xref: /freebsd/sys/kern/vfs_subr.c (revision 5b5b7e2ca2fa9a2418dd51749f4ef6f881ae7179)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	@(#)vfs_subr.c	8.31 (Berkeley) 5/26/95
37  */
38 
39 /*
40  * External virtual filesystem routines
41  */
42 
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
45 
46 #include "opt_ddb.h"
47 #include "opt_watchdog.h"
48 
49 #include <sys/param.h>
50 #include <sys/systm.h>
51 #include <sys/asan.h>
52 #include <sys/bio.h>
53 #include <sys/buf.h>
54 #include <sys/capsicum.h>
55 #include <sys/condvar.h>
56 #include <sys/conf.h>
57 #include <sys/counter.h>
58 #include <sys/dirent.h>
59 #include <sys/event.h>
60 #include <sys/eventhandler.h>
61 #include <sys/extattr.h>
62 #include <sys/file.h>
63 #include <sys/fcntl.h>
64 #include <sys/jail.h>
65 #include <sys/kdb.h>
66 #include <sys/kernel.h>
67 #include <sys/kthread.h>
68 #include <sys/ktr.h>
69 #include <sys/lockf.h>
70 #include <sys/malloc.h>
71 #include <sys/mount.h>
72 #include <sys/namei.h>
73 #include <sys/pctrie.h>
74 #include <sys/priv.h>
75 #include <sys/reboot.h>
76 #include <sys/refcount.h>
77 #include <sys/rwlock.h>
78 #include <sys/sched.h>
79 #include <sys/sleepqueue.h>
80 #include <sys/smr.h>
81 #include <sys/smp.h>
82 #include <sys/stat.h>
83 #include <sys/sysctl.h>
84 #include <sys/syslog.h>
85 #include <sys/vmmeter.h>
86 #include <sys/vnode.h>
87 #include <sys/watchdog.h>
88 
89 #include <machine/stdarg.h>
90 
91 #include <security/mac/mac_framework.h>
92 
93 #include <vm/vm.h>
94 #include <vm/vm_object.h>
95 #include <vm/vm_extern.h>
96 #include <vm/pmap.h>
97 #include <vm/vm_map.h>
98 #include <vm/vm_page.h>
99 #include <vm/vm_kern.h>
100 #include <vm/uma.h>
101 
102 #ifdef DDB
103 #include <ddb/ddb.h>
104 #endif
105 
106 static void	delmntque(struct vnode *vp);
107 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
108 		    int slpflag, int slptimeo);
109 static void	syncer_shutdown(void *arg, int howto);
110 static int	vtryrecycle(struct vnode *vp);
111 static void	v_init_counters(struct vnode *);
112 static void	vn_seqc_init(struct vnode *);
113 static void	vn_seqc_write_end_free(struct vnode *vp);
114 static void	vgonel(struct vnode *);
115 static bool	vhold_recycle_free(struct vnode *);
116 static void	vdropl_recycle(struct vnode *vp);
117 static void	vdrop_recycle(struct vnode *vp);
118 static void	vfs_knllock(void *arg);
119 static void	vfs_knlunlock(void *arg);
120 static void	vfs_knl_assert_lock(void *arg, int what);
121 static void	destroy_vpollinfo(struct vpollinfo *vi);
122 static int	v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
123 		    daddr_t startlbn, daddr_t endlbn);
124 static void	vnlru_recalc(void);
125 
126 /*
127  * Number of vnodes in existence.  Increased whenever getnewvnode()
128  * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
129  */
130 static u_long __exclusive_cache_line numvnodes;
131 
132 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
133     "Number of vnodes in existence");
134 
135 static counter_u64_t vnodes_created;
136 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
137     "Number of vnodes created by getnewvnode");
138 
139 /*
140  * Conversion tables for conversion from vnode types to inode formats
141  * and back.
142  */
143 enum vtype iftovt_tab[16] = {
144 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
145 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
146 };
147 int vttoif_tab[10] = {
148 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
149 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
150 };
151 
152 /*
153  * List of allocates vnodes in the system.
154  */
155 static TAILQ_HEAD(freelst, vnode) vnode_list;
156 static struct vnode *vnode_list_free_marker;
157 static struct vnode *vnode_list_reclaim_marker;
158 
159 /*
160  * "Free" vnode target.  Free vnodes are rarely completely free, but are
161  * just ones that are cheap to recycle.  Usually they are for files which
162  * have been stat'd but not read; these usually have inode and namecache
163  * data attached to them.  This target is the preferred minimum size of a
164  * sub-cache consisting mostly of such files. The system balances the size
165  * of this sub-cache with its complement to try to prevent either from
166  * thrashing while the other is relatively inactive.  The targets express
167  * a preference for the best balance.
168  *
169  * "Above" this target there are 2 further targets (watermarks) related
170  * to recyling of free vnodes.  In the best-operating case, the cache is
171  * exactly full, the free list has size between vlowat and vhiwat above the
172  * free target, and recycling from it and normal use maintains this state.
173  * Sometimes the free list is below vlowat or even empty, but this state
174  * is even better for immediate use provided the cache is not full.
175  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
176  * ones) to reach one of these states.  The watermarks are currently hard-
177  * coded as 4% and 9% of the available space higher.  These and the default
178  * of 25% for wantfreevnodes are too large if the memory size is large.
179  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
180  * whenever vnlru_proc() becomes active.
181  */
182 static long wantfreevnodes;
183 static long __exclusive_cache_line freevnodes;
184 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
185     &freevnodes, 0, "Number of \"free\" vnodes");
186 static long freevnodes_old;
187 
188 static counter_u64_t recycles_count;
189 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
190     "Number of vnodes recycled to meet vnode cache targets");
191 
192 static counter_u64_t recycles_free_count;
193 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count,
194     "Number of free vnodes recycled to meet vnode cache targets");
195 
196 static counter_u64_t deferred_inact;
197 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact,
198     "Number of times inactive processing was deferred");
199 
200 /* To keep more than one thread at a time from running vfs_getnewfsid */
201 static struct mtx mntid_mtx;
202 
203 /*
204  * Lock for any access to the following:
205  *	vnode_list
206  *	numvnodes
207  *	freevnodes
208  */
209 static struct mtx __exclusive_cache_line vnode_list_mtx;
210 
211 /* Publicly exported FS */
212 struct nfs_public nfs_pub;
213 
214 static uma_zone_t buf_trie_zone;
215 static smr_t buf_trie_smr;
216 
217 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
218 static uma_zone_t vnode_zone;
219 MALLOC_DEFINE(M_VNODEPOLL, "VN POLL", "vnode poll");
220 
221 __read_frequently smr_t vfs_smr;
222 
223 /*
224  * The workitem queue.
225  *
226  * It is useful to delay writes of file data and filesystem metadata
227  * for tens of seconds so that quickly created and deleted files need
228  * not waste disk bandwidth being created and removed. To realize this,
229  * we append vnodes to a "workitem" queue. When running with a soft
230  * updates implementation, most pending metadata dependencies should
231  * not wait for more than a few seconds. Thus, mounted on block devices
232  * are delayed only about a half the time that file data is delayed.
233  * Similarly, directory updates are more critical, so are only delayed
234  * about a third the time that file data is delayed. Thus, there are
235  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
236  * one each second (driven off the filesystem syncer process). The
237  * syncer_delayno variable indicates the next queue that is to be processed.
238  * Items that need to be processed soon are placed in this queue:
239  *
240  *	syncer_workitem_pending[syncer_delayno]
241  *
242  * A delay of fifteen seconds is done by placing the request fifteen
243  * entries later in the queue:
244  *
245  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
246  *
247  */
248 static int syncer_delayno;
249 static long syncer_mask;
250 LIST_HEAD(synclist, bufobj);
251 static struct synclist *syncer_workitem_pending;
252 /*
253  * The sync_mtx protects:
254  *	bo->bo_synclist
255  *	sync_vnode_count
256  *	syncer_delayno
257  *	syncer_state
258  *	syncer_workitem_pending
259  *	syncer_worklist_len
260  *	rushjob
261  */
262 static struct mtx sync_mtx;
263 static struct cv sync_wakeup;
264 
265 #define SYNCER_MAXDELAY		32
266 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
267 static int syncdelay = 30;		/* max time to delay syncing data */
268 static int filedelay = 30;		/* time to delay syncing files */
269 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
270     "Time to delay syncing files (in seconds)");
271 static int dirdelay = 29;		/* time to delay syncing directories */
272 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
273     "Time to delay syncing directories (in seconds)");
274 static int metadelay = 28;		/* time to delay syncing metadata */
275 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
276     "Time to delay syncing metadata (in seconds)");
277 static int rushjob;		/* number of slots to run ASAP */
278 static int stat_rush_requests;	/* number of times I/O speeded up */
279 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
280     "Number of times I/O speeded up (rush requests)");
281 
282 #define	VDBATCH_SIZE 8
283 struct vdbatch {
284 	u_int index;
285 	long freevnodes;
286 	struct mtx lock;
287 	struct vnode *tab[VDBATCH_SIZE];
288 };
289 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
290 
291 static void	vdbatch_dequeue(struct vnode *vp);
292 
293 /*
294  * When shutting down the syncer, run it at four times normal speed.
295  */
296 #define SYNCER_SHUTDOWN_SPEEDUP		4
297 static int sync_vnode_count;
298 static int syncer_worklist_len;
299 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
300     syncer_state;
301 
302 /* Target for maximum number of vnodes. */
303 u_long desiredvnodes;
304 static u_long gapvnodes;		/* gap between wanted and desired */
305 static u_long vhiwat;		/* enough extras after expansion */
306 static u_long vlowat;		/* minimal extras before expansion */
307 static u_long vstir;		/* nonzero to stir non-free vnodes */
308 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
309 
310 static u_long vnlru_read_freevnodes(void);
311 
312 /*
313  * Note that no attempt is made to sanitize these parameters.
314  */
315 static int
316 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
317 {
318 	u_long val;
319 	int error;
320 
321 	val = desiredvnodes;
322 	error = sysctl_handle_long(oidp, &val, 0, req);
323 	if (error != 0 || req->newptr == NULL)
324 		return (error);
325 
326 	if (val == desiredvnodes)
327 		return (0);
328 	mtx_lock(&vnode_list_mtx);
329 	desiredvnodes = val;
330 	wantfreevnodes = desiredvnodes / 4;
331 	vnlru_recalc();
332 	mtx_unlock(&vnode_list_mtx);
333 	/*
334 	 * XXX There is no protection against multiple threads changing
335 	 * desiredvnodes at the same time. Locking above only helps vnlru and
336 	 * getnewvnode.
337 	 */
338 	vfs_hash_changesize(desiredvnodes);
339 	cache_changesize(desiredvnodes);
340 	return (0);
341 }
342 
343 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
344     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
345     "LU", "Target for maximum number of vnodes");
346 
347 static int
348 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
349 {
350 	u_long val;
351 	int error;
352 
353 	val = wantfreevnodes;
354 	error = sysctl_handle_long(oidp, &val, 0, req);
355 	if (error != 0 || req->newptr == NULL)
356 		return (error);
357 
358 	if (val == wantfreevnodes)
359 		return (0);
360 	mtx_lock(&vnode_list_mtx);
361 	wantfreevnodes = val;
362 	vnlru_recalc();
363 	mtx_unlock(&vnode_list_mtx);
364 	return (0);
365 }
366 
367 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
368     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
369     "LU", "Target for minimum number of \"free\" vnodes");
370 
371 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
372     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
373 static int vnlru_nowhere;
374 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
375     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
376 
377 static int
378 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
379 {
380 	struct vnode *vp;
381 	struct nameidata nd;
382 	char *buf;
383 	unsigned long ndflags;
384 	int error;
385 
386 	if (req->newptr == NULL)
387 		return (EINVAL);
388 	if (req->newlen >= PATH_MAX)
389 		return (E2BIG);
390 
391 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
392 	error = SYSCTL_IN(req, buf, req->newlen);
393 	if (error != 0)
394 		goto out;
395 
396 	buf[req->newlen] = '\0';
397 
398 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1;
399 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf);
400 	if ((error = namei(&nd)) != 0)
401 		goto out;
402 	vp = nd.ni_vp;
403 
404 	if (VN_IS_DOOMED(vp)) {
405 		/*
406 		 * This vnode is being recycled.  Return != 0 to let the caller
407 		 * know that the sysctl had no effect.  Return EAGAIN because a
408 		 * subsequent call will likely succeed (since namei will create
409 		 * a new vnode if necessary)
410 		 */
411 		error = EAGAIN;
412 		goto putvnode;
413 	}
414 
415 	counter_u64_add(recycles_count, 1);
416 	vgone(vp);
417 putvnode:
418 	NDFREE(&nd, 0);
419 out:
420 	free(buf, M_TEMP);
421 	return (error);
422 }
423 
424 static int
425 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
426 {
427 	struct thread *td = curthread;
428 	struct vnode *vp;
429 	struct file *fp;
430 	int error;
431 	int fd;
432 
433 	if (req->newptr == NULL)
434 		return (EBADF);
435 
436         error = sysctl_handle_int(oidp, &fd, 0, req);
437         if (error != 0)
438                 return (error);
439 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
440 	if (error != 0)
441 		return (error);
442 	vp = fp->f_vnode;
443 
444 	error = vn_lock(vp, LK_EXCLUSIVE);
445 	if (error != 0)
446 		goto drop;
447 
448 	counter_u64_add(recycles_count, 1);
449 	vgone(vp);
450 	VOP_UNLOCK(vp);
451 drop:
452 	fdrop(fp, td);
453 	return (error);
454 }
455 
456 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
457     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
458     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
459 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
460     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
461     sysctl_ftry_reclaim_vnode, "I",
462     "Try to reclaim a vnode by its file descriptor");
463 
464 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
465 #define vnsz2log 8
466 #ifndef DEBUG_LOCKS
467 _Static_assert(sizeof(struct vnode) >= 1UL << vnsz2log &&
468     sizeof(struct vnode) < 1UL << (vnsz2log + 1),
469     "vnsz2log needs to be updated");
470 #endif
471 
472 /*
473  * Support for the bufobj clean & dirty pctrie.
474  */
475 static void *
476 buf_trie_alloc(struct pctrie *ptree)
477 {
478 	return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT));
479 }
480 
481 static void
482 buf_trie_free(struct pctrie *ptree, void *node)
483 {
484 	uma_zfree_smr(buf_trie_zone, node);
485 }
486 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free,
487     buf_trie_smr);
488 
489 /*
490  * Initialize the vnode management data structures.
491  *
492  * Reevaluate the following cap on the number of vnodes after the physical
493  * memory size exceeds 512GB.  In the limit, as the physical memory size
494  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
495  */
496 #ifndef	MAXVNODES_MAX
497 #define	MAXVNODES_MAX	(512UL * 1024 * 1024 / 64)	/* 8M */
498 #endif
499 
500 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
501 
502 static struct vnode *
503 vn_alloc_marker(struct mount *mp)
504 {
505 	struct vnode *vp;
506 
507 	vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
508 	vp->v_type = VMARKER;
509 	vp->v_mount = mp;
510 
511 	return (vp);
512 }
513 
514 static void
515 vn_free_marker(struct vnode *vp)
516 {
517 
518 	MPASS(vp->v_type == VMARKER);
519 	free(vp, M_VNODE_MARKER);
520 }
521 
522 #ifdef KASAN
523 static int
524 vnode_ctor(void *mem, int size, void *arg __unused, int flags __unused)
525 {
526 	kasan_mark(mem, size, roundup2(size, UMA_ALIGN_PTR + 1), 0);
527 	return (0);
528 }
529 
530 static void
531 vnode_dtor(void *mem, int size, void *arg __unused)
532 {
533 	size_t end1, end2, off1, off2;
534 
535 	_Static_assert(offsetof(struct vnode, v_vnodelist) <
536 	    offsetof(struct vnode, v_dbatchcpu),
537 	    "KASAN marks require updating");
538 
539 	off1 = offsetof(struct vnode, v_vnodelist);
540 	off2 = offsetof(struct vnode, v_dbatchcpu);
541 	end1 = off1 + sizeof(((struct vnode *)NULL)->v_vnodelist);
542 	end2 = off2 + sizeof(((struct vnode *)NULL)->v_dbatchcpu);
543 
544 	/*
545 	 * Access to the v_vnodelist and v_dbatchcpu fields are permitted even
546 	 * after the vnode has been freed.  Try to get some KASAN coverage by
547 	 * marking everything except those two fields as invalid.  Because
548 	 * KASAN's tracking is not byte-granular, any preceding fields sharing
549 	 * the same 8-byte aligned word must also be marked valid.
550 	 */
551 
552 	/* Handle the area from the start until v_vnodelist... */
553 	off1 = rounddown2(off1, KASAN_SHADOW_SCALE);
554 	kasan_mark(mem, off1, off1, KASAN_UMA_FREED);
555 
556 	/* ... then the area between v_vnodelist and v_dbatchcpu ... */
557 	off1 = roundup2(end1, KASAN_SHADOW_SCALE);
558 	off2 = rounddown2(off2, KASAN_SHADOW_SCALE);
559 	if (off2 > off1)
560 		kasan_mark((void *)((char *)mem + off1), off2 - off1,
561 		    off2 - off1, KASAN_UMA_FREED);
562 
563 	/* ... and finally the area from v_dbatchcpu to the end. */
564 	off2 = roundup2(end2, KASAN_SHADOW_SCALE);
565 	kasan_mark((void *)((char *)mem + off2), size - off2, size - off2,
566 	    KASAN_UMA_FREED);
567 }
568 #endif /* KASAN */
569 
570 /*
571  * Initialize a vnode as it first enters the zone.
572  */
573 static int
574 vnode_init(void *mem, int size, int flags)
575 {
576 	struct vnode *vp;
577 
578 	vp = mem;
579 	bzero(vp, size);
580 	/*
581 	 * Setup locks.
582 	 */
583 	vp->v_vnlock = &vp->v_lock;
584 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
585 	/*
586 	 * By default, don't allow shared locks unless filesystems opt-in.
587 	 */
588 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
589 	    LK_NOSHARE | LK_IS_VNODE);
590 	/*
591 	 * Initialize bufobj.
592 	 */
593 	bufobj_init(&vp->v_bufobj, vp);
594 	/*
595 	 * Initialize namecache.
596 	 */
597 	cache_vnode_init(vp);
598 	/*
599 	 * Initialize rangelocks.
600 	 */
601 	rangelock_init(&vp->v_rl);
602 
603 	vp->v_dbatchcpu = NOCPU;
604 
605 	/*
606 	 * Check vhold_recycle_free for an explanation.
607 	 */
608 	vp->v_holdcnt = VHOLD_NO_SMR;
609 	vp->v_type = VNON;
610 	mtx_lock(&vnode_list_mtx);
611 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
612 	mtx_unlock(&vnode_list_mtx);
613 	return (0);
614 }
615 
616 /*
617  * Free a vnode when it is cleared from the zone.
618  */
619 static void
620 vnode_fini(void *mem, int size)
621 {
622 	struct vnode *vp;
623 	struct bufobj *bo;
624 
625 	vp = mem;
626 	vdbatch_dequeue(vp);
627 	mtx_lock(&vnode_list_mtx);
628 	TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
629 	mtx_unlock(&vnode_list_mtx);
630 	rangelock_destroy(&vp->v_rl);
631 	lockdestroy(vp->v_vnlock);
632 	mtx_destroy(&vp->v_interlock);
633 	bo = &vp->v_bufobj;
634 	rw_destroy(BO_LOCKPTR(bo));
635 
636 	kasan_mark(mem, size, size, 0);
637 }
638 
639 /*
640  * Provide the size of NFS nclnode and NFS fh for calculation of the
641  * vnode memory consumption.  The size is specified directly to
642  * eliminate dependency on NFS-private header.
643  *
644  * Other filesystems may use bigger or smaller (like UFS and ZFS)
645  * private inode data, but the NFS-based estimation is ample enough.
646  * Still, we care about differences in the size between 64- and 32-bit
647  * platforms.
648  *
649  * Namecache structure size is heuristically
650  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
651  */
652 #ifdef _LP64
653 #define	NFS_NCLNODE_SZ	(528 + 64)
654 #define	NC_SZ		148
655 #else
656 #define	NFS_NCLNODE_SZ	(360 + 32)
657 #define	NC_SZ		92
658 #endif
659 
660 static void
661 vntblinit(void *dummy __unused)
662 {
663 	struct vdbatch *vd;
664 	uma_ctor ctor;
665 	uma_dtor dtor;
666 	int cpu, physvnodes, virtvnodes;
667 
668 	/*
669 	 * Desiredvnodes is a function of the physical memory size and the
670 	 * kernel's heap size.  Generally speaking, it scales with the
671 	 * physical memory size.  The ratio of desiredvnodes to the physical
672 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
673 	 * Thereafter, the
674 	 * marginal ratio of desiredvnodes to the physical memory size is
675 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
676 	 * size.  The memory required by desiredvnodes vnodes and vm objects
677 	 * must not exceed 1/10th of the kernel's heap size.
678 	 */
679 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
680 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
681 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
682 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
683 	desiredvnodes = min(physvnodes, virtvnodes);
684 	if (desiredvnodes > MAXVNODES_MAX) {
685 		if (bootverbose)
686 			printf("Reducing kern.maxvnodes %lu -> %lu\n",
687 			    desiredvnodes, MAXVNODES_MAX);
688 		desiredvnodes = MAXVNODES_MAX;
689 	}
690 	wantfreevnodes = desiredvnodes / 4;
691 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
692 	TAILQ_INIT(&vnode_list);
693 	mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
694 	/*
695 	 * The lock is taken to appease WITNESS.
696 	 */
697 	mtx_lock(&vnode_list_mtx);
698 	vnlru_recalc();
699 	mtx_unlock(&vnode_list_mtx);
700 	vnode_list_free_marker = vn_alloc_marker(NULL);
701 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
702 	vnode_list_reclaim_marker = vn_alloc_marker(NULL);
703 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
704 
705 #ifdef KASAN
706 	ctor = vnode_ctor;
707 	dtor = vnode_dtor;
708 #else
709 	ctor = NULL;
710 	dtor = NULL;
711 #endif
712 	vnode_zone = uma_zcreate("VNODE", sizeof(struct vnode), ctor, dtor,
713 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, UMA_ZONE_NOKASAN);
714 	uma_zone_set_smr(vnode_zone, vfs_smr);
715 
716 	/*
717 	 * Preallocate enough nodes to support one-per buf so that
718 	 * we can not fail an insert.  reassignbuf() callers can not
719 	 * tolerate the insertion failure.
720 	 */
721 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
722 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
723 	    UMA_ZONE_NOFREE | UMA_ZONE_SMR);
724 	buf_trie_smr = uma_zone_get_smr(buf_trie_zone);
725 	uma_prealloc(buf_trie_zone, nbuf);
726 
727 	vnodes_created = counter_u64_alloc(M_WAITOK);
728 	recycles_count = counter_u64_alloc(M_WAITOK);
729 	recycles_free_count = counter_u64_alloc(M_WAITOK);
730 	deferred_inact = counter_u64_alloc(M_WAITOK);
731 
732 	/*
733 	 * Initialize the filesystem syncer.
734 	 */
735 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
736 	    &syncer_mask);
737 	syncer_maxdelay = syncer_mask + 1;
738 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
739 	cv_init(&sync_wakeup, "syncer");
740 
741 	CPU_FOREACH(cpu) {
742 		vd = DPCPU_ID_PTR((cpu), vd);
743 		bzero(vd, sizeof(*vd));
744 		mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
745 	}
746 }
747 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
748 
749 /*
750  * Mark a mount point as busy. Used to synchronize access and to delay
751  * unmounting. Eventually, mountlist_mtx is not released on failure.
752  *
753  * vfs_busy() is a custom lock, it can block the caller.
754  * vfs_busy() only sleeps if the unmount is active on the mount point.
755  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
756  * vnode belonging to mp.
757  *
758  * Lookup uses vfs_busy() to traverse mount points.
759  * root fs			var fs
760  * / vnode lock		A	/ vnode lock (/var)		D
761  * /var vnode lock	B	/log vnode lock(/var/log)	E
762  * vfs_busy lock	C	vfs_busy lock			F
763  *
764  * Within each file system, the lock order is C->A->B and F->D->E.
765  *
766  * When traversing across mounts, the system follows that lock order:
767  *
768  *        C->A->B
769  *              |
770  *              +->F->D->E
771  *
772  * The lookup() process for namei("/var") illustrates the process:
773  *  VOP_LOOKUP() obtains B while A is held
774  *  vfs_busy() obtains a shared lock on F while A and B are held
775  *  vput() releases lock on B
776  *  vput() releases lock on A
777  *  VFS_ROOT() obtains lock on D while shared lock on F is held
778  *  vfs_unbusy() releases shared lock on F
779  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
780  *    Attempt to lock A (instead of vp_crossmp) while D is held would
781  *    violate the global order, causing deadlocks.
