xref: /freebsd/sys/kern/vfs_subr.c (revision 21795c374aceb685dbdb1bd18c7a5c41c3cf0baf)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	@(#)vfs_subr.c	8.31 (Berkeley) 5/26/95
37  */
38 
39 /*
40  * External virtual filesystem routines
41  */
42 
43 #include <sys/cdefs.h>
44 #include "opt_ddb.h"
45 #include "opt_watchdog.h"
46 
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/asan.h>
50 #include <sys/bio.h>
51 #include <sys/buf.h>
52 #include <sys/capsicum.h>
53 #include <sys/condvar.h>
54 #include <sys/conf.h>
55 #include <sys/counter.h>
56 #include <sys/dirent.h>
57 #include <sys/event.h>
58 #include <sys/eventhandler.h>
59 #include <sys/extattr.h>
60 #include <sys/file.h>
61 #include <sys/fcntl.h>
62 #include <sys/jail.h>
63 #include <sys/kdb.h>
64 #include <sys/kernel.h>
65 #include <sys/kthread.h>
66 #include <sys/ktr.h>
67 #include <sys/limits.h>
68 #include <sys/lockf.h>
69 #include <sys/malloc.h>
70 #include <sys/mount.h>
71 #include <sys/namei.h>
72 #include <sys/pctrie.h>
73 #include <sys/priv.h>
74 #include <sys/reboot.h>
75 #include <sys/refcount.h>
76 #include <sys/rwlock.h>
77 #include <sys/sched.h>
78 #include <sys/sleepqueue.h>
79 #include <sys/smr.h>
80 #include <sys/smp.h>
81 #include <sys/stat.h>
82 #include <sys/sysctl.h>
83 #include <sys/syslog.h>
84 #include <sys/vmmeter.h>
85 #include <sys/vnode.h>
86 #include <sys/watchdog.h>
87 
88 #include <machine/stdarg.h>
89 
90 #include <security/mac/mac_framework.h>
91 
92 #include <vm/vm.h>
93 #include <vm/vm_object.h>
94 #include <vm/vm_extern.h>
95 #include <vm/pmap.h>
96 #include <vm/vm_map.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_kern.h>
99 #include <vm/uma.h>
100 
101 #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 u_long deferred_inact;
227 SYSCTL_ULONG(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD,
228     &deferred_inact, 0, "Number of times inactive processing was deferred");
229 
230 /* To keep more than one thread at a time from running vfs_getnewfsid */
231 static struct mtx mntid_mtx;
232 
233 /*
234  * Lock for any access to the following:
235  *	vnode_list
236  *	numvnodes
237  *	freevnodes
238  */
239 static struct mtx __exclusive_cache_line vnode_list_mtx;
240 
241 /* Publicly exported FS */
242 struct nfs_public nfs_pub;
243 
244 static uma_zone_t buf_trie_zone;
245 static smr_t buf_trie_smr;
246 
247 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
248 static uma_zone_t vnode_zone;
249 MALLOC_DEFINE(M_VNODEPOLL, "VN POLL", "vnode poll");
250 
251 __read_frequently smr_t vfs_smr;
252 
253 /*
254  * The workitem queue.
255  *
256  * It is useful to delay writes of file data and filesystem metadata
257  * for tens of seconds so that quickly created and deleted files need
258  * not waste disk bandwidth being created and removed. To realize this,
259  * we append vnodes to a "workitem" queue. When running with a soft
260  * updates implementation, most pending metadata dependencies should
261  * not wait for more than a few seconds. Thus, mounted on block devices
262  * are delayed only about a half the time that file data is delayed.
263  * Similarly, directory updates are more critical, so are only delayed
264  * about a third the time that file data is delayed. Thus, there are
265  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
266  * one each second (driven off the filesystem syncer process). The
267  * syncer_delayno variable indicates the next queue that is to be processed.
268  * Items that need to be processed soon are placed in this queue:
269  *
270  *	syncer_workitem_pending[syncer_delayno]
271  *
272  * A delay of fifteen seconds is done by placing the request fifteen
273  * entries later in the queue:
274  *
275  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
276  *
277  */
278 static int syncer_delayno;
279 static long syncer_mask;
280 LIST_HEAD(synclist, bufobj);
281 static struct synclist *syncer_workitem_pending;
282 /*
283  * The sync_mtx protects:
284  *	bo->bo_synclist
285  *	sync_vnode_count
286  *	syncer_delayno
287  *	syncer_state
288  *	syncer_workitem_pending
289  *	syncer_worklist_len
290  *	rushjob
291  */
292 static struct mtx sync_mtx;
293 static struct cv sync_wakeup;
294 
295 #define SYNCER_MAXDELAY		32
296 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
297 static int syncdelay = 30;		/* max time to delay syncing data */
298 static int filedelay = 30;		/* time to delay syncing files */
299 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
300     "Time to delay syncing files (in seconds)");
301 static int dirdelay = 29;		/* time to delay syncing directories */
302 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
303     "Time to delay syncing directories (in seconds)");
304 static int metadelay = 28;		/* time to delay syncing metadata */
305 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
306     "Time to delay syncing metadata (in seconds)");
307 static int rushjob;		/* number of slots to run ASAP */
308 static int stat_rush_requests;	/* number of times I/O speeded up */
309 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
310     "Number of times I/O speeded up (rush requests)");
311 
312 #define	VDBATCH_SIZE 8
313 struct vdbatch {
314 	u_int index;
315 	struct mtx lock;
316 	struct vnode *tab[VDBATCH_SIZE];
317 };
318 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
319 
320 static void	vdbatch_dequeue(struct vnode *vp);
321 
322 /*
323  * When shutting down the syncer, run it at four times normal speed.
324  */
325 #define SYNCER_SHUTDOWN_SPEEDUP		4
326 static int sync_vnode_count;
327 static int syncer_worklist_len;
328 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
329     syncer_state;
330 
331 /* Target for maximum number of vnodes. */
332 u_long desiredvnodes;
333 static u_long gapvnodes;		/* gap between wanted and desired */
334 static u_long vhiwat;		/* enough extras after expansion */
335 static u_long vlowat;		/* minimal extras before expansion */
336 static bool vstir;		/* nonzero to stir non-free vnodes */
337 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
338 
339 static u_long vnlru_read_freevnodes(void);
340 
341 /*
342  * Note that no attempt is made to sanitize these parameters.
343  */
344 static int
345 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
346 {
347 	u_long val;
348 	int error;
349 
350 	val = desiredvnodes;
351 	error = sysctl_handle_long(oidp, &val, 0, req);
352 	if (error != 0 || req->newptr == NULL)
353 		return (error);
354 
355 	if (val == desiredvnodes)
356 		return (0);
357 	mtx_lock(&vnode_list_mtx);
358 	desiredvnodes = val;
359 	wantfreevnodes = desiredvnodes / 4;
360 	vnlru_recalc();
361 	mtx_unlock(&vnode_list_mtx);
362 	/*
363 	 * XXX There is no protection against multiple threads changing
364 	 * desiredvnodes at the same time. Locking above only helps vnlru and
365 	 * getnewvnode.
366 	 */
367 	vfs_hash_changesize(desiredvnodes);
368 	cache_changesize(desiredvnodes);
369 	return (0);
370 }
371 
372 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
373     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
374     "LU", "Target for maximum number of vnodes (legacy)");
375 SYSCTL_PROC(_vfs_vnode_param, OID_AUTO, limit,
376     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
377     "LU", "Target for maximum number of vnodes");
378 
379 static int
380 sysctl_freevnodes(SYSCTL_HANDLER_ARGS)
381 {
382 	u_long rfreevnodes;
383 
384 	rfreevnodes = vnlru_read_freevnodes();
385 	return (sysctl_handle_long(oidp, &rfreevnodes, 0, req));
386 }
387 
388 SYSCTL_PROC(_vfs, OID_AUTO, freevnodes,
389     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_freevnodes,
390     "LU", "Number of \"free\" vnodes (legacy)");
391 SYSCTL_PROC(_vfs_vnode_stats, OID_AUTO, free,
392     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_freevnodes,
393     "LU", "Number of \"free\" vnodes");
394 
395 static int
396 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
397 {
398 	u_long val;
399 	int error;
400 
401 	val = wantfreevnodes;
402 	error = sysctl_handle_long(oidp, &val, 0, req);
403 	if (error != 0 || req->newptr == NULL)
404 		return (error);
405 
406 	if (val == wantfreevnodes)
407 		return (0);
408 	mtx_lock(&vnode_list_mtx);
409 	wantfreevnodes = val;
410 	vnlru_recalc();
411 	mtx_unlock(&vnode_list_mtx);
412 	return (0);
413 }
414 
415 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
416     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
417     "LU", "Target for minimum number of \"free\" vnodes (legacy)");
418 SYSCTL_PROC(_vfs_vnode_param, OID_AUTO, wantfree,
419     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
420     "LU", "Target for minimum number of \"free\" vnodes");
421 
422 static int vnlru_nowhere;
423 SYSCTL_INT(_vfs_vnode_vnlru, OID_AUTO, failed_runs, CTLFLAG_RD | CTLFLAG_STATS,
424     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
425 
426 static int
427 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
428 {
429 	struct vnode *vp;
430 	struct nameidata nd;
431 	char *buf;
432 	unsigned long ndflags;
433 	int error;
434 
435 	if (req->newptr == NULL)
436 		return (EINVAL);
437 	if (req->newlen >= PATH_MAX)
438 		return (E2BIG);
439 
440 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
441 	error = SYSCTL_IN(req, buf, req->newlen);
442 	if (error != 0)
443 		goto out;
444 
445 	buf[req->newlen] = '\0';
446 
447 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1;
448 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf);
449 	if ((error = namei(&nd)) != 0)
450 		goto out;
451 	vp = nd.ni_vp;
452 
453 	if (VN_IS_DOOMED(vp)) {
454 		/*
455 		 * This vnode is being recycled.  Return != 0 to let the caller
456 		 * know that the sysctl had no effect.  Return EAGAIN because a
457 		 * subsequent call will likely succeed (since namei will create
458 		 * a new vnode if necessary)
459 		 */
460 		error = EAGAIN;
461 		goto putvnode;
462 	}
463 
464 	vgone(vp);
465 putvnode:
466 	vput(vp);
467 	NDFREE_PNBUF(&nd);
468 out:
469 	free(buf, M_TEMP);
470 	return (error);
471 }
472 
473 static int
474 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
475 {
476 	struct thread *td = curthread;
477 	struct vnode *vp;
478 	struct file *fp;
479 	int error;
480 	int fd;
481 
482 	if (req->newptr == NULL)
483 		return (EBADF);
484 
485         error = sysctl_handle_int(oidp, &fd, 0, req);
486         if (error != 0)
487                 return (error);
488 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
489 	if (error != 0)
490 		return (error);
491 	vp = fp->f_vnode;
492 
493 	error = vn_lock(vp, LK_EXCLUSIVE);
494 	if (error != 0)
495 		goto drop;
496 
497 	vgone(vp);
498 	VOP_UNLOCK(vp);
499 drop:
500 	fdrop(fp, td);
501 	return (error);
502 }
503 
504 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
505     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
506     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
507 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
508     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
509     sysctl_ftry_reclaim_vnode, "I",
510     "Try to reclaim a vnode by its file descriptor");
511 
512 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
513 #define vnsz2log 8
514 #ifndef DEBUG_LOCKS
515 _Static_assert(sizeof(struct vnode) >= 1UL << vnsz2log &&
516     sizeof(struct vnode) < 1UL << (vnsz2log + 1),
517     "vnsz2log needs to be updated");
518 #endif
519 
520 /*
521  * Support for the bufobj clean & dirty pctrie.
522  */
523 static void *
524 buf_trie_alloc(struct pctrie *ptree)
525 {
526 	return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT));
527 }
528 
529 static void
530 buf_trie_free(struct pctrie *ptree, void *node)
531 {
532 	uma_zfree_smr(buf_trie_zone, node);
533 }
534 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free,
535     buf_trie_smr);
536 
537 /*
538  * Initialize the vnode management data structures.
539  *
540  * Reevaluate the following cap on the number of vnodes after the physical
541  * memory size exceeds 512GB.  In the limit, as the physical memory size
542  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
543  */
544 #ifndef	MAXVNODES_MAX
545 #define	MAXVNODES_MAX	(512UL * 1024 * 1024 / 64)	/* 8M */
546 #endif
547 
548 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
549 
550 static struct vnode *
551 vn_alloc_marker(struct mount *mp)
552 {
553 	struct vnode *vp;
554 
555 	vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
556 	vp->v_type = VMARKER;
557 	vp->v_mount = mp;
558 
559 	return (vp);
560 }
561 
562 static void
563 vn_free_marker(struct vnode *vp)
564 {
565 
566 	MPASS(vp->v_type == VMARKER);
567 	free(vp, M_VNODE_MARKER);
568 }
569 
570 #ifdef KASAN
571 static int
572 vnode_ctor(void *mem, int size, void *arg __unused, int flags __unused)
573 {
574 	kasan_mark(mem, size, roundup2(size, UMA_ALIGN_PTR + 1), 0);
575 	return (0);
576 }
577 
578 static void
579 vnode_dtor(void *mem, int size, void *arg __unused)
580 {
581 	size_t end1, end2, off1, off2;
582 
583 	_Static_assert(offsetof(struct vnode, v_vnodelist) <
584 	    offsetof(struct vnode, v_dbatchcpu),
585 	    "KASAN marks require updating");
586 
587 	off1 = offsetof(struct vnode, v_vnodelist);
588 	off2 = offsetof(struct vnode, v_dbatchcpu);
589 	end1 = off1 + sizeof(((struct vnode *)NULL)->v_vnodelist);
590 	end2 = off2 + sizeof(((struct vnode *)NULL)->v_dbatchcpu);
591 
592 	/*
593 	 * Access to the v_vnodelist and v_dbatchcpu fields are permitted even
594 	 * after the vnode has been freed.  Try to get some KASAN coverage by
595 	 * marking everything except those two fields as invalid.  Because
596 	 * KASAN's tracking is not byte-granular, any preceding fields sharing
597 	 * the same 8-byte aligned word must also be marked valid.
598 	 */
599 
600 	/* Handle the area from the start until v_vnodelist... */
601 	off1 = rounddown2(off1, KASAN_SHADOW_SCALE);
602 	kasan_mark(mem, off1, off1, KASAN_UMA_FREED);
603 
604 	/* ... then the area between v_vnodelist and v_dbatchcpu ... */
605 	off1 = roundup2(end1, KASAN_SHADOW_SCALE);
606 	off2 = rounddown2(off2, KASAN_SHADOW_SCALE);
607 	if (off2 > off1)
608 		kasan_mark((void *)((char *)mem + off1), off2 - off1,
609 		    off2 - off1, KASAN_UMA_FREED);
610 
611 	/* ... and finally the area from v_dbatchcpu to the end. */
612 	off2 = roundup2(end2, KASAN_SHADOW_SCALE);
613 	kasan_mark((void *)((char *)mem + off2), size - off2, size - off2,
614 	    KASAN_UMA_FREED);
615 }
616 #endif /* KASAN */
617 
618 /*
619  * Initialize a vnode as it first enters the zone.
620  */
621 static int
622 vnode_init(void *mem, int size, int flags)
623 {
624 	struct vnode *vp;
625 
626 	vp = mem;
627 	bzero(vp, size);
628 	/*
629 	 * Setup locks.
630 	 */
631 	vp->v_vnlock = &vp->v_lock;
632 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
633 	/*
634 	 * By default, don't allow shared locks unless filesystems opt-in.
635 	 */
636 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
637 	    LK_NOSHARE | LK_IS_VNODE);
638 	/*
639 	 * Initialize bufobj.
640 	 */
641 	bufobj_init(&vp->v_bufobj, vp);
642 	/*
643 	 * Initialize namecache.
644 	 */
645 	cache_vnode_init(vp);
646 	/*
647 	 * Initialize rangelocks.
648 	 */
649 	rangelock_init(&vp->v_rl);
650 
651 	vp->v_dbatchcpu = NOCPU;
652 
653 	vp->v_state = VSTATE_DEAD;
654 
655 	/*
656 	 * Check vhold_recycle_free for an explanation.
657 	 */
658 	vp->v_holdcnt = VHOLD_NO_SMR;
659 	vp->v_type = VNON;
660 	mtx_lock(&vnode_list_mtx);
661 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
662 	mtx_unlock(&vnode_list_mtx);
663 	return (0);
664 }
665 
666 /*
667  * Free a vnode when it is cleared from the zone.
668  */
669 static void
670 vnode_fini(void *mem, int size)
671 {
672 	struct vnode *vp;
673 	struct bufobj *bo;
674 
675 	vp = mem;
676 	vdbatch_dequeue(vp);
677 	mtx_lock(&vnode_list_mtx);
678 	TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
679 	mtx_unlock(&vnode_list_mtx);
680 	rangelock_destroy(&vp->v_rl);
681 	lockdestroy(vp->v_vnlock);
682 	mtx_destroy(&vp->v_interlock);
683 	bo = &vp->v_bufobj;
684 	rw_destroy(BO_LOCKPTR(bo));
685 
686 	kasan_mark(mem, size, size, 0);
687 }
688 
689 /*
690  * Provide the size of NFS nclnode and NFS fh for calculation of the
691  * vnode memory consumption.  The size is specified directly to
692  * eliminate dependency on NFS-private header.
693  *
694  * Other filesystems may use bigger or smaller (like UFS and ZFS)
695  * private inode data, but the NFS-based estimation is ample enough.
696  * Still, we care about differences in the size between 64- and 32-bit
697  * platforms.
698  *
699  * Namecache structure size is heuristically
700  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
701  */
702 #ifdef _LP64
703 #define	NFS_NCLNODE_SZ	(528 + 64)
704 #define	NC_SZ		148
705 #else
706 #define	NFS_NCLNODE_SZ	(360 + 32)
707 #define	NC_SZ		92
708 #endif
709 
710 static void
711 vntblinit(void *dummy __unused)
712 {
713 	struct vdbatch *vd;
714 	uma_ctor ctor;
715 	uma_dtor dtor;
716 	int cpu, physvnodes, virtvnodes;
717 
718 	/*
719 	 * Desiredvnodes is a function of the physical memory size and the
720 	 * kernel's heap size.  Generally speaking, it scales with the
721 	 * physical memory size.  The ratio of desiredvnodes to the physical
722 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
723 	 * Thereafter, the
724 	 * marginal ratio of desiredvnodes to the physical memory size is
725 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
726 	 * size.  The memory required by desiredvnodes vnodes and vm objects
727 	 * must not exceed 1/10th of the kernel's heap size.
728 	 */
729 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
730 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
731 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
732 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
733 	desiredvnodes = min(physvnodes, virtvnodes);
734 	if (desiredvnodes > MAXVNODES_MAX) {
735 		if (bootverbose)
736 			printf("Reducing kern.maxvnodes %lu -> %lu\n",
737 			    desiredvnodes, MAXVNODES_MAX);
738 		desiredvnodes = MAXVNODES_MAX;
739 	}
740 	wantfreevnodes = desiredvnodes / 4;
741 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
742 	TAILQ_INIT(&vnode_list);
743 	mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
744 	/*
745 	 * The lock is taken to appease WITNESS.
746 	 */
747 	mtx_lock(&vnode_list_mtx);
748 	vnlru_recalc();
749 	mtx_unlock(&vnode_list_mtx);
750 	vnode_list_free_marker = vn_alloc_marker(NULL);
751 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
752 	vnode_list_reclaim_marker = vn_alloc_marker(NULL);
753 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
754 
755 #ifdef KASAN
756 	ctor = vnode_ctor;
757 	dtor = vnode_dtor;
758 #else
759 	ctor = NULL;
760 	dtor = NULL;
761 #endif
762 	vnode_zone = uma_zcreate("VNODE", sizeof(struct vnode), ctor, dtor,
763 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, UMA_ZONE_NOKASAN);
764 	uma_zone_set_smr(vnode_zone, vfs_smr);
765 
766 	/*
767 	 * Preallocate enough nodes to support one-per buf so that
768 	 * we can not fail an insert.  reassignbuf() callers can not
769 	 * tolerate the insertion failure.
770 	 */
771 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
772 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
773 	    UMA_ZONE_NOFREE | UMA_ZONE_SMR);
774 	buf_trie_smr = uma_zone_get_smr(buf_trie_zone);
775 	uma_prealloc(buf_trie_zone, nbuf);
776 
777 	vnodes_created = counter_u64_alloc(M_WAITOK);
778 	direct_recycles_free_count = counter_u64_alloc(M_WAITOK);
779 	vnode_skipped_requeues = counter_u64_alloc(M_WAITOK);
780 
781 	/*
782 	 * Initialize the filesystem syncer.
783 	 */
784 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
785 	    &syncer_mask);
786 	syncer_maxdelay = syncer_mask + 1;
787 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
788 	cv_init(&sync_wakeup, "syncer");
789 
790 	CPU_FOREACH(cpu) {
791 		vd = DPCPU_ID_PTR((cpu), vd);
792 		bzero(vd, sizeof(*vd));
793 		mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
794 	}
795 }
796 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
797 
798 /*
799  * Mark a mount point as busy. Used to synchronize access and to delay
800  * unmounting. Eventually, mountlist_mtx is not released on failure.
801  *
802  * vfs_busy() is a custom lock, it can block the caller.
803  * vfs_busy() only sleeps if the unmount is active on the mount point.
804  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
805  * vnode belonging to mp.
806  *
807  * Lookup uses vfs_busy() to traverse mount points.
808  * root fs			var fs
809  * / vnode lock		A	/ vnode lock (/var)		D
810  * /var vnode lock	B	/log vnode lock(/var/log)	E
811  * vfs_busy lock	C	vfs_busy lock			F
812  *
813  * Within each file system, the lock order is C->A->B and F->D->E.
814  *
815  * When traversing across mounts, the system follows that lock order:
816  *
817  *        C->A->B
818  *              |
819  *              +->F->D->E
820  *
821  * The lookup() process for namei("/var") illustrates the process:
822  *  1. VOP_LOOKUP() obtains B while A is held
823  *  2. vfs_busy() obtains a shared lock on F while A and B are held
824  *  3. vput() releases lock on B
825  *  4. vput() releases lock on A
826  *  5. VFS_ROOT() obtains lock on D while shared lock on F is held
827  *  6. vfs_unbusy() releases shared lock on F
828  *  7. vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
829  *     Attempt to lock A (instead of vp_crossmp) while D is held would
830  *     violate the global order, causing deadlocks.
831  *
832  * dounmount() locks B while F is drained.  Note that for stacked
833  * filesystems, D and B in the example above may be the same lock,
834  * which introdues potential lock order reversal deadlock between
835  * dounmount() and step 5 above.  These filesystems may avoid the LOR
836  * by setting VV_CROSSLOCK on the covered vnode so that lock B will
837  * remain held until after step 5.
838  */
839 int
840 vfs_busy(struct mount *mp, int flags)
841 {
842 	struct mount_pcpu *mpcpu;
843 
844 	MPASS((flags & ~MBF_MASK) == 0);
845 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
846 
847 	if (vfs_op_thread_enter(mp, mpcpu)) {
848 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
849 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
850 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
851 		vfs_mp_count_add_pcpu(mpcpu, ref, 1);
852 		vfs_mp_count_add_pcpu(mpcpu, lockref, 1);
853 		vfs_op_thread_exit(mp, mpcpu);
854 		if (flags & MBF_MNTLSTLOCK)
855 			mtx_unlock(&mountlist_mtx);
856 		return (0);
857 	}
858 
859 	MNT_ILOCK(mp);
860 	vfs_assert_mount_counters(mp);
861 	MNT_REF(mp);
862 	/*
863 	 * If mount point is currently being unmounted, sleep until the
864 	 * mount point fate is decided.  If thread doing the unmounting fails,
865 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
866 	 * that this mount point has survived the unmount attempt and vfs_busy
867 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
868 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
869 	 * about to be really destroyed.  vfs_busy needs to release its
870 	 * reference on the mount point in this case and return with ENOENT,
871 	 * telling the caller the mount it tried to busy is no longer valid.
872 	 */
873 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
874 		KASSERT(TAILQ_EMPTY(&mp->mnt_uppers),
875 		    ("%s: non-empty upper mount list with pending unmount",
876 		    __func__));
877 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
878 			MNT_REL(mp);
879 			MNT_IUNLOCK(mp);
880 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
881 			    __func__);
882 			return (ENOENT);
883 		}
884 		if (flags & MBF_MNTLSTLOCK)
885 			mtx_unlock(&mountlist_mtx);
886 		mp->mnt_kern_flag |= MNTK_MWAIT;
887 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
888 		if (flags & MBF_MNTLSTLOCK)
889 			mtx_lock(&mountlist_mtx);
890 		MNT_ILOCK(mp);
891 	}
892 	if (flags & MBF_MNTLSTLOCK)
893 		mtx_unlock(&mountlist_mtx);
894 	mp->mnt_lockref++;
895 	MNT_IUNLOCK(mp);
896 	return (0);
897 }
898 
899 /*
900  * Free a busy filesystem.
901  */
902 void
903 vfs_unbusy(struct mount *mp)
904 {
905 	struct mount_pcpu *mpcpu;
906 	int c;
907 
908 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
909 
910 	if (vfs_op_thread_enter(mp, mpcpu)) {
911 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
912 		vfs_mp_count_sub_pcpu(mpcpu, lockref, 1);
913 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
914 		vfs_op_thread_exit(mp, mpcpu);
915 		return;
916 	}
917 
918 	MNT_ILOCK(mp);
919 	vfs_assert_mount_counters(mp);
920 	MNT_REL(mp);
921 	c = --mp->mnt_lockref;
922 	if (mp->mnt_vfs_ops == 0) {
923 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
924 		MNT_IUNLOCK(mp);
925 		return;
926 	}
927 	if (c < 0)
928 		vfs_dump_mount_counters(mp);
929 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
930 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
931 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
932 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
933 		wakeup(&mp->mnt_lockref);
934 	}
935 	MNT_IUNLOCK(mp);
936 }
937 
938 /*
939  * Lookup a mount point by filesystem identifier.
940  */
941 struct mount *
942 vfs_getvfs(fsid_t *fsid)
943 {
944 	struct mount *mp;
945 
946 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
947 	mtx_lock(&mountlist_mtx);
948 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
949 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
950 			vfs_ref(mp);
951 			mtx_unlock(&mountlist_mtx);
952 			return (mp);
953 		}
954 	}
955 	mtx_unlock(&mountlist_mtx);
956 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
957 	return ((struct mount *) 0);
958 }
959 
960 /*
961  * Lookup a mount point by filesystem identifier, busying it before
962  * returning.
963  *
964  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
965  * cache for popular filesystem identifiers.  The cache is lockess, using
966  * the fact that struct mount's are never freed.  In worst case we may
967  * get pointer to unmounted or even different filesystem, so we have to
968  * check what we got, and go slow way if so.
969  */
970 struct mount *
971 vfs_busyfs(fsid_t *fsid)
972 {
973 #define	FSID_CACHE_SIZE	256
974 	typedef struct mount * volatile vmp_t;
975 	static vmp_t cache[FSID_CACHE_SIZE];
976 	struct mount *mp;
977 	int error;
978 	uint32_t hash;
979 
980 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
981 	hash = fsid->val[0] ^ fsid->val[1];
982 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
983 	mp = cache[hash];
984 	if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0)
985 		goto slow;
986 	if (vfs_busy(mp, 0) != 0) {
987 		cache[hash] = NULL;
988 		goto slow;
989 	}
990 	if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0)
991 		return (mp);
992 	else
993 	    vfs_unbusy(mp);
994 
995 slow:
996 	mtx_lock(&mountlist_mtx);
997 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
998 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
999 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
1000 			if (error) {
1001 				cache[hash] = NULL;
1002 				mtx_unlock(&mountlist_mtx);
1003 				return (NULL);
1004 			}
1005 			cache[hash] = mp;
1006 			return (mp);
1007 		}
1008 	}
1009 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
1010 	mtx_unlock(&mountlist_mtx);
1011 	return ((struct mount *) 0);
1012 }
1013 
1014 /*
1015  * Check if a user can access privileged mount options.
1016  */
1017 int
1018 vfs_suser(struct mount *mp, struct thread *td)
1019 {
1020 	int error;
1021 
1022 	if (jailed(td->td_ucred)) {
1023 		/*
1024 		 * If the jail of the calling thread lacks permission for
1025 		 * this type of file system, deny immediately.
