xref: /freebsd/sys/kern/vfs_subr.c (revision 9b37d84c87e69dabc69d818aa4d2fea718bd8b74)
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 
37 /*
38  * External virtual filesystem routines
39  */
40 
41 #include <sys/cdefs.h>
42 #include "opt_ddb.h"
43 #include "opt_watchdog.h"
44 
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/asan.h>
48 #include <sys/bio.h>
49 #include <sys/buf.h>
50 #include <sys/capsicum.h>
51 #include <sys/condvar.h>
52 #include <sys/conf.h>
53 #include <sys/counter.h>
54 #include <sys/dirent.h>
55 #include <sys/event.h>
56 #include <sys/eventhandler.h>
57 #include <sys/extattr.h>
58 #include <sys/file.h>
59 #include <sys/fcntl.h>
60 #include <sys/jail.h>
61 #include <sys/kdb.h>
62 #include <sys/kernel.h>
63 #include <sys/kthread.h>
64 #include <sys/ktr.h>
65 #include <sys/limits.h>
66 #include <sys/lockf.h>
67 #include <sys/malloc.h>
68 #include <sys/mount.h>
69 #include <sys/namei.h>
70 #include <sys/pctrie.h>
71 #include <sys/priv.h>
72 #include <sys/reboot.h>
73 #include <sys/refcount.h>
74 #include <sys/rwlock.h>
75 #include <sys/sched.h>
76 #include <sys/sleepqueue.h>
77 #include <sys/smr.h>
78 #include <sys/smp.h>
79 #include <sys/stat.h>
80 #include <sys/sysctl.h>
81 #include <sys/syslog.h>
82 #include <sys/user.h>
83 #include <sys/vmmeter.h>
84 #include <sys/vnode.h>
85 #include <sys/watchdog.h>
86 
87 #include <machine/stdarg.h>
88 
89 #include <security/mac/mac_framework.h>
90 
91 #include <vm/vm.h>
92 #include <vm/vm_object.h>
93 #include <vm/vm_extern.h>
94 #include <vm/pmap.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_page.h>
97 #include <vm/vm_kern.h>
98 #include <vm/vnode_pager.h>
99 #include <vm/uma.h>
100 
101 #if defined(DEBUG_VFS_LOCKS) && (!defined(INVARIANTS) || !defined(WITNESS))
102 #error DEBUG_VFS_LOCKS requires INVARIANTS and WITNESS
103 #endif
104 
105 #ifdef DDB
106 #include <ddb/ddb.h>
107 #endif
108 
109 static void	delmntque(struct vnode *vp);
110 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
111 		    int slpflag, int slptimeo);
112 static void	syncer_shutdown(void *arg, int howto);
113 static int	vtryrecycle(struct vnode *vp, bool isvnlru);
114 static void	v_init_counters(struct vnode *);
115 static void	vn_seqc_init(struct vnode *);
116 static void	vn_seqc_write_end_free(struct vnode *vp);
117 static void	vgonel(struct vnode *);
118 static bool	vhold_recycle_free(struct vnode *);
119 static void	vdropl_recycle(struct vnode *vp);
120 static void	vdrop_recycle(struct vnode *vp);
121 static void	vfs_knllock(void *arg);
122 static void	vfs_knlunlock(void *arg);
123 static void	vfs_knl_assert_lock(void *arg, int what);
124 static void	destroy_vpollinfo(struct vpollinfo *vi);
125 static int	v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
126 		    daddr_t startlbn, daddr_t endlbn);
127 static void	vnlru_recalc(void);
128 
129 static SYSCTL_NODE(_vfs, OID_AUTO, vnode, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
130     "vnode configuration and statistics");
131 static SYSCTL_NODE(_vfs_vnode, OID_AUTO, param, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
132     "vnode configuration");
133 static SYSCTL_NODE(_vfs_vnode, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
134     "vnode statistics");
135 static SYSCTL_NODE(_vfs_vnode, OID_AUTO, vnlru, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
136     "vnode recycling");
137 
138 /*
139  * Number of vnodes in existence.  Increased whenever getnewvnode()
140  * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
141  */
142 static u_long __exclusive_cache_line numvnodes;
143 
144 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
145     "Number of vnodes in existence (legacy)");
146 SYSCTL_ULONG(_vfs_vnode_stats, OID_AUTO, count, CTLFLAG_RD, &numvnodes, 0,
147     "Number of vnodes in existence");
148 
149 static counter_u64_t vnodes_created;
150 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
151     "Number of vnodes created by getnewvnode (legacy)");
152 SYSCTL_COUNTER_U64(_vfs_vnode_stats, OID_AUTO, created, CTLFLAG_RD, &vnodes_created,
153     "Number of vnodes created by getnewvnode");
154 
155 /*
156  * Conversion tables for conversion from vnode types to inode formats
157  * and back.
158  */
159 __enum_uint8(vtype) iftovt_tab[16] = {
160 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
161 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
162 };
163 int vttoif_tab[10] = {
164 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
165 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
166 };
167 
168 /*
169  * List of allocates vnodes in the system.
170  */
171 static TAILQ_HEAD(freelst, vnode) vnode_list;
172 static struct vnode *vnode_list_free_marker;
173 static struct vnode *vnode_list_reclaim_marker;
174 
175 /*
176  * "Free" vnode target.  Free vnodes are rarely completely free, but are
177  * just ones that are cheap to recycle.  Usually they are for files which
178  * have been stat'd but not read; these usually have inode and namecache
179  * data attached to them.  This target is the preferred minimum size of a
180  * sub-cache consisting mostly of such files. The system balances the size
181  * of this sub-cache with its complement to try to prevent either from
182  * thrashing while the other is relatively inactive.  The targets express
183  * a preference for the best balance.
184  *
185  * "Above" this target there are 2 further targets (watermarks) related
186  * to recyling of free vnodes.  In the best-operating case, the cache is
187  * exactly full, the free list has size between vlowat and vhiwat above the
188  * free target, and recycling from it and normal use maintains this state.
189  * Sometimes the free list is below vlowat or even empty, but this state
190  * is even better for immediate use provided the cache is not full.
191  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
192  * ones) to reach one of these states.  The watermarks are currently hard-
193  * coded as 4% and 9% of the available space higher.  These and the default
194  * of 25% for wantfreevnodes are too large if the memory size is large.
195  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
196  * whenever vnlru_proc() becomes active.
197  */
198 static long wantfreevnodes;
199 static long __exclusive_cache_line freevnodes;
200 static long freevnodes_old;
201 
202 static u_long recycles_count;
203 SYSCTL_ULONG(_vfs, OID_AUTO, recycles, CTLFLAG_RD | CTLFLAG_STATS, &recycles_count, 0,
204     "Number of vnodes recycled to meet vnode cache targets (legacy)");
205 SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, recycles, CTLFLAG_RD | CTLFLAG_STATS,
206     &recycles_count, 0,
207     "Number of vnodes recycled to meet vnode cache targets");
208 
209 static u_long recycles_free_count;
210 SYSCTL_ULONG(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD | CTLFLAG_STATS,
211     &recycles_free_count, 0,
212     "Number of free vnodes recycled to meet vnode cache targets (legacy)");
213 SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, recycles_free, CTLFLAG_RD | CTLFLAG_STATS,
214     &recycles_free_count, 0,
215     "Number of free vnodes recycled to meet vnode cache targets");
216 
217 static counter_u64_t direct_recycles_free_count;
218 SYSCTL_COUNTER_U64(_vfs_vnode_vnlru, OID_AUTO, direct_recycles_free, CTLFLAG_RD,
219     &direct_recycles_free_count,
220     "Number of free vnodes recycled by vn_alloc callers to meet vnode cache targets");
221 
222 static counter_u64_t vnode_skipped_requeues;
223 SYSCTL_COUNTER_U64(_vfs_vnode_stats, OID_AUTO, skipped_requeues, CTLFLAG_RD, &vnode_skipped_requeues,
224     "Number of times LRU requeue was skipped due to lock contention");
225 
226 static __read_mostly bool vnode_can_skip_requeue;
227 SYSCTL_BOOL(_vfs_vnode_param, OID_AUTO, can_skip_requeue, CTLFLAG_RW,
228     &vnode_can_skip_requeue, 0, "Is LRU requeue skippable");
229 
230 static u_long deferred_inact;
231 SYSCTL_ULONG(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD,
232     &deferred_inact, 0, "Number of times inactive processing was deferred");
233 
234 /* To keep more than one thread at a time from running vfs_getnewfsid */
235 static struct mtx mntid_mtx;
236 
237 /*
238  * Lock for any access to the following:
239  *	vnode_list
240  *	numvnodes
241  *	freevnodes
242  */
243 static struct mtx __exclusive_cache_line vnode_list_mtx;
244 
245 /* Publicly exported FS */
246 struct nfs_public nfs_pub;
247 
248 static uma_zone_t buf_trie_zone;
249 static smr_t buf_trie_smr;
250 
251 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
252 static uma_zone_t vnode_zone;
253 MALLOC_DEFINE(M_VNODEPOLL, "VN POLL", "vnode poll");
254 
255 __read_frequently smr_t vfs_smr;
256 
257 /*
258  * The workitem queue.
259  *
260  * It is useful to delay writes of file data and filesystem metadata
261  * for tens of seconds so that quickly created and deleted files need
262  * not waste disk bandwidth being created and removed. To realize this,
263  * we append vnodes to a "workitem" queue. When running with a soft
264  * updates implementation, most pending metadata dependencies should
265  * not wait for more than a few seconds. Thus, mounted on block devices
266  * are delayed only about a half the time that file data is delayed.
267  * Similarly, directory updates are more critical, so are only delayed
268  * about a third the time that file data is delayed. Thus, there are
269  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
270  * one each second (driven off the filesystem syncer process). The
271  * syncer_delayno variable indicates the next queue that is to be processed.
272  * Items that need to be processed soon are placed in this queue:
273  *
274  *	syncer_workitem_pending[syncer_delayno]
275  *
276  * A delay of fifteen seconds is done by placing the request fifteen
277  * entries later in the queue:
278  *
279  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
280  *
281  */
282 static int syncer_delayno;
283 static long syncer_mask;
284 LIST_HEAD(synclist, bufobj);
285 static struct synclist *syncer_workitem_pending;
286 /*
287  * The sync_mtx protects:
288  *	bo->bo_synclist
289  *	sync_vnode_count
290  *	syncer_delayno
291  *	syncer_state
292  *	syncer_workitem_pending
293  *	syncer_worklist_len
294  *	rushjob
295  */
296 static struct mtx sync_mtx;
297 static struct cv sync_wakeup;
298 
299 #define SYNCER_MAXDELAY		32
300 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
301 static int syncdelay = 30;		/* max time to delay syncing data */
302 static int filedelay = 30;		/* time to delay syncing files */
303 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
304     "Time to delay syncing files (in seconds)");
305 static int dirdelay = 29;		/* time to delay syncing directories */
306 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
307     "Time to delay syncing directories (in seconds)");
308 static int metadelay = 28;		/* time to delay syncing metadata */
309 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
310     "Time to delay syncing metadata (in seconds)");
311 static int rushjob;		/* number of slots to run ASAP */
312 static int stat_rush_requests;	/* number of times I/O speeded up */
313 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
314     "Number of times I/O speeded up (rush requests)");
315 
316 #define	VDBATCH_SIZE 8
317 struct vdbatch {
318 	u_int index;
319 	struct mtx lock;
320 	struct vnode *tab[VDBATCH_SIZE];
321 };
322 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
323 
324 static void	vdbatch_dequeue(struct vnode *vp);
325 
326 /*
327  * The syncer will require at least SYNCER_MAXDELAY iterations to shutdown;
328  * we probably don't want to pause for the whole second each time.
329  */
330 #define SYNCER_SHUTDOWN_SPEEDUP		32
331 static int sync_vnode_count;
332 static int syncer_worklist_len;
333 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
334     syncer_state;
335 
336 /* Target for maximum number of vnodes. */
337 u_long desiredvnodes;
338 static u_long gapvnodes;		/* gap between wanted and desired */
339 static u_long vhiwat;		/* enough extras after expansion */
340 static u_long vlowat;		/* minimal extras before expansion */
341 static bool vstir;		/* nonzero to stir non-free vnodes */
342 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
343 
344 static u_long vnlru_read_freevnodes(void);
345 
346 /*
347  * Note that no attempt is made to sanitize these parameters.
348  */
349 static int
350 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
351 {
352 	u_long val;
353 	int error;
354 
355 	val = desiredvnodes;
356 	error = sysctl_handle_long(oidp, &val, 0, req);
357 	if (error != 0 || req->newptr == NULL)
358 		return (error);
359 
360 	if (val == desiredvnodes)
361 		return (0);
362 	mtx_lock(&vnode_list_mtx);
363 	desiredvnodes = val;
364 	wantfreevnodes = desiredvnodes / 4;
365 	vnlru_recalc();
366 	mtx_unlock(&vnode_list_mtx);
367 	/*
368 	 * XXX There is no protection against multiple threads changing
369 	 * desiredvnodes at the same time. Locking above only helps vnlru and
370 	 * getnewvnode.
371 	 */
372 	vfs_hash_changesize(desiredvnodes);
373 	cache_changesize(desiredvnodes);
374 	return (0);
375 }
376 
377 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
378     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
379     "LU", "Target for maximum number of vnodes (legacy)");
380 SYSCTL_PROC(_vfs_vnode_param, OID_AUTO, limit,
381     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
382     "LU", "Target for maximum number of vnodes");
383 
384 static int
385 sysctl_freevnodes(SYSCTL_HANDLER_ARGS)
386 {
387 	u_long rfreevnodes;
388 
389 	rfreevnodes = vnlru_read_freevnodes();
390 	return (sysctl_handle_long(oidp, &rfreevnodes, 0, req));
391 }
392 
393 SYSCTL_PROC(_vfs, OID_AUTO, freevnodes,
394     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_freevnodes,
395     "LU", "Number of \"free\" vnodes (legacy)");
396 SYSCTL_PROC(_vfs_vnode_stats, OID_AUTO, free,
397     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_freevnodes,
398     "LU", "Number of \"free\" vnodes");
399 
400 static int
401 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
402 {
403 	u_long val;
404 	int error;
405 
406 	val = wantfreevnodes;
407 	error = sysctl_handle_long(oidp, &val, 0, req);
408 	if (error != 0 || req->newptr == NULL)
409 		return (error);
410 
411 	if (val == wantfreevnodes)
412 		return (0);
413 	mtx_lock(&vnode_list_mtx);
414 	wantfreevnodes = val;
415 	vnlru_recalc();
416 	mtx_unlock(&vnode_list_mtx);
417 	return (0);
418 }
419 
420 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
421     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
422     "LU", "Target for minimum number of \"free\" vnodes (legacy)");
423 SYSCTL_PROC(_vfs_vnode_param, OID_AUTO, wantfree,
424     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
425     "LU", "Target for minimum number of \"free\" vnodes");
426 
427 static int vnlru_nowhere;
428 SYSCTL_INT(_vfs_vnode_vnlru, OID_AUTO, failed_runs, CTLFLAG_RD | CTLFLAG_STATS,
429     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
430 
431 static int
432 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
433 {
434 	struct vnode *vp;
435 	struct nameidata nd;
436 	char *buf;
437 	unsigned long ndflags;
438 	int error;
439 
440 	if (req->newptr == NULL)
441 		return (EINVAL);
442 	if (req->newlen >= PATH_MAX)
443 		return (E2BIG);
444 
445 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
446 	error = SYSCTL_IN(req, buf, req->newlen);
447 	if (error != 0)
448 		goto out;
449 
450 	buf[req->newlen] = '\0';
451 
452 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1;
453 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf);
454 	if ((error = namei(&nd)) != 0)
455 		goto out;
456 	vp = nd.ni_vp;
457 
458 	if (VN_IS_DOOMED(vp)) {
459 		/*
460 		 * This vnode is being recycled.  Return != 0 to let the caller
461 		 * know that the sysctl had no effect.  Return EAGAIN because a
462 		 * subsequent call will likely succeed (since namei will create
463 		 * a new vnode if necessary)
464 		 */
465 		error = EAGAIN;
466 		goto putvnode;
467 	}
468 
469 	vgone(vp);
470 putvnode:
471 	vput(vp);
472 	NDFREE_PNBUF(&nd);
473 out:
474 	free(buf, M_TEMP);
475 	return (error);
476 }
477 
478 static int
479 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
480 {
481 	struct thread *td = curthread;
482 	struct vnode *vp;
483 	struct file *fp;
484 	int error;
485 	int fd;
486 
487 	if (req->newptr == NULL)
488 		return (EBADF);
489 
490         error = sysctl_handle_int(oidp, &fd, 0, req);
491         if (error != 0)
492                 return (error);
493 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
494 	if (error != 0)
495 		return (error);
496 	vp = fp->f_vnode;
497 
498 	error = vn_lock(vp, LK_EXCLUSIVE);
499 	if (error != 0)
500 		goto drop;
501 
502 	vgone(vp);
503 	VOP_UNLOCK(vp);
504 drop:
505 	fdrop(fp, td);
506 	return (error);
507 }
508 
509 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
510     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
511     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
512 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
513     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
514     sysctl_ftry_reclaim_vnode, "I",
515     "Try to reclaim a vnode by its file descriptor");
516 
517 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
518 #define vnsz2log 8
519 #ifndef DEBUG_LOCKS
520 _Static_assert(sizeof(struct vnode) >= 1UL << vnsz2log &&
521     sizeof(struct vnode) < 1UL << (vnsz2log + 1),
522     "vnsz2log needs to be updated");
523 #endif
524 
525 /*
526  * Support for the bufobj clean & dirty pctrie.
527  */
528 static void *
529 buf_trie_alloc(struct pctrie *ptree)
530 {
531 	return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT));
532 }
533 
534 static void
535 buf_trie_free(struct pctrie *ptree, void *node)
536 {
537 	uma_zfree_smr(buf_trie_zone, node);
538 }
539 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free,
540     buf_trie_smr);
541 
542 /*
543  * Lookup the next element greater than or equal to lblkno, accounting for the
544  * fact that, for pctries, negative values are greater than nonnegative ones.
545  */
546 static struct buf *
547 buf_lookup_ge(struct bufv *bv, daddr_t lblkno)
548 {
549 	struct buf *bp;
550 
551 	bp = BUF_PCTRIE_LOOKUP_GE(&bv->bv_root, lblkno);
552 	if (bp == NULL && lblkno < 0)
553 		bp = BUF_PCTRIE_LOOKUP_GE(&bv->bv_root, 0);
554 	if (bp != NULL && bp->b_lblkno < lblkno)
555 		bp = NULL;
556 	return (bp);
557 }
558 
559 /*
560  * Insert bp, and find the next element smaller than bp, accounting for the fact
561  * that, for pctries, negative values are greater than nonnegative ones.
562  */
563 static int
564 buf_insert_lookup_le(struct bufv *bv, struct buf *bp, struct buf **n)
565 {
566 	int error;
567 
568 	error = BUF_PCTRIE_INSERT_LOOKUP_LE(&bv->bv_root, bp, n);
569 	if (error != EEXIST) {
570 		if (*n == NULL && bp->b_lblkno >= 0)
571 			*n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, ~0L);
572 		if (*n != NULL && (*n)->b_lblkno >= bp->b_lblkno)
573 			*n = NULL;
574 	}
575 	return (error);
576 }
577 
578 /*
579  * Initialize the vnode management data structures.
580  *
581  * Reevaluate the following cap on the number of vnodes after the physical
582  * memory size exceeds 512GB.  In the limit, as the physical memory size
583  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
584  */
585 #ifndef	MAXVNODES_MAX
586 #define	MAXVNODES_MAX	(512UL * 1024 * 1024 / 64)	/* 8M */
587 #endif
588 
589 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
590 
591 static struct vnode *
592 vn_alloc_marker(struct mount *mp)
593 {
594 	struct vnode *vp;
595 
596 	vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
597 	vp->v_type = VMARKER;
598 	vp->v_mount = mp;
599 
600 	return (vp);
601 }
602 
603 static void
604 vn_free_marker(struct vnode *vp)
605 {
606 
607 	MPASS(vp->v_type == VMARKER);
608 	free(vp, M_VNODE_MARKER);
609 }
610 
611 #ifdef KASAN
612 static int
613 vnode_ctor(void *mem, int size, void *arg __unused, int flags __unused)
614 {
615 	kasan_mark(mem, size, roundup2(size, UMA_ALIGN_PTR + 1), 0);
616 	return (0);
617 }
618 
619 static void
620 vnode_dtor(void *mem, int size, void *arg __unused)
621 {
622 	size_t end1, end2, off1, off2;
623 
624 	_Static_assert(offsetof(struct vnode, v_vnodelist) <
625 	    offsetof(struct vnode, v_dbatchcpu),
626 	    "KASAN marks require updating");
627 
628 	off1 = offsetof(struct vnode, v_vnodelist);
629 	off2 = offsetof(struct vnode, v_dbatchcpu);
630 	end1 = off1 + sizeof(((struct vnode *)NULL)->v_vnodelist);
631 	end2 = off2 + sizeof(((struct vnode *)NULL)->v_dbatchcpu);
632 
633 	/*
634 	 * Access to the v_vnodelist and v_dbatchcpu fields are permitted even
635 	 * after the vnode has been freed.  Try to get some KASAN coverage by
636 	 * marking everything except those two fields as invalid.  Because
637 	 * KASAN's tracking is not byte-granular, any preceding fields sharing
638 	 * the same 8-byte aligned word must also be marked valid.
639 	 */
640 
641 	/* Handle the area from the start until v_vnodelist... */
642 	off1 = rounddown2(off1, KASAN_SHADOW_SCALE);
643 	kasan_mark(mem, off1, off1, KASAN_UMA_FREED);
644 
645 	/* ... then the area between v_vnodelist and v_dbatchcpu ... */
646 	off1 = roundup2(end1, KASAN_SHADOW_SCALE);
647 	off2 = rounddown2(off2, KASAN_SHADOW_SCALE);
648 	if (off2 > off1)
649 		kasan_mark((void *)((char *)mem + off1), off2 - off1,
650 		    off2 - off1, KASAN_UMA_FREED);
651 
652 	/* ... and finally the area from v_dbatchcpu to the end. */
653 	off2 = roundup2(end2, KASAN_SHADOW_SCALE);
654 	kasan_mark((void *)((char *)mem + off2), size - off2, size - off2,
655 	    KASAN_UMA_FREED);
656 }
657 #endif /* KASAN */
658 
659 /*
660  * Initialize a vnode as it first enters the zone.
661  */
662 static int
663 vnode_init(void *mem, int size, int flags)
664 {
665 	struct vnode *vp;
666 
667 	vp = mem;
668 	bzero(vp, size);
669 	/*
670 	 * Setup locks.
671 	 */
672 	vp->v_vnlock = &vp->v_lock;
673 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
674 	/*
675 	 * By default, don't allow shared locks unless filesystems opt-in.
676 	 */
677 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
678 	    LK_NOSHARE | LK_IS_VNODE);
679 	/*
680 	 * Initialize bufobj.
681 	 */
682 	bufobj_init(&vp->v_bufobj, vp);
683 	/*
684 	 * Initialize namecache.
685 	 */
686 	cache_vnode_init(vp);
687 	/*
688 	 * Initialize rangelocks.
689 	 */
690 	rangelock_init(&vp->v_rl);
691 
692 	vp->v_dbatchcpu = NOCPU;
693 
694 	vp->v_state = VSTATE_DEAD;
695 
696 	/*
697 	 * Check vhold_recycle_free for an explanation.
698 	 */
699 	vp->v_holdcnt = VHOLD_NO_SMR;
700 	vp->v_type = VNON;
701 	mtx_lock(&vnode_list_mtx);
702 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
703 	mtx_unlock(&vnode_list_mtx);
704 	return (0);
705 }
706 
707 /*
708  * Free a vnode when it is cleared from the zone.
709  */
710 static void
711 vnode_fini(void *mem, int size)
712 {
713 	struct vnode *vp;
714 	struct bufobj *bo;
715 
716 	vp = mem;
717 	vdbatch_dequeue(vp);
718 	mtx_lock(&vnode_list_mtx);
719 	TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
720 	mtx_unlock(&vnode_list_mtx);
721 	rangelock_destroy(&vp->v_rl);
722 	lockdestroy(vp->v_vnlock);
723 	mtx_destroy(&vp->v_interlock);
724 	bo = &vp->v_bufobj;
725 	rw_destroy(BO_LOCKPTR(bo));
726 
727 	kasan_mark(mem, size, size, 0);
728 }
729 
730 /*
731  * Provide the size of NFS nclnode and NFS fh for calculation of the
732  * vnode memory consumption.  The size is specified directly to
733  * eliminate dependency on NFS-private header.
734  *
735  * Other filesystems may use bigger or smaller (like UFS and ZFS)
736  * private inode data, but the NFS-based estimation is ample enough.
737  * Still, we care about differences in the size between 64- and 32-bit
738  * platforms.
739  *
740  * Namecache structure size is heuristically
741  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
742  */
743 #ifdef _LP64
744 #define	NFS_NCLNODE_SZ	(528 + 64)
745 #define	NC_SZ		148
746 #else
747 #define	NFS_NCLNODE_SZ	(360 + 32)
748 #define	NC_SZ		92
749 #endif
750 
751 static void
752 vntblinit(void *dummy __unused)
753 {
754 	struct vdbatch *vd;
755 	uma_ctor ctor;
756 	uma_dtor dtor;
757 	int cpu, physvnodes, virtvnodes;
758 
759 	/*
760 	 * 'desiredvnodes' is the minimum of a function of the physical memory
761 	 * size and another of the kernel heap size (UMA limit, a portion of the
762 	 * KVA).
763 	 *
764 	 * Currently, on 64-bit platforms, 'desiredvnodes' is set to
765 	 * 'virtvnodes' up to a physical memory cutoff of ~1722MB, after which
766 	 * 'physvnodes' applies instead.  With the current automatic tuning for
767 	 * 'maxfiles' (32 files/MB), 'desiredvnodes' is always greater than it.
768 	 */
769 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 32 +
770 	    min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 32;
771 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
772 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
773 	desiredvnodes = min(physvnodes, virtvnodes);
774 	if (desiredvnodes > MAXVNODES_MAX) {
775 		if (bootverbose)
776 			printf("Reducing kern.maxvnodes %lu -> %lu\n",
777 			    desiredvnodes, MAXVNODES_MAX);
778 		desiredvnodes = MAXVNODES_MAX;
779 	}
780 	wantfreevnodes = desiredvnodes / 4;
781 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
782 	TAILQ_INIT(&vnode_list);
783 	mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
784 	/*
785 	 * The lock is taken to appease WITNESS.
786 	 */
787 	mtx_lock(&vnode_list_mtx);
788 	vnlru_recalc();
789 	mtx_unlock(&vnode_list_mtx);
790 	vnode_list_free_marker = vn_alloc_marker(NULL);
791 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
792 	vnode_list_reclaim_marker = vn_alloc_marker(NULL);
793 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
794 
795 #ifdef KASAN
796 	ctor = vnode_ctor;
797 	dtor = vnode_dtor;
798 #else
799 	ctor = NULL;
800 	dtor = NULL;
801 #endif
802 	vnode_zone = uma_zcreate("VNODE", sizeof(struct vnode), ctor, dtor,
803 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, UMA_ZONE_NOKASAN);
804 	uma_zone_set_smr(vnode_zone, vfs_smr);
805 
806 	/*
807 	 * Preallocate enough nodes to support one-per buf so that
808 	 * we can not fail an insert.  reassignbuf() callers can not
809 	 * tolerate the insertion failure.
810 	 */
811 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
812 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
813 	    UMA_ZONE_NOFREE | UMA_ZONE_SMR);
814 	buf_trie_smr = uma_zone_get_smr(buf_trie_zone);
815 	uma_prealloc(buf_trie_zone, nbuf);
816 
817 	vnodes_created = counter_u64_alloc(M_WAITOK);
818 	direct_recycles_free_count = counter_u64_alloc(M_WAITOK);
819 	vnode_skipped_requeues = counter_u64_alloc(M_WAITOK);
820 
821 	/*
822 	 * Initialize the filesystem syncer.
823 	 */
824 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
825 	    &syncer_mask);
826 	syncer_maxdelay = syncer_mask + 1;
827 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
828 	cv_init(&sync_wakeup, "syncer");
829 
830 	CPU_FOREACH(cpu) {
831 		vd = DPCPU_ID_PTR((cpu), vd);
832 		bzero(vd, sizeof(*vd));
833 		mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
834 	}
835 }
836 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
837 
838 /*
839  * Mark a mount point as busy. Used to synchronize access and to delay
840  * unmounting. Eventually, mountlist_mtx is not released on failure.
841  *
842  * vfs_busy() is a custom lock, it can block the caller.
843  * vfs_busy() only sleeps if the unmount is active on the mount point.
844  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
845  * vnode belonging to mp.
846  *
847  * Lookup uses vfs_busy() to traverse mount points.
848  * root fs			var fs
849  * / vnode lock		A	/ vnode lock (/var)		D
850  * /var vnode lock	B	/log vnode lock(/var/log)	E
851  * vfs_busy lock	C	vfs_busy lock			F
852  *
853  * Within each file system, the lock order is C->A->B and F->D->E.
854  *
855  * When traversing across mounts, the system follows that lock order:
856  *
857  *        C->A->B
858  *              |
859  *              +->F->D->E
860  *
861  * The lookup() process for namei("/var") illustrates the process:
862  *  1. VOP_LOOKUP() obtains B while A is held
863  *  2. vfs_busy() obtains a shared lock on F while A and B are held
864  *  3. vput() releases lock on B
865  *  4. vput() releases lock on A
866  *  5. VFS_ROOT() obtains lock on D while shared lock on F is held
867  *  6. vfs_unbusy() releases shared lock on F
868  *  7. vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
869  *     Attempt to lock A (instead of vp_crossmp) while D is held would
870  *     violate the global order, causing deadlocks.
871  *
872  * dounmount() locks B while F is drained.  Note that for stacked
873  * filesystems, D and B in the example above may be the same lock,
874  * which introdues potential lock order reversal deadlock between
875  * dounmount() and step 5 above.  These filesystems may avoid the LOR
876  * by setting VV_CROSSLOCK on the covered vnode so that lock B will
877  * remain held until after step 5.