782  *
783  * dounmount() locks B while F is drained.
784  */
785 int
786 vfs_busy(struct mount *mp, int flags)
787 {
788 	struct mount_pcpu *mpcpu;
789 
790 	MPASS((flags & ~MBF_MASK) == 0);
791 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
792 
793 	if (vfs_op_thread_enter(mp, mpcpu)) {
794 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
795 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
796 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
797 		vfs_mp_count_add_pcpu(mpcpu, ref, 1);
798 		vfs_mp_count_add_pcpu(mpcpu, lockref, 1);
799 		vfs_op_thread_exit(mp, mpcpu);
800 		if (flags & MBF_MNTLSTLOCK)
801 			mtx_unlock(&mountlist_mtx);
802 		return (0);
803 	}
804 
805 	MNT_ILOCK(mp);
806 	vfs_assert_mount_counters(mp);
807 	MNT_REF(mp);
808 	/*
809 	 * If mount point is currently being unmounted, sleep until the
810 	 * mount point fate is decided.  If thread doing the unmounting fails,
811 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
812 	 * that this mount point has survived the unmount attempt and vfs_busy
813 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
814 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
815 	 * about to be really destroyed.  vfs_busy needs to release its
816 	 * reference on the mount point in this case and return with ENOENT,
817 	 * telling the caller that mount mount it tried to busy is no longer
818 	 * valid.
819 	 */
820 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
821 		KASSERT(TAILQ_EMPTY(&mp->mnt_uppers),
822 		    ("%s: non-empty upper mount list with pending unmount",
823 		    __func__));
824 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
825 			MNT_REL(mp);
826 			MNT_IUNLOCK(mp);
827 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
828 			    __func__);
829 			return (ENOENT);
830 		}
831 		if (flags & MBF_MNTLSTLOCK)
832 			mtx_unlock(&mountlist_mtx);
833 		mp->mnt_kern_flag |= MNTK_MWAIT;
834 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
835 		if (flags & MBF_MNTLSTLOCK)
836 			mtx_lock(&mountlist_mtx);
837 		MNT_ILOCK(mp);
838 	}
839 	if (flags & MBF_MNTLSTLOCK)
840 		mtx_unlock(&mountlist_mtx);
841 	mp->mnt_lockref++;
842 	MNT_IUNLOCK(mp);
843 	return (0);
844 }
845 
846 /*
847  * Free a busy filesystem.
848  */
849 void
850 vfs_unbusy(struct mount *mp)
851 {
852 	struct mount_pcpu *mpcpu;
853 	int c;
854 
855 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
856 
857 	if (vfs_op_thread_enter(mp, mpcpu)) {
858 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
859 		vfs_mp_count_sub_pcpu(mpcpu, lockref, 1);
860 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
861 		vfs_op_thread_exit(mp, mpcpu);
862 		return;
863 	}
864 
865 	MNT_ILOCK(mp);
866 	vfs_assert_mount_counters(mp);
867 	MNT_REL(mp);
868 	c = --mp->mnt_lockref;
869 	if (mp->mnt_vfs_ops == 0) {
870 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
871 		MNT_IUNLOCK(mp);
872 		return;
873 	}
874 	if (c < 0)
875 		vfs_dump_mount_counters(mp);
876 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
877 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
878 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
879 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
880 		wakeup(&mp->mnt_lockref);
881 	}
882 	MNT_IUNLOCK(mp);
883 }
884 
885 /*
886  * Lookup a mount point by filesystem identifier.
887  */
888 struct mount *
889 vfs_getvfs(fsid_t *fsid)
890 {
891 	struct mount *mp;
892 
893 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
894 	mtx_lock(&mountlist_mtx);
895 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
896 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
897 			vfs_ref(mp);
898 			mtx_unlock(&mountlist_mtx);
899 			return (mp);
900 		}
901 	}
902 	mtx_unlock(&mountlist_mtx);
903 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
904 	return ((struct mount *) 0);
905 }
906 
907 /*
908  * Lookup a mount point by filesystem identifier, busying it before
909  * returning.
910  *
911  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
912  * cache for popular filesystem identifiers.  The cache is lockess, using
913  * the fact that struct mount's are never freed.  In worst case we may
914  * get pointer to unmounted or even different filesystem, so we have to
915  * check what we got, and go slow way if so.
916  */
917 struct mount *
918 vfs_busyfs(fsid_t *fsid)
919 {
920 #define	FSID_CACHE_SIZE	256
921 	typedef struct mount * volatile vmp_t;
922 	static vmp_t cache[FSID_CACHE_SIZE];
923 	struct mount *mp;
924 	int error;
925 	uint32_t hash;
926 
927 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
928 	hash = fsid->val[0] ^ fsid->val[1];
929 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
930 	mp = cache[hash];
931 	if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0)
932 		goto slow;
933 	if (vfs_busy(mp, 0) != 0) {
934 		cache[hash] = NULL;
935 		goto slow;
936 	}
937 	if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0)
938 		return (mp);
939 	else
940 	    vfs_unbusy(mp);
941 
942 slow:
943 	mtx_lock(&mountlist_mtx);
944 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
945 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
946 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
947 			if (error) {
948 				cache[hash] = NULL;
949 				mtx_unlock(&mountlist_mtx);
950 				return (NULL);
951 			}
952 			cache[hash] = mp;
953 			return (mp);
954 		}
955 	}
956 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
957 	mtx_unlock(&mountlist_mtx);
958 	return ((struct mount *) 0);
959 }
960 
961 /*
962  * Check if a user can access privileged mount options.
963  */
964 int
965 vfs_suser(struct mount *mp, struct thread *td)
966 {
967 	int error;
968 
969 	if (jailed(td->td_ucred)) {
970 		/*
971 		 * If the jail of the calling thread lacks permission for
972 		 * this type of file system, deny immediately.
973 		 */
974 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
975 			return (EPERM);
976 
977 		/*
978 		 * If the file system was mounted outside the jail of the
979 		 * calling thread, deny immediately.
980 		 */
981 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
982 			return (EPERM);
983 	}
984 
985 	/*
986 	 * If file system supports delegated administration, we don't check
987 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
988 	 * by the file system itself.
989 	 * If this is not the user that did original mount, we check for
990 	 * the PRIV_VFS_MOUNT_OWNER privilege.
991 	 */
992 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
993 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
994 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
995 			return (error);
996 	}
997 	return (0);
998 }
999 
1000 /*
1001  * Get a new unique fsid.  Try to make its val[0] unique, since this value
1002  * will be used to create fake device numbers for stat().  Also try (but
1003  * not so hard) make its val[0] unique mod 2^16, since some emulators only
1004  * support 16-bit device numbers.  We end up with unique val[0]'s for the
1005  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
1006  *
1007  * Keep in mind that several mounts may be running in parallel.  Starting
1008  * the search one past where the previous search terminated is both a
1009  * micro-optimization and a defense against returning the same fsid to
1010  * different mounts.
1011  */
1012 void
1013 vfs_getnewfsid(struct mount *mp)
1014 {
1015 	static uint16_t mntid_base;
1016 	struct mount *nmp;
1017 	fsid_t tfsid;
1018 	int mtype;
1019 
1020 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
1021 	mtx_lock(&mntid_mtx);
1022 	mtype = mp->mnt_vfc->vfc_typenum;
1023 	tfsid.val[1] = mtype;
1024 	mtype = (mtype & 0xFF) << 24;
1025 	for (;;) {
1026 		tfsid.val[0] = makedev(255,
1027 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
1028 		mntid_base++;
1029 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
1030 			break;
1031 		vfs_rel(nmp);
1032 	}
1033 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
1034 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
1035 	mtx_unlock(&mntid_mtx);
1036 }
1037 
1038 /*
1039  * Knob to control the precision of file timestamps:
1040  *
1041  *   0 = seconds only; nanoseconds zeroed.
1042  *   1 = seconds and nanoseconds, accurate within 1/HZ.
1043  *   2 = seconds and nanoseconds, truncated to microseconds.
1044  * >=3 = seconds and nanoseconds, maximum precision.
1045  */
1046 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
1047 
1048 static int timestamp_precision = TSP_USEC;
1049 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
1050     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
1051     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1052     "3+: sec + ns (max. precision))");
1053 
1054 /*
1055  * Get a current timestamp.
1056  */
1057 void
1058 vfs_timestamp(struct timespec *tsp)
1059 {
1060 	struct timeval tv;
1061 
1062 	switch (timestamp_precision) {
1063 	case TSP_SEC:
1064 		tsp->tv_sec = time_second;
1065 		tsp->tv_nsec = 0;
1066 		break;
1067 	case TSP_HZ:
1068 		getnanotime(tsp);
1069 		break;
1070 	case TSP_USEC:
1071 		microtime(&tv);
1072 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1073 		break;
1074 	case TSP_NSEC:
1075 	default:
1076 		nanotime(tsp);
1077 		break;
1078 	}
1079 }
1080 
1081 /*
1082  * Set vnode attributes to VNOVAL
1083  */
1084 void
1085 vattr_null(struct vattr *vap)
1086 {
1087 
1088 	vap->va_type = VNON;
1089 	vap->va_size = VNOVAL;
1090 	vap->va_bytes = VNOVAL;
1091 	vap->va_mode = VNOVAL;
1092 	vap->va_nlink = VNOVAL;
1093 	vap->va_uid = VNOVAL;
1094 	vap->va_gid = VNOVAL;
1095 	vap->va_fsid = VNOVAL;
1096 	vap->va_fileid = VNOVAL;
1097 	vap->va_blocksize = VNOVAL;
1098 	vap->va_rdev = VNOVAL;
1099 	vap->va_atime.tv_sec = VNOVAL;
1100 	vap->va_atime.tv_nsec = VNOVAL;
1101 	vap->va_mtime.tv_sec = VNOVAL;
1102 	vap->va_mtime.tv_nsec = VNOVAL;
1103 	vap->va_ctime.tv_sec = VNOVAL;
1104 	vap->va_ctime.tv_nsec = VNOVAL;
1105 	vap->va_birthtime.tv_sec = VNOVAL;
1106 	vap->va_birthtime.tv_nsec = VNOVAL;
1107 	vap->va_flags = VNOVAL;
1108 	vap->va_gen = VNOVAL;
1109 	vap->va_vaflags = 0;
1110 }
1111 
1112 /*
1113  * Try to reduce the total number of vnodes.
1114  *
1115  * This routine (and its user) are buggy in at least the following ways:
1116  * - all parameters were picked years ago when RAM sizes were significantly
1117  *   smaller
1118  * - it can pick vnodes based on pages used by the vm object, but filesystems
1119  *   like ZFS don't use it making the pick broken
1120  * - since ZFS has its own aging policy it gets partially combated by this one
1121  * - a dedicated method should be provided for filesystems to let them decide
1122  *   whether the vnode should be recycled
1123  *
1124  * This routine is called when we have too many vnodes.  It attempts
1125  * to free <count> vnodes and will potentially free vnodes that still
1126  * have VM backing store (VM backing store is typically the cause
1127  * of a vnode blowout so we want to do this).  Therefore, this operation
1128  * is not considered cheap.
1129  *
1130  * A number of conditions may prevent a vnode from being reclaimed.
1131  * the buffer cache may have references on the vnode, a directory
1132  * vnode may still have references due to the namei cache representing
1133  * underlying files, or the vnode may be in active use.   It is not
1134  * desirable to reuse such vnodes.  These conditions may cause the
1135  * number of vnodes to reach some minimum value regardless of what
1136  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
1137  *
1138  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1139  * 			 entries if this argument is strue
1140  * @param trigger	 Only reclaim vnodes with fewer than this many resident
1141  *			 pages.
1142  * @param target	 How many vnodes to reclaim.
1143  * @return		 The number of vnodes that were reclaimed.
1144  */
1145 static int
1146 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1147 {
1148 	struct vnode *vp, *mvp;
1149 	struct mount *mp;
1150 	struct vm_object *object;
1151 	u_long done;
1152 	bool retried;
1153 
1154 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1155 
1156 	retried = false;
1157 	done = 0;
1158 
1159 	mvp = vnode_list_reclaim_marker;
1160 restart:
1161 	vp = mvp;
1162 	while (done < target) {
1163 		vp = TAILQ_NEXT(vp, v_vnodelist);
1164 		if (__predict_false(vp == NULL))
1165 			break;
1166 
1167 		if (__predict_false(vp->v_type == VMARKER))
1168 			continue;
1169 
1170 		/*
1171 		 * If it's been deconstructed already, it's still
1172 		 * referenced, or it exceeds the trigger, skip it.
1173 		 * Also skip free vnodes.  We are trying to make space
1174 		 * to expand the free list, not reduce it.
1175 		 */
1176 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1177 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1178 			goto next_iter;
1179 
1180 		if (vp->v_type == VBAD || vp->v_type == VNON)
1181 			goto next_iter;
1182 
1183 		object = atomic_load_ptr(&vp->v_object);
1184 		if (object == NULL || object->resident_page_count > trigger) {
1185 			goto next_iter;
1186 		}
1187 
1188 		/*
1189 		 * Handle races against vnode allocation. Filesystems lock the
1190 		 * vnode some time after it gets returned from getnewvnode,
1191 		 * despite type and hold count being manipulated earlier.
1192 		 * Resorting to checking v_mount restores guarantees present
1193 		 * before the global list was reworked to contain all vnodes.
1194 		 */
1195 		if (!VI_TRYLOCK(vp))
1196 			goto next_iter;
1197 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1198 			VI_UNLOCK(vp);
1199 			goto next_iter;
1200 		}
1201 		if (vp->v_mount == NULL) {
1202 			VI_UNLOCK(vp);
1203 			goto next_iter;
1204 		}
1205 		vholdl(vp);
1206 		VI_UNLOCK(vp);
1207 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1208 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1209 		mtx_unlock(&vnode_list_mtx);
1210 
1211 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1212 			vdrop_recycle(vp);
1213 			goto next_iter_unlocked;
1214 		}
1215 		if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1216 			vdrop_recycle(vp);
1217 			vn_finished_write(mp);
1218 			goto next_iter_unlocked;
1219 		}
1220 
1221 		VI_LOCK(vp);
1222 		if (vp->v_usecount > 0 ||
1223 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1224 		    (vp->v_object != NULL && vp->v_object->handle == vp &&
1225 		    vp->v_object->resident_page_count > trigger)) {
1226 			VOP_UNLOCK(vp);
1227 			vdropl_recycle(vp);
1228 			vn_finished_write(mp);
1229 			goto next_iter_unlocked;
1230 		}
1231 		counter_u64_add(recycles_count, 1);
1232 		vgonel(vp);
1233 		VOP_UNLOCK(vp);
1234 		vdropl_recycle(vp);
1235 		vn_finished_write(mp);
1236 		done++;
1237 next_iter_unlocked:
1238 		if (should_yield())
1239 			kern_yield(PRI_USER);
1240 		mtx_lock(&vnode_list_mtx);
1241 		goto restart;
1242 next_iter:
1243 		MPASS(vp->v_type != VMARKER);
1244 		if (!should_yield())
1245 			continue;
1246 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1247 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1248 		mtx_unlock(&vnode_list_mtx);
1249 		kern_yield(PRI_USER);
1250 		mtx_lock(&vnode_list_mtx);
1251 		goto restart;
1252 	}
1253 	if (done == 0 && !retried) {
1254 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1255 		TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1256 		retried = true;
1257 		goto restart;
1258 	}
1259 	return (done);
1260 }
1261 
1262 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1263 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1264     0,
1265     "limit on vnode free requests per call to the vnlru_free routine");
1266 
1267 /*
1268  * Attempt to reduce the free list by the requested amount.
1269  */
1270 static int
1271 vnlru_free_impl(int count, struct vfsops *mnt_op, struct vnode *mvp)
1272 {
1273 	struct vnode *vp;
1274 	struct mount *mp;
1275 	int ocount;
1276 
1277 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1278 	if (count > max_vnlru_free)
1279 		count = max_vnlru_free;
1280 	ocount = count;
1281 	vp = mvp;
1282 	for (;;) {
1283 		if (count == 0) {
1284 			break;
1285 		}
1286 		vp = TAILQ_NEXT(vp, v_vnodelist);
1287 		if (__predict_false(vp == NULL)) {
1288 			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1289 			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1290 			break;
1291 		}
1292 		if (__predict_false(vp->v_type == VMARKER))
1293 			continue;
1294 		if (vp->v_holdcnt > 0)
1295 			continue;
1296 		/*
1297 		 * Don't recycle if our vnode is from different type
1298 		 * of mount point.  Note that mp is type-safe, the
1299 		 * check does not reach unmapped address even if
1300 		 * vnode is reclaimed.
1301 		 */
1302 		if (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1303 		    mp->mnt_op != mnt_op) {
1304 			continue;
1305 		}
1306 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1307 			continue;
1308 		}
1309 		if (!vhold_recycle_free(vp))
1310 			continue;
1311 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1312 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1313 		mtx_unlock(&vnode_list_mtx);
1314 		/*
1315 		 * FIXME: ignores the return value, meaning it may be nothing
1316 		 * got recycled but it claims otherwise to the caller.
1317 		 *
1318 		 * Originally the value started being ignored in 2005 with
1319 		 * 114a1006a8204aa156e1f9ad6476cdff89cada7f .
1320 		 *
1321 		 * Respecting the value can run into significant stalls if most
1322 		 * vnodes belong to one file system and it has writes
1323 		 * suspended.  In presence of many threads and millions of
1324 		 * vnodes they keep contending on the vnode_list_mtx lock only
1325 		 * to find vnodes they can't recycle.
1326 		 *
1327 		 * The solution would be to pre-check if the vnode is likely to
1328 		 * be recycle-able, but it needs to happen with the
1329 		 * vnode_list_mtx lock held. This runs into a problem where
1330 		 * VOP_GETWRITEMOUNT (currently needed to find out about if
1331 		 * writes are frozen) can take locks which LOR against it.
1332 		 *
1333 		 * Check nullfs for one example (null_getwritemount).
1334 		 */
1335 		vtryrecycle(vp);
1336 		count--;
1337 		mtx_lock(&vnode_list_mtx);
1338 		vp = mvp;
1339 	}
1340 	return (ocount - count);
1341 }
1342 
1343 static int
1344 vnlru_free_locked(int count)
1345 {
1346 
1347 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1348 	return (vnlru_free_impl(count, NULL, vnode_list_free_marker));
1349 }
1350 
1351 void
1352 vnlru_free_vfsops(int count, struct vfsops *mnt_op, struct vnode *mvp)
1353 {
1354 
1355 	MPASS(mnt_op != NULL);
1356 	MPASS(mvp != NULL);
1357 	VNPASS(mvp->v_type == VMARKER, mvp);
1358 	mtx_lock(&vnode_list_mtx);
1359 	vnlru_free_impl(count, mnt_op, mvp);
1360 	mtx_unlock(&vnode_list_mtx);
1361 }
1362 
1363 struct vnode *
1364 vnlru_alloc_marker(void)
1365 {
1366 	struct vnode *mvp;
1367 
1368 	mvp = vn_alloc_marker(NULL);
1369 	mtx_lock(&vnode_list_mtx);
1370 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, mvp, v_vnodelist);
1371 	mtx_unlock(&vnode_list_mtx);
1372 	return (mvp);
1373 }
1374 
1375 void
1376 vnlru_free_marker(struct vnode *mvp)
1377 {
1378 	mtx_lock(&vnode_list_mtx);
1379 	TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1380 	mtx_unlock(&vnode_list_mtx);
1381 	vn_free_marker(mvp);
1382 }
1383 
1384 static void
1385 vnlru_recalc(void)
1386 {
1387 
1388 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1389 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1390 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1391 	vlowat = vhiwat / 2;
1392 }
1393 
1394 /*
1395  * Attempt to recycle vnodes in a context that is always safe to block.
1396  * Calling vlrurecycle() from the bowels of filesystem code has some
1397  * interesting deadlock problems.
1398  */
1399 static struct proc *vnlruproc;
1400 static int vnlruproc_sig;
1401 
1402 /*
1403  * The main freevnodes counter is only updated when threads requeue their vnode
1404  * batches. CPUs are conditionally walked to compute a more accurate total.
1405  *
1406  * Limit how much of a slop are we willing to tolerate. Note: the actual value
1407  * at any given moment can still exceed slop, but it should not be by significant
1408  * margin in practice.
1409  */
1410 #define VNLRU_FREEVNODES_SLOP 128
1411 
1412 static __inline void
1413 vfs_freevnodes_inc(void)
1414 {
1415 	struct vdbatch *vd;
1416 
1417 	critical_enter();
1418 	vd = DPCPU_PTR(vd);
1419 	vd->freevnodes++;
1420 	critical_exit();
1421 }
1422 
1423 static __inline void
1424 vfs_freevnodes_dec(void)
1425 {
1426 	struct vdbatch *vd;
1427 
1428 	critical_enter();
1429 	vd = DPCPU_PTR(vd);
1430 	vd->freevnodes--;
1431 	critical_exit();
1432 }
1433 
1434 static u_long
1435 vnlru_read_freevnodes(void)
1436 {
1437 	struct vdbatch *vd;
1438 	long slop;
1439 	int cpu;
1440 
1441 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1442 	if (freevnodes > freevnodes_old)
1443 		slop = freevnodes - freevnodes_old;
1444 	else
1445 		slop = freevnodes_old - freevnodes;
1446 	if (slop < VNLRU_FREEVNODES_SLOP)
1447 		return (freevnodes >= 0 ? freevnodes : 0);
1448 	freevnodes_old = freevnodes;
1449 	CPU_FOREACH(cpu) {
1450 		vd = DPCPU_ID_PTR((cpu), vd);
1451 		freevnodes_old += vd->freevnodes;
1452 	}
1453 	return (freevnodes_old >= 0 ? freevnodes_old : 0);
1454 }
1455 
1456 static bool
1457 vnlru_under(u_long rnumvnodes, u_long limit)
1458 {
1459 	u_long rfreevnodes, space;
1460 
1461 	if (__predict_false(rnumvnodes > desiredvnodes))
1462 		return (true);
1463 
1464 	space = desiredvnodes - rnumvnodes;
1465 	if (space < limit) {
1466 		rfreevnodes = vnlru_read_freevnodes();
1467 		if (rfreevnodes > wantfreevnodes)
1468 			space += rfreevnodes - wantfreevnodes;
1469 	}
1470 	return (space < limit);
1471 }
1472 
1473 static bool
1474 vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1475 {
1476 	long rfreevnodes, space;
1477 
1478 	if (__predict_false(rnumvnodes > desiredvnodes))
1479 		return (true);
1480 
1481 	space = desiredvnodes - rnumvnodes;
1482 	if (space < limit) {
1483 		rfreevnodes = atomic_load_long(&freevnodes);
1484 		if (rfreevnodes > wantfreevnodes)
1485 			space += rfreevnodes - wantfreevnodes;
1486 	}
1487 	return (space < limit);
1488 }
1489 
1490 static void
1491 vnlru_kick(void)
1492 {
1493 
1494 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1495 	if (vnlruproc_sig == 0) {
1496 		vnlruproc_sig = 1;
1497 		wakeup(vnlruproc);
1498 	}
1499 }
1500 
1501 static void
1502 vnlru_proc(void)
1503 {
1504 	u_long rnumvnodes, rfreevnodes, target;
1505 	unsigned long onumvnodes;
1506 	int done, force, trigger, usevnodes;
1507 	bool reclaim_nc_src, want_reread;
1508 
1509 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1510 	    SHUTDOWN_PRI_FIRST);
1511 
1512 	force = 0;
1513 	want_reread = false;
1514 	for (;;) {
1515 		kproc_suspend_check(vnlruproc);
1516 		mtx_lock(&vnode_list_mtx);
1517 		rnumvnodes = atomic_load_long(&numvnodes);
1518 
1519 		if (want_reread) {
1520 			force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1521 			want_reread = false;
1522 		}
1523 
1524 		/*
1525 		 * If numvnodes is too large (due to desiredvnodes being
1526 		 * adjusted using its sysctl, or emergency growth), first
1527 		 * try to reduce it by discarding from the free list.
1528 		 */
1529 		if (rnumvnodes > desiredvnodes) {
1530 			vnlru_free_locked(rnumvnodes - desiredvnodes);
1531 			rnumvnodes = atomic_load_long(&numvnodes);
1532 		}
1533 		/*
1534 		 * Sleep if the vnode cache is in a good state.  This is
1535 		 * when it is not over-full and has space for about a 4%
1536 		 * or 9% expansion (by growing its size or inexcessively
1537 		 * reducing its free list).  Otherwise, try to reclaim
1538 		 * space for a 10% expansion.
