevent a vnode from being reclaimed.
1072  * the buffer cache may have references on the vnode, a directory
1073  * vnode may still have references due to the namei cache representing
1074  * underlying files, or the vnode may be in active use.   It is not
1075  * desirable to reuse such vnodes.  These conditions may cause the
1076  * number of vnodes to reach some minimum value regardless of what
1077  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
1078  *
1079  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1080  * 			 entries if this argument is strue
1081  * @param trigger	 Only reclaim vnodes with fewer than this many resident
1082  *			 pages.
1083  * @param target	 How many vnodes to reclaim.
1084  * @return		 The number of vnodes that were reclaimed.
1085  */
1086 static int
1087 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1088 {
1089 	struct vnode *vp, *mvp;
1090 	struct mount *mp;
1091 	struct vm_object *object;
1092 	u_long done;
1093 	bool retried;
1094 
1095 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1096 
1097 	retried = false;
1098 	done = 0;
1099 
1100 	mvp = vnode_list_reclaim_marker;
1101 restart:
1102 	vp = mvp;
1103 	while (done < target) {
1104 		vp = TAILQ_NEXT(vp, v_vnodelist);
1105 		if (__predict_false(vp == NULL))
1106 			break;
1107 
1108 		if (__predict_false(vp->v_type == VMARKER))
1109 			continue;
1110 
1111 		/*
1112 		 * If it's been deconstructed already, it's still
1113 		 * referenced, or it exceeds the trigger, skip it.
1114 		 * Also skip free vnodes.  We are trying to make space
1115 		 * to expand the free list, not reduce it.
1116 		 */
1117 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1118 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1119 			goto next_iter;
1120 
1121 		if (vp->v_type == VBAD || vp->v_type == VNON)
1122 			goto next_iter;
1123 
1124 		object = atomic_load_ptr(&vp->v_object);
1125 		if (object == NULL || object->resident_page_count > trigger) {
1126 			goto next_iter;
1127 		}
1128 
1129 		vhold(vp);
1130 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1131 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1132 		mtx_unlock(&vnode_list_mtx);
1133 
1134 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1135 			vdrop(vp);
1136 			goto next_iter_unlocked;
1137 		}
1138 		if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1139 			vdrop(vp);
1140 			vn_finished_write(mp);
1141 			goto next_iter_unlocked;
1142 		}
1143 
1144 		VI_LOCK(vp);
1145 		if (vp->v_usecount > 0 ||
1146 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1147 		    (vp->v_object != NULL &&
1148 		    vp->v_object->resident_page_count > trigger)) {
1149 			VOP_UNLOCK(vp);
1150 			vdropl(vp);
1151 			vn_finished_write(mp);
1152 			goto next_iter_unlocked;
1153 		}
1154 		counter_u64_add(recycles_count, 1);
1155 		vgonel(vp);
1156 		VOP_UNLOCK(vp);
1157 		vdropl(vp);
1158 		vn_finished_write(mp);
1159 		done++;
1160 next_iter_unlocked:
1161 		if (should_yield())
1162 			kern_yield(PRI_USER);
1163 		mtx_lock(&vnode_list_mtx);
1164 		goto restart;
1165 next_iter:
1166 		MPASS(vp->v_type != VMARKER);
1167 		if (!should_yield())
1168 			continue;
1169 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1170 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1171 		mtx_unlock(&vnode_list_mtx);
1172 		kern_yield(PRI_USER);
1173 		mtx_lock(&vnode_list_mtx);
1174 		goto restart;
1175 	}
1176 	if (done == 0 && !retried) {
1177 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1178 		TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1179 		retried = true;
1180 		goto restart;
1181 	}
1182 	return (done);
1183 }
1184 
1185 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1186 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1187     0,
1188     "limit on vnode free requests per call to the vnlru_free routine");
1189 
1190 /*
1191  * Attempt to reduce the free list by the requested amount.
1192  */
1193 static int
1194 vnlru_free_locked(int count, struct vfsops *mnt_op)
1195 {
1196 	struct vnode *vp, *mvp;
1197 	struct mount *mp;
1198 	int ocount;
1199 
1200 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1201 	if (count > max_vnlru_free)
1202 		count = max_vnlru_free;
1203 	ocount = count;
1204 	mvp = vnode_list_free_marker;
1205 restart:
1206 	vp = mvp;
1207 	while (count > 0) {
1208 		vp = TAILQ_NEXT(vp, v_vnodelist);
1209 		if (__predict_false(vp == NULL)) {
1210 			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1211 			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1212 			break;
1213 		}
1214 		if (__predict_false(vp->v_type == VMARKER))
1215 			continue;
1216 
1217 		/*
1218 		 * Don't recycle if our vnode is from different type
1219 		 * of mount point.  Note that mp is type-safe, the
1220 		 * check does not reach unmapped address even if
1221 		 * vnode is reclaimed.
1222 		 * Don't recycle if we can't get the interlock without
1223 		 * blocking.
1224 		 */
1225 		if (vp->v_holdcnt > 0 || (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1226 		    mp->mnt_op != mnt_op)) {
1227 			continue;
1228 		}
1229 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1230 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1231 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1232 			continue;
1233 		}
1234 		vhold(vp);
1235 		count--;
1236 		mtx_unlock(&vnode_list_mtx);
1237 		vtryrecycle(vp);
1238 		mtx_lock(&vnode_list_mtx);
1239 		goto restart;
1240 	}
1241 	return (ocount - count);
1242 }
1243 
1244 void
1245 vnlru_free(int count, struct vfsops *mnt_op)
1246 {
1247 
1248 	mtx_lock(&vnode_list_mtx);
1249 	vnlru_free_locked(count, mnt_op);
1250 	mtx_unlock(&vnode_list_mtx);
1251 }
1252 
1253 static void
1254 vnlru_recalc(void)
1255 {
1256 
1257 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1258 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1259 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1260 	vlowat = vhiwat / 2;
1261 }
1262 
1263 /*
1264  * Attempt to recycle vnodes in a context that is always safe to block.
1265  * Calling vlrurecycle() from the bowels of filesystem code has some
1266  * interesting deadlock problems.
1267  */
1268 static struct proc *vnlruproc;
1269 static int vnlruproc_sig;
1270 
1271 /*
1272  * The main freevnodes counter is only updated when threads requeue their vnode
1273  * batches. CPUs are conditionally walked to compute a more accurate total.
1274  *
1275  * Limit how much of a slop are we willing to tolerate. Note: the actual value
1276  * at any given moment can still exceed slop, but it should not be by significant
1277  * margin in practice.
1278  */
1279 #define VNLRU_FREEVNODES_SLOP 128
1280 
1281 static __inline void
1282 vn_freevnodes_inc(void)
1283 {
1284 	struct vdbatch *vd;
1285 
1286 	critical_enter();
1287 	vd = DPCPU_PTR(vd);
1288 	vd->freevnodes++;
1289 	critical_exit();
1290 }
1291 
1292 static __inline void
1293 vn_freevnodes_dec(void)
1294 {
1295 	struct vdbatch *vd;
1296 
1297 	critical_enter();
1298 	vd = DPCPU_PTR(vd);
1299 	vd->freevnodes--;
1300 	critical_exit();
1301 }
1302 
1303 static u_long
1304 vnlru_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 
1325 static bool
1326 vnlru_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 
1342 static bool
1343 vnlru_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 
1359 static void
1360 vnlru_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 
1370 static void
1371 vnlru_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 
1474 static struct kproc_desc vnlru_kp = {
1475 	"vnlru",
1476 	vnlru_proc,
1477 	&vnlruproc
1478 };
1479 SYSINIT(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  */
1492 static int
1493 vtryrecycle(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 		vdrop(vp);
1509 		return (EWOULDBLOCK);
1510 	}
1511 	/*
1512 	 * Don't recycle if its filesystem is being suspended.
1513 	 */
1514 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1515 		VOP_UNLOCK(vp);
1516 		CTR2(KTR_VFS,
1517 		    "%s: impossible to recycle, cannot start the write for %p",
1518 		    __func__, vp);
1519 		vdrop(vp);
1520 		return (EBUSY);
1521 	}
1522 	/*
1523 	 * If we got this far, we need to acquire the interlock and see if
1524 	 * anyone picked up this vnode from another list.  If not, we will
1525 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1526 	 * will skip over it.
1527 	 */
1528 	VI_LOCK(vp);
1529 	if (vp->v_usecount) {
1530 		VOP_UNLOCK(vp);
1531 		vdropl(vp);
1532 		vn_finished_write(vnmp);
1533 		CTR2(KTR_VFS,
1534 		    "%s: impossible to recycle, %p is already referenced",
1535 		    __func__, vp);
1536 		return (EBUSY);
1537 	}
1538 	if (!VN_IS_DOOMED(vp)) {
1539 		counter_u64_add(recycles_free_count, 1);
1540 		vgonel(vp);
1541 	}
1542 	VOP_UNLOCK(vp);
1543 	vdropl(vp);
1544 	vn_finished_write(vnmp);
1545 	return (0);
1546 }
1547 
1548 /*
1549  * Allocate a new vnode.
1550  *
1551  * The operation never returns an error. Returning an error was disabled
1552  * in r145385 (dated 2005) with the following comment:
1553  *
1554  * XXX Not all VFS_VGET/ffs_vget callers check returns.
1555  *
1556  * Given the age of this commit (almost 15 years at the time of writing this
1557  * comment) restoring the ability to fail requires a significant audit of
1558  * all codepaths.
1559  *
1560  * The routine can try to free a vnode or stall for up to 1 second waiting for
1561  * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1562  */
1563 static u_long vn_alloc_cyclecount;
1564 
1565 static struct vnode * __noinline
1566 vn_alloc_hard(struct mount *mp)
1567 {
1568 	u_long rnumvnodes, rfreevnodes;
1569 
1570 	mtx_lock(&vnode_list_mtx);
1571 	rnumvnodes = atomic_load_long(&numvnodes);
1572 	if (rnumvnodes + 1 < desiredvnodes) {
1573 		vn_alloc_cyclecount = 0;
1574 		goto alloc;
1575 	}
1576 	rfreevnodes = vnlru_read_freevnodes();
1577 	if (vn_alloc_cyclecount++ >= rfreevnodes) {
1578 		vn_alloc_cyclecount = 0;
1579 		vstir = 1;
1580 	}
1581 	/*
1582 	 * Grow the vnode cache if it will not be above its target max
1583 	 * after growing.  Otherwise, if the free list is nonempty, try
1584 	 * to reclaim 1 item from it before growing the cache (possibly
1585 	 * above its target max if the reclamation failed or is delayed).
1586 	 * Otherwise, wait for some space.  In all cases, schedule
1587 	 * vnlru_proc() if we are getting short of space.  The watermarks
1588 	 * should be chosen so that we never wait or even reclaim from
1589 	 * the free list to below its target minimum.
1590 	 */
1591 	if (vnlru_free_locked(1, NULL) > 0)
1592 		goto alloc;
1593 	if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1594 		/*
1595 		 * Wait for space for a new vnode.
1596 		 */
1597 		vnlru_kick();
1598 		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1599 		if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1600 		    vnlru_read_freevnodes() > 1)
1601 			vnlru_free_locked(1, NULL);
1602 	}
1603 alloc:
1604 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1605 	if (vnlru_under(rnumvnodes, vlowat))
1606 		vnlru_kick();
1607 	mtx_unlock(&vnode_list_mtx);
1608 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1609 }
1610 
1611 static struct vnode *
1612 vn_alloc(struct mount *mp)
1613 {
1614 	u_long rnumvnodes;
1615 
1616 	if (__predict_false(vn_alloc_cyclecount != 0))
1617 		return (vn_alloc_hard(mp));
1618 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1619 	if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1620 		atomic_subtract_long(&numvnodes, 1);
1621 		return (vn_alloc_hard(mp));
1622 	}
1623 
1624 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1625 }
1626 
1627 static void
1628 vn_free(struct vnode *vp)
1629 {
1630 
1631 	atomic_subtract_long(&numvnodes, 1);
1632 	uma_zfree_smr(vnode_zone, vp);
1633 }
1634 
1635 /*
1636  * Return the next vnode from the free list.
1637  */
1638 int
1639 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1640     struct vnode **vpp)
1641 {
1642 	struct vnode *vp;
1643 	struct thread *td;
1644 	struct lock_object *lo;
1645 
1646 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1647 
1648 	KASSERT(vops->registered,
1649 	    ("%s: not registered vector op %p\n", __func__, vops));
1650 
1651 	td = curthread;
1652 	if (td->td_vp_reserved != NULL) {
1653 		vp = td->td_vp_reserved;
1654 		td->td_vp_reserved = NULL;
1655 	} else {
1656 		vp = vn_alloc(mp);
1657 	}
1658 	counter_u64_add(vnodes_created, 1);
1659 	/*
1660 	 * Locks are given the generic name "vnode" when created.
1661 	 * Follow the historic practice of using the filesystem
1662 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1663 	 *
1664 	 * Locks live in a witness group keyed on their name. Thus,
1665 	 * when a lock is renamed, it must also move from the witness
1666 	 * group of its old name to the witness group of its new name.
1667 	 *
1668 	 * The change only needs to be made when the vnode moves
1669 	 * from one filesystem type to another. We ensure that each
1670 	 * filesystem use a single static name pointer for its tag so
1671 	 * that we can compare pointers rather than doing a strcmp().
1672 	 */
1673 	lo = &vp->v_vnlock->lock_object;
1674 #ifdef WITNESS
1675 	if (lo->lo_name != tag) {
1676 #endif
1677 		lo->lo_name = tag;
1678 #ifdef WITNESS
1679 		WITNESS_DESTROY(lo);
1680 		WITNESS_INIT(lo, tag);
1681 	}
1682 #endif
1683 	/*
1684 	 * By default, don't allow shared locks unless filesystems opt-in.
1685 	 */
1686 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1687 	/*
1688 	 * Finalize various vnode identity bits.
1689 	 */
1690 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1691 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1692 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1693 	vp->v_type = VNON;
1694 	vp->v_op = vops;
1695 	v_init_counters(vp);
1696 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1697 #ifdef DIAGNOSTIC
1698 	if (mp == NULL && vops != &dead_vnodeops)
1699 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1700 #endif
1701 #ifdef MAC
1702 	mac_vnode_init(vp);
1703 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1704 		mac_vnode_associate_singlelabel(mp, vp);
1705 #endif
1706 	if (mp != NULL) {
1707 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1708 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1709 			vp->v_vflag |= VV_NOKNOTE;
1710 	}
1711 
1712 	/*
1713 	 * For the filesystems which do not use vfs_hash_insert(),
1714 	 * still initialize v_hash to have vfs_hash_index() useful.
1715 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1716 	 * its own hashing.
1717 	 */
1718 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1719 
1720 	*vpp = vp;
1721 	return (0);
1722 }
1723 
1724 void
1725 getnewvnode_reserve(void)
1726 {
1727 	struct thread *td;
1728 
1729 	td = curthread;
1730 	MPASS(td->td_vp_reserved == NULL);
1731 	td->td_vp_reserved = vn_alloc(NULL);
1732 }
1733 
1734 void
1735 getnewvnode_drop_reserve(void)
1736 {
1737 	struct thread *td;
1738 
1739 	td = curthread;
1740 	if (td->td_vp_reserved != NULL) {
1741 		vn_free(td->td_vp_reserved);
1742 		td->td_vp_reserved = NULL;
1743 	}
1744 }
1745 
1746 static void __noinline
1747 freevnode(struct vnode *vp)
1748 {
1749 	struct bufobj *bo;
1750 
1751 	/*
1752 	 * The vnode has been marked for destruction, so free it.
1753 	 *
1754 	 * The vnode will be returned to the zone where it will
1755 	 * normally remain until it is needed for another vnode. We
1756 	 * need to cleanup (or verify that the cleanup has already
1757 	 * been done) any residual data left from its current use
1758 	 * so as not to contaminate the freshly allocated vnode.
1759 	 */
1760 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1761 	/*
1762 	 * Paired with vgone.
1763 	 */
1764 	vn_seqc_write_end_locked(vp);
1765 	VNPASS(vp->v_seqc_users == 0, vp);
1766 
1767 	bo = &vp->v_bufobj;
1768 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1769 	VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp);
1770 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1771 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1772 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1773 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1774 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1775 	    ("clean blk trie not empty"));
1776 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1777 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1778 	    ("dirty blk trie not empty"));
1779 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1780 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1781 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1782 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1783 	    ("Dangling rangelock waiters"));
1784 	VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp,
1785 	    ("Leaked inactivation"));
1786 	VI_UNLOCK(vp);
1787 #ifdef MAC
1788 	mac_vnode_destroy(vp);
1789 #endif
1790 	if (vp->v_pollinfo != NULL) {
1791 		destroy_vpollinfo(vp->v_pollinfo);
1792 		vp->v_pollinfo = NULL;
1793 	}
1794 #ifdef INVARIANTS
1795 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
1796 	vp->v_op = NULL;
1797 #endif
1798 	vp->v_mountedhere = NULL;
1799 	vp->v_unpcb = NULL;
1800 	vp->v_rdev = NULL;
1801 	vp->v_fifoinfo = NULL;
1802 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
1803 	vp->v_irflag = 0;
1804 	vp->v_iflag = 0;
1805 	vp->v_vflag = 0;
1806 	bo->bo_flag = 0;
1807 	vn_free(vp);
1808 }
1809 
1810 /*
1811  * Delete from old mount point vnode list, if on one.
1812  */
1813 static void
1814 delmntque(struct vnode *vp)
1815 {
1816 	struct mount *mp;
1817 
1818 	VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
1819 
1820 	mp = vp->v_mount;
1821 	if (mp == NULL)
1822 		return;
1823 	MNT_ILOCK(mp);
1824 	VI_LOCK(vp);
1825 	vp->v_mount = NULL;
1826 	VI_UNLOCK(vp);
1827 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1828 		("bad mount point vnode list size"));
1829 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1830 	mp->mnt_nvnodelistsize--;
1831 	MNT_REL(mp);
1832 	MNT_IUNLOCK(mp);
1833 }
1834 
1835 static void
1836 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1837 {
1838 
1839 	vp->v_data = NULL;
1840 	vp->v_op = &dead_vnodeops;
1841 	vgone(vp);
1842 	vput(vp);
1843 }
1844 
1845 /*
1846  * Insert into list of vnodes for the new mount point, if available.
1847  */
1848 int
1849 insmntque1(struct vnode *vp, struct mount *mp,
1850 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1851 {
1852 
1853 	KASSERT(vp->v_mount == NULL,
1854 		("insmntque: vnode already on per mount vnode list"));
1855 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1856 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1857 
1858 	/*
1859 	 * We acquire the vnode interlock early to ensure that the
1860 	 * vnode cannot be recycled by another process releasing a
1861 	 * holdcnt on it before we get it on both the vnode list
1862 	 * and the active vnode list. The mount mutex protects only
1863 	 * manipulation of the vnode list and the vnode freelist
1864 	 * mutex protects only manipulation of the active vnode list.
1865 	 * Hence the need to hold the vnode interlock throughout.
1866 	 */
1867 	MNT_ILOCK(mp);
1868 	VI_LOCK(vp);
1869 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1870 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1871 	    mp->mnt_nvnodelistsize == 0)) &&
1872 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1873 		VI_UNLOCK(vp);
1874 		MNT_IUNLOCK(mp);
1875 		if (dtr != NULL)
1876 			dtr(vp, dtr_arg);
1877 		return (EBUSY);
1878 	}
1879 	vp->v_mount = mp;
1880 	MNT_REF(mp);
1881 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1882 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1883 		("neg mount point vnode list size"));
1884 	mp->mnt_nvnodelistsize++;
1885 	VI_UNLOCK(vp);
1886 	MNT_IUNLOCK(mp);
1887 	return (0);
1888 }
1889 
1890 int
1891 insmntque(struct vnode *vp, struct mount *mp)
1892 {
1893 
1894 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1895 }
1896 
1897 /*
1898  * Flush out and invalidate all buffers associated with a bufobj
1899  * Called with the underlying object locked.
1900  */
1901 int
1902 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1903 {
1904 	int error;
1905 
1906 	BO_LOCK(bo);
1907 	if (flags & V_SAVE) {
1908 		error = bufobj_wwait(bo, slpflag, slptimeo);
1909 		if (error) {
1910 			BO_UNLOCK(bo);
1911 			return (error);
1912 		}
1913 		if (bo->bo_dirty.bv_cnt > 0) {
1914 			BO_UNLOCK(bo);
1915 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1916 				return (error);
1917 			/*
1918 			 * XXX We could save a lock/unlock if this was only
1919 			 * enabled under INVARIANTS
1920 			 */
1921 			BO_LOCK(bo);
1922 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1923 				panic("vinvalbuf: dirty bufs");
1924 		}
1925 	}
1926 	/*
1927 	 * If you alter this loop please notice that interlock is dropped and
1928 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1929 	 * no race conditions occur from this.