878  */
879 int
880 vfs_busy(struct mount *mp, int flags)
881 {
882 	struct mount_pcpu *mpcpu;
883 
884 	MPASS((flags & ~MBF_MASK) == 0);
885 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
886 
887 	if (vfs_op_thread_enter(mp, mpcpu)) {
888 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
889 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
890 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
891 		vfs_mp_count_add_pcpu(mpcpu, ref, 1);
892 		vfs_mp_count_add_pcpu(mpcpu, lockref, 1);
893 		vfs_op_thread_exit(mp, mpcpu);
894 		if (flags & MBF_MNTLSTLOCK)
895 			mtx_unlock(&mountlist_mtx);
896 		return (0);
897 	}
898 
899 	MNT_ILOCK(mp);
900 	vfs_assert_mount_counters(mp);
901 	MNT_REF(mp);
902 	/*
903 	 * If mount point is currently being unmounted, sleep until the
904 	 * mount point fate is decided.  If thread doing the unmounting fails,
905 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
906 	 * that this mount point has survived the unmount attempt and vfs_busy
907 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
908 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
909 	 * about to be really destroyed.  vfs_busy needs to release its
910 	 * reference on the mount point in this case and return with ENOENT,
911 	 * telling the caller the mount it tried to busy is no longer valid.
912 	 */
913 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
914 		KASSERT(TAILQ_EMPTY(&mp->mnt_uppers),
915 		    ("%s: non-empty upper mount list with pending unmount",
916 		    __func__));
917 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
918 			MNT_REL(mp);
919 			MNT_IUNLOCK(mp);
920 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
921 			    __func__);
922 			return (ENOENT);
923 		}
924 		if (flags & MBF_MNTLSTLOCK)
925 			mtx_unlock(&mountlist_mtx);
926 		mp->mnt_kern_flag |= MNTK_MWAIT;
927 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
928 		if (flags & MBF_MNTLSTLOCK)
929 			mtx_lock(&mountlist_mtx);
930 		MNT_ILOCK(mp);
931 	}
932 	if (flags & MBF_MNTLSTLOCK)
933 		mtx_unlock(&mountlist_mtx);
934 	mp->mnt_lockref++;
935 	MNT_IUNLOCK(mp);
936 	return (0);
937 }
938 
939 /*
940  * Free a busy filesystem.
941  */
942 void
943 vfs_unbusy(struct mount *mp)
944 {
945 	struct mount_pcpu *mpcpu;
946 	int c;
947 
948 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
949 
950 	if (vfs_op_thread_enter(mp, mpcpu)) {
951 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
952 		vfs_mp_count_sub_pcpu(mpcpu, lockref, 1);
953 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
954 		vfs_op_thread_exit(mp, mpcpu);
955 		return;
956 	}
957 
958 	MNT_ILOCK(mp);
959 	vfs_assert_mount_counters(mp);
960 	MNT_REL(mp);
961 	c = --mp->mnt_lockref;
962 	if (mp->mnt_vfs_ops == 0) {
963 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
964 		MNT_IUNLOCK(mp);
965 		return;
966 	}
967 	if (c < 0)
968 		vfs_dump_mount_counters(mp);
969 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
970 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
971 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
972 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
973 		wakeup(&mp->mnt_lockref);
974 	}
975 	MNT_IUNLOCK(mp);
976 }
977 
978 /*
979  * Lookup a mount point by filesystem identifier.
980  */
981 struct mount *
982 vfs_getvfs(fsid_t *fsid)
983 {
984 	struct mount *mp;
985 
986 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
987 	mtx_lock(&mountlist_mtx);
988 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
989 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
990 			vfs_ref(mp);
991 			mtx_unlock(&mountlist_mtx);
992 			return (mp);
993 		}
994 	}
995 	mtx_unlock(&mountlist_mtx);
996 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
997 	return ((struct mount *) 0);
998 }
999 
1000 /*
1001  * Lookup a mount point by filesystem identifier, busying it before
1002  * returning.
1003  *
1004  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
1005  * cache for popular filesystem identifiers.  The cache is lockess, using
1006  * the fact that struct mount's are never freed.  In worst case we may
1007  * get pointer to unmounted or even different filesystem, so we have to
1008  * check what we got, and go slow way if so.
1009  */
1010 struct mount *
1011 vfs_busyfs(fsid_t *fsid)
1012 {
1013 #define	FSID_CACHE_SIZE	256
1014 	typedef struct mount * volatile vmp_t;
1015 	static vmp_t cache[FSID_CACHE_SIZE];
1016 	struct mount *mp;
1017 	int error;
1018 	uint32_t hash;
1019 
1020 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
1021 	hash = fsid->val[0] ^ fsid->val[1];
1022 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
1023 	mp = cache[hash];
1024 	if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0)
1025 		goto slow;
1026 	if (vfs_busy(mp, 0) != 0) {
1027 		cache[hash] = NULL;
1028 		goto slow;
1029 	}
1030 	if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0)
1031 		return (mp);
1032 	else
1033 	    vfs_unbusy(mp);
1034 
1035 slow:
1036 	mtx_lock(&mountlist_mtx);
1037 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
1038 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
1039 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
1040 			if (error) {
1041 				cache[hash] = NULL;
1042 				mtx_unlock(&mountlist_mtx);
1043 				return (NULL);
1044 			}
1045 			cache[hash] = mp;
1046 			return (mp);
1047 		}
1048 	}
1049 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
1050 	mtx_unlock(&mountlist_mtx);
1051 	return ((struct mount *) 0);
1052 }
1053 
1054 /*
1055  * Check if a user can access privileged mount options.
1056  */
1057 int
1058 vfs_suser(struct mount *mp, struct thread *td)
1059 {
1060 	int error;
1061 
1062 	if (jailed(td->td_ucred)) {
1063 		/*
1064 		 * If the jail of the calling thread lacks permission for
1065 		 * this type of file system, deny immediately.
1066 		 */
1067 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
1068 			return (EPERM);
1069 
1070 		/*
1071 		 * If the file system was mounted outside the jail of the
1072 		 * calling thread, deny immediately.
1073 		 */
1074 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
1075 			return (EPERM);
1076 	}
1077 
1078 	/*
1079 	 * If file system supports delegated administration, we don't check
1080 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
1081 	 * by the file system itself.
1082 	 * If this is not the user that did original mount, we check for
1083 	 * the PRIV_VFS_MOUNT_OWNER privilege.
1084 	 */
1085 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
1086 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
1087 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
1088 			return (error);
1089 	}
1090 	return (0);
1091 }
1092 
1093 /*
1094  * Get a new unique fsid.  Try to make its val[0] unique, since this value
1095  * will be used to create fake device numbers for stat().  Also try (but
1096  * not so hard) make its val[0] unique mod 2^16, since some emulators only
1097  * support 16-bit device numbers.  We end up with unique val[0]'s for the
1098  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
1099  *
1100  * Keep in mind that several mounts may be running in parallel.  Starting
1101  * the search one past where the previous search terminated is both a
1102  * micro-optimization and a defense against returning the same fsid to
1103  * different mounts.
1104  */
1105 void
1106 vfs_getnewfsid(struct mount *mp)
1107 {
1108 	static uint16_t mntid_base;
1109 	struct mount *nmp;
1110 	fsid_t tfsid;
1111 	int mtype;
1112 
1113 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
1114 	mtx_lock(&mntid_mtx);
1115 	mtype = mp->mnt_vfc->vfc_typenum;
1116 	tfsid.val[1] = mtype;
1117 	mtype = (mtype & 0xFF) << 24;
1118 	for (;;) {
1119 		tfsid.val[0] = makedev(255,
1120 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
1121 		mntid_base++;
1122 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
1123 			break;
1124 		vfs_rel(nmp);
1125 	}
1126 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
1127 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
1128 	mtx_unlock(&mntid_mtx);
1129 }
1130 
1131 /*
1132  * Knob to control the precision of file timestamps:
1133  *
1134  *   0 = seconds only; nanoseconds zeroed.
1135  *   1 = seconds and nanoseconds, accurate within 1/HZ.
1136  *   2 = seconds and nanoseconds, truncated to microseconds.
1137  * >=3 = seconds and nanoseconds, maximum precision.
1138  */
1139 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
1140 
1141 static int timestamp_precision = TSP_USEC;
1142 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
1143     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
1144     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1145     "3+: sec + ns (max. precision))");
1146 
1147 /*
1148  * Get a current timestamp.
1149  */
1150 void
1151 vfs_timestamp(struct timespec *tsp)
1152 {
1153 	struct timeval tv;
1154 
1155 	switch (timestamp_precision) {
1156 	case TSP_SEC:
1157 		tsp->tv_sec = time_second;
1158 		tsp->tv_nsec = 0;
1159 		break;
1160 	case TSP_HZ:
1161 		getnanotime(tsp);
1162 		break;
1163 	case TSP_USEC:
1164 		microtime(&tv);
1165 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1166 		break;
1167 	case TSP_NSEC:
1168 	default:
1169 		nanotime(tsp);
1170 		break;
1171 	}
1172 }
1173 
1174 /*
1175  * Set vnode attributes to VNOVAL
1176  */
1177 void
1178 vattr_null(struct vattr *vap)
1179 {
1180 
1181 	vap->va_type = VNON;
1182 	vap->va_size = VNOVAL;
1183 	vap->va_bytes = VNOVAL;
1184 	vap->va_mode = VNOVAL;
1185 	vap->va_nlink = VNOVAL;
1186 	vap->va_uid = VNOVAL;
1187 	vap->va_gid = VNOVAL;
1188 	vap->va_fsid = VNOVAL;
1189 	vap->va_fileid = VNOVAL;
1190 	vap->va_blocksize = VNOVAL;
1191 	vap->va_rdev = VNOVAL;
1192 	vap->va_atime.tv_sec = VNOVAL;
1193 	vap->va_atime.tv_nsec = VNOVAL;
1194 	vap->va_mtime.tv_sec = VNOVAL;
1195 	vap->va_mtime.tv_nsec = VNOVAL;
1196 	vap->va_ctime.tv_sec = VNOVAL;
1197 	vap->va_ctime.tv_nsec = VNOVAL;
1198 	vap->va_birthtime.tv_sec = VNOVAL;
1199 	vap->va_birthtime.tv_nsec = VNOVAL;
1200 	vap->va_flags = VNOVAL;
1201 	vap->va_gen = VNOVAL;
1202 	vap->va_vaflags = 0;
1203 	vap->va_filerev = VNOVAL;
1204 	vap->va_bsdflags = 0;
1205 }
1206 
1207 /*
1208  * Try to reduce the total number of vnodes.
1209  *
1210  * This routine (and its user) are buggy in at least the following ways:
1211  * - all parameters were picked years ago when RAM sizes were significantly
1212  *   smaller
1213  * - it can pick vnodes based on pages used by the vm object, but filesystems
1214  *   like ZFS don't use it making the pick broken
1215  * - since ZFS has its own aging policy it gets partially combated by this one
1216  * - a dedicated method should be provided for filesystems to let them decide
1217  *   whether the vnode should be recycled
1218  *
1219  * This routine is called when we have too many vnodes.  It attempts
1220  * to free <count> vnodes and will potentially free vnodes that still
1221  * have VM backing store (VM backing store is typically the cause
1222  * of a vnode blowout so we want to do this).  Therefore, this operation
1223  * is not considered cheap.
1224  *
1225  * A number of conditions may prevent a vnode from being reclaimed.
1226  * the buffer cache may have references on the vnode, a directory
1227  * vnode may still have references due to the namei cache representing
1228  * underlying files, or the vnode may be in active use.   It is not
1229  * desirable to reuse such vnodes.  These conditions may cause the
1230  * number of vnodes to reach some minimum value regardless of what
1231  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
1232  *
1233  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1234  * 			 entries if this argument is strue
1235  * @param trigger	 Only reclaim vnodes with fewer than this many resident
1236  *			 pages.
1237  * @param target	 How many vnodes to reclaim.
1238  * @return		 The number of vnodes that were reclaimed.
1239  */
1240 static int
1241 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1242 {
1243 	struct vnode *vp, *mvp;
1244 	struct mount *mp;
1245 	struct vm_object *object;
1246 	u_long done;
1247 	bool retried;
1248 
1249 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1250 
1251 	retried = false;
1252 	done = 0;
1253 
1254 	mvp = vnode_list_reclaim_marker;
1255 restart:
1256 	vp = mvp;
1257 	while (done < target) {
1258 		vp = TAILQ_NEXT(vp, v_vnodelist);
1259 		if (__predict_false(vp == NULL))
1260 			break;
1261 
1262 		if (__predict_false(vp->v_type == VMARKER))
1263 			continue;
1264 
1265 		/*
1266 		 * If it's been deconstructed already, it's still
1267 		 * referenced, or it exceeds the trigger, skip it.
1268 		 * Also skip free vnodes.  We are trying to make space
1269 		 * for more free vnodes, not reduce their count.
1270 		 */
1271 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1272 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1273 			goto next_iter;
1274 
1275 		if (vp->v_type == VBAD || vp->v_type == VNON)
1276 			goto next_iter;
1277 
1278 		object = atomic_load_ptr(&vp->v_object);
1279 		if (object == NULL || object->resident_page_count > trigger) {
1280 			goto next_iter;
1281 		}
1282 
1283 		/*
1284 		 * Handle races against vnode allocation. Filesystems lock the
1285 		 * vnode some time after it gets returned from getnewvnode,
1286 		 * despite type and hold count being manipulated earlier.
1287 		 * Resorting to checking v_mount restores guarantees present
1288 		 * before the global list was reworked to contain all vnodes.
1289 		 */
1290 		if (!VI_TRYLOCK(vp))
1291 			goto next_iter;
1292 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1293 			VI_UNLOCK(vp);
1294 			goto next_iter;
1295 		}
1296 		if (vp->v_mount == NULL) {
1297 			VI_UNLOCK(vp);
1298 			goto next_iter;
1299 		}
1300 		vholdl(vp);
1301 		VI_UNLOCK(vp);
1302 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1303 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1304 		mtx_unlock(&vnode_list_mtx);
1305 
1306 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1307 			vdrop_recycle(vp);
1308 			goto next_iter_unlocked;
1309 		}
1310 		if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1311 			vdrop_recycle(vp);
1312 			vn_finished_write(mp);
1313 			goto next_iter_unlocked;
1314 		}
1315 
1316 		VI_LOCK(vp);
1317 		if (vp->v_usecount > 0 ||
1318 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1319 		    (vp->v_object != NULL && vp->v_object->handle == vp &&
1320 		    vp->v_object->resident_page_count > trigger)) {
1321 			VOP_UNLOCK(vp);
1322 			vdropl_recycle(vp);
1323 			vn_finished_write(mp);
1324 			goto next_iter_unlocked;
1325 		}
1326 		recycles_count++;
1327 		vgonel(vp);
1328 		VOP_UNLOCK(vp);
1329 		vdropl_recycle(vp);
1330 		vn_finished_write(mp);
1331 		done++;
1332 next_iter_unlocked:
1333 		maybe_yield();
1334 		mtx_lock(&vnode_list_mtx);
1335 		goto restart;
1336 next_iter:
1337 		MPASS(vp->v_type != VMARKER);
1338 		if (!should_yield())
1339 			continue;
1340 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1341 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1342 		mtx_unlock(&vnode_list_mtx);
1343 		kern_yield(PRI_USER);
1344 		mtx_lock(&vnode_list_mtx);
1345 		goto restart;
1346 	}
1347 	if (done == 0 && !retried) {
1348 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1349 		TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1350 		retried = true;
1351 		goto restart;
1352 	}
1353 	return (done);
1354 }
1355 
1356 static int max_free_per_call = 10000;
1357 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_free_per_call, 0,
1358     "limit on vnode free requests per call to the vnlru_free routine (legacy)");
1359 SYSCTL_INT(_vfs_vnode_vnlru, OID_AUTO, max_free_per_call, CTLFLAG_RW,
1360     &max_free_per_call, 0,
1361     "limit on vnode free requests per call to the vnlru_free routine");
1362 
1363 /*
1364  * Attempt to recycle requested amount of free vnodes.
1365  */
1366 static int
1367 vnlru_free_impl(int count, struct vfsops *mnt_op, struct vnode *mvp, bool isvnlru)
1368 {
1369 	struct vnode *vp;
1370 	struct mount *mp;
1371 	int ocount;
1372 	bool retried;
1373 
1374 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1375 	if (count > max_free_per_call)
1376 		count = max_free_per_call;
1377 	if (count == 0) {
1378 		mtx_unlock(&vnode_list_mtx);
1379 		return (0);
1380 	}
1381 	ocount = count;
1382 	retried = false;
1383 	vp = mvp;
1384 	for (;;) {
1385 		vp = TAILQ_NEXT(vp, v_vnodelist);
1386 		if (__predict_false(vp == NULL)) {
1387 			/*
1388 			 * The free vnode marker can be past eligible vnodes:
1389 			 * 1. if vdbatch_process trylock failed
1390 			 * 2. if vtryrecycle failed
1391 			 *
1392 			 * If so, start the scan from scratch.
1393 			 */
1394 			if (!retried && vnlru_read_freevnodes() > 0) {
1395 				TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1396 				TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1397 				vp = mvp;
1398 				retried = true;
1399 				continue;
1400 			}
1401 
1402 			/*
1403 			 * Give up
1404 			 */
1405 			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1406 			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1407 			mtx_unlock(&vnode_list_mtx);
1408 			break;
1409 		}
1410 		if (__predict_false(vp->v_type == VMARKER))
1411 			continue;
1412 		if (vp->v_holdcnt > 0)
1413 			continue;
1414 		/*
1415 		 * Don't recycle if our vnode is from different type
1416 		 * of mount point.  Note that mp is type-safe, the
1417 		 * check does not reach unmapped address even if
1418 		 * vnode is reclaimed.
1419 		 */
1420 		if (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1421 		    mp->mnt_op != mnt_op) {
1422 			continue;
1423 		}
1424 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1425 			continue;
1426 		}
1427 		if (!vhold_recycle_free(vp))
1428 			continue;
1429 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1430 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1431 		mtx_unlock(&vnode_list_mtx);
1432 		/*
1433 		 * FIXME: ignores the return value, meaning it may be nothing
1434 		 * got recycled but it claims otherwise to the caller.
1435 		 *
1436 		 * Originally the value started being ignored in 2005 with
1437 		 * 114a1006a8204aa156e1f9ad6476cdff89cada7f .
1438 		 *
1439 		 * Respecting the value can run into significant stalls if most
1440 		 * vnodes belong to one file system and it has writes
1441 		 * suspended.  In presence of many threads and millions of
1442 		 * vnodes they keep contending on the vnode_list_mtx lock only
1443 		 * to find vnodes they can't recycle.
1444 		 *
1445 		 * The solution would be to pre-check if the vnode is likely to
1446 		 * be recycle-able, but it needs to happen with the
1447 		 * vnode_list_mtx lock held. This runs into a problem where
1448 		 * VOP_GETWRITEMOUNT (currently needed to find out about if
1449 		 * writes are frozen) can take locks which LOR against it.
1450 		 *
1451 		 * Check nullfs for one example (null_getwritemount).
1452 		 */
1453 		vtryrecycle(vp, isvnlru);
1454 		count--;
1455 		if (count == 0) {
1456 			break;
1457 		}
1458 		mtx_lock(&vnode_list_mtx);
1459 		vp = mvp;
1460 	}
1461 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1462 	return (ocount - count);
1463 }
1464 
1465 /*
1466  * XXX: returns without vnode_list_mtx locked!
1467  */
1468 static int
1469 vnlru_free_locked_direct(int count)
1470 {
1471 	int ret;
1472 
1473 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1474 	ret = vnlru_free_impl(count, NULL, vnode_list_free_marker, false);
1475 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1476 	return (ret);
1477 }
1478 
1479 static int
1480 vnlru_free_locked_vnlru(int count)
1481 {
1482 	int ret;
1483 
1484 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1485 	ret = vnlru_free_impl(count, NULL, vnode_list_free_marker, true);
1486 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1487 	return (ret);
1488 }
1489 
1490 static int
1491 vnlru_free_vnlru(int count)
1492 {
1493 
1494 	mtx_lock(&vnode_list_mtx);
1495 	return (vnlru_free_locked_vnlru(count));
1496 }
1497 
1498 void
1499 vnlru_free_vfsops(int count, struct vfsops *mnt_op, struct vnode *mvp)
1500 {
1501 
1502 	MPASS(mnt_op != NULL);
1503 	MPASS(mvp != NULL);
1504 	VNPASS(mvp->v_type == VMARKER, mvp);
1505 	mtx_lock(&vnode_list_mtx);
1506 	vnlru_free_impl(count, mnt_op, mvp, true);
1507 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1508 }
1509 
1510 struct vnode *
1511 vnlru_alloc_marker(void)
1512 {
1513 	struct vnode *mvp;
1514 
1515 	mvp = vn_alloc_marker(NULL);
1516 	mtx_lock(&vnode_list_mtx);
1517 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, mvp, v_vnodelist);
1518 	mtx_unlock(&vnode_list_mtx);
1519 	return (mvp);
1520 }
1521 
1522 void
1523 vnlru_free_marker(struct vnode *mvp)
1524 {
1525 	mtx_lock(&vnode_list_mtx);
1526 	TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1527 	mtx_unlock(&vnode_list_mtx);
1528 	vn_free_marker(mvp);
1529 }
1530 
1531 static void
1532 vnlru_recalc(void)
1533 {
1534 
1535 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1536 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1537 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1538 	vlowat = vhiwat / 2;
1539 }
1540 
1541 /*
1542  * Attempt to recycle vnodes in a context that is always safe to block.
1543  * Calling vlrurecycle() from the bowels of filesystem code has some
1544  * interesting deadlock problems.
1545  */
1546 static struct proc *vnlruproc;
1547 static int vnlruproc_sig;
1548 static u_long vnlruproc_kicks;
1549 
1550 SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, kicks, CTLFLAG_RD, &vnlruproc_kicks, 0,
1551     "Number of times vnlru awakened due to vnode shortage");
1552 
1553 #define VNLRU_COUNT_SLOP 100
1554 
1555 /*
1556  * The main freevnodes counter is only updated when a counter local to CPU
1557  * diverges from 0 by more than VNLRU_FREEVNODES_SLOP. CPUs are conditionally
1558  * walked to compute a more accurate total.
1559  *
1560  * Note: the actual value at any given moment can still exceed slop, but it
1561  * should not be by significant margin in practice.
1562  */
1563 #define VNLRU_FREEVNODES_SLOP 126
1564 
1565 static void __noinline
1566 vfs_freevnodes_rollup(int8_t *lfreevnodes)
1567 {
1568 
1569 	atomic_add_long(&freevnodes, *lfreevnodes);
1570 	*lfreevnodes = 0;
1571 	critical_exit();
1572 }
1573 
1574 static __inline void
1575 vfs_freevnodes_inc(void)
1576 {
1577 	int8_t *lfreevnodes;
1578 
1579 	critical_enter();
1580 	lfreevnodes = PCPU_PTR(vfs_freevnodes);
1581 	(*lfreevnodes)++;
1582 	if (__predict_false(*lfreevnodes == VNLRU_FREEVNODES_SLOP))
1583 		vfs_freevnodes_rollup(lfreevnodes);
1584 	else
1585 		critical_exit();
1586 }
1587 
1588 static __inline void
1589 vfs_freevnodes_dec(void)
1590 {
1591 	int8_t *lfreevnodes;
1592 
1593 	critical_enter();
1594 	lfreevnodes = PCPU_PTR(vfs_freevnodes);
1595 	(*lfreevnodes)--;
1596 	if (__predict_false(*lfreevnodes == -VNLRU_FREEVNODES_SLOP))
1597 		vfs_freevnodes_rollup(lfreevnodes);
1598 	else
1599 		critical_exit();
1600 }
1601 
1602 static u_long
1603 vnlru_read_freevnodes(void)
1604 {
1605 	long slop, rfreevnodes, rfreevnodes_old;
1606 	int cpu;
1607 
1608 	rfreevnodes = atomic_load_long(&freevnodes);
1609 	rfreevnodes_old = atomic_load_long(&freevnodes_old);
1610 
1611 	if (rfreevnodes > rfreevnodes_old)
1612 		slop = rfreevnodes - rfreevnodes_old;
1613 	else
1614 		slop = rfreevnodes_old - rfreevnodes;
1615 	if (slop < VNLRU_FREEVNODES_SLOP)
1616 		return (rfreevnodes >= 0 ? rfreevnodes : 0);
1617 	CPU_FOREACH(cpu) {
1618 		rfreevnodes += cpuid_to_pcpu[cpu]->pc_vfs_freevnodes;
1619 	}
1620 	atomic_store_long(&freevnodes_old, rfreevnodes);
1621 	return (freevnodes_old >= 0 ? freevnodes_old : 0);
1622 }
1623 
1624 static bool
1625 vnlru_under(u_long rnumvnodes, u_long limit)
1626 {
1627 	u_long rfreevnodes, space;
1628 
1629 	if (__predict_false(rnumvnodes > desiredvnodes))
1630 		return (true);
1631 
1632 	space = desiredvnodes - rnumvnodes;
1633 	if (space < limit) {
1634 		rfreevnodes = vnlru_read_freevnodes();
1635 		if (rfreevnodes > wantfreevnodes)
1636 			space += rfreevnodes - wantfreevnodes;
1637 	}
1638 	return (space < limit);
1639 }
1640 
1641 static void
1642 vnlru_kick_locked(void)
1643 {
1644 
1645 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1646 	if (vnlruproc_sig == 0) {
1647 		vnlruproc_sig = 1;
1648 		vnlruproc_kicks++;
1649 		wakeup(vnlruproc);
1650 	}
1651 }
1652 
1653 static void
1654 vnlru_kick_cond(void)
1655 {
1656 
1657 	if (vnlru_read_freevnodes() > wantfreevnodes)
1658 		return;
1659 
1660 	if (vnlruproc_sig)
1661 		return;
1662 	mtx_lock(&vnode_list_mtx);
1663 	vnlru_kick_locked();
1664 	mtx_unlock(&vnode_list_mtx);
1665 }
1666 
1667 static void
1668 vnlru_proc_sleep(void)
1669 {
1670 
1671 	if (vnlruproc_sig) {
1672 		vnlruproc_sig = 0;
1673 		wakeup(&vnlruproc_sig);
1674 	}
1675 	msleep(vnlruproc, &vnode_list_mtx, PVFS|PDROP, "vlruwt", hz);
1676 }
1677 
1678 /*
1679  * A lighter version of the machinery below.
1680  *
1681  * Tries to reach goals only by recycling free vnodes and does not invoke
1682  * uma_reclaim(UMA_RECLAIM_DRAIN).
1683  *
1684  * This works around pathological behavior in vnlru in presence of tons of free
1685  * vnodes, but without having to rewrite the machinery at this time. Said
1686  * behavior boils down to continuously trying to reclaim all kinds of vnodes
1687  * (cycling through all levels of "force") when the count is transiently above
1688  * limit. This happens a lot when all vnodes are used up and vn_alloc
1689  * speculatively increments the counter.
1690  *
1691  * Sample testcase: vnode limit 8388608, 20 separate directory trees each with
1692  * 1 million files in total and 20 find(1) processes stating them in parallel
1693  * (one per each tree).
1694  *
1695  * On a kernel with only stock machinery this needs anywhere between 60 and 120
1696  * seconds to execute (time varies *wildly* between runs). With the workaround
1697  * it consistently stays around 20 seconds [it got further down with later
1698  * changes].
1699  *
1700  * That is to say the entire thing needs a fundamental redesign (most notably
1701  * to accommodate faster recycling), the above only tries to get it ouf the way.
1702  *
1703  * Return values are:
1704  * -1 -- fallback to regular vnlru loop
1705  *  0 -- do nothing, go to sleep
1706  * >0 -- recycle this many vnodes
1707  */
1708 static long
1709 vnlru_proc_light_pick(void)
1710 {
1711 	u_long rnumvnodes, rfreevnodes;
1712 
1713 	if (vstir || vnlruproc_sig == 1)
1714 		return (-1);
1715 
1716 	rnumvnodes = atomic_load_long(&numvnodes);
1717 	rfreevnodes = vnlru_read_freevnodes();
1718 
1719 	/*
1720 	 * vnode limit might have changed and now we may be at a significant
1721 	 * excess. Bail if we can't sort it out with free vnodes.
1722 	 *
1723 	 * Due to atomic updates the count can legitimately go above
1724 	 * the limit for a short period, don't bother doing anything in
1725 	 * that case.
1726 	 */
1727 	if (rnumvnodes > desiredvnodes + VNLRU_COUNT_SLOP + 10) {
1728 		if (rnumvnodes - rfreevnodes >= desiredvnodes ||
1729 		    rfreevnodes <= wantfreevnodes) {
1730 			return (-1);
1731 		}
1732 
1733 		return (rnumvnodes - desiredvnodes);
1734 	}
1735 
1736 	/*
1737 	 * Don't try to reach wantfreevnodes target if there are too few vnodes
1738 	 * to begin with.
1739 	 */
1740 	if (rnumvnodes < wantfreevnodes) {
1741 		return (0);
1742 	}
1743 
1744 	if (rfreevnodes < wantfreevnodes) {
1745 		return (-1);
1746 	}
1747 
1748 	return (0);
1749 }
1750 
1751 static bool
1752 vnlru_proc_light(void)
1753 {
1754 	long freecount;
1755 
1756 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1757 
1758 	freecount = vnlru_proc_light_pick();
1759 	if (freecount == -1)
1760 		return (false);
1761 
1762 	if (freecount != 0) {
1763 		vnlru_free_vnlru(freecount);
1764 	}
1765 
1766 	mtx_lock(&vnode_list_mtx);
1767 	vnlru_proc_sleep();
1768 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1769 	return (true);
1770 }
1771 
1772 static u_long uma_reclaim_calls;
1773 SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, uma_reclaim_calls, CTLFLAG_RD | CTLFLAG_STATS,
1774     &uma_reclaim_calls, 0, "Number of calls to uma_reclaim");
1775 
1776 static void
1777 vnlru_proc(void)
1778 {
1779 	u_long rnumvnodes, rfreevnodes, target;
1780 	unsigned long onumvnodes;
1781 	int done, force, trigger, usevnodes;
1782 	bool reclaim_nc_src, want_reread;
1783 
1784 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1785 	    SHUTDOWN_PRI_FIRST);
1786 
1787 	force = 0;
1788 	want_reread = false;
1789 	for (;;) {
1790 		kproc_suspend_check(vnlruproc);
1791 
1792 		if (force == 0 && vnlru_proc_light())
1793 			continue;
1794 
1795 		mtx_lock(&vnode_list_mtx);
1796 		rnumvnodes = atomic_load_long(&numvnodes);
1797 
1798 		if (want_reread) {
1799 			force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1800 			want_reread = false;
1801 		}
1802 
1803 		/*
1804 		 * If numvnodes is too large (due to desiredvnodes being
1805 		 * adjusted using its sysctl, or emergency growth), first
1806 		 * try to reduce it by discarding free vnodes.
