xref: /freebsd/sys/kern/vfs_subr.c (revision ee7077f24f5b02bde8cf5c202848128f18733398)
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);
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 /*
130  * Number of vnodes in existence.  Increased whenever getnewvnode()
131  * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
132  */
133 static u_long __exclusive_cache_line numvnodes;
134 
135 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
136     "Number of vnodes in existence");
137 
138 static counter_u64_t vnodes_created;
139 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
140     "Number of vnodes created by getnewvnode");
141 
142 /*
143  * Conversion tables for conversion from vnode types to inode formats
144  * and back.
145  */
146 __enum_uint8(vtype) iftovt_tab[16] = {
147 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
148 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
149 };
150 int vttoif_tab[10] = {
151 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
152 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
153 };
154 
155 /*
156  * List of allocates vnodes in the system.
157  */
158 static TAILQ_HEAD(freelst, vnode) vnode_list;
159 static struct vnode *vnode_list_free_marker;
160 static struct vnode *vnode_list_reclaim_marker;
161 
162 /*
163  * "Free" vnode target.  Free vnodes are rarely completely free, but are
164  * just ones that are cheap to recycle.  Usually they are for files which
165  * have been stat'd but not read; these usually have inode and namecache
166  * data attached to them.  This target is the preferred minimum size of a
167  * sub-cache consisting mostly of such files. The system balances the size
168  * of this sub-cache with its complement to try to prevent either from
169  * thrashing while the other is relatively inactive.  The targets express
170  * a preference for the best balance.
171  *
172  * "Above" this target there are 2 further targets (watermarks) related
173  * to recyling of free vnodes.  In the best-operating case, the cache is
174  * exactly full, the free list has size between vlowat and vhiwat above the
175  * free target, and recycling from it and normal use maintains this state.
176  * Sometimes the free list is below vlowat or even empty, but this state
177  * is even better for immediate use provided the cache is not full.
178  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
179  * ones) to reach one of these states.  The watermarks are currently hard-
180  * coded as 4% and 9% of the available space higher.  These and the default
181  * of 25% for wantfreevnodes are too large if the memory size is large.
182  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
183  * whenever vnlru_proc() becomes active.
184  */
185 static long wantfreevnodes;
186 static long __exclusive_cache_line freevnodes;
187 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
188     &freevnodes, 0, "Number of \"free\" vnodes");
189 static long freevnodes_old;
190 
191 static counter_u64_t recycles_count;
192 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
193     "Number of vnodes recycled to meet vnode cache targets");
194 
195 static counter_u64_t recycles_free_count;
196 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count,
197     "Number of free vnodes recycled to meet vnode cache targets");
198 
199 static counter_u64_t vnode_skipped_requeues;
200 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnode_skipped_requeues, CTLFLAG_RD, &vnode_skipped_requeues,
201     "Number of times LRU requeue was skipped due to lock contention");
202 
203 static u_long deferred_inact;
204 SYSCTL_ULONG(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD,
205     &deferred_inact, 0, "Number of times inactive processing was deferred");
206 
207 /* To keep more than one thread at a time from running vfs_getnewfsid */
208 static struct mtx mntid_mtx;
209 
210 /*
211  * Lock for any access to the following:
212  *	vnode_list
213  *	numvnodes
214  *	freevnodes
215  */
216 static struct mtx __exclusive_cache_line vnode_list_mtx;
217 
218 /* Publicly exported FS */
219 struct nfs_public nfs_pub;
220 
221 static uma_zone_t buf_trie_zone;
222 static smr_t buf_trie_smr;
223 
224 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
225 static uma_zone_t vnode_zone;
226 MALLOC_DEFINE(M_VNODEPOLL, "VN POLL", "vnode poll");
227 
228 __read_frequently smr_t vfs_smr;
229 
230 /*
231  * The workitem queue.
232  *
233  * It is useful to delay writes of file data and filesystem metadata
234  * for tens of seconds so that quickly created and deleted files need
235  * not waste disk bandwidth being created and removed. To realize this,
236  * we append vnodes to a "workitem" queue. When running with a soft
237  * updates implementation, most pending metadata dependencies should
238  * not wait for more than a few seconds. Thus, mounted on block devices
239  * are delayed only about a half the time that file data is delayed.
240  * Similarly, directory updates are more critical, so are only delayed
241  * about a third the time that file data is delayed. Thus, there are
242  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
243  * one each second (driven off the filesystem syncer process). The
244  * syncer_delayno variable indicates the next queue that is to be processed.
245  * Items that need to be processed soon are placed in this queue:
246  *
247  *	syncer_workitem_pending[syncer_delayno]
248  *
249  * A delay of fifteen seconds is done by placing the request fifteen
250  * entries later in the queue:
251  *
252  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
253  *
254  */
255 static int syncer_delayno;
256 static long syncer_mask;
257 LIST_HEAD(synclist, bufobj);
258 static struct synclist *syncer_workitem_pending;
259 /*
260  * The sync_mtx protects:
261  *	bo->bo_synclist
262  *	sync_vnode_count
263  *	syncer_delayno
264  *	syncer_state
265  *	syncer_workitem_pending
266  *	syncer_worklist_len
267  *	rushjob
268  */
269 static struct mtx sync_mtx;
270 static struct cv sync_wakeup;
271 
272 #define SYNCER_MAXDELAY		32
273 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
274 static int syncdelay = 30;		/* max time to delay syncing data */
275 static int filedelay = 30;		/* time to delay syncing files */
276 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
277     "Time to delay syncing files (in seconds)");
278 static int dirdelay = 29;		/* time to delay syncing directories */
279 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
280     "Time to delay syncing directories (in seconds)");
281 static int metadelay = 28;		/* time to delay syncing metadata */
282 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
283     "Time to delay syncing metadata (in seconds)");
284 static int rushjob;		/* number of slots to run ASAP */
285 static int stat_rush_requests;	/* number of times I/O speeded up */
286 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
287     "Number of times I/O speeded up (rush requests)");
288 
289 #define	VDBATCH_SIZE 8
290 struct vdbatch {
291 	u_int index;
292 	struct mtx lock;
293 	struct vnode *tab[VDBATCH_SIZE];
294 };
295 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
296 
297 static void	vdbatch_dequeue(struct vnode *vp);
298 
299 /*
300  * When shutting down the syncer, run it at four times normal speed.
301  */
302 #define SYNCER_SHUTDOWN_SPEEDUP		4
303 static int sync_vnode_count;
304 static int syncer_worklist_len;
305 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
306     syncer_state;
307 
308 /* Target for maximum number of vnodes. */
309 u_long desiredvnodes;
310 static u_long gapvnodes;		/* gap between wanted and desired */
311 static u_long vhiwat;		/* enough extras after expansion */
312 static u_long vlowat;		/* minimal extras before expansion */
313 static u_long vstir;		/* nonzero to stir non-free vnodes */
314 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
315 
316 static u_long vnlru_read_freevnodes(void);
317 
318 /*
319  * Note that no attempt is made to sanitize these parameters.
320  */
321 static int
322 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
323 {
324 	u_long val;
325 	int error;
326 
327 	val = desiredvnodes;
328 	error = sysctl_handle_long(oidp, &val, 0, req);
329 	if (error != 0 || req->newptr == NULL)
330 		return (error);
331 
332 	if (val == desiredvnodes)
333 		return (0);
334 	mtx_lock(&vnode_list_mtx);
335 	desiredvnodes = val;
336 	wantfreevnodes = desiredvnodes / 4;
337 	vnlru_recalc();
338 	mtx_unlock(&vnode_list_mtx);
339 	/*
340 	 * XXX There is no protection against multiple threads changing
341 	 * desiredvnodes at the same time. Locking above only helps vnlru and
342 	 * getnewvnode.
343 	 */
344 	vfs_hash_changesize(desiredvnodes);
345 	cache_changesize(desiredvnodes);
346 	return (0);
347 }
348 
349 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
350     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
351     "LU", "Target for maximum number of vnodes");
352 
353 static int
354 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
355 {
356 	u_long val;
357 	int error;
358 
359 	val = wantfreevnodes;
360 	error = sysctl_handle_long(oidp, &val, 0, req);
361 	if (error != 0 || req->newptr == NULL)
362 		return (error);
363 
364 	if (val == wantfreevnodes)
365 		return (0);
366 	mtx_lock(&vnode_list_mtx);
367 	wantfreevnodes = val;
368 	vnlru_recalc();
369 	mtx_unlock(&vnode_list_mtx);
370 	return (0);
371 }
372 
373 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
374     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
375     "LU", "Target for minimum number of \"free\" vnodes");
376 
377 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
378     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
379 static int vnlru_nowhere;
380 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW | CTLFLAG_STATS,
381     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
382 
383 static int
384 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
385 {
386 	struct vnode *vp;
387 	struct nameidata nd;
388 	char *buf;
389 	unsigned long ndflags;
390 	int error;
391 
392 	if (req->newptr == NULL)
393 		return (EINVAL);
394 	if (req->newlen >= PATH_MAX)
395 		return (E2BIG);
396 
397 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
398 	error = SYSCTL_IN(req, buf, req->newlen);
399 	if (error != 0)
400 		goto out;
401 
402 	buf[req->newlen] = '\0';
403 
404 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1;
405 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf);
406 	if ((error = namei(&nd)) != 0)
407 		goto out;
408 	vp = nd.ni_vp;
409 
410 	if (VN_IS_DOOMED(vp)) {
411 		/*
412 		 * This vnode is being recycled.  Return != 0 to let the caller
413 		 * know that the sysctl had no effect.  Return EAGAIN because a
414 		 * subsequent call will likely succeed (since namei will create
415 		 * a new vnode if necessary)
416 		 */
417 		error = EAGAIN;
418 		goto putvnode;
419 	}
420 
421 	counter_u64_add(recycles_count, 1);
422 	vgone(vp);
423 putvnode:
424 	vput(vp);
425 	NDFREE_PNBUF(&nd);
426 out:
427 	free(buf, M_TEMP);
428 	return (error);
429 }
430 
431 static int
432 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
433 {
434 	struct thread *td = curthread;
435 	struct vnode *vp;
436 	struct file *fp;
437 	int error;
438 	int fd;
439 
440 	if (req->newptr == NULL)
441 		return (EBADF);
442 
443         error = sysctl_handle_int(oidp, &fd, 0, req);
444         if (error != 0)
445                 return (error);
446 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
447 	if (error != 0)
448 		return (error);
449 	vp = fp->f_vnode;
450 
451 	error = vn_lock(vp, LK_EXCLUSIVE);
452 	if (error != 0)
453 		goto drop;
454 
455 	counter_u64_add(recycles_count, 1);
456 	vgone(vp);
457 	VOP_UNLOCK(vp);
458 drop:
459 	fdrop(fp, td);
460 	return (error);
461 }
462 
463 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
464     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
465     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
466 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
467     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
468     sysctl_ftry_reclaim_vnode, "I",
469     "Try to reclaim a vnode by its file descriptor");
470 
471 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
472 #define vnsz2log 8
473 #ifndef DEBUG_LOCKS
474 _Static_assert(sizeof(struct vnode) >= 1UL << vnsz2log &&
475     sizeof(struct vnode) < 1UL << (vnsz2log + 1),
476     "vnsz2log needs to be updated");
477 #endif
478 
479 /*
480  * Support for the bufobj clean & dirty pctrie.
481  */
482 static void *
483 buf_trie_alloc(struct pctrie *ptree)
484 {
485 	return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT));
486 }
487 
488 static void
489 buf_trie_free(struct pctrie *ptree, void *node)
490 {
491 	uma_zfree_smr(buf_trie_zone, node);
492 }
493 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free,
494     buf_trie_smr);
495 
496 /*
497  * Initialize the vnode management data structures.
498  *
499  * Reevaluate the following cap on the number of vnodes after the physical
500  * memory size exceeds 512GB.  In the limit, as the physical memory size
501  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
502  */
503 #ifndef	MAXVNODES_MAX
504 #define	MAXVNODES_MAX	(512UL * 1024 * 1024 / 64)	/* 8M */
505 #endif
506 
507 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
508 
509 static struct vnode *
510 vn_alloc_marker(struct mount *mp)
511 {
512 	struct vnode *vp;
513 
514 	vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
515 	vp->v_type = VMARKER;
516 	vp->v_mount = mp;
517 
518 	return (vp);
519 }
520 
521 static void
522 vn_free_marker(struct vnode *vp)
523 {
524 
525 	MPASS(vp->v_type == VMARKER);
526 	free(vp, M_VNODE_MARKER);
527 }
528 
529 #ifdef KASAN
530 static int
531 vnode_ctor(void *mem, int size, void *arg __unused, int flags __unused)
532 {
533 	kasan_mark(mem, size, roundup2(size, UMA_ALIGN_PTR + 1), 0);
534 	return (0);
535 }
536 
537 static void
538 vnode_dtor(void *mem, int size, void *arg __unused)
539 {
540 	size_t end1, end2, off1, off2;
541 
542 	_Static_assert(offsetof(struct vnode, v_vnodelist) <
543 	    offsetof(struct vnode, v_dbatchcpu),
544 	    "KASAN marks require updating");
545 
546 	off1 = offsetof(struct vnode, v_vnodelist);
547 	off2 = offsetof(struct vnode, v_dbatchcpu);
548 	end1 = off1 + sizeof(((struct vnode *)NULL)->v_vnodelist);
549 	end2 = off2 + sizeof(((struct vnode *)NULL)->v_dbatchcpu);
550 
551 	/*
552 	 * Access to the v_vnodelist and v_dbatchcpu fields are permitted even
553 	 * after the vnode has been freed.  Try to get some KASAN coverage by
554 	 * marking everything except those two fields as invalid.  Because
555 	 * KASAN's tracking is not byte-granular, any preceding fields sharing
556 	 * the same 8-byte aligned word must also be marked valid.
557 	 */
558 
559 	/* Handle the area from the start until v_vnodelist... */
560 	off1 = rounddown2(off1, KASAN_SHADOW_SCALE);
561 	kasan_mark(mem, off1, off1, KASAN_UMA_FREED);
562 
563 	/* ... then the area between v_vnodelist and v_dbatchcpu ... */
564 	off1 = roundup2(end1, KASAN_SHADOW_SCALE);
565 	off2 = rounddown2(off2, KASAN_SHADOW_SCALE);
566 	if (off2 > off1)
567 		kasan_mark((void *)((char *)mem + off1), off2 - off1,
568 		    off2 - off1, KASAN_UMA_FREED);
569 
570 	/* ... and finally the area from v_dbatchcpu to the end. */
571 	off2 = roundup2(end2, KASAN_SHADOW_SCALE);
572 	kasan_mark((void *)((char *)mem + off2), size - off2, size - off2,
573 	    KASAN_UMA_FREED);
574 }
575 #endif /* KASAN */
576 
577 /*
578  * Initialize a vnode as it first enters the zone.
579  */
580 static int
581 vnode_init(void *mem, int size, int flags)
582 {
583 	struct vnode *vp;
584 
585 	vp = mem;
586 	bzero(vp, size);
587 	/*
588 	 * Setup locks.
589 	 */
590 	vp->v_vnlock = &vp->v_lock;
591 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
592 	/*
593 	 * By default, don't allow shared locks unless filesystems opt-in.
594 	 */
595 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
596 	    LK_NOSHARE | LK_IS_VNODE);
597 	/*
598 	 * Initialize bufobj.
599 	 */
600 	bufobj_init(&vp->v_bufobj, vp);
601 	/*
602 	 * Initialize namecache.
603 	 */
604 	cache_vnode_init(vp);
605 	/*
606 	 * Initialize rangelocks.
607 	 */
608 	rangelock_init(&vp->v_rl);
609 
610 	vp->v_dbatchcpu = NOCPU;
611 
612 	vp->v_state = VSTATE_DEAD;
613 
614 	/*
615 	 * Check vhold_recycle_free for an explanation.
616 	 */
617 	vp->v_holdcnt = VHOLD_NO_SMR;
618 	vp->v_type = VNON;
619 	mtx_lock(&vnode_list_mtx);
620 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
621 	mtx_unlock(&vnode_list_mtx);
622 	return (0);
623 }
624 
625 /*
626  * Free a vnode when it is cleared from the zone.
627  */
628 static void
629 vnode_fini(void *mem, int size)
630 {
631 	struct vnode *vp;
632 	struct bufobj *bo;
633 
634 	vp = mem;
635 	vdbatch_dequeue(vp);
636 	mtx_lock(&vnode_list_mtx);
637 	TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
638 	mtx_unlock(&vnode_list_mtx);
639 	rangelock_destroy(&vp->v_rl);
640 	lockdestroy(vp->v_vnlock);
641 	mtx_destroy(&vp->v_interlock);
642 	bo = &vp->v_bufobj;
643 	rw_destroy(BO_LOCKPTR(bo));
644 
645 	kasan_mark(mem, size, size, 0);
646 }
647 
648 /*
649  * Provide the size of NFS nclnode and NFS fh for calculation of the
650  * vnode memory consumption.  The size is specified directly to
651  * eliminate dependency on NFS-private header.
652  *
653  * Other filesystems may use bigger or smaller (like UFS and ZFS)
654  * private inode data, but the NFS-based estimation is ample enough.
655  * Still, we care about differences in the size between 64- and 32-bit
656  * platforms.
657  *
658  * Namecache structure size is heuristically
659  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
660  */
661 #ifdef _LP64
662 #define	NFS_NCLNODE_SZ	(528 + 64)
663 #define	NC_SZ		148
664 #else
665 #define	NFS_NCLNODE_SZ	(360 + 32)
666 #define	NC_SZ		92
667 #endif
668 
669 static void
670 vntblinit(void *dummy __unused)
671 {
672 	struct vdbatch *vd;
673 	uma_ctor ctor;
674 	uma_dtor dtor;
675 	int cpu, physvnodes, virtvnodes;
676 
677 	/*
678 	 * Desiredvnodes is a function of the physical memory size and the
679 	 * kernel's heap size.  Generally speaking, it scales with the
680 	 * physical memory size.  The ratio of desiredvnodes to the physical
681 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
682 	 * Thereafter, the
683 	 * marginal ratio of desiredvnodes to the physical memory size is
684 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
685 	 * size.  The memory required by desiredvnodes vnodes and vm objects
686 	 * must not exceed 1/10th of the kernel's heap size.
687 	 */
688 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
689 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
690 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
691 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
692 	desiredvnodes = min(physvnodes, virtvnodes);
693 	if (desiredvnodes > MAXVNODES_MAX) {
694 		if (bootverbose)
695 			printf("Reducing kern.maxvnodes %lu -> %lu\n",
696 			    desiredvnodes, MAXVNODES_MAX);
697 		desiredvnodes = MAXVNODES_MAX;
698 	}
699 	wantfreevnodes = desiredvnodes / 4;
700 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
701 	TAILQ_INIT(&vnode_list);
702 	mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
703 	/*
704 	 * The lock is taken to appease WITNESS.
705 	 */
706 	mtx_lock(&vnode_list_mtx);
707 	vnlru_recalc();
708 	mtx_unlock(&vnode_list_mtx);
709 	vnode_list_free_marker = vn_alloc_marker(NULL);
710 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
711 	vnode_list_reclaim_marker = vn_alloc_marker(NULL);
712 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
713 
714 #ifdef KASAN
715 	ctor = vnode_ctor;
716 	dtor = vnode_dtor;
717 #else
718 	ctor = NULL;
719 	dtor = NULL;
720 #endif
721 	vnode_zone = uma_zcreate("VNODE", sizeof(struct vnode), ctor, dtor,
722 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, UMA_ZONE_NOKASAN);
723 	uma_zone_set_smr(vnode_zone, vfs_smr);
724 
725 	/*
726 	 * Preallocate enough nodes to support one-per buf so that
727 	 * we can not fail an insert.  reassignbuf() callers can not
728 	 * tolerate the insertion failure.
729 	 */
730 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
731 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
732 	    UMA_ZONE_NOFREE | UMA_ZONE_SMR);
733 	buf_trie_smr = uma_zone_get_smr(buf_trie_zone);
734 	uma_prealloc(buf_trie_zone, nbuf);
735 
736 	vnodes_created = counter_u64_alloc(M_WAITOK);
737 	recycles_count = counter_u64_alloc(M_WAITOK);
738 	recycles_free_count = counter_u64_alloc(M_WAITOK);
739 	vnode_skipped_requeues = counter_u64_alloc(M_WAITOK);
740 
741 	/*
742 	 * Initialize the filesystem syncer.
743 	 */
744 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
745 	    &syncer_mask);
746 	syncer_maxdelay = syncer_mask + 1;
747 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
748 	cv_init(&sync_wakeup, "syncer");
749 
750 	CPU_FOREACH(cpu) {
751 		vd = DPCPU_ID_PTR((cpu), vd);
752 		bzero(vd, sizeof(*vd));
753 		mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
754 	}
755 }
756 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
757 
758 /*
759  * Mark a mount point as busy. Used to synchronize access and to delay
760  * unmounting. Eventually, mountlist_mtx is not released on failure.
761  *
762  * vfs_busy() is a custom lock, it can block the caller.
763  * vfs_busy() only sleeps if the unmount is active on the mount point.
764  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
765  * vnode belonging to mp.
766  *
767  * Lookup uses vfs_busy() to traverse mount points.
768  * root fs			var fs
769  * / vnode lock		A	/ vnode lock (/var)		D
770  * /var vnode lock	B	/log vnode lock(/var/log)	E
771  * vfs_busy lock	C	vfs_busy lock			F
772  *
773  * Within each file system, the lock order is C->A->B and F->D->E.
774  *
775  * When traversing across mounts, the system follows that lock order:
776  *
777  *        C->A->B
778  *              |
779  *              +->F->D->E
780  *
781  * The lookup() process for namei("/var") illustrates the process:
782  *  1. VOP_LOOKUP() obtains B while A is held
783  *  2. vfs_busy() obtains a shared lock on F while A and B are held
784  *  3. vput() releases lock on B
785  *  4. vput() releases lock on A
786  *  5. VFS_ROOT() obtains lock on D while shared lock on F is held
787  *  6. vfs_unbusy() releases shared lock on F
788  *  7. vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
789  *     Attempt to lock A (instead of vp_crossmp) while D is held would
790  *     violate the global order, causing deadlocks.
791  *
792  * dounmount() locks B while F is drained.  Note that for stacked
793  * filesystems, D and B in the example above may be the same lock,
794  * which introdues potential lock order reversal deadlock between
795  * dounmount() and step 5 above.  These filesystems may avoid the LOR
796  * by setting VV_CROSSLOCK on the covered vnode so that lock B will
797  * remain held until after step 5.
798  */
799 int
800 vfs_busy(struct mount *mp, int flags)
801 {
802 	struct mount_pcpu *mpcpu;
803 
804 	MPASS((flags & ~MBF_MASK) == 0);
805 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
806 
807 	if (vfs_op_thread_enter(mp, mpcpu)) {
808 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
809 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
810 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
811 		vfs_mp_count_add_pcpu(mpcpu, ref, 1);
812 		vfs_mp_count_add_pcpu(mpcpu, lockref, 1);
813 		vfs_op_thread_exit(mp, mpcpu);
814 		if (flags & MBF_MNTLSTLOCK)
815 			mtx_unlock(&mountlist_mtx);
816 		return (0);
817 	}
818 
819 	MNT_ILOCK(mp);
820 	vfs_assert_mount_counters(mp);
821 	MNT_REF(mp);
822 	/*
823 	 * If mount point is currently being unmounted, sleep until the
824 	 * mount point fate is decided.  If thread doing the unmounting fails,
825 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
826 	 * that this mount point has survived the unmount attempt and vfs_busy
827 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
828 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
829 	 * about to be really destroyed.  vfs_busy needs to release its
830 	 * reference on the mount point in this case and return with ENOENT,
831 	 * telling the caller the mount it tried to busy is no longer valid.
832 	 */
833 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
834 		KASSERT(TAILQ_EMPTY(&mp->mnt_uppers),
835 		    ("%s: non-empty upper mount list with pending unmount",
836 		    __func__));
837 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
838 			MNT_REL(mp);
839 			MNT_IUNLOCK(mp);
840 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
841 			    __func__);
842 			return (ENOENT);
843 		}
844 		if (flags & MBF_MNTLSTLOCK)
845 			mtx_unlock(&mountlist_mtx);
846 		mp->mnt_kern_flag |= MNTK_MWAIT;
847 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
848 		if (flags & MBF_MNTLSTLOCK)
849 			mtx_lock(&mountlist_mtx);
850 		MNT_ILOCK(mp);
851 	}
852 	if (flags & MBF_MNTLSTLOCK)
853 		mtx_unlock(&mountlist_mtx);
854 	mp->mnt_lockref++;
855 	MNT_IUNLOCK(mp);
856 	return (0);
857 }
858 
859 /*
860  * Free a busy filesystem.
861  */
862 void
863 vfs_unbusy(struct mount *mp)
864 {
865 	struct mount_pcpu *mpcpu;
866 	int c;
867 
868 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
869 
870 	if (vfs_op_thread_enter(mp, mpcpu)) {
871 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
872 		vfs_mp_count_sub_pcpu(mpcpu, lockref, 1);
873 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
874 		vfs_op_thread_exit(mp, mpcpu);
875 		return;
876 	}
877 
878 	MNT_ILOCK(mp);
879 	vfs_assert_mount_counters(mp);
880 	MNT_REL(mp);
881 	c = --mp->mnt_lockref;
882 	if (mp->mnt_vfs_ops == 0) {
883 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
884 		MNT_IUNLOCK(mp);
885 		return;
886 	}
887 	if (c < 0)
888 		vfs_dump_mount_counters(mp);
889 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
890 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
891 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
892 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
893 		wakeup(&mp->mnt_lockref);
894 	}
895 	MNT_IUNLOCK(mp);
896 }
897 
898 /*
899  * Lookup a mount point by filesystem identifier.
900  */
901 struct mount *
902 vfs_getvfs(fsid_t *fsid)
903 {
904 	struct mount *mp;
905 
906 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
907 	mtx_lock(&mountlist_mtx);
908 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
909 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
910 			vfs_ref(mp);
911 			mtx_unlock(&mountlist_mtx);
912 			return (mp);
913 		}
914 	}
915 	mtx_unlock(&mountlist_mtx);
916 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
917 	return ((struct mount *) 0);
918 }
919 
920 /*
921  * Lookup a mount point by filesystem identifier, busying it before
922  * returning.
923  *
924  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
925  * cache for popular filesystem identifiers.  The cache is lockess, using
926  * the fact that struct mount's are never freed.  In worst case we may
927  * get pointer to unmounted or even different filesystem, so we have to
928  * check what we got, and go slow way if so.
929  */
930 struct mount *
931 vfs_busyfs(fsid_t *fsid)
932 {
933 #define	FSID_CACHE_SIZE	256
934 	typedef struct mount * volatile vmp_t;
935 	static vmp_t cache[FSID_CACHE_SIZE];
936 	struct mount *mp;
937 	int error;
938 	uint32_t hash;
939 
940 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
941 	hash = fsid->val[0] ^ fsid->val[1];
942 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
943 	mp = cache[hash];
944 	if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0)
945 		goto slow;
946 	if (vfs_busy(mp, 0) != 0) {
947 		cache[hash] = NULL;
948 		goto slow;
949 	}
950 	if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0)
951 		return (mp);
952 	else
953 	    vfs_unbusy(mp);
954 
955 slow:
956 	mtx_lock(&mountlist_mtx);
957 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
958 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
959 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
960 			if (error) {
961 				cache[hash] = NULL;
962 				mtx_unlock(&mountlist_mtx);
963 				return (NULL);
964 			}
965 			cache[hash] = mp;
966 			return (mp);
967 		}
968 	}
969 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
970 	mtx_unlock(&mountlist_mtx);
971 	return ((struct mount *) 0);
972 }
973 
974 /*
975  * Check if a user can access privileged mount options.