1930 	 */
1931 	do {
1932 		error = flushbuflist(&bo->bo_clean,
1933 		    flags, bo, slpflag, slptimeo);
1934 		if (error == 0 && !(flags & V_CLEANONLY))
1935 			error = flushbuflist(&bo->bo_dirty,
1936 			    flags, bo, slpflag, slptimeo);
1937 		if (error != 0 && error != EAGAIN) {
1938 			BO_UNLOCK(bo);
1939 			return (error);
1940 		}
1941 	} while (error != 0);
1942 
1943 	/*
1944 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1945 	 * have write I/O in-progress but if there is a VM object then the
1946 	 * VM object can also have read-I/O in-progress.
1947 	 */
1948 	do {
1949 		bufobj_wwait(bo, 0, 0);
1950 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
1951 			BO_UNLOCK(bo);
1952 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
1953 			BO_LOCK(bo);
1954 		}
1955 	} while (bo->bo_numoutput > 0);
1956 	BO_UNLOCK(bo);
1957 
1958 	/*
1959 	 * Destroy the copy in the VM cache, too.
1960 	 */
1961 	if (bo->bo_object != NULL &&
1962 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1963 		VM_OBJECT_WLOCK(bo->bo_object);
1964 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1965 		    OBJPR_CLEANONLY : 0);
1966 		VM_OBJECT_WUNLOCK(bo->bo_object);
1967 	}
1968 
1969 #ifdef INVARIANTS
1970 	BO_LOCK(bo);
1971 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1972 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1973 	    bo->bo_clean.bv_cnt > 0))
1974 		panic("vinvalbuf: flush failed");
1975 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1976 	    bo->bo_dirty.bv_cnt > 0)
1977 		panic("vinvalbuf: flush dirty failed");
1978 	BO_UNLOCK(bo);
1979 #endif
1980 	return (0);
1981 }
1982 
1983 /*
1984  * Flush out and invalidate all buffers associated with a vnode.
1985  * Called with the underlying object locked.
1986  */
1987 int
1988 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1989 {
1990 
1991 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1992 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1993 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1994 		return (0);
1995 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1996 }
1997 
1998 /*
1999  * Flush out buffers on the specified list.
2000  *
2001  */
2002 static int
2003 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2004     int slptimeo)
2005 {
2006 	struct buf *bp, *nbp;
2007 	int retval, error;
2008 	daddr_t lblkno;
2009 	b_xflags_t xflags;
2010 
2011 	ASSERT_BO_WLOCKED(bo);
2012 
2013 	retval = 0;
2014 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2015 		/*
2016 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2017 		 * do not skip any buffers. If we are flushing only V_NORMAL
2018 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2019 		 * flushing only V_ALT buffers then skip buffers not marked
2020 		 * as BX_ALTDATA.
2021 		 */
2022 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2023 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2024 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2025 			continue;
2026 		}
2027 		if (nbp != NULL) {
2028 			lblkno = nbp->b_lblkno;
2029 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2030 		}
2031 		retval = EAGAIN;
2032 		error = BUF_TIMELOCK(bp,
2033 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2034 		    "flushbuf", slpflag, slptimeo);
2035 		if (error) {
2036 			BO_LOCK(bo);
2037 			return (error != ENOLCK ? error : EAGAIN);
2038 		}
2039 		KASSERT(bp->b_bufobj == bo,
2040 		    ("bp %p wrong b_bufobj %p should be %p",
2041 		    bp, bp->b_bufobj, bo));
2042 		/*
2043 		 * XXX Since there are no node locks for NFS, I
2044 		 * believe there is a slight chance that a delayed
2045 		 * write will occur while sleeping just above, so
2046 		 * check for it.
2047 		 */
2048 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2049 		    (flags & V_SAVE)) {
2050 			bremfree(bp);
2051 			bp->b_flags |= B_ASYNC;
2052 			bwrite(bp);
2053 			BO_LOCK(bo);
2054 			return (EAGAIN);	/* XXX: why not loop ? */
2055 		}
2056 		bremfree(bp);
2057 		bp->b_flags |= (B_INVAL | B_RELBUF);
2058 		bp->b_flags &= ~B_ASYNC;
2059 		brelse(bp);
2060 		BO_LOCK(bo);
2061 		if (nbp == NULL)
2062 			break;
2063 		nbp = gbincore(bo, lblkno);
2064 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2065 		    != xflags)
2066 			break;			/* nbp invalid */
2067 	}
2068 	return (retval);
2069 }
2070 
2071 int
2072 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2073 {
2074 	struct buf *bp;
2075 	int error;
2076 	daddr_t lblkno;
2077 
2078 	ASSERT_BO_LOCKED(bo);
2079 
2080 	for (lblkno = startn;;) {
2081 again:
2082 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2083 		if (bp == NULL || bp->b_lblkno >= endn ||
2084 		    bp->b_lblkno < startn)
2085 			break;
2086 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2087 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2088 		if (error != 0) {
2089 			BO_RLOCK(bo);
2090 			if (error == ENOLCK)
2091 				goto again;
2092 			return (error);
2093 		}
2094 		KASSERT(bp->b_bufobj == bo,
2095 		    ("bp %p wrong b_bufobj %p should be %p",
2096 		    bp, bp->b_bufobj, bo));
2097 		lblkno = bp->b_lblkno + 1;
2098 		if ((bp->b_flags & B_MANAGED) == 0)
2099 			bremfree(bp);
2100 		bp->b_flags |= B_RELBUF;
2101 		/*
2102 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2103 		 * pages backing each buffer in the range are unlikely to be
2104 		 * reused.  Dirty buffers will have the hint applied once
2105 		 * they've been written.
2106 		 */
2107 		if ((bp->b_flags & B_VMIO) != 0)
2108 			bp->b_flags |= B_NOREUSE;
2109 		brelse(bp);
2110 		BO_RLOCK(bo);
2111 	}
2112 	return (0);
2113 }
2114 
2115 /*
2116  * Truncate a file's buffer and pages to a specified length.  This
2117  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2118  * sync activity.
2119  */
2120 int
2121 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2122 {
2123 	struct buf *bp, *nbp;
2124 	struct bufobj *bo;
2125 	daddr_t startlbn;
2126 
2127 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2128 	    vp, blksize, (uintmax_t)length);
2129 
2130 	/*
2131 	 * Round up to the *next* lbn.
2132 	 */
2133 	startlbn = howmany(length, blksize);
2134 
2135 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2136 
2137 	bo = &vp->v_bufobj;
2138 restart_unlocked:
2139 	BO_LOCK(bo);
2140 
2141 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2142 		;
2143 
2144 	if (length > 0) {
2145 restartsync:
2146 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2147 			if (bp->b_lblkno > 0)
2148 				continue;
2149 			/*
2150 			 * Since we hold the vnode lock this should only
2151 			 * fail if we're racing with the buf daemon.
2152 			 */
2153 			if (BUF_LOCK(bp,
2154 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2155 			    BO_LOCKPTR(bo)) == ENOLCK)
2156 				goto restart_unlocked;
2157 
2158 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2159 			    ("buf(%p) on dirty queue without DELWRI", bp));
2160 
2161 			bremfree(bp);
2162 			bawrite(bp);
2163 			BO_LOCK(bo);
2164 			goto restartsync;
2165 		}
2166 	}
2167 
2168 	bufobj_wwait(bo, 0, 0);
2169 	BO_UNLOCK(bo);
2170 	vnode_pager_setsize(vp, length);
2171 
2172 	return (0);
2173 }
2174 
2175 /*
2176  * Invalidate the cached pages of a file's buffer within the range of block
2177  * numbers [startlbn, endlbn).
2178  */
2179 void
2180 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2181     int blksize)
2182 {
2183 	struct bufobj *bo;
2184 	off_t start, end;
2185 
2186 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2187 
2188 	start = blksize * startlbn;
2189 	end = blksize * endlbn;
2190 
2191 	bo = &vp->v_bufobj;
2192 	BO_LOCK(bo);
2193 	MPASS(blksize == bo->bo_bsize);
2194 
2195 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2196 		;
2197 
2198 	BO_UNLOCK(bo);
2199 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2200 }
2201 
2202 static int
2203 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2204     daddr_t startlbn, daddr_t endlbn)
2205 {
2206 	struct buf *bp, *nbp;
2207 	bool anyfreed;
2208 
2209 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2210 	ASSERT_BO_LOCKED(bo);
2211 
2212 	do {
2213 		anyfreed = false;
2214 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2215 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2216 				continue;
2217 			if (BUF_LOCK(bp,
2218 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2219 			    BO_LOCKPTR(bo)) == ENOLCK) {
2220 				BO_LOCK(bo);
2221 				return (EAGAIN);
2222 			}
2223 
2224 			bremfree(bp);
2225 			bp->b_flags |= B_INVAL | B_RELBUF;
2226 			bp->b_flags &= ~B_ASYNC;
2227 			brelse(bp);
2228 			anyfreed = true;
2229 
2230 			BO_LOCK(bo);
2231 			if (nbp != NULL &&
2232 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2233 			    nbp->b_vp != vp ||
2234 			    (nbp->b_flags & B_DELWRI) != 0))
2235 				return (EAGAIN);
2236 		}
2237 
2238 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2239 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2240 				continue;
2241 			if (BUF_LOCK(bp,
2242 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2243 			    BO_LOCKPTR(bo)) == ENOLCK) {
2244 				BO_LOCK(bo);
2245 				return (EAGAIN);
2246 			}
2247 			bremfree(bp);
2248 			bp->b_flags |= B_INVAL | B_RELBUF;
2249 			bp->b_flags &= ~B_ASYNC;
2250 			brelse(bp);
2251 			anyfreed = true;
2252 
2253 			BO_LOCK(bo);
2254 			if (nbp != NULL &&
2255 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2256 			    (nbp->b_vp != vp) ||
2257 			    (nbp->b_flags & B_DELWRI) == 0))
2258 				return (EAGAIN);
2259 		}
2260 	} while (anyfreed);
2261 	return (0);
2262 }
2263 
2264 static void
2265 buf_vlist_remove(struct buf *bp)
2266 {
2267 	struct bufv *bv;
2268 	b_xflags_t flags;
2269 
2270 	flags = bp->b_xflags;
2271 
2272 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2273 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2274 	KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 &&
2275 	    (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN),
2276 	    ("%s: buffer %p has invalid queue state", __func__, bp));
2277 
2278 	if ((flags & BX_VNDIRTY) != 0)
2279 		bv = &bp->b_bufobj->bo_dirty;
2280 	else
2281 		bv = &bp->b_bufobj->bo_clean;
2282 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2283 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2284 	bv->bv_cnt--;
2285 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2286 }
2287 
2288 /*
2289  * Add the buffer to the sorted clean or dirty block list.
2290  *
2291  * NOTE: xflags is passed as a constant, optimizing this inline function!
2292  */
2293 static void
2294 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2295 {
2296 	struct bufv *bv;
2297 	struct buf *n;
2298 	int error;
2299 
2300 	ASSERT_BO_WLOCKED(bo);
2301 	KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2302 	    ("buf_vlist_add: bo %p does not allow bufs", bo));
2303 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2304 	    ("dead bo %p", bo));
2305 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2306 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2307 	bp->b_xflags |= xflags;
2308 	if (xflags & BX_VNDIRTY)
2309 		bv = &bo->bo_dirty;
2310 	else
2311 		bv = &bo->bo_clean;
2312 
2313 	/*
2314 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2315 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2316 	 * than _ge.
2317 	 */
2318 	if (bv->bv_cnt == 0 ||
2319 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2320 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2321 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2322 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2323 	else
2324 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2325 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2326 	if (error)
2327 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2328 	bv->bv_cnt++;
2329 }
2330 
2331 /*
2332  * Look up a buffer using the buffer tries.
2333  */
2334 struct buf *
2335 gbincore(struct bufobj *bo, daddr_t lblkno)
2336 {
2337 	struct buf *bp;
2338 
2339 	ASSERT_BO_LOCKED(bo);
2340 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2341 	if (bp != NULL)
2342 		return (bp);
2343 	return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno));
2344 }
2345 
2346 /*
2347  * Look up a buf using the buffer tries, without the bufobj lock.  This relies
2348  * on SMR for safe lookup, and bufs being in a no-free zone to provide type
2349  * stability of the result.  Like other lockless lookups, the found buf may
2350  * already be invalid by the time this function returns.
2351  */
2352 struct buf *
2353 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno)
2354 {
2355 	struct buf *bp;
2356 
2357 	ASSERT_BO_UNLOCKED(bo);
2358 	bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno);
2359 	if (bp != NULL)
2360 		return (bp);
2361 	return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno));
2362 }
2363 
2364 /*
2365  * Associate a buffer with a vnode.
2366  */
2367 void
2368 bgetvp(struct vnode *vp, struct buf *bp)
2369 {
2370 	struct bufobj *bo;
2371 
2372 	bo = &vp->v_bufobj;
2373 	ASSERT_BO_WLOCKED(bo);
2374 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2375 
2376 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2377 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2378 	    ("bgetvp: bp already attached! %p", bp));
2379 
2380 	vhold(vp);
2381 	bp->b_vp = vp;
2382 	bp->b_bufobj = bo;
2383 	/*
2384 	 * Insert onto list for new vnode.
2385 	 */
2386 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2387 }
2388 
2389 /*
2390  * Disassociate a buffer from a vnode.
2391  */
2392 void
2393 brelvp(struct buf *bp)
2394 {
2395 	struct bufobj *bo;
2396 	struct vnode *vp;
2397 
2398 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2399 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2400 
2401 	/*
2402 	 * Delete from old vnode list, if on one.
2403 	 */
2404 	vp = bp->b_vp;		/* XXX */
2405 	bo = bp->b_bufobj;
2406 	BO_LOCK(bo);
2407 	buf_vlist_remove(bp);
2408 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2409 		bo->bo_flag &= ~BO_ONWORKLST;
2410 		mtx_lock(&sync_mtx);
2411 		LIST_REMOVE(bo, bo_synclist);
2412 		syncer_worklist_len--;
2413 		mtx_unlock(&sync_mtx);
2414 	}
2415 	bp->b_vp = NULL;
2416 	bp->b_bufobj = NULL;
2417 	BO_UNLOCK(bo);
2418 	vdrop(vp);
2419 }
2420 
2421 /*
2422  * Add an item to the syncer work queue.
2423  */
2424 static void
2425 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2426 {
2427 	int slot;
2428 
2429 	ASSERT_BO_WLOCKED(bo);
2430 
2431 	mtx_lock(&sync_mtx);
2432 	if (bo->bo_flag & BO_ONWORKLST)
2433 		LIST_REMOVE(bo, bo_synclist);
2434 	else {
2435 		bo->bo_flag |= BO_ONWORKLST;
2436 		syncer_worklist_len++;
2437 	}
2438 
2439 	if (delay > syncer_maxdelay - 2)
2440 		delay = syncer_maxdelay - 2;
2441 	slot = (syncer_delayno + delay) & syncer_mask;
2442 
2443 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2444 	mtx_unlock(&sync_mtx);
2445 }
2446 
2447 static int
2448 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2449 {
2450 	int error, len;
2451 
2452 	mtx_lock(&sync_mtx);
2453 	len = syncer_worklist_len - sync_vnode_count;
2454 	mtx_unlock(&sync_mtx);
2455 	error = SYSCTL_OUT(req, &len, sizeof(len));
2456 	return (error);
2457 }
2458 
2459 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2460     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2461     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2462 
2463 static struct proc *updateproc;
2464 static void sched_sync(void);
2465 static struct kproc_desc up_kp = {
2466 	"syncer",
2467 	sched_sync,
2468 	&updateproc
2469 };
2470 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2471 
2472 static int
2473 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2474 {
2475 	struct vnode *vp;
2476 	struct mount *mp;
2477 
2478 	*bo = LIST_FIRST(slp);
2479 	if (*bo == NULL)
2480 		return (0);
2481 	vp = bo2vnode(*bo);
2482 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2483 		return (1);
2484 	/*
2485 	 * We use vhold in case the vnode does not
2486 	 * successfully sync.  vhold prevents the vnode from
2487 	 * going away when we unlock the sync_mtx so that
2488 	 * we can acquire the vnode interlock.
2489 	 */
2490 	vholdl(vp);
2491 	mtx_unlock(&sync_mtx);
2492 	VI_UNLOCK(vp);
2493 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2494 		vdrop(vp);
2495 		mtx_lock(&sync_mtx);
2496 		return (*bo == LIST_FIRST(slp));
2497 	}
2498 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2499 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2500 	VOP_UNLOCK(vp);
2501 	vn_finished_write(mp);
2502 	BO_LOCK(*bo);
2503 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2504 		/*
2505 		 * Put us back on the worklist.  The worklist
2506 		 * routine will remove us from our current
2507 		 * position and then add us back in at a later
2508 		 * position.
2509 		 */
2510 		vn_syncer_add_to_worklist(*bo, syncdelay);
2511 	}
2512 	BO_UNLOCK(*bo);
2513 	vdrop(vp);
2514 	mtx_lock(&sync_mtx);
2515 	return (0);
2516 }
2517 
2518 static int first_printf = 1;
2519 
2520 /*
2521  * System filesystem synchronizer daemon.
2522  */
2523 static void
2524 sched_sync(void)
2525 {
2526 	struct synclist *next, *slp;
2527 	struct bufobj *bo;
2528 	long starttime;
2529 	struct thread *td = curthread;
2530 	int last_work_seen;
2531 	int net_worklist_len;
2532 	int syncer_final_iter;
2533 	int error;
2534 
2535 	last_work_seen = 0;
2536 	syncer_final_iter = 0;
2537 	syncer_state = SYNCER_RUNNING;
2538 	starttime = time_uptime;
2539 	td->td_pflags |= TDP_NORUNNINGBUF;
2540 
2541 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2542 	    SHUTDOWN_PRI_LAST);
2543 
2544 	mtx_lock(&sync_mtx);
2545 	for (;;) {
2546 		if (syncer_state == SYNCER_FINAL_DELAY &&
2547 		    syncer_final_iter == 0) {
2548 			mtx_unlock(&sync_mtx);
2549 			kproc_suspend_check(td->td_proc);
2550 			mtx_lock(&sync_mtx);
2551 		}
2552 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2553 		if (syncer_state != SYNCER_RUNNING &&
2554 		    starttime != time_uptime) {
2555 			if (first_printf) {
2556 				printf("\nSyncing disks, vnodes remaining... ");
2557 				first_printf = 0;
2558 			}
2559 			printf("%d ", net_worklist_len);
2560 		}
2561 		starttime = time_uptime;
2562 
2563 		/*
2564 		 * Push files whose dirty time has expired.  Be careful
2565 		 * of interrupt race on slp queue.
2566 		 *
2567 		 * Skip over empty worklist slots when shutting down.
2568 		 */
2569 		do {
2570 			slp = &syncer_workitem_pending[syncer_delayno];
2571 			syncer_delayno += 1;
2572 			if (syncer_delayno == syncer_maxdelay)
2573 				syncer_delayno = 0;
2574 			next = &syncer_workitem_pending[syncer_delayno];
2575 			/*
2576 			 * If the worklist has wrapped since the
2577 			 * it was emptied of all but syncer vnodes,
2578 			 * switch to the FINAL_DELAY state and run
2579 			 * for one more second.
2580 			 */
2581 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2582 			    net_worklist_len == 0 &&
2583 			    last_work_seen == syncer_delayno) {
2584 				syncer_state = SYNCER_FINAL_DELAY;
2585 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2586 			}
2587 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2588 		    syncer_worklist_len > 0);
2589 
2590 		/*
2591 		 * Keep track of the last time there was anything
2592 		 * on the worklist other than syncer vnodes.
2593 		 * Return to the SHUTTING_DOWN state if any
2594 		 * new work appears.
2595 		 */
2596 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2597 			last_work_seen = syncer_delayno;
2598 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2599 			syncer_state = SYNCER_SHUTTING_DOWN;
2600 		while (!LIST_EMPTY(slp)) {
2601 			error = sync_vnode(slp, &bo, td);
2602 			if (error == 1) {
2603 				LIST_REMOVE(bo, bo_synclist);
2604 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2605 				continue;
2606 			}
2607 
2608 			if (first_printf == 0) {
2609 				/*
2610 				 * Drop the sync mutex, because some watchdog
2611 				 * drivers need to sleep while patting
2612 				 */
2613 				mtx_unlock(&sync_mtx);
2614 				wdog_kern_pat(WD_LASTVAL);
2615 				mtx_lock(&sync_mtx);
2616 			}
2617 		}
2618 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2619 			syncer_final_iter--;
2620 		/*
2621 		 * The variable rushjob allows the kernel to speed up the
2622 		 * processing of the filesystem syncer process. A rushjob
2623 		 * value of N tells the filesystem syncer to process the next
2624 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2625 		 * is used by the soft update code to speed up the filesystem
2626 		 * syncer process when the incore state is getting so far
2627 		 * ahead of the disk that the kernel memory pool is being
2628 		 * threatened with exhaustion.