1807 		 */
1808 		if (rnumvnodes > desiredvnodes + 10) {
1809 			vnlru_free_locked_vnlru(rnumvnodes - desiredvnodes);
1810 			mtx_lock(&vnode_list_mtx);
1811 			rnumvnodes = atomic_load_long(&numvnodes);
1812 		}
1813 		/*
1814 		 * Sleep if the vnode cache is in a good state.  This is
1815 		 * when it is not over-full and has space for about a 4%
1816 		 * or 9% expansion (by growing its size or inexcessively
1817 		 * reducing free vnode count).  Otherwise, try to reclaim
1818 		 * space for a 10% expansion.
1819 		 */
1820 		if (vstir && force == 0) {
1821 			force = 1;
1822 			vstir = false;
1823 		}
1824 		if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1825 			vnlru_proc_sleep();
1826 			continue;
1827 		}
1828 		rfreevnodes = vnlru_read_freevnodes();
1829 
1830 		onumvnodes = rnumvnodes;
1831 		/*
1832 		 * Calculate parameters for recycling.  These are the same
1833 		 * throughout the loop to give some semblance of fairness.
1834 		 * The trigger point is to avoid recycling vnodes with lots
1835 		 * of resident pages.  We aren't trying to free memory; we
1836 		 * are trying to recycle or at least free vnodes.
1837 		 */
1838 		if (rnumvnodes <= desiredvnodes)
1839 			usevnodes = rnumvnodes - rfreevnodes;
1840 		else
1841 			usevnodes = rnumvnodes;
1842 		if (usevnodes <= 0)
1843 			usevnodes = 1;
1844 		/*
1845 		 * The trigger value is chosen to give a conservatively
1846 		 * large value to ensure that it alone doesn't prevent
1847 		 * making progress.  The value can easily be so large that
1848 		 * it is effectively infinite in some congested and
1849 		 * misconfigured cases, and this is necessary.  Normally
1850 		 * it is about 8 to 100 (pages), which is quite large.
1851 		 */
1852 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1853 		if (force < 2)
1854 			trigger = vsmalltrigger;
1855 		reclaim_nc_src = force >= 3;
1856 		target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1857 		target = target / 10 + 1;
1858 		done = vlrureclaim(reclaim_nc_src, trigger, target);
1859 		mtx_unlock(&vnode_list_mtx);
1860 		/*
1861 		 * Total number of vnodes can transiently go slightly above the
1862 		 * limit (see vn_alloc_hard), no need to call uma_reclaim if
1863 		 * this happens.
1864 		 */
1865 		if (onumvnodes + VNLRU_COUNT_SLOP + 1000 > desiredvnodes &&
1866 		    numvnodes <= desiredvnodes) {
1867 			uma_reclaim_calls++;
1868 			uma_reclaim(UMA_RECLAIM_DRAIN);
1869 		}
1870 		if (done == 0) {
1871 			if (force == 0 || force == 1) {
1872 				force = 2;
1873 				continue;
1874 			}
1875 			if (force == 2) {
1876 				force = 3;
1877 				continue;
1878 			}
1879 			want_reread = true;
1880 			force = 0;
1881 			vnlru_nowhere++;
1882 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1883 		} else {
1884 			want_reread = true;
1885 			kern_yield(PRI_USER);
1886 		}
1887 	}
1888 }
1889 
1890 static struct kproc_desc vnlru_kp = {
1891 	"vnlru",
1892 	vnlru_proc,
1893 	&vnlruproc
1894 };
1895 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1896     &vnlru_kp);
1897 
1898 /*
1899  * Routines having to do with the management of the vnode table.
1900  */
1901 
1902 /*
1903  * Try to recycle a freed vnode.
1904  */
1905 static int
1906 vtryrecycle(struct vnode *vp, bool isvnlru)
1907 {
1908 	struct mount *vnmp;
1909 
1910 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1911 	VNPASS(vp->v_holdcnt > 0, vp);
1912 	/*
1913 	 * This vnode may found and locked via some other list, if so we
1914 	 * can't recycle it yet.
1915 	 */
1916 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1917 		CTR2(KTR_VFS,
1918 		    "%s: impossible to recycle, vp %p lock is already held",
1919 		    __func__, vp);
1920 		vdrop_recycle(vp);
1921 		return (EWOULDBLOCK);
1922 	}
1923 	/*
1924 	 * Don't recycle if its filesystem is being suspended.
1925 	 */
1926 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1927 		VOP_UNLOCK(vp);
1928 		CTR2(KTR_VFS,
1929 		    "%s: impossible to recycle, cannot start the write for %p",
1930 		    __func__, vp);
1931 		vdrop_recycle(vp);
1932 		return (EBUSY);
1933 	}
1934 	/*
1935 	 * If we got this far, we need to acquire the interlock and see if
1936 	 * anyone picked up this vnode from another list.  If not, we will
1937 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1938 	 * will skip over it.
1939 	 */
1940 	VI_LOCK(vp);
1941 	if (vp->v_usecount) {
1942 		VOP_UNLOCK(vp);
1943 		vdropl_recycle(vp);
1944 		vn_finished_write(vnmp);
1945 		CTR2(KTR_VFS,
1946 		    "%s: impossible to recycle, %p is already referenced",
1947 		    __func__, vp);
1948 		return (EBUSY);
1949 	}
1950 	if (!VN_IS_DOOMED(vp)) {
1951 		if (isvnlru)
1952 			recycles_free_count++;
1953 		else
1954 			counter_u64_add(direct_recycles_free_count, 1);
1955 		vgonel(vp);
1956 	}
1957 	VOP_UNLOCK(vp);
1958 	vdropl_recycle(vp);
1959 	vn_finished_write(vnmp);
1960 	return (0);
1961 }
1962 
1963 /*
1964  * Allocate a new vnode.
1965  *
1966  * The operation never returns an error. Returning an error was disabled
1967  * in r145385 (dated 2005) with the following comment:
1968  *
1969  * XXX Not all VFS_VGET/ffs_vget callers check returns.
1970  *
1971  * Given the age of this commit (almost 15 years at the time of writing this
1972  * comment) restoring the ability to fail requires a significant audit of
1973  * all codepaths.
1974  *
1975  * The routine can try to free a vnode or stall for up to 1 second waiting for
1976  * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1977  */
1978 static u_long vn_alloc_cyclecount;
1979 static u_long vn_alloc_sleeps;
1980 
1981 SYSCTL_ULONG(_vfs_vnode_stats, OID_AUTO, alloc_sleeps, CTLFLAG_RD, &vn_alloc_sleeps, 0,
1982     "Number of times vnode allocation blocked waiting on vnlru");
1983 
1984 static struct vnode * __noinline
1985 vn_alloc_hard(struct mount *mp, u_long rnumvnodes, bool bumped)
1986 {
1987 	u_long rfreevnodes;
1988 
1989 	if (bumped) {
1990 		if (rnumvnodes > desiredvnodes + VNLRU_COUNT_SLOP) {
1991 			atomic_subtract_long(&numvnodes, 1);
1992 			bumped = false;
1993 		}
1994 	}
1995 
1996 	mtx_lock(&vnode_list_mtx);
1997 
1998 	/*
1999 	 * Reload 'numvnodes', as since we acquired the lock, it may have
2000 	 * changed significantly if we waited, and 'rnumvnodes' above was only
2001 	 * actually passed if 'bumped' is true (else it is 0).
2002 	 */
2003 	rnumvnodes = atomic_load_long(&numvnodes);
2004 	if (rnumvnodes + !bumped < desiredvnodes) {
2005 		vn_alloc_cyclecount = 0;
2006 		mtx_unlock(&vnode_list_mtx);
2007 		goto alloc;
2008 	}
2009 
2010 	rfreevnodes = vnlru_read_freevnodes();
2011 	if (vn_alloc_cyclecount++ >= rfreevnodes) {
2012 		vn_alloc_cyclecount = 0;
2013 		vstir = true;
2014 	}
2015 
2016 	/*
2017 	 * Grow the vnode cache if it will not be above its target max after
2018 	 * growing.  Otherwise, if there is at least one free vnode, try to
2019 	 * reclaim 1 item from it before growing the cache (possibly above its
2020 	 * target max if the reclamation failed or is delayed).
2021 	 */
2022 	if (vnlru_free_locked_direct(1) > 0)
2023 		goto alloc;
2024 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
2025 	if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
2026 		/*
2027 		 * Wait for space for a new vnode.
2028 		 */
2029 		if (bumped) {
2030 			atomic_subtract_long(&numvnodes, 1);
2031 			bumped = false;
2032 		}
2033 		mtx_lock(&vnode_list_mtx);
2034 		vnlru_kick_locked();
2035 		vn_alloc_sleeps++;
2036 		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
2037 		if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
2038 		    vnlru_read_freevnodes() > 1)
2039 			vnlru_free_locked_direct(1);
2040 		else
2041 			mtx_unlock(&vnode_list_mtx);
2042 	}
2043 alloc:
2044 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
2045 	if (!bumped)
2046 		atomic_add_long(&numvnodes, 1);
2047 	vnlru_kick_cond();
2048 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
2049 }
2050 
2051 static struct vnode *
2052 vn_alloc(struct mount *mp)
2053 {
2054 	u_long rnumvnodes;
2055 
2056 	if (__predict_false(vn_alloc_cyclecount != 0))
2057 		return (vn_alloc_hard(mp, 0, false));
2058 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
2059 	if (__predict_false(vnlru_under(rnumvnodes, vlowat))) {
2060 		return (vn_alloc_hard(mp, rnumvnodes, true));
2061 	}
2062 
2063 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
2064 }
2065 
2066 static void
2067 vn_free(struct vnode *vp)
2068 {
2069 
2070 	atomic_subtract_long(&numvnodes, 1);
2071 	uma_zfree_smr(vnode_zone, vp);
2072 }
2073 
2074 /*
2075  * Allocate a new vnode.
2076  */
2077 int
2078 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
2079     struct vnode **vpp)
2080 {
2081 	struct vnode *vp;
2082 	struct thread *td;
2083 	struct lock_object *lo;
2084 
2085 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
2086 
2087 	KASSERT(vops->registered,
2088 	    ("%s: not registered vector op %p\n", __func__, vops));
2089 	cache_validate_vop_vector(mp, vops);
2090 
2091 	td = curthread;
2092 	if (td->td_vp_reserved != NULL) {
2093 		vp = td->td_vp_reserved;
2094 		td->td_vp_reserved = NULL;
2095 	} else {
2096 		vp = vn_alloc(mp);
2097 	}
2098 	counter_u64_add(vnodes_created, 1);
2099 
2100 	vn_set_state(vp, VSTATE_UNINITIALIZED);
2101 
2102 	/*
2103 	 * Locks are given the generic name "vnode" when created.
2104 	 * Follow the historic practice of using the filesystem
2105 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
2106 	 *
2107 	 * Locks live in a witness group keyed on their name. Thus,
2108 	 * when a lock is renamed, it must also move from the witness
2109 	 * group of its old name to the witness group of its new name.
2110 	 *
2111 	 * The change only needs to be made when the vnode moves
2112 	 * from one filesystem type to another. We ensure that each
2113 	 * filesystem use a single static name pointer for its tag so
2114 	 * that we can compare pointers rather than doing a strcmp().
2115 	 */
2116 	lo = &vp->v_vnlock->lock_object;
2117 #ifdef WITNESS
2118 	if (lo->lo_name != tag) {
2119 #endif
2120 		lo->lo_name = tag;
2121 #ifdef WITNESS
2122 		WITNESS_DESTROY(lo);
2123 		WITNESS_INIT(lo, tag);
2124 	}
2125 #endif
2126 	/*
2127 	 * By default, don't allow shared locks unless filesystems opt-in.
2128 	 */
2129 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
2130 	/*
2131 	 * Finalize various vnode identity bits.
2132 	 */
2133 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
2134 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
2135 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
2136 	vp->v_type = VNON;
2137 	vp->v_op = vops;
2138 	vp->v_irflag = 0;
2139 	v_init_counters(vp);
2140 	vn_seqc_init(vp);
2141 	vp->v_bufobj.bo_ops = &buf_ops_bio;
2142 #ifdef DIAGNOSTIC
2143 	if (mp == NULL && vops != &dead_vnodeops)
2144 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
2145 #endif
2146 #ifdef MAC
2147 	mac_vnode_init(vp);
2148 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
2149 		mac_vnode_associate_singlelabel(mp, vp);
2150 #endif
2151 	if (mp != NULL) {
2152 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
2153 	}
2154 
2155 	/*
2156 	 * For the filesystems which do not use vfs_hash_insert(),
2157 	 * still initialize v_hash to have vfs_hash_index() useful.
2158 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
2159 	 * its own hashing.
2160 	 */
2161 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
2162 
2163 	*vpp = vp;
2164 	return (0);
2165 }
2166 
2167 void
2168 getnewvnode_reserve(void)
2169 {
2170 	struct thread *td;
2171 
2172 	td = curthread;
2173 	MPASS(td->td_vp_reserved == NULL);
2174 	td->td_vp_reserved = vn_alloc(NULL);
2175 }
2176 
2177 void
2178 getnewvnode_drop_reserve(void)
2179 {
2180 	struct thread *td;
2181 
2182 	td = curthread;
2183 	if (td->td_vp_reserved != NULL) {
2184 		vn_free(td->td_vp_reserved);
2185 		td->td_vp_reserved = NULL;
2186 	}
2187 }
2188 
2189 static void __noinline
2190 freevnode(struct vnode *vp)
2191 {
2192 	struct bufobj *bo;
2193 
2194 	/*
2195 	 * The vnode has been marked for destruction, so free it.
2196 	 *
2197 	 * The vnode will be returned to the zone where it will
2198 	 * normally remain until it is needed for another vnode. We
2199 	 * need to cleanup (or verify that the cleanup has already
2200 	 * been done) any residual data left from its current use
2201 	 * so as not to contaminate the freshly allocated vnode.
2202 	 */
2203 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
2204 	/*
2205 	 * Paired with vgone.
2206 	 */
2207 	vn_seqc_write_end_free(vp);
2208 
2209 	bo = &vp->v_bufobj;
2210 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
2211 	VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp);
2212 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
2213 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
2214 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
2215 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
2216 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
2217 	    ("clean blk trie not empty"));
2218 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
2219 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
2220 	    ("dirty blk trie not empty"));
2221 	VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp,
2222 	    ("Leaked inactivation"));
2223 	VI_UNLOCK(vp);
2224 	cache_assert_no_entries(vp);
2225 
2226 #ifdef MAC
2227 	mac_vnode_destroy(vp);
2228 #endif
2229 	if (vp->v_pollinfo != NULL) {
2230 		/*
2231 		 * Use LK_NOWAIT to shut up witness about the lock. We may get
2232 		 * here while having another vnode locked when trying to
2233 		 * satisfy a lookup and needing to recycle.
2234 		 */
2235 		VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT);
2236 		destroy_vpollinfo(vp->v_pollinfo);
2237 		VOP_UNLOCK(vp);
2238 		vp->v_pollinfo = NULL;
2239 	}
2240 	vp->v_mountedhere = NULL;
2241 	vp->v_unpcb = NULL;
2242 	vp->v_rdev = NULL;
2243 	vp->v_fifoinfo = NULL;
2244 	vp->v_iflag = 0;
2245 	vp->v_vflag = 0;
2246 	bo->bo_flag = 0;
2247 	vn_free(vp);
2248 }
2249 
2250 /*
2251  * Delete from old mount point vnode list, if on one.
2252  */
2253 static void
2254 delmntque(struct vnode *vp)
2255 {
2256 	struct mount *mp;
2257 
2258 	VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
2259 
2260 	mp = vp->v_mount;
2261 	MNT_ILOCK(mp);
2262 	VI_LOCK(vp);
2263 	vp->v_mount = NULL;
2264 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
2265 		("bad mount point vnode list size"));
2266 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2267 	mp->mnt_nvnodelistsize--;
2268 	MNT_REL(mp);
2269 	MNT_IUNLOCK(mp);
2270 	/*
2271 	 * The caller expects the interlock to be still held.
2272 	 */
2273 	ASSERT_VI_LOCKED(vp, __func__);
2274 }
2275 
2276 static int
2277 insmntque1_int(struct vnode *vp, struct mount *mp, bool dtr)
2278 {
2279 
2280 	KASSERT(vp->v_mount == NULL,
2281 		("insmntque: vnode already on per mount vnode list"));
2282 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
2283 	if ((mp->mnt_kern_flag & MNTK_UNLOCKED_INSMNTQUE) == 0) {
2284 		ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
2285 	} else {
2286 		KASSERT(!dtr,
2287 		    ("%s: can't have MNTK_UNLOCKED_INSMNTQUE and cleanup",
2288 		    __func__));
2289 	}
2290 
2291 	/*
2292 	 * We acquire the vnode interlock early to ensure that the
2293 	 * vnode cannot be recycled by another process releasing a
2294 	 * holdcnt on it before we get it on both the vnode list
2295 	 * and the active vnode list. The mount mutex protects only
2296 	 * manipulation of the vnode list and the vnode freelist
2297 	 * mutex protects only manipulation of the active vnode list.
2298 	 * Hence the need to hold the vnode interlock throughout.
2299 	 */
2300 	MNT_ILOCK(mp);
2301 	VI_LOCK(vp);
2302 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
2303 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
2304 	    mp->mnt_nvnodelistsize == 0)) &&
2305 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
2306 		VI_UNLOCK(vp);
2307 		MNT_IUNLOCK(mp);
2308 		if (dtr) {
2309 			vp->v_data = NULL;
2310 			vp->v_op = &dead_vnodeops;
2311 			vgone(vp);
2312 			vput(vp);
2313 		}
2314 		return (EBUSY);
2315 	}
2316 	vp->v_mount = mp;
2317 	MNT_REF(mp);
2318 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2319 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
2320 		("neg mount point vnode list size"));
2321 	mp->mnt_nvnodelistsize++;
2322 	VI_UNLOCK(vp);
2323 	MNT_IUNLOCK(mp);
2324 	return (0);
2325 }
2326 
2327 /*
2328  * Insert into list of vnodes for the new mount point, if available.
2329  * insmntque() reclaims the vnode on insertion failure, insmntque1()
2330  * leaves handling of the vnode to the caller.
2331  */
2332 int
2333 insmntque(struct vnode *vp, struct mount *mp)
2334 {
2335 	return (insmntque1_int(vp, mp, true));
2336 }
2337 
2338 int
2339 insmntque1(struct vnode *vp, struct mount *mp)
2340 {
2341 	return (insmntque1_int(vp, mp, false));
2342 }
2343 
2344 /*
2345  * Flush out and invalidate all buffers associated with a bufobj
2346  * Called with the underlying object locked.
2347  */
2348 int
2349 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
2350 {
2351 	int error;
2352 
2353 	BO_LOCK(bo);
2354 	if (flags & V_SAVE) {
2355 		error = bufobj_wwait(bo, slpflag, slptimeo);
2356 		if (error) {
2357 			BO_UNLOCK(bo);
2358 			return (error);
2359 		}
2360 		if (bo->bo_dirty.bv_cnt > 0) {
2361 			BO_UNLOCK(bo);
2362 			do {
2363 				error = BO_SYNC(bo, MNT_WAIT);
2364 			} while (error == ERELOOKUP);
2365 			if (error != 0)
2366 				return (error);
2367 			BO_LOCK(bo);
2368 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) {
2369 				BO_UNLOCK(bo);
2370 				return (EBUSY);
2371 			}
2372 		}
2373 	}
2374 	/*
2375 	 * If you alter this loop please notice that interlock is dropped and
2376 	 * reacquired in flushbuflist.  Special care is needed to ensure that
2377 	 * no race conditions occur from this.
2378 	 */
2379 	do {
2380 		error = flushbuflist(&bo->bo_clean,
2381 		    flags, bo, slpflag, slptimeo);
2382 		if (error == 0 && !(flags & V_CLEANONLY))
2383 			error = flushbuflist(&bo->bo_dirty,
2384 			    flags, bo, slpflag, slptimeo);
2385 		if (error != 0 && error != EAGAIN) {
2386 			BO_UNLOCK(bo);
2387 			return (error);
2388 		}
2389 	} while (error != 0);
2390 
2391 	/*
2392 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
2393 	 * have write I/O in-progress but if there is a VM object then the
2394 	 * VM object can also have read-I/O in-progress.
2395 	 */
2396 	do {
2397 		bufobj_wwait(bo, 0, 0);
2398 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
2399 			BO_UNLOCK(bo);
2400 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
2401 			BO_LOCK(bo);
2402 		}
2403 	} while (bo->bo_numoutput > 0);
2404 	BO_UNLOCK(bo);
2405 
2406 	/*
2407 	 * Destroy the copy in the VM cache, too.
2408 	 */
2409 	if (bo->bo_object != NULL &&
2410 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
2411 		VM_OBJECT_WLOCK(bo->bo_object);
2412 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
2413 		    OBJPR_CLEANONLY : 0);
2414 		VM_OBJECT_WUNLOCK(bo->bo_object);
2415 	}
2416 
2417 #ifdef INVARIANTS
2418 	BO_LOCK(bo);
2419 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
2420 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
2421 	    bo->bo_clean.bv_cnt > 0))
2422 		panic("vinvalbuf: flush failed");
2423 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
2424 	    bo->bo_dirty.bv_cnt > 0)
2425 		panic("vinvalbuf: flush dirty failed");
2426 	BO_UNLOCK(bo);
2427 #endif
2428 	return (0);
2429 }
2430 
2431 /*
2432  * Flush out and invalidate all buffers associated with a vnode.
2433  * Called with the underlying object locked.
2434  */
2435 int
2436 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
2437 {
2438 
2439 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2440 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
2441 	if (vp->v_object != NULL && vp->v_object->handle != vp)
2442 		return (0);
2443 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
2444 }
2445 
2446 /*
2447  * Flush out buffers on the specified list.
2448  *
2449  */
2450 static int
2451 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2452     int slptimeo)
2453 {
2454 	struct buf *bp, *nbp;
2455 	int retval, error;
2456 	daddr_t lblkno;
2457 	b_xflags_t xflags;
2458 
2459 	ASSERT_BO_WLOCKED(bo);
2460 
2461 	retval = 0;
2462 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2463 		/*
2464 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2465 		 * do not skip any buffers. If we are flushing only V_NORMAL
2466 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2467 		 * flushing only V_ALT buffers then skip buffers not marked
2468 		 * as BX_ALTDATA.
2469 		 */
2470 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2471 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2472 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2473 			continue;
2474 		}
2475 		if (nbp != NULL) {
2476 			lblkno = nbp->b_lblkno;
2477 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2478 		}
2479 		retval = EAGAIN;
2480 		error = BUF_TIMELOCK(bp,
2481 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2482 		    "flushbuf", slpflag, slptimeo);
2483 		if (error) {
2484 			BO_LOCK(bo);
2485 			return (error != ENOLCK ? error : EAGAIN);
2486 		}
2487 		KASSERT(bp->b_bufobj == bo,
2488 		    ("bp %p wrong b_bufobj %p should be %p",
2489 		    bp, bp->b_bufobj, bo));
2490 		/*
2491 		 * XXX Since there are no node locks for NFS, I
2492 		 * believe there is a slight chance that a delayed
2493 		 * write will occur while sleeping just above, so
2494 		 * check for it.
2495 		 */
2496 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2497 		    (flags & V_SAVE)) {
2498 			bremfree(bp);
2499 			bp->b_flags |= B_ASYNC;
2500 			bwrite(bp);
2501 			BO_LOCK(bo);
2502 			return (EAGAIN);	/* XXX: why not loop ? */
2503 		}
2504 		bremfree(bp);
2505 		bp->b_flags |= (B_INVAL | B_RELBUF);
2506 		bp->b_flags &= ~B_ASYNC;
2507 		brelse(bp);
2508 		BO_LOCK(bo);
2509 		if (nbp == NULL)
2510 			break;
2511 		nbp = gbincore(bo, lblkno);
2512 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2513 		    != xflags)
2514 			break;			/* nbp invalid */
2515 	}
2516 	return (retval);
2517 }
2518 
2519 int
2520 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2521 {
2522 	struct buf *bp;
2523 	int error;
2524 	daddr_t lblkno;
2525 
2526 	ASSERT_BO_LOCKED(bo);
2527 
2528 	for (lblkno = startn;;) {
2529 again:
2530 		bp = buf_lookup_ge(bufv, lblkno);
2531 		if (bp == NULL || bp->b_lblkno >= endn)
2532 			break;
2533 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2534 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2535 		if (error != 0) {
2536 			BO_RLOCK(bo);
2537 			if (error == ENOLCK)
2538 				goto again;
2539 			return (error);
2540 		}
2541 		KASSERT(bp->b_bufobj == bo,
2542 		    ("bp %p wrong b_bufobj %p should be %p",
2543 		    bp, bp->b_bufobj, bo));
2544 		lblkno = bp->b_lblkno + 1;
2545 		if ((bp->b_flags & B_MANAGED) == 0)
2546 			bremfree(bp);
2547 		bp->b_flags |= B_RELBUF;
2548 		/*
2549 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2550 		 * pages backing each buffer in the range are unlikely to be
2551 		 * reused.  Dirty buffers will have the hint applied once
2552 		 * they've been written.
2553 		 */
2554 		if ((bp->b_flags & B_VMIO) != 0)
2555 			bp->b_flags |= B_NOREUSE;
2556 		brelse(bp);
2557 		BO_RLOCK(bo);
2558 	}
2559 	return (0);
2560 }
2561 
2562 /*
2563  * Truncate a file's buffer and pages to a specified length.  This
2564  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2565  * sync activity.
2566  */
2567 int
2568 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2569 {
2570 	struct buf *bp, *nbp;
2571 	struct bufobj *bo;
2572 	daddr_t startlbn;
2573 
2574 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2575 	    vp, blksize, (uintmax_t)length);
2576 
2577 	/*
2578 	 * Round up to the *next* lbn.
2579 	 */
2580 	startlbn = howmany(length, blksize);
2581 
2582 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2583 
2584 	bo = &vp->v_bufobj;
2585 restart_unlocked:
2586 	BO_LOCK(bo);
2587 
2588 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2589 		;
2590 
2591 	if (length > 0) {
2592 		/*
2593 		 * Write out vnode metadata, e.g. indirect blocks.
2594 		 */
2595 restartsync:
2596 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2597 			if (bp->b_lblkno >= 0)
2598 				continue;
2599 			/*
2600 			 * Since we hold the vnode lock this should only
2601 			 * fail if we're racing with the buf daemon.
2602 			 */
2603 			if (BUF_LOCK(bp,
2604 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2605 			    BO_LOCKPTR(bo)) == ENOLCK)
2606 				goto restart_unlocked;
2607 
2608 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2609 			    ("buf(%p) on dirty queue without DELWRI", bp));
2610 
2611 			bremfree(bp);
2612 			bawrite(bp);
2613 			BO_LOCK(bo);
2614 			goto restartsync;
2615 		}
2616 	}
2617 
2618 	bufobj_wwait(bo, 0, 0);
2619 	BO_UNLOCK(bo);
2620 	vnode_pager_setsize(vp, length);
2621 
2622 	return (0);
2623 }
2624 
2625 /*
2626  * Invalidate the cached pages of a file's buffer within the range of block
2627  * numbers [startlbn, endlbn).
2628  */
2629 void
2630 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2631     int blksize)
2632 {
2633 	struct bufobj *bo;
2634 	off_t start, end;
2635 
2636 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2637 
2638 	start = blksize * startlbn;
2639 	end = blksize * endlbn;
2640 
2641 	bo = &vp->v_bufobj;
2642 	BO_LOCK(bo);
2643 	MPASS(blksize == bo->bo_bsize);
2644 
2645 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2646 		;
2647 
2648 	BO_UNLOCK(bo);
2649 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2650 }
2651 
2652 static int
2653 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2654     daddr_t startlbn, daddr_t endlbn)
2655 {
2656 	struct bufv *bv;
2657 	struct buf *bp, *nbp;
2658 	uint8_t anyfreed;
2659 	bool clean;
2660 
2661 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2662 	ASSERT_BO_LOCKED(bo);
2663 
2664 	anyfreed = 1;
2665 	clean = true;
2666 	do {
2667 		bv = clean ? &bo->bo_clean : &bo->bo_dirty;
2668 		bp = buf_lookup_ge(bv, startlbn);
2669 		if (bp == NULL)
2670 			continue;
2671 		TAILQ_FOREACH_FROM_SAFE(bp, &bv->bv_hd, b_bobufs, nbp) {
2672 			if (bp->b_lblkno >= endlbn)
2673 				break;
2674 			if (BUF_LOCK(bp,
2675 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2676 			    BO_LOCKPTR(bo)) == ENOLCK) {
2677 				BO_LOCK(bo);
2678 				return (EAGAIN);
2679 			}
2680 
2681 			bremfree(bp);
2682 			bp->b_flags |= B_INVAL | B_RELBUF;
2683 			bp->b_flags &= ~B_ASYNC;
2684 			brelse(bp);
2685 			anyfreed = 2;
2686 
2687 			BO_LOCK(bo);
2688 			if (nbp != NULL &&
2689 			    (((nbp->b_xflags &
2690 			    (clean ? BX_VNCLEAN : BX_VNDIRTY)) == 0) ||
2691 			    nbp->b_vp != vp ||
2692 			    (nbp->b_flags & B_DELWRI) == (clean? B_DELWRI: 0)))
2693 				return (EAGAIN);
2694 		}
2695 	} while (clean = !clean, anyfreed-- > 0);
2696 	return (0);
2697 }
2698 
2699 static void
2700 buf_vlist_remove(struct buf *bp)
2701 {
2702 	struct bufv *bv;
2703 	b_xflags_t flags;
2704 
2705 	flags = bp->b_xflags;
2706 
2707 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2708 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2709 	KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 &&
2710 	    (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN),
2711 	    ("%s: buffer %p has invalid queue state", __func__, bp));
2712 
2713 	if ((flags & BX_VNDIRTY) != 0)
2714 		bv = &bp->b_bufobj->bo_dirty;
2715 	else
2716 		bv = &bp->b_bufobj->bo_clean;
2717 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2718 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2719 	bv->bv_cnt--;
2720 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2721 }
2722 
2723 /*
2724  * Add the buffer to the sorted clean or dirty block list.  Return zero on
2725  * success, EEXIST if a buffer with this identity already exists, or another
2726  * error on allocation failure.