976  */
977 int
978 vfs_suser(struct mount *mp, struct thread *td)
979 {
980 	int error;
981 
982 	if (jailed(td->td_ucred)) {
983 		/*
984 		 * If the jail of the calling thread lacks permission for
985 		 * this type of file system, deny immediately.
986 		 */
987 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
988 			return (EPERM);
989 
990 		/*
991 		 * If the file system was mounted outside the jail of the
992 		 * calling thread, deny immediately.
993 		 */
994 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
995 			return (EPERM);
996 	}
997 
998 	/*
999 	 * If file system supports delegated administration, we don't check
1000 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
1001 	 * by the file system itself.
1002 	 * If this is not the user that did original mount, we check for
1003 	 * the PRIV_VFS_MOUNT_OWNER privilege.
1004 	 */
1005 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
1006 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
1007 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
1008 			return (error);
1009 	}
1010 	return (0);
1011 }
1012 
1013 /*
1014  * Get a new unique fsid.  Try to make its val[0] unique, since this value
1015  * will be used to create fake device numbers for stat().  Also try (but
1016  * not so hard) make its val[0] unique mod 2^16, since some emulators only
1017  * support 16-bit device numbers.  We end up with unique val[0]'s for the
1018  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
1019  *
1020  * Keep in mind that several mounts may be running in parallel.  Starting
1021  * the search one past where the previous search terminated is both a
1022  * micro-optimization and a defense against returning the same fsid to
1023  * different mounts.
1024  */
1025 void
1026 vfs_getnewfsid(struct mount *mp)
1027 {
1028 	static uint16_t mntid_base;
1029 	struct mount *nmp;
1030 	fsid_t tfsid;
1031 	int mtype;
1032 
1033 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
1034 	mtx_lock(&mntid_mtx);
1035 	mtype = mp->mnt_vfc->vfc_typenum;
1036 	tfsid.val[1] = mtype;
1037 	mtype = (mtype & 0xFF) << 24;
1038 	for (;;) {
1039 		tfsid.val[0] = makedev(255,
1040 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
1041 		mntid_base++;
1042 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
1043 			break;
1044 		vfs_rel(nmp);
1045 	}
1046 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
1047 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
1048 	mtx_unlock(&mntid_mtx);
1049 }
1050 
1051 /*
1052  * Knob to control the precision of file timestamps:
1053  *
1054  *   0 = seconds only; nanoseconds zeroed.
1055  *   1 = seconds and nanoseconds, accurate within 1/HZ.
1056  *   2 = seconds and nanoseconds, truncated to microseconds.
1057  * >=3 = seconds and nanoseconds, maximum precision.
1058  */
1059 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
1060 
1061 static int timestamp_precision = TSP_USEC;
1062 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
1063     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
1064     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1065     "3+: sec + ns (max. precision))");
1066 
1067 /*
1068  * Get a current timestamp.
1069  */
1070 void
1071 vfs_timestamp(struct timespec *tsp)
1072 {
1073 	struct timeval tv;
1074 
1075 	switch (timestamp_precision) {
1076 	case TSP_SEC:
1077 		tsp->tv_sec = time_second;
1078 		tsp->tv_nsec = 0;
1079 		break;
1080 	case TSP_HZ:
1081 		getnanotime(tsp);
1082 		break;
1083 	case TSP_USEC:
1084 		microtime(&tv);
1085 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1086 		break;
1087 	case TSP_NSEC:
1088 	default:
1089 		nanotime(tsp);
1090 		break;
1091 	}
1092 }
1093 
1094 /*
1095  * Set vnode attributes to VNOVAL
1096  */
1097 void
1098 vattr_null(struct vattr *vap)
1099 {
1100 
1101 	vap->va_type = VNON;
1102 	vap->va_size = VNOVAL;
1103 	vap->va_bytes = VNOVAL;
1104 	vap->va_mode = VNOVAL;
1105 	vap->va_nlink = VNOVAL;
1106 	vap->va_uid = VNOVAL;
1107 	vap->va_gid = VNOVAL;
1108 	vap->va_fsid = VNOVAL;
1109 	vap->va_fileid = VNOVAL;
1110 	vap->va_blocksize = VNOVAL;
1111 	vap->va_rdev = VNOVAL;
1112 	vap->va_atime.tv_sec = VNOVAL;
1113 	vap->va_atime.tv_nsec = VNOVAL;
1114 	vap->va_mtime.tv_sec = VNOVAL;
1115 	vap->va_mtime.tv_nsec = VNOVAL;
1116 	vap->va_ctime.tv_sec = VNOVAL;
1117 	vap->va_ctime.tv_nsec = VNOVAL;
1118 	vap->va_birthtime.tv_sec = VNOVAL;
1119 	vap->va_birthtime.tv_nsec = VNOVAL;
1120 	vap->va_flags = VNOVAL;
1121 	vap->va_gen = VNOVAL;
1122 	vap->va_vaflags = 0;
1123 }
1124 
1125 /*
1126  * Try to reduce the total number of vnodes.
1127  *
1128  * This routine (and its user) are buggy in at least the following ways:
1129  * - all parameters were picked years ago when RAM sizes were significantly
1130  *   smaller
1131  * - it can pick vnodes based on pages used by the vm object, but filesystems
1132  *   like ZFS don't use it making the pick broken
1133  * - since ZFS has its own aging policy it gets partially combated by this one
1134  * - a dedicated method should be provided for filesystems to let them decide
1135  *   whether the vnode should be recycled
1136  *
1137  * This routine is called when we have too many vnodes.  It attempts
1138  * to free <count> vnodes and will potentially free vnodes that still
1139  * have VM backing store (VM backing store is typically the cause
1140  * of a vnode blowout so we want to do this).  Therefore, this operation
1141  * is not considered cheap.
1142  *
1143  * A number of conditions may prevent a vnode from being reclaimed.
1144  * the buffer cache may have references on the vnode, a directory
1145  * vnode may still have references due to the namei cache representing
1146  * underlying files, or the vnode may be in active use.   It is not
1147  * desirable to reuse such vnodes.  These conditions may cause the
1148  * number of vnodes to reach some minimum value regardless of what
1149  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
1150  *
1151  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1152  * 			 entries if this argument is strue
1153  * @param trigger	 Only reclaim vnodes with fewer than this many resident
1154  *			 pages.
1155  * @param target	 How many vnodes to reclaim.
1156  * @return		 The number of vnodes that were reclaimed.
1157  */
1158 static int
1159 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1160 {
1161 	struct vnode *vp, *mvp;
1162 	struct mount *mp;
1163 	struct vm_object *object;
1164 	u_long done;
1165 	bool retried;
1166 
1167 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1168 
1169 	retried = false;
1170 	done = 0;
1171 
1172 	mvp = vnode_list_reclaim_marker;
1173 restart:
1174 	vp = mvp;
1175 	while (done < target) {
1176 		vp = TAILQ_NEXT(vp, v_vnodelist);
1177 		if (__predict_false(vp == NULL))
1178 			break;
1179 
1180 		if (__predict_false(vp->v_type == VMARKER))
1181 			continue;
1182 
1183 		/*
1184 		 * If it's been deconstructed already, it's still
1185 		 * referenced, or it exceeds the trigger, skip it.
1186 		 * Also skip free vnodes.  We are trying to make space
1187 		 * to expand the free list, not reduce it.
1188 		 */
1189 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1190 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1191 			goto next_iter;
1192 
1193 		if (vp->v_type == VBAD || vp->v_type == VNON)
1194 			goto next_iter;
1195 
1196 		object = atomic_load_ptr(&vp->v_object);
1197 		if (object == NULL || object->resident_page_count > trigger) {
1198 			goto next_iter;
1199 		}
1200 
1201 		/*
1202 		 * Handle races against vnode allocation. Filesystems lock the
1203 		 * vnode some time after it gets returned from getnewvnode,
1204 		 * despite type and hold count being manipulated earlier.
1205 		 * Resorting to checking v_mount restores guarantees present
1206 		 * before the global list was reworked to contain all vnodes.
1207 		 */
1208 		if (!VI_TRYLOCK(vp))
1209 			goto next_iter;
1210 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1211 			VI_UNLOCK(vp);
1212 			goto next_iter;
1213 		}
1214 		if (vp->v_mount == NULL) {
1215 			VI_UNLOCK(vp);
1216 			goto next_iter;
1217 		}
1218 		vholdl(vp);
1219 		VI_UNLOCK(vp);
1220 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1221 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1222 		mtx_unlock(&vnode_list_mtx);
1223 
1224 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1225 			vdrop_recycle(vp);
1226 			goto next_iter_unlocked;
1227 		}
1228 		if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1229 			vdrop_recycle(vp);
1230 			vn_finished_write(mp);
1231 			goto next_iter_unlocked;
1232 		}
1233 
1234 		VI_LOCK(vp);
1235 		if (vp->v_usecount > 0 ||
1236 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1237 		    (vp->v_object != NULL && vp->v_object->handle == vp &&
1238 		    vp->v_object->resident_page_count > trigger)) {
1239 			VOP_UNLOCK(vp);
1240 			vdropl_recycle(vp);
1241 			vn_finished_write(mp);
1242 			goto next_iter_unlocked;
1243 		}
1244 		counter_u64_add(recycles_count, 1);
1245 		vgonel(vp);
1246 		VOP_UNLOCK(vp);
1247 		vdropl_recycle(vp);
1248 		vn_finished_write(mp);
1249 		done++;
1250 next_iter_unlocked:
1251 		maybe_yield();
1252 		mtx_lock(&vnode_list_mtx);
1253 		goto restart;
1254 next_iter:
1255 		MPASS(vp->v_type != VMARKER);
1256 		if (!should_yield())
1257 			continue;
1258 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1259 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1260 		mtx_unlock(&vnode_list_mtx);
1261 		kern_yield(PRI_USER);
1262 		mtx_lock(&vnode_list_mtx);
1263 		goto restart;
1264 	}
1265 	if (done == 0 && !retried) {
1266 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1267 		TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1268 		retried = true;
1269 		goto restart;
1270 	}
1271 	return (done);
1272 }
1273 
1274 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1275 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1276     0,
1277     "limit on vnode free requests per call to the vnlru_free routine");
1278 
1279 /*
1280  * Attempt to reduce the free list by the requested amount.
1281  */
1282 static int
1283 vnlru_free_impl(int count, struct vfsops *mnt_op, struct vnode *mvp)
1284 {
1285 	struct vnode *vp;
1286 	struct mount *mp;
1287 	int ocount;
1288 	bool retried;
1289 
1290 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1291 	if (count > max_vnlru_free)
1292 		count = max_vnlru_free;
1293 	if (count == 0) {
1294 		mtx_unlock(&vnode_list_mtx);
1295 		return (0);
1296 	}
1297 	ocount = count;
1298 	retried = false;
1299 	vp = mvp;
1300 	for (;;) {
1301 		vp = TAILQ_NEXT(vp, v_vnodelist);
1302 		if (__predict_false(vp == NULL)) {
1303 			/*
1304 			 * The free vnode marker can be past eligible vnodes:
1305 			 * 1. if vdbatch_process trylock failed
1306 			 * 2. if vtryrecycle failed
1307 			 *
1308 			 * If so, start the scan from scratch.
1309 			 */
1310 			if (!retried && vnlru_read_freevnodes() > 0) {
1311 				TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1312 				TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1313 				vp = mvp;
1314 				retried = true;
1315 				continue;
1316 			}
1317 
1318 			/*
1319 			 * Give up
1320 			 */
1321 			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1322 			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1323 			mtx_unlock(&vnode_list_mtx);
1324 			break;
1325 		}
1326 		if (__predict_false(vp->v_type == VMARKER))
1327 			continue;
1328 		if (vp->v_holdcnt > 0)
1329 			continue;
1330 		/*
1331 		 * Don't recycle if our vnode is from different type
1332 		 * of mount point.  Note that mp is type-safe, the
1333 		 * check does not reach unmapped address even if
1334 		 * vnode is reclaimed.
1335 		 */
1336 		if (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1337 		    mp->mnt_op != mnt_op) {
1338 			continue;
1339 		}
1340 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1341 			continue;
1342 		}
1343 		if (!vhold_recycle_free(vp))
1344 			continue;
1345 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1346 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1347 		mtx_unlock(&vnode_list_mtx);
1348 		/*
1349 		 * FIXME: ignores the return value, meaning it may be nothing
1350 		 * got recycled but it claims otherwise to the caller.
1351 		 *
1352 		 * Originally the value started being ignored in 2005 with
1353 		 * 114a1006a8204aa156e1f9ad6476cdff89cada7f .
1354 		 *
1355 		 * Respecting the value can run into significant stalls if most
1356 		 * vnodes belong to one file system and it has writes
1357 		 * suspended.  In presence of many threads and millions of
1358 		 * vnodes they keep contending on the vnode_list_mtx lock only
1359 		 * to find vnodes they can't recycle.
1360 		 *
1361 		 * The solution would be to pre-check if the vnode is likely to
1362 		 * be recycle-able, but it needs to happen with the
1363 		 * vnode_list_mtx lock held. This runs into a problem where
1364 		 * VOP_GETWRITEMOUNT (currently needed to find out about if
1365 		 * writes are frozen) can take locks which LOR against it.
1366 		 *
1367 		 * Check nullfs for one example (null_getwritemount).
1368 		 */
1369 		vtryrecycle(vp);
1370 		count--;
1371 		if (count == 0) {
1372 			break;
1373 		}
1374 		mtx_lock(&vnode_list_mtx);
1375 		vp = mvp;
1376 	}
1377 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1378 	return (ocount - count);
1379 }
1380 
1381 /*
1382  * XXX: returns without vnode_list_mtx locked!
1383  */
1384 static int
1385 vnlru_free_locked(int count)
1386 {
1387 	int ret;
1388 
1389 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1390 	ret = vnlru_free_impl(count, NULL, vnode_list_free_marker);
1391 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1392 	return (ret);
1393 }
1394 
1395 void
1396 vnlru_free_vfsops(int count, struct vfsops *mnt_op, struct vnode *mvp)
1397 {
1398 
1399 	MPASS(mnt_op != NULL);
1400 	MPASS(mvp != NULL);
1401 	VNPASS(mvp->v_type == VMARKER, mvp);
1402 	mtx_lock(&vnode_list_mtx);
1403 	vnlru_free_impl(count, mnt_op, mvp);
1404 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1405 }
1406 
1407 struct vnode *
1408 vnlru_alloc_marker(void)
1409 {
1410 	struct vnode *mvp;
1411 
1412 	mvp = vn_alloc_marker(NULL);
1413 	mtx_lock(&vnode_list_mtx);
1414 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, mvp, v_vnodelist);
1415 	mtx_unlock(&vnode_list_mtx);
1416 	return (mvp);
1417 }
1418 
1419 void
1420 vnlru_free_marker(struct vnode *mvp)
1421 {
1422 	mtx_lock(&vnode_list_mtx);
1423 	TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1424 	mtx_unlock(&vnode_list_mtx);
1425 	vn_free_marker(mvp);
1426 }
1427 
1428 static void
1429 vnlru_recalc(void)
1430 {
1431 
1432 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1433 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1434 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1435 	vlowat = vhiwat / 2;
1436 }
1437 
1438 /*
1439  * Attempt to recycle vnodes in a context that is always safe to block.
1440  * Calling vlrurecycle() from the bowels of filesystem code has some
1441  * interesting deadlock problems.
1442  */
1443 static struct proc *vnlruproc;
1444 static int vnlruproc_sig;
1445 
1446 /*
1447  * The main freevnodes counter is only updated when a counter local to CPU
1448  * diverges from 0 by more than VNLRU_FREEVNODES_SLOP. CPUs are conditionally
1449  * walked to compute a more accurate total.
1450  *
1451  * Note: the actual value at any given moment can still exceed slop, but it
1452  * should not be by significant margin in practice.
1453  */
1454 #define VNLRU_FREEVNODES_SLOP 126
1455 
1456 static void __noinline
1457 vfs_freevnodes_rollup(int8_t *lfreevnodes)
1458 {
1459 
1460 	atomic_add_long(&freevnodes, *lfreevnodes);
1461 	*lfreevnodes = 0;
1462 	critical_exit();
1463 }
1464 
1465 static __inline void
1466 vfs_freevnodes_inc(void)
1467 {
1468 	int8_t *lfreevnodes;
1469 
1470 	critical_enter();
1471 	lfreevnodes = PCPU_PTR(vfs_freevnodes);
1472 	(*lfreevnodes)++;
1473 	if (__predict_false(*lfreevnodes == VNLRU_FREEVNODES_SLOP))
1474 		vfs_freevnodes_rollup(lfreevnodes);
1475 	else
1476 		critical_exit();
1477 }
1478 
1479 static __inline void
1480 vfs_freevnodes_dec(void)
1481 {
1482 	int8_t *lfreevnodes;
1483 
1484 	critical_enter();
1485 	lfreevnodes = PCPU_PTR(vfs_freevnodes);
1486 	(*lfreevnodes)--;
1487 	if (__predict_false(*lfreevnodes == -VNLRU_FREEVNODES_SLOP))
1488 		vfs_freevnodes_rollup(lfreevnodes);
1489 	else
1490 		critical_exit();
1491 }
1492 
1493 static u_long
1494 vnlru_read_freevnodes(void)
1495 {
1496 	long slop, rfreevnodes;
1497 	int cpu;
1498 
1499 	rfreevnodes = atomic_load_long(&freevnodes);
1500 
1501 	if (rfreevnodes > freevnodes_old)
1502 		slop = rfreevnodes - freevnodes_old;
1503 	else
1504 		slop = freevnodes_old - rfreevnodes;
1505 	if (slop < VNLRU_FREEVNODES_SLOP)
1506 		return (rfreevnodes >= 0 ? rfreevnodes : 0);
1507 	freevnodes_old = rfreevnodes;
1508 	CPU_FOREACH(cpu) {
1509 		freevnodes_old += cpuid_to_pcpu[cpu]->pc_vfs_freevnodes;
1510 	}
1511 	return (freevnodes_old >= 0 ? freevnodes_old : 0);
1512 }
1513 
1514 static bool
1515 vnlru_under(u_long rnumvnodes, u_long limit)
1516 {
1517 	u_long rfreevnodes, space;
1518 
1519 	if (__predict_false(rnumvnodes > desiredvnodes))
1520 		return (true);
1521 
1522 	space = desiredvnodes - rnumvnodes;
1523 	if (space < limit) {
1524 		rfreevnodes = vnlru_read_freevnodes();
1525 		if (rfreevnodes > wantfreevnodes)
1526 			space += rfreevnodes - wantfreevnodes;
1527 	}
1528 	return (space < limit);
1529 }
1530 
1531 static bool
1532 vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1533 {
1534 	long rfreevnodes, space;
1535 
1536 	if (__predict_false(rnumvnodes > desiredvnodes))
1537 		return (true);
1538 
1539 	space = desiredvnodes - rnumvnodes;
1540 	if (space < limit) {
1541 		rfreevnodes = atomic_load_long(&freevnodes);
1542 		if (rfreevnodes > wantfreevnodes)
1543 			space += rfreevnodes - wantfreevnodes;
1544 	}
1545 	return (space < limit);
1546 }
1547 
1548 static void
1549 vnlru_kick_locked(void)
1550 {
1551 
1552 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1553 	if (vnlruproc_sig == 0) {
1554 		vnlruproc_sig = 1;
1555 		wakeup(vnlruproc);
1556 	}
1557 }
1558 
1559 static void
1560 vnlru_kick_cond(void)
1561 {
1562 
1563 	if (vnlruproc_sig)
1564 		return;
1565 	mtx_lock(&vnode_list_mtx);
1566 	vnlru_kick_locked();
1567 	mtx_unlock(&vnode_list_mtx);
1568 }
1569 
1570 static void
1571 vnlru_proc(void)
1572 {
1573 	u_long rnumvnodes, rfreevnodes, target;
1574 	unsigned long onumvnodes;
1575 	int done, force, trigger, usevnodes;
1576 	bool reclaim_nc_src, want_reread;
1577 
1578 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1579 	    SHUTDOWN_PRI_FIRST);
1580 
1581 	force = 0;
1582 	want_reread = false;
1583 	for (;;) {
1584 		kproc_suspend_check(vnlruproc);
1585 		mtx_lock(&vnode_list_mtx);
1586 		rnumvnodes = atomic_load_long(&numvnodes);
1587 
1588 		if (want_reread) {
1589 			force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1590 			want_reread = false;
1591 		}
1592 
1593 		/*
1594 		 * If numvnodes is too large (due to desiredvnodes being
1595 		 * adjusted using its sysctl, or emergency growth), first
1596 		 * try to reduce it by discarding from the free list.
1597 		 */
1598 		if (rnumvnodes > desiredvnodes) {
1599 			vnlru_free_locked(rnumvnodes - desiredvnodes);
1600 			mtx_lock(&vnode_list_mtx);
1601 			rnumvnodes = atomic_load_long(&numvnodes);
1602 		}
1603 		/*
1604 		 * Sleep if the vnode cache is in a good state.  This is
1605 		 * when it is not over-full and has space for about a 4%
1606 		 * or 9% expansion (by growing its size or inexcessively
1607 		 * reducing its free list).  Otherwise, try to reclaim
1608 		 * space for a 10% expansion.
1609 		 */
1610 		if (vstir && force == 0) {
1611 			force = 1;
1612 			vstir = 0;
1613 		}
1614 		if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1615 			vnlruproc_sig = 0;
1616 			wakeup(&vnlruproc_sig);
1617 			msleep(vnlruproc, &vnode_list_mtx,
1618 			    PVFS|PDROP, "vlruwt", hz);
1619 			continue;
1620 		}
1621 		rfreevnodes = vnlru_read_freevnodes();
1622 
1623 		onumvnodes = rnumvnodes;
1624 		/*
1625 		 * Calculate parameters for recycling.  These are the same
1626 		 * throughout the loop to give some semblance of fairness.
1627 		 * The trigger point is to avoid recycling vnodes with lots
1628 		 * of resident pages.  We aren't trying to free memory; we
1629 		 * are trying to recycle or at least free vnodes.
1630 		 */
1631 		if (rnumvnodes <= desiredvnodes)
1632 			usevnodes = rnumvnodes - rfreevnodes;
1633 		else
1634 			usevnodes = rnumvnodes;
1635 		if (usevnodes <= 0)
1636 			usevnodes = 1;
1637 		/*
1638 		 * The trigger value is chosen to give a conservatively
1639 		 * large value to ensure that it alone doesn't prevent
1640 		 * making progress.  The value can easily be so large that
1641 		 * it is effectively infinite in some congested and
1642 		 * misconfigured cases, and this is necessary.  Normally
1643 		 * it is about 8 to 100 (pages), which is quite large.
1644 		 */
1645 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1646 		if (force < 2)
1647 			trigger = vsmalltrigger;
1648 		reclaim_nc_src = force >= 3;
1649 		target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1650 		target = target / 10 + 1;
1651 		done = vlrureclaim(reclaim_nc_src, trigger, target);
1652 		mtx_unlock(&vnode_list_mtx);
1653 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1654 			uma_reclaim(UMA_RECLAIM_DRAIN);
1655 		if (done == 0) {
1656 			if (force == 0 || force == 1) {
1657 				force = 2;
1658 				continue;
1659 			}
1660 			if (force == 2) {
1661 				force = 3;
1662 				continue;
1663 			}
1664 			want_reread = true;
1665 			force = 0;
1666 			vnlru_nowhere++;
1667 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1668 		} else {
1669 			want_reread = true;
1670 			kern_yield(PRI_USER);
1671 		}
1672 	}
1673 }
1674 
1675 static struct kproc_desc vnlru_kp = {
1676 	"vnlru",
1677 	vnlru_proc,
1678 	&vnlruproc
1679 };
1680 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1681     &vnlru_kp);
1682 
1683 /*
1684  * Routines having to do with the management of the vnode table.
1685  */
1686 
1687 /*
1688  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1689  * before we actually vgone().  This function must be called with the vnode
1690  * held to prevent the vnode from being returned to the free list midway
1691  * through vgone().
1692  */
1693 static int
1694 vtryrecycle(struct vnode *vp)
1695 {
1696 	struct mount *vnmp;
1697 
1698 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1699 	VNPASS(vp->v_holdcnt > 0, vp);
1700 	/*
1701 	 * This vnode may found and locked via some other list, if so we
1702 	 * can't recycle it yet.
1703 	 */
1704 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1705 		CTR2(KTR_VFS,
1706 		    "%s: impossible to recycle, vp %p lock is already held",
1707 		    __func__, vp);
1708 		vdrop_recycle(vp);
1709 		return (EWOULDBLOCK);
1710 	}
1711 	/*
1712 	 * Don't recycle if its filesystem is being suspended.
1713 	 */
1714 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1715 		VOP_UNLOCK(vp);
1716 		CTR2(KTR_VFS,
1717 		    "%s: impossible to recycle, cannot start the write for %p",
1718 		    __func__, vp);
1719 		vdrop_recycle(vp);
1720 		return (EBUSY);
1721 	}
1722 	/*
1723 	 * If we got this far, we need to acquire the interlock and see if
1724 	 * anyone picked up this vnode from another list.  If not, we will
1725 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1726 	 * will skip over it.
1727 	 */
1728 	VI_LOCK(vp);
1729 	if (vp->v_usecount) {
1730 		VOP_UNLOCK(vp);
1731 		vdropl_recycle(vp);
1732 		vn_finished_write(vnmp);
1733 		CTR2(KTR_VFS,
1734 		    "%s: impossible to recycle, %p is already referenced",
1735 		    __func__, vp);
1736 		return (EBUSY);
1737 	}
1738 	if (!VN_IS_DOOMED(vp)) {
1739 		counter_u64_add(recycles_free_count, 1);
1740 		vgonel(vp);
1741 	}
1742 	VOP_UNLOCK(vp);
1743 	vdropl_recycle(vp);
1744 	vn_finished_write(vnmp);
1745 	return (0);
1746 }
1747 
1748 /*
1749  * Allocate a new vnode.
1750  *
1751  * The operation never returns an error. Returning an error was disabled
1752  * in r145385 (dated 2005) with the following comment:
1753  *
1754  * XXX Not all VFS_VGET/ffs_vget callers check returns.
1755  *
1756  * Given the age of this commit (almost 15 years at the time of writing this
1757  * comment) restoring the ability to fail requires a significant audit of
1758  * all codepaths.
1759  *
1760  * The routine can try to free a vnode or stall for up to 1 second waiting for
1761  * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1762  */
1763 static u_long vn_alloc_cyclecount;
1764 static u_long vn_alloc_sleeps;
1765 
1766 SYSCTL_ULONG(_vfs, OID_AUTO, vnode_alloc_sleeps, CTLFLAG_RD, &vn_alloc_sleeps, 0,
1767     "Number of times vnode allocation blocked waiting on vnlru");
1768 
1769 static struct vnode * __noinline
1770 vn_alloc_hard(struct mount *mp)
1771 {
1772 	u_long rnumvnodes, rfreevnodes;
1773 
1774 	mtx_lock(&vnode_list_mtx);
1775 	rnumvnodes = atomic_load_long(&numvnodes);
1776 	if (rnumvnodes + 1 < desiredvnodes) {
1777 		vn_alloc_cyclecount = 0;
1778 		mtx_unlock(&vnode_list_mtx);
1779 		goto alloc;
1780 	}
1781 	rfreevnodes = vnlru_read_freevnodes();
1782 	if (vn_alloc_cyclecount++ >= rfreevnodes) {
1783 		vn_alloc_cyclecount = 0;
1784 		vstir = 1;
1785 	}
1786 	/*
1787 	 * Grow the vnode cache if it will not be above its target max
1788 	 * after growing.  Otherwise, if the free list is nonempty, try
1789 	 * to reclaim 1 item from it before growing the cache (possibly
1790 	 * above its target max if the reclamation failed or is delayed).