2629 		 */
2630 		if (rushjob > 0) {
2631 			rushjob -= 1;
2632 			continue;
2633 		}
2634 		/*
2635 		 * Just sleep for a short period of time between
2636 		 * iterations when shutting down to allow some I/O
2637 		 * to happen.
2638 		 *
2639 		 * If it has taken us less than a second to process the
2640 		 * current work, then wait. Otherwise start right over
2641 		 * again. We can still lose time if any single round
2642 		 * takes more than two seconds, but it does not really
2643 		 * matter as we are just trying to generally pace the
2644 		 * filesystem activity.
2645 		 */
2646 		if (syncer_state != SYNCER_RUNNING ||
2647 		    time_uptime == starttime) {
2648 			thread_lock(td);
2649 			sched_prio(td, PPAUSE);
2650 			thread_unlock(td);
2651 		}
2652 		if (syncer_state != SYNCER_RUNNING)
2653 			cv_timedwait(&sync_wakeup, &sync_mtx,
2654 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2655 		else if (time_uptime == starttime)
2656 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2657 	}
2658 }
2659 
2660 /*
2661  * Request the syncer daemon to speed up its work.
2662  * We never push it to speed up more than half of its
2663  * normal turn time, otherwise it could take over the cpu.
2664  */
2665 int
2666 speedup_syncer(void)
2667 {
2668 	int ret = 0;
2669 
2670 	mtx_lock(&sync_mtx);
2671 	if (rushjob < syncdelay / 2) {
2672 		rushjob += 1;
2673 		stat_rush_requests += 1;
2674 		ret = 1;
2675 	}
2676 	mtx_unlock(&sync_mtx);
2677 	cv_broadcast(&sync_wakeup);
2678 	return (ret);
2679 }
2680 
2681 /*
2682  * Tell the syncer to speed up its work and run though its work
2683  * list several times, then tell it to shut down.
2684  */
2685 static void
2686 syncer_shutdown(void *arg, int howto)
2687 {
2688 
2689 	if (howto & RB_NOSYNC)
2690 		return;
2691 	mtx_lock(&sync_mtx);
2692 	syncer_state = SYNCER_SHUTTING_DOWN;
2693 	rushjob = 0;
2694 	mtx_unlock(&sync_mtx);
2695 	cv_broadcast(&sync_wakeup);
2696 	kproc_shutdown(arg, howto);
2697 }
2698 
2699 void
2700 syncer_suspend(void)
2701 {
2702 
2703 	syncer_shutdown(updateproc, 0);
2704 }
2705 
2706 void
2707 syncer_resume(void)
2708 {
2709 
2710 	mtx_lock(&sync_mtx);
2711 	first_printf = 1;
2712 	syncer_state = SYNCER_RUNNING;
2713 	mtx_unlock(&sync_mtx);
2714 	cv_broadcast(&sync_wakeup);
2715 	kproc_resume(updateproc);
2716 }
2717 
2718 /*
2719  * Move the buffer between the clean and dirty lists of its vnode.
2720  */
2721 void
2722 reassignbuf(struct buf *bp)
2723 {
2724 	struct vnode *vp;
2725 	struct bufobj *bo;
2726 	int delay;
2727 #ifdef INVARIANTS
2728 	struct bufv *bv;
2729 #endif
2730 
2731 	vp = bp->b_vp;
2732 	bo = bp->b_bufobj;
2733 
2734 	KASSERT((bp->b_flags & B_PAGING) == 0,
2735 	    ("%s: cannot reassign paging buffer %p", __func__, bp));
2736 
2737 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2738 	    bp, bp->b_vp, bp->b_flags);
2739 
2740 	BO_LOCK(bo);
2741 	buf_vlist_remove(bp);
2742 
2743 	/*
2744 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2745 	 * of clean buffers.
2746 	 */
2747 	if (bp->b_flags & B_DELWRI) {
2748 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2749 			switch (vp->v_type) {
2750 			case VDIR:
2751 				delay = dirdelay;
2752 				break;
2753 			case VCHR:
2754 				delay = metadelay;
2755 				break;
2756 			default:
2757 				delay = filedelay;
2758 			}
2759 			vn_syncer_add_to_worklist(bo, delay);
2760 		}
2761 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2762 	} else {
2763 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2764 
2765 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2766 			mtx_lock(&sync_mtx);
2767 			LIST_REMOVE(bo, bo_synclist);
2768 			syncer_worklist_len--;
2769 			mtx_unlock(&sync_mtx);
2770 			bo->bo_flag &= ~BO_ONWORKLST;
2771 		}
2772 	}
2773 #ifdef INVARIANTS
2774 	bv = &bo->bo_clean;
2775 	bp = TAILQ_FIRST(&bv->bv_hd);
2776 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2777 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2778 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2779 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2780 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2781 	bv = &bo->bo_dirty;
2782 	bp = TAILQ_FIRST(&bv->bv_hd);
2783 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2784 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2785 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2786 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2787 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2788 #endif
2789 	BO_UNLOCK(bo);
2790 }
2791 
2792 static void
2793 v_init_counters(struct vnode *vp)
2794 {
2795 
2796 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2797 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2798 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2799 
2800 	refcount_init(&vp->v_holdcnt, 1);
2801 	refcount_init(&vp->v_usecount, 1);
2802 }
2803 
2804 /*
2805  * Grab a particular vnode from the free list, increment its
2806  * reference count and lock it.  VIRF_DOOMED is set if the vnode
2807  * is being destroyed.  Only callers who specify LK_RETRY will
2808  * see doomed vnodes.  If inactive processing was delayed in
2809  * vput try to do it here.
2810  *
2811  * usecount is manipulated using atomics without holding any locks.
2812  *
2813  * holdcnt can be manipulated using atomics without holding any locks,
2814  * except when transitioning 1<->0, in which case the interlock is held.
2815  *
2816  * Consumers which don't guarantee liveness of the vnode can use SMR to
2817  * try to get a reference. Note this operation can fail since the vnode
2818  * may be awaiting getting freed by the time they get to it.
2819  */
2820 enum vgetstate
2821 vget_prep_smr(struct vnode *vp)
2822 {
2823 	enum vgetstate vs;
2824 
2825 	VFS_SMR_ASSERT_ENTERED();
2826 
2827 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2828 		vs = VGET_USECOUNT;
2829 	} else {
2830 		if (vhold_smr(vp))
2831 			vs = VGET_HOLDCNT;
2832 		else
2833 			vs = VGET_NONE;
2834 	}
2835 	return (vs);
2836 }
2837 
2838 enum vgetstate
2839 vget_prep(struct vnode *vp)
2840 {
2841 	enum vgetstate vs;
2842 
2843 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2844 		vs = VGET_USECOUNT;
2845 	} else {
2846 		vhold(vp);
2847 		vs = VGET_HOLDCNT;
2848 	}
2849 	return (vs);
2850 }
2851 
2852 void
2853 vget_abort(struct vnode *vp, enum vgetstate vs)
2854 {
2855 
2856 	switch (vs) {
2857 	case VGET_USECOUNT:
2858 		vrele(vp);
2859 		break;
2860 	case VGET_HOLDCNT:
2861 		vdrop(vp);
2862 		break;
2863 	default:
2864 		__assert_unreachable();
2865 	}
2866 }
2867 
2868 int
2869 vget(struct vnode *vp, int flags)
2870 {
2871 	enum vgetstate vs;
2872 
2873 	vs = vget_prep(vp);
2874 	return (vget_finish(vp, flags, vs));
2875 }
2876 
2877 int
2878 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2879 {
2880 	int error;
2881 
2882 	if ((flags & LK_INTERLOCK) != 0)
2883 		ASSERT_VI_LOCKED(vp, __func__);
2884 	else
2885 		ASSERT_VI_UNLOCKED(vp, __func__);
2886 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
2887 	VNPASS(vp->v_holdcnt > 0, vp);
2888 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
2889 
2890 	error = vn_lock(vp, flags);
2891 	if (__predict_false(error != 0)) {
2892 		vget_abort(vp, vs);
2893 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2894 		    vp);
2895 		return (error);
2896 	}
2897 
2898 	vget_finish_ref(vp, vs);
2899 	return (0);
2900 }
2901 
2902 void
2903 vget_finish_ref(struct vnode *vp, enum vgetstate vs)
2904 {
2905 	int old;
2906 
2907 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
2908 	VNPASS(vp->v_holdcnt > 0, vp);
2909 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
2910 
2911 	if (vs == VGET_USECOUNT)
2912 		return;
2913 
2914 	/*
2915 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
2916 	 * the vnode around. Otherwise someone else lended their hold count and
2917 	 * we have to drop ours.
2918 	 */
2919 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
2920 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
2921 	if (old != 0) {
2922 #ifdef INVARIANTS
2923 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2924 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2925 #else
2926 		refcount_release(&vp->v_holdcnt);
2927 #endif
2928 	}
2929 }
2930 
2931 void
2932 vref(struct vnode *vp)
2933 {
2934 	enum vgetstate vs;
2935 
2936 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2937 	vs = vget_prep(vp);
2938 	vget_finish_ref(vp, vs);
2939 }
2940 
2941 void
2942 vrefact(struct vnode *vp)
2943 {
2944 
2945 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2946 #ifdef INVARIANTS
2947 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
2948 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
2949 #else
2950 	refcount_acquire(&vp->v_usecount);
2951 #endif
2952 }
2953 
2954 void
2955 vlazy(struct vnode *vp)
2956 {
2957 	struct mount *mp;
2958 
2959 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
2960 
2961 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
2962 		return;
2963 	/*
2964 	 * We may get here for inactive routines after the vnode got doomed.
2965 	 */
2966 	if (VN_IS_DOOMED(vp))
2967 		return;
2968 	mp = vp->v_mount;
2969 	mtx_lock(&mp->mnt_listmtx);
2970 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
2971 		vp->v_mflag |= VMP_LAZYLIST;
2972 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
2973 		mp->mnt_lazyvnodelistsize++;
2974 	}
2975 	mtx_unlock(&mp->mnt_listmtx);
2976 }
2977 
2978 /*
2979  * This routine is only meant to be called from vgonel prior to dooming
2980  * the vnode.
2981  */
2982 static void
2983 vunlazy_gone(struct vnode *vp)
2984 {
2985 	struct mount *mp;
2986 
2987 	ASSERT_VOP_ELOCKED(vp, __func__);
2988 	ASSERT_VI_LOCKED(vp, __func__);
2989 	VNPASS(!VN_IS_DOOMED(vp), vp);
2990 
2991 	if (vp->v_mflag & VMP_LAZYLIST) {
2992 		mp = vp->v_mount;
2993 		mtx_lock(&mp->mnt_listmtx);
2994 		VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
2995 		vp->v_mflag &= ~VMP_LAZYLIST;
2996 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
2997 		mp->mnt_lazyvnodelistsize--;
2998 		mtx_unlock(&mp->mnt_listmtx);
2999 	}
3000 }
3001 
3002 static void
3003 vdefer_inactive(struct vnode *vp)
3004 {
3005 
3006 	ASSERT_VI_LOCKED(vp, __func__);
3007 	VNASSERT(vp->v_holdcnt > 0, vp,
3008 	    ("%s: vnode without hold count", __func__));
3009 	if (VN_IS_DOOMED(vp)) {
3010 		vdropl(vp);
3011 		return;
3012 	}
3013 	if (vp->v_iflag & VI_DEFINACT) {
3014 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3015 		vdropl(vp);
3016 		return;
3017 	}
3018 	if (vp->v_usecount > 0) {
3019 		vp->v_iflag &= ~VI_OWEINACT;
3020 		vdropl(vp);
3021 		return;
3022 	}
3023 	vlazy(vp);
3024 	vp->v_iflag |= VI_DEFINACT;
3025 	VI_UNLOCK(vp);
3026 	counter_u64_add(deferred_inact, 1);
3027 }
3028 
3029 static void
3030 vdefer_inactive_unlocked(struct vnode *vp)
3031 {
3032 
3033 	VI_LOCK(vp);
3034 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3035 		vdropl(vp);
3036 		return;
3037 	}
3038 	vdefer_inactive(vp);
3039 }
3040 
3041 enum vput_op { VRELE, VPUT, VUNREF };
3042 
3043 /*
3044  * Handle ->v_usecount transitioning to 0.
3045  *
3046  * By releasing the last usecount we take ownership of the hold count which
3047  * provides liveness of the vnode, meaning we have to vdrop.
3048  *
3049  * For all vnodes we may need to perform inactive processing. It requires an
3050  * exclusive lock on the vnode, while it is legal to call here with only a
3051  * shared lock (or no locks). If locking the vnode in an expected manner fails,
3052  * inactive processing gets deferred to the syncer.
3053  *
3054  * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3055  * on the lock being held all the way until VOP_INACTIVE. This in particular
3056  * happens with UFS which adds half-constructed vnodes to the hash, where they
3057  * can be found by other code.
3058  */
3059 static void
3060 vput_final(struct vnode *vp, enum vput_op func)
3061 {
3062 	int error;
3063 	bool want_unlock;
3064 
3065 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3066 	VNPASS(vp->v_holdcnt > 0, vp);
3067 
3068 	VI_LOCK(vp);
3069 
3070 	/*
3071 	 * By the time we got here someone else might have transitioned
3072 	 * the count back to > 0.
3073 	 */
3074 	if (vp->v_usecount > 0)
3075 		goto out;
3076 
3077 	/*
3078 	 * If the vnode is doomed vgone already performed inactive processing
3079 	 * (if needed).
3080 	 */
3081 	if (VN_IS_DOOMED(vp))
3082 		goto out;
3083 
3084 	if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3085 		goto out;
3086 
3087 	if (vp->v_iflag & VI_DOINGINACT)
3088 		goto out;
3089 
3090 	/*
3091 	 * Locking operations here will drop the interlock and possibly the
3092 	 * vnode lock, opening a window where the vnode can get doomed all the
3093 	 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3094 	 * perform inactive.
3095 	 */
3096 	vp->v_iflag |= VI_OWEINACT;
3097 	want_unlock = false;
3098 	error = 0;
3099 	switch (func) {
3100 	case VRELE:
3101 		switch (VOP_ISLOCKED(vp)) {
3102 		case LK_EXCLUSIVE:
3103 			break;
3104 		case LK_EXCLOTHER:
3105 		case 0:
3106 			want_unlock = true;
3107 			error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3108 			VI_LOCK(vp);
3109 			break;
3110 		default:
3111 			/*
3112 			 * The lock has at least one sharer, but we have no way
3113 			 * to conclude whether this is us. Play it safe and
3114 			 * defer processing.
3115 			 */
3116 			error = EAGAIN;
3117 			break;
3118 		}
3119 		break;
3120 	case VPUT:
3121 		want_unlock = true;
3122 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3123 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3124 			    LK_NOWAIT);
3125 			VI_LOCK(vp);
3126 		}
3127 		break;
3128 	case VUNREF:
3129 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3130 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3131 			VI_LOCK(vp);
3132 		}
3133 		break;
3134 	}
3135 	if (error == 0) {
3136 		vinactive(vp);
3137 		if (want_unlock)
3138 			VOP_UNLOCK(vp);
3139 		vdropl(vp);
3140 	} else {
3141 		vdefer_inactive(vp);
3142 	}
3143 	return;
3144 out:
3145 	if (func == VPUT)
3146 		VOP_UNLOCK(vp);
3147 	vdropl(vp);
3148 }
3149 
3150 /*
3151  * Decrement ->v_usecount for a vnode.
3152  *
3153  * Releasing the last use count requires additional processing, see vput_final
3154  * above for details.
3155  *
3156  * Comment above each variant denotes lock state on entry and exit.
3157  */
3158 
3159 /*
3160  * in: any
3161  * out: same as passed in
3162  */
3163 void
3164 vrele(struct vnode *vp)
3165 {
3166 
3167 	ASSERT_VI_UNLOCKED(vp, __func__);
3168 	if (!refcount_release(&vp->v_usecount))
3169 		return;
3170 	vput_final(vp, VRELE);
3171 }
3172 
3173 /*
3174  * in: locked
3175  * out: unlocked
3176  */
3177 void
3178 vput(struct vnode *vp)
3179 {
3180 
3181 	ASSERT_VOP_LOCKED(vp, __func__);
3182 	ASSERT_VI_UNLOCKED(vp, __func__);
3183 	if (!refcount_release(&vp->v_usecount)) {
3184 		VOP_UNLOCK(vp);
3185 		return;
3186 	}
3187 	vput_final(vp, VPUT);
3188 }
3189 
3190 /*
3191  * in: locked
3192  * out: locked
3193  */
3194 void
3195 vunref(struct vnode *vp)
3196 {
3197 
3198 	ASSERT_VOP_LOCKED(vp, __func__);
3199 	ASSERT_VI_UNLOCKED(vp, __func__);
3200 	if (!refcount_release(&vp->v_usecount))
3201 		return;
3202 	vput_final(vp, VUNREF);
3203 }
3204 
3205 void
3206 vhold(struct vnode *vp)
3207 {
3208 	int old;
3209 
3210 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3211 	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3212 	VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3213 	    ("%s: wrong hold count %d", __func__, old));
3214 	if (old == 0)
3215 		vn_freevnodes_dec();
3216 }
3217 
3218 void
3219 vholdnz(struct vnode *vp)
3220 {
3221 
3222 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3223 #ifdef INVARIANTS
3224 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3225 	VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3226 	    ("%s: wrong hold count %d", __func__, old));
3227 #else
3228 	atomic_add_int(&vp->v_holdcnt, 1);
3229 #endif
3230 }
3231 
3232 /*
3233  * Grab a hold count unless the vnode is freed.
3234  *
3235  * Only use this routine if vfs smr is the only protection you have against
3236  * freeing the vnode.
3237  *
3238  * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag
3239  * is not set.  After the flag is set the vnode becomes immutable to anyone but
3240  * the thread which managed to set the flag.
3241  *
3242  * It may be tempting to replace the loop with:
3243  * count = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3244  * if (count & VHOLD_NO_SMR) {
3245  *     backpedal and error out;
3246  * }
3247  *
3248  * However, while this is more performant, it hinders debugging by eliminating
3249  * the previously mentioned invariant.
3250  */
3251 bool
3252 vhold_smr(struct vnode *vp)
3253 {
3254 	int count;
3255 
3256 	VFS_SMR_ASSERT_ENTERED();
3257 
3258 	count = atomic_load_int(&vp->v_holdcnt);
3259 	for (;;) {
3260 		if (count & VHOLD_NO_SMR) {
3261 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3262 			    ("non-zero hold count with flags %d\n", count));
3263 			return (false);
3264 		}
3265 
3266 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3267 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3268 			if (count == 0)
3269 				vn_freevnodes_dec();
3270 			return (true);
3271 		}
3272 	}
3273 }
3274 
3275 static void __noinline
3276 vdbatch_process(struct vdbatch *vd)
3277 {
3278 	struct vnode *vp;
3279 	int i;
3280 
3281 	mtx_assert(&vd->lock, MA_OWNED);
3282 	MPASS(curthread->td_pinned > 0);
3283 	MPASS(vd->index == VDBATCH_SIZE);
3284 
3285 	mtx_lock(&vnode_list_mtx);
3286 	critical_enter();
3287 	freevnodes += vd->freevnodes;
3288 	for (i = 0; i < VDBATCH_SIZE; i++) {
3289 		vp = vd->tab[i];
3290 		TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3291 		TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3292 		MPASS(vp->v_dbatchcpu != NOCPU);
3293 		vp->v_dbatchcpu = NOCPU;
3294 	}
3295 	mtx_unlock(&vnode_list_mtx);
3296 	vd->freevnodes = 0;
3297 	bzero(vd->tab, sizeof(vd->tab));
3298 	vd->index = 0;
3299 	critical_exit();
3300 }
3301 
3302 static void
3303 vdbatch_enqueue(struct vnode *vp)
3304 {
3305 	struct vdbatch *vd;
3306 
3307 	ASSERT_VI_LOCKED(vp, __func__);
3308 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3309 	    ("%s: deferring requeue of a doomed vnode", __func__));
3310 
3311 	if (vp->v_dbatchcpu != NOCPU) {
3312 		VI_UNLOCK(vp);
3313 		return;
3314 	}
3315 
3316 	sched_pin();
3317 	vd = DPCPU_PTR(vd);
3318 	mtx_lock(&vd->lock);
3319 	MPASS(vd->index < VDBATCH_SIZE);
3320 	MPASS(vd->tab[vd->index] == NULL);
3321 	/*
3322 	 * A hack: we depend on being pinned so that we know what to put in
3323 	 * ->v_dbatchcpu.