2727  */
2728 static inline int
2729 buf_vlist_find_or_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2730 {
2731 	struct bufv *bv;
2732 	struct buf *n;
2733 	int error;
2734 
2735 	ASSERT_BO_WLOCKED(bo);
2736 	KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2737 	    ("buf_vlist_add: bo %p does not allow bufs", bo));
2738 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2739 	    ("dead bo %p", bo));
2740 	KASSERT((bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) == xflags,
2741 	    ("buf_vlist_add: b_xflags %#x not set on bp %p", xflags, bp));
2742 
2743 	if (xflags & BX_VNDIRTY)
2744 		bv = &bo->bo_dirty;
2745 	else
2746 		bv = &bo->bo_clean;
2747 
2748 	error = buf_insert_lookup_le(bv, bp, &n);
2749 	if (n == NULL) {
2750 		KASSERT(error != EEXIST,
2751 		    ("buf_vlist_add: EEXIST but no existing buf found: bp %p",
2752 		    bp));
2753 	} else {
2754 		KASSERT(n->b_lblkno <= bp->b_lblkno,
2755 		    ("buf_vlist_add: out of order insert/lookup: bp %p n %p",
2756 		    bp, n));
2757 		KASSERT((n->b_lblkno == bp->b_lblkno) == (error == EEXIST),
2758 		    ("buf_vlist_add: inconsistent result for existing buf: "
2759 		    "error %d bp %p n %p", error, bp, n));
2760 	}
2761 	if (error != 0)
2762 		return (error);
2763 
2764 	/* Keep the list ordered. */
2765 	if (n == NULL) {
2766 		KASSERT(TAILQ_EMPTY(&bv->bv_hd) ||
2767 		    bp->b_lblkno < TAILQ_FIRST(&bv->bv_hd)->b_lblkno,
2768 		    ("buf_vlist_add: queue order: "
2769 		    "%p should be before first %p",
2770 		    bp, TAILQ_FIRST(&bv->bv_hd)));
2771 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2772 	} else {
2773 		KASSERT(TAILQ_NEXT(n, b_bobufs) == NULL ||
2774 		    bp->b_lblkno < TAILQ_NEXT(n, b_bobufs)->b_lblkno,
2775 		    ("buf_vlist_add: queue order: "
2776 		    "%p should be before next %p",
2777 		    bp, TAILQ_NEXT(n, b_bobufs)));
2778 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2779 	}
2780 
2781 	bv->bv_cnt++;
2782 	return (0);
2783 }
2784 
2785 /*
2786  * Add the buffer to the sorted clean or dirty block list.
2787  *
2788  * NOTE: xflags is passed as a constant, optimizing this inline function!
2789  */
2790 static void
2791 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2792 {
2793 	int error;
2794 
2795 	KASSERT((bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) == 0,
2796 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2797 	bp->b_xflags |= xflags;
2798 	error = buf_vlist_find_or_add(bp, bo, xflags);
2799 	if (error)
2800 		panic("buf_vlist_add: error=%d", error);
2801 }
2802 
2803 /*
2804  * Look up a buffer using the buffer tries.
2805  */
2806 struct buf *
2807 gbincore(struct bufobj *bo, daddr_t lblkno)
2808 {
2809 	struct buf *bp;
2810 
2811 	ASSERT_BO_LOCKED(bo);
2812 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2813 	if (bp != NULL)
2814 		return (bp);
2815 	return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno));
2816 }
2817 
2818 /*
2819  * Look up a buf using the buffer tries, without the bufobj lock.  This relies
2820  * on SMR for safe lookup, and bufs being in a no-free zone to provide type
2821  * stability of the result.  Like other lockless lookups, the found buf may
2822  * already be invalid by the time this function returns.
2823  */
2824 struct buf *
2825 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno)
2826 {
2827 	struct buf *bp;
2828 
2829 	ASSERT_BO_UNLOCKED(bo);
2830 	bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno);
2831 	if (bp != NULL)
2832 		return (bp);
2833 	return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno));
2834 }
2835 
2836 /*
2837  * Associate a buffer with a vnode.
2838  */
2839 int
2840 bgetvp(struct vnode *vp, struct buf *bp)
2841 {
2842 	struct bufobj *bo;
2843 	int error;
2844 
2845 	bo = &vp->v_bufobj;
2846 	ASSERT_BO_UNLOCKED(bo);
2847 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2848 
2849 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2850 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2851 	    ("bgetvp: bp already attached! %p", bp));
2852 
2853 	/*
2854 	 * Add the buf to the vnode's clean list unless we lost a race and find
2855 	 * an existing buf in either dirty or clean.
2856 	 */
2857 	bp->b_vp = vp;
2858 	bp->b_bufobj = bo;
2859 	bp->b_xflags |= BX_VNCLEAN;
2860 	error = EEXIST;
2861 	BO_LOCK(bo);
2862 	if (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, bp->b_lblkno) == NULL)
2863 		error = buf_vlist_find_or_add(bp, bo, BX_VNCLEAN);
2864 	BO_UNLOCK(bo);
2865 	if (__predict_true(error == 0)) {
2866 		vhold(vp);
2867 		return (0);
2868 	}
2869 	if (error != EEXIST)
2870 		panic("bgetvp: buf_vlist_add error: %d", error);
2871 	bp->b_vp = NULL;
2872 	bp->b_bufobj = NULL;
2873 	bp->b_xflags &= ~BX_VNCLEAN;
2874 	return (error);
2875 }
2876 
2877 /*
2878  * Disassociate a buffer from a vnode.
2879  */
2880 void
2881 brelvp(struct buf *bp)
2882 {
2883 	struct bufobj *bo;
2884 	struct vnode *vp;
2885 
2886 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2887 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2888 
2889 	/*
2890 	 * Delete from old vnode list, if on one.
2891 	 */
2892 	vp = bp->b_vp;		/* XXX */
2893 	bo = bp->b_bufobj;
2894 	BO_LOCK(bo);
2895 	buf_vlist_remove(bp);
2896 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2897 		bo->bo_flag &= ~BO_ONWORKLST;
2898 		mtx_lock(&sync_mtx);
2899 		LIST_REMOVE(bo, bo_synclist);
2900 		syncer_worklist_len--;
2901 		mtx_unlock(&sync_mtx);
2902 	}
2903 	bp->b_vp = NULL;
2904 	bp->b_bufobj = NULL;
2905 	BO_UNLOCK(bo);
2906 	vdrop(vp);
2907 }
2908 
2909 /*
2910  * Add an item to the syncer work queue.
2911  */
2912 static void
2913 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2914 {
2915 	int slot;
2916 
2917 	ASSERT_BO_WLOCKED(bo);
2918 
2919 	mtx_lock(&sync_mtx);
2920 	if (bo->bo_flag & BO_ONWORKLST)
2921 		LIST_REMOVE(bo, bo_synclist);
2922 	else {
2923 		bo->bo_flag |= BO_ONWORKLST;
2924 		syncer_worklist_len++;
2925 	}
2926 
2927 	if (delay > syncer_maxdelay - 2)
2928 		delay = syncer_maxdelay - 2;
2929 	slot = (syncer_delayno + delay) & syncer_mask;
2930 
2931 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2932 	mtx_unlock(&sync_mtx);
2933 }
2934 
2935 static int
2936 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2937 {
2938 	int error, len;
2939 
2940 	mtx_lock(&sync_mtx);
2941 	len = syncer_worklist_len - sync_vnode_count;
2942 	mtx_unlock(&sync_mtx);
2943 	error = SYSCTL_OUT(req, &len, sizeof(len));
2944 	return (error);
2945 }
2946 
2947 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2948     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2949     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2950 
2951 static struct proc *updateproc;
2952 static void sched_sync(void);
2953 static struct kproc_desc up_kp = {
2954 	"syncer",
2955 	sched_sync,
2956 	&updateproc
2957 };
2958 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2959 
2960 static int
2961 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2962 {
2963 	struct vnode *vp;
2964 	struct mount *mp;
2965 
2966 	*bo = LIST_FIRST(slp);
2967 	if (*bo == NULL)
2968 		return (0);
2969 	vp = bo2vnode(*bo);
2970 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2971 		return (1);
2972 	/*
2973 	 * We use vhold in case the vnode does not
2974 	 * successfully sync.  vhold prevents the vnode from
2975 	 * going away when we unlock the sync_mtx so that
2976 	 * we can acquire the vnode interlock.
2977 	 */
2978 	vholdl(vp);
2979 	mtx_unlock(&sync_mtx);
2980 	VI_UNLOCK(vp);
2981 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2982 		vdrop(vp);
2983 		mtx_lock(&sync_mtx);
2984 		return (*bo == LIST_FIRST(slp));
2985 	}
2986 	MPASSERT(mp == NULL || (curthread->td_pflags & TDP_IGNSUSP) != 0 ||
2987 	    (mp->mnt_kern_flag & MNTK_SUSPENDED) == 0, mp,
2988 	    ("suspended mp syncing vp %p", vp));
2989 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2990 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2991 	VOP_UNLOCK(vp);
2992 	vn_finished_write(mp);
2993 	BO_LOCK(*bo);
2994 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2995 		/*
2996 		 * Put us back on the worklist.  The worklist
2997 		 * routine will remove us from our current
2998 		 * position and then add us back in at a later
2999 		 * position.
3000 		 */
3001 		vn_syncer_add_to_worklist(*bo, syncdelay);
3002 	}
3003 	BO_UNLOCK(*bo);
3004 	vdrop(vp);
3005 	mtx_lock(&sync_mtx);
3006 	return (0);
3007 }
3008 
3009 static int first_printf = 1;
3010 
3011 /*
3012  * System filesystem synchronizer daemon.
3013  */
3014 static void
3015 sched_sync(void)
3016 {
3017 	struct synclist *next, *slp;
3018 	struct bufobj *bo;
3019 	long starttime;
3020 	struct thread *td = curthread;
3021 	int last_work_seen;
3022 	int net_worklist_len;
3023 	int syncer_final_iter;
3024 	int error;
3025 
3026 	last_work_seen = 0;
3027 	syncer_final_iter = 0;
3028 	syncer_state = SYNCER_RUNNING;
3029 	starttime = time_uptime;
3030 	td->td_pflags |= TDP_NORUNNINGBUF;
3031 
3032 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
3033 	    SHUTDOWN_PRI_LAST);
3034 
3035 	mtx_lock(&sync_mtx);
3036 	for (;;) {
3037 		if (syncer_state == SYNCER_FINAL_DELAY &&
3038 		    syncer_final_iter == 0) {
3039 			mtx_unlock(&sync_mtx);
3040 			kproc_suspend_check(td->td_proc);
3041 			mtx_lock(&sync_mtx);
3042 		}
3043 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
3044 		if (syncer_state != SYNCER_RUNNING &&
3045 		    starttime != time_uptime) {
3046 			if (first_printf) {
3047 				printf("\nSyncing disks, vnodes remaining... ");
3048 				first_printf = 0;
3049 			}
3050 			printf("%d ", net_worklist_len);
3051 		}
3052 		starttime = time_uptime;
3053 
3054 		/*
3055 		 * Push files whose dirty time has expired.  Be careful
3056 		 * of interrupt race on slp queue.
3057 		 *
3058 		 * Skip over empty worklist slots when shutting down.
3059 		 */
3060 		do {
3061 			slp = &syncer_workitem_pending[syncer_delayno];
3062 			syncer_delayno += 1;
3063 			if (syncer_delayno == syncer_maxdelay)
3064 				syncer_delayno = 0;
3065 			next = &syncer_workitem_pending[syncer_delayno];
3066 			/*
3067 			 * If the worklist has wrapped since the
3068 			 * it was emptied of all but syncer vnodes,
3069 			 * switch to the FINAL_DELAY state and run
3070 			 * for one more second.
3071 			 */
3072 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
3073 			    net_worklist_len == 0 &&
3074 			    last_work_seen == syncer_delayno) {
3075 				syncer_state = SYNCER_FINAL_DELAY;
3076 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
3077 			}
3078 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
3079 		    syncer_worklist_len > 0);
3080 
3081 		/*
3082 		 * Keep track of the last time there was anything
3083 		 * on the worklist other than syncer vnodes.
3084 		 * Return to the SHUTTING_DOWN state if any
3085 		 * new work appears.
3086 		 */
3087 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
3088 			last_work_seen = syncer_delayno;
3089 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
3090 			syncer_state = SYNCER_SHUTTING_DOWN;
3091 		while (!LIST_EMPTY(slp)) {
3092 			error = sync_vnode(slp, &bo, td);
3093 			if (error == 1) {
3094 				LIST_REMOVE(bo, bo_synclist);
3095 				LIST_INSERT_HEAD(next, bo, bo_synclist);
3096 				continue;
3097 			}
3098 
3099 			if (first_printf == 0) {
3100 				/*
3101 				 * Drop the sync mutex, because some watchdog
3102 				 * drivers need to sleep while patting
3103 				 */
3104 				mtx_unlock(&sync_mtx);
3105 				wdog_kern_pat(WD_LASTVAL);
3106 				mtx_lock(&sync_mtx);
3107 			}
3108 		}
3109 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
3110 			syncer_final_iter--;
3111 		/*
3112 		 * The variable rushjob allows the kernel to speed up the
3113 		 * processing of the filesystem syncer process. A rushjob
3114 		 * value of N tells the filesystem syncer to process the next
3115 		 * N seconds worth of work on its queue ASAP. Currently rushjob
3116 		 * is used by the soft update code to speed up the filesystem
3117 		 * syncer process when the incore state is getting so far
3118 		 * ahead of the disk that the kernel memory pool is being
3119 		 * threatened with exhaustion.
3120 		 */
3121 		if (rushjob > 0) {
3122 			rushjob -= 1;
3123 			continue;
3124 		}
3125 		/*
3126 		 * Just sleep for a short period of time between
3127 		 * iterations when shutting down to allow some I/O
3128 		 * to happen.
3129 		 *
3130 		 * If it has taken us less than a second to process the
3131 		 * current work, then wait. Otherwise start right over
3132 		 * again. We can still lose time if any single round
3133 		 * takes more than two seconds, but it does not really
3134 		 * matter as we are just trying to generally pace the
3135 		 * filesystem activity.
3136 		 */
3137 		if (syncer_state != SYNCER_RUNNING ||
3138 		    time_uptime == starttime) {
3139 			thread_lock(td);
3140 			sched_prio(td, PPAUSE);
3141 			thread_unlock(td);
3142 		}
3143 		if (syncer_state != SYNCER_RUNNING)
3144 			cv_timedwait(&sync_wakeup, &sync_mtx,
3145 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
3146 		else if (time_uptime == starttime)
3147 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
3148 	}
3149 }
3150 
3151 /*
3152  * Request the syncer daemon to speed up its work.
3153  * We never push it to speed up more than half of its
3154  * normal turn time, otherwise it could take over the cpu.
3155  */
3156 int
3157 speedup_syncer(void)
3158 {
3159 	int ret = 0;
3160 
3161 	mtx_lock(&sync_mtx);
3162 	if (rushjob < syncdelay / 2) {
3163 		rushjob += 1;
3164 		stat_rush_requests += 1;
3165 		ret = 1;
3166 	}
3167 	mtx_unlock(&sync_mtx);
3168 	cv_broadcast(&sync_wakeup);
3169 	return (ret);
3170 }
3171 
3172 /*
3173  * Tell the syncer to speed up its work and run though its work
3174  * list several times, then tell it to shut down.
3175  */
3176 static void
3177 syncer_shutdown(void *arg, int howto)
3178 {
3179 
3180 	if (howto & RB_NOSYNC)
3181 		return;
3182 	mtx_lock(&sync_mtx);
3183 	syncer_state = SYNCER_SHUTTING_DOWN;
3184 	rushjob = 0;
3185 	mtx_unlock(&sync_mtx);
3186 	cv_broadcast(&sync_wakeup);
3187 	kproc_shutdown(arg, howto);
3188 }
3189 
3190 void
3191 syncer_suspend(void)
3192 {
3193 
3194 	syncer_shutdown(updateproc, 0);
3195 }
3196 
3197 void
3198 syncer_resume(void)
3199 {
3200 
3201 	mtx_lock(&sync_mtx);
3202 	first_printf = 1;
3203 	syncer_state = SYNCER_RUNNING;
3204 	mtx_unlock(&sync_mtx);
3205 	cv_broadcast(&sync_wakeup);
3206 	kproc_resume(updateproc);
3207 }
3208 
3209 /*
3210  * Move the buffer between the clean and dirty lists of its vnode.
3211  */
3212 void
3213 reassignbuf(struct buf *bp)
3214 {
3215 	struct vnode *vp;
3216 	struct bufobj *bo;
3217 	int delay;
3218 #ifdef INVARIANTS
3219 	struct bufv *bv;
3220 #endif
3221 
3222 	vp = bp->b_vp;
3223 	bo = bp->b_bufobj;
3224 
3225 	KASSERT((bp->b_flags & B_PAGING) == 0,
3226 	    ("%s: cannot reassign paging buffer %p", __func__, bp));
3227 
3228 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
3229 	    bp, bp->b_vp, bp->b_flags);
3230 
3231 	BO_LOCK(bo);
3232 	if ((bo->bo_flag & BO_NONSTERILE) == 0) {
3233 		/*
3234 		 * Coordinate with getblk's unlocked lookup.  Make
3235 		 * BO_NONSTERILE visible before the first reassignbuf produces
3236 		 * any side effect.  This could be outside the bo lock if we
3237 		 * used a separate atomic flag field.
3238 		 */
3239 		bo->bo_flag |= BO_NONSTERILE;
3240 		atomic_thread_fence_rel();
3241 	}
3242 	buf_vlist_remove(bp);
3243 
3244 	/*
3245 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
3246 	 * of clean buffers.
3247 	 */
3248 	if (bp->b_flags & B_DELWRI) {
3249 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
3250 			switch (vp->v_type) {
3251 			case VDIR:
3252 				delay = dirdelay;
3253 				break;
3254 			case VCHR:
3255 				delay = metadelay;
3256 				break;
3257 			default:
3258 				delay = filedelay;
3259 			}
3260 			vn_syncer_add_to_worklist(bo, delay);
3261 		}
3262 		buf_vlist_add(bp, bo, BX_VNDIRTY);
3263 	} else {
3264 		buf_vlist_add(bp, bo, BX_VNCLEAN);
3265 
3266 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
3267 			mtx_lock(&sync_mtx);
3268 			LIST_REMOVE(bo, bo_synclist);
3269 			syncer_worklist_len--;
3270 			mtx_unlock(&sync_mtx);
3271 			bo->bo_flag &= ~BO_ONWORKLST;
3272 		}
3273 	}
3274 #ifdef INVARIANTS
3275 	bv = &bo->bo_clean;
3276 	bp = TAILQ_FIRST(&bv->bv_hd);
3277 	KASSERT(bp == NULL || bp->b_bufobj == bo,
3278 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3279 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
3280 	KASSERT(bp == NULL || bp->b_bufobj == bo,
3281 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3282 	bv = &bo->bo_dirty;
3283 	bp = TAILQ_FIRST(&bv->bv_hd);
3284 	KASSERT(bp == NULL || bp->b_bufobj == bo,
3285 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3286 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
3287 	KASSERT(bp == NULL || bp->b_bufobj == bo,
3288 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3289 #endif
3290 	BO_UNLOCK(bo);
3291 }
3292 
3293 static void
3294 v_init_counters(struct vnode *vp)
3295 {
3296 
3297 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
3298 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
3299 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
3300 
3301 	refcount_init(&vp->v_holdcnt, 1);
3302 	refcount_init(&vp->v_usecount, 1);
3303 }
3304 
3305 /*
3306  * Get a usecount on a vnode.
3307  *
3308  * vget and vget_finish may fail to lock the vnode if they lose a race against
3309  * it being doomed. LK_RETRY can be passed in flags to lock it anyway.
3310  *
3311  * Consumers which don't guarantee liveness of the vnode can use SMR to
3312  * try to get a reference. Note this operation can fail since the vnode
3313  * may be awaiting getting freed by the time they get to it.
3314  */
3315 enum vgetstate
3316 vget_prep_smr(struct vnode *vp)
3317 {
3318 	enum vgetstate vs;
3319 
3320 	VFS_SMR_ASSERT_ENTERED();
3321 
3322 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
3323 		vs = VGET_USECOUNT;
3324 	} else {
3325 		if (vhold_smr(vp))
3326 			vs = VGET_HOLDCNT;
3327 		else
3328 			vs = VGET_NONE;
3329 	}
3330 	return (vs);
3331 }
3332 
3333 enum vgetstate
3334 vget_prep(struct vnode *vp)
3335 {
3336 	enum vgetstate vs;
3337 
3338 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
3339 		vs = VGET_USECOUNT;
3340 	} else {
3341 		vhold(vp);
3342 		vs = VGET_HOLDCNT;
3343 	}
3344 	return (vs);
3345 }
3346 
3347 void
3348 vget_abort(struct vnode *vp, enum vgetstate vs)
3349 {
3350 
3351 	switch (vs) {
3352 	case VGET_USECOUNT:
3353 		vrele(vp);
3354 		break;
3355 	case VGET_HOLDCNT:
3356 		vdrop(vp);
3357 		break;
3358 	default:
3359 		__assert_unreachable();
3360 	}
3361 }
3362 
3363 int
3364 vget(struct vnode *vp, int flags)
3365 {
3366 	enum vgetstate vs;
3367 
3368 	vs = vget_prep(vp);
3369 	return (vget_finish(vp, flags, vs));
3370 }
3371 
3372 int
3373 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
3374 {
3375 	int error;
3376 
3377 	if ((flags & LK_INTERLOCK) != 0)
3378 		ASSERT_VI_LOCKED(vp, __func__);
3379 	else
3380 		ASSERT_VI_UNLOCKED(vp, __func__);
3381 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3382 	VNPASS(vp->v_holdcnt > 0, vp);
3383 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3384 
3385 	error = vn_lock(vp, flags);
3386 	if (__predict_false(error != 0)) {
3387 		vget_abort(vp, vs);
3388 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
3389 		    vp);
3390 		return (error);
3391 	}
3392 
3393 	vget_finish_ref(vp, vs);
3394 	return (0);
3395 }
3396 
3397 void
3398 vget_finish_ref(struct vnode *vp, enum vgetstate vs)
3399 {
3400 	int old;
3401 
3402 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3403 	VNPASS(vp->v_holdcnt > 0, vp);
3404 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3405 
3406 	if (vs == VGET_USECOUNT)
3407 		return;
3408 
3409 	/*
3410 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
3411 	 * the vnode around. Otherwise someone else lended their hold count and
3412 	 * we have to drop ours.
3413 	 */
3414 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
3415 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3416 	if (old != 0) {
3417 #ifdef INVARIANTS
3418 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3419 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3420 #else
3421 		refcount_release(&vp->v_holdcnt);
3422 #endif
3423 	}
3424 }
3425 
3426 void
3427 vref(struct vnode *vp)
3428 {
3429 	enum vgetstate vs;
3430 
3431 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3432 	vs = vget_prep(vp);
3433 	vget_finish_ref(vp, vs);
3434 }
3435 
3436 void
3437 vrefact(struct vnode *vp)
3438 {
3439 	int old __diagused;
3440 
3441 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3442 	old = refcount_acquire(&vp->v_usecount);
3443 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3444 }
3445 
3446 void
3447 vlazy(struct vnode *vp)
3448 {
3449 	struct mount *mp;
3450 
3451 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3452 
3453 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3454 		return;
3455 	/*
3456 	 * We may get here for inactive routines after the vnode got doomed.
3457 	 */
3458 	if (VN_IS_DOOMED(vp))
3459 		return;
3460 	mp = vp->v_mount;
3461 	mtx_lock(&mp->mnt_listmtx);
3462 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3463 		vp->v_mflag |= VMP_LAZYLIST;
3464 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3465 		mp->mnt_lazyvnodelistsize++;
3466 	}
3467 	mtx_unlock(&mp->mnt_listmtx);
3468 }
3469 
3470 static void
3471 vunlazy(struct vnode *vp)
3472 {
3473 	struct mount *mp;
3474 
3475 	ASSERT_VI_LOCKED(vp, __func__);
3476 	VNPASS(!VN_IS_DOOMED(vp), vp);
3477 
3478 	mp = vp->v_mount;
3479 	mtx_lock(&mp->mnt_listmtx);
3480 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3481 	/*
3482 	 * Don't remove the vnode from the lazy list if another thread
3483 	 * has increased the hold count. It may have re-enqueued the
3484 	 * vnode to the lazy list and is now responsible for its
3485 	 * removal.
3486 	 */
3487 	if (vp->v_holdcnt == 0) {
3488 		vp->v_mflag &= ~VMP_LAZYLIST;
3489 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3490 		mp->mnt_lazyvnodelistsize--;
3491 	}
3492 	mtx_unlock(&mp->mnt_listmtx);
3493 }
3494 
3495 /*
3496  * This routine is only meant to be called from vgonel prior to dooming
3497  * the vnode.
3498  */
3499 static void
3500 vunlazy_gone(struct vnode *vp)
3501 {
3502 	struct mount *mp;
3503 
3504 	ASSERT_VOP_ELOCKED(vp, __func__);
3505 	ASSERT_VI_LOCKED(vp, __func__);
3506 	VNPASS(!VN_IS_DOOMED(vp), vp);
3507 
3508 	if (vp->v_mflag & VMP_LAZYLIST) {
3509 		mp = vp->v_mount;
3510 		mtx_lock(&mp->mnt_listmtx);
3511 		VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3512 		vp->v_mflag &= ~VMP_LAZYLIST;
3513 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3514 		mp->mnt_lazyvnodelistsize--;
3515 		mtx_unlock(&mp->mnt_listmtx);
3516 	}
3517 }
3518 
3519 static void
3520 vdefer_inactive(struct vnode *vp)
3521 {
3522 
3523 	ASSERT_VI_LOCKED(vp, __func__);
3524 	VNPASS(vp->v_holdcnt > 0, vp);
3525 	if (VN_IS_DOOMED(vp)) {
3526 		vdropl(vp);
3527 		return;
3528 	}
3529 	if (vp->v_iflag & VI_DEFINACT) {
3530 		VNPASS(vp->v_holdcnt > 1, vp);
3531 		vdropl(vp);
3532 		return;
3533 	}
3534 	if (vp->v_usecount > 0) {
3535 		vp->v_iflag &= ~VI_OWEINACT;
3536 		vdropl(vp);
3537 		return;
3538 	}
3539 	vlazy(vp);
3540 	vp->v_iflag |= VI_DEFINACT;
3541 	VI_UNLOCK(vp);
3542 	atomic_add_long(&deferred_inact, 1);
3543 }
3544 
3545 static void
3546 vdefer_inactive_unlocked(struct vnode *vp)
3547 {
3548 
3549 	VI_LOCK(vp);
3550 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3551 		vdropl(vp);
3552 		return;
3553 	}
3554 	vdefer_inactive(vp);
3555 }
3556 
3557 enum vput_op { VRELE, VPUT, VUNREF };
3558 
3559 /*
3560  * Handle ->v_usecount transitioning to 0.
3561  *
3562  * By releasing the last usecount we take ownership of the hold count which
3563  * provides liveness of the vnode, meaning we have to vdrop.
3564  *
3565  * For all vnodes we may need to perform inactive processing. It requires an
3566  * exclusive lock on the vnode, while it is legal to call here with only a
3567  * shared lock (or no locks). If locking the vnode in an expected manner fails,
3568  * inactive processing gets deferred to the syncer.
3569  *
3570  * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3571  * on the lock being held all the way until VOP_INACTIVE. This in particular
3572  * happens with UFS which adds half-constructed vnodes to the hash, where they
3573  * can be found by other code.
3574  */
3575 static void
3576 vput_final(struct vnode *vp, enum vput_op func)
3577 {
3578 	int error;
3579 	bool want_unlock;
3580 
3581 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3582 	VNPASS(vp->v_holdcnt > 0, vp);
3583 
3584 	VI_LOCK(vp);
3585 
3586 	/*
3587 	 * By the time we got here someone else might have transitioned
3588 	 * the count back to > 0.
3589 	 */
3590 	if (vp->v_usecount > 0)
3591 		goto out;
3592 
3593 	/*
3594 	 * If the vnode is doomed vgone already performed inactive processing
3595 	 * (if needed).
3596 	 */
3597 	if (VN_IS_DOOMED(vp))
3598 		goto out;
3599 
3600 	if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3601 		goto out;
3602 
3603 	if (vp->v_iflag & VI_DOINGINACT)
3604 		goto out;
3605 
3606 	/*
3607 	 * Locking operations here will drop the interlock and possibly the
3608 	 * vnode lock, opening a window where the vnode can get doomed all the
3609 	 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3610 	 * perform inactive.
3611 	 */
3612 	vp->v_iflag |= VI_OWEINACT;
3613 	want_unlock = false;
3614 	error = 0;
3615 	switch (func) {
3616 	case VRELE:
3617 		switch (VOP_ISLOCKED(vp)) {
3618 		case LK_EXCLUSIVE:
3619 			break;
3620 		case LK_EXCLOTHER:
3621 		case 0:
3622 			want_unlock = true;
3623 			error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3624 			VI_LOCK(vp);
3625 			break;
3626 		default:
3627 			/*
3628 			 * The lock has at least one sharer, but we have no way
3629 			 * to conclude whether this is us. Play it safe and
3630 			 * defer processing.