1539 		 */
1540 		if (vstir && force == 0) {
1541 			force = 1;
1542 			vstir = 0;
1543 		}
1544 		if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1545 			vnlruproc_sig = 0;
1546 			wakeup(&vnlruproc_sig);
1547 			msleep(vnlruproc, &vnode_list_mtx,
1548 			    PVFS|PDROP, "vlruwt", hz);
1549 			continue;
1550 		}
1551 		rfreevnodes = vnlru_read_freevnodes();
1552 
1553 		onumvnodes = rnumvnodes;
1554 		/*
1555 		 * Calculate parameters for recycling.  These are the same
1556 		 * throughout the loop to give some semblance of fairness.
1557 		 * The trigger point is to avoid recycling vnodes with lots
1558 		 * of resident pages.  We aren't trying to free memory; we
1559 		 * are trying to recycle or at least free vnodes.
1560 		 */
1561 		if (rnumvnodes <= desiredvnodes)
1562 			usevnodes = rnumvnodes - rfreevnodes;
1563 		else
1564 			usevnodes = rnumvnodes;
1565 		if (usevnodes <= 0)
1566 			usevnodes = 1;
1567 		/*
1568 		 * The trigger value is chosen to give a conservatively
1569 		 * large value to ensure that it alone doesn't prevent
1570 		 * making progress.  The value can easily be so large that
1571 		 * it is effectively infinite in some congested and
1572 		 * misconfigured cases, and this is necessary.  Normally
1573 		 * it is about 8 to 100 (pages), which is quite large.
1574 		 */
1575 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1576 		if (force < 2)
1577 			trigger = vsmalltrigger;
1578 		reclaim_nc_src = force >= 3;
1579 		target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1580 		target = target / 10 + 1;
1581 		done = vlrureclaim(reclaim_nc_src, trigger, target);
1582 		mtx_unlock(&vnode_list_mtx);
1583 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1584 			uma_reclaim(UMA_RECLAIM_DRAIN);
1585 		if (done == 0) {
1586 			if (force == 0 || force == 1) {
1587 				force = 2;
1588 				continue;
1589 			}
1590 			if (force == 2) {
1591 				force = 3;
1592 				continue;
1593 			}
1594 			want_reread = true;
1595 			force = 0;
1596 			vnlru_nowhere++;
1597 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1598 		} else {
1599 			want_reread = true;
1600 			kern_yield(PRI_USER);
1601 		}
1602 	}
1603 }
1604 
1605 static struct kproc_desc vnlru_kp = {
1606 	"vnlru",
1607 	vnlru_proc,
1608 	&vnlruproc
1609 };
1610 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1611     &vnlru_kp);
1612 
1613 /*
1614  * Routines having to do with the management of the vnode table.
1615  */
1616 
1617 /*
1618  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1619  * before we actually vgone().  This function must be called with the vnode
1620  * held to prevent the vnode from being returned to the free list midway
1621  * through vgone().
1622  */
1623 static int
1624 vtryrecycle(struct vnode *vp)
1625 {
1626 	struct mount *vnmp;
1627 
1628 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1629 	VNASSERT(vp->v_holdcnt, vp,
1630 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1631 	/*
1632 	 * This vnode may found and locked via some other list, if so we
1633 	 * can't recycle it yet.
1634 	 */
1635 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1636 		CTR2(KTR_VFS,
1637 		    "%s: impossible to recycle, vp %p lock is already held",
1638 		    __func__, vp);
1639 		vdrop_recycle(vp);
1640 		return (EWOULDBLOCK);
1641 	}
1642 	/*
1643 	 * Don't recycle if its filesystem is being suspended.
1644 	 */
1645 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1646 		VOP_UNLOCK(vp);
1647 		CTR2(KTR_VFS,
1648 		    "%s: impossible to recycle, cannot start the write for %p",
1649 		    __func__, vp);
1650 		vdrop_recycle(vp);
1651 		return (EBUSY);
1652 	}
1653 	/*
1654 	 * If we got this far, we need to acquire the interlock and see if
1655 	 * anyone picked up this vnode from another list.  If not, we will
1656 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1657 	 * will skip over it.
1658 	 */
1659 	VI_LOCK(vp);
1660 	if (vp->v_usecount) {
1661 		VOP_UNLOCK(vp);
1662 		vdropl_recycle(vp);
1663 		vn_finished_write(vnmp);
1664 		CTR2(KTR_VFS,
1665 		    "%s: impossible to recycle, %p is already referenced",
1666 		    __func__, vp);
1667 		return (EBUSY);
1668 	}
1669 	if (!VN_IS_DOOMED(vp)) {
1670 		counter_u64_add(recycles_free_count, 1);
1671 		vgonel(vp);
1672 	}
1673 	VOP_UNLOCK(vp);
1674 	vdropl_recycle(vp);
1675 	vn_finished_write(vnmp);
1676 	return (0);
1677 }
1678 
1679 /*
1680  * Allocate a new vnode.
1681  *
1682  * The operation never returns an error. Returning an error was disabled
1683  * in r145385 (dated 2005) with the following comment:
1684  *
1685  * XXX Not all VFS_VGET/ffs_vget callers check returns.
1686  *
1687  * Given the age of this commit (almost 15 years at the time of writing this
1688  * comment) restoring the ability to fail requires a significant audit of
1689  * all codepaths.
1690  *
1691  * The routine can try to free a vnode or stall for up to 1 second waiting for
1692  * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1693  */
1694 static u_long vn_alloc_cyclecount;
1695 
1696 static struct vnode * __noinline
1697 vn_alloc_hard(struct mount *mp)
1698 {
1699 	u_long rnumvnodes, rfreevnodes;
1700 
1701 	mtx_lock(&vnode_list_mtx);
1702 	rnumvnodes = atomic_load_long(&numvnodes);
1703 	if (rnumvnodes + 1 < desiredvnodes) {
1704 		vn_alloc_cyclecount = 0;
1705 		goto alloc;
1706 	}
1707 	rfreevnodes = vnlru_read_freevnodes();
1708 	if (vn_alloc_cyclecount++ >= rfreevnodes) {
1709 		vn_alloc_cyclecount = 0;
1710 		vstir = 1;
1711 	}
1712 	/*
1713 	 * Grow the vnode cache if it will not be above its target max
1714 	 * after growing.  Otherwise, if the free list is nonempty, try
1715 	 * to reclaim 1 item from it before growing the cache (possibly
1716 	 * above its target max if the reclamation failed or is delayed).
1717 	 * Otherwise, wait for some space.  In all cases, schedule
1718 	 * vnlru_proc() if we are getting short of space.  The watermarks
1719 	 * should be chosen so that we never wait or even reclaim from
1720 	 * the free list to below its target minimum.
1721 	 */
1722 	if (vnlru_free_locked(1) > 0)
1723 		goto alloc;
1724 	if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1725 		/*
1726 		 * Wait for space for a new vnode.
1727 		 */
1728 		vnlru_kick();
1729 		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1730 		if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1731 		    vnlru_read_freevnodes() > 1)
1732 			vnlru_free_locked(1);
1733 	}
1734 alloc:
1735 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1736 	if (vnlru_under(rnumvnodes, vlowat))
1737 		vnlru_kick();
1738 	mtx_unlock(&vnode_list_mtx);
1739 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1740 }
1741 
1742 static struct vnode *
1743 vn_alloc(struct mount *mp)
1744 {
1745 	u_long rnumvnodes;
1746 
1747 	if (__predict_false(vn_alloc_cyclecount != 0))
1748 		return (vn_alloc_hard(mp));
1749 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1750 	if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1751 		atomic_subtract_long(&numvnodes, 1);
1752 		return (vn_alloc_hard(mp));
1753 	}
1754 
1755 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1756 }
1757 
1758 static void
1759 vn_free(struct vnode *vp)
1760 {
1761 
1762 	atomic_subtract_long(&numvnodes, 1);
1763 	uma_zfree_smr(vnode_zone, vp);
1764 }
1765 
1766 /*
1767  * Return the next vnode from the free list.
1768  */
1769 int
1770 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1771     struct vnode **vpp)
1772 {
1773 	struct vnode *vp;
1774 	struct thread *td;
1775 	struct lock_object *lo;
1776 
1777 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1778 
1779 	KASSERT(vops->registered,
1780 	    ("%s: not registered vector op %p\n", __func__, vops));
1781 
1782 	td = curthread;
1783 	if (td->td_vp_reserved != NULL) {
1784 		vp = td->td_vp_reserved;
1785 		td->td_vp_reserved = NULL;
1786 	} else {
1787 		vp = vn_alloc(mp);
1788 	}
1789 	counter_u64_add(vnodes_created, 1);
1790 	/*
1791 	 * Locks are given the generic name "vnode" when created.
1792 	 * Follow the historic practice of using the filesystem
1793 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1794 	 *
1795 	 * Locks live in a witness group keyed on their name. Thus,
1796 	 * when a lock is renamed, it must also move from the witness
1797 	 * group of its old name to the witness group of its new name.
1798 	 *
1799 	 * The change only needs to be made when the vnode moves
1800 	 * from one filesystem type to another. We ensure that each
1801 	 * filesystem use a single static name pointer for its tag so
1802 	 * that we can compare pointers rather than doing a strcmp().
1803 	 */
1804 	lo = &vp->v_vnlock->lock_object;
1805 #ifdef WITNESS
1806 	if (lo->lo_name != tag) {
1807 #endif
1808 		lo->lo_name = tag;
1809 #ifdef WITNESS
1810 		WITNESS_DESTROY(lo);
1811 		WITNESS_INIT(lo, tag);
1812 	}
1813 #endif
1814 	/*
1815 	 * By default, don't allow shared locks unless filesystems opt-in.
1816 	 */
1817 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1818 	/*
1819 	 * Finalize various vnode identity bits.
1820 	 */
1821 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1822 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1823 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1824 	vp->v_type = VNON;
1825 	vp->v_op = vops;
1826 	vp->v_irflag = 0;
1827 	v_init_counters(vp);
1828 	vn_seqc_init(vp);
1829 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1830 #ifdef DIAGNOSTIC
1831 	if (mp == NULL && vops != &dead_vnodeops)
1832 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1833 #endif
1834 #ifdef MAC
1835 	mac_vnode_init(vp);
1836 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1837 		mac_vnode_associate_singlelabel(mp, vp);
1838 #endif
1839 	if (mp != NULL) {
1840 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1841 	}
1842 
1843 	/*
1844 	 * For the filesystems which do not use vfs_hash_insert(),
1845 	 * still initialize v_hash to have vfs_hash_index() useful.
1846 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1847 	 * its own hashing.
1848 	 */
1849 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1850 
1851 	*vpp = vp;
1852 	return (0);
1853 }
1854 
1855 void
1856 getnewvnode_reserve(void)
1857 {
1858 	struct thread *td;
1859 
1860 	td = curthread;
1861 	MPASS(td->td_vp_reserved == NULL);
1862 	td->td_vp_reserved = vn_alloc(NULL);
1863 }
1864 
1865 void
1866 getnewvnode_drop_reserve(void)
1867 {
1868 	struct thread *td;
1869 
1870 	td = curthread;
1871 	if (td->td_vp_reserved != NULL) {
1872 		vn_free(td->td_vp_reserved);
1873 		td->td_vp_reserved = NULL;
1874 	}
1875 }
1876 
1877 static void __noinline
1878 freevnode(struct vnode *vp)
1879 {
1880 	struct bufobj *bo;
1881 
1882 	/*
1883 	 * The vnode has been marked for destruction, so free it.
1884 	 *
1885 	 * The vnode will be returned to the zone where it will
1886 	 * normally remain until it is needed for another vnode. We
1887 	 * need to cleanup (or verify that the cleanup has already
1888 	 * been done) any residual data left from its current use
1889 	 * so as not to contaminate the freshly allocated vnode.
1890 	 */
1891 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1892 	/*
1893 	 * Paired with vgone.
1894 	 */
1895 	vn_seqc_write_end_free(vp);
1896 
1897 	bo = &vp->v_bufobj;
1898 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1899 	VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp);
1900 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1901 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1902 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1903 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1904 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1905 	    ("clean blk trie not empty"));
1906 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1907 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1908 	    ("dirty blk trie not empty"));
1909 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1910 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1911 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1912 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1913 	    ("Dangling rangelock waiters"));
1914 	VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp,
1915 	    ("Leaked inactivation"));
1916 	VI_UNLOCK(vp);
1917 #ifdef MAC
1918 	mac_vnode_destroy(vp);
1919 #endif
1920 	if (vp->v_pollinfo != NULL) {
1921 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1922 		destroy_vpollinfo(vp->v_pollinfo);
1923 		VOP_UNLOCK(vp);
1924 		vp->v_pollinfo = NULL;
1925 	}
1926 	vp->v_mountedhere = NULL;
1927 	vp->v_unpcb = NULL;
1928 	vp->v_rdev = NULL;
1929 	vp->v_fifoinfo = NULL;
1930 	vp->v_iflag = 0;
1931 	vp->v_vflag = 0;
1932 	bo->bo_flag = 0;
1933 	vn_free(vp);
1934 }
1935 
1936 /*
1937  * Delete from old mount point vnode list, if on one.
1938  */
1939 static void
1940 delmntque(struct vnode *vp)
1941 {
1942 	struct mount *mp;
1943 
1944 	VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
1945 
1946 	mp = vp->v_mount;
1947 	MNT_ILOCK(mp);
1948 	VI_LOCK(vp);
1949 	vp->v_mount = NULL;
1950 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1951 		("bad mount point vnode list size"));
1952 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1953 	mp->mnt_nvnodelistsize--;
1954 	MNT_REL(mp);
1955 	MNT_IUNLOCK(mp);
1956 	/*
1957 	 * The caller expects the interlock to be still held.
1958 	 */
1959 	ASSERT_VI_LOCKED(vp, __func__);
1960 }
1961 
1962 static int
1963 insmntque1_int(struct vnode *vp, struct mount *mp, bool dtr)
1964 {
1965 
1966 	KASSERT(vp->v_mount == NULL,
1967 		("insmntque: vnode already on per mount vnode list"));
1968 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1969 	if ((mp->mnt_kern_flag & MNTK_UNLOCKED_INSMNTQUE) == 0) {
1970 		ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1971 	} else {
1972 		KASSERT(!dtr,
1973 		    ("%s: can't have MNTK_UNLOCKED_INSMNTQUE and cleanup",
1974 		    __func__));
1975 	}
1976 
1977 	/*
1978 	 * We acquire the vnode interlock early to ensure that the
1979 	 * vnode cannot be recycled by another process releasing a
1980 	 * holdcnt on it before we get it on both the vnode list
1981 	 * and the active vnode list. The mount mutex protects only
1982 	 * manipulation of the vnode list and the vnode freelist
1983 	 * mutex protects only manipulation of the active vnode list.
1984 	 * Hence the need to hold the vnode interlock throughout.
1985 	 */
1986 	MNT_ILOCK(mp);
1987 	VI_LOCK(vp);
1988 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1989 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1990 	    mp->mnt_nvnodelistsize == 0)) &&
1991 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1992 		VI_UNLOCK(vp);
1993 		MNT_IUNLOCK(mp);
1994 		if (dtr) {
1995 			vp->v_data = NULL;
1996 			vp->v_op = &dead_vnodeops;
1997 			vgone(vp);
1998 			vput(vp);
1999 		}
2000 		return (EBUSY);
2001 	}
2002 	vp->v_mount = mp;
2003 	MNT_REF(mp);
2004 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2005 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
2006 		("neg mount point vnode list size"));
2007 	mp->mnt_nvnodelistsize++;
2008 	VI_UNLOCK(vp);
2009 	MNT_IUNLOCK(mp);
2010 	return (0);
2011 }
2012 
2013 /*
2014  * Insert into list of vnodes for the new mount point, if available.
2015  * insmntque() reclaims the vnode on insertion failure, insmntque1()
2016  * leaves handling of the vnode to the caller.
2017  */
2018 int
2019 insmntque(struct vnode *vp, struct mount *mp)
2020 {
2021 	return (insmntque1_int(vp, mp, true));
2022 }
2023 
2024 int
2025 insmntque1(struct vnode *vp, struct mount *mp)
2026 {
2027 	return (insmntque1_int(vp, mp, false));
2028 }
2029 
2030 /*
2031  * Flush out and invalidate all buffers associated with a bufobj
2032  * Called with the underlying object locked.
2033  */
2034 int
2035 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
2036 {
2037 	int error;
2038 
2039 	BO_LOCK(bo);
2040 	if (flags & V_SAVE) {
2041 		error = bufobj_wwait(bo, slpflag, slptimeo);
2042 		if (error) {
2043 			BO_UNLOCK(bo);
2044 			return (error);
2045 		}
2046 		if (bo->bo_dirty.bv_cnt > 0) {
2047 			BO_UNLOCK(bo);
2048 			do {
2049 				error = BO_SYNC(bo, MNT_WAIT);
2050 			} while (error == ERELOOKUP);
2051 			if (error != 0)
2052 				return (error);
2053 			BO_LOCK(bo);
2054 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) {
2055 				BO_UNLOCK(bo);
2056 				return (EBUSY);
2057 			}
2058 		}
2059 	}
2060 	/*
2061 	 * If you alter this loop please notice that interlock is dropped and
2062 	 * reacquired in flushbuflist.  Special care is needed to ensure that
2063 	 * no race conditions occur from this.
2064 	 */
2065 	do {
2066 		error = flushbuflist(&bo->bo_clean,
2067 		    flags, bo, slpflag, slptimeo);
2068 		if (error == 0 && !(flags & V_CLEANONLY))
2069 			error = flushbuflist(&bo->bo_dirty,
2070 			    flags, bo, slpflag, slptimeo);
2071 		if (error != 0 && error != EAGAIN) {
2072 			BO_UNLOCK(bo);
2073 			return (error);
2074 		}
2075 	} while (error != 0);
2076 
2077 	/*
2078 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
2079 	 * have write I/O in-progress but if there is a VM object then the
2080 	 * VM object can also have read-I/O in-progress.
2081 	 */
2082 	do {
2083 		bufobj_wwait(bo, 0, 0);
2084 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
2085 			BO_UNLOCK(bo);
2086 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
2087 			BO_LOCK(bo);
2088 		}
2089 	} while (bo->bo_numoutput > 0);
2090 	BO_UNLOCK(bo);
2091 
2092 	/*
2093 	 * Destroy the copy in the VM cache, too.
2094 	 */
2095 	if (bo->bo_object != NULL &&
2096 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
2097 		VM_OBJECT_WLOCK(bo->bo_object);
2098 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
2099 		    OBJPR_CLEANONLY : 0);
2100 		VM_OBJECT_WUNLOCK(bo->bo_object);
2101 	}
2102 
2103 #ifdef INVARIANTS
2104 	BO_LOCK(bo);
2105 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
2106 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
2107 	    bo->bo_clean.bv_cnt > 0))
2108 		panic("vinvalbuf: flush failed");
2109 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
2110 	    bo->bo_dirty.bv_cnt > 0)
2111 		panic("vinvalbuf: flush dirty failed");
2112 	BO_UNLOCK(bo);
2113 #endif
2114 	return (0);
2115 }
2116 
2117 /*
2118  * Flush out and invalidate all buffers associated with a vnode.
2119  * Called with the underlying object locked.
2120  */
2121 int
2122 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
2123 {
2124 
2125 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2126 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
2127 	if (vp->v_object != NULL && vp->v_object->handle != vp)
2128 		return (0);
2129 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
2130 }
2131 
2132 /*
2133  * Flush out buffers on the specified list.
2134  *
2135  */
2136 static int
2137 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2138     int slptimeo)
2139 {
2140 	struct buf *bp, *nbp;
2141 	int retval, error;
2142 	daddr_t lblkno;
2143 	b_xflags_t xflags;
2144 
2145 	ASSERT_BO_WLOCKED(bo);
2146 
2147 	retval = 0;
2148 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2149 		/*
2150 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2151 		 * do not skip any buffers. If we are flushing only V_NORMAL
2152 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2153 		 * flushing only V_ALT buffers then skip buffers not marked
2154 		 * as BX_ALTDATA.
2155 		 */
2156 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2157 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2158 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2159 			continue;
2160 		}
2161 		if (nbp != NULL) {
2162 			lblkno = nbp->b_lblkno;
2163 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2164 		}
2165 		retval = EAGAIN;
2166 		error = BUF_TIMELOCK(bp,
2167 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2168 		    "flushbuf", slpflag, slptimeo);
2169 		if (error) {
2170 			BO_LOCK(bo);
2171 			return (error != ENOLCK ? error : EAGAIN);
2172 		}
2173 		KASSERT(bp->b_bufobj == bo,
2174 		    ("bp %p wrong b_bufobj %p should be %p",
2175 		    bp, bp->b_bufobj, bo));
2176 		/*
2177 		 * XXX Since there are no node locks for NFS, I
2178 		 * believe there is a slight chance that a delayed
2179 		 * write will occur while sleeping just above, so
2180 		 * check for it.
2181 		 */
2182 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2183 		    (flags & V_SAVE)) {
2184 			bremfree(bp);
2185 			bp->b_flags |= B_ASYNC;
2186 			bwrite(bp);
2187 			BO_LOCK(bo);
2188 			return (EAGAIN);	/* XXX: why not loop ? */
2189 		}
2190 		bremfree(bp);
2191 		bp->b_flags |= (B_INVAL | B_RELBUF);
2192 		bp->b_flags &= ~B_ASYNC;
2193 		brelse(bp);
2194 		BO_LOCK(bo);
2195 		if (nbp == NULL)
2196 			break;
2197 		nbp = gbincore(bo, lblkno);
2198 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2199 		    != xflags)
2200 			break;			/* nbp invalid */
2201 	}
2202 	return (retval);
2203 }
2204 
2205 int
2206 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2207 {
2208 	struct buf *bp;
2209 	int error;
2210 	daddr_t lblkno;
2211 
2212 	ASSERT_BO_LOCKED(bo);
2213 
2214 	for (lblkno = startn;;) {
2215 again:
2216 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2217 		if (bp == NULL || bp->b_lblkno >= endn ||
2218 		    bp->b_lblkno < startn)
2219 			break;
2220 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2221 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2222 		if (error != 0) {
2223 			BO_RLOCK(bo);
2224 			if (error == ENOLCK)
2225 				goto again;
2226 			return (error);
2227 		}
2228 		KASSERT(bp->b_bufobj == bo,
2229 		    ("bp %p wrong b_bufobj %p should be %p",
2230 		    bp, bp->b_bufobj, bo));
2231 		lblkno = bp->b_lblkno + 1;
2232 		if ((bp->b_flags & B_MANAGED) == 0)
2233 			bremfree(bp);
2234 		bp->b_flags |= B_RELBUF;
2235 		/*
2236 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2237 		 * pages backing each buffer in the range are unlikely to be
2238 		 * reused.  Dirty buffers will have the hint applied once
2239 		 * they've been written.
2240 		 */
2241 		if ((bp->b_flags & B_VMIO) != 0)
2242 			bp->b_flags |= B_NOREUSE;
2243 		brelse(bp);
2244 		BO_RLOCK(bo);
2245 	}
2246 	return (0);
2247 }
2248 
2249 /*
2250  * Truncate a file's buffer and pages to a specified length.  This
2251  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2252  * sync activity.
2253  */
2254 int
2255 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2256 {
2257 	struct buf *bp, *nbp;
2258 	struct bufobj *bo;
2259 	daddr_t startlbn;
2260 
2261 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2262 	    vp, blksize, (uintmax_t)length);
2263 
2264 	/*
2265 	 * Round up to the *next* lbn.
2266 	 */
2267 	startlbn = howmany(length, blksize);
2268 
2269 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2270 
2271 	bo = &vp->v_bufobj;
2272 restart_unlocked:
2273 	BO_LOCK(bo);
2274 
2275 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2276 		;
2277 
2278 	if (length > 0) {
2279 restartsync:
2280 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2281 			if (bp->b_lblkno > 0)
2282 				continue;
2283 			/*
2284 			 * Since we hold the vnode lock this should only
2285 			 * fail if we're racing with the buf daemon.
2286 			 */
2287 			if (BUF_LOCK(bp,
2288 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2289 			    BO_LOCKPTR(bo)) == ENOLCK)
2290 				goto restart_unlocked;
2291 
2292 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2293 			    ("buf(%p) on dirty queue without DELWRI", bp));
2294 
2295 			bremfree(bp);
2296 			bawrite(bp);
2297 			BO_LOCK(bo);
2298 			goto restartsync;
2299 		}
2300 	}
2301 
2302 	bufobj_wwait(bo, 0, 0);
2303 	BO_UNLOCK(bo);
2304 	vnode_pager_setsize(vp, length);
2305 
2306 	return (0);
2307 }
2308 
2309 /*
2310  * Invalidate the cached pages of a file's buffer within the range of block
2311  * numbers [startlbn, endlbn).
