xref: /freebsd/sys/kern/vfs_subr.c (revision 734e82fe33aa764367791a7d603b383996c6b40b)
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 u_long deferred_inact;
200 SYSCTL_ULONG(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD,
201     &deferred_inact, 0, "Number of times inactive processing was deferred");
202 
203 /* To keep more than one thread at a time from running vfs_getnewfsid */
204 static struct mtx mntid_mtx;
205 
206 /*
207  * Lock for any access to the following:
208  *	vnode_list
209  *	numvnodes
210  *	freevnodes
211  */
212 static struct mtx __exclusive_cache_line vnode_list_mtx;
213 
214 /* Publicly exported FS */
215 struct nfs_public nfs_pub;
216 
217 static uma_zone_t buf_trie_zone;
218 static smr_t buf_trie_smr;
219 
220 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
221 static uma_zone_t vnode_zone;
222 MALLOC_DEFINE(M_VNODEPOLL, "VN POLL", "vnode poll");
223 
224 __read_frequently smr_t vfs_smr;
225 
226 /*
227  * The workitem queue.
228  *
229  * It is useful to delay writes of file data and filesystem metadata
230  * for tens of seconds so that quickly created and deleted files need
231  * not waste disk bandwidth being created and removed. To realize this,
232  * we append vnodes to a "workitem" queue. When running with a soft
233  * updates implementation, most pending metadata dependencies should
234  * not wait for more than a few seconds. Thus, mounted on block devices
235  * are delayed only about a half the time that file data is delayed.
236  * Similarly, directory updates are more critical, so are only delayed
237  * about a third the time that file data is delayed. Thus, there are
238  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
239  * one each second (driven off the filesystem syncer process). The
240  * syncer_delayno variable indicates the next queue that is to be processed.
241  * Items that need to be processed soon are placed in this queue:
242  *
243  *	syncer_workitem_pending[syncer_delayno]
244  *
245  * A delay of fifteen seconds is done by placing the request fifteen
246  * entries later in the queue:
247  *
248  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
249  *
250  */
251 static int syncer_delayno;
252 static long syncer_mask;
253 LIST_HEAD(synclist, bufobj);
254 static struct synclist *syncer_workitem_pending;
255 /*
256  * The sync_mtx protects:
257  *	bo->bo_synclist
258  *	sync_vnode_count
259  *	syncer_delayno
260  *	syncer_state
261  *	syncer_workitem_pending
262  *	syncer_worklist_len
263  *	rushjob
264  */
265 static struct mtx sync_mtx;
266 static struct cv sync_wakeup;
267 
268 #define SYNCER_MAXDELAY		32
269 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
270 static int syncdelay = 30;		/* max time to delay syncing data */
271 static int filedelay = 30;		/* time to delay syncing files */
272 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
273     "Time to delay syncing files (in seconds)");
274 static int dirdelay = 29;		/* time to delay syncing directories */
275 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
276     "Time to delay syncing directories (in seconds)");
277 static int metadelay = 28;		/* time to delay syncing metadata */
278 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
279     "Time to delay syncing metadata (in seconds)");
280 static int rushjob;		/* number of slots to run ASAP */
281 static int stat_rush_requests;	/* number of times I/O speeded up */
282 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
283     "Number of times I/O speeded up (rush requests)");
284 
285 #define	VDBATCH_SIZE 8
286 struct vdbatch {
287 	u_int index;
288 	struct mtx lock;
289 	struct vnode *tab[VDBATCH_SIZE];
290 };
291 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
292 
293 static void	vdbatch_dequeue(struct vnode *vp);
294 
295 /*
296  * When shutting down the syncer, run it at four times normal speed.
297  */
298 #define SYNCER_SHUTDOWN_SPEEDUP		4
299 static int sync_vnode_count;
300 static int syncer_worklist_len;
301 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
302     syncer_state;
303 
304 /* Target for maximum number of vnodes. */
305 u_long desiredvnodes;
306 static u_long gapvnodes;		/* gap between wanted and desired */
307 static u_long vhiwat;		/* enough extras after expansion */
308 static u_long vlowat;		/* minimal extras before expansion */
309 static u_long vstir;		/* nonzero to stir non-free vnodes */
310 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
311 
312 static u_long vnlru_read_freevnodes(void);
313 
314 /*
315  * Note that no attempt is made to sanitize these parameters.
316  */
317 static int
318 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
319 {
320 	u_long val;
321 	int error;
322 
323 	val = desiredvnodes;
324 	error = sysctl_handle_long(oidp, &val, 0, req);
325 	if (error != 0 || req->newptr == NULL)
326 		return (error);
327 
328 	if (val == desiredvnodes)
329 		return (0);
330 	mtx_lock(&vnode_list_mtx);
331 	desiredvnodes = val;
332 	wantfreevnodes = desiredvnodes / 4;
333 	vnlru_recalc();
334 	mtx_unlock(&vnode_list_mtx);
335 	/*
336 	 * XXX There is no protection against multiple threads changing
337 	 * desiredvnodes at the same time. Locking above only helps vnlru and
338 	 * getnewvnode.
339 	 */
340 	vfs_hash_changesize(desiredvnodes);
341 	cache_changesize(desiredvnodes);
342 	return (0);
343 }
344 
345 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
346     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
347     "LU", "Target for maximum number of vnodes");
348 
349 static int
350 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
351 {
352 	u_long val;
353 	int error;
354 
355 	val = wantfreevnodes;
356 	error = sysctl_handle_long(oidp, &val, 0, req);
357 	if (error != 0 || req->newptr == NULL)
358 		return (error);
359 
360 	if (val == wantfreevnodes)
361 		return (0);
362 	mtx_lock(&vnode_list_mtx);
363 	wantfreevnodes = val;
364 	vnlru_recalc();
365 	mtx_unlock(&vnode_list_mtx);
366 	return (0);
367 }
368 
369 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
370     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
371     "LU", "Target for minimum number of \"free\" vnodes");
372 
373 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
374     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
375 static int vnlru_nowhere;
376 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW | CTLFLAG_STATS,
377     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
378 
379 static int
380 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
381 {
382 	struct vnode *vp;
383 	struct nameidata nd;
384 	char *buf;
385 	unsigned long ndflags;
386 	int error;
387 
388 	if (req->newptr == NULL)
389 		return (EINVAL);
390 	if (req->newlen >= PATH_MAX)
391 		return (E2BIG);
392 
393 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
394 	error = SYSCTL_IN(req, buf, req->newlen);
395 	if (error != 0)
396 		goto out;
397 
398 	buf[req->newlen] = '\0';
399 
400 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1;
401 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf);
402 	if ((error = namei(&nd)) != 0)
403 		goto out;
404 	vp = nd.ni_vp;
405 
406 	if (VN_IS_DOOMED(vp)) {
407 		/*
408 		 * This vnode is being recycled.  Return != 0 to let the caller
409 		 * know that the sysctl had no effect.  Return EAGAIN because a
410 		 * subsequent call will likely succeed (since namei will create
411 		 * a new vnode if necessary)
412 		 */
413 		error = EAGAIN;
414 		goto putvnode;
415 	}
416 
417 	counter_u64_add(recycles_count, 1);
418 	vgone(vp);
419 putvnode:
420 	vput(vp);
421 	NDFREE_PNBUF(&nd);
422 out:
423 	free(buf, M_TEMP);
424 	return (error);
425 }
426 
427 static int
428 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
429 {
430 	struct thread *td = curthread;
431 	struct vnode *vp;
432 	struct file *fp;
433 	int error;
434 	int fd;
435 
436 	if (req->newptr == NULL)
437 		return (EBADF);
438 
439         error = sysctl_handle_int(oidp, &fd, 0, req);
440         if (error != 0)
441                 return (error);
442 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
443 	if (error != 0)
444 		return (error);
445 	vp = fp->f_vnode;
446 
447 	error = vn_lock(vp, LK_EXCLUSIVE);
448 	if (error != 0)
449 		goto drop;
450 
451 	counter_u64_add(recycles_count, 1);
452 	vgone(vp);
453 	VOP_UNLOCK(vp);
454 drop:
455 	fdrop(fp, td);
456 	return (error);
457 }
458 
459 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
460     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
461     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
462 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
463     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
464     sysctl_ftry_reclaim_vnode, "I",
465     "Try to reclaim a vnode by its file descriptor");
466 
467 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
468 #define vnsz2log 8
469 #ifndef DEBUG_LOCKS
470 _Static_assert(sizeof(struct vnode) >= 1UL << vnsz2log &&
471     sizeof(struct vnode) < 1UL << (vnsz2log + 1),
472     "vnsz2log needs to be updated");
473 #endif
474 
475 /*
476  * Support for the bufobj clean & dirty pctrie.
477  */
478 static void *
479 buf_trie_alloc(struct pctrie *ptree)
480 {
481 	return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT));
482 }
483 
484 static void
485 buf_trie_free(struct pctrie *ptree, void *node)
486 {
487 	uma_zfree_smr(buf_trie_zone, node);
488 }
489 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free,
490     buf_trie_smr);
491 
492 /*
493  * Initialize the vnode management data structures.
494  *
495  * Reevaluate the following cap on the number of vnodes after the physical
496  * memory size exceeds 512GB.  In the limit, as the physical memory size
497  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
498  */
499 #ifndef	MAXVNODES_MAX
500 #define	MAXVNODES_MAX	(512UL * 1024 * 1024 / 64)	/* 8M */
501 #endif
502 
503 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
504 
505 static struct vnode *
506 vn_alloc_marker(struct mount *mp)
507 {
508 	struct vnode *vp;
509 
510 	vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
511 	vp->v_type = VMARKER;
512 	vp->v_mount = mp;
513 
514 	return (vp);
515 }
516 
517 static void
518 vn_free_marker(struct vnode *vp)
519 {
520 
521 	MPASS(vp->v_type == VMARKER);
522 	free(vp, M_VNODE_MARKER);
523 }
524 
525 #ifdef KASAN
526 static int
527 vnode_ctor(void *mem, int size, void *arg __unused, int flags __unused)
528 {
529 	kasan_mark(mem, size, roundup2(size, UMA_ALIGN_PTR + 1), 0);
530 	return (0);
531 }
532 
533 static void
534 vnode_dtor(void *mem, int size, void *arg __unused)
535 {
536 	size_t end1, end2, off1, off2;
537 
538 	_Static_assert(offsetof(struct vnode, v_vnodelist) <
539 	    offsetof(struct vnode, v_dbatchcpu),
540 	    "KASAN marks require updating");
541 
542 	off1 = offsetof(struct vnode, v_vnodelist);
543 	off2 = offsetof(struct vnode, v_dbatchcpu);
544 	end1 = off1 + sizeof(((struct vnode *)NULL)->v_vnodelist);
545 	end2 = off2 + sizeof(((struct vnode *)NULL)->v_dbatchcpu);
546 
547 	/*
548 	 * Access to the v_vnodelist and v_dbatchcpu fields are permitted even
549 	 * after the vnode has been freed.  Try to get some KASAN coverage by
550 	 * marking everything except those two fields as invalid.  Because
551 	 * KASAN's tracking is not byte-granular, any preceding fields sharing
552 	 * the same 8-byte aligned word must also be marked valid.
553 	 */
554 
555 	/* Handle the area from the start until v_vnodelist... */
556 	off1 = rounddown2(off1, KASAN_SHADOW_SCALE);
557 	kasan_mark(mem, off1, off1, KASAN_UMA_FREED);
558 
559 	/* ... then the area between v_vnodelist and v_dbatchcpu ... */
560 	off1 = roundup2(end1, KASAN_SHADOW_SCALE);
561 	off2 = rounddown2(off2, KASAN_SHADOW_SCALE);
562 	if (off2 > off1)
563 		kasan_mark((void *)((char *)mem + off1), off2 - off1,
564 		    off2 - off1, KASAN_UMA_FREED);
565 
566 	/* ... and finally the area from v_dbatchcpu to the end. */
567 	off2 = roundup2(end2, KASAN_SHADOW_SCALE);
568 	kasan_mark((void *)((char *)mem + off2), size - off2, size - off2,
569 	    KASAN_UMA_FREED);
570 }
571 #endif /* KASAN */
572 
573 /*
574  * Initialize a vnode as it first enters the zone.
575  */
576 static int
577 vnode_init(void *mem, int size, int flags)
578 {
579 	struct vnode *vp;
580 
581 	vp = mem;
582 	bzero(vp, size);
583 	/*
584 	 * Setup locks.
585 	 */
586 	vp->v_vnlock = &vp->v_lock;
587 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
588 	/*
589 	 * By default, don't allow shared locks unless filesystems opt-in.
590 	 */
591 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
592 	    LK_NOSHARE | LK_IS_VNODE);
593 	/*
594 	 * Initialize bufobj.
595 	 */
596 	bufobj_init(&vp->v_bufobj, vp);
597 	/*
598 	 * Initialize namecache.
599 	 */
600 	cache_vnode_init(vp);
601 	/*
602 	 * Initialize rangelocks.
603 	 */
604 	rangelock_init(&vp->v_rl);
605 
606 	vp->v_dbatchcpu = NOCPU;
607 
608 	vp->v_state = VSTATE_DEAD;
609 
610 	/*
611 	 * Check vhold_recycle_free for an explanation.
612 	 */
613 	vp->v_holdcnt = VHOLD_NO_SMR;
614 	vp->v_type = VNON;
615 	mtx_lock(&vnode_list_mtx);
616 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
617 	mtx_unlock(&vnode_list_mtx);
618 	return (0);
619 }
620 
621 /*
622  * Free a vnode when it is cleared from the zone.
623  */
624 static void
625 vnode_fini(void *mem, int size)
626 {
627 	struct vnode *vp;
628 	struct bufobj *bo;
629 
630 	vp = mem;
631 	vdbatch_dequeue(vp);
632 	mtx_lock(&vnode_list_mtx);
633 	TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
634 	mtx_unlock(&vnode_list_mtx);
635 	rangelock_destroy(&vp->v_rl);
636 	lockdestroy(vp->v_vnlock);
637 	mtx_destroy(&vp->v_interlock);
638 	bo = &vp->v_bufobj;
639 	rw_destroy(BO_LOCKPTR(bo));
640 
641 	kasan_mark(mem, size, size, 0);
642 }
643 
644 /*
645  * Provide the size of NFS nclnode and NFS fh for calculation of the
646  * vnode memory consumption.  The size is specified directly to
647  * eliminate dependency on NFS-private header.
648  *
649  * Other filesystems may use bigger or smaller (like UFS and ZFS)
650  * private inode data, but the NFS-based estimation is ample enough.
651  * Still, we care about differences in the size between 64- and 32-bit
652  * platforms.
653  *
654  * Namecache structure size is heuristically
655  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
656  */
657 #ifdef _LP64
658 #define	NFS_NCLNODE_SZ	(528 + 64)
659 #define	NC_SZ		148
660 #else
661 #define	NFS_NCLNODE_SZ	(360 + 32)
662 #define	NC_SZ		92
663 #endif
664 
665 static void
666 vntblinit(void *dummy __unused)
667 {
668 	struct vdbatch *vd;
669 	uma_ctor ctor;
670 	uma_dtor dtor;
671 	int cpu, physvnodes, virtvnodes;
672 
673 	/*
674 	 * Desiredvnodes is a function of the physical memory size and the
675 	 * kernel's heap size.  Generally speaking, it scales with the
676 	 * physical memory size.  The ratio of desiredvnodes to the physical
677 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
678 	 * Thereafter, the
679 	 * marginal ratio of desiredvnodes to the physical memory size is
680 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
681 	 * size.  The memory required by desiredvnodes vnodes and vm objects
682 	 * must not exceed 1/10th of the kernel's heap size.
683 	 */
684 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
685 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
686 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
687 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
688 	desiredvnodes = min(physvnodes, virtvnodes);
689 	if (desiredvnodes > MAXVNODES_MAX) {
690 		if (bootverbose)
691 			printf("Reducing kern.maxvnodes %lu -> %lu\n",
692 			    desiredvnodes, MAXVNODES_MAX);
693 		desiredvnodes = MAXVNODES_MAX;
694 	}
695 	wantfreevnodes = desiredvnodes / 4;
696 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
697 	TAILQ_INIT(&vnode_list);
698 	mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
699 	/*
700 	 * The lock is taken to appease WITNESS.
701 	 */
702 	mtx_lock(&vnode_list_mtx);
703 	vnlru_recalc();
704 	mtx_unlock(&vnode_list_mtx);
705 	vnode_list_free_marker = vn_alloc_marker(NULL);
706 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
707 	vnode_list_reclaim_marker = vn_alloc_marker(NULL);
708 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
709 
710 #ifdef KASAN
711 	ctor = vnode_ctor;
712 	dtor = vnode_dtor;
713 #else
714 	ctor = NULL;
715 	dtor = NULL;
716 #endif
717 	vnode_zone = uma_zcreate("VNODE", sizeof(struct vnode), ctor, dtor,
718 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, UMA_ZONE_NOKASAN);
719 	uma_zone_set_smr(vnode_zone, vfs_smr);
720 
721 	/*
722 	 * Preallocate enough nodes to support one-per buf so that
723 	 * we can not fail an insert.  reassignbuf() callers can not
724 	 * tolerate the insertion failure.
725 	 */
726 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
727 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
728 	    UMA_ZONE_NOFREE | UMA_ZONE_SMR);
729 	buf_trie_smr = uma_zone_get_smr(buf_trie_zone);
730 	uma_prealloc(buf_trie_zone, nbuf);
731 
732 	vnodes_created = counter_u64_alloc(M_WAITOK);
733 	recycles_count = counter_u64_alloc(M_WAITOK);
734 	recycles_free_count = counter_u64_alloc(M_WAITOK);
735 
736 	/*
737 	 * Initialize the filesystem syncer.
738 	 */
739 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
740 	    &syncer_mask);
741 	syncer_maxdelay = syncer_mask + 1;
742 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
743 	cv_init(&sync_wakeup, "syncer");
744 
745 	CPU_FOREACH(cpu) {
746 		vd = DPCPU_ID_PTR((cpu), vd);
747 		bzero(vd, sizeof(*vd));
748 		mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
749 	}
750 }
751 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
752 
753 /*
754  * Mark a mount point as busy. Used to synchronize access and to delay
755  * unmounting. Eventually, mountlist_mtx is not released on failure.
756  *
757  * vfs_busy() is a custom lock, it can block the caller.
758  * vfs_busy() only sleeps if the unmount is active on the mount point.
759  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
760  * vnode belonging to mp.
761  *
762  * Lookup uses vfs_busy() to traverse mount points.
763  * root fs			var fs
764  * / vnode lock		A	/ vnode lock (/var)		D
765  * /var vnode lock	B	/log vnode lock(/var/log)	E
766  * vfs_busy lock	C	vfs_busy lock			F
767  *
768  * Within each file system, the lock order is C->A->B and F->D->E.
769  *
770  * When traversing across mounts, the system follows that lock order:
771  *
772  *        C->A->B
773  *              |
774  *              +->F->D->E
775  *
776  * The lookup() process for namei("/var") illustrates the process:
777  *  1. VOP_LOOKUP() obtains B while A is held
778  *  2. vfs_busy() obtains a shared lock on F while A and B are held
779  *  3. vput() releases lock on B
780  *  4. vput() releases lock on A
781  *  5. VFS_ROOT() obtains lock on D while shared lock on F is held
782  *  6. vfs_unbusy() releases shared lock on F
783  *  7. vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
784  *     Attempt to lock A (instead of vp_crossmp) while D is held would
785  *     violate the global order, causing deadlocks.
786  *
787  * dounmount() locks B while F is drained.  Note that for stacked
788  * filesystems, D and B in the example above may be the same lock,
789  * which introdues potential lock order reversal deadlock between
790  * dounmount() and step 5 above.  These filesystems may avoid the LOR
791  * by setting VV_CROSSLOCK on the covered vnode so that lock B will
792  * remain held until after step 5.
793  */
794 int
795 vfs_busy(struct mount *mp, int flags)
796 {
797 	struct mount_pcpu *mpcpu;
798 
799 	MPASS((flags & ~MBF_MASK) == 0);
800 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
801 
802 	if (vfs_op_thread_enter(mp, mpcpu)) {
803 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
804 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
805 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
806 		vfs_mp_count_add_pcpu(mpcpu, ref, 1);
807 		vfs_mp_count_add_pcpu(mpcpu, lockref, 1);
808 		vfs_op_thread_exit(mp, mpcpu);
809 		if (flags & MBF_MNTLSTLOCK)
810 			mtx_unlock(&mountlist_mtx);
811 		return (0);
812 	}
813 
814 	MNT_ILOCK(mp);
815 	vfs_assert_mount_counters(mp);
816 	MNT_REF(mp);
817 	/*
818 	 * If mount point is currently being unmounted, sleep until the
819 	 * mount point fate is decided.  If thread doing the unmounting fails,
820 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
821 	 * that this mount point has survived the unmount attempt and vfs_busy
822 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
823 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
824 	 * about to be really destroyed.  vfs_busy needs to release its
825 	 * reference on the mount point in this case and return with ENOENT,
826 	 * telling the caller the mount it tried to busy is no longer valid.
827 	 */
828 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
829 		KASSERT(TAILQ_EMPTY(&mp->mnt_uppers),
830 		    ("%s: non-empty upper mount list with pending unmount",
831 		    __func__));
832 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
833 			MNT_REL(mp);
834 			MNT_IUNLOCK(mp);
835 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
836 			    __func__);
837 			return (ENOENT);
838 		}
839 		if (flags & MBF_MNTLSTLOCK)
840 			mtx_unlock(&mountlist_mtx);
841 		mp->mnt_kern_flag |= MNTK_MWAIT;
842 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
843 		if (flags & MBF_MNTLSTLOCK)
844 			mtx_lock(&mountlist_mtx);
845 		MNT_ILOCK(mp);
846 	}
847 	if (flags & MBF_MNTLSTLOCK)
848 		mtx_unlock(&mountlist_mtx);
849 	mp->mnt_lockref++;
850 	MNT_IUNLOCK(mp);
851 	return (0);
852 }
853 
854 /*
855  * Free a busy filesystem.
856  */
857 void
858 vfs_unbusy(struct mount *mp)
859 {
860 	struct mount_pcpu *mpcpu;
861 	int c;
862 
863 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
864 
865 	if (vfs_op_thread_enter(mp, mpcpu)) {
866 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
867 		vfs_mp_count_sub_pcpu(mpcpu, lockref, 1);
868 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
869 		vfs_op_thread_exit(mp, mpcpu);
870 		return;
871 	}
872 
873 	MNT_ILOCK(mp);
874 	vfs_assert_mount_counters(mp);
875 	MNT_REL(mp);
876 	c = --mp->mnt_lockref;
877 	if (mp->mnt_vfs_ops == 0) {
878 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
879 		MNT_IUNLOCK(mp);
880 		return;
881 	}
882 	if (c < 0)
883 		vfs_dump_mount_counters(mp);
884 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
885 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
886 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
887 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
888 		wakeup(&mp->mnt_lockref);
889 	}
890 	MNT_IUNLOCK(mp);
891 }
892 
893 /*
894  * Lookup a mount point by filesystem identifier.
895  */
896 struct mount *
897 vfs_getvfs(fsid_t *fsid)
898 {
899 	struct mount *mp;
900 
901 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
902 	mtx_lock(&mountlist_mtx);
903 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
904 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
905 			vfs_ref(mp);
906 			mtx_unlock(&mountlist_mtx);
907 			return (mp);
908 		}
909 	}
910 	mtx_unlock(&mountlist_mtx);
911 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
912 	return ((struct mount *) 0);
913 }
914 
915 /*
916  * Lookup a mount point by filesystem identifier, busying it before
917  * returning.
918  *
919  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
920  * cache for popular filesystem identifiers.  The cache is lockess, using
921  * the fact that struct mount's are never freed.  In worst case we may
922  * get pointer to unmounted or even different filesystem, so we have to
923  * check what we got, and go slow way if so.
924  */
925 struct mount *
926 vfs_busyfs(fsid_t *fsid)
927 {
928 #define	FSID_CACHE_SIZE	256
929 	typedef struct mount * volatile vmp_t;
930 	static vmp_t cache[FSID_CACHE_SIZE];
931 	struct mount *mp;
932 	int error;
933 	uint32_t hash;
934 
935 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
936 	hash = fsid->val[0] ^ fsid->val[1];
937 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
938 	mp = cache[hash];
939 	if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0)
940 		goto slow;
941 	if (vfs_busy(mp, 0) != 0) {
942 		cache[hash] = NULL;
943 		goto slow;
944 	}
945 	if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0)
946 		return (mp);
947 	else
948 	    vfs_unbusy(mp);
949 
950 slow:
951 	mtx_lock(&mountlist_mtx);
952 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
953 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
954 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
955 			if (error) {
956 				cache[hash] = NULL;
957 				mtx_unlock(&mountlist_mtx);
958 				return (NULL);
959 			}
960 			cache[hash] = mp;
961 			return (mp);
962 		}
963 	}
964 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
965 	mtx_unlock(&mountlist_mtx);
966 	return ((struct mount *) 0);
967 }
968 
969 /*
970  * Check if a user can access privileged mount options.
971  */
972 int
973 vfs_suser(struct mount *mp, struct thread *td)
974 {
975 	int error;
976 
977 	if (jailed(td->td_ucred)) {
978 		/*
979 		 * If the jail of the calling thread lacks permission for
980 		 * this type of file system, deny immediately.
981 		 */
982 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
983 			return (EPERM);
984 
985 		/*
986 		 * If the file system was mounted outside the jail of the
987 		 * calling thread, deny immediately.
988 		 */
989 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
990 			return (EPERM);
991 	}
992 
993 	/*
994 	 * If file system supports delegated administration, we don't check
995 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
996 	 * by the file system itself.
997 	 * If this is not the user that did original mount, we check for
998 	 * the PRIV_VFS_MOUNT_OWNER privilege.
999 	 */
1000 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
1001 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
1002 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
1003 			return (error);
1004 	}
1005 	return (0);
1006 }
1007 
1008 /*
1009  * Get a new unique fsid.  Try to make its val[0] unique, since this value
1010  * will be used to create fake device numbers for stat().  Also try (but
1011  * not so hard) make its val[0] unique mod 2^16, since some emulators only
1012  * support 16-bit device numbers.  We end up with unique val[0]'s for the
1013  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
1014  *
1015  * Keep in mind that several mounts may be running in parallel.  Starting
1016  * the search one past where the previous search terminated is both a
1017  * micro-optimization and a defense against returning the same fsid to
1018  * different mounts.
1019  */
1020 void
1021 vfs_getnewfsid(struct mount *mp)
1022 {
1023 	static uint16_t mntid_base;
1024 	struct mount *nmp;
1025 	fsid_t tfsid;
1026 	int mtype;
1027 
1028 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
1029 	mtx_lock(&mntid_mtx);
1030 	mtype = mp->mnt_vfc->vfc_typenum;
1031 	tfsid.val[1] = mtype;
1032 	mtype = (mtype & 0xFF) << 24;
1033 	for (;;) {
1034 		tfsid.val[0] = makedev(255,
1035 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
1036 		mntid_base++;
1037 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
1038 			break;
1039 		vfs_rel(nmp);
1040 	}
1041 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
1042 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
1043 	mtx_unlock(&mntid_mtx);
1044 }
1045 
1046 /*
1047  * Knob to control the precision of file timestamps:
1048  *
1049  *   0 = seconds only; nanoseconds zeroed.