1791 	 * Otherwise, wait for some space.  In all cases, schedule
1792 	 * vnlru_proc() if we are getting short of space.  The watermarks
1793 	 * should be chosen so that we never wait or even reclaim from
1794 	 * the free list to below its target minimum.
1795 	 */
1796 	if (vnlru_free_locked(1) > 0)
1797 		goto alloc;
1798 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1799 	if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1800 		/*
1801 		 * Wait for space for a new vnode.
1802 		 */
1803 		mtx_lock(&vnode_list_mtx);
1804 		vnlru_kick_locked();
1805 		vn_alloc_sleeps++;
1806 		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1807 		if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1808 		    vnlru_read_freevnodes() > 1)
1809 			vnlru_free_locked(1);
1810 		else
1811 			mtx_unlock(&vnode_list_mtx);
1812 	}
1813 alloc:
1814 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1815 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1816 	if (vnlru_under(rnumvnodes, vlowat))
1817 		vnlru_kick_cond();
1818 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1819 }
1820 
1821 static struct vnode *
1822 vn_alloc(struct mount *mp)
1823 {
1824 	u_long rnumvnodes;
1825 
1826 	if (__predict_false(vn_alloc_cyclecount != 0))
1827 		return (vn_alloc_hard(mp));
1828 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1829 	if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1830 		atomic_subtract_long(&numvnodes, 1);
1831 		return (vn_alloc_hard(mp));
1832 	}
1833 
1834 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1835 }
1836 
1837 static void
1838 vn_free(struct vnode *vp)
1839 {
1840 
1841 	atomic_subtract_long(&numvnodes, 1);
1842 	uma_zfree_smr(vnode_zone, vp);
1843 }
1844 
1845 /*
1846  * Return the next vnode from the free list.
1847  */
1848 int
1849 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1850     struct vnode **vpp)
1851 {
1852 	struct vnode *vp;
1853 	struct thread *td;
1854 	struct lock_object *lo;
1855 
1856 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1857 
1858 	KASSERT(vops->registered,
1859 	    ("%s: not registered vector op %p\n", __func__, vops));
1860 	cache_validate_vop_vector(mp, vops);
1861 
1862 	td = curthread;
1863 	if (td->td_vp_reserved != NULL) {
1864 		vp = td->td_vp_reserved;
1865 		td->td_vp_reserved = NULL;
1866 	} else {
1867 		vp = vn_alloc(mp);
1868 	}
1869 	counter_u64_add(vnodes_created, 1);
1870 
1871 	vn_set_state(vp, VSTATE_UNINITIALIZED);
1872 
1873 	/*
1874 	 * Locks are given the generic name "vnode" when created.
1875 	 * Follow the historic practice of using the filesystem
1876 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1877 	 *
1878 	 * Locks live in a witness group keyed on their name. Thus,
1879 	 * when a lock is renamed, it must also move from the witness
1880 	 * group of its old name to the witness group of its new name.
1881 	 *
1882 	 * The change only needs to be made when the vnode moves
1883 	 * from one filesystem type to another. We ensure that each
1884 	 * filesystem use a single static name pointer for its tag so
1885 	 * that we can compare pointers rather than doing a strcmp().
1886 	 */
1887 	lo = &vp->v_vnlock->lock_object;
1888 #ifdef WITNESS
1889 	if (lo->lo_name != tag) {
1890 #endif
1891 		lo->lo_name = tag;
1892 #ifdef WITNESS
1893 		WITNESS_DESTROY(lo);
1894 		WITNESS_INIT(lo, tag);
1895 	}
1896 #endif
1897 	/*
1898 	 * By default, don't allow shared locks unless filesystems opt-in.
1899 	 */
1900 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1901 	/*
1902 	 * Finalize various vnode identity bits.
1903 	 */
1904 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1905 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1906 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1907 	vp->v_type = VNON;
1908 	vp->v_op = vops;
1909 	vp->v_irflag = 0;
1910 	v_init_counters(vp);
1911 	vn_seqc_init(vp);
1912 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1913 #ifdef DIAGNOSTIC
1914 	if (mp == NULL && vops != &dead_vnodeops)
1915 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1916 #endif
1917 #ifdef MAC
1918 	mac_vnode_init(vp);
1919 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1920 		mac_vnode_associate_singlelabel(mp, vp);
1921 #endif
1922 	if (mp != NULL) {
1923 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1924 	}
1925 
1926 	/*
1927 	 * For the filesystems which do not use vfs_hash_insert(),
1928 	 * still initialize v_hash to have vfs_hash_index() useful.
1929 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1930 	 * its own hashing.
1931 	 */
1932 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1933 
1934 	*vpp = vp;
1935 	return (0);
1936 }
1937 
1938 void
1939 getnewvnode_reserve(void)
1940 {
1941 	struct thread *td;
1942 
1943 	td = curthread;
1944 	MPASS(td->td_vp_reserved == NULL);
1945 	td->td_vp_reserved = vn_alloc(NULL);
1946 }
1947 
1948 void
1949 getnewvnode_drop_reserve(void)
1950 {
1951 	struct thread *td;
1952 
1953 	td = curthread;
1954 	if (td->td_vp_reserved != NULL) {
1955 		vn_free(td->td_vp_reserved);
1956 		td->td_vp_reserved = NULL;
1957 	}
1958 }
1959 
1960 static void __noinline
1961 freevnode(struct vnode *vp)
1962 {
1963 	struct bufobj *bo;
1964 
1965 	/*
1966 	 * The vnode has been marked for destruction, so free it.
1967 	 *
1968 	 * The vnode will be returned to the zone where it will
1969 	 * normally remain until it is needed for another vnode. We
1970 	 * need to cleanup (or verify that the cleanup has already
1971 	 * been done) any residual data left from its current use
1972 	 * so as not to contaminate the freshly allocated vnode.
1973 	 */
1974 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1975 	/*
1976 	 * Paired with vgone.
1977 	 */
1978 	vn_seqc_write_end_free(vp);
1979 
1980 	bo = &vp->v_bufobj;
1981 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1982 	VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp);
1983 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1984 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1985 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1986 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1987 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1988 	    ("clean blk trie not empty"));
1989 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1990 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1991 	    ("dirty blk trie not empty"));
1992 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1993 	    ("Dangling rangelock waiters"));
1994 	VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp,
1995 	    ("Leaked inactivation"));
1996 	VI_UNLOCK(vp);
1997 	cache_assert_no_entries(vp);
1998 
1999 #ifdef MAC
2000 	mac_vnode_destroy(vp);
2001 #endif
2002 	if (vp->v_pollinfo != NULL) {
2003 		/*
2004 		 * Use LK_NOWAIT to shut up witness about the lock. We may get
2005 		 * here while having another vnode locked when trying to
2006 		 * satisfy a lookup and needing to recycle.
2007 		 */
2008 		VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT);
2009 		destroy_vpollinfo(vp->v_pollinfo);
2010 		VOP_UNLOCK(vp);
2011 		vp->v_pollinfo = NULL;
2012 	}
2013 	vp->v_mountedhere = NULL;
2014 	vp->v_unpcb = NULL;
2015 	vp->v_rdev = NULL;
2016 	vp->v_fifoinfo = NULL;
2017 	vp->v_iflag = 0;
2018 	vp->v_vflag = 0;
2019 	bo->bo_flag = 0;
2020 	vn_free(vp);
2021 }
2022 
2023 /*
2024  * Delete from old mount point vnode list, if on one.
2025  */
2026 static void
2027 delmntque(struct vnode *vp)
2028 {
2029 	struct mount *mp;
2030 
2031 	VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
2032 
2033 	mp = vp->v_mount;
2034 	MNT_ILOCK(mp);
2035 	VI_LOCK(vp);
2036 	vp->v_mount = NULL;
2037 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
2038 		("bad mount point vnode list size"));
2039 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2040 	mp->mnt_nvnodelistsize--;
2041 	MNT_REL(mp);
2042 	MNT_IUNLOCK(mp);
2043 	/*
2044 	 * The caller expects the interlock to be still held.
2045 	 */
2046 	ASSERT_VI_LOCKED(vp, __func__);
2047 }
2048 
2049 static int
2050 insmntque1_int(struct vnode *vp, struct mount *mp, bool dtr)
2051 {
2052 
2053 	KASSERT(vp->v_mount == NULL,
2054 		("insmntque: vnode already on per mount vnode list"));
2055 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
2056 	if ((mp->mnt_kern_flag & MNTK_UNLOCKED_INSMNTQUE) == 0) {
2057 		ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
2058 	} else {
2059 		KASSERT(!dtr,
2060 		    ("%s: can't have MNTK_UNLOCKED_INSMNTQUE and cleanup",
2061 		    __func__));
2062 	}
2063 
2064 	/*
2065 	 * We acquire the vnode interlock early to ensure that the
2066 	 * vnode cannot be recycled by another process releasing a
2067 	 * holdcnt on it before we get it on both the vnode list
2068 	 * and the active vnode list. The mount mutex protects only
2069 	 * manipulation of the vnode list and the vnode freelist
2070 	 * mutex protects only manipulation of the active vnode list.
2071 	 * Hence the need to hold the vnode interlock throughout.
2072 	 */
2073 	MNT_ILOCK(mp);
2074 	VI_LOCK(vp);
2075 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
2076 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
2077 	    mp->mnt_nvnodelistsize == 0)) &&
2078 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
2079 		VI_UNLOCK(vp);
2080 		MNT_IUNLOCK(mp);
2081 		if (dtr) {
2082 			vp->v_data = NULL;
2083 			vp->v_op = &dead_vnodeops;
2084 			vgone(vp);
2085 			vput(vp);
2086 		}
2087 		return (EBUSY);
2088 	}
2089 	vp->v_mount = mp;
2090 	MNT_REF(mp);
2091 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2092 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
2093 		("neg mount point vnode list size"));
2094 	mp->mnt_nvnodelistsize++;
2095 	VI_UNLOCK(vp);
2096 	MNT_IUNLOCK(mp);
2097 	return (0);
2098 }
2099 
2100 /*
2101  * Insert into list of vnodes for the new mount point, if available.
2102  * insmntque() reclaims the vnode on insertion failure, insmntque1()
2103  * leaves handling of the vnode to the caller.
2104  */
2105 int
2106 insmntque(struct vnode *vp, struct mount *mp)
2107 {
2108 	return (insmntque1_int(vp, mp, true));
2109 }
2110 
2111 int
2112 insmntque1(struct vnode *vp, struct mount *mp)
2113 {
2114 	return (insmntque1_int(vp, mp, false));
2115 }
2116 
2117 /*
2118  * Flush out and invalidate all buffers associated with a bufobj
2119  * Called with the underlying object locked.
2120  */
2121 int
2122 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
2123 {
2124 	int error;
2125 
2126 	BO_LOCK(bo);
2127 	if (flags & V_SAVE) {
2128 		error = bufobj_wwait(bo, slpflag, slptimeo);
2129 		if (error) {
2130 			BO_UNLOCK(bo);
2131 			return (error);
2132 		}
2133 		if (bo->bo_dirty.bv_cnt > 0) {
2134 			BO_UNLOCK(bo);
2135 			do {
2136 				error = BO_SYNC(bo, MNT_WAIT);
2137 			} while (error == ERELOOKUP);
2138 			if (error != 0)
2139 				return (error);
2140 			BO_LOCK(bo);
2141 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) {
2142 				BO_UNLOCK(bo);
2143 				return (EBUSY);
2144 			}
2145 		}
2146 	}
2147 	/*
2148 	 * If you alter this loop please notice that interlock is dropped and
2149 	 * reacquired in flushbuflist.  Special care is needed to ensure that
2150 	 * no race conditions occur from this.
2151 	 */
2152 	do {
2153 		error = flushbuflist(&bo->bo_clean,
2154 		    flags, bo, slpflag, slptimeo);
2155 		if (error == 0 && !(flags & V_CLEANONLY))
2156 			error = flushbuflist(&bo->bo_dirty,
2157 			    flags, bo, slpflag, slptimeo);
2158 		if (error != 0 && error != EAGAIN) {
2159 			BO_UNLOCK(bo);
2160 			return (error);
2161 		}
2162 	} while (error != 0);
2163 
2164 	/*
2165 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
2166 	 * have write I/O in-progress but if there is a VM object then the
2167 	 * VM object can also have read-I/O in-progress.
2168 	 */
2169 	do {
2170 		bufobj_wwait(bo, 0, 0);
2171 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
2172 			BO_UNLOCK(bo);
2173 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
2174 			BO_LOCK(bo);
2175 		}
2176 	} while (bo->bo_numoutput > 0);
2177 	BO_UNLOCK(bo);
2178 
2179 	/*
2180 	 * Destroy the copy in the VM cache, too.
2181 	 */
2182 	if (bo->bo_object != NULL &&
2183 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
2184 		VM_OBJECT_WLOCK(bo->bo_object);
2185 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
2186 		    OBJPR_CLEANONLY : 0);
2187 		VM_OBJECT_WUNLOCK(bo->bo_object);
2188 	}
2189 
2190 #ifdef INVARIANTS
2191 	BO_LOCK(bo);
2192 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
2193 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
2194 	    bo->bo_clean.bv_cnt > 0))
2195 		panic("vinvalbuf: flush failed");
2196 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
2197 	    bo->bo_dirty.bv_cnt > 0)
2198 		panic("vinvalbuf: flush dirty failed");
2199 	BO_UNLOCK(bo);
2200 #endif
2201 	return (0);
2202 }
2203 
2204 /*
2205  * Flush out and invalidate all buffers associated with a vnode.
2206  * Called with the underlying object locked.
2207  */
2208 int
2209 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
2210 {
2211 
2212 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2213 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
2214 	if (vp->v_object != NULL && vp->v_object->handle != vp)
2215 		return (0);
2216 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
2217 }
2218 
2219 /*
2220  * Flush out buffers on the specified list.
2221  *
2222  */
2223 static int
2224 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2225     int slptimeo)
2226 {
2227 	struct buf *bp, *nbp;
2228 	int retval, error;
2229 	daddr_t lblkno;
2230 	b_xflags_t xflags;
2231 
2232 	ASSERT_BO_WLOCKED(bo);
2233 
2234 	retval = 0;
2235 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2236 		/*
2237 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2238 		 * do not skip any buffers. If we are flushing only V_NORMAL
2239 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2240 		 * flushing only V_ALT buffers then skip buffers not marked
2241 		 * as BX_ALTDATA.
2242 		 */
2243 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2244 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2245 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2246 			continue;
2247 		}
2248 		if (nbp != NULL) {
2249 			lblkno = nbp->b_lblkno;
2250 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2251 		}
2252 		retval = EAGAIN;
2253 		error = BUF_TIMELOCK(bp,
2254 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2255 		    "flushbuf", slpflag, slptimeo);
2256 		if (error) {
2257 			BO_LOCK(bo);
2258 			return (error != ENOLCK ? error : EAGAIN);
2259 		}
2260 		KASSERT(bp->b_bufobj == bo,
2261 		    ("bp %p wrong b_bufobj %p should be %p",
2262 		    bp, bp->b_bufobj, bo));
2263 		/*
2264 		 * XXX Since there are no node locks for NFS, I
2265 		 * believe there is a slight chance that a delayed
2266 		 * write will occur while sleeping just above, so
2267 		 * check for it.
2268 		 */
2269 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2270 		    (flags & V_SAVE)) {
2271 			bremfree(bp);
2272 			bp->b_flags |= B_ASYNC;
2273 			bwrite(bp);
2274 			BO_LOCK(bo);
2275 			return (EAGAIN);	/* XXX: why not loop ? */
2276 		}
2277 		bremfree(bp);
2278 		bp->b_flags |= (B_INVAL | B_RELBUF);
2279 		bp->b_flags &= ~B_ASYNC;
2280 		brelse(bp);
2281 		BO_LOCK(bo);
2282 		if (nbp == NULL)
2283 			break;
2284 		nbp = gbincore(bo, lblkno);
2285 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2286 		    != xflags)
2287 			break;			/* nbp invalid */
2288 	}
2289 	return (retval);
2290 }
2291 
2292 int
2293 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2294 {
2295 	struct buf *bp;
2296 	int error;
2297 	daddr_t lblkno;
2298 
2299 	ASSERT_BO_LOCKED(bo);
2300 
2301 	for (lblkno = startn;;) {
2302 again:
2303 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2304 		if (bp == NULL || bp->b_lblkno >= endn ||
2305 		    bp->b_lblkno < startn)
2306 			break;
2307 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2308 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2309 		if (error != 0) {
2310 			BO_RLOCK(bo);
2311 			if (error == ENOLCK)
2312 				goto again;
2313 			return (error);
2314 		}
2315 		KASSERT(bp->b_bufobj == bo,
2316 		    ("bp %p wrong b_bufobj %p should be %p",
2317 		    bp, bp->b_bufobj, bo));
2318 		lblkno = bp->b_lblkno + 1;
2319 		if ((bp->b_flags & B_MANAGED) == 0)
2320 			bremfree(bp);
2321 		bp->b_flags |= B_RELBUF;
2322 		/*
2323 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2324 		 * pages backing each buffer in the range are unlikely to be
2325 		 * reused.  Dirty buffers will have the hint applied once
2326 		 * they've been written.
2327 		 */
2328 		if ((bp->b_flags & B_VMIO) != 0)
2329 			bp->b_flags |= B_NOREUSE;
2330 		brelse(bp);
2331 		BO_RLOCK(bo);
2332 	}
2333 	return (0);
2334 }
2335 
2336 /*
2337  * Truncate a file's buffer and pages to a specified length.  This
2338  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2339  * sync activity.
2340  */
2341 int
2342 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2343 {
2344 	struct buf *bp, *nbp;
2345 	struct bufobj *bo;
2346 	daddr_t startlbn;
2347 
2348 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2349 	    vp, blksize, (uintmax_t)length);
2350 
2351 	/*
2352 	 * Round up to the *next* lbn.
2353 	 */
2354 	startlbn = howmany(length, blksize);
2355 
2356 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2357 
2358 	bo = &vp->v_bufobj;
2359 restart_unlocked:
2360 	BO_LOCK(bo);
2361 
2362 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2363 		;
2364 
2365 	if (length > 0) {
2366 restartsync:
2367 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2368 			if (bp->b_lblkno > 0)
2369 				continue;
2370 			/*
2371 			 * Since we hold the vnode lock this should only
2372 			 * fail if we're racing with the buf daemon.
2373 			 */
2374 			if (BUF_LOCK(bp,
2375 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2376 			    BO_LOCKPTR(bo)) == ENOLCK)
2377 				goto restart_unlocked;
2378 
2379 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2380 			    ("buf(%p) on dirty queue without DELWRI", bp));
2381 
2382 			bremfree(bp);
2383 			bawrite(bp);
2384 			BO_LOCK(bo);
2385 			goto restartsync;
2386 		}
2387 	}
2388 
2389 	bufobj_wwait(bo, 0, 0);
2390 	BO_UNLOCK(bo);
2391 	vnode_pager_setsize(vp, length);
2392 
2393 	return (0);
2394 }
2395 
2396 /*
2397  * Invalidate the cached pages of a file's buffer within the range of block
2398  * numbers [startlbn, endlbn).
2399  */
2400 void
2401 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2402     int blksize)
2403 {
2404 	struct bufobj *bo;
2405 	off_t start, end;
2406 
2407 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2408 
2409 	start = blksize * startlbn;
2410 	end = blksize * endlbn;
2411 
2412 	bo = &vp->v_bufobj;
2413 	BO_LOCK(bo);
2414 	MPASS(blksize == bo->bo_bsize);
2415 
2416 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2417 		;
2418 
2419 	BO_UNLOCK(bo);
2420 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2421 }
2422 
2423 static int
2424 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2425     daddr_t startlbn, daddr_t endlbn)
2426 {
2427 	struct buf *bp, *nbp;
2428 	bool anyfreed;
2429 
2430 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2431 	ASSERT_BO_LOCKED(bo);
2432 
2433 	do {
2434 		anyfreed = false;
2435 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2436 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2437 				continue;
2438 			if (BUF_LOCK(bp,
2439 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2440 			    BO_LOCKPTR(bo)) == ENOLCK) {
2441 				BO_LOCK(bo);
2442 				return (EAGAIN);
2443 			}
2444 
2445 			bremfree(bp);
2446 			bp->b_flags |= B_INVAL | B_RELBUF;
2447 			bp->b_flags &= ~B_ASYNC;
2448 			brelse(bp);
2449 			anyfreed = true;
2450 
2451 			BO_LOCK(bo);
2452 			if (nbp != NULL &&
2453 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2454 			    nbp->b_vp != vp ||
2455 			    (nbp->b_flags & B_DELWRI) != 0))
2456 				return (EAGAIN);
2457 		}
2458 
2459 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2460 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2461 				continue;
2462 			if (BUF_LOCK(bp,
2463 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2464 			    BO_LOCKPTR(bo)) == ENOLCK) {
2465 				BO_LOCK(bo);
2466 				return (EAGAIN);
2467 			}
2468 			bremfree(bp);
2469 			bp->b_flags |= B_INVAL | B_RELBUF;
2470 			bp->b_flags &= ~B_ASYNC;
2471 			brelse(bp);
2472 			anyfreed = true;
2473 
2474 			BO_LOCK(bo);
2475 			if (nbp != NULL &&
2476 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2477 			    (nbp->b_vp != vp) ||
2478 			    (nbp->b_flags & B_DELWRI) == 0))
2479 				return (EAGAIN);
2480 		}
2481 	} while (anyfreed);
2482 	return (0);
2483 }
2484 
2485 static void
2486 buf_vlist_remove(struct buf *bp)
2487 {
2488 	struct bufv *bv;
2489 	b_xflags_t flags;
2490 
2491 	flags = bp->b_xflags;
2492 
2493 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2494 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2495 	KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 &&
2496 	    (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN),
2497 	    ("%s: buffer %p has invalid queue state", __func__, bp));
2498 
2499 	if ((flags & BX_VNDIRTY) != 0)
2500 		bv = &bp->b_bufobj->bo_dirty;
2501 	else
2502 		bv = &bp->b_bufobj->bo_clean;
2503 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2504 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2505 	bv->bv_cnt--;
2506 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2507 }
2508 
2509 /*
2510  * Add the buffer to the sorted clean or dirty block list.
2511  *
2512  * NOTE: xflags is passed as a constant, optimizing this inline function!
2513  */
2514 static void
2515 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2516 {
2517 	struct bufv *bv;
2518 	struct buf *n;
2519 	int error;
2520 
2521 	ASSERT_BO_WLOCKED(bo);
2522 	KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2523 	    ("buf_vlist_add: bo %p does not allow bufs", bo));
2524 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2525 	    ("dead bo %p", bo));
2526 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2527 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2528 	bp->b_xflags |= xflags;
2529 	if (xflags & BX_VNDIRTY)
2530 		bv = &bo->bo_dirty;
2531 	else
2532 		bv = &bo->bo_clean;
2533 
2534 	/*
2535 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2536 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2537 	 * than _ge.
2538 	 */
2539 	if (bv->bv_cnt == 0 ||
2540 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2541 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2542 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2543 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2544 	else
2545 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2546 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2547 	if (error)
2548 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2549 	bv->bv_cnt++;
2550 }
2551 
2552 /*
2553  * Look up a buffer using the buffer tries.
2554  */
2555 struct buf *
2556 gbincore(struct bufobj *bo, daddr_t lblkno)
2557 {
2558 	struct buf *bp;
2559 
2560 	ASSERT_BO_LOCKED(bo);
2561 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2562 	if (bp != NULL)
2563 		return (bp);
2564 	return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno));
2565 }
2566 
2567 /*
2568  * Look up a buf using the buffer tries, without the bufobj lock.  This relies
2569  * on SMR for safe lookup, and bufs being in a no-free zone to provide type
2570  * stability of the result.  Like other lockless lookups, the found buf may
2571  * already be invalid by the time this function returns.
2572  */
2573 struct buf *
2574 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno)
2575 {
2576 	struct buf *bp;
2577 
2578 	ASSERT_BO_UNLOCKED(bo);
2579 	bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno);
2580 	if (bp != NULL)
2581 		return (bp);
2582 	return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno));
2583 }
2584 
2585 /*
2586  * Associate a buffer with a vnode.
2587  */
2588 void
2589 bgetvp(struct vnode *vp, struct buf *bp)
2590 {
2591 	struct bufobj *bo;
2592 
2593 	bo = &vp->v_bufobj;
2594 	ASSERT_BO_WLOCKED(bo);
2595 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2596 
2597 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2598 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2599 	    ("bgetvp: bp already attached! %p", bp));
2600 
2601 	vhold(vp);
2602 	bp->b_vp = vp;
2603 	bp->b_bufobj = bo;
2604 	/*
2605 	 * Insert onto list for new vnode.
2606 	 */
2607 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2608 }
2609 
2610 /*
2611  * Disassociate a buffer from a vnode.
2612  */
2613 void
2614 brelvp(struct buf *bp)
2615 {
2616 	struct bufobj *bo;
2617 	struct vnode *vp;
2618 
2619 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2620 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2621 
2622 	/*
2623 	 * Delete from old vnode list, if on one.
2624 	 */
2625 	vp = bp->b_vp;		/* XXX */
2626 	bo = bp->b_bufobj;
2627 	BO_LOCK(bo);
2628 	buf_vlist_remove(bp);
2629 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2630 		bo->bo_flag &= ~BO_ONWORKLST;
2631 		mtx_lock(&sync_mtx);
2632 		LIST_REMOVE(bo, bo_synclist);
2633 		syncer_worklist_len--;
2634 		mtx_unlock(&sync_mtx);
2635 	}
2636 	bp->b_vp = NULL;
2637 	bp->b_bufobj = NULL;
2638 	BO_UNLOCK(bo);
2639 	vdrop(vp);
2640 }
2641 
2642 /*
2643  * Add an item to the syncer work queue.
2644  */
2645 static void
2646 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2647 {
2648 	int slot;
2649 
2650 	ASSERT_BO_WLOCKED(bo);
2651 
2652 	mtx_lock(&sync_mtx);
2653 	if (bo->bo_flag & BO_ONWORKLST)
2654 		LIST_REMOVE(bo, bo_synclist);
2655 	else {
2656 		bo->bo_flag |= BO_ONWORKLST;
2657 		syncer_worklist_len++;
2658 	}
2659 
2660 	if (delay > syncer_maxdelay - 2)
2661 		delay = syncer_maxdelay - 2;
2662 	slot = (syncer_delayno + delay) & syncer_mask;
2663 
2664 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2665 	mtx_unlock(&sync_mtx);
2666 }
2667 
2668 static int
2669 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2670 {
2671 	int error, len;
2672 
2673 	mtx_lock(&sync_mtx);
2674 	len = syncer_worklist_len - sync_vnode_count;
2675 	mtx_unlock(&sync_mtx);
2676 	error = SYSCTL_OUT(req, &len, sizeof(len));
2677 	return (error);
2678 }
2679 
2680 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2681     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2682     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2683 
2684 static struct proc *updateproc;
2685 static void sched_sync(void);
2686 static struct kproc_desc up_kp = {
2687 	"syncer",
2688 	sched_sync,
2689 	&updateproc
2690 };
2691 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2692 
2693 static int
2694 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2695 {
2696 	struct vnode *vp;
2697 	struct mount *mp;
2698 
2699 	*bo = LIST_FIRST(slp);
2700 	if (*bo == NULL)
2701 		return (0);
2702 	vp = bo2vnode(*bo);
2703 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2704 		return (1);
2705 	/*
2706 	 * We use vhold in case the vnode does not
2707 	 * successfully sync.  vhold prevents the vnode from
2708 	 * going away when we unlock the sync_mtx so that
2709 	 * we can acquire the vnode interlock.