3324 	 */
3325 	vp->v_dbatchcpu = curcpu;
3326 	vd->tab[vd->index] = vp;
3327 	vd->index++;
3328 	VI_UNLOCK(vp);
3329 	if (vd->index == VDBATCH_SIZE)
3330 		vdbatch_process(vd);
3331 	mtx_unlock(&vd->lock);
3332 	sched_unpin();
3333 }
3334 
3335 /*
3336  * This routine must only be called for vnodes which are about to be
3337  * deallocated. Supporting dequeue for arbitrary vndoes would require
3338  * validating that the locked batch matches.
3339  */
3340 static void
3341 vdbatch_dequeue(struct vnode *vp)
3342 {
3343 	struct vdbatch *vd;
3344 	int i;
3345 	short cpu;
3346 
3347 	VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3348 	    ("%s: called for a used vnode\n", __func__));
3349 
3350 	cpu = vp->v_dbatchcpu;
3351 	if (cpu == NOCPU)
3352 		return;
3353 
3354 	vd = DPCPU_ID_PTR(cpu, vd);
3355 	mtx_lock(&vd->lock);
3356 	for (i = 0; i < vd->index; i++) {
3357 		if (vd->tab[i] != vp)
3358 			continue;
3359 		vp->v_dbatchcpu = NOCPU;
3360 		vd->index--;
3361 		vd->tab[i] = vd->tab[vd->index];
3362 		vd->tab[vd->index] = NULL;
3363 		break;
3364 	}
3365 	mtx_unlock(&vd->lock);
3366 	/*
3367 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3368 	 */
3369 	MPASS(vp->v_dbatchcpu == NOCPU);
3370 }
3371 
3372 /*
3373  * Drop the hold count of the vnode.  If this is the last reference to
3374  * the vnode we place it on the free list unless it has been vgone'd
3375  * (marked VIRF_DOOMED) in which case we will free it.
3376  *
3377  * Because the vnode vm object keeps a hold reference on the vnode if
3378  * there is at least one resident non-cached page, the vnode cannot
3379  * leave the active list without the page cleanup done.
3380  */
3381 static void
3382 vdrop_deactivate(struct vnode *vp)
3383 {
3384 	struct mount *mp;
3385 
3386 	ASSERT_VI_LOCKED(vp, __func__);
3387 	/*
3388 	 * Mark a vnode as free: remove it from its active list
3389 	 * and put it up for recycling on the freelist.
3390 	 */
3391 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3392 	    ("vdrop: returning doomed vnode"));
3393 	VNASSERT(vp->v_op != NULL, vp,
3394 	    ("vdrop: vnode already reclaimed."));
3395 	VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
3396 	    ("vnode with VI_OWEINACT set"));
3397 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp,
3398 	    ("vnode with VI_DEFINACT set"));
3399 	if (vp->v_mflag & VMP_LAZYLIST) {
3400 		mp = vp->v_mount;
3401 		mtx_lock(&mp->mnt_listmtx);
3402 		VNASSERT(vp->v_mflag & VMP_LAZYLIST, vp, ("lost VMP_LAZYLIST"));
3403 		/*
3404 		 * Don't remove the vnode from the lazy list if another thread
3405 		 * has increased the hold count. It may have re-enqueued the
3406 		 * vnode to the lazy list and is now responsible for its
3407 		 * removal.
3408 		 */
3409 		if (vp->v_holdcnt == 0) {
3410 			vp->v_mflag &= ~VMP_LAZYLIST;
3411 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3412 			mp->mnt_lazyvnodelistsize--;
3413 		}
3414 		mtx_unlock(&mp->mnt_listmtx);
3415 	}
3416 	vdbatch_enqueue(vp);
3417 }
3418 
3419 static void __noinline
3420 vdropl_final(struct vnode *vp)
3421 {
3422 
3423 	ASSERT_VI_LOCKED(vp, __func__);
3424 	VNPASS(VN_IS_DOOMED(vp), vp);
3425 	/*
3426 	 * Set the VHOLD_NO_SMR flag.
3427 	 *
3428 	 * We may be racing against vhold_smr. If they win we can just pretend
3429 	 * we never got this far, they will vdrop later.
3430 	 */
3431 	if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) {
3432 		vn_freevnodes_inc();
3433 		VI_UNLOCK(vp);
3434 		/*
3435 		 * We lost the aforementioned race. Any subsequent access is
3436 		 * invalid as they might have managed to vdropl on their own.
3437 		 */
3438 		return;
3439 	}
3440 	/*
3441 	 * Don't bump freevnodes as this one is going away.
3442 	 */
3443 	freevnode(vp);
3444 }
3445 
3446 void
3447 vdrop(struct vnode *vp)
3448 {
3449 
3450 	ASSERT_VI_UNLOCKED(vp, __func__);
3451 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3452 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3453 		return;
3454 	VI_LOCK(vp);
3455 	vdropl(vp);
3456 }
3457 
3458 void
3459 vdropl(struct vnode *vp)
3460 {
3461 
3462 	ASSERT_VI_LOCKED(vp, __func__);
3463 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3464 	if (!refcount_release(&vp->v_holdcnt)) {
3465 		VI_UNLOCK(vp);
3466 		return;
3467 	}
3468 	if (!VN_IS_DOOMED(vp)) {
3469 		vn_freevnodes_inc();
3470 		vdrop_deactivate(vp);
3471 		/*
3472 		 * Also unlocks the interlock. We can't assert on it as we
3473 		 * released our hold and by now the vnode might have been
3474 		 * freed.
3475 		 */
3476 		return;
3477 	}
3478 	vdropl_final(vp);
3479 }
3480 
3481 /*
3482  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3483  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3484  */
3485 static void
3486 vinactivef(struct vnode *vp)
3487 {
3488 	struct vm_object *obj;
3489 
3490 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3491 	ASSERT_VI_LOCKED(vp, "vinactive");
3492 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3493 	    ("vinactive: recursed on VI_DOINGINACT"));
3494 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3495 	vp->v_iflag |= VI_DOINGINACT;
3496 	vp->v_iflag &= ~VI_OWEINACT;
3497 	VI_UNLOCK(vp);
3498 	/*
3499 	 * Before moving off the active list, we must be sure that any
3500 	 * modified pages are converted into the vnode's dirty
3501 	 * buffers, since these will no longer be checked once the
3502 	 * vnode is on the inactive list.
3503 	 *
3504 	 * The write-out of the dirty pages is asynchronous.  At the
3505 	 * point that VOP_INACTIVE() is called, there could still be
3506 	 * pending I/O and dirty pages in the object.
3507 	 */
3508 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3509 	    vm_object_mightbedirty(obj)) {
3510 		VM_OBJECT_WLOCK(obj);
3511 		vm_object_page_clean(obj, 0, 0, 0);
3512 		VM_OBJECT_WUNLOCK(obj);
3513 	}
3514 	VOP_INACTIVE(vp);
3515 	VI_LOCK(vp);
3516 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3517 	    ("vinactive: lost VI_DOINGINACT"));
3518 	vp->v_iflag &= ~VI_DOINGINACT;
3519 }
3520 
3521 void
3522 vinactive(struct vnode *vp)
3523 {
3524 
3525 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3526 	ASSERT_VI_LOCKED(vp, "vinactive");
3527 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3528 
3529 	if ((vp->v_iflag & VI_OWEINACT) == 0)
3530 		return;
3531 	if (vp->v_iflag & VI_DOINGINACT)
3532 		return;
3533 	if (vp->v_usecount > 0) {
3534 		vp->v_iflag &= ~VI_OWEINACT;
3535 		return;
3536 	}
3537 	vinactivef(vp);
3538 }
3539 
3540 /*
3541  * Remove any vnodes in the vnode table belonging to mount point mp.
3542  *
3543  * If FORCECLOSE is not specified, there should not be any active ones,
3544  * return error if any are found (nb: this is a user error, not a
3545  * system error). If FORCECLOSE is specified, detach any active vnodes
3546  * that are found.
3547  *
3548  * If WRITECLOSE is set, only flush out regular file vnodes open for
3549  * writing.
3550  *
3551  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3552  *
3553  * `rootrefs' specifies the base reference count for the root vnode
3554  * of this filesystem. The root vnode is considered busy if its
3555  * v_usecount exceeds this value. On a successful return, vflush(, td)
3556  * will call vrele() on the root vnode exactly rootrefs times.
3557  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3558  * be zero.
3559  */
3560 #ifdef DIAGNOSTIC
3561 static int busyprt = 0;		/* print out busy vnodes */
3562 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3563 #endif
3564 
3565 int
3566 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3567 {
3568 	struct vnode *vp, *mvp, *rootvp = NULL;
3569 	struct vattr vattr;
3570 	int busy = 0, error;
3571 
3572 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3573 	    rootrefs, flags);
3574 	if (rootrefs > 0) {
3575 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3576 		    ("vflush: bad args"));
3577 		/*
3578 		 * Get the filesystem root vnode. We can vput() it
3579 		 * immediately, since with rootrefs > 0, it won't go away.
3580 		 */
3581 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3582 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3583 			    __func__, error);
3584 			return (error);
3585 		}
3586 		vput(rootvp);
3587 	}
3588 loop:
3589 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3590 		vholdl(vp);
3591 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3592 		if (error) {
3593 			vdrop(vp);
3594 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3595 			goto loop;
3596 		}
3597 		/*
3598 		 * Skip over a vnodes marked VV_SYSTEM.
3599 		 */
3600 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3601 			VOP_UNLOCK(vp);
3602 			vdrop(vp);
3603 			continue;
3604 		}
3605 		/*
3606 		 * If WRITECLOSE is set, flush out unlinked but still open
3607 		 * files (even if open only for reading) and regular file
3608 		 * vnodes open for writing.
3609 		 */
3610 		if (flags & WRITECLOSE) {
3611 			if (vp->v_object != NULL) {
3612 				VM_OBJECT_WLOCK(vp->v_object);
3613 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3614 				VM_OBJECT_WUNLOCK(vp->v_object);
3615 			}
3616 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3617 			if (error != 0) {
3618 				VOP_UNLOCK(vp);
3619 				vdrop(vp);
3620 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3621 				return (error);
3622 			}
3623 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3624 			VI_LOCK(vp);
3625 
3626 			if ((vp->v_type == VNON ||
3627 			    (error == 0 && vattr.va_nlink > 0)) &&
3628 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3629 				VOP_UNLOCK(vp);
3630 				vdropl(vp);
3631 				continue;
3632 			}
3633 		} else
3634 			VI_LOCK(vp);
3635 		/*
3636 		 * With v_usecount == 0, all we need to do is clear out the
3637 		 * vnode data structures and we are done.
3638 		 *
3639 		 * If FORCECLOSE is set, forcibly close the vnode.
3640 		 */
3641 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3642 			vgonel(vp);
3643 		} else {
3644 			busy++;
3645 #ifdef DIAGNOSTIC
3646 			if (busyprt)
3647 				vn_printf(vp, "vflush: busy vnode ");
3648 #endif
3649 		}
3650 		VOP_UNLOCK(vp);
3651 		vdropl(vp);
3652 	}
3653 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3654 		/*
3655 		 * If just the root vnode is busy, and if its refcount
3656 		 * is equal to `rootrefs', then go ahead and kill it.
3657 		 */
3658 		VI_LOCK(rootvp);
3659 		KASSERT(busy > 0, ("vflush: not busy"));
3660 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3661 		    ("vflush: usecount %d < rootrefs %d",
3662 		     rootvp->v_usecount, rootrefs));
3663 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3664 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3665 			vgone(rootvp);
3666 			VOP_UNLOCK(rootvp);
3667 			busy = 0;
3668 		} else
3669 			VI_UNLOCK(rootvp);
3670 	}
3671 	if (busy) {
3672 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3673 		    busy);
3674 		return (EBUSY);
3675 	}
3676 	for (; rootrefs > 0; rootrefs--)
3677 		vrele(rootvp);
3678 	return (0);
3679 }
3680 
3681 /*
3682  * Recycle an unused vnode to the front of the free list.
3683  */
3684 int
3685 vrecycle(struct vnode *vp)
3686 {
3687 	int recycled;
3688 
3689 	VI_LOCK(vp);
3690 	recycled = vrecyclel(vp);
3691 	VI_UNLOCK(vp);
3692 	return (recycled);
3693 }
3694 
3695 /*
3696  * vrecycle, with the vp interlock held.
3697  */
3698 int
3699 vrecyclel(struct vnode *vp)
3700 {
3701 	int recycled;
3702 
3703 	ASSERT_VOP_ELOCKED(vp, __func__);
3704 	ASSERT_VI_LOCKED(vp, __func__);
3705 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3706 	recycled = 0;
3707 	if (vp->v_usecount == 0) {
3708 		recycled = 1;
3709 		vgonel(vp);
3710 	}
3711 	return (recycled);
3712 }
3713 
3714 /*
3715  * Eliminate all activity associated with a vnode
3716  * in preparation for reuse.
3717  */
3718 void
3719 vgone(struct vnode *vp)
3720 {
3721 	VI_LOCK(vp);
3722 	vgonel(vp);
3723 	VI_UNLOCK(vp);
3724 }
3725 
3726 static void
3727 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3728     struct vnode *lowervp __unused)
3729 {
3730 }
3731 
3732 /*
3733  * Notify upper mounts about reclaimed or unlinked vnode.
3734  */
3735 void
3736 vfs_notify_upper(struct vnode *vp, int event)
3737 {
3738 	static struct vfsops vgonel_vfsops = {
3739 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3740 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3741 	};
3742 	struct mount *mp, *ump, *mmp;
3743 
3744 	mp = vp->v_mount;
3745 	if (mp == NULL)
3746 		return;
3747 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3748 		return;
3749 
3750 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3751 	mmp->mnt_op = &vgonel_vfsops;
3752 	mmp->mnt_kern_flag |= MNTK_MARKER;
3753 	MNT_ILOCK(mp);
3754 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3755 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3756 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3757 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3758 			continue;
3759 		}
3760 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3761 		MNT_IUNLOCK(mp);
3762 		switch (event) {
3763 		case VFS_NOTIFY_UPPER_RECLAIM:
3764 			VFS_RECLAIM_LOWERVP(ump, vp);
3765 			break;
3766 		case VFS_NOTIFY_UPPER_UNLINK:
3767 			VFS_UNLINK_LOWERVP(ump, vp);
3768 			break;
3769 		default:
3770 			KASSERT(0, ("invalid event %d", event));
3771 			break;
3772 		}
3773 		MNT_ILOCK(mp);
3774 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3775 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3776 	}
3777 	free(mmp, M_TEMP);
3778 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3779 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3780 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3781 		wakeup(&mp->mnt_uppers);
3782 	}
3783 	MNT_IUNLOCK(mp);
3784 }
3785 
3786 /*
3787  * vgone, with the vp interlock held.
3788  */
3789 static void
3790 vgonel(struct vnode *vp)
3791 {
3792 	struct thread *td;
3793 	struct mount *mp;
3794 	vm_object_t object;
3795 	bool active, doinginact, oweinact;
3796 
3797 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3798 	ASSERT_VI_LOCKED(vp, "vgonel");
3799 	VNASSERT(vp->v_holdcnt, vp,
3800 	    ("vgonel: vp %p has no reference.", vp));
3801 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3802 	td = curthread;
3803 
3804 	/*
3805 	 * Don't vgonel if we're already doomed.
3806 	 */
3807 	if (vp->v_irflag & VIRF_DOOMED)
3808 		return;
3809 	/*
3810 	 * Paired with freevnode.
3811 	 */
3812 	vn_seqc_write_begin_locked(vp);
3813 	vunlazy_gone(vp);
3814 	vp->v_irflag |= VIRF_DOOMED;
3815 
3816 	/*
3817 	 * Check to see if the vnode is in use.  If so, we have to
3818 	 * call VOP_CLOSE() and VOP_INACTIVE().
3819 	 *
3820 	 * It could be that VOP_INACTIVE() requested reclamation, in
3821 	 * which case we should avoid recursion, so check
3822 	 * VI_DOINGINACT.  This is not precise but good enough.
3823 	 */
3824 	active = vp->v_usecount > 0;
3825 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3826 	doinginact = (vp->v_iflag & VI_DOINGINACT) != 0;
3827 
3828 	/*
3829 	 * If we need to do inactive VI_OWEINACT will be set.
3830 	 */
3831 	if (vp->v_iflag & VI_DEFINACT) {
3832 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3833 		vp->v_iflag &= ~VI_DEFINACT;
3834 		vdropl(vp);
3835 	} else {
3836 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
3837 		VI_UNLOCK(vp);
3838 	}
3839 	cache_purge_vgone(vp);
3840 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3841 
3842 	/*
3843 	 * If purging an active vnode, it must be closed and
3844 	 * deactivated before being reclaimed.
3845 	 */
3846 	if (active)
3847 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3848 	if ((oweinact || active) && !doinginact) {
3849 		VI_LOCK(vp);
3850 		vinactivef(vp);
3851 		VI_UNLOCK(vp);
3852 	}
3853 	if (vp->v_type == VSOCK)
3854 		vfs_unp_reclaim(vp);
3855 
3856 	/*
3857 	 * Clean out any buffers associated with the vnode.
3858 	 * If the flush fails, just toss the buffers.
3859 	 */
3860 	mp = NULL;
3861 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3862 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3863 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3864 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3865 			;
3866 	}
3867 
3868 	BO_LOCK(&vp->v_bufobj);
3869 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3870 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3871 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3872 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3873 	    ("vp %p bufobj not invalidated", vp));
3874 
3875 	/*
3876 	 * For VMIO bufobj, BO_DEAD is set later, or in
3877 	 * vm_object_terminate() after the object's page queue is
3878 	 * flushed.
3879 	 */
3880 	object = vp->v_bufobj.bo_object;
3881 	if (object == NULL)
3882 		vp->v_bufobj.bo_flag |= BO_DEAD;
3883 	BO_UNLOCK(&vp->v_bufobj);
3884 
3885 	/*
3886 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
3887 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
3888 	 * should not touch the object borrowed from the lower vnode
3889 	 * (the handle check).
3890 	 */
3891 	if (object != NULL && object->type == OBJT_VNODE &&
3892 	    object->handle == vp)
3893 		vnode_destroy_vobject(vp);
3894 
3895 	/*
3896 	 * Reclaim the vnode.
3897 	 */
3898 	if (VOP_RECLAIM(vp))
3899 		panic("vgone: cannot reclaim");
3900 	if (mp != NULL)
3901 		vn_finished_secondary_write(mp);
3902 	VNASSERT(vp->v_object == NULL, vp,
3903 	    ("vop_reclaim left v_object vp=%p", vp));
3904 	/*
3905 	 * Clear the advisory locks and wake up waiting threads.
3906 	 */
3907 	(void)VOP_ADVLOCKPURGE(vp);
3908 	vp->v_lockf = NULL;
3909 	/*
3910 	 * Delete from old mount point vnode list.
3911 	 */
3912 	delmntque(vp);
3913 	/*
3914 	 * Done with purge, reset to the standard lock and invalidate
3915 	 * the vnode.
3916 	 */
3917 	VI_LOCK(vp);
3918 	vp->v_vnlock = &vp->v_lock;
3919 	vp->v_op = &dead_vnodeops;
3920 	vp->v_type = VBAD;
3921 }
3922 
3923 /*
3924  * Print out a description of a vnode.
3925  */
3926 static const char * const typename[] =
3927 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3928  "VMARKER"};
3929 
3930 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0,
3931     "new hold count flag not added to vn_printf");
3932 
3933 void
3934 vn_printf(struct vnode *vp, const char *fmt, ...)