1050  *   1 = seconds and nanoseconds, accurate within 1/HZ.
1051  *   2 = seconds and nanoseconds, truncated to microseconds.
1052  * >=3 = seconds and nanoseconds, maximum precision.
1053  */
1054 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
1055 
1056 static int timestamp_precision = TSP_USEC;
1057 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
1058     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
1059     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1060     "3+: sec + ns (max. precision))");
1061 
1062 /*
1063  * Get a current timestamp.
1064  */
1065 void
1066 vfs_timestamp(struct timespec *tsp)
1067 {
1068 	struct timeval tv;
1069 
1070 	switch (timestamp_precision) {
1071 	case TSP_SEC:
1072 		tsp->tv_sec = time_second;
1073 		tsp->tv_nsec = 0;
1074 		break;
1075 	case TSP_HZ:
1076 		getnanotime(tsp);
1077 		break;
1078 	case TSP_USEC:
1079 		microtime(&tv);
1080 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1081 		break;
1082 	case TSP_NSEC:
1083 	default:
1084 		nanotime(tsp);
1085 		break;
1086 	}
1087 }
1088 
1089 /*
1090  * Set vnode attributes to VNOVAL
1091  */
1092 void
1093 vattr_null(struct vattr *vap)
1094 {
1095 
1096 	vap->va_type = VNON;
1097 	vap->va_size = VNOVAL;
1098 	vap->va_bytes = VNOVAL;
1099 	vap->va_mode = VNOVAL;
1100 	vap->va_nlink = VNOVAL;
1101 	vap->va_uid = VNOVAL;
1102 	vap->va_gid = VNOVAL;
1103 	vap->va_fsid = VNOVAL;
1104 	vap->va_fileid = VNOVAL;
1105 	vap->va_blocksize = VNOVAL;
1106 	vap->va_rdev = VNOVAL;
1107 	vap->va_atime.tv_sec = VNOVAL;
1108 	vap->va_atime.tv_nsec = VNOVAL;
1109 	vap->va_mtime.tv_sec = VNOVAL;
1110 	vap->va_mtime.tv_nsec = VNOVAL;
1111 	vap->va_ctime.tv_sec = VNOVAL;
1112 	vap->va_ctime.tv_nsec = VNOVAL;
1113 	vap->va_birthtime.tv_sec = VNOVAL;
1114 	vap->va_birthtime.tv_nsec = VNOVAL;
1115 	vap->va_flags = VNOVAL;
1116 	vap->va_gen = VNOVAL;
1117 	vap->va_vaflags = 0;
1118 }
1119 
1120 /*
1121  * Try to reduce the total number of vnodes.
1122  *
1123  * This routine (and its user) are buggy in at least the following ways:
1124  * - all parameters were picked years ago when RAM sizes were significantly
1125  *   smaller
1126  * - it can pick vnodes based on pages used by the vm object, but filesystems
1127  *   like ZFS don't use it making the pick broken
1128  * - since ZFS has its own aging policy it gets partially combated by this one
1129  * - a dedicated method should be provided for filesystems to let them decide
1130  *   whether the vnode should be recycled
1131  *
1132  * This routine is called when we have too many vnodes.  It attempts
1133  * to free <count> vnodes and will potentially free vnodes that still
1134  * have VM backing store (VM backing store is typically the cause
1135  * of a vnode blowout so we want to do this).  Therefore, this operation
1136  * is not considered cheap.
1137  *
1138  * A number of conditions may prevent a vnode from being reclaimed.
1139  * the buffer cache may have references on the vnode, a directory
1140  * vnode may still have references due to the namei cache representing
1141  * underlying files, or the vnode may be in active use.   It is not
1142  * desirable to reuse such vnodes.  These conditions may cause the
1143  * number of vnodes to reach some minimum value regardless of what
1144  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
1145  *
1146  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1147  * 			 entries if this argument is strue
1148  * @param trigger	 Only reclaim vnodes with fewer than this many resident
1149  *			 pages.
1150  * @param target	 How many vnodes to reclaim.
1151  * @return		 The number of vnodes that were reclaimed.
1152  */
1153 static int
1154 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1155 {
1156 	struct vnode *vp, *mvp;
1157 	struct mount *mp;
1158 	struct vm_object *object;
1159 	u_long done;
1160 	bool retried;
1161 
1162 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1163 
1164 	retried = false;
1165 	done = 0;
1166 
1167 	mvp = vnode_list_reclaim_marker;
1168 restart:
1169 	vp = mvp;
1170 	while (done < target) {
1171 		vp = TAILQ_NEXT(vp, v_vnodelist);
1172 		if (__predict_false(vp == NULL))
1173 			break;
1174 
1175 		if (__predict_false(vp->v_type == VMARKER))
1176 			continue;
1177 
1178 		/*
1179 		 * If it's been deconstructed already, it's still
1180 		 * referenced, or it exceeds the trigger, skip it.
1181 		 * Also skip free vnodes.  We are trying to make space
1182 		 * to expand the free list, not reduce it.
1183 		 */
1184 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1185 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1186 			goto next_iter;
1187 
1188 		if (vp->v_type == VBAD || vp->v_type == VNON)
1189 			goto next_iter;
1190 
1191 		object = atomic_load_ptr(&vp->v_object);
1192 		if (object == NULL || object->resident_page_count > trigger) {
1193 			goto next_iter;
1194 		}
1195 
1196 		/*
1197 		 * Handle races against vnode allocation. Filesystems lock the
1198 		 * vnode some time after it gets returned from getnewvnode,
1199 		 * despite type and hold count being manipulated earlier.
1200 		 * Resorting to checking v_mount restores guarantees present
1201 		 * before the global list was reworked to contain all vnodes.
1202 		 */
1203 		if (!VI_TRYLOCK(vp))
1204 			goto next_iter;
1205 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1206 			VI_UNLOCK(vp);
1207 			goto next_iter;
1208 		}
1209 		if (vp->v_mount == NULL) {
1210 			VI_UNLOCK(vp);
1211 			goto next_iter;
1212 		}
1213 		vholdl(vp);
1214 		VI_UNLOCK(vp);
1215 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1216 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1217 		mtx_unlock(&vnode_list_mtx);
1218 
1219 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1220 			vdrop_recycle(vp);
1221 			goto next_iter_unlocked;
1222 		}
1223 		if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1224 			vdrop_recycle(vp);
1225 			vn_finished_write(mp);
1226 			goto next_iter_unlocked;
1227 		}
1228 
1229 		VI_LOCK(vp);
1230 		if (vp->v_usecount > 0 ||
1231 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1232 		    (vp->v_object != NULL && vp->v_object->handle == vp &&
1233 		    vp->v_object->resident_page_count > trigger)) {
1234 			VOP_UNLOCK(vp);
1235 			vdropl_recycle(vp);
1236 			vn_finished_write(mp);
1237 			goto next_iter_unlocked;
1238 		}
1239 		counter_u64_add(recycles_count, 1);
1240 		vgonel(vp);
1241 		VOP_UNLOCK(vp);
1242 		vdropl_recycle(vp);
1243 		vn_finished_write(mp);
1244 		done++;
1245 next_iter_unlocked:
1246 		maybe_yield();
1247 		mtx_lock(&vnode_list_mtx);
1248 		goto restart;
1249 next_iter:
1250 		MPASS(vp->v_type != VMARKER);
1251 		if (!should_yield())
1252 			continue;
1253 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1254 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1255 		mtx_unlock(&vnode_list_mtx);
1256 		kern_yield(PRI_USER);
1257 		mtx_lock(&vnode_list_mtx);
1258 		goto restart;
1259 	}
1260 	if (done == 0 && !retried) {
1261 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1262 		TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1263 		retried = true;
1264 		goto restart;
1265 	}
1266 	return (done);
1267 }
1268 
1269 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1270 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1271     0,
1272     "limit on vnode free requests per call to the vnlru_free routine");
1273 
1274 /*
1275  * Attempt to reduce the free list by the requested amount.
1276  */
1277 static int
1278 vnlru_free_impl(int count, struct vfsops *mnt_op, struct vnode *mvp)
1279 {
1280 	struct vnode *vp;
1281 	struct mount *mp;
1282 	int ocount;
1283 
1284 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1285 	if (count > max_vnlru_free)
1286 		count = max_vnlru_free;
1287 	ocount = count;
1288 	vp = mvp;
1289 	for (;;) {
1290 		if (count == 0) {
1291 			break;
1292 		}
1293 		vp = TAILQ_NEXT(vp, v_vnodelist);
1294 		if (__predict_false(vp == NULL)) {
1295 			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1296 			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1297 			break;
1298 		}
1299 		if (__predict_false(vp->v_type == VMARKER))
1300 			continue;
1301 		if (vp->v_holdcnt > 0)
1302 			continue;
1303 		/*
1304 		 * Don't recycle if our vnode is from different type
1305 		 * of mount point.  Note that mp is type-safe, the
1306 		 * check does not reach unmapped address even if
1307 		 * vnode is reclaimed.
1308 		 */
1309 		if (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1310 		    mp->mnt_op != mnt_op) {
1311 			continue;
1312 		}
1313 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1314 			continue;
1315 		}
1316 		if (!vhold_recycle_free(vp))
1317 			continue;
1318 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1319 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1320 		mtx_unlock(&vnode_list_mtx);
1321 		/*
1322 		 * FIXME: ignores the return value, meaning it may be nothing
1323 		 * got recycled but it claims otherwise to the caller.
1324 		 *
1325 		 * Originally the value started being ignored in 2005 with
1326 		 * 114a1006a8204aa156e1f9ad6476cdff89cada7f .
1327 		 *
1328 		 * Respecting the value can run into significant stalls if most
1329 		 * vnodes belong to one file system and it has writes
1330 		 * suspended.  In presence of many threads and millions of
1331 		 * vnodes they keep contending on the vnode_list_mtx lock only
1332 		 * to find vnodes they can't recycle.
1333 		 *
1334 		 * The solution would be to pre-check if the vnode is likely to
1335 		 * be recycle-able, but it needs to happen with the
1336 		 * vnode_list_mtx lock held. This runs into a problem where
1337 		 * VOP_GETWRITEMOUNT (currently needed to find out about if
1338 		 * writes are frozen) can take locks which LOR against it.
1339 		 *
1340 		 * Check nullfs for one example (null_getwritemount).
1341 		 */
1342 		vtryrecycle(vp);
1343 		count--;
1344 		mtx_lock(&vnode_list_mtx);
1345 		vp = mvp;
1346 	}
1347 	return (ocount - count);
1348 }
1349 
1350 static int
1351 vnlru_free_locked(int count)
1352 {
1353 
1354 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1355 	return (vnlru_free_impl(count, NULL, vnode_list_free_marker));
1356 }
1357 
1358 void
1359 vnlru_free_vfsops(int count, struct vfsops *mnt_op, struct vnode *mvp)
1360 {
1361 
1362 	MPASS(mnt_op != NULL);
1363 	MPASS(mvp != NULL);
1364 	VNPASS(mvp->v_type == VMARKER, mvp);
1365 	mtx_lock(&vnode_list_mtx);
1366 	vnlru_free_impl(count, mnt_op, mvp);
1367 	mtx_unlock(&vnode_list_mtx);
1368 }
1369 
1370 struct vnode *
1371 vnlru_alloc_marker(void)
1372 {
1373 	struct vnode *mvp;
1374 
1375 	mvp = vn_alloc_marker(NULL);
1376 	mtx_lock(&vnode_list_mtx);
1377 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, mvp, v_vnodelist);
1378 	mtx_unlock(&vnode_list_mtx);
1379 	return (mvp);
1380 }
1381 
1382 void
1383 vnlru_free_marker(struct vnode *mvp)
1384 {
1385 	mtx_lock(&vnode_list_mtx);
1386 	TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1387 	mtx_unlock(&vnode_list_mtx);
1388 	vn_free_marker(mvp);
1389 }
1390 
1391 static void
1392 vnlru_recalc(void)
1393 {
1394 
1395 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1396 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1397 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1398 	vlowat = vhiwat / 2;
1399 }
1400 
1401 /*
1402  * Attempt to recycle vnodes in a context that is always safe to block.
1403  * Calling vlrurecycle() from the bowels of filesystem code has some
1404  * interesting deadlock problems.
1405  */
1406 static struct proc *vnlruproc;
1407 static int vnlruproc_sig;
1408 
1409 /*
1410  * The main freevnodes counter is only updated when threads requeue their vnode
1411  * batches. CPUs are conditionally walked to compute a more accurate total.
1412  *
1413  * Limit how much of a slop are we willing to tolerate. Note: the actual value
1414  * at any given moment can still exceed slop, but it should not be by significant
1415  * margin in practice.
1416  */
1417 #define VNLRU_FREEVNODES_SLOP 126
1418 
1419 static void __noinline
1420 vfs_freevnodes_rollup(int8_t *lfreevnodes)
1421 {
1422 
1423 	atomic_add_long(&freevnodes, *lfreevnodes);
1424 	*lfreevnodes = 0;
1425 	critical_exit();
1426 }
1427 
1428 static __inline void
1429 vfs_freevnodes_inc(void)
1430 {
1431 	int8_t *lfreevnodes;
1432 
1433 	critical_enter();
1434 	lfreevnodes = PCPU_PTR(vfs_freevnodes);
1435 	(*lfreevnodes)++;
1436 	if (__predict_false(*lfreevnodes == VNLRU_FREEVNODES_SLOP))
1437 		vfs_freevnodes_rollup(lfreevnodes);
1438 	else
1439 		critical_exit();
1440 }
1441 
1442 static __inline void
1443 vfs_freevnodes_dec(void)
1444 {
1445 	int8_t *lfreevnodes;
1446 
1447 	critical_enter();
1448 	lfreevnodes = PCPU_PTR(vfs_freevnodes);
1449 	(*lfreevnodes)--;
1450 	if (__predict_false(*lfreevnodes == -VNLRU_FREEVNODES_SLOP))
1451 		vfs_freevnodes_rollup(lfreevnodes);
1452 	else
1453 		critical_exit();
1454 }
1455 
1456 static u_long
1457 vnlru_read_freevnodes(void)
1458 {
1459 	long slop, rfreevnodes;
1460 	int cpu;
1461 
1462 	rfreevnodes = atomic_load_long(&freevnodes);
1463 
1464 	if (rfreevnodes > freevnodes_old)
1465 		slop = rfreevnodes - freevnodes_old;
1466 	else
1467 		slop = freevnodes_old - rfreevnodes;
1468 	if (slop < VNLRU_FREEVNODES_SLOP)
1469 		return (rfreevnodes >= 0 ? rfreevnodes : 0);
1470 	freevnodes_old = rfreevnodes;
1471 	CPU_FOREACH(cpu) {
1472 		freevnodes_old += cpuid_to_pcpu[cpu]->pc_vfs_freevnodes;
1473 	}
1474 	return (freevnodes_old >= 0 ? freevnodes_old : 0);
1475 }
1476 
1477 static bool
1478 vnlru_under(u_long rnumvnodes, u_long limit)
1479 {
1480 	u_long rfreevnodes, space;
1481 
1482 	if (__predict_false(rnumvnodes > desiredvnodes))
1483 		return (true);
1484 
1485 	space = desiredvnodes - rnumvnodes;
1486 	if (space < limit) {
1487 		rfreevnodes = vnlru_read_freevnodes();
1488 		if (rfreevnodes > wantfreevnodes)
1489 			space += rfreevnodes - wantfreevnodes;
1490 	}
1491 	return (space < limit);
1492 }
1493 
1494 static bool
1495 vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1496 {
1497 	long rfreevnodes, space;
1498 
1499 	if (__predict_false(rnumvnodes > desiredvnodes))
1500 		return (true);
1501 
1502 	space = desiredvnodes - rnumvnodes;
1503 	if (space < limit) {
1504 		rfreevnodes = atomic_load_long(&freevnodes);
1505 		if (rfreevnodes > wantfreevnodes)
1506 			space += rfreevnodes - wantfreevnodes;
1507 	}
1508 	return (space < limit);
1509 }
1510 
1511 static void
1512 vnlru_kick(void)
1513 {
1514 
1515 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1516 	if (vnlruproc_sig == 0) {
1517 		vnlruproc_sig = 1;
1518 		wakeup(vnlruproc);
1519 	}
1520 }
1521 
1522 static void
1523 vnlru_proc(void)
1524 {
1525 	u_long rnumvnodes, rfreevnodes, target;
1526 	unsigned long onumvnodes;
1527 	int done, force, trigger, usevnodes;
1528 	bool reclaim_nc_src, want_reread;
1529 
1530 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1531 	    SHUTDOWN_PRI_FIRST);
1532 
1533 	force = 0;
1534 	want_reread = false;
1535 	for (;;) {
1536 		kproc_suspend_check(vnlruproc);
1537 		mtx_lock(&vnode_list_mtx);
1538 		rnumvnodes = atomic_load_long(&numvnodes);
1539 
1540 		if (want_reread) {
1541 			force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1542 			want_reread = false;
1543 		}
1544 
1545 		/*
1546 		 * If numvnodes is too large (due to desiredvnodes being
1547 		 * adjusted using its sysctl, or emergency growth), first
1548 		 * try to reduce it by discarding from the free list.
1549 		 */
1550 		if (rnumvnodes > desiredvnodes) {
1551 			vnlru_free_locked(rnumvnodes - desiredvnodes);
1552 			rnumvnodes = atomic_load_long(&numvnodes);
1553 		}
1554 		/*
1555 		 * Sleep if the vnode cache is in a good state.  This is
1556 		 * when it is not over-full and has space for about a 4%
1557 		 * or 9% expansion (by growing its size or inexcessively
1558 		 * reducing its free list).  Otherwise, try to reclaim
1559 		 * space for a 10% expansion.
1560 		 */
1561 		if (vstir && force == 0) {
1562 			force = 1;
1563 			vstir = 0;
1564 		}
1565 		if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1566 			vnlruproc_sig = 0;
1567 			wakeup(&vnlruproc_sig);
1568 			msleep(vnlruproc, &vnode_list_mtx,
1569 			    PVFS|PDROP, "vlruwt", hz);
1570 			continue;
1571 		}
1572 		rfreevnodes = vnlru_read_freevnodes();
1573 
1574 		onumvnodes = rnumvnodes;
1575 		/*
1576 		 * Calculate parameters for recycling.  These are the same
1577 		 * throughout the loop to give some semblance of fairness.
1578 		 * The trigger point is to avoid recycling vnodes with lots
1579 		 * of resident pages.  We aren't trying to free memory; we
1580 		 * are trying to recycle or at least free vnodes.
1581 		 */
1582 		if (rnumvnodes <= desiredvnodes)
1583 			usevnodes = rnumvnodes - rfreevnodes;
1584 		else
1585 			usevnodes = rnumvnodes;
1586 		if (usevnodes <= 0)
1587 			usevnodes = 1;
1588 		/*
1589 		 * The trigger value is chosen to give a conservatively
1590 		 * large value to ensure that it alone doesn't prevent
1591 		 * making progress.  The value can easily be so large that
1592 		 * it is effectively infinite in some congested and
1593 		 * misconfigured cases, and this is necessary.  Normally
1594 		 * it is about 8 to 100 (pages), which is quite large.
1595 		 */
1596 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1597 		if (force < 2)
1598 			trigger = vsmalltrigger;
1599 		reclaim_nc_src = force >= 3;
1600 		target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1601 		target = target / 10 + 1;
1602 		done = vlrureclaim(reclaim_nc_src, trigger, target);
1603 		mtx_unlock(&vnode_list_mtx);
1604 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1605 			uma_reclaim(UMA_RECLAIM_DRAIN);
1606 		if (done == 0) {
1607 			if (force == 0 || force == 1) {
1608 				force = 2;
1609 				continue;
1610 			}
1611 			if (force == 2) {
1612 				force = 3;
1613 				continue;
1614 			}
1615 			want_reread = true;
1616 			force = 0;
1617 			vnlru_nowhere++;
1618 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1619 		} else {
1620 			want_reread = true;
1621 			kern_yield(PRI_USER);
1622 		}
1623 	}
1624 }
1625 
1626 static struct kproc_desc vnlru_kp = {
1627 	"vnlru",
1628 	vnlru_proc,
1629 	&vnlruproc
1630 };
1631 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1632     &vnlru_kp);
1633 
1634 /*
1635  * Routines having to do with the management of the vnode table.
1636  */
1637 
1638 /*
1639  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1640  * before we actually vgone().  This function must be called with the vnode
1641  * held to prevent the vnode from being returned to the free list midway
1642  * through vgone().
1643  */
1644 static int
1645 vtryrecycle(struct vnode *vp)
1646 {
1647 	struct mount *vnmp;
1648 
1649 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1650 	VNPASS(vp->v_holdcnt > 0, vp);
1651 	/*
1652 	 * This vnode may found and locked via some other list, if so we
1653 	 * can't recycle it yet.
1654 	 */
1655 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1656 		CTR2(KTR_VFS,
1657 		    "%s: impossible to recycle, vp %p lock is already held",
1658 		    __func__, vp);
1659 		vdrop_recycle(vp);
1660 		return (EWOULDBLOCK);
1661 	}
1662 	/*
1663 	 * Don't recycle if its filesystem is being suspended.
1664 	 */
1665 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1666 		VOP_UNLOCK(vp);
1667 		CTR2(KTR_VFS,
1668 		    "%s: impossible to recycle, cannot start the write for %p",
1669 		    __func__, vp);
1670 		vdrop_recycle(vp);
1671 		return (EBUSY);
1672 	}
1673 	/*
1674 	 * If we got this far, we need to acquire the interlock and see if
1675 	 * anyone picked up this vnode from another list.  If not, we will
1676 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1677 	 * will skip over it.
1678 	 */
1679 	VI_LOCK(vp);
1680 	if (vp->v_usecount) {
1681 		VOP_UNLOCK(vp);
1682 		vdropl_recycle(vp);
1683 		vn_finished_write(vnmp);
1684 		CTR2(KTR_VFS,
1685 		    "%s: impossible to recycle, %p is already referenced",
1686 		    __func__, vp);
1687 		return (EBUSY);
1688 	}
1689 	if (!VN_IS_DOOMED(vp)) {
1690 		counter_u64_add(recycles_free_count, 1);
1691 		vgonel(vp);
1692 	}
1693 	VOP_UNLOCK(vp);
1694 	vdropl_recycle(vp);
1695 	vn_finished_write(vnmp);
1696 	return (0);
1697 }
1698 
1699 /*
1700  * Allocate a new vnode.
1701  *
1702  * The operation never returns an error. Returning an error was disabled
1703  * in r145385 (dated 2005) with the following comment:
1704  *
1705  * XXX Not all VFS_VGET/ffs_vget callers check returns.
1706  *
1707  * Given the age of this commit (almost 15 years at the time of writing this
1708  * comment) restoring the ability to fail requires a significant audit of
1709  * all codepaths.
1710  *
1711  * The routine can try to free a vnode or stall for up to 1 second waiting for
1712  * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1713  */
1714 static u_long vn_alloc_cyclecount;
1715 static u_long vn_alloc_sleeps;
1716 
1717 SYSCTL_ULONG(_vfs, OID_AUTO, vnode_alloc_sleeps, CTLFLAG_RD, &vn_alloc_sleeps, 0,
1718     "Number of times vnode allocation blocked waiting on vnlru");
1719 
1720 static struct vnode * __noinline
1721 vn_alloc_hard(struct mount *mp)
1722 {
1723 	u_long rnumvnodes, rfreevnodes;
1724 
1725 	mtx_lock(&vnode_list_mtx);
1726 	rnumvnodes = atomic_load_long(&numvnodes);
1727 	if (rnumvnodes + 1 < desiredvnodes) {
1728 		vn_alloc_cyclecount = 0;
1729 		goto alloc;
1730 	}
1731 	rfreevnodes = vnlru_read_freevnodes();
1732 	if (vn_alloc_cyclecount++ >= rfreevnodes) {
1733 		vn_alloc_cyclecount = 0;
1734 		vstir = 1;
1735 	}
1736 	/*
1737 	 * Grow the vnode cache if it will not be above its target max
1738 	 * after growing.  Otherwise, if the free list is nonempty, try
1739 	 * to reclaim 1 item from it before growing the cache (possibly
1740 	 * above its target max if the reclamation failed or is delayed).
1741 	 * Otherwise, wait for some space.  In all cases, schedule
1742 	 * vnlru_proc() if we are getting short of space.  The watermarks
1743 	 * should be chosen so that we never wait or even reclaim from
1744 	 * the free list to below its target minimum.
1745 	 */
1746 	if (vnlru_free_locked(1) > 0)
1747 		goto alloc;
1748 	if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1749 		/*
1750 		 * Wait for space for a new vnode.
1751 		 */
1752 		vnlru_kick();
1753 		vn_alloc_sleeps++;
1754 		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1755 		if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1756 		    vnlru_read_freevnodes() > 1)
1757 			vnlru_free_locked(1);
1758 	}
1759 alloc:
1760 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1761 	if (vnlru_under(rnumvnodes, vlowat))
1762 		vnlru_kick();
1763 	mtx_unlock(&vnode_list_mtx);
1764 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1765 }
1766 
1767 static struct vnode *
1768 vn_alloc(struct mount *mp)
1769 {
1770 	u_long rnumvnodes;
1771 
1772 	if (__predict_false(vn_alloc_cyclecount != 0))
1773 		return (vn_alloc_hard(mp));
1774 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1775 	if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1776 		atomic_subtract_long(&numvnodes, 1);
1777 		return (vn_alloc_hard(mp));
1778 	}
1779 
1780 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1781 }
1782 
1783 static void
1784 vn_free(struct vnode *vp)
1785 {
1786 
1787 	atomic_subtract_long(&numvnodes, 1);
1788 	uma_zfree_smr(vnode_zone, vp);
1789 }
1790 
1791 /*
1792  * Return the next vnode from the free list.
1793  */
1794 int
1795 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1796     struct vnode **vpp)
1797 {
1798 	struct vnode *vp;
1799 	struct thread *td;
1800 	struct lock_object *lo;
1801 
1802 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1803 
1804 	KASSERT(vops->registered,
1805 	    ("%s: not registered vector op %p\n", __func__, vops));
1806 	cache_validate_vop_vector(mp, vops);
1807 
1808 	td = curthread;
1809 	if (td->td_vp_reserved != NULL) {
1810 		vp = td->td_vp_reserved;
1811 		td->td_vp_reserved = NULL;
1812 	} else {
1813 		vp = vn_alloc(mp);
1814 	}
1815 	counter_u64_add(vnodes_created, 1);
1816 
1817 	vn_set_state(vp, VSTATE_UNINITIALIZED);
1818 
1819 	/*
1820 	 * Locks are given the generic name "vnode" when created.