2312  */
2313 void
2314 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2315     int blksize)
2316 {
2317 	struct bufobj *bo;
2318 	off_t start, end;
2319 
2320 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2321 
2322 	start = blksize * startlbn;
2323 	end = blksize * endlbn;
2324 
2325 	bo = &vp->v_bufobj;
2326 	BO_LOCK(bo);
2327 	MPASS(blksize == bo->bo_bsize);
2328 
2329 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2330 		;
2331 
2332 	BO_UNLOCK(bo);
2333 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2334 }
2335 
2336 static int
2337 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2338     daddr_t startlbn, daddr_t endlbn)
2339 {
2340 	struct buf *bp, *nbp;
2341 	bool anyfreed;
2342 
2343 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2344 	ASSERT_BO_LOCKED(bo);
2345 
2346 	do {
2347 		anyfreed = false;
2348 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2349 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2350 				continue;
2351 			if (BUF_LOCK(bp,
2352 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2353 			    BO_LOCKPTR(bo)) == ENOLCK) {
2354 				BO_LOCK(bo);
2355 				return (EAGAIN);
2356 			}
2357 
2358 			bremfree(bp);
2359 			bp->b_flags |= B_INVAL | B_RELBUF;
2360 			bp->b_flags &= ~B_ASYNC;
2361 			brelse(bp);
2362 			anyfreed = true;
2363 
2364 			BO_LOCK(bo);
2365 			if (nbp != NULL &&
2366 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2367 			    nbp->b_vp != vp ||
2368 			    (nbp->b_flags & B_DELWRI) != 0))
2369 				return (EAGAIN);
2370 		}
2371 
2372 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2373 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2374 				continue;
2375 			if (BUF_LOCK(bp,
2376 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2377 			    BO_LOCKPTR(bo)) == ENOLCK) {
2378 				BO_LOCK(bo);
2379 				return (EAGAIN);
2380 			}
2381 			bremfree(bp);
2382 			bp->b_flags |= B_INVAL | B_RELBUF;
2383 			bp->b_flags &= ~B_ASYNC;
2384 			brelse(bp);
2385 			anyfreed = true;
2386 
2387 			BO_LOCK(bo);
2388 			if (nbp != NULL &&
2389 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2390 			    (nbp->b_vp != vp) ||
2391 			    (nbp->b_flags & B_DELWRI) == 0))
2392 				return (EAGAIN);
2393 		}
2394 	} while (anyfreed);
2395 	return (0);
2396 }
2397 
2398 static void
2399 buf_vlist_remove(struct buf *bp)
2400 {
2401 	struct bufv *bv;
2402 	b_xflags_t flags;
2403 
2404 	flags = bp->b_xflags;
2405 
2406 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2407 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2408 	KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 &&
2409 	    (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN),
2410 	    ("%s: buffer %p has invalid queue state", __func__, bp));
2411 
2412 	if ((flags & BX_VNDIRTY) != 0)
2413 		bv = &bp->b_bufobj->bo_dirty;
2414 	else
2415 		bv = &bp->b_bufobj->bo_clean;
2416 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2417 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2418 	bv->bv_cnt--;
2419 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2420 }
2421 
2422 /*
2423  * Add the buffer to the sorted clean or dirty block list.
2424  *
2425  * NOTE: xflags is passed as a constant, optimizing this inline function!
2426  */
2427 static void
2428 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2429 {
2430 	struct bufv *bv;
2431 	struct buf *n;
2432 	int error;
2433 
2434 	ASSERT_BO_WLOCKED(bo);
2435 	KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2436 	    ("buf_vlist_add: bo %p does not allow bufs", bo));
2437 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2438 	    ("dead bo %p", bo));
2439 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2440 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2441 	bp->b_xflags |= xflags;
2442 	if (xflags & BX_VNDIRTY)
2443 		bv = &bo->bo_dirty;
2444 	else
2445 		bv = &bo->bo_clean;
2446 
2447 	/*
2448 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2449 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2450 	 * than _ge.
2451 	 */
2452 	if (bv->bv_cnt == 0 ||
2453 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2454 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2455 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2456 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2457 	else
2458 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2459 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2460 	if (error)
2461 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2462 	bv->bv_cnt++;
2463 }
2464 
2465 /*
2466  * Look up a buffer using the buffer tries.
2467  */
2468 struct buf *
2469 gbincore(struct bufobj *bo, daddr_t lblkno)
2470 {
2471 	struct buf *bp;
2472 
2473 	ASSERT_BO_LOCKED(bo);
2474 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2475 	if (bp != NULL)
2476 		return (bp);
2477 	return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno));
2478 }
2479 
2480 /*
2481  * Look up a buf using the buffer tries, without the bufobj lock.  This relies
2482  * on SMR for safe lookup, and bufs being in a no-free zone to provide type
2483  * stability of the result.  Like other lockless lookups, the found buf may
2484  * already be invalid by the time this function returns.
2485  */
2486 struct buf *
2487 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno)
2488 {
2489 	struct buf *bp;
2490 
2491 	ASSERT_BO_UNLOCKED(bo);
2492 	bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno);
2493 	if (bp != NULL)
2494 		return (bp);
2495 	return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno));
2496 }
2497 
2498 /*
2499  * Associate a buffer with a vnode.
2500  */
2501 void
2502 bgetvp(struct vnode *vp, struct buf *bp)
2503 {
2504 	struct bufobj *bo;
2505 
2506 	bo = &vp->v_bufobj;
2507 	ASSERT_BO_WLOCKED(bo);
2508 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2509 
2510 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2511 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2512 	    ("bgetvp: bp already attached! %p", bp));
2513 
2514 	vhold(vp);
2515 	bp->b_vp = vp;
2516 	bp->b_bufobj = bo;
2517 	/*
2518 	 * Insert onto list for new vnode.
2519 	 */
2520 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2521 }
2522 
2523 /*
2524  * Disassociate a buffer from a vnode.
2525  */
2526 void
2527 brelvp(struct buf *bp)
2528 {
2529 	struct bufobj *bo;
2530 	struct vnode *vp;
2531 
2532 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2533 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2534 
2535 	/*
2536 	 * Delete from old vnode list, if on one.
2537 	 */
2538 	vp = bp->b_vp;		/* XXX */
2539 	bo = bp->b_bufobj;
2540 	BO_LOCK(bo);
2541 	buf_vlist_remove(bp);
2542 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2543 		bo->bo_flag &= ~BO_ONWORKLST;
2544 		mtx_lock(&sync_mtx);
2545 		LIST_REMOVE(bo, bo_synclist);
2546 		syncer_worklist_len--;
2547 		mtx_unlock(&sync_mtx);
2548 	}
2549 	bp->b_vp = NULL;
2550 	bp->b_bufobj = NULL;
2551 	BO_UNLOCK(bo);
2552 	vdrop(vp);
2553 }
2554 
2555 /*
2556  * Add an item to the syncer work queue.
2557  */
2558 static void
2559 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2560 {
2561 	int slot;
2562 
2563 	ASSERT_BO_WLOCKED(bo);
2564 
2565 	mtx_lock(&sync_mtx);
2566 	if (bo->bo_flag & BO_ONWORKLST)
2567 		LIST_REMOVE(bo, bo_synclist);
2568 	else {
2569 		bo->bo_flag |= BO_ONWORKLST;
2570 		syncer_worklist_len++;
2571 	}
2572 
2573 	if (delay > syncer_maxdelay - 2)
2574 		delay = syncer_maxdelay - 2;
2575 	slot = (syncer_delayno + delay) & syncer_mask;
2576 
2577 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2578 	mtx_unlock(&sync_mtx);
2579 }
2580 
2581 static int
2582 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2583 {
2584 	int error, len;
2585 
2586 	mtx_lock(&sync_mtx);
2587 	len = syncer_worklist_len - sync_vnode_count;
2588 	mtx_unlock(&sync_mtx);
2589 	error = SYSCTL_OUT(req, &len, sizeof(len));
2590 	return (error);
2591 }
2592 
2593 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2594     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2595     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2596 
2597 static struct proc *updateproc;
2598 static void sched_sync(void);
2599 static struct kproc_desc up_kp = {
2600 	"syncer",
2601 	sched_sync,
2602 	&updateproc
2603 };
2604 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2605 
2606 static int
2607 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2608 {
2609 	struct vnode *vp;
2610 	struct mount *mp;
2611 
2612 	*bo = LIST_FIRST(slp);
2613 	if (*bo == NULL)
2614 		return (0);
2615 	vp = bo2vnode(*bo);
2616 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2617 		return (1);
2618 	/*
2619 	 * We use vhold in case the vnode does not
2620 	 * successfully sync.  vhold prevents the vnode from
2621 	 * going away when we unlock the sync_mtx so that
2622 	 * we can acquire the vnode interlock.
2623 	 */
2624 	vholdl(vp);
2625 	mtx_unlock(&sync_mtx);
2626 	VI_UNLOCK(vp);
2627 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2628 		vdrop(vp);
2629 		mtx_lock(&sync_mtx);
2630 		return (*bo == LIST_FIRST(slp));
2631 	}
2632 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2633 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2634 	VOP_UNLOCK(vp);
2635 	vn_finished_write(mp);
2636 	BO_LOCK(*bo);
2637 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2638 		/*
2639 		 * Put us back on the worklist.  The worklist
2640 		 * routine will remove us from our current
2641 		 * position and then add us back in at a later
2642 		 * position.
2643 		 */
2644 		vn_syncer_add_to_worklist(*bo, syncdelay);
2645 	}
2646 	BO_UNLOCK(*bo);
2647 	vdrop(vp);
2648 	mtx_lock(&sync_mtx);
2649 	return (0);
2650 }
2651 
2652 static int first_printf = 1;
2653 
2654 /*
2655  * System filesystem synchronizer daemon.
2656  */
2657 static void
2658 sched_sync(void)
2659 {
2660 	struct synclist *next, *slp;
2661 	struct bufobj *bo;
2662 	long starttime;
2663 	struct thread *td = curthread;
2664 	int last_work_seen;
2665 	int net_worklist_len;
2666 	int syncer_final_iter;
2667 	int error;
2668 
2669 	last_work_seen = 0;
2670 	syncer_final_iter = 0;
2671 	syncer_state = SYNCER_RUNNING;
2672 	starttime = time_uptime;
2673 	td->td_pflags |= TDP_NORUNNINGBUF;
2674 
2675 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2676 	    SHUTDOWN_PRI_LAST);
2677 
2678 	mtx_lock(&sync_mtx);
2679 	for (;;) {
2680 		if (syncer_state == SYNCER_FINAL_DELAY &&
2681 		    syncer_final_iter == 0) {
2682 			mtx_unlock(&sync_mtx);
2683 			kproc_suspend_check(td->td_proc);
2684 			mtx_lock(&sync_mtx);
2685 		}
2686 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2687 		if (syncer_state != SYNCER_RUNNING &&
2688 		    starttime != time_uptime) {
2689 			if (first_printf) {
2690 				printf("\nSyncing disks, vnodes remaining... ");
2691 				first_printf = 0;
2692 			}
2693 			printf("%d ", net_worklist_len);
2694 		}
2695 		starttime = time_uptime;
2696 
2697 		/*
2698 		 * Push files whose dirty time has expired.  Be careful
2699 		 * of interrupt race on slp queue.
2700 		 *
2701 		 * Skip over empty worklist slots when shutting down.
2702 		 */
2703 		do {
2704 			slp = &syncer_workitem_pending[syncer_delayno];
2705 			syncer_delayno += 1;
2706 			if (syncer_delayno == syncer_maxdelay)
2707 				syncer_delayno = 0;
2708 			next = &syncer_workitem_pending[syncer_delayno];
2709 			/*
2710 			 * If the worklist has wrapped since the
2711 			 * it was emptied of all but syncer vnodes,
2712 			 * switch to the FINAL_DELAY state and run
2713 			 * for one more second.
2714 			 */
2715 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2716 			    net_worklist_len == 0 &&
2717 			    last_work_seen == syncer_delayno) {
2718 				syncer_state = SYNCER_FINAL_DELAY;
2719 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2720 			}
2721 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2722 		    syncer_worklist_len > 0);
2723 
2724 		/*
2725 		 * Keep track of the last time there was anything
2726 		 * on the worklist other than syncer vnodes.
2727 		 * Return to the SHUTTING_DOWN state if any
2728 		 * new work appears.
2729 		 */
2730 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2731 			last_work_seen = syncer_delayno;
2732 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2733 			syncer_state = SYNCER_SHUTTING_DOWN;
2734 		while (!LIST_EMPTY(slp)) {
2735 			error = sync_vnode(slp, &bo, td);
2736 			if (error == 1) {
2737 				LIST_REMOVE(bo, bo_synclist);
2738 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2739 				continue;
2740 			}
2741 
2742 			if (first_printf == 0) {
2743 				/*
2744 				 * Drop the sync mutex, because some watchdog
2745 				 * drivers need to sleep while patting
2746 				 */
2747 				mtx_unlock(&sync_mtx);
2748 				wdog_kern_pat(WD_LASTVAL);
2749 				mtx_lock(&sync_mtx);
2750 			}
2751 		}
2752 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2753 			syncer_final_iter--;
2754 		/*
2755 		 * The variable rushjob allows the kernel to speed up the
2756 		 * processing of the filesystem syncer process. A rushjob
2757 		 * value of N tells the filesystem syncer to process the next
2758 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2759 		 * is used by the soft update code to speed up the filesystem
2760 		 * syncer process when the incore state is getting so far
2761 		 * ahead of the disk that the kernel memory pool is being
2762 		 * threatened with exhaustion.
2763 		 */
2764 		if (rushjob > 0) {
2765 			rushjob -= 1;
2766 			continue;
2767 		}
2768 		/*
2769 		 * Just sleep for a short period of time between
2770 		 * iterations when shutting down to allow some I/O
2771 		 * to happen.
2772 		 *
2773 		 * If it has taken us less than a second to process the
2774 		 * current work, then wait. Otherwise start right over
2775 		 * again. We can still lose time if any single round
2776 		 * takes more than two seconds, but it does not really
2777 		 * matter as we are just trying to generally pace the
2778 		 * filesystem activity.
2779 		 */
2780 		if (syncer_state != SYNCER_RUNNING ||
2781 		    time_uptime == starttime) {
2782 			thread_lock(td);
2783 			sched_prio(td, PPAUSE);
2784 			thread_unlock(td);
2785 		}
2786 		if (syncer_state != SYNCER_RUNNING)
2787 			cv_timedwait(&sync_wakeup, &sync_mtx,
2788 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2789 		else if (time_uptime == starttime)
2790 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2791 	}
2792 }
2793 
2794 /*
2795  * Request the syncer daemon to speed up its work.
2796  * We never push it to speed up more than half of its
2797  * normal turn time, otherwise it could take over the cpu.
2798  */
2799 int
2800 speedup_syncer(void)
2801 {
2802 	int ret = 0;
2803 
2804 	mtx_lock(&sync_mtx);
2805 	if (rushjob < syncdelay / 2) {
2806 		rushjob += 1;
2807 		stat_rush_requests += 1;
2808 		ret = 1;
2809 	}
2810 	mtx_unlock(&sync_mtx);
2811 	cv_broadcast(&sync_wakeup);
2812 	return (ret);
2813 }
2814 
2815 /*
2816  * Tell the syncer to speed up its work and run though its work
2817  * list several times, then tell it to shut down.
2818  */
2819 static void
2820 syncer_shutdown(void *arg, int howto)
2821 {
2822 
2823 	if (howto & RB_NOSYNC)
2824 		return;
2825 	mtx_lock(&sync_mtx);
2826 	syncer_state = SYNCER_SHUTTING_DOWN;
2827 	rushjob = 0;
2828 	mtx_unlock(&sync_mtx);
2829 	cv_broadcast(&sync_wakeup);
2830 	kproc_shutdown(arg, howto);
2831 }
2832 
2833 void
2834 syncer_suspend(void)
2835 {
2836 
2837 	syncer_shutdown(updateproc, 0);
2838 }
2839 
2840 void
2841 syncer_resume(void)
2842 {
2843 
2844 	mtx_lock(&sync_mtx);
2845 	first_printf = 1;
2846 	syncer_state = SYNCER_RUNNING;
2847 	mtx_unlock(&sync_mtx);
2848 	cv_broadcast(&sync_wakeup);
2849 	kproc_resume(updateproc);
2850 }
2851 
2852 /*
2853  * Move the buffer between the clean and dirty lists of its vnode.
2854  */
2855 void
2856 reassignbuf(struct buf *bp)
2857 {
2858 	struct vnode *vp;
2859 	struct bufobj *bo;
2860 	int delay;
2861 #ifdef INVARIANTS
2862 	struct bufv *bv;
2863 #endif
2864 
2865 	vp = bp->b_vp;
2866 	bo = bp->b_bufobj;
2867 
2868 	KASSERT((bp->b_flags & B_PAGING) == 0,
2869 	    ("%s: cannot reassign paging buffer %p", __func__, bp));
2870 
2871 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2872 	    bp, bp->b_vp, bp->b_flags);
2873 
2874 	BO_LOCK(bo);
2875 	buf_vlist_remove(bp);
2876 
2877 	/*
2878 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2879 	 * of clean buffers.
2880 	 */
2881 	if (bp->b_flags & B_DELWRI) {
2882 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2883 			switch (vp->v_type) {
2884 			case VDIR:
2885 				delay = dirdelay;
2886 				break;
2887 			case VCHR:
2888 				delay = metadelay;
2889 				break;
2890 			default:
2891 				delay = filedelay;
2892 			}
2893 			vn_syncer_add_to_worklist(bo, delay);
2894 		}
2895 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2896 	} else {
2897 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2898 
2899 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2900 			mtx_lock(&sync_mtx);
2901 			LIST_REMOVE(bo, bo_synclist);
2902 			syncer_worklist_len--;
2903 			mtx_unlock(&sync_mtx);
2904 			bo->bo_flag &= ~BO_ONWORKLST;
2905 		}
2906 	}
2907 #ifdef INVARIANTS
2908 	bv = &bo->bo_clean;
2909 	bp = TAILQ_FIRST(&bv->bv_hd);
2910 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2911 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2912 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2913 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2914 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2915 	bv = &bo->bo_dirty;
2916 	bp = TAILQ_FIRST(&bv->bv_hd);
2917 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2918 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2919 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2920 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2921 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2922 #endif
2923 	BO_UNLOCK(bo);
2924 }
2925 
2926 static void
2927 v_init_counters(struct vnode *vp)
2928 {
2929 
2930 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2931 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2932 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2933 
2934 	refcount_init(&vp->v_holdcnt, 1);
2935 	refcount_init(&vp->v_usecount, 1);
2936 }
2937 
2938 /*
2939  * Grab a particular vnode from the free list, increment its
2940  * reference count and lock it.  VIRF_DOOMED is set if the vnode
2941  * is being destroyed.  Only callers who specify LK_RETRY will
2942  * see doomed vnodes.  If inactive processing was delayed in
2943  * vput try to do it here.
2944  *
2945  * usecount is manipulated using atomics without holding any locks.
2946  *
2947  * holdcnt can be manipulated using atomics without holding any locks,
2948  * except when transitioning 1<->0, in which case the interlock is held.
2949  *
2950  * Consumers which don't guarantee liveness of the vnode can use SMR to
2951  * try to get a reference. Note this operation can fail since the vnode
2952  * may be awaiting getting freed by the time they get to it.
2953  */
2954 enum vgetstate
2955 vget_prep_smr(struct vnode *vp)
2956 {
2957 	enum vgetstate vs;
2958 
2959 	VFS_SMR_ASSERT_ENTERED();
2960 
2961 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2962 		vs = VGET_USECOUNT;
2963 	} else {
2964 		if (vhold_smr(vp))
2965 			vs = VGET_HOLDCNT;
2966 		else
2967 			vs = VGET_NONE;
2968 	}
2969 	return (vs);
2970 }
2971 
2972 enum vgetstate
2973 vget_prep(struct vnode *vp)
2974 {
2975 	enum vgetstate vs;
2976 
2977 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2978 		vs = VGET_USECOUNT;
2979 	} else {
2980 		vhold(vp);
2981 		vs = VGET_HOLDCNT;
2982 	}
2983 	return (vs);
2984 }
2985 
2986 void
2987 vget_abort(struct vnode *vp, enum vgetstate vs)
2988 {
2989 
2990 	switch (vs) {
2991 	case VGET_USECOUNT:
2992 		vrele(vp);
2993 		break;
2994 	case VGET_HOLDCNT:
2995 		vdrop(vp);
2996 		break;
2997 	default:
2998 		__assert_unreachable();
2999 	}
3000 }
3001 
3002 int
3003 vget(struct vnode *vp, int flags)
3004 {
3005 	enum vgetstate vs;
3006 
3007 	vs = vget_prep(vp);
3008 	return (vget_finish(vp, flags, vs));
3009 }
3010 
3011 int
3012 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
3013 {
3014 	int error;
3015 
3016 	if ((flags & LK_INTERLOCK) != 0)
3017 		ASSERT_VI_LOCKED(vp, __func__);
3018 	else
3019 		ASSERT_VI_UNLOCKED(vp, __func__);
3020 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3021 	VNPASS(vp->v_holdcnt > 0, vp);
3022 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3023 
3024 	error = vn_lock(vp, flags);
3025 	if (__predict_false(error != 0)) {
3026 		vget_abort(vp, vs);
3027 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
3028 		    vp);
3029 		return (error);
3030 	}
3031 
3032 	vget_finish_ref(vp, vs);
3033 	return (0);
3034 }
3035 
3036 void
3037 vget_finish_ref(struct vnode *vp, enum vgetstate vs)
3038 {
3039 	int old;
3040 
3041 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3042 	VNPASS(vp->v_holdcnt > 0, vp);
3043 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3044 
3045 	if (vs == VGET_USECOUNT)
3046 		return;
3047 
3048 	/*
3049 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
3050 	 * the vnode around. Otherwise someone else lended their hold count and
3051 	 * we have to drop ours.
3052 	 */
3053 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
3054 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3055 	if (old != 0) {
3056 #ifdef INVARIANTS
3057 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3058 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3059 #else
3060 		refcount_release(&vp->v_holdcnt);
3061 #endif
3062 	}
3063 }
3064 
3065 void
3066 vref(struct vnode *vp)
3067 {
3068 	enum vgetstate vs;
3069 
3070 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3071 	vs = vget_prep(vp);
3072 	vget_finish_ref(vp, vs);
3073 }
3074 
3075 void
3076 vrefact(struct vnode *vp)
3077 {
3078 
3079 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3080 #ifdef INVARIANTS
3081 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3082 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3083 #else
3084 	refcount_acquire(&vp->v_usecount);
3085 #endif
3086 }
3087 
3088 void
3089 vlazy(struct vnode *vp)
3090 {
3091 	struct mount *mp;
3092 
3093 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3094 
3095 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3096 		return;
3097 	/*
3098 	 * We may get here for inactive routines after the vnode got doomed.
3099 	 */
3100 	if (VN_IS_DOOMED(vp))
3101 		return;
3102 	mp = vp->v_mount;
3103 	mtx_lock(&mp->mnt_listmtx);
3104 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3105 		vp->v_mflag |= VMP_LAZYLIST;
3106 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3107 		mp->mnt_lazyvnodelistsize++;
3108 	}
3109 	mtx_unlock(&mp->mnt_listmtx);
3110 }
3111 
3112 static void
3113 vunlazy(struct vnode *vp)
3114 {
3115 	struct mount *mp;
3116 
3117 	ASSERT_VI_LOCKED(vp, __func__);
3118 	VNPASS(!VN_IS_DOOMED(vp), vp);
3119 
3120 	mp = vp->v_mount;
3121 	mtx_lock(&mp->mnt_listmtx);
3122 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3123 	/*
3124 	 * Don't remove the vnode from the lazy list if another thread
3125 	 * has increased the hold count. It may have re-enqueued the
3126 	 * vnode to the lazy list and is now responsible for its
3127 	 * removal.
3128 	 */
3129 	if (vp->v_holdcnt == 0) {
3130 		vp->v_mflag &= ~VMP_LAZYLIST;
3131 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3132 		mp->mnt_lazyvnodelistsize--;
3133 	}
3134 	mtx_unlock(&mp->mnt_listmtx);
3135 }
3136 
3137 /*
3138  * This routine is only meant to be called from vgonel prior to dooming
3139  * the vnode.