1026 		 */
1027 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
1028 			return (EPERM);
1029 
1030 		/*
1031 		 * If the file system was mounted outside the jail of the
1032 		 * calling thread, deny immediately.
1033 		 */
1034 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
1035 			return (EPERM);
1036 	}
1037 
1038 	/*
1039 	 * If file system supports delegated administration, we don't check
1040 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
1041 	 * by the file system itself.
1042 	 * If this is not the user that did original mount, we check for
1043 	 * the PRIV_VFS_MOUNT_OWNER privilege.
1044 	 */
1045 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
1046 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
1047 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
1048 			return (error);
1049 	}
1050 	return (0);
1051 }
1052 
1053 /*
1054  * Get a new unique fsid.  Try to make its val[0] unique, since this value
1055  * will be used to create fake device numbers for stat().  Also try (but
1056  * not so hard) make its val[0] unique mod 2^16, since some emulators only
1057  * support 16-bit device numbers.  We end up with unique val[0]'s for the
1058  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
1059  *
1060  * Keep in mind that several mounts may be running in parallel.  Starting
1061  * the search one past where the previous search terminated is both a
1062  * micro-optimization and a defense against returning the same fsid to
1063  * different mounts.
1064  */
1065 void
1066 vfs_getnewfsid(struct mount *mp)
1067 {
1068 	static uint16_t mntid_base;
1069 	struct mount *nmp;
1070 	fsid_t tfsid;
1071 	int mtype;
1072 
1073 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
1074 	mtx_lock(&mntid_mtx);
1075 	mtype = mp->mnt_vfc->vfc_typenum;
1076 	tfsid.val[1] = mtype;
1077 	mtype = (mtype & 0xFF) << 24;
1078 	for (;;) {
1079 		tfsid.val[0] = makedev(255,
1080 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
1081 		mntid_base++;
1082 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
1083 			break;
1084 		vfs_rel(nmp);
1085 	}
1086 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
1087 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
1088 	mtx_unlock(&mntid_mtx);
1089 }
1090 
1091 /*
1092  * Knob to control the precision of file timestamps:
1093  *
1094  *   0 = seconds only; nanoseconds zeroed.
1095  *   1 = seconds and nanoseconds, accurate within 1/HZ.
1096  *   2 = seconds and nanoseconds, truncated to microseconds.
1097  * >=3 = seconds and nanoseconds, maximum precision.
1098  */
1099 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
1100 
1101 static int timestamp_precision = TSP_USEC;
1102 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
1103     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
1104     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1105     "3+: sec + ns (max. precision))");
1106 
1107 /*
1108  * Get a current timestamp.
1109  */
1110 void
1111 vfs_timestamp(struct timespec *tsp)
1112 {
1113 	struct timeval tv;
1114 
1115 	switch (timestamp_precision) {
1116 	case TSP_SEC:
1117 		tsp->tv_sec = time_second;
1118 		tsp->tv_nsec = 0;
1119 		break;
1120 	case TSP_HZ:
1121 		getnanotime(tsp);
1122 		break;
1123 	case TSP_USEC:
1124 		microtime(&tv);
1125 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1126 		break;
1127 	case TSP_NSEC:
1128 	default:
1129 		nanotime(tsp);
1130 		break;
1131 	}
1132 }
1133 
1134 /*
1135  * Set vnode attributes to VNOVAL
1136  */
1137 void
1138 vattr_null(struct vattr *vap)
1139 {
1140 
1141 	vap->va_type = VNON;
1142 	vap->va_size = VNOVAL;
1143 	vap->va_bytes = VNOVAL;
1144 	vap->va_mode = VNOVAL;
1145 	vap->va_nlink = VNOVAL;
1146 	vap->va_uid = VNOVAL;
1147 	vap->va_gid = VNOVAL;
1148 	vap->va_fsid = VNOVAL;
1149 	vap->va_fileid = VNOVAL;
1150 	vap->va_blocksize = VNOVAL;
1151 	vap->va_rdev = VNOVAL;
1152 	vap->va_atime.tv_sec = VNOVAL;
1153 	vap->va_atime.tv_nsec = VNOVAL;
1154 	vap->va_mtime.tv_sec = VNOVAL;
1155 	vap->va_mtime.tv_nsec = VNOVAL;
1156 	vap->va_ctime.tv_sec = VNOVAL;
1157 	vap->va_ctime.tv_nsec = VNOVAL;
1158 	vap->va_birthtime.tv_sec = VNOVAL;
1159 	vap->va_birthtime.tv_nsec = VNOVAL;
1160 	vap->va_flags = VNOVAL;
1161 	vap->va_gen = VNOVAL;
1162 	vap->va_vaflags = 0;
1163 }
1164 
1165 /*
1166  * Try to reduce the total number of vnodes.
1167  *
1168  * This routine (and its user) are buggy in at least the following ways:
1169  * - all parameters were picked years ago when RAM sizes were significantly
1170  *   smaller
1171  * - it can pick vnodes based on pages used by the vm object, but filesystems
1172  *   like ZFS don't use it making the pick broken
1173  * - since ZFS has its own aging policy it gets partially combated by this one
1174  * - a dedicated method should be provided for filesystems to let them decide
1175  *   whether the vnode should be recycled
1176  *
1177  * This routine is called when we have too many vnodes.  It attempts
1178  * to free <count> vnodes and will potentially free vnodes that still
1179  * have VM backing store (VM backing store is typically the cause
1180  * of a vnode blowout so we want to do this).  Therefore, this operation
1181  * is not considered cheap.
1182  *
1183  * A number of conditions may prevent a vnode from being reclaimed.
1184  * the buffer cache may have references on the vnode, a directory
1185  * vnode may still have references due to the namei cache representing
1186  * underlying files, or the vnode may be in active use.   It is not
1187  * desirable to reuse such vnodes.  These conditions may cause the
1188  * number of vnodes to reach some minimum value regardless of what
1189  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
1190  *
1191  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1192  * 			 entries if this argument is strue
1193  * @param trigger	 Only reclaim vnodes with fewer than this many resident
1194  *			 pages.
1195  * @param target	 How many vnodes to reclaim.
1196  * @return		 The number of vnodes that were reclaimed.
1197  */
1198 static int
1199 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1200 {
1201 	struct vnode *vp, *mvp;
1202 	struct mount *mp;
1203 	struct vm_object *object;
1204 	u_long done;
1205 	bool retried;
1206 
1207 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1208 
1209 	retried = false;
1210 	done = 0;
1211 
1212 	mvp = vnode_list_reclaim_marker;
1213 restart:
1214 	vp = mvp;
1215 	while (done < target) {
1216 		vp = TAILQ_NEXT(vp, v_vnodelist);
1217 		if (__predict_false(vp == NULL))
1218 			break;
1219 
1220 		if (__predict_false(vp->v_type == VMARKER))
1221 			continue;
1222 
1223 		/*
1224 		 * If it's been deconstructed already, it's still
1225 		 * referenced, or it exceeds the trigger, skip it.
1226 		 * Also skip free vnodes.  We are trying to make space
1227 		 * for more free vnodes, not reduce their count.
1228 		 */
1229 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1230 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1231 			goto next_iter;
1232 
1233 		if (vp->v_type == VBAD || vp->v_type == VNON)
1234 			goto next_iter;
1235 
1236 		object = atomic_load_ptr(&vp->v_object);
1237 		if (object == NULL || object->resident_page_count > trigger) {
1238 			goto next_iter;
1239 		}
1240 
1241 		/*
1242 		 * Handle races against vnode allocation. Filesystems lock the
1243 		 * vnode some time after it gets returned from getnewvnode,
1244 		 * despite type and hold count being manipulated earlier.
1245 		 * Resorting to checking v_mount restores guarantees present
1246 		 * before the global list was reworked to contain all vnodes.
1247 		 */
1248 		if (!VI_TRYLOCK(vp))
1249 			goto next_iter;
1250 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1251 			VI_UNLOCK(vp);
1252 			goto next_iter;
1253 		}
1254 		if (vp->v_mount == NULL) {
1255 			VI_UNLOCK(vp);
1256 			goto next_iter;
1257 		}
1258 		vholdl(vp);
1259 		VI_UNLOCK(vp);
1260 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1261 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1262 		mtx_unlock(&vnode_list_mtx);
1263 
1264 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1265 			vdrop_recycle(vp);
1266 			goto next_iter_unlocked;
1267 		}
1268 		if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1269 			vdrop_recycle(vp);
1270 			vn_finished_write(mp);
1271 			goto next_iter_unlocked;
1272 		}
1273 
1274 		VI_LOCK(vp);
1275 		if (vp->v_usecount > 0 ||
1276 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1277 		    (vp->v_object != NULL && vp->v_object->handle == vp &&
1278 		    vp->v_object->resident_page_count > trigger)) {
1279 			VOP_UNLOCK(vp);
1280 			vdropl_recycle(vp);
1281 			vn_finished_write(mp);
1282 			goto next_iter_unlocked;
1283 		}
1284 		recycles_count++;
1285 		vgonel(vp);
1286 		VOP_UNLOCK(vp);
1287 		vdropl_recycle(vp);
1288 		vn_finished_write(mp);
1289 		done++;
1290 next_iter_unlocked:
1291 		maybe_yield();
1292 		mtx_lock(&vnode_list_mtx);
1293 		goto restart;
1294 next_iter:
1295 		MPASS(vp->v_type != VMARKER);
1296 		if (!should_yield())
1297 			continue;
1298 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1299 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1300 		mtx_unlock(&vnode_list_mtx);
1301 		kern_yield(PRI_USER);
1302 		mtx_lock(&vnode_list_mtx);
1303 		goto restart;
1304 	}
1305 	if (done == 0 && !retried) {
1306 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1307 		TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1308 		retried = true;
1309 		goto restart;
1310 	}
1311 	return (done);
1312 }
1313 
1314 static int max_free_per_call = 10000;
1315 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_free_per_call, 0,
1316     "limit on vnode free requests per call to the vnlru_free routine (legacy)");
1317 SYSCTL_INT(_vfs_vnode_vnlru, OID_AUTO, max_free_per_call, CTLFLAG_RW,
1318     &max_free_per_call, 0,
1319     "limit on vnode free requests per call to the vnlru_free routine");
1320 
1321 /*
1322  * Attempt to recycle requested amount of free vnodes.
1323  */
1324 static int
1325 vnlru_free_impl(int count, struct vfsops *mnt_op, struct vnode *mvp, bool isvnlru)
1326 {
1327 	struct vnode *vp;
1328 	struct mount *mp;
1329 	int ocount;
1330 	bool retried;
1331 
1332 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1333 	if (count > max_free_per_call)
1334 		count = max_free_per_call;
1335 	if (count == 0) {
1336 		mtx_unlock(&vnode_list_mtx);
1337 		return (0);
1338 	}
1339 	ocount = count;
1340 	retried = false;
1341 	vp = mvp;
1342 	for (;;) {
1343 		vp = TAILQ_NEXT(vp, v_vnodelist);
1344 		if (__predict_false(vp == NULL)) {
1345 			/*
1346 			 * The free vnode marker can be past eligible vnodes:
1347 			 * 1. if vdbatch_process trylock failed
1348 			 * 2. if vtryrecycle failed
1349 			 *
1350 			 * If so, start the scan from scratch.
1351 			 */
1352 			if (!retried && vnlru_read_freevnodes() > 0) {
1353 				TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1354 				TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1355 				vp = mvp;
1356 				retried = true;
1357 				continue;
1358 			}
1359 
1360 			/*
1361 			 * Give up
1362 			 */
1363 			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1364 			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1365 			mtx_unlock(&vnode_list_mtx);
1366 			break;
1367 		}
1368 		if (__predict_false(vp->v_type == VMARKER))
1369 			continue;
1370 		if (vp->v_holdcnt > 0)
1371 			continue;
1372 		/*
1373 		 * Don't recycle if our vnode is from different type
1374 		 * of mount point.  Note that mp is type-safe, the
1375 		 * check does not reach unmapped address even if
1376 		 * vnode is reclaimed.
1377 		 */
1378 		if (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1379 		    mp->mnt_op != mnt_op) {
1380 			continue;
1381 		}
1382 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1383 			continue;
1384 		}
1385 		if (!vhold_recycle_free(vp))
1386 			continue;
1387 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1388 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1389 		mtx_unlock(&vnode_list_mtx);
1390 		/*
1391 		 * FIXME: ignores the return value, meaning it may be nothing
1392 		 * got recycled but it claims otherwise to the caller.
1393 		 *
1394 		 * Originally the value started being ignored in 2005 with
1395 		 * 114a1006a8204aa156e1f9ad6476cdff89cada7f .
1396 		 *
1397 		 * Respecting the value can run into significant stalls if most
1398 		 * vnodes belong to one file system and it has writes
1399 		 * suspended.  In presence of many threads and millions of
1400 		 * vnodes they keep contending on the vnode_list_mtx lock only
1401 		 * to find vnodes they can't recycle.
1402 		 *
1403 		 * The solution would be to pre-check if the vnode is likely to
1404 		 * be recycle-able, but it needs to happen with the
1405 		 * vnode_list_mtx lock held. This runs into a problem where
1406 		 * VOP_GETWRITEMOUNT (currently needed to find out about if
1407 		 * writes are frozen) can take locks which LOR against it.
1408 		 *
1409 		 * Check nullfs for one example (null_getwritemount).
1410 		 */
1411 		vtryrecycle(vp, isvnlru);
1412 		count--;
1413 		if (count == 0) {
1414 			break;
1415 		}
1416 		mtx_lock(&vnode_list_mtx);
1417 		vp = mvp;
1418 	}
1419 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1420 	return (ocount - count);
1421 }
1422 
1423 /*
1424  * XXX: returns without vnode_list_mtx locked!
1425  */
1426 static int
1427 vnlru_free_locked_direct(int count)
1428 {
1429 	int ret;
1430 
1431 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1432 	ret = vnlru_free_impl(count, NULL, vnode_list_free_marker, false);
1433 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1434 	return (ret);
1435 }
1436 
1437 static int
1438 vnlru_free_locked_vnlru(int count)
1439 {
1440 	int ret;
1441 
1442 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1443 	ret = vnlru_free_impl(count, NULL, vnode_list_free_marker, true);
1444 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1445 	return (ret);
1446 }
1447 
1448 static int
1449 vnlru_free_vnlru(int count)
1450 {
1451 
1452 	mtx_lock(&vnode_list_mtx);
1453 	return (vnlru_free_locked_vnlru(count));
1454 }
1455 
1456 void
1457 vnlru_free_vfsops(int count, struct vfsops *mnt_op, struct vnode *mvp)
1458 {
1459 
1460 	MPASS(mnt_op != NULL);
1461 	MPASS(mvp != NULL);
1462 	VNPASS(mvp->v_type == VMARKER, mvp);
1463 	mtx_lock(&vnode_list_mtx);
1464 	vnlru_free_impl(count, mnt_op, mvp, true);
1465 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1466 }
1467 
1468 struct vnode *
1469 vnlru_alloc_marker(void)
1470 {
1471 	struct vnode *mvp;
1472 
1473 	mvp = vn_alloc_marker(NULL);
1474 	mtx_lock(&vnode_list_mtx);
1475 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, mvp, v_vnodelist);
1476 	mtx_unlock(&vnode_list_mtx);
1477 	return (mvp);
1478 }
1479 
1480 void
1481 vnlru_free_marker(struct vnode *mvp)
1482 {
1483 	mtx_lock(&vnode_list_mtx);
1484 	TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1485 	mtx_unlock(&vnode_list_mtx);
1486 	vn_free_marker(mvp);
1487 }
1488 
1489 static void
1490 vnlru_recalc(void)
1491 {
1492 
1493 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1494 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1495 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1496 	vlowat = vhiwat / 2;
1497 }
1498 
1499 /*
1500  * Attempt to recycle vnodes in a context that is always safe to block.
1501  * Calling vlrurecycle() from the bowels of filesystem code has some
1502  * interesting deadlock problems.
1503  */
1504 static struct proc *vnlruproc;
1505 static int vnlruproc_sig;
1506 static u_long vnlruproc_kicks;
1507 
1508 SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, kicks, CTLFLAG_RD, &vnlruproc_kicks, 0,
1509     "Number of times vnlru awakened due to vnode shortage");
1510 
1511 #define VNLRU_COUNT_SLOP 100
1512 
1513 /*
1514  * The main freevnodes counter is only updated when a counter local to CPU
1515  * diverges from 0 by more than VNLRU_FREEVNODES_SLOP. CPUs are conditionally
1516  * walked to compute a more accurate total.
1517  *
1518  * Note: the actual value at any given moment can still exceed slop, but it
1519  * should not be by significant margin in practice.
1520  */
1521 #define VNLRU_FREEVNODES_SLOP 126
1522 
1523 static void __noinline
1524 vfs_freevnodes_rollup(int8_t *lfreevnodes)
1525 {
1526 
1527 	atomic_add_long(&freevnodes, *lfreevnodes);
1528 	*lfreevnodes = 0;
1529 	critical_exit();
1530 }
1531 
1532 static __inline void
1533 vfs_freevnodes_inc(void)
1534 {
1535 	int8_t *lfreevnodes;
1536 
1537 	critical_enter();
1538 	lfreevnodes = PCPU_PTR(vfs_freevnodes);
1539 	(*lfreevnodes)++;
1540 	if (__predict_false(*lfreevnodes == VNLRU_FREEVNODES_SLOP))
1541 		vfs_freevnodes_rollup(lfreevnodes);
1542 	else
1543 		critical_exit();
1544 }
1545 
1546 static __inline void
1547 vfs_freevnodes_dec(void)
1548 {
1549 	int8_t *lfreevnodes;
1550 
1551 	critical_enter();
1552 	lfreevnodes = PCPU_PTR(vfs_freevnodes);
1553 	(*lfreevnodes)--;
1554 	if (__predict_false(*lfreevnodes == -VNLRU_FREEVNODES_SLOP))
1555 		vfs_freevnodes_rollup(lfreevnodes);
1556 	else
1557 		critical_exit();
1558 }
1559 
1560 static u_long
1561 vnlru_read_freevnodes(void)
1562 {
1563 	long slop, rfreevnodes, rfreevnodes_old;
1564 	int cpu;
1565 
1566 	rfreevnodes = atomic_load_long(&freevnodes);
1567 	rfreevnodes_old = atomic_load_long(&freevnodes_old);
1568 
1569 	if (rfreevnodes > rfreevnodes_old)
1570 		slop = rfreevnodes - rfreevnodes_old;
1571 	else
1572 		slop = rfreevnodes_old - rfreevnodes;
1573 	if (slop < VNLRU_FREEVNODES_SLOP)
1574 		return (rfreevnodes >= 0 ? rfreevnodes : 0);
1575 	CPU_FOREACH(cpu) {
1576 		rfreevnodes += cpuid_to_pcpu[cpu]->pc_vfs_freevnodes;
1577 	}
1578 	atomic_store_long(&freevnodes_old, rfreevnodes);
1579 	return (freevnodes_old >= 0 ? freevnodes_old : 0);
1580 }
1581 
1582 static bool
1583 vnlru_under(u_long rnumvnodes, u_long limit)
1584 {
1585 	u_long rfreevnodes, space;
1586 
1587 	if (__predict_false(rnumvnodes > desiredvnodes))
1588 		return (true);
1589 
1590 	space = desiredvnodes - rnumvnodes;
1591 	if (space < limit) {
1592 		rfreevnodes = vnlru_read_freevnodes();
1593 		if (rfreevnodes > wantfreevnodes)
1594 			space += rfreevnodes - wantfreevnodes;
1595 	}
1596 	return (space < limit);
1597 }
1598 
1599 static void
1600 vnlru_kick_locked(void)
1601 {
1602 
1603 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1604 	if (vnlruproc_sig == 0) {
1605 		vnlruproc_sig = 1;
1606 		vnlruproc_kicks++;
1607 		wakeup(vnlruproc);
1608 	}
1609 }
1610 
1611 static void
1612 vnlru_kick_cond(void)
1613 {
1614 
1615 	if (vnlru_read_freevnodes() > wantfreevnodes)
1616 		return;
1617 
1618 	if (vnlruproc_sig)
1619 		return;
1620 	mtx_lock(&vnode_list_mtx);
1621 	vnlru_kick_locked();
1622 	mtx_unlock(&vnode_list_mtx);
1623 }
1624 
1625 static void
1626 vnlru_proc_sleep(void)
1627 {
1628 
1629 	if (vnlruproc_sig) {
1630 		vnlruproc_sig = 0;
1631 		wakeup(&vnlruproc_sig);
1632 	}
1633 	msleep(vnlruproc, &vnode_list_mtx, PVFS|PDROP, "vlruwt", hz);
1634 }
1635 
1636 /*
1637  * A lighter version of the machinery below.
1638  *
1639  * Tries to reach goals only by recycling free vnodes and does not invoke
1640  * uma_reclaim(UMA_RECLAIM_DRAIN).
1641  *
1642  * This works around pathological behavior in vnlru in presence of tons of free
1643  * vnodes, but without having to rewrite the machinery at this time. Said
1644  * behavior boils down to continuously trying to reclaim all kinds of vnodes
1645  * (cycling through all levels of "force") when the count is transiently above
1646  * limit. This happens a lot when all vnodes are used up and vn_alloc
1647  * speculatively increments the counter.
1648  *
1649  * Sample testcase: vnode limit 8388608, 20 separate directory trees each with
1650  * 1 million files in total and 20 find(1) processes stating them in parallel
1651  * (one per each tree).
1652  *
1653  * On a kernel with only stock machinery this needs anywhere between 60 and 120
1654  * seconds to execute (time varies *wildly* between runs). With the workaround
1655  * it consistently stays around 20 seconds [it got further down with later
1656  * changes].
1657  *
1658  * That is to say the entire thing needs a fundamental redesign (most notably
1659  * to accommodate faster recycling), the above only tries to get it ouf the way.
1660  *
1661  * Return values are:
1662  * -1 -- fallback to regular vnlru loop
1663  *  0 -- do nothing, go to sleep
1664  * >0 -- recycle this many vnodes
1665  */
1666 static long
1667 vnlru_proc_light_pick(void)
1668 {
1669 	u_long rnumvnodes, rfreevnodes;
1670 
1671 	if (vstir || vnlruproc_sig == 1)
1672 		return (-1);
1673 
1674 	rnumvnodes = atomic_load_long(&numvnodes);
1675 	rfreevnodes = vnlru_read_freevnodes();
1676 
1677 	/*
1678 	 * vnode limit might have changed and now we may be at a significant
1679 	 * excess. Bail if we can't sort it out with free vnodes.
1680 	 *
1681 	 * Due to atomic updates the count can legitimately go above
1682 	 * the limit for a short period, don't bother doing anything in
1683 	 * that case.
1684 	 */
1685 	if (rnumvnodes > desiredvnodes + VNLRU_COUNT_SLOP + 10) {
1686 		if (rnumvnodes - rfreevnodes >= desiredvnodes ||
1687 		    rfreevnodes <= wantfreevnodes) {
1688 			return (-1);
1689 		}
1690 
1691 		return (rnumvnodes - desiredvnodes);
1692 	}
1693 
1694 	/*
1695 	 * Don't try to reach wantfreevnodes target if there are too few vnodes
1696 	 * to begin with.
1697 	 */
1698 	if (rnumvnodes < wantfreevnodes) {
1699 		return (0);
1700 	}
1701 
1702 	if (rfreevnodes < wantfreevnodes) {
1703 		return (-1);
1704 	}
1705 
1706 	return (0);
1707 }
1708 
1709 static bool
1710 vnlru_proc_light(void)
1711 {
1712 	long freecount;
1713 
1714 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1715 
1716 	freecount = vnlru_proc_light_pick();
1717 	if (freecount == -1)
1718 		return (false);
1719 
1720 	if (freecount != 0) {
1721 		vnlru_free_vnlru(freecount);
1722 	}
1723 
1724 	mtx_lock(&vnode_list_mtx);
1725 	vnlru_proc_sleep();
1726 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1727 	return (true);
1728 }
1729 
1730 static u_long uma_reclaim_calls;
1731 SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, uma_reclaim_calls, CTLFLAG_RD | CTLFLAG_STATS,
1732     &uma_reclaim_calls, 0, "Number of calls to uma_reclaim");
1733 
1734 static void
1735 vnlru_proc(void)
1736 {
1737 	u_long rnumvnodes, rfreevnodes, target;
1738 	unsigned long onumvnodes;
1739 	int done, force, trigger, usevnodes;
1740 	bool reclaim_nc_src, want_reread;
1741 
1742 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1743 	    SHUTDOWN_PRI_FIRST);
1744 
1745 	force = 0;
1746 	want_reread = false;
1747 	for (;;) {
1748 		kproc_suspend_check(vnlruproc);
1749 
1750 		if (force == 0 && vnlru_proc_light())
1751 			continue;
1752 
1753 		mtx_lock(&vnode_list_mtx);
1754 		rnumvnodes = atomic_load_long(&numvnodes);
1755 
1756 		if (want_reread) {
1757 			force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1758 			want_reread = false;
1759 		}
1760 
1761 		/*
1762 		 * If numvnodes is too large (due to desiredvnodes being
1763 		 * adjusted using its sysctl, or emergency growth), first
1764 		 * try to reduce it by discarding free vnodes.
1765 		 */
1766 		if (rnumvnodes > desiredvnodes + 10) {
1767 			vnlru_free_locked_vnlru(rnumvnodes - desiredvnodes);
1768 			mtx_lock(&vnode_list_mtx);
1769 			rnumvnodes = atomic_load_long(&numvnodes);
1770 		}
1771 		/*
1772 		 * Sleep if the vnode cache is in a good state.  This is
1773 		 * when it is not over-full and has space for about a 4%
1774 		 * or 9% expansion (by growing its size or inexcessively
1775 		 * reducing free vnode count).  Otherwise, try to reclaim
1776 		 * space for a 10% expansion.
1777 		 */
1778 		if (vstir && force == 0) {
1779 			force = 1;
1780 			vstir = false;
1781 		}
1782 		if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1783 			vnlru_proc_sleep();
1784 			continue;
1785 		}
1786 		rfreevnodes = vnlru_read_freevnodes();
1787 
1788 		onumvnodes = rnumvnodes;
1789 		/*
1790 		 * Calculate parameters for recycling.  These are the same
1791 		 * throughout the loop to give some semblance of fairness.
1792 		 * The trigger point is to avoid recycling vnodes with lots
1793 		 * of resident pages.  We aren't trying to free memory; we
1794 		 * are trying to recycle or at least free vnodes.
1795 		 */
1796 		if (rnumvnodes <= desiredvnodes)
1797 			usevnodes = rnumvnodes - rfreevnodes;
1798 		else
1799 			usevnodes = rnumvnodes;
1800 		if (usevnodes <= 0)
1801 			usevnodes = 1;
1802 		/*
1803 		 * The trigger value is chosen to give a conservatively
1804 		 * large value to ensure that it alone doesn't prevent
1805 		 * making progress.  The value can easily be so large that
1806 		 * it is effectively infinite in some congested and
1807 		 * misconfigured cases, and this is necessary.  Normally
1808 		 * it is about 8 to 100 (pages), which is quite large.
1809 		 */
1810 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1811 		if (force < 2)
1812 			trigger = vsmalltrigger;
1813 		reclaim_nc_src = force >= 3;
1814 		target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1815 		target = target / 10 + 1;
1816 		done = vlrureclaim(reclaim_nc_src, trigger, target);
1817 		mtx_unlock(&vnode_list_mtx);
1818 		/*
1819 		 * Total number of vnodes can transiently go slightly above the
1820 		 * limit (see vn_alloc_hard), no need to call uma_reclaim if
1821 		 * this happens.
1822 		 */
1823 		if (onumvnodes + VNLRU_COUNT_SLOP + 1000 > desiredvnodes &&
1824 		    numvnodes <= desiredvnodes) {
1825 			uma_reclaim_calls++;
1826 			uma_reclaim(UMA_RECLAIM_DRAIN);
1827 		}
1828 		if (done == 0) {
1829 			if (force == 0 || force == 1) {
1830 				force = 2;
1831 				continue;
1832 			}
1833 			if (force == 2) {
1834 				force = 3;
1835 				continue;
1836 			}
1837 			want_reread = true;
1838 			force = 0;
1839 			vnlru_nowhere++;
1840 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1841 		} else {
1842 			want_reread = true;
1843 			kern_yield(PRI_USER);
1844 		}
1845 	}
1846 }
1847 
1848 static struct kproc_desc vnlru_kp = {
1849 	"vnlru",
1850 	vnlru_proc,
1851 	&vnlruproc
1852 };
1853 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1854     &vnlru_kp);
1855 
1856 /*
1857  * Routines having to do with the management of the vnode table.