1821 	 * Follow the historic practice of using the filesystem
1822 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1823 	 *
1824 	 * Locks live in a witness group keyed on their name. Thus,
1825 	 * when a lock is renamed, it must also move from the witness
1826 	 * group of its old name to the witness group of its new name.
1827 	 *
1828 	 * The change only needs to be made when the vnode moves
1829 	 * from one filesystem type to another. We ensure that each
1830 	 * filesystem use a single static name pointer for its tag so
1831 	 * that we can compare pointers rather than doing a strcmp().
1832 	 */
1833 	lo = &vp->v_vnlock->lock_object;
1834 #ifdef WITNESS
1835 	if (lo->lo_name != tag) {
1836 #endif
1837 		lo->lo_name = tag;
1838 #ifdef WITNESS
1839 		WITNESS_DESTROY(lo);
1840 		WITNESS_INIT(lo, tag);
1841 	}
1842 #endif
1843 	/*
1844 	 * By default, don't allow shared locks unless filesystems opt-in.
1845 	 */
1846 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1847 	/*
1848 	 * Finalize various vnode identity bits.
1849 	 */
1850 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1851 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1852 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1853 	vp->v_type = VNON;
1854 	vp->v_op = vops;
1855 	vp->v_irflag = 0;
1856 	v_init_counters(vp);
1857 	vn_seqc_init(vp);
1858 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1859 #ifdef DIAGNOSTIC
1860 	if (mp == NULL && vops != &dead_vnodeops)
1861 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1862 #endif
1863 #ifdef MAC
1864 	mac_vnode_init(vp);
1865 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1866 		mac_vnode_associate_singlelabel(mp, vp);
1867 #endif
1868 	if (mp != NULL) {
1869 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1870 	}
1871 
1872 	/*
1873 	 * For the filesystems which do not use vfs_hash_insert(),
1874 	 * still initialize v_hash to have vfs_hash_index() useful.
1875 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1876 	 * its own hashing.
1877 	 */
1878 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1879 
1880 	*vpp = vp;
1881 	return (0);
1882 }
1883 
1884 void
1885 getnewvnode_reserve(void)
1886 {
1887 	struct thread *td;
1888 
1889 	td = curthread;
1890 	MPASS(td->td_vp_reserved == NULL);
1891 	td->td_vp_reserved = vn_alloc(NULL);
1892 }
1893 
1894 void
1895 getnewvnode_drop_reserve(void)
1896 {
1897 	struct thread *td;
1898 
1899 	td = curthread;
1900 	if (td->td_vp_reserved != NULL) {
1901 		vn_free(td->td_vp_reserved);
1902 		td->td_vp_reserved = NULL;
1903 	}
1904 }
1905 
1906 static void __noinline
1907 freevnode(struct vnode *vp)
1908 {
1909 	struct bufobj *bo;
1910 
1911 	/*
1912 	 * The vnode has been marked for destruction, so free it.
1913 	 *
1914 	 * The vnode will be returned to the zone where it will
1915 	 * normally remain until it is needed for another vnode. We
1916 	 * need to cleanup (or verify that the cleanup has already
1917 	 * been done) any residual data left from its current use
1918 	 * so as not to contaminate the freshly allocated vnode.
1919 	 */
1920 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1921 	/*
1922 	 * Paired with vgone.
1923 	 */
1924 	vn_seqc_write_end_free(vp);
1925 
1926 	bo = &vp->v_bufobj;
1927 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1928 	VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp);
1929 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1930 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1931 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1932 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1933 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1934 	    ("clean blk trie not empty"));
1935 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1936 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1937 	    ("dirty blk trie not empty"));
1938 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1939 	    ("Dangling rangelock waiters"));
1940 	VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp,
1941 	    ("Leaked inactivation"));
1942 	VI_UNLOCK(vp);
1943 	cache_assert_no_entries(vp);
1944 
1945 #ifdef MAC
1946 	mac_vnode_destroy(vp);
1947 #endif
1948 	if (vp->v_pollinfo != NULL) {
1949 		/*
1950 		 * Use LK_NOWAIT to shut up witness about the lock. We may get
1951 		 * here while having another vnode locked when trying to
1952 		 * satisfy a lookup and needing to recycle.
1953 		 */
1954 		VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT);
1955 		destroy_vpollinfo(vp->v_pollinfo);
1956 		VOP_UNLOCK(vp);
1957 		vp->v_pollinfo = NULL;
1958 	}
1959 	vp->v_mountedhere = NULL;
1960 	vp->v_unpcb = NULL;
1961 	vp->v_rdev = NULL;
1962 	vp->v_fifoinfo = NULL;
1963 	vp->v_iflag = 0;
1964 	vp->v_vflag = 0;
1965 	bo->bo_flag = 0;
1966 	vn_free(vp);
1967 }
1968 
1969 /*
1970  * Delete from old mount point vnode list, if on one.
1971  */
1972 static void
1973 delmntque(struct vnode *vp)
1974 {
1975 	struct mount *mp;
1976 
1977 	VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
1978 
1979 	mp = vp->v_mount;
1980 	MNT_ILOCK(mp);
1981 	VI_LOCK(vp);
1982 	vp->v_mount = NULL;
1983 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1984 		("bad mount point vnode list size"));
1985 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1986 	mp->mnt_nvnodelistsize--;
1987 	MNT_REL(mp);
1988 	MNT_IUNLOCK(mp);
1989 	/*
1990 	 * The caller expects the interlock to be still held.
1991 	 */
1992 	ASSERT_VI_LOCKED(vp, __func__);
1993 }
1994 
1995 static int
1996 insmntque1_int(struct vnode *vp, struct mount *mp, bool dtr)
1997 {
1998 
1999 	KASSERT(vp->v_mount == NULL,
2000 		("insmntque: vnode already on per mount vnode list"));
2001 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
2002 	if ((mp->mnt_kern_flag & MNTK_UNLOCKED_INSMNTQUE) == 0) {
2003 		ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
2004 	} else {
2005 		KASSERT(!dtr,
2006 		    ("%s: can't have MNTK_UNLOCKED_INSMNTQUE and cleanup",
2007 		    __func__));
2008 	}
2009 
2010 	/*
2011 	 * We acquire the vnode interlock early to ensure that the
2012 	 * vnode cannot be recycled by another process releasing a
2013 	 * holdcnt on it before we get it on both the vnode list
2014 	 * and the active vnode list. The mount mutex protects only
2015 	 * manipulation of the vnode list and the vnode freelist
2016 	 * mutex protects only manipulation of the active vnode list.
2017 	 * Hence the need to hold the vnode interlock throughout.
2018 	 */
2019 	MNT_ILOCK(mp);
2020 	VI_LOCK(vp);
2021 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
2022 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
2023 	    mp->mnt_nvnodelistsize == 0)) &&
2024 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
2025 		VI_UNLOCK(vp);
2026 		MNT_IUNLOCK(mp);
2027 		if (dtr) {
2028 			vp->v_data = NULL;
2029 			vp->v_op = &dead_vnodeops;
2030 			vgone(vp);
2031 			vput(vp);
2032 		}
2033 		return (EBUSY);
2034 	}
2035 	vp->v_mount = mp;
2036 	MNT_REF(mp);
2037 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2038 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
2039 		("neg mount point vnode list size"));
2040 	mp->mnt_nvnodelistsize++;
2041 	VI_UNLOCK(vp);
2042 	MNT_IUNLOCK(mp);
2043 	return (0);
2044 }
2045 
2046 /*
2047  * Insert into list of vnodes for the new mount point, if available.
2048  * insmntque() reclaims the vnode on insertion failure, insmntque1()
2049  * leaves handling of the vnode to the caller.
2050  */
2051 int
2052 insmntque(struct vnode *vp, struct mount *mp)
2053 {
2054 	return (insmntque1_int(vp, mp, true));
2055 }
2056 
2057 int
2058 insmntque1(struct vnode *vp, struct mount *mp)
2059 {
2060 	return (insmntque1_int(vp, mp, false));
2061 }
2062 
2063 /*
2064  * Flush out and invalidate all buffers associated with a bufobj
2065  * Called with the underlying object locked.
2066  */
2067 int
2068 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
2069 {
2070 	int error;
2071 
2072 	BO_LOCK(bo);
2073 	if (flags & V_SAVE) {
2074 		error = bufobj_wwait(bo, slpflag, slptimeo);
2075 		if (error) {
2076 			BO_UNLOCK(bo);
2077 			return (error);
2078 		}
2079 		if (bo->bo_dirty.bv_cnt > 0) {
2080 			BO_UNLOCK(bo);
2081 			do {
2082 				error = BO_SYNC(bo, MNT_WAIT);
2083 			} while (error == ERELOOKUP);
2084 			if (error != 0)
2085 				return (error);
2086 			BO_LOCK(bo);
2087 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) {
2088 				BO_UNLOCK(bo);
2089 				return (EBUSY);
2090 			}
2091 		}
2092 	}
2093 	/*
2094 	 * If you alter this loop please notice that interlock is dropped and
2095 	 * reacquired in flushbuflist.  Special care is needed to ensure that
2096 	 * no race conditions occur from this.
2097 	 */
2098 	do {
2099 		error = flushbuflist(&bo->bo_clean,
2100 		    flags, bo, slpflag, slptimeo);
2101 		if (error == 0 && !(flags & V_CLEANONLY))
2102 			error = flushbuflist(&bo->bo_dirty,
2103 			    flags, bo, slpflag, slptimeo);
2104 		if (error != 0 && error != EAGAIN) {
2105 			BO_UNLOCK(bo);
2106 			return (error);
2107 		}
2108 	} while (error != 0);
2109 
2110 	/*
2111 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
2112 	 * have write I/O in-progress but if there is a VM object then the
2113 	 * VM object can also have read-I/O in-progress.
2114 	 */
2115 	do {
2116 		bufobj_wwait(bo, 0, 0);
2117 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
2118 			BO_UNLOCK(bo);
2119 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
2120 			BO_LOCK(bo);
2121 		}
2122 	} while (bo->bo_numoutput > 0);
2123 	BO_UNLOCK(bo);
2124 
2125 	/*
2126 	 * Destroy the copy in the VM cache, too.
2127 	 */
2128 	if (bo->bo_object != NULL &&
2129 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
2130 		VM_OBJECT_WLOCK(bo->bo_object);
2131 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
2132 		    OBJPR_CLEANONLY : 0);
2133 		VM_OBJECT_WUNLOCK(bo->bo_object);
2134 	}
2135 
2136 #ifdef INVARIANTS
2137 	BO_LOCK(bo);
2138 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
2139 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
2140 	    bo->bo_clean.bv_cnt > 0))
2141 		panic("vinvalbuf: flush failed");
2142 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
2143 	    bo->bo_dirty.bv_cnt > 0)
2144 		panic("vinvalbuf: flush dirty failed");
2145 	BO_UNLOCK(bo);
2146 #endif
2147 	return (0);
2148 }
2149 
2150 /*
2151  * Flush out and invalidate all buffers associated with a vnode.
2152  * Called with the underlying object locked.
2153  */
2154 int
2155 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
2156 {
2157 
2158 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2159 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
2160 	if (vp->v_object != NULL && vp->v_object->handle != vp)
2161 		return (0);
2162 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
2163 }
2164 
2165 /*
2166  * Flush out buffers on the specified list.
2167  *
2168  */
2169 static int
2170 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2171     int slptimeo)
2172 {
2173 	struct buf *bp, *nbp;
2174 	int retval, error;
2175 	daddr_t lblkno;
2176 	b_xflags_t xflags;
2177 
2178 	ASSERT_BO_WLOCKED(bo);
2179 
2180 	retval = 0;
2181 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2182 		/*
2183 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2184 		 * do not skip any buffers. If we are flushing only V_NORMAL
2185 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2186 		 * flushing only V_ALT buffers then skip buffers not marked
2187 		 * as BX_ALTDATA.
2188 		 */
2189 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2190 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2191 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2192 			continue;
2193 		}
2194 		if (nbp != NULL) {
2195 			lblkno = nbp->b_lblkno;
2196 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2197 		}
2198 		retval = EAGAIN;
2199 		error = BUF_TIMELOCK(bp,
2200 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2201 		    "flushbuf", slpflag, slptimeo);
2202 		if (error) {
2203 			BO_LOCK(bo);
2204 			return (error != ENOLCK ? error : EAGAIN);
2205 		}
2206 		KASSERT(bp->b_bufobj == bo,
2207 		    ("bp %p wrong b_bufobj %p should be %p",
2208 		    bp, bp->b_bufobj, bo));
2209 		/*
2210 		 * XXX Since there are no node locks for NFS, I
2211 		 * believe there is a slight chance that a delayed
2212 		 * write will occur while sleeping just above, so
2213 		 * check for it.
2214 		 */
2215 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2216 		    (flags & V_SAVE)) {
2217 			bremfree(bp);
2218 			bp->b_flags |= B_ASYNC;
2219 			bwrite(bp);
2220 			BO_LOCK(bo);
2221 			return (EAGAIN);	/* XXX: why not loop ? */
2222 		}
2223 		bremfree(bp);
2224 		bp->b_flags |= (B_INVAL | B_RELBUF);
2225 		bp->b_flags &= ~B_ASYNC;
2226 		brelse(bp);
2227 		BO_LOCK(bo);
2228 		if (nbp == NULL)
2229 			break;
2230 		nbp = gbincore(bo, lblkno);
2231 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2232 		    != xflags)
2233 			break;			/* nbp invalid */
2234 	}
2235 	return (retval);
2236 }
2237 
2238 int
2239 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2240 {
2241 	struct buf *bp;
2242 	int error;
2243 	daddr_t lblkno;
2244 
2245 	ASSERT_BO_LOCKED(bo);
2246 
2247 	for (lblkno = startn;;) {
2248 again:
2249 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2250 		if (bp == NULL || bp->b_lblkno >= endn ||
2251 		    bp->b_lblkno < startn)
2252 			break;
2253 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2254 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2255 		if (error != 0) {
2256 			BO_RLOCK(bo);
2257 			if (error == ENOLCK)
2258 				goto again;
2259 			return (error);
2260 		}
2261 		KASSERT(bp->b_bufobj == bo,
2262 		    ("bp %p wrong b_bufobj %p should be %p",
2263 		    bp, bp->b_bufobj, bo));
2264 		lblkno = bp->b_lblkno + 1;
2265 		if ((bp->b_flags & B_MANAGED) == 0)
2266 			bremfree(bp);
2267 		bp->b_flags |= B_RELBUF;
2268 		/*
2269 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2270 		 * pages backing each buffer in the range are unlikely to be
2271 		 * reused.  Dirty buffers will have the hint applied once
2272 		 * they've been written.
2273 		 */
2274 		if ((bp->b_flags & B_VMIO) != 0)
2275 			bp->b_flags |= B_NOREUSE;
2276 		brelse(bp);
2277 		BO_RLOCK(bo);
2278 	}
2279 	return (0);
2280 }
2281 
2282 /*
2283  * Truncate a file's buffer and pages to a specified length.  This
2284  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2285  * sync activity.
2286  */
2287 int
2288 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2289 {
2290 	struct buf *bp, *nbp;
2291 	struct bufobj *bo;
2292 	daddr_t startlbn;
2293 
2294 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2295 	    vp, blksize, (uintmax_t)length);
2296 
2297 	/*
2298 	 * Round up to the *next* lbn.
2299 	 */
2300 	startlbn = howmany(length, blksize);
2301 
2302 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2303 
2304 	bo = &vp->v_bufobj;
2305 restart_unlocked:
2306 	BO_LOCK(bo);
2307 
2308 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2309 		;
2310 
2311 	if (length > 0) {
2312 restartsync:
2313 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2314 			if (bp->b_lblkno > 0)
2315 				continue;
2316 			/*
2317 			 * Since we hold the vnode lock this should only
2318 			 * fail if we're racing with the buf daemon.
2319 			 */
2320 			if (BUF_LOCK(bp,
2321 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2322 			    BO_LOCKPTR(bo)) == ENOLCK)
2323 				goto restart_unlocked;
2324 
2325 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2326 			    ("buf(%p) on dirty queue without DELWRI", bp));
2327 
2328 			bremfree(bp);
2329 			bawrite(bp);
2330 			BO_LOCK(bo);
2331 			goto restartsync;
2332 		}
2333 	}
2334 
2335 	bufobj_wwait(bo, 0, 0);
2336 	BO_UNLOCK(bo);
2337 	vnode_pager_setsize(vp, length);
2338 
2339 	return (0);
2340 }
2341 
2342 /*
2343  * Invalidate the cached pages of a file's buffer within the range of block
2344  * numbers [startlbn, endlbn).
2345  */
2346 void
2347 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2348     int blksize)
2349 {
2350 	struct bufobj *bo;
2351 	off_t start, end;
2352 
2353 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2354 
2355 	start = blksize * startlbn;
2356 	end = blksize * endlbn;
2357 
2358 	bo = &vp->v_bufobj;
2359 	BO_LOCK(bo);
2360 	MPASS(blksize == bo->bo_bsize);
2361 
2362 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2363 		;
2364 
2365 	BO_UNLOCK(bo);
2366 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2367 }
2368 
2369 static int
2370 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2371     daddr_t startlbn, daddr_t endlbn)
2372 {
2373 	struct buf *bp, *nbp;
2374 	bool anyfreed;
2375 
2376 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2377 	ASSERT_BO_LOCKED(bo);
2378 
2379 	do {
2380 		anyfreed = false;
2381 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2382 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2383 				continue;
2384 			if (BUF_LOCK(bp,
2385 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2386 			    BO_LOCKPTR(bo)) == ENOLCK) {
2387 				BO_LOCK(bo);
2388 				return (EAGAIN);
2389 			}
2390 
2391 			bremfree(bp);
2392 			bp->b_flags |= B_INVAL | B_RELBUF;
2393 			bp->b_flags &= ~B_ASYNC;
2394 			brelse(bp);
2395 			anyfreed = true;
2396 
2397 			BO_LOCK(bo);
2398 			if (nbp != NULL &&
2399 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2400 			    nbp->b_vp != vp ||
2401 			    (nbp->b_flags & B_DELWRI) != 0))
2402 				return (EAGAIN);
2403 		}
2404 
2405 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2406 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2407 				continue;
2408 			if (BUF_LOCK(bp,
2409 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2410 			    BO_LOCKPTR(bo)) == ENOLCK) {
2411 				BO_LOCK(bo);
2412 				return (EAGAIN);
2413 			}
2414 			bremfree(bp);
2415 			bp->b_flags |= B_INVAL | B_RELBUF;
2416 			bp->b_flags &= ~B_ASYNC;
2417 			brelse(bp);
2418 			anyfreed = true;
2419 
2420 			BO_LOCK(bo);
2421 			if (nbp != NULL &&
2422 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2423 			    (nbp->b_vp != vp) ||
2424 			    (nbp->b_flags & B_DELWRI) == 0))
2425 				return (EAGAIN);
2426 		}
2427 	} while (anyfreed);
2428 	return (0);
2429 }
2430 
2431 static void
2432 buf_vlist_remove(struct buf *bp)
2433 {
2434 	struct bufv *bv;
2435 	b_xflags_t flags;
2436 
2437 	flags = bp->b_xflags;
2438 
2439 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2440 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2441 	KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 &&
2442 	    (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN),
2443 	    ("%s: buffer %p has invalid queue state", __func__, bp));
2444 
2445 	if ((flags & BX_VNDIRTY) != 0)
2446 		bv = &bp->b_bufobj->bo_dirty;
2447 	else
2448 		bv = &bp->b_bufobj->bo_clean;
2449 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2450 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2451 	bv->bv_cnt--;
2452 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2453 }
2454 
2455 /*
2456  * Add the buffer to the sorted clean or dirty block list.
2457  *
2458  * NOTE: xflags is passed as a constant, optimizing this inline function!
2459  */
2460 static void
2461 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2462 {
2463 	struct bufv *bv;
2464 	struct buf *n;
2465 	int error;
2466 
2467 	ASSERT_BO_WLOCKED(bo);
2468 	KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2469 	    ("buf_vlist_add: bo %p does not allow bufs", bo));
2470 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2471 	    ("dead bo %p", bo));
2472 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2473 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2474 	bp->b_xflags |= xflags;
2475 	if (xflags & BX_VNDIRTY)
2476 		bv = &bo->bo_dirty;
2477 	else
2478 		bv = &bo->bo_clean;
2479 
2480 	/*
2481 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2482 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2483 	 * than _ge.
2484 	 */
2485 	if (bv->bv_cnt == 0 ||
2486 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2487 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2488 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2489 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2490 	else
2491 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2492 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2493 	if (error)
2494 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2495 	bv->bv_cnt++;
2496 }
2497 
2498 /*
2499  * Look up a buffer using the buffer tries.
2500  */
2501 struct buf *
2502 gbincore(struct bufobj *bo, daddr_t lblkno)
2503 {
2504 	struct buf *bp;
2505 
2506 	ASSERT_BO_LOCKED(bo);
2507 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2508 	if (bp != NULL)
2509 		return (bp);
2510 	return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno));
2511 }
2512 
2513 /*
2514  * Look up a buf using the buffer tries, without the bufobj lock.  This relies
2515  * on SMR for safe lookup, and bufs being in a no-free zone to provide type
2516  * stability of the result.  Like other lockless lookups, the found buf may
2517  * already be invalid by the time this function returns.
2518  */
2519 struct buf *
2520 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno)
2521 {
2522 	struct buf *bp;
2523 
2524 	ASSERT_BO_UNLOCKED(bo);
2525 	bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno);
2526 	if (bp != NULL)
2527 		return (bp);
2528 	return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno));
2529 }
2530 
2531 /*
2532  * Associate a buffer with a vnode.
2533  */
2534 void
2535 bgetvp(struct vnode *vp, struct buf *bp)
2536 {
2537 	struct bufobj *bo;
2538 
2539 	bo = &vp->v_bufobj;
2540 	ASSERT_BO_WLOCKED(bo);
2541 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2542 
2543 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2544 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2545 	    ("bgetvp: bp already attached! %p", bp));
2546 
2547 	vhold(vp);
2548 	bp->b_vp = vp;
2549 	bp->b_bufobj = bo;
2550 	/*
2551 	 * Insert onto list for new vnode.
2552 	 */
2553 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2554 }
2555 
2556 /*
2557  * Disassociate a buffer from a vnode.
2558  */
2559 void
2560 brelvp(struct buf *bp)
2561 {
2562 	struct bufobj *bo;
2563 	struct vnode *vp;
2564 
2565 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2566 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2567 
2568 	/*
2569 	 * Delete from old vnode list, if on one.
2570 	 */
2571 	vp = bp->b_vp;		/* XXX */
2572 	bo = bp->b_bufobj;
2573 	BO_LOCK(bo);
2574 	buf_vlist_remove(bp);
2575 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2576 		bo->bo_flag &= ~BO_ONWORKLST;
2577 		mtx_lock(&sync_mtx);
2578 		LIST_REMOVE(bo, bo_synclist);
2579 		syncer_worklist_len--;
2580 		mtx_unlock(&sync_mtx);
2581 	}
2582 	bp->b_vp = NULL;
2583 	bp->b_bufobj = NULL;
2584 	BO_UNLOCK(bo);
2585 	vdrop(vp);
2586 }
2587 
2588 /*
2589  * Add an item to the syncer work queue.
2590  */
2591 static void
2592 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2593 {
2594 	int slot;
2595 
2596 	ASSERT_BO_WLOCKED(bo);
2597 
2598 	mtx_lock(&sync_mtx);
2599 	if (bo->bo_flag & BO_ONWORKLST)
2600 		LIST_REMOVE(bo, bo_synclist);
2601 	else {
2602 		bo->bo_flag |= BO_ONWORKLST;
2603 		syncer_worklist_len++;
2604 	}
2605 
2606 	if (delay > syncer_maxdelay - 2)
2607 		delay = syncer_maxdelay - 2;
2608 	slot = (syncer_delayno + delay) & syncer_mask;
2609 
2610 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2611 	mtx_unlock(&sync_mtx);
2612 }
2613 
2614 static int
2615 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2616 {
2617 	int error, len;
2618 
2619 	mtx_lock(&sync_mtx);
2620 	len = syncer_worklist_len - sync_vnode_count;
2621 	mtx_unlock(&sync_mtx);
2622 	error = SYSCTL_OUT(req, &len, sizeof(len));
2623 	return (error);
2624 }
2625 
2626 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2627     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2628     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2629 
2630 static struct proc *updateproc;
2631 static void sched_sync(void);
2632 static struct kproc_desc up_kp = {
2633 	"syncer",
2634 	sched_sync,
2635 	&updateproc
2636 };
2637 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2638 
2639 static int
2640 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2641 {
2642 	struct vnode *vp;
2643 	struct mount *mp;
2644 
2645 	*bo = LIST_FIRST(slp);
2646 	if (*bo == NULL)
2647 		return (0);
2648 	vp = bo2vnode(*bo);
2649 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2650 		return (1);
2651 	/*
2652 	 * We use vhold in case the vnode does not
2653 	 * successfully sync.  vhold prevents the vnode from
2654 	 * going away when we unlock the sync_mtx so that
2655 	 * we can acquire the vnode interlock.
2656 	 */
2657 	vholdl(vp);
2658 	mtx_unlock(&sync_mtx);
2659 	VI_UNLOCK(vp);
2660 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2661 		vdrop(vp);
2662 		mtx_lock(&sync_mtx);
2663 		return (*bo == LIST_FIRST(slp));
2664 	}
2665 	MPASSERT(mp == NULL || (curthread->td_pflags & TDP_IGNSUSP) != 0 ||
2666 	    (mp->mnt_kern_flag & MNTK_SUSPENDED) == 0, mp,
2667 	    ("suspended mp syncing vp %p", vp));
2668 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2669 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2670 	VOP_UNLOCK(vp);
2671 	vn_finished_write(mp);
2672 	BO_LOCK(*bo);
2673 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2674 		/*
2675 		 * Put us back on the worklist.  The worklist
2676 		 * routine will remove us from our current
2677 		 * position and then add us back in at a later
2678 		 * position.
2679 		 */
2680 		vn_syncer_add_to_worklist(*bo, syncdelay);
2681 	}
2682 	BO_UNLOCK(*bo);
2683 	vdrop(vp);
2684 	mtx_lock(&sync_mtx);
2685 	return (0);
2686 }
2687 
2688 static int first_printf = 1;
2689 
2690 /*
2691  * System filesystem synchronizer daemon.
2692  */
2693 static void
2694 sched_sync(void)
2695 {
2696 	struct synclist *next, *slp;
2697 	struct bufobj *bo;
2698 	long starttime;
2699 	struct thread *td = curthread;
2700 	int last_work_seen;
2701 	int net_worklist_len;
2702 	int syncer_final_iter;
2703 	int error;
2704 
2705 	last_work_seen = 0;
2706 	syncer_final_iter = 0;
2707 	syncer_state = SYNCER_RUNNING;
2708 	starttime = time_uptime;
2709 	td->td_pflags |= TDP_NORUNNINGBUF;
2710 
2711 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2712 	    SHUTDOWN_PRI_LAST);
2713 
2714 	mtx_lock(&sync_mtx);
2715 	for (;;) {
2716 		if (syncer_state == SYNCER_FINAL_DELAY &&
2717 		    syncer_final_iter == 0) {
2718 			mtx_unlock(&sync_mtx);
2719 			kproc_suspend_check(td->td_proc);
2720 			mtx_lock(&sync_mtx);
2721 		}
2722 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2723 		if (syncer_state != SYNCER_RUNNING &&
2724 		    starttime != time_uptime) {
2725 			if (first_printf) {
2726 				printf("\nSyncing disks, vnodes remaining... ");
2727 				first_printf = 0;
2728 			}
2729 			printf("%d ", net_worklist_len);
2730 		}
2731 		starttime = time_uptime;
2732 
2733 		/*
2734 		 * Push files whose dirty time has expired.  Be careful
2735 		 * of interrupt race on slp queue.