3631 			 */
3632 			error = EAGAIN;
3633 			break;
3634 		}
3635 		break;
3636 	case VPUT:
3637 		want_unlock = true;
3638 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3639 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3640 			    LK_NOWAIT);
3641 			VI_LOCK(vp);
3642 		}
3643 		break;
3644 	case VUNREF:
3645 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3646 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3647 			VI_LOCK(vp);
3648 		}
3649 		break;
3650 	}
3651 	if (error == 0) {
3652 		if (func == VUNREF) {
3653 			VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp,
3654 			    ("recursive vunref"));
3655 			vp->v_vflag |= VV_UNREF;
3656 		}
3657 		for (;;) {
3658 			error = vinactive(vp);
3659 			if (want_unlock)
3660 				VOP_UNLOCK(vp);
3661 			if (error != ERELOOKUP || !want_unlock)
3662 				break;
3663 			VOP_LOCK(vp, LK_EXCLUSIVE);
3664 		}
3665 		if (func == VUNREF)
3666 			vp->v_vflag &= ~VV_UNREF;
3667 		vdropl(vp);
3668 	} else {
3669 		vdefer_inactive(vp);
3670 	}
3671 	return;
3672 out:
3673 	if (func == VPUT)
3674 		VOP_UNLOCK(vp);
3675 	vdropl(vp);
3676 }
3677 
3678 /*
3679  * Decrement ->v_usecount for a vnode.
3680  *
3681  * Releasing the last use count requires additional processing, see vput_final
3682  * above for details.
3683  *
3684  * Comment above each variant denotes lock state on entry and exit.
3685  */
3686 
3687 /*
3688  * in: any
3689  * out: same as passed in
3690  */
3691 void
3692 vrele(struct vnode *vp)
3693 {
3694 
3695 	ASSERT_VI_UNLOCKED(vp, __func__);
3696 	if (!refcount_release(&vp->v_usecount))
3697 		return;
3698 	vput_final(vp, VRELE);
3699 }
3700 
3701 /*
3702  * in: locked
3703  * out: unlocked
3704  */
3705 void
3706 vput(struct vnode *vp)
3707 {
3708 
3709 	ASSERT_VOP_LOCKED(vp, __func__);
3710 	ASSERT_VI_UNLOCKED(vp, __func__);
3711 	if (!refcount_release(&vp->v_usecount)) {
3712 		VOP_UNLOCK(vp);
3713 		return;
3714 	}
3715 	vput_final(vp, VPUT);
3716 }
3717 
3718 /*
3719  * in: locked
3720  * out: locked
3721  */
3722 void
3723 vunref(struct vnode *vp)
3724 {
3725 
3726 	ASSERT_VOP_LOCKED(vp, __func__);
3727 	ASSERT_VI_UNLOCKED(vp, __func__);
3728 	if (!refcount_release(&vp->v_usecount))
3729 		return;
3730 	vput_final(vp, VUNREF);
3731 }
3732 
3733 void
3734 vhold(struct vnode *vp)
3735 {
3736 	int old;
3737 
3738 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3739 	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3740 	VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3741 	    ("%s: wrong hold count %d", __func__, old));
3742 	if (old == 0)
3743 		vfs_freevnodes_dec();
3744 }
3745 
3746 void
3747 vholdnz(struct vnode *vp)
3748 {
3749 
3750 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3751 #ifdef INVARIANTS
3752 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3753 	VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3754 	    ("%s: wrong hold count %d", __func__, old));
3755 #else
3756 	atomic_add_int(&vp->v_holdcnt, 1);
3757 #endif
3758 }
3759 
3760 /*
3761  * Grab a hold count unless the vnode is freed.
3762  *
3763  * Only use this routine if vfs smr is the only protection you have against
3764  * freeing the vnode.
3765  *
3766  * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag
3767  * is not set.  After the flag is set the vnode becomes immutable to anyone but
3768  * the thread which managed to set the flag.
3769  *
3770  * It may be tempting to replace the loop with:
3771  * count = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3772  * if (count & VHOLD_NO_SMR) {
3773  *     backpedal and error out;
3774  * }
3775  *
3776  * However, while this is more performant, it hinders debugging by eliminating
3777  * the previously mentioned invariant.
3778  */
3779 bool
3780 vhold_smr(struct vnode *vp)
3781 {
3782 	int count;
3783 
3784 	VFS_SMR_ASSERT_ENTERED();
3785 
3786 	count = atomic_load_int(&vp->v_holdcnt);
3787 	for (;;) {
3788 		if (count & VHOLD_NO_SMR) {
3789 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3790 			    ("non-zero hold count with flags %d\n", count));
3791 			return (false);
3792 		}
3793 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3794 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3795 			if (count == 0)
3796 				vfs_freevnodes_dec();
3797 			return (true);
3798 		}
3799 	}
3800 }
3801 
3802 /*
3803  * Hold a free vnode for recycling.
3804  *
3805  * Note: vnode_init references this comment.
3806  *
3807  * Attempts to recycle only need the global vnode list lock and have no use for
3808  * SMR.
3809  *
3810  * However, vnodes get inserted into the global list before they get fully
3811  * initialized and stay there until UMA decides to free the memory. This in
3812  * particular means the target can be found before it becomes usable and after
3813  * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to
3814  * VHOLD_NO_SMR.
3815  *
3816  * Note: the vnode may gain more references after we transition the count 0->1.
3817  */
3818 static bool
3819 vhold_recycle_free(struct vnode *vp)
3820 {
3821 	int count;
3822 
3823 	mtx_assert(&vnode_list_mtx, MA_OWNED);
3824 
3825 	count = atomic_load_int(&vp->v_holdcnt);
3826 	for (;;) {
3827 		if (count & VHOLD_NO_SMR) {
3828 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3829 			    ("non-zero hold count with flags %d\n", count));
3830 			return (false);
3831 		}
3832 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3833 		if (count > 0) {
3834 			return (false);
3835 		}
3836 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3837 			vfs_freevnodes_dec();
3838 			return (true);
3839 		}
3840 	}
3841 }
3842 
3843 static void __noinline
3844 vdbatch_process(struct vdbatch *vd)
3845 {
3846 	struct vnode *vp;
3847 	int i;
3848 
3849 	mtx_assert(&vd->lock, MA_OWNED);
3850 	MPASS(curthread->td_pinned > 0);
3851 	MPASS(vd->index == VDBATCH_SIZE);
3852 
3853 	/*
3854 	 * Attempt to requeue the passed batch, but give up easily.
3855 	 *
3856 	 * Despite batching the mechanism is prone to transient *significant*
3857 	 * lock contention, where vnode_list_mtx becomes the primary bottleneck
3858 	 * if multiple CPUs get here (one real-world example is highly parallel
3859 	 * do-nothing make , which will stat *tons* of vnodes). Since it is
3860 	 * quasi-LRU (read: not that great even if fully honoured) provide an
3861 	 * option to just dodge the problem. Parties which don't like it are
3862 	 * welcome to implement something better.
3863 	 */
3864 	if (vnode_can_skip_requeue) {
3865 		if (!mtx_trylock(&vnode_list_mtx)) {
3866 			counter_u64_add(vnode_skipped_requeues, 1);
3867 			critical_enter();
3868 			for (i = 0; i < VDBATCH_SIZE; i++) {
3869 				vp = vd->tab[i];
3870 				vd->tab[i] = NULL;
3871 				MPASS(vp->v_dbatchcpu != NOCPU);
3872 				vp->v_dbatchcpu = NOCPU;
3873 			}
3874 			vd->index = 0;
3875 			critical_exit();
3876 			return;
3877 
3878 		}
3879 		/* fallthrough to locked processing */
3880 	} else {
3881 		mtx_lock(&vnode_list_mtx);
3882 	}
3883 
3884 	mtx_assert(&vnode_list_mtx, MA_OWNED);
3885 	critical_enter();
3886 	for (i = 0; i < VDBATCH_SIZE; i++) {
3887 		vp = vd->tab[i];
3888 		vd->tab[i] = NULL;
3889 		TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3890 		TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3891 		MPASS(vp->v_dbatchcpu != NOCPU);
3892 		vp->v_dbatchcpu = NOCPU;
3893 	}
3894 	mtx_unlock(&vnode_list_mtx);
3895 	vd->index = 0;
3896 	critical_exit();
3897 }
3898 
3899 static void
3900 vdbatch_enqueue(struct vnode *vp)
3901 {
3902 	struct vdbatch *vd;
3903 
3904 	ASSERT_VI_LOCKED(vp, __func__);
3905 	VNPASS(!VN_IS_DOOMED(vp), vp);
3906 
3907 	if (vp->v_dbatchcpu != NOCPU) {
3908 		VI_UNLOCK(vp);
3909 		return;
3910 	}
3911 
3912 	sched_pin();
3913 	vd = DPCPU_PTR(vd);
3914 	mtx_lock(&vd->lock);
3915 	MPASS(vd->index < VDBATCH_SIZE);
3916 	MPASS(vd->tab[vd->index] == NULL);
3917 	/*
3918 	 * A hack: we depend on being pinned so that we know what to put in
3919 	 * ->v_dbatchcpu.
3920 	 */
3921 	vp->v_dbatchcpu = curcpu;
3922 	vd->tab[vd->index] = vp;
3923 	vd->index++;
3924 	VI_UNLOCK(vp);
3925 	if (vd->index == VDBATCH_SIZE)
3926 		vdbatch_process(vd);
3927 	mtx_unlock(&vd->lock);
3928 	sched_unpin();
3929 }
3930 
3931 /*
3932  * This routine must only be called for vnodes which are about to be
3933  * deallocated. Supporting dequeue for arbitrary vndoes would require
3934  * validating that the locked batch matches.
3935  */
3936 static void
3937 vdbatch_dequeue(struct vnode *vp)
3938 {
3939 	struct vdbatch *vd;
3940 	int i;
3941 	short cpu;
3942 
3943 	VNPASS(vp->v_type == VBAD || vp->v_type == VNON, vp);
3944 
3945 	cpu = vp->v_dbatchcpu;
3946 	if (cpu == NOCPU)
3947 		return;
3948 
3949 	vd = DPCPU_ID_PTR(cpu, vd);
3950 	mtx_lock(&vd->lock);
3951 	for (i = 0; i < vd->index; i++) {
3952 		if (vd->tab[i] != vp)
3953 			continue;
3954 		vp->v_dbatchcpu = NOCPU;
3955 		vd->index--;
3956 		vd->tab[i] = vd->tab[vd->index];
3957 		vd->tab[vd->index] = NULL;
3958 		break;
3959 	}
3960 	mtx_unlock(&vd->lock);
3961 	/*
3962 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3963 	 */
3964 	MPASS(vp->v_dbatchcpu == NOCPU);
3965 }
3966 
3967 /*
3968  * Drop the hold count of the vnode.
3969  *
3970  * It will only get freed if this is the last hold *and* it has been vgone'd.
3971  *
3972  * Because the vnode vm object keeps a hold reference on the vnode if
3973  * there is at least one resident non-cached page, the vnode cannot
3974  * leave the active list without the page cleanup done.
3975  */
3976 static void __noinline
3977 vdropl_final(struct vnode *vp)
3978 {
3979 
3980 	ASSERT_VI_LOCKED(vp, __func__);
3981 	VNPASS(VN_IS_DOOMED(vp), vp);
3982 	/*
3983 	 * Set the VHOLD_NO_SMR flag.
3984 	 *
3985 	 * We may be racing against vhold_smr. If they win we can just pretend
3986 	 * we never got this far, they will vdrop later.
3987 	 */
3988 	if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) {
3989 		vfs_freevnodes_inc();
3990 		VI_UNLOCK(vp);
3991 		/*
3992 		 * We lost the aforementioned race. Any subsequent access is
3993 		 * invalid as they might have managed to vdropl on their own.
3994 		 */
3995 		return;
3996 	}
3997 	/*
3998 	 * Don't bump freevnodes as this one is going away.
3999 	 */
4000 	freevnode(vp);
4001 }
4002 
4003 void
4004 vdrop(struct vnode *vp)
4005 {
4006 
4007 	ASSERT_VI_UNLOCKED(vp, __func__);
4008 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4009 	if (refcount_release_if_not_last(&vp->v_holdcnt))
4010 		return;
4011 	VI_LOCK(vp);
4012 	vdropl(vp);
4013 }
4014 
4015 static __always_inline void
4016 vdropl_impl(struct vnode *vp, bool enqueue)
4017 {
4018 
4019 	ASSERT_VI_LOCKED(vp, __func__);
4020 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4021 	if (!refcount_release(&vp->v_holdcnt)) {
4022 		VI_UNLOCK(vp);
4023 		return;
4024 	}
4025 	VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp);
4026 	VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
4027 	if (VN_IS_DOOMED(vp)) {
4028 		vdropl_final(vp);
4029 		return;
4030 	}
4031 
4032 	vfs_freevnodes_inc();
4033 	if (vp->v_mflag & VMP_LAZYLIST) {
4034 		vunlazy(vp);
4035 	}
4036 
4037 	if (!enqueue) {
4038 		VI_UNLOCK(vp);
4039 		return;
4040 	}
4041 
4042 	/*
4043 	 * Also unlocks the interlock. We can't assert on it as we
4044 	 * released our hold and by now the vnode might have been
4045 	 * freed.
4046 	 */
4047 	vdbatch_enqueue(vp);
4048 }
4049 
4050 void
4051 vdropl(struct vnode *vp)
4052 {
4053 
4054 	vdropl_impl(vp, true);
4055 }
4056 
4057 /*
4058  * vdrop a vnode when recycling
4059  *
4060  * This is a special case routine only to be used when recycling, differs from
4061  * regular vdrop by not requeieing the vnode on LRU.
4062  *
4063  * Consider a case where vtryrecycle continuously fails with all vnodes (due to
4064  * e.g., frozen writes on the filesystem), filling the batch and causing it to
4065  * be requeued. Then vnlru will end up revisiting the same vnodes. This is a
4066  * loop which can last for as long as writes are frozen.
4067  */
4068 static void
4069 vdropl_recycle(struct vnode *vp)
4070 {
4071 
4072 	vdropl_impl(vp, false);
4073 }
4074 
4075 static void
4076 vdrop_recycle(struct vnode *vp)
4077 {
4078 
4079 	VI_LOCK(vp);
4080 	vdropl_recycle(vp);
4081 }
4082 
4083 /*
4084  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
4085  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
4086  */
4087 static int
4088 vinactivef(struct vnode *vp)
4089 {
4090 	int error;
4091 
4092 	ASSERT_VOP_ELOCKED(vp, "vinactive");
4093 	ASSERT_VI_LOCKED(vp, "vinactive");
4094 	VNPASS((vp->v_iflag & VI_DOINGINACT) == 0, vp);
4095 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4096 	vp->v_iflag |= VI_DOINGINACT;
4097 	vp->v_iflag &= ~VI_OWEINACT;
4098 	VI_UNLOCK(vp);
4099 
4100 	/*
4101 	 * Before moving off the active list, we must be sure that any
4102 	 * modified pages are converted into the vnode's dirty
4103 	 * buffers, since these will no longer be checked once the
4104 	 * vnode is on the inactive list.
4105 	 *
4106 	 * The write-out of the dirty pages is asynchronous.  At the
4107 	 * point that VOP_INACTIVE() is called, there could still be
4108 	 * pending I/O and dirty pages in the object.
4109 	 */
4110 	if ((vp->v_vflag & VV_NOSYNC) == 0)
4111 		vnode_pager_clean_async(vp);
4112 
4113 	error = VOP_INACTIVE(vp);
4114 	VI_LOCK(vp);
4115 	VNPASS(vp->v_iflag & VI_DOINGINACT, vp);
4116 	vp->v_iflag &= ~VI_DOINGINACT;
4117 	return (error);
4118 }
4119 
4120 int
4121 vinactive(struct vnode *vp)
4122 {
4123 
4124 	ASSERT_VOP_ELOCKED(vp, "vinactive");
4125 	ASSERT_VI_LOCKED(vp, "vinactive");
4126 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4127 
4128 	if ((vp->v_iflag & VI_OWEINACT) == 0)
4129 		return (0);
4130 	if (vp->v_iflag & VI_DOINGINACT)
4131 		return (0);
4132 	if (vp->v_usecount > 0) {
4133 		vp->v_iflag &= ~VI_OWEINACT;
4134 		return (0);
4135 	}
4136 	return (vinactivef(vp));
4137 }
4138 
4139 /*
4140  * Remove any vnodes in the vnode table belonging to mount point mp.
4141  *
4142  * If FORCECLOSE is not specified, there should not be any active ones,
4143  * return error if any are found (nb: this is a user error, not a
4144  * system error). If FORCECLOSE is specified, detach any active vnodes
4145  * that are found.
4146  *
4147  * If WRITECLOSE is set, only flush out regular file vnodes open for
4148  * writing.
4149  *
4150  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
4151  *
4152  * `rootrefs' specifies the base reference count for the root vnode
4153  * of this filesystem. The root vnode is considered busy if its
4154  * v_usecount exceeds this value. On a successful return, vflush(, td)
4155  * will call vrele() on the root vnode exactly rootrefs times.
4156  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
4157  * be zero.
4158  */
4159 #ifdef DIAGNOSTIC
4160 static int busyprt = 0;		/* print out busy vnodes */
4161 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
4162 #endif
4163 
4164 int
4165 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
4166 {
4167 	struct vnode *vp, *mvp, *rootvp = NULL;
4168 	struct vattr vattr;
4169 	int busy = 0, error;
4170 
4171 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
4172 	    rootrefs, flags);
4173 	if (rootrefs > 0) {
4174 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
4175 		    ("vflush: bad args"));
4176 		/*
4177 		 * Get the filesystem root vnode. We can vput() it
4178 		 * immediately, since with rootrefs > 0, it won't go away.
4179 		 */
4180 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
4181 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
4182 			    __func__, error);
4183 			return (error);
4184 		}
4185 		vput(rootvp);
4186 	}
4187 loop:
4188 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
4189 		vholdl(vp);
4190 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
4191 		if (error) {
4192 			vdrop(vp);
4193 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
4194 			goto loop;
4195 		}
4196 		/*
4197 		 * Skip over a vnodes marked VV_SYSTEM.
4198 		 */
4199 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
4200 			VOP_UNLOCK(vp);
4201 			vdrop(vp);
4202 			continue;
4203 		}
4204 		/*
4205 		 * If WRITECLOSE is set, flush out unlinked but still open
4206 		 * files (even if open only for reading) and regular file
4207 		 * vnodes open for writing.
4208 		 */
4209 		if (flags & WRITECLOSE) {
4210 			vnode_pager_clean_async(vp);
4211 			do {
4212 				error = VOP_FSYNC(vp, MNT_WAIT, td);
4213 			} while (error == ERELOOKUP);
4214 			if (error != 0) {
4215 				VOP_UNLOCK(vp);
4216 				vdrop(vp);
4217 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
4218 				return (error);
4219 			}
4220 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
4221 			VI_LOCK(vp);
4222 
4223 			if ((vp->v_type == VNON ||
4224 			    (error == 0 && vattr.va_nlink > 0)) &&
4225 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
4226 				VOP_UNLOCK(vp);
4227 				vdropl(vp);
4228 				continue;
4229 			}
4230 		} else
4231 			VI_LOCK(vp);
4232 		/*
4233 		 * With v_usecount == 0, all we need to do is clear out the
4234 		 * vnode data structures and we are done.
4235 		 *
4236 		 * If FORCECLOSE is set, forcibly close the vnode.
4237 		 */
4238 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
4239 			vgonel(vp);
4240 		} else {
4241 			busy++;
4242 #ifdef DIAGNOSTIC
4243 			if (busyprt)
4244 				vn_printf(vp, "vflush: busy vnode ");
4245 #endif
4246 		}
4247 		VOP_UNLOCK(vp);
4248 		vdropl(vp);
4249 	}
4250 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
4251 		/*
4252 		 * If just the root vnode is busy, and if its refcount
4253 		 * is equal to `rootrefs', then go ahead and kill it.
4254 		 */
4255 		VI_LOCK(rootvp);
4256 		KASSERT(busy > 0, ("vflush: not busy"));
4257 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
4258 		    ("vflush: usecount %d < rootrefs %d",
4259 		     rootvp->v_usecount, rootrefs));
4260 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
4261 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
4262 			vgone(rootvp);
4263 			VOP_UNLOCK(rootvp);
4264 			busy = 0;
4265 		} else
4266 			VI_UNLOCK(rootvp);
4267 	}
4268 	if (busy) {
4269 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
4270 		    busy);
4271 		return (EBUSY);
4272 	}
4273 	for (; rootrefs > 0; rootrefs--)
4274 		vrele(rootvp);
4275 	return (0);
4276 }
4277 
4278 /*
4279  * Recycle an unused vnode.
4280  */
4281 int
4282 vrecycle(struct vnode *vp)
4283 {
4284 	int recycled;
4285 
4286 	VI_LOCK(vp);
4287 	recycled = vrecyclel(vp);
4288 	VI_UNLOCK(vp);
4289 	return (recycled);
4290 }
4291 
4292 /*
4293  * vrecycle, with the vp interlock held.
4294  */
4295 int
4296 vrecyclel(struct vnode *vp)
4297 {
4298 	int recycled;
4299 
4300 	ASSERT_VOP_ELOCKED(vp, __func__);
4301 	ASSERT_VI_LOCKED(vp, __func__);
4302 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4303 	recycled = 0;
4304 	if (vp->v_usecount == 0) {
4305 		recycled = 1;
4306 		vgonel(vp);
4307 	}
4308 	return (recycled);
4309 }
4310 
4311 /*
4312  * Eliminate all activity associated with a vnode
4313  * in preparation for reuse.
4314  */
4315 void
4316 vgone(struct vnode *vp)
4317 {
4318 	VI_LOCK(vp);
4319 	vgonel(vp);
4320 	VI_UNLOCK(vp);
4321 }
4322 
4323 /*
4324  * Notify upper mounts about reclaimed or unlinked vnode.
4325  */
4326 void
4327 vfs_notify_upper(struct vnode *vp, enum vfs_notify_upper_type event)
4328 {
4329 	struct mount *mp;
4330 	struct mount_upper_node *ump;
4331 
4332 	mp = atomic_load_ptr(&vp->v_mount);
4333 	if (mp == NULL)
4334 		return;
4335 	if (TAILQ_EMPTY(&mp->mnt_notify))
4336 		return;
4337 
4338 	MNT_ILOCK(mp);
4339 	mp->mnt_upper_pending++;
4340 	KASSERT(mp->mnt_upper_pending > 0,
4341 	    ("%s: mnt_upper_pending %d", __func__, mp->mnt_upper_pending));
4342 	TAILQ_FOREACH(ump, &mp->mnt_notify, mnt_upper_link) {
4343 		MNT_IUNLOCK(mp);
4344 		switch (event) {
4345 		case VFS_NOTIFY_UPPER_RECLAIM:
4346 			VFS_RECLAIM_LOWERVP(ump->mp, vp);
4347 			break;
4348 		case VFS_NOTIFY_UPPER_UNLINK:
4349 			VFS_UNLINK_LOWERVP(ump->mp, vp);
4350 			break;
4351 		}
4352 		MNT_ILOCK(mp);
4353 	}
4354 	mp->mnt_upper_pending--;
4355 	if ((mp->mnt_kern_flag & MNTK_UPPER_WAITER) != 0 &&
4356 	    mp->mnt_upper_pending == 0) {
4357 		mp->mnt_kern_flag &= ~MNTK_UPPER_WAITER;
4358 		wakeup(&mp->mnt_uppers);
4359 	}
4360 	MNT_IUNLOCK(mp);
4361 }
4362 
4363 /*
4364  * vgone, with the vp interlock held.
4365  */
4366 static void
4367 vgonel(struct vnode *vp)
4368 {
4369 	struct thread *td;
4370 	struct mount *mp;
4371 	vm_object_t object;
4372 	bool active, doinginact, oweinact;
4373 
4374 	ASSERT_VOP_ELOCKED(vp, "vgonel");
4375 	ASSERT_VI_LOCKED(vp, "vgonel");
4376 	VNASSERT(vp->v_holdcnt, vp,
4377 	    ("vgonel: vp %p has no reference.", vp));
4378 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4379 	td = curthread;
4380 
4381 	/*
4382 	 * Don't vgonel if we're already doomed.
4383 	 */
4384 	if (VN_IS_DOOMED(vp)) {
4385 		VNPASS(vn_get_state(vp) == VSTATE_DESTROYING || \
4386 		    vn_get_state(vp) == VSTATE_DEAD, vp);
4387 		return;
4388 	}
4389 	/*
4390 	 * Paired with freevnode.
4391 	 */
4392 	vn_seqc_write_begin_locked(vp);
4393 	vunlazy_gone(vp);
4394 	vn_irflag_set_locked(vp, VIRF_DOOMED);
4395 	vn_set_state(vp, VSTATE_DESTROYING);
4396 
4397 	/*
4398 	 * Check to see if the vnode is in use.  If so, we have to
4399 	 * call VOP_CLOSE() and VOP_INACTIVE().
4400 	 *
4401 	 * It could be that VOP_INACTIVE() requested reclamation, in
4402 	 * which case we should avoid recursion, so check
4403 	 * VI_DOINGINACT.  This is not precise but good enough.
4404 	 */
4405 	active = vp->v_usecount > 0;
4406 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4407 	doinginact = (vp->v_iflag & VI_DOINGINACT) != 0;
4408 
4409 	/*
4410 	 * If we need to do inactive VI_OWEINACT will be set.
4411 	 */
4412 	if (vp->v_iflag & VI_DEFINACT) {
4413 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
4414 		vp->v_iflag &= ~VI_DEFINACT;
4415 		vdropl(vp);
4416 	} else {
4417 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
4418 		VI_UNLOCK(vp);
4419 	}
4420 	cache_purge_vgone(vp);
4421 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
4422 
4423 	/*
4424 	 * If purging an active vnode, it must be closed and
4425 	 * deactivated before being reclaimed.
4426 	 */
4427 	if (active)
4428 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
4429 	if (!doinginact) {
4430 		do {
4431 			if (oweinact || active) {
4432 				VI_LOCK(vp);
4433 				vinactivef(vp);
4434 				oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4435 				VI_UNLOCK(vp);
4436 			}
4437 		} while (oweinact);
4438 	}
4439 	if (vp->v_type == VSOCK)
4440 		vfs_unp_reclaim(vp);
4441 
4442 	/*
4443 	 * Clean out any buffers associated with the vnode.
4444 	 * If the flush fails, just toss the buffers.
4445 	 */
4446 	mp = NULL;
4447 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
4448 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
4449 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
4450 		while (vinvalbuf(vp, 0, 0, 0) != 0)
4451 			;
4452 	}
4453 
4454 	BO_LOCK(&vp->v_bufobj);
4455 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
4456 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
4457 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
4458 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
4459 	    ("vp %p bufobj not invalidated", vp));
4460 
4461 	/*
4462 	 * For VMIO bufobj, BO_DEAD is set later, or in
4463 	 * vm_object_terminate() after the object's page queue is
4464 	 * flushed.
4465 	 */
4466 	object = vp->v_bufobj.bo_object;
4467 	if (object == NULL)
4468 		vp->v_bufobj.bo_flag |= BO_DEAD;
4469 	BO_UNLOCK(&vp->v_bufobj);
4470 
4471 	/*
4472 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
4473 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
4474 	 * should not touch the object borrowed from the lower vnode
4475 	 * (the handle check).
4476 	 */
4477 	if (object != NULL && object->type == OBJT_VNODE &&
4478 	    object->handle == vp)
4479 		vnode_destroy_vobject(vp);
4480 
4481 	/*
4482 	 * Reclaim the vnode.
4483 	 */
4484 	if (VOP_RECLAIM(vp))
4485 		panic("vgone: cannot reclaim");
4486 	if (mp != NULL)
4487 		vn_finished_secondary_write(mp);
4488 	VNASSERT(vp->v_object == NULL, vp,
4489 	    ("vop_reclaim left v_object vp=%p", vp));
4490 	/*
4491 	 * Clear the advisory locks and wake up waiting threads.
4492 	 */
4493 	if (vp->v_lockf != NULL) {
4494 		(void)VOP_ADVLOCKPURGE(vp);
4495 		vp->v_lockf = NULL;
4496 	}
4497 	/*
4498 	 * Delete from old mount point vnode list.
4499 	 */
4500 	if (vp->v_mount == NULL) {
4501 		VI_LOCK(vp);
4502 	} else {
4503 		delmntque(vp);
4504 		ASSERT_VI_LOCKED(vp, "vgonel 2");
4505 	}
4506 	/*
4507 	 * Done with purge, reset to the standard lock and invalidate
4508 	 * the vnode.
4509 	 */
4510 	vp->v_vnlock = &vp->v_lock;
4511 	vp->v_op = &dead_vnodeops;
4512 	vp->v_type = VBAD;
4513 	vn_set_state(vp, VSTATE_DEAD);
4514 }
4515 
4516 /*
4517  * Print out a description of a vnode.
4518  */
4519 static const char *const vtypename[] = {
4520 	[VNON] = "VNON",
4521 	[VREG] = "VREG",
4522 	[VDIR] = "VDIR",
4523 	[VBLK] = "VBLK",
4524 	[VCHR] = "VCHR",
4525 	[VLNK] = "VLNK",
4526 	[VSOCK] = "VSOCK",
4527 	[VFIFO] = "VFIFO",
4528 	[VBAD] = "VBAD",
4529 	[VMARKER] = "VMARKER",
4530 };
4531 _Static_assert(nitems(vtypename) == VLASTTYPE + 1,
4532     "vnode type name not added to vtypename");
4533 
4534 static const char *const vstatename[] = {
4535 	[VSTATE_UNINITIALIZED] = "VSTATE_UNINITIALIZED",
4536 	[VSTATE_CONSTRUCTED] = "VSTATE_CONSTRUCTED",
4537 	[VSTATE_DESTROYING] = "VSTATE_DESTROYING",
4538 	[VSTATE_DEAD] = "VSTATE_DEAD",
4539 };
4540 _Static_assert(nitems(vstatename) == VLASTSTATE + 1,
4541     "vnode state name not added to vstatename");
4542 
4543 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0,
4544     "new hold count flag not added to vn_printf");
4545 
4546 void
4547 vn_printf(struct vnode *vp, const char *fmt, ...)