1858  */
1859 
1860 /*
1861  * Try to recycle a freed vnode.
1862  */
1863 static int
1864 vtryrecycle(struct vnode *vp, bool isvnlru)
1865 {
1866 	struct mount *vnmp;
1867 
1868 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1869 	VNPASS(vp->v_holdcnt > 0, vp);
1870 	/*
1871 	 * This vnode may found and locked via some other list, if so we
1872 	 * can't recycle it yet.
1873 	 */
1874 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1875 		CTR2(KTR_VFS,
1876 		    "%s: impossible to recycle, vp %p lock is already held",
1877 		    __func__, vp);
1878 		vdrop_recycle(vp);
1879 		return (EWOULDBLOCK);
1880 	}
1881 	/*
1882 	 * Don't recycle if its filesystem is being suspended.
1883 	 */
1884 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1885 		VOP_UNLOCK(vp);
1886 		CTR2(KTR_VFS,
1887 		    "%s: impossible to recycle, cannot start the write for %p",
1888 		    __func__, vp);
1889 		vdrop_recycle(vp);
1890 		return (EBUSY);
1891 	}
1892 	/*
1893 	 * If we got this far, we need to acquire the interlock and see if
1894 	 * anyone picked up this vnode from another list.  If not, we will
1895 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1896 	 * will skip over it.
1897 	 */
1898 	VI_LOCK(vp);
1899 	if (vp->v_usecount) {
1900 		VOP_UNLOCK(vp);
1901 		vdropl_recycle(vp);
1902 		vn_finished_write(vnmp);
1903 		CTR2(KTR_VFS,
1904 		    "%s: impossible to recycle, %p is already referenced",
1905 		    __func__, vp);
1906 		return (EBUSY);
1907 	}
1908 	if (!VN_IS_DOOMED(vp)) {
1909 		if (isvnlru)
1910 			recycles_free_count++;
1911 		else
1912 			counter_u64_add(direct_recycles_free_count, 1);
1913 		vgonel(vp);
1914 	}
1915 	VOP_UNLOCK(vp);
1916 	vdropl_recycle(vp);
1917 	vn_finished_write(vnmp);
1918 	return (0);
1919 }
1920 
1921 /*
1922  * Allocate a new vnode.
1923  *
1924  * The operation never returns an error. Returning an error was disabled
1925  * in r145385 (dated 2005) with the following comment:
1926  *
1927  * XXX Not all VFS_VGET/ffs_vget callers check returns.
1928  *
1929  * Given the age of this commit (almost 15 years at the time of writing this
1930  * comment) restoring the ability to fail requires a significant audit of
1931  * all codepaths.
1932  *
1933  * The routine can try to free a vnode or stall for up to 1 second waiting for
1934  * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1935  */
1936 static u_long vn_alloc_cyclecount;
1937 static u_long vn_alloc_sleeps;
1938 
1939 SYSCTL_ULONG(_vfs_vnode_stats, OID_AUTO, alloc_sleeps, CTLFLAG_RD, &vn_alloc_sleeps, 0,
1940     "Number of times vnode allocation blocked waiting on vnlru");
1941 
1942 static struct vnode * __noinline
1943 vn_alloc_hard(struct mount *mp, u_long rnumvnodes, bool bumped)
1944 {
1945 	u_long rfreevnodes;
1946 
1947 	if (bumped) {
1948 		if (rnumvnodes > desiredvnodes + VNLRU_COUNT_SLOP) {
1949 			atomic_subtract_long(&numvnodes, 1);
1950 			bumped = false;
1951 		}
1952 	}
1953 
1954 	mtx_lock(&vnode_list_mtx);
1955 
1956 	if (vn_alloc_cyclecount != 0) {
1957 		rnumvnodes = atomic_load_long(&numvnodes);
1958 		if (rnumvnodes + 1 < desiredvnodes) {
1959 			vn_alloc_cyclecount = 0;
1960 			mtx_unlock(&vnode_list_mtx);
1961 			goto alloc;
1962 		}
1963 
1964 		rfreevnodes = vnlru_read_freevnodes();
1965 		if (rfreevnodes < wantfreevnodes) {
1966 			if (vn_alloc_cyclecount++ >= rfreevnodes) {
1967 				vn_alloc_cyclecount = 0;
1968 				vstir = true;
1969 			}
1970 		} else {
1971 			vn_alloc_cyclecount = 0;
1972 		}
1973 	}
1974 
1975 	/*
1976 	 * Grow the vnode cache if it will not be above its target max after
1977 	 * growing.  Otherwise, if there is at least one free vnode, try to
1978 	 * reclaim 1 item from it before growing the cache (possibly above its
1979 	 * target max if the reclamation failed or is delayed).
1980 	 */
1981 	if (vnlru_free_locked_direct(1) > 0)
1982 		goto alloc;
1983 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1984 	if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1985 		/*
1986 		 * Wait for space for a new vnode.
1987 		 */
1988 		if (bumped) {
1989 			atomic_subtract_long(&numvnodes, 1);
1990 			bumped = false;
1991 		}
1992 		mtx_lock(&vnode_list_mtx);
1993 		vnlru_kick_locked();
1994 		vn_alloc_sleeps++;
1995 		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1996 		if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1997 		    vnlru_read_freevnodes() > 1)
1998 			vnlru_free_locked_direct(1);
1999 		else
2000 			mtx_unlock(&vnode_list_mtx);
2001 	}
2002 alloc:
2003 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
2004 	if (!bumped)
2005 		atomic_add_long(&numvnodes, 1);
2006 	vnlru_kick_cond();
2007 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
2008 }
2009 
2010 static struct vnode *
2011 vn_alloc(struct mount *mp)
2012 {
2013 	u_long rnumvnodes;
2014 
2015 	if (__predict_false(vn_alloc_cyclecount != 0))
2016 		return (vn_alloc_hard(mp, 0, false));
2017 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
2018 	if (__predict_false(vnlru_under(rnumvnodes, vlowat))) {
2019 		return (vn_alloc_hard(mp, rnumvnodes, true));
2020 	}
2021 
2022 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
2023 }
2024 
2025 static void
2026 vn_free(struct vnode *vp)
2027 {
2028 
2029 	atomic_subtract_long(&numvnodes, 1);
2030 	uma_zfree_smr(vnode_zone, vp);
2031 }
2032 
2033 /*
2034  * Allocate a new vnode.
2035  */
2036 int
2037 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
2038     struct vnode **vpp)
2039 {
2040 	struct vnode *vp;
2041 	struct thread *td;
2042 	struct lock_object *lo;
2043 
2044 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
2045 
2046 	KASSERT(vops->registered,
2047 	    ("%s: not registered vector op %p\n", __func__, vops));
2048 	cache_validate_vop_vector(mp, vops);
2049 
2050 	td = curthread;
2051 	if (td->td_vp_reserved != NULL) {
2052 		vp = td->td_vp_reserved;
2053 		td->td_vp_reserved = NULL;
2054 	} else {
2055 		vp = vn_alloc(mp);
2056 	}
2057 	counter_u64_add(vnodes_created, 1);
2058 
2059 	vn_set_state(vp, VSTATE_UNINITIALIZED);
2060 
2061 	/*
2062 	 * Locks are given the generic name "vnode" when created.
2063 	 * Follow the historic practice of using the filesystem
2064 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
2065 	 *
2066 	 * Locks live in a witness group keyed on their name. Thus,
2067 	 * when a lock is renamed, it must also move from the witness
2068 	 * group of its old name to the witness group of its new name.
2069 	 *
2070 	 * The change only needs to be made when the vnode moves
2071 	 * from one filesystem type to another. We ensure that each
2072 	 * filesystem use a single static name pointer for its tag so
2073 	 * that we can compare pointers rather than doing a strcmp().
2074 	 */
2075 	lo = &vp->v_vnlock->lock_object;
2076 #ifdef WITNESS
2077 	if (lo->lo_name != tag) {
2078 #endif
2079 		lo->lo_name = tag;
2080 #ifdef WITNESS
2081 		WITNESS_DESTROY(lo);
2082 		WITNESS_INIT(lo, tag);
2083 	}
2084 #endif
2085 	/*
2086 	 * By default, don't allow shared locks unless filesystems opt-in.
2087 	 */
2088 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
2089 	/*
2090 	 * Finalize various vnode identity bits.
2091 	 */
2092 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
2093 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
2094 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
2095 	vp->v_type = VNON;
2096 	vp->v_op = vops;
2097 	vp->v_irflag = 0;
2098 	v_init_counters(vp);
2099 	vn_seqc_init(vp);
2100 	vp->v_bufobj.bo_ops = &buf_ops_bio;
2101 #ifdef DIAGNOSTIC
2102 	if (mp == NULL && vops != &dead_vnodeops)
2103 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
2104 #endif
2105 #ifdef MAC
2106 	mac_vnode_init(vp);
2107 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
2108 		mac_vnode_associate_singlelabel(mp, vp);
2109 #endif
2110 	if (mp != NULL) {
2111 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
2112 	}
2113 
2114 	/*
2115 	 * For the filesystems which do not use vfs_hash_insert(),
2116 	 * still initialize v_hash to have vfs_hash_index() useful.
2117 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
2118 	 * its own hashing.
2119 	 */
2120 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
2121 
2122 	*vpp = vp;
2123 	return (0);
2124 }
2125 
2126 void
2127 getnewvnode_reserve(void)
2128 {
2129 	struct thread *td;
2130 
2131 	td = curthread;
2132 	MPASS(td->td_vp_reserved == NULL);
2133 	td->td_vp_reserved = vn_alloc(NULL);
2134 }
2135 
2136 void
2137 getnewvnode_drop_reserve(void)
2138 {
2139 	struct thread *td;
2140 
2141 	td = curthread;
2142 	if (td->td_vp_reserved != NULL) {
2143 		vn_free(td->td_vp_reserved);
2144 		td->td_vp_reserved = NULL;
2145 	}
2146 }
2147 
2148 static void __noinline
2149 freevnode(struct vnode *vp)
2150 {
2151 	struct bufobj *bo;
2152 
2153 	/*
2154 	 * The vnode has been marked for destruction, so free it.
2155 	 *
2156 	 * The vnode will be returned to the zone where it will
2157 	 * normally remain until it is needed for another vnode. We
2158 	 * need to cleanup (or verify that the cleanup has already
2159 	 * been done) any residual data left from its current use
2160 	 * so as not to contaminate the freshly allocated vnode.
2161 	 */
2162 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
2163 	/*
2164 	 * Paired with vgone.
2165 	 */
2166 	vn_seqc_write_end_free(vp);
2167 
2168 	bo = &vp->v_bufobj;
2169 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
2170 	VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp);
2171 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
2172 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
2173 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
2174 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
2175 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
2176 	    ("clean blk trie not empty"));
2177 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
2178 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
2179 	    ("dirty blk trie not empty"));
2180 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
2181 	    ("Dangling rangelock waiters"));
2182 	VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp,
2183 	    ("Leaked inactivation"));
2184 	VI_UNLOCK(vp);
2185 	cache_assert_no_entries(vp);
2186 
2187 #ifdef MAC
2188 	mac_vnode_destroy(vp);
2189 #endif
2190 	if (vp->v_pollinfo != NULL) {
2191 		/*
2192 		 * Use LK_NOWAIT to shut up witness about the lock. We may get
2193 		 * here while having another vnode locked when trying to
2194 		 * satisfy a lookup and needing to recycle.
2195 		 */
2196 		VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT);
2197 		destroy_vpollinfo(vp->v_pollinfo);
2198 		VOP_UNLOCK(vp);
2199 		vp->v_pollinfo = NULL;
2200 	}
2201 	vp->v_mountedhere = NULL;
2202 	vp->v_unpcb = NULL;
2203 	vp->v_rdev = NULL;
2204 	vp->v_fifoinfo = NULL;
2205 	vp->v_iflag = 0;
2206 	vp->v_vflag = 0;
2207 	bo->bo_flag = 0;
2208 	vn_free(vp);
2209 }
2210 
2211 /*
2212  * Delete from old mount point vnode list, if on one.
2213  */
2214 static void
2215 delmntque(struct vnode *vp)
2216 {
2217 	struct mount *mp;
2218 
2219 	VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
2220 
2221 	mp = vp->v_mount;
2222 	MNT_ILOCK(mp);
2223 	VI_LOCK(vp);
2224 	vp->v_mount = NULL;
2225 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
2226 		("bad mount point vnode list size"));
2227 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2228 	mp->mnt_nvnodelistsize--;
2229 	MNT_REL(mp);
2230 	MNT_IUNLOCK(mp);
2231 	/*
2232 	 * The caller expects the interlock to be still held.
2233 	 */
2234 	ASSERT_VI_LOCKED(vp, __func__);
2235 }
2236 
2237 static int
2238 insmntque1_int(struct vnode *vp, struct mount *mp, bool dtr)
2239 {
2240 
2241 	KASSERT(vp->v_mount == NULL,
2242 		("insmntque: vnode already on per mount vnode list"));
2243 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
2244 	if ((mp->mnt_kern_flag & MNTK_UNLOCKED_INSMNTQUE) == 0) {
2245 		ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
2246 	} else {
2247 		KASSERT(!dtr,
2248 		    ("%s: can't have MNTK_UNLOCKED_INSMNTQUE and cleanup",
2249 		    __func__));
2250 	}
2251 
2252 	/*
2253 	 * We acquire the vnode interlock early to ensure that the
2254 	 * vnode cannot be recycled by another process releasing a
2255 	 * holdcnt on it before we get it on both the vnode list
2256 	 * and the active vnode list. The mount mutex protects only
2257 	 * manipulation of the vnode list and the vnode freelist
2258 	 * mutex protects only manipulation of the active vnode list.
2259 	 * Hence the need to hold the vnode interlock throughout.
2260 	 */
2261 	MNT_ILOCK(mp);
2262 	VI_LOCK(vp);
2263 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
2264 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
2265 	    mp->mnt_nvnodelistsize == 0)) &&
2266 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
2267 		VI_UNLOCK(vp);
2268 		MNT_IUNLOCK(mp);
2269 		if (dtr) {
2270 			vp->v_data = NULL;
2271 			vp->v_op = &dead_vnodeops;
2272 			vgone(vp);
2273 			vput(vp);
2274 		}
2275 		return (EBUSY);
2276 	}
2277 	vp->v_mount = mp;
2278 	MNT_REF(mp);
2279 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2280 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
2281 		("neg mount point vnode list size"));
2282 	mp->mnt_nvnodelistsize++;
2283 	VI_UNLOCK(vp);
2284 	MNT_IUNLOCK(mp);
2285 	return (0);
2286 }
2287 
2288 /*
2289  * Insert into list of vnodes for the new mount point, if available.
2290  * insmntque() reclaims the vnode on insertion failure, insmntque1()
2291  * leaves handling of the vnode to the caller.
2292  */
2293 int
2294 insmntque(struct vnode *vp, struct mount *mp)
2295 {
2296 	return (insmntque1_int(vp, mp, true));
2297 }
2298 
2299 int
2300 insmntque1(struct vnode *vp, struct mount *mp)
2301 {
2302 	return (insmntque1_int(vp, mp, false));
2303 }
2304 
2305 /*
2306  * Flush out and invalidate all buffers associated with a bufobj
2307  * Called with the underlying object locked.
2308  */
2309 int
2310 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
2311 {
2312 	int error;
2313 
2314 	BO_LOCK(bo);
2315 	if (flags & V_SAVE) {
2316 		error = bufobj_wwait(bo, slpflag, slptimeo);
2317 		if (error) {
2318 			BO_UNLOCK(bo);
2319 			return (error);
2320 		}
2321 		if (bo->bo_dirty.bv_cnt > 0) {
2322 			BO_UNLOCK(bo);
2323 			do {
2324 				error = BO_SYNC(bo, MNT_WAIT);
2325 			} while (error == ERELOOKUP);
2326 			if (error != 0)
2327 				return (error);
2328 			BO_LOCK(bo);
2329 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) {
2330 				BO_UNLOCK(bo);
2331 				return (EBUSY);
2332 			}
2333 		}
2334 	}
2335 	/*
2336 	 * If you alter this loop please notice that interlock is dropped and
2337 	 * reacquired in flushbuflist.  Special care is needed to ensure that
2338 	 * no race conditions occur from this.
2339 	 */
2340 	do {
2341 		error = flushbuflist(&bo->bo_clean,
2342 		    flags, bo, slpflag, slptimeo);
2343 		if (error == 0 && !(flags & V_CLEANONLY))
2344 			error = flushbuflist(&bo->bo_dirty,
2345 			    flags, bo, slpflag, slptimeo);
2346 		if (error != 0 && error != EAGAIN) {
2347 			BO_UNLOCK(bo);
2348 			return (error);
2349 		}
2350 	} while (error != 0);
2351 
2352 	/*
2353 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
2354 	 * have write I/O in-progress but if there is a VM object then the
2355 	 * VM object can also have read-I/O in-progress.
2356 	 */
2357 	do {
2358 		bufobj_wwait(bo, 0, 0);
2359 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
2360 			BO_UNLOCK(bo);
2361 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
2362 			BO_LOCK(bo);
2363 		}
2364 	} while (bo->bo_numoutput > 0);
2365 	BO_UNLOCK(bo);
2366 
2367 	/*
2368 	 * Destroy the copy in the VM cache, too.
2369 	 */
2370 	if (bo->bo_object != NULL &&
2371 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
2372 		VM_OBJECT_WLOCK(bo->bo_object);
2373 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
2374 		    OBJPR_CLEANONLY : 0);
2375 		VM_OBJECT_WUNLOCK(bo->bo_object);
2376 	}
2377 
2378 #ifdef INVARIANTS
2379 	BO_LOCK(bo);
2380 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
2381 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
2382 	    bo->bo_clean.bv_cnt > 0))
2383 		panic("vinvalbuf: flush failed");
2384 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
2385 	    bo->bo_dirty.bv_cnt > 0)
2386 		panic("vinvalbuf: flush dirty failed");
2387 	BO_UNLOCK(bo);
2388 #endif
2389 	return (0);
2390 }
2391 
2392 /*
2393  * Flush out and invalidate all buffers associated with a vnode.
2394  * Called with the underlying object locked.
2395  */
2396 int
2397 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
2398 {
2399 
2400 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2401 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
2402 	if (vp->v_object != NULL && vp->v_object->handle != vp)
2403 		return (0);
2404 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
2405 }
2406 
2407 /*
2408  * Flush out buffers on the specified list.
2409  *
2410  */
2411 static int
2412 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2413     int slptimeo)
2414 {
2415 	struct buf *bp, *nbp;
2416 	int retval, error;
2417 	daddr_t lblkno;
2418 	b_xflags_t xflags;
2419 
2420 	ASSERT_BO_WLOCKED(bo);
2421 
2422 	retval = 0;
2423 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2424 		/*
2425 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2426 		 * do not skip any buffers. If we are flushing only V_NORMAL
2427 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2428 		 * flushing only V_ALT buffers then skip buffers not marked
2429 		 * as BX_ALTDATA.
2430 		 */
2431 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2432 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2433 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2434 			continue;
2435 		}
2436 		if (nbp != NULL) {
2437 			lblkno = nbp->b_lblkno;
2438 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2439 		}
2440 		retval = EAGAIN;
2441 		error = BUF_TIMELOCK(bp,
2442 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2443 		    "flushbuf", slpflag, slptimeo);
2444 		if (error) {
2445 			BO_LOCK(bo);
2446 			return (error != ENOLCK ? error : EAGAIN);
2447 		}
2448 		KASSERT(bp->b_bufobj == bo,
2449 		    ("bp %p wrong b_bufobj %p should be %p",
2450 		    bp, bp->b_bufobj, bo));
2451 		/*
2452 		 * XXX Since there are no node locks for NFS, I
2453 		 * believe there is a slight chance that a delayed
2454 		 * write will occur while sleeping just above, so
2455 		 * check for it.
2456 		 */
2457 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2458 		    (flags & V_SAVE)) {
2459 			bremfree(bp);
2460 			bp->b_flags |= B_ASYNC;
2461 			bwrite(bp);
2462 			BO_LOCK(bo);
2463 			return (EAGAIN);	/* XXX: why not loop ? */
2464 		}
2465 		bremfree(bp);
2466 		bp->b_flags |= (B_INVAL | B_RELBUF);
2467 		bp->b_flags &= ~B_ASYNC;
2468 		brelse(bp);
2469 		BO_LOCK(bo);
2470 		if (nbp == NULL)
2471 			break;
2472 		nbp = gbincore(bo, lblkno);
2473 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2474 		    != xflags)
2475 			break;			/* nbp invalid */
2476 	}
2477 	return (retval);
2478 }
2479 
2480 int
2481 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2482 {
2483 	struct buf *bp;
2484 	int error;
2485 	daddr_t lblkno;
2486 
2487 	ASSERT_BO_LOCKED(bo);
2488 
2489 	for (lblkno = startn;;) {
2490 again:
2491 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2492 		if (bp == NULL || bp->b_lblkno >= endn ||
2493 		    bp->b_lblkno < startn)
2494 			break;
2495 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2496 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2497 		if (error != 0) {
2498 			BO_RLOCK(bo);
2499 			if (error == ENOLCK)
2500 				goto again;
2501 			return (error);
2502 		}
2503 		KASSERT(bp->b_bufobj == bo,
2504 		    ("bp %p wrong b_bufobj %p should be %p",
2505 		    bp, bp->b_bufobj, bo));
2506 		lblkno = bp->b_lblkno + 1;
2507 		if ((bp->b_flags & B_MANAGED) == 0)
2508 			bremfree(bp);
2509 		bp->b_flags |= B_RELBUF;
2510 		/*
2511 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2512 		 * pages backing each buffer in the range are unlikely to be
2513 		 * reused.  Dirty buffers will have the hint applied once
2514 		 * they've been written.
2515 		 */
2516 		if ((bp->b_flags & B_VMIO) != 0)
2517 			bp->b_flags |= B_NOREUSE;
2518 		brelse(bp);
2519 		BO_RLOCK(bo);
2520 	}
2521 	return (0);
2522 }
2523 
2524 /*
2525  * Truncate a file's buffer and pages to a specified length.  This
2526  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2527  * sync activity.
2528  */
2529 int
2530 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2531 {
2532 	struct buf *bp, *nbp;
2533 	struct bufobj *bo;
2534 	daddr_t startlbn;
2535 
2536 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2537 	    vp, blksize, (uintmax_t)length);
2538 
2539 	/*
2540 	 * Round up to the *next* lbn.
2541 	 */
2542 	startlbn = howmany(length, blksize);
2543 
2544 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2545 
2546 	bo = &vp->v_bufobj;
2547 restart_unlocked:
2548 	BO_LOCK(bo);
2549 
2550 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2551 		;
2552 
2553 	if (length > 0) {
2554 restartsync:
2555 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2556 			if (bp->b_lblkno > 0)
2557 				continue;
2558 			/*
2559 			 * Since we hold the vnode lock this should only
2560 			 * fail if we're racing with the buf daemon.
2561 			 */
2562 			if (BUF_LOCK(bp,
2563 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2564 			    BO_LOCKPTR(bo)) == ENOLCK)
2565 				goto restart_unlocked;
2566 
2567 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2568 			    ("buf(%p) on dirty queue without DELWRI", bp));
2569 
2570 			bremfree(bp);
2571 			bawrite(bp);
2572 			BO_LOCK(bo);
2573 			goto restartsync;
2574 		}
2575 	}
2576 
2577 	bufobj_wwait(bo, 0, 0);
2578 	BO_UNLOCK(bo);
2579 	vnode_pager_setsize(vp, length);
2580 
2581 	return (0);
2582 }
2583 
2584 /*
2585  * Invalidate the cached pages of a file's buffer within the range of block
2586  * numbers [startlbn, endlbn).
2587  */
2588 void
2589 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2590     int blksize)
2591 {
2592 	struct bufobj *bo;
2593 	off_t start, end;
2594 
2595 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2596 
2597 	start = blksize * startlbn;
2598 	end = blksize * endlbn;
2599 
2600 	bo = &vp->v_bufobj;
2601 	BO_LOCK(bo);
2602 	MPASS(blksize == bo->bo_bsize);
2603 
2604 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2605 		;
2606 
2607 	BO_UNLOCK(bo);
2608 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2609 }
2610 
2611 static int
2612 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2613     daddr_t startlbn, daddr_t endlbn)
2614 {
2615 	struct buf *bp, *nbp;
2616 	bool anyfreed;
2617 
2618 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2619 	ASSERT_BO_LOCKED(bo);
2620 
2621 	do {
2622 		anyfreed = false;
2623 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2624 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2625 				continue;
2626 			if (BUF_LOCK(bp,
2627 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2628 			    BO_LOCKPTR(bo)) == ENOLCK) {
2629 				BO_LOCK(bo);
2630 				return (EAGAIN);
2631 			}
2632 
2633 			bremfree(bp);
2634 			bp->b_flags |= B_INVAL | B_RELBUF;
2635 			bp->b_flags &= ~B_ASYNC;
2636 			brelse(bp);
2637 			anyfreed = true;
2638 
2639 			BO_LOCK(bo);
2640 			if (nbp != NULL &&
2641 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2642 			    nbp->b_vp != vp ||
2643 			    (nbp->b_flags & B_DELWRI) != 0))
2644 				return (EAGAIN);
2645 		}
2646 
2647 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2648 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2649 				continue;
2650 			if (BUF_LOCK(bp,
2651 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2652 			    BO_LOCKPTR(bo)) == ENOLCK) {
2653 				BO_LOCK(bo);
2654 				return (EAGAIN);
2655 			}
2656 			bremfree(bp);
2657 			bp->b_flags |= B_INVAL | B_RELBUF;
2658 			bp->b_flags &= ~B_ASYNC;
2659 			brelse(bp);
2660 			anyfreed = true;
2661 
2662 			BO_LOCK(bo);
2663 			if (nbp != NULL &&
2664 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2665 			    (nbp->b_vp != vp) ||
2666 			    (nbp->b_flags & B_DELWRI) == 0))
2667 				return (EAGAIN);
2668 		}
2669 	} while (anyfreed);
2670 	return (0);
2671 }
2672 
2673 static void
2674 buf_vlist_remove(struct buf *bp)
2675 {
2676 	struct bufv *bv;
2677 	b_xflags_t flags;
2678 
2679 	flags = bp->b_xflags;
2680 
2681 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2682 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2683 	KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 &&
2684 	    (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN),
2685 	    ("%s: buffer %p has invalid queue state", __func__, bp));
2686 
2687 	if ((flags & BX_VNDIRTY) != 0)
2688 		bv = &bp->b_bufobj->bo_dirty;
2689 	else
2690 		bv = &bp->b_bufobj->bo_clean;
2691 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2692 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2693 	bv->bv_cnt--;
2694 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2695 }
2696 
2697 /*
2698  * Add the buffer to the sorted clean or dirty block list.
2699  *
2700  * NOTE: xflags is passed as a constant, optimizing this inline function!
2701  */
2702 static void
2703 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2704 {
2705 	struct bufv *bv;
2706 	struct buf *n;
2707 	int error;
2708 
2709 	ASSERT_BO_WLOCKED(bo);
2710 	KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2711 	    ("buf_vlist_add: bo %p does not allow bufs", bo));
2712 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2713 	    ("dead bo %p", bo));
2714 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2715 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2716 	bp->b_xflags |= xflags;
2717 	if (xflags & BX_VNDIRTY)
2718 		bv = &bo->bo_dirty;
2719 	else
2720 		bv = &bo->bo_clean;
2721 
2722 	/*
2723 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2724 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2725 	 * than _ge.
2726 	 */
2727 	if (bv->bv_cnt == 0 ||
2728 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2729 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2730 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2731 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2732 	else
2733 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2734 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2735 	if (error)
2736 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2737 	bv->bv_cnt++;
2738 }
2739 
2740 /*
2741  * Look up a buffer using the buffer tries.
2742  */
2743 struct buf *
2744 gbincore(struct bufobj *bo, daddr_t lblkno)
2745 {
2746 	struct buf *bp;
2747 
2748 	ASSERT_BO_LOCKED(bo);
2749 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2750 	if (bp != NULL)
2751 		return (bp);
2752 	return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno));
2753 }
2754 
2755 /*
2756  * Look up a buf using the buffer tries, without the bufobj lock.  This relies
2757  * on SMR for safe lookup, and bufs being in a no-free zone to provide type
2758  * stability of the result.  Like other lockless lookups, the found buf may
2759  * already be invalid by the time this function returns.