2736 		 *
2737 		 * Skip over empty worklist slots when shutting down.
2738 		 */
2739 		do {
2740 			slp = &syncer_workitem_pending[syncer_delayno];
2741 			syncer_delayno += 1;
2742 			if (syncer_delayno == syncer_maxdelay)
2743 				syncer_delayno = 0;
2744 			next = &syncer_workitem_pending[syncer_delayno];
2745 			/*
2746 			 * If the worklist has wrapped since the
2747 			 * it was emptied of all but syncer vnodes,
2748 			 * switch to the FINAL_DELAY state and run
2749 			 * for one more second.
2750 			 */
2751 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2752 			    net_worklist_len == 0 &&
2753 			    last_work_seen == syncer_delayno) {
2754 				syncer_state = SYNCER_FINAL_DELAY;
2755 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2756 			}
2757 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2758 		    syncer_worklist_len > 0);
2759 
2760 		/*
2761 		 * Keep track of the last time there was anything
2762 		 * on the worklist other than syncer vnodes.
2763 		 * Return to the SHUTTING_DOWN state if any
2764 		 * new work appears.
2765 		 */
2766 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2767 			last_work_seen = syncer_delayno;
2768 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2769 			syncer_state = SYNCER_SHUTTING_DOWN;
2770 		while (!LIST_EMPTY(slp)) {
2771 			error = sync_vnode(slp, &bo, td);
2772 			if (error == 1) {
2773 				LIST_REMOVE(bo, bo_synclist);
2774 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2775 				continue;
2776 			}
2777 
2778 			if (first_printf == 0) {
2779 				/*
2780 				 * Drop the sync mutex, because some watchdog
2781 				 * drivers need to sleep while patting
2782 				 */
2783 				mtx_unlock(&sync_mtx);
2784 				wdog_kern_pat(WD_LASTVAL);
2785 				mtx_lock(&sync_mtx);
2786 			}
2787 		}
2788 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2789 			syncer_final_iter--;
2790 		/*
2791 		 * The variable rushjob allows the kernel to speed up the
2792 		 * processing of the filesystem syncer process. A rushjob
2793 		 * value of N tells the filesystem syncer to process the next
2794 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2795 		 * is used by the soft update code to speed up the filesystem
2796 		 * syncer process when the incore state is getting so far
2797 		 * ahead of the disk that the kernel memory pool is being
2798 		 * threatened with exhaustion.
2799 		 */
2800 		if (rushjob > 0) {
2801 			rushjob -= 1;
2802 			continue;
2803 		}
2804 		/*
2805 		 * Just sleep for a short period of time between
2806 		 * iterations when shutting down to allow some I/O
2807 		 * to happen.
2808 		 *
2809 		 * If it has taken us less than a second to process the
2810 		 * current work, then wait. Otherwise start right over
2811 		 * again. We can still lose time if any single round
2812 		 * takes more than two seconds, but it does not really
2813 		 * matter as we are just trying to generally pace the
2814 		 * filesystem activity.
2815 		 */
2816 		if (syncer_state != SYNCER_RUNNING ||
2817 		    time_uptime == starttime) {
2818 			thread_lock(td);
2819 			sched_prio(td, PPAUSE);
2820 			thread_unlock(td);
2821 		}
2822 		if (syncer_state != SYNCER_RUNNING)
2823 			cv_timedwait(&sync_wakeup, &sync_mtx,
2824 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2825 		else if (time_uptime == starttime)
2826 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2827 	}
2828 }
2829 
2830 /*
2831  * Request the syncer daemon to speed up its work.
2832  * We never push it to speed up more than half of its
2833  * normal turn time, otherwise it could take over the cpu.
2834  */
2835 int
2836 speedup_syncer(void)
2837 {
2838 	int ret = 0;
2839 
2840 	mtx_lock(&sync_mtx);
2841 	if (rushjob < syncdelay / 2) {
2842 		rushjob += 1;
2843 		stat_rush_requests += 1;
2844 		ret = 1;
2845 	}
2846 	mtx_unlock(&sync_mtx);
2847 	cv_broadcast(&sync_wakeup);
2848 	return (ret);
2849 }
2850 
2851 /*
2852  * Tell the syncer to speed up its work and run though its work
2853  * list several times, then tell it to shut down.
2854  */
2855 static void
2856 syncer_shutdown(void *arg, int howto)
2857 {
2858 
2859 	if (howto & RB_NOSYNC)
2860 		return;
2861 	mtx_lock(&sync_mtx);
2862 	syncer_state = SYNCER_SHUTTING_DOWN;
2863 	rushjob = 0;
2864 	mtx_unlock(&sync_mtx);
2865 	cv_broadcast(&sync_wakeup);
2866 	kproc_shutdown(arg, howto);
2867 }
2868 
2869 void
2870 syncer_suspend(void)
2871 {
2872 
2873 	syncer_shutdown(updateproc, 0);
2874 }
2875 
2876 void
2877 syncer_resume(void)
2878 {
2879 
2880 	mtx_lock(&sync_mtx);
2881 	first_printf = 1;
2882 	syncer_state = SYNCER_RUNNING;
2883 	mtx_unlock(&sync_mtx);
2884 	cv_broadcast(&sync_wakeup);
2885 	kproc_resume(updateproc);
2886 }
2887 
2888 /*
2889  * Move the buffer between the clean and dirty lists of its vnode.
2890  */
2891 void
2892 reassignbuf(struct buf *bp)
2893 {
2894 	struct vnode *vp;
2895 	struct bufobj *bo;
2896 	int delay;
2897 #ifdef INVARIANTS
2898 	struct bufv *bv;
2899 #endif
2900 
2901 	vp = bp->b_vp;
2902 	bo = bp->b_bufobj;
2903 
2904 	KASSERT((bp->b_flags & B_PAGING) == 0,
2905 	    ("%s: cannot reassign paging buffer %p", __func__, bp));
2906 
2907 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2908 	    bp, bp->b_vp, bp->b_flags);
2909 
2910 	BO_LOCK(bo);
2911 	buf_vlist_remove(bp);
2912 
2913 	/*
2914 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2915 	 * of clean buffers.
2916 	 */
2917 	if (bp->b_flags & B_DELWRI) {
2918 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2919 			switch (vp->v_type) {
2920 			case VDIR:
2921 				delay = dirdelay;
2922 				break;
2923 			case VCHR:
2924 				delay = metadelay;
2925 				break;
2926 			default:
2927 				delay = filedelay;
2928 			}
2929 			vn_syncer_add_to_worklist(bo, delay);
2930 		}
2931 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2932 	} else {
2933 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2934 
2935 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2936 			mtx_lock(&sync_mtx);
2937 			LIST_REMOVE(bo, bo_synclist);
2938 			syncer_worklist_len--;
2939 			mtx_unlock(&sync_mtx);
2940 			bo->bo_flag &= ~BO_ONWORKLST;
2941 		}
2942 	}
2943 #ifdef INVARIANTS
2944 	bv = &bo->bo_clean;
2945 	bp = TAILQ_FIRST(&bv->bv_hd);
2946 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2947 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2948 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2949 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2950 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2951 	bv = &bo->bo_dirty;
2952 	bp = TAILQ_FIRST(&bv->bv_hd);
2953 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2954 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2955 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2956 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2957 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2958 #endif
2959 	BO_UNLOCK(bo);
2960 }
2961 
2962 static void
2963 v_init_counters(struct vnode *vp)
2964 {
2965 
2966 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2967 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2968 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2969 
2970 	refcount_init(&vp->v_holdcnt, 1);
2971 	refcount_init(&vp->v_usecount, 1);
2972 }
2973 
2974 /*
2975  * Grab a particular vnode from the free list, increment its
2976  * reference count and lock it.  VIRF_DOOMED is set if the vnode
2977  * is being destroyed.  Only callers who specify LK_RETRY will
2978  * see doomed vnodes.  If inactive processing was delayed in
2979  * vput try to do it here.
2980  *
2981  * usecount is manipulated using atomics without holding any locks.
2982  *
2983  * holdcnt can be manipulated using atomics without holding any locks,
2984  * except when transitioning 1<->0, in which case the interlock is held.
2985  *
2986  * Consumers which don't guarantee liveness of the vnode can use SMR to
2987  * try to get a reference. Note this operation can fail since the vnode
2988  * may be awaiting getting freed by the time they get to it.
2989  */
2990 enum vgetstate
2991 vget_prep_smr(struct vnode *vp)
2992 {
2993 	enum vgetstate vs;
2994 
2995 	VFS_SMR_ASSERT_ENTERED();
2996 
2997 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2998 		vs = VGET_USECOUNT;
2999 	} else {
3000 		if (vhold_smr(vp))
3001 			vs = VGET_HOLDCNT;
3002 		else
3003 			vs = VGET_NONE;
3004 	}
3005 	return (vs);
3006 }
3007 
3008 enum vgetstate
3009 vget_prep(struct vnode *vp)
3010 {
3011 	enum vgetstate vs;
3012 
3013 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
3014 		vs = VGET_USECOUNT;
3015 	} else {
3016 		vhold(vp);
3017 		vs = VGET_HOLDCNT;
3018 	}
3019 	return (vs);
3020 }
3021 
3022 void
3023 vget_abort(struct vnode *vp, enum vgetstate vs)
3024 {
3025 
3026 	switch (vs) {
3027 	case VGET_USECOUNT:
3028 		vrele(vp);
3029 		break;
3030 	case VGET_HOLDCNT:
3031 		vdrop(vp);
3032 		break;
3033 	default:
3034 		__assert_unreachable();
3035 	}
3036 }
3037 
3038 int
3039 vget(struct vnode *vp, int flags)
3040 {
3041 	enum vgetstate vs;
3042 
3043 	vs = vget_prep(vp);
3044 	return (vget_finish(vp, flags, vs));
3045 }
3046 
3047 int
3048 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
3049 {
3050 	int error;
3051 
3052 	if ((flags & LK_INTERLOCK) != 0)
3053 		ASSERT_VI_LOCKED(vp, __func__);
3054 	else
3055 		ASSERT_VI_UNLOCKED(vp, __func__);
3056 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3057 	VNPASS(vp->v_holdcnt > 0, vp);
3058 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3059 
3060 	error = vn_lock(vp, flags);
3061 	if (__predict_false(error != 0)) {
3062 		vget_abort(vp, vs);
3063 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
3064 		    vp);
3065 		return (error);
3066 	}
3067 
3068 	vget_finish_ref(vp, vs);
3069 	return (0);
3070 }
3071 
3072 void
3073 vget_finish_ref(struct vnode *vp, enum vgetstate vs)
3074 {
3075 	int old;
3076 
3077 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3078 	VNPASS(vp->v_holdcnt > 0, vp);
3079 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3080 
3081 	if (vs == VGET_USECOUNT)
3082 		return;
3083 
3084 	/*
3085 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
3086 	 * the vnode around. Otherwise someone else lended their hold count and
3087 	 * we have to drop ours.
3088 	 */
3089 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
3090 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3091 	if (old != 0) {
3092 #ifdef INVARIANTS
3093 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3094 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3095 #else
3096 		refcount_release(&vp->v_holdcnt);
3097 #endif
3098 	}
3099 }
3100 
3101 void
3102 vref(struct vnode *vp)
3103 {
3104 	enum vgetstate vs;
3105 
3106 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3107 	vs = vget_prep(vp);
3108 	vget_finish_ref(vp, vs);
3109 }
3110 
3111 void
3112 vrefact(struct vnode *vp)
3113 {
3114 
3115 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3116 #ifdef INVARIANTS
3117 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3118 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3119 #else
3120 	refcount_acquire(&vp->v_usecount);
3121 #endif
3122 }
3123 
3124 void
3125 vlazy(struct vnode *vp)
3126 {
3127 	struct mount *mp;
3128 
3129 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3130 
3131 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3132 		return;
3133 	/*
3134 	 * We may get here for inactive routines after the vnode got doomed.
3135 	 */
3136 	if (VN_IS_DOOMED(vp))
3137 		return;
3138 	mp = vp->v_mount;
3139 	mtx_lock(&mp->mnt_listmtx);
3140 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3141 		vp->v_mflag |= VMP_LAZYLIST;
3142 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3143 		mp->mnt_lazyvnodelistsize++;
3144 	}
3145 	mtx_unlock(&mp->mnt_listmtx);
3146 }
3147 
3148 static void
3149 vunlazy(struct vnode *vp)
3150 {
3151 	struct mount *mp;
3152 
3153 	ASSERT_VI_LOCKED(vp, __func__);
3154 	VNPASS(!VN_IS_DOOMED(vp), vp);
3155 
3156 	mp = vp->v_mount;
3157 	mtx_lock(&mp->mnt_listmtx);
3158 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3159 	/*
3160 	 * Don't remove the vnode from the lazy list if another thread
3161 	 * has increased the hold count. It may have re-enqueued the
3162 	 * vnode to the lazy list and is now responsible for its
3163 	 * removal.
3164 	 */
3165 	if (vp->v_holdcnt == 0) {
3166 		vp->v_mflag &= ~VMP_LAZYLIST;
3167 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3168 		mp->mnt_lazyvnodelistsize--;
3169 	}
3170 	mtx_unlock(&mp->mnt_listmtx);
3171 }
3172 
3173 /*
3174  * This routine is only meant to be called from vgonel prior to dooming
3175  * the vnode.
3176  */
3177 static void
3178 vunlazy_gone(struct vnode *vp)
3179 {
3180 	struct mount *mp;
3181 
3182 	ASSERT_VOP_ELOCKED(vp, __func__);
3183 	ASSERT_VI_LOCKED(vp, __func__);
3184 	VNPASS(!VN_IS_DOOMED(vp), vp);
3185 
3186 	if (vp->v_mflag & VMP_LAZYLIST) {
3187 		mp = vp->v_mount;
3188 		mtx_lock(&mp->mnt_listmtx);
3189 		VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3190 		vp->v_mflag &= ~VMP_LAZYLIST;
3191 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3192 		mp->mnt_lazyvnodelistsize--;
3193 		mtx_unlock(&mp->mnt_listmtx);
3194 	}
3195 }
3196 
3197 static void
3198 vdefer_inactive(struct vnode *vp)
3199 {
3200 
3201 	ASSERT_VI_LOCKED(vp, __func__);
3202 	VNPASS(vp->v_holdcnt > 0, vp);
3203 	if (VN_IS_DOOMED(vp)) {
3204 		vdropl(vp);
3205 		return;
3206 	}
3207 	if (vp->v_iflag & VI_DEFINACT) {
3208 		VNPASS(vp->v_holdcnt > 1, vp);
3209 		vdropl(vp);
3210 		return;
3211 	}
3212 	if (vp->v_usecount > 0) {
3213 		vp->v_iflag &= ~VI_OWEINACT;
3214 		vdropl(vp);
3215 		return;
3216 	}
3217 	vlazy(vp);
3218 	vp->v_iflag |= VI_DEFINACT;
3219 	VI_UNLOCK(vp);
3220 	atomic_add_long(&deferred_inact, 1);
3221 }
3222 
3223 static void
3224 vdefer_inactive_unlocked(struct vnode *vp)
3225 {
3226 
3227 	VI_LOCK(vp);
3228 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3229 		vdropl(vp);
3230 		return;
3231 	}
3232 	vdefer_inactive(vp);
3233 }
3234 
3235 enum vput_op { VRELE, VPUT, VUNREF };
3236 
3237 /*
3238  * Handle ->v_usecount transitioning to 0.
3239  *
3240  * By releasing the last usecount we take ownership of the hold count which
3241  * provides liveness of the vnode, meaning we have to vdrop.
3242  *
3243  * For all vnodes we may need to perform inactive processing. It requires an
3244  * exclusive lock on the vnode, while it is legal to call here with only a
3245  * shared lock (or no locks). If locking the vnode in an expected manner fails,
3246  * inactive processing gets deferred to the syncer.
3247  *
3248  * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3249  * on the lock being held all the way until VOP_INACTIVE. This in particular
3250  * happens with UFS which adds half-constructed vnodes to the hash, where they
3251  * can be found by other code.
3252  */
3253 static void
3254 vput_final(struct vnode *vp, enum vput_op func)
3255 {
3256 	int error;
3257 	bool want_unlock;
3258 
3259 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3260 	VNPASS(vp->v_holdcnt > 0, vp);
3261 
3262 	VI_LOCK(vp);
3263 
3264 	/*
3265 	 * By the time we got here someone else might have transitioned
3266 	 * the count back to > 0.
3267 	 */
3268 	if (vp->v_usecount > 0)
3269 		goto out;
3270 
3271 	/*
3272 	 * If the vnode is doomed vgone already performed inactive processing
3273 	 * (if needed).
3274 	 */
3275 	if (VN_IS_DOOMED(vp))
3276 		goto out;
3277 
3278 	if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3279 		goto out;
3280 
3281 	if (vp->v_iflag & VI_DOINGINACT)
3282 		goto out;
3283 
3284 	/*
3285 	 * Locking operations here will drop the interlock and possibly the
3286 	 * vnode lock, opening a window where the vnode can get doomed all the
3287 	 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3288 	 * perform inactive.
3289 	 */
3290 	vp->v_iflag |= VI_OWEINACT;
3291 	want_unlock = false;
3292 	error = 0;
3293 	switch (func) {
3294 	case VRELE:
3295 		switch (VOP_ISLOCKED(vp)) {
3296 		case LK_EXCLUSIVE:
3297 			break;
3298 		case LK_EXCLOTHER:
3299 		case 0:
3300 			want_unlock = true;
3301 			error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3302 			VI_LOCK(vp);
3303 			break;
3304 		default:
3305 			/*
3306 			 * The lock has at least one sharer, but we have no way
3307 			 * to conclude whether this is us. Play it safe and
3308 			 * defer processing.
3309 			 */
3310 			error = EAGAIN;
3311 			break;
3312 		}
3313 		break;
3314 	case VPUT:
3315 		want_unlock = true;
3316 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3317 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3318 			    LK_NOWAIT);
3319 			VI_LOCK(vp);
3320 		}
3321 		break;
3322 	case VUNREF:
3323 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3324 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3325 			VI_LOCK(vp);
3326 		}
3327 		break;
3328 	}
3329 	if (error == 0) {
3330 		if (func == VUNREF) {
3331 			VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp,
3332 			    ("recursive vunref"));
3333 			vp->v_vflag |= VV_UNREF;
3334 		}
3335 		for (;;) {
3336 			error = vinactive(vp);
3337 			if (want_unlock)
3338 				VOP_UNLOCK(vp);
3339 			if (error != ERELOOKUP || !want_unlock)
3340 				break;
3341 			VOP_LOCK(vp, LK_EXCLUSIVE);
3342 		}
3343 		if (func == VUNREF)
3344 			vp->v_vflag &= ~VV_UNREF;
3345 		vdropl(vp);
3346 	} else {
3347 		vdefer_inactive(vp);
3348 	}
3349 	return;
3350 out:
3351 	if (func == VPUT)
3352 		VOP_UNLOCK(vp);
3353 	vdropl(vp);
3354 }
3355 
3356 /*
3357  * Decrement ->v_usecount for a vnode.
3358  *
3359  * Releasing the last use count requires additional processing, see vput_final
3360  * above for details.
3361  *
3362  * Comment above each variant denotes lock state on entry and exit.
3363  */
3364 
3365 /*
3366  * in: any
3367  * out: same as passed in
3368  */
3369 void
3370 vrele(struct vnode *vp)
3371 {
3372 
3373 	ASSERT_VI_UNLOCKED(vp, __func__);
3374 	if (!refcount_release(&vp->v_usecount))
3375 		return;
3376 	vput_final(vp, VRELE);
3377 }
3378 
3379 /*
3380  * in: locked
3381  * out: unlocked
3382  */
3383 void
3384 vput(struct vnode *vp)
3385 {
3386 
3387 	ASSERT_VOP_LOCKED(vp, __func__);
3388 	ASSERT_VI_UNLOCKED(vp, __func__);
3389 	if (!refcount_release(&vp->v_usecount)) {
3390 		VOP_UNLOCK(vp);
3391 		return;
3392 	}
3393 	vput_final(vp, VPUT);
3394 }
3395 
3396 /*
3397  * in: locked
3398  * out: locked
3399  */
3400 void
3401 vunref(struct vnode *vp)
3402 {
3403 
3404 	ASSERT_VOP_LOCKED(vp, __func__);
3405 	ASSERT_VI_UNLOCKED(vp, __func__);
3406 	if (!refcount_release(&vp->v_usecount))
3407 		return;
3408 	vput_final(vp, VUNREF);
3409 }
3410 
3411 void
3412 vhold(struct vnode *vp)
3413 {
3414 	int old;
3415 
3416 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3417 	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3418 	VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3419 	    ("%s: wrong hold count %d", __func__, old));
3420 	if (old == 0)
3421 		vfs_freevnodes_dec();
3422 }
3423 
3424 void
3425 vholdnz(struct vnode *vp)
3426 {
3427 
3428 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3429 #ifdef INVARIANTS
3430 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3431 	VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3432 	    ("%s: wrong hold count %d", __func__, old));
3433 #else
3434 	atomic_add_int(&vp->v_holdcnt, 1);
3435 #endif
3436 }
3437 
3438 /*
3439  * Grab a hold count unless the vnode is freed.
3440  *
3441  * Only use this routine if vfs smr is the only protection you have against
3442  * freeing the vnode.
3443  *
3444  * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag
3445  * is not set.  After the flag is set the vnode becomes immutable to anyone but
3446  * the thread which managed to set the flag.
3447  *
3448  * It may be tempting to replace the loop with:
3449  * count = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3450  * if (count & VHOLD_NO_SMR) {
3451  *     backpedal and error out;
3452  * }
3453  *
3454  * However, while this is more performant, it hinders debugging by eliminating
3455  * the previously mentioned invariant.
3456  */
3457 bool
3458 vhold_smr(struct vnode *vp)
3459 {
3460 	int count;
3461 
3462 	VFS_SMR_ASSERT_ENTERED();
3463 
3464 	count = atomic_load_int(&vp->v_holdcnt);
3465 	for (;;) {
3466 		if (count & VHOLD_NO_SMR) {
3467 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3468 			    ("non-zero hold count with flags %d\n", count));
3469 			return (false);
3470 		}
3471 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3472 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3473 			if (count == 0)
3474 				vfs_freevnodes_dec();
3475 			return (true);
3476 		}
3477 	}
3478 }
3479 
3480 /*
3481  * Hold a free vnode for recycling.
3482  *
3483  * Note: vnode_init references this comment.
3484  *
3485  * Attempts to recycle only need the global vnode list lock and have no use for
3486  * SMR.
3487  *
3488  * However, vnodes get inserted into the global list before they get fully
3489  * initialized and stay there until UMA decides to free the memory. This in
3490  * particular means the target can be found before it becomes usable and after
3491  * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to
3492  * VHOLD_NO_SMR.
3493  *
3494  * Note: the vnode may gain more references after we transition the count 0->1.
3495  */
3496 static bool
3497 vhold_recycle_free(struct vnode *vp)
3498 {
3499 	int count;
3500 
3501 	mtx_assert(&vnode_list_mtx, MA_OWNED);
3502 
3503 	count = atomic_load_int(&vp->v_holdcnt);
3504 	for (;;) {
3505 		if (count & VHOLD_NO_SMR) {
3506 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3507 			    ("non-zero hold count with flags %d\n", count));
3508 			return (false);
3509 		}
3510 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3511 		if (count > 0) {
3512 			return (false);
3513 		}
3514 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3515 			vfs_freevnodes_dec();
3516 			return (true);
3517 		}
3518 	}
3519 }
3520 
3521 static void __noinline
3522 vdbatch_process(struct vdbatch *vd)
3523 {
3524 	struct vnode *vp;
3525 	int i;
3526 
3527 	mtx_assert(&vd->lock, MA_OWNED);
3528 	MPASS(curthread->td_pinned > 0);
3529 	MPASS(vd->index == VDBATCH_SIZE);
3530 
3531 	critical_enter();
3532 	if (mtx_trylock(&vnode_list_mtx)) {
3533 		for (i = 0; i < VDBATCH_SIZE; i++) {
3534 			vp = vd->tab[i];
3535 			vd->tab[i] = NULL;
3536 			TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3537 			TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3538 			MPASS(vp->v_dbatchcpu != NOCPU);
3539 			vp->v_dbatchcpu = NOCPU;
3540 		}
3541 		mtx_unlock(&vnode_list_mtx);
3542 	} else {
3543 		for (i = 0; i < VDBATCH_SIZE; i++) {
3544 			vp = vd->tab[i];
3545 			vd->tab[i] = NULL;
3546 			MPASS(vp->v_dbatchcpu != NOCPU);
3547 			vp->v_dbatchcpu = NOCPU;
3548 		}
3549 	}
3550 	vd->index = 0;
3551 	critical_exit();
3552 }
3553 
3554 static void
3555 vdbatch_enqueue(struct vnode *vp)
3556 {
3557 	struct vdbatch *vd;
3558 
3559 	ASSERT_VI_LOCKED(vp, __func__);
3560 	VNPASS(!VN_IS_DOOMED(vp), vp);
3561 
3562 	if (vp->v_dbatchcpu != NOCPU) {
3563 		VI_UNLOCK(vp);
3564 		return;
3565 	}
3566 
3567 	sched_pin();
3568 	vd = DPCPU_PTR(vd);
3569 	mtx_lock(&vd->lock);
3570 	MPASS(vd->index < VDBATCH_SIZE);
3571 	MPASS(vd->tab[vd->index] == NULL);
3572 	/*
3573 	 * A hack: we depend on being pinned so that we know what to put in
3574 	 * ->v_dbatchcpu.
3575 	 */
3576 	vp->v_dbatchcpu = curcpu;
3577 	vd->tab[vd->index] = vp;
3578 	vd->index++;
3579 	VI_UNLOCK(vp);
3580 	if (vd->index == VDBATCH_SIZE)
3581 		vdbatch_process(vd);
3582 	mtx_unlock(&vd->lock);
3583 	sched_unpin();
3584 }
3585 
3586 /*
3587  * This routine must only be called for vnodes which are about to be
3588  * deallocated. Supporting dequeue for arbitrary vndoes would require
3589  * validating that the locked batch matches.