2710 	 */
2711 	vholdl(vp);
2712 	mtx_unlock(&sync_mtx);
2713 	VI_UNLOCK(vp);
2714 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2715 		vdrop(vp);
2716 		mtx_lock(&sync_mtx);
2717 		return (*bo == LIST_FIRST(slp));
2718 	}
2719 	MPASSERT(mp == NULL || (curthread->td_pflags & TDP_IGNSUSP) != 0 ||
2720 	    (mp->mnt_kern_flag & MNTK_SUSPENDED) == 0, mp,
2721 	    ("suspended mp syncing vp %p", vp));
2722 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2723 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2724 	VOP_UNLOCK(vp);
2725 	vn_finished_write(mp);
2726 	BO_LOCK(*bo);
2727 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2728 		/*
2729 		 * Put us back on the worklist.  The worklist
2730 		 * routine will remove us from our current
2731 		 * position and then add us back in at a later
2732 		 * position.
2733 		 */
2734 		vn_syncer_add_to_worklist(*bo, syncdelay);
2735 	}
2736 	BO_UNLOCK(*bo);
2737 	vdrop(vp);
2738 	mtx_lock(&sync_mtx);
2739 	return (0);
2740 }
2741 
2742 static int first_printf = 1;
2743 
2744 /*
2745  * System filesystem synchronizer daemon.
2746  */
2747 static void
2748 sched_sync(void)
2749 {
2750 	struct synclist *next, *slp;
2751 	struct bufobj *bo;
2752 	long starttime;
2753 	struct thread *td = curthread;
2754 	int last_work_seen;
2755 	int net_worklist_len;
2756 	int syncer_final_iter;
2757 	int error;
2758 
2759 	last_work_seen = 0;
2760 	syncer_final_iter = 0;
2761 	syncer_state = SYNCER_RUNNING;
2762 	starttime = time_uptime;
2763 	td->td_pflags |= TDP_NORUNNINGBUF;
2764 
2765 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2766 	    SHUTDOWN_PRI_LAST);
2767 
2768 	mtx_lock(&sync_mtx);
2769 	for (;;) {
2770 		if (syncer_state == SYNCER_FINAL_DELAY &&
2771 		    syncer_final_iter == 0) {
2772 			mtx_unlock(&sync_mtx);
2773 			kproc_suspend_check(td->td_proc);
2774 			mtx_lock(&sync_mtx);
2775 		}
2776 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2777 		if (syncer_state != SYNCER_RUNNING &&
2778 		    starttime != time_uptime) {
2779 			if (first_printf) {
2780 				printf("\nSyncing disks, vnodes remaining... ");
2781 				first_printf = 0;
2782 			}
2783 			printf("%d ", net_worklist_len);
2784 		}
2785 		starttime = time_uptime;
2786 
2787 		/*
2788 		 * Push files whose dirty time has expired.  Be careful
2789 		 * of interrupt race on slp queue.
2790 		 *
2791 		 * Skip over empty worklist slots when shutting down.
2792 		 */
2793 		do {
2794 			slp = &syncer_workitem_pending[syncer_delayno];
2795 			syncer_delayno += 1;
2796 			if (syncer_delayno == syncer_maxdelay)
2797 				syncer_delayno = 0;
2798 			next = &syncer_workitem_pending[syncer_delayno];
2799 			/*
2800 			 * If the worklist has wrapped since the
2801 			 * it was emptied of all but syncer vnodes,
2802 			 * switch to the FINAL_DELAY state and run
2803 			 * for one more second.
2804 			 */
2805 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2806 			    net_worklist_len == 0 &&
2807 			    last_work_seen == syncer_delayno) {
2808 				syncer_state = SYNCER_FINAL_DELAY;
2809 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2810 			}
2811 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2812 		    syncer_worklist_len > 0);
2813 
2814 		/*
2815 		 * Keep track of the last time there was anything
2816 		 * on the worklist other than syncer vnodes.
2817 		 * Return to the SHUTTING_DOWN state if any
2818 		 * new work appears.
2819 		 */
2820 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2821 			last_work_seen = syncer_delayno;
2822 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2823 			syncer_state = SYNCER_SHUTTING_DOWN;
2824 		while (!LIST_EMPTY(slp)) {
2825 			error = sync_vnode(slp, &bo, td);
2826 			if (error == 1) {
2827 				LIST_REMOVE(bo, bo_synclist);
2828 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2829 				continue;
2830 			}
2831 
2832 			if (first_printf == 0) {
2833 				/*
2834 				 * Drop the sync mutex, because some watchdog
2835 				 * drivers need to sleep while patting
2836 				 */
2837 				mtx_unlock(&sync_mtx);
2838 				wdog_kern_pat(WD_LASTVAL);
2839 				mtx_lock(&sync_mtx);
2840 			}
2841 		}
2842 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2843 			syncer_final_iter--;
2844 		/*
2845 		 * The variable rushjob allows the kernel to speed up the
2846 		 * processing of the filesystem syncer process. A rushjob
2847 		 * value of N tells the filesystem syncer to process the next
2848 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2849 		 * is used by the soft update code to speed up the filesystem
2850 		 * syncer process when the incore state is getting so far
2851 		 * ahead of the disk that the kernel memory pool is being
2852 		 * threatened with exhaustion.
2853 		 */
2854 		if (rushjob > 0) {
2855 			rushjob -= 1;
2856 			continue;
2857 		}
2858 		/*
2859 		 * Just sleep for a short period of time between
2860 		 * iterations when shutting down to allow some I/O
2861 		 * to happen.
2862 		 *
2863 		 * If it has taken us less than a second to process the
2864 		 * current work, then wait. Otherwise start right over
2865 		 * again. We can still lose time if any single round
2866 		 * takes more than two seconds, but it does not really
2867 		 * matter as we are just trying to generally pace the
2868 		 * filesystem activity.
2869 		 */
2870 		if (syncer_state != SYNCER_RUNNING ||
2871 		    time_uptime == starttime) {
2872 			thread_lock(td);
2873 			sched_prio(td, PPAUSE);
2874 			thread_unlock(td);
2875 		}
2876 		if (syncer_state != SYNCER_RUNNING)
2877 			cv_timedwait(&sync_wakeup, &sync_mtx,
2878 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2879 		else if (time_uptime == starttime)
2880 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2881 	}
2882 }
2883 
2884 /*
2885  * Request the syncer daemon to speed up its work.
2886  * We never push it to speed up more than half of its
2887  * normal turn time, otherwise it could take over the cpu.
2888  */
2889 int
2890 speedup_syncer(void)
2891 {
2892 	int ret = 0;
2893 
2894 	mtx_lock(&sync_mtx);
2895 	if (rushjob < syncdelay / 2) {
2896 		rushjob += 1;
2897 		stat_rush_requests += 1;
2898 		ret = 1;
2899 	}
2900 	mtx_unlock(&sync_mtx);
2901 	cv_broadcast(&sync_wakeup);
2902 	return (ret);
2903 }
2904 
2905 /*
2906  * Tell the syncer to speed up its work and run though its work
2907  * list several times, then tell it to shut down.
2908  */
2909 static void
2910 syncer_shutdown(void *arg, int howto)
2911 {
2912 
2913 	if (howto & RB_NOSYNC)
2914 		return;
2915 	mtx_lock(&sync_mtx);
2916 	syncer_state = SYNCER_SHUTTING_DOWN;
2917 	rushjob = 0;
2918 	mtx_unlock(&sync_mtx);
2919 	cv_broadcast(&sync_wakeup);
2920 	kproc_shutdown(arg, howto);
2921 }
2922 
2923 void
2924 syncer_suspend(void)
2925 {
2926 
2927 	syncer_shutdown(updateproc, 0);
2928 }
2929 
2930 void
2931 syncer_resume(void)
2932 {
2933 
2934 	mtx_lock(&sync_mtx);
2935 	first_printf = 1;
2936 	syncer_state = SYNCER_RUNNING;
2937 	mtx_unlock(&sync_mtx);
2938 	cv_broadcast(&sync_wakeup);
2939 	kproc_resume(updateproc);
2940 }
2941 
2942 /*
2943  * Move the buffer between the clean and dirty lists of its vnode.
2944  */
2945 void
2946 reassignbuf(struct buf *bp)
2947 {
2948 	struct vnode *vp;
2949 	struct bufobj *bo;
2950 	int delay;
2951 #ifdef INVARIANTS
2952 	struct bufv *bv;
2953 #endif
2954 
2955 	vp = bp->b_vp;
2956 	bo = bp->b_bufobj;
2957 
2958 	KASSERT((bp->b_flags & B_PAGING) == 0,
2959 	    ("%s: cannot reassign paging buffer %p", __func__, bp));
2960 
2961 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2962 	    bp, bp->b_vp, bp->b_flags);
2963 
2964 	BO_LOCK(bo);
2965 	buf_vlist_remove(bp);
2966 
2967 	/*
2968 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2969 	 * of clean buffers.
2970 	 */
2971 	if (bp->b_flags & B_DELWRI) {
2972 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2973 			switch (vp->v_type) {
2974 			case VDIR:
2975 				delay = dirdelay;
2976 				break;
2977 			case VCHR:
2978 				delay = metadelay;
2979 				break;
2980 			default:
2981 				delay = filedelay;
2982 			}
2983 			vn_syncer_add_to_worklist(bo, delay);
2984 		}
2985 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2986 	} else {
2987 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2988 
2989 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2990 			mtx_lock(&sync_mtx);
2991 			LIST_REMOVE(bo, bo_synclist);
2992 			syncer_worklist_len--;
2993 			mtx_unlock(&sync_mtx);
2994 			bo->bo_flag &= ~BO_ONWORKLST;
2995 		}
2996 	}
2997 #ifdef INVARIANTS
2998 	bv = &bo->bo_clean;
2999 	bp = TAILQ_FIRST(&bv->bv_hd);
3000 	KASSERT(bp == NULL || bp->b_bufobj == bo,
3001 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3002 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
3003 	KASSERT(bp == NULL || bp->b_bufobj == bo,
3004 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3005 	bv = &bo->bo_dirty;
3006 	bp = TAILQ_FIRST(&bv->bv_hd);
3007 	KASSERT(bp == NULL || bp->b_bufobj == bo,
3008 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3009 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
3010 	KASSERT(bp == NULL || bp->b_bufobj == bo,
3011 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3012 #endif
3013 	BO_UNLOCK(bo);
3014 }
3015 
3016 static void
3017 v_init_counters(struct vnode *vp)
3018 {
3019 
3020 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
3021 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
3022 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
3023 
3024 	refcount_init(&vp->v_holdcnt, 1);
3025 	refcount_init(&vp->v_usecount, 1);
3026 }
3027 
3028 /*
3029  * Grab a particular vnode from the free list, increment its
3030  * reference count and lock it.  VIRF_DOOMED is set if the vnode
3031  * is being destroyed.  Only callers who specify LK_RETRY will
3032  * see doomed vnodes.  If inactive processing was delayed in
3033  * vput try to do it here.
3034  *
3035  * usecount is manipulated using atomics without holding any locks.
3036  *
3037  * holdcnt can be manipulated using atomics without holding any locks,
3038  * except when transitioning 1<->0, in which case the interlock is held.
3039  *
3040  * Consumers which don't guarantee liveness of the vnode can use SMR to
3041  * try to get a reference. Note this operation can fail since the vnode
3042  * may be awaiting getting freed by the time they get to it.
3043  */
3044 enum vgetstate
3045 vget_prep_smr(struct vnode *vp)
3046 {
3047 	enum vgetstate vs;
3048 
3049 	VFS_SMR_ASSERT_ENTERED();
3050 
3051 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
3052 		vs = VGET_USECOUNT;
3053 	} else {
3054 		if (vhold_smr(vp))
3055 			vs = VGET_HOLDCNT;
3056 		else
3057 			vs = VGET_NONE;
3058 	}
3059 	return (vs);
3060 }
3061 
3062 enum vgetstate
3063 vget_prep(struct vnode *vp)
3064 {
3065 	enum vgetstate vs;
3066 
3067 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
3068 		vs = VGET_USECOUNT;
3069 	} else {
3070 		vhold(vp);
3071 		vs = VGET_HOLDCNT;
3072 	}
3073 	return (vs);
3074 }
3075 
3076 void
3077 vget_abort(struct vnode *vp, enum vgetstate vs)
3078 {
3079 
3080 	switch (vs) {
3081 	case VGET_USECOUNT:
3082 		vrele(vp);
3083 		break;
3084 	case VGET_HOLDCNT:
3085 		vdrop(vp);
3086 		break;
3087 	default:
3088 		__assert_unreachable();
3089 	}
3090 }
3091 
3092 int
3093 vget(struct vnode *vp, int flags)
3094 {
3095 	enum vgetstate vs;
3096 
3097 	vs = vget_prep(vp);
3098 	return (vget_finish(vp, flags, vs));
3099 }
3100 
3101 int
3102 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
3103 {
3104 	int error;
3105 
3106 	if ((flags & LK_INTERLOCK) != 0)
3107 		ASSERT_VI_LOCKED(vp, __func__);
3108 	else
3109 		ASSERT_VI_UNLOCKED(vp, __func__);
3110 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3111 	VNPASS(vp->v_holdcnt > 0, vp);
3112 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3113 
3114 	error = vn_lock(vp, flags);
3115 	if (__predict_false(error != 0)) {
3116 		vget_abort(vp, vs);
3117 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
3118 		    vp);
3119 		return (error);
3120 	}
3121 
3122 	vget_finish_ref(vp, vs);
3123 	return (0);
3124 }
3125 
3126 void
3127 vget_finish_ref(struct vnode *vp, enum vgetstate vs)
3128 {
3129 	int old;
3130 
3131 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3132 	VNPASS(vp->v_holdcnt > 0, vp);
3133 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3134 
3135 	if (vs == VGET_USECOUNT)
3136 		return;
3137 
3138 	/*
3139 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
3140 	 * the vnode around. Otherwise someone else lended their hold count and
3141 	 * we have to drop ours.
3142 	 */
3143 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
3144 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3145 	if (old != 0) {
3146 #ifdef INVARIANTS
3147 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3148 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3149 #else
3150 		refcount_release(&vp->v_holdcnt);
3151 #endif
3152 	}
3153 }
3154 
3155 void
3156 vref(struct vnode *vp)
3157 {
3158 	enum vgetstate vs;
3159 
3160 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3161 	vs = vget_prep(vp);
3162 	vget_finish_ref(vp, vs);
3163 }
3164 
3165 void
3166 vrefact(struct vnode *vp)
3167 {
3168 
3169 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3170 #ifdef INVARIANTS
3171 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3172 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3173 #else
3174 	refcount_acquire(&vp->v_usecount);
3175 #endif
3176 }
3177 
3178 void
3179 vlazy(struct vnode *vp)
3180 {
3181 	struct mount *mp;
3182 
3183 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3184 
3185 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3186 		return;
3187 	/*
3188 	 * We may get here for inactive routines after the vnode got doomed.
3189 	 */
3190 	if (VN_IS_DOOMED(vp))
3191 		return;
3192 	mp = vp->v_mount;
3193 	mtx_lock(&mp->mnt_listmtx);
3194 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3195 		vp->v_mflag |= VMP_LAZYLIST;
3196 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3197 		mp->mnt_lazyvnodelistsize++;
3198 	}
3199 	mtx_unlock(&mp->mnt_listmtx);
3200 }
3201 
3202 static void
3203 vunlazy(struct vnode *vp)
3204 {
3205 	struct mount *mp;
3206 
3207 	ASSERT_VI_LOCKED(vp, __func__);
3208 	VNPASS(!VN_IS_DOOMED(vp), vp);
3209 
3210 	mp = vp->v_mount;
3211 	mtx_lock(&mp->mnt_listmtx);
3212 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3213 	/*
3214 	 * Don't remove the vnode from the lazy list if another thread
3215 	 * has increased the hold count. It may have re-enqueued the
3216 	 * vnode to the lazy list and is now responsible for its
3217 	 * removal.
3218 	 */
3219 	if (vp->v_holdcnt == 0) {
3220 		vp->v_mflag &= ~VMP_LAZYLIST;
3221 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3222 		mp->mnt_lazyvnodelistsize--;
3223 	}
3224 	mtx_unlock(&mp->mnt_listmtx);
3225 }
3226 
3227 /*
3228  * This routine is only meant to be called from vgonel prior to dooming
3229  * the vnode.
3230  */
3231 static void
3232 vunlazy_gone(struct vnode *vp)
3233 {
3234 	struct mount *mp;
3235 
3236 	ASSERT_VOP_ELOCKED(vp, __func__);
3237 	ASSERT_VI_LOCKED(vp, __func__);
3238 	VNPASS(!VN_IS_DOOMED(vp), vp);
3239 
3240 	if (vp->v_mflag & VMP_LAZYLIST) {
3241 		mp = vp->v_mount;
3242 		mtx_lock(&mp->mnt_listmtx);
3243 		VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3244 		vp->v_mflag &= ~VMP_LAZYLIST;
3245 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3246 		mp->mnt_lazyvnodelistsize--;
3247 		mtx_unlock(&mp->mnt_listmtx);
3248 	}
3249 }
3250 
3251 static void
3252 vdefer_inactive(struct vnode *vp)
3253 {
3254 
3255 	ASSERT_VI_LOCKED(vp, __func__);
3256 	VNPASS(vp->v_holdcnt > 0, vp);
3257 	if (VN_IS_DOOMED(vp)) {
3258 		vdropl(vp);
3259 		return;
3260 	}
3261 	if (vp->v_iflag & VI_DEFINACT) {
3262 		VNPASS(vp->v_holdcnt > 1, vp);
3263 		vdropl(vp);
3264 		return;
3265 	}
3266 	if (vp->v_usecount > 0) {
3267 		vp->v_iflag &= ~VI_OWEINACT;
3268 		vdropl(vp);
3269 		return;
3270 	}
3271 	vlazy(vp);
3272 	vp->v_iflag |= VI_DEFINACT;
3273 	VI_UNLOCK(vp);
3274 	atomic_add_long(&deferred_inact, 1);
3275 }
3276 
3277 static void
3278 vdefer_inactive_unlocked(struct vnode *vp)
3279 {
3280 
3281 	VI_LOCK(vp);
3282 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3283 		vdropl(vp);
3284 		return;
3285 	}
3286 	vdefer_inactive(vp);
3287 }
3288 
3289 enum vput_op { VRELE, VPUT, VUNREF };
3290 
3291 /*
3292  * Handle ->v_usecount transitioning to 0.
3293  *
3294  * By releasing the last usecount we take ownership of the hold count which
3295  * provides liveness of the vnode, meaning we have to vdrop.
3296  *
3297  * For all vnodes we may need to perform inactive processing. It requires an
3298  * exclusive lock on the vnode, while it is legal to call here with only a
3299  * shared lock (or no locks). If locking the vnode in an expected manner fails,
3300  * inactive processing gets deferred to the syncer.
3301  *
3302  * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3303  * on the lock being held all the way until VOP_INACTIVE. This in particular
3304  * happens with UFS which adds half-constructed vnodes to the hash, where they
3305  * can be found by other code.
3306  */
3307 static void
3308 vput_final(struct vnode *vp, enum vput_op func)
3309 {
3310 	int error;
3311 	bool want_unlock;
3312 
3313 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3314 	VNPASS(vp->v_holdcnt > 0, vp);
3315 
3316 	VI_LOCK(vp);
3317 
3318 	/*
3319 	 * By the time we got here someone else might have transitioned
3320 	 * the count back to > 0.
3321 	 */
3322 	if (vp->v_usecount > 0)
3323 		goto out;
3324 
3325 	/*
3326 	 * If the vnode is doomed vgone already performed inactive processing
3327 	 * (if needed).
3328 	 */
3329 	if (VN_IS_DOOMED(vp))
3330 		goto out;
3331 
3332 	if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3333 		goto out;
3334 
3335 	if (vp->v_iflag & VI_DOINGINACT)
3336 		goto out;
3337 
3338 	/*
3339 	 * Locking operations here will drop the interlock and possibly the
3340 	 * vnode lock, opening a window where the vnode can get doomed all the
3341 	 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3342 	 * perform inactive.
3343 	 */
3344 	vp->v_iflag |= VI_OWEINACT;
3345 	want_unlock = false;
3346 	error = 0;
3347 	switch (func) {
3348 	case VRELE:
3349 		switch (VOP_ISLOCKED(vp)) {
3350 		case LK_EXCLUSIVE:
3351 			break;
3352 		case LK_EXCLOTHER:
3353 		case 0:
3354 			want_unlock = true;
3355 			error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3356 			VI_LOCK(vp);
3357 			break;
3358 		default:
3359 			/*
3360 			 * The lock has at least one sharer, but we have no way
3361 			 * to conclude whether this is us. Play it safe and
3362 			 * defer processing.
3363 			 */
3364 			error = EAGAIN;
3365 			break;
3366 		}
3367 		break;
3368 	case VPUT:
3369 		want_unlock = true;
3370 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3371 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3372 			    LK_NOWAIT);
3373 			VI_LOCK(vp);
3374 		}
3375 		break;
3376 	case VUNREF:
3377 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3378 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3379 			VI_LOCK(vp);
3380 		}
3381 		break;
3382 	}
3383 	if (error == 0) {
3384 		if (func == VUNREF) {
3385 			VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp,
3386 			    ("recursive vunref"));
3387 			vp->v_vflag |= VV_UNREF;
3388 		}
3389 		for (;;) {
3390 			error = vinactive(vp);
3391 			if (want_unlock)
3392 				VOP_UNLOCK(vp);
3393 			if (error != ERELOOKUP || !want_unlock)
3394 				break;
3395 			VOP_LOCK(vp, LK_EXCLUSIVE);
3396 		}
3397 		if (func == VUNREF)
3398 			vp->v_vflag &= ~VV_UNREF;
3399 		vdropl(vp);
3400 	} else {
3401 		vdefer_inactive(vp);
3402 	}
3403 	return;
3404 out:
3405 	if (func == VPUT)
3406 		VOP_UNLOCK(vp);
3407 	vdropl(vp);
3408 }
3409 
3410 /*
3411  * Decrement ->v_usecount for a vnode.
3412  *
3413  * Releasing the last use count requires additional processing, see vput_final
3414  * above for details.
3415  *
3416  * Comment above each variant denotes lock state on entry and exit.
3417  */
3418 
3419 /*
3420  * in: any
3421  * out: same as passed in
3422  */
3423 void
3424 vrele(struct vnode *vp)
3425 {
3426 
3427 	ASSERT_VI_UNLOCKED(vp, __func__);
3428 	if (!refcount_release(&vp->v_usecount))
3429 		return;
3430 	vput_final(vp, VRELE);
3431 }
3432 
3433 /*
3434  * in: locked
3435  * out: unlocked
3436  */
3437 void
3438 vput(struct vnode *vp)
3439 {
3440 
3441 	ASSERT_VOP_LOCKED(vp, __func__);
3442 	ASSERT_VI_UNLOCKED(vp, __func__);
3443 	if (!refcount_release(&vp->v_usecount)) {
3444 		VOP_UNLOCK(vp);
3445 		return;
3446 	}
3447 	vput_final(vp, VPUT);
3448 }
3449 
3450 /*
3451  * in: locked
3452  * out: locked
3453  */
3454 void
3455 vunref(struct vnode *vp)
3456 {
3457 
3458 	ASSERT_VOP_LOCKED(vp, __func__);
3459 	ASSERT_VI_UNLOCKED(vp, __func__);
3460 	if (!refcount_release(&vp->v_usecount))
3461 		return;
3462 	vput_final(vp, VUNREF);
3463 }
3464 
3465 void
3466 vhold(struct vnode *vp)
3467 {
3468 	int old;
3469 
3470 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3471 	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3472 	VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3473 	    ("%s: wrong hold count %d", __func__, old));
3474 	if (old == 0)
3475 		vfs_freevnodes_dec();
3476 }
3477 
3478 void
3479 vholdnz(struct vnode *vp)
3480 {
3481 
3482 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3483 #ifdef INVARIANTS
3484 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3485 	VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3486 	    ("%s: wrong hold count %d", __func__, old));
3487 #else
3488 	atomic_add_int(&vp->v_holdcnt, 1);
3489 #endif
3490 }
3491 
3492 /*
3493  * Grab a hold count unless the vnode is freed.
3494  *
3495  * Only use this routine if vfs smr is the only protection you have against
3496  * freeing the vnode.
3497  *
3498  * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag
3499  * is not set.  After the flag is set the vnode becomes immutable to anyone but
3500  * the thread which managed to set the flag.
3501  *
3502  * It may be tempting to replace the loop with:
3503  * count = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3504  * if (count & VHOLD_NO_SMR) {
3505  *     backpedal and error out;
3506  * }
3507  *
3508  * However, while this is more performant, it hinders debugging by eliminating
3509  * the previously mentioned invariant.
3510  */
3511 bool
3512 vhold_smr(struct vnode *vp)
3513 {
3514 	int count;
3515 
3516 	VFS_SMR_ASSERT_ENTERED();
3517 
3518 	count = atomic_load_int(&vp->v_holdcnt);
3519 	for (;;) {
3520 		if (count & VHOLD_NO_SMR) {
3521 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3522 			    ("non-zero hold count with flags %d\n", count));
3523 			return (false);
3524 		}
3525 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3526 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3527 			if (count == 0)
3528 				vfs_freevnodes_dec();
3529 			return (true);
3530 		}
3531 	}
3532 }
3533 
3534 /*
3535  * Hold a free vnode for recycling.
3536  *
3537  * Note: vnode_init references this comment.
3538  *
3539  * Attempts to recycle only need the global vnode list lock and have no use for
3540  * SMR.
3541  *
3542  * However, vnodes get inserted into the global list before they get fully
3543  * initialized and stay there until UMA decides to free the memory. This in
3544  * particular means the target can be found before it becomes usable and after
3545  * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to
3546  * VHOLD_NO_SMR.
3547  *
3548  * Note: the vnode may gain more references after we transition the count 0->1.
3549  */
3550 static bool
3551 vhold_recycle_free(struct vnode *vp)
3552 {
3553 	int count;
3554 
3555 	mtx_assert(&vnode_list_mtx, MA_OWNED);
3556 
3557 	count = atomic_load_int(&vp->v_holdcnt);
3558 	for (;;) {
3559 		if (count & VHOLD_NO_SMR) {
3560 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3561 			    ("non-zero hold count with flags %d\n", count));
3562 			return (false);
3563 		}
3564 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3565 		if (count > 0) {
3566 			return (false);
3567 		}
3568 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3569 			vfs_freevnodes_dec();
3570 			return (true);
3571 		}
3572 	}
3573 }
3574 
3575 static void __noinline
3576 vdbatch_process(struct vdbatch *vd)
3577 {
3578 	struct vnode *vp;
3579 	int i;
3580 
3581 	mtx_assert(&vd->lock, MA_OWNED);
3582 	MPASS(curthread->td_pinned > 0);
3583 	MPASS(vd->index == VDBATCH_SIZE);
3584 
3585 	/*
3586 	 * Attempt to requeue the passed batch, but give up easily.
3587 	 *
3588 	 * Despite batching the mechanism is prone to transient *significant*
3589 	 * lock contention, where vnode_list_mtx becomes the primary bottleneck
3590 	 * if multiple CPUs get here (one real-world example is highly parallel
3591 	 * do-nothing make , which will stat *tons* of vnodes). Since it is
3592 	 * quasi-LRU (read: not that great even if fully honoured) just dodge
3593 	 * the problem. Parties which don't like it are welcome to implement
3594 	 * something better.