3140  */
3141 static void
3142 vunlazy_gone(struct vnode *vp)
3143 {
3144 	struct mount *mp;
3145 
3146 	ASSERT_VOP_ELOCKED(vp, __func__);
3147 	ASSERT_VI_LOCKED(vp, __func__);
3148 	VNPASS(!VN_IS_DOOMED(vp), vp);
3149 
3150 	if (vp->v_mflag & VMP_LAZYLIST) {
3151 		mp = vp->v_mount;
3152 		mtx_lock(&mp->mnt_listmtx);
3153 		VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3154 		vp->v_mflag &= ~VMP_LAZYLIST;
3155 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3156 		mp->mnt_lazyvnodelistsize--;
3157 		mtx_unlock(&mp->mnt_listmtx);
3158 	}
3159 }
3160 
3161 static void
3162 vdefer_inactive(struct vnode *vp)
3163 {
3164 
3165 	ASSERT_VI_LOCKED(vp, __func__);
3166 	VNASSERT(vp->v_holdcnt > 0, vp,
3167 	    ("%s: vnode without hold count", __func__));
3168 	if (VN_IS_DOOMED(vp)) {
3169 		vdropl(vp);
3170 		return;
3171 	}
3172 	if (vp->v_iflag & VI_DEFINACT) {
3173 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3174 		vdropl(vp);
3175 		return;
3176 	}
3177 	if (vp->v_usecount > 0) {
3178 		vp->v_iflag &= ~VI_OWEINACT;
3179 		vdropl(vp);
3180 		return;
3181 	}
3182 	vlazy(vp);
3183 	vp->v_iflag |= VI_DEFINACT;
3184 	VI_UNLOCK(vp);
3185 	counter_u64_add(deferred_inact, 1);
3186 }
3187 
3188 static void
3189 vdefer_inactive_unlocked(struct vnode *vp)
3190 {
3191 
3192 	VI_LOCK(vp);
3193 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3194 		vdropl(vp);
3195 		return;
3196 	}
3197 	vdefer_inactive(vp);
3198 }
3199 
3200 enum vput_op { VRELE, VPUT, VUNREF };
3201 
3202 /*
3203  * Handle ->v_usecount transitioning to 0.
3204  *
3205  * By releasing the last usecount we take ownership of the hold count which
3206  * provides liveness of the vnode, meaning we have to vdrop.
3207  *
3208  * For all vnodes we may need to perform inactive processing. It requires an
3209  * exclusive lock on the vnode, while it is legal to call here with only a
3210  * shared lock (or no locks). If locking the vnode in an expected manner fails,
3211  * inactive processing gets deferred to the syncer.
3212  *
3213  * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3214  * on the lock being held all the way until VOP_INACTIVE. This in particular
3215  * happens with UFS which adds half-constructed vnodes to the hash, where they
3216  * can be found by other code.
3217  */
3218 static void
3219 vput_final(struct vnode *vp, enum vput_op func)
3220 {
3221 	int error;
3222 	bool want_unlock;
3223 
3224 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3225 	VNPASS(vp->v_holdcnt > 0, vp);
3226 
3227 	VI_LOCK(vp);
3228 
3229 	/*
3230 	 * By the time we got here someone else might have transitioned
3231 	 * the count back to > 0.
3232 	 */
3233 	if (vp->v_usecount > 0)
3234 		goto out;
3235 
3236 	/*
3237 	 * If the vnode is doomed vgone already performed inactive processing
3238 	 * (if needed).
3239 	 */
3240 	if (VN_IS_DOOMED(vp))
3241 		goto out;
3242 
3243 	if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3244 		goto out;
3245 
3246 	if (vp->v_iflag & VI_DOINGINACT)
3247 		goto out;
3248 
3249 	/*
3250 	 * Locking operations here will drop the interlock and possibly the
3251 	 * vnode lock, opening a window where the vnode can get doomed all the
3252 	 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3253 	 * perform inactive.
3254 	 */
3255 	vp->v_iflag |= VI_OWEINACT;
3256 	want_unlock = false;
3257 	error = 0;
3258 	switch (func) {
3259 	case VRELE:
3260 		switch (VOP_ISLOCKED(vp)) {
3261 		case LK_EXCLUSIVE:
3262 			break;
3263 		case LK_EXCLOTHER:
3264 		case 0:
3265 			want_unlock = true;
3266 			error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3267 			VI_LOCK(vp);
3268 			break;
3269 		default:
3270 			/*
3271 			 * The lock has at least one sharer, but we have no way
3272 			 * to conclude whether this is us. Play it safe and
3273 			 * defer processing.
3274 			 */
3275 			error = EAGAIN;
3276 			break;
3277 		}
3278 		break;
3279 	case VPUT:
3280 		want_unlock = true;
3281 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3282 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3283 			    LK_NOWAIT);
3284 			VI_LOCK(vp);
3285 		}
3286 		break;
3287 	case VUNREF:
3288 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3289 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3290 			VI_LOCK(vp);
3291 		}
3292 		break;
3293 	}
3294 	if (error == 0) {
3295 		if (func == VUNREF) {
3296 			VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp,
3297 			    ("recursive vunref"));
3298 			vp->v_vflag |= VV_UNREF;
3299 		}
3300 		for (;;) {
3301 			error = vinactive(vp);
3302 			if (want_unlock)
3303 				VOP_UNLOCK(vp);
3304 			if (error != ERELOOKUP || !want_unlock)
3305 				break;
3306 			VOP_LOCK(vp, LK_EXCLUSIVE);
3307 		}
3308 		if (func == VUNREF)
3309 			vp->v_vflag &= ~VV_UNREF;
3310 		vdropl(vp);
3311 	} else {
3312 		vdefer_inactive(vp);
3313 	}
3314 	return;
3315 out:
3316 	if (func == VPUT)
3317 		VOP_UNLOCK(vp);
3318 	vdropl(vp);
3319 }
3320 
3321 /*
3322  * Decrement ->v_usecount for a vnode.
3323  *
3324  * Releasing the last use count requires additional processing, see vput_final
3325  * above for details.
3326  *
3327  * Comment above each variant denotes lock state on entry and exit.
3328  */
3329 
3330 /*
3331  * in: any
3332  * out: same as passed in
3333  */
3334 void
3335 vrele(struct vnode *vp)
3336 {
3337 
3338 	ASSERT_VI_UNLOCKED(vp, __func__);
3339 	if (!refcount_release(&vp->v_usecount))
3340 		return;
3341 	vput_final(vp, VRELE);
3342 }
3343 
3344 /*
3345  * in: locked
3346  * out: unlocked
3347  */
3348 void
3349 vput(struct vnode *vp)
3350 {
3351 
3352 	ASSERT_VOP_LOCKED(vp, __func__);
3353 	ASSERT_VI_UNLOCKED(vp, __func__);
3354 	if (!refcount_release(&vp->v_usecount)) {
3355 		VOP_UNLOCK(vp);
3356 		return;
3357 	}
3358 	vput_final(vp, VPUT);
3359 }
3360 
3361 /*
3362  * in: locked
3363  * out: locked
3364  */
3365 void
3366 vunref(struct vnode *vp)
3367 {
3368 
3369 	ASSERT_VOP_LOCKED(vp, __func__);
3370 	ASSERT_VI_UNLOCKED(vp, __func__);
3371 	if (!refcount_release(&vp->v_usecount))
3372 		return;
3373 	vput_final(vp, VUNREF);
3374 }
3375 
3376 void
3377 vhold(struct vnode *vp)
3378 {
3379 	int old;
3380 
3381 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3382 	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3383 	VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3384 	    ("%s: wrong hold count %d", __func__, old));
3385 	if (old == 0)
3386 		vfs_freevnodes_dec();
3387 }
3388 
3389 void
3390 vholdnz(struct vnode *vp)
3391 {
3392 
3393 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3394 #ifdef INVARIANTS
3395 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3396 	VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3397 	    ("%s: wrong hold count %d", __func__, old));
3398 #else
3399 	atomic_add_int(&vp->v_holdcnt, 1);
3400 #endif
3401 }
3402 
3403 /*
3404  * Grab a hold count unless the vnode is freed.
3405  *
3406  * Only use this routine if vfs smr is the only protection you have against
3407  * freeing the vnode.
3408  *
3409  * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag
3410  * is not set.  After the flag is set the vnode becomes immutable to anyone but
3411  * the thread which managed to set the flag.
3412  *
3413  * It may be tempting to replace the loop with:
3414  * count = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3415  * if (count & VHOLD_NO_SMR) {
3416  *     backpedal and error out;
3417  * }
3418  *
3419  * However, while this is more performant, it hinders debugging by eliminating
3420  * the previously mentioned invariant.
3421  */
3422 bool
3423 vhold_smr(struct vnode *vp)
3424 {
3425 	int count;
3426 
3427 	VFS_SMR_ASSERT_ENTERED();
3428 
3429 	count = atomic_load_int(&vp->v_holdcnt);
3430 	for (;;) {
3431 		if (count & VHOLD_NO_SMR) {
3432 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3433 			    ("non-zero hold count with flags %d\n", count));
3434 			return (false);
3435 		}
3436 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3437 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3438 			if (count == 0)
3439 				vfs_freevnodes_dec();
3440 			return (true);
3441 		}
3442 	}
3443 }
3444 
3445 /*
3446  * Hold a free vnode for recycling.
3447  *
3448  * Note: vnode_init references this comment.
3449  *
3450  * Attempts to recycle only need the global vnode list lock and have no use for
3451  * SMR.
3452  *
3453  * However, vnodes get inserted into the global list before they get fully
3454  * initialized and stay there until UMA decides to free the memory. This in
3455  * particular means the target can be found before it becomes usable and after
3456  * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to
3457  * VHOLD_NO_SMR.
3458  *
3459  * Note: the vnode may gain more references after we transition the count 0->1.
3460  */
3461 static bool
3462 vhold_recycle_free(struct vnode *vp)
3463 {
3464 	int count;
3465 
3466 	mtx_assert(&vnode_list_mtx, MA_OWNED);
3467 
3468 	count = atomic_load_int(&vp->v_holdcnt);
3469 	for (;;) {
3470 		if (count & VHOLD_NO_SMR) {
3471 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3472 			    ("non-zero hold count with flags %d\n", count));
3473 			return (false);
3474 		}
3475 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3476 		if (count > 0) {
3477 			return (false);
3478 		}
3479 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3480 			vfs_freevnodes_dec();
3481 			return (true);
3482 		}
3483 	}
3484 }
3485 
3486 static void __noinline
3487 vdbatch_process(struct vdbatch *vd)
3488 {
3489 	struct vnode *vp;
3490 	int i;
3491 
3492 	mtx_assert(&vd->lock, MA_OWNED);
3493 	MPASS(curthread->td_pinned > 0);
3494 	MPASS(vd->index == VDBATCH_SIZE);
3495 
3496 	mtx_lock(&vnode_list_mtx);
3497 	critical_enter();
3498 	freevnodes += vd->freevnodes;
3499 	for (i = 0; i < VDBATCH_SIZE; i++) {
3500 		vp = vd->tab[i];
3501 		TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3502 		TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3503 		MPASS(vp->v_dbatchcpu != NOCPU);
3504 		vp->v_dbatchcpu = NOCPU;
3505 	}
3506 	mtx_unlock(&vnode_list_mtx);
3507 	vd->freevnodes = 0;
3508 	bzero(vd->tab, sizeof(vd->tab));
3509 	vd->index = 0;
3510 	critical_exit();
3511 }
3512 
3513 static void
3514 vdbatch_enqueue(struct vnode *vp)
3515 {
3516 	struct vdbatch *vd;
3517 
3518 	ASSERT_VI_LOCKED(vp, __func__);
3519 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3520 	    ("%s: deferring requeue of a doomed vnode", __func__));
3521 
3522 	if (vp->v_dbatchcpu != NOCPU) {
3523 		VI_UNLOCK(vp);
3524 		return;
3525 	}
3526 
3527 	sched_pin();
3528 	vd = DPCPU_PTR(vd);
3529 	mtx_lock(&vd->lock);
3530 	MPASS(vd->index < VDBATCH_SIZE);
3531 	MPASS(vd->tab[vd->index] == NULL);
3532 	/*
3533 	 * A hack: we depend on being pinned so that we know what to put in
3534 	 * ->v_dbatchcpu.
3535 	 */
3536 	vp->v_dbatchcpu = curcpu;
3537 	vd->tab[vd->index] = vp;
3538 	vd->index++;
3539 	VI_UNLOCK(vp);
3540 	if (vd->index == VDBATCH_SIZE)
3541 		vdbatch_process(vd);
3542 	mtx_unlock(&vd->lock);
3543 	sched_unpin();
3544 }
3545 
3546 /*
3547  * This routine must only be called for vnodes which are about to be
3548  * deallocated. Supporting dequeue for arbitrary vndoes would require
3549  * validating that the locked batch matches.
3550  */
3551 static void
3552 vdbatch_dequeue(struct vnode *vp)
3553 {
3554 	struct vdbatch *vd;
3555 	int i;
3556 	short cpu;
3557 
3558 	VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3559 	    ("%s: called for a used vnode\n", __func__));
3560 
3561 	cpu = vp->v_dbatchcpu;
3562 	if (cpu == NOCPU)
3563 		return;
3564 
3565 	vd = DPCPU_ID_PTR(cpu, vd);
3566 	mtx_lock(&vd->lock);
3567 	for (i = 0; i < vd->index; i++) {
3568 		if (vd->tab[i] != vp)
3569 			continue;
3570 		vp->v_dbatchcpu = NOCPU;
3571 		vd->index--;
3572 		vd->tab[i] = vd->tab[vd->index];
3573 		vd->tab[vd->index] = NULL;
3574 		break;
3575 	}
3576 	mtx_unlock(&vd->lock);
3577 	/*
3578 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3579 	 */
3580 	MPASS(vp->v_dbatchcpu == NOCPU);
3581 }
3582 
3583 /*
3584  * Drop the hold count of the vnode.  If this is the last reference to
3585  * the vnode we place it on the free list unless it has been vgone'd
3586  * (marked VIRF_DOOMED) in which case we will free it.
3587  *
3588  * Because the vnode vm object keeps a hold reference on the vnode if
3589  * there is at least one resident non-cached page, the vnode cannot
3590  * leave the active list without the page cleanup done.
3591  */
3592 static void __noinline
3593 vdropl_final(struct vnode *vp)
3594 {
3595 
3596 	ASSERT_VI_LOCKED(vp, __func__);
3597 	VNPASS(VN_IS_DOOMED(vp), vp);
3598 	/*
3599 	 * Set the VHOLD_NO_SMR flag.
3600 	 *
3601 	 * We may be racing against vhold_smr. If they win we can just pretend
3602 	 * we never got this far, they will vdrop later.
3603 	 */
3604 	if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) {
3605 		vfs_freevnodes_inc();
3606 		VI_UNLOCK(vp);
3607 		/*
3608 		 * We lost the aforementioned race. Any subsequent access is
3609 		 * invalid as they might have managed to vdropl on their own.
3610 		 */
3611 		return;
3612 	}
3613 	/*
3614 	 * Don't bump freevnodes as this one is going away.
3615 	 */
3616 	freevnode(vp);
3617 }
3618 
3619 void
3620 vdrop(struct vnode *vp)
3621 {
3622 
3623 	ASSERT_VI_UNLOCKED(vp, __func__);
3624 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3625 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3626 		return;
3627 	VI_LOCK(vp);
3628 	vdropl(vp);
3629 }
3630 
3631 static void __always_inline
3632 vdropl_impl(struct vnode *vp, bool enqueue)
3633 {
3634 
3635 	ASSERT_VI_LOCKED(vp, __func__);
3636 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3637 	if (!refcount_release(&vp->v_holdcnt)) {
3638 		VI_UNLOCK(vp);
3639 		return;
3640 	}
3641 	VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp);
3642 	VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
3643 	if (VN_IS_DOOMED(vp)) {
3644 		vdropl_final(vp);
3645 		return;
3646 	}
3647 
3648 	vfs_freevnodes_inc();
3649 	if (vp->v_mflag & VMP_LAZYLIST) {
3650 		vunlazy(vp);
3651 	}
3652 
3653 	if (!enqueue) {
3654 		VI_UNLOCK(vp);
3655 		return;
3656 	}
3657 
3658 	/*
3659 	 * Also unlocks the interlock. We can't assert on it as we
3660 	 * released our hold and by now the vnode might have been
3661 	 * freed.
3662 	 */
3663 	vdbatch_enqueue(vp);
3664 }
3665 
3666 void
3667 vdropl(struct vnode *vp)
3668 {
3669 
3670 	vdropl_impl(vp, true);
3671 }
3672 
3673 /*
3674  * vdrop a vnode when recycling
3675  *
3676  * This is a special case routine only to be used when recycling, differs from
3677  * regular vdrop by not requeieing the vnode on LRU.
3678  *
3679  * Consider a case where vtryrecycle continuously fails with all vnodes (due to
3680  * e.g., frozen writes on the filesystem), filling the batch and causing it to
3681  * be requeued. Then vnlru will end up revisiting the same vnodes. This is a
3682  * loop which can last for as long as writes are frozen.
3683  */
3684 static void
3685 vdropl_recycle(struct vnode *vp)
3686 {
3687 
3688 	vdropl_impl(vp, false);
3689 }
3690 
3691 static void
3692 vdrop_recycle(struct vnode *vp)
3693 {
3694 
3695 	VI_LOCK(vp);
3696 	vdropl_recycle(vp);
3697 }
3698 
3699 /*
3700  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3701  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3702  */
3703 static int
3704 vinactivef(struct vnode *vp)
3705 {
3706 	struct vm_object *obj;
3707 	int error;
3708 
3709 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3710 	ASSERT_VI_LOCKED(vp, "vinactive");
3711 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3712 	    ("vinactive: recursed on VI_DOINGINACT"));
3713 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3714 	vp->v_iflag |= VI_DOINGINACT;
3715 	vp->v_iflag &= ~VI_OWEINACT;
3716 	VI_UNLOCK(vp);
3717 	/*
3718 	 * Before moving off the active list, we must be sure that any
3719 	 * modified pages are converted into the vnode's dirty
3720 	 * buffers, since these will no longer be checked once the
3721 	 * vnode is on the inactive list.
3722 	 *
3723 	 * The write-out of the dirty pages is asynchronous.  At the
3724 	 * point that VOP_INACTIVE() is called, there could still be
3725 	 * pending I/O and dirty pages in the object.
3726 	 */
3727 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3728 	    vm_object_mightbedirty(obj)) {
3729 		VM_OBJECT_WLOCK(obj);
3730 		vm_object_page_clean(obj, 0, 0, 0);
3731 		VM_OBJECT_WUNLOCK(obj);
3732 	}
3733 	error = VOP_INACTIVE(vp);
3734 	VI_LOCK(vp);
3735 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3736 	    ("vinactive: lost VI_DOINGINACT"));
3737 	vp->v_iflag &= ~VI_DOINGINACT;
3738 	return (error);
3739 }
3740 
3741 int
3742 vinactive(struct vnode *vp)
3743 {
3744 
3745 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3746 	ASSERT_VI_LOCKED(vp, "vinactive");
3747 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3748 
3749 	if ((vp->v_iflag & VI_OWEINACT) == 0)
3750 		return (0);
3751 	if (vp->v_iflag & VI_DOINGINACT)
3752 		return (0);
3753 	if (vp->v_usecount > 0) {
3754 		vp->v_iflag &= ~VI_OWEINACT;
3755 		return (0);
3756 	}
3757 	return (vinactivef(vp));
3758 }
3759 
3760 /*
3761  * Remove any vnodes in the vnode table belonging to mount point mp.
3762  *
3763  * If FORCECLOSE is not specified, there should not be any active ones,
3764  * return error if any are found (nb: this is a user error, not a
3765  * system error). If FORCECLOSE is specified, detach any active vnodes
3766  * that are found.
3767  *
3768  * If WRITECLOSE is set, only flush out regular file vnodes open for
3769  * writing.
3770  *
3771  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3772  *
3773  * `rootrefs' specifies the base reference count for the root vnode
3774  * of this filesystem. The root vnode is considered busy if its
3775  * v_usecount exceeds this value. On a successful return, vflush(, td)
3776  * will call vrele() on the root vnode exactly rootrefs times.
3777  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3778  * be zero.
3779  */
3780 #ifdef DIAGNOSTIC
3781 static int busyprt = 0;		/* print out busy vnodes */
3782 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3783 #endif
3784 
3785 int
3786 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3787 {
3788 	struct vnode *vp, *mvp, *rootvp = NULL;
3789 	struct vattr vattr;
3790 	int busy = 0, error;
3791 
3792 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3793 	    rootrefs, flags);
3794 	if (rootrefs > 0) {
3795 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3796 		    ("vflush: bad args"));
3797 		/*
3798 		 * Get the filesystem root vnode. We can vput() it
3799 		 * immediately, since with rootrefs > 0, it won't go away.
3800 		 */
3801 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3802 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3803 			    __func__, error);
3804 			return (error);
3805 		}
3806 		vput(rootvp);
3807 	}
3808 loop:
3809 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3810 		vholdl(vp);
3811 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3812 		if (error) {
3813 			vdrop(vp);
3814 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3815 			goto loop;
3816 		}
3817 		/*
3818 		 * Skip over a vnodes marked VV_SYSTEM.
3819 		 */
3820 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3821 			VOP_UNLOCK(vp);
3822 			vdrop(vp);
3823 			continue;
3824 		}
3825 		/*
3826 		 * If WRITECLOSE is set, flush out unlinked but still open
3827 		 * files (even if open only for reading) and regular file
3828 		 * vnodes open for writing.
3829 		 */
3830 		if (flags & WRITECLOSE) {
3831 			if (vp->v_object != NULL) {
3832 				VM_OBJECT_WLOCK(vp->v_object);
3833 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3834 				VM_OBJECT_WUNLOCK(vp->v_object);
3835 			}
3836 			do {
3837 				error = VOP_FSYNC(vp, MNT_WAIT, td);
3838 			} while (error == ERELOOKUP);
3839 			if (error != 0) {
3840 				VOP_UNLOCK(vp);
3841 				vdrop(vp);
3842 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3843 				return (error);
3844 			}
3845 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3846 			VI_LOCK(vp);
3847 
3848 			if ((vp->v_type == VNON ||
3849 			    (error == 0 && vattr.va_nlink > 0)) &&
3850 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3851 				VOP_UNLOCK(vp);
3852 				vdropl(vp);
3853 				continue;
3854 			}
3855 		} else
3856 			VI_LOCK(vp);
3857 		/*
3858 		 * With v_usecount == 0, all we need to do is clear out the
3859 		 * vnode data structures and we are done.
3860 		 *
3861 		 * If FORCECLOSE is set, forcibly close the vnode.
3862 		 */
3863 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3864 			vgonel(vp);
3865 		} else {
3866 			busy++;
3867 #ifdef DIAGNOSTIC
3868 			if (busyprt)
3869 				vn_printf(vp, "vflush: busy vnode ");
3870 #endif
3871 		}
3872 		VOP_UNLOCK(vp);
3873 		vdropl(vp);
3874 	}
3875 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3876 		/*
3877 		 * If just the root vnode is busy, and if its refcount
3878 		 * is equal to `rootrefs', then go ahead and kill it.
3879 		 */
3880 		VI_LOCK(rootvp);
3881 		KASSERT(busy > 0, ("vflush: not busy"));
3882 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3883 		    ("vflush: usecount %d < rootrefs %d",
3884 		     rootvp->v_usecount, rootrefs));
3885 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3886 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3887 			vgone(rootvp);
3888 			VOP_UNLOCK(rootvp);
3889 			busy = 0;
3890 		} else
3891 			VI_UNLOCK(rootvp);
3892 	}
3893 	if (busy) {
3894 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3895 		    busy);
3896 		return (EBUSY);
3897 	}
3898 	for (; rootrefs > 0; rootrefs--)
3899 		vrele(rootvp);
3900 	return (0);
3901 }
3902 
3903 /*
3904  * Recycle an unused vnode to the front of the free list.
3905  */
3906 int
3907 vrecycle(struct vnode *vp)
3908 {
3909 	int recycled;
3910 
3911 	VI_LOCK(vp);
3912 	recycled = vrecyclel(vp);
3913 	VI_UNLOCK(vp);
3914 	return (recycled);
3915 }
3916 
3917 /*
3918  * vrecycle, with the vp interlock held.
3919  */
3920 int
3921 vrecyclel(struct vnode *vp)
3922 {
3923 	int recycled;
3924 
3925 	ASSERT_VOP_ELOCKED(vp, __func__);
3926 	ASSERT_VI_LOCKED(vp, __func__);
3927 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3928 	recycled = 0;
3929 	if (vp->v_usecount == 0) {
3930 		recycled = 1;
3931 		vgonel(vp);
3932 	}
3933 	return (recycled);
3934 }
3935 
3936 /*
3937  * Eliminate all activity associated with a vnode
3938  * in preparation for reuse.
3939  */
3940 void
3941 vgone(struct vnode *vp)
3942 {
3943 	VI_LOCK(vp);
3944 	vgonel(vp);
3945 	VI_UNLOCK(vp);
3946 }
3947 
3948 /*
3949  * Notify upper mounts about reclaimed or unlinked vnode.