3590  */
3591 static void
3592 vdbatch_dequeue(struct vnode *vp)
3593 {
3594 	struct vdbatch *vd;
3595 	int i;
3596 	short cpu;
3597 
3598 	VNPASS(vp->v_type == VBAD || vp->v_type == VNON, vp);
3599 
3600 	cpu = vp->v_dbatchcpu;
3601 	if (cpu == NOCPU)
3602 		return;
3603 
3604 	vd = DPCPU_ID_PTR(cpu, vd);
3605 	mtx_lock(&vd->lock);
3606 	for (i = 0; i < vd->index; i++) {
3607 		if (vd->tab[i] != vp)
3608 			continue;
3609 		vp->v_dbatchcpu = NOCPU;
3610 		vd->index--;
3611 		vd->tab[i] = vd->tab[vd->index];
3612 		vd->tab[vd->index] = NULL;
3613 		break;
3614 	}
3615 	mtx_unlock(&vd->lock);
3616 	/*
3617 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3618 	 */
3619 	MPASS(vp->v_dbatchcpu == NOCPU);
3620 }
3621 
3622 /*
3623  * Drop the hold count of the vnode.  If this is the last reference to
3624  * the vnode we place it on the free list unless it has been vgone'd
3625  * (marked VIRF_DOOMED) in which case we will free it.
3626  *
3627  * Because the vnode vm object keeps a hold reference on the vnode if
3628  * there is at least one resident non-cached page, the vnode cannot
3629  * leave the active list without the page cleanup done.
3630  */
3631 static void __noinline
3632 vdropl_final(struct vnode *vp)
3633 {
3634 
3635 	ASSERT_VI_LOCKED(vp, __func__);
3636 	VNPASS(VN_IS_DOOMED(vp), vp);
3637 	/*
3638 	 * Set the VHOLD_NO_SMR flag.
3639 	 *
3640 	 * We may be racing against vhold_smr. If they win we can just pretend
3641 	 * we never got this far, they will vdrop later.
3642 	 */
3643 	if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) {
3644 		vfs_freevnodes_inc();
3645 		VI_UNLOCK(vp);
3646 		/*
3647 		 * We lost the aforementioned race. Any subsequent access is
3648 		 * invalid as they might have managed to vdropl on their own.
3649 		 */
3650 		return;
3651 	}
3652 	/*
3653 	 * Don't bump freevnodes as this one is going away.
3654 	 */
3655 	freevnode(vp);
3656 }
3657 
3658 void
3659 vdrop(struct vnode *vp)
3660 {
3661 
3662 	ASSERT_VI_UNLOCKED(vp, __func__);
3663 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3664 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3665 		return;
3666 	VI_LOCK(vp);
3667 	vdropl(vp);
3668 }
3669 
3670 static void __always_inline
3671 vdropl_impl(struct vnode *vp, bool enqueue)
3672 {
3673 
3674 	ASSERT_VI_LOCKED(vp, __func__);
3675 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3676 	if (!refcount_release(&vp->v_holdcnt)) {
3677 		VI_UNLOCK(vp);
3678 		return;
3679 	}
3680 	VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp);
3681 	VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
3682 	if (VN_IS_DOOMED(vp)) {
3683 		vdropl_final(vp);
3684 		return;
3685 	}
3686 
3687 	vfs_freevnodes_inc();
3688 	if (vp->v_mflag & VMP_LAZYLIST) {
3689 		vunlazy(vp);
3690 	}
3691 
3692 	if (!enqueue) {
3693 		VI_UNLOCK(vp);
3694 		return;
3695 	}
3696 
3697 	/*
3698 	 * Also unlocks the interlock. We can't assert on it as we
3699 	 * released our hold and by now the vnode might have been
3700 	 * freed.
3701 	 */
3702 	vdbatch_enqueue(vp);
3703 }
3704 
3705 void
3706 vdropl(struct vnode *vp)
3707 {
3708 
3709 	vdropl_impl(vp, true);
3710 }
3711 
3712 /*
3713  * vdrop a vnode when recycling
3714  *
3715  * This is a special case routine only to be used when recycling, differs from
3716  * regular vdrop by not requeieing the vnode on LRU.
3717  *
3718  * Consider a case where vtryrecycle continuously fails with all vnodes (due to
3719  * e.g., frozen writes on the filesystem), filling the batch and causing it to
3720  * be requeued. Then vnlru will end up revisiting the same vnodes. This is a
3721  * loop which can last for as long as writes are frozen.
3722  */
3723 static void
3724 vdropl_recycle(struct vnode *vp)
3725 {
3726 
3727 	vdropl_impl(vp, false);
3728 }
3729 
3730 static void
3731 vdrop_recycle(struct vnode *vp)
3732 {
3733 
3734 	VI_LOCK(vp);
3735 	vdropl_recycle(vp);
3736 }
3737 
3738 /*
3739  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3740  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3741  */
3742 static int
3743 vinactivef(struct vnode *vp)
3744 {
3745 	struct vm_object *obj;
3746 	int error;
3747 
3748 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3749 	ASSERT_VI_LOCKED(vp, "vinactive");
3750 	VNPASS((vp->v_iflag & VI_DOINGINACT) == 0, vp);
3751 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3752 	vp->v_iflag |= VI_DOINGINACT;
3753 	vp->v_iflag &= ~VI_OWEINACT;
3754 	VI_UNLOCK(vp);
3755 	/*
3756 	 * Before moving off the active list, we must be sure that any
3757 	 * modified pages are converted into the vnode's dirty
3758 	 * buffers, since these will no longer be checked once the
3759 	 * vnode is on the inactive list.
3760 	 *
3761 	 * The write-out of the dirty pages is asynchronous.  At the
3762 	 * point that VOP_INACTIVE() is called, there could still be
3763 	 * pending I/O and dirty pages in the object.
3764 	 */
3765 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3766 	    vm_object_mightbedirty(obj)) {
3767 		VM_OBJECT_WLOCK(obj);
3768 		vm_object_page_clean(obj, 0, 0, 0);
3769 		VM_OBJECT_WUNLOCK(obj);
3770 	}
3771 	error = VOP_INACTIVE(vp);
3772 	VI_LOCK(vp);
3773 	VNPASS(vp->v_iflag & VI_DOINGINACT, vp);
3774 	vp->v_iflag &= ~VI_DOINGINACT;
3775 	return (error);
3776 }
3777 
3778 int
3779 vinactive(struct vnode *vp)
3780 {
3781 
3782 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3783 	ASSERT_VI_LOCKED(vp, "vinactive");
3784 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3785 
3786 	if ((vp->v_iflag & VI_OWEINACT) == 0)
3787 		return (0);
3788 	if (vp->v_iflag & VI_DOINGINACT)
3789 		return (0);
3790 	if (vp->v_usecount > 0) {
3791 		vp->v_iflag &= ~VI_OWEINACT;
3792 		return (0);
3793 	}
3794 	return (vinactivef(vp));
3795 }
3796 
3797 /*
3798  * Remove any vnodes in the vnode table belonging to mount point mp.
3799  *
3800  * If FORCECLOSE is not specified, there should not be any active ones,
3801  * return error if any are found (nb: this is a user error, not a
3802  * system error). If FORCECLOSE is specified, detach any active vnodes
3803  * that are found.
3804  *
3805  * If WRITECLOSE is set, only flush out regular file vnodes open for
3806  * writing.
3807  *
3808  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3809  *
3810  * `rootrefs' specifies the base reference count for the root vnode
3811  * of this filesystem. The root vnode is considered busy if its
3812  * v_usecount exceeds this value. On a successful return, vflush(, td)
3813  * will call vrele() on the root vnode exactly rootrefs times.
3814  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3815  * be zero.
3816  */
3817 #ifdef DIAGNOSTIC
3818 static int busyprt = 0;		/* print out busy vnodes */
3819 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3820 #endif
3821 
3822 int
3823 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3824 {
3825 	struct vnode *vp, *mvp, *rootvp = NULL;
3826 	struct vattr vattr;
3827 	int busy = 0, error;
3828 
3829 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3830 	    rootrefs, flags);
3831 	if (rootrefs > 0) {
3832 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3833 		    ("vflush: bad args"));
3834 		/*
3835 		 * Get the filesystem root vnode. We can vput() it
3836 		 * immediately, since with rootrefs > 0, it won't go away.
3837 		 */
3838 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3839 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3840 			    __func__, error);
3841 			return (error);
3842 		}
3843 		vput(rootvp);
3844 	}
3845 loop:
3846 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3847 		vholdl(vp);
3848 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3849 		if (error) {
3850 			vdrop(vp);
3851 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3852 			goto loop;
3853 		}
3854 		/*
3855 		 * Skip over a vnodes marked VV_SYSTEM.
3856 		 */
3857 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3858 			VOP_UNLOCK(vp);
3859 			vdrop(vp);
3860 			continue;
3861 		}
3862 		/*
3863 		 * If WRITECLOSE is set, flush out unlinked but still open
3864 		 * files (even if open only for reading) and regular file
3865 		 * vnodes open for writing.
3866 		 */
3867 		if (flags & WRITECLOSE) {
3868 			if (vp->v_object != NULL) {
3869 				VM_OBJECT_WLOCK(vp->v_object);
3870 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3871 				VM_OBJECT_WUNLOCK(vp->v_object);
3872 			}
3873 			do {
3874 				error = VOP_FSYNC(vp, MNT_WAIT, td);
3875 			} while (error == ERELOOKUP);
3876 			if (error != 0) {
3877 				VOP_UNLOCK(vp);
3878 				vdrop(vp);
3879 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3880 				return (error);
3881 			}
3882 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3883 			VI_LOCK(vp);
3884 
3885 			if ((vp->v_type == VNON ||
3886 			    (error == 0 && vattr.va_nlink > 0)) &&
3887 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3888 				VOP_UNLOCK(vp);
3889 				vdropl(vp);
3890 				continue;
3891 			}
3892 		} else
3893 			VI_LOCK(vp);
3894 		/*
3895 		 * With v_usecount == 0, all we need to do is clear out the
3896 		 * vnode data structures and we are done.
3897 		 *
3898 		 * If FORCECLOSE is set, forcibly close the vnode.
3899 		 */
3900 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3901 			vgonel(vp);
3902 		} else {
3903 			busy++;
3904 #ifdef DIAGNOSTIC
3905 			if (busyprt)
3906 				vn_printf(vp, "vflush: busy vnode ");
3907 #endif
3908 		}
3909 		VOP_UNLOCK(vp);
3910 		vdropl(vp);
3911 	}
3912 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3913 		/*
3914 		 * If just the root vnode is busy, and if its refcount
3915 		 * is equal to `rootrefs', then go ahead and kill it.
3916 		 */
3917 		VI_LOCK(rootvp);
3918 		KASSERT(busy > 0, ("vflush: not busy"));
3919 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3920 		    ("vflush: usecount %d < rootrefs %d",
3921 		     rootvp->v_usecount, rootrefs));
3922 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3923 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3924 			vgone(rootvp);
3925 			VOP_UNLOCK(rootvp);
3926 			busy = 0;
3927 		} else
3928 			VI_UNLOCK(rootvp);
3929 	}
3930 	if (busy) {
3931 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3932 		    busy);
3933 		return (EBUSY);
3934 	}
3935 	for (; rootrefs > 0; rootrefs--)
3936 		vrele(rootvp);
3937 	return (0);
3938 }
3939 
3940 /*
3941  * Recycle an unused vnode to the front of the free list.
3942  */
3943 int
3944 vrecycle(struct vnode *vp)
3945 {
3946 	int recycled;
3947 
3948 	VI_LOCK(vp);
3949 	recycled = vrecyclel(vp);
3950 	VI_UNLOCK(vp);
3951 	return (recycled);
3952 }
3953 
3954 /*
3955  * vrecycle, with the vp interlock held.
3956  */
3957 int
3958 vrecyclel(struct vnode *vp)
3959 {
3960 	int recycled;
3961 
3962 	ASSERT_VOP_ELOCKED(vp, __func__);
3963 	ASSERT_VI_LOCKED(vp, __func__);
3964 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3965 	recycled = 0;
3966 	if (vp->v_usecount == 0) {
3967 		recycled = 1;
3968 		vgonel(vp);
3969 	}
3970 	return (recycled);
3971 }
3972 
3973 /*
3974  * Eliminate all activity associated with a vnode
3975  * in preparation for reuse.
3976  */
3977 void
3978 vgone(struct vnode *vp)
3979 {
3980 	VI_LOCK(vp);
3981 	vgonel(vp);
3982 	VI_UNLOCK(vp);
3983 }
3984 
3985 /*
3986  * Notify upper mounts about reclaimed or unlinked vnode.
3987  */
3988 void
3989 vfs_notify_upper(struct vnode *vp, enum vfs_notify_upper_type event)
3990 {
3991 	struct mount *mp;
3992 	struct mount_upper_node *ump;
3993 
3994 	mp = atomic_load_ptr(&vp->v_mount);
3995 	if (mp == NULL)
3996 		return;
3997 	if (TAILQ_EMPTY(&mp->mnt_notify))
3998 		return;
3999 
4000 	MNT_ILOCK(mp);
4001 	mp->mnt_upper_pending++;
4002 	KASSERT(mp->mnt_upper_pending > 0,
4003 	    ("%s: mnt_upper_pending %d", __func__, mp->mnt_upper_pending));
4004 	TAILQ_FOREACH(ump, &mp->mnt_notify, mnt_upper_link) {
4005 		MNT_IUNLOCK(mp);
4006 		switch (event) {
4007 		case VFS_NOTIFY_UPPER_RECLAIM:
4008 			VFS_RECLAIM_LOWERVP(ump->mp, vp);
4009 			break;
4010 		case VFS_NOTIFY_UPPER_UNLINK:
4011 			VFS_UNLINK_LOWERVP(ump->mp, vp);
4012 			break;
4013 		}
4014 		MNT_ILOCK(mp);
4015 	}
4016 	mp->mnt_upper_pending--;
4017 	if ((mp->mnt_kern_flag & MNTK_UPPER_WAITER) != 0 &&
4018 	    mp->mnt_upper_pending == 0) {
4019 		mp->mnt_kern_flag &= ~MNTK_UPPER_WAITER;
4020 		wakeup(&mp->mnt_uppers);
4021 	}
4022 	MNT_IUNLOCK(mp);
4023 }
4024 
4025 /*
4026  * vgone, with the vp interlock held.
4027  */
4028 static void
4029 vgonel(struct vnode *vp)
4030 {
4031 	struct thread *td;
4032 	struct mount *mp;
4033 	vm_object_t object;
4034 	bool active, doinginact, oweinact;
4035 
4036 	ASSERT_VOP_ELOCKED(vp, "vgonel");
4037 	ASSERT_VI_LOCKED(vp, "vgonel");
4038 	VNASSERT(vp->v_holdcnt, vp,
4039 	    ("vgonel: vp %p has no reference.", vp));
4040 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4041 	td = curthread;
4042 
4043 	/*
4044 	 * Don't vgonel if we're already doomed.
4045 	 */
4046 	if (VN_IS_DOOMED(vp)) {
4047 		VNPASS(vn_get_state(vp) == VSTATE_DESTROYING || \
4048 		    vn_get_state(vp) == VSTATE_DEAD, vp);
4049 		return;
4050 	}
4051 	/*
4052 	 * Paired with freevnode.
4053 	 */
4054 	vn_seqc_write_begin_locked(vp);
4055 	vunlazy_gone(vp);
4056 	vn_irflag_set_locked(vp, VIRF_DOOMED);
4057 	vn_set_state(vp, VSTATE_DESTROYING);
4058 
4059 	/*
4060 	 * Check to see if the vnode is in use.  If so, we have to
4061 	 * call VOP_CLOSE() and VOP_INACTIVE().
4062 	 *
4063 	 * It could be that VOP_INACTIVE() requested reclamation, in
4064 	 * which case we should avoid recursion, so check
4065 	 * VI_DOINGINACT.  This is not precise but good enough.
4066 	 */
4067 	active = vp->v_usecount > 0;
4068 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4069 	doinginact = (vp->v_iflag & VI_DOINGINACT) != 0;
4070 
4071 	/*
4072 	 * If we need to do inactive VI_OWEINACT will be set.
4073 	 */
4074 	if (vp->v_iflag & VI_DEFINACT) {
4075 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
4076 		vp->v_iflag &= ~VI_DEFINACT;
4077 		vdropl(vp);
4078 	} else {
4079 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
4080 		VI_UNLOCK(vp);
4081 	}
4082 	cache_purge_vgone(vp);
4083 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
4084 
4085 	/*
4086 	 * If purging an active vnode, it must be closed and
4087 	 * deactivated before being reclaimed.
4088 	 */
4089 	if (active)
4090 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
4091 	if (!doinginact) {
4092 		do {
4093 			if (oweinact || active) {
4094 				VI_LOCK(vp);
4095 				vinactivef(vp);
4096 				oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4097 				VI_UNLOCK(vp);
4098 			}
4099 		} while (oweinact);
4100 	}
4101 	if (vp->v_type == VSOCK)
4102 		vfs_unp_reclaim(vp);
4103 
4104 	/*
4105 	 * Clean out any buffers associated with the vnode.
4106 	 * If the flush fails, just toss the buffers.
4107 	 */
4108 	mp = NULL;
4109 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
4110 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
4111 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
4112 		while (vinvalbuf(vp, 0, 0, 0) != 0)
4113 			;
4114 	}
4115 
4116 	BO_LOCK(&vp->v_bufobj);
4117 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
4118 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
4119 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
4120 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
4121 	    ("vp %p bufobj not invalidated", vp));
4122 
4123 	/*
4124 	 * For VMIO bufobj, BO_DEAD is set later, or in
4125 	 * vm_object_terminate() after the object's page queue is
4126 	 * flushed.
4127 	 */
4128 	object = vp->v_bufobj.bo_object;
4129 	if (object == NULL)
4130 		vp->v_bufobj.bo_flag |= BO_DEAD;
4131 	BO_UNLOCK(&vp->v_bufobj);
4132 
4133 	/*
4134 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
4135 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
4136 	 * should not touch the object borrowed from the lower vnode
4137 	 * (the handle check).
4138 	 */
4139 	if (object != NULL && object->type == OBJT_VNODE &&
4140 	    object->handle == vp)
4141 		vnode_destroy_vobject(vp);
4142 
4143 	/*
4144 	 * Reclaim the vnode.
4145 	 */
4146 	if (VOP_RECLAIM(vp))
4147 		panic("vgone: cannot reclaim");
4148 	if (mp != NULL)
4149 		vn_finished_secondary_write(mp);
4150 	VNASSERT(vp->v_object == NULL, vp,
4151 	    ("vop_reclaim left v_object vp=%p", vp));
4152 	/*
4153 	 * Clear the advisory locks and wake up waiting threads.
4154 	 */
4155 	if (vp->v_lockf != NULL) {
4156 		(void)VOP_ADVLOCKPURGE(vp);
4157 		vp->v_lockf = NULL;
4158 	}
4159 	/*
4160 	 * Delete from old mount point vnode list.
4161 	 */
4162 	if (vp->v_mount == NULL) {
4163 		VI_LOCK(vp);
4164 	} else {
4165 		delmntque(vp);
4166 		ASSERT_VI_LOCKED(vp, "vgonel 2");
4167 	}
4168 	/*
4169 	 * Done with purge, reset to the standard lock and invalidate
4170 	 * the vnode.
4171 	 */
4172 	vp->v_vnlock = &vp->v_lock;
4173 	vp->v_op = &dead_vnodeops;
4174 	vp->v_type = VBAD;
4175 	vn_set_state(vp, VSTATE_DEAD);
4176 }
4177 
4178 /*
4179  * Print out a description of a vnode.
4180  */
4181 static const char *const vtypename[] = {
4182 	[VNON] = "VNON",
4183 	[VREG] = "VREG",
4184 	[VDIR] = "VDIR",
4185 	[VBLK] = "VBLK",
4186 	[VCHR] = "VCHR",
4187 	[VLNK] = "VLNK",
4188 	[VSOCK] = "VSOCK",
4189 	[VFIFO] = "VFIFO",
4190 	[VBAD] = "VBAD",
4191 	[VMARKER] = "VMARKER",
4192 };
4193 _Static_assert(nitems(vtypename) == VLASTTYPE + 1,
4194     "vnode type name not added to vtypename");
4195 
4196 static const char *const vstatename[] = {
4197 	[VSTATE_UNINITIALIZED] = "VSTATE_UNINITIALIZED",
4198 	[VSTATE_CONSTRUCTED] = "VSTATE_CONSTRUCTED",
4199 	[VSTATE_DESTROYING] = "VSTATE_DESTROYING",
4200 	[VSTATE_DEAD] = "VSTATE_DEAD",
4201 };
4202 _Static_assert(nitems(vstatename) == VLASTSTATE + 1,
4203     "vnode state name not added to vstatename");
4204 
4205 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0,
4206     "new hold count flag not added to vn_printf");
4207 
4208 void
4209 vn_printf(struct vnode *vp, const char *fmt, ...)
4210 {
4211 	va_list ap;
4212 	char buf[256], buf2[16];
4213 	u_long flags;
4214 	u_int holdcnt;
4215 	short irflag;
4216 
4217 	va_start(ap, fmt);
4218 	vprintf(fmt, ap);
4219 	va_end(ap);
4220 	printf("%p: ", (void *)vp);
4221 	printf("type %s state %s op %p\n", vtypename[vp->v_type],
4222 	    vstatename[vp->v_state], vp->v_op);
4223 	holdcnt = atomic_load_int(&vp->v_holdcnt);
4224 	printf("    usecount %d, writecount %d, refcount %d seqc users %d",
4225 	    vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS,
4226 	    vp->v_seqc_users);
4227 	switch (vp->v_type) {
4228 	case VDIR:
4229 		printf(" mountedhere %p\n", vp->v_mountedhere);
4230 		break;
4231 	case VCHR:
4232 		printf(" rdev %p\n", vp->v_rdev);
4233 		break;
4234 	case VSOCK:
4235 		printf(" socket %p\n", vp->v_unpcb);
4236 		break;
4237 	case VFIFO:
4238 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
4239 		break;
4240 	default:
4241 		printf("\n");
4242 		break;
4243 	}
4244 	buf[0] = '\0';
4245 	buf[1] = '\0';
4246 	if (holdcnt & VHOLD_NO_SMR)
4247 		strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf));
4248 	printf("    hold count flags (%s)\n", buf + 1);
4249 
4250 	buf[0] = '\0';
4251 	buf[1] = '\0';
4252 	irflag = vn_irflag_read(vp);
4253 	if (irflag & VIRF_DOOMED)
4254 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4255 	if (irflag & VIRF_PGREAD)
4256 		strlcat(buf, "|VIRF_PGREAD", sizeof(buf));
4257 	if (irflag & VIRF_MOUNTPOINT)
4258 		strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf));
4259 	if (irflag & VIRF_TEXT_REF)
4260 		strlcat(buf, "|VIRF_TEXT_REF", sizeof(buf));
4261 	flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT | VIRF_TEXT_REF);
4262 	if (flags != 0) {
4263 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4264 		strlcat(buf, buf2, sizeof(buf));
4265 	}
4266 	if (vp->v_vflag & VV_ROOT)
4267 		strlcat(buf, "|VV_ROOT", sizeof(buf));
4268 	if (vp->v_vflag & VV_ISTTY)
4269 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
4270 	if (vp->v_vflag & VV_NOSYNC)
4271 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4272 	if (vp->v_vflag & VV_ETERNALDEV)
4273 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4274 	if (vp->v_vflag & VV_CACHEDLABEL)
4275 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4276 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
4277 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4278 	if (vp->v_vflag & VV_COPYONWRITE)
4279 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4280 	if (vp->v_vflag & VV_SYSTEM)
4281 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4282 	if (vp->v_vflag & VV_PROCDEP)
4283 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4284 	if (vp->v_vflag & VV_DELETED)
4285 		strlcat(buf, "|VV_DELETED", sizeof(buf));
4286 	if (vp->v_vflag & VV_MD)
4287 		strlcat(buf, "|VV_MD", sizeof(buf));
4288 	if (vp->v_vflag & VV_FORCEINSMQ)
4289 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4290 	if (vp->v_vflag & VV_READLINK)
4291 		strlcat(buf, "|VV_READLINK", sizeof(buf));
4292 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4293 	    VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM |
4294 	    VV_PROCDEP | VV_DELETED | VV_MD | VV_FORCEINSMQ | VV_READLINK);
4295 	if (flags != 0) {
4296 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4297 		strlcat(buf, buf2, sizeof(buf));
4298 	}
4299 	if (vp->v_iflag & VI_MOUNT)
4300 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
4301 	if (vp->v_iflag & VI_DOINGINACT)
4302 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4303 	if (vp->v_iflag & VI_OWEINACT)
4304 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4305 	if (vp->v_iflag & VI_DEFINACT)
4306 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4307 	if (vp->v_iflag & VI_FOPENING)
4308 		strlcat(buf, "|VI_FOPENING", sizeof(buf));
4309 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOINGINACT |
4310 	    VI_OWEINACT | VI_DEFINACT | VI_FOPENING);
4311 	if (flags != 0) {
4312 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4313 		strlcat(buf, buf2, sizeof(buf));
4314 	}
4315 	if (vp->v_mflag & VMP_LAZYLIST)
4316 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4317 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
4318 	if (flags != 0) {
4319 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4320 		strlcat(buf, buf2, sizeof(buf));
4321 	}
4322 	printf("    flags (%s)", buf + 1);
4323 	if (mtx_owned(VI_MTX(vp)))
4324 		printf(" VI_LOCKed");
4325 	printf("\n");
4326 	if (vp->v_object != NULL)
4327 		printf("    v_object %p ref %d pages %d "
4328 		    "cleanbuf %d dirtybuf %d\n",
4329 		    vp->v_object, vp->v_object->ref_count,
4330 		    vp->v_object->resident_page_count,
4331 		    vp->v_bufobj.bo_clean.bv_cnt,
4332 		    vp->v_bufobj.bo_dirty.bv_cnt);
4333 	printf("    ");
4334 	lockmgr_printinfo(vp->v_vnlock);
4335 	if (vp->v_data != NULL)
4336 		VOP_PRINT(vp);
4337 }
4338 
4339 #ifdef DDB
4340 /*
4341  * List all of the locked vnodes in the system.
4342  * Called when debugging the kernel.
4343  */
4344 DB_SHOW_COMMAND_FLAGS(lockedvnods, lockedvnodes, DB_CMD_MEMSAFE)
4345 {
4346 	struct mount *mp;
4347 	struct vnode *vp;
4348 
4349 	/*
4350 	 * Note: because this is DDB, we can't obey the locking semantics
4351 	 * for these structures, which means we could catch an inconsistent
4352 	 * state and dereference a nasty pointer.  Not much to be done
4353 	 * about that.
4354 	 */
4355 	db_printf("Locked vnodes\n");
4356 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4357 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4358 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4359 				vn_printf(vp, "vnode ");
4360 		}
4361 	}
4362 }
4363 
4364 /*
4365  * Show details about the given vnode.
4366  */
4367 DB_SHOW_COMMAND(vnode, db_show_vnode)
4368 {
4369 	struct vnode *vp;
4370 
4371 	if (!have_addr)
4372 		return;
4373 	vp = (struct vnode *)addr;
4374 	vn_printf(vp, "vnode ");
4375 }
4376 
4377 /*
4378  * Show details about the given mount point.