3595 	 */
3596 	critical_enter();
3597 	if (mtx_trylock(&vnode_list_mtx)) {
3598 		for (i = 0; i < VDBATCH_SIZE; i++) {
3599 			vp = vd->tab[i];
3600 			vd->tab[i] = NULL;
3601 			TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3602 			TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3603 			MPASS(vp->v_dbatchcpu != NOCPU);
3604 			vp->v_dbatchcpu = NOCPU;
3605 		}
3606 		mtx_unlock(&vnode_list_mtx);
3607 	} else {
3608 		counter_u64_add(vnode_skipped_requeues, 1);
3609 
3610 		for (i = 0; i < VDBATCH_SIZE; i++) {
3611 			vp = vd->tab[i];
3612 			vd->tab[i] = NULL;
3613 			MPASS(vp->v_dbatchcpu != NOCPU);
3614 			vp->v_dbatchcpu = NOCPU;
3615 		}
3616 	}
3617 	vd->index = 0;
3618 	critical_exit();
3619 }
3620 
3621 static void
3622 vdbatch_enqueue(struct vnode *vp)
3623 {
3624 	struct vdbatch *vd;
3625 
3626 	ASSERT_VI_LOCKED(vp, __func__);
3627 	VNPASS(!VN_IS_DOOMED(vp), vp);
3628 
3629 	if (vp->v_dbatchcpu != NOCPU) {
3630 		VI_UNLOCK(vp);
3631 		return;
3632 	}
3633 
3634 	sched_pin();
3635 	vd = DPCPU_PTR(vd);
3636 	mtx_lock(&vd->lock);
3637 	MPASS(vd->index < VDBATCH_SIZE);
3638 	MPASS(vd->tab[vd->index] == NULL);
3639 	/*
3640 	 * A hack: we depend on being pinned so that we know what to put in
3641 	 * ->v_dbatchcpu.
3642 	 */
3643 	vp->v_dbatchcpu = curcpu;
3644 	vd->tab[vd->index] = vp;
3645 	vd->index++;
3646 	VI_UNLOCK(vp);
3647 	if (vd->index == VDBATCH_SIZE)
3648 		vdbatch_process(vd);
3649 	mtx_unlock(&vd->lock);
3650 	sched_unpin();
3651 }
3652 
3653 /*
3654  * This routine must only be called for vnodes which are about to be
3655  * deallocated. Supporting dequeue for arbitrary vndoes would require
3656  * validating that the locked batch matches.
3657  */
3658 static void
3659 vdbatch_dequeue(struct vnode *vp)
3660 {
3661 	struct vdbatch *vd;
3662 	int i;
3663 	short cpu;
3664 
3665 	VNPASS(vp->v_type == VBAD || vp->v_type == VNON, vp);
3666 
3667 	cpu = vp->v_dbatchcpu;
3668 	if (cpu == NOCPU)
3669 		return;
3670 
3671 	vd = DPCPU_ID_PTR(cpu, vd);
3672 	mtx_lock(&vd->lock);
3673 	for (i = 0; i < vd->index; i++) {
3674 		if (vd->tab[i] != vp)
3675 			continue;
3676 		vp->v_dbatchcpu = NOCPU;
3677 		vd->index--;
3678 		vd->tab[i] = vd->tab[vd->index];
3679 		vd->tab[vd->index] = NULL;
3680 		break;
3681 	}
3682 	mtx_unlock(&vd->lock);
3683 	/*
3684 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3685 	 */
3686 	MPASS(vp->v_dbatchcpu == NOCPU);
3687 }
3688 
3689 /*
3690  * Drop the hold count of the vnode.  If this is the last reference to
3691  * the vnode we place it on the free list unless it has been vgone'd
3692  * (marked VIRF_DOOMED) in which case we will free it.
3693  *
3694  * Because the vnode vm object keeps a hold reference on the vnode if
3695  * there is at least one resident non-cached page, the vnode cannot
3696  * leave the active list without the page cleanup done.
3697  */
3698 static void __noinline
3699 vdropl_final(struct vnode *vp)
3700 {
3701 
3702 	ASSERT_VI_LOCKED(vp, __func__);
3703 	VNPASS(VN_IS_DOOMED(vp), vp);
3704 	/*
3705 	 * Set the VHOLD_NO_SMR flag.
3706 	 *
3707 	 * We may be racing against vhold_smr. If they win we can just pretend
3708 	 * we never got this far, they will vdrop later.
3709 	 */
3710 	if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) {
3711 		vfs_freevnodes_inc();
3712 		VI_UNLOCK(vp);
3713 		/*
3714 		 * We lost the aforementioned race. Any subsequent access is
3715 		 * invalid as they might have managed to vdropl on their own.
3716 		 */
3717 		return;
3718 	}
3719 	/*
3720 	 * Don't bump freevnodes as this one is going away.
3721 	 */
3722 	freevnode(vp);
3723 }
3724 
3725 void
3726 vdrop(struct vnode *vp)
3727 {
3728 
3729 	ASSERT_VI_UNLOCKED(vp, __func__);
3730 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3731 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3732 		return;
3733 	VI_LOCK(vp);
3734 	vdropl(vp);
3735 }
3736 
3737 static void __always_inline
3738 vdropl_impl(struct vnode *vp, bool enqueue)
3739 {
3740 
3741 	ASSERT_VI_LOCKED(vp, __func__);
3742 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3743 	if (!refcount_release(&vp->v_holdcnt)) {
3744 		VI_UNLOCK(vp);
3745 		return;
3746 	}
3747 	VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp);
3748 	VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
3749 	if (VN_IS_DOOMED(vp)) {
3750 		vdropl_final(vp);
3751 		return;
3752 	}
3753 
3754 	vfs_freevnodes_inc();
3755 	if (vp->v_mflag & VMP_LAZYLIST) {
3756 		vunlazy(vp);
3757 	}
3758 
3759 	if (!enqueue) {
3760 		VI_UNLOCK(vp);
3761 		return;
3762 	}
3763 
3764 	/*
3765 	 * Also unlocks the interlock. We can't assert on it as we
3766 	 * released our hold and by now the vnode might have been
3767 	 * freed.
3768 	 */
3769 	vdbatch_enqueue(vp);
3770 }
3771 
3772 void
3773 vdropl(struct vnode *vp)
3774 {
3775 
3776 	vdropl_impl(vp, true);
3777 }
3778 
3779 /*
3780  * vdrop a vnode when recycling
3781  *
3782  * This is a special case routine only to be used when recycling, differs from
3783  * regular vdrop by not requeieing the vnode on LRU.
3784  *
3785  * Consider a case where vtryrecycle continuously fails with all vnodes (due to
3786  * e.g., frozen writes on the filesystem), filling the batch and causing it to
3787  * be requeued. Then vnlru will end up revisiting the same vnodes. This is a
3788  * loop which can last for as long as writes are frozen.
3789  */
3790 static void
3791 vdropl_recycle(struct vnode *vp)
3792 {
3793 
3794 	vdropl_impl(vp, false);
3795 }
3796 
3797 static void
3798 vdrop_recycle(struct vnode *vp)
3799 {
3800 
3801 	VI_LOCK(vp);
3802 	vdropl_recycle(vp);
3803 }
3804 
3805 /*
3806  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3807  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3808  */
3809 static int
3810 vinactivef(struct vnode *vp)
3811 {
3812 	struct vm_object *obj;
3813 	int error;
3814 
3815 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3816 	ASSERT_VI_LOCKED(vp, "vinactive");
3817 	VNPASS((vp->v_iflag & VI_DOINGINACT) == 0, vp);
3818 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3819 	vp->v_iflag |= VI_DOINGINACT;
3820 	vp->v_iflag &= ~VI_OWEINACT;
3821 	VI_UNLOCK(vp);
3822 	/*
3823 	 * Before moving off the active list, we must be sure that any
3824 	 * modified pages are converted into the vnode's dirty
3825 	 * buffers, since these will no longer be checked once the
3826 	 * vnode is on the inactive list.
3827 	 *
3828 	 * The write-out of the dirty pages is asynchronous.  At the
3829 	 * point that VOP_INACTIVE() is called, there could still be
3830 	 * pending I/O and dirty pages in the object.
3831 	 */
3832 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3833 	    vm_object_mightbedirty(obj)) {
3834 		VM_OBJECT_WLOCK(obj);
3835 		vm_object_page_clean(obj, 0, 0, 0);
3836 		VM_OBJECT_WUNLOCK(obj);
3837 	}
3838 	error = VOP_INACTIVE(vp);
3839 	VI_LOCK(vp);
3840 	VNPASS(vp->v_iflag & VI_DOINGINACT, vp);
3841 	vp->v_iflag &= ~VI_DOINGINACT;
3842 	return (error);
3843 }
3844 
3845 int
3846 vinactive(struct vnode *vp)
3847 {
3848 
3849 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3850 	ASSERT_VI_LOCKED(vp, "vinactive");
3851 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3852 
3853 	if ((vp->v_iflag & VI_OWEINACT) == 0)
3854 		return (0);
3855 	if (vp->v_iflag & VI_DOINGINACT)
3856 		return (0);
3857 	if (vp->v_usecount > 0) {
3858 		vp->v_iflag &= ~VI_OWEINACT;
3859 		return (0);
3860 	}
3861 	return (vinactivef(vp));
3862 }
3863 
3864 /*
3865  * Remove any vnodes in the vnode table belonging to mount point mp.
3866  *
3867  * If FORCECLOSE is not specified, there should not be any active ones,
3868  * return error if any are found (nb: this is a user error, not a
3869  * system error). If FORCECLOSE is specified, detach any active vnodes
3870  * that are found.
3871  *
3872  * If WRITECLOSE is set, only flush out regular file vnodes open for
3873  * writing.
3874  *
3875  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3876  *
3877  * `rootrefs' specifies the base reference count for the root vnode
3878  * of this filesystem. The root vnode is considered busy if its
3879  * v_usecount exceeds this value. On a successful return, vflush(, td)
3880  * will call vrele() on the root vnode exactly rootrefs times.
3881  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3882  * be zero.
3883  */
3884 #ifdef DIAGNOSTIC
3885 static int busyprt = 0;		/* print out busy vnodes */
3886 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3887 #endif
3888 
3889 int
3890 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3891 {
3892 	struct vnode *vp, *mvp, *rootvp = NULL;
3893 	struct vattr vattr;
3894 	int busy = 0, error;
3895 
3896 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3897 	    rootrefs, flags);
3898 	if (rootrefs > 0) {
3899 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3900 		    ("vflush: bad args"));
3901 		/*
3902 		 * Get the filesystem root vnode. We can vput() it
3903 		 * immediately, since with rootrefs > 0, it won't go away.
3904 		 */
3905 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3906 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3907 			    __func__, error);
3908 			return (error);
3909 		}
3910 		vput(rootvp);
3911 	}
3912 loop:
3913 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3914 		vholdl(vp);
3915 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3916 		if (error) {
3917 			vdrop(vp);
3918 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3919 			goto loop;
3920 		}
3921 		/*
3922 		 * Skip over a vnodes marked VV_SYSTEM.
3923 		 */
3924 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3925 			VOP_UNLOCK(vp);
3926 			vdrop(vp);
3927 			continue;
3928 		}
3929 		/*
3930 		 * If WRITECLOSE is set, flush out unlinked but still open
3931 		 * files (even if open only for reading) and regular file
3932 		 * vnodes open for writing.
3933 		 */
3934 		if (flags & WRITECLOSE) {
3935 			if (vp->v_object != NULL) {
3936 				VM_OBJECT_WLOCK(vp->v_object);
3937 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3938 				VM_OBJECT_WUNLOCK(vp->v_object);
3939 			}
3940 			do {
3941 				error = VOP_FSYNC(vp, MNT_WAIT, td);
3942 			} while (error == ERELOOKUP);
3943 			if (error != 0) {
3944 				VOP_UNLOCK(vp);
3945 				vdrop(vp);
3946 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3947 				return (error);
3948 			}
3949 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3950 			VI_LOCK(vp);
3951 
3952 			if ((vp->v_type == VNON ||
3953 			    (error == 0 && vattr.va_nlink > 0)) &&
3954 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3955 				VOP_UNLOCK(vp);
3956 				vdropl(vp);
3957 				continue;
3958 			}
3959 		} else
3960 			VI_LOCK(vp);
3961 		/*
3962 		 * With v_usecount == 0, all we need to do is clear out the
3963 		 * vnode data structures and we are done.
3964 		 *
3965 		 * If FORCECLOSE is set, forcibly close the vnode.
3966 		 */
3967 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3968 			vgonel(vp);
3969 		} else {
3970 			busy++;
3971 #ifdef DIAGNOSTIC
3972 			if (busyprt)
3973 				vn_printf(vp, "vflush: busy vnode ");
3974 #endif
3975 		}
3976 		VOP_UNLOCK(vp);
3977 		vdropl(vp);
3978 	}
3979 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3980 		/*
3981 		 * If just the root vnode is busy, and if its refcount
3982 		 * is equal to `rootrefs', then go ahead and kill it.
3983 		 */
3984 		VI_LOCK(rootvp);
3985 		KASSERT(busy > 0, ("vflush: not busy"));
3986 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3987 		    ("vflush: usecount %d < rootrefs %d",
3988 		     rootvp->v_usecount, rootrefs));
3989 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3990 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3991 			vgone(rootvp);
3992 			VOP_UNLOCK(rootvp);
3993 			busy = 0;
3994 		} else
3995 			VI_UNLOCK(rootvp);
3996 	}
3997 	if (busy) {
3998 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3999 		    busy);
4000 		return (EBUSY);
4001 	}
4002 	for (; rootrefs > 0; rootrefs--)
4003 		vrele(rootvp);
4004 	return (0);
4005 }
4006 
4007 /*
4008  * Recycle an unused vnode to the front of the free list.
4009  */
4010 int
4011 vrecycle(struct vnode *vp)
4012 {
4013 	int recycled;
4014 
4015 	VI_LOCK(vp);
4016 	recycled = vrecyclel(vp);
4017 	VI_UNLOCK(vp);
4018 	return (recycled);
4019 }
4020 
4021 /*
4022  * vrecycle, with the vp interlock held.
4023  */
4024 int
4025 vrecyclel(struct vnode *vp)
4026 {
4027 	int recycled;
4028 
4029 	ASSERT_VOP_ELOCKED(vp, __func__);
4030 	ASSERT_VI_LOCKED(vp, __func__);
4031 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4032 	recycled = 0;
4033 	if (vp->v_usecount == 0) {
4034 		recycled = 1;
4035 		vgonel(vp);
4036 	}
4037 	return (recycled);
4038 }
4039 
4040 /*
4041  * Eliminate all activity associated with a vnode
4042  * in preparation for reuse.
4043  */
4044 void
4045 vgone(struct vnode *vp)
4046 {
4047 	VI_LOCK(vp);
4048 	vgonel(vp);
4049 	VI_UNLOCK(vp);
4050 }
4051 
4052 /*
4053  * Notify upper mounts about reclaimed or unlinked vnode.
4054  */
4055 void
4056 vfs_notify_upper(struct vnode *vp, enum vfs_notify_upper_type event)
4057 {
4058 	struct mount *mp;
4059 	struct mount_upper_node *ump;
4060 
4061 	mp = atomic_load_ptr(&vp->v_mount);
4062 	if (mp == NULL)
4063 		return;
4064 	if (TAILQ_EMPTY(&mp->mnt_notify))
4065 		return;
4066 
4067 	MNT_ILOCK(mp);
4068 	mp->mnt_upper_pending++;
4069 	KASSERT(mp->mnt_upper_pending > 0,
4070 	    ("%s: mnt_upper_pending %d", __func__, mp->mnt_upper_pending));
4071 	TAILQ_FOREACH(ump, &mp->mnt_notify, mnt_upper_link) {
4072 		MNT_IUNLOCK(mp);
4073 		switch (event) {
4074 		case VFS_NOTIFY_UPPER_RECLAIM:
4075 			VFS_RECLAIM_LOWERVP(ump->mp, vp);
4076 			break;
4077 		case VFS_NOTIFY_UPPER_UNLINK:
4078 			VFS_UNLINK_LOWERVP(ump->mp, vp);
4079 			break;
4080 		}
4081 		MNT_ILOCK(mp);
4082 	}
4083 	mp->mnt_upper_pending--;
4084 	if ((mp->mnt_kern_flag & MNTK_UPPER_WAITER) != 0 &&
4085 	    mp->mnt_upper_pending == 0) {
4086 		mp->mnt_kern_flag &= ~MNTK_UPPER_WAITER;
4087 		wakeup(&mp->mnt_uppers);
4088 	}
4089 	MNT_IUNLOCK(mp);
4090 }
4091 
4092 /*
4093  * vgone, with the vp interlock held.
4094  */
4095 static void
4096 vgonel(struct vnode *vp)
4097 {
4098 	struct thread *td;
4099 	struct mount *mp;
4100 	vm_object_t object;
4101 	bool active, doinginact, oweinact;
4102 
4103 	ASSERT_VOP_ELOCKED(vp, "vgonel");
4104 	ASSERT_VI_LOCKED(vp, "vgonel");
4105 	VNASSERT(vp->v_holdcnt, vp,
4106 	    ("vgonel: vp %p has no reference.", vp));
4107 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4108 	td = curthread;
4109 
4110 	/*
4111 	 * Don't vgonel if we're already doomed.
4112 	 */
4113 	if (VN_IS_DOOMED(vp)) {
4114 		VNPASS(vn_get_state(vp) == VSTATE_DESTROYING || \
4115 		    vn_get_state(vp) == VSTATE_DEAD, vp);
4116 		return;
4117 	}
4118 	/*
4119 	 * Paired with freevnode.
4120 	 */
4121 	vn_seqc_write_begin_locked(vp);
4122 	vunlazy_gone(vp);
4123 	vn_irflag_set_locked(vp, VIRF_DOOMED);
4124 	vn_set_state(vp, VSTATE_DESTROYING);
4125 
4126 	/*
4127 	 * Check to see if the vnode is in use.  If so, we have to
4128 	 * call VOP_CLOSE() and VOP_INACTIVE().
4129 	 *
4130 	 * It could be that VOP_INACTIVE() requested reclamation, in
4131 	 * which case we should avoid recursion, so check
4132 	 * VI_DOINGINACT.  This is not precise but good enough.
4133 	 */
4134 	active = vp->v_usecount > 0;
4135 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4136 	doinginact = (vp->v_iflag & VI_DOINGINACT) != 0;
4137 
4138 	/*
4139 	 * If we need to do inactive VI_OWEINACT will be set.
4140 	 */
4141 	if (vp->v_iflag & VI_DEFINACT) {
4142 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
4143 		vp->v_iflag &= ~VI_DEFINACT;
4144 		vdropl(vp);
4145 	} else {
4146 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
4147 		VI_UNLOCK(vp);
4148 	}
4149 	cache_purge_vgone(vp);
4150 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
4151 
4152 	/*
4153 	 * If purging an active vnode, it must be closed and
4154 	 * deactivated before being reclaimed.
4155 	 */
4156 	if (active)
4157 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
4158 	if (!doinginact) {
4159 		do {
4160 			if (oweinact || active) {
4161 				VI_LOCK(vp);
4162 				vinactivef(vp);
4163 				oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4164 				VI_UNLOCK(vp);
4165 			}
4166 		} while (oweinact);
4167 	}
4168 	if (vp->v_type == VSOCK)
4169 		vfs_unp_reclaim(vp);
4170 
4171 	/*
4172 	 * Clean out any buffers associated with the vnode.
4173 	 * If the flush fails, just toss the buffers.
4174 	 */
4175 	mp = NULL;
4176 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
4177 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
4178 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
4179 		while (vinvalbuf(vp, 0, 0, 0) != 0)
4180 			;
4181 	}
4182 
4183 	BO_LOCK(&vp->v_bufobj);
4184 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
4185 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
4186 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
4187 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
4188 	    ("vp %p bufobj not invalidated", vp));
4189 
4190 	/*
4191 	 * For VMIO bufobj, BO_DEAD is set later, or in
4192 	 * vm_object_terminate() after the object's page queue is
4193 	 * flushed.
4194 	 */
4195 	object = vp->v_bufobj.bo_object;
4196 	if (object == NULL)
4197 		vp->v_bufobj.bo_flag |= BO_DEAD;
4198 	BO_UNLOCK(&vp->v_bufobj);
4199 
4200 	/*
4201 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
4202 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
4203 	 * should not touch the object borrowed from the lower vnode
4204 	 * (the handle check).
4205 	 */
4206 	if (object != NULL && object->type == OBJT_VNODE &&
4207 	    object->handle == vp)
4208 		vnode_destroy_vobject(vp);
4209 
4210 	/*
4211 	 * Reclaim the vnode.
4212 	 */
4213 	if (VOP_RECLAIM(vp))
4214 		panic("vgone: cannot reclaim");
4215 	if (mp != NULL)
4216 		vn_finished_secondary_write(mp);
4217 	VNASSERT(vp->v_object == NULL, vp,
4218 	    ("vop_reclaim left v_object vp=%p", vp));
4219 	/*
4220 	 * Clear the advisory locks and wake up waiting threads.
4221 	 */
4222 	if (vp->v_lockf != NULL) {
4223 		(void)VOP_ADVLOCKPURGE(vp);
4224 		vp->v_lockf = NULL;
4225 	}
4226 	/*
4227 	 * Delete from old mount point vnode list.
4228 	 */
4229 	if (vp->v_mount == NULL) {
4230 		VI_LOCK(vp);
4231 	} else {
4232 		delmntque(vp);
4233 		ASSERT_VI_LOCKED(vp, "vgonel 2");
4234 	}
4235 	/*
4236 	 * Done with purge, reset to the standard lock and invalidate
4237 	 * the vnode.
4238 	 */
4239 	vp->v_vnlock = &vp->v_lock;
4240 	vp->v_op = &dead_vnodeops;
4241 	vp->v_type = VBAD;
4242 	vn_set_state(vp, VSTATE_DEAD);
4243 }
4244 
4245 /*
4246  * Print out a description of a vnode.
4247  */
4248 static const char *const vtypename[] = {
4249 	[VNON] = "VNON",
4250 	[VREG] = "VREG",
4251 	[VDIR] = "VDIR",
4252 	[VBLK] = "VBLK",
4253 	[VCHR] = "VCHR",
4254 	[VLNK] = "VLNK",
4255 	[VSOCK] = "VSOCK",
4256 	[VFIFO] = "VFIFO",
4257 	[VBAD] = "VBAD",
4258 	[VMARKER] = "VMARKER",
4259 };
4260 _Static_assert(nitems(vtypename) == VLASTTYPE + 1,
4261     "vnode type name not added to vtypename");
4262 
4263 static const char *const vstatename[] = {
4264 	[VSTATE_UNINITIALIZED] = "VSTATE_UNINITIALIZED",
4265 	[VSTATE_CONSTRUCTED] = "VSTATE_CONSTRUCTED",
4266 	[VSTATE_DESTROYING] = "VSTATE_DESTROYING",
4267 	[VSTATE_DEAD] = "VSTATE_DEAD",
4268 };
4269 _Static_assert(nitems(vstatename) == VLASTSTATE + 1,
4270     "vnode state name not added to vstatename");
4271 
4272 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0,
4273     "new hold count flag not added to vn_printf");
4274 
4275 void
4276 vn_printf(struct vnode *vp, const char *fmt, ...)
4277 {
4278 	va_list ap;
4279 	char buf[256], buf2[16];
4280 	u_long flags;
4281 	u_int holdcnt;
4282 	short irflag;
4283 
4284 	va_start(ap, fmt);
4285 	vprintf(fmt, ap);
4286 	va_end(ap);
4287 	printf("%p: ", (void *)vp);
4288 	printf("type %s state %s op %p\n", vtypename[vp->v_type],
4289 	    vstatename[vp->v_state], vp->v_op);
4290 	holdcnt = atomic_load_int(&vp->v_holdcnt);
4291 	printf("    usecount %d, writecount %d, refcount %d seqc users %d",
4292 	    vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS,
4293 	    vp->v_seqc_users);
4294 	switch (vp->v_type) {
4295 	case VDIR:
4296 		printf(" mountedhere %p\n", vp->v_mountedhere);
4297 		break;
4298 	case VCHR:
4299 		printf(" rdev %p\n", vp->v_rdev);
4300 		break;
4301 	case VSOCK:
4302 		printf(" socket %p\n", vp->v_unpcb);
4303 		break;
4304 	case VFIFO:
4305 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
4306 		break;
4307 	default:
4308 		printf("\n");
4309 		break;
4310 	}
4311 	buf[0] = '\0';
4312 	buf[1] = '\0';
4313 	if (holdcnt & VHOLD_NO_SMR)
4314 		strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf));
4315 	printf("    hold count flags (%s)\n", buf + 1);
4316 
4317 	buf[0] = '\0';
4318 	buf[1] = '\0';
4319 	irflag = vn_irflag_read(vp);
4320 	if (irflag & VIRF_DOOMED)
4321 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4322 	if (irflag & VIRF_PGREAD)
4323 		strlcat(buf, "|VIRF_PGREAD", sizeof(buf));
4324 	if (irflag & VIRF_MOUNTPOINT)
4325 		strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf));
4326 	if (irflag & VIRF_TEXT_REF)
4327 		strlcat(buf, "|VIRF_TEXT_REF", sizeof(buf));
4328 	flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT | VIRF_TEXT_REF);
4329 	if (flags != 0) {
4330 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4331 		strlcat(buf, buf2, sizeof(buf));
4332 	}
4333 	if (vp->v_vflag & VV_ROOT)
4334 		strlcat(buf, "|VV_ROOT", sizeof(buf));
4335 	if (vp->v_vflag & VV_ISTTY)
4336 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
4337 	if (vp->v_vflag & VV_NOSYNC)
4338 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4339 	if (vp->v_vflag & VV_ETERNALDEV)
4340 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4341 	if (vp->v_vflag & VV_CACHEDLABEL)
4342 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4343 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
4344 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4345 	if (vp->v_vflag & VV_COPYONWRITE)
4346 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4347 	if (vp->v_vflag & VV_SYSTEM)
4348 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4349 	if (vp->v_vflag & VV_PROCDEP)
4350 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4351 	if (vp->v_vflag & VV_DELETED)
4352 		strlcat(buf, "|VV_DELETED", sizeof(buf));
4353 	if (vp->v_vflag & VV_MD)
4354 		strlcat(buf, "|VV_MD", sizeof(buf));
4355 	if (vp->v_vflag & VV_FORCEINSMQ)
4356 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4357 	if (vp->v_vflag & VV_READLINK)
4358 		strlcat(buf, "|VV_READLINK", sizeof(buf));
4359 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4360 	    VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM |
4361 	    VV_PROCDEP | VV_DELETED | VV_MD | VV_FORCEINSMQ | VV_READLINK);
4362 	if (flags != 0) {
4363 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4364 		strlcat(buf, buf2, sizeof(buf));
4365 	}
4366 	if (vp->v_iflag & VI_MOUNT)
4367 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
4368 	if (vp->v_iflag & VI_DOINGINACT)
4369 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4370 	if (vp->v_iflag & VI_OWEINACT)
4371 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4372 	if (vp->v_iflag & VI_DEFINACT)
4373 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4374 	if (vp->v_iflag & VI_FOPENING)
4375 		strlcat(buf, "|VI_FOPENING", sizeof(buf));
4376 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOINGINACT |
4377 	    VI_OWEINACT | VI_DEFINACT | VI_FOPENING);
4378 	if (flags != 0) {
4379 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4380 		strlcat(buf, buf2, sizeof(buf));
4381 	}
4382 	if (vp->v_mflag & VMP_LAZYLIST)
4383 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4384 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
4385 	if (flags != 0) {
4386 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4387 		strlcat(buf, buf2, sizeof(buf));
4388 	}
4389 	printf("    flags (%s)", buf + 1);
4390 	if (mtx_owned(VI_MTX(vp)))
4391 		printf(" VI_LOCKed");
4392 	printf("\n");
4393 	if (vp->v_object != NULL)
4394 		printf("    v_object %p ref %d pages %d "
4395 		    "cleanbuf %d dirtybuf %d\n",
4396 		    vp->v_object, vp->v_object->ref_count,
4397 		    vp->v_object->resident_page_count,
4398 		    vp->v_bufobj.bo_clean.bv_cnt,
4399 		    vp->v_bufobj.bo_dirty.bv_cnt);
4400 	printf("    ");
4401 	lockmgr_printinfo(vp->v_vnlock);
4402 	if (vp->v_data != NULL)
4403 		VOP_PRINT(vp);
4404 }
4405 
4406 #ifdef DDB
4407 /*
4408  * List all of the locked vnodes in the system.