3935 {
3936 	va_list ap;
3937 	char buf[256], buf2[16];
3938 	u_long flags;
3939 	u_int holdcnt;
3940 
3941 	va_start(ap, fmt);
3942 	vprintf(fmt, ap);
3943 	va_end(ap);
3944 	printf("%p: ", (void *)vp);
3945 	printf("type %s\n", typename[vp->v_type]);
3946 	holdcnt = atomic_load_int(&vp->v_holdcnt);
3947 	printf("    usecount %d, writecount %d, refcount %d seqc users %d",
3948 	    vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS,
3949 	    vp->v_seqc_users);
3950 	switch (vp->v_type) {
3951 	case VDIR:
3952 		printf(" mountedhere %p\n", vp->v_mountedhere);
3953 		break;
3954 	case VCHR:
3955 		printf(" rdev %p\n", vp->v_rdev);
3956 		break;
3957 	case VSOCK:
3958 		printf(" socket %p\n", vp->v_unpcb);
3959 		break;
3960 	case VFIFO:
3961 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
3962 		break;
3963 	default:
3964 		printf("\n");
3965 		break;
3966 	}
3967 	buf[0] = '\0';
3968 	buf[1] = '\0';
3969 	if (holdcnt & VHOLD_NO_SMR)
3970 		strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf));
3971 	printf("    hold count flags (%s)\n", buf + 1);
3972 
3973 	buf[0] = '\0';
3974 	buf[1] = '\0';
3975 	if (vp->v_irflag & VIRF_DOOMED)
3976 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
3977 	if (vp->v_irflag & VIRF_PGREAD)
3978 		strlcat(buf, "|VIRF_PGREAD", sizeof(buf));
3979 	flags = vp->v_irflag & ~(VIRF_DOOMED | VIRF_PGREAD);
3980 	if (flags != 0) {
3981 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
3982 		strlcat(buf, buf2, sizeof(buf));
3983 	}
3984 	if (vp->v_vflag & VV_ROOT)
3985 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3986 	if (vp->v_vflag & VV_ISTTY)
3987 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3988 	if (vp->v_vflag & VV_NOSYNC)
3989 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3990 	if (vp->v_vflag & VV_ETERNALDEV)
3991 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3992 	if (vp->v_vflag & VV_CACHEDLABEL)
3993 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3994 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
3995 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
3996 	if (vp->v_vflag & VV_COPYONWRITE)
3997 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3998 	if (vp->v_vflag & VV_SYSTEM)
3999 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4000 	if (vp->v_vflag & VV_PROCDEP)
4001 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4002 	if (vp->v_vflag & VV_NOKNOTE)
4003 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
4004 	if (vp->v_vflag & VV_DELETED)
4005 		strlcat(buf, "|VV_DELETED", sizeof(buf));
4006 	if (vp->v_vflag & VV_MD)
4007 		strlcat(buf, "|VV_MD", sizeof(buf));
4008 	if (vp->v_vflag & VV_FORCEINSMQ)
4009 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4010 	if (vp->v_vflag & VV_READLINK)
4011 		strlcat(buf, "|VV_READLINK", sizeof(buf));
4012 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4013 	    VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
4014 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
4015 	if (flags != 0) {
4016 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4017 		strlcat(buf, buf2, sizeof(buf));
4018 	}
4019 	if (vp->v_iflag & VI_TEXT_REF)
4020 		strlcat(buf, "|VI_TEXT_REF", sizeof(buf));
4021 	if (vp->v_iflag & VI_MOUNT)
4022 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
4023 	if (vp->v_iflag & VI_DOINGINACT)
4024 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4025 	if (vp->v_iflag & VI_OWEINACT)
4026 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4027 	if (vp->v_iflag & VI_DEFINACT)
4028 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4029 	flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_DOINGINACT |
4030 	    VI_OWEINACT | VI_DEFINACT);
4031 	if (flags != 0) {
4032 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4033 		strlcat(buf, buf2, sizeof(buf));
4034 	}
4035 	if (vp->v_mflag & VMP_LAZYLIST)
4036 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4037 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
4038 	if (flags != 0) {
4039 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4040 		strlcat(buf, buf2, sizeof(buf));
4041 	}
4042 	printf("    flags (%s)\n", buf + 1);
4043 	if (mtx_owned(VI_MTX(vp)))
4044 		printf(" VI_LOCKed");
4045 	if (vp->v_object != NULL)
4046 		printf("    v_object %p ref %d pages %d "
4047 		    "cleanbuf %d dirtybuf %d\n",
4048 		    vp->v_object, vp->v_object->ref_count,
4049 		    vp->v_object->resident_page_count,
4050 		    vp->v_bufobj.bo_clean.bv_cnt,
4051 		    vp->v_bufobj.bo_dirty.bv_cnt);
4052 	printf("    ");
4053 	lockmgr_printinfo(vp->v_vnlock);
4054 	if (vp->v_data != NULL)
4055 		VOP_PRINT(vp);
4056 }
4057 
4058 #ifdef DDB
4059 /*
4060  * List all of the locked vnodes in the system.
4061  * Called when debugging the kernel.
4062  */
4063 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
4064 {
4065 	struct mount *mp;
4066 	struct vnode *vp;
4067 
4068 	/*
4069 	 * Note: because this is DDB, we can't obey the locking semantics
4070 	 * for these structures, which means we could catch an inconsistent
4071 	 * state and dereference a nasty pointer.  Not much to be done
4072 	 * about that.
4073 	 */
4074 	db_printf("Locked vnodes\n");
4075 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4076 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4077 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4078 				vn_printf(vp, "vnode ");
4079 		}
4080 	}
4081 }
4082 
4083 /*
4084  * Show details about the given vnode.
4085  */
4086 DB_SHOW_COMMAND(vnode, db_show_vnode)
4087 {
4088 	struct vnode *vp;
4089 
4090 	if (!have_addr)
4091 		return;
4092 	vp = (struct vnode *)addr;
4093 	vn_printf(vp, "vnode ");
4094 }
4095 
4096 /*
4097  * Show details about the given mount point.
4098  */
4099 DB_SHOW_COMMAND(mount, db_show_mount)
4100 {
4101 	struct mount *mp;
4102 	struct vfsopt *opt;
4103 	struct statfs *sp;
4104 	struct vnode *vp;
4105 	char buf[512];
4106 	uint64_t mflags;
4107 	u_int flags;
4108 
4109 	if (!have_addr) {
4110 		/* No address given, print short info about all mount points. */
4111 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4112 			db_printf("%p %s on %s (%s)\n", mp,
4113 			    mp->mnt_stat.f_mntfromname,
4114 			    mp->mnt_stat.f_mntonname,
4115 			    mp->mnt_stat.f_fstypename);
4116 			if (db_pager_quit)
4117 				break;
4118 		}
4119 		db_printf("\nMore info: show mount <addr>\n");
4120 		return;
4121 	}
4122 
4123 	mp = (struct mount *)addr;
4124 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4125 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4126 
4127 	buf[0] = '\0';
4128 	mflags = mp->mnt_flag;
4129 #define	MNT_FLAG(flag)	do {						\
4130 	if (mflags & (flag)) {						\
4131 		if (buf[0] != '\0')					\
4132 			strlcat(buf, ", ", sizeof(buf));		\
4133 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4134 		mflags &= ~(flag);					\
4135 	}								\
4136 } while (0)
4137 	MNT_FLAG(MNT_RDONLY);
4138 	MNT_FLAG(MNT_SYNCHRONOUS);
4139 	MNT_FLAG(MNT_NOEXEC);
4140 	MNT_FLAG(MNT_NOSUID);
4141 	MNT_FLAG(MNT_NFS4ACLS);
4142 	MNT_FLAG(MNT_UNION);
4143 	MNT_FLAG(MNT_ASYNC);
4144 	MNT_FLAG(MNT_SUIDDIR);
4145 	MNT_FLAG(MNT_SOFTDEP);
4146 	MNT_FLAG(MNT_NOSYMFOLLOW);
4147 	MNT_FLAG(MNT_GJOURNAL);
4148 	MNT_FLAG(MNT_MULTILABEL);
4149 	MNT_FLAG(MNT_ACLS);
4150 	MNT_FLAG(MNT_NOATIME);
4151 	MNT_FLAG(MNT_NOCLUSTERR);
4152 	MNT_FLAG(MNT_NOCLUSTERW);
4153 	MNT_FLAG(MNT_SUJ);
4154 	MNT_FLAG(MNT_EXRDONLY);
4155 	MNT_FLAG(MNT_EXPORTED);
4156 	MNT_FLAG(MNT_DEFEXPORTED);
4157 	MNT_FLAG(MNT_EXPORTANON);
4158 	MNT_FLAG(MNT_EXKERB);
4159 	MNT_FLAG(MNT_EXPUBLIC);
4160 	MNT_FLAG(MNT_LOCAL);
4161 	MNT_FLAG(MNT_QUOTA);
4162 	MNT_FLAG(MNT_ROOTFS);
4163 	MNT_FLAG(MNT_USER);
4164 	MNT_FLAG(MNT_IGNORE);
4165 	MNT_FLAG(MNT_UPDATE);
4166 	MNT_FLAG(MNT_DELEXPORT);
4167 	MNT_FLAG(MNT_RELOAD);
4168 	MNT_FLAG(MNT_FORCE);
4169 	MNT_FLAG(MNT_SNAPSHOT);
4170 	MNT_FLAG(MNT_BYFSID);
4171 #undef MNT_FLAG
4172 	if (mflags != 0) {
4173 		if (buf[0] != '\0')
4174 			strlcat(buf, ", ", sizeof(buf));
4175 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4176 		    "0x%016jx", mflags);
4177 	}
4178 	db_printf("    mnt_flag = %s\n", buf);
4179 
4180 	buf[0] = '\0';
4181 	flags = mp->mnt_kern_flag;
4182 #define	MNT_KERN_FLAG(flag)	do {					\
4183 	if (flags & (flag)) {						\
4184 		if (buf[0] != '\0')					\
4185 			strlcat(buf, ", ", sizeof(buf));		\
4186 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4187 		flags &= ~(flag);					\
4188 	}								\
4189 } while (0)
4190 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4191 	MNT_KERN_FLAG(MNTK_ASYNC);
4192 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4193 	MNT_KERN_FLAG(MNTK_DRAINING);
4194 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4195 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4196 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4197 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4198 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
4199 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
4200 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
4201 	MNT_KERN_FLAG(MNTK_MARKER);
4202 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4203 	MNT_KERN_FLAG(MNTK_FPLOOKUP);
4204 	MNT_KERN_FLAG(MNTK_NOASYNC);
4205 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4206 	MNT_KERN_FLAG(MNTK_MWAIT);
4207 	MNT_KERN_FLAG(MNTK_SUSPEND);
4208 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4209 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4210 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4211 	MNT_KERN_FLAG(MNTK_NOKNOTE);
4212 #undef MNT_KERN_FLAG
4213 	if (flags != 0) {
4214 		if (buf[0] != '\0')
4215 			strlcat(buf, ", ", sizeof(buf));
4216 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4217 		    "0x%08x", flags);
4218 	}
4219 	db_printf("    mnt_kern_flag = %s\n", buf);
4220 
4221 	db_printf("    mnt_opt = ");
4222 	opt = TAILQ_FIRST(mp->mnt_opt);
4223 	if (opt != NULL) {
4224 		db_printf("%s", opt->name);
4225 		opt = TAILQ_NEXT(opt, link);
4226 		while (opt != NULL) {
4227 			db_printf(", %s", opt->name);
4228 			opt = TAILQ_NEXT(opt, link);
4229 		}
4230 	}
4231 	db_printf("\n");
4232 
4233 	sp = &mp->mnt_stat;
4234 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4235 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4236 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4237 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4238 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4239 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4240 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4241 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4242 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4243 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4244 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4245 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4246 
4247 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4248 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4249 	if (jailed(mp->mnt_cred))
4250 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4251 	db_printf(" }\n");
4252 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4253 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4254 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4255 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4256 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4257 	    mp->mnt_lazyvnodelistsize);
4258 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4259 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4260 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
4261 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4262 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4263 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4264 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4265 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4266 	db_printf("    mnt_secondary_accwrites = %d\n",
4267 	    mp->mnt_secondary_accwrites);
4268 	db_printf("    mnt_gjprovider = %s\n",
4269 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4270 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4271 
4272 	db_printf("\n\nList of active vnodes\n");
4273 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4274 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4275 			vn_printf(vp, "vnode ");
4276 			if (db_pager_quit)
4277 				break;
4278 		}
4279 	}
4280 	db_printf("\n\nList of inactive vnodes\n");
4281 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4282 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4283 			vn_printf(vp, "vnode ");
4284 			if (db_pager_quit)
4285 				break;
4286 		}
4287 	}
4288 }
4289 #endif	/* DDB */
4290 
4291 /*
4292  * Fill in a struct xvfsconf based on a struct vfsconf.
4293  */
4294 static int
4295 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4296 {
4297 	struct xvfsconf xvfsp;
4298 
4299 	bzero(&xvfsp, sizeof(xvfsp));
4300 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4301 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4302 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4303 	xvfsp.vfc_flags = vfsp->vfc_flags;
4304 	/*
4305 	 * These are unused in userland, we keep them
4306 	 * to not break binary compatibility.
4307 	 */
4308 	xvfsp.vfc_vfsops = NULL;
4309 	xvfsp.vfc_next = NULL;
4310 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4311 }
4312 
4313 #ifdef COMPAT_FREEBSD32
4314 struct xvfsconf32 {
4315 	uint32_t	vfc_vfsops;
4316 	char		vfc_name[MFSNAMELEN];
4317 	int32_t		vfc_typenum;
4318 	int32_t		vfc_refcount;
4319 	int32_t		vfc_flags;
4320 	uint32_t	vfc_next;
4321 };
4322 
4323 static int
4324 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4325 {
4326 	struct xvfsconf32 xvfsp;
4327 
4328 	bzero(&xvfsp, sizeof(xvfsp));
4329 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4330 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4331 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4332 	xvfsp.vfc_flags = vfsp->vfc_flags;
4333 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4334 }
4335 #endif
4336 
4337 /*
4338  * Top level filesystem related information gathering.
4339  */
4340 static int
4341 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4342 {
4343 	struct vfsconf *vfsp;
4344 	int error;
4345 
4346 	error = 0;
4347 	vfsconf_slock();
4348 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4349 #ifdef COMPAT_FREEBSD32
4350 		if (req->flags & SCTL_MASK32)
4351 			error = vfsconf2x32(req, vfsp);
4352 		else
4353 #endif
4354 			error = vfsconf2x(req, vfsp);
4355 		if (error)
4356 			break;
4357 	}
4358 	vfsconf_sunlock();
4359 	return (error);
4360 }
4361 
4362 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4363     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4364     "S,xvfsconf", "List of all configured filesystems");
4365 
4366 #ifndef BURN_BRIDGES
4367 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4368 
4369 static int
4370 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4371 {
4372 	int *name = (int *)arg1 - 1;	/* XXX */
4373 	u_int namelen = arg2 + 1;	/* XXX */
4374 	struct vfsconf *vfsp;
4375 
4376 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4377 	    "please rebuild world\n");
4378 
4379 #if 1 || defined(COMPAT_PRELITE2)
4380 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4381 	if (namelen == 1)
4382 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4383 #endif
4384 
4385 	switch (name[1]) {
4386 	case VFS_MAXTYPENUM:
4387 		if (namelen != 2)
4388 			return (ENOTDIR);
4389 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4390 	case VFS_CONF:
4391 		if (namelen != 3)
4392 			return (ENOTDIR);	/* overloaded */
4393 		vfsconf_slock();
4394 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4395 			if (vfsp->vfc_typenum == name[2])
4396 				break;
4397 		}
4398 		vfsconf_sunlock();
4399 		if (vfsp == NULL)
4400 			return (EOPNOTSUPP);
4401 #ifdef COMPAT_FREEBSD32
4402 		if (req->flags & SCTL_MASK32)
4403 			return (vfsconf2x32(req, vfsp));
4404 		else
4405 #endif
4406 			return (vfsconf2x(req, vfsp));
4407 	}
4408 	return (EOPNOTSUPP);
4409 }
4410 
4411 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4412     CTLFLAG_MPSAFE, vfs_sysctl,
4413     "Generic filesystem");
4414 
4415 #if 1 || defined(COMPAT_PRELITE2)
4416 
4417 static int
4418 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4419 {
4420 	int error;
4421 	struct vfsconf *vfsp;
4422 	struct ovfsconf ovfs;
4423 
4424 	vfsconf_slock();
4425 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4426 		bzero(&ovfs, sizeof(ovfs));
4427 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4428 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4429 		ovfs.vfc_index = vfsp->vfc_typenum;
4430 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4431 		ovfs.vfc_flags = vfsp->vfc_flags;
4432 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4433 		if (error != 0) {
4434 			vfsconf_sunlock();
4435 			return (error);
4436 		}
4437 	}
4438 	vfsconf_sunlock();
4439 	return (0);
4440 }
4441 
4442 #endif /* 1 || COMPAT_PRELITE2 */
4443 #endif /* !BURN_BRIDGES */
4444 
4445 #define KINFO_VNODESLOP		10
4446 #ifdef notyet
4447 /*
4448  * Dump vnode list (via sysctl).
4449  */
4450 /* ARGSUSED */
4451 static int
4452 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4453 {
4454 	struct xvnode *xvn;
4455 	struct mount *mp;
4456 	struct vnode *vp;
4457 	int error, len, n;
4458 
4459 	/*
4460 	 * Stale numvnodes access is not fatal here.
4461 	 */
4462 	req->lock = 0;
4463 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4464 	if (!req->oldptr)
4465 		/* Make an estimate */
4466 		return (SYSCTL_OUT(req, 0, len));
4467 
4468 	error = sysctl_wire_old_buffer(req, 0);
4469 	if (error != 0)
4470 		return (error);
4471 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4472 	n = 0;
4473 	mtx_lock(&mountlist_mtx);
4474 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4475 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4476 			continue;
4477 		MNT_ILOCK(mp);
4478 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4479 			if (n == len)
4480 				break;
4481 			vref(vp);
4482 			xvn[n].xv_size = sizeof *xvn;
4483 			xvn[n].xv_vnode = vp;
4484 			xvn[n].xv_id = 0;	/* XXX compat */
4485 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4486 			XV_COPY(usecount);
4487 			XV_COPY(writecount);
4488 			XV_COPY(holdcnt);
4489 			XV_COPY(mount);
4490 			XV_COPY(numoutput);
4491 			XV_COPY(type);
4492 #undef XV_COPY
4493 			xvn[n].xv_flag = vp->v_vflag;
4494 
4495 			switch (vp->v_type) {
4496 			case VREG:
4497 			case VDIR:
4498 			case VLNK:
4499 				break;
4500 			case VBLK:
4501 			case VCHR:
4502 				if (vp->v_rdev == NULL) {
4503 					vrele(vp);
4504 					continue;
4505 				}
4506 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4507 				break;
4508 			case VSOCK:
4509 				xvn[n].xv_socket = vp->v_socket;
4510 				break;
4511 			case VFIFO:
4512 				xvn[n].xv_fifo = vp->v_fifoinfo;
4513 				break;
4514 			case VNON:
4515 			case VBAD:
4516 			default:
4517 				/* shouldn't happen? */
4518 				vrele(vp);
4519 				continue;
4520 			}
4521 			vrele(vp);
4522 			++n;
4523 		}
4524 		MNT_IUNLOCK(mp);
4525 		mtx_lock(&mountlist_mtx);
4526 		vfs_unbusy(mp);
4527 		if (n == len)
4528 			break;
4529 	}
4530 	mtx_unlock(&mountlist_mtx);
4531 
4532 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4533 	free(xvn, M_TEMP);
4534 	return (error);
4535 }
4536 
4537 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4538     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4539     "");
4540 #endif
4541 
4542 static void
4543 unmount_or_warn(struct mount *mp)
4544 {
4545 	int error;
4546 
4547 	error = dounmount(mp, MNT_FORCE, curthread);
4548 	if (error != 0) {
4549 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4550 		if (error == EBUSY)
4551 			printf("BUSY)\n");
4552 		else
4553 			printf("%d)\n", error);
4554 	}
4555 }
4556 
4557 /*
4558  * Unmount all filesystems. The list is traversed in reverse order
4559  * of mounting to avoid dependencies.
4560  */
4561 void
4562 vfs_unmountall(void)
4563 {
4564 	struct mount *mp, *tmp;
4565 
4566 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4567 
4568 	/*
4569 	 * Since this only runs when rebooting, it is not interlocked.
4570 	 */
4571 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4572 		vfs_ref(mp);
4573 
4574 		/*
4575 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4576 		 * unmount of the latter.
4577 		 */
4578 		if (mp == rootdevmp)
4579 			continue;
4580 
4581 		unmount_or_warn(mp);
4582 	}
4583 
4584 	if (rootdevmp != NULL)
4585 		unmount_or_warn(rootdevmp);
4586 }
4587 
4588 static void
4589 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4590 {
4591 
4592 	ASSERT_VI_LOCKED(vp, __func__);
4593 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4594 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
4595 		vdropl(vp);
4596 		return;
4597 	}
4598 	if (vn_lock(vp, lkflags) == 0) {
4599 		VI_LOCK(vp);
4600 		vinactive(vp);
4601 		VOP_UNLOCK(vp);
4602 		vdropl(vp);
4603 		return;
4604 	}
4605 	vdefer_inactive_unlocked(vp);
4606 }
4607 
4608 static int
4609 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4610 {
4611 
4612 	return (vp->v_iflag & VI_DEFINACT);
4613 }
4614 
4615 static void __noinline
4616 vfs_periodic_inactive(struct mount *mp, int flags)
4617 {
4618 	struct vnode *vp, *mvp;
4619 	int lkflags;
4620 
4621 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4622 	if (flags != MNT_WAIT)
4623 		lkflags |= LK_NOWAIT;
4624 
4625 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4626 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
4627 			VI_UNLOCK(vp);
4628 			continue;
4629 		}
4630 		vp->v_iflag &= ~VI_DEFINACT;
4631 		vfs_deferred_inactive(vp, lkflags);
4632 	}
4633 }
4634 
4635 static inline bool
4636 vfs_want_msync(struct vnode *vp)
4637 {
4638 	struct vm_object *obj;
4639 
4640 	/*
4641 	 * This test may be performed without any locks held.