4379  */
4380 DB_SHOW_COMMAND(mount, db_show_mount)
4381 {
4382 	struct mount *mp;
4383 	struct vfsopt *opt;
4384 	struct statfs *sp;
4385 	struct vnode *vp;
4386 	char buf[512];
4387 	uint64_t mflags;
4388 	u_int flags;
4389 
4390 	if (!have_addr) {
4391 		/* No address given, print short info about all mount points. */
4392 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4393 			db_printf("%p %s on %s (%s)\n", mp,
4394 			    mp->mnt_stat.f_mntfromname,
4395 			    mp->mnt_stat.f_mntonname,
4396 			    mp->mnt_stat.f_fstypename);
4397 			if (db_pager_quit)
4398 				break;
4399 		}
4400 		db_printf("\nMore info: show mount <addr>\n");
4401 		return;
4402 	}
4403 
4404 	mp = (struct mount *)addr;
4405 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4406 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4407 
4408 	buf[0] = '\0';
4409 	mflags = mp->mnt_flag;
4410 #define	MNT_FLAG(flag)	do {						\
4411 	if (mflags & (flag)) {						\
4412 		if (buf[0] != '\0')					\
4413 			strlcat(buf, ", ", sizeof(buf));		\
4414 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4415 		mflags &= ~(flag);					\
4416 	}								\
4417 } while (0)
4418 	MNT_FLAG(MNT_RDONLY);
4419 	MNT_FLAG(MNT_SYNCHRONOUS);
4420 	MNT_FLAG(MNT_NOEXEC);
4421 	MNT_FLAG(MNT_NOSUID);
4422 	MNT_FLAG(MNT_NFS4ACLS);
4423 	MNT_FLAG(MNT_UNION);
4424 	MNT_FLAG(MNT_ASYNC);
4425 	MNT_FLAG(MNT_SUIDDIR);
4426 	MNT_FLAG(MNT_SOFTDEP);
4427 	MNT_FLAG(MNT_NOSYMFOLLOW);
4428 	MNT_FLAG(MNT_GJOURNAL);
4429 	MNT_FLAG(MNT_MULTILABEL);
4430 	MNT_FLAG(MNT_ACLS);
4431 	MNT_FLAG(MNT_NOATIME);
4432 	MNT_FLAG(MNT_NOCLUSTERR);
4433 	MNT_FLAG(MNT_NOCLUSTERW);
4434 	MNT_FLAG(MNT_SUJ);
4435 	MNT_FLAG(MNT_EXRDONLY);
4436 	MNT_FLAG(MNT_EXPORTED);
4437 	MNT_FLAG(MNT_DEFEXPORTED);
4438 	MNT_FLAG(MNT_EXPORTANON);
4439 	MNT_FLAG(MNT_EXKERB);
4440 	MNT_FLAG(MNT_EXPUBLIC);
4441 	MNT_FLAG(MNT_LOCAL);
4442 	MNT_FLAG(MNT_QUOTA);
4443 	MNT_FLAG(MNT_ROOTFS);
4444 	MNT_FLAG(MNT_USER);
4445 	MNT_FLAG(MNT_IGNORE);
4446 	MNT_FLAG(MNT_UPDATE);
4447 	MNT_FLAG(MNT_DELEXPORT);
4448 	MNT_FLAG(MNT_RELOAD);
4449 	MNT_FLAG(MNT_FORCE);
4450 	MNT_FLAG(MNT_SNAPSHOT);
4451 	MNT_FLAG(MNT_BYFSID);
4452 #undef MNT_FLAG
4453 	if (mflags != 0) {
4454 		if (buf[0] != '\0')
4455 			strlcat(buf, ", ", sizeof(buf));
4456 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4457 		    "0x%016jx", mflags);
4458 	}
4459 	db_printf("    mnt_flag = %s\n", buf);
4460 
4461 	buf[0] = '\0';
4462 	flags = mp->mnt_kern_flag;
4463 #define	MNT_KERN_FLAG(flag)	do {					\
4464 	if (flags & (flag)) {						\
4465 		if (buf[0] != '\0')					\
4466 			strlcat(buf, ", ", sizeof(buf));		\
4467 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4468 		flags &= ~(flag);					\
4469 	}								\
4470 } while (0)
4471 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4472 	MNT_KERN_FLAG(MNTK_ASYNC);
4473 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4474 	MNT_KERN_FLAG(MNTK_NOMSYNC);
4475 	MNT_KERN_FLAG(MNTK_DRAINING);
4476 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4477 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4478 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4479 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4480 	MNT_KERN_FLAG(MNTK_RECURSE);
4481 	MNT_KERN_FLAG(MNTK_UPPER_WAITER);
4482 	MNT_KERN_FLAG(MNTK_UNLOCKED_INSMNTQUE);
4483 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4484 	MNT_KERN_FLAG(MNTK_VMSETSIZE_BUG);
4485 	MNT_KERN_FLAG(MNTK_FPLOOKUP);
4486 	MNT_KERN_FLAG(MNTK_TASKQUEUE_WAITER);
4487 	MNT_KERN_FLAG(MNTK_NOASYNC);
4488 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4489 	MNT_KERN_FLAG(MNTK_MWAIT);
4490 	MNT_KERN_FLAG(MNTK_SUSPEND);
4491 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4492 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4493 	MNT_KERN_FLAG(MNTK_NULL_NOCACHE);
4494 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4495 #undef MNT_KERN_FLAG
4496 	if (flags != 0) {
4497 		if (buf[0] != '\0')
4498 			strlcat(buf, ", ", sizeof(buf));
4499 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4500 		    "0x%08x", flags);
4501 	}
4502 	db_printf("    mnt_kern_flag = %s\n", buf);
4503 
4504 	db_printf("    mnt_opt = ");
4505 	opt = TAILQ_FIRST(mp->mnt_opt);
4506 	if (opt != NULL) {
4507 		db_printf("%s", opt->name);
4508 		opt = TAILQ_NEXT(opt, link);
4509 		while (opt != NULL) {
4510 			db_printf(", %s", opt->name);
4511 			opt = TAILQ_NEXT(opt, link);
4512 		}
4513 	}
4514 	db_printf("\n");
4515 
4516 	sp = &mp->mnt_stat;
4517 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4518 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4519 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4520 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4521 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4522 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4523 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4524 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4525 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4526 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4527 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4528 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4529 
4530 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4531 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4532 	if (jailed(mp->mnt_cred))
4533 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4534 	db_printf(" }\n");
4535 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4536 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4537 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4538 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4539 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4540 	    mp->mnt_lazyvnodelistsize);
4541 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4542 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4543 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4544 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4545 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4546 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4547 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4548 	db_printf("    mnt_secondary_accwrites = %d\n",
4549 	    mp->mnt_secondary_accwrites);
4550 	db_printf("    mnt_gjprovider = %s\n",
4551 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4552 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4553 
4554 	db_printf("\n\nList of active vnodes\n");
4555 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4556 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4557 			vn_printf(vp, "vnode ");
4558 			if (db_pager_quit)
4559 				break;
4560 		}
4561 	}
4562 	db_printf("\n\nList of inactive vnodes\n");
4563 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4564 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4565 			vn_printf(vp, "vnode ");
4566 			if (db_pager_quit)
4567 				break;
4568 		}
4569 	}
4570 }
4571 #endif	/* DDB */
4572 
4573 /*
4574  * Fill in a struct xvfsconf based on a struct vfsconf.
4575  */
4576 static int
4577 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4578 {
4579 	struct xvfsconf xvfsp;
4580 
4581 	bzero(&xvfsp, sizeof(xvfsp));
4582 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4583 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4584 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4585 	xvfsp.vfc_flags = vfsp->vfc_flags;
4586 	/*
4587 	 * These are unused in userland, we keep them
4588 	 * to not break binary compatibility.
4589 	 */
4590 	xvfsp.vfc_vfsops = NULL;
4591 	xvfsp.vfc_next = NULL;
4592 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4593 }
4594 
4595 #ifdef COMPAT_FREEBSD32
4596 struct xvfsconf32 {
4597 	uint32_t	vfc_vfsops;
4598 	char		vfc_name[MFSNAMELEN];
4599 	int32_t		vfc_typenum;
4600 	int32_t		vfc_refcount;
4601 	int32_t		vfc_flags;
4602 	uint32_t	vfc_next;
4603 };
4604 
4605 static int
4606 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4607 {
4608 	struct xvfsconf32 xvfsp;
4609 
4610 	bzero(&xvfsp, sizeof(xvfsp));
4611 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4612 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4613 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4614 	xvfsp.vfc_flags = vfsp->vfc_flags;
4615 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4616 }
4617 #endif
4618 
4619 /*
4620  * Top level filesystem related information gathering.
4621  */
4622 static int
4623 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4624 {
4625 	struct vfsconf *vfsp;
4626 	int error;
4627 
4628 	error = 0;
4629 	vfsconf_slock();
4630 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4631 #ifdef COMPAT_FREEBSD32
4632 		if (req->flags & SCTL_MASK32)
4633 			error = vfsconf2x32(req, vfsp);
4634 		else
4635 #endif
4636 			error = vfsconf2x(req, vfsp);
4637 		if (error)
4638 			break;
4639 	}
4640 	vfsconf_sunlock();
4641 	return (error);
4642 }
4643 
4644 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4645     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4646     "S,xvfsconf", "List of all configured filesystems");
4647 
4648 #ifndef BURN_BRIDGES
4649 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4650 
4651 static int
4652 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4653 {
4654 	int *name = (int *)arg1 - 1;	/* XXX */
4655 	u_int namelen = arg2 + 1;	/* XXX */
4656 	struct vfsconf *vfsp;
4657 
4658 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4659 	    "please rebuild world\n");
4660 
4661 #if 1 || defined(COMPAT_PRELITE2)
4662 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4663 	if (namelen == 1)
4664 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4665 #endif
4666 
4667 	switch (name[1]) {
4668 	case VFS_MAXTYPENUM:
4669 		if (namelen != 2)
4670 			return (ENOTDIR);
4671 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4672 	case VFS_CONF:
4673 		if (namelen != 3)
4674 			return (ENOTDIR);	/* overloaded */
4675 		vfsconf_slock();
4676 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4677 			if (vfsp->vfc_typenum == name[2])
4678 				break;
4679 		}
4680 		vfsconf_sunlock();
4681 		if (vfsp == NULL)
4682 			return (EOPNOTSUPP);
4683 #ifdef COMPAT_FREEBSD32
4684 		if (req->flags & SCTL_MASK32)
4685 			return (vfsconf2x32(req, vfsp));
4686 		else
4687 #endif
4688 			return (vfsconf2x(req, vfsp));
4689 	}
4690 	return (EOPNOTSUPP);
4691 }
4692 
4693 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4694     CTLFLAG_MPSAFE, vfs_sysctl,
4695     "Generic filesystem");
4696 
4697 #if 1 || defined(COMPAT_PRELITE2)
4698 
4699 static int
4700 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4701 {
4702 	int error;
4703 	struct vfsconf *vfsp;
4704 	struct ovfsconf ovfs;
4705 
4706 	vfsconf_slock();
4707 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4708 		bzero(&ovfs, sizeof(ovfs));
4709 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4710 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4711 		ovfs.vfc_index = vfsp->vfc_typenum;
4712 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4713 		ovfs.vfc_flags = vfsp->vfc_flags;
4714 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4715 		if (error != 0) {
4716 			vfsconf_sunlock();
4717 			return (error);
4718 		}
4719 	}
4720 	vfsconf_sunlock();
4721 	return (0);
4722 }
4723 
4724 #endif /* 1 || COMPAT_PRELITE2 */
4725 #endif /* !BURN_BRIDGES */
4726 
4727 static void
4728 unmount_or_warn(struct mount *mp)
4729 {
4730 	int error;
4731 
4732 	error = dounmount(mp, MNT_FORCE, curthread);
4733 	if (error != 0) {
4734 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4735 		if (error == EBUSY)
4736 			printf("BUSY)\n");
4737 		else
4738 			printf("%d)\n", error);
4739 	}
4740 }
4741 
4742 /*
4743  * Unmount all filesystems. The list is traversed in reverse order
4744  * of mounting to avoid dependencies.
4745  */
4746 void
4747 vfs_unmountall(void)
4748 {
4749 	struct mount *mp, *tmp;
4750 
4751 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4752 
4753 	/*
4754 	 * Since this only runs when rebooting, it is not interlocked.
4755 	 */
4756 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4757 		vfs_ref(mp);
4758 
4759 		/*
4760 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4761 		 * unmount of the latter.
4762 		 */
4763 		if (mp == rootdevmp)
4764 			continue;
4765 
4766 		unmount_or_warn(mp);
4767 	}
4768 
4769 	if (rootdevmp != NULL)
4770 		unmount_or_warn(rootdevmp);
4771 }
4772 
4773 static void
4774 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4775 {
4776 
4777 	ASSERT_VI_LOCKED(vp, __func__);
4778 	VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
4779 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
4780 		vdropl(vp);
4781 		return;
4782 	}
4783 	if (vn_lock(vp, lkflags) == 0) {
4784 		VI_LOCK(vp);
4785 		vinactive(vp);
4786 		VOP_UNLOCK(vp);
4787 		vdropl(vp);
4788 		return;
4789 	}
4790 	vdefer_inactive_unlocked(vp);
4791 }
4792 
4793 static int
4794 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4795 {
4796 
4797 	return (vp->v_iflag & VI_DEFINACT);
4798 }
4799 
4800 static void __noinline
4801 vfs_periodic_inactive(struct mount *mp, int flags)
4802 {
4803 	struct vnode *vp, *mvp;
4804 	int lkflags;
4805 
4806 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4807 	if (flags != MNT_WAIT)
4808 		lkflags |= LK_NOWAIT;
4809 
4810 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4811 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
4812 			VI_UNLOCK(vp);
4813 			continue;
4814 		}
4815 		vp->v_iflag &= ~VI_DEFINACT;
4816 		vfs_deferred_inactive(vp, lkflags);
4817 	}
4818 }
4819 
4820 static inline bool
4821 vfs_want_msync(struct vnode *vp)
4822 {
4823 	struct vm_object *obj;
4824 
4825 	/*
4826 	 * This test may be performed without any locks held.
4827 	 * We rely on vm_object's type stability.
4828 	 */
4829 	if (vp->v_vflag & VV_NOSYNC)
4830 		return (false);
4831 	obj = vp->v_object;
4832 	return (obj != NULL && vm_object_mightbedirty(obj));
4833 }
4834 
4835 static int
4836 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4837 {
4838 
4839 	if (vp->v_vflag & VV_NOSYNC)
4840 		return (false);
4841 	if (vp->v_iflag & VI_DEFINACT)
4842 		return (true);
4843 	return (vfs_want_msync(vp));
4844 }
4845 
4846 static void __noinline
4847 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4848 {
4849 	struct vnode *vp, *mvp;
4850 	struct vm_object *obj;
4851 	int lkflags, objflags;
4852 	bool seen_defer;
4853 
4854 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4855 	if (flags != MNT_WAIT) {
4856 		lkflags |= LK_NOWAIT;
4857 		objflags = OBJPC_NOSYNC;
4858 	} else {
4859 		objflags = OBJPC_SYNC;
4860 	}
4861 
4862 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4863 		seen_defer = false;
4864 		if (vp->v_iflag & VI_DEFINACT) {
4865 			vp->v_iflag &= ~VI_DEFINACT;
4866 			seen_defer = true;
4867 		}
4868 		if (!vfs_want_msync(vp)) {
4869 			if (seen_defer)
4870 				vfs_deferred_inactive(vp, lkflags);
4871 			else
4872 				VI_UNLOCK(vp);
4873 			continue;
4874 		}
4875 		if (vget(vp, lkflags) == 0) {
4876 			obj = vp->v_object;
4877 			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4878 				VM_OBJECT_WLOCK(obj);
4879 				vm_object_page_clean(obj, 0, 0, objflags);
4880 				VM_OBJECT_WUNLOCK(obj);
4881 			}
4882 			vput(vp);
4883 			if (seen_defer)
4884 				vdrop(vp);
4885 		} else {
4886 			if (seen_defer)
4887 				vdefer_inactive_unlocked(vp);
4888 		}
4889 	}
4890 }
4891 
4892 void
4893 vfs_periodic(struct mount *mp, int flags)
4894 {
4895 
4896 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4897 
4898 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4899 		vfs_periodic_inactive(mp, flags);
4900 	else
4901 		vfs_periodic_msync_inactive(mp, flags);
4902 }
4903 
4904 static void
4905 destroy_vpollinfo_free(struct vpollinfo *vi)
4906 {
4907 
4908 	knlist_destroy(&vi->vpi_selinfo.si_note);
4909 	mtx_destroy(&vi->vpi_lock);
4910 	free(vi, M_VNODEPOLL);
4911 }
4912 
4913 static void
4914 destroy_vpollinfo(struct vpollinfo *vi)
4915 {
4916 
4917 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4918 	seldrain(&vi->vpi_selinfo);
4919 	destroy_vpollinfo_free(vi);
4920 }
4921 
4922 /*
4923  * Initialize per-vnode helper structure to hold poll-related state.
4924  */
4925 void
4926 v_addpollinfo(struct vnode *vp)
4927 {
4928 	struct vpollinfo *vi;
4929 
4930 	if (vp->v_pollinfo != NULL)
4931 		return;
4932 	vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO);
4933 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4934 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4935 	    vfs_knlunlock, vfs_knl_assert_lock);
4936 	VI_LOCK(vp);
4937 	if (vp->v_pollinfo != NULL) {
4938 		VI_UNLOCK(vp);
4939 		destroy_vpollinfo_free(vi);
4940 		return;
4941 	}
4942 	vp->v_pollinfo = vi;
4943 	VI_UNLOCK(vp);
4944 }
4945 
4946 /*
4947  * Record a process's interest in events which might happen to
4948  * a vnode.  Because poll uses the historic select-style interface
4949  * internally, this routine serves as both the ``check for any
4950  * pending events'' and the ``record my interest in future events''
4951  * functions.  (These are done together, while the lock is held,
4952  * to avoid race conditions.)
4953  */
4954 int
4955 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4956 {
4957 
4958 	v_addpollinfo(vp);
4959 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4960 	if (vp->v_pollinfo->vpi_revents & events) {
4961 		/*
4962 		 * This leaves events we are not interested
4963 		 * in available for the other process which
4964 		 * which presumably had requested them
4965 		 * (otherwise they would never have been
4966 		 * recorded).
4967 		 */
4968 		events &= vp->v_pollinfo->vpi_revents;
4969 		vp->v_pollinfo->vpi_revents &= ~events;
4970 
4971 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4972 		return (events);
4973 	}
4974 	vp->v_pollinfo->vpi_events |= events;
4975 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4976 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4977 	return (0);
4978 }
4979 
4980 /*
4981  * Routine to create and manage a filesystem syncer vnode.
4982  */
4983 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4984 static int	sync_fsync(struct  vop_fsync_args *);
4985 static int	sync_inactive(struct  vop_inactive_args *);
4986 static int	sync_reclaim(struct  vop_reclaim_args *);
4987 
4988 static struct vop_vector sync_vnodeops = {
4989 	.vop_bypass =	VOP_EOPNOTSUPP,
4990 	.vop_close =	sync_close,
4991 	.vop_fsync =	sync_fsync,
4992 	.vop_getwritemount = vop_stdgetwritemount,
4993 	.vop_inactive =	sync_inactive,
4994 	.vop_need_inactive = vop_stdneed_inactive,
4995 	.vop_reclaim =	sync_reclaim,
4996 	.vop_lock1 =	vop_stdlock,
4997 	.vop_unlock =	vop_stdunlock,
4998 	.vop_islocked =	vop_stdislocked,
4999 	.vop_fplookup_vexec = VOP_EAGAIN,
5000 	.vop_fplookup_symlink = VOP_EAGAIN,
5001 };
5002 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
5003 
5004 /*
5005  * Create a new filesystem syncer vnode for the specified mount point.
5006  */
5007 void
5008 vfs_allocate_syncvnode(struct mount *mp)
5009 {
5010 	struct vnode *vp;
5011 	struct bufobj *bo;
5012 	static long start, incr, next;
5013 	int error;
5014 
5015 	/* Allocate a new vnode */
5016 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
5017 	if (error != 0)
5018 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
5019 	vp->v_type = VNON;
5020 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5021 	vp->v_vflag |= VV_FORCEINSMQ;
5022 	error = insmntque1(vp, mp);
5023 	if (error != 0)
5024 		panic("vfs_allocate_syncvnode: insmntque() failed");
5025 	vp->v_vflag &= ~VV_FORCEINSMQ;
5026 	vn_set_state(vp, VSTATE_CONSTRUCTED);
5027 	VOP_UNLOCK(vp);
5028 	/*
5029 	 * Place the vnode onto the syncer worklist. We attempt to
5030 	 * scatter them about on the list so that they will go off
5031 	 * at evenly distributed times even if all the filesystems
5032 	 * are mounted at once.
5033 	 */
5034 	next += incr;
5035 	if (next == 0 || next > syncer_maxdelay) {
5036 		start /= 2;
5037 		incr /= 2;
5038 		if (start == 0) {
5039 			start = syncer_maxdelay / 2;
5040 			incr = syncer_maxdelay;
5041 		}
5042 		next = start;
5043 	}
5044 	bo = &vp->v_bufobj;
5045 	BO_LOCK(bo);
5046 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
5047 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
5048 	mtx_lock(&sync_mtx);
5049 	sync_vnode_count++;
5050 	if (mp->mnt_syncer == NULL) {
5051 		mp->mnt_syncer = vp;
5052 		vp = NULL;
5053 	}
5054 	mtx_unlock(&sync_mtx);
5055 	BO_UNLOCK(bo);
5056 	if (vp != NULL) {
5057 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5058 		vgone(vp);
5059 		vput(vp);
5060 	}
5061 }
5062 
5063 void
5064 vfs_deallocate_syncvnode(struct mount *mp)
5065 {
5066 	struct vnode *vp;
5067 
5068 	mtx_lock(&sync_mtx);
5069 	vp = mp->mnt_syncer;
5070 	if (vp != NULL)
5071 		mp->mnt_syncer = NULL;
5072 	mtx_unlock(&sync_mtx);
5073 	if (vp != NULL)
5074 		vrele(vp);
5075 }
5076 
5077 /*
5078  * Do a lazy sync of the filesystem.
5079  */
5080 static int
5081 sync_fsync(struct vop_fsync_args *ap)
5082 {
5083 	struct vnode *syncvp = ap->a_vp;
5084 	struct mount *mp = syncvp->v_mount;
5085 	int error, save;
5086 	struct bufobj *bo;
5087 
5088 	/*
5089 	 * We only need to do something if this is a lazy evaluation.
5090 	 */
5091 	if (ap->a_waitfor != MNT_LAZY)
5092 		return (0);
5093 
5094 	/*
5095 	 * Move ourselves to the back of the sync list.
5096 	 */
5097 	bo = &syncvp->v_bufobj;
5098 	BO_LOCK(bo);
5099 	vn_syncer_add_to_worklist(bo, syncdelay);
5100 	BO_UNLOCK(bo);
5101 
5102 	/*
5103 	 * Walk the list of vnodes pushing all that are dirty and
5104 	 * not already on the sync list.
5105 	 */
5106 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
5107 		return (0);
5108 	VOP_UNLOCK(syncvp);
5109 	save = curthread_pflags_set(TDP_SYNCIO);
5110 	/*
5111 	 * The filesystem at hand may be idle with free vnodes stored in the
5112 	 * batch.  Return them instead of letting them stay there indefinitely.
5113 	 */
5114 	vfs_periodic(mp, MNT_NOWAIT);
5115 	error = VFS_SYNC(mp, MNT_LAZY);
5116 	curthread_pflags_restore(save);
5117 	vn_lock(syncvp, LK_EXCLUSIVE | LK_RETRY);
5118 	vfs_unbusy(mp);
5119 	return (error);
5120 }
5121 
5122 /*
5123  * The syncer vnode is no referenced.
5124  */
5125 static int
5126 sync_inactive(struct vop_inactive_args *ap)
5127 {
5128 
5129 	vgone(ap->a_vp);
5130 	return (0);
5131 }
5132 
5133 /*
5134  * The syncer vnode is no longer needed and is being decommissioned.
5135  *
5136  * Modifications to the worklist must be protected by sync_mtx.
5137  */
5138 static int
5139 sync_reclaim(struct vop_reclaim_args *ap)
5140 {
5141 	struct vnode *vp = ap->a_vp;
5142 	struct bufobj *bo;
5143 
5144 	bo = &vp->v_bufobj;
5145 	BO_LOCK(bo);
5146 	mtx_lock(&sync_mtx);
5147 	if (vp->v_mount->mnt_syncer == vp)
5148 		vp->v_mount->mnt_syncer = NULL;
5149 	if (bo->bo_flag & BO_ONWORKLST) {
5150 		LIST_REMOVE(bo, bo_synclist);
5151 		syncer_worklist_len--;
5152 		sync_vnode_count--;
5153 		bo->bo_flag &= ~BO_ONWORKLST;
5154 	}
5155 	mtx_unlock(&sync_mtx);
5156 	BO_UNLOCK(bo);
5157 
5158 	return (0);
5159 }
5160 
5161 int
5162 vn_need_pageq_flush(struct vnode *vp)
5163 {
5164 	struct vm_object *obj;
5165 
5166 	obj = vp->v_object;
5167 	return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5168 	    vm_object_mightbedirty(obj));
5169 }
5170 
5171 /*
5172  * Check if vnode represents a disk device
5173  */
5174 bool
5175 vn_isdisk_error(struct vnode *vp, int *errp)
5176 {
5177 	int error;
5178 
5179 	if (vp->v_type != VCHR) {
5180 		error = ENOTBLK;
5181 		goto out;
5182 	}
5183 	error = 0;
5184 	dev_lock();
5185 	if (vp->v_rdev == NULL)
5186 		error = ENXIO;
5187 	else if (vp->v_rdev->si_devsw == NULL)
5188 		error = ENXIO;
5189 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5190 		error = ENOTBLK;
5191 	dev_unlock();
5192 out:
5193 	*errp = error;
5194 	return (error == 0);
5195 }
5196 
5197 bool
5198 vn_isdisk(struct vnode *vp)
5199 {
5200 	int error;
5201 
5202 	return (vn_isdisk_error(vp, &error));
5203 }
5204 
5205 /*
5206  * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
5207  * the comment above cache_fplookup for details.
5208  */
5209 int
5210 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred)
5211 {
5212 	int error;
5213 
5214 	VFS_SMR_ASSERT_ENTERED();
5215 
5216 	/* Check the owner. */
5217 	if (cred->cr_uid == file_uid) {
5218 		if (file_mode & S_IXUSR)
5219 			return (0);
5220 		goto out_error;
5221 	}
5222 
5223 	/* Otherwise, check the groups (first match) */
5224 	if (groupmember(file_gid, cred)) {
5225 		if (file_mode & S_IXGRP)
5226 			return (0);
5227 		goto out_error;
5228 	}
5229 
5230 	/* Otherwise, check everyone else. */
5231 	if (file_mode & S_IXOTH)
5232 		return (0);
5233 out_error:
5234 	/*
5235 	 * Permission check failed, but it is possible denial will get overwritten
5236 	 * (e.g., when root is traversing through a 700 directory owned by someone
5237 	 * else).