2760  */
2761 struct buf *
2762 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno)
2763 {
2764 	struct buf *bp;
2765 
2766 	ASSERT_BO_UNLOCKED(bo);
2767 	bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno);
2768 	if (bp != NULL)
2769 		return (bp);
2770 	return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno));
2771 }
2772 
2773 /*
2774  * Associate a buffer with a vnode.
2775  */
2776 void
2777 bgetvp(struct vnode *vp, struct buf *bp)
2778 {
2779 	struct bufobj *bo;
2780 
2781 	bo = &vp->v_bufobj;
2782 	ASSERT_BO_WLOCKED(bo);
2783 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2784 
2785 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2786 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2787 	    ("bgetvp: bp already attached! %p", bp));
2788 
2789 	vhold(vp);
2790 	bp->b_vp = vp;
2791 	bp->b_bufobj = bo;
2792 	/*
2793 	 * Insert onto list for new vnode.
2794 	 */
2795 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2796 }
2797 
2798 /*
2799  * Disassociate a buffer from a vnode.
2800  */
2801 void
2802 brelvp(struct buf *bp)
2803 {
2804 	struct bufobj *bo;
2805 	struct vnode *vp;
2806 
2807 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2808 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2809 
2810 	/*
2811 	 * Delete from old vnode list, if on one.
2812 	 */
2813 	vp = bp->b_vp;		/* XXX */
2814 	bo = bp->b_bufobj;
2815 	BO_LOCK(bo);
2816 	buf_vlist_remove(bp);
2817 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2818 		bo->bo_flag &= ~BO_ONWORKLST;
2819 		mtx_lock(&sync_mtx);
2820 		LIST_REMOVE(bo, bo_synclist);
2821 		syncer_worklist_len--;
2822 		mtx_unlock(&sync_mtx);
2823 	}
2824 	bp->b_vp = NULL;
2825 	bp->b_bufobj = NULL;
2826 	BO_UNLOCK(bo);
2827 	vdrop(vp);
2828 }
2829 
2830 /*
2831  * Add an item to the syncer work queue.
2832  */
2833 static void
2834 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2835 {
2836 	int slot;
2837 
2838 	ASSERT_BO_WLOCKED(bo);
2839 
2840 	mtx_lock(&sync_mtx);
2841 	if (bo->bo_flag & BO_ONWORKLST)
2842 		LIST_REMOVE(bo, bo_synclist);
2843 	else {
2844 		bo->bo_flag |= BO_ONWORKLST;
2845 		syncer_worklist_len++;
2846 	}
2847 
2848 	if (delay > syncer_maxdelay - 2)
2849 		delay = syncer_maxdelay - 2;
2850 	slot = (syncer_delayno + delay) & syncer_mask;
2851 
2852 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2853 	mtx_unlock(&sync_mtx);
2854 }
2855 
2856 static int
2857 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2858 {
2859 	int error, len;
2860 
2861 	mtx_lock(&sync_mtx);
2862 	len = syncer_worklist_len - sync_vnode_count;
2863 	mtx_unlock(&sync_mtx);
2864 	error = SYSCTL_OUT(req, &len, sizeof(len));
2865 	return (error);
2866 }
2867 
2868 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2869     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2870     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2871 
2872 static struct proc *updateproc;
2873 static void sched_sync(void);
2874 static struct kproc_desc up_kp = {
2875 	"syncer",
2876 	sched_sync,
2877 	&updateproc
2878 };
2879 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2880 
2881 static int
2882 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2883 {
2884 	struct vnode *vp;
2885 	struct mount *mp;
2886 
2887 	*bo = LIST_FIRST(slp);
2888 	if (*bo == NULL)
2889 		return (0);
2890 	vp = bo2vnode(*bo);
2891 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2892 		return (1);
2893 	/*
2894 	 * We use vhold in case the vnode does not
2895 	 * successfully sync.  vhold prevents the vnode from
2896 	 * going away when we unlock the sync_mtx so that
2897 	 * we can acquire the vnode interlock.
2898 	 */
2899 	vholdl(vp);
2900 	mtx_unlock(&sync_mtx);
2901 	VI_UNLOCK(vp);
2902 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2903 		vdrop(vp);
2904 		mtx_lock(&sync_mtx);
2905 		return (*bo == LIST_FIRST(slp));
2906 	}
2907 	MPASSERT(mp == NULL || (curthread->td_pflags & TDP_IGNSUSP) != 0 ||
2908 	    (mp->mnt_kern_flag & MNTK_SUSPENDED) == 0, mp,
2909 	    ("suspended mp syncing vp %p", vp));
2910 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2911 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2912 	VOP_UNLOCK(vp);
2913 	vn_finished_write(mp);
2914 	BO_LOCK(*bo);
2915 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2916 		/*
2917 		 * Put us back on the worklist.  The worklist
2918 		 * routine will remove us from our current
2919 		 * position and then add us back in at a later
2920 		 * position.
2921 		 */
2922 		vn_syncer_add_to_worklist(*bo, syncdelay);
2923 	}
2924 	BO_UNLOCK(*bo);
2925 	vdrop(vp);
2926 	mtx_lock(&sync_mtx);
2927 	return (0);
2928 }
2929 
2930 static int first_printf = 1;
2931 
2932 /*
2933  * System filesystem synchronizer daemon.
2934  */
2935 static void
2936 sched_sync(void)
2937 {
2938 	struct synclist *next, *slp;
2939 	struct bufobj *bo;
2940 	long starttime;
2941 	struct thread *td = curthread;
2942 	int last_work_seen;
2943 	int net_worklist_len;
2944 	int syncer_final_iter;
2945 	int error;
2946 
2947 	last_work_seen = 0;
2948 	syncer_final_iter = 0;
2949 	syncer_state = SYNCER_RUNNING;
2950 	starttime = time_uptime;
2951 	td->td_pflags |= TDP_NORUNNINGBUF;
2952 
2953 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2954 	    SHUTDOWN_PRI_LAST);
2955 
2956 	mtx_lock(&sync_mtx);
2957 	for (;;) {
2958 		if (syncer_state == SYNCER_FINAL_DELAY &&
2959 		    syncer_final_iter == 0) {
2960 			mtx_unlock(&sync_mtx);
2961 			kproc_suspend_check(td->td_proc);
2962 			mtx_lock(&sync_mtx);
2963 		}
2964 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2965 		if (syncer_state != SYNCER_RUNNING &&
2966 		    starttime != time_uptime) {
2967 			if (first_printf) {
2968 				printf("\nSyncing disks, vnodes remaining... ");
2969 				first_printf = 0;
2970 			}
2971 			printf("%d ", net_worklist_len);
2972 		}
2973 		starttime = time_uptime;
2974 
2975 		/*
2976 		 * Push files whose dirty time has expired.  Be careful
2977 		 * of interrupt race on slp queue.
2978 		 *
2979 		 * Skip over empty worklist slots when shutting down.
2980 		 */
2981 		do {
2982 			slp = &syncer_workitem_pending[syncer_delayno];
2983 			syncer_delayno += 1;
2984 			if (syncer_delayno == syncer_maxdelay)
2985 				syncer_delayno = 0;
2986 			next = &syncer_workitem_pending[syncer_delayno];
2987 			/*
2988 			 * If the worklist has wrapped since the
2989 			 * it was emptied of all but syncer vnodes,
2990 			 * switch to the FINAL_DELAY state and run
2991 			 * for one more second.
2992 			 */
2993 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2994 			    net_worklist_len == 0 &&
2995 			    last_work_seen == syncer_delayno) {
2996 				syncer_state = SYNCER_FINAL_DELAY;
2997 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2998 			}
2999 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
3000 		    syncer_worklist_len > 0);
3001 
3002 		/*
3003 		 * Keep track of the last time there was anything
3004 		 * on the worklist other than syncer vnodes.
3005 		 * Return to the SHUTTING_DOWN state if any
3006 		 * new work appears.
3007 		 */
3008 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
3009 			last_work_seen = syncer_delayno;
3010 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
3011 			syncer_state = SYNCER_SHUTTING_DOWN;
3012 		while (!LIST_EMPTY(slp)) {
3013 			error = sync_vnode(slp, &bo, td);
3014 			if (error == 1) {
3015 				LIST_REMOVE(bo, bo_synclist);
3016 				LIST_INSERT_HEAD(next, bo, bo_synclist);
3017 				continue;
3018 			}
3019 
3020 			if (first_printf == 0) {
3021 				/*
3022 				 * Drop the sync mutex, because some watchdog
3023 				 * drivers need to sleep while patting
3024 				 */
3025 				mtx_unlock(&sync_mtx);
3026 				wdog_kern_pat(WD_LASTVAL);
3027 				mtx_lock(&sync_mtx);
3028 			}
3029 		}
3030 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
3031 			syncer_final_iter--;
3032 		/*
3033 		 * The variable rushjob allows the kernel to speed up the
3034 		 * processing of the filesystem syncer process. A rushjob
3035 		 * value of N tells the filesystem syncer to process the next
3036 		 * N seconds worth of work on its queue ASAP. Currently rushjob
3037 		 * is used by the soft update code to speed up the filesystem
3038 		 * syncer process when the incore state is getting so far
3039 		 * ahead of the disk that the kernel memory pool is being
3040 		 * threatened with exhaustion.
3041 		 */
3042 		if (rushjob > 0) {
3043 			rushjob -= 1;
3044 			continue;
3045 		}
3046 		/*
3047 		 * Just sleep for a short period of time between
3048 		 * iterations when shutting down to allow some I/O
3049 		 * to happen.
3050 		 *
3051 		 * If it has taken us less than a second to process the
3052 		 * current work, then wait. Otherwise start right over
3053 		 * again. We can still lose time if any single round
3054 		 * takes more than two seconds, but it does not really
3055 		 * matter as we are just trying to generally pace the
3056 		 * filesystem activity.
3057 		 */
3058 		if (syncer_state != SYNCER_RUNNING ||
3059 		    time_uptime == starttime) {
3060 			thread_lock(td);
3061 			sched_prio(td, PPAUSE);
3062 			thread_unlock(td);
3063 		}
3064 		if (syncer_state != SYNCER_RUNNING)
3065 			cv_timedwait(&sync_wakeup, &sync_mtx,
3066 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
3067 		else if (time_uptime == starttime)
3068 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
3069 	}
3070 }
3071 
3072 /*
3073  * Request the syncer daemon to speed up its work.
3074  * We never push it to speed up more than half of its
3075  * normal turn time, otherwise it could take over the cpu.
3076  */
3077 int
3078 speedup_syncer(void)
3079 {
3080 	int ret = 0;
3081 
3082 	mtx_lock(&sync_mtx);
3083 	if (rushjob < syncdelay / 2) {
3084 		rushjob += 1;
3085 		stat_rush_requests += 1;
3086 		ret = 1;
3087 	}
3088 	mtx_unlock(&sync_mtx);
3089 	cv_broadcast(&sync_wakeup);
3090 	return (ret);
3091 }
3092 
3093 /*
3094  * Tell the syncer to speed up its work and run though its work
3095  * list several times, then tell it to shut down.
3096  */
3097 static void
3098 syncer_shutdown(void *arg, int howto)
3099 {
3100 
3101 	if (howto & RB_NOSYNC)
3102 		return;
3103 	mtx_lock(&sync_mtx);
3104 	syncer_state = SYNCER_SHUTTING_DOWN;
3105 	rushjob = 0;
3106 	mtx_unlock(&sync_mtx);
3107 	cv_broadcast(&sync_wakeup);
3108 	kproc_shutdown(arg, howto);
3109 }
3110 
3111 void
3112 syncer_suspend(void)
3113 {
3114 
3115 	syncer_shutdown(updateproc, 0);
3116 }
3117 
3118 void
3119 syncer_resume(void)
3120 {
3121 
3122 	mtx_lock(&sync_mtx);
3123 	first_printf = 1;
3124 	syncer_state = SYNCER_RUNNING;
3125 	mtx_unlock(&sync_mtx);
3126 	cv_broadcast(&sync_wakeup);
3127 	kproc_resume(updateproc);
3128 }
3129 
3130 /*
3131  * Move the buffer between the clean and dirty lists of its vnode.
3132  */
3133 void
3134 reassignbuf(struct buf *bp)
3135 {
3136 	struct vnode *vp;
3137 	struct bufobj *bo;
3138 	int delay;
3139 #ifdef INVARIANTS
3140 	struct bufv *bv;
3141 #endif
3142 
3143 	vp = bp->b_vp;
3144 	bo = bp->b_bufobj;
3145 
3146 	KASSERT((bp->b_flags & B_PAGING) == 0,
3147 	    ("%s: cannot reassign paging buffer %p", __func__, bp));
3148 
3149 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
3150 	    bp, bp->b_vp, bp->b_flags);
3151 
3152 	BO_LOCK(bo);
3153 	buf_vlist_remove(bp);
3154 
3155 	/*
3156 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
3157 	 * of clean buffers.
3158 	 */
3159 	if (bp->b_flags & B_DELWRI) {
3160 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
3161 			switch (vp->v_type) {
3162 			case VDIR:
3163 				delay = dirdelay;
3164 				break;
3165 			case VCHR:
3166 				delay = metadelay;
3167 				break;
3168 			default:
3169 				delay = filedelay;
3170 			}
3171 			vn_syncer_add_to_worklist(bo, delay);
3172 		}
3173 		buf_vlist_add(bp, bo, BX_VNDIRTY);
3174 	} else {
3175 		buf_vlist_add(bp, bo, BX_VNCLEAN);
3176 
3177 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
3178 			mtx_lock(&sync_mtx);
3179 			LIST_REMOVE(bo, bo_synclist);
3180 			syncer_worklist_len--;
3181 			mtx_unlock(&sync_mtx);
3182 			bo->bo_flag &= ~BO_ONWORKLST;
3183 		}
3184 	}
3185 #ifdef INVARIANTS
3186 	bv = &bo->bo_clean;
3187 	bp = TAILQ_FIRST(&bv->bv_hd);
3188 	KASSERT(bp == NULL || bp->b_bufobj == bo,
3189 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3190 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
3191 	KASSERT(bp == NULL || bp->b_bufobj == bo,
3192 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3193 	bv = &bo->bo_dirty;
3194 	bp = TAILQ_FIRST(&bv->bv_hd);
3195 	KASSERT(bp == NULL || bp->b_bufobj == bo,
3196 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3197 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
3198 	KASSERT(bp == NULL || bp->b_bufobj == bo,
3199 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3200 #endif
3201 	BO_UNLOCK(bo);
3202 }
3203 
3204 static void
3205 v_init_counters(struct vnode *vp)
3206 {
3207 
3208 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
3209 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
3210 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
3211 
3212 	refcount_init(&vp->v_holdcnt, 1);
3213 	refcount_init(&vp->v_usecount, 1);
3214 }
3215 
3216 /*
3217  * Get a usecount on a vnode.
3218  *
3219  * vget and vget_finish may fail to lock the vnode if they lose a race against
3220  * it being doomed. LK_RETRY can be passed in flags to lock it anyway.
3221  *
3222  * Consumers which don't guarantee liveness of the vnode can use SMR to
3223  * try to get a reference. Note this operation can fail since the vnode
3224  * may be awaiting getting freed by the time they get to it.
3225  */
3226 enum vgetstate
3227 vget_prep_smr(struct vnode *vp)
3228 {
3229 	enum vgetstate vs;
3230 
3231 	VFS_SMR_ASSERT_ENTERED();
3232 
3233 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
3234 		vs = VGET_USECOUNT;
3235 	} else {
3236 		if (vhold_smr(vp))
3237 			vs = VGET_HOLDCNT;
3238 		else
3239 			vs = VGET_NONE;
3240 	}
3241 	return (vs);
3242 }
3243 
3244 enum vgetstate
3245 vget_prep(struct vnode *vp)
3246 {
3247 	enum vgetstate vs;
3248 
3249 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
3250 		vs = VGET_USECOUNT;
3251 	} else {
3252 		vhold(vp);
3253 		vs = VGET_HOLDCNT;
3254 	}
3255 	return (vs);
3256 }
3257 
3258 void
3259 vget_abort(struct vnode *vp, enum vgetstate vs)
3260 {
3261 
3262 	switch (vs) {
3263 	case VGET_USECOUNT:
3264 		vrele(vp);
3265 		break;
3266 	case VGET_HOLDCNT:
3267 		vdrop(vp);
3268 		break;
3269 	default:
3270 		__assert_unreachable();
3271 	}
3272 }
3273 
3274 int
3275 vget(struct vnode *vp, int flags)
3276 {
3277 	enum vgetstate vs;
3278 
3279 	vs = vget_prep(vp);
3280 	return (vget_finish(vp, flags, vs));
3281 }
3282 
3283 int
3284 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
3285 {
3286 	int error;
3287 
3288 	if ((flags & LK_INTERLOCK) != 0)
3289 		ASSERT_VI_LOCKED(vp, __func__);
3290 	else
3291 		ASSERT_VI_UNLOCKED(vp, __func__);
3292 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3293 	VNPASS(vp->v_holdcnt > 0, vp);
3294 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3295 
3296 	error = vn_lock(vp, flags);
3297 	if (__predict_false(error != 0)) {
3298 		vget_abort(vp, vs);
3299 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
3300 		    vp);
3301 		return (error);
3302 	}
3303 
3304 	vget_finish_ref(vp, vs);
3305 	return (0);
3306 }
3307 
3308 void
3309 vget_finish_ref(struct vnode *vp, enum vgetstate vs)
3310 {
3311 	int old;
3312 
3313 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3314 	VNPASS(vp->v_holdcnt > 0, vp);
3315 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3316 
3317 	if (vs == VGET_USECOUNT)
3318 		return;
3319 
3320 	/*
3321 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
3322 	 * the vnode around. Otherwise someone else lended their hold count and
3323 	 * we have to drop ours.
3324 	 */
3325 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
3326 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3327 	if (old != 0) {
3328 #ifdef INVARIANTS
3329 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3330 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3331 #else
3332 		refcount_release(&vp->v_holdcnt);
3333 #endif
3334 	}
3335 }
3336 
3337 void
3338 vref(struct vnode *vp)
3339 {
3340 	enum vgetstate vs;
3341 
3342 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3343 	vs = vget_prep(vp);
3344 	vget_finish_ref(vp, vs);
3345 }
3346 
3347 void
3348 vrefact(struct vnode *vp)
3349 {
3350 
3351 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3352 #ifdef INVARIANTS
3353 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3354 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3355 #else
3356 	refcount_acquire(&vp->v_usecount);
3357 #endif
3358 }
3359 
3360 void
3361 vlazy(struct vnode *vp)
3362 {
3363 	struct mount *mp;
3364 
3365 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3366 
3367 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3368 		return;
3369 	/*
3370 	 * We may get here for inactive routines after the vnode got doomed.
3371 	 */
3372 	if (VN_IS_DOOMED(vp))
3373 		return;
3374 	mp = vp->v_mount;
3375 	mtx_lock(&mp->mnt_listmtx);
3376 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3377 		vp->v_mflag |= VMP_LAZYLIST;
3378 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3379 		mp->mnt_lazyvnodelistsize++;
3380 	}
3381 	mtx_unlock(&mp->mnt_listmtx);
3382 }
3383 
3384 static void
3385 vunlazy(struct vnode *vp)
3386 {
3387 	struct mount *mp;
3388 
3389 	ASSERT_VI_LOCKED(vp, __func__);
3390 	VNPASS(!VN_IS_DOOMED(vp), vp);
3391 
3392 	mp = vp->v_mount;
3393 	mtx_lock(&mp->mnt_listmtx);
3394 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3395 	/*
3396 	 * Don't remove the vnode from the lazy list if another thread
3397 	 * has increased the hold count. It may have re-enqueued the
3398 	 * vnode to the lazy list and is now responsible for its
3399 	 * removal.
3400 	 */
3401 	if (vp->v_holdcnt == 0) {
3402 		vp->v_mflag &= ~VMP_LAZYLIST;
3403 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3404 		mp->mnt_lazyvnodelistsize--;
3405 	}
3406 	mtx_unlock(&mp->mnt_listmtx);
3407 }
3408 
3409 /*
3410  * This routine is only meant to be called from vgonel prior to dooming
3411  * the vnode.
3412  */
3413 static void
3414 vunlazy_gone(struct vnode *vp)
3415 {
3416 	struct mount *mp;
3417 
3418 	ASSERT_VOP_ELOCKED(vp, __func__);
3419 	ASSERT_VI_LOCKED(vp, __func__);
3420 	VNPASS(!VN_IS_DOOMED(vp), vp);
3421 
3422 	if (vp->v_mflag & VMP_LAZYLIST) {
3423 		mp = vp->v_mount;
3424 		mtx_lock(&mp->mnt_listmtx);
3425 		VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3426 		vp->v_mflag &= ~VMP_LAZYLIST;
3427 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3428 		mp->mnt_lazyvnodelistsize--;
3429 		mtx_unlock(&mp->mnt_listmtx);
3430 	}
3431 }
3432 
3433 static void
3434 vdefer_inactive(struct vnode *vp)
3435 {
3436 
3437 	ASSERT_VI_LOCKED(vp, __func__);
3438 	VNPASS(vp->v_holdcnt > 0, vp);
3439 	if (VN_IS_DOOMED(vp)) {
3440 		vdropl(vp);
3441 		return;
3442 	}
3443 	if (vp->v_iflag & VI_DEFINACT) {
3444 		VNPASS(vp->v_holdcnt > 1, vp);
3445 		vdropl(vp);
3446 		return;
3447 	}
3448 	if (vp->v_usecount > 0) {
3449 		vp->v_iflag &= ~VI_OWEINACT;
3450 		vdropl(vp);
3451 		return;
3452 	}
3453 	vlazy(vp);
3454 	vp->v_iflag |= VI_DEFINACT;
3455 	VI_UNLOCK(vp);
3456 	atomic_add_long(&deferred_inact, 1);
3457 }
3458 
3459 static void
3460 vdefer_inactive_unlocked(struct vnode *vp)
3461 {
3462 
3463 	VI_LOCK(vp);
3464 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3465 		vdropl(vp);
3466 		return;
3467 	}
3468 	vdefer_inactive(vp);
3469 }
3470 
3471 enum vput_op { VRELE, VPUT, VUNREF };
3472 
3473 /*
3474  * Handle ->v_usecount transitioning to 0.
3475  *
3476  * By releasing the last usecount we take ownership of the hold count which
3477  * provides liveness of the vnode, meaning we have to vdrop.
3478  *
3479  * For all vnodes we may need to perform inactive processing. It requires an
3480  * exclusive lock on the vnode, while it is legal to call here with only a
3481  * shared lock (or no locks). If locking the vnode in an expected manner fails,
3482  * inactive processing gets deferred to the syncer.
3483  *
3484  * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3485  * on the lock being held all the way until VOP_INACTIVE. This in particular
3486  * happens with UFS which adds half-constructed vnodes to the hash, where they
3487  * can be found by other code.
3488  */
3489 static void
3490 vput_final(struct vnode *vp, enum vput_op func)
3491 {
3492 	int error;
3493 	bool want_unlock;
3494 
3495 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3496 	VNPASS(vp->v_holdcnt > 0, vp);
3497 
3498 	VI_LOCK(vp);
3499 
3500 	/*
3501 	 * By the time we got here someone else might have transitioned
3502 	 * the count back to > 0.
3503 	 */
3504 	if (vp->v_usecount > 0)
3505 		goto out;
3506 
3507 	/*
3508 	 * If the vnode is doomed vgone already performed inactive processing
3509 	 * (if needed).
3510 	 */
3511 	if (VN_IS_DOOMED(vp))
3512 		goto out;
3513 
3514 	if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3515 		goto out;
3516 
3517 	if (vp->v_iflag & VI_DOINGINACT)
3518 		goto out;
3519 
3520 	/*
3521 	 * Locking operations here will drop the interlock and possibly the
3522 	 * vnode lock, opening a window where the vnode can get doomed all the
3523 	 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3524 	 * perform inactive.
3525 	 */
3526 	vp->v_iflag |= VI_OWEINACT;
3527 	want_unlock = false;
3528 	error = 0;
3529 	switch (func) {
3530 	case VRELE:
3531 		switch (VOP_ISLOCKED(vp)) {
3532 		case LK_EXCLUSIVE:
3533 			break;
3534 		case LK_EXCLOTHER:
3535 		case 0:
3536 			want_unlock = true;
3537 			error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3538 			VI_LOCK(vp);
3539 			break;
3540 		default:
3541 			/*
3542 			 * The lock has at least one sharer, but we have no way
3543 			 * to conclude whether this is us. Play it safe and
3544 			 * defer processing.
3545 			 */
3546 			error = EAGAIN;
3547 			break;
3548 		}
3549 		break;
3550 	case VPUT:
3551 		want_unlock = true;
3552 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3553 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3554 			    LK_NOWAIT);
3555 			VI_LOCK(vp);
3556 		}
3557 		break;
3558 	case VUNREF:
3559 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3560 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3561 			VI_LOCK(vp);
3562 		}
3563 		break;
3564 	}
3565 	if (error == 0) {
3566 		if (func == VUNREF) {
3567 			VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp,
3568 			    ("recursive vunref"));
3569 			vp->v_vflag |= VV_UNREF;
3570 		}
3571 		for (;;) {
3572 			error = vinactive(vp);
3573 			if (want_unlock)
3574 				VOP_UNLOCK(vp);
3575 			if (error != ERELOOKUP || !want_unlock)
3576 				break;
3577 			VOP_LOCK(vp, LK_EXCLUSIVE);
3578 		}
3579 		if (func == VUNREF)
3580 			vp->v_vflag &= ~VV_UNREF;
3581 		vdropl(vp);
3582 	} else {
3583 		vdefer_inactive(vp);
3584 	}
3585 	return;
3586 out:
3587 	if (func == VPUT)
3588 		VOP_UNLOCK(vp);
3589 	vdropl(vp);
3590 }
3591 
3592 /*
3593  * Decrement ->v_usecount for a vnode.
3594  *
3595  * Releasing the last use count requires additional processing, see vput_final
3596  * above for details.
3597  *
3598  * Comment above each variant denotes lock state on entry and exit.
3599  */
3600 
3601 /*
3602  * in: any
3603  * out: same as passed in
3604  */
3605 void
3606 vrele(struct vnode *vp)
3607 {
3608 
3609 	ASSERT_VI_UNLOCKED(vp, __func__);
3610 	if (!refcount_release(&vp->v_usecount))
3611 		return;
3612 	vput_final(vp, VRELE);
3613 }
3614 
3615 /*
3616  * in: locked
3617  * out: unlocked
3618  */
3619 void
3620 vput(struct vnode *vp)
3621 {
3622 
3623 	ASSERT_VOP_LOCKED(vp, __func__);
3624 	ASSERT_VI_UNLOCKED(vp, __func__);
3625 	if (!refcount_release(&vp->v_usecount)) {
3626 		VOP_UNLOCK(vp);
3627 		return;
3628 	}
3629 	vput_final(vp, VPUT);
3630 }
3631 
3632 /*
3633  * in: locked
3634  * out: locked
3635  */
3636 void
3637 vunref(struct vnode *vp)
3638 {
3639 
3640 	ASSERT_VOP_LOCKED(vp, __func__);
3641 	ASSERT_VI_UNLOCKED(vp, __func__);
3642 	if (!refcount_release(&vp->v_usecount))
3643 		return;
3644 	vput_final(vp, VUNREF);
3645 }
3646 
3647 void
3648 vhold(struct vnode *vp)
3649 {
3650 	int old;
3651 
3652 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3653 	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3654 	VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3655 	    ("%s: wrong hold count %d", __func__, old));
3656 	if (old == 0)
3657 		vfs_freevnodes_dec();
3658 }
3659 
3660 void
3661 vholdnz(struct vnode *vp)
3662 {
3663 
3664 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3665 #ifdef INVARIANTS
3666 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3667 	VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3668 	    ("%s: wrong hold count %d", __func__, old));
3669 #else
3670 	atomic_add_int(&vp->v_holdcnt, 1);
3671 #endif
3672 }
3673 
3674 /*
3675  * Grab a hold count unless the vnode is freed.
3676  *
3677  * Only use this routine if vfs smr is the only protection you have against
3678  * freeing the vnode.
3679  *
3680  * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag
3681  * is not set.  After the flag is set the vnode becomes immutable to anyone but
3682  * the thread which managed to set the flag.