4548 {
4549 	va_list ap;
4550 	char buf[256], buf2[16];
4551 	u_long flags;
4552 	u_int holdcnt;
4553 	short irflag;
4554 
4555 	va_start(ap, fmt);
4556 	vprintf(fmt, ap);
4557 	va_end(ap);
4558 	printf("%p: ", (void *)vp);
4559 	printf("type %s state %s op %p\n", vtypename[vp->v_type],
4560 	    vstatename[vp->v_state], vp->v_op);
4561 	holdcnt = atomic_load_int(&vp->v_holdcnt);
4562 	printf("    usecount %d, writecount %d, refcount %d seqc users %d",
4563 	    vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS,
4564 	    vp->v_seqc_users);
4565 	switch (vp->v_type) {
4566 	case VDIR:
4567 		printf(" mountedhere %p\n", vp->v_mountedhere);
4568 		break;
4569 	case VCHR:
4570 		printf(" rdev %p\n", vp->v_rdev);
4571 		break;
4572 	case VSOCK:
4573 		printf(" socket %p\n", vp->v_unpcb);
4574 		break;
4575 	case VFIFO:
4576 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
4577 		break;
4578 	default:
4579 		printf("\n");
4580 		break;
4581 	}
4582 	buf[0] = '\0';
4583 	buf[1] = '\0';
4584 	if (holdcnt & VHOLD_NO_SMR)
4585 		strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf));
4586 	printf("    hold count flags (%s)\n", buf + 1);
4587 
4588 	buf[0] = '\0';
4589 	buf[1] = '\0';
4590 	irflag = vn_irflag_read(vp);
4591 	if (irflag & VIRF_DOOMED)
4592 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4593 	if (irflag & VIRF_PGREAD)
4594 		strlcat(buf, "|VIRF_PGREAD", sizeof(buf));
4595 	if (irflag & VIRF_MOUNTPOINT)
4596 		strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf));
4597 	if (irflag & VIRF_TEXT_REF)
4598 		strlcat(buf, "|VIRF_TEXT_REF", sizeof(buf));
4599 	flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT | VIRF_TEXT_REF);
4600 	if (flags != 0) {
4601 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4602 		strlcat(buf, buf2, sizeof(buf));
4603 	}
4604 	if (vp->v_vflag & VV_ROOT)
4605 		strlcat(buf, "|VV_ROOT", sizeof(buf));
4606 	if (vp->v_vflag & VV_ISTTY)
4607 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
4608 	if (vp->v_vflag & VV_NOSYNC)
4609 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4610 	if (vp->v_vflag & VV_ETERNALDEV)
4611 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4612 	if (vp->v_vflag & VV_CACHEDLABEL)
4613 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4614 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
4615 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4616 	if (vp->v_vflag & VV_COPYONWRITE)
4617 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4618 	if (vp->v_vflag & VV_SYSTEM)
4619 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4620 	if (vp->v_vflag & VV_PROCDEP)
4621 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4622 	if (vp->v_vflag & VV_DELETED)
4623 		strlcat(buf, "|VV_DELETED", sizeof(buf));
4624 	if (vp->v_vflag & VV_MD)
4625 		strlcat(buf, "|VV_MD", sizeof(buf));
4626 	if (vp->v_vflag & VV_FORCEINSMQ)
4627 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4628 	if (vp->v_vflag & VV_READLINK)
4629 		strlcat(buf, "|VV_READLINK", sizeof(buf));
4630 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4631 	    VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM |
4632 	    VV_PROCDEP | VV_DELETED | VV_MD | VV_FORCEINSMQ | VV_READLINK);
4633 	if (flags != 0) {
4634 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4635 		strlcat(buf, buf2, sizeof(buf));
4636 	}
4637 	if (vp->v_iflag & VI_MOUNT)
4638 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
4639 	if (vp->v_iflag & VI_DOINGINACT)
4640 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4641 	if (vp->v_iflag & VI_OWEINACT)
4642 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4643 	if (vp->v_iflag & VI_DEFINACT)
4644 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4645 	if (vp->v_iflag & VI_FOPENING)
4646 		strlcat(buf, "|VI_FOPENING", sizeof(buf));
4647 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOINGINACT |
4648 	    VI_OWEINACT | VI_DEFINACT | VI_FOPENING);
4649 	if (flags != 0) {
4650 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4651 		strlcat(buf, buf2, sizeof(buf));
4652 	}
4653 	if (vp->v_mflag & VMP_LAZYLIST)
4654 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4655 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
4656 	if (flags != 0) {
4657 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4658 		strlcat(buf, buf2, sizeof(buf));
4659 	}
4660 	printf("    flags (%s)", buf + 1);
4661 	if (mtx_owned(VI_MTX(vp)))
4662 		printf(" VI_LOCKed");
4663 	printf("\n");
4664 	if (vp->v_object != NULL)
4665 		printf("    v_object %p ref %d pages %d "
4666 		    "cleanbuf %d dirtybuf %d\n",
4667 		    vp->v_object, vp->v_object->ref_count,
4668 		    vp->v_object->resident_page_count,
4669 		    vp->v_bufobj.bo_clean.bv_cnt,
4670 		    vp->v_bufobj.bo_dirty.bv_cnt);
4671 	printf("    ");
4672 	lockmgr_printinfo(vp->v_vnlock);
4673 	if (vp->v_data != NULL)
4674 		VOP_PRINT(vp);
4675 }
4676 
4677 #ifdef DDB
4678 /*
4679  * List all of the locked vnodes in the system.
4680  * Called when debugging the kernel.
4681  */
4682 DB_SHOW_COMMAND_FLAGS(lockedvnods, lockedvnodes, DB_CMD_MEMSAFE)
4683 {
4684 	struct mount *mp;
4685 	struct vnode *vp;
4686 
4687 	/*
4688 	 * Note: because this is DDB, we can't obey the locking semantics
4689 	 * for these structures, which means we could catch an inconsistent
4690 	 * state and dereference a nasty pointer.  Not much to be done
4691 	 * about that.
4692 	 */
4693 	db_printf("Locked vnodes\n");
4694 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4695 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4696 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4697 				vn_printf(vp, "vnode ");
4698 		}
4699 	}
4700 }
4701 
4702 /*
4703  * Show details about the given vnode.
4704  */
4705 DB_SHOW_COMMAND(vnode, db_show_vnode)
4706 {
4707 	struct vnode *vp;
4708 
4709 	if (!have_addr)
4710 		return;
4711 	vp = (struct vnode *)addr;
4712 	vn_printf(vp, "vnode ");
4713 }
4714 
4715 /*
4716  * Show details about the given mount point.
4717  */
4718 DB_SHOW_COMMAND(mount, db_show_mount)
4719 {
4720 	struct mount *mp;
4721 	struct vfsopt *opt;
4722 	struct statfs *sp;
4723 	struct vnode *vp;
4724 	char buf[512];
4725 	uint64_t mflags;
4726 	u_int flags;
4727 
4728 	if (!have_addr) {
4729 		/* No address given, print short info about all mount points. */
4730 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4731 			db_printf("%p %s on %s (%s)\n", mp,
4732 			    mp->mnt_stat.f_mntfromname,
4733 			    mp->mnt_stat.f_mntonname,
4734 			    mp->mnt_stat.f_fstypename);
4735 			if (db_pager_quit)
4736 				break;
4737 		}
4738 		db_printf("\nMore info: show mount <addr>\n");
4739 		return;
4740 	}
4741 
4742 	mp = (struct mount *)addr;
4743 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4744 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4745 
4746 	buf[0] = '\0';
4747 	mflags = mp->mnt_flag;
4748 #define	MNT_FLAG(flag)	do {						\
4749 	if (mflags & (flag)) {						\
4750 		if (buf[0] != '\0')					\
4751 			strlcat(buf, ", ", sizeof(buf));		\
4752 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4753 		mflags &= ~(flag);					\
4754 	}								\
4755 } while (0)
4756 	MNT_FLAG(MNT_RDONLY);
4757 	MNT_FLAG(MNT_SYNCHRONOUS);
4758 	MNT_FLAG(MNT_NOEXEC);
4759 	MNT_FLAG(MNT_NOSUID);
4760 	MNT_FLAG(MNT_NFS4ACLS);
4761 	MNT_FLAG(MNT_UNION);
4762 	MNT_FLAG(MNT_ASYNC);
4763 	MNT_FLAG(MNT_SUIDDIR);
4764 	MNT_FLAG(MNT_SOFTDEP);
4765 	MNT_FLAG(MNT_NOSYMFOLLOW);
4766 	MNT_FLAG(MNT_GJOURNAL);
4767 	MNT_FLAG(MNT_MULTILABEL);
4768 	MNT_FLAG(MNT_ACLS);
4769 	MNT_FLAG(MNT_NOATIME);
4770 	MNT_FLAG(MNT_NOCLUSTERR);
4771 	MNT_FLAG(MNT_NOCLUSTERW);
4772 	MNT_FLAG(MNT_SUJ);
4773 	MNT_FLAG(MNT_EXRDONLY);
4774 	MNT_FLAG(MNT_EXPORTED);
4775 	MNT_FLAG(MNT_DEFEXPORTED);
4776 	MNT_FLAG(MNT_EXPORTANON);
4777 	MNT_FLAG(MNT_EXKERB);
4778 	MNT_FLAG(MNT_EXPUBLIC);
4779 	MNT_FLAG(MNT_LOCAL);
4780 	MNT_FLAG(MNT_QUOTA);
4781 	MNT_FLAG(MNT_ROOTFS);
4782 	MNT_FLAG(MNT_USER);
4783 	MNT_FLAG(MNT_IGNORE);
4784 	MNT_FLAG(MNT_UPDATE);
4785 	MNT_FLAG(MNT_DELEXPORT);
4786 	MNT_FLAG(MNT_RELOAD);
4787 	MNT_FLAG(MNT_FORCE);
4788 	MNT_FLAG(MNT_SNAPSHOT);
4789 	MNT_FLAG(MNT_BYFSID);
4790 	MNT_FLAG(MNT_NAMEDATTR);
4791 #undef MNT_FLAG
4792 	if (mflags != 0) {
4793 		if (buf[0] != '\0')
4794 			strlcat(buf, ", ", sizeof(buf));
4795 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4796 		    "0x%016jx", mflags);
4797 	}
4798 	db_printf("    mnt_flag = %s\n", buf);
4799 
4800 	buf[0] = '\0';
4801 	flags = mp->mnt_kern_flag;
4802 #define	MNT_KERN_FLAG(flag)	do {					\
4803 	if (flags & (flag)) {						\
4804 		if (buf[0] != '\0')					\
4805 			strlcat(buf, ", ", sizeof(buf));		\
4806 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4807 		flags &= ~(flag);					\
4808 	}								\
4809 } while (0)
4810 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4811 	MNT_KERN_FLAG(MNTK_ASYNC);
4812 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4813 	MNT_KERN_FLAG(MNTK_NOMSYNC);
4814 	MNT_KERN_FLAG(MNTK_DRAINING);
4815 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4816 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4817 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4818 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4819 	MNT_KERN_FLAG(MNTK_RECURSE);
4820 	MNT_KERN_FLAG(MNTK_UPPER_WAITER);
4821 	MNT_KERN_FLAG(MNTK_UNLOCKED_INSMNTQUE);
4822 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4823 	MNT_KERN_FLAG(MNTK_VMSETSIZE_BUG);
4824 	MNT_KERN_FLAG(MNTK_FPLOOKUP);
4825 	MNT_KERN_FLAG(MNTK_TASKQUEUE_WAITER);
4826 	MNT_KERN_FLAG(MNTK_NOASYNC);
4827 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4828 	MNT_KERN_FLAG(MNTK_MWAIT);
4829 	MNT_KERN_FLAG(MNTK_SUSPEND);
4830 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4831 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4832 	MNT_KERN_FLAG(MNTK_NULL_NOCACHE);
4833 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4834 #undef MNT_KERN_FLAG
4835 	if (flags != 0) {
4836 		if (buf[0] != '\0')
4837 			strlcat(buf, ", ", sizeof(buf));
4838 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4839 		    "0x%08x", flags);
4840 	}
4841 	db_printf("    mnt_kern_flag = %s\n", buf);
4842 
4843 	db_printf("    mnt_opt = ");
4844 	opt = TAILQ_FIRST(mp->mnt_opt);
4845 	if (opt != NULL) {
4846 		db_printf("%s", opt->name);
4847 		opt = TAILQ_NEXT(opt, link);
4848 		while (opt != NULL) {
4849 			db_printf(", %s", opt->name);
4850 			opt = TAILQ_NEXT(opt, link);
4851 		}
4852 	}
4853 	db_printf("\n");
4854 
4855 	sp = &mp->mnt_stat;
4856 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4857 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4858 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4859 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4860 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4861 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4862 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4863 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4864 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4865 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4866 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4867 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4868 
4869 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4870 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4871 	if (jailed(mp->mnt_cred))
4872 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4873 	db_printf(" }\n");
4874 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4875 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4876 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4877 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4878 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4879 	    mp->mnt_lazyvnodelistsize);
4880 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4881 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4882 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4883 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4884 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4885 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4886 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4887 	db_printf("    mnt_secondary_accwrites = %d\n",
4888 	    mp->mnt_secondary_accwrites);
4889 	db_printf("    mnt_gjprovider = %s\n",
4890 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4891 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4892 
4893 	db_printf("\n\nList of active vnodes\n");
4894 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4895 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4896 			vn_printf(vp, "vnode ");
4897 			if (db_pager_quit)
4898 				break;
4899 		}
4900 	}
4901 	db_printf("\n\nList of inactive vnodes\n");
4902 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4903 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4904 			vn_printf(vp, "vnode ");
4905 			if (db_pager_quit)
4906 				break;
4907 		}
4908 	}
4909 }
4910 #endif	/* DDB */
4911 
4912 /*
4913  * Fill in a struct xvfsconf based on a struct vfsconf.
4914  */
4915 static int
4916 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4917 {
4918 	struct xvfsconf xvfsp;
4919 
4920 	bzero(&xvfsp, sizeof(xvfsp));
4921 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4922 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4923 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4924 	xvfsp.vfc_flags = vfsp->vfc_flags;
4925 	/*
4926 	 * These are unused in userland, we keep them
4927 	 * to not break binary compatibility.
4928 	 */
4929 	xvfsp.vfc_vfsops = NULL;
4930 	xvfsp.vfc_next = NULL;
4931 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4932 }
4933 
4934 #ifdef COMPAT_FREEBSD32
4935 struct xvfsconf32 {
4936 	uint32_t	vfc_vfsops;
4937 	char		vfc_name[MFSNAMELEN];
4938 	int32_t		vfc_typenum;
4939 	int32_t		vfc_refcount;
4940 	int32_t		vfc_flags;
4941 	uint32_t	vfc_next;
4942 };
4943 
4944 static int
4945 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4946 {
4947 	struct xvfsconf32 xvfsp;
4948 
4949 	bzero(&xvfsp, sizeof(xvfsp));
4950 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4951 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4952 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4953 	xvfsp.vfc_flags = vfsp->vfc_flags;
4954 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4955 }
4956 #endif
4957 
4958 /*
4959  * Top level filesystem related information gathering.
4960  */
4961 static int
4962 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4963 {
4964 	struct vfsconf *vfsp;
4965 	int error;
4966 
4967 	error = 0;
4968 	vfsconf_slock();
4969 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4970 #ifdef COMPAT_FREEBSD32
4971 		if (req->flags & SCTL_MASK32)
4972 			error = vfsconf2x32(req, vfsp);
4973 		else
4974 #endif
4975 			error = vfsconf2x(req, vfsp);
4976 		if (error)
4977 			break;
4978 	}
4979 	vfsconf_sunlock();
4980 	return (error);
4981 }
4982 
4983 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4984     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4985     "S,xvfsconf", "List of all configured filesystems");
4986 
4987 #ifndef BURN_BRIDGES
4988 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4989 
4990 static int
4991 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4992 {
4993 	int *name = (int *)arg1 - 1;	/* XXX */
4994 	u_int namelen = arg2 + 1;	/* XXX */
4995 	struct vfsconf *vfsp;
4996 
4997 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4998 	    "please rebuild world\n");
4999 
5000 #if 1 || defined(COMPAT_PRELITE2)
5001 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
5002 	if (namelen == 1)
5003 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
5004 #endif
5005 
5006 	switch (name[1]) {
5007 	case VFS_MAXTYPENUM:
5008 		if (namelen != 2)
5009 			return (ENOTDIR);
5010 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
5011 	case VFS_CONF:
5012 		if (namelen != 3)
5013 			return (ENOTDIR);	/* overloaded */
5014 		vfsconf_slock();
5015 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
5016 			if (vfsp->vfc_typenum == name[2])
5017 				break;
5018 		}
5019 		vfsconf_sunlock();
5020 		if (vfsp == NULL)
5021 			return (EOPNOTSUPP);
5022 #ifdef COMPAT_FREEBSD32
5023 		if (req->flags & SCTL_MASK32)
5024 			return (vfsconf2x32(req, vfsp));
5025 		else
5026 #endif
5027 			return (vfsconf2x(req, vfsp));
5028 	}
5029 	return (EOPNOTSUPP);
5030 }
5031 
5032 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
5033     CTLFLAG_MPSAFE, vfs_sysctl,
5034     "Generic filesystem");
5035 
5036 #if 1 || defined(COMPAT_PRELITE2)
5037 
5038 static int
5039 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
5040 {
5041 	int error;
5042 	struct vfsconf *vfsp;
5043 	struct ovfsconf ovfs;
5044 
5045 	vfsconf_slock();
5046 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
5047 		bzero(&ovfs, sizeof(ovfs));
5048 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
5049 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
5050 		ovfs.vfc_index = vfsp->vfc_typenum;
5051 		ovfs.vfc_refcount = vfsp->vfc_refcount;
5052 		ovfs.vfc_flags = vfsp->vfc_flags;
5053 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
5054 		if (error != 0) {
5055 			vfsconf_sunlock();
5056 			return (error);
5057 		}
5058 	}
5059 	vfsconf_sunlock();
5060 	return (0);
5061 }
5062 
5063 #endif /* 1 || COMPAT_PRELITE2 */
5064 #endif /* !BURN_BRIDGES */
5065 
5066 static void
5067 unmount_or_warn(struct mount *mp)
5068 {
5069 	int error;
5070 
5071 	error = dounmount(mp, MNT_FORCE, curthread);
5072 	if (error != 0) {
5073 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
5074 		if (error == EBUSY)
5075 			printf("BUSY)\n");
5076 		else
5077 			printf("%d)\n", error);
5078 	}
5079 }
5080 
5081 /*
5082  * Unmount all filesystems. The list is traversed in reverse order
5083  * of mounting to avoid dependencies.
5084  */
5085 void
5086 vfs_unmountall(void)
5087 {
5088 	struct mount *mp, *tmp;
5089 
5090 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
5091 
5092 	/*
5093 	 * Since this only runs when rebooting, it is not interlocked.
5094 	 */
5095 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
5096 		vfs_ref(mp);
5097 
5098 		/*
5099 		 * Forcibly unmounting "/dev" before "/" would prevent clean
5100 		 * unmount of the latter.
5101 		 */
5102 		if (mp == rootdevmp)
5103 			continue;
5104 
5105 		unmount_or_warn(mp);
5106 	}
5107 
5108 	if (rootdevmp != NULL)
5109 		unmount_or_warn(rootdevmp);
5110 }
5111 
5112 static void
5113 vfs_deferred_inactive(struct vnode *vp, int lkflags)
5114 {
5115 
5116 	ASSERT_VI_LOCKED(vp, __func__);
5117 	VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
5118 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
5119 		vdropl(vp);
5120 		return;
5121 	}
5122 	if (vn_lock(vp, lkflags) == 0) {
5123 		VI_LOCK(vp);
5124 		vinactive(vp);
5125 		VOP_UNLOCK(vp);
5126 		vdropl(vp);
5127 		return;
5128 	}
5129 	vdefer_inactive_unlocked(vp);
5130 }
5131 
5132 static int
5133 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
5134 {
5135 
5136 	return (vp->v_iflag & VI_DEFINACT);
5137 }
5138 
5139 static void __noinline
5140 vfs_periodic_inactive(struct mount *mp, int flags)
5141 {
5142 	struct vnode *vp, *mvp;
5143 	int lkflags;
5144 
5145 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
5146 	if (flags != MNT_WAIT)
5147 		lkflags |= LK_NOWAIT;
5148 
5149 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
5150 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
5151 			VI_UNLOCK(vp);
5152 			continue;
5153 		}
5154 		vp->v_iflag &= ~VI_DEFINACT;
5155 		vfs_deferred_inactive(vp, lkflags);
5156 	}
5157 }
5158 
5159 static inline bool
5160 vfs_want_msync(struct vnode *vp)
5161 {
5162 	struct vm_object *obj;
5163 
5164 	/*
5165 	 * This test may be performed without any locks held.
5166 	 * We rely on vm_object's type stability.
5167 	 */
5168 	if (vp->v_vflag & VV_NOSYNC)
5169 		return (false);
5170 	obj = vp->v_object;
5171 	return (obj != NULL && vm_object_mightbedirty(obj));
5172 }
5173 
5174 static int
5175 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
5176 {
5177 
5178 	if (vp->v_vflag & VV_NOSYNC)
5179 		return (false);
5180 	if (vp->v_iflag & VI_DEFINACT)
5181 		return (true);
5182 	return (vfs_want_msync(vp));
5183 }
5184 
5185 static void __noinline
5186 vfs_periodic_msync_inactive(struct mount *mp, int flags)
5187 {
5188 	struct vnode *vp, *mvp;
5189 	int lkflags;
5190 	bool seen_defer;
5191 
5192 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
5193 	if (flags != MNT_WAIT)
5194 		lkflags |= LK_NOWAIT;
5195 
5196 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
5197 		seen_defer = false;
5198 		if (vp->v_iflag & VI_DEFINACT) {
5199 			vp->v_iflag &= ~VI_DEFINACT;
5200 			seen_defer = true;
5201 		}
5202 		if (!vfs_want_msync(vp)) {
5203 			if (seen_defer)
5204 				vfs_deferred_inactive(vp, lkflags);
5205 			else
5206 				VI_UNLOCK(vp);
5207 			continue;
5208 		}
5209 		if (vget(vp, lkflags) == 0) {
5210 			if ((vp->v_vflag & VV_NOSYNC) == 0) {
5211 				if (flags == MNT_WAIT)
5212 					vnode_pager_clean_sync(vp);
5213 				else
5214 					vnode_pager_clean_async(vp);
5215 			}
5216 			vput(vp);
5217 			if (seen_defer)
5218 				vdrop(vp);
5219 		} else {
5220 			if (seen_defer)
5221 				vdefer_inactive_unlocked(vp);
5222 		}
5223 	}
5224 }
5225 
5226 void
5227 vfs_periodic(struct mount *mp, int flags)
5228 {
5229 
5230 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
5231 
5232 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
5233 		vfs_periodic_inactive(mp, flags);
5234 	else
5235 		vfs_periodic_msync_inactive(mp, flags);
5236 }
5237 
5238 static void
5239 destroy_vpollinfo_free(struct vpollinfo *vi)
5240 {
5241 
5242 	knlist_destroy(&vi->vpi_selinfo.si_note);
5243 	mtx_destroy(&vi->vpi_lock);
5244 	free(vi, M_VNODEPOLL);
5245 }
5246 
5247 static void
5248 destroy_vpollinfo(struct vpollinfo *vi)
5249 {
5250 
5251 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
5252 	seldrain(&vi->vpi_selinfo);
5253 	destroy_vpollinfo_free(vi);
5254 }
5255 
5256 /*
5257  * Initialize per-vnode helper structure to hold poll-related state.
5258  */
5259 void
5260 v_addpollinfo(struct vnode *vp)
5261 {
5262 	struct vpollinfo *vi;
5263 
5264 	if (vp->v_pollinfo != NULL)
5265 		return;
5266 	vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO);
5267 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
5268 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
5269 	    vfs_knlunlock, vfs_knl_assert_lock);
5270 	VI_LOCK(vp);
5271 	if (vp->v_pollinfo != NULL) {
5272 		VI_UNLOCK(vp);
5273 		destroy_vpollinfo_free(vi);
5274 		return;
5275 	}
5276 	vp->v_pollinfo = vi;
5277 	VI_UNLOCK(vp);
5278 }
5279 
5280 /*
5281  * Record a process's interest in events which might happen to
5282  * a vnode.  Because poll uses the historic select-style interface
5283  * internally, this routine serves as both the ``check for any
5284  * pending events'' and the ``record my interest in future events''
5285  * functions.  (These are done together, while the lock is held,
5286  * to avoid race conditions.)
5287  */
5288 int
5289 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
5290 {
5291 
5292 	v_addpollinfo(vp);
5293 	mtx_lock(&vp->v_pollinfo->vpi_lock);
5294 	if (vp->v_pollinfo->vpi_revents & events) {
5295 		/*
5296 		 * This leaves events we are not interested
5297 		 * in available for the other process which
5298 		 * which presumably had requested them
5299 		 * (otherwise they would never have been
5300 		 * recorded).
5301 		 */
5302 		events &= vp->v_pollinfo->vpi_revents;
5303 		vp->v_pollinfo->vpi_revents &= ~events;
5304 
5305 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
5306 		return (events);
5307 	}
5308 	vp->v_pollinfo->vpi_events |= events;
5309 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
5310 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
5311 	return (0);
5312 }
5313 
5314 /*
5315  * Routine to create and manage a filesystem syncer vnode.
5316  */
5317 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
5318 static int	sync_fsync(struct  vop_fsync_args *);
5319 static int	sync_inactive(struct  vop_inactive_args *);
5320 static int	sync_reclaim(struct  vop_reclaim_args *);
5321 
5322 static struct vop_vector sync_vnodeops = {
5323 	.vop_bypass =	VOP_EOPNOTSUPP,
5324 	.vop_close =	sync_close,
5325 	.vop_fsync =	sync_fsync,
5326 	.vop_getwritemount = vop_stdgetwritemount,
5327 	.vop_inactive =	sync_inactive,
5328 	.vop_need_inactive = vop_stdneed_inactive,
5329 	.vop_reclaim =	sync_reclaim,
5330 	.vop_lock1 =	vop_stdlock,
5331 	.vop_unlock =	vop_stdunlock,
5332 	.vop_islocked =	vop_stdislocked,
5333 	.vop_fplookup_vexec = VOP_EAGAIN,
5334 	.vop_fplookup_symlink = VOP_EAGAIN,
5335 };
5336 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
5337 
5338 /*
5339  * Create a new filesystem syncer vnode for the specified mount point.
5340  */
5341 void
5342 vfs_allocate_syncvnode(struct mount *mp)
5343 {
5344 	struct vnode *vp;
5345 	struct bufobj *bo;
5346 	static long start, incr, next;
5347 	int error;
5348 
5349 	/* Allocate a new vnode */
5350 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
5351 	if (error != 0)
5352 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
5353 	vp->v_type = VNON;
5354 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5355 	vp->v_vflag |= VV_FORCEINSMQ;
5356 	error = insmntque1(vp, mp);
5357 	if (error != 0)
5358 		panic("vfs_allocate_syncvnode: insmntque() failed");
5359 	vp->v_vflag &= ~VV_FORCEINSMQ;
5360 	vn_set_state(vp, VSTATE_CONSTRUCTED);
5361 	VOP_UNLOCK(vp);
5362 	/*
5363 	 * Place the vnode onto the syncer worklist. We attempt to
5364 	 * scatter them about on the list so that they will go off
5365 	 * at evenly distributed times even if all the filesystems
5366 	 * are mounted at once.
5367 	 */
5368 	next += incr;
5369 	if (next == 0 || next > syncer_maxdelay) {
5370 		start /= 2;
5371 		incr /= 2;
5372 		if (start == 0) {
5373 			start = syncer_maxdelay / 2;
5374 			incr = syncer_maxdelay;
5375 		}
5376 		next = start;
5377 	}
5378 	bo = &vp->v_bufobj;
5379 	BO_LOCK(bo);
5380 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
5381 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
5382 	mtx_lock(&sync_mtx);
5383 	sync_vnode_count++;
5384 	if (mp->mnt_syncer == NULL) {
5385 		mp->mnt_syncer = vp;
5386 		vp = NULL;
5387 	}
5388 	mtx_unlock(&sync_mtx);
5389 	BO_UNLOCK(bo);
5390 	if (vp != NULL) {
5391 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5392 		vgone(vp);
5393 		vput(vp);
5394 	}
5395 }
5396 
5397 void
5398 vfs_deallocate_syncvnode(struct mount *mp)
5399 {
5400 	struct vnode *vp;
5401 
5402 	mtx_lock(&sync_mtx);
5403 	vp = mp->mnt_syncer;
5404 	if (vp != NULL)
5405 		mp->mnt_syncer = NULL;
5406 	mtx_unlock(&sync_mtx);
5407 	if (vp != NULL)
5408 		vrele(vp);
5409 }
5410 
5411 /*
5412  * Do a lazy sync of the filesystem.
5413  */
5414 static int
5415 sync_fsync(struct vop_fsync_args *ap)
5416 {
5417 	struct vnode *syncvp = ap->a_vp;
5418 	struct mount *mp = syncvp->v_mount;
5419 	int error, save;
5420 	struct bufobj *bo;
5421 
5422 	/*
5423 	 * We only need to do something if this is a lazy evaluation.
5424 	 */
5425 	if (ap->a_waitfor != MNT_LAZY)
5426 		return (0);
5427 
5428 	/*
5429 	 * Move ourselves to the back of the sync list.
5430 	 */
5431 	bo = &syncvp->v_bufobj;
5432 	BO_LOCK(bo);
5433 	vn_syncer_add_to_worklist(bo, syncdelay);
5434 	BO_UNLOCK(bo);
5435 
5436 	/*
5437 	 * Walk the list of vnodes pushing all that are dirty and
5438 	 * not already on the sync list.
5439 	 */
5440 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
5441 		return (0);
5442 	VOP_UNLOCK(syncvp);
5443 	save = curthread_pflags_set(TDP_SYNCIO);
5444 	/*
5445 	 * The filesystem at hand may be idle with free vnodes stored in the
5446 	 * batch.  Return them instead of letting them stay there indefinitely.