5238 	 *
5239 	 * vaccess() calls priv_check_cred which in turn can descent into MAC
5240 	 * modules overriding this result. It's quite unclear what semantics
5241 	 * are allowed for them to operate, thus for safety we don't call them
5242 	 * from within the SMR section. This also means if any such modules
5243 	 * are present, we have to let the regular lookup decide.
5244 	 */
5245 	error = priv_check_cred_vfs_lookup_nomac(cred);
5246 	switch (error) {
5247 	case 0:
5248 		return (0);
5249 	case EAGAIN:
5250 		/*
5251 		 * MAC modules present.
5252 		 */
5253 		return (EAGAIN);
5254 	case EPERM:
5255 		return (EACCES);
5256 	default:
5257 		return (error);
5258 	}
5259 }
5260 
5261 /*
5262  * Common filesystem object access control check routine.  Accepts a
5263  * vnode's type, "mode", uid and gid, requested access mode, and credentials.
5264  * Returns 0 on success, or an errno on failure.
5265  */
5266 int
5267 vaccess(__enum_uint8(vtype) type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5268     accmode_t accmode, struct ucred *cred)
5269 {
5270 	accmode_t dac_granted;
5271 	accmode_t priv_granted;
5272 
5273 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5274 	    ("invalid bit in accmode"));
5275 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5276 	    ("VAPPEND without VWRITE"));
5277 
5278 	/*
5279 	 * Look for a normal, non-privileged way to access the file/directory
5280 	 * as requested.  If it exists, go with that.
5281 	 */
5282 
5283 	dac_granted = 0;
5284 
5285 	/* Check the owner. */
5286 	if (cred->cr_uid == file_uid) {
5287 		dac_granted |= VADMIN;
5288 		if (file_mode & S_IXUSR)
5289 			dac_granted |= VEXEC;
5290 		if (file_mode & S_IRUSR)
5291 			dac_granted |= VREAD;
5292 		if (file_mode & S_IWUSR)
5293 			dac_granted |= (VWRITE | VAPPEND);
5294 
5295 		if ((accmode & dac_granted) == accmode)
5296 			return (0);
5297 
5298 		goto privcheck;
5299 	}
5300 
5301 	/* Otherwise, check the groups (first match) */
5302 	if (groupmember(file_gid, cred)) {
5303 		if (file_mode & S_IXGRP)
5304 			dac_granted |= VEXEC;
5305 		if (file_mode & S_IRGRP)
5306 			dac_granted |= VREAD;
5307 		if (file_mode & S_IWGRP)
5308 			dac_granted |= (VWRITE | VAPPEND);
5309 
5310 		if ((accmode & dac_granted) == accmode)
5311 			return (0);
5312 
5313 		goto privcheck;
5314 	}
5315 
5316 	/* Otherwise, check everyone else. */
5317 	if (file_mode & S_IXOTH)
5318 		dac_granted |= VEXEC;
5319 	if (file_mode & S_IROTH)
5320 		dac_granted |= VREAD;
5321 	if (file_mode & S_IWOTH)
5322 		dac_granted |= (VWRITE | VAPPEND);
5323 	if ((accmode & dac_granted) == accmode)
5324 		return (0);
5325 
5326 privcheck:
5327 	/*
5328 	 * Build a privilege mask to determine if the set of privileges
5329 	 * satisfies the requirements when combined with the granted mask
5330 	 * from above.  For each privilege, if the privilege is required,
5331 	 * bitwise or the request type onto the priv_granted mask.
5332 	 */
5333 	priv_granted = 0;
5334 
5335 	if (type == VDIR) {
5336 		/*
5337 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5338 		 * requests, instead of PRIV_VFS_EXEC.
5339 		 */
5340 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5341 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5342 			priv_granted |= VEXEC;
5343 	} else {
5344 		/*
5345 		 * Ensure that at least one execute bit is on. Otherwise,
5346 		 * a privileged user will always succeed, and we don't want
5347 		 * this to happen unless the file really is executable.
5348 		 */
5349 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5350 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5351 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5352 			priv_granted |= VEXEC;
5353 	}
5354 
5355 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5356 	    !priv_check_cred(cred, PRIV_VFS_READ))
5357 		priv_granted |= VREAD;
5358 
5359 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5360 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5361 		priv_granted |= (VWRITE | VAPPEND);
5362 
5363 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5364 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5365 		priv_granted |= VADMIN;
5366 
5367 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5368 		return (0);
5369 	}
5370 
5371 	return ((accmode & VADMIN) ? EPERM : EACCES);
5372 }
5373 
5374 /*
5375  * Credential check based on process requesting service, and per-attribute
5376  * permissions.
5377  */
5378 int
5379 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5380     struct thread *td, accmode_t accmode)
5381 {
5382 
5383 	/*
5384 	 * Kernel-invoked always succeeds.
5385 	 */
5386 	if (cred == NOCRED)
5387 		return (0);
5388 
5389 	/*
5390 	 * Do not allow privileged processes in jail to directly manipulate
5391 	 * system attributes.
5392 	 */
5393 	switch (attrnamespace) {
5394 	case EXTATTR_NAMESPACE_SYSTEM:
5395 		/* Potentially should be: return (EPERM); */
5396 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5397 	case EXTATTR_NAMESPACE_USER:
5398 		return (VOP_ACCESS(vp, accmode, cred, td));
5399 	default:
5400 		return (EPERM);
5401 	}
5402 }
5403 
5404 #ifdef DEBUG_VFS_LOCKS
5405 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5406 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5407     "Drop into debugger on lock violation");
5408 
5409 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5410 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5411     0, "Check for interlock across VOPs");
5412 
5413 int vfs_badlock_print = 1;	/* Print lock violations. */
5414 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5415     0, "Print lock violations");
5416 
5417 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5418 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5419     0, "Print vnode details on lock violations");
5420 
5421 #ifdef KDB
5422 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5423 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5424     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5425 #endif
5426 
5427 static void
5428 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5429 {
5430 
5431 #ifdef KDB
5432 	if (vfs_badlock_backtrace)
5433 		kdb_backtrace();
5434 #endif
5435 	if (vfs_badlock_vnode)
5436 		vn_printf(vp, "vnode ");
5437 	if (vfs_badlock_print)
5438 		printf("%s: %p %s\n", str, (void *)vp, msg);
5439 	if (vfs_badlock_ddb)
5440 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5441 }
5442 
5443 void
5444 assert_vi_locked(struct vnode *vp, const char *str)
5445 {
5446 
5447 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5448 		vfs_badlock("interlock is not locked but should be", str, vp);
5449 }
5450 
5451 void
5452 assert_vi_unlocked(struct vnode *vp, const char *str)
5453 {
5454 
5455 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5456 		vfs_badlock("interlock is locked but should not be", str, vp);
5457 }
5458 
5459 void
5460 assert_vop_locked(struct vnode *vp, const char *str)
5461 {
5462 	if (KERNEL_PANICKED() || vp == NULL)
5463 		return;
5464 
5465 #ifdef WITNESS
5466 	if ((vp->v_irflag & VIRF_CROSSMP) == 0 &&
5467 	    witness_is_owned(&vp->v_vnlock->lock_object) == -1)
5468 #else
5469 	int locked = VOP_ISLOCKED(vp);
5470 	if (locked == 0 || locked == LK_EXCLOTHER)
5471 #endif
5472 		vfs_badlock("is not locked but should be", str, vp);
5473 }
5474 
5475 void
5476 assert_vop_unlocked(struct vnode *vp, const char *str)
5477 {
5478 	if (KERNEL_PANICKED() || vp == NULL)
5479 		return;
5480 
5481 #ifdef WITNESS
5482 	if ((vp->v_irflag & VIRF_CROSSMP) == 0 &&
5483 	    witness_is_owned(&vp->v_vnlock->lock_object) == 1)
5484 #else
5485 	if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5486 #endif
5487 		vfs_badlock("is locked but should not be", str, vp);
5488 }
5489 
5490 void
5491 assert_vop_elocked(struct vnode *vp, const char *str)
5492 {
5493 	if (KERNEL_PANICKED() || vp == NULL)
5494 		return;
5495 
5496 	if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5497 		vfs_badlock("is not exclusive locked but should be", str, vp);
5498 }
5499 #endif /* DEBUG_VFS_LOCKS */
5500 
5501 void
5502 vop_rename_fail(struct vop_rename_args *ap)
5503 {
5504 
5505 	if (ap->a_tvp != NULL)
5506 		vput(ap->a_tvp);
5507 	if (ap->a_tdvp == ap->a_tvp)
5508 		vrele(ap->a_tdvp);
5509 	else
5510 		vput(ap->a_tdvp);
5511 	vrele(ap->a_fdvp);
5512 	vrele(ap->a_fvp);
5513 }
5514 
5515 void
5516 vop_rename_pre(void *ap)
5517 {
5518 	struct vop_rename_args *a = ap;
5519 
5520 #ifdef DEBUG_VFS_LOCKS
5521 	if (a->a_tvp)
5522 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5523 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5524 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5525 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5526 
5527 	/* Check the source (from). */
5528 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5529 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5530 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5531 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5532 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5533 
5534 	/* Check the target. */
5535 	if (a->a_tvp)
5536 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5537 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5538 #endif
5539 	/*
5540 	 * It may be tempting to add vn_seqc_write_begin/end calls here and
5541 	 * in vop_rename_post but that's not going to work out since some
5542 	 * filesystems relookup vnodes mid-rename. This is probably a bug.
5543 	 *
5544 	 * For now filesystems are expected to do the relevant calls after they
5545 	 * decide what vnodes to operate on.
5546 	 */
5547 	if (a->a_tdvp != a->a_fdvp)
5548 		vhold(a->a_fdvp);
5549 	if (a->a_tvp != a->a_fvp)
5550 		vhold(a->a_fvp);
5551 	vhold(a->a_tdvp);
5552 	if (a->a_tvp)
5553 		vhold(a->a_tvp);
5554 }
5555 
5556 #ifdef DEBUG_VFS_LOCKS
5557 void
5558 vop_fplookup_vexec_debugpre(void *ap __unused)
5559 {
5560 
5561 	VFS_SMR_ASSERT_ENTERED();
5562 }
5563 
5564 void
5565 vop_fplookup_vexec_debugpost(void *ap, int rc)
5566 {
5567 	struct vop_fplookup_vexec_args *a;
5568 	struct vnode *vp;
5569 
5570 	a = ap;
5571 	vp = a->a_vp;
5572 
5573 	VFS_SMR_ASSERT_ENTERED();
5574 	if (rc == EOPNOTSUPP)
5575 		VNPASS(VN_IS_DOOMED(vp), vp);
5576 }
5577 
5578 void
5579 vop_fplookup_symlink_debugpre(void *ap __unused)
5580 {
5581 
5582 	VFS_SMR_ASSERT_ENTERED();
5583 }
5584 
5585 void
5586 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused)
5587 {
5588 
5589 	VFS_SMR_ASSERT_ENTERED();
5590 }
5591 
5592 static void
5593 vop_fsync_debugprepost(struct vnode *vp, const char *name)
5594 {
5595 	if (vp->v_type == VCHR)
5596 		;
5597 	else if (MNT_EXTENDED_SHARED(vp->v_mount))
5598 		ASSERT_VOP_LOCKED(vp, name);
5599 	else
5600 		ASSERT_VOP_ELOCKED(vp, name);
5601 }
5602 
5603 void
5604 vop_fsync_debugpre(void *a)
5605 {
5606 	struct vop_fsync_args *ap;
5607 
5608 	ap = a;
5609 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5610 }
5611 
5612 void
5613 vop_fsync_debugpost(void *a, int rc __unused)
5614 {
5615 	struct vop_fsync_args *ap;
5616 
5617 	ap = a;
5618 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5619 }
5620 
5621 void
5622 vop_fdatasync_debugpre(void *a)
5623 {
5624 	struct vop_fdatasync_args *ap;
5625 
5626 	ap = a;
5627 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5628 }
5629 
5630 void
5631 vop_fdatasync_debugpost(void *a, int rc __unused)
5632 {
5633 	struct vop_fdatasync_args *ap;
5634 
5635 	ap = a;
5636 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5637 }
5638 
5639 void
5640 vop_strategy_debugpre(void *ap)
5641 {
5642 	struct vop_strategy_args *a;
5643 	struct buf *bp;
5644 
5645 	a = ap;
5646 	bp = a->a_bp;
5647 
5648 	/*
5649 	 * Cluster ops lock their component buffers but not the IO container.
5650 	 */
5651 	if ((bp->b_flags & B_CLUSTER) != 0)
5652 		return;
5653 
5654 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5655 		if (vfs_badlock_print)
5656 			printf(
5657 			    "VOP_STRATEGY: bp is not locked but should be\n");
5658 		if (vfs_badlock_ddb)
5659 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5660 	}
5661 }
5662 
5663 void
5664 vop_lock_debugpre(void *ap)
5665 {
5666 	struct vop_lock1_args *a = ap;
5667 
5668 	if ((a->a_flags & LK_INTERLOCK) == 0)
5669 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5670 	else
5671 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5672 }
5673 
5674 void
5675 vop_lock_debugpost(void *ap, int rc)
5676 {
5677 	struct vop_lock1_args *a = ap;
5678 
5679 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5680 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5681 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5682 }
5683 
5684 void
5685 vop_unlock_debugpre(void *ap)
5686 {
5687 	struct vop_unlock_args *a = ap;
5688 	struct vnode *vp = a->a_vp;
5689 
5690 	VNPASS(vn_get_state(vp) != VSTATE_UNINITIALIZED, vp);
5691 	ASSERT_VOP_LOCKED(vp, "VOP_UNLOCK");
5692 }
5693 
5694 void
5695 vop_need_inactive_debugpre(void *ap)
5696 {
5697 	struct vop_need_inactive_args *a = ap;
5698 
5699 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5700 }
5701 
5702 void
5703 vop_need_inactive_debugpost(void *ap, int rc)
5704 {
5705 	struct vop_need_inactive_args *a = ap;
5706 
5707 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5708 }
5709 #endif
5710 
5711 void
5712 vop_create_pre(void *ap)
5713 {
5714 	struct vop_create_args *a;
5715 	struct vnode *dvp;
5716 
5717 	a = ap;
5718 	dvp = a->a_dvp;
5719 	vn_seqc_write_begin(dvp);
5720 }
5721 
5722 void
5723 vop_create_post(void *ap, int rc)
5724 {
5725 	struct vop_create_args *a;
5726 	struct vnode *dvp;
5727 
5728 	a = ap;
5729 	dvp = a->a_dvp;
5730 	vn_seqc_write_end(dvp);
5731 	if (!rc)
5732 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5733 }
5734 
5735 void
5736 vop_whiteout_pre(void *ap)
5737 {
5738 	struct vop_whiteout_args *a;
5739 	struct vnode *dvp;
5740 
5741 	a = ap;
5742 	dvp = a->a_dvp;
5743 	vn_seqc_write_begin(dvp);
5744 }
5745 
5746 void
5747 vop_whiteout_post(void *ap, int rc)
5748 {
5749 	struct vop_whiteout_args *a;
5750 	struct vnode *dvp;
5751 
5752 	a = ap;
5753 	dvp = a->a_dvp;
5754 	vn_seqc_write_end(dvp);
5755 }
5756 
5757 void
5758 vop_deleteextattr_pre(void *ap)
5759 {
5760 	struct vop_deleteextattr_args *a;
5761 	struct vnode *vp;
5762 
5763 	a = ap;
5764 	vp = a->a_vp;
5765 	vn_seqc_write_begin(vp);
5766 }
5767 
5768 void
5769 vop_deleteextattr_post(void *ap, int rc)
5770 {
5771 	struct vop_deleteextattr_args *a;
5772 	struct vnode *vp;
5773 
5774 	a = ap;
5775 	vp = a->a_vp;
5776 	vn_seqc_write_end(vp);
5777 	if (!rc)
5778 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5779 }
5780 
5781 void
5782 vop_link_pre(void *ap)
5783 {
5784 	struct vop_link_args *a;
5785 	struct vnode *vp, *tdvp;
5786 
5787 	a = ap;
5788 	vp = a->a_vp;
5789 	tdvp = a->a_tdvp;
5790 	vn_seqc_write_begin(vp);
5791 	vn_seqc_write_begin(tdvp);
5792 }
5793 
5794 void
5795 vop_link_post(void *ap, int rc)
5796 {
5797 	struct vop_link_args *a;
5798 	struct vnode *vp, *tdvp;
5799 
5800 	a = ap;
5801 	vp = a->a_vp;
5802 	tdvp = a->a_tdvp;
5803 	vn_seqc_write_end(vp);
5804 	vn_seqc_write_end(tdvp);
5805 	if (!rc) {
5806 		VFS_KNOTE_LOCKED(vp, NOTE_LINK);
5807 		VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE);
5808 	}
5809 }
5810 
5811 void
5812 vop_mkdir_pre(void *ap)
5813 {
5814 	struct vop_mkdir_args *a;
5815 	struct vnode *dvp;
5816 
5817 	a = ap;
5818 	dvp = a->a_dvp;
5819 	vn_seqc_write_begin(dvp);
5820 }
5821 
5822 void
5823 vop_mkdir_post(void *ap, int rc)
5824 {
5825 	struct vop_mkdir_args *a;
5826 	struct vnode *dvp;
5827 
5828 	a = ap;
5829 	dvp = a->a_dvp;
5830 	vn_seqc_write_end(dvp);
5831 	if (!rc)
5832 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5833 }
5834 
5835 #ifdef DEBUG_VFS_LOCKS
5836 void
5837 vop_mkdir_debugpost(void *ap, int rc)
5838 {
5839 	struct vop_mkdir_args *a;
5840 
5841 	a = ap;
5842 	if (!rc)
5843 		cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp);
5844 }
5845 #endif
5846 
5847 void
5848 vop_mknod_pre(void *ap)
5849 {
5850 	struct vop_mknod_args *a;
5851 	struct vnode *dvp;
5852 
5853 	a = ap;
5854 	dvp = a->a_dvp;
5855 	vn_seqc_write_begin(dvp);
5856 }
5857 
5858 void
5859 vop_mknod_post(void *ap, int rc)
5860 {
5861 	struct vop_mknod_args *a;
5862 	struct vnode *dvp;
5863 
5864 	a = ap;
5865 	dvp = a->a_dvp;
5866 	vn_seqc_write_end(dvp);
5867 	if (!rc)
5868 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5869 }
5870 
5871 void
5872 vop_reclaim_post(void *ap, int rc)
5873 {
5874 	struct vop_reclaim_args *a;
5875 	struct vnode *vp;
5876 
5877 	a = ap;
5878 	vp = a->a_vp;
5879 	ASSERT_VOP_IN_SEQC(vp);
5880 	if (!rc)
5881 		VFS_KNOTE_LOCKED(vp, NOTE_REVOKE);
5882 }
5883 
5884 void
5885 vop_remove_pre(void *ap)
5886 {
5887 	struct vop_remove_args *a;
5888 	struct vnode *dvp, *vp;
5889 
5890 	a = ap;
5891 	dvp = a->a_dvp;
5892 	vp = a->a_vp;
5893 	vn_seqc_write_begin(dvp);
5894 	vn_seqc_write_begin(vp);
5895 }
5896 
5897 void
5898 vop_remove_post(void *ap, int rc)
5899 {
5900 	struct vop_remove_args *a;
5901 	struct vnode *dvp, *vp;
5902 
5903 	a = ap;
5904 	dvp = a->a_dvp;
5905 	vp = a->a_vp;
5906 	vn_seqc_write_end(dvp);
5907 	vn_seqc_write_end(vp);
5908 	if (!rc) {
5909 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5910 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5911 	}
5912 }
5913 
5914 void
5915 vop_rename_post(void *ap, int rc)
5916 {
5917 	struct vop_rename_args *a = ap;
5918 	long hint;
5919 
5920 	if (!rc) {
5921 		hint = NOTE_WRITE;
5922 		if (a->a_fdvp == a->a_tdvp) {
5923 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5924 				hint |= NOTE_LINK;
5925 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5926 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5927 		} else {
5928 			hint |= NOTE_EXTEND;
5929 			if (a->a_fvp->v_type == VDIR)
5930 				hint |= NOTE_LINK;
5931 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5932 
5933 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5934 			    a->a_tvp->v_type == VDIR)
5935 				hint &= ~NOTE_LINK;
5936 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5937 		}
5938 
5939 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5940 		if (a->a_tvp)
5941 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5942 	}
5943 	if (a->a_tdvp != a->a_fdvp)
5944 		vdrop(a->a_fdvp);
5945 	if (a->a_tvp != a->a_fvp)
5946 		vdrop(a->a_fvp);
5947 	vdrop(a->a_tdvp);
5948 	if (a->a_tvp)
5949 		vdrop(a->a_tvp);
5950 }
5951 
5952 void
5953 vop_rmdir_pre(void *ap)
5954 {
5955 	struct vop_rmdir_args *a;
5956 	struct vnode *dvp, *vp;
5957 
5958 	a = ap;
5959 	dvp = a->a_dvp;
5960 	vp = a->a_vp;
5961 	vn_seqc_write_begin(dvp);
5962 	vn_seqc_write_begin(vp);
5963 }
5964 
5965 void
5966 vop_rmdir_post(void *ap, int rc)
5967 {
5968 	struct vop_rmdir_args *a;
5969 	struct vnode *dvp, *vp;
5970 
5971 	a = ap;
5972 	dvp = a->a_dvp;
5973 	vp = a->a_vp;
5974 	vn_seqc_write_end(dvp);
5975 	vn_seqc_write_end(vp);
5976 	if (!rc) {
5977 		vp->v_vflag |= VV_UNLINKED;
5978 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5979 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5980 	}
5981 }
5982 
5983 void
5984 vop_setattr_pre(void *ap)
5985 {
5986 	struct vop_setattr_args *a;
5987 	struct vnode *vp;
5988 
5989 	a = ap;
5990 	vp = a->a_vp;
5991 	vn_seqc_write_begin(vp);
5992 }
5993 
5994 void
5995 vop_setattr_post(void *ap, int rc)
5996 {
5997 	struct vop_setattr_args *a;
5998 	struct vnode *vp;
5999 
6000 	a = ap;
6001 	vp = a->a_vp;
6002 	vn_seqc_write_end(vp);
6003 	if (!rc)
6004 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
6005 }
6006 
6007 void
6008 vop_setacl_pre(void *ap)
6009 {
6010 	struct vop_setacl_args *a;
6011 	struct vnode *vp;
6012 
6013 	a = ap;
6014 	vp = a->a_vp;
6015 	vn_seqc_write_begin(vp);
6016 }
6017 
6018 void
6019 vop_setacl_post(void *ap, int rc __unused)
6020 {
6021 	struct vop_setacl_args *a;
6022 	struct vnode *vp;
6023 
6024 	a = ap;
6025 	vp = a->a_vp;
6026 	vn_seqc_write_end(vp);
6027 }
6028 
6029 void
6030 vop_setextattr_pre(void *ap)
6031 {
6032 	struct vop_setextattr_args *a;
6033 	struct vnode *vp;
6034 
6035 	a = ap;
6036 	vp = a->a_vp;
6037 	vn_seqc_write_begin(vp);
6038 }
6039 
6040 void
6041 vop_setextattr_post(void *ap, int rc)
6042 {
6043 	struct vop_setextattr_args *a;
6044 	struct vnode *vp;
6045 
6046 	a = ap;
6047 	vp = a->a_vp;
6048 	vn_seqc_write_end(vp);
6049 	if (!rc)
6050 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
6051 }
6052 
6053 void
6054 vop_symlink_pre(void *ap)
6055 {
6056 	struct vop_symlink_args *a;
6057 	struct vnode *dvp;
6058 
6059 	a = ap;
6060 	dvp = a->a_dvp;
6061 	vn_seqc_write_begin(dvp);
6062 }
6063 
6064 void
6065 vop_symlink_post(void *ap, int rc)
6066 {
6067 	struct vop_symlink_args *a;
6068 	struct vnode *dvp;
6069 
6070 	a = ap;
6071 	dvp = a->a_dvp;
6072 	vn_seqc_write_end(dvp);
6073 	if (!rc)
6074 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6075 }
6076 
6077 void
6078 vop_open_post(void *ap, int rc)
6079 {
6080 	struct vop_open_args *a = ap;
6081 
6082 	if (!rc)
6083 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
6084 }
6085 
6086 void
6087 vop_close_post(void *ap, int rc)
6088 {
6089 	struct vop_close_args *a = ap;
6090 
6091 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
6092 	    !VN_IS_DOOMED(a->a_vp))) {
6093 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
6094 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
6095 	}
6096 }
6097 
6098 void
6099 vop_read_post(void *ap, int rc)
6100 {
6101 	struct vop_read_args *a = ap;
6102 
6103 	if (!rc)
6104 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6105 }
6106 
6107 void
6108 vop_read_pgcache_post(void *ap, int rc)
6109 {
6110 	struct vop_read_pgcache_args *a = ap;
6111 
6112 	if (!rc)
6113 		VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ);
6114 }
6115 
6116 void
6117 vop_readdir_post(void *ap, int rc)
6118 {
6119 	struct vop_readdir_args *a = ap;
6120 
6121 	if (!rc)
6122 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6123 }
6124 
6125 static struct knlist fs_knlist;
6126 
6127 static void
6128 vfs_event_init(void *arg)
6129 {
6130 	knlist_init_mtx(&fs_knlist, NULL);
6131 }
6132 /* XXX - correct order? */
6133 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
6134 
6135 void
6136 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
6137 {
6138 
6139 	KNOTE_UNLOCKED(&fs_knlist, event);
6140 }
6141 
6142 static int	filt_fsattach(struct knote *kn);
6143 static void	filt_fsdetach(struct knote *kn);
6144 static int	filt_fsevent(struct knote *kn, long hint);
6145 
6146 struct filterops fs_filtops = {
6147 	.f_isfd = 0,
6148 	.f_attach = filt_fsattach,
6149 	.f_detach = filt_fsdetach,
6150 	.f_event = filt_fsevent
6151 };
6152 
6153 static int
6154 filt_fsattach(struct knote *kn)
6155 {
6156 
6157 	kn->kn_flags |= EV_CLEAR;
6158 	knlist_add(&fs_knlist, kn, 0);
6159 	return (0);
6160 }
6161 
6162 static void
6163 filt_fsdetach(struct knote *kn)
6164 {
6165 
6166 	knlist_remove(&fs_knlist, kn, 0);
6167 }
6168 
6169 static int
6170 filt_fsevent(struct knote *kn, long hint)
6171 {
6172 
6173 	kn->kn_fflags |= kn->kn_sfflags & hint;
6174 
6175 	return (kn->kn_fflags != 0);
6176 }
6177 
6178 static int
6179 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
6180 {
6181 	struct vfsidctl vc;
6182 	int error;
6183 	struct mount *mp;
6184 
6185 	error = SYSCTL_IN(req, &vc, sizeof(vc));
6186 	if (error)
6187 		return (error);
6188 	if (vc.vc_vers != VFS_CTL_VERS1)
6189 		return (EINVAL);
6190 	mp = vfs_getvfs(&vc.vc_fsid);
6191 	if (mp == NULL)
6192 		return (ENOENT);
6193 	/* ensure that a specific sysctl goes to the right filesystem. */
6194 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
6195 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
6196 		vfs_rel(mp);
6197 		return (EINVAL);
6198 	}
6199 	VCTLTOREQ(&vc, req);
6200 	error = VFS_SYSCTL(mp, vc.vc_op, req);
6201 	vfs_rel(mp);
6202 	return (error);
6203 }
6204 
6205 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
6206     NULL, 0, sysctl_vfs_ctl, "",
6207     "Sysctl by fsid");
6208 
6209 /*
6210  * Function to initialize a va_filerev field sensibly.