3683  *
3684  * It may be tempting to replace the loop with:
3685  * count = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3686  * if (count & VHOLD_NO_SMR) {
3687  *     backpedal and error out;
3688  * }
3689  *
3690  * However, while this is more performant, it hinders debugging by eliminating
3691  * the previously mentioned invariant.
3692  */
3693 bool
3694 vhold_smr(struct vnode *vp)
3695 {
3696 	int count;
3697 
3698 	VFS_SMR_ASSERT_ENTERED();
3699 
3700 	count = atomic_load_int(&vp->v_holdcnt);
3701 	for (;;) {
3702 		if (count & VHOLD_NO_SMR) {
3703 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3704 			    ("non-zero hold count with flags %d\n", count));
3705 			return (false);
3706 		}
3707 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3708 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3709 			if (count == 0)
3710 				vfs_freevnodes_dec();
3711 			return (true);
3712 		}
3713 	}
3714 }
3715 
3716 /*
3717  * Hold a free vnode for recycling.
3718  *
3719  * Note: vnode_init references this comment.
3720  *
3721  * Attempts to recycle only need the global vnode list lock and have no use for
3722  * SMR.
3723  *
3724  * However, vnodes get inserted into the global list before they get fully
3725  * initialized and stay there until UMA decides to free the memory. This in
3726  * particular means the target can be found before it becomes usable and after
3727  * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to
3728  * VHOLD_NO_SMR.
3729  *
3730  * Note: the vnode may gain more references after we transition the count 0->1.
3731  */
3732 static bool
3733 vhold_recycle_free(struct vnode *vp)
3734 {
3735 	int count;
3736 
3737 	mtx_assert(&vnode_list_mtx, MA_OWNED);
3738 
3739 	count = atomic_load_int(&vp->v_holdcnt);
3740 	for (;;) {
3741 		if (count & VHOLD_NO_SMR) {
3742 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3743 			    ("non-zero hold count with flags %d\n", count));
3744 			return (false);
3745 		}
3746 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3747 		if (count > 0) {
3748 			return (false);
3749 		}
3750 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3751 			vfs_freevnodes_dec();
3752 			return (true);
3753 		}
3754 	}
3755 }
3756 
3757 static void __noinline
3758 vdbatch_process(struct vdbatch *vd)
3759 {
3760 	struct vnode *vp;
3761 	int i;
3762 
3763 	mtx_assert(&vd->lock, MA_OWNED);
3764 	MPASS(curthread->td_pinned > 0);
3765 	MPASS(vd->index == VDBATCH_SIZE);
3766 
3767 	/*
3768 	 * Attempt to requeue the passed batch, but give up easily.
3769 	 *
3770 	 * Despite batching the mechanism is prone to transient *significant*
3771 	 * lock contention, where vnode_list_mtx becomes the primary bottleneck
3772 	 * if multiple CPUs get here (one real-world example is highly parallel
3773 	 * do-nothing make , which will stat *tons* of vnodes). Since it is
3774 	 * quasi-LRU (read: not that great even if fully honoured) just dodge
3775 	 * the problem. Parties which don't like it are welcome to implement
3776 	 * something better.
3777 	 */
3778 	critical_enter();
3779 	if (mtx_trylock(&vnode_list_mtx)) {
3780 		for (i = 0; i < VDBATCH_SIZE; i++) {
3781 			vp = vd->tab[i];
3782 			vd->tab[i] = NULL;
3783 			TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3784 			TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3785 			MPASS(vp->v_dbatchcpu != NOCPU);
3786 			vp->v_dbatchcpu = NOCPU;
3787 		}
3788 		mtx_unlock(&vnode_list_mtx);
3789 	} else {
3790 		counter_u64_add(vnode_skipped_requeues, 1);
3791 
3792 		for (i = 0; i < VDBATCH_SIZE; i++) {
3793 			vp = vd->tab[i];
3794 			vd->tab[i] = NULL;
3795 			MPASS(vp->v_dbatchcpu != NOCPU);
3796 			vp->v_dbatchcpu = NOCPU;
3797 		}
3798 	}
3799 	vd->index = 0;
3800 	critical_exit();
3801 }
3802 
3803 static void
3804 vdbatch_enqueue(struct vnode *vp)
3805 {
3806 	struct vdbatch *vd;
3807 
3808 	ASSERT_VI_LOCKED(vp, __func__);
3809 	VNPASS(!VN_IS_DOOMED(vp), vp);
3810 
3811 	if (vp->v_dbatchcpu != NOCPU) {
3812 		VI_UNLOCK(vp);
3813 		return;
3814 	}
3815 
3816 	sched_pin();
3817 	vd = DPCPU_PTR(vd);
3818 	mtx_lock(&vd->lock);
3819 	MPASS(vd->index < VDBATCH_SIZE);
3820 	MPASS(vd->tab[vd->index] == NULL);
3821 	/*
3822 	 * A hack: we depend on being pinned so that we know what to put in
3823 	 * ->v_dbatchcpu.
3824 	 */
3825 	vp->v_dbatchcpu = curcpu;
3826 	vd->tab[vd->index] = vp;
3827 	vd->index++;
3828 	VI_UNLOCK(vp);
3829 	if (vd->index == VDBATCH_SIZE)
3830 		vdbatch_process(vd);
3831 	mtx_unlock(&vd->lock);
3832 	sched_unpin();
3833 }
3834 
3835 /*
3836  * This routine must only be called for vnodes which are about to be
3837  * deallocated. Supporting dequeue for arbitrary vndoes would require
3838  * validating that the locked batch matches.
3839  */
3840 static void
3841 vdbatch_dequeue(struct vnode *vp)
3842 {
3843 	struct vdbatch *vd;
3844 	int i;
3845 	short cpu;
3846 
3847 	VNPASS(vp->v_type == VBAD || vp->v_type == VNON, vp);
3848 
3849 	cpu = vp->v_dbatchcpu;
3850 	if (cpu == NOCPU)
3851 		return;
3852 
3853 	vd = DPCPU_ID_PTR(cpu, vd);
3854 	mtx_lock(&vd->lock);
3855 	for (i = 0; i < vd->index; i++) {
3856 		if (vd->tab[i] != vp)
3857 			continue;
3858 		vp->v_dbatchcpu = NOCPU;
3859 		vd->index--;
3860 		vd->tab[i] = vd->tab[vd->index];
3861 		vd->tab[vd->index] = NULL;
3862 		break;
3863 	}
3864 	mtx_unlock(&vd->lock);
3865 	/*
3866 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3867 	 */
3868 	MPASS(vp->v_dbatchcpu == NOCPU);
3869 }
3870 
3871 /*
3872  * Drop the hold count of the vnode.
3873  *
3874  * It will only get freed if this is the last hold *and* it has been vgone'd.
3875  *
3876  * Because the vnode vm object keeps a hold reference on the vnode if
3877  * there is at least one resident non-cached page, the vnode cannot
3878  * leave the active list without the page cleanup done.
3879  */
3880 static void __noinline
3881 vdropl_final(struct vnode *vp)
3882 {
3883 
3884 	ASSERT_VI_LOCKED(vp, __func__);
3885 	VNPASS(VN_IS_DOOMED(vp), vp);
3886 	/*
3887 	 * Set the VHOLD_NO_SMR flag.
3888 	 *
3889 	 * We may be racing against vhold_smr. If they win we can just pretend
3890 	 * we never got this far, they will vdrop later.
3891 	 */
3892 	if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) {
3893 		vfs_freevnodes_inc();
3894 		VI_UNLOCK(vp);
3895 		/*
3896 		 * We lost the aforementioned race. Any subsequent access is
3897 		 * invalid as they might have managed to vdropl on their own.
3898 		 */
3899 		return;
3900 	}
3901 	/*
3902 	 * Don't bump freevnodes as this one is going away.
3903 	 */
3904 	freevnode(vp);
3905 }
3906 
3907 void
3908 vdrop(struct vnode *vp)
3909 {
3910 
3911 	ASSERT_VI_UNLOCKED(vp, __func__);
3912 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3913 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3914 		return;
3915 	VI_LOCK(vp);
3916 	vdropl(vp);
3917 }
3918 
3919 static void __always_inline
3920 vdropl_impl(struct vnode *vp, bool enqueue)
3921 {
3922 
3923 	ASSERT_VI_LOCKED(vp, __func__);
3924 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3925 	if (!refcount_release(&vp->v_holdcnt)) {
3926 		VI_UNLOCK(vp);
3927 		return;
3928 	}
3929 	VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp);
3930 	VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
3931 	if (VN_IS_DOOMED(vp)) {
3932 		vdropl_final(vp);
3933 		return;
3934 	}
3935 
3936 	vfs_freevnodes_inc();
3937 	if (vp->v_mflag & VMP_LAZYLIST) {
3938 		vunlazy(vp);
3939 	}
3940 
3941 	if (!enqueue) {
3942 		VI_UNLOCK(vp);
3943 		return;
3944 	}
3945 
3946 	/*
3947 	 * Also unlocks the interlock. We can't assert on it as we
3948 	 * released our hold and by now the vnode might have been
3949 	 * freed.
3950 	 */
3951 	vdbatch_enqueue(vp);
3952 }
3953 
3954 void
3955 vdropl(struct vnode *vp)
3956 {
3957 
3958 	vdropl_impl(vp, true);
3959 }
3960 
3961 /*
3962  * vdrop a vnode when recycling
3963  *
3964  * This is a special case routine only to be used when recycling, differs from
3965  * regular vdrop by not requeieing the vnode on LRU.
3966  *
3967  * Consider a case where vtryrecycle continuously fails with all vnodes (due to
3968  * e.g., frozen writes on the filesystem), filling the batch and causing it to
3969  * be requeued. Then vnlru will end up revisiting the same vnodes. This is a
3970  * loop which can last for as long as writes are frozen.
3971  */
3972 static void
3973 vdropl_recycle(struct vnode *vp)
3974 {
3975 
3976 	vdropl_impl(vp, false);
3977 }
3978 
3979 static void
3980 vdrop_recycle(struct vnode *vp)
3981 {
3982 
3983 	VI_LOCK(vp);
3984 	vdropl_recycle(vp);
3985 }
3986 
3987 /*
3988  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3989  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3990  */
3991 static int
3992 vinactivef(struct vnode *vp)
3993 {
3994 	struct vm_object *obj;
3995 	int error;
3996 
3997 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3998 	ASSERT_VI_LOCKED(vp, "vinactive");
3999 	VNPASS((vp->v_iflag & VI_DOINGINACT) == 0, vp);
4000 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4001 	vp->v_iflag |= VI_DOINGINACT;
4002 	vp->v_iflag &= ~VI_OWEINACT;
4003 	VI_UNLOCK(vp);
4004 	/*
4005 	 * Before moving off the active list, we must be sure that any
4006 	 * modified pages are converted into the vnode's dirty
4007 	 * buffers, since these will no longer be checked once the
4008 	 * vnode is on the inactive list.
4009 	 *
4010 	 * The write-out of the dirty pages is asynchronous.  At the
4011 	 * point that VOP_INACTIVE() is called, there could still be
4012 	 * pending I/O and dirty pages in the object.
4013 	 */
4014 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
4015 	    vm_object_mightbedirty(obj)) {
4016 		VM_OBJECT_WLOCK(obj);
4017 		vm_object_page_clean(obj, 0, 0, 0);
4018 		VM_OBJECT_WUNLOCK(obj);
4019 	}
4020 	error = VOP_INACTIVE(vp);
4021 	VI_LOCK(vp);
4022 	VNPASS(vp->v_iflag & VI_DOINGINACT, vp);
4023 	vp->v_iflag &= ~VI_DOINGINACT;
4024 	return (error);
4025 }
4026 
4027 int
4028 vinactive(struct vnode *vp)
4029 {
4030 
4031 	ASSERT_VOP_ELOCKED(vp, "vinactive");
4032 	ASSERT_VI_LOCKED(vp, "vinactive");
4033 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4034 
4035 	if ((vp->v_iflag & VI_OWEINACT) == 0)
4036 		return (0);
4037 	if (vp->v_iflag & VI_DOINGINACT)
4038 		return (0);
4039 	if (vp->v_usecount > 0) {
4040 		vp->v_iflag &= ~VI_OWEINACT;
4041 		return (0);
4042 	}
4043 	return (vinactivef(vp));
4044 }
4045 
4046 /*
4047  * Remove any vnodes in the vnode table belonging to mount point mp.
4048  *
4049  * If FORCECLOSE is not specified, there should not be any active ones,
4050  * return error if any are found (nb: this is a user error, not a
4051  * system error). If FORCECLOSE is specified, detach any active vnodes
4052  * that are found.
4053  *
4054  * If WRITECLOSE is set, only flush out regular file vnodes open for
4055  * writing.
4056  *
4057  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
4058  *
4059  * `rootrefs' specifies the base reference count for the root vnode
4060  * of this filesystem. The root vnode is considered busy if its
4061  * v_usecount exceeds this value. On a successful return, vflush(, td)
4062  * will call vrele() on the root vnode exactly rootrefs times.
4063  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
4064  * be zero.
4065  */
4066 #ifdef DIAGNOSTIC
4067 static int busyprt = 0;		/* print out busy vnodes */
4068 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
4069 #endif
4070 
4071 int
4072 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
4073 {
4074 	struct vnode *vp, *mvp, *rootvp = NULL;
4075 	struct vattr vattr;
4076 	int busy = 0, error;
4077 
4078 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
4079 	    rootrefs, flags);
4080 	if (rootrefs > 0) {
4081 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
4082 		    ("vflush: bad args"));
4083 		/*
4084 		 * Get the filesystem root vnode. We can vput() it
4085 		 * immediately, since with rootrefs > 0, it won't go away.
4086 		 */
4087 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
4088 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
4089 			    __func__, error);
4090 			return (error);
4091 		}
4092 		vput(rootvp);
4093 	}
4094 loop:
4095 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
4096 		vholdl(vp);
4097 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
4098 		if (error) {
4099 			vdrop(vp);
4100 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
4101 			goto loop;
4102 		}
4103 		/*
4104 		 * Skip over a vnodes marked VV_SYSTEM.
4105 		 */
4106 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
4107 			VOP_UNLOCK(vp);
4108 			vdrop(vp);
4109 			continue;
4110 		}
4111 		/*
4112 		 * If WRITECLOSE is set, flush out unlinked but still open
4113 		 * files (even if open only for reading) and regular file
4114 		 * vnodes open for writing.
4115 		 */
4116 		if (flags & WRITECLOSE) {
4117 			if (vp->v_object != NULL) {
4118 				VM_OBJECT_WLOCK(vp->v_object);
4119 				vm_object_page_clean(vp->v_object, 0, 0, 0);
4120 				VM_OBJECT_WUNLOCK(vp->v_object);
4121 			}
4122 			do {
4123 				error = VOP_FSYNC(vp, MNT_WAIT, td);
4124 			} while (error == ERELOOKUP);
4125 			if (error != 0) {
4126 				VOP_UNLOCK(vp);
4127 				vdrop(vp);
4128 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
4129 				return (error);
4130 			}
4131 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
4132 			VI_LOCK(vp);
4133 
4134 			if ((vp->v_type == VNON ||
4135 			    (error == 0 && vattr.va_nlink > 0)) &&
4136 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
4137 				VOP_UNLOCK(vp);
4138 				vdropl(vp);
4139 				continue;
4140 			}
4141 		} else
4142 			VI_LOCK(vp);
4143 		/*
4144 		 * With v_usecount == 0, all we need to do is clear out the
4145 		 * vnode data structures and we are done.
4146 		 *
4147 		 * If FORCECLOSE is set, forcibly close the vnode.
4148 		 */
4149 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
4150 			vgonel(vp);
4151 		} else {
4152 			busy++;
4153 #ifdef DIAGNOSTIC
4154 			if (busyprt)
4155 				vn_printf(vp, "vflush: busy vnode ");
4156 #endif
4157 		}
4158 		VOP_UNLOCK(vp);
4159 		vdropl(vp);
4160 	}
4161 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
4162 		/*
4163 		 * If just the root vnode is busy, and if its refcount
4164 		 * is equal to `rootrefs', then go ahead and kill it.
4165 		 */
4166 		VI_LOCK(rootvp);
4167 		KASSERT(busy > 0, ("vflush: not busy"));
4168 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
4169 		    ("vflush: usecount %d < rootrefs %d",
4170 		     rootvp->v_usecount, rootrefs));
4171 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
4172 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
4173 			vgone(rootvp);
4174 			VOP_UNLOCK(rootvp);
4175 			busy = 0;
4176 		} else
4177 			VI_UNLOCK(rootvp);
4178 	}
4179 	if (busy) {
4180 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
4181 		    busy);
4182 		return (EBUSY);
4183 	}
4184 	for (; rootrefs > 0; rootrefs--)
4185 		vrele(rootvp);
4186 	return (0);
4187 }
4188 
4189 /*
4190  * Recycle an unused vnode.
4191  */
4192 int
4193 vrecycle(struct vnode *vp)
4194 {
4195 	int recycled;
4196 
4197 	VI_LOCK(vp);
4198 	recycled = vrecyclel(vp);
4199 	VI_UNLOCK(vp);
4200 	return (recycled);
4201 }
4202 
4203 /*
4204  * vrecycle, with the vp interlock held.
4205  */
4206 int
4207 vrecyclel(struct vnode *vp)
4208 {
4209 	int recycled;
4210 
4211 	ASSERT_VOP_ELOCKED(vp, __func__);
4212 	ASSERT_VI_LOCKED(vp, __func__);
4213 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4214 	recycled = 0;
4215 	if (vp->v_usecount == 0) {
4216 		recycled = 1;
4217 		vgonel(vp);
4218 	}
4219 	return (recycled);
4220 }
4221 
4222 /*
4223  * Eliminate all activity associated with a vnode
4224  * in preparation for reuse.
4225  */
4226 void
4227 vgone(struct vnode *vp)
4228 {
4229 	VI_LOCK(vp);
4230 	vgonel(vp);
4231 	VI_UNLOCK(vp);
4232 }
4233 
4234 /*
4235  * Notify upper mounts about reclaimed or unlinked vnode.
4236  */
4237 void
4238 vfs_notify_upper(struct vnode *vp, enum vfs_notify_upper_type event)
4239 {
4240 	struct mount *mp;
4241 	struct mount_upper_node *ump;
4242 
4243 	mp = atomic_load_ptr(&vp->v_mount);
4244 	if (mp == NULL)
4245 		return;
4246 	if (TAILQ_EMPTY(&mp->mnt_notify))
4247 		return;
4248 
4249 	MNT_ILOCK(mp);
4250 	mp->mnt_upper_pending++;
4251 	KASSERT(mp->mnt_upper_pending > 0,
4252 	    ("%s: mnt_upper_pending %d", __func__, mp->mnt_upper_pending));
4253 	TAILQ_FOREACH(ump, &mp->mnt_notify, mnt_upper_link) {
4254 		MNT_IUNLOCK(mp);
4255 		switch (event) {
4256 		case VFS_NOTIFY_UPPER_RECLAIM:
4257 			VFS_RECLAIM_LOWERVP(ump->mp, vp);
4258 			break;
4259 		case VFS_NOTIFY_UPPER_UNLINK:
4260 			VFS_UNLINK_LOWERVP(ump->mp, vp);
4261 			break;
4262 		}
4263 		MNT_ILOCK(mp);
4264 	}
4265 	mp->mnt_upper_pending--;
4266 	if ((mp->mnt_kern_flag & MNTK_UPPER_WAITER) != 0 &&
4267 	    mp->mnt_upper_pending == 0) {
4268 		mp->mnt_kern_flag &= ~MNTK_UPPER_WAITER;
4269 		wakeup(&mp->mnt_uppers);
4270 	}
4271 	MNT_IUNLOCK(mp);
4272 }
4273 
4274 /*
4275  * vgone, with the vp interlock held.
4276  */
4277 static void
4278 vgonel(struct vnode *vp)
4279 {
4280 	struct thread *td;
4281 	struct mount *mp;
4282 	vm_object_t object;
4283 	bool active, doinginact, oweinact;
4284 
4285 	ASSERT_VOP_ELOCKED(vp, "vgonel");
4286 	ASSERT_VI_LOCKED(vp, "vgonel");
4287 	VNASSERT(vp->v_holdcnt, vp,
4288 	    ("vgonel: vp %p has no reference.", vp));
4289 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4290 	td = curthread;
4291 
4292 	/*
4293 	 * Don't vgonel if we're already doomed.
4294 	 */
4295 	if (VN_IS_DOOMED(vp)) {
4296 		VNPASS(vn_get_state(vp) == VSTATE_DESTROYING || \
4297 		    vn_get_state(vp) == VSTATE_DEAD, vp);
4298 		return;
4299 	}
4300 	/*
4301 	 * Paired with freevnode.
4302 	 */
4303 	vn_seqc_write_begin_locked(vp);
4304 	vunlazy_gone(vp);
4305 	vn_irflag_set_locked(vp, VIRF_DOOMED);
4306 	vn_set_state(vp, VSTATE_DESTROYING);
4307 
4308 	/*
4309 	 * Check to see if the vnode is in use.  If so, we have to
4310 	 * call VOP_CLOSE() and VOP_INACTIVE().
4311 	 *
4312 	 * It could be that VOP_INACTIVE() requested reclamation, in
4313 	 * which case we should avoid recursion, so check
4314 	 * VI_DOINGINACT.  This is not precise but good enough.
4315 	 */
4316 	active = vp->v_usecount > 0;
4317 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4318 	doinginact = (vp->v_iflag & VI_DOINGINACT) != 0;
4319 
4320 	/*
4321 	 * If we need to do inactive VI_OWEINACT will be set.
4322 	 */
4323 	if (vp->v_iflag & VI_DEFINACT) {
4324 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
4325 		vp->v_iflag &= ~VI_DEFINACT;
4326 		vdropl(vp);
4327 	} else {
4328 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
4329 		VI_UNLOCK(vp);
4330 	}
4331 	cache_purge_vgone(vp);
4332 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
4333 
4334 	/*
4335 	 * If purging an active vnode, it must be closed and
4336 	 * deactivated before being reclaimed.
4337 	 */
4338 	if (active)
4339 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
4340 	if (!doinginact) {
4341 		do {
4342 			if (oweinact || active) {
4343 				VI_LOCK(vp);
4344 				vinactivef(vp);
4345 				oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4346 				VI_UNLOCK(vp);
4347 			}
4348 		} while (oweinact);
4349 	}
4350 	if (vp->v_type == VSOCK)
4351 		vfs_unp_reclaim(vp);
4352 
4353 	/*
4354 	 * Clean out any buffers associated with the vnode.
4355 	 * If the flush fails, just toss the buffers.
4356 	 */
4357 	mp = NULL;
4358 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
4359 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
4360 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
4361 		while (vinvalbuf(vp, 0, 0, 0) != 0)
4362 			;
4363 	}
4364 
4365 	BO_LOCK(&vp->v_bufobj);
4366 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
4367 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
4368 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
4369 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
4370 	    ("vp %p bufobj not invalidated", vp));
4371 
4372 	/*
4373 	 * For VMIO bufobj, BO_DEAD is set later, or in
4374 	 * vm_object_terminate() after the object's page queue is
4375 	 * flushed.
4376 	 */
4377 	object = vp->v_bufobj.bo_object;
4378 	if (object == NULL)
4379 		vp->v_bufobj.bo_flag |= BO_DEAD;
4380 	BO_UNLOCK(&vp->v_bufobj);
4381 
4382 	/*
4383 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
4384 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
4385 	 * should not touch the object borrowed from the lower vnode
4386 	 * (the handle check).
4387 	 */
4388 	if (object != NULL && object->type == OBJT_VNODE &&
4389 	    object->handle == vp)
4390 		vnode_destroy_vobject(vp);
4391 
4392 	/*
4393 	 * Reclaim the vnode.
4394 	 */
4395 	if (VOP_RECLAIM(vp))
4396 		panic("vgone: cannot reclaim");
4397 	if (mp != NULL)
4398 		vn_finished_secondary_write(mp);
4399 	VNASSERT(vp->v_object == NULL, vp,
4400 	    ("vop_reclaim left v_object vp=%p", vp));
4401 	/*
4402 	 * Clear the advisory locks and wake up waiting threads.
4403 	 */
4404 	if (vp->v_lockf != NULL) {
4405 		(void)VOP_ADVLOCKPURGE(vp);
4406 		vp->v_lockf = NULL;
4407 	}
4408 	/*
4409 	 * Delete from old mount point vnode list.
4410 	 */
4411 	if (vp->v_mount == NULL) {
4412 		VI_LOCK(vp);
4413 	} else {
4414 		delmntque(vp);
4415 		ASSERT_VI_LOCKED(vp, "vgonel 2");
4416 	}
4417 	/*
4418 	 * Done with purge, reset to the standard lock and invalidate
4419 	 * the vnode.
4420 	 */
4421 	vp->v_vnlock = &vp->v_lock;
4422 	vp->v_op = &dead_vnodeops;
4423 	vp->v_type = VBAD;
4424 	vn_set_state(vp, VSTATE_DEAD);
4425 }
4426 
4427 /*
4428  * Print out a description of a vnode.
4429  */
4430 static const char *const vtypename[] = {
4431 	[VNON] = "VNON",
4432 	[VREG] = "VREG",
4433 	[VDIR] = "VDIR",
4434 	[VBLK] = "VBLK",
4435 	[VCHR] = "VCHR",
4436 	[VLNK] = "VLNK",
4437 	[VSOCK] = "VSOCK",
4438 	[VFIFO] = "VFIFO",
4439 	[VBAD] = "VBAD",
4440 	[VMARKER] = "VMARKER",
4441 };
4442 _Static_assert(nitems(vtypename) == VLASTTYPE + 1,
4443     "vnode type name not added to vtypename");
4444 
4445 static const char *const vstatename[] = {
4446 	[VSTATE_UNINITIALIZED] = "VSTATE_UNINITIALIZED",
4447 	[VSTATE_CONSTRUCTED] = "VSTATE_CONSTRUCTED",
4448 	[VSTATE_DESTROYING] = "VSTATE_DESTROYING",
4449 	[VSTATE_DEAD] = "VSTATE_DEAD",
4450 };
4451 _Static_assert(nitems(vstatename) == VLASTSTATE + 1,
4452     "vnode state name not added to vstatename");
4453 
4454 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0,
4455     "new hold count flag not added to vn_printf");
4456 
4457 void
4458 vn_printf(struct vnode *vp, const char *fmt, ...)