5447 	 */
5448 	vfs_periodic(mp, MNT_NOWAIT);
5449 	error = VFS_SYNC(mp, MNT_LAZY);
5450 	curthread_pflags_restore(save);
5451 	vn_lock(syncvp, LK_EXCLUSIVE | LK_RETRY);
5452 	vfs_unbusy(mp);
5453 	return (error);
5454 }
5455 
5456 /*
5457  * The syncer vnode is no referenced.
5458  */
5459 static int
5460 sync_inactive(struct vop_inactive_args *ap)
5461 {
5462 
5463 	vgone(ap->a_vp);
5464 	return (0);
5465 }
5466 
5467 /*
5468  * The syncer vnode is no longer needed and is being decommissioned.
5469  *
5470  * Modifications to the worklist must be protected by sync_mtx.
5471  */
5472 static int
5473 sync_reclaim(struct vop_reclaim_args *ap)
5474 {
5475 	struct vnode *vp = ap->a_vp;
5476 	struct bufobj *bo;
5477 
5478 	bo = &vp->v_bufobj;
5479 	BO_LOCK(bo);
5480 	mtx_lock(&sync_mtx);
5481 	if (vp->v_mount->mnt_syncer == vp)
5482 		vp->v_mount->mnt_syncer = NULL;
5483 	if (bo->bo_flag & BO_ONWORKLST) {
5484 		LIST_REMOVE(bo, bo_synclist);
5485 		syncer_worklist_len--;
5486 		sync_vnode_count--;
5487 		bo->bo_flag &= ~BO_ONWORKLST;
5488 	}
5489 	mtx_unlock(&sync_mtx);
5490 	BO_UNLOCK(bo);
5491 
5492 	return (0);
5493 }
5494 
5495 int
5496 vn_need_pageq_flush(struct vnode *vp)
5497 {
5498 	struct vm_object *obj;
5499 
5500 	obj = vp->v_object;
5501 	return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5502 	    vm_object_mightbedirty(obj));
5503 }
5504 
5505 /*
5506  * Check if vnode represents a disk device
5507  */
5508 bool
5509 vn_isdisk_error(struct vnode *vp, int *errp)
5510 {
5511 	int error;
5512 
5513 	if (vp->v_type != VCHR) {
5514 		error = ENOTBLK;
5515 		goto out;
5516 	}
5517 	error = 0;
5518 	dev_lock();
5519 	if (vp->v_rdev == NULL)
5520 		error = ENXIO;
5521 	else if (vp->v_rdev->si_devsw == NULL)
5522 		error = ENXIO;
5523 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5524 		error = ENOTBLK;
5525 	dev_unlock();
5526 out:
5527 	*errp = error;
5528 	return (error == 0);
5529 }
5530 
5531 bool
5532 vn_isdisk(struct vnode *vp)
5533 {
5534 	int error;
5535 
5536 	return (vn_isdisk_error(vp, &error));
5537 }
5538 
5539 /*
5540  * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
5541  * the comment above cache_fplookup for details.
5542  */
5543 int
5544 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred)
5545 {
5546 	int error;
5547 
5548 	VFS_SMR_ASSERT_ENTERED();
5549 
5550 	/* Check the owner. */
5551 	if (cred->cr_uid == file_uid) {
5552 		if (file_mode & S_IXUSR)
5553 			return (0);
5554 		goto out_error;
5555 	}
5556 
5557 	/* Otherwise, check the groups (first match) */
5558 	if (groupmember(file_gid, cred)) {
5559 		if (file_mode & S_IXGRP)
5560 			return (0);
5561 		goto out_error;
5562 	}
5563 
5564 	/* Otherwise, check everyone else. */
5565 	if (file_mode & S_IXOTH)
5566 		return (0);
5567 out_error:
5568 	/*
5569 	 * Permission check failed, but it is possible denial will get overwritten
5570 	 * (e.g., when root is traversing through a 700 directory owned by someone
5571 	 * else).
5572 	 *
5573 	 * vaccess() calls priv_check_cred which in turn can descent into MAC
5574 	 * modules overriding this result. It's quite unclear what semantics
5575 	 * are allowed for them to operate, thus for safety we don't call them
5576 	 * from within the SMR section. This also means if any such modules
5577 	 * are present, we have to let the regular lookup decide.
5578 	 */
5579 	error = priv_check_cred_vfs_lookup_nomac(cred);
5580 	switch (error) {
5581 	case 0:
5582 		return (0);
5583 	case EAGAIN:
5584 		/*
5585 		 * MAC modules present.
5586 		 */
5587 		return (EAGAIN);
5588 	case EPERM:
5589 		return (EACCES);
5590 	default:
5591 		return (error);
5592 	}
5593 }
5594 
5595 /*
5596  * Common filesystem object access control check routine.  Accepts a
5597  * vnode's type, "mode", uid and gid, requested access mode, and credentials.
5598  * Returns 0 on success, or an errno on failure.
5599  */
5600 int
5601 vaccess(__enum_uint8(vtype) type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5602     accmode_t accmode, struct ucred *cred)
5603 {
5604 	accmode_t dac_granted;
5605 	accmode_t priv_granted;
5606 
5607 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5608 	    ("invalid bit in accmode"));
5609 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5610 	    ("VAPPEND without VWRITE"));
5611 
5612 	/*
5613 	 * Look for a normal, non-privileged way to access the file/directory
5614 	 * as requested.  If it exists, go with that.
5615 	 */
5616 
5617 	dac_granted = 0;
5618 
5619 	/* Check the owner. */
5620 	if (cred->cr_uid == file_uid) {
5621 		dac_granted |= VADMIN;
5622 		if (file_mode & S_IXUSR)
5623 			dac_granted |= VEXEC;
5624 		if (file_mode & S_IRUSR)
5625 			dac_granted |= VREAD;
5626 		if (file_mode & S_IWUSR)
5627 			dac_granted |= (VWRITE | VAPPEND);
5628 
5629 		if ((accmode & dac_granted) == accmode)
5630 			return (0);
5631 
5632 		goto privcheck;
5633 	}
5634 
5635 	/* Otherwise, check the groups (first match) */
5636 	if (groupmember(file_gid, cred)) {
5637 		if (file_mode & S_IXGRP)
5638 			dac_granted |= VEXEC;
5639 		if (file_mode & S_IRGRP)
5640 			dac_granted |= VREAD;
5641 		if (file_mode & S_IWGRP)
5642 			dac_granted |= (VWRITE | VAPPEND);
5643 
5644 		if ((accmode & dac_granted) == accmode)
5645 			return (0);
5646 
5647 		goto privcheck;
5648 	}
5649 
5650 	/* Otherwise, check everyone else. */
5651 	if (file_mode & S_IXOTH)
5652 		dac_granted |= VEXEC;
5653 	if (file_mode & S_IROTH)
5654 		dac_granted |= VREAD;
5655 	if (file_mode & S_IWOTH)
5656 		dac_granted |= (VWRITE | VAPPEND);
5657 	if ((accmode & dac_granted) == accmode)
5658 		return (0);
5659 
5660 privcheck:
5661 	/*
5662 	 * Build a privilege mask to determine if the set of privileges
5663 	 * satisfies the requirements when combined with the granted mask
5664 	 * from above.  For each privilege, if the privilege is required,
5665 	 * bitwise or the request type onto the priv_granted mask.
5666 	 */
5667 	priv_granted = 0;
5668 
5669 	if (type == VDIR) {
5670 		/*
5671 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5672 		 * requests, instead of PRIV_VFS_EXEC.
5673 		 */
5674 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5675 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5676 			priv_granted |= VEXEC;
5677 	} else {
5678 		/*
5679 		 * Ensure that at least one execute bit is on. Otherwise,
5680 		 * a privileged user will always succeed, and we don't want
5681 		 * this to happen unless the file really is executable.
5682 		 */
5683 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5684 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5685 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5686 			priv_granted |= VEXEC;
5687 	}
5688 
5689 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5690 	    !priv_check_cred(cred, PRIV_VFS_READ))
5691 		priv_granted |= VREAD;
5692 
5693 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5694 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5695 		priv_granted |= (VWRITE | VAPPEND);
5696 
5697 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5698 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5699 		priv_granted |= VADMIN;
5700 
5701 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5702 		return (0);
5703 	}
5704 
5705 	return ((accmode & VADMIN) ? EPERM : EACCES);
5706 }
5707 
5708 /*
5709  * Credential check based on process requesting service, and per-attribute
5710  * permissions.
5711  */
5712 int
5713 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5714     struct thread *td, accmode_t accmode)
5715 {
5716 
5717 	/*
5718 	 * Kernel-invoked always succeeds.
5719 	 */
5720 	if (cred == NOCRED)
5721 		return (0);
5722 
5723 	/*
5724 	 * Do not allow privileged processes in jail to directly manipulate
5725 	 * system attributes.
5726 	 */
5727 	switch (attrnamespace) {
5728 	case EXTATTR_NAMESPACE_SYSTEM:
5729 		/* Potentially should be: return (EPERM); */
5730 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5731 	case EXTATTR_NAMESPACE_USER:
5732 		return (VOP_ACCESS(vp, accmode, cred, td));
5733 	default:
5734 		return (EPERM);
5735 	}
5736 }
5737 
5738 #ifdef DEBUG_VFS_LOCKS
5739 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5740 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5741     "Drop into debugger on lock violation");
5742 
5743 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5744 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5745     0, "Check for interlock across VOPs");
5746 
5747 int vfs_badlock_print = 1;	/* Print lock violations. */
5748 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5749     0, "Print lock violations");
5750 
5751 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5752 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5753     0, "Print vnode details on lock violations");
5754 
5755 #ifdef KDB
5756 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5757 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5758     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5759 #endif
5760 
5761 static void
5762 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5763 {
5764 
5765 #ifdef KDB
5766 	if (vfs_badlock_backtrace)
5767 		kdb_backtrace();
5768 #endif
5769 	if (vfs_badlock_vnode)
5770 		vn_printf(vp, "vnode ");
5771 	if (vfs_badlock_print)
5772 		printf("%s: %p %s\n", str, (void *)vp, msg);
5773 	if (vfs_badlock_ddb)
5774 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5775 }
5776 
5777 void
5778 assert_vi_locked(struct vnode *vp, const char *str)
5779 {
5780 
5781 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5782 		vfs_badlock("interlock is not locked but should be", str, vp);
5783 }
5784 
5785 void
5786 assert_vi_unlocked(struct vnode *vp, const char *str)
5787 {
5788 
5789 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5790 		vfs_badlock("interlock is locked but should not be", str, vp);
5791 }
5792 
5793 void
5794 assert_vop_locked(struct vnode *vp, const char *str)
5795 {
5796 	if (KERNEL_PANICKED() || vp == NULL)
5797 		return;
5798 
5799 #ifdef WITNESS
5800 	if ((vp->v_irflag & VIRF_CROSSMP) == 0 &&
5801 	    witness_is_owned(&vp->v_vnlock->lock_object) == -1)
5802 #else
5803 	int locked = VOP_ISLOCKED(vp);
5804 	if (locked == 0 || locked == LK_EXCLOTHER)
5805 #endif
5806 		vfs_badlock("is not locked but should be", str, vp);
5807 }
5808 
5809 void
5810 assert_vop_unlocked(struct vnode *vp, const char *str)
5811 {
5812 	if (KERNEL_PANICKED() || vp == NULL)
5813 		return;
5814 
5815 #ifdef WITNESS
5816 	if ((vp->v_irflag & VIRF_CROSSMP) == 0 &&
5817 	    witness_is_owned(&vp->v_vnlock->lock_object) == 1)
5818 #else
5819 	if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5820 #endif
5821 		vfs_badlock("is locked but should not be", str, vp);
5822 }
5823 
5824 void
5825 assert_vop_elocked(struct vnode *vp, const char *str)
5826 {
5827 	if (KERNEL_PANICKED() || vp == NULL)
5828 		return;
5829 
5830 	if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5831 		vfs_badlock("is not exclusive locked but should be", str, vp);
5832 }
5833 #endif /* DEBUG_VFS_LOCKS */
5834 
5835 void
5836 vop_rename_fail(struct vop_rename_args *ap)
5837 {
5838 
5839 	if (ap->a_tvp != NULL)
5840 		vput(ap->a_tvp);
5841 	if (ap->a_tdvp == ap->a_tvp)
5842 		vrele(ap->a_tdvp);
5843 	else
5844 		vput(ap->a_tdvp);
5845 	vrele(ap->a_fdvp);
5846 	vrele(ap->a_fvp);
5847 }
5848 
5849 void
5850 vop_rename_pre(void *ap)
5851 {
5852 	struct vop_rename_args *a = ap;
5853 
5854 #ifdef DEBUG_VFS_LOCKS
5855 	if (a->a_tvp)
5856 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5857 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5858 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5859 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5860 
5861 	/* Check the source (from). */
5862 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5863 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5864 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5865 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5866 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5867 
5868 	/* Check the target. */
5869 	if (a->a_tvp)
5870 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5871 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5872 #endif
5873 	/*
5874 	 * It may be tempting to add vn_seqc_write_begin/end calls here and
5875 	 * in vop_rename_post but that's not going to work out since some
5876 	 * filesystems relookup vnodes mid-rename. This is probably a bug.
5877 	 *
5878 	 * For now filesystems are expected to do the relevant calls after they
5879 	 * decide what vnodes to operate on.
5880 	 */
5881 	if (a->a_tdvp != a->a_fdvp)
5882 		vhold(a->a_fdvp);
5883 	if (a->a_tvp != a->a_fvp)
5884 		vhold(a->a_fvp);
5885 	vhold(a->a_tdvp);
5886 	if (a->a_tvp)
5887 		vhold(a->a_tvp);
5888 }
5889 
5890 #ifdef DEBUG_VFS_LOCKS
5891 void
5892 vop_fplookup_vexec_debugpre(void *ap __unused)
5893 {
5894 
5895 	VFS_SMR_ASSERT_ENTERED();
5896 }
5897 
5898 void
5899 vop_fplookup_vexec_debugpost(void *ap, int rc)
5900 {
5901 	struct vop_fplookup_vexec_args *a;
5902 	struct vnode *vp;
5903 
5904 	a = ap;
5905 	vp = a->a_vp;
5906 
5907 	VFS_SMR_ASSERT_ENTERED();
5908 	if (rc == EOPNOTSUPP)
5909 		VNPASS(VN_IS_DOOMED(vp), vp);
5910 }
5911 
5912 void
5913 vop_fplookup_symlink_debugpre(void *ap __unused)
5914 {
5915 
5916 	VFS_SMR_ASSERT_ENTERED();
5917 }
5918 
5919 void
5920 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused)
5921 {
5922 
5923 	VFS_SMR_ASSERT_ENTERED();
5924 }
5925 
5926 static void
5927 vop_fsync_debugprepost(struct vnode *vp, const char *name)
5928 {
5929 	if (vp->v_type == VCHR)
5930 		;
5931 	/*
5932 	 * The shared vs. exclusive locking policy for fsync()
5933 	 * is actually determined by vp's write mount as indicated
5934 	 * by VOP_GETWRITEMOUNT(), which for stacked filesystems
5935 	 * may not be the same as vp->v_mount.  However, if the
5936 	 * underlying filesystem which really handles the fsync()
5937 	 * supports shared locking, the stacked filesystem must also
5938 	 * be prepared for its VOP_FSYNC() operation to be called
5939 	 * with only a shared lock.  On the other hand, if the
5940 	 * stacked filesystem claims support for shared write
5941 	 * locking but the underlying filesystem does not, and the
5942 	 * caller incorrectly uses a shared lock, this condition
5943 	 * should still be caught when the stacked filesystem
5944 	 * invokes VOP_FSYNC() on the underlying filesystem.
5945 	 */
5946 	else if (MNT_SHARED_WRITES(vp->v_mount))
5947 		ASSERT_VOP_LOCKED(vp, name);
5948 	else
5949 		ASSERT_VOP_ELOCKED(vp, name);
5950 }
5951 
5952 void
5953 vop_fsync_debugpre(void *a)
5954 {
5955 	struct vop_fsync_args *ap;
5956 
5957 	ap = a;
5958 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5959 }
5960 
5961 void
5962 vop_fsync_debugpost(void *a, int rc __unused)
5963 {
5964 	struct vop_fsync_args *ap;
5965 
5966 	ap = a;
5967 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5968 }
5969 
5970 void
5971 vop_fdatasync_debugpre(void *a)
5972 {
5973 	struct vop_fdatasync_args *ap;
5974 
5975 	ap = a;
5976 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5977 }
5978 
5979 void
5980 vop_fdatasync_debugpost(void *a, int rc __unused)
5981 {
5982 	struct vop_fdatasync_args *ap;
5983 
5984 	ap = a;
5985 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5986 }
5987 
5988 void
5989 vop_strategy_debugpre(void *ap)
5990 {
5991 	struct vop_strategy_args *a;
5992 	struct buf *bp;
5993 
5994 	a = ap;
5995 	bp = a->a_bp;
5996 
5997 	/*
5998 	 * Cluster ops lock their component buffers but not the IO container.
5999 	 */
6000 	if ((bp->b_flags & B_CLUSTER) != 0)
6001 		return;
6002 
6003 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
6004 		if (vfs_badlock_print)
6005 			printf(
6006 			    "VOP_STRATEGY: bp is not locked but should be\n");
6007 		if (vfs_badlock_ddb)
6008 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
6009 	}
6010 }
6011 
6012 void
6013 vop_lock_debugpre(void *ap)
6014 {
6015 	struct vop_lock1_args *a = ap;
6016 
6017 	if ((a->a_flags & LK_INTERLOCK) == 0)
6018 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
6019 	else
6020 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
6021 }
6022 
6023 void
6024 vop_lock_debugpost(void *ap, int rc)
6025 {
6026 	struct vop_lock1_args *a = ap;
6027 
6028 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
6029 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
6030 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
6031 }
6032 
6033 void
6034 vop_unlock_debugpre(void *ap)
6035 {
6036 	struct vop_unlock_args *a = ap;
6037 	struct vnode *vp = a->a_vp;
6038 
6039 	VNPASS(vn_get_state(vp) != VSTATE_UNINITIALIZED, vp);
6040 	ASSERT_VOP_LOCKED(vp, "VOP_UNLOCK");
6041 }
6042 
6043 void
6044 vop_need_inactive_debugpre(void *ap)
6045 {
6046 	struct vop_need_inactive_args *a = ap;
6047 
6048 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
6049 }
6050 
6051 void
6052 vop_need_inactive_debugpost(void *ap, int rc)
6053 {
6054 	struct vop_need_inactive_args *a = ap;
6055 
6056 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
6057 }
6058 #endif
6059 
6060 void
6061 vop_create_pre(void *ap)
6062 {
6063 	struct vop_create_args *a;
6064 	struct vnode *dvp;
6065 
6066 	a = ap;
6067 	dvp = a->a_dvp;
6068 	vn_seqc_write_begin(dvp);
6069 }
6070 
6071 void
6072 vop_create_post(void *ap, int rc)
6073 {
6074 	struct vop_create_args *a;
6075 	struct vnode *dvp;
6076 
6077 	a = ap;
6078 	dvp = a->a_dvp;
6079 	vn_seqc_write_end(dvp);
6080 	if (!rc)
6081 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6082 }
6083 
6084 void
6085 vop_whiteout_pre(void *ap)
6086 {
6087 	struct vop_whiteout_args *a;
6088 	struct vnode *dvp;
6089 
6090 	a = ap;
6091 	dvp = a->a_dvp;
6092 	vn_seqc_write_begin(dvp);
6093 }
6094 
6095 void
6096 vop_whiteout_post(void *ap, int rc)
6097 {
6098 	struct vop_whiteout_args *a;
6099 	struct vnode *dvp;
6100 
6101 	a = ap;
6102 	dvp = a->a_dvp;
6103 	vn_seqc_write_end(dvp);
6104 }
6105 
6106 void
6107 vop_deleteextattr_pre(void *ap)
6108 {
6109 	struct vop_deleteextattr_args *a;
6110 	struct vnode *vp;
6111 
6112 	a = ap;
6113 	vp = a->a_vp;
6114 	vn_seqc_write_begin(vp);
6115 }
6116 
6117 void
6118 vop_deleteextattr_post(void *ap, int rc)
6119 {
6120 	struct vop_deleteextattr_args *a;
6121 	struct vnode *vp;
6122 
6123 	a = ap;
6124 	vp = a->a_vp;
6125 	vn_seqc_write_end(vp);
6126 	if (!rc)
6127 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
6128 }
6129 
6130 void
6131 vop_link_pre(void *ap)
6132 {
6133 	struct vop_link_args *a;
6134 	struct vnode *vp, *tdvp;
6135 
6136 	a = ap;
6137 	vp = a->a_vp;
6138 	tdvp = a->a_tdvp;
6139 	vn_seqc_write_begin(vp);
6140 	vn_seqc_write_begin(tdvp);
6141 }
6142 
6143 void
6144 vop_link_post(void *ap, int rc)
6145 {
6146 	struct vop_link_args *a;
6147 	struct vnode *vp, *tdvp;
6148 
6149 	a = ap;
6150 	vp = a->a_vp;
6151 	tdvp = a->a_tdvp;
6152 	vn_seqc_write_end(vp);
6153 	vn_seqc_write_end(tdvp);
6154 	if (!rc) {
6155 		VFS_KNOTE_LOCKED(vp, NOTE_LINK);
6156 		VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE);
6157 	}
6158 }
6159 
6160 void
6161 vop_mkdir_pre(void *ap)
6162 {
6163 	struct vop_mkdir_args *a;
6164 	struct vnode *dvp;
6165 
6166 	a = ap;
6167 	dvp = a->a_dvp;
6168 	vn_seqc_write_begin(dvp);
6169 }
6170 
6171 void
6172 vop_mkdir_post(void *ap, int rc)
6173 {
6174 	struct vop_mkdir_args *a;
6175 	struct vnode *dvp;
6176 
6177 	a = ap;
6178 	dvp = a->a_dvp;
6179 	vn_seqc_write_end(dvp);
6180 	if (!rc)
6181 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
6182 }
6183 
6184 #ifdef DEBUG_VFS_LOCKS
6185 void
6186 vop_mkdir_debugpost(void *ap, int rc)
6187 {
6188 	struct vop_mkdir_args *a;
6189 
6190 	a = ap;
6191 	if (!rc)
6192 		cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp);
6193 }
6194 #endif
6195 
6196 void
6197 vop_mknod_pre(void *ap)
6198 {
6199 	struct vop_mknod_args *a;
6200 	struct vnode *dvp;
6201 
6202 	a = ap;
6203 	dvp = a->a_dvp;
6204 	vn_seqc_write_begin(dvp);
6205 }
6206 
6207 void
6208 vop_mknod_post(void *ap, int rc)
6209 {
6210 	struct vop_mknod_args *a;
6211 	struct vnode *dvp;
6212 
6213 	a = ap;
6214 	dvp = a->a_dvp;
6215 	vn_seqc_write_end(dvp);
6216 	if (!rc)
6217 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6218 }
6219 
6220 void
6221 vop_reclaim_post(void *ap, int rc)
6222 {
6223 	struct vop_reclaim_args *a;
6224 	struct vnode *vp;
6225 
6226 	a = ap;
6227 	vp = a->a_vp;
6228 	ASSERT_VOP_IN_SEQC(vp);
6229 	if (!rc)
6230 		VFS_KNOTE_LOCKED(vp, NOTE_REVOKE);
6231 }
6232 
6233 void
6234 vop_remove_pre(void *ap)
6235 {
6236 	struct vop_remove_args *a;
6237 	struct vnode *dvp, *vp;
6238 
6239 	a = ap;
6240 	dvp = a->a_dvp;
6241 	vp = a->a_vp;
6242 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_UNLINK);
6243 	vn_seqc_write_begin(dvp);
6244 	vn_seqc_write_begin(vp);
6245 }
6246 
6247 void
6248 vop_remove_post(void *ap, int rc)
6249 {
6250 	struct vop_remove_args *a;
6251 	struct vnode *dvp, *vp;
6252 
6253 	a = ap;
6254 	dvp = a->a_dvp;
6255 	vp = a->a_vp;
6256 	vn_seqc_write_end(dvp);
6257 	vn_seqc_write_end(vp);
6258 	if (!rc) {
6259 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6260 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
6261 	}
6262 }
6263 
6264 void
6265 vop_rename_post(void *ap, int rc)
6266 {
6267 	struct vop_rename_args *a = ap;
6268 	long hint;
6269 
6270 	if (!rc) {
6271 		hint = NOTE_WRITE;
6272 		if (a->a_fdvp == a->a_tdvp) {
6273 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
6274 				hint |= NOTE_LINK;
6275 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
6276 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
6277 		} else {
6278 			hint |= NOTE_EXTEND;
6279 			if (a->a_fvp->v_type == VDIR)
6280 				hint |= NOTE_LINK;
6281 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
6282 
6283 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
6284 			    a->a_tvp->v_type == VDIR)
6285 				hint &= ~NOTE_LINK;
6286 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
6287 		}
6288 
6289 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
6290 		if (a->a_tvp)
6291 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
6292 	}
6293 	if (a->a_tdvp != a->a_fdvp)
6294 		vdrop(a->a_fdvp);
6295 	if (a->a_tvp != a->a_fvp)
6296 		vdrop(a->a_fvp);
6297 	vdrop(a->a_tdvp);
6298 	if (a->a_tvp)
6299 		vdrop(a->a_tvp);
6300 }
6301 
6302 void
6303 vop_rmdir_pre(void *ap)
6304 {
6305 	struct vop_rmdir_args *a;
6306 	struct vnode *dvp, *vp;
6307 
6308 	a = ap;
6309 	dvp = a->a_dvp;
6310 	vp = a->a_vp;
6311 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_UNLINK);
6312 	vn_seqc_write_begin(dvp);
6313 	vn_seqc_write_begin(vp);
6314 }
6315 
6316 void
6317 vop_rmdir_post(void *ap, int rc)
6318 {
6319 	struct vop_rmdir_args *a;
6320 	struct vnode *dvp, *vp;
6321 
6322 	a = ap;
6323 	dvp = a->a_dvp;
6324 	vp = a->a_vp;
6325 	vn_seqc_write_end(dvp);
6326 	vn_seqc_write_end(vp);
6327 	if (!rc) {
6328 		vp->v_vflag |= VV_UNLINKED;
6329 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
6330 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
6331 	}
6332 }
6333 
6334 void
6335 vop_setattr_pre(void *ap)
6336 {
6337 	struct vop_setattr_args *a;
6338 	struct vnode *vp;
6339 
6340 	a = ap;
6341 	vp = a->a_vp;
6342 	vn_seqc_write_begin(vp);
6343 }
6344 
6345 void
6346 vop_setattr_post(void *ap, int rc)
6347 {
6348 	struct vop_setattr_args *a;
6349 	struct vnode *vp;
6350 
6351 	a = ap;
6352 	vp = a->a_vp;
6353 	vn_seqc_write_end(vp);
6354 	if (!rc)
6355 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
6356 }
6357 
6358 void
6359 vop_setacl_pre(void *ap)
6360 {
6361 	struct vop_setacl_args *a;
6362 	struct vnode *vp;
6363 
6364 	a = ap;
6365 	vp = a->a_vp;
6366 	vn_seqc_write_begin(vp);
6367 }
6368 
6369 void
6370 vop_setacl_post(void *ap, int rc __unused)
6371 {
6372 	struct vop_setacl_args *a;
6373 	struct vnode *vp;
6374 
6375 	a = ap;
6376 	vp = a->a_vp;
6377 	vn_seqc_write_end(vp);
6378 }
6379 
6380 void
6381 vop_setextattr_pre(void *ap)
6382 {
6383 	struct vop_setextattr_args *a;
6384 	struct vnode *vp;
6385 
6386 	a = ap;
6387 	vp = a->a_vp;
6388 	vn_seqc_write_begin(vp);
6389 }
6390 
6391 void
6392 vop_setextattr_post(void *ap, int rc)
6393 {
6394 	struct vop_setextattr_args *a;
6395 	struct vnode *vp;
6396 
6397 	a = ap;
6398 	vp = a->a_vp;
6399 	vn_seqc_write_end(vp);
6400 	if (!rc)
6401 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
6402 }
6403 
6404 void
6405 vop_symlink_pre(void *ap)
6406 {
6407 	struct vop_symlink_args *a;
6408 	struct vnode *dvp;
6409 
6410 	a = ap;
6411 	dvp = a->a_dvp;
6412 	vn_seqc_write_begin(dvp);
6413 }
6414 
6415 void
6416 vop_symlink_post(void *ap, int rc)
6417 {
6418 	struct vop_symlink_args *a;
6419 	struct vnode *dvp;
6420 
6421 	a = ap;
6422 	dvp = a->a_dvp;
6423 	vn_seqc_write_end(dvp);
6424 	if (!rc)
6425 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6426 }
6427 
6428 void
6429 vop_open_post(void *ap, int rc)
6430 {
6431 	struct vop_open_args *a = ap;
6432 
6433 	if (!rc)
6434 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
6435 }
6436 
6437 void
6438 vop_close_post(void *ap, int rc)
6439 {
6440 	struct vop_close_args *a = ap;
6441 
6442 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
6443 	    !VN_IS_DOOMED(a->a_vp))) {
6444 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
6445 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
6446 	}
6447 }
6448 
6449 void
6450 vop_read_post(void *ap, int rc)
6451 {
6452 	struct vop_read_args *a = ap;
6453 
6454 	if (!rc)
6455 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6456 }
6457 
6458 void
6459 vop_read_pgcache_post(void *ap, int rc)
6460 {
6461 	struct vop_read_pgcache_args *a = ap;
6462 
6463 	if (!rc)
6464 		VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ);
6465 }
6466 
6467 void
6468 vop_readdir_post(void *ap, int rc)
6469 {
6470 	struct vop_readdir_args *a = ap;
6471 
6472 	if (!rc)
6473 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6474 }
6475 
6476 static struct knlist fs_knlist;
6477 
6478 static void
6479 vfs_event_init(void *arg)
6480 {
6481 	knlist_init_mtx(&fs_knlist, NULL);
6482 }
6483 /* XXX - correct order? */
6484 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
6485 
6486 void
6487 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
6488 {
6489 
6490 	KNOTE_UNLOCKED(&fs_knlist, event);
6491 }
6492 
6493 static int	filt_fsattach(struct knote *kn);
6494 static void	filt_fsdetach(struct knote *kn);
6495 static int	filt_fsevent(struct knote *kn, long hint);
6496 
6497 const struct filterops fs_filtops = {
6498 	.f_isfd = 0,
6499 	.f_attach = filt_fsattach,
6500 	.f_detach = filt_fsdetach,
6501 	.f_event = filt_fsevent,
6502 };
6503 
6504 static int
6505 filt_fsattach(struct knote *kn)
6506 {
6507 
6508 	kn->kn_flags |= EV_CLEAR;
6509 	knlist_add(&fs_knlist, kn, 0);
6510 	return (0);
6511 }
6512 
6513 static void
6514 filt_fsdetach(struct knote *kn)
6515 {
6516 
6517 	knlist_remove(&fs_knlist, kn, 0);
6518 }
6519 
6520 static int
6521 filt_fsevent(struct knote *kn, long hint)
6522 {
6523 
6524 	kn->kn_fflags |= kn->kn_sfflags & hint;
6525 
6526 	return (kn->kn_fflags != 0);
6527 }
6528 
6529 static int
6530 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
6531 {
6532 	struct vfsidctl vc;
6533 	int error;
6534 	struct mount *mp;
6535 
6536 	if (req->newptr == NULL)
6537 		return (EINVAL);
6538 	error = SYSCTL_IN(req, &vc, sizeof(vc));
6539 	if (error)
6540 		return (error);
6541 	if (vc.vc_vers != VFS_CTL_VERS1)
6542 		return (EINVAL);
6543 	mp = vfs_getvfs(&vc.vc_fsid);
6544 	if (mp == NULL)
6545 		return (ENOENT);
6546 	/* ensure that a specific sysctl goes to the right filesystem. */
6547 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
6548 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
6549 		vfs_rel(mp);
6550 		return (EINVAL);
6551 	}
6552 	VCTLTOREQ(&vc, req);
6553 	error = VFS_SYSCTL(mp, vc.vc_op, req);
6554 	vfs_rel(mp);
6555 	return (error);
6556 }
6557 
6558 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
6559     NULL, 0, sysctl_vfs_ctl, "",
6560     "Sysctl by fsid");
6561 
6562 /*
6563  * Function to initialize a va_filerev field sensibly.