4409  * Called when debugging the kernel.
4410  */
4411 DB_SHOW_COMMAND_FLAGS(lockedvnods, lockedvnodes, DB_CMD_MEMSAFE)
4412 {
4413 	struct mount *mp;
4414 	struct vnode *vp;
4415 
4416 	/*
4417 	 * Note: because this is DDB, we can't obey the locking semantics
4418 	 * for these structures, which means we could catch an inconsistent
4419 	 * state and dereference a nasty pointer.  Not much to be done
4420 	 * about that.
4421 	 */
4422 	db_printf("Locked vnodes\n");
4423 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4424 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4425 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4426 				vn_printf(vp, "vnode ");
4427 		}
4428 	}
4429 }
4430 
4431 /*
4432  * Show details about the given vnode.
4433  */
4434 DB_SHOW_COMMAND(vnode, db_show_vnode)
4435 {
4436 	struct vnode *vp;
4437 
4438 	if (!have_addr)
4439 		return;
4440 	vp = (struct vnode *)addr;
4441 	vn_printf(vp, "vnode ");
4442 }
4443 
4444 /*
4445  * Show details about the given mount point.
4446  */
4447 DB_SHOW_COMMAND(mount, db_show_mount)
4448 {
4449 	struct mount *mp;
4450 	struct vfsopt *opt;
4451 	struct statfs *sp;
4452 	struct vnode *vp;
4453 	char buf[512];
4454 	uint64_t mflags;
4455 	u_int flags;
4456 
4457 	if (!have_addr) {
4458 		/* No address given, print short info about all mount points. */
4459 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4460 			db_printf("%p %s on %s (%s)\n", mp,
4461 			    mp->mnt_stat.f_mntfromname,
4462 			    mp->mnt_stat.f_mntonname,
4463 			    mp->mnt_stat.f_fstypename);
4464 			if (db_pager_quit)
4465 				break;
4466 		}
4467 		db_printf("\nMore info: show mount <addr>\n");
4468 		return;
4469 	}
4470 
4471 	mp = (struct mount *)addr;
4472 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4473 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4474 
4475 	buf[0] = '\0';
4476 	mflags = mp->mnt_flag;
4477 #define	MNT_FLAG(flag)	do {						\
4478 	if (mflags & (flag)) {						\
4479 		if (buf[0] != '\0')					\
4480 			strlcat(buf, ", ", sizeof(buf));		\
4481 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4482 		mflags &= ~(flag);					\
4483 	}								\
4484 } while (0)
4485 	MNT_FLAG(MNT_RDONLY);
4486 	MNT_FLAG(MNT_SYNCHRONOUS);
4487 	MNT_FLAG(MNT_NOEXEC);
4488 	MNT_FLAG(MNT_NOSUID);
4489 	MNT_FLAG(MNT_NFS4ACLS);
4490 	MNT_FLAG(MNT_UNION);
4491 	MNT_FLAG(MNT_ASYNC);
4492 	MNT_FLAG(MNT_SUIDDIR);
4493 	MNT_FLAG(MNT_SOFTDEP);
4494 	MNT_FLAG(MNT_NOSYMFOLLOW);
4495 	MNT_FLAG(MNT_GJOURNAL);
4496 	MNT_FLAG(MNT_MULTILABEL);
4497 	MNT_FLAG(MNT_ACLS);
4498 	MNT_FLAG(MNT_NOATIME);
4499 	MNT_FLAG(MNT_NOCLUSTERR);
4500 	MNT_FLAG(MNT_NOCLUSTERW);
4501 	MNT_FLAG(MNT_SUJ);
4502 	MNT_FLAG(MNT_EXRDONLY);
4503 	MNT_FLAG(MNT_EXPORTED);
4504 	MNT_FLAG(MNT_DEFEXPORTED);
4505 	MNT_FLAG(MNT_EXPORTANON);
4506 	MNT_FLAG(MNT_EXKERB);
4507 	MNT_FLAG(MNT_EXPUBLIC);
4508 	MNT_FLAG(MNT_LOCAL);
4509 	MNT_FLAG(MNT_QUOTA);
4510 	MNT_FLAG(MNT_ROOTFS);
4511 	MNT_FLAG(MNT_USER);
4512 	MNT_FLAG(MNT_IGNORE);
4513 	MNT_FLAG(MNT_UPDATE);
4514 	MNT_FLAG(MNT_DELEXPORT);
4515 	MNT_FLAG(MNT_RELOAD);
4516 	MNT_FLAG(MNT_FORCE);
4517 	MNT_FLAG(MNT_SNAPSHOT);
4518 	MNT_FLAG(MNT_BYFSID);
4519 #undef MNT_FLAG
4520 	if (mflags != 0) {
4521 		if (buf[0] != '\0')
4522 			strlcat(buf, ", ", sizeof(buf));
4523 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4524 		    "0x%016jx", mflags);
4525 	}
4526 	db_printf("    mnt_flag = %s\n", buf);
4527 
4528 	buf[0] = '\0';
4529 	flags = mp->mnt_kern_flag;
4530 #define	MNT_KERN_FLAG(flag)	do {					\
4531 	if (flags & (flag)) {						\
4532 		if (buf[0] != '\0')					\
4533 			strlcat(buf, ", ", sizeof(buf));		\
4534 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4535 		flags &= ~(flag);					\
4536 	}								\
4537 } while (0)
4538 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4539 	MNT_KERN_FLAG(MNTK_ASYNC);
4540 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4541 	MNT_KERN_FLAG(MNTK_NOMSYNC);
4542 	MNT_KERN_FLAG(MNTK_DRAINING);
4543 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4544 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4545 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4546 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4547 	MNT_KERN_FLAG(MNTK_RECURSE);
4548 	MNT_KERN_FLAG(MNTK_UPPER_WAITER);
4549 	MNT_KERN_FLAG(MNTK_UNLOCKED_INSMNTQUE);
4550 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4551 	MNT_KERN_FLAG(MNTK_VMSETSIZE_BUG);
4552 	MNT_KERN_FLAG(MNTK_FPLOOKUP);
4553 	MNT_KERN_FLAG(MNTK_TASKQUEUE_WAITER);
4554 	MNT_KERN_FLAG(MNTK_NOASYNC);
4555 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4556 	MNT_KERN_FLAG(MNTK_MWAIT);
4557 	MNT_KERN_FLAG(MNTK_SUSPEND);
4558 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4559 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4560 	MNT_KERN_FLAG(MNTK_NULL_NOCACHE);
4561 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4562 #undef MNT_KERN_FLAG
4563 	if (flags != 0) {
4564 		if (buf[0] != '\0')
4565 			strlcat(buf, ", ", sizeof(buf));
4566 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4567 		    "0x%08x", flags);
4568 	}
4569 	db_printf("    mnt_kern_flag = %s\n", buf);
4570 
4571 	db_printf("    mnt_opt = ");
4572 	opt = TAILQ_FIRST(mp->mnt_opt);
4573 	if (opt != NULL) {
4574 		db_printf("%s", opt->name);
4575 		opt = TAILQ_NEXT(opt, link);
4576 		while (opt != NULL) {
4577 			db_printf(", %s", opt->name);
4578 			opt = TAILQ_NEXT(opt, link);
4579 		}
4580 	}
4581 	db_printf("\n");
4582 
4583 	sp = &mp->mnt_stat;
4584 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4585 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4586 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4587 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4588 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4589 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4590 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4591 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4592 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4593 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4594 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4595 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4596 
4597 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4598 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4599 	if (jailed(mp->mnt_cred))
4600 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4601 	db_printf(" }\n");
4602 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4603 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4604 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4605 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4606 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4607 	    mp->mnt_lazyvnodelistsize);
4608 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4609 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4610 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4611 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4612 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4613 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4614 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4615 	db_printf("    mnt_secondary_accwrites = %d\n",
4616 	    mp->mnt_secondary_accwrites);
4617 	db_printf("    mnt_gjprovider = %s\n",
4618 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4619 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4620 
4621 	db_printf("\n\nList of active vnodes\n");
4622 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4623 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4624 			vn_printf(vp, "vnode ");
4625 			if (db_pager_quit)
4626 				break;
4627 		}
4628 	}
4629 	db_printf("\n\nList of inactive vnodes\n");
4630 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4631 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4632 			vn_printf(vp, "vnode ");
4633 			if (db_pager_quit)
4634 				break;
4635 		}
4636 	}
4637 }
4638 #endif	/* DDB */
4639 
4640 /*
4641  * Fill in a struct xvfsconf based on a struct vfsconf.
4642  */
4643 static int
4644 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4645 {
4646 	struct xvfsconf xvfsp;
4647 
4648 	bzero(&xvfsp, sizeof(xvfsp));
4649 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4650 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4651 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4652 	xvfsp.vfc_flags = vfsp->vfc_flags;
4653 	/*
4654 	 * These are unused in userland, we keep them
4655 	 * to not break binary compatibility.
4656 	 */
4657 	xvfsp.vfc_vfsops = NULL;
4658 	xvfsp.vfc_next = NULL;
4659 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4660 }
4661 
4662 #ifdef COMPAT_FREEBSD32
4663 struct xvfsconf32 {
4664 	uint32_t	vfc_vfsops;
4665 	char		vfc_name[MFSNAMELEN];
4666 	int32_t		vfc_typenum;
4667 	int32_t		vfc_refcount;
4668 	int32_t		vfc_flags;
4669 	uint32_t	vfc_next;
4670 };
4671 
4672 static int
4673 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4674 {
4675 	struct xvfsconf32 xvfsp;
4676 
4677 	bzero(&xvfsp, sizeof(xvfsp));
4678 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4679 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4680 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4681 	xvfsp.vfc_flags = vfsp->vfc_flags;
4682 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4683 }
4684 #endif
4685 
4686 /*
4687  * Top level filesystem related information gathering.
4688  */
4689 static int
4690 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4691 {
4692 	struct vfsconf *vfsp;
4693 	int error;
4694 
4695 	error = 0;
4696 	vfsconf_slock();
4697 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4698 #ifdef COMPAT_FREEBSD32
4699 		if (req->flags & SCTL_MASK32)
4700 			error = vfsconf2x32(req, vfsp);
4701 		else
4702 #endif
4703 			error = vfsconf2x(req, vfsp);
4704 		if (error)
4705 			break;
4706 	}
4707 	vfsconf_sunlock();
4708 	return (error);
4709 }
4710 
4711 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4712     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4713     "S,xvfsconf", "List of all configured filesystems");
4714 
4715 #ifndef BURN_BRIDGES
4716 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4717 
4718 static int
4719 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4720 {
4721 	int *name = (int *)arg1 - 1;	/* XXX */
4722 	u_int namelen = arg2 + 1;	/* XXX */
4723 	struct vfsconf *vfsp;
4724 
4725 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4726 	    "please rebuild world\n");
4727 
4728 #if 1 || defined(COMPAT_PRELITE2)
4729 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4730 	if (namelen == 1)
4731 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4732 #endif
4733 
4734 	switch (name[1]) {
4735 	case VFS_MAXTYPENUM:
4736 		if (namelen != 2)
4737 			return (ENOTDIR);
4738 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4739 	case VFS_CONF:
4740 		if (namelen != 3)
4741 			return (ENOTDIR);	/* overloaded */
4742 		vfsconf_slock();
4743 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4744 			if (vfsp->vfc_typenum == name[2])
4745 				break;
4746 		}
4747 		vfsconf_sunlock();
4748 		if (vfsp == NULL)
4749 			return (EOPNOTSUPP);
4750 #ifdef COMPAT_FREEBSD32
4751 		if (req->flags & SCTL_MASK32)
4752 			return (vfsconf2x32(req, vfsp));
4753 		else
4754 #endif
4755 			return (vfsconf2x(req, vfsp));
4756 	}
4757 	return (EOPNOTSUPP);
4758 }
4759 
4760 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4761     CTLFLAG_MPSAFE, vfs_sysctl,
4762     "Generic filesystem");
4763 
4764 #if 1 || defined(COMPAT_PRELITE2)
4765 
4766 static int
4767 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4768 {
4769 	int error;
4770 	struct vfsconf *vfsp;
4771 	struct ovfsconf ovfs;
4772 
4773 	vfsconf_slock();
4774 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4775 		bzero(&ovfs, sizeof(ovfs));
4776 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4777 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4778 		ovfs.vfc_index = vfsp->vfc_typenum;
4779 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4780 		ovfs.vfc_flags = vfsp->vfc_flags;
4781 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4782 		if (error != 0) {
4783 			vfsconf_sunlock();
4784 			return (error);
4785 		}
4786 	}
4787 	vfsconf_sunlock();
4788 	return (0);
4789 }
4790 
4791 #endif /* 1 || COMPAT_PRELITE2 */
4792 #endif /* !BURN_BRIDGES */
4793 
4794 static void
4795 unmount_or_warn(struct mount *mp)
4796 {
4797 	int error;
4798 
4799 	error = dounmount(mp, MNT_FORCE, curthread);
4800 	if (error != 0) {
4801 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4802 		if (error == EBUSY)
4803 			printf("BUSY)\n");
4804 		else
4805 			printf("%d)\n", error);
4806 	}
4807 }
4808 
4809 /*
4810  * Unmount all filesystems. The list is traversed in reverse order
4811  * of mounting to avoid dependencies.
4812  */
4813 void
4814 vfs_unmountall(void)
4815 {
4816 	struct mount *mp, *tmp;
4817 
4818 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4819 
4820 	/*
4821 	 * Since this only runs when rebooting, it is not interlocked.
4822 	 */
4823 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4824 		vfs_ref(mp);
4825 
4826 		/*
4827 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4828 		 * unmount of the latter.
4829 		 */
4830 		if (mp == rootdevmp)
4831 			continue;
4832 
4833 		unmount_or_warn(mp);
4834 	}
4835 
4836 	if (rootdevmp != NULL)
4837 		unmount_or_warn(rootdevmp);
4838 }
4839 
4840 static void
4841 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4842 {
4843 
4844 	ASSERT_VI_LOCKED(vp, __func__);
4845 	VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
4846 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
4847 		vdropl(vp);
4848 		return;
4849 	}
4850 	if (vn_lock(vp, lkflags) == 0) {
4851 		VI_LOCK(vp);
4852 		vinactive(vp);
4853 		VOP_UNLOCK(vp);
4854 		vdropl(vp);
4855 		return;
4856 	}
4857 	vdefer_inactive_unlocked(vp);
4858 }
4859 
4860 static int
4861 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4862 {
4863 
4864 	return (vp->v_iflag & VI_DEFINACT);
4865 }
4866 
4867 static void __noinline
4868 vfs_periodic_inactive(struct mount *mp, int flags)
4869 {
4870 	struct vnode *vp, *mvp;
4871 	int lkflags;
4872 
4873 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4874 	if (flags != MNT_WAIT)
4875 		lkflags |= LK_NOWAIT;
4876 
4877 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4878 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
4879 			VI_UNLOCK(vp);
4880 			continue;
4881 		}
4882 		vp->v_iflag &= ~VI_DEFINACT;
4883 		vfs_deferred_inactive(vp, lkflags);
4884 	}
4885 }
4886 
4887 static inline bool
4888 vfs_want_msync(struct vnode *vp)
4889 {
4890 	struct vm_object *obj;
4891 
4892 	/*
4893 	 * This test may be performed without any locks held.
4894 	 * We rely on vm_object's type stability.
4895 	 */
4896 	if (vp->v_vflag & VV_NOSYNC)
4897 		return (false);
4898 	obj = vp->v_object;
4899 	return (obj != NULL && vm_object_mightbedirty(obj));
4900 }
4901 
4902 static int
4903 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4904 {
4905 
4906 	if (vp->v_vflag & VV_NOSYNC)
4907 		return (false);
4908 	if (vp->v_iflag & VI_DEFINACT)
4909 		return (true);
4910 	return (vfs_want_msync(vp));
4911 }
4912 
4913 static void __noinline
4914 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4915 {
4916 	struct vnode *vp, *mvp;
4917 	struct vm_object *obj;
4918 	int lkflags, objflags;
4919 	bool seen_defer;
4920 
4921 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4922 	if (flags != MNT_WAIT) {
4923 		lkflags |= LK_NOWAIT;
4924 		objflags = OBJPC_NOSYNC;
4925 	} else {
4926 		objflags = OBJPC_SYNC;
4927 	}
4928 
4929 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4930 		seen_defer = false;
4931 		if (vp->v_iflag & VI_DEFINACT) {
4932 			vp->v_iflag &= ~VI_DEFINACT;
4933 			seen_defer = true;
4934 		}
4935 		if (!vfs_want_msync(vp)) {
4936 			if (seen_defer)
4937 				vfs_deferred_inactive(vp, lkflags);
4938 			else
4939 				VI_UNLOCK(vp);
4940 			continue;
4941 		}
4942 		if (vget(vp, lkflags) == 0) {
4943 			obj = vp->v_object;
4944 			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4945 				VM_OBJECT_WLOCK(obj);
4946 				vm_object_page_clean(obj, 0, 0, objflags);
4947 				VM_OBJECT_WUNLOCK(obj);
4948 			}
4949 			vput(vp);
4950 			if (seen_defer)
4951 				vdrop(vp);
4952 		} else {
4953 			if (seen_defer)
4954 				vdefer_inactive_unlocked(vp);
4955 		}
4956 	}
4957 }
4958 
4959 void
4960 vfs_periodic(struct mount *mp, int flags)
4961 {
4962 
4963 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4964 
4965 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4966 		vfs_periodic_inactive(mp, flags);
4967 	else
4968 		vfs_periodic_msync_inactive(mp, flags);
4969 }
4970 
4971 static void
4972 destroy_vpollinfo_free(struct vpollinfo *vi)
4973 {
4974 
4975 	knlist_destroy(&vi->vpi_selinfo.si_note);
4976 	mtx_destroy(&vi->vpi_lock);
4977 	free(vi, M_VNODEPOLL);
4978 }
4979 
4980 static void
4981 destroy_vpollinfo(struct vpollinfo *vi)
4982 {
4983 
4984 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4985 	seldrain(&vi->vpi_selinfo);
4986 	destroy_vpollinfo_free(vi);
4987 }
4988 
4989 /*
4990  * Initialize per-vnode helper structure to hold poll-related state.
4991  */
4992 void
4993 v_addpollinfo(struct vnode *vp)
4994 {
4995 	struct vpollinfo *vi;
4996 
4997 	if (vp->v_pollinfo != NULL)
4998 		return;
4999 	vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO);
5000 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
5001 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
5002 	    vfs_knlunlock, vfs_knl_assert_lock);
5003 	VI_LOCK(vp);
5004 	if (vp->v_pollinfo != NULL) {
5005 		VI_UNLOCK(vp);
5006 		destroy_vpollinfo_free(vi);
5007 		return;
5008 	}
5009 	vp->v_pollinfo = vi;
5010 	VI_UNLOCK(vp);
5011 }
5012 
5013 /*
5014  * Record a process's interest in events which might happen to
5015  * a vnode.  Because poll uses the historic select-style interface
5016  * internally, this routine serves as both the ``check for any
5017  * pending events'' and the ``record my interest in future events''
5018  * functions.  (These are done together, while the lock is held,
5019  * to avoid race conditions.)
5020  */
5021 int
5022 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
5023 {
5024 
5025 	v_addpollinfo(vp);
5026 	mtx_lock(&vp->v_pollinfo->vpi_lock);
5027 	if (vp->v_pollinfo->vpi_revents & events) {
5028 		/*
5029 		 * This leaves events we are not interested
5030 		 * in available for the other process which
5031 		 * which presumably had requested them
5032 		 * (otherwise they would never have been
5033 		 * recorded).
5034 		 */
5035 		events &= vp->v_pollinfo->vpi_revents;
5036 		vp->v_pollinfo->vpi_revents &= ~events;
5037 
5038 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
5039 		return (events);
5040 	}
5041 	vp->v_pollinfo->vpi_events |= events;
5042 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
5043 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
5044 	return (0);
5045 }
5046 
5047 /*
5048  * Routine to create and manage a filesystem syncer vnode.
5049  */
5050 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
5051 static int	sync_fsync(struct  vop_fsync_args *);
5052 static int	sync_inactive(struct  vop_inactive_args *);
5053 static int	sync_reclaim(struct  vop_reclaim_args *);
5054 
5055 static struct vop_vector sync_vnodeops = {
5056 	.vop_bypass =	VOP_EOPNOTSUPP,
5057 	.vop_close =	sync_close,
5058 	.vop_fsync =	sync_fsync,
5059 	.vop_getwritemount = vop_stdgetwritemount,
5060 	.vop_inactive =	sync_inactive,
5061 	.vop_need_inactive = vop_stdneed_inactive,
5062 	.vop_reclaim =	sync_reclaim,
5063 	.vop_lock1 =	vop_stdlock,
5064 	.vop_unlock =	vop_stdunlock,
5065 	.vop_islocked =	vop_stdislocked,
5066 	.vop_fplookup_vexec = VOP_EAGAIN,
5067 	.vop_fplookup_symlink = VOP_EAGAIN,
5068 };
5069 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
5070 
5071 /*
5072  * Create a new filesystem syncer vnode for the specified mount point.
5073  */
5074 void
5075 vfs_allocate_syncvnode(struct mount *mp)
5076 {
5077 	struct vnode *vp;
5078 	struct bufobj *bo;
5079 	static long start, incr, next;
5080 	int error;
5081 
5082 	/* Allocate a new vnode */
5083 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
5084 	if (error != 0)
5085 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
5086 	vp->v_type = VNON;
5087 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5088 	vp->v_vflag |= VV_FORCEINSMQ;
5089 	error = insmntque1(vp, mp);
5090 	if (error != 0)
5091 		panic("vfs_allocate_syncvnode: insmntque() failed");
5092 	vp->v_vflag &= ~VV_FORCEINSMQ;
5093 	vn_set_state(vp, VSTATE_CONSTRUCTED);
5094 	VOP_UNLOCK(vp);
5095 	/*
5096 	 * Place the vnode onto the syncer worklist. We attempt to
5097 	 * scatter them about on the list so that they will go off
5098 	 * at evenly distributed times even if all the filesystems
5099 	 * are mounted at once.
5100 	 */
5101 	next += incr;
5102 	if (next == 0 || next > syncer_maxdelay) {
5103 		start /= 2;
5104 		incr /= 2;
5105 		if (start == 0) {
5106 			start = syncer_maxdelay / 2;
5107 			incr = syncer_maxdelay;
5108 		}
5109 		next = start;
5110 	}
5111 	bo = &vp->v_bufobj;
5112 	BO_LOCK(bo);
5113 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
5114 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
5115 	mtx_lock(&sync_mtx);
5116 	sync_vnode_count++;
5117 	if (mp->mnt_syncer == NULL) {
5118 		mp->mnt_syncer = vp;
5119 		vp = NULL;
5120 	}
5121 	mtx_unlock(&sync_mtx);
5122 	BO_UNLOCK(bo);
5123 	if (vp != NULL) {
5124 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5125 		vgone(vp);
5126 		vput(vp);
5127 	}
5128 }
5129 
5130 void
5131 vfs_deallocate_syncvnode(struct mount *mp)
5132 {
5133 	struct vnode *vp;
5134 
5135 	mtx_lock(&sync_mtx);
5136 	vp = mp->mnt_syncer;
5137 	if (vp != NULL)
5138 		mp->mnt_syncer = NULL;
5139 	mtx_unlock(&sync_mtx);
5140 	if (vp != NULL)
5141 		vrele(vp);
5142 }
5143 
5144 /*
5145  * Do a lazy sync of the filesystem.
5146  */
5147 static int
5148 sync_fsync(struct vop_fsync_args *ap)
5149 {
5150 	struct vnode *syncvp = ap->a_vp;
5151 	struct mount *mp = syncvp->v_mount;
5152 	int error, save;
5153 	struct bufobj *bo;
5154 
5155 	/*
5156 	 * We only need to do something if this is a lazy evaluation.
5157 	 */
5158 	if (ap->a_waitfor != MNT_LAZY)
5159 		return (0);
5160 
5161 	/*
5162 	 * Move ourselves to the back of the sync list.
5163 	 */
5164 	bo = &syncvp->v_bufobj;
5165 	BO_LOCK(bo);
5166 	vn_syncer_add_to_worklist(bo, syncdelay);
5167 	BO_UNLOCK(bo);
5168 
5169 	/*
5170 	 * Walk the list of vnodes pushing all that are dirty and
5171 	 * not already on the sync list.
5172 	 */
5173 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
5174 		return (0);
5175 	VOP_UNLOCK(syncvp);
5176 	save = curthread_pflags_set(TDP_SYNCIO);
5177 	/*
5178 	 * The filesystem at hand may be idle with free vnodes stored in the
5179 	 * batch.  Return them instead of letting them stay there indefinitely.
5180 	 */
5181 	vfs_periodic(mp, MNT_NOWAIT);
5182 	error = VFS_SYNC(mp, MNT_LAZY);
5183 	curthread_pflags_restore(save);
5184 	vn_lock(syncvp, LK_EXCLUSIVE | LK_RETRY);
5185 	vfs_unbusy(mp);
5186 	return (error);
5187 }
5188 
5189 /*
5190  * The syncer vnode is no referenced.
5191  */
5192 static int
5193 sync_inactive(struct vop_inactive_args *ap)
5194 {
5195 
5196 	vgone(ap->a_vp);
5197 	return (0);
5198 }
5199 
5200 /*
5201  * The syncer vnode is no longer needed and is being decommissioned.
5202  *
5203  * Modifications to the worklist must be protected by sync_mtx.
5204  */
5205 static int
5206 sync_reclaim(struct vop_reclaim_args *ap)
5207 {
5208 	struct vnode *vp = ap->a_vp;
5209 	struct bufobj *bo;
5210 
5211 	bo = &vp->v_bufobj;
5212 	BO_LOCK(bo);
5213 	mtx_lock(&sync_mtx);
5214 	if (vp->v_mount->mnt_syncer == vp)
5215 		vp->v_mount->mnt_syncer = NULL;
5216 	if (bo->bo_flag & BO_ONWORKLST) {
5217 		LIST_REMOVE(bo, bo_synclist);
5218 		syncer_worklist_len--;
5219 		sync_vnode_count--;
5220 		bo->bo_flag &= ~BO_ONWORKLST;
5221 	}
5222 	mtx_unlock(&sync_mtx);
5223 	BO_UNLOCK(bo);
5224 
5225 	return (0);
5226 }
5227 
5228 int
5229 vn_need_pageq_flush(struct vnode *vp)
5230 {
5231 	struct vm_object *obj;
5232 
5233 	obj = vp->v_object;
5234 	return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5235 	    vm_object_mightbedirty(obj));
5236 }
5237 
5238 /*
5239  * Check if vnode represents a disk device
5240  */
5241 bool
5242 vn_isdisk_error(struct vnode *vp, int *errp)
5243 {
5244 	int error;
5245 
5246 	if (vp->v_type != VCHR) {
5247 		error = ENOTBLK;
5248 		goto out;
5249 	}
5250 	error = 0;
5251 	dev_lock();
5252 	if (vp->v_rdev == NULL)
5253 		error = ENXIO;
5254 	else if (vp->v_rdev->si_devsw == NULL)
5255 		error = ENXIO;
5256 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5257 		error = ENOTBLK;
5258 	dev_unlock();
5259 out:
5260 	*errp = error;
5261 	return (error == 0);
5262 }
5263 
5264 bool
5265 vn_isdisk(struct vnode *vp)
5266 {
5267 	int error;
5268 
5269 	return (vn_isdisk_error(vp, &error));
5270 }
5271 
5272 /*
5273  * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
5274  * the comment above cache_fplookup for details.