4459 {
4460 	va_list ap;
4461 	char buf[256], buf2[16];
4462 	u_long flags;
4463 	u_int holdcnt;
4464 	short irflag;
4465 
4466 	va_start(ap, fmt);
4467 	vprintf(fmt, ap);
4468 	va_end(ap);
4469 	printf("%p: ", (void *)vp);
4470 	printf("type %s state %s op %p\n", vtypename[vp->v_type],
4471 	    vstatename[vp->v_state], vp->v_op);
4472 	holdcnt = atomic_load_int(&vp->v_holdcnt);
4473 	printf("    usecount %d, writecount %d, refcount %d seqc users %d",
4474 	    vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS,
4475 	    vp->v_seqc_users);
4476 	switch (vp->v_type) {
4477 	case VDIR:
4478 		printf(" mountedhere %p\n", vp->v_mountedhere);
4479 		break;
4480 	case VCHR:
4481 		printf(" rdev %p\n", vp->v_rdev);
4482 		break;
4483 	case VSOCK:
4484 		printf(" socket %p\n", vp->v_unpcb);
4485 		break;
4486 	case VFIFO:
4487 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
4488 		break;
4489 	default:
4490 		printf("\n");
4491 		break;
4492 	}
4493 	buf[0] = '\0';
4494 	buf[1] = '\0';
4495 	if (holdcnt & VHOLD_NO_SMR)
4496 		strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf));
4497 	printf("    hold count flags (%s)\n", buf + 1);
4498 
4499 	buf[0] = '\0';
4500 	buf[1] = '\0';
4501 	irflag = vn_irflag_read(vp);
4502 	if (irflag & VIRF_DOOMED)
4503 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4504 	if (irflag & VIRF_PGREAD)
4505 		strlcat(buf, "|VIRF_PGREAD", sizeof(buf));
4506 	if (irflag & VIRF_MOUNTPOINT)
4507 		strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf));
4508 	if (irflag & VIRF_TEXT_REF)
4509 		strlcat(buf, "|VIRF_TEXT_REF", sizeof(buf));
4510 	flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT | VIRF_TEXT_REF);
4511 	if (flags != 0) {
4512 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4513 		strlcat(buf, buf2, sizeof(buf));
4514 	}
4515 	if (vp->v_vflag & VV_ROOT)
4516 		strlcat(buf, "|VV_ROOT", sizeof(buf));
4517 	if (vp->v_vflag & VV_ISTTY)
4518 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
4519 	if (vp->v_vflag & VV_NOSYNC)
4520 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4521 	if (vp->v_vflag & VV_ETERNALDEV)
4522 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4523 	if (vp->v_vflag & VV_CACHEDLABEL)
4524 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4525 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
4526 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4527 	if (vp->v_vflag & VV_COPYONWRITE)
4528 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4529 	if (vp->v_vflag & VV_SYSTEM)
4530 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4531 	if (vp->v_vflag & VV_PROCDEP)
4532 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4533 	if (vp->v_vflag & VV_DELETED)
4534 		strlcat(buf, "|VV_DELETED", sizeof(buf));
4535 	if (vp->v_vflag & VV_MD)
4536 		strlcat(buf, "|VV_MD", sizeof(buf));
4537 	if (vp->v_vflag & VV_FORCEINSMQ)
4538 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4539 	if (vp->v_vflag & VV_READLINK)
4540 		strlcat(buf, "|VV_READLINK", sizeof(buf));
4541 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4542 	    VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM |
4543 	    VV_PROCDEP | VV_DELETED | VV_MD | VV_FORCEINSMQ | VV_READLINK);
4544 	if (flags != 0) {
4545 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4546 		strlcat(buf, buf2, sizeof(buf));
4547 	}
4548 	if (vp->v_iflag & VI_MOUNT)
4549 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
4550 	if (vp->v_iflag & VI_DOINGINACT)
4551 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4552 	if (vp->v_iflag & VI_OWEINACT)
4553 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4554 	if (vp->v_iflag & VI_DEFINACT)
4555 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4556 	if (vp->v_iflag & VI_FOPENING)
4557 		strlcat(buf, "|VI_FOPENING", sizeof(buf));
4558 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOINGINACT |
4559 	    VI_OWEINACT | VI_DEFINACT | VI_FOPENING);
4560 	if (flags != 0) {
4561 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4562 		strlcat(buf, buf2, sizeof(buf));
4563 	}
4564 	if (vp->v_mflag & VMP_LAZYLIST)
4565 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4566 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
4567 	if (flags != 0) {
4568 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4569 		strlcat(buf, buf2, sizeof(buf));
4570 	}
4571 	printf("    flags (%s)", buf + 1);
4572 	if (mtx_owned(VI_MTX(vp)))
4573 		printf(" VI_LOCKed");
4574 	printf("\n");
4575 	if (vp->v_object != NULL)
4576 		printf("    v_object %p ref %d pages %d "
4577 		    "cleanbuf %d dirtybuf %d\n",
4578 		    vp->v_object, vp->v_object->ref_count,
4579 		    vp->v_object->resident_page_count,
4580 		    vp->v_bufobj.bo_clean.bv_cnt,
4581 		    vp->v_bufobj.bo_dirty.bv_cnt);
4582 	printf("    ");
4583 	lockmgr_printinfo(vp->v_vnlock);
4584 	if (vp->v_data != NULL)
4585 		VOP_PRINT(vp);
4586 }
4587 
4588 #ifdef DDB
4589 /*
4590  * List all of the locked vnodes in the system.
4591  * Called when debugging the kernel.
4592  */
4593 DB_SHOW_COMMAND_FLAGS(lockedvnods, lockedvnodes, DB_CMD_MEMSAFE)
4594 {
4595 	struct mount *mp;
4596 	struct vnode *vp;
4597 
4598 	/*
4599 	 * Note: because this is DDB, we can't obey the locking semantics
4600 	 * for these structures, which means we could catch an inconsistent
4601 	 * state and dereference a nasty pointer.  Not much to be done
4602 	 * about that.
4603 	 */
4604 	db_printf("Locked vnodes\n");
4605 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4606 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4607 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4608 				vn_printf(vp, "vnode ");
4609 		}
4610 	}
4611 }
4612 
4613 /*
4614  * Show details about the given vnode.
4615  */
4616 DB_SHOW_COMMAND(vnode, db_show_vnode)
4617 {
4618 	struct vnode *vp;
4619 
4620 	if (!have_addr)
4621 		return;
4622 	vp = (struct vnode *)addr;
4623 	vn_printf(vp, "vnode ");
4624 }
4625 
4626 /*
4627  * Show details about the given mount point.
4628  */
4629 DB_SHOW_COMMAND(mount, db_show_mount)
4630 {
4631 	struct mount *mp;
4632 	struct vfsopt *opt;
4633 	struct statfs *sp;
4634 	struct vnode *vp;
4635 	char buf[512];
4636 	uint64_t mflags;
4637 	u_int flags;
4638 
4639 	if (!have_addr) {
4640 		/* No address given, print short info about all mount points. */
4641 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4642 			db_printf("%p %s on %s (%s)\n", mp,
4643 			    mp->mnt_stat.f_mntfromname,
4644 			    mp->mnt_stat.f_mntonname,
4645 			    mp->mnt_stat.f_fstypename);
4646 			if (db_pager_quit)
4647 				break;
4648 		}
4649 		db_printf("\nMore info: show mount <addr>\n");
4650 		return;
4651 	}
4652 
4653 	mp = (struct mount *)addr;
4654 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4655 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4656 
4657 	buf[0] = '\0';
4658 	mflags = mp->mnt_flag;
4659 #define	MNT_FLAG(flag)	do {						\
4660 	if (mflags & (flag)) {						\
4661 		if (buf[0] != '\0')					\
4662 			strlcat(buf, ", ", sizeof(buf));		\
4663 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4664 		mflags &= ~(flag);					\
4665 	}								\
4666 } while (0)
4667 	MNT_FLAG(MNT_RDONLY);
4668 	MNT_FLAG(MNT_SYNCHRONOUS);
4669 	MNT_FLAG(MNT_NOEXEC);
4670 	MNT_FLAG(MNT_NOSUID);
4671 	MNT_FLAG(MNT_NFS4ACLS);
4672 	MNT_FLAG(MNT_UNION);
4673 	MNT_FLAG(MNT_ASYNC);
4674 	MNT_FLAG(MNT_SUIDDIR);
4675 	MNT_FLAG(MNT_SOFTDEP);
4676 	MNT_FLAG(MNT_NOSYMFOLLOW);
4677 	MNT_FLAG(MNT_GJOURNAL);
4678 	MNT_FLAG(MNT_MULTILABEL);
4679 	MNT_FLAG(MNT_ACLS);
4680 	MNT_FLAG(MNT_NOATIME);
4681 	MNT_FLAG(MNT_NOCLUSTERR);
4682 	MNT_FLAG(MNT_NOCLUSTERW);
4683 	MNT_FLAG(MNT_SUJ);
4684 	MNT_FLAG(MNT_EXRDONLY);
4685 	MNT_FLAG(MNT_EXPORTED);
4686 	MNT_FLAG(MNT_DEFEXPORTED);
4687 	MNT_FLAG(MNT_EXPORTANON);
4688 	MNT_FLAG(MNT_EXKERB);
4689 	MNT_FLAG(MNT_EXPUBLIC);
4690 	MNT_FLAG(MNT_LOCAL);
4691 	MNT_FLAG(MNT_QUOTA);
4692 	MNT_FLAG(MNT_ROOTFS);
4693 	MNT_FLAG(MNT_USER);
4694 	MNT_FLAG(MNT_IGNORE);
4695 	MNT_FLAG(MNT_UPDATE);
4696 	MNT_FLAG(MNT_DELEXPORT);
4697 	MNT_FLAG(MNT_RELOAD);
4698 	MNT_FLAG(MNT_FORCE);
4699 	MNT_FLAG(MNT_SNAPSHOT);
4700 	MNT_FLAG(MNT_BYFSID);
4701 #undef MNT_FLAG
4702 	if (mflags != 0) {
4703 		if (buf[0] != '\0')
4704 			strlcat(buf, ", ", sizeof(buf));
4705 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4706 		    "0x%016jx", mflags);
4707 	}
4708 	db_printf("    mnt_flag = %s\n", buf);
4709 
4710 	buf[0] = '\0';
4711 	flags = mp->mnt_kern_flag;
4712 #define	MNT_KERN_FLAG(flag)	do {					\
4713 	if (flags & (flag)) {						\
4714 		if (buf[0] != '\0')					\
4715 			strlcat(buf, ", ", sizeof(buf));		\
4716 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4717 		flags &= ~(flag);					\
4718 	}								\
4719 } while (0)
4720 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4721 	MNT_KERN_FLAG(MNTK_ASYNC);
4722 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4723 	MNT_KERN_FLAG(MNTK_NOMSYNC);
4724 	MNT_KERN_FLAG(MNTK_DRAINING);
4725 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4726 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4727 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4728 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4729 	MNT_KERN_FLAG(MNTK_RECURSE);
4730 	MNT_KERN_FLAG(MNTK_UPPER_WAITER);
4731 	MNT_KERN_FLAG(MNTK_UNLOCKED_INSMNTQUE);
4732 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4733 	MNT_KERN_FLAG(MNTK_VMSETSIZE_BUG);
4734 	MNT_KERN_FLAG(MNTK_FPLOOKUP);
4735 	MNT_KERN_FLAG(MNTK_TASKQUEUE_WAITER);
4736 	MNT_KERN_FLAG(MNTK_NOASYNC);
4737 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4738 	MNT_KERN_FLAG(MNTK_MWAIT);
4739 	MNT_KERN_FLAG(MNTK_SUSPEND);
4740 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4741 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4742 	MNT_KERN_FLAG(MNTK_NULL_NOCACHE);
4743 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4744 #undef MNT_KERN_FLAG
4745 	if (flags != 0) {
4746 		if (buf[0] != '\0')
4747 			strlcat(buf, ", ", sizeof(buf));
4748 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4749 		    "0x%08x", flags);
4750 	}
4751 	db_printf("    mnt_kern_flag = %s\n", buf);
4752 
4753 	db_printf("    mnt_opt = ");
4754 	opt = TAILQ_FIRST(mp->mnt_opt);
4755 	if (opt != NULL) {
4756 		db_printf("%s", opt->name);
4757 		opt = TAILQ_NEXT(opt, link);
4758 		while (opt != NULL) {
4759 			db_printf(", %s", opt->name);
4760 			opt = TAILQ_NEXT(opt, link);
4761 		}
4762 	}
4763 	db_printf("\n");
4764 
4765 	sp = &mp->mnt_stat;
4766 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4767 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4768 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4769 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4770 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4771 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4772 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4773 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4774 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4775 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4776 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4777 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4778 
4779 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4780 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4781 	if (jailed(mp->mnt_cred))
4782 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4783 	db_printf(" }\n");
4784 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4785 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4786 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4787 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4788 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4789 	    mp->mnt_lazyvnodelistsize);
4790 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4791 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4792 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4793 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4794 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4795 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4796 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4797 	db_printf("    mnt_secondary_accwrites = %d\n",
4798 	    mp->mnt_secondary_accwrites);
4799 	db_printf("    mnt_gjprovider = %s\n",
4800 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4801 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4802 
4803 	db_printf("\n\nList of active vnodes\n");
4804 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4805 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4806 			vn_printf(vp, "vnode ");
4807 			if (db_pager_quit)
4808 				break;
4809 		}
4810 	}
4811 	db_printf("\n\nList of inactive vnodes\n");
4812 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4813 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4814 			vn_printf(vp, "vnode ");
4815 			if (db_pager_quit)
4816 				break;
4817 		}
4818 	}
4819 }
4820 #endif	/* DDB */
4821 
4822 /*
4823  * Fill in a struct xvfsconf based on a struct vfsconf.
4824  */
4825 static int
4826 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4827 {
4828 	struct xvfsconf xvfsp;
4829 
4830 	bzero(&xvfsp, sizeof(xvfsp));
4831 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4832 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4833 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4834 	xvfsp.vfc_flags = vfsp->vfc_flags;
4835 	/*
4836 	 * These are unused in userland, we keep them
4837 	 * to not break binary compatibility.
4838 	 */
4839 	xvfsp.vfc_vfsops = NULL;
4840 	xvfsp.vfc_next = NULL;
4841 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4842 }
4843 
4844 #ifdef COMPAT_FREEBSD32
4845 struct xvfsconf32 {
4846 	uint32_t	vfc_vfsops;
4847 	char		vfc_name[MFSNAMELEN];
4848 	int32_t		vfc_typenum;
4849 	int32_t		vfc_refcount;
4850 	int32_t		vfc_flags;
4851 	uint32_t	vfc_next;
4852 };
4853 
4854 static int
4855 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4856 {
4857 	struct xvfsconf32 xvfsp;
4858 
4859 	bzero(&xvfsp, sizeof(xvfsp));
4860 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4861 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4862 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4863 	xvfsp.vfc_flags = vfsp->vfc_flags;
4864 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4865 }
4866 #endif
4867 
4868 /*
4869  * Top level filesystem related information gathering.
4870  */
4871 static int
4872 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4873 {
4874 	struct vfsconf *vfsp;
4875 	int error;
4876 
4877 	error = 0;
4878 	vfsconf_slock();
4879 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4880 #ifdef COMPAT_FREEBSD32
4881 		if (req->flags & SCTL_MASK32)
4882 			error = vfsconf2x32(req, vfsp);
4883 		else
4884 #endif
4885 			error = vfsconf2x(req, vfsp);
4886 		if (error)
4887 			break;
4888 	}
4889 	vfsconf_sunlock();
4890 	return (error);
4891 }
4892 
4893 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4894     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4895     "S,xvfsconf", "List of all configured filesystems");
4896 
4897 #ifndef BURN_BRIDGES
4898 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4899 
4900 static int
4901 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4902 {
4903 	int *name = (int *)arg1 - 1;	/* XXX */
4904 	u_int namelen = arg2 + 1;	/* XXX */
4905 	struct vfsconf *vfsp;
4906 
4907 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4908 	    "please rebuild world\n");
4909 
4910 #if 1 || defined(COMPAT_PRELITE2)
4911 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4912 	if (namelen == 1)
4913 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4914 #endif
4915 
4916 	switch (name[1]) {
4917 	case VFS_MAXTYPENUM:
4918 		if (namelen != 2)
4919 			return (ENOTDIR);
4920 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4921 	case VFS_CONF:
4922 		if (namelen != 3)
4923 			return (ENOTDIR);	/* overloaded */
4924 		vfsconf_slock();
4925 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4926 			if (vfsp->vfc_typenum == name[2])
4927 				break;
4928 		}
4929 		vfsconf_sunlock();
4930 		if (vfsp == NULL)
4931 			return (EOPNOTSUPP);
4932 #ifdef COMPAT_FREEBSD32
4933 		if (req->flags & SCTL_MASK32)
4934 			return (vfsconf2x32(req, vfsp));
4935 		else
4936 #endif
4937 			return (vfsconf2x(req, vfsp));
4938 	}
4939 	return (EOPNOTSUPP);
4940 }
4941 
4942 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4943     CTLFLAG_MPSAFE, vfs_sysctl,
4944     "Generic filesystem");
4945 
4946 #if 1 || defined(COMPAT_PRELITE2)
4947 
4948 static int
4949 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4950 {
4951 	int error;
4952 	struct vfsconf *vfsp;
4953 	struct ovfsconf ovfs;
4954 
4955 	vfsconf_slock();
4956 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4957 		bzero(&ovfs, sizeof(ovfs));
4958 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4959 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4960 		ovfs.vfc_index = vfsp->vfc_typenum;
4961 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4962 		ovfs.vfc_flags = vfsp->vfc_flags;
4963 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4964 		if (error != 0) {
4965 			vfsconf_sunlock();
4966 			return (error);
4967 		}
4968 	}
4969 	vfsconf_sunlock();
4970 	return (0);
4971 }
4972 
4973 #endif /* 1 || COMPAT_PRELITE2 */
4974 #endif /* !BURN_BRIDGES */
4975 
4976 static void
4977 unmount_or_warn(struct mount *mp)
4978 {
4979 	int error;
4980 
4981 	error = dounmount(mp, MNT_FORCE, curthread);
4982 	if (error != 0) {
4983 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4984 		if (error == EBUSY)
4985 			printf("BUSY)\n");
4986 		else
4987 			printf("%d)\n", error);
4988 	}
4989 }
4990 
4991 /*
4992  * Unmount all filesystems. The list is traversed in reverse order
4993  * of mounting to avoid dependencies.
4994  */
4995 void
4996 vfs_unmountall(void)
4997 {
4998 	struct mount *mp, *tmp;
4999 
5000 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
5001 
5002 	/*
5003 	 * Since this only runs when rebooting, it is not interlocked.
5004 	 */
5005 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
5006 		vfs_ref(mp);
5007 
5008 		/*
5009 		 * Forcibly unmounting "/dev" before "/" would prevent clean
5010 		 * unmount of the latter.
5011 		 */
5012 		if (mp == rootdevmp)
5013 			continue;
5014 
5015 		unmount_or_warn(mp);
5016 	}
5017 
5018 	if (rootdevmp != NULL)
5019 		unmount_or_warn(rootdevmp);
5020 }
5021 
5022 static void
5023 vfs_deferred_inactive(struct vnode *vp, int lkflags)
5024 {
5025 
5026 	ASSERT_VI_LOCKED(vp, __func__);
5027 	VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
5028 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
5029 		vdropl(vp);
5030 		return;
5031 	}
5032 	if (vn_lock(vp, lkflags) == 0) {
5033 		VI_LOCK(vp);
5034 		vinactive(vp);
5035 		VOP_UNLOCK(vp);
5036 		vdropl(vp);
5037 		return;
5038 	}
5039 	vdefer_inactive_unlocked(vp);
5040 }
5041 
5042 static int
5043 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
5044 {
5045 
5046 	return (vp->v_iflag & VI_DEFINACT);
5047 }
5048 
5049 static void __noinline
5050 vfs_periodic_inactive(struct mount *mp, int flags)
5051 {
5052 	struct vnode *vp, *mvp;
5053 	int lkflags;
5054 
5055 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
5056 	if (flags != MNT_WAIT)
5057 		lkflags |= LK_NOWAIT;
5058 
5059 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
5060 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
5061 			VI_UNLOCK(vp);
5062 			continue;
5063 		}
5064 		vp->v_iflag &= ~VI_DEFINACT;
5065 		vfs_deferred_inactive(vp, lkflags);
5066 	}
5067 }
5068 
5069 static inline bool
5070 vfs_want_msync(struct vnode *vp)
5071 {
5072 	struct vm_object *obj;
5073 
5074 	/*
5075 	 * This test may be performed without any locks held.
5076 	 * We rely on vm_object's type stability.
5077 	 */
5078 	if (vp->v_vflag & VV_NOSYNC)
5079 		return (false);
5080 	obj = vp->v_object;
5081 	return (obj != NULL && vm_object_mightbedirty(obj));
5082 }
5083 
5084 static int
5085 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
5086 {
5087 
5088 	if (vp->v_vflag & VV_NOSYNC)
5089 		return (false);
5090 	if (vp->v_iflag & VI_DEFINACT)
5091 		return (true);
5092 	return (vfs_want_msync(vp));
5093 }
5094 
5095 static void __noinline
5096 vfs_periodic_msync_inactive(struct mount *mp, int flags)
5097 {
5098 	struct vnode *vp, *mvp;
5099 	struct vm_object *obj;
5100 	int lkflags, objflags;
5101 	bool seen_defer;
5102 
5103 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
5104 	if (flags != MNT_WAIT) {
5105 		lkflags |= LK_NOWAIT;
5106 		objflags = OBJPC_NOSYNC;
5107 	} else {
5108 		objflags = OBJPC_SYNC;
5109 	}
5110 
5111 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
5112 		seen_defer = false;
5113 		if (vp->v_iflag & VI_DEFINACT) {
5114 			vp->v_iflag &= ~VI_DEFINACT;
5115 			seen_defer = true;
5116 		}
5117 		if (!vfs_want_msync(vp)) {
5118 			if (seen_defer)
5119 				vfs_deferred_inactive(vp, lkflags);
5120 			else
5121 				VI_UNLOCK(vp);
5122 			continue;
5123 		}
5124 		if (vget(vp, lkflags) == 0) {
5125 			obj = vp->v_object;
5126 			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
5127 				VM_OBJECT_WLOCK(obj);
5128 				vm_object_page_clean(obj, 0, 0, objflags);
5129 				VM_OBJECT_WUNLOCK(obj);
5130 			}
5131 			vput(vp);
5132 			if (seen_defer)
5133 				vdrop(vp);
5134 		} else {
5135 			if (seen_defer)
5136 				vdefer_inactive_unlocked(vp);
5137 		}
5138 	}
5139 }
5140 
5141 void
5142 vfs_periodic(struct mount *mp, int flags)
5143 {
5144 
5145 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
5146 
5147 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
5148 		vfs_periodic_inactive(mp, flags);
5149 	else
5150 		vfs_periodic_msync_inactive(mp, flags);
5151 }
5152 
5153 static void
5154 destroy_vpollinfo_free(struct vpollinfo *vi)
5155 {
5156 
5157 	knlist_destroy(&vi->vpi_selinfo.si_note);
5158 	mtx_destroy(&vi->vpi_lock);
5159 	free(vi, M_VNODEPOLL);
5160 }
5161 
5162 static void
5163 destroy_vpollinfo(struct vpollinfo *vi)
5164 {
5165 
5166 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
5167 	seldrain(&vi->vpi_selinfo);
5168 	destroy_vpollinfo_free(vi);
5169 }
5170 
5171 /*
5172  * Initialize per-vnode helper structure to hold poll-related state.
5173  */
5174 void
5175 v_addpollinfo(struct vnode *vp)
5176 {
5177 	struct vpollinfo *vi;
5178 
5179 	if (vp->v_pollinfo != NULL)
5180 		return;
5181 	vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO);
5182 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
5183 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
5184 	    vfs_knlunlock, vfs_knl_assert_lock);
5185 	VI_LOCK(vp);
5186 	if (vp->v_pollinfo != NULL) {
5187 		VI_UNLOCK(vp);
5188 		destroy_vpollinfo_free(vi);
5189 		return;
5190 	}
5191 	vp->v_pollinfo = vi;
5192 	VI_UNLOCK(vp);
5193 }
5194 
5195 /*
5196  * Record a process's interest in events which might happen to
5197  * a vnode.  Because poll uses the historic select-style interface
5198  * internally, this routine serves as both the ``check for any
5199  * pending events'' and the ``record my interest in future events''
5200  * functions.  (These are done together, while the lock is held,
5201  * to avoid race conditions.)
5202  */
5203 int
5204 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
5205 {
5206 
5207 	v_addpollinfo(vp);
5208 	mtx_lock(&vp->v_pollinfo->vpi_lock);
5209 	if (vp->v_pollinfo->vpi_revents & events) {
5210 		/*
5211 		 * This leaves events we are not interested
5212 		 * in available for the other process which
5213 		 * which presumably had requested them
5214 		 * (otherwise they would never have been
5215 		 * recorded).
5216 		 */
5217 		events &= vp->v_pollinfo->vpi_revents;
5218 		vp->v_pollinfo->vpi_revents &= ~events;
5219 
5220 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
5221 		return (events);
5222 	}
5223 	vp->v_pollinfo->vpi_events |= events;
5224 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
5225 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
5226 	return (0);
5227 }
5228 
5229 /*
5230  * Routine to create and manage a filesystem syncer vnode.
5231  */
5232 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
5233 static int	sync_fsync(struct  vop_fsync_args *);
5234 static int	sync_inactive(struct  vop_inactive_args *);
5235 static int	sync_reclaim(struct  vop_reclaim_args *);
5236 
5237 static struct vop_vector sync_vnodeops = {
5238 	.vop_bypass =	VOP_EOPNOTSUPP,
5239 	.vop_close =	sync_close,
5240 	.vop_fsync =	sync_fsync,
5241 	.vop_getwritemount = vop_stdgetwritemount,
5242 	.vop_inactive =	sync_inactive,
5243 	.vop_need_inactive = vop_stdneed_inactive,
5244 	.vop_reclaim =	sync_reclaim,
5245 	.vop_lock1 =	vop_stdlock,
5246 	.vop_unlock =	vop_stdunlock,
5247 	.vop_islocked =	vop_stdislocked,
5248 	.vop_fplookup_vexec = VOP_EAGAIN,
5249 	.vop_fplookup_symlink = VOP_EAGAIN,
5250 };
5251 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
5252 
5253 /*
5254  * Create a new filesystem syncer vnode for the specified mount point.
5255  */
5256 void
5257 vfs_allocate_syncvnode(struct mount *mp)
5258 {
5259 	struct vnode *vp;
5260 	struct bufobj *bo;
5261 	static long start, incr, next;
5262 	int error;
5263 
5264 	/* Allocate a new vnode */
5265 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
5266 	if (error != 0)
5267 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
5268 	vp->v_type = VNON;
5269 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5270 	vp->v_vflag |= VV_FORCEINSMQ;
5271 	error = insmntque1(vp, mp);
5272 	if (error != 0)
5273 		panic("vfs_allocate_syncvnode: insmntque() failed");
5274 	vp->v_vflag &= ~VV_FORCEINSMQ;
5275 	vn_set_state(vp, VSTATE_CONSTRUCTED);
5276 	VOP_UNLOCK(vp);
5277 	/*
5278 	 * Place the vnode onto the syncer worklist. We attempt to
5279 	 * scatter them about on the list so that they will go off
5280 	 * at evenly distributed times even if all the filesystems
5281 	 * are mounted at once.
5282 	 */
5283 	next += incr;
5284 	if (next == 0 || next > syncer_maxdelay) {
5285 		start /= 2;
5286 		incr /= 2;
5287 		if (start == 0) {
5288 			start = syncer_maxdelay / 2;
5289 			incr = syncer_maxdelay;
5290 		}
5291 		next = start;
5292 	}
5293 	bo = &vp->v_bufobj;
5294 	BO_LOCK(bo);
5295 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
5296 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
5297 	mtx_lock(&sync_mtx);
5298 	sync_vnode_count++;
5299 	if (mp->mnt_syncer == NULL) {
5300 		mp->mnt_syncer = vp;
5301 		vp = NULL;
5302 	}
5303 	mtx_unlock(&sync_mtx);
5304 	BO_UNLOCK(bo);
5305 	if (vp != NULL) {
5306 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5307 		vgone(vp);
5308 		vput(vp);
5309 	}
5310 }
5311 
5312 void
5313 vfs_deallocate_syncvnode(struct mount *mp)
5314 {
5315 	struct vnode *vp;
5316 
5317 	mtx_lock(&sync_mtx);
5318 	vp = mp->mnt_syncer;
5319 	if (vp != NULL)
5320 		mp->mnt_syncer = NULL;
5321 	mtx_unlock(&sync_mtx);
5322 	if (vp != NULL)
5323 		vrele(vp);
5324 }
5325 
5326 /*
5327  * Do a lazy sync of the filesystem.
5328  */
5329 static int
5330 sync_fsync(struct vop_fsync_args *ap)
5331 {
5332 	struct vnode *syncvp = ap->a_vp;
5333 	struct mount *mp = syncvp->v_mount;
5334 	int error, save;
5335 	struct bufobj *bo;
5336 
5337 	/*
5338 	 * We only need to do something if this is a lazy evaluation.
5339 	 */
5340 	if (ap->a_waitfor != MNT_LAZY)
5341 		return (0);
5342 
5343 	/*
5344 	 * Move ourselves to the back of the sync list.
5345 	 */
5346 	bo = &syncvp->v_bufobj;
5347 	BO_LOCK(bo);
5348 	vn_syncer_add_to_worklist(bo, syncdelay);
5349 	BO_UNLOCK(bo);
5350 
5351 	/*
5352 	 * Walk the list of vnodes pushing all that are dirty and
5353 	 * not already on the sync list.
5354 	 */
5355 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
5356 		return (0);
5357 	VOP_UNLOCK(syncvp);
5358 	save = curthread_pflags_set(TDP_SYNCIO);
5359 	/*
5360 	 * The filesystem at hand may be idle with free vnodes stored in the
5361 	 * batch.  Return them instead of letting them stay there indefinitely.
5362 	 */
5363 	vfs_periodic(mp, MNT_NOWAIT);
5364 	error = VFS_SYNC(mp, MNT_LAZY);
5365 	curthread_pflags_restore(save);
5366 	vn_lock(syncvp, LK_EXCLUSIVE | LK_RETRY);
5367 	vfs_unbusy(mp);
5368 	return (error);
5369 }
5370 
5371 /*
5372  * The syncer vnode is no referenced.