3950  */
3951 void
3952 vfs_notify_upper(struct vnode *vp, enum vfs_notify_upper_type event)
3953 {
3954 	struct mount *mp;
3955 	struct mount_upper_node *ump;
3956 
3957 	mp = atomic_load_ptr(&vp->v_mount);
3958 	if (mp == NULL)
3959 		return;
3960 	if (TAILQ_EMPTY(&mp->mnt_notify))
3961 		return;
3962 
3963 	MNT_ILOCK(mp);
3964 	mp->mnt_upper_pending++;
3965 	KASSERT(mp->mnt_upper_pending > 0,
3966 	    ("%s: mnt_upper_pending %d", __func__, mp->mnt_upper_pending));
3967 	TAILQ_FOREACH(ump, &mp->mnt_notify, mnt_upper_link) {
3968 		MNT_IUNLOCK(mp);
3969 		switch (event) {
3970 		case VFS_NOTIFY_UPPER_RECLAIM:
3971 			VFS_RECLAIM_LOWERVP(ump->mp, vp);
3972 			break;
3973 		case VFS_NOTIFY_UPPER_UNLINK:
3974 			VFS_UNLINK_LOWERVP(ump->mp, vp);
3975 			break;
3976 		}
3977 		MNT_ILOCK(mp);
3978 	}
3979 	mp->mnt_upper_pending--;
3980 	if ((mp->mnt_kern_flag & MNTK_UPPER_WAITER) != 0 &&
3981 	    mp->mnt_upper_pending == 0) {
3982 		mp->mnt_kern_flag &= ~MNTK_UPPER_WAITER;
3983 		wakeup(&mp->mnt_uppers);
3984 	}
3985 	MNT_IUNLOCK(mp);
3986 }
3987 
3988 /*
3989  * vgone, with the vp interlock held.
3990  */
3991 static void
3992 vgonel(struct vnode *vp)
3993 {
3994 	struct thread *td;
3995 	struct mount *mp;
3996 	vm_object_t object;
3997 	bool active, doinginact, oweinact;
3998 
3999 	ASSERT_VOP_ELOCKED(vp, "vgonel");
4000 	ASSERT_VI_LOCKED(vp, "vgonel");
4001 	VNASSERT(vp->v_holdcnt, vp,
4002 	    ("vgonel: vp %p has no reference.", vp));
4003 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4004 	td = curthread;
4005 
4006 	/*
4007 	 * Don't vgonel if we're already doomed.
4008 	 */
4009 	if (VN_IS_DOOMED(vp))
4010 		return;
4011 	/*
4012 	 * Paired with freevnode.
4013 	 */
4014 	vn_seqc_write_begin_locked(vp);
4015 	vunlazy_gone(vp);
4016 	vn_irflag_set_locked(vp, VIRF_DOOMED);
4017 
4018 	/*
4019 	 * Check to see if the vnode is in use.  If so, we have to
4020 	 * call VOP_CLOSE() and VOP_INACTIVE().
4021 	 *
4022 	 * It could be that VOP_INACTIVE() requested reclamation, in
4023 	 * which case we should avoid recursion, so check
4024 	 * VI_DOINGINACT.  This is not precise but good enough.
4025 	 */
4026 	active = vp->v_usecount > 0;
4027 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4028 	doinginact = (vp->v_iflag & VI_DOINGINACT) != 0;
4029 
4030 	/*
4031 	 * If we need to do inactive VI_OWEINACT will be set.
4032 	 */
4033 	if (vp->v_iflag & VI_DEFINACT) {
4034 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
4035 		vp->v_iflag &= ~VI_DEFINACT;
4036 		vdropl(vp);
4037 	} else {
4038 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
4039 		VI_UNLOCK(vp);
4040 	}
4041 	cache_purge_vgone(vp);
4042 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
4043 
4044 	/*
4045 	 * If purging an active vnode, it must be closed and
4046 	 * deactivated before being reclaimed.
4047 	 */
4048 	if (active)
4049 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
4050 	if (!doinginact) {
4051 		do {
4052 			if (oweinact || active) {
4053 				VI_LOCK(vp);
4054 				vinactivef(vp);
4055 				oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4056 				VI_UNLOCK(vp);
4057 			}
4058 		} while (oweinact);
4059 	}
4060 	if (vp->v_type == VSOCK)
4061 		vfs_unp_reclaim(vp);
4062 
4063 	/*
4064 	 * Clean out any buffers associated with the vnode.
4065 	 * If the flush fails, just toss the buffers.
4066 	 */
4067 	mp = NULL;
4068 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
4069 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
4070 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
4071 		while (vinvalbuf(vp, 0, 0, 0) != 0)
4072 			;
4073 	}
4074 
4075 	BO_LOCK(&vp->v_bufobj);
4076 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
4077 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
4078 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
4079 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
4080 	    ("vp %p bufobj not invalidated", vp));
4081 
4082 	/*
4083 	 * For VMIO bufobj, BO_DEAD is set later, or in
4084 	 * vm_object_terminate() after the object's page queue is
4085 	 * flushed.
4086 	 */
4087 	object = vp->v_bufobj.bo_object;
4088 	if (object == NULL)
4089 		vp->v_bufobj.bo_flag |= BO_DEAD;
4090 	BO_UNLOCK(&vp->v_bufobj);
4091 
4092 	/*
4093 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
4094 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
4095 	 * should not touch the object borrowed from the lower vnode
4096 	 * (the handle check).
4097 	 */
4098 	if (object != NULL && object->type == OBJT_VNODE &&
4099 	    object->handle == vp)
4100 		vnode_destroy_vobject(vp);
4101 
4102 	/*
4103 	 * Reclaim the vnode.
4104 	 */
4105 	if (VOP_RECLAIM(vp))
4106 		panic("vgone: cannot reclaim");
4107 	if (mp != NULL)
4108 		vn_finished_secondary_write(mp);
4109 	VNASSERT(vp->v_object == NULL, vp,
4110 	    ("vop_reclaim left v_object vp=%p", vp));
4111 	/*
4112 	 * Clear the advisory locks and wake up waiting threads.
4113 	 */
4114 	if (vp->v_lockf != NULL) {
4115 		(void)VOP_ADVLOCKPURGE(vp);
4116 		vp->v_lockf = NULL;
4117 	}
4118 	/*
4119 	 * Delete from old mount point vnode list.
4120 	 */
4121 	if (vp->v_mount == NULL) {
4122 		VI_LOCK(vp);
4123 	} else {
4124 		delmntque(vp);
4125 		ASSERT_VI_LOCKED(vp, "vgonel 2");
4126 	}
4127 	/*
4128 	 * Done with purge, reset to the standard lock and invalidate
4129 	 * the vnode.
4130 	 */
4131 	vp->v_vnlock = &vp->v_lock;
4132 	vp->v_op = &dead_vnodeops;
4133 	vp->v_type = VBAD;
4134 }
4135 
4136 /*
4137  * Print out a description of a vnode.
4138  */
4139 static const char * const typename[] =
4140 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
4141  "VMARKER"};
4142 
4143 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0,
4144     "new hold count flag not added to vn_printf");
4145 
4146 void
4147 vn_printf(struct vnode *vp, const char *fmt, ...)
4148 {
4149 	va_list ap;
4150 	char buf[256], buf2[16];
4151 	u_long flags;
4152 	u_int holdcnt;
4153 	short irflag;
4154 
4155 	va_start(ap, fmt);
4156 	vprintf(fmt, ap);
4157 	va_end(ap);
4158 	printf("%p: ", (void *)vp);
4159 	printf("type %s\n", typename[vp->v_type]);
4160 	holdcnt = atomic_load_int(&vp->v_holdcnt);
4161 	printf("    usecount %d, writecount %d, refcount %d seqc users %d",
4162 	    vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS,
4163 	    vp->v_seqc_users);
4164 	switch (vp->v_type) {
4165 	case VDIR:
4166 		printf(" mountedhere %p\n", vp->v_mountedhere);
4167 		break;
4168 	case VCHR:
4169 		printf(" rdev %p\n", vp->v_rdev);
4170 		break;
4171 	case VSOCK:
4172 		printf(" socket %p\n", vp->v_unpcb);
4173 		break;
4174 	case VFIFO:
4175 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
4176 		break;
4177 	default:
4178 		printf("\n");
4179 		break;
4180 	}
4181 	buf[0] = '\0';
4182 	buf[1] = '\0';
4183 	if (holdcnt & VHOLD_NO_SMR)
4184 		strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf));
4185 	printf("    hold count flags (%s)\n", buf + 1);
4186 
4187 	buf[0] = '\0';
4188 	buf[1] = '\0';
4189 	irflag = vn_irflag_read(vp);
4190 	if (irflag & VIRF_DOOMED)
4191 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4192 	if (irflag & VIRF_PGREAD)
4193 		strlcat(buf, "|VIRF_PGREAD", sizeof(buf));
4194 	if (irflag & VIRF_MOUNTPOINT)
4195 		strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf));
4196 	if (irflag & VIRF_TEXT_REF)
4197 		strlcat(buf, "|VIRF_TEXT_REF", sizeof(buf));
4198 	flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT | VIRF_TEXT_REF);
4199 	if (flags != 0) {
4200 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4201 		strlcat(buf, buf2, sizeof(buf));
4202 	}
4203 	if (vp->v_vflag & VV_ROOT)
4204 		strlcat(buf, "|VV_ROOT", sizeof(buf));
4205 	if (vp->v_vflag & VV_ISTTY)
4206 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
4207 	if (vp->v_vflag & VV_NOSYNC)
4208 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4209 	if (vp->v_vflag & VV_ETERNALDEV)
4210 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4211 	if (vp->v_vflag & VV_CACHEDLABEL)
4212 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4213 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
4214 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4215 	if (vp->v_vflag & VV_COPYONWRITE)
4216 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4217 	if (vp->v_vflag & VV_SYSTEM)
4218 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4219 	if (vp->v_vflag & VV_PROCDEP)
4220 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4221 	if (vp->v_vflag & VV_DELETED)
4222 		strlcat(buf, "|VV_DELETED", sizeof(buf));
4223 	if (vp->v_vflag & VV_MD)
4224 		strlcat(buf, "|VV_MD", sizeof(buf));
4225 	if (vp->v_vflag & VV_FORCEINSMQ)
4226 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4227 	if (vp->v_vflag & VV_READLINK)
4228 		strlcat(buf, "|VV_READLINK", sizeof(buf));
4229 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4230 	    VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM |
4231 	    VV_PROCDEP | VV_DELETED | VV_MD | VV_FORCEINSMQ | VV_READLINK);
4232 	if (flags != 0) {
4233 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4234 		strlcat(buf, buf2, sizeof(buf));
4235 	}
4236 	if (vp->v_iflag & VI_MOUNT)
4237 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
4238 	if (vp->v_iflag & VI_DOINGINACT)
4239 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4240 	if (vp->v_iflag & VI_OWEINACT)
4241 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4242 	if (vp->v_iflag & VI_DEFINACT)
4243 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4244 	if (vp->v_iflag & VI_FOPENING)
4245 		strlcat(buf, "|VI_FOPENING", sizeof(buf));
4246 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOINGINACT |
4247 	    VI_OWEINACT | VI_DEFINACT | VI_FOPENING);
4248 	if (flags != 0) {
4249 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4250 		strlcat(buf, buf2, sizeof(buf));
4251 	}
4252 	if (vp->v_mflag & VMP_LAZYLIST)
4253 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4254 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
4255 	if (flags != 0) {
4256 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4257 		strlcat(buf, buf2, sizeof(buf));
4258 	}
4259 	printf("    flags (%s)", buf + 1);
4260 	if (mtx_owned(VI_MTX(vp)))
4261 		printf(" VI_LOCKed");
4262 	printf("\n");
4263 	if (vp->v_object != NULL)
4264 		printf("    v_object %p ref %d pages %d "
4265 		    "cleanbuf %d dirtybuf %d\n",
4266 		    vp->v_object, vp->v_object->ref_count,
4267 		    vp->v_object->resident_page_count,
4268 		    vp->v_bufobj.bo_clean.bv_cnt,
4269 		    vp->v_bufobj.bo_dirty.bv_cnt);
4270 	printf("    ");
4271 	lockmgr_printinfo(vp->v_vnlock);
4272 	if (vp->v_data != NULL)
4273 		VOP_PRINT(vp);
4274 }
4275 
4276 #ifdef DDB
4277 /*
4278  * List all of the locked vnodes in the system.
4279  * Called when debugging the kernel.
4280  */
4281 DB_SHOW_COMMAND_FLAGS(lockedvnods, lockedvnodes, DB_CMD_MEMSAFE)
4282 {
4283 	struct mount *mp;
4284 	struct vnode *vp;
4285 
4286 	/*
4287 	 * Note: because this is DDB, we can't obey the locking semantics
4288 	 * for these structures, which means we could catch an inconsistent
4289 	 * state and dereference a nasty pointer.  Not much to be done
4290 	 * about that.
4291 	 */
4292 	db_printf("Locked vnodes\n");
4293 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4294 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4295 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4296 				vn_printf(vp, "vnode ");
4297 		}
4298 	}
4299 }
4300 
4301 /*
4302  * Show details about the given vnode.
4303  */
4304 DB_SHOW_COMMAND(vnode, db_show_vnode)
4305 {
4306 	struct vnode *vp;
4307 
4308 	if (!have_addr)
4309 		return;
4310 	vp = (struct vnode *)addr;
4311 	vn_printf(vp, "vnode ");
4312 }
4313 
4314 /*
4315  * Show details about the given mount point.
4316  */
4317 DB_SHOW_COMMAND(mount, db_show_mount)
4318 {
4319 	struct mount *mp;
4320 	struct vfsopt *opt;
4321 	struct statfs *sp;
4322 	struct vnode *vp;
4323 	char buf[512];
4324 	uint64_t mflags;
4325 	u_int flags;
4326 
4327 	if (!have_addr) {
4328 		/* No address given, print short info about all mount points. */
4329 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4330 			db_printf("%p %s on %s (%s)\n", mp,
4331 			    mp->mnt_stat.f_mntfromname,
4332 			    mp->mnt_stat.f_mntonname,
4333 			    mp->mnt_stat.f_fstypename);
4334 			if (db_pager_quit)
4335 				break;
4336 		}
4337 		db_printf("\nMore info: show mount <addr>\n");
4338 		return;
4339 	}
4340 
4341 	mp = (struct mount *)addr;
4342 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4343 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4344 
4345 	buf[0] = '\0';
4346 	mflags = mp->mnt_flag;
4347 #define	MNT_FLAG(flag)	do {						\
4348 	if (mflags & (flag)) {						\
4349 		if (buf[0] != '\0')					\
4350 			strlcat(buf, ", ", sizeof(buf));		\
4351 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4352 		mflags &= ~(flag);					\
4353 	}								\
4354 } while (0)
4355 	MNT_FLAG(MNT_RDONLY);
4356 	MNT_FLAG(MNT_SYNCHRONOUS);
4357 	MNT_FLAG(MNT_NOEXEC);
4358 	MNT_FLAG(MNT_NOSUID);
4359 	MNT_FLAG(MNT_NFS4ACLS);
4360 	MNT_FLAG(MNT_UNION);
4361 	MNT_FLAG(MNT_ASYNC);
4362 	MNT_FLAG(MNT_SUIDDIR);
4363 	MNT_FLAG(MNT_SOFTDEP);
4364 	MNT_FLAG(MNT_NOSYMFOLLOW);
4365 	MNT_FLAG(MNT_GJOURNAL);
4366 	MNT_FLAG(MNT_MULTILABEL);
4367 	MNT_FLAG(MNT_ACLS);
4368 	MNT_FLAG(MNT_NOATIME);
4369 	MNT_FLAG(MNT_NOCLUSTERR);
4370 	MNT_FLAG(MNT_NOCLUSTERW);
4371 	MNT_FLAG(MNT_SUJ);
4372 	MNT_FLAG(MNT_EXRDONLY);
4373 	MNT_FLAG(MNT_EXPORTED);
4374 	MNT_FLAG(MNT_DEFEXPORTED);
4375 	MNT_FLAG(MNT_EXPORTANON);
4376 	MNT_FLAG(MNT_EXKERB);
4377 	MNT_FLAG(MNT_EXPUBLIC);
4378 	MNT_FLAG(MNT_LOCAL);
4379 	MNT_FLAG(MNT_QUOTA);
4380 	MNT_FLAG(MNT_ROOTFS);
4381 	MNT_FLAG(MNT_USER);
4382 	MNT_FLAG(MNT_IGNORE);
4383 	MNT_FLAG(MNT_UPDATE);
4384 	MNT_FLAG(MNT_DELEXPORT);
4385 	MNT_FLAG(MNT_RELOAD);
4386 	MNT_FLAG(MNT_FORCE);
4387 	MNT_FLAG(MNT_SNAPSHOT);
4388 	MNT_FLAG(MNT_BYFSID);
4389 #undef MNT_FLAG
4390 	if (mflags != 0) {
4391 		if (buf[0] != '\0')
4392 			strlcat(buf, ", ", sizeof(buf));
4393 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4394 		    "0x%016jx", mflags);
4395 	}
4396 	db_printf("    mnt_flag = %s\n", buf);
4397 
4398 	buf[0] = '\0';
4399 	flags = mp->mnt_kern_flag;
4400 #define	MNT_KERN_FLAG(flag)	do {					\
4401 	if (flags & (flag)) {						\
4402 		if (buf[0] != '\0')					\
4403 			strlcat(buf, ", ", sizeof(buf));		\
4404 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4405 		flags &= ~(flag);					\
4406 	}								\
4407 } while (0)
4408 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4409 	MNT_KERN_FLAG(MNTK_ASYNC);
4410 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4411 	MNT_KERN_FLAG(MNTK_NOMSYNC);
4412 	MNT_KERN_FLAG(MNTK_DRAINING);
4413 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4414 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4415 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4416 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4417 	MNT_KERN_FLAG(MNTK_RECURSE);
4418 	MNT_KERN_FLAG(MNTK_UPPER_WAITER);
4419 	MNT_KERN_FLAG(MNTK_UNLOCKED_INSMNTQUE);
4420 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4421 	MNT_KERN_FLAG(MNTK_VMSETSIZE_BUG);
4422 	MNT_KERN_FLAG(MNTK_FPLOOKUP);
4423 	MNT_KERN_FLAG(MNTK_TASKQUEUE_WAITER);
4424 	MNT_KERN_FLAG(MNTK_NOASYNC);
4425 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4426 	MNT_KERN_FLAG(MNTK_MWAIT);
4427 	MNT_KERN_FLAG(MNTK_SUSPEND);
4428 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4429 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4430 	MNT_KERN_FLAG(MNTK_NULL_NOCACHE);
4431 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4432 #undef MNT_KERN_FLAG
4433 	if (flags != 0) {
4434 		if (buf[0] != '\0')
4435 			strlcat(buf, ", ", sizeof(buf));
4436 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4437 		    "0x%08x", flags);
4438 	}
4439 	db_printf("    mnt_kern_flag = %s\n", buf);
4440 
4441 	db_printf("    mnt_opt = ");
4442 	opt = TAILQ_FIRST(mp->mnt_opt);
4443 	if (opt != NULL) {
4444 		db_printf("%s", opt->name);
4445 		opt = TAILQ_NEXT(opt, link);
4446 		while (opt != NULL) {
4447 			db_printf(", %s", opt->name);
4448 			opt = TAILQ_NEXT(opt, link);
4449 		}
4450 	}
4451 	db_printf("\n");
4452 
4453 	sp = &mp->mnt_stat;
4454 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4455 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4456 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4457 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4458 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4459 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4460 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4461 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4462 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4463 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4464 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4465 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4466 
4467 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4468 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4469 	if (jailed(mp->mnt_cred))
4470 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4471 	db_printf(" }\n");
4472 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4473 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4474 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4475 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4476 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4477 	    mp->mnt_lazyvnodelistsize);
4478 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4479 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4480 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4481 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4482 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4483 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4484 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4485 	db_printf("    mnt_secondary_accwrites = %d\n",
4486 	    mp->mnt_secondary_accwrites);
4487 	db_printf("    mnt_gjprovider = %s\n",
4488 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4489 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4490 
4491 	db_printf("\n\nList of active vnodes\n");
4492 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4493 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4494 			vn_printf(vp, "vnode ");
4495 			if (db_pager_quit)
4496 				break;
4497 		}
4498 	}
4499 	db_printf("\n\nList of inactive vnodes\n");
4500 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4501 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4502 			vn_printf(vp, "vnode ");
4503 			if (db_pager_quit)
4504 				break;
4505 		}
4506 	}
4507 }
4508 #endif	/* DDB */
4509 
4510 /*
4511  * Fill in a struct xvfsconf based on a struct vfsconf.
4512  */
4513 static int
4514 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4515 {
4516 	struct xvfsconf xvfsp;
4517 
4518 	bzero(&xvfsp, sizeof(xvfsp));
4519 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4520 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4521 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4522 	xvfsp.vfc_flags = vfsp->vfc_flags;
4523 	/*
4524 	 * These are unused in userland, we keep them
4525 	 * to not break binary compatibility.
4526 	 */
4527 	xvfsp.vfc_vfsops = NULL;
4528 	xvfsp.vfc_next = NULL;
4529 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4530 }
4531 
4532 #ifdef COMPAT_FREEBSD32
4533 struct xvfsconf32 {
4534 	uint32_t	vfc_vfsops;
4535 	char		vfc_name[MFSNAMELEN];
4536 	int32_t		vfc_typenum;
4537 	int32_t		vfc_refcount;
4538 	int32_t		vfc_flags;
4539 	uint32_t	vfc_next;
4540 };
4541 
4542 static int
4543 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4544 {
4545 	struct xvfsconf32 xvfsp;
4546 
4547 	bzero(&xvfsp, sizeof(xvfsp));
4548 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4549 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4550 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4551 	xvfsp.vfc_flags = vfsp->vfc_flags;
4552 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4553 }
4554 #endif
4555 
4556 /*
4557  * Top level filesystem related information gathering.
4558  */
4559 static int
4560 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4561 {
4562 	struct vfsconf *vfsp;
4563 	int error;
4564 
4565 	error = 0;
4566 	vfsconf_slock();
4567 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4568 #ifdef COMPAT_FREEBSD32
4569 		if (req->flags & SCTL_MASK32)
4570 			error = vfsconf2x32(req, vfsp);
4571 		else
4572 #endif
4573 			error = vfsconf2x(req, vfsp);
4574 		if (error)
4575 			break;
4576 	}
4577 	vfsconf_sunlock();
4578 	return (error);
4579 }
4580 
4581 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4582     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4583     "S,xvfsconf", "List of all configured filesystems");
4584 
4585 #ifndef BURN_BRIDGES
4586 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4587 
4588 static int
4589 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4590 {
4591 	int *name = (int *)arg1 - 1;	/* XXX */
4592 	u_int namelen = arg2 + 1;	/* XXX */
4593 	struct vfsconf *vfsp;
4594 
4595 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4596 	    "please rebuild world\n");
4597 
4598 #if 1 || defined(COMPAT_PRELITE2)
4599 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4600 	if (namelen == 1)
4601 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4602 #endif
4603 
4604 	switch (name[1]) {
4605 	case VFS_MAXTYPENUM:
4606 		if (namelen != 2)
4607 			return (ENOTDIR);
4608 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4609 	case VFS_CONF:
4610 		if (namelen != 3)
4611 			return (ENOTDIR);	/* overloaded */
4612 		vfsconf_slock();
4613 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4614 			if (vfsp->vfc_typenum == name[2])
4615 				break;
4616 		}
4617 		vfsconf_sunlock();
4618 		if (vfsp == NULL)
4619 			return (EOPNOTSUPP);
4620 #ifdef COMPAT_FREEBSD32
4621 		if (req->flags & SCTL_MASK32)
4622 			return (vfsconf2x32(req, vfsp));
4623 		else
4624 #endif
4625 			return (vfsconf2x(req, vfsp));
4626 	}
4627 	return (EOPNOTSUPP);
4628 }
4629 
4630 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4631     CTLFLAG_MPSAFE, vfs_sysctl,
4632     "Generic filesystem");
4633 
4634 #if 1 || defined(COMPAT_PRELITE2)
4635 
4636 static int
4637 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4638 {
4639 	int error;
4640 	struct vfsconf *vfsp;
4641 	struct ovfsconf ovfs;
4642 
4643 	vfsconf_slock();
4644 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4645 		bzero(&ovfs, sizeof(ovfs));
4646 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4647 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4648 		ovfs.vfc_index = vfsp->vfc_typenum;
4649 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4650 		ovfs.vfc_flags = vfsp->vfc_flags;
4651 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4652 		if (error != 0) {
4653 			vfsconf_sunlock();
4654 			return (error);
4655 		}
4656 	}
4657 	vfsconf_sunlock();
4658 	return (0);
4659 }
4660 
4661 #endif /* 1 || COMPAT_PRELITE2 */
4662 #endif /* !BURN_BRIDGES */
4663 
4664 #define KINFO_VNODESLOP		10
4665 #ifdef notyet
4666 /*
4667  * Dump vnode list (via sysctl).