4642 	 * We rely on vm_object's type stability.
4643 	 */
4644 	if (vp->v_vflag & VV_NOSYNC)
4645 		return (false);
4646 	obj = vp->v_object;
4647 	return (obj != NULL && vm_object_mightbedirty(obj));
4648 }
4649 
4650 static int
4651 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4652 {
4653 
4654 	if (vp->v_vflag & VV_NOSYNC)
4655 		return (false);
4656 	if (vp->v_iflag & VI_DEFINACT)
4657 		return (true);
4658 	return (vfs_want_msync(vp));
4659 }
4660 
4661 static void __noinline
4662 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4663 {
4664 	struct vnode *vp, *mvp;
4665 	struct vm_object *obj;
4666 	int lkflags, objflags;
4667 	bool seen_defer;
4668 
4669 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4670 	if (flags != MNT_WAIT) {
4671 		lkflags |= LK_NOWAIT;
4672 		objflags = OBJPC_NOSYNC;
4673 	} else {
4674 		objflags = OBJPC_SYNC;
4675 	}
4676 
4677 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4678 		seen_defer = false;
4679 		if (vp->v_iflag & VI_DEFINACT) {
4680 			vp->v_iflag &= ~VI_DEFINACT;
4681 			seen_defer = true;
4682 		}
4683 		if (!vfs_want_msync(vp)) {
4684 			if (seen_defer)
4685 				vfs_deferred_inactive(vp, lkflags);
4686 			else
4687 				VI_UNLOCK(vp);
4688 			continue;
4689 		}
4690 		if (vget(vp, lkflags) == 0) {
4691 			obj = vp->v_object;
4692 			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4693 				VM_OBJECT_WLOCK(obj);
4694 				vm_object_page_clean(obj, 0, 0, objflags);
4695 				VM_OBJECT_WUNLOCK(obj);
4696 			}
4697 			vput(vp);
4698 			if (seen_defer)
4699 				vdrop(vp);
4700 		} else {
4701 			if (seen_defer)
4702 				vdefer_inactive_unlocked(vp);
4703 		}
4704 	}
4705 }
4706 
4707 void
4708 vfs_periodic(struct mount *mp, int flags)
4709 {
4710 
4711 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4712 
4713 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4714 		vfs_periodic_inactive(mp, flags);
4715 	else
4716 		vfs_periodic_msync_inactive(mp, flags);
4717 }
4718 
4719 static void
4720 destroy_vpollinfo_free(struct vpollinfo *vi)
4721 {
4722 
4723 	knlist_destroy(&vi->vpi_selinfo.si_note);
4724 	mtx_destroy(&vi->vpi_lock);
4725 	uma_zfree(vnodepoll_zone, vi);
4726 }
4727 
4728 static void
4729 destroy_vpollinfo(struct vpollinfo *vi)
4730 {
4731 
4732 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4733 	seldrain(&vi->vpi_selinfo);
4734 	destroy_vpollinfo_free(vi);
4735 }
4736 
4737 /*
4738  * Initialize per-vnode helper structure to hold poll-related state.
4739  */
4740 void
4741 v_addpollinfo(struct vnode *vp)
4742 {
4743 	struct vpollinfo *vi;
4744 
4745 	if (vp->v_pollinfo != NULL)
4746 		return;
4747 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4748 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4749 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4750 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4751 	VI_LOCK(vp);
4752 	if (vp->v_pollinfo != NULL) {
4753 		VI_UNLOCK(vp);
4754 		destroy_vpollinfo_free(vi);
4755 		return;
4756 	}
4757 	vp->v_pollinfo = vi;
4758 	VI_UNLOCK(vp);
4759 }
4760 
4761 /*
4762  * Record a process's interest in events which might happen to
4763  * a vnode.  Because poll uses the historic select-style interface
4764  * internally, this routine serves as both the ``check for any
4765  * pending events'' and the ``record my interest in future events''
4766  * functions.  (These are done together, while the lock is held,
4767  * to avoid race conditions.)
4768  */
4769 int
4770 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4771 {
4772 
4773 	v_addpollinfo(vp);
4774 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4775 	if (vp->v_pollinfo->vpi_revents & events) {
4776 		/*
4777 		 * This leaves events we are not interested
4778 		 * in available for the other process which
4779 		 * which presumably had requested them
4780 		 * (otherwise they would never have been
4781 		 * recorded).
4782 		 */
4783 		events &= vp->v_pollinfo->vpi_revents;
4784 		vp->v_pollinfo->vpi_revents &= ~events;
4785 
4786 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4787 		return (events);
4788 	}
4789 	vp->v_pollinfo->vpi_events |= events;
4790 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4791 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4792 	return (0);
4793 }
4794 
4795 /*
4796  * Routine to create and manage a filesystem syncer vnode.
4797  */
4798 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4799 static int	sync_fsync(struct  vop_fsync_args *);
4800 static int	sync_inactive(struct  vop_inactive_args *);
4801 static int	sync_reclaim(struct  vop_reclaim_args *);
4802 
4803 static struct vop_vector sync_vnodeops = {
4804 	.vop_bypass =	VOP_EOPNOTSUPP,
4805 	.vop_close =	sync_close,		/* close */
4806 	.vop_fsync =	sync_fsync,		/* fsync */
4807 	.vop_inactive =	sync_inactive,	/* inactive */
4808 	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4809 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4810 	.vop_lock1 =	vop_stdlock,	/* lock */
4811 	.vop_unlock =	vop_stdunlock,	/* unlock */
4812 	.vop_islocked =	vop_stdislocked,	/* islocked */
4813 };
4814 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4815 
4816 /*
4817  * Create a new filesystem syncer vnode for the specified mount point.
4818  */
4819 void
4820 vfs_allocate_syncvnode(struct mount *mp)
4821 {
4822 	struct vnode *vp;
4823 	struct bufobj *bo;
4824 	static long start, incr, next;
4825 	int error;
4826 
4827 	/* Allocate a new vnode */
4828 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4829 	if (error != 0)
4830 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4831 	vp->v_type = VNON;
4832 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4833 	vp->v_vflag |= VV_FORCEINSMQ;
4834 	error = insmntque(vp, mp);
4835 	if (error != 0)
4836 		panic("vfs_allocate_syncvnode: insmntque() failed");
4837 	vp->v_vflag &= ~VV_FORCEINSMQ;
4838 	VOP_UNLOCK(vp);
4839 	/*
4840 	 * Place the vnode onto the syncer worklist. We attempt to
4841 	 * scatter them about on the list so that they will go off
4842 	 * at evenly distributed times even if all the filesystems
4843 	 * are mounted at once.
4844 	 */
4845 	next += incr;
4846 	if (next == 0 || next > syncer_maxdelay) {
4847 		start /= 2;
4848 		incr /= 2;
4849 		if (start == 0) {
4850 			start = syncer_maxdelay / 2;
4851 			incr = syncer_maxdelay;
4852 		}
4853 		next = start;
4854 	}
4855 	bo = &vp->v_bufobj;
4856 	BO_LOCK(bo);
4857 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4858 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4859 	mtx_lock(&sync_mtx);
4860 	sync_vnode_count++;
4861 	if (mp->mnt_syncer == NULL) {
4862 		mp->mnt_syncer = vp;
4863 		vp = NULL;
4864 	}
4865 	mtx_unlock(&sync_mtx);
4866 	BO_UNLOCK(bo);
4867 	if (vp != NULL) {
4868 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4869 		vgone(vp);
4870 		vput(vp);
4871 	}
4872 }
4873 
4874 void
4875 vfs_deallocate_syncvnode(struct mount *mp)
4876 {
4877 	struct vnode *vp;
4878 
4879 	mtx_lock(&sync_mtx);
4880 	vp = mp->mnt_syncer;
4881 	if (vp != NULL)
4882 		mp->mnt_syncer = NULL;
4883 	mtx_unlock(&sync_mtx);
4884 	if (vp != NULL)
4885 		vrele(vp);
4886 }
4887 
4888 /*
4889  * Do a lazy sync of the filesystem.
4890  */
4891 static int
4892 sync_fsync(struct vop_fsync_args *ap)
4893 {
4894 	struct vnode *syncvp = ap->a_vp;
4895 	struct mount *mp = syncvp->v_mount;
4896 	int error, save;
4897 	struct bufobj *bo;
4898 
4899 	/*
4900 	 * We only need to do something if this is a lazy evaluation.
4901 	 */
4902 	if (ap->a_waitfor != MNT_LAZY)
4903 		return (0);
4904 
4905 	/*
4906 	 * Move ourselves to the back of the sync list.
4907 	 */
4908 	bo = &syncvp->v_bufobj;
4909 	BO_LOCK(bo);
4910 	vn_syncer_add_to_worklist(bo, syncdelay);
4911 	BO_UNLOCK(bo);
4912 
4913 	/*
4914 	 * Walk the list of vnodes pushing all that are dirty and
4915 	 * not already on the sync list.
4916 	 */
4917 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4918 		return (0);
4919 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4920 		vfs_unbusy(mp);
4921 		return (0);
4922 	}
4923 	save = curthread_pflags_set(TDP_SYNCIO);
4924 	/*
4925 	 * The filesystem at hand may be idle with free vnodes stored in the
4926 	 * batch.  Return them instead of letting them stay there indefinitely.
4927 	 */
4928 	vfs_periodic(mp, MNT_NOWAIT);
4929 	error = VFS_SYNC(mp, MNT_LAZY);
4930 	curthread_pflags_restore(save);
4931 	vn_finished_write(mp);
4932 	vfs_unbusy(mp);
4933 	return (error);
4934 }
4935 
4936 /*
4937  * The syncer vnode is no referenced.
4938  */
4939 static int
4940 sync_inactive(struct vop_inactive_args *ap)
4941 {
4942 
4943 	vgone(ap->a_vp);
4944 	return (0);
4945 }
4946 
4947 /*
4948  * The syncer vnode is no longer needed and is being decommissioned.
4949  *
4950  * Modifications to the worklist must be protected by sync_mtx.
4951  */
4952 static int
4953 sync_reclaim(struct vop_reclaim_args *ap)
4954 {
4955 	struct vnode *vp = ap->a_vp;
4956 	struct bufobj *bo;
4957 
4958 	bo = &vp->v_bufobj;
4959 	BO_LOCK(bo);
4960 	mtx_lock(&sync_mtx);
4961 	if (vp->v_mount->mnt_syncer == vp)
4962 		vp->v_mount->mnt_syncer = NULL;
4963 	if (bo->bo_flag & BO_ONWORKLST) {
4964 		LIST_REMOVE(bo, bo_synclist);
4965 		syncer_worklist_len--;
4966 		sync_vnode_count--;
4967 		bo->bo_flag &= ~BO_ONWORKLST;
4968 	}
4969 	mtx_unlock(&sync_mtx);
4970 	BO_UNLOCK(bo);
4971 
4972 	return (0);
4973 }
4974 
4975 int
4976 vn_need_pageq_flush(struct vnode *vp)
4977 {
4978 	struct vm_object *obj;
4979 	int need;
4980 
4981 	MPASS(mtx_owned(VI_MTX(vp)));
4982 	need = 0;
4983 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
4984 	    vm_object_mightbedirty(obj))
4985 		need = 1;
4986 	return (need);
4987 }
4988 
4989 /*
4990  * Check if vnode represents a disk device
4991  */
4992 bool
4993 vn_isdisk_error(struct vnode *vp, int *errp)
4994 {
4995 	int error;
4996 
4997 	if (vp->v_type != VCHR) {
4998 		error = ENOTBLK;
4999 		goto out;
5000 	}
5001 	error = 0;
5002 	dev_lock();
5003 	if (vp->v_rdev == NULL)
5004 		error = ENXIO;
5005 	else if (vp->v_rdev->si_devsw == NULL)
5006 		error = ENXIO;
5007 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5008 		error = ENOTBLK;
5009 	dev_unlock();
5010 out:
5011 	*errp = error;
5012 	return (error == 0);
5013 }
5014 
5015 bool
5016 vn_isdisk(struct vnode *vp)
5017 {
5018 	int error;
5019 
5020 	return (vn_isdisk_error(vp, &error));
5021 }
5022 
5023 /*
5024  * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
5025  * the comment above cache_fplookup for details.
5026  */
5027 int
5028 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred)
5029 {
5030 	int error;
5031 
5032 	VFS_SMR_ASSERT_ENTERED();
5033 
5034 	/* Check the owner. */
5035 	if (cred->cr_uid == file_uid) {
5036 		if (file_mode & S_IXUSR)
5037 			return (0);
5038 		goto out_error;
5039 	}
5040 
5041 	/* Otherwise, check the groups (first match) */
5042 	if (groupmember(file_gid, cred)) {
5043 		if (file_mode & S_IXGRP)
5044 			return (0);
5045 		goto out_error;
5046 	}
5047 
5048 	/* Otherwise, check everyone else. */
5049 	if (file_mode & S_IXOTH)
5050 		return (0);
5051 out_error:
5052 	/*
5053 	 * Permission check failed, but it is possible denial will get overwritten
5054 	 * (e.g., when root is traversing through a 700 directory owned by someone
5055 	 * else).
5056 	 *
5057 	 * vaccess() calls priv_check_cred which in turn can descent into MAC
5058 	 * modules overriding this result. It's quite unclear what semantics
5059 	 * are allowed for them to operate, thus for safety we don't call them
5060 	 * from within the SMR section. This also means if any such modules
5061 	 * are present, we have to let the regular lookup decide.
5062 	 */
5063 	error = priv_check_cred_vfs_lookup_nomac(cred);
5064 	switch (error) {
5065 	case 0:
5066 		return (0);
5067 	case EAGAIN:
5068 		/*
5069 		 * MAC modules present.
5070 		 */
5071 		return (EAGAIN);
5072 	case EPERM:
5073 		return (EACCES);
5074 	default:
5075 		return (error);
5076 	}
5077 }
5078 
5079 /*
5080  * Common filesystem object access control check routine.  Accepts a
5081  * vnode's type, "mode", uid and gid, requested access mode, and credentials.
5082  * Returns 0 on success, or an errno on failure.
5083  */
5084 int
5085 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5086     accmode_t accmode, struct ucred *cred)
5087 {
5088 	accmode_t dac_granted;
5089 	accmode_t priv_granted;
5090 
5091 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5092 	    ("invalid bit in accmode"));
5093 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5094 	    ("VAPPEND without VWRITE"));
5095 
5096 	/*
5097 	 * Look for a normal, non-privileged way to access the file/directory
5098 	 * as requested.  If it exists, go with that.
5099 	 */
5100 
5101 	dac_granted = 0;
5102 
5103 	/* Check the owner. */
5104 	if (cred->cr_uid == file_uid) {
5105 		dac_granted |= VADMIN;
5106 		if (file_mode & S_IXUSR)
5107 			dac_granted |= VEXEC;
5108 		if (file_mode & S_IRUSR)
5109 			dac_granted |= VREAD;
5110 		if (file_mode & S_IWUSR)
5111 			dac_granted |= (VWRITE | VAPPEND);
5112 
5113 		if ((accmode & dac_granted) == accmode)
5114 			return (0);
5115 
5116 		goto privcheck;
5117 	}
5118 
5119 	/* Otherwise, check the groups (first match) */
5120 	if (groupmember(file_gid, cred)) {
5121 		if (file_mode & S_IXGRP)
5122 			dac_granted |= VEXEC;
5123 		if (file_mode & S_IRGRP)
5124 			dac_granted |= VREAD;
5125 		if (file_mode & S_IWGRP)
5126 			dac_granted |= (VWRITE | VAPPEND);
5127 
5128 		if ((accmode & dac_granted) == accmode)
5129 			return (0);
5130 
5131 		goto privcheck;
5132 	}
5133 
5134 	/* Otherwise, check everyone else. */
5135 	if (file_mode & S_IXOTH)
5136 		dac_granted |= VEXEC;
5137 	if (file_mode & S_IROTH)
5138 		dac_granted |= VREAD;
5139 	if (file_mode & S_IWOTH)
5140 		dac_granted |= (VWRITE | VAPPEND);
5141 	if ((accmode & dac_granted) == accmode)
5142 		return (0);
5143 
5144 privcheck:
5145 	/*
5146 	 * Build a privilege mask to determine if the set of privileges
5147 	 * satisfies the requirements when combined with the granted mask
5148 	 * from above.  For each privilege, if the privilege is required,
5149 	 * bitwise or the request type onto the priv_granted mask.
5150 	 */
5151 	priv_granted = 0;
5152 
5153 	if (type == VDIR) {
5154 		/*
5155 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5156 		 * requests, instead of PRIV_VFS_EXEC.
5157 		 */
5158 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5159 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5160 			priv_granted |= VEXEC;
5161 	} else {
5162 		/*
5163 		 * Ensure that at least one execute bit is on. Otherwise,
5164 		 * a privileged user will always succeed, and we don't want
5165 		 * this to happen unless the file really is executable.
5166 		 */
5167 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5168 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5169 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5170 			priv_granted |= VEXEC;
5171 	}
5172 
5173 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5174 	    !priv_check_cred(cred, PRIV_VFS_READ))
5175 		priv_granted |= VREAD;
5176 
5177 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5178 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5179 		priv_granted |= (VWRITE | VAPPEND);
5180 
5181 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5182 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5183 		priv_granted |= VADMIN;
5184 
5185 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5186 		return (0);
5187 	}
5188 
5189 	return ((accmode & VADMIN) ? EPERM : EACCES);
5190 }
5191 
5192 /*
5193  * Credential check based on process requesting service, and per-attribute
5194  * permissions.
5195  */
5196 int
5197 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5198     struct thread *td, accmode_t accmode)
5199 {
5200 
5201 	/*
5202 	 * Kernel-invoked always succeeds.
5203 	 */
5204 	if (cred == NOCRED)
5205 		return (0);
5206 
5207 	/*
5208 	 * Do not allow privileged processes in jail to directly manipulate
5209 	 * system attributes.
5210 	 */
5211 	switch (attrnamespace) {
5212 	case EXTATTR_NAMESPACE_SYSTEM:
5213 		/* Potentially should be: return (EPERM); */
5214 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5215 	case EXTATTR_NAMESPACE_USER:
5216 		return (VOP_ACCESS(vp, accmode, cred, td));
5217 	default:
5218 		return (EPERM);
5219 	}
5220 }
5221 
5222 #ifdef DEBUG_VFS_LOCKS
5223 /*
5224  * This only exists to suppress warnings from unlocked specfs accesses.  It is
5225  * no longer ok to have an unlocked VFS.
5226  */
5227 #define	IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL ||		\
5228 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
5229 
5230 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5231 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5232     "Drop into debugger on lock violation");
5233 
5234 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5235 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5236     0, "Check for interlock across VOPs");
5237 
5238 int vfs_badlock_print = 1;	/* Print lock violations. */
5239 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5240     0, "Print lock violations");
5241 
5242 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5243 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5244     0, "Print vnode details on lock violations");
5245 
5246 #ifdef KDB
5247 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5248 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5249     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5250 #endif
5251 
5252 static void
5253 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5254 {
5255 
5256 #ifdef KDB
5257 	if (vfs_badlock_backtrace)
5258 		kdb_backtrace();
5259 #endif
5260 	if (vfs_badlock_vnode)
5261 		vn_printf(vp, "vnode ");
5262 	if (vfs_badlock_print)
5263 		printf("%s: %p %s\n", str, (void *)vp, msg);
5264 	if (vfs_badlock_ddb)
5265 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5266 }
5267 
5268 void
5269 assert_vi_locked(struct vnode *vp, const char *str)
5270 {
5271 
5272 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5273 		vfs_badlock("interlock is not locked but should be", str, vp);
5274 }
5275 
5276 void
5277 assert_vi_unlocked(struct vnode *vp, const char *str)
5278 {
5279 
5280 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5281 		vfs_badlock("interlock is locked but should not be", str, vp);
5282 }
5283 
5284 void
5285 assert_vop_locked(struct vnode *vp, const char *str)
5286 {
5287 	int locked;
5288 
5289 	if (!IGNORE_LOCK(vp)) {
5290 		locked = VOP_ISLOCKED(vp);
5291 		if (locked == 0 || locked == LK_EXCLOTHER)
5292 			vfs_badlock("is not locked but should be", str, vp);
5293 	}
5294 }
5295 
5296 void
5297 assert_vop_unlocked(struct vnode *vp, const char *str)
5298 {
5299 
5300 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5301 		vfs_badlock("is locked but should not be", str, vp);
5302 }
5303 
5304 void
5305 assert_vop_elocked(struct vnode *vp, const char *str)
5306 {
5307 
5308 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5309 		vfs_badlock("is not exclusive locked but should be", str, vp);
5310 }
5311 #endif /* DEBUG_VFS_LOCKS */
5312 
5313 void
5314 vop_rename_fail(struct vop_rename_args *ap)
5315 {
5316 
5317 	if (ap->a_tvp != NULL)
5318 		vput(ap->a_tvp);
5319 	if (ap->a_tdvp == ap->a_tvp)
5320 		vrele(ap->a_tdvp);
5321 	else
5322 		vput(ap->a_tdvp);
5323 	vrele(ap->a_fdvp);
5324 	vrele(ap->a_fvp);
5325 }
5326 
5327 void
5328 vop_rename_pre(void *ap)
5329 {
5330 	struct vop_rename_args *a = ap;
5331 
5332 #ifdef DEBUG_VFS_LOCKS
5333 	if (a->a_tvp)
5334 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5335 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5336 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5337 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5338 
5339 	/* Check the source (from). */
5340 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5341 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5342 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5343 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5344 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5345 
5346 	/* Check the target. */
5347 	if (a->a_tvp)
5348 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5349 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5350 #endif
5351 	/*
5352 	 * It may be tempting to add vn_seqc_write_begin/end calls here and
5353 	 * in vop_rename_post but that's not going to work out since some
5354 	 * filesystems relookup vnodes mid-rename. This is probably a bug.