5275  */
5276 int
5277 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred)
5278 {
5279 	int error;
5280 
5281 	VFS_SMR_ASSERT_ENTERED();
5282 
5283 	/* Check the owner. */
5284 	if (cred->cr_uid == file_uid) {
5285 		if (file_mode & S_IXUSR)
5286 			return (0);
5287 		goto out_error;
5288 	}
5289 
5290 	/* Otherwise, check the groups (first match) */
5291 	if (groupmember(file_gid, cred)) {
5292 		if (file_mode & S_IXGRP)
5293 			return (0);
5294 		goto out_error;
5295 	}
5296 
5297 	/* Otherwise, check everyone else. */
5298 	if (file_mode & S_IXOTH)
5299 		return (0);
5300 out_error:
5301 	/*
5302 	 * Permission check failed, but it is possible denial will get overwritten
5303 	 * (e.g., when root is traversing through a 700 directory owned by someone
5304 	 * else).
5305 	 *
5306 	 * vaccess() calls priv_check_cred which in turn can descent into MAC
5307 	 * modules overriding this result. It's quite unclear what semantics
5308 	 * are allowed for them to operate, thus for safety we don't call them
5309 	 * from within the SMR section. This also means if any such modules
5310 	 * are present, we have to let the regular lookup decide.
5311 	 */
5312 	error = priv_check_cred_vfs_lookup_nomac(cred);
5313 	switch (error) {
5314 	case 0:
5315 		return (0);
5316 	case EAGAIN:
5317 		/*
5318 		 * MAC modules present.
5319 		 */
5320 		return (EAGAIN);
5321 	case EPERM:
5322 		return (EACCES);
5323 	default:
5324 		return (error);
5325 	}
5326 }
5327 
5328 /*
5329  * Common filesystem object access control check routine.  Accepts a
5330  * vnode's type, "mode", uid and gid, requested access mode, and credentials.
5331  * Returns 0 on success, or an errno on failure.
5332  */
5333 int
5334 vaccess(__enum_uint8(vtype) type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5335     accmode_t accmode, struct ucred *cred)
5336 {
5337 	accmode_t dac_granted;
5338 	accmode_t priv_granted;
5339 
5340 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5341 	    ("invalid bit in accmode"));
5342 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5343 	    ("VAPPEND without VWRITE"));
5344 
5345 	/*
5346 	 * Look for a normal, non-privileged way to access the file/directory
5347 	 * as requested.  If it exists, go with that.
5348 	 */
5349 
5350 	dac_granted = 0;
5351 
5352 	/* Check the owner. */
5353 	if (cred->cr_uid == file_uid) {
5354 		dac_granted |= VADMIN;
5355 		if (file_mode & S_IXUSR)
5356 			dac_granted |= VEXEC;
5357 		if (file_mode & S_IRUSR)
5358 			dac_granted |= VREAD;
5359 		if (file_mode & S_IWUSR)
5360 			dac_granted |= (VWRITE | VAPPEND);
5361 
5362 		if ((accmode & dac_granted) == accmode)
5363 			return (0);
5364 
5365 		goto privcheck;
5366 	}
5367 
5368 	/* Otherwise, check the groups (first match) */
5369 	if (groupmember(file_gid, cred)) {
5370 		if (file_mode & S_IXGRP)
5371 			dac_granted |= VEXEC;
5372 		if (file_mode & S_IRGRP)
5373 			dac_granted |= VREAD;
5374 		if (file_mode & S_IWGRP)
5375 			dac_granted |= (VWRITE | VAPPEND);
5376 
5377 		if ((accmode & dac_granted) == accmode)
5378 			return (0);
5379 
5380 		goto privcheck;
5381 	}
5382 
5383 	/* Otherwise, check everyone else. */
5384 	if (file_mode & S_IXOTH)
5385 		dac_granted |= VEXEC;
5386 	if (file_mode & S_IROTH)
5387 		dac_granted |= VREAD;
5388 	if (file_mode & S_IWOTH)
5389 		dac_granted |= (VWRITE | VAPPEND);
5390 	if ((accmode & dac_granted) == accmode)
5391 		return (0);
5392 
5393 privcheck:
5394 	/*
5395 	 * Build a privilege mask to determine if the set of privileges
5396 	 * satisfies the requirements when combined with the granted mask
5397 	 * from above.  For each privilege, if the privilege is required,
5398 	 * bitwise or the request type onto the priv_granted mask.
5399 	 */
5400 	priv_granted = 0;
5401 
5402 	if (type == VDIR) {
5403 		/*
5404 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5405 		 * requests, instead of PRIV_VFS_EXEC.
5406 		 */
5407 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5408 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5409 			priv_granted |= VEXEC;
5410 	} else {
5411 		/*
5412 		 * Ensure that at least one execute bit is on. Otherwise,
5413 		 * a privileged user will always succeed, and we don't want
5414 		 * this to happen unless the file really is executable.
5415 		 */
5416 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5417 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5418 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5419 			priv_granted |= VEXEC;
5420 	}
5421 
5422 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5423 	    !priv_check_cred(cred, PRIV_VFS_READ))
5424 		priv_granted |= VREAD;
5425 
5426 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5427 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5428 		priv_granted |= (VWRITE | VAPPEND);
5429 
5430 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5431 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5432 		priv_granted |= VADMIN;
5433 
5434 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5435 		return (0);
5436 	}
5437 
5438 	return ((accmode & VADMIN) ? EPERM : EACCES);
5439 }
5440 
5441 /*
5442  * Credential check based on process requesting service, and per-attribute
5443  * permissions.
5444  */
5445 int
5446 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5447     struct thread *td, accmode_t accmode)
5448 {
5449 
5450 	/*
5451 	 * Kernel-invoked always succeeds.
5452 	 */
5453 	if (cred == NOCRED)
5454 		return (0);
5455 
5456 	/*
5457 	 * Do not allow privileged processes in jail to directly manipulate
5458 	 * system attributes.
5459 	 */
5460 	switch (attrnamespace) {
5461 	case EXTATTR_NAMESPACE_SYSTEM:
5462 		/* Potentially should be: return (EPERM); */
5463 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5464 	case EXTATTR_NAMESPACE_USER:
5465 		return (VOP_ACCESS(vp, accmode, cred, td));
5466 	default:
5467 		return (EPERM);
5468 	}
5469 }
5470 
5471 #ifdef DEBUG_VFS_LOCKS
5472 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5473 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5474     "Drop into debugger on lock violation");
5475 
5476 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5477 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5478     0, "Check for interlock across VOPs");
5479 
5480 int vfs_badlock_print = 1;	/* Print lock violations. */
5481 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5482     0, "Print lock violations");
5483 
5484 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5485 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5486     0, "Print vnode details on lock violations");
5487 
5488 #ifdef KDB
5489 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5490 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5491     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5492 #endif
5493 
5494 static void
5495 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5496 {
5497 
5498 #ifdef KDB
5499 	if (vfs_badlock_backtrace)
5500 		kdb_backtrace();
5501 #endif
5502 	if (vfs_badlock_vnode)
5503 		vn_printf(vp, "vnode ");
5504 	if (vfs_badlock_print)
5505 		printf("%s: %p %s\n", str, (void *)vp, msg);
5506 	if (vfs_badlock_ddb)
5507 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5508 }
5509 
5510 void
5511 assert_vi_locked(struct vnode *vp, const char *str)
5512 {
5513 
5514 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5515 		vfs_badlock("interlock is not locked but should be", str, vp);
5516 }
5517 
5518 void
5519 assert_vi_unlocked(struct vnode *vp, const char *str)
5520 {
5521 
5522 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5523 		vfs_badlock("interlock is locked but should not be", str, vp);
5524 }
5525 
5526 void
5527 assert_vop_locked(struct vnode *vp, const char *str)
5528 {
5529 	if (KERNEL_PANICKED() || vp == NULL)
5530 		return;
5531 
5532 #ifdef WITNESS
5533 	if ((vp->v_irflag & VIRF_CROSSMP) == 0 &&
5534 	    witness_is_owned(&vp->v_vnlock->lock_object) == -1)
5535 #else
5536 	int locked = VOP_ISLOCKED(vp);
5537 	if (locked == 0 || locked == LK_EXCLOTHER)
5538 #endif
5539 		vfs_badlock("is not locked but should be", str, vp);
5540 }
5541 
5542 void
5543 assert_vop_unlocked(struct vnode *vp, const char *str)
5544 {
5545 	if (KERNEL_PANICKED() || vp == NULL)
5546 		return;
5547 
5548 #ifdef WITNESS
5549 	if ((vp->v_irflag & VIRF_CROSSMP) == 0 &&
5550 	    witness_is_owned(&vp->v_vnlock->lock_object) == 1)
5551 #else
5552 	if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5553 #endif
5554 		vfs_badlock("is locked but should not be", str, vp);
5555 }
5556 
5557 void
5558 assert_vop_elocked(struct vnode *vp, const char *str)
5559 {
5560 	if (KERNEL_PANICKED() || vp == NULL)
5561 		return;
5562 
5563 	if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5564 		vfs_badlock("is not exclusive locked but should be", str, vp);
5565 }
5566 #endif /* DEBUG_VFS_LOCKS */
5567 
5568 void
5569 vop_rename_fail(struct vop_rename_args *ap)
5570 {
5571 
5572 	if (ap->a_tvp != NULL)
5573 		vput(ap->a_tvp);
5574 	if (ap->a_tdvp == ap->a_tvp)
5575 		vrele(ap->a_tdvp);
5576 	else
5577 		vput(ap->a_tdvp);
5578 	vrele(ap->a_fdvp);
5579 	vrele(ap->a_fvp);
5580 }
5581 
5582 void
5583 vop_rename_pre(void *ap)
5584 {
5585 	struct vop_rename_args *a = ap;
5586 
5587 #ifdef DEBUG_VFS_LOCKS
5588 	if (a->a_tvp)
5589 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5590 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5591 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5592 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5593 
5594 	/* Check the source (from). */
5595 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5596 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5597 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5598 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5599 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5600 
5601 	/* Check the target. */
5602 	if (a->a_tvp)
5603 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5604 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5605 #endif
5606 	/*
5607 	 * It may be tempting to add vn_seqc_write_begin/end calls here and
5608 	 * in vop_rename_post but that's not going to work out since some
5609 	 * filesystems relookup vnodes mid-rename. This is probably a bug.
5610 	 *
5611 	 * For now filesystems are expected to do the relevant calls after they
5612 	 * decide what vnodes to operate on.
5613 	 */
5614 	if (a->a_tdvp != a->a_fdvp)
5615 		vhold(a->a_fdvp);
5616 	if (a->a_tvp != a->a_fvp)
5617 		vhold(a->a_fvp);
5618 	vhold(a->a_tdvp);
5619 	if (a->a_tvp)
5620 		vhold(a->a_tvp);
5621 }
5622 
5623 #ifdef DEBUG_VFS_LOCKS
5624 void
5625 vop_fplookup_vexec_debugpre(void *ap __unused)
5626 {
5627 
5628 	VFS_SMR_ASSERT_ENTERED();
5629 }
5630 
5631 void
5632 vop_fplookup_vexec_debugpost(void *ap, int rc)
5633 {
5634 	struct vop_fplookup_vexec_args *a;
5635 	struct vnode *vp;
5636 
5637 	a = ap;
5638 	vp = a->a_vp;
5639 
5640 	VFS_SMR_ASSERT_ENTERED();
5641 	if (rc == EOPNOTSUPP)
5642 		VNPASS(VN_IS_DOOMED(vp), vp);
5643 }
5644 
5645 void
5646 vop_fplookup_symlink_debugpre(void *ap __unused)
5647 {
5648 
5649 	VFS_SMR_ASSERT_ENTERED();
5650 }
5651 
5652 void
5653 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused)
5654 {
5655 
5656 	VFS_SMR_ASSERT_ENTERED();
5657 }
5658 
5659 static void
5660 vop_fsync_debugprepost(struct vnode *vp, const char *name)
5661 {
5662 	if (vp->v_type == VCHR)
5663 		;
5664 	else if (MNT_EXTENDED_SHARED(vp->v_mount))
5665 		ASSERT_VOP_LOCKED(vp, name);
5666 	else
5667 		ASSERT_VOP_ELOCKED(vp, name);
5668 }
5669 
5670 void
5671 vop_fsync_debugpre(void *a)
5672 {
5673 	struct vop_fsync_args *ap;
5674 
5675 	ap = a;
5676 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5677 }
5678 
5679 void
5680 vop_fsync_debugpost(void *a, int rc __unused)
5681 {
5682 	struct vop_fsync_args *ap;
5683 
5684 	ap = a;
5685 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5686 }
5687 
5688 void
5689 vop_fdatasync_debugpre(void *a)
5690 {
5691 	struct vop_fdatasync_args *ap;
5692 
5693 	ap = a;
5694 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5695 }
5696 
5697 void
5698 vop_fdatasync_debugpost(void *a, int rc __unused)
5699 {
5700 	struct vop_fdatasync_args *ap;
5701 
5702 	ap = a;
5703 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5704 }
5705 
5706 void
5707 vop_strategy_debugpre(void *ap)
5708 {
5709 	struct vop_strategy_args *a;
5710 	struct buf *bp;
5711 
5712 	a = ap;
5713 	bp = a->a_bp;
5714 
5715 	/*
5716 	 * Cluster ops lock their component buffers but not the IO container.
5717 	 */
5718 	if ((bp->b_flags & B_CLUSTER) != 0)
5719 		return;
5720 
5721 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5722 		if (vfs_badlock_print)
5723 			printf(
5724 			    "VOP_STRATEGY: bp is not locked but should be\n");
5725 		if (vfs_badlock_ddb)
5726 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5727 	}
5728 }
5729 
5730 void
5731 vop_lock_debugpre(void *ap)
5732 {
5733 	struct vop_lock1_args *a = ap;
5734 
5735 	if ((a->a_flags & LK_INTERLOCK) == 0)
5736 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5737 	else
5738 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5739 }
5740 
5741 void
5742 vop_lock_debugpost(void *ap, int rc)
5743 {
5744 	struct vop_lock1_args *a = ap;
5745 
5746 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5747 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5748 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5749 }
5750 
5751 void
5752 vop_unlock_debugpre(void *ap)
5753 {
5754 	struct vop_unlock_args *a = ap;
5755 	struct vnode *vp = a->a_vp;
5756 
5757 	VNPASS(vn_get_state(vp) != VSTATE_UNINITIALIZED, vp);
5758 	ASSERT_VOP_LOCKED(vp, "VOP_UNLOCK");
5759 }
5760 
5761 void
5762 vop_need_inactive_debugpre(void *ap)
5763 {
5764 	struct vop_need_inactive_args *a = ap;
5765 
5766 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5767 }
5768 
5769 void
5770 vop_need_inactive_debugpost(void *ap, int rc)
5771 {
5772 	struct vop_need_inactive_args *a = ap;
5773 
5774 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5775 }
5776 #endif
5777 
5778 void
5779 vop_create_pre(void *ap)
5780 {
5781 	struct vop_create_args *a;
5782 	struct vnode *dvp;
5783 
5784 	a = ap;
5785 	dvp = a->a_dvp;
5786 	vn_seqc_write_begin(dvp);
5787 }
5788 
5789 void
5790 vop_create_post(void *ap, int rc)
5791 {
5792 	struct vop_create_args *a;
5793 	struct vnode *dvp;
5794 
5795 	a = ap;
5796 	dvp = a->a_dvp;
5797 	vn_seqc_write_end(dvp);
5798 	if (!rc)
5799 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5800 }
5801 
5802 void
5803 vop_whiteout_pre(void *ap)
5804 {
5805 	struct vop_whiteout_args *a;
5806 	struct vnode *dvp;
5807 
5808 	a = ap;
5809 	dvp = a->a_dvp;
5810 	vn_seqc_write_begin(dvp);
5811 }
5812 
5813 void
5814 vop_whiteout_post(void *ap, int rc)
5815 {
5816 	struct vop_whiteout_args *a;
5817 	struct vnode *dvp;
5818 
5819 	a = ap;
5820 	dvp = a->a_dvp;
5821 	vn_seqc_write_end(dvp);
5822 }
5823 
5824 void
5825 vop_deleteextattr_pre(void *ap)
5826 {
5827 	struct vop_deleteextattr_args *a;
5828 	struct vnode *vp;
5829 
5830 	a = ap;
5831 	vp = a->a_vp;
5832 	vn_seqc_write_begin(vp);
5833 }
5834 
5835 void
5836 vop_deleteextattr_post(void *ap, int rc)
5837 {
5838 	struct vop_deleteextattr_args *a;
5839 	struct vnode *vp;
5840 
5841 	a = ap;
5842 	vp = a->a_vp;
5843 	vn_seqc_write_end(vp);
5844 	if (!rc)
5845 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5846 }
5847 
5848 void
5849 vop_link_pre(void *ap)
5850 {
5851 	struct vop_link_args *a;
5852 	struct vnode *vp, *tdvp;
5853 
5854 	a = ap;
5855 	vp = a->a_vp;
5856 	tdvp = a->a_tdvp;
5857 	vn_seqc_write_begin(vp);
5858 	vn_seqc_write_begin(tdvp);
5859 }
5860 
5861 void
5862 vop_link_post(void *ap, int rc)
5863 {
5864 	struct vop_link_args *a;
5865 	struct vnode *vp, *tdvp;
5866 
5867 	a = ap;
5868 	vp = a->a_vp;
5869 	tdvp = a->a_tdvp;
5870 	vn_seqc_write_end(vp);
5871 	vn_seqc_write_end(tdvp);
5872 	if (!rc) {
5873 		VFS_KNOTE_LOCKED(vp, NOTE_LINK);
5874 		VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE);
5875 	}
5876 }
5877 
5878 void
5879 vop_mkdir_pre(void *ap)
5880 {
5881 	struct vop_mkdir_args *a;
5882 	struct vnode *dvp;
5883 
5884 	a = ap;
5885 	dvp = a->a_dvp;
5886 	vn_seqc_write_begin(dvp);
5887 }
5888 
5889 void
5890 vop_mkdir_post(void *ap, int rc)
5891 {
5892 	struct vop_mkdir_args *a;
5893 	struct vnode *dvp;
5894 
5895 	a = ap;
5896 	dvp = a->a_dvp;
5897 	vn_seqc_write_end(dvp);
5898 	if (!rc)
5899 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5900 }
5901 
5902 #ifdef DEBUG_VFS_LOCKS
5903 void
5904 vop_mkdir_debugpost(void *ap, int rc)
5905 {
5906 	struct vop_mkdir_args *a;
5907 
5908 	a = ap;
5909 	if (!rc)
5910 		cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp);
5911 }
5912 #endif
5913 
5914 void
5915 vop_mknod_pre(void *ap)
5916 {
5917 	struct vop_mknod_args *a;
5918 	struct vnode *dvp;
5919 
5920 	a = ap;
5921 	dvp = a->a_dvp;
5922 	vn_seqc_write_begin(dvp);
5923 }
5924 
5925 void
5926 vop_mknod_post(void *ap, int rc)
5927 {
5928 	struct vop_mknod_args *a;
5929 	struct vnode *dvp;
5930 
5931 	a = ap;
5932 	dvp = a->a_dvp;
5933 	vn_seqc_write_end(dvp);
5934 	if (!rc)
5935 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5936 }
5937 
5938 void
5939 vop_reclaim_post(void *ap, int rc)
5940 {
5941 	struct vop_reclaim_args *a;
5942 	struct vnode *vp;
5943 
5944 	a = ap;
5945 	vp = a->a_vp;
5946 	ASSERT_VOP_IN_SEQC(vp);
5947 	if (!rc)
5948 		VFS_KNOTE_LOCKED(vp, NOTE_REVOKE);
5949 }
5950 
5951 void
5952 vop_remove_pre(void *ap)
5953 {
5954 	struct vop_remove_args *a;
5955 	struct vnode *dvp, *vp;
5956 
5957 	a = ap;
5958 	dvp = a->a_dvp;
5959 	vp = a->a_vp;
5960 	vn_seqc_write_begin(dvp);
5961 	vn_seqc_write_begin(vp);
5962 }
5963 
5964 void
5965 vop_remove_post(void *ap, int rc)
5966 {
5967 	struct vop_remove_args *a;
5968 	struct vnode *dvp, *vp;
5969 
5970 	a = ap;
5971 	dvp = a->a_dvp;
5972 	vp = a->a_vp;
5973 	vn_seqc_write_end(dvp);
5974 	vn_seqc_write_end(vp);
5975 	if (!rc) {
5976 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5977 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5978 	}
5979 }
5980 
5981 void
5982 vop_rename_post(void *ap, int rc)
5983 {
5984 	struct vop_rename_args *a = ap;
5985 	long hint;
5986 
5987 	if (!rc) {
5988 		hint = NOTE_WRITE;
5989 		if (a->a_fdvp == a->a_tdvp) {
5990 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5991 				hint |= NOTE_LINK;
5992 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5993 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5994 		} else {
5995 			hint |= NOTE_EXTEND;
5996 			if (a->a_fvp->v_type == VDIR)
5997 				hint |= NOTE_LINK;
5998 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5999 
6000 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
6001 			    a->a_tvp->v_type == VDIR)
6002 				hint &= ~NOTE_LINK;
6003 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
6004 		}
6005 
6006 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
6007 		if (a->a_tvp)
6008 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
6009 	}
6010 	if (a->a_tdvp != a->a_fdvp)
6011 		vdrop(a->a_fdvp);
6012 	if (a->a_tvp != a->a_fvp)
6013 		vdrop(a->a_fvp);
6014 	vdrop(a->a_tdvp);
6015 	if (a->a_tvp)
6016 		vdrop(a->a_tvp);
6017 }
6018 
6019 void
6020 vop_rmdir_pre(void *ap)
6021 {
6022 	struct vop_rmdir_args *a;
6023 	struct vnode *dvp, *vp;
6024 
6025 	a = ap;
6026 	dvp = a->a_dvp;
6027 	vp = a->a_vp;
6028 	vn_seqc_write_begin(dvp);
6029 	vn_seqc_write_begin(vp);
6030 }
6031 
6032 void
6033 vop_rmdir_post(void *ap, int rc)
6034 {
6035 	struct vop_rmdir_args *a;
6036 	struct vnode *dvp, *vp;
6037 
6038 	a = ap;
6039 	dvp = a->a_dvp;
6040 	vp = a->a_vp;
6041 	vn_seqc_write_end(dvp);
6042 	vn_seqc_write_end(vp);
6043 	if (!rc) {
6044 		vp->v_vflag |= VV_UNLINKED;
6045 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
6046 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
6047 	}
6048 }
6049 
6050 void
6051 vop_setattr_pre(void *ap)
6052 {
6053 	struct vop_setattr_args *a;
6054 	struct vnode *vp;
6055 
6056 	a = ap;
6057 	vp = a->a_vp;
6058 	vn_seqc_write_begin(vp);
6059 }
6060 
6061 void
6062 vop_setattr_post(void *ap, int rc)
6063 {
6064 	struct vop_setattr_args *a;
6065 	struct vnode *vp;
6066 
6067 	a = ap;
6068 	vp = a->a_vp;
6069 	vn_seqc_write_end(vp);
6070 	if (!rc)
6071 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
6072 }
6073 
6074 void
6075 vop_setacl_pre(void *ap)
6076 {
6077 	struct vop_setacl_args *a;
6078 	struct vnode *vp;
6079 
6080 	a = ap;
6081 	vp = a->a_vp;
6082 	vn_seqc_write_begin(vp);
6083 }
6084 
6085 void
6086 vop_setacl_post(void *ap, int rc __unused)
6087 {
6088 	struct vop_setacl_args *a;
6089 	struct vnode *vp;
6090 
6091 	a = ap;
6092 	vp = a->a_vp;
6093 	vn_seqc_write_end(vp);
6094 }
6095 
6096 void
6097 vop_setextattr_pre(void *ap)
6098 {
6099 	struct vop_setextattr_args *a;
6100 	struct vnode *vp;
6101 
6102 	a = ap;
6103 	vp = a->a_vp;
6104 	vn_seqc_write_begin(vp);
6105 }
6106 
6107 void
6108 vop_setextattr_post(void *ap, int rc)
6109 {
6110 	struct vop_setextattr_args *a;
6111 	struct vnode *vp;
6112 
6113 	a = ap;
6114 	vp = a->a_vp;
6115 	vn_seqc_write_end(vp);
6116 	if (!rc)
6117 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
6118 }
6119 
6120 void
6121 vop_symlink_pre(void *ap)
6122 {
6123 	struct vop_symlink_args *a;
6124 	struct vnode *dvp;
6125 
6126 	a = ap;
6127 	dvp = a->a_dvp;
6128 	vn_seqc_write_begin(dvp);
6129 }
6130 
6131 void
6132 vop_symlink_post(void *ap, int rc)
6133 {
6134 	struct vop_symlink_args *a;
6135 	struct vnode *dvp;
6136 
6137 	a = ap;
6138 	dvp = a->a_dvp;
6139 	vn_seqc_write_end(dvp);
6140 	if (!rc)
6141 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6142 }
6143 
6144 void
6145 vop_open_post(void *ap, int rc)
6146 {
6147 	struct vop_open_args *a = ap;
6148 
6149 	if (!rc)
6150 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
6151 }
6152 
6153 void
6154 vop_close_post(void *ap, int rc)
6155 {
6156 	struct vop_close_args *a = ap;
6157 
6158 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
6159 	    !VN_IS_DOOMED(a->a_vp))) {
6160 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
6161 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
6162 	}
6163 }
6164 
6165 void
6166 vop_read_post(void *ap, int rc)
6167 {
6168 	struct vop_read_args *a = ap;
6169 
6170 	if (!rc)
6171 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6172 }
6173 
6174 void
6175 vop_read_pgcache_post(void *ap, int rc)
6176 {
6177 	struct vop_read_pgcache_args *a = ap;
6178 
6179 	if (!rc)
6180 		VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ);
6181 }
6182 
6183 void
6184 vop_readdir_post(void *ap, int rc)
6185 {
6186 	struct vop_readdir_args *a = ap;
6187 
6188 	if (!rc)
6189 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6190 }
6191 
6192 static struct knlist fs_knlist;
6193 
6194 static void
6195 vfs_event_init(void *arg)
6196 {
6197 	knlist_init_mtx(&fs_knlist, NULL);
6198 }
6199 /* XXX - correct order? */
6200 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
6201 
6202 void
6203 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
6204 {
6205 
6206 	KNOTE_UNLOCKED(&fs_knlist, event);
6207 }
6208 
6209 static int	filt_fsattach(struct knote *kn);
6210 static void	filt_fsdetach(struct knote *kn);
6211 static int	filt_fsevent(struct knote *kn, long hint);
6212 
6213 struct filterops fs_filtops = {
6214 	.f_isfd = 0,
6215 	.f_attach = filt_fsattach,
6216 	.f_detach = filt_fsdetach,
6217 	.f_event = filt_fsevent
6218 };
6219 
6220 static int
6221 filt_fsattach(struct knote *kn)
6222 {
6223 
6224 	kn->kn_flags |= EV_CLEAR;
6225 	knlist_add(&fs_knlist, kn, 0);
6226 	return (0);
6227 }
6228 
6229 static void
6230 filt_fsdetach(struct knote *kn)
6231 {
6232 
6233 	knlist_remove(&fs_knlist, kn, 0);
6234 }
6235 
6236 static int
6237 filt_fsevent(struct knote *kn, long hint)
6238 {
6239 
6240 	kn->kn_fflags |= kn->kn_sfflags & hint;
6241 
6242 	return (kn->kn_fflags != 0);
6243 }
6244 
6245 static int
6246 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
6247 {
6248 	struct vfsidctl vc;
6249 	int error;
6250 	struct mount *mp;
6251 
6252 	error = SYSCTL_IN(req, &vc, sizeof(vc));
6253 	if (error)
6254 		return (error);
6255 	if (vc.vc_vers != VFS_CTL_VERS1)
6256 		return (EINVAL);
6257 	mp = vfs_getvfs(&vc.vc_fsid);
6258 	if (mp == NULL)
6259 		return (ENOENT);
6260 	/* ensure that a specific sysctl goes to the right filesystem. */
6261 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
6262 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
6263 		vfs_rel(mp);
6264 		return (EINVAL);
6265 	}
6266 	VCTLTOREQ(&vc, req);
6267 	error = VFS_SYSCTL(mp, vc.vc_op, req);
6268 	vfs_rel(mp);
6269 	return (error);
6270 }
6271 
6272 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
6273     NULL, 0, sysctl_vfs_ctl, "",
6274     "Sysctl by fsid");
6275 
6276 /*
6277  * Function to initialize a va_filerev field sensibly.