5373  */
5374 static int
5375 sync_inactive(struct vop_inactive_args *ap)
5376 {
5377 
5378 	vgone(ap->a_vp);
5379 	return (0);
5380 }
5381 
5382 /*
5383  * The syncer vnode is no longer needed and is being decommissioned.
5384  *
5385  * Modifications to the worklist must be protected by sync_mtx.
5386  */
5387 static int
5388 sync_reclaim(struct vop_reclaim_args *ap)
5389 {
5390 	struct vnode *vp = ap->a_vp;
5391 	struct bufobj *bo;
5392 
5393 	bo = &vp->v_bufobj;
5394 	BO_LOCK(bo);
5395 	mtx_lock(&sync_mtx);
5396 	if (vp->v_mount->mnt_syncer == vp)
5397 		vp->v_mount->mnt_syncer = NULL;
5398 	if (bo->bo_flag & BO_ONWORKLST) {
5399 		LIST_REMOVE(bo, bo_synclist);
5400 		syncer_worklist_len--;
5401 		sync_vnode_count--;
5402 		bo->bo_flag &= ~BO_ONWORKLST;
5403 	}
5404 	mtx_unlock(&sync_mtx);
5405 	BO_UNLOCK(bo);
5406 
5407 	return (0);
5408 }
5409 
5410 int
5411 vn_need_pageq_flush(struct vnode *vp)
5412 {
5413 	struct vm_object *obj;
5414 
5415 	obj = vp->v_object;
5416 	return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5417 	    vm_object_mightbedirty(obj));
5418 }
5419 
5420 /*
5421  * Check if vnode represents a disk device
5422  */
5423 bool
5424 vn_isdisk_error(struct vnode *vp, int *errp)
5425 {
5426 	int error;
5427 
5428 	if (vp->v_type != VCHR) {
5429 		error = ENOTBLK;
5430 		goto out;
5431 	}
5432 	error = 0;
5433 	dev_lock();
5434 	if (vp->v_rdev == NULL)
5435 		error = ENXIO;
5436 	else if (vp->v_rdev->si_devsw == NULL)
5437 		error = ENXIO;
5438 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5439 		error = ENOTBLK;
5440 	dev_unlock();
5441 out:
5442 	*errp = error;
5443 	return (error == 0);
5444 }
5445 
5446 bool
5447 vn_isdisk(struct vnode *vp)
5448 {
5449 	int error;
5450 
5451 	return (vn_isdisk_error(vp, &error));
5452 }
5453 
5454 /*
5455  * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
5456  * the comment above cache_fplookup for details.
5457  */
5458 int
5459 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred)
5460 {
5461 	int error;
5462 
5463 	VFS_SMR_ASSERT_ENTERED();
5464 
5465 	/* Check the owner. */
5466 	if (cred->cr_uid == file_uid) {
5467 		if (file_mode & S_IXUSR)
5468 			return (0);
5469 		goto out_error;
5470 	}
5471 
5472 	/* Otherwise, check the groups (first match) */
5473 	if (groupmember(file_gid, cred)) {
5474 		if (file_mode & S_IXGRP)
5475 			return (0);
5476 		goto out_error;
5477 	}
5478 
5479 	/* Otherwise, check everyone else. */
5480 	if (file_mode & S_IXOTH)
5481 		return (0);
5482 out_error:
5483 	/*
5484 	 * Permission check failed, but it is possible denial will get overwritten
5485 	 * (e.g., when root is traversing through a 700 directory owned by someone
5486 	 * else).
5487 	 *
5488 	 * vaccess() calls priv_check_cred which in turn can descent into MAC
5489 	 * modules overriding this result. It's quite unclear what semantics
5490 	 * are allowed for them to operate, thus for safety we don't call them
5491 	 * from within the SMR section. This also means if any such modules
5492 	 * are present, we have to let the regular lookup decide.
5493 	 */
5494 	error = priv_check_cred_vfs_lookup_nomac(cred);
5495 	switch (error) {
5496 	case 0:
5497 		return (0);
5498 	case EAGAIN:
5499 		/*
5500 		 * MAC modules present.
5501 		 */
5502 		return (EAGAIN);
5503 	case EPERM:
5504 		return (EACCES);
5505 	default:
5506 		return (error);
5507 	}
5508 }
5509 
5510 /*
5511  * Common filesystem object access control check routine.  Accepts a
5512  * vnode's type, "mode", uid and gid, requested access mode, and credentials.
5513  * Returns 0 on success, or an errno on failure.
5514  */
5515 int
5516 vaccess(__enum_uint8(vtype) type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5517     accmode_t accmode, struct ucred *cred)
5518 {
5519 	accmode_t dac_granted;
5520 	accmode_t priv_granted;
5521 
5522 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5523 	    ("invalid bit in accmode"));
5524 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5525 	    ("VAPPEND without VWRITE"));
5526 
5527 	/*
5528 	 * Look for a normal, non-privileged way to access the file/directory
5529 	 * as requested.  If it exists, go with that.
5530 	 */
5531 
5532 	dac_granted = 0;
5533 
5534 	/* Check the owner. */
5535 	if (cred->cr_uid == file_uid) {
5536 		dac_granted |= VADMIN;
5537 		if (file_mode & S_IXUSR)
5538 			dac_granted |= VEXEC;
5539 		if (file_mode & S_IRUSR)
5540 			dac_granted |= VREAD;
5541 		if (file_mode & S_IWUSR)
5542 			dac_granted |= (VWRITE | VAPPEND);
5543 
5544 		if ((accmode & dac_granted) == accmode)
5545 			return (0);
5546 
5547 		goto privcheck;
5548 	}
5549 
5550 	/* Otherwise, check the groups (first match) */
5551 	if (groupmember(file_gid, cred)) {
5552 		if (file_mode & S_IXGRP)
5553 			dac_granted |= VEXEC;
5554 		if (file_mode & S_IRGRP)
5555 			dac_granted |= VREAD;
5556 		if (file_mode & S_IWGRP)
5557 			dac_granted |= (VWRITE | VAPPEND);
5558 
5559 		if ((accmode & dac_granted) == accmode)
5560 			return (0);
5561 
5562 		goto privcheck;
5563 	}
5564 
5565 	/* Otherwise, check everyone else. */
5566 	if (file_mode & S_IXOTH)
5567 		dac_granted |= VEXEC;
5568 	if (file_mode & S_IROTH)
5569 		dac_granted |= VREAD;
5570 	if (file_mode & S_IWOTH)
5571 		dac_granted |= (VWRITE | VAPPEND);
5572 	if ((accmode & dac_granted) == accmode)
5573 		return (0);
5574 
5575 privcheck:
5576 	/*
5577 	 * Build a privilege mask to determine if the set of privileges
5578 	 * satisfies the requirements when combined with the granted mask
5579 	 * from above.  For each privilege, if the privilege is required,
5580 	 * bitwise or the request type onto the priv_granted mask.
5581 	 */
5582 	priv_granted = 0;
5583 
5584 	if (type == VDIR) {
5585 		/*
5586 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5587 		 * requests, instead of PRIV_VFS_EXEC.
5588 		 */
5589 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5590 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5591 			priv_granted |= VEXEC;
5592 	} else {
5593 		/*
5594 		 * Ensure that at least one execute bit is on. Otherwise,
5595 		 * a privileged user will always succeed, and we don't want
5596 		 * this to happen unless the file really is executable.
5597 		 */
5598 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5599 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5600 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5601 			priv_granted |= VEXEC;
5602 	}
5603 
5604 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5605 	    !priv_check_cred(cred, PRIV_VFS_READ))
5606 		priv_granted |= VREAD;
5607 
5608 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5609 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5610 		priv_granted |= (VWRITE | VAPPEND);
5611 
5612 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5613 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5614 		priv_granted |= VADMIN;
5615 
5616 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5617 		return (0);
5618 	}
5619 
5620 	return ((accmode & VADMIN) ? EPERM : EACCES);
5621 }
5622 
5623 /*
5624  * Credential check based on process requesting service, and per-attribute
5625  * permissions.
5626  */
5627 int
5628 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5629     struct thread *td, accmode_t accmode)
5630 {
5631 
5632 	/*
5633 	 * Kernel-invoked always succeeds.
5634 	 */
5635 	if (cred == NOCRED)
5636 		return (0);
5637 
5638 	/*
5639 	 * Do not allow privileged processes in jail to directly manipulate
5640 	 * system attributes.
5641 	 */
5642 	switch (attrnamespace) {
5643 	case EXTATTR_NAMESPACE_SYSTEM:
5644 		/* Potentially should be: return (EPERM); */
5645 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5646 	case EXTATTR_NAMESPACE_USER:
5647 		return (VOP_ACCESS(vp, accmode, cred, td));
5648 	default:
5649 		return (EPERM);
5650 	}
5651 }
5652 
5653 #ifdef DEBUG_VFS_LOCKS
5654 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5655 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5656     "Drop into debugger on lock violation");
5657 
5658 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5659 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5660     0, "Check for interlock across VOPs");
5661 
5662 int vfs_badlock_print = 1;	/* Print lock violations. */
5663 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5664     0, "Print lock violations");
5665 
5666 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5667 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5668     0, "Print vnode details on lock violations");
5669 
5670 #ifdef KDB
5671 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5672 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5673     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5674 #endif
5675 
5676 static void
5677 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5678 {
5679 
5680 #ifdef KDB
5681 	if (vfs_badlock_backtrace)
5682 		kdb_backtrace();
5683 #endif
5684 	if (vfs_badlock_vnode)
5685 		vn_printf(vp, "vnode ");
5686 	if (vfs_badlock_print)
5687 		printf("%s: %p %s\n", str, (void *)vp, msg);
5688 	if (vfs_badlock_ddb)
5689 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5690 }
5691 
5692 void
5693 assert_vi_locked(struct vnode *vp, const char *str)
5694 {
5695 
5696 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5697 		vfs_badlock("interlock is not locked but should be", str, vp);
5698 }
5699 
5700 void
5701 assert_vi_unlocked(struct vnode *vp, const char *str)
5702 {
5703 
5704 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5705 		vfs_badlock("interlock is locked but should not be", str, vp);
5706 }
5707 
5708 void
5709 assert_vop_locked(struct vnode *vp, const char *str)
5710 {
5711 	if (KERNEL_PANICKED() || vp == NULL)
5712 		return;
5713 
5714 #ifdef WITNESS
5715 	if ((vp->v_irflag & VIRF_CROSSMP) == 0 &&
5716 	    witness_is_owned(&vp->v_vnlock->lock_object) == -1)
5717 #else
5718 	int locked = VOP_ISLOCKED(vp);
5719 	if (locked == 0 || locked == LK_EXCLOTHER)
5720 #endif
5721 		vfs_badlock("is not locked but should be", str, vp);
5722 }
5723 
5724 void
5725 assert_vop_unlocked(struct vnode *vp, const char *str)
5726 {
5727 	if (KERNEL_PANICKED() || vp == NULL)
5728 		return;
5729 
5730 #ifdef WITNESS
5731 	if ((vp->v_irflag & VIRF_CROSSMP) == 0 &&
5732 	    witness_is_owned(&vp->v_vnlock->lock_object) == 1)
5733 #else
5734 	if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5735 #endif
5736 		vfs_badlock("is locked but should not be", str, vp);
5737 }
5738 
5739 void
5740 assert_vop_elocked(struct vnode *vp, const char *str)
5741 {
5742 	if (KERNEL_PANICKED() || vp == NULL)
5743 		return;
5744 
5745 	if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5746 		vfs_badlock("is not exclusive locked but should be", str, vp);
5747 }
5748 #endif /* DEBUG_VFS_LOCKS */
5749 
5750 void
5751 vop_rename_fail(struct vop_rename_args *ap)
5752 {
5753 
5754 	if (ap->a_tvp != NULL)
5755 		vput(ap->a_tvp);
5756 	if (ap->a_tdvp == ap->a_tvp)
5757 		vrele(ap->a_tdvp);
5758 	else
5759 		vput(ap->a_tdvp);
5760 	vrele(ap->a_fdvp);
5761 	vrele(ap->a_fvp);
5762 }
5763 
5764 void
5765 vop_rename_pre(void *ap)
5766 {
5767 	struct vop_rename_args *a = ap;
5768 
5769 #ifdef DEBUG_VFS_LOCKS
5770 	if (a->a_tvp)
5771 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5772 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5773 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5774 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5775 
5776 	/* Check the source (from). */
5777 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5778 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5779 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5780 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5781 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5782 
5783 	/* Check the target. */
5784 	if (a->a_tvp)
5785 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5786 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5787 #endif
5788 	/*
5789 	 * It may be tempting to add vn_seqc_write_begin/end calls here and
5790 	 * in vop_rename_post but that's not going to work out since some
5791 	 * filesystems relookup vnodes mid-rename. This is probably a bug.
5792 	 *
5793 	 * For now filesystems are expected to do the relevant calls after they
5794 	 * decide what vnodes to operate on.
5795 	 */
5796 	if (a->a_tdvp != a->a_fdvp)
5797 		vhold(a->a_fdvp);
5798 	if (a->a_tvp != a->a_fvp)
5799 		vhold(a->a_fvp);
5800 	vhold(a->a_tdvp);
5801 	if (a->a_tvp)
5802 		vhold(a->a_tvp);
5803 }
5804 
5805 #ifdef DEBUG_VFS_LOCKS
5806 void
5807 vop_fplookup_vexec_debugpre(void *ap __unused)
5808 {
5809 
5810 	VFS_SMR_ASSERT_ENTERED();
5811 }
5812 
5813 void
5814 vop_fplookup_vexec_debugpost(void *ap, int rc)
5815 {
5816 	struct vop_fplookup_vexec_args *a;
5817 	struct vnode *vp;
5818 
5819 	a = ap;
5820 	vp = a->a_vp;
5821 
5822 	VFS_SMR_ASSERT_ENTERED();
5823 	if (rc == EOPNOTSUPP)
5824 		VNPASS(VN_IS_DOOMED(vp), vp);
5825 }
5826 
5827 void
5828 vop_fplookup_symlink_debugpre(void *ap __unused)
5829 {
5830 
5831 	VFS_SMR_ASSERT_ENTERED();
5832 }
5833 
5834 void
5835 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused)
5836 {
5837 
5838 	VFS_SMR_ASSERT_ENTERED();
5839 }
5840 
5841 static void
5842 vop_fsync_debugprepost(struct vnode *vp, const char *name)
5843 {
5844 	if (vp->v_type == VCHR)
5845 		;
5846 	else if (MNT_EXTENDED_SHARED(vp->v_mount))
5847 		ASSERT_VOP_LOCKED(vp, name);
5848 	else
5849 		ASSERT_VOP_ELOCKED(vp, name);
5850 }
5851 
5852 void
5853 vop_fsync_debugpre(void *a)
5854 {
5855 	struct vop_fsync_args *ap;
5856 
5857 	ap = a;
5858 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5859 }
5860 
5861 void
5862 vop_fsync_debugpost(void *a, int rc __unused)
5863 {
5864 	struct vop_fsync_args *ap;
5865 
5866 	ap = a;
5867 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5868 }
5869 
5870 void
5871 vop_fdatasync_debugpre(void *a)
5872 {
5873 	struct vop_fdatasync_args *ap;
5874 
5875 	ap = a;
5876 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5877 }
5878 
5879 void
5880 vop_fdatasync_debugpost(void *a, int rc __unused)
5881 {
5882 	struct vop_fdatasync_args *ap;
5883 
5884 	ap = a;
5885 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5886 }
5887 
5888 void
5889 vop_strategy_debugpre(void *ap)
5890 {
5891 	struct vop_strategy_args *a;
5892 	struct buf *bp;
5893 
5894 	a = ap;
5895 	bp = a->a_bp;
5896 
5897 	/*
5898 	 * Cluster ops lock their component buffers but not the IO container.
5899 	 */
5900 	if ((bp->b_flags & B_CLUSTER) != 0)
5901 		return;
5902 
5903 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5904 		if (vfs_badlock_print)
5905 			printf(
5906 			    "VOP_STRATEGY: bp is not locked but should be\n");
5907 		if (vfs_badlock_ddb)
5908 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5909 	}
5910 }
5911 
5912 void
5913 vop_lock_debugpre(void *ap)
5914 {
5915 	struct vop_lock1_args *a = ap;
5916 
5917 	if ((a->a_flags & LK_INTERLOCK) == 0)
5918 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5919 	else
5920 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5921 }
5922 
5923 void
5924 vop_lock_debugpost(void *ap, int rc)
5925 {
5926 	struct vop_lock1_args *a = ap;
5927 
5928 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5929 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5930 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5931 }
5932 
5933 void
5934 vop_unlock_debugpre(void *ap)
5935 {
5936 	struct vop_unlock_args *a = ap;
5937 	struct vnode *vp = a->a_vp;
5938 
5939 	VNPASS(vn_get_state(vp) != VSTATE_UNINITIALIZED, vp);
5940 	ASSERT_VOP_LOCKED(vp, "VOP_UNLOCK");
5941 }
5942 
5943 void
5944 vop_need_inactive_debugpre(void *ap)
5945 {
5946 	struct vop_need_inactive_args *a = ap;
5947 
5948 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5949 }
5950 
5951 void
5952 vop_need_inactive_debugpost(void *ap, int rc)
5953 {
5954 	struct vop_need_inactive_args *a = ap;
5955 
5956 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5957 }
5958 #endif
5959 
5960 void
5961 vop_create_pre(void *ap)
5962 {
5963 	struct vop_create_args *a;
5964 	struct vnode *dvp;
5965 
5966 	a = ap;
5967 	dvp = a->a_dvp;
5968 	vn_seqc_write_begin(dvp);
5969 }
5970 
5971 void
5972 vop_create_post(void *ap, int rc)
5973 {
5974 	struct vop_create_args *a;
5975 	struct vnode *dvp;
5976 
5977 	a = ap;
5978 	dvp = a->a_dvp;
5979 	vn_seqc_write_end(dvp);
5980 	if (!rc)
5981 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5982 }
5983 
5984 void
5985 vop_whiteout_pre(void *ap)
5986 {
5987 	struct vop_whiteout_args *a;
5988 	struct vnode *dvp;
5989 
5990 	a = ap;
5991 	dvp = a->a_dvp;
5992 	vn_seqc_write_begin(dvp);
5993 }
5994 
5995 void
5996 vop_whiteout_post(void *ap, int rc)
5997 {
5998 	struct vop_whiteout_args *a;
5999 	struct vnode *dvp;
6000 
6001 	a = ap;
6002 	dvp = a->a_dvp;
6003 	vn_seqc_write_end(dvp);
6004 }
6005 
6006 void
6007 vop_deleteextattr_pre(void *ap)
6008 {
6009 	struct vop_deleteextattr_args *a;
6010 	struct vnode *vp;
6011 
6012 	a = ap;
6013 	vp = a->a_vp;
6014 	vn_seqc_write_begin(vp);
6015 }
6016 
6017 void
6018 vop_deleteextattr_post(void *ap, int rc)
6019 {
6020 	struct vop_deleteextattr_args *a;
6021 	struct vnode *vp;
6022 
6023 	a = ap;
6024 	vp = a->a_vp;
6025 	vn_seqc_write_end(vp);
6026 	if (!rc)
6027 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
6028 }
6029 
6030 void
6031 vop_link_pre(void *ap)
6032 {
6033 	struct vop_link_args *a;
6034 	struct vnode *vp, *tdvp;
6035 
6036 	a = ap;
6037 	vp = a->a_vp;
6038 	tdvp = a->a_tdvp;
6039 	vn_seqc_write_begin(vp);
6040 	vn_seqc_write_begin(tdvp);
6041 }
6042 
6043 void
6044 vop_link_post(void *ap, int rc)
6045 {
6046 	struct vop_link_args *a;
6047 	struct vnode *vp, *tdvp;
6048 
6049 	a = ap;
6050 	vp = a->a_vp;
6051 	tdvp = a->a_tdvp;
6052 	vn_seqc_write_end(vp);
6053 	vn_seqc_write_end(tdvp);
6054 	if (!rc) {
6055 		VFS_KNOTE_LOCKED(vp, NOTE_LINK);
6056 		VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE);
6057 	}
6058 }
6059 
6060 void
6061 vop_mkdir_pre(void *ap)
6062 {
6063 	struct vop_mkdir_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_mkdir_post(void *ap, int rc)
6073 {
6074 	struct vop_mkdir_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 | NOTE_LINK);
6082 }
6083 
6084 #ifdef DEBUG_VFS_LOCKS
6085 void
6086 vop_mkdir_debugpost(void *ap, int rc)
6087 {
6088 	struct vop_mkdir_args *a;
6089 
6090 	a = ap;
6091 	if (!rc)
6092 		cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp);
6093 }
6094 #endif
6095 
6096 void
6097 vop_mknod_pre(void *ap)
6098 {
6099 	struct vop_mknod_args *a;
6100 	struct vnode *dvp;
6101 
6102 	a = ap;
6103 	dvp = a->a_dvp;
6104 	vn_seqc_write_begin(dvp);
6105 }
6106 
6107 void
6108 vop_mknod_post(void *ap, int rc)
6109 {
6110 	struct vop_mknod_args *a;
6111 	struct vnode *dvp;
6112 
6113 	a = ap;
6114 	dvp = a->a_dvp;
6115 	vn_seqc_write_end(dvp);
6116 	if (!rc)
6117 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6118 }
6119 
6120 void
6121 vop_reclaim_post(void *ap, int rc)
6122 {
6123 	struct vop_reclaim_args *a;
6124 	struct vnode *vp;
6125 
6126 	a = ap;
6127 	vp = a->a_vp;
6128 	ASSERT_VOP_IN_SEQC(vp);
6129 	if (!rc)
6130 		VFS_KNOTE_LOCKED(vp, NOTE_REVOKE);
6131 }
6132 
6133 void
6134 vop_remove_pre(void *ap)
6135 {
6136 	struct vop_remove_args *a;
6137 	struct vnode *dvp, *vp;
6138 
6139 	a = ap;
6140 	dvp = a->a_dvp;
6141 	vp = a->a_vp;
6142 	vn_seqc_write_begin(dvp);
6143 	vn_seqc_write_begin(vp);
6144 }
6145 
6146 void
6147 vop_remove_post(void *ap, int rc)
6148 {
6149 	struct vop_remove_args *a;
6150 	struct vnode *dvp, *vp;
6151 
6152 	a = ap;
6153 	dvp = a->a_dvp;
6154 	vp = a->a_vp;
6155 	vn_seqc_write_end(dvp);
6156 	vn_seqc_write_end(vp);
6157 	if (!rc) {
6158 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6159 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
6160 	}
6161 }
6162 
6163 void
6164 vop_rename_post(void *ap, int rc)
6165 {
6166 	struct vop_rename_args *a = ap;
6167 	long hint;
6168 
6169 	if (!rc) {
6170 		hint = NOTE_WRITE;
6171 		if (a->a_fdvp == a->a_tdvp) {
6172 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
6173 				hint |= NOTE_LINK;
6174 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
6175 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
6176 		} else {
6177 			hint |= NOTE_EXTEND;
6178 			if (a->a_fvp->v_type == VDIR)
6179 				hint |= NOTE_LINK;
6180 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
6181 
6182 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
6183 			    a->a_tvp->v_type == VDIR)
6184 				hint &= ~NOTE_LINK;
6185 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
6186 		}
6187 
6188 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
6189 		if (a->a_tvp)
6190 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
6191 	}
6192 	if (a->a_tdvp != a->a_fdvp)
6193 		vdrop(a->a_fdvp);
6194 	if (a->a_tvp != a->a_fvp)
6195 		vdrop(a->a_fvp);
6196 	vdrop(a->a_tdvp);
6197 	if (a->a_tvp)
6198 		vdrop(a->a_tvp);
6199 }
6200 
6201 void
6202 vop_rmdir_pre(void *ap)
6203 {
6204 	struct vop_rmdir_args *a;
6205 	struct vnode *dvp, *vp;
6206 
6207 	a = ap;
6208 	dvp = a->a_dvp;
6209 	vp = a->a_vp;
6210 	vn_seqc_write_begin(dvp);
6211 	vn_seqc_write_begin(vp);
6212 }
6213 
6214 void
6215 vop_rmdir_post(void *ap, int rc)
6216 {
6217 	struct vop_rmdir_args *a;
6218 	struct vnode *dvp, *vp;
6219 
6220 	a = ap;
6221 	dvp = a->a_dvp;
6222 	vp = a->a_vp;
6223 	vn_seqc_write_end(dvp);
6224 	vn_seqc_write_end(vp);
6225 	if (!rc) {
6226 		vp->v_vflag |= VV_UNLINKED;
6227 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
6228 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
6229 	}
6230 }
6231 
6232 void
6233 vop_setattr_pre(void *ap)
6234 {
6235 	struct vop_setattr_args *a;
6236 	struct vnode *vp;
6237 
6238 	a = ap;
6239 	vp = a->a_vp;
6240 	vn_seqc_write_begin(vp);
6241 }
6242 
6243 void
6244 vop_setattr_post(void *ap, int rc)
6245 {
6246 	struct vop_setattr_args *a;
6247 	struct vnode *vp;
6248 
6249 	a = ap;
6250 	vp = a->a_vp;
6251 	vn_seqc_write_end(vp);
6252 	if (!rc)
6253 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
6254 }
6255 
6256 void
6257 vop_setacl_pre(void *ap)
6258 {
6259 	struct vop_setacl_args *a;
6260 	struct vnode *vp;
6261 
6262 	a = ap;
6263 	vp = a->a_vp;
6264 	vn_seqc_write_begin(vp);
6265 }
6266 
6267 void
6268 vop_setacl_post(void *ap, int rc __unused)
6269 {
6270 	struct vop_setacl_args *a;
6271 	struct vnode *vp;
6272 
6273 	a = ap;
6274 	vp = a->a_vp;
6275 	vn_seqc_write_end(vp);
6276 }
6277 
6278 void
6279 vop_setextattr_pre(void *ap)
6280 {
6281 	struct vop_setextattr_args *a;
6282 	struct vnode *vp;
6283 
6284 	a = ap;
6285 	vp = a->a_vp;
6286 	vn_seqc_write_begin(vp);
6287 }
6288 
6289 void
6290 vop_setextattr_post(void *ap, int rc)
6291 {
6292 	struct vop_setextattr_args *a;
6293 	struct vnode *vp;
6294 
6295 	a = ap;
6296 	vp = a->a_vp;
6297 	vn_seqc_write_end(vp);
6298 	if (!rc)
6299 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
6300 }
6301 
6302 void
6303 vop_symlink_pre(void *ap)
6304 {
6305 	struct vop_symlink_args *a;
6306 	struct vnode *dvp;
6307 
6308 	a = ap;
6309 	dvp = a->a_dvp;
6310 	vn_seqc_write_begin(dvp);
6311 }
6312 
6313 void
6314 vop_symlink_post(void *ap, int rc)
6315 {
6316 	struct vop_symlink_args *a;
6317 	struct vnode *dvp;
6318 
6319 	a = ap;
6320 	dvp = a->a_dvp;
6321 	vn_seqc_write_end(dvp);
6322 	if (!rc)
6323 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6324 }
6325 
6326 void
6327 vop_open_post(void *ap, int rc)
6328 {
6329 	struct vop_open_args *a = ap;
6330 
6331 	if (!rc)
6332 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
6333 }
6334 
6335 void
6336 vop_close_post(void *ap, int rc)
6337 {
6338 	struct vop_close_args *a = ap;
6339 
6340 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
6341 	    !VN_IS_DOOMED(a->a_vp))) {
6342 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
6343 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
6344 	}
6345 }
6346 
6347 void
6348 vop_read_post(void *ap, int rc)
6349 {
6350 	struct vop_read_args *a = ap;
6351 
6352 	if (!rc)
6353 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6354 }
6355 
6356 void
6357 vop_read_pgcache_post(void *ap, int rc)
6358 {
6359 	struct vop_read_pgcache_args *a = ap;
6360 
6361 	if (!rc)
6362 		VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ);
6363 }
6364 
6365 void
6366 vop_readdir_post(void *ap, int rc)
6367 {
6368 	struct vop_readdir_args *a = ap;
6369 
6370 	if (!rc)
6371 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6372 }
6373 
6374 static struct knlist fs_knlist;
6375 
6376 static void
6377 vfs_event_init(void *arg)
6378 {
6379 	knlist_init_mtx(&fs_knlist, NULL);
6380 }
6381 /* XXX - correct order? */
6382 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
6383 
6384 void
6385 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
6386 {
6387 
6388 	KNOTE_UNLOCKED(&fs_knlist, event);
6389 }
6390 
6391 static int	filt_fsattach(struct knote *kn);
6392 static void	filt_fsdetach(struct knote *kn);
6393 static int	filt_fsevent(struct knote *kn, long hint);
6394 
6395 struct filterops fs_filtops = {
6396 	.f_isfd = 0,
6397 	.f_attach = filt_fsattach,
6398 	.f_detach = filt_fsdetach,
6399 	.f_event = filt_fsevent
6400 };
6401 
6402 static int
6403 filt_fsattach(struct knote *kn)
6404 {
6405 
6406 	kn->kn_flags |= EV_CLEAR;
6407 	knlist_add(&fs_knlist, kn, 0);
6408 	return (0);
6409 }
6410 
6411 static void
6412 filt_fsdetach(struct knote *kn)
6413 {
6414 
6415 	knlist_remove(&fs_knlist, kn, 0);
6416 }
6417 
6418 static int
6419 filt_fsevent(struct knote *kn, long hint)
6420 {
6421 
6422 	kn->kn_fflags |= kn->kn_sfflags & hint;
6423 
6424 	return (kn->kn_fflags != 0);
6425 }
6426 
6427 static int
6428 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
6429 {
6430 	struct vfsidctl vc;
6431 	int error;
6432 	struct mount *mp;
6433 
6434 	error = SYSCTL_IN(req, &vc, sizeof(vc));
6435 	if (error)
6436 		return (error);
6437 	if (vc.vc_vers != VFS_CTL_VERS1)
6438 		return (EINVAL);
6439 	mp = vfs_getvfs(&vc.vc_fsid);
6440 	if (mp == NULL)
6441 		return (ENOENT);
6442 	/* ensure that a specific sysctl goes to the right filesystem. */
6443 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
6444 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
6445 		vfs_rel(mp);
6446 		return (EINVAL);
6447 	}
6448 	VCTLTOREQ(&vc, req);
6449 	error = VFS_SYSCTL(mp, vc.vc_op, req);
6450 	vfs_rel(mp);
6451 	return (error);
6452 }
6453 
6454 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
6455     NULL, 0, sysctl_vfs_ctl, "",
6456     "Sysctl by fsid");
6457 
6458 /*
6459  * Function to initialize a va_filerev field sensibly.