5355 	 *
5356 	 * For now filesystems are expected to do the relevant calls after they
5357 	 * decide what vnodes to operate on.
5358 	 */
5359 	if (a->a_tdvp != a->a_fdvp)
5360 		vhold(a->a_fdvp);
5361 	if (a->a_tvp != a->a_fvp)
5362 		vhold(a->a_fvp);
5363 	vhold(a->a_tdvp);
5364 	if (a->a_tvp)
5365 		vhold(a->a_tvp);
5366 }
5367 
5368 #ifdef DEBUG_VFS_LOCKS
5369 void
5370 vop_fplookup_vexec_debugpre(void *ap __unused)
5371 {
5372 
5373 	VFS_SMR_ASSERT_ENTERED();
5374 }
5375 
5376 void
5377 vop_fplookup_vexec_debugpost(void *ap __unused, int rc __unused)
5378 {
5379 
5380 	VFS_SMR_ASSERT_ENTERED();
5381 }
5382 
5383 void
5384 vop_strategy_debugpre(void *ap)
5385 {
5386 	struct vop_strategy_args *a;
5387 	struct buf *bp;
5388 
5389 	a = ap;
5390 	bp = a->a_bp;
5391 
5392 	/*
5393 	 * Cluster ops lock their component buffers but not the IO container.
5394 	 */
5395 	if ((bp->b_flags & B_CLUSTER) != 0)
5396 		return;
5397 
5398 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5399 		if (vfs_badlock_print)
5400 			printf(
5401 			    "VOP_STRATEGY: bp is not locked but should be\n");
5402 		if (vfs_badlock_ddb)
5403 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5404 	}
5405 }
5406 
5407 void
5408 vop_lock_debugpre(void *ap)
5409 {
5410 	struct vop_lock1_args *a = ap;
5411 
5412 	if ((a->a_flags & LK_INTERLOCK) == 0)
5413 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5414 	else
5415 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5416 }
5417 
5418 void
5419 vop_lock_debugpost(void *ap, int rc)
5420 {
5421 	struct vop_lock1_args *a = ap;
5422 
5423 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5424 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5425 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5426 }
5427 
5428 void
5429 vop_unlock_debugpre(void *ap)
5430 {
5431 	struct vop_unlock_args *a = ap;
5432 
5433 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5434 }
5435 
5436 void
5437 vop_need_inactive_debugpre(void *ap)
5438 {
5439 	struct vop_need_inactive_args *a = ap;
5440 
5441 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5442 }
5443 
5444 void
5445 vop_need_inactive_debugpost(void *ap, int rc)
5446 {
5447 	struct vop_need_inactive_args *a = ap;
5448 
5449 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5450 }
5451 #endif
5452 
5453 void
5454 vop_create_pre(void *ap)
5455 {
5456 	struct vop_create_args *a;
5457 	struct vnode *dvp;
5458 
5459 	a = ap;
5460 	dvp = a->a_dvp;
5461 	vn_seqc_write_begin(dvp);
5462 }
5463 
5464 void
5465 vop_create_post(void *ap, int rc)
5466 {
5467 	struct vop_create_args *a;
5468 	struct vnode *dvp;
5469 
5470 	a = ap;
5471 	dvp = a->a_dvp;
5472 	vn_seqc_write_end(dvp);
5473 	if (!rc)
5474 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5475 }
5476 
5477 void
5478 vop_whiteout_pre(void *ap)
5479 {
5480 	struct vop_whiteout_args *a;
5481 	struct vnode *dvp;
5482 
5483 	a = ap;
5484 	dvp = a->a_dvp;
5485 	vn_seqc_write_begin(dvp);
5486 }
5487 
5488 void
5489 vop_whiteout_post(void *ap, int rc)
5490 {
5491 	struct vop_whiteout_args *a;
5492 	struct vnode *dvp;
5493 
5494 	a = ap;
5495 	dvp = a->a_dvp;
5496 	vn_seqc_write_end(dvp);
5497 }
5498 
5499 void
5500 vop_deleteextattr_pre(void *ap)
5501 {
5502 	struct vop_deleteextattr_args *a;
5503 	struct vnode *vp;
5504 
5505 	a = ap;
5506 	vp = a->a_vp;
5507 	vn_seqc_write_begin(vp);
5508 }
5509 
5510 void
5511 vop_deleteextattr_post(void *ap, int rc)
5512 {
5513 	struct vop_deleteextattr_args *a;
5514 	struct vnode *vp;
5515 
5516 	a = ap;
5517 	vp = a->a_vp;
5518 	vn_seqc_write_end(vp);
5519 	if (!rc)
5520 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5521 }
5522 
5523 void
5524 vop_link_pre(void *ap)
5525 {
5526 	struct vop_link_args *a;
5527 	struct vnode *vp, *tdvp;
5528 
5529 	a = ap;
5530 	vp = a->a_vp;
5531 	tdvp = a->a_tdvp;
5532 	vn_seqc_write_begin(vp);
5533 	vn_seqc_write_begin(tdvp);
5534 }
5535 
5536 void
5537 vop_link_post(void *ap, int rc)
5538 {
5539 	struct vop_link_args *a;
5540 	struct vnode *vp, *tdvp;
5541 
5542 	a = ap;
5543 	vp = a->a_vp;
5544 	tdvp = a->a_tdvp;
5545 	vn_seqc_write_end(vp);
5546 	vn_seqc_write_end(tdvp);
5547 	if (!rc) {
5548 		VFS_KNOTE_LOCKED(vp, NOTE_LINK);
5549 		VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE);
5550 	}
5551 }
5552 
5553 void
5554 vop_mkdir_pre(void *ap)
5555 {
5556 	struct vop_mkdir_args *a;
5557 	struct vnode *dvp;
5558 
5559 	a = ap;
5560 	dvp = a->a_dvp;
5561 	vn_seqc_write_begin(dvp);
5562 }
5563 
5564 void
5565 vop_mkdir_post(void *ap, int rc)
5566 {
5567 	struct vop_mkdir_args *a;
5568 	struct vnode *dvp;
5569 
5570 	a = ap;
5571 	dvp = a->a_dvp;
5572 	vn_seqc_write_end(dvp);
5573 	if (!rc)
5574 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5575 }
5576 
5577 #ifdef DEBUG_VFS_LOCKS
5578 void
5579 vop_mkdir_debugpost(void *ap, int rc)
5580 {
5581 	struct vop_mkdir_args *a;
5582 
5583 	a = ap;
5584 	if (!rc)
5585 		cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp);
5586 }
5587 #endif
5588 
5589 void
5590 vop_mknod_pre(void *ap)
5591 {
5592 	struct vop_mknod_args *a;
5593 	struct vnode *dvp;
5594 
5595 	a = ap;
5596 	dvp = a->a_dvp;
5597 	vn_seqc_write_begin(dvp);
5598 }
5599 
5600 void
5601 vop_mknod_post(void *ap, int rc)
5602 {
5603 	struct vop_mknod_args *a;
5604 	struct vnode *dvp;
5605 
5606 	a = ap;
5607 	dvp = a->a_dvp;
5608 	vn_seqc_write_end(dvp);
5609 	if (!rc)
5610 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5611 }
5612 
5613 void
5614 vop_reclaim_post(void *ap, int rc)
5615 {
5616 	struct vop_reclaim_args *a;
5617 	struct vnode *vp;
5618 
5619 	a = ap;
5620 	vp = a->a_vp;
5621 	ASSERT_VOP_IN_SEQC(vp);
5622 	if (!rc)
5623 		VFS_KNOTE_LOCKED(vp, NOTE_REVOKE);
5624 }
5625 
5626 void
5627 vop_remove_pre(void *ap)
5628 {
5629 	struct vop_remove_args *a;
5630 	struct vnode *dvp, *vp;
5631 
5632 	a = ap;
5633 	dvp = a->a_dvp;
5634 	vp = a->a_vp;
5635 	vn_seqc_write_begin(dvp);
5636 	vn_seqc_write_begin(vp);
5637 }
5638 
5639 void
5640 vop_remove_post(void *ap, int rc)
5641 {
5642 	struct vop_remove_args *a;
5643 	struct vnode *dvp, *vp;
5644 
5645 	a = ap;
5646 	dvp = a->a_dvp;
5647 	vp = a->a_vp;
5648 	vn_seqc_write_end(dvp);
5649 	vn_seqc_write_end(vp);
5650 	if (!rc) {
5651 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5652 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5653 	}
5654 }
5655 
5656 void
5657 vop_rename_post(void *ap, int rc)
5658 {
5659 	struct vop_rename_args *a = ap;
5660 	long hint;
5661 
5662 	if (!rc) {
5663 		hint = NOTE_WRITE;
5664 		if (a->a_fdvp == a->a_tdvp) {
5665 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5666 				hint |= NOTE_LINK;
5667 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5668 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5669 		} else {
5670 			hint |= NOTE_EXTEND;
5671 			if (a->a_fvp->v_type == VDIR)
5672 				hint |= NOTE_LINK;
5673 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5674 
5675 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5676 			    a->a_tvp->v_type == VDIR)
5677 				hint &= ~NOTE_LINK;
5678 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5679 		}
5680 
5681 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5682 		if (a->a_tvp)
5683 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5684 	}
5685 	if (a->a_tdvp != a->a_fdvp)
5686 		vdrop(a->a_fdvp);
5687 	if (a->a_tvp != a->a_fvp)
5688 		vdrop(a->a_fvp);
5689 	vdrop(a->a_tdvp);
5690 	if (a->a_tvp)
5691 		vdrop(a->a_tvp);
5692 }
5693 
5694 void
5695 vop_rmdir_pre(void *ap)
5696 {
5697 	struct vop_rmdir_args *a;
5698 	struct vnode *dvp, *vp;
5699 
5700 	a = ap;
5701 	dvp = a->a_dvp;
5702 	vp = a->a_vp;
5703 	vn_seqc_write_begin(dvp);
5704 	vn_seqc_write_begin(vp);
5705 }
5706 
5707 void
5708 vop_rmdir_post(void *ap, int rc)
5709 {
5710 	struct vop_rmdir_args *a;
5711 	struct vnode *dvp, *vp;
5712 
5713 	a = ap;
5714 	dvp = a->a_dvp;
5715 	vp = a->a_vp;
5716 	vn_seqc_write_end(dvp);
5717 	vn_seqc_write_end(vp);
5718 	if (!rc) {
5719 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5720 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5721 	}
5722 }
5723 
5724 void
5725 vop_setattr_pre(void *ap)
5726 {
5727 	struct vop_setattr_args *a;
5728 	struct vnode *vp;
5729 
5730 	a = ap;
5731 	vp = a->a_vp;
5732 	vn_seqc_write_begin(vp);
5733 }
5734 
5735 void
5736 vop_setattr_post(void *ap, int rc)
5737 {
5738 	struct vop_setattr_args *a;
5739 	struct vnode *vp;
5740 
5741 	a = ap;
5742 	vp = a->a_vp;
5743 	vn_seqc_write_end(vp);
5744 	if (!rc)
5745 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5746 }
5747 
5748 void
5749 vop_setacl_pre(void *ap)
5750 {
5751 	struct vop_setacl_args *a;
5752 	struct vnode *vp;
5753 
5754 	a = ap;
5755 	vp = a->a_vp;
5756 	vn_seqc_write_begin(vp);
5757 }
5758 
5759 void
5760 vop_setacl_post(void *ap, int rc __unused)
5761 {
5762 	struct vop_setacl_args *a;
5763 	struct vnode *vp;
5764 
5765 	a = ap;
5766 	vp = a->a_vp;
5767 	vn_seqc_write_end(vp);
5768 }
5769 
5770 void
5771 vop_setextattr_pre(void *ap)
5772 {
5773 	struct vop_setextattr_args *a;
5774 	struct vnode *vp;
5775 
5776 	a = ap;
5777 	vp = a->a_vp;
5778 	vn_seqc_write_begin(vp);
5779 }
5780 
5781 void
5782 vop_setextattr_post(void *ap, int rc)
5783 {
5784 	struct vop_setextattr_args *a;
5785 	struct vnode *vp;
5786 
5787 	a = ap;
5788 	vp = a->a_vp;
5789 	vn_seqc_write_end(vp);
5790 	if (!rc)
5791 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5792 }
5793 
5794 void
5795 vop_symlink_pre(void *ap)
5796 {
5797 	struct vop_symlink_args *a;
5798 	struct vnode *dvp;
5799 
5800 	a = ap;
5801 	dvp = a->a_dvp;
5802 	vn_seqc_write_begin(dvp);
5803 }
5804 
5805 void
5806 vop_symlink_post(void *ap, int rc)
5807 {
5808 	struct vop_symlink_args *a;
5809 	struct vnode *dvp;
5810 
5811 	a = ap;
5812 	dvp = a->a_dvp;
5813 	vn_seqc_write_end(dvp);
5814 	if (!rc)
5815 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5816 }
5817 
5818 void
5819 vop_open_post(void *ap, int rc)
5820 {
5821 	struct vop_open_args *a = ap;
5822 
5823 	if (!rc)
5824 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5825 }
5826 
5827 void
5828 vop_close_post(void *ap, int rc)
5829 {
5830 	struct vop_close_args *a = ap;
5831 
5832 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5833 	    !VN_IS_DOOMED(a->a_vp))) {
5834 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5835 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
5836 	}
5837 }
5838 
5839 void
5840 vop_read_post(void *ap, int rc)
5841 {
5842 	struct vop_read_args *a = ap;
5843 
5844 	if (!rc)
5845 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5846 }
5847 
5848 void
5849 vop_read_pgcache_post(void *ap, int rc)
5850 {
5851 	struct vop_read_pgcache_args *a = ap;
5852 
5853 	if (!rc)
5854 		VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ);
5855 }
5856 
5857 void
5858 vop_readdir_post(void *ap, int rc)
5859 {
5860 	struct vop_readdir_args *a = ap;
5861 
5862 	if (!rc)
5863 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5864 }
5865 
5866 static struct knlist fs_knlist;
5867 
5868 static void
5869 vfs_event_init(void *arg)
5870 {
5871 	knlist_init_mtx(&fs_knlist, NULL);
5872 }
5873 /* XXX - correct order? */
5874 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5875 
5876 void
5877 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5878 {
5879 
5880 	KNOTE_UNLOCKED(&fs_knlist, event);
5881 }
5882 
5883 static int	filt_fsattach(struct knote *kn);
5884 static void	filt_fsdetach(struct knote *kn);
5885 static int	filt_fsevent(struct knote *kn, long hint);
5886 
5887 struct filterops fs_filtops = {
5888 	.f_isfd = 0,
5889 	.f_attach = filt_fsattach,
5890 	.f_detach = filt_fsdetach,
5891 	.f_event = filt_fsevent
5892 };
5893 
5894 static int
5895 filt_fsattach(struct knote *kn)
5896 {
5897 
5898 	kn->kn_flags |= EV_CLEAR;
5899 	knlist_add(&fs_knlist, kn, 0);
5900 	return (0);
5901 }
5902 
5903 static void
5904 filt_fsdetach(struct knote *kn)
5905 {
5906 
5907 	knlist_remove(&fs_knlist, kn, 0);
5908 }
5909 
5910 static int
5911 filt_fsevent(struct knote *kn, long hint)
5912 {
5913 
5914 	kn->kn_fflags |= hint;
5915 	return (kn->kn_fflags != 0);
5916 }
5917 
5918 static int
5919 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5920 {
5921 	struct vfsidctl vc;
5922 	int error;
5923 	struct mount *mp;
5924 
5925 	error = SYSCTL_IN(req, &vc, sizeof(vc));
5926 	if (error)
5927 		return (error);
5928 	if (vc.vc_vers != VFS_CTL_VERS1)
5929 		return (EINVAL);
5930 	mp = vfs_getvfs(&vc.vc_fsid);
5931 	if (mp == NULL)
5932 		return (ENOENT);
5933 	/* ensure that a specific sysctl goes to the right filesystem. */
5934 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
5935 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
5936 		vfs_rel(mp);
5937 		return (EINVAL);
5938 	}
5939 	VCTLTOREQ(&vc, req);
5940 	error = VFS_SYSCTL(mp, vc.vc_op, req);
5941 	vfs_rel(mp);
5942 	return (error);
5943 }
5944 
5945 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
5946     NULL, 0, sysctl_vfs_ctl, "",
5947     "Sysctl by fsid");
5948 
5949 /*
5950  * Function to initialize a va_filerev field sensibly.
5951  * XXX: Wouldn't a random number make a lot more sense ??
5952  */
5953 u_quad_t
5954 init_va_filerev(void)
5955 {
5956 	struct bintime bt;
5957 
5958 	getbinuptime(&bt);
5959 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5960 }
5961 
5962 static int	filt_vfsread(struct knote *kn, long hint);
5963 static int	filt_vfswrite(struct knote *kn, long hint);
5964 static int	filt_vfsvnode(struct knote *kn, long hint);
5965 static void	filt_vfsdetach(struct knote *kn);
5966 static struct filterops vfsread_filtops = {
5967 	.f_isfd = 1,
5968 	.f_detach = filt_vfsdetach,
5969 	.f_event = filt_vfsread
5970 };
5971 static struct filterops vfswrite_filtops = {
5972 	.f_isfd = 1,
5973 	.f_detach = filt_vfsdetach,
5974 	.f_event = filt_vfswrite
5975 };
5976 static struct filterops vfsvnode_filtops = {
5977 	.f_isfd = 1,
5978 	.f_detach = filt_vfsdetach,
5979 	.f_event = filt_vfsvnode
5980 };
5981 
5982 static void
5983 vfs_knllock(void *arg)
5984 {
5985 	struct vnode *vp = arg;
5986 
5987 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5988 }
5989 
5990 static void
5991 vfs_knlunlock(void *arg)
5992 {
5993 	struct vnode *vp = arg;
5994 
5995 	VOP_UNLOCK(vp);
5996 }
5997 
5998 static void
5999 vfs_knl_assert_locked(void *arg)
6000 {
6001 #ifdef DEBUG_VFS_LOCKS
6002 	struct vnode *vp = arg;
6003 
6004 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
6005 #endif
6006 }
6007 
6008 static void
6009 vfs_knl_assert_unlocked(void *arg)
6010 {
6011 #ifdef DEBUG_VFS_LOCKS
6012 	struct vnode *vp = arg;
6013 
6014 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
6015 #endif
6016 }
6017 
6018 int
6019 vfs_kqfilter(struct vop_kqfilter_args *ap)
6020 {
6021 	struct vnode *vp = ap->a_vp;
6022 	struct knote *kn = ap->a_kn;
6023 	struct knlist *knl;
6024 
6025 	switch (kn->kn_filter) {
6026 	case EVFILT_READ:
6027 		kn->kn_fop = &vfsread_filtops;
6028 		break;
6029 	case EVFILT_WRITE:
6030 		kn->kn_fop = &vfswrite_filtops;
6031 		break;
6032 	case EVFILT_VNODE:
6033 		kn->kn_fop = &vfsvnode_filtops;
6034 		break;
6035 	default:
6036 		return (EINVAL);
6037 	}
6038 
6039 	kn->kn_hook = (caddr_t)vp;
6040 
6041 	v_addpollinfo(vp);
6042 	if (vp->v_pollinfo == NULL)
6043 		return (ENOMEM);
6044 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
6045 	vhold(vp);
6046 	knlist_add(knl, kn, 0);
6047 
6048 	return (0);
6049 }
6050 
6051 /*
6052  * Detach knote from vnode
6053  */
6054 static void
6055 filt_vfsdetach(struct knote *kn)
6056 {
6057 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6058 
6059 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
6060 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
6061 	vdrop(vp);
6062 }
6063 
6064 /*ARGSUSED*/
6065 static int
6066 filt_vfsread(struct knote *kn, long hint)
6067 {
6068 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6069 	struct vattr va;
6070 	int res;
6071 
6072 	/*
6073 	 * filesystem is gone, so set the EOF flag and schedule
6074 	 * the knote for deletion.