6564  * XXX: Wouldn't a random number make a lot more sense ??
6565  */
6566 u_quad_t
6567 init_va_filerev(void)
6568 {
6569 	struct bintime bt;
6570 
6571 	getbinuptime(&bt);
6572 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
6573 }
6574 
6575 static int	filt_vfsread(struct knote *kn, long hint);
6576 static int	filt_vfswrite(struct knote *kn, long hint);
6577 static int	filt_vfsvnode(struct knote *kn, long hint);
6578 static void	filt_vfsdetach(struct knote *kn);
6579 static int	filt_vfsdump(struct proc *p, struct knote *kn,
6580 		    struct kinfo_knote *kin);
6581 
6582 static const struct filterops vfsread_filtops = {
6583 	.f_isfd = 1,
6584 	.f_detach = filt_vfsdetach,
6585 	.f_event = filt_vfsread,
6586 	.f_userdump = filt_vfsdump,
6587 };
6588 static const struct filterops vfswrite_filtops = {
6589 	.f_isfd = 1,
6590 	.f_detach = filt_vfsdetach,
6591 	.f_event = filt_vfswrite,
6592 	.f_userdump = filt_vfsdump,
6593 };
6594 static const struct filterops vfsvnode_filtops = {
6595 	.f_isfd = 1,
6596 	.f_detach = filt_vfsdetach,
6597 	.f_event = filt_vfsvnode,
6598 	.f_userdump = filt_vfsdump,
6599 };
6600 
6601 static void
6602 vfs_knllock(void *arg)
6603 {
6604 	struct vnode *vp = arg;
6605 
6606 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
6607 }
6608 
6609 static void
6610 vfs_knlunlock(void *arg)
6611 {
6612 	struct vnode *vp = arg;
6613 
6614 	VOP_UNLOCK(vp);
6615 }
6616 
6617 static void
6618 vfs_knl_assert_lock(void *arg, int what)
6619 {
6620 #ifdef DEBUG_VFS_LOCKS
6621 	struct vnode *vp = arg;
6622 
6623 	if (what == LA_LOCKED)
6624 		ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
6625 	else
6626 		ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
6627 #endif
6628 }
6629 
6630 int
6631 vfs_kqfilter(struct vop_kqfilter_args *ap)
6632 {
6633 	struct vnode *vp = ap->a_vp;
6634 	struct knote *kn = ap->a_kn;
6635 	struct knlist *knl;
6636 
6637 	KASSERT(vp->v_type != VFIFO || (kn->kn_filter != EVFILT_READ &&
6638 	    kn->kn_filter != EVFILT_WRITE),
6639 	    ("READ/WRITE filter on a FIFO leaked through"));
6640 	switch (kn->kn_filter) {
6641 	case EVFILT_READ:
6642 		kn->kn_fop = &vfsread_filtops;
6643 		break;
6644 	case EVFILT_WRITE:
6645 		kn->kn_fop = &vfswrite_filtops;
6646 		break;
6647 	case EVFILT_VNODE:
6648 		kn->kn_fop = &vfsvnode_filtops;
6649 		break;
6650 	default:
6651 		return (EINVAL);
6652 	}
6653 
6654 	kn->kn_hook = (caddr_t)vp;
6655 
6656 	v_addpollinfo(vp);
6657 	if (vp->v_pollinfo == NULL)
6658 		return (ENOMEM);
6659 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
6660 	vhold(vp);
6661 	knlist_add(knl, kn, 0);
6662 
6663 	return (0);
6664 }
6665 
6666 /*
6667  * Detach knote from vnode
6668  */
6669 static void
6670 filt_vfsdetach(struct knote *kn)
6671 {
6672 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6673 
6674 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
6675 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
6676 	vdrop(vp);
6677 }
6678 
6679 /*ARGSUSED*/
6680 static int
6681 filt_vfsread(struct knote *kn, long hint)
6682 {
6683 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6684 	off_t size;
6685 	int res;
6686 
6687 	/*
6688 	 * filesystem is gone, so set the EOF flag and schedule
6689 	 * the knote for deletion.
6690 	 */
6691 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6692 		VI_LOCK(vp);
6693 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6694 		VI_UNLOCK(vp);
6695 		return (1);
6696 	}
6697 
6698 	if (vn_getsize_locked(vp, &size, curthread->td_ucred) != 0)
6699 		return (0);
6700 
6701 	VI_LOCK(vp);
6702 	kn->kn_data = size - kn->kn_fp->f_offset;
6703 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
6704 	VI_UNLOCK(vp);
6705 	return (res);
6706 }
6707 
6708 /*ARGSUSED*/
6709 static int
6710 filt_vfswrite(struct knote *kn, long hint)
6711 {
6712 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6713 
6714 	VI_LOCK(vp);
6715 
6716 	/*
6717 	 * filesystem is gone, so set the EOF flag and schedule
6718 	 * the knote for deletion.
6719 	 */
6720 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
6721 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6722 
6723 	kn->kn_data = 0;
6724 	VI_UNLOCK(vp);
6725 	return (1);
6726 }
6727 
6728 static int
6729 filt_vfsvnode(struct knote *kn, long hint)
6730 {
6731 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6732 	int res;
6733 
6734 	VI_LOCK(vp);
6735 	if (kn->kn_sfflags & hint)
6736 		kn->kn_fflags |= hint;
6737 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6738 		kn->kn_flags |= EV_EOF;
6739 		VI_UNLOCK(vp);
6740 		return (1);
6741 	}
6742 	res = (kn->kn_fflags != 0);
6743 	VI_UNLOCK(vp);
6744 	return (res);
6745 }
6746 
6747 static int
6748 filt_vfsdump(struct proc *p, struct knote *kn, struct kinfo_knote *kin)
6749 {
6750 	struct vattr va;
6751 	struct vnode *vp;
6752 	char *fullpath, *freepath;
6753 	int error;
6754 
6755 	kin->knt_extdata = KNOTE_EXTDATA_VNODE;
6756 
6757 	vp = kn->kn_fp->f_vnode;
6758 	kin->knt_vnode.knt_vnode_type = vntype_to_kinfo(vp->v_type);
6759 
6760 	va.va_fsid = VNOVAL;
6761 	vn_lock(vp, LK_SHARED | LK_RETRY);
6762 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
6763 	VOP_UNLOCK(vp);
6764 	if (error != 0)
6765 		return (error);
6766 	kin->knt_vnode.knt_vnode_fsid = va.va_fsid;
6767 	kin->knt_vnode.knt_vnode_fileid = va.va_fileid;
6768 
6769 	freepath = NULL;
6770 	fullpath = "-";
6771 	error = vn_fullpath(vp, &fullpath, &freepath);
6772 	if (error == 0) {
6773 		strlcpy(kin->knt_vnode.knt_vnode_fullpath, fullpath,
6774 		    sizeof(kin->knt_vnode.knt_vnode_fullpath));
6775 	}
6776 	if (freepath != NULL)
6777 		free(freepath, M_TEMP);
6778 
6779 	return (0);
6780 }
6781 
6782 int
6783 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6784 {
6785 	int error;
6786 
6787 	if (dp->d_reclen > ap->a_uio->uio_resid)
6788 		return (ENAMETOOLONG);
6789 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
6790 	if (error) {
6791 		if (ap->a_ncookies != NULL) {
6792 			if (ap->a_cookies != NULL)
6793 				free(ap->a_cookies, M_TEMP);
6794 			ap->a_cookies = NULL;
6795 			*ap->a_ncookies = 0;
6796 		}
6797 		return (error);
6798 	}
6799 	if (ap->a_ncookies == NULL)
6800 		return (0);
6801 
6802 	KASSERT(ap->a_cookies,
6803 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6804 
6805 	*ap->a_cookies = realloc(*ap->a_cookies,
6806 	    (*ap->a_ncookies + 1) * sizeof(uint64_t), M_TEMP, M_WAITOK | M_ZERO);
6807 	(*ap->a_cookies)[*ap->a_ncookies] = off;
6808 	*ap->a_ncookies += 1;
6809 	return (0);
6810 }
6811 
6812 /*
6813  * The purpose of this routine is to remove granularity from accmode_t,
6814  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6815  * VADMIN and VAPPEND.
6816  *
6817  * If it returns 0, the caller is supposed to continue with the usual
6818  * access checks using 'accmode' as modified by this routine.  If it
6819  * returns nonzero value, the caller is supposed to return that value
6820  * as errno.
6821  *
6822  * Note that after this routine runs, accmode may be zero.
6823  */
6824 int
6825 vfs_unixify_accmode(accmode_t *accmode)
6826 {
6827 	/*
6828 	 * There is no way to specify explicit "deny" rule using
6829 	 * file mode or POSIX.1e ACLs.
6830 	 */
6831 	if (*accmode & VEXPLICIT_DENY) {
6832 		*accmode = 0;
6833 		return (0);
6834 	}
6835 
6836 	/*
6837 	 * None of these can be translated into usual access bits.
6838 	 * Also, the common case for NFSv4 ACLs is to not contain
6839 	 * either of these bits. Caller should check for VWRITE
6840 	 * on the containing directory instead.
6841 	 */
6842 	if (*accmode & (VDELETE_CHILD | VDELETE))
6843 		return (EPERM);
6844 
6845 	if (*accmode & VADMIN_PERMS) {
6846 		*accmode &= ~VADMIN_PERMS;
6847 		*accmode |= VADMIN;
6848 	}
6849 
6850 	/*
6851 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6852 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6853 	 */
6854 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6855 
6856 	return (0);
6857 }
6858 
6859 /*
6860  * Clear out a doomed vnode (if any) and replace it with a new one as long
6861  * as the fs is not being unmounted. Return the root vnode to the caller.
6862  */
6863 static int __noinline
6864 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6865 {
6866 	struct vnode *vp;
6867 	int error;
6868 
6869 restart:
6870 	if (mp->mnt_rootvnode != NULL) {
6871 		MNT_ILOCK(mp);
6872 		vp = mp->mnt_rootvnode;
6873 		if (vp != NULL) {
6874 			if (!VN_IS_DOOMED(vp)) {
6875 				vrefact(vp);
6876 				MNT_IUNLOCK(mp);
6877 				error = vn_lock(vp, flags);
6878 				if (error == 0) {
6879 					*vpp = vp;
6880 					return (0);
6881 				}
6882 				vrele(vp);
6883 				goto restart;
6884 			}
6885 			/*
6886 			 * Clear the old one.
6887 			 */
6888 			mp->mnt_rootvnode = NULL;
6889 		}
6890 		MNT_IUNLOCK(mp);
6891 		if (vp != NULL) {
6892 			vfs_op_barrier_wait(mp);
6893 			vrele(vp);
6894 		}
6895 	}
6896 	error = VFS_CACHEDROOT(mp, flags, vpp);
6897 	if (error != 0)
6898 		return (error);
6899 	if (mp->mnt_vfs_ops == 0) {
6900 		MNT_ILOCK(mp);
6901 		if (mp->mnt_vfs_ops != 0) {
6902 			MNT_IUNLOCK(mp);
6903 			return (0);
6904 		}
6905 		if (mp->mnt_rootvnode == NULL) {
6906 			vrefact(*vpp);
6907 			mp->mnt_rootvnode = *vpp;
6908 		} else {
6909 			if (mp->mnt_rootvnode != *vpp) {
6910 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6911 					panic("%s: mismatch between vnode returned "
6912 					    " by VFS_CACHEDROOT and the one cached "
6913 					    " (%p != %p)",
6914 					    __func__, *vpp, mp->mnt_rootvnode);
6915 				}
6916 			}
6917 		}
6918 		MNT_IUNLOCK(mp);
6919 	}
6920 	return (0);
6921 }
6922 
6923 int
6924 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6925 {
6926 	struct mount_pcpu *mpcpu;
6927 	struct vnode *vp;
6928 	int error;
6929 
6930 	if (!vfs_op_thread_enter(mp, mpcpu))
6931 		return (vfs_cache_root_fallback(mp, flags, vpp));
6932 	vp = atomic_load_ptr(&mp->mnt_rootvnode);
6933 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6934 		vfs_op_thread_exit(mp, mpcpu);
6935 		return (vfs_cache_root_fallback(mp, flags, vpp));
6936 	}
6937 	vrefact(vp);
6938 	vfs_op_thread_exit(mp, mpcpu);
6939 	error = vn_lock(vp, flags);
6940 	if (error != 0) {
6941 		vrele(vp);
6942 		return (vfs_cache_root_fallback(mp, flags, vpp));
6943 	}
6944 	*vpp = vp;
6945 	return (0);
6946 }
6947 
6948 struct vnode *
6949 vfs_cache_root_clear(struct mount *mp)
6950 {
6951 	struct vnode *vp;
6952 
6953 	/*
6954 	 * ops > 0 guarantees there is nobody who can see this vnode
6955 	 */
6956 	MPASS(mp->mnt_vfs_ops > 0);
6957 	vp = mp->mnt_rootvnode;
6958 	if (vp != NULL)
6959 		vn_seqc_write_begin(vp);
6960 	mp->mnt_rootvnode = NULL;
6961 	return (vp);
6962 }
6963 
6964 void
6965 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6966 {
6967 
6968 	MPASS(mp->mnt_vfs_ops > 0);
6969 	vrefact(vp);
6970 	mp->mnt_rootvnode = vp;
6971 }
6972 
6973 /*
6974  * These are helper functions for filesystems to traverse all
6975  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6976  *
6977  * This interface replaces MNT_VNODE_FOREACH.
6978  */
6979 
6980 struct vnode *
6981 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6982 {
6983 	struct vnode *vp;
6984 
6985 	maybe_yield();
6986 	MNT_ILOCK(mp);
6987 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6988 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6989 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6990 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6991 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6992 			continue;
6993 		VI_LOCK(vp);
6994 		if (VN_IS_DOOMED(vp)) {
6995 			VI_UNLOCK(vp);
6996 			continue;
6997 		}
6998 		break;
6999 	}
7000 	if (vp == NULL) {
7001 		__mnt_vnode_markerfree_all(mvp, mp);
7002 		/* MNT_IUNLOCK(mp); -- done in above function */
7003 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
7004 		return (NULL);
7005 	}
7006 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
7007 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
7008 	MNT_IUNLOCK(mp);
7009 	return (vp);
7010 }
7011 
7012 struct vnode *
7013 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
7014 {
7015 	struct vnode *vp;
7016 
7017 	*mvp = vn_alloc_marker(mp);
7018 	MNT_ILOCK(mp);
7019 	MNT_REF(mp);
7020 
7021 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
7022 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
7023 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
7024 			continue;
7025 		VI_LOCK(vp);
7026 		if (VN_IS_DOOMED(vp)) {
7027 			VI_UNLOCK(vp);
7028 			continue;
7029 		}
7030 		break;
7031 	}
7032 	if (vp == NULL) {
7033 		MNT_REL(mp);
7034 		MNT_IUNLOCK(mp);
7035 		vn_free_marker(*mvp);
7036 		*mvp = NULL;
7037 		return (NULL);
7038 	}
7039 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
7040 	MNT_IUNLOCK(mp);
7041 	return (vp);
7042 }
7043 
7044 void
7045 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
7046 {
7047 
7048 	if (*mvp == NULL) {
7049 		MNT_IUNLOCK(mp);
7050 		return;
7051 	}
7052 
7053 	mtx_assert(MNT_MTX(mp), MA_OWNED);
7054 
7055 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
7056 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
7057 	MNT_REL(mp);
7058 	MNT_IUNLOCK(mp);
7059 	vn_free_marker(*mvp);
7060 	*mvp = NULL;
7061 }
7062 
7063 /*
7064  * These are helper functions for filesystems to traverse their
7065  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
7066  */
7067 static void
7068 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
7069 {
7070 
7071 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
7072 
7073 	MNT_ILOCK(mp);
7074 	MNT_REL(mp);
7075 	MNT_IUNLOCK(mp);
7076 	vn_free_marker(*mvp);
7077 	*mvp = NULL;
7078 }
7079 
7080 /*
7081  * Relock the mp mount vnode list lock with the vp vnode interlock in the
7082  * conventional lock order during mnt_vnode_next_lazy iteration.
7083  *
7084  * On entry, the mount vnode list lock is held and the vnode interlock is not.
7085  * The list lock is dropped and reacquired.  On success, both locks are held.
7086  * On failure, the mount vnode list lock is held but the vnode interlock is
7087  * not, and the procedure may have yielded.
7088  */
7089 static bool
7090 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
7091     struct vnode *vp)
7092 {
7093 
7094 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
7095 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
7096 	    ("%s: bad marker", __func__));
7097 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
7098 	    ("%s: inappropriate vnode", __func__));
7099 	ASSERT_VI_UNLOCKED(vp, __func__);
7100 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
7101 
7102 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
7103 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
7104 
7105 	/*
7106 	 * Note we may be racing against vdrop which transitioned the hold
7107 	 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
7108 	 * if we are the only user after we get the interlock we will just
7109 	 * vdrop.
7110 	 */
7111 	vhold(vp);
7112 	mtx_unlock(&mp->mnt_listmtx);
7113 	VI_LOCK(vp);
7114 	if (VN_IS_DOOMED(vp)) {
7115 		VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
7116 		goto out_lost;
7117 	}
7118 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
7119 	/*
7120 	 * There is nothing to do if we are the last user.
7121 	 */
7122 	if (!refcount_release_if_not_last(&vp->v_holdcnt))
7123 		goto out_lost;
7124 	mtx_lock(&mp->mnt_listmtx);
7125 	return (true);
7126 out_lost:
7127 	vdropl(vp);
7128 	maybe_yield();
7129 	mtx_lock(&mp->mnt_listmtx);
7130 	return (false);
7131 }
7132 
7133 static struct vnode *
7134 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
7135     void *cbarg)
7136 {
7137 	struct vnode *vp;
7138 
7139 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
7140 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
7141 restart:
7142 	vp = TAILQ_NEXT(*mvp, v_lazylist);
7143 	while (vp != NULL) {
7144 		if (vp->v_type == VMARKER) {
7145 			vp = TAILQ_NEXT(vp, v_lazylist);
7146 			continue;
7147 		}
7148 		/*
7149 		 * See if we want to process the vnode. Note we may encounter a
7150 		 * long string of vnodes we don't care about and hog the list
7151 		 * as a result. Check for it and requeue the marker.
7152 		 */
7153 		VNPASS(!VN_IS_DOOMED(vp), vp);
7154 		if (!cb(vp, cbarg)) {
7155 			if (!should_yield()) {
7156 				vp = TAILQ_NEXT(vp, v_lazylist);
7157 				continue;
7158 			}
7159 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
7160 			    v_lazylist);
7161 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
7162 			    v_lazylist);
7163 			mtx_unlock(&mp->mnt_listmtx);
7164 			kern_yield(PRI_USER);
7165 			mtx_lock(&mp->mnt_listmtx);
7166 			goto restart;
7167 		}
7168 		/*
7169 		 * Try-lock because this is the wrong lock order.
7170 		 */
7171 		if (!VI_TRYLOCK(vp) &&
7172 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
7173 			goto restart;
7174 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
7175 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
7176 		    ("alien vnode on the lazy list %p %p", vp, mp));
7177 		VNPASS(vp->v_mount == mp, vp);
7178 		VNPASS(!VN_IS_DOOMED(vp), vp);
7179 		break;
7180 	}
7181 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
7182 
7183 	/* Check if we are done */
7184 	if (vp == NULL) {
7185 		mtx_unlock(&mp->mnt_listmtx);
7186 		mnt_vnode_markerfree_lazy(mvp, mp);
7187 		return (NULL);
7188 	}
7189 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
7190 	mtx_unlock(&mp->mnt_listmtx);
7191 	ASSERT_VI_LOCKED(vp, "lazy iter");
7192 	return (vp);
7193 }
7194 
7195 struct vnode *
7196 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
7197     void *cbarg)
7198 {
7199 
7200 	maybe_yield();
7201 	mtx_lock(&mp->mnt_listmtx);
7202 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
7203 }
7204 
7205 struct vnode *
7206 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
7207     void *cbarg)
7208 {
7209 	struct vnode *vp;
7210 
7211 	if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
7212 		return (NULL);
7213 
7214 	*mvp = vn_alloc_marker(mp);
7215 	MNT_ILOCK(mp);
7216 	MNT_REF(mp);
7217 	MNT_IUNLOCK(mp);
7218 
7219 	mtx_lock(&mp->mnt_listmtx);
7220 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
7221 	if (vp == NULL) {
7222 		mtx_unlock(&mp->mnt_listmtx);
7223 		mnt_vnode_markerfree_lazy(mvp, mp);
7224 		return (NULL);
7225 	}
7226 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
7227 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
7228 }
7229 
7230 void
7231 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
7232 {
7233 
7234 	if (*mvp == NULL)
7235 		return;
7236 
7237 	mtx_lock(&mp->mnt_listmtx);
7238 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
7239 	mtx_unlock(&mp->mnt_listmtx);
7240 	mnt_vnode_markerfree_lazy(mvp, mp);
7241 }
7242 
7243 int
7244 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
7245 {
7246 
7247 	if ((cnp->cn_flags & NOEXECCHECK) != 0) {
7248 		cnp->cn_flags &= ~NOEXECCHECK;
7249 		return (0);
7250 	}
7251 
7252 	return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, curthread));
7253 }
7254 
7255 /*
7256  * Do not use this variant unless you have means other than the hold count
7257  * to prevent the vnode from getting freed.
7258  */
7259 void
7260 vn_seqc_write_begin_locked(struct vnode *vp)
7261 {
7262 
7263 	ASSERT_VI_LOCKED(vp, __func__);
7264 	VNPASS(vp->v_holdcnt > 0, vp);
7265 	VNPASS(vp->v_seqc_users >= 0, vp);
7266 	vp->v_seqc_users++;
7267 	if (vp->v_seqc_users == 1)
7268 		seqc_sleepable_write_begin(&vp->v_seqc);
7269 }
7270 
7271 void
7272 vn_seqc_write_begin(struct vnode *vp)
7273 {
7274 
7275 	VI_LOCK(vp);
7276 	vn_seqc_write_begin_locked(vp);
7277 	VI_UNLOCK(vp);
7278 }
7279 
7280 void
7281 vn_seqc_write_end_locked(struct vnode *vp)
7282 {
7283 
7284 	ASSERT_VI_LOCKED(vp, __func__);
7285 	VNPASS(vp->v_seqc_users > 0, vp);
7286 	vp->v_seqc_users--;
7287 	if (vp->v_seqc_users == 0)
7288 		seqc_sleepable_write_end(&vp->v_seqc);
7289 }
7290 
7291 void
7292 vn_seqc_write_end(struct vnode *vp)
7293 {
7294 
7295 	VI_LOCK(vp);
7296 	vn_seqc_write_end_locked(vp);
7297 	VI_UNLOCK(vp);
7298 }
7299 
7300 /*
7301  * Special case handling for allocating and freeing vnodes.
7302  *
7303  * The counter remains unchanged on free so that a doomed vnode will
7304  * keep testing as in modify as long as it is accessible with SMR.
7305  */
7306 static void
7307 vn_seqc_init(struct vnode *vp)
7308 {
7309 
7310 	vp->v_seqc = 0;
7311 	vp->v_seqc_users = 0;
7312 }
7313 
7314 static void
7315 vn_seqc_write_end_free(struct vnode *vp)
7316 {
7317 
7318 	VNPASS(seqc_in_modify(vp->v_seqc), vp);
7319 	VNPASS(vp->v_seqc_users == 1, vp);
7320 }
7321 
7322 void
7323 vn_irflag_set_locked(struct vnode *vp, short toset)
7324 {
7325 	short flags;
7326 
7327 	ASSERT_VI_LOCKED(vp, __func__);
7328 	flags = vn_irflag_read(vp);
7329 	VNASSERT((flags & toset) == 0, vp,
7330 	    ("%s: some of the passed flags already set (have %d, passed %d)\n",
7331 	    __func__, flags, toset));
7332 	atomic_store_short(&vp->v_irflag, flags | toset);
7333 }
7334 
7335 void
7336 vn_irflag_set(struct vnode *vp, short toset)
7337 {
7338 
7339 	VI_LOCK(vp);
7340 	vn_irflag_set_locked(vp, toset);
7341 	VI_UNLOCK(vp);
7342 }
7343 
7344 void
7345 vn_irflag_set_cond_locked(struct vnode *vp, short toset)
7346 {
7347 	short flags;
7348 
7349 	ASSERT_VI_LOCKED(vp, __func__);
7350 	flags = vn_irflag_read(vp);
7351 	atomic_store_short(&vp->v_irflag, flags | toset);
7352 }
7353 
7354 void
7355 vn_irflag_set_cond(struct vnode *vp, short toset)
7356 {
7357 
7358 	VI_LOCK(vp);
7359 	vn_irflag_set_cond_locked(vp, toset);
7360 	VI_UNLOCK(vp);
7361 }
7362 
7363 void
7364 vn_irflag_unset_locked(struct vnode *vp, short tounset)
7365 {
7366 	short flags;
7367 
7368 	ASSERT_VI_LOCKED(vp, __func__);
7369 	flags = vn_irflag_read(vp);
7370 	VNASSERT((flags & tounset) == tounset, vp,
7371 	    ("%s: some of the passed flags not set (have %d, passed %d)\n",
7372 	    __func__, flags, tounset));
7373 	atomic_store_short(&vp->v_irflag, flags & ~tounset);
7374 }
7375 
7376 void
7377 vn_irflag_unset(struct vnode *vp, short tounset)
7378 {
7379 
7380 	VI_LOCK(vp);
7381 	vn_irflag_unset_locked(vp, tounset);
7382 	VI_UNLOCK(vp);
7383 }
7384 
7385 int
7386 vn_getsize_locked(struct vnode *vp, off_t *size, struct ucred *cred)
7387 {
7388 	struct vattr vattr;
7389 	int error;
7390 
7391 	ASSERT_VOP_LOCKED(vp, __func__);
7392 	error = VOP_GETATTR(vp, &vattr, cred);
7393 	if (__predict_true(error == 0)) {
7394 		if (vattr.va_size <= OFF_MAX)
7395 			*size = vattr.va_size;
7396 		else
7397 			error = EFBIG;
7398 	}
7399 	return (error);
7400 }
7401 
7402 int
7403 vn_getsize(struct vnode *vp, off_t *size, struct ucred *cred)
7404 {
7405 	int error;
7406 
7407 	VOP_LOCK(vp, LK_SHARED);
7408 	error = vn_getsize_locked(vp, size, cred);
7409 	VOP_UNLOCK(vp);
7410 	return (error);
7411 }
7412 
7413 #ifdef INVARIANTS
7414 void
7415 vn_set_state_validate(struct vnode *vp, __enum_uint8(vstate) state)
7416 {
7417 
7418 	switch (vp->v_state) {
7419 	case VSTATE_UNINITIALIZED:
7420 		switch (state) {
7421 		case VSTATE_CONSTRUCTED:
7422 		case VSTATE_DESTROYING:
7423 			return;
7424 		default:
7425 			break;
7426 		}
7427 		break;
7428 	case VSTATE_CONSTRUCTED:
7429 		ASSERT_VOP_ELOCKED(vp, __func__);
7430 		switch (state) {
7431 		case VSTATE_DESTROYING:
7432 			return;
7433 		default:
7434 			break;
7435 		}
7436 		break;
7437 	case VSTATE_DESTROYING:
7438 		ASSERT_VOP_ELOCKED(vp, __func__);
7439 		switch (state) {
7440 		case VSTATE_DEAD:
7441 			return;
7442 		default:
7443 			break;
7444 		}
7445 		break;
7446 	case VSTATE_DEAD:
7447 		switch (state) {
7448 		case VSTATE_UNINITIALIZED:
7449 			return;
7450 		default:
7451 			break;
7452 		}
7453 		break;
7454 	}
7455 
7456 	vn_printf(vp, "invalid state transition %d -> %d\n", vp->v_state, state);
7457 	panic("invalid state transition %d -> %d\n", vp->v_state, state);
7458 }
7459 #endif
7460