6211  * XXX: Wouldn't a random number make a lot more sense ??
6212  */
6213 u_quad_t
6214 init_va_filerev(void)
6215 {
6216 	struct bintime bt;
6217 
6218 	getbinuptime(&bt);
6219 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
6220 }
6221 
6222 static int	filt_vfsread(struct knote *kn, long hint);
6223 static int	filt_vfswrite(struct knote *kn, long hint);
6224 static int	filt_vfsvnode(struct knote *kn, long hint);
6225 static void	filt_vfsdetach(struct knote *kn);
6226 static struct filterops vfsread_filtops = {
6227 	.f_isfd = 1,
6228 	.f_detach = filt_vfsdetach,
6229 	.f_event = filt_vfsread
6230 };
6231 static struct filterops vfswrite_filtops = {
6232 	.f_isfd = 1,
6233 	.f_detach = filt_vfsdetach,
6234 	.f_event = filt_vfswrite
6235 };
6236 static struct filterops vfsvnode_filtops = {
6237 	.f_isfd = 1,
6238 	.f_detach = filt_vfsdetach,
6239 	.f_event = filt_vfsvnode
6240 };
6241 
6242 static void
6243 vfs_knllock(void *arg)
6244 {
6245 	struct vnode *vp = arg;
6246 
6247 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
6248 }
6249 
6250 static void
6251 vfs_knlunlock(void *arg)
6252 {
6253 	struct vnode *vp = arg;
6254 
6255 	VOP_UNLOCK(vp);
6256 }
6257 
6258 static void
6259 vfs_knl_assert_lock(void *arg, int what)
6260 {
6261 #ifdef DEBUG_VFS_LOCKS
6262 	struct vnode *vp = arg;
6263 
6264 	if (what == LA_LOCKED)
6265 		ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
6266 	else
6267 		ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
6268 #endif
6269 }
6270 
6271 int
6272 vfs_kqfilter(struct vop_kqfilter_args *ap)
6273 {
6274 	struct vnode *vp = ap->a_vp;
6275 	struct knote *kn = ap->a_kn;
6276 	struct knlist *knl;
6277 
6278 	KASSERT(vp->v_type != VFIFO || (kn->kn_filter != EVFILT_READ &&
6279 	    kn->kn_filter != EVFILT_WRITE),
6280 	    ("READ/WRITE filter on a FIFO leaked through"));
6281 	switch (kn->kn_filter) {
6282 	case EVFILT_READ:
6283 		kn->kn_fop = &vfsread_filtops;
6284 		break;
6285 	case EVFILT_WRITE:
6286 		kn->kn_fop = &vfswrite_filtops;
6287 		break;
6288 	case EVFILT_VNODE:
6289 		kn->kn_fop = &vfsvnode_filtops;
6290 		break;
6291 	default:
6292 		return (EINVAL);
6293 	}
6294 
6295 	kn->kn_hook = (caddr_t)vp;
6296 
6297 	v_addpollinfo(vp);
6298 	if (vp->v_pollinfo == NULL)
6299 		return (ENOMEM);
6300 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
6301 	vhold(vp);
6302 	knlist_add(knl, kn, 0);
6303 
6304 	return (0);
6305 }
6306 
6307 /*
6308  * Detach knote from vnode
6309  */
6310 static void
6311 filt_vfsdetach(struct knote *kn)
6312 {
6313 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6314 
6315 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
6316 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
6317 	vdrop(vp);
6318 }
6319 
6320 /*ARGSUSED*/
6321 static int
6322 filt_vfsread(struct knote *kn, long hint)
6323 {
6324 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6325 	off_t size;
6326 	int res;
6327 
6328 	/*
6329 	 * filesystem is gone, so set the EOF flag and schedule
6330 	 * the knote for deletion.
6331 	 */
6332 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6333 		VI_LOCK(vp);
6334 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6335 		VI_UNLOCK(vp);
6336 		return (1);
6337 	}
6338 
6339 	if (vn_getsize_locked(vp, &size, curthread->td_ucred) != 0)
6340 		return (0);
6341 
6342 	VI_LOCK(vp);
6343 	kn->kn_data = size - kn->kn_fp->f_offset;
6344 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
6345 	VI_UNLOCK(vp);
6346 	return (res);
6347 }
6348 
6349 /*ARGSUSED*/
6350 static int
6351 filt_vfswrite(struct knote *kn, long hint)
6352 {
6353 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6354 
6355 	VI_LOCK(vp);
6356 
6357 	/*
6358 	 * filesystem is gone, so set the EOF flag and schedule
6359 	 * the knote for deletion.
6360 	 */
6361 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
6362 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6363 
6364 	kn->kn_data = 0;
6365 	VI_UNLOCK(vp);
6366 	return (1);
6367 }
6368 
6369 static int
6370 filt_vfsvnode(struct knote *kn, long hint)
6371 {
6372 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6373 	int res;
6374 
6375 	VI_LOCK(vp);
6376 	if (kn->kn_sfflags & hint)
6377 		kn->kn_fflags |= hint;
6378 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6379 		kn->kn_flags |= EV_EOF;
6380 		VI_UNLOCK(vp);
6381 		return (1);
6382 	}
6383 	res = (kn->kn_fflags != 0);
6384 	VI_UNLOCK(vp);
6385 	return (res);
6386 }
6387 
6388 int
6389 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6390 {
6391 	int error;
6392 
6393 	if (dp->d_reclen > ap->a_uio->uio_resid)
6394 		return (ENAMETOOLONG);
6395 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
6396 	if (error) {
6397 		if (ap->a_ncookies != NULL) {
6398 			if (ap->a_cookies != NULL)
6399 				free(ap->a_cookies, M_TEMP);
6400 			ap->a_cookies = NULL;
6401 			*ap->a_ncookies = 0;
6402 		}
6403 		return (error);
6404 	}
6405 	if (ap->a_ncookies == NULL)
6406 		return (0);
6407 
6408 	KASSERT(ap->a_cookies,
6409 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6410 
6411 	*ap->a_cookies = realloc(*ap->a_cookies,
6412 	    (*ap->a_ncookies + 1) * sizeof(uint64_t), M_TEMP, M_WAITOK | M_ZERO);
6413 	(*ap->a_cookies)[*ap->a_ncookies] = off;
6414 	*ap->a_ncookies += 1;
6415 	return (0);
6416 }
6417 
6418 /*
6419  * The purpose of this routine is to remove granularity from accmode_t,
6420  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6421  * VADMIN and VAPPEND.
6422  *
6423  * If it returns 0, the caller is supposed to continue with the usual
6424  * access checks using 'accmode' as modified by this routine.  If it
6425  * returns nonzero value, the caller is supposed to return that value
6426  * as errno.
6427  *
6428  * Note that after this routine runs, accmode may be zero.
6429  */
6430 int
6431 vfs_unixify_accmode(accmode_t *accmode)
6432 {
6433 	/*
6434 	 * There is no way to specify explicit "deny" rule using
6435 	 * file mode or POSIX.1e ACLs.
6436 	 */
6437 	if (*accmode & VEXPLICIT_DENY) {
6438 		*accmode = 0;
6439 		return (0);
6440 	}
6441 
6442 	/*
6443 	 * None of these can be translated into usual access bits.
6444 	 * Also, the common case for NFSv4 ACLs is to not contain
6445 	 * either of these bits. Caller should check for VWRITE
6446 	 * on the containing directory instead.
6447 	 */
6448 	if (*accmode & (VDELETE_CHILD | VDELETE))
6449 		return (EPERM);
6450 
6451 	if (*accmode & VADMIN_PERMS) {
6452 		*accmode &= ~VADMIN_PERMS;
6453 		*accmode |= VADMIN;
6454 	}
6455 
6456 	/*
6457 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6458 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6459 	 */
6460 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6461 
6462 	return (0);
6463 }
6464 
6465 /*
6466  * Clear out a doomed vnode (if any) and replace it with a new one as long
6467  * as the fs is not being unmounted. Return the root vnode to the caller.
6468  */
6469 static int __noinline
6470 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6471 {
6472 	struct vnode *vp;
6473 	int error;
6474 
6475 restart:
6476 	if (mp->mnt_rootvnode != NULL) {
6477 		MNT_ILOCK(mp);
6478 		vp = mp->mnt_rootvnode;
6479 		if (vp != NULL) {
6480 			if (!VN_IS_DOOMED(vp)) {
6481 				vrefact(vp);
6482 				MNT_IUNLOCK(mp);
6483 				error = vn_lock(vp, flags);
6484 				if (error == 0) {
6485 					*vpp = vp;
6486 					return (0);
6487 				}
6488 				vrele(vp);
6489 				goto restart;
6490 			}
6491 			/*
6492 			 * Clear the old one.
6493 			 */
6494 			mp->mnt_rootvnode = NULL;
6495 		}
6496 		MNT_IUNLOCK(mp);
6497 		if (vp != NULL) {
6498 			vfs_op_barrier_wait(mp);
6499 			vrele(vp);
6500 		}
6501 	}
6502 	error = VFS_CACHEDROOT(mp, flags, vpp);
6503 	if (error != 0)
6504 		return (error);
6505 	if (mp->mnt_vfs_ops == 0) {
6506 		MNT_ILOCK(mp);
6507 		if (mp->mnt_vfs_ops != 0) {
6508 			MNT_IUNLOCK(mp);
6509 			return (0);
6510 		}
6511 		if (mp->mnt_rootvnode == NULL) {
6512 			vrefact(*vpp);
6513 			mp->mnt_rootvnode = *vpp;
6514 		} else {
6515 			if (mp->mnt_rootvnode != *vpp) {
6516 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6517 					panic("%s: mismatch between vnode returned "
6518 					    " by VFS_CACHEDROOT and the one cached "
6519 					    " (%p != %p)",
6520 					    __func__, *vpp, mp->mnt_rootvnode);
6521 				}
6522 			}
6523 		}
6524 		MNT_IUNLOCK(mp);
6525 	}
6526 	return (0);
6527 }
6528 
6529 int
6530 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6531 {
6532 	struct mount_pcpu *mpcpu;
6533 	struct vnode *vp;
6534 	int error;
6535 
6536 	if (!vfs_op_thread_enter(mp, mpcpu))
6537 		return (vfs_cache_root_fallback(mp, flags, vpp));
6538 	vp = atomic_load_ptr(&mp->mnt_rootvnode);
6539 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6540 		vfs_op_thread_exit(mp, mpcpu);
6541 		return (vfs_cache_root_fallback(mp, flags, vpp));
6542 	}
6543 	vrefact(vp);
6544 	vfs_op_thread_exit(mp, mpcpu);
6545 	error = vn_lock(vp, flags);
6546 	if (error != 0) {
6547 		vrele(vp);
6548 		return (vfs_cache_root_fallback(mp, flags, vpp));
6549 	}
6550 	*vpp = vp;
6551 	return (0);
6552 }
6553 
6554 struct vnode *
6555 vfs_cache_root_clear(struct mount *mp)
6556 {
6557 	struct vnode *vp;
6558 
6559 	/*
6560 	 * ops > 0 guarantees there is nobody who can see this vnode
6561 	 */
6562 	MPASS(mp->mnt_vfs_ops > 0);
6563 	vp = mp->mnt_rootvnode;
6564 	if (vp != NULL)
6565 		vn_seqc_write_begin(vp);
6566 	mp->mnt_rootvnode = NULL;
6567 	return (vp);
6568 }
6569 
6570 void
6571 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6572 {
6573 
6574 	MPASS(mp->mnt_vfs_ops > 0);
6575 	vrefact(vp);
6576 	mp->mnt_rootvnode = vp;
6577 }
6578 
6579 /*
6580  * These are helper functions for filesystems to traverse all
6581  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6582  *
6583  * This interface replaces MNT_VNODE_FOREACH.
6584  */
6585 
6586 struct vnode *
6587 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6588 {
6589 	struct vnode *vp;
6590 
6591 	maybe_yield();
6592 	MNT_ILOCK(mp);
6593 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6594 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6595 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6596 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6597 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6598 			continue;
6599 		VI_LOCK(vp);
6600 		if (VN_IS_DOOMED(vp)) {
6601 			VI_UNLOCK(vp);
6602 			continue;
6603 		}
6604 		break;
6605 	}
6606 	if (vp == NULL) {
6607 		__mnt_vnode_markerfree_all(mvp, mp);
6608 		/* MNT_IUNLOCK(mp); -- done in above function */
6609 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6610 		return (NULL);
6611 	}
6612 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6613 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6614 	MNT_IUNLOCK(mp);
6615 	return (vp);
6616 }
6617 
6618 struct vnode *
6619 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6620 {
6621 	struct vnode *vp;
6622 
6623 	*mvp = vn_alloc_marker(mp);
6624 	MNT_ILOCK(mp);
6625 	MNT_REF(mp);
6626 
6627 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6628 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6629 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6630 			continue;
6631 		VI_LOCK(vp);
6632 		if (VN_IS_DOOMED(vp)) {
6633 			VI_UNLOCK(vp);
6634 			continue;
6635 		}
6636 		break;
6637 	}
6638 	if (vp == NULL) {
6639 		MNT_REL(mp);
6640 		MNT_IUNLOCK(mp);
6641 		vn_free_marker(*mvp);
6642 		*mvp = NULL;
6643 		return (NULL);
6644 	}
6645 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6646 	MNT_IUNLOCK(mp);
6647 	return (vp);
6648 }
6649 
6650 void
6651 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6652 {
6653 
6654 	if (*mvp == NULL) {
6655 		MNT_IUNLOCK(mp);
6656 		return;
6657 	}
6658 
6659 	mtx_assert(MNT_MTX(mp), MA_OWNED);
6660 
6661 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6662 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6663 	MNT_REL(mp);
6664 	MNT_IUNLOCK(mp);
6665 	vn_free_marker(*mvp);
6666 	*mvp = NULL;
6667 }
6668 
6669 /*
6670  * These are helper functions for filesystems to traverse their
6671  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6672  */
6673 static void
6674 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6675 {
6676 
6677 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6678 
6679 	MNT_ILOCK(mp);
6680 	MNT_REL(mp);
6681 	MNT_IUNLOCK(mp);
6682 	vn_free_marker(*mvp);
6683 	*mvp = NULL;
6684 }
6685 
6686 /*
6687  * Relock the mp mount vnode list lock with the vp vnode interlock in the
6688  * conventional lock order during mnt_vnode_next_lazy iteration.
6689  *
6690  * On entry, the mount vnode list lock is held and the vnode interlock is not.
6691  * The list lock is dropped and reacquired.  On success, both locks are held.
6692  * On failure, the mount vnode list lock is held but the vnode interlock is
6693  * not, and the procedure may have yielded.
6694  */
6695 static bool
6696 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6697     struct vnode *vp)
6698 {
6699 
6700 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6701 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6702 	    ("%s: bad marker", __func__));
6703 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6704 	    ("%s: inappropriate vnode", __func__));
6705 	ASSERT_VI_UNLOCKED(vp, __func__);
6706 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6707 
6708 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6709 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6710 
6711 	/*
6712 	 * Note we may be racing against vdrop which transitioned the hold
6713 	 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6714 	 * if we are the only user after we get the interlock we will just
6715 	 * vdrop.
6716 	 */
6717 	vhold(vp);
6718 	mtx_unlock(&mp->mnt_listmtx);
6719 	VI_LOCK(vp);
6720 	if (VN_IS_DOOMED(vp)) {
6721 		VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6722 		goto out_lost;
6723 	}
6724 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6725 	/*
6726 	 * There is nothing to do if we are the last user.
6727 	 */
6728 	if (!refcount_release_if_not_last(&vp->v_holdcnt))
6729 		goto out_lost;
6730 	mtx_lock(&mp->mnt_listmtx);
6731 	return (true);
6732 out_lost:
6733 	vdropl(vp);
6734 	maybe_yield();
6735 	mtx_lock(&mp->mnt_listmtx);
6736 	return (false);
6737 }
6738 
6739 static struct vnode *
6740 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6741     void *cbarg)
6742 {
6743 	struct vnode *vp;
6744 
6745 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6746 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6747 restart:
6748 	vp = TAILQ_NEXT(*mvp, v_lazylist);
6749 	while (vp != NULL) {
6750 		if (vp->v_type == VMARKER) {
6751 			vp = TAILQ_NEXT(vp, v_lazylist);
6752 			continue;
6753 		}
6754 		/*
6755 		 * See if we want to process the vnode. Note we may encounter a
6756 		 * long string of vnodes we don't care about and hog the list
6757 		 * as a result. Check for it and requeue the marker.
6758 		 */
6759 		VNPASS(!VN_IS_DOOMED(vp), vp);
6760 		if (!cb(vp, cbarg)) {
6761 			if (!should_yield()) {
6762 				vp = TAILQ_NEXT(vp, v_lazylist);
6763 				continue;
6764 			}
6765 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6766 			    v_lazylist);
6767 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6768 			    v_lazylist);
6769 			mtx_unlock(&mp->mnt_listmtx);
6770 			kern_yield(PRI_USER);
6771 			mtx_lock(&mp->mnt_listmtx);
6772 			goto restart;
6773 		}
6774 		/*
6775 		 * Try-lock because this is the wrong lock order.
6776 		 */
6777 		if (!VI_TRYLOCK(vp) &&
6778 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6779 			goto restart;
6780 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6781 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6782 		    ("alien vnode on the lazy list %p %p", vp, mp));
6783 		VNPASS(vp->v_mount == mp, vp);
6784 		VNPASS(!VN_IS_DOOMED(vp), vp);
6785 		break;
6786 	}
6787 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6788 
6789 	/* Check if we are done */
6790 	if (vp == NULL) {
6791 		mtx_unlock(&mp->mnt_listmtx);
6792 		mnt_vnode_markerfree_lazy(mvp, mp);
6793 		return (NULL);
6794 	}
6795 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6796 	mtx_unlock(&mp->mnt_listmtx);
6797 	ASSERT_VI_LOCKED(vp, "lazy iter");
6798 	return (vp);
6799 }
6800 
6801 struct vnode *
6802 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6803     void *cbarg)
6804 {
6805 
6806 	maybe_yield();
6807 	mtx_lock(&mp->mnt_listmtx);
6808 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6809 }
6810 
6811 struct vnode *
6812 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6813     void *cbarg)
6814 {
6815 	struct vnode *vp;
6816 
6817 	if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6818 		return (NULL);
6819 
6820 	*mvp = vn_alloc_marker(mp);
6821 	MNT_ILOCK(mp);
6822 	MNT_REF(mp);
6823 	MNT_IUNLOCK(mp);
6824 
6825 	mtx_lock(&mp->mnt_listmtx);
6826 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6827 	if (vp == NULL) {
6828 		mtx_unlock(&mp->mnt_listmtx);
6829 		mnt_vnode_markerfree_lazy(mvp, mp);
6830 		return (NULL);
6831 	}
6832 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6833 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6834 }
6835 
6836 void
6837 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6838 {
6839 
6840 	if (*mvp == NULL)
6841 		return;
6842 
6843 	mtx_lock(&mp->mnt_listmtx);
6844 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6845 	mtx_unlock(&mp->mnt_listmtx);
6846 	mnt_vnode_markerfree_lazy(mvp, mp);
6847 }
6848 
6849 int
6850 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6851 {
6852 
6853 	if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6854 		cnp->cn_flags &= ~NOEXECCHECK;
6855 		return (0);
6856 	}
6857 
6858 	return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, curthread));
6859 }
6860 
6861 /*
6862  * Do not use this variant unless you have means other than the hold count
6863  * to prevent the vnode from getting freed.
6864  */
6865 void
6866 vn_seqc_write_begin_locked(struct vnode *vp)
6867 {
6868 
6869 	ASSERT_VI_LOCKED(vp, __func__);
6870 	VNPASS(vp->v_holdcnt > 0, vp);
6871 	VNPASS(vp->v_seqc_users >= 0, vp);
6872 	vp->v_seqc_users++;
6873 	if (vp->v_seqc_users == 1)
6874 		seqc_sleepable_write_begin(&vp->v_seqc);
6875 }
6876 
6877 void
6878 vn_seqc_write_begin(struct vnode *vp)
6879 {
6880 
6881 	VI_LOCK(vp);
6882 	vn_seqc_write_begin_locked(vp);
6883 	VI_UNLOCK(vp);
6884 }
6885 
6886 void
6887 vn_seqc_write_end_locked(struct vnode *vp)
6888 {
6889 
6890 	ASSERT_VI_LOCKED(vp, __func__);
6891 	VNPASS(vp->v_seqc_users > 0, vp);
6892 	vp->v_seqc_users--;
6893 	if (vp->v_seqc_users == 0)
6894 		seqc_sleepable_write_end(&vp->v_seqc);
6895 }
6896 
6897 void
6898 vn_seqc_write_end(struct vnode *vp)
6899 {
6900 
6901 	VI_LOCK(vp);
6902 	vn_seqc_write_end_locked(vp);
6903 	VI_UNLOCK(vp);
6904 }
6905 
6906 /*
6907  * Special case handling for allocating and freeing vnodes.
6908  *
6909  * The counter remains unchanged on free so that a doomed vnode will
6910  * keep testing as in modify as long as it is accessible with SMR.
6911  */
6912 static void
6913 vn_seqc_init(struct vnode *vp)
6914 {
6915 
6916 	vp->v_seqc = 0;
6917 	vp->v_seqc_users = 0;
6918 }
6919 
6920 static void
6921 vn_seqc_write_end_free(struct vnode *vp)
6922 {
6923 
6924 	VNPASS(seqc_in_modify(vp->v_seqc), vp);
6925 	VNPASS(vp->v_seqc_users == 1, vp);
6926 }
6927 
6928 void
6929 vn_irflag_set_locked(struct vnode *vp, short toset)
6930 {
6931 	short flags;
6932 
6933 	ASSERT_VI_LOCKED(vp, __func__);
6934 	flags = vn_irflag_read(vp);
6935 	VNASSERT((flags & toset) == 0, vp,
6936 	    ("%s: some of the passed flags already set (have %d, passed %d)\n",
6937 	    __func__, flags, toset));
6938 	atomic_store_short(&vp->v_irflag, flags | toset);
6939 }
6940 
6941 void
6942 vn_irflag_set(struct vnode *vp, short toset)
6943 {
6944 
6945 	VI_LOCK(vp);
6946 	vn_irflag_set_locked(vp, toset);
6947 	VI_UNLOCK(vp);
6948 }
6949 
6950 void
6951 vn_irflag_set_cond_locked(struct vnode *vp, short toset)
6952 {
6953 	short flags;
6954 
6955 	ASSERT_VI_LOCKED(vp, __func__);
6956 	flags = vn_irflag_read(vp);
6957 	atomic_store_short(&vp->v_irflag, flags | toset);
6958 }
6959 
6960 void
6961 vn_irflag_set_cond(struct vnode *vp, short toset)
6962 {
6963 
6964 	VI_LOCK(vp);
6965 	vn_irflag_set_cond_locked(vp, toset);
6966 	VI_UNLOCK(vp);
6967 }
6968 
6969 void
6970 vn_irflag_unset_locked(struct vnode *vp, short tounset)
6971 {
6972 	short flags;
6973 
6974 	ASSERT_VI_LOCKED(vp, __func__);
6975 	flags = vn_irflag_read(vp);
6976 	VNASSERT((flags & tounset) == tounset, vp,
6977 	    ("%s: some of the passed flags not set (have %d, passed %d)\n",
6978 	    __func__, flags, tounset));
6979 	atomic_store_short(&vp->v_irflag, flags & ~tounset);
6980 }
6981 
6982 void
6983 vn_irflag_unset(struct vnode *vp, short tounset)
6984 {
6985 
6986 	VI_LOCK(vp);
6987 	vn_irflag_unset_locked(vp, tounset);
6988 	VI_UNLOCK(vp);
6989 }
6990 
6991 int
6992 vn_getsize_locked(struct vnode *vp, off_t *size, struct ucred *cred)
6993 {
6994 	struct vattr vattr;
6995 	int error;
6996 
6997 	ASSERT_VOP_LOCKED(vp, __func__);
6998 	error = VOP_GETATTR(vp, &vattr, cred);
6999 	if (__predict_true(error == 0)) {
7000 		if (vattr.va_size <= OFF_MAX)
7001 			*size = vattr.va_size;
7002 		else
7003 			error = EFBIG;
7004 	}
7005 	return (error);
7006 }
7007 
7008 int
7009 vn_getsize(struct vnode *vp, off_t *size, struct ucred *cred)
7010 {
7011 	int error;
7012 
7013 	VOP_LOCK(vp, LK_SHARED);
7014 	error = vn_getsize_locked(vp, size, cred);
7015 	VOP_UNLOCK(vp);
7016 	return (error);
7017 }
7018 
7019 #ifdef INVARIANTS
7020 void
7021 vn_set_state_validate(struct vnode *vp, __enum_uint8(vstate) state)
7022 {
7023 
7024 	switch (vp->v_state) {
7025 	case VSTATE_UNINITIALIZED:
7026 		switch (state) {
7027 		case VSTATE_CONSTRUCTED:
7028 		case VSTATE_DESTROYING:
7029 			return;
7030 		default:
7031 			break;
7032 		}
7033 		break;
7034 	case VSTATE_CONSTRUCTED:
7035 		ASSERT_VOP_ELOCKED(vp, __func__);
7036 		switch (state) {
7037 		case VSTATE_DESTROYING:
7038 			return;
7039 		default:
7040 			break;
7041 		}
7042 		break;
7043 	case VSTATE_DESTROYING:
7044 		ASSERT_VOP_ELOCKED(vp, __func__);
7045 		switch (state) {
7046 		case VSTATE_DEAD:
7047 			return;
7048 		default:
7049 			break;
7050 		}
7051 		break;
7052 	case VSTATE_DEAD:
7053 		switch (state) {
7054 		case VSTATE_UNINITIALIZED:
7055 			return;
7056 		default:
7057 			break;
7058 		}
7059 		break;
7060 	}
7061 
7062 	vn_printf(vp, "invalid state transition %d -> %d\n", vp->v_state, state);
7063 	panic("invalid state transition %d -> %d\n", vp->v_state, state);
7064 }
7065 #endif
7066