6075 	 */
6076 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6077 		VI_LOCK(vp);
6078 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6079 		VI_UNLOCK(vp);
6080 		return (1);
6081 	}
6082 
6083 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
6084 		return (0);
6085 
6086 	VI_LOCK(vp);
6087 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
6088 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
6089 	VI_UNLOCK(vp);
6090 	return (res);
6091 }
6092 
6093 /*ARGSUSED*/
6094 static int
6095 filt_vfswrite(struct knote *kn, long hint)
6096 {
6097 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6098 
6099 	VI_LOCK(vp);
6100 
6101 	/*
6102 	 * filesystem is gone, so set the EOF flag and schedule
6103 	 * the knote for deletion.
6104 	 */
6105 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
6106 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6107 
6108 	kn->kn_data = 0;
6109 	VI_UNLOCK(vp);
6110 	return (1);
6111 }
6112 
6113 static int
6114 filt_vfsvnode(struct knote *kn, long hint)
6115 {
6116 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6117 	int res;
6118 
6119 	VI_LOCK(vp);
6120 	if (kn->kn_sfflags & hint)
6121 		kn->kn_fflags |= hint;
6122 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6123 		kn->kn_flags |= EV_EOF;
6124 		VI_UNLOCK(vp);
6125 		return (1);
6126 	}
6127 	res = (kn->kn_fflags != 0);
6128 	VI_UNLOCK(vp);
6129 	return (res);
6130 }
6131 
6132 /*
6133  * Returns whether the directory is empty or not.
6134  * If it is empty, the return value is 0; otherwise
6135  * the return value is an error value (which may
6136  * be ENOTEMPTY).
6137  */
6138 int
6139 vfs_emptydir(struct vnode *vp)
6140 {
6141 	struct uio uio;
6142 	struct iovec iov;
6143 	struct dirent *dirent, *dp, *endp;
6144 	int error, eof;
6145 
6146 	error = 0;
6147 	eof = 0;
6148 
6149 	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
6150 
6151 	dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
6152 	iov.iov_base = dirent;
6153 	iov.iov_len = sizeof(struct dirent);
6154 
6155 	uio.uio_iov = &iov;
6156 	uio.uio_iovcnt = 1;
6157 	uio.uio_offset = 0;
6158 	uio.uio_resid = sizeof(struct dirent);
6159 	uio.uio_segflg = UIO_SYSSPACE;
6160 	uio.uio_rw = UIO_READ;
6161 	uio.uio_td = curthread;
6162 
6163 	while (eof == 0 && error == 0) {
6164 		error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
6165 		    NULL, NULL);
6166 		if (error != 0)
6167 			break;
6168 		endp = (void *)((uint8_t *)dirent +
6169 		    sizeof(struct dirent) - uio.uio_resid);
6170 		for (dp = dirent; dp < endp;
6171 		     dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
6172 			if (dp->d_type == DT_WHT)
6173 				continue;
6174 			if (dp->d_namlen == 0)
6175 				continue;
6176 			if (dp->d_type != DT_DIR &&
6177 			    dp->d_type != DT_UNKNOWN) {
6178 				error = ENOTEMPTY;
6179 				break;
6180 			}
6181 			if (dp->d_namlen > 2) {
6182 				error = ENOTEMPTY;
6183 				break;
6184 			}
6185 			if (dp->d_namlen == 1 &&
6186 			    dp->d_name[0] != '.') {
6187 				error = ENOTEMPTY;
6188 				break;
6189 			}
6190 			if (dp->d_namlen == 2 &&
6191 			    dp->d_name[1] != '.') {
6192 				error = ENOTEMPTY;
6193 				break;
6194 			}
6195 			uio.uio_resid = sizeof(struct dirent);
6196 		}
6197 	}
6198 	free(dirent, M_TEMP);
6199 	return (error);
6200 }
6201 
6202 int
6203 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6204 {
6205 	int error;
6206 
6207 	if (dp->d_reclen > ap->a_uio->uio_resid)
6208 		return (ENAMETOOLONG);
6209 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
6210 	if (error) {
6211 		if (ap->a_ncookies != NULL) {
6212 			if (ap->a_cookies != NULL)
6213 				free(ap->a_cookies, M_TEMP);
6214 			ap->a_cookies = NULL;
6215 			*ap->a_ncookies = 0;
6216 		}
6217 		return (error);
6218 	}
6219 	if (ap->a_ncookies == NULL)
6220 		return (0);
6221 
6222 	KASSERT(ap->a_cookies,
6223 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6224 
6225 	*ap->a_cookies = realloc(*ap->a_cookies,
6226 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
6227 	(*ap->a_cookies)[*ap->a_ncookies] = off;
6228 	*ap->a_ncookies += 1;
6229 	return (0);
6230 }
6231 
6232 /*
6233  * The purpose of this routine is to remove granularity from accmode_t,
6234  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6235  * VADMIN and VAPPEND.
6236  *
6237  * If it returns 0, the caller is supposed to continue with the usual
6238  * access checks using 'accmode' as modified by this routine.  If it
6239  * returns nonzero value, the caller is supposed to return that value
6240  * as errno.
6241  *
6242  * Note that after this routine runs, accmode may be zero.
6243  */
6244 int
6245 vfs_unixify_accmode(accmode_t *accmode)
6246 {
6247 	/*
6248 	 * There is no way to specify explicit "deny" rule using
6249 	 * file mode or POSIX.1e ACLs.
6250 	 */
6251 	if (*accmode & VEXPLICIT_DENY) {
6252 		*accmode = 0;
6253 		return (0);
6254 	}
6255 
6256 	/*
6257 	 * None of these can be translated into usual access bits.
6258 	 * Also, the common case for NFSv4 ACLs is to not contain
6259 	 * either of these bits. Caller should check for VWRITE
6260 	 * on the containing directory instead.
6261 	 */
6262 	if (*accmode & (VDELETE_CHILD | VDELETE))
6263 		return (EPERM);
6264 
6265 	if (*accmode & VADMIN_PERMS) {
6266 		*accmode &= ~VADMIN_PERMS;
6267 		*accmode |= VADMIN;
6268 	}
6269 
6270 	/*
6271 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6272 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6273 	 */
6274 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6275 
6276 	return (0);
6277 }
6278 
6279 /*
6280  * Clear out a doomed vnode (if any) and replace it with a new one as long
6281  * as the fs is not being unmounted. Return the root vnode to the caller.
6282  */
6283 static int __noinline
6284 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6285 {
6286 	struct vnode *vp;
6287 	int error;
6288 
6289 restart:
6290 	if (mp->mnt_rootvnode != NULL) {
6291 		MNT_ILOCK(mp);
6292 		vp = mp->mnt_rootvnode;
6293 		if (vp != NULL) {
6294 			if (!VN_IS_DOOMED(vp)) {
6295 				vrefact(vp);
6296 				MNT_IUNLOCK(mp);
6297 				error = vn_lock(vp, flags);
6298 				if (error == 0) {
6299 					*vpp = vp;
6300 					return (0);
6301 				}
6302 				vrele(vp);
6303 				goto restart;
6304 			}
6305 			/*
6306 			 * Clear the old one.
6307 			 */
6308 			mp->mnt_rootvnode = NULL;
6309 		}
6310 		MNT_IUNLOCK(mp);
6311 		if (vp != NULL) {
6312 			vfs_op_barrier_wait(mp);
6313 			vrele(vp);
6314 		}
6315 	}
6316 	error = VFS_CACHEDROOT(mp, flags, vpp);
6317 	if (error != 0)
6318 		return (error);
6319 	if (mp->mnt_vfs_ops == 0) {
6320 		MNT_ILOCK(mp);
6321 		if (mp->mnt_vfs_ops != 0) {
6322 			MNT_IUNLOCK(mp);
6323 			return (0);
6324 		}
6325 		if (mp->mnt_rootvnode == NULL) {
6326 			vrefact(*vpp);
6327 			mp->mnt_rootvnode = *vpp;
6328 		} else {
6329 			if (mp->mnt_rootvnode != *vpp) {
6330 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6331 					panic("%s: mismatch between vnode returned "
6332 					    " by VFS_CACHEDROOT and the one cached "
6333 					    " (%p != %p)",
6334 					    __func__, *vpp, mp->mnt_rootvnode);
6335 				}
6336 			}
6337 		}
6338 		MNT_IUNLOCK(mp);
6339 	}
6340 	return (0);
6341 }
6342 
6343 int
6344 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6345 {
6346 	struct vnode *vp;
6347 	int error;
6348 
6349 	if (!vfs_op_thread_enter(mp))
6350 		return (vfs_cache_root_fallback(mp, flags, vpp));
6351 	vp = atomic_load_ptr(&mp->mnt_rootvnode);
6352 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6353 		vfs_op_thread_exit(mp);
6354 		return (vfs_cache_root_fallback(mp, flags, vpp));
6355 	}
6356 	vrefact(vp);
6357 	vfs_op_thread_exit(mp);
6358 	error = vn_lock(vp, flags);
6359 	if (error != 0) {
6360 		vrele(vp);
6361 		return (vfs_cache_root_fallback(mp, flags, vpp));
6362 	}
6363 	*vpp = vp;
6364 	return (0);
6365 }
6366 
6367 struct vnode *
6368 vfs_cache_root_clear(struct mount *mp)
6369 {
6370 	struct vnode *vp;
6371 
6372 	/*
6373 	 * ops > 0 guarantees there is nobody who can see this vnode
6374 	 */
6375 	MPASS(mp->mnt_vfs_ops > 0);
6376 	vp = mp->mnt_rootvnode;
6377 	if (vp != NULL)
6378 		vn_seqc_write_begin(vp);
6379 	mp->mnt_rootvnode = NULL;
6380 	return (vp);
6381 }
6382 
6383 void
6384 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6385 {
6386 
6387 	MPASS(mp->mnt_vfs_ops > 0);
6388 	vrefact(vp);
6389 	mp->mnt_rootvnode = vp;
6390 }
6391 
6392 /*
6393  * These are helper functions for filesystems to traverse all
6394  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6395  *
6396  * This interface replaces MNT_VNODE_FOREACH.
6397  */
6398 
6399 struct vnode *
6400 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6401 {
6402 	struct vnode *vp;
6403 
6404 	if (should_yield())
6405 		kern_yield(PRI_USER);
6406 	MNT_ILOCK(mp);
6407 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6408 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6409 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6410 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6411 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6412 			continue;
6413 		VI_LOCK(vp);
6414 		if (VN_IS_DOOMED(vp)) {
6415 			VI_UNLOCK(vp);
6416 			continue;
6417 		}
6418 		break;
6419 	}
6420 	if (vp == NULL) {
6421 		__mnt_vnode_markerfree_all(mvp, mp);
6422 		/* MNT_IUNLOCK(mp); -- done in above function */
6423 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6424 		return (NULL);
6425 	}
6426 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6427 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6428 	MNT_IUNLOCK(mp);
6429 	return (vp);
6430 }
6431 
6432 struct vnode *
6433 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6434 {
6435 	struct vnode *vp;
6436 
6437 	*mvp = vn_alloc_marker(mp);
6438 	MNT_ILOCK(mp);
6439 	MNT_REF(mp);
6440 
6441 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6442 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6443 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6444 			continue;
6445 		VI_LOCK(vp);
6446 		if (VN_IS_DOOMED(vp)) {
6447 			VI_UNLOCK(vp);
6448 			continue;
6449 		}
6450 		break;
6451 	}
6452 	if (vp == NULL) {
6453 		MNT_REL(mp);
6454 		MNT_IUNLOCK(mp);
6455 		vn_free_marker(*mvp);
6456 		*mvp = NULL;
6457 		return (NULL);
6458 	}
6459 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6460 	MNT_IUNLOCK(mp);
6461 	return (vp);
6462 }
6463 
6464 void
6465 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6466 {
6467 
6468 	if (*mvp == NULL) {
6469 		MNT_IUNLOCK(mp);
6470 		return;
6471 	}
6472 
6473 	mtx_assert(MNT_MTX(mp), MA_OWNED);
6474 
6475 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6476 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6477 	MNT_REL(mp);
6478 	MNT_IUNLOCK(mp);
6479 	vn_free_marker(*mvp);
6480 	*mvp = NULL;
6481 }
6482 
6483 /*
6484  * These are helper functions for filesystems to traverse their
6485  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6486  */
6487 static void
6488 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6489 {
6490 
6491 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6492 
6493 	MNT_ILOCK(mp);
6494 	MNT_REL(mp);
6495 	MNT_IUNLOCK(mp);
6496 	vn_free_marker(*mvp);
6497 	*mvp = NULL;
6498 }
6499 
6500 /*
6501  * Relock the mp mount vnode list lock with the vp vnode interlock in the
6502  * conventional lock order during mnt_vnode_next_lazy iteration.
6503  *
6504  * On entry, the mount vnode list lock is held and the vnode interlock is not.
6505  * The list lock is dropped and reacquired.  On success, both locks are held.
6506  * On failure, the mount vnode list lock is held but the vnode interlock is
6507  * not, and the procedure may have yielded.
6508  */
6509 static bool
6510 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6511     struct vnode *vp)
6512 {
6513 
6514 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6515 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6516 	    ("%s: bad marker", __func__));
6517 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6518 	    ("%s: inappropriate vnode", __func__));
6519 	ASSERT_VI_UNLOCKED(vp, __func__);
6520 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6521 
6522 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6523 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6524 
6525 	/*
6526 	 * Note we may be racing against vdrop which transitioned the hold
6527 	 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6528 	 * if we are the only user after we get the interlock we will just
6529 	 * vdrop.
6530 	 */
6531 	vhold(vp);
6532 	mtx_unlock(&mp->mnt_listmtx);
6533 	VI_LOCK(vp);
6534 	if (VN_IS_DOOMED(vp)) {
6535 		VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6536 		goto out_lost;
6537 	}
6538 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6539 	/*
6540 	 * There is nothing to do if we are the last user.
6541 	 */
6542 	if (!refcount_release_if_not_last(&vp->v_holdcnt))
6543 		goto out_lost;
6544 	mtx_lock(&mp->mnt_listmtx);
6545 	return (true);
6546 out_lost:
6547 	vdropl(vp);
6548 	maybe_yield();
6549 	mtx_lock(&mp->mnt_listmtx);
6550 	return (false);
6551 }
6552 
6553 static struct vnode *
6554 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6555     void *cbarg)
6556 {
6557 	struct vnode *vp;
6558 
6559 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6560 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6561 restart:
6562 	vp = TAILQ_NEXT(*mvp, v_lazylist);
6563 	while (vp != NULL) {
6564 		if (vp->v_type == VMARKER) {
6565 			vp = TAILQ_NEXT(vp, v_lazylist);
6566 			continue;
6567 		}
6568 		/*
6569 		 * See if we want to process the vnode. Note we may encounter a
6570 		 * long string of vnodes we don't care about and hog the list
6571 		 * as a result. Check for it and requeue the marker.
6572 		 */
6573 		VNPASS(!VN_IS_DOOMED(vp), vp);
6574 		if (!cb(vp, cbarg)) {
6575 			if (!should_yield()) {
6576 				vp = TAILQ_NEXT(vp, v_lazylist);
6577 				continue;
6578 			}
6579 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6580 			    v_lazylist);
6581 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6582 			    v_lazylist);
6583 			mtx_unlock(&mp->mnt_listmtx);
6584 			kern_yield(PRI_USER);
6585 			mtx_lock(&mp->mnt_listmtx);
6586 			goto restart;
6587 		}
6588 		/*
6589 		 * Try-lock because this is the wrong lock order.
6590 		 */
6591 		if (!VI_TRYLOCK(vp) &&
6592 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6593 			goto restart;
6594 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6595 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6596 		    ("alien vnode on the lazy list %p %p", vp, mp));
6597 		VNPASS(vp->v_mount == mp, vp);
6598 		VNPASS(!VN_IS_DOOMED(vp), vp);
6599 		break;
6600 	}
6601 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6602 
6603 	/* Check if we are done */
6604 	if (vp == NULL) {
6605 		mtx_unlock(&mp->mnt_listmtx);
6606 		mnt_vnode_markerfree_lazy(mvp, mp);
6607 		return (NULL);
6608 	}
6609 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6610 	mtx_unlock(&mp->mnt_listmtx);
6611 	ASSERT_VI_LOCKED(vp, "lazy iter");
6612 	return (vp);
6613 }
6614 
6615 struct vnode *
6616 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6617     void *cbarg)
6618 {
6619 
6620 	if (should_yield())
6621 		kern_yield(PRI_USER);
6622 	mtx_lock(&mp->mnt_listmtx);
6623 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6624 }
6625 
6626 struct vnode *
6627 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6628     void *cbarg)
6629 {
6630 	struct vnode *vp;
6631 
6632 	if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6633 		return (NULL);
6634 
6635 	*mvp = vn_alloc_marker(mp);
6636 	MNT_ILOCK(mp);
6637 	MNT_REF(mp);
6638 	MNT_IUNLOCK(mp);
6639 
6640 	mtx_lock(&mp->mnt_listmtx);
6641 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6642 	if (vp == NULL) {
6643 		mtx_unlock(&mp->mnt_listmtx);
6644 		mnt_vnode_markerfree_lazy(mvp, mp);
6645 		return (NULL);
6646 	}
6647 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6648 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6649 }
6650 
6651 void
6652 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6653 {
6654 
6655 	if (*mvp == NULL)
6656 		return;
6657 
6658 	mtx_lock(&mp->mnt_listmtx);
6659 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6660 	mtx_unlock(&mp->mnt_listmtx);
6661 	mnt_vnode_markerfree_lazy(mvp, mp);
6662 }
6663 
6664 int
6665 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6666 {
6667 
6668 	if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6669 		cnp->cn_flags &= ~NOEXECCHECK;
6670 		return (0);
6671 	}
6672 
6673 	return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, cnp->cn_thread));
6674 }
6675 
6676 /*
6677  * Do not use this variant unless you have means other than the hold count
6678  * to prevent the vnode from getting freed.
6679  */
6680 void
6681 vn_seqc_write_begin_unheld_locked(struct vnode *vp)
6682 {
6683 
6684 	ASSERT_VI_LOCKED(vp, __func__);
6685 	VNPASS(vp->v_seqc_users >= 0, vp);
6686 	vp->v_seqc_users++;
6687 	if (vp->v_seqc_users == 1)
6688 		seqc_sleepable_write_begin(&vp->v_seqc);
6689 }
6690 
6691 void
6692 vn_seqc_write_begin_locked(struct vnode *vp)
6693 {
6694 
6695 	ASSERT_VI_LOCKED(vp, __func__);
6696 	VNPASS(vp->v_holdcnt > 0, vp);
6697 	vn_seqc_write_begin_unheld_locked(vp);
6698 }
6699 
6700 void
6701 vn_seqc_write_begin(struct vnode *vp)
6702 {
6703 
6704 	VI_LOCK(vp);
6705 	vn_seqc_write_begin_locked(vp);
6706 	VI_UNLOCK(vp);
6707 }
6708 
6709 void
6710 vn_seqc_write_begin_unheld(struct vnode *vp)
6711 {
6712 
6713 	VI_LOCK(vp);
6714 	vn_seqc_write_begin_unheld_locked(vp);
6715 	VI_UNLOCK(vp);
6716 }
6717 
6718 void
6719 vn_seqc_write_end_locked(struct vnode *vp)
6720 {
6721 
6722 	ASSERT_VI_LOCKED(vp, __func__);
6723 	VNPASS(vp->v_seqc_users > 0, vp);
6724 	vp->v_seqc_users--;
6725 	if (vp->v_seqc_users == 0)
6726 		seqc_sleepable_write_end(&vp->v_seqc);
6727 }
6728 
6729 void
6730 vn_seqc_write_end(struct vnode *vp)
6731 {
6732 
6733 	VI_LOCK(vp);
6734 	vn_seqc_write_end_locked(vp);
6735 	VI_UNLOCK(vp);
6736 }
6737