6460  * XXX: Wouldn't a random number make a lot more sense ??
6461  */
6462 u_quad_t
6463 init_va_filerev(void)
6464 {
6465 	struct bintime bt;
6466 
6467 	getbinuptime(&bt);
6468 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
6469 }
6470 
6471 static int	filt_vfsread(struct knote *kn, long hint);
6472 static int	filt_vfswrite(struct knote *kn, long hint);
6473 static int	filt_vfsvnode(struct knote *kn, long hint);
6474 static void	filt_vfsdetach(struct knote *kn);
6475 static struct filterops vfsread_filtops = {
6476 	.f_isfd = 1,
6477 	.f_detach = filt_vfsdetach,
6478 	.f_event = filt_vfsread
6479 };
6480 static struct filterops vfswrite_filtops = {
6481 	.f_isfd = 1,
6482 	.f_detach = filt_vfsdetach,
6483 	.f_event = filt_vfswrite
6484 };
6485 static struct filterops vfsvnode_filtops = {
6486 	.f_isfd = 1,
6487 	.f_detach = filt_vfsdetach,
6488 	.f_event = filt_vfsvnode
6489 };
6490 
6491 static void
6492 vfs_knllock(void *arg)
6493 {
6494 	struct vnode *vp = arg;
6495 
6496 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
6497 }
6498 
6499 static void
6500 vfs_knlunlock(void *arg)
6501 {
6502 	struct vnode *vp = arg;
6503 
6504 	VOP_UNLOCK(vp);
6505 }
6506 
6507 static void
6508 vfs_knl_assert_lock(void *arg, int what)
6509 {
6510 #ifdef DEBUG_VFS_LOCKS
6511 	struct vnode *vp = arg;
6512 
6513 	if (what == LA_LOCKED)
6514 		ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
6515 	else
6516 		ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
6517 #endif
6518 }
6519 
6520 int
6521 vfs_kqfilter(struct vop_kqfilter_args *ap)
6522 {
6523 	struct vnode *vp = ap->a_vp;
6524 	struct knote *kn = ap->a_kn;
6525 	struct knlist *knl;
6526 
6527 	KASSERT(vp->v_type != VFIFO || (kn->kn_filter != EVFILT_READ &&
6528 	    kn->kn_filter != EVFILT_WRITE),
6529 	    ("READ/WRITE filter on a FIFO leaked through"));
6530 	switch (kn->kn_filter) {
6531 	case EVFILT_READ:
6532 		kn->kn_fop = &vfsread_filtops;
6533 		break;
6534 	case EVFILT_WRITE:
6535 		kn->kn_fop = &vfswrite_filtops;
6536 		break;
6537 	case EVFILT_VNODE:
6538 		kn->kn_fop = &vfsvnode_filtops;
6539 		break;
6540 	default:
6541 		return (EINVAL);
6542 	}
6543 
6544 	kn->kn_hook = (caddr_t)vp;
6545 
6546 	v_addpollinfo(vp);
6547 	if (vp->v_pollinfo == NULL)
6548 		return (ENOMEM);
6549 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
6550 	vhold(vp);
6551 	knlist_add(knl, kn, 0);
6552 
6553 	return (0);
6554 }
6555 
6556 /*
6557  * Detach knote from vnode
6558  */
6559 static void
6560 filt_vfsdetach(struct knote *kn)
6561 {
6562 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6563 
6564 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
6565 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
6566 	vdrop(vp);
6567 }
6568 
6569 /*ARGSUSED*/
6570 static int
6571 filt_vfsread(struct knote *kn, long hint)
6572 {
6573 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6574 	off_t size;
6575 	int res;
6576 
6577 	/*
6578 	 * filesystem is gone, so set the EOF flag and schedule
6579 	 * the knote for deletion.
6580 	 */
6581 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6582 		VI_LOCK(vp);
6583 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6584 		VI_UNLOCK(vp);
6585 		return (1);
6586 	}
6587 
6588 	if (vn_getsize_locked(vp, &size, curthread->td_ucred) != 0)
6589 		return (0);
6590 
6591 	VI_LOCK(vp);
6592 	kn->kn_data = size - kn->kn_fp->f_offset;
6593 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
6594 	VI_UNLOCK(vp);
6595 	return (res);
6596 }
6597 
6598 /*ARGSUSED*/
6599 static int
6600 filt_vfswrite(struct knote *kn, long hint)
6601 {
6602 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6603 
6604 	VI_LOCK(vp);
6605 
6606 	/*
6607 	 * filesystem is gone, so set the EOF flag and schedule
6608 	 * the knote for deletion.
6609 	 */
6610 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
6611 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6612 
6613 	kn->kn_data = 0;
6614 	VI_UNLOCK(vp);
6615 	return (1);
6616 }
6617 
6618 static int
6619 filt_vfsvnode(struct knote *kn, long hint)
6620 {
6621 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6622 	int res;
6623 
6624 	VI_LOCK(vp);
6625 	if (kn->kn_sfflags & hint)
6626 		kn->kn_fflags |= hint;
6627 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6628 		kn->kn_flags |= EV_EOF;
6629 		VI_UNLOCK(vp);
6630 		return (1);
6631 	}
6632 	res = (kn->kn_fflags != 0);
6633 	VI_UNLOCK(vp);
6634 	return (res);
6635 }
6636 
6637 int
6638 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6639 {
6640 	int error;
6641 
6642 	if (dp->d_reclen > ap->a_uio->uio_resid)
6643 		return (ENAMETOOLONG);
6644 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
6645 	if (error) {
6646 		if (ap->a_ncookies != NULL) {
6647 			if (ap->a_cookies != NULL)
6648 				free(ap->a_cookies, M_TEMP);
6649 			ap->a_cookies = NULL;
6650 			*ap->a_ncookies = 0;
6651 		}
6652 		return (error);
6653 	}
6654 	if (ap->a_ncookies == NULL)
6655 		return (0);
6656 
6657 	KASSERT(ap->a_cookies,
6658 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6659 
6660 	*ap->a_cookies = realloc(*ap->a_cookies,
6661 	    (*ap->a_ncookies + 1) * sizeof(uint64_t), M_TEMP, M_WAITOK | M_ZERO);
6662 	(*ap->a_cookies)[*ap->a_ncookies] = off;
6663 	*ap->a_ncookies += 1;
6664 	return (0);
6665 }
6666 
6667 /*
6668  * The purpose of this routine is to remove granularity from accmode_t,
6669  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6670  * VADMIN and VAPPEND.
6671  *
6672  * If it returns 0, the caller is supposed to continue with the usual
6673  * access checks using 'accmode' as modified by this routine.  If it
6674  * returns nonzero value, the caller is supposed to return that value
6675  * as errno.
6676  *
6677  * Note that after this routine runs, accmode may be zero.
6678  */
6679 int
6680 vfs_unixify_accmode(accmode_t *accmode)
6681 {
6682 	/*
6683 	 * There is no way to specify explicit "deny" rule using
6684 	 * file mode or POSIX.1e ACLs.
6685 	 */
6686 	if (*accmode & VEXPLICIT_DENY) {
6687 		*accmode = 0;
6688 		return (0);
6689 	}
6690 
6691 	/*
6692 	 * None of these can be translated into usual access bits.
6693 	 * Also, the common case for NFSv4 ACLs is to not contain
6694 	 * either of these bits. Caller should check for VWRITE
6695 	 * on the containing directory instead.
6696 	 */
6697 	if (*accmode & (VDELETE_CHILD | VDELETE))
6698 		return (EPERM);
6699 
6700 	if (*accmode & VADMIN_PERMS) {
6701 		*accmode &= ~VADMIN_PERMS;
6702 		*accmode |= VADMIN;
6703 	}
6704 
6705 	/*
6706 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6707 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6708 	 */
6709 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6710 
6711 	return (0);
6712 }
6713 
6714 /*
6715  * Clear out a doomed vnode (if any) and replace it with a new one as long
6716  * as the fs is not being unmounted. Return the root vnode to the caller.
6717  */
6718 static int __noinline
6719 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6720 {
6721 	struct vnode *vp;
6722 	int error;
6723 
6724 restart:
6725 	if (mp->mnt_rootvnode != NULL) {
6726 		MNT_ILOCK(mp);
6727 		vp = mp->mnt_rootvnode;
6728 		if (vp != NULL) {
6729 			if (!VN_IS_DOOMED(vp)) {
6730 				vrefact(vp);
6731 				MNT_IUNLOCK(mp);
6732 				error = vn_lock(vp, flags);
6733 				if (error == 0) {
6734 					*vpp = vp;
6735 					return (0);
6736 				}
6737 				vrele(vp);
6738 				goto restart;
6739 			}
6740 			/*
6741 			 * Clear the old one.
6742 			 */
6743 			mp->mnt_rootvnode = NULL;
6744 		}
6745 		MNT_IUNLOCK(mp);
6746 		if (vp != NULL) {
6747 			vfs_op_barrier_wait(mp);
6748 			vrele(vp);
6749 		}
6750 	}
6751 	error = VFS_CACHEDROOT(mp, flags, vpp);
6752 	if (error != 0)
6753 		return (error);
6754 	if (mp->mnt_vfs_ops == 0) {
6755 		MNT_ILOCK(mp);
6756 		if (mp->mnt_vfs_ops != 0) {
6757 			MNT_IUNLOCK(mp);
6758 			return (0);
6759 		}
6760 		if (mp->mnt_rootvnode == NULL) {
6761 			vrefact(*vpp);
6762 			mp->mnt_rootvnode = *vpp;
6763 		} else {
6764 			if (mp->mnt_rootvnode != *vpp) {
6765 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6766 					panic("%s: mismatch between vnode returned "
6767 					    " by VFS_CACHEDROOT and the one cached "
6768 					    " (%p != %p)",
6769 					    __func__, *vpp, mp->mnt_rootvnode);
6770 				}
6771 			}
6772 		}
6773 		MNT_IUNLOCK(mp);
6774 	}
6775 	return (0);
6776 }
6777 
6778 int
6779 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6780 {
6781 	struct mount_pcpu *mpcpu;
6782 	struct vnode *vp;
6783 	int error;
6784 
6785 	if (!vfs_op_thread_enter(mp, mpcpu))
6786 		return (vfs_cache_root_fallback(mp, flags, vpp));
6787 	vp = atomic_load_ptr(&mp->mnt_rootvnode);
6788 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6789 		vfs_op_thread_exit(mp, mpcpu);
6790 		return (vfs_cache_root_fallback(mp, flags, vpp));
6791 	}
6792 	vrefact(vp);
6793 	vfs_op_thread_exit(mp, mpcpu);
6794 	error = vn_lock(vp, flags);
6795 	if (error != 0) {
6796 		vrele(vp);
6797 		return (vfs_cache_root_fallback(mp, flags, vpp));
6798 	}
6799 	*vpp = vp;
6800 	return (0);
6801 }
6802 
6803 struct vnode *
6804 vfs_cache_root_clear(struct mount *mp)
6805 {
6806 	struct vnode *vp;
6807 
6808 	/*
6809 	 * ops > 0 guarantees there is nobody who can see this vnode
6810 	 */
6811 	MPASS(mp->mnt_vfs_ops > 0);
6812 	vp = mp->mnt_rootvnode;
6813 	if (vp != NULL)
6814 		vn_seqc_write_begin(vp);
6815 	mp->mnt_rootvnode = NULL;
6816 	return (vp);
6817 }
6818 
6819 void
6820 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6821 {
6822 
6823 	MPASS(mp->mnt_vfs_ops > 0);
6824 	vrefact(vp);
6825 	mp->mnt_rootvnode = vp;
6826 }
6827 
6828 /*
6829  * These are helper functions for filesystems to traverse all
6830  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6831  *
6832  * This interface replaces MNT_VNODE_FOREACH.
6833  */
6834 
6835 struct vnode *
6836 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6837 {
6838 	struct vnode *vp;
6839 
6840 	maybe_yield();
6841 	MNT_ILOCK(mp);
6842 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6843 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6844 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6845 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6846 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6847 			continue;
6848 		VI_LOCK(vp);
6849 		if (VN_IS_DOOMED(vp)) {
6850 			VI_UNLOCK(vp);
6851 			continue;
6852 		}
6853 		break;
6854 	}
6855 	if (vp == NULL) {
6856 		__mnt_vnode_markerfree_all(mvp, mp);
6857 		/* MNT_IUNLOCK(mp); -- done in above function */
6858 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6859 		return (NULL);
6860 	}
6861 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6862 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6863 	MNT_IUNLOCK(mp);
6864 	return (vp);
6865 }
6866 
6867 struct vnode *
6868 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6869 {
6870 	struct vnode *vp;
6871 
6872 	*mvp = vn_alloc_marker(mp);
6873 	MNT_ILOCK(mp);
6874 	MNT_REF(mp);
6875 
6876 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6877 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6878 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6879 			continue;
6880 		VI_LOCK(vp);
6881 		if (VN_IS_DOOMED(vp)) {
6882 			VI_UNLOCK(vp);
6883 			continue;
6884 		}
6885 		break;
6886 	}
6887 	if (vp == NULL) {
6888 		MNT_REL(mp);
6889 		MNT_IUNLOCK(mp);
6890 		vn_free_marker(*mvp);
6891 		*mvp = NULL;
6892 		return (NULL);
6893 	}
6894 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6895 	MNT_IUNLOCK(mp);
6896 	return (vp);
6897 }
6898 
6899 void
6900 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6901 {
6902 
6903 	if (*mvp == NULL) {
6904 		MNT_IUNLOCK(mp);
6905 		return;
6906 	}
6907 
6908 	mtx_assert(MNT_MTX(mp), MA_OWNED);
6909 
6910 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6911 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6912 	MNT_REL(mp);
6913 	MNT_IUNLOCK(mp);
6914 	vn_free_marker(*mvp);
6915 	*mvp = NULL;
6916 }
6917 
6918 /*
6919  * These are helper functions for filesystems to traverse their
6920  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6921  */
6922 static void
6923 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6924 {
6925 
6926 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6927 
6928 	MNT_ILOCK(mp);
6929 	MNT_REL(mp);
6930 	MNT_IUNLOCK(mp);
6931 	vn_free_marker(*mvp);
6932 	*mvp = NULL;
6933 }
6934 
6935 /*
6936  * Relock the mp mount vnode list lock with the vp vnode interlock in the
6937  * conventional lock order during mnt_vnode_next_lazy iteration.
6938  *
6939  * On entry, the mount vnode list lock is held and the vnode interlock is not.
6940  * The list lock is dropped and reacquired.  On success, both locks are held.
6941  * On failure, the mount vnode list lock is held but the vnode interlock is
6942  * not, and the procedure may have yielded.
6943  */
6944 static bool
6945 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6946     struct vnode *vp)
6947 {
6948 
6949 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6950 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6951 	    ("%s: bad marker", __func__));
6952 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6953 	    ("%s: inappropriate vnode", __func__));
6954 	ASSERT_VI_UNLOCKED(vp, __func__);
6955 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6956 
6957 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6958 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6959 
6960 	/*
6961 	 * Note we may be racing against vdrop which transitioned the hold
6962 	 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6963 	 * if we are the only user after we get the interlock we will just
6964 	 * vdrop.
6965 	 */
6966 	vhold(vp);
6967 	mtx_unlock(&mp->mnt_listmtx);
6968 	VI_LOCK(vp);
6969 	if (VN_IS_DOOMED(vp)) {
6970 		VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6971 		goto out_lost;
6972 	}
6973 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6974 	/*
6975 	 * There is nothing to do if we are the last user.
6976 	 */
6977 	if (!refcount_release_if_not_last(&vp->v_holdcnt))
6978 		goto out_lost;
6979 	mtx_lock(&mp->mnt_listmtx);
6980 	return (true);
6981 out_lost:
6982 	vdropl(vp);
6983 	maybe_yield();
6984 	mtx_lock(&mp->mnt_listmtx);
6985 	return (false);
6986 }
6987 
6988 static struct vnode *
6989 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6990     void *cbarg)
6991 {
6992 	struct vnode *vp;
6993 
6994 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6995 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6996 restart:
6997 	vp = TAILQ_NEXT(*mvp, v_lazylist);
6998 	while (vp != NULL) {
6999 		if (vp->v_type == VMARKER) {
7000 			vp = TAILQ_NEXT(vp, v_lazylist);
7001 			continue;
7002 		}
7003 		/*
7004 		 * See if we want to process the vnode. Note we may encounter a
7005 		 * long string of vnodes we don't care about and hog the list
7006 		 * as a result. Check for it and requeue the marker.
7007 		 */
7008 		VNPASS(!VN_IS_DOOMED(vp), vp);
7009 		if (!cb(vp, cbarg)) {
7010 			if (!should_yield()) {
7011 				vp = TAILQ_NEXT(vp, v_lazylist);
7012 				continue;
7013 			}
7014 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
7015 			    v_lazylist);
7016 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
7017 			    v_lazylist);
7018 			mtx_unlock(&mp->mnt_listmtx);
7019 			kern_yield(PRI_USER);
7020 			mtx_lock(&mp->mnt_listmtx);
7021 			goto restart;
7022 		}
7023 		/*
7024 		 * Try-lock because this is the wrong lock order.
7025 		 */
7026 		if (!VI_TRYLOCK(vp) &&
7027 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
7028 			goto restart;
7029 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
7030 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
7031 		    ("alien vnode on the lazy list %p %p", vp, mp));
7032 		VNPASS(vp->v_mount == mp, vp);
7033 		VNPASS(!VN_IS_DOOMED(vp), vp);
7034 		break;
7035 	}
7036 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
7037 
7038 	/* Check if we are done */
7039 	if (vp == NULL) {
7040 		mtx_unlock(&mp->mnt_listmtx);
7041 		mnt_vnode_markerfree_lazy(mvp, mp);
7042 		return (NULL);
7043 	}
7044 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
7045 	mtx_unlock(&mp->mnt_listmtx);
7046 	ASSERT_VI_LOCKED(vp, "lazy iter");
7047 	return (vp);
7048 }
7049 
7050 struct vnode *
7051 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
7052     void *cbarg)
7053 {
7054 
7055 	maybe_yield();
7056 	mtx_lock(&mp->mnt_listmtx);
7057 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
7058 }
7059 
7060 struct vnode *
7061 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
7062     void *cbarg)
7063 {
7064 	struct vnode *vp;
7065 
7066 	if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
7067 		return (NULL);
7068 
7069 	*mvp = vn_alloc_marker(mp);
7070 	MNT_ILOCK(mp);
7071 	MNT_REF(mp);
7072 	MNT_IUNLOCK(mp);
7073 
7074 	mtx_lock(&mp->mnt_listmtx);
7075 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
7076 	if (vp == NULL) {
7077 		mtx_unlock(&mp->mnt_listmtx);
7078 		mnt_vnode_markerfree_lazy(mvp, mp);
7079 		return (NULL);
7080 	}
7081 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
7082 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
7083 }
7084 
7085 void
7086 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
7087 {
7088 
7089 	if (*mvp == NULL)
7090 		return;
7091 
7092 	mtx_lock(&mp->mnt_listmtx);
7093 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
7094 	mtx_unlock(&mp->mnt_listmtx);
7095 	mnt_vnode_markerfree_lazy(mvp, mp);
7096 }
7097 
7098 int
7099 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
7100 {
7101 
7102 	if ((cnp->cn_flags & NOEXECCHECK) != 0) {
7103 		cnp->cn_flags &= ~NOEXECCHECK;
7104 		return (0);
7105 	}
7106 
7107 	return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, curthread));
7108 }
7109 
7110 /*
7111  * Do not use this variant unless you have means other than the hold count
7112  * to prevent the vnode from getting freed.
7113  */
7114 void
7115 vn_seqc_write_begin_locked(struct vnode *vp)
7116 {
7117 
7118 	ASSERT_VI_LOCKED(vp, __func__);
7119 	VNPASS(vp->v_holdcnt > 0, vp);
7120 	VNPASS(vp->v_seqc_users >= 0, vp);
7121 	vp->v_seqc_users++;
7122 	if (vp->v_seqc_users == 1)
7123 		seqc_sleepable_write_begin(&vp->v_seqc);
7124 }
7125 
7126 void
7127 vn_seqc_write_begin(struct vnode *vp)
7128 {
7129 
7130 	VI_LOCK(vp);
7131 	vn_seqc_write_begin_locked(vp);
7132 	VI_UNLOCK(vp);
7133 }
7134 
7135 void
7136 vn_seqc_write_end_locked(struct vnode *vp)
7137 {
7138 
7139 	ASSERT_VI_LOCKED(vp, __func__);
7140 	VNPASS(vp->v_seqc_users > 0, vp);
7141 	vp->v_seqc_users--;
7142 	if (vp->v_seqc_users == 0)
7143 		seqc_sleepable_write_end(&vp->v_seqc);
7144 }
7145 
7146 void
7147 vn_seqc_write_end(struct vnode *vp)
7148 {
7149 
7150 	VI_LOCK(vp);
7151 	vn_seqc_write_end_locked(vp);
7152 	VI_UNLOCK(vp);
7153 }
7154 
7155 /*
7156  * Special case handling for allocating and freeing vnodes.
7157  *
7158  * The counter remains unchanged on free so that a doomed vnode will
7159  * keep testing as in modify as long as it is accessible with SMR.
7160  */
7161 static void
7162 vn_seqc_init(struct vnode *vp)
7163 {
7164 
7165 	vp->v_seqc = 0;
7166 	vp->v_seqc_users = 0;
7167 }
7168 
7169 static void
7170 vn_seqc_write_end_free(struct vnode *vp)
7171 {
7172 
7173 	VNPASS(seqc_in_modify(vp->v_seqc), vp);
7174 	VNPASS(vp->v_seqc_users == 1, vp);
7175 }
7176 
7177 void
7178 vn_irflag_set_locked(struct vnode *vp, short toset)
7179 {
7180 	short flags;
7181 
7182 	ASSERT_VI_LOCKED(vp, __func__);
7183 	flags = vn_irflag_read(vp);
7184 	VNASSERT((flags & toset) == 0, vp,
7185 	    ("%s: some of the passed flags already set (have %d, passed %d)\n",
7186 	    __func__, flags, toset));
7187 	atomic_store_short(&vp->v_irflag, flags | toset);
7188 }
7189 
7190 void
7191 vn_irflag_set(struct vnode *vp, short toset)
7192 {
7193 
7194 	VI_LOCK(vp);
7195 	vn_irflag_set_locked(vp, toset);
7196 	VI_UNLOCK(vp);
7197 }
7198 
7199 void
7200 vn_irflag_set_cond_locked(struct vnode *vp, short toset)
7201 {
7202 	short flags;
7203 
7204 	ASSERT_VI_LOCKED(vp, __func__);
7205 	flags = vn_irflag_read(vp);
7206 	atomic_store_short(&vp->v_irflag, flags | toset);
7207 }
7208 
7209 void
7210 vn_irflag_set_cond(struct vnode *vp, short toset)
7211 {
7212 
7213 	VI_LOCK(vp);
7214 	vn_irflag_set_cond_locked(vp, toset);
7215 	VI_UNLOCK(vp);
7216 }
7217 
7218 void
7219 vn_irflag_unset_locked(struct vnode *vp, short tounset)
7220 {
7221 	short flags;
7222 
7223 	ASSERT_VI_LOCKED(vp, __func__);
7224 	flags = vn_irflag_read(vp);
7225 	VNASSERT((flags & tounset) == tounset, vp,
7226 	    ("%s: some of the passed flags not set (have %d, passed %d)\n",
7227 	    __func__, flags, tounset));
7228 	atomic_store_short(&vp->v_irflag, flags & ~tounset);
7229 }
7230 
7231 void
7232 vn_irflag_unset(struct vnode *vp, short tounset)
7233 {
7234 
7235 	VI_LOCK(vp);
7236 	vn_irflag_unset_locked(vp, tounset);
7237 	VI_UNLOCK(vp);
7238 }
7239 
7240 int
7241 vn_getsize_locked(struct vnode *vp, off_t *size, struct ucred *cred)
7242 {
7243 	struct vattr vattr;
7244 	int error;
7245 
7246 	ASSERT_VOP_LOCKED(vp, __func__);
7247 	error = VOP_GETATTR(vp, &vattr, cred);
7248 	if (__predict_true(error == 0)) {
7249 		if (vattr.va_size <= OFF_MAX)
7250 			*size = vattr.va_size;
7251 		else
7252 			error = EFBIG;
7253 	}
7254 	return (error);
7255 }
7256 
7257 int
7258 vn_getsize(struct vnode *vp, off_t *size, struct ucred *cred)
7259 {
7260 	int error;
7261 
7262 	VOP_LOCK(vp, LK_SHARED);
7263 	error = vn_getsize_locked(vp, size, cred);
7264 	VOP_UNLOCK(vp);
7265 	return (error);
7266 }
7267 
7268 #ifdef INVARIANTS
7269 void
7270 vn_set_state_validate(struct vnode *vp, __enum_uint8(vstate) state)
7271 {
7272 
7273 	switch (vp->v_state) {
7274 	case VSTATE_UNINITIALIZED:
7275 		switch (state) {
7276 		case VSTATE_CONSTRUCTED:
7277 		case VSTATE_DESTROYING:
7278 			return;
7279 		default:
7280 			break;
7281 		}
7282 		break;
7283 	case VSTATE_CONSTRUCTED:
7284 		ASSERT_VOP_ELOCKED(vp, __func__);
7285 		switch (state) {
7286 		case VSTATE_DESTROYING:
7287 			return;
7288 		default:
7289 			break;
7290 		}
7291 		break;
7292 	case VSTATE_DESTROYING:
7293 		ASSERT_VOP_ELOCKED(vp, __func__);
7294 		switch (state) {
7295 		case VSTATE_DEAD:
7296 			return;
7297 		default:
7298 			break;
7299 		}
7300 		break;
7301 	case VSTATE_DEAD:
7302 		switch (state) {
7303 		case VSTATE_UNINITIALIZED:
7304 			return;
7305 		default:
7306 			break;
7307 		}
7308 		break;
7309 	}
7310 
7311 	vn_printf(vp, "invalid state transition %d -> %d\n", vp->v_state, state);
7312 	panic("invalid state transition %d -> %d\n", vp->v_state, state);
7313 }
7314 #endif
7315