4668  */
4669 /* ARGSUSED */
4670 static int
4671 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4672 {
4673 	struct xvnode *xvn;
4674 	struct mount *mp;
4675 	struct vnode *vp;
4676 	int error, len, n;
4677 
4678 	/*
4679 	 * Stale numvnodes access is not fatal here.
4680 	 */
4681 	req->lock = 0;
4682 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4683 	if (!req->oldptr)
4684 		/* Make an estimate */
4685 		return (SYSCTL_OUT(req, 0, len));
4686 
4687 	error = sysctl_wire_old_buffer(req, 0);
4688 	if (error != 0)
4689 		return (error);
4690 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4691 	n = 0;
4692 	mtx_lock(&mountlist_mtx);
4693 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4694 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4695 			continue;
4696 		MNT_ILOCK(mp);
4697 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4698 			if (n == len)
4699 				break;
4700 			vref(vp);
4701 			xvn[n].xv_size = sizeof *xvn;
4702 			xvn[n].xv_vnode = vp;
4703 			xvn[n].xv_id = 0;	/* XXX compat */
4704 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4705 			XV_COPY(usecount);
4706 			XV_COPY(writecount);
4707 			XV_COPY(holdcnt);
4708 			XV_COPY(mount);
4709 			XV_COPY(numoutput);
4710 			XV_COPY(type);
4711 #undef XV_COPY
4712 			xvn[n].xv_flag = vp->v_vflag;
4713 
4714 			switch (vp->v_type) {
4715 			case VREG:
4716 			case VDIR:
4717 			case VLNK:
4718 				break;
4719 			case VBLK:
4720 			case VCHR:
4721 				if (vp->v_rdev == NULL) {
4722 					vrele(vp);
4723 					continue;
4724 				}
4725 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4726 				break;
4727 			case VSOCK:
4728 				xvn[n].xv_socket = vp->v_socket;
4729 				break;
4730 			case VFIFO:
4731 				xvn[n].xv_fifo = vp->v_fifoinfo;
4732 				break;
4733 			case VNON:
4734 			case VBAD:
4735 			default:
4736 				/* shouldn't happen? */
4737 				vrele(vp);
4738 				continue;
4739 			}
4740 			vrele(vp);
4741 			++n;
4742 		}
4743 		MNT_IUNLOCK(mp);
4744 		mtx_lock(&mountlist_mtx);
4745 		vfs_unbusy(mp);
4746 		if (n == len)
4747 			break;
4748 	}
4749 	mtx_unlock(&mountlist_mtx);
4750 
4751 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4752 	free(xvn, M_TEMP);
4753 	return (error);
4754 }
4755 
4756 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4757     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4758     "");
4759 #endif
4760 
4761 static void
4762 unmount_or_warn(struct mount *mp)
4763 {
4764 	int error;
4765 
4766 	error = dounmount(mp, MNT_FORCE, curthread);
4767 	if (error != 0) {
4768 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4769 		if (error == EBUSY)
4770 			printf("BUSY)\n");
4771 		else
4772 			printf("%d)\n", error);
4773 	}
4774 }
4775 
4776 /*
4777  * Unmount all filesystems. The list is traversed in reverse order
4778  * of mounting to avoid dependencies.
4779  */
4780 void
4781 vfs_unmountall(void)
4782 {
4783 	struct mount *mp, *tmp;
4784 
4785 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4786 
4787 	/*
4788 	 * Since this only runs when rebooting, it is not interlocked.
4789 	 */
4790 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4791 		vfs_ref(mp);
4792 
4793 		/*
4794 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4795 		 * unmount of the latter.
4796 		 */
4797 		if (mp == rootdevmp)
4798 			continue;
4799 
4800 		unmount_or_warn(mp);
4801 	}
4802 
4803 	if (rootdevmp != NULL)
4804 		unmount_or_warn(rootdevmp);
4805 }
4806 
4807 static void
4808 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4809 {
4810 
4811 	ASSERT_VI_LOCKED(vp, __func__);
4812 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4813 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
4814 		vdropl(vp);
4815 		return;
4816 	}
4817 	if (vn_lock(vp, lkflags) == 0) {
4818 		VI_LOCK(vp);
4819 		vinactive(vp);
4820 		VOP_UNLOCK(vp);
4821 		vdropl(vp);
4822 		return;
4823 	}
4824 	vdefer_inactive_unlocked(vp);
4825 }
4826 
4827 static int
4828 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4829 {
4830 
4831 	return (vp->v_iflag & VI_DEFINACT);
4832 }
4833 
4834 static void __noinline
4835 vfs_periodic_inactive(struct mount *mp, int flags)
4836 {
4837 	struct vnode *vp, *mvp;
4838 	int lkflags;
4839 
4840 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4841 	if (flags != MNT_WAIT)
4842 		lkflags |= LK_NOWAIT;
4843 
4844 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4845 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
4846 			VI_UNLOCK(vp);
4847 			continue;
4848 		}
4849 		vp->v_iflag &= ~VI_DEFINACT;
4850 		vfs_deferred_inactive(vp, lkflags);
4851 	}
4852 }
4853 
4854 static inline bool
4855 vfs_want_msync(struct vnode *vp)
4856 {
4857 	struct vm_object *obj;
4858 
4859 	/*
4860 	 * This test may be performed without any locks held.
4861 	 * We rely on vm_object's type stability.
4862 	 */
4863 	if (vp->v_vflag & VV_NOSYNC)
4864 		return (false);
4865 	obj = vp->v_object;
4866 	return (obj != NULL && vm_object_mightbedirty(obj));
4867 }
4868 
4869 static int
4870 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4871 {
4872 
4873 	if (vp->v_vflag & VV_NOSYNC)
4874 		return (false);
4875 	if (vp->v_iflag & VI_DEFINACT)
4876 		return (true);
4877 	return (vfs_want_msync(vp));
4878 }
4879 
4880 static void __noinline
4881 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4882 {
4883 	struct vnode *vp, *mvp;
4884 	struct vm_object *obj;
4885 	int lkflags, objflags;
4886 	bool seen_defer;
4887 
4888 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4889 	if (flags != MNT_WAIT) {
4890 		lkflags |= LK_NOWAIT;
4891 		objflags = OBJPC_NOSYNC;
4892 	} else {
4893 		objflags = OBJPC_SYNC;
4894 	}
4895 
4896 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4897 		seen_defer = false;
4898 		if (vp->v_iflag & VI_DEFINACT) {
4899 			vp->v_iflag &= ~VI_DEFINACT;
4900 			seen_defer = true;
4901 		}
4902 		if (!vfs_want_msync(vp)) {
4903 			if (seen_defer)
4904 				vfs_deferred_inactive(vp, lkflags);
4905 			else
4906 				VI_UNLOCK(vp);
4907 			continue;
4908 		}
4909 		if (vget(vp, lkflags) == 0) {
4910 			obj = vp->v_object;
4911 			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4912 				VM_OBJECT_WLOCK(obj);
4913 				vm_object_page_clean(obj, 0, 0, objflags);
4914 				VM_OBJECT_WUNLOCK(obj);
4915 			}
4916 			vput(vp);
4917 			if (seen_defer)
4918 				vdrop(vp);
4919 		} else {
4920 			if (seen_defer)
4921 				vdefer_inactive_unlocked(vp);
4922 		}
4923 	}
4924 }
4925 
4926 void
4927 vfs_periodic(struct mount *mp, int flags)
4928 {
4929 
4930 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4931 
4932 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4933 		vfs_periodic_inactive(mp, flags);
4934 	else
4935 		vfs_periodic_msync_inactive(mp, flags);
4936 }
4937 
4938 static void
4939 destroy_vpollinfo_free(struct vpollinfo *vi)
4940 {
4941 
4942 	knlist_destroy(&vi->vpi_selinfo.si_note);
4943 	mtx_destroy(&vi->vpi_lock);
4944 	free(vi, M_VNODEPOLL);
4945 }
4946 
4947 static void
4948 destroy_vpollinfo(struct vpollinfo *vi)
4949 {
4950 
4951 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4952 	seldrain(&vi->vpi_selinfo);
4953 	destroy_vpollinfo_free(vi);
4954 }
4955 
4956 /*
4957  * Initialize per-vnode helper structure to hold poll-related state.
4958  */
4959 void
4960 v_addpollinfo(struct vnode *vp)
4961 {
4962 	struct vpollinfo *vi;
4963 
4964 	if (vp->v_pollinfo != NULL)
4965 		return;
4966 	vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO);
4967 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4968 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4969 	    vfs_knlunlock, vfs_knl_assert_lock);
4970 	VI_LOCK(vp);
4971 	if (vp->v_pollinfo != NULL) {
4972 		VI_UNLOCK(vp);
4973 		destroy_vpollinfo_free(vi);
4974 		return;
4975 	}
4976 	vp->v_pollinfo = vi;
4977 	VI_UNLOCK(vp);
4978 }
4979 
4980 /*
4981  * Record a process's interest in events which might happen to
4982  * a vnode.  Because poll uses the historic select-style interface
4983  * internally, this routine serves as both the ``check for any
4984  * pending events'' and the ``record my interest in future events''
4985  * functions.  (These are done together, while the lock is held,
4986  * to avoid race conditions.)
4987  */
4988 int
4989 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4990 {
4991 
4992 	v_addpollinfo(vp);
4993 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4994 	if (vp->v_pollinfo->vpi_revents & events) {
4995 		/*
4996 		 * This leaves events we are not interested
4997 		 * in available for the other process which
4998 		 * which presumably had requested them
4999 		 * (otherwise they would never have been
5000 		 * recorded).
5001 		 */
5002 		events &= vp->v_pollinfo->vpi_revents;
5003 		vp->v_pollinfo->vpi_revents &= ~events;
5004 
5005 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
5006 		return (events);
5007 	}
5008 	vp->v_pollinfo->vpi_events |= events;
5009 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
5010 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
5011 	return (0);
5012 }
5013 
5014 /*
5015  * Routine to create and manage a filesystem syncer vnode.
5016  */
5017 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
5018 static int	sync_fsync(struct  vop_fsync_args *);
5019 static int	sync_inactive(struct  vop_inactive_args *);
5020 static int	sync_reclaim(struct  vop_reclaim_args *);
5021 
5022 static struct vop_vector sync_vnodeops = {
5023 	.vop_bypass =	VOP_EOPNOTSUPP,
5024 	.vop_close =	sync_close,		/* close */
5025 	.vop_fsync =	sync_fsync,		/* fsync */
5026 	.vop_inactive =	sync_inactive,	/* inactive */
5027 	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
5028 	.vop_reclaim =	sync_reclaim,	/* reclaim */
5029 	.vop_lock1 =	vop_stdlock,	/* lock */
5030 	.vop_unlock =	vop_stdunlock,	/* unlock */
5031 	.vop_islocked =	vop_stdislocked,	/* islocked */
5032 };
5033 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
5034 
5035 /*
5036  * Create a new filesystem syncer vnode for the specified mount point.
5037  */
5038 void
5039 vfs_allocate_syncvnode(struct mount *mp)
5040 {
5041 	struct vnode *vp;
5042 	struct bufobj *bo;
5043 	static long start, incr, next;
5044 	int error;
5045 
5046 	/* Allocate a new vnode */
5047 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
5048 	if (error != 0)
5049 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
5050 	vp->v_type = VNON;
5051 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5052 	vp->v_vflag |= VV_FORCEINSMQ;
5053 	error = insmntque1(vp, mp);
5054 	if (error != 0)
5055 		panic("vfs_allocate_syncvnode: insmntque() failed");
5056 	vp->v_vflag &= ~VV_FORCEINSMQ;
5057 	VOP_UNLOCK(vp);
5058 	/*
5059 	 * Place the vnode onto the syncer worklist. We attempt to
5060 	 * scatter them about on the list so that they will go off
5061 	 * at evenly distributed times even if all the filesystems
5062 	 * are mounted at once.
5063 	 */
5064 	next += incr;
5065 	if (next == 0 || next > syncer_maxdelay) {
5066 		start /= 2;
5067 		incr /= 2;
5068 		if (start == 0) {
5069 			start = syncer_maxdelay / 2;
5070 			incr = syncer_maxdelay;
5071 		}
5072 		next = start;
5073 	}
5074 	bo = &vp->v_bufobj;
5075 	BO_LOCK(bo);
5076 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
5077 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
5078 	mtx_lock(&sync_mtx);
5079 	sync_vnode_count++;
5080 	if (mp->mnt_syncer == NULL) {
5081 		mp->mnt_syncer = vp;
5082 		vp = NULL;
5083 	}
5084 	mtx_unlock(&sync_mtx);
5085 	BO_UNLOCK(bo);
5086 	if (vp != NULL) {
5087 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5088 		vgone(vp);
5089 		vput(vp);
5090 	}
5091 }
5092 
5093 void
5094 vfs_deallocate_syncvnode(struct mount *mp)
5095 {
5096 	struct vnode *vp;
5097 
5098 	mtx_lock(&sync_mtx);
5099 	vp = mp->mnt_syncer;
5100 	if (vp != NULL)
5101 		mp->mnt_syncer = NULL;
5102 	mtx_unlock(&sync_mtx);
5103 	if (vp != NULL)
5104 		vrele(vp);
5105 }
5106 
5107 /*
5108  * Do a lazy sync of the filesystem.
5109  */
5110 static int
5111 sync_fsync(struct vop_fsync_args *ap)
5112 {
5113 	struct vnode *syncvp = ap->a_vp;
5114 	struct mount *mp = syncvp->v_mount;
5115 	int error, save;
5116 	struct bufobj *bo;
5117 
5118 	/*
5119 	 * We only need to do something if this is a lazy evaluation.
5120 	 */
5121 	if (ap->a_waitfor != MNT_LAZY)
5122 		return (0);
5123 
5124 	/*
5125 	 * Move ourselves to the back of the sync list.
5126 	 */
5127 	bo = &syncvp->v_bufobj;
5128 	BO_LOCK(bo);
5129 	vn_syncer_add_to_worklist(bo, syncdelay);
5130 	BO_UNLOCK(bo);
5131 
5132 	/*
5133 	 * Walk the list of vnodes pushing all that are dirty and
5134 	 * not already on the sync list.
5135 	 */
5136 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
5137 		return (0);
5138 	VOP_UNLOCK(syncvp);
5139 	save = curthread_pflags_set(TDP_SYNCIO);
5140 	/*
5141 	 * The filesystem at hand may be idle with free vnodes stored in the
5142 	 * batch.  Return them instead of letting them stay there indefinitely.
5143 	 */
5144 	vfs_periodic(mp, MNT_NOWAIT);
5145 	error = VFS_SYNC(mp, MNT_LAZY);
5146 	curthread_pflags_restore(save);
5147 	vn_lock(syncvp, LK_EXCLUSIVE | LK_RETRY);
5148 	vfs_unbusy(mp);
5149 	return (error);
5150 }
5151 
5152 /*
5153  * The syncer vnode is no referenced.
5154  */
5155 static int
5156 sync_inactive(struct vop_inactive_args *ap)
5157 {
5158 
5159 	vgone(ap->a_vp);
5160 	return (0);
5161 }
5162 
5163 /*
5164  * The syncer vnode is no longer needed and is being decommissioned.
5165  *
5166  * Modifications to the worklist must be protected by sync_mtx.
5167  */
5168 static int
5169 sync_reclaim(struct vop_reclaim_args *ap)
5170 {
5171 	struct vnode *vp = ap->a_vp;
5172 	struct bufobj *bo;
5173 
5174 	bo = &vp->v_bufobj;
5175 	BO_LOCK(bo);
5176 	mtx_lock(&sync_mtx);
5177 	if (vp->v_mount->mnt_syncer == vp)
5178 		vp->v_mount->mnt_syncer = NULL;
5179 	if (bo->bo_flag & BO_ONWORKLST) {
5180 		LIST_REMOVE(bo, bo_synclist);
5181 		syncer_worklist_len--;
5182 		sync_vnode_count--;
5183 		bo->bo_flag &= ~BO_ONWORKLST;
5184 	}
5185 	mtx_unlock(&sync_mtx);
5186 	BO_UNLOCK(bo);
5187 
5188 	return (0);
5189 }
5190 
5191 int
5192 vn_need_pageq_flush(struct vnode *vp)
5193 {
5194 	struct vm_object *obj;
5195 
5196 	obj = vp->v_object;
5197 	return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5198 	    vm_object_mightbedirty(obj));
5199 }
5200 
5201 /*
5202  * Check if vnode represents a disk device
5203  */
5204 bool
5205 vn_isdisk_error(struct vnode *vp, int *errp)
5206 {
5207 	int error;
5208 
5209 	if (vp->v_type != VCHR) {
5210 		error = ENOTBLK;
5211 		goto out;
5212 	}
5213 	error = 0;
5214 	dev_lock();
5215 	if (vp->v_rdev == NULL)
5216 		error = ENXIO;
5217 	else if (vp->v_rdev->si_devsw == NULL)
5218 		error = ENXIO;
5219 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5220 		error = ENOTBLK;
5221 	dev_unlock();
5222 out:
5223 	*errp = error;
5224 	return (error == 0);
5225 }
5226 
5227 bool
5228 vn_isdisk(struct vnode *vp)
5229 {
5230 	int error;
5231 
5232 	return (vn_isdisk_error(vp, &error));
5233 }
5234 
5235 /*
5236  * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
5237  * the comment above cache_fplookup for details.
5238  */
5239 int
5240 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred)
5241 {
5242 	int error;
5243 
5244 	VFS_SMR_ASSERT_ENTERED();
5245 
5246 	/* Check the owner. */
5247 	if (cred->cr_uid == file_uid) {
5248 		if (file_mode & S_IXUSR)
5249 			return (0);
5250 		goto out_error;
5251 	}
5252 
5253 	/* Otherwise, check the groups (first match) */
5254 	if (groupmember(file_gid, cred)) {
5255 		if (file_mode & S_IXGRP)
5256 			return (0);
5257 		goto out_error;
5258 	}
5259 
5260 	/* Otherwise, check everyone else. */
5261 	if (file_mode & S_IXOTH)
5262 		return (0);
5263 out_error:
5264 	/*
5265 	 * Permission check failed, but it is possible denial will get overwritten
5266 	 * (e.g., when root is traversing through a 700 directory owned by someone
5267 	 * else).
5268 	 *
5269 	 * vaccess() calls priv_check_cred which in turn can descent into MAC
5270 	 * modules overriding this result. It's quite unclear what semantics
5271 	 * are allowed for them to operate, thus for safety we don't call them
5272 	 * from within the SMR section. This also means if any such modules
5273 	 * are present, we have to let the regular lookup decide.
5274 	 */
5275 	error = priv_check_cred_vfs_lookup_nomac(cred);
5276 	switch (error) {
5277 	case 0:
5278 		return (0);
5279 	case EAGAIN:
5280 		/*
5281 		 * MAC modules present.
5282 		 */
5283 		return (EAGAIN);
5284 	case EPERM:
5285 		return (EACCES);
5286 	default:
5287 		return (error);
5288 	}
5289 }
5290 
5291 /*
5292  * Common filesystem object access control check routine.  Accepts a
5293  * vnode's type, "mode", uid and gid, requested access mode, and credentials.
5294  * Returns 0 on success, or an errno on failure.
5295  */
5296 int
5297 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5298     accmode_t accmode, struct ucred *cred)
5299 {
5300 	accmode_t dac_granted;
5301 	accmode_t priv_granted;
5302 
5303 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5304 	    ("invalid bit in accmode"));
5305 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5306 	    ("VAPPEND without VWRITE"));
5307 
5308 	/*
5309 	 * Look for a normal, non-privileged way to access the file/directory
5310 	 * as requested.  If it exists, go with that.
5311 	 */
5312 
5313 	dac_granted = 0;
5314 
5315 	/* Check the owner. */
5316 	if (cred->cr_uid == file_uid) {
5317 		dac_granted |= VADMIN;
5318 		if (file_mode & S_IXUSR)
5319 			dac_granted |= VEXEC;
5320 		if (file_mode & S_IRUSR)
5321 			dac_granted |= VREAD;
5322 		if (file_mode & S_IWUSR)
5323 			dac_granted |= (VWRITE | VAPPEND);
5324 
5325 		if ((accmode & dac_granted) == accmode)
5326 			return (0);
5327 
5328 		goto privcheck;
5329 	}
5330 
5331 	/* Otherwise, check the groups (first match) */
5332 	if (groupmember(file_gid, cred)) {
5333 		if (file_mode & S_IXGRP)
5334 			dac_granted |= VEXEC;
5335 		if (file_mode & S_IRGRP)
5336 			dac_granted |= VREAD;
5337 		if (file_mode & S_IWGRP)
5338 			dac_granted |= (VWRITE | VAPPEND);
5339 
5340 		if ((accmode & dac_granted) == accmode)
5341 			return (0);
5342 
5343 		goto privcheck;
5344 	}
5345 
5346 	/* Otherwise, check everyone else. */
5347 	if (file_mode & S_IXOTH)
5348 		dac_granted |= VEXEC;
5349 	if (file_mode & S_IROTH)
5350 		dac_granted |= VREAD;
5351 	if (file_mode & S_IWOTH)
5352 		dac_granted |= (VWRITE | VAPPEND);
5353 	if ((accmode & dac_granted) == accmode)
5354 		return (0);
5355 
5356 privcheck:
5357 	/*
5358 	 * Build a privilege mask to determine if the set of privileges
5359 	 * satisfies the requirements when combined with the granted mask
5360 	 * from above.  For each privilege, if the privilege is required,
5361 	 * bitwise or the request type onto the priv_granted mask.
5362 	 */
5363 	priv_granted = 0;
5364 
5365 	if (type == VDIR) {
5366 		/*
5367 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5368 		 * requests, instead of PRIV_VFS_EXEC.
5369 		 */
5370 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5371 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5372 			priv_granted |= VEXEC;
5373 	} else {
5374 		/*
5375 		 * Ensure that at least one execute bit is on. Otherwise,
5376 		 * a privileged user will always succeed, and we don't want
5377 		 * this to happen unless the file really is executable.
5378 		 */
5379 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5380 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5381 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5382 			priv_granted |= VEXEC;
5383 	}
5384 
5385 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5386 	    !priv_check_cred(cred, PRIV_VFS_READ))
5387 		priv_granted |= VREAD;
5388 
5389 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5390 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5391 		priv_granted |= (VWRITE | VAPPEND);
5392 
5393 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5394 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5395 		priv_granted |= VADMIN;
5396 
5397 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5398 		return (0);
5399 	}
5400 
5401 	return ((accmode & VADMIN) ? EPERM : EACCES);
5402 }
5403 
5404 /*
5405  * Credential check based on process requesting service, and per-attribute
5406  * permissions.
5407  */
5408 int
5409 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5410     struct thread *td, accmode_t accmode)
5411 {
5412 
5413 	/*
5414 	 * Kernel-invoked always succeeds.
5415 	 */
5416 	if (cred == NOCRED)
5417 		return (0);
5418 
5419 	/*
5420 	 * Do not allow privileged processes in jail to directly manipulate
5421 	 * system attributes.