6278  * XXX: Wouldn't a random number make a lot more sense ??
6279  */
6280 u_quad_t
6281 init_va_filerev(void)
6282 {
6283 	struct bintime bt;
6284 
6285 	getbinuptime(&bt);
6286 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
6287 }
6288 
6289 static int	filt_vfsread(struct knote *kn, long hint);
6290 static int	filt_vfswrite(struct knote *kn, long hint);
6291 static int	filt_vfsvnode(struct knote *kn, long hint);
6292 static void	filt_vfsdetach(struct knote *kn);
6293 static struct filterops vfsread_filtops = {
6294 	.f_isfd = 1,
6295 	.f_detach = filt_vfsdetach,
6296 	.f_event = filt_vfsread
6297 };
6298 static struct filterops vfswrite_filtops = {
6299 	.f_isfd = 1,
6300 	.f_detach = filt_vfsdetach,
6301 	.f_event = filt_vfswrite
6302 };
6303 static struct filterops vfsvnode_filtops = {
6304 	.f_isfd = 1,
6305 	.f_detach = filt_vfsdetach,
6306 	.f_event = filt_vfsvnode
6307 };
6308 
6309 static void
6310 vfs_knllock(void *arg)
6311 {
6312 	struct vnode *vp = arg;
6313 
6314 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
6315 }
6316 
6317 static void
6318 vfs_knlunlock(void *arg)
6319 {
6320 	struct vnode *vp = arg;
6321 
6322 	VOP_UNLOCK(vp);
6323 }
6324 
6325 static void
6326 vfs_knl_assert_lock(void *arg, int what)
6327 {
6328 #ifdef DEBUG_VFS_LOCKS
6329 	struct vnode *vp = arg;
6330 
6331 	if (what == LA_LOCKED)
6332 		ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
6333 	else
6334 		ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
6335 #endif
6336 }
6337 
6338 int
6339 vfs_kqfilter(struct vop_kqfilter_args *ap)
6340 {
6341 	struct vnode *vp = ap->a_vp;
6342 	struct knote *kn = ap->a_kn;
6343 	struct knlist *knl;
6344 
6345 	KASSERT(vp->v_type != VFIFO || (kn->kn_filter != EVFILT_READ &&
6346 	    kn->kn_filter != EVFILT_WRITE),
6347 	    ("READ/WRITE filter on a FIFO leaked through"));
6348 	switch (kn->kn_filter) {
6349 	case EVFILT_READ:
6350 		kn->kn_fop = &vfsread_filtops;
6351 		break;
6352 	case EVFILT_WRITE:
6353 		kn->kn_fop = &vfswrite_filtops;
6354 		break;
6355 	case EVFILT_VNODE:
6356 		kn->kn_fop = &vfsvnode_filtops;
6357 		break;
6358 	default:
6359 		return (EINVAL);
6360 	}
6361 
6362 	kn->kn_hook = (caddr_t)vp;
6363 
6364 	v_addpollinfo(vp);
6365 	if (vp->v_pollinfo == NULL)
6366 		return (ENOMEM);
6367 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
6368 	vhold(vp);
6369 	knlist_add(knl, kn, 0);
6370 
6371 	return (0);
6372 }
6373 
6374 /*
6375  * Detach knote from vnode
6376  */
6377 static void
6378 filt_vfsdetach(struct knote *kn)
6379 {
6380 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6381 
6382 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
6383 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
6384 	vdrop(vp);
6385 }
6386 
6387 /*ARGSUSED*/
6388 static int
6389 filt_vfsread(struct knote *kn, long hint)
6390 {
6391 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6392 	off_t size;
6393 	int res;
6394 
6395 	/*
6396 	 * filesystem is gone, so set the EOF flag and schedule
6397 	 * the knote for deletion.
6398 	 */
6399 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6400 		VI_LOCK(vp);
6401 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6402 		VI_UNLOCK(vp);
6403 		return (1);
6404 	}
6405 
6406 	if (vn_getsize_locked(vp, &size, curthread->td_ucred) != 0)
6407 		return (0);
6408 
6409 	VI_LOCK(vp);
6410 	kn->kn_data = size - kn->kn_fp->f_offset;
6411 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
6412 	VI_UNLOCK(vp);
6413 	return (res);
6414 }
6415 
6416 /*ARGSUSED*/
6417 static int
6418 filt_vfswrite(struct knote *kn, long hint)
6419 {
6420 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6421 
6422 	VI_LOCK(vp);
6423 
6424 	/*
6425 	 * filesystem is gone, so set the EOF flag and schedule
6426 	 * the knote for deletion.
6427 	 */
6428 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
6429 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6430 
6431 	kn->kn_data = 0;
6432 	VI_UNLOCK(vp);
6433 	return (1);
6434 }
6435 
6436 static int
6437 filt_vfsvnode(struct knote *kn, long hint)
6438 {
6439 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6440 	int res;
6441 
6442 	VI_LOCK(vp);
6443 	if (kn->kn_sfflags & hint)
6444 		kn->kn_fflags |= hint;
6445 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6446 		kn->kn_flags |= EV_EOF;
6447 		VI_UNLOCK(vp);
6448 		return (1);
6449 	}
6450 	res = (kn->kn_fflags != 0);
6451 	VI_UNLOCK(vp);
6452 	return (res);
6453 }
6454 
6455 int
6456 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6457 {
6458 	int error;
6459 
6460 	if (dp->d_reclen > ap->a_uio->uio_resid)
6461 		return (ENAMETOOLONG);
6462 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
6463 	if (error) {
6464 		if (ap->a_ncookies != NULL) {
6465 			if (ap->a_cookies != NULL)
6466 				free(ap->a_cookies, M_TEMP);
6467 			ap->a_cookies = NULL;
6468 			*ap->a_ncookies = 0;
6469 		}
6470 		return (error);
6471 	}
6472 	if (ap->a_ncookies == NULL)
6473 		return (0);
6474 
6475 	KASSERT(ap->a_cookies,
6476 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6477 
6478 	*ap->a_cookies = realloc(*ap->a_cookies,
6479 	    (*ap->a_ncookies + 1) * sizeof(uint64_t), M_TEMP, M_WAITOK | M_ZERO);
6480 	(*ap->a_cookies)[*ap->a_ncookies] = off;
6481 	*ap->a_ncookies += 1;
6482 	return (0);
6483 }
6484 
6485 /*
6486  * The purpose of this routine is to remove granularity from accmode_t,
6487  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6488  * VADMIN and VAPPEND.
6489  *
6490  * If it returns 0, the caller is supposed to continue with the usual
6491  * access checks using 'accmode' as modified by this routine.  If it
6492  * returns nonzero value, the caller is supposed to return that value
6493  * as errno.
6494  *
6495  * Note that after this routine runs, accmode may be zero.
6496  */
6497 int
6498 vfs_unixify_accmode(accmode_t *accmode)
6499 {
6500 	/*
6501 	 * There is no way to specify explicit "deny" rule using
6502 	 * file mode or POSIX.1e ACLs.
6503 	 */
6504 	if (*accmode & VEXPLICIT_DENY) {
6505 		*accmode = 0;
6506 		return (0);
6507 	}
6508 
6509 	/*
6510 	 * None of these can be translated into usual access bits.
6511 	 * Also, the common case for NFSv4 ACLs is to not contain
6512 	 * either of these bits. Caller should check for VWRITE
6513 	 * on the containing directory instead.
6514 	 */
6515 	if (*accmode & (VDELETE_CHILD | VDELETE))
6516 		return (EPERM);
6517 
6518 	if (*accmode & VADMIN_PERMS) {
6519 		*accmode &= ~VADMIN_PERMS;
6520 		*accmode |= VADMIN;
6521 	}
6522 
6523 	/*
6524 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6525 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6526 	 */
6527 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6528 
6529 	return (0);
6530 }
6531 
6532 /*
6533  * Clear out a doomed vnode (if any) and replace it with a new one as long
6534  * as the fs is not being unmounted. Return the root vnode to the caller.
6535  */
6536 static int __noinline
6537 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6538 {
6539 	struct vnode *vp;
6540 	int error;
6541 
6542 restart:
6543 	if (mp->mnt_rootvnode != NULL) {
6544 		MNT_ILOCK(mp);
6545 		vp = mp->mnt_rootvnode;
6546 		if (vp != NULL) {
6547 			if (!VN_IS_DOOMED(vp)) {
6548 				vrefact(vp);
6549 				MNT_IUNLOCK(mp);
6550 				error = vn_lock(vp, flags);
6551 				if (error == 0) {
6552 					*vpp = vp;
6553 					return (0);
6554 				}
6555 				vrele(vp);
6556 				goto restart;
6557 			}
6558 			/*
6559 			 * Clear the old one.
6560 			 */
6561 			mp->mnt_rootvnode = NULL;
6562 		}
6563 		MNT_IUNLOCK(mp);
6564 		if (vp != NULL) {
6565 			vfs_op_barrier_wait(mp);
6566 			vrele(vp);
6567 		}
6568 	}
6569 	error = VFS_CACHEDROOT(mp, flags, vpp);
6570 	if (error != 0)
6571 		return (error);
6572 	if (mp->mnt_vfs_ops == 0) {
6573 		MNT_ILOCK(mp);
6574 		if (mp->mnt_vfs_ops != 0) {
6575 			MNT_IUNLOCK(mp);
6576 			return (0);
6577 		}
6578 		if (mp->mnt_rootvnode == NULL) {
6579 			vrefact(*vpp);
6580 			mp->mnt_rootvnode = *vpp;
6581 		} else {
6582 			if (mp->mnt_rootvnode != *vpp) {
6583 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6584 					panic("%s: mismatch between vnode returned "
6585 					    " by VFS_CACHEDROOT and the one cached "
6586 					    " (%p != %p)",
6587 					    __func__, *vpp, mp->mnt_rootvnode);
6588 				}
6589 			}
6590 		}
6591 		MNT_IUNLOCK(mp);
6592 	}
6593 	return (0);
6594 }
6595 
6596 int
6597 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6598 {
6599 	struct mount_pcpu *mpcpu;
6600 	struct vnode *vp;
6601 	int error;
6602 
6603 	if (!vfs_op_thread_enter(mp, mpcpu))
6604 		return (vfs_cache_root_fallback(mp, flags, vpp));
6605 	vp = atomic_load_ptr(&mp->mnt_rootvnode);
6606 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6607 		vfs_op_thread_exit(mp, mpcpu);
6608 		return (vfs_cache_root_fallback(mp, flags, vpp));
6609 	}
6610 	vrefact(vp);
6611 	vfs_op_thread_exit(mp, mpcpu);
6612 	error = vn_lock(vp, flags);
6613 	if (error != 0) {
6614 		vrele(vp);
6615 		return (vfs_cache_root_fallback(mp, flags, vpp));
6616 	}
6617 	*vpp = vp;
6618 	return (0);
6619 }
6620 
6621 struct vnode *
6622 vfs_cache_root_clear(struct mount *mp)
6623 {
6624 	struct vnode *vp;
6625 
6626 	/*
6627 	 * ops > 0 guarantees there is nobody who can see this vnode
6628 	 */
6629 	MPASS(mp->mnt_vfs_ops > 0);
6630 	vp = mp->mnt_rootvnode;
6631 	if (vp != NULL)
6632 		vn_seqc_write_begin(vp);
6633 	mp->mnt_rootvnode = NULL;
6634 	return (vp);
6635 }
6636 
6637 void
6638 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6639 {
6640 
6641 	MPASS(mp->mnt_vfs_ops > 0);
6642 	vrefact(vp);
6643 	mp->mnt_rootvnode = vp;
6644 }
6645 
6646 /*
6647  * These are helper functions for filesystems to traverse all
6648  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6649  *
6650  * This interface replaces MNT_VNODE_FOREACH.
6651  */
6652 
6653 struct vnode *
6654 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6655 {
6656 	struct vnode *vp;
6657 
6658 	maybe_yield();
6659 	MNT_ILOCK(mp);
6660 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6661 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6662 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6663 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6664 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6665 			continue;
6666 		VI_LOCK(vp);
6667 		if (VN_IS_DOOMED(vp)) {
6668 			VI_UNLOCK(vp);
6669 			continue;
6670 		}
6671 		break;
6672 	}
6673 	if (vp == NULL) {
6674 		__mnt_vnode_markerfree_all(mvp, mp);
6675 		/* MNT_IUNLOCK(mp); -- done in above function */
6676 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6677 		return (NULL);
6678 	}
6679 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6680 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6681 	MNT_IUNLOCK(mp);
6682 	return (vp);
6683 }
6684 
6685 struct vnode *
6686 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6687 {
6688 	struct vnode *vp;
6689 
6690 	*mvp = vn_alloc_marker(mp);
6691 	MNT_ILOCK(mp);
6692 	MNT_REF(mp);
6693 
6694 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6695 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6696 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6697 			continue;
6698 		VI_LOCK(vp);
6699 		if (VN_IS_DOOMED(vp)) {
6700 			VI_UNLOCK(vp);
6701 			continue;
6702 		}
6703 		break;
6704 	}
6705 	if (vp == NULL) {
6706 		MNT_REL(mp);
6707 		MNT_IUNLOCK(mp);
6708 		vn_free_marker(*mvp);
6709 		*mvp = NULL;
6710 		return (NULL);
6711 	}
6712 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6713 	MNT_IUNLOCK(mp);
6714 	return (vp);
6715 }
6716 
6717 void
6718 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6719 {
6720 
6721 	if (*mvp == NULL) {
6722 		MNT_IUNLOCK(mp);
6723 		return;
6724 	}
6725 
6726 	mtx_assert(MNT_MTX(mp), MA_OWNED);
6727 
6728 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6729 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6730 	MNT_REL(mp);
6731 	MNT_IUNLOCK(mp);
6732 	vn_free_marker(*mvp);
6733 	*mvp = NULL;
6734 }
6735 
6736 /*
6737  * These are helper functions for filesystems to traverse their
6738  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6739  */
6740 static void
6741 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6742 {
6743 
6744 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6745 
6746 	MNT_ILOCK(mp);
6747 	MNT_REL(mp);
6748 	MNT_IUNLOCK(mp);
6749 	vn_free_marker(*mvp);
6750 	*mvp = NULL;
6751 }
6752 
6753 /*
6754  * Relock the mp mount vnode list lock with the vp vnode interlock in the
6755  * conventional lock order during mnt_vnode_next_lazy iteration.
6756  *
6757  * On entry, the mount vnode list lock is held and the vnode interlock is not.
6758  * The list lock is dropped and reacquired.  On success, both locks are held.
6759  * On failure, the mount vnode list lock is held but the vnode interlock is
6760  * not, and the procedure may have yielded.
6761  */
6762 static bool
6763 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6764     struct vnode *vp)
6765 {
6766 
6767 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6768 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6769 	    ("%s: bad marker", __func__));
6770 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6771 	    ("%s: inappropriate vnode", __func__));
6772 	ASSERT_VI_UNLOCKED(vp, __func__);
6773 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6774 
6775 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6776 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6777 
6778 	/*
6779 	 * Note we may be racing against vdrop which transitioned the hold
6780 	 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6781 	 * if we are the only user after we get the interlock we will just
6782 	 * vdrop.
6783 	 */
6784 	vhold(vp);
6785 	mtx_unlock(&mp->mnt_listmtx);
6786 	VI_LOCK(vp);
6787 	if (VN_IS_DOOMED(vp)) {
6788 		VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6789 		goto out_lost;
6790 	}
6791 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6792 	/*
6793 	 * There is nothing to do if we are the last user.
6794 	 */
6795 	if (!refcount_release_if_not_last(&vp->v_holdcnt))
6796 		goto out_lost;
6797 	mtx_lock(&mp->mnt_listmtx);
6798 	return (true);
6799 out_lost:
6800 	vdropl(vp);
6801 	maybe_yield();
6802 	mtx_lock(&mp->mnt_listmtx);
6803 	return (false);
6804 }
6805 
6806 static struct vnode *
6807 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6808     void *cbarg)
6809 {
6810 	struct vnode *vp;
6811 
6812 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6813 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6814 restart:
6815 	vp = TAILQ_NEXT(*mvp, v_lazylist);
6816 	while (vp != NULL) {
6817 		if (vp->v_type == VMARKER) {
6818 			vp = TAILQ_NEXT(vp, v_lazylist);
6819 			continue;
6820 		}
6821 		/*
6822 		 * See if we want to process the vnode. Note we may encounter a
6823 		 * long string of vnodes we don't care about and hog the list
6824 		 * as a result. Check for it and requeue the marker.
6825 		 */
6826 		VNPASS(!VN_IS_DOOMED(vp), vp);
6827 		if (!cb(vp, cbarg)) {
6828 			if (!should_yield()) {
6829 				vp = TAILQ_NEXT(vp, v_lazylist);
6830 				continue;
6831 			}
6832 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6833 			    v_lazylist);
6834 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6835 			    v_lazylist);
6836 			mtx_unlock(&mp->mnt_listmtx);
6837 			kern_yield(PRI_USER);
6838 			mtx_lock(&mp->mnt_listmtx);
6839 			goto restart;
6840 		}
6841 		/*
6842 		 * Try-lock because this is the wrong lock order.
6843 		 */
6844 		if (!VI_TRYLOCK(vp) &&
6845 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6846 			goto restart;
6847 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6848 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6849 		    ("alien vnode on the lazy list %p %p", vp, mp));
6850 		VNPASS(vp->v_mount == mp, vp);
6851 		VNPASS(!VN_IS_DOOMED(vp), vp);
6852 		break;
6853 	}
6854 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6855 
6856 	/* Check if we are done */
6857 	if (vp == NULL) {
6858 		mtx_unlock(&mp->mnt_listmtx);
6859 		mnt_vnode_markerfree_lazy(mvp, mp);
6860 		return (NULL);
6861 	}
6862 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6863 	mtx_unlock(&mp->mnt_listmtx);
6864 	ASSERT_VI_LOCKED(vp, "lazy iter");
6865 	return (vp);
6866 }
6867 
6868 struct vnode *
6869 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6870     void *cbarg)
6871 {
6872 
6873 	maybe_yield();
6874 	mtx_lock(&mp->mnt_listmtx);
6875 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6876 }
6877 
6878 struct vnode *
6879 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6880     void *cbarg)
6881 {
6882 	struct vnode *vp;
6883 
6884 	if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6885 		return (NULL);
6886 
6887 	*mvp = vn_alloc_marker(mp);
6888 	MNT_ILOCK(mp);
6889 	MNT_REF(mp);
6890 	MNT_IUNLOCK(mp);
6891 
6892 	mtx_lock(&mp->mnt_listmtx);
6893 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6894 	if (vp == NULL) {
6895 		mtx_unlock(&mp->mnt_listmtx);
6896 		mnt_vnode_markerfree_lazy(mvp, mp);
6897 		return (NULL);
6898 	}
6899 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6900 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6901 }
6902 
6903 void
6904 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6905 {
6906 
6907 	if (*mvp == NULL)
6908 		return;
6909 
6910 	mtx_lock(&mp->mnt_listmtx);
6911 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6912 	mtx_unlock(&mp->mnt_listmtx);
6913 	mnt_vnode_markerfree_lazy(mvp, mp);
6914 }
6915 
6916 int
6917 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6918 {
6919 
6920 	if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6921 		cnp->cn_flags &= ~NOEXECCHECK;
6922 		return (0);
6923 	}
6924 
6925 	return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, curthread));
6926 }
6927 
6928 /*
6929  * Do not use this variant unless you have means other than the hold count
6930  * to prevent the vnode from getting freed.
6931  */
6932 void
6933 vn_seqc_write_begin_locked(struct vnode *vp)
6934 {
6935 
6936 	ASSERT_VI_LOCKED(vp, __func__);
6937 	VNPASS(vp->v_holdcnt > 0, vp);
6938 	VNPASS(vp->v_seqc_users >= 0, vp);
6939 	vp->v_seqc_users++;
6940 	if (vp->v_seqc_users == 1)
6941 		seqc_sleepable_write_begin(&vp->v_seqc);
6942 }
6943 
6944 void
6945 vn_seqc_write_begin(struct vnode *vp)
6946 {
6947 
6948 	VI_LOCK(vp);
6949 	vn_seqc_write_begin_locked(vp);
6950 	VI_UNLOCK(vp);
6951 }
6952 
6953 void
6954 vn_seqc_write_end_locked(struct vnode *vp)
6955 {
6956 
6957 	ASSERT_VI_LOCKED(vp, __func__);
6958 	VNPASS(vp->v_seqc_users > 0, vp);
6959 	vp->v_seqc_users--;
6960 	if (vp->v_seqc_users == 0)
6961 		seqc_sleepable_write_end(&vp->v_seqc);
6962 }
6963 
6964 void
6965 vn_seqc_write_end(struct vnode *vp)
6966 {
6967 
6968 	VI_LOCK(vp);
6969 	vn_seqc_write_end_locked(vp);
6970 	VI_UNLOCK(vp);
6971 }
6972 
6973 /*
6974  * Special case handling for allocating and freeing vnodes.
6975  *
6976  * The counter remains unchanged on free so that a doomed vnode will
6977  * keep testing as in modify as long as it is accessible with SMR.
6978  */
6979 static void
6980 vn_seqc_init(struct vnode *vp)
6981 {
6982 
6983 	vp->v_seqc = 0;
6984 	vp->v_seqc_users = 0;
6985 }
6986 
6987 static void
6988 vn_seqc_write_end_free(struct vnode *vp)
6989 {
6990 
6991 	VNPASS(seqc_in_modify(vp->v_seqc), vp);
6992 	VNPASS(vp->v_seqc_users == 1, vp);
6993 }
6994 
6995 void
6996 vn_irflag_set_locked(struct vnode *vp, short toset)
6997 {
6998 	short flags;
6999 
7000 	ASSERT_VI_LOCKED(vp, __func__);
7001 	flags = vn_irflag_read(vp);
7002 	VNASSERT((flags & toset) == 0, vp,
7003 	    ("%s: some of the passed flags already set (have %d, passed %d)\n",
7004 	    __func__, flags, toset));
7005 	atomic_store_short(&vp->v_irflag, flags | toset);
7006 }
7007 
7008 void
7009 vn_irflag_set(struct vnode *vp, short toset)
7010 {
7011 
7012 	VI_LOCK(vp);
7013 	vn_irflag_set_locked(vp, toset);
7014 	VI_UNLOCK(vp);
7015 }
7016 
7017 void
7018 vn_irflag_set_cond_locked(struct vnode *vp, short toset)
7019 {
7020 	short flags;
7021 
7022 	ASSERT_VI_LOCKED(vp, __func__);
7023 	flags = vn_irflag_read(vp);
7024 	atomic_store_short(&vp->v_irflag, flags | toset);
7025 }
7026 
7027 void
7028 vn_irflag_set_cond(struct vnode *vp, short toset)
7029 {
7030 
7031 	VI_LOCK(vp);
7032 	vn_irflag_set_cond_locked(vp, toset);
7033 	VI_UNLOCK(vp);
7034 }
7035 
7036 void
7037 vn_irflag_unset_locked(struct vnode *vp, short tounset)
7038 {
7039 	short flags;
7040 
7041 	ASSERT_VI_LOCKED(vp, __func__);
7042 	flags = vn_irflag_read(vp);
7043 	VNASSERT((flags & tounset) == tounset, vp,
7044 	    ("%s: some of the passed flags not set (have %d, passed %d)\n",
7045 	    __func__, flags, tounset));
7046 	atomic_store_short(&vp->v_irflag, flags & ~tounset);
7047 }
7048 
7049 void
7050 vn_irflag_unset(struct vnode *vp, short tounset)
7051 {
7052 
7053 	VI_LOCK(vp);
7054 	vn_irflag_unset_locked(vp, tounset);
7055 	VI_UNLOCK(vp);
7056 }
7057 
7058 int
7059 vn_getsize_locked(struct vnode *vp, off_t *size, struct ucred *cred)
7060 {
7061 	struct vattr vattr;
7062 	int error;
7063 
7064 	ASSERT_VOP_LOCKED(vp, __func__);
7065 	error = VOP_GETATTR(vp, &vattr, cred);
7066 	if (__predict_true(error == 0)) {
7067 		if (vattr.va_size <= OFF_MAX)
7068 			*size = vattr.va_size;
7069 		else
7070 			error = EFBIG;
7071 	}
7072 	return (error);
7073 }
7074 
7075 int
7076 vn_getsize(struct vnode *vp, off_t *size, struct ucred *cred)
7077 {
7078 	int error;
7079 
7080 	VOP_LOCK(vp, LK_SHARED);
7081 	error = vn_getsize_locked(vp, size, cred);
7082 	VOP_UNLOCK(vp);
7083 	return (error);
7084 }
7085 
7086 #ifdef INVARIANTS
7087 void
7088 vn_set_state_validate(struct vnode *vp, __enum_uint8(vstate) state)
7089 {
7090 
7091 	switch (vp->v_state) {
7092 	case VSTATE_UNINITIALIZED:
7093 		switch (state) {
7094 		case VSTATE_CONSTRUCTED:
7095 		case VSTATE_DESTROYING:
7096 			return;
7097 		default:
7098 			break;
7099 		}
7100 		break;
7101 	case VSTATE_CONSTRUCTED:
7102 		ASSERT_VOP_ELOCKED(vp, __func__);
7103 		switch (state) {
7104 		case VSTATE_DESTROYING:
7105 			return;
7106 		default:
7107 			break;
7108 		}
7109 		break;
7110 	case VSTATE_DESTROYING:
7111 		ASSERT_VOP_ELOCKED(vp, __func__);
7112 		switch (state) {
7113 		case VSTATE_DEAD:
7114 			return;
7115 		default:
7116 			break;
7117 		}
7118 		break;
7119 	case VSTATE_DEAD:
7120 		switch (state) {
7121 		case VSTATE_UNINITIALIZED:
7122 			return;
7123 		default:
7124 			break;
7125 		}
7126 		break;
7127 	}
7128 
7129 	vn_printf(vp, "invalid state transition %d -> %d\n", vp->v_state, state);
7130 	panic("invalid state transition %d -> %d\n", vp->v_state, state);
7131 }
7132 #endif
7133