5422 	 */
5423 	switch (attrnamespace) {
5424 	case EXTATTR_NAMESPACE_SYSTEM:
5425 		/* Potentially should be: return (EPERM); */
5426 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5427 	case EXTATTR_NAMESPACE_USER:
5428 		return (VOP_ACCESS(vp, accmode, cred, td));
5429 	default:
5430 		return (EPERM);
5431 	}
5432 }
5433 
5434 #ifdef DEBUG_VFS_LOCKS
5435 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5436 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5437     "Drop into debugger on lock violation");
5438 
5439 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5440 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5441     0, "Check for interlock across VOPs");
5442 
5443 int vfs_badlock_print = 1;	/* Print lock violations. */
5444 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5445     0, "Print lock violations");
5446 
5447 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5448 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5449     0, "Print vnode details on lock violations");
5450 
5451 #ifdef KDB
5452 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5453 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5454     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5455 #endif
5456 
5457 static void
5458 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5459 {
5460 
5461 #ifdef KDB
5462 	if (vfs_badlock_backtrace)
5463 		kdb_backtrace();
5464 #endif
5465 	if (vfs_badlock_vnode)
5466 		vn_printf(vp, "vnode ");
5467 	if (vfs_badlock_print)
5468 		printf("%s: %p %s\n", str, (void *)vp, msg);
5469 	if (vfs_badlock_ddb)
5470 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5471 }
5472 
5473 void
5474 assert_vi_locked(struct vnode *vp, const char *str)
5475 {
5476 
5477 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5478 		vfs_badlock("interlock is not locked but should be", str, vp);
5479 }
5480 
5481 void
5482 assert_vi_unlocked(struct vnode *vp, const char *str)
5483 {
5484 
5485 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5486 		vfs_badlock("interlock is locked but should not be", str, vp);
5487 }
5488 
5489 void
5490 assert_vop_locked(struct vnode *vp, const char *str)
5491 {
5492 	int locked;
5493 
5494 	if (KERNEL_PANICKED() || vp == NULL)
5495 		return;
5496 
5497 	locked = VOP_ISLOCKED(vp);
5498 	if (locked == 0 || locked == LK_EXCLOTHER)
5499 		vfs_badlock("is not locked but should be", str, vp);
5500 }
5501 
5502 void
5503 assert_vop_unlocked(struct vnode *vp, const char *str)
5504 {
5505 	if (KERNEL_PANICKED() || vp == NULL)
5506 		return;
5507 
5508 	if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5509 		vfs_badlock("is locked but should not be", str, vp);
5510 }
5511 
5512 void
5513 assert_vop_elocked(struct vnode *vp, const char *str)
5514 {
5515 	if (KERNEL_PANICKED() || vp == NULL)
5516 		return;
5517 
5518 	if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5519 		vfs_badlock("is not exclusive locked but should be", str, vp);
5520 }
5521 #endif /* DEBUG_VFS_LOCKS */
5522 
5523 void
5524 vop_rename_fail(struct vop_rename_args *ap)
5525 {
5526 
5527 	if (ap->a_tvp != NULL)
5528 		vput(ap->a_tvp);
5529 	if (ap->a_tdvp == ap->a_tvp)
5530 		vrele(ap->a_tdvp);
5531 	else
5532 		vput(ap->a_tdvp);
5533 	vrele(ap->a_fdvp);
5534 	vrele(ap->a_fvp);
5535 }
5536 
5537 void
5538 vop_rename_pre(void *ap)
5539 {
5540 	struct vop_rename_args *a = ap;
5541 
5542 #ifdef DEBUG_VFS_LOCKS
5543 	if (a->a_tvp)
5544 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5545 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5546 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5547 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5548 
5549 	/* Check the source (from). */
5550 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5551 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5552 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5553 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5554 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5555 
5556 	/* Check the target. */
5557 	if (a->a_tvp)
5558 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5559 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5560 #endif
5561 	/*
5562 	 * It may be tempting to add vn_seqc_write_begin/end calls here and
5563 	 * in vop_rename_post but that's not going to work out since some
5564 	 * filesystems relookup vnodes mid-rename. This is probably a bug.
5565 	 *
5566 	 * For now filesystems are expected to do the relevant calls after they
5567 	 * decide what vnodes to operate on.
5568 	 */
5569 	if (a->a_tdvp != a->a_fdvp)
5570 		vhold(a->a_fdvp);
5571 	if (a->a_tvp != a->a_fvp)
5572 		vhold(a->a_fvp);
5573 	vhold(a->a_tdvp);
5574 	if (a->a_tvp)
5575 		vhold(a->a_tvp);
5576 }
5577 
5578 #ifdef DEBUG_VFS_LOCKS
5579 void
5580 vop_fplookup_vexec_debugpre(void *ap __unused)
5581 {
5582 
5583 	VFS_SMR_ASSERT_ENTERED();
5584 }
5585 
5586 void
5587 vop_fplookup_vexec_debugpost(void *ap __unused, int rc __unused)
5588 {
5589 
5590 	VFS_SMR_ASSERT_ENTERED();
5591 }
5592 
5593 void
5594 vop_fplookup_symlink_debugpre(void *ap __unused)
5595 {
5596 
5597 	VFS_SMR_ASSERT_ENTERED();
5598 }
5599 
5600 void
5601 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused)
5602 {
5603 
5604 	VFS_SMR_ASSERT_ENTERED();
5605 }
5606 
5607 static void
5608 vop_fsync_debugprepost(struct vnode *vp, const char *name)
5609 {
5610 	if (vp->v_type == VCHR)
5611 		;
5612 	else if (MNT_EXTENDED_SHARED(vp->v_mount))
5613 		ASSERT_VOP_LOCKED(vp, name);
5614 	else
5615 		ASSERT_VOP_ELOCKED(vp, name);
5616 }
5617 
5618 void
5619 vop_fsync_debugpre(void *a)
5620 {
5621 	struct vop_fsync_args *ap;
5622 
5623 	ap = a;
5624 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5625 }
5626 
5627 void
5628 vop_fsync_debugpost(void *a, int rc __unused)
5629 {
5630 	struct vop_fsync_args *ap;
5631 
5632 	ap = a;
5633 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5634 }
5635 
5636 void
5637 vop_fdatasync_debugpre(void *a)
5638 {
5639 	struct vop_fdatasync_args *ap;
5640 
5641 	ap = a;
5642 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5643 }
5644 
5645 void
5646 vop_fdatasync_debugpost(void *a, int rc __unused)
5647 {
5648 	struct vop_fdatasync_args *ap;
5649 
5650 	ap = a;
5651 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5652 }
5653 
5654 void
5655 vop_strategy_debugpre(void *ap)
5656 {
5657 	struct vop_strategy_args *a;
5658 	struct buf *bp;
5659 
5660 	a = ap;
5661 	bp = a->a_bp;
5662 
5663 	/*
5664 	 * Cluster ops lock their component buffers but not the IO container.
5665 	 */
5666 	if ((bp->b_flags & B_CLUSTER) != 0)
5667 		return;
5668 
5669 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5670 		if (vfs_badlock_print)
5671 			printf(
5672 			    "VOP_STRATEGY: bp is not locked but should be\n");
5673 		if (vfs_badlock_ddb)
5674 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5675 	}
5676 }
5677 
5678 void
5679 vop_lock_debugpre(void *ap)
5680 {
5681 	struct vop_lock1_args *a = ap;
5682 
5683 	if ((a->a_flags & LK_INTERLOCK) == 0)
5684 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5685 	else
5686 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5687 }
5688 
5689 void
5690 vop_lock_debugpost(void *ap, int rc)
5691 {
5692 	struct vop_lock1_args *a = ap;
5693 
5694 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5695 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5696 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5697 }
5698 
5699 void
5700 vop_unlock_debugpre(void *ap)
5701 {
5702 	struct vop_unlock_args *a = ap;
5703 
5704 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5705 }
5706 
5707 void
5708 vop_need_inactive_debugpre(void *ap)
5709 {
5710 	struct vop_need_inactive_args *a = ap;
5711 
5712 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5713 }
5714 
5715 void
5716 vop_need_inactive_debugpost(void *ap, int rc)
5717 {
5718 	struct vop_need_inactive_args *a = ap;
5719 
5720 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5721 }
5722 #endif
5723 
5724 void
5725 vop_create_pre(void *ap)
5726 {
5727 	struct vop_create_args *a;
5728 	struct vnode *dvp;
5729 
5730 	a = ap;
5731 	dvp = a->a_dvp;
5732 	vn_seqc_write_begin(dvp);
5733 }
5734 
5735 void
5736 vop_create_post(void *ap, int rc)
5737 {
5738 	struct vop_create_args *a;
5739 	struct vnode *dvp;
5740 
5741 	a = ap;
5742 	dvp = a->a_dvp;
5743 	vn_seqc_write_end(dvp);
5744 	if (!rc)
5745 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5746 }
5747 
5748 void
5749 vop_whiteout_pre(void *ap)
5750 {
5751 	struct vop_whiteout_args *a;
5752 	struct vnode *dvp;
5753 
5754 	a = ap;
5755 	dvp = a->a_dvp;
5756 	vn_seqc_write_begin(dvp);
5757 }
5758 
5759 void
5760 vop_whiteout_post(void *ap, int rc)
5761 {
5762 	struct vop_whiteout_args *a;
5763 	struct vnode *dvp;
5764 
5765 	a = ap;
5766 	dvp = a->a_dvp;
5767 	vn_seqc_write_end(dvp);
5768 }
5769 
5770 void
5771 vop_deleteextattr_pre(void *ap)
5772 {
5773 	struct vop_deleteextattr_args *a;
5774 	struct vnode *vp;
5775 
5776 	a = ap;
5777 	vp = a->a_vp;
5778 	vn_seqc_write_begin(vp);
5779 }
5780 
5781 void
5782 vop_deleteextattr_post(void *ap, int rc)
5783 {
5784 	struct vop_deleteextattr_args *a;
5785 	struct vnode *vp;
5786 
5787 	a = ap;
5788 	vp = a->a_vp;
5789 	vn_seqc_write_end(vp);
5790 	if (!rc)
5791 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5792 }
5793 
5794 void
5795 vop_link_pre(void *ap)
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_begin(vp);
5804 	vn_seqc_write_begin(tdvp);
5805 }
5806 
5807 void
5808 vop_link_post(void *ap, int rc)
5809 {
5810 	struct vop_link_args *a;
5811 	struct vnode *vp, *tdvp;
5812 
5813 	a = ap;
5814 	vp = a->a_vp;
5815 	tdvp = a->a_tdvp;
5816 	vn_seqc_write_end(vp);
5817 	vn_seqc_write_end(tdvp);
5818 	if (!rc) {
5819 		VFS_KNOTE_LOCKED(vp, NOTE_LINK);
5820 		VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE);
5821 	}
5822 }
5823 
5824 void
5825 vop_mkdir_pre(void *ap)
5826 {
5827 	struct vop_mkdir_args *a;
5828 	struct vnode *dvp;
5829 
5830 	a = ap;
5831 	dvp = a->a_dvp;
5832 	vn_seqc_write_begin(dvp);
5833 }
5834 
5835 void
5836 vop_mkdir_post(void *ap, int rc)
5837 {
5838 	struct vop_mkdir_args *a;
5839 	struct vnode *dvp;
5840 
5841 	a = ap;
5842 	dvp = a->a_dvp;
5843 	vn_seqc_write_end(dvp);
5844 	if (!rc)
5845 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5846 }
5847 
5848 #ifdef DEBUG_VFS_LOCKS
5849 void
5850 vop_mkdir_debugpost(void *ap, int rc)
5851 {
5852 	struct vop_mkdir_args *a;
5853 
5854 	a = ap;
5855 	if (!rc)
5856 		cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp);
5857 }
5858 #endif
5859 
5860 void
5861 vop_mknod_pre(void *ap)
5862 {
5863 	struct vop_mknod_args *a;
5864 	struct vnode *dvp;
5865 
5866 	a = ap;
5867 	dvp = a->a_dvp;
5868 	vn_seqc_write_begin(dvp);
5869 }
5870 
5871 void
5872 vop_mknod_post(void *ap, int rc)
5873 {
5874 	struct vop_mknod_args *a;
5875 	struct vnode *dvp;
5876 
5877 	a = ap;
5878 	dvp = a->a_dvp;
5879 	vn_seqc_write_end(dvp);
5880 	if (!rc)
5881 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5882 }
5883 
5884 void
5885 vop_reclaim_post(void *ap, int rc)
5886 {
5887 	struct vop_reclaim_args *a;
5888 	struct vnode *vp;
5889 
5890 	a = ap;
5891 	vp = a->a_vp;
5892 	ASSERT_VOP_IN_SEQC(vp);
5893 	if (!rc)
5894 		VFS_KNOTE_LOCKED(vp, NOTE_REVOKE);
5895 }
5896 
5897 void
5898 vop_remove_pre(void *ap)
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_begin(dvp);
5907 	vn_seqc_write_begin(vp);
5908 }
5909 
5910 void
5911 vop_remove_post(void *ap, int rc)
5912 {
5913 	struct vop_remove_args *a;
5914 	struct vnode *dvp, *vp;
5915 
5916 	a = ap;
5917 	dvp = a->a_dvp;
5918 	vp = a->a_vp;
5919 	vn_seqc_write_end(dvp);
5920 	vn_seqc_write_end(vp);
5921 	if (!rc) {
5922 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5923 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5924 	}
5925 }
5926 
5927 void
5928 vop_rename_post(void *ap, int rc)
5929 {
5930 	struct vop_rename_args *a = ap;
5931 	long hint;
5932 
5933 	if (!rc) {
5934 		hint = NOTE_WRITE;
5935 		if (a->a_fdvp == a->a_tdvp) {
5936 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5937 				hint |= NOTE_LINK;
5938 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5939 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5940 		} else {
5941 			hint |= NOTE_EXTEND;
5942 			if (a->a_fvp->v_type == VDIR)
5943 				hint |= NOTE_LINK;
5944 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5945 
5946 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5947 			    a->a_tvp->v_type == VDIR)
5948 				hint &= ~NOTE_LINK;
5949 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5950 		}
5951 
5952 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5953 		if (a->a_tvp)
5954 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5955 	}
5956 	if (a->a_tdvp != a->a_fdvp)
5957 		vdrop(a->a_fdvp);
5958 	if (a->a_tvp != a->a_fvp)
5959 		vdrop(a->a_fvp);
5960 	vdrop(a->a_tdvp);
5961 	if (a->a_tvp)
5962 		vdrop(a->a_tvp);
5963 }
5964 
5965 void
5966 vop_rmdir_pre(void *ap)
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_begin(dvp);
5975 	vn_seqc_write_begin(vp);
5976 }
5977 
5978 void
5979 vop_rmdir_post(void *ap, int rc)
5980 {
5981 	struct vop_rmdir_args *a;
5982 	struct vnode *dvp, *vp;
5983 
5984 	a = ap;
5985 	dvp = a->a_dvp;
5986 	vp = a->a_vp;
5987 	vn_seqc_write_end(dvp);
5988 	vn_seqc_write_end(vp);
5989 	if (!rc) {
5990 		vp->v_vflag |= VV_UNLINKED;
5991 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5992 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5993 	}
5994 }
5995 
5996 void
5997 vop_setattr_pre(void *ap)
5998 {
5999 	struct vop_setattr_args *a;
6000 	struct vnode *vp;
6001 
6002 	a = ap;
6003 	vp = a->a_vp;
6004 	vn_seqc_write_begin(vp);
6005 }
6006 
6007 void
6008 vop_setattr_post(void *ap, int rc)
6009 {
6010 	struct vop_setattr_args *a;
6011 	struct vnode *vp;
6012 
6013 	a = ap;
6014 	vp = a->a_vp;
6015 	vn_seqc_write_end(vp);
6016 	if (!rc)
6017 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
6018 }
6019 
6020 void
6021 vop_setacl_pre(void *ap)
6022 {
6023 	struct vop_setacl_args *a;
6024 	struct vnode *vp;
6025 
6026 	a = ap;
6027 	vp = a->a_vp;
6028 	vn_seqc_write_begin(vp);
6029 }
6030 
6031 void
6032 vop_setacl_post(void *ap, int rc __unused)
6033 {
6034 	struct vop_setacl_args *a;
6035 	struct vnode *vp;
6036 
6037 	a = ap;
6038 	vp = a->a_vp;
6039 	vn_seqc_write_end(vp);
6040 }
6041 
6042 void
6043 vop_setextattr_pre(void *ap)
6044 {
6045 	struct vop_setextattr_args *a;
6046 	struct vnode *vp;
6047 
6048 	a = ap;
6049 	vp = a->a_vp;
6050 	vn_seqc_write_begin(vp);
6051 }
6052 
6053 void
6054 vop_setextattr_post(void *ap, int rc)
6055 {
6056 	struct vop_setextattr_args *a;
6057 	struct vnode *vp;
6058 
6059 	a = ap;
6060 	vp = a->a_vp;
6061 	vn_seqc_write_end(vp);
6062 	if (!rc)
6063 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
6064 }
6065 
6066 void
6067 vop_symlink_pre(void *ap)
6068 {
6069 	struct vop_symlink_args *a;
6070 	struct vnode *dvp;
6071 
6072 	a = ap;
6073 	dvp = a->a_dvp;
6074 	vn_seqc_write_begin(dvp);
6075 }
6076 
6077 void
6078 vop_symlink_post(void *ap, int rc)
6079 {
6080 	struct vop_symlink_args *a;
6081 	struct vnode *dvp;
6082 
6083 	a = ap;
6084 	dvp = a->a_dvp;
6085 	vn_seqc_write_end(dvp);
6086 	if (!rc)
6087 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6088 }
6089 
6090 void
6091 vop_open_post(void *ap, int rc)
6092 {
6093 	struct vop_open_args *a = ap;
6094 
6095 	if (!rc)
6096 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
6097 }
6098 
6099 void
6100 vop_close_post(void *ap, int rc)
6101 {
6102 	struct vop_close_args *a = ap;
6103 
6104 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
6105 	    !VN_IS_DOOMED(a->a_vp))) {
6106 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
6107 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
6108 	}
6109 }
6110 
6111 void
6112 vop_read_post(void *ap, int rc)
6113 {
6114 	struct vop_read_args *a = ap;
6115 
6116 	if (!rc)
6117 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6118 }
6119 
6120 void
6121 vop_read_pgcache_post(void *ap, int rc)
6122 {
6123 	struct vop_read_pgcache_args *a = ap;
6124 
6125 	if (!rc)
6126 		VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ);
6127 }
6128 
6129 void
6130 vop_readdir_post(void *ap, int rc)
6131 {
6132 	struct vop_readdir_args *a = ap;
6133 
6134 	if (!rc)
6135 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6136 }
6137 
6138 static struct knlist fs_knlist;
6139 
6140 static void
6141 vfs_event_init(void *arg)
6142 {
6143 	knlist_init_mtx(&fs_knlist, NULL);
6144 }
6145 /* XXX - correct order? */
6146 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
6147 
6148 void
6149 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
6150 {
6151 
6152 	KNOTE_UNLOCKED(&fs_knlist, event);
6153 }
6154 
6155 static int	filt_fsattach(struct knote *kn);
6156 static void	filt_fsdetach(struct knote *kn);
6157 static int	filt_fsevent(struct knote *kn, long hint);
6158 
6159 struct filterops fs_filtops = {
6160 	.f_isfd = 0,
6161 	.f_attach = filt_fsattach,
6162 	.f_detach = filt_fsdetach,
6163 	.f_event = filt_fsevent
6164 };
6165 
6166 static int
6167 filt_fsattach(struct knote *kn)
6168 {
6169 
6170 	kn->kn_flags |= EV_CLEAR;
6171 	knlist_add(&fs_knlist, kn, 0);
6172 	return (0);
6173 }
6174 
6175 static void
6176 filt_fsdetach(struct knote *kn)
6177 {
6178 
6179 	knlist_remove(&fs_knlist, kn, 0);
6180 }
6181 
6182 static int
6183 filt_fsevent(struct knote *kn, long hint)
6184 {
6185 
6186 	kn->kn_fflags |= kn->kn_sfflags & hint;
6187 
6188 	return (kn->kn_fflags != 0);
6189 }
6190 
6191 static int
6192 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
6193 {
6194 	struct vfsidctl vc;
6195 	int error;
6196 	struct mount *mp;
6197 
6198 	error = SYSCTL_IN(req, &vc, sizeof(vc));
6199 	if (error)
6200 		return (error);
6201 	if (vc.vc_vers != VFS_CTL_VERS1)
6202 		return (EINVAL);
6203 	mp = vfs_getvfs(&vc.vc_fsid);
6204 	if (mp == NULL)
6205 		return (ENOENT);
6206 	/* ensure that a specific sysctl goes to the right filesystem. */
6207 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
6208 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
6209 		vfs_rel(mp);
6210 		return (EINVAL);
6211 	}
6212 	VCTLTOREQ(&vc, req);
6213 	error = VFS_SYSCTL(mp, vc.vc_op, req);
6214 	vfs_rel(mp);
6215 	return (error);
6216 }
6217 
6218 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
6219     NULL, 0, sysctl_vfs_ctl, "",
6220     "Sysctl by fsid");
6221 
6222 /*
6223  * Function to initialize a va_filerev field sensibly.
6224  * XXX: Wouldn't a random number make a lot more sense ??
6225  */
6226 u_quad_t
6227 init_va_filerev(void)
6228 {
6229 	struct bintime bt;
6230 
6231 	getbinuptime(&bt);
6232 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
6233 }
6234 
6235 static int	filt_vfsread(struct knote *kn, long hint);
6236 static int	filt_vfswrite(struct knote *kn, long hint);
6237 static int	filt_vfsvnode(struct knote *kn, long hint);
6238 static void	filt_vfsdetach(struct knote *kn);
6239 static struct filterops vfsread_filtops = {
6240 	.f_isfd = 1,
6241 	.f_detach = filt_vfsdetach,
6242 	.f_event = filt_vfsread
6243 };
6244 static struct filterops vfswrite_filtops = {
6245 	.f_isfd = 1,
6246 	.f_detach = filt_vfsdetach,
6247 	.f_event = filt_vfswrite
6248 };
6249 static struct filterops vfsvnode_filtops = {
6250 	.f_isfd = 1,
6251 	.f_detach = filt_vfsdetach,
6252 	.f_event = filt_vfsvnode
6253 };
6254 
6255 static void
6256 vfs_knllock(void *arg)
6257 {
6258 	struct vnode *vp = arg;
6259 
6260 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
6261 }
6262 
6263 static void
6264 vfs_knlunlock(void *arg)
6265 {
6266 	struct vnode *vp = arg;
6267 
6268 	VOP_UNLOCK(vp);
6269 }
6270 
6271 static void
6272 vfs_knl_assert_lock(void *arg, int what)
6273 {
6274 #ifdef DEBUG_VFS_LOCKS
6275 	struct vnode *vp = arg;
6276 
6277 	if (what == LA_LOCKED)
6278 		ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
6279 	else
6280 		ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
6281 #endif
6282 }
6283 
6284 int
6285 vfs_kqfilter(struct vop_kqfilter_args *ap)
6286 {
6287 	struct vnode *vp = ap->a_vp;
6288 	struct knote *kn = ap->a_kn;
6289 	struct knlist *knl;
6290 
6291 	KASSERT(vp->v_type != VFIFO || (kn->kn_filter != EVFILT_READ &&
6292 	    kn->kn_filter != EVFILT_WRITE),
6293 	    ("READ/WRITE filter on a FIFO leaked through"));
6294 	switch (kn->kn_filter) {
6295 	case EVFILT_READ:
6296 		kn->kn_fop = &vfsread_filtops;
6297 		break;
6298 	case EVFILT_WRITE:
6299 		kn->kn_fop = &vfswrite_filtops;
6300 		break;
6301 	case EVFILT_VNODE:
6302 		kn->kn_fop = &vfsvnode_filtops;
6303 		break;
6304 	default:
6305 		return (EINVAL);
6306 	}
6307 
6308 	kn->kn_hook = (caddr_t)vp;
6309 
6310 	v_addpollinfo(vp);
6311 	if (vp->v_pollinfo == NULL)
6312 		return (ENOMEM);
6313 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
6314 	vhold(vp);
6315 	knlist_add(knl, kn, 0);
6316 
6317 	return (0);
6318 }
6319 
6320 /*
6321  * Detach knote from vnode
6322  */
6323 static void
6324 filt_vfsdetach(struct knote *kn)
6325 {
6326 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6327 
6328 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
6329 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
6330 	vdrop(vp);
6331 }
6332 
6333 /*ARGSUSED*/
6334 static int
6335 filt_vfsread(struct knote *kn, long hint)
6336 {
6337 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6338 	struct vattr va;
6339 	int res;
6340 
6341 	/*
6342 	 * filesystem is gone, so set the EOF flag and schedule
6343 	 * the knote for deletion.
6344 	 */
6345 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6346 		VI_LOCK(vp);
6347 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6348 		VI_UNLOCK(vp);
6349 		return (1);
6350 	}
6351 
6352 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
6353 		return (0);
6354 
6355 	VI_LOCK(vp);
6356 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
6357 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
6358 	VI_UNLOCK(vp);
6359 	return (res);
6360 }
6361 
6362 /*ARGSUSED*/
6363 static int
6364 filt_vfswrite(struct knote *kn, long hint)
6365 {
6366 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6367 
6368 	VI_LOCK(vp);
6369 
6370 	/*
6371 	 * filesystem is gone, so set the EOF flag and schedule
6372 	 * the knote for deletion.
6373 	 */
6374 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
6375 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6376 
6377 	kn->kn_data = 0;
6378 	VI_UNLOCK(vp);
6379 	return (1);
6380 }
6381 
6382 static int
6383 filt_vfsvnode(struct knote *kn, long hint)
6384 {
6385 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6386 	int res;
6387 
6388 	VI_LOCK(vp);
6389 	if (kn->kn_sfflags & hint)
6390 		kn->kn_fflags |= hint;
6391 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6392 		kn->kn_flags |= EV_EOF;
6393 		VI_UNLOCK(vp);
6394 		return (1);
6395 	}
6396 	res = (kn->kn_fflags != 0);
6397 	VI_UNLOCK(vp);
6398 	return (res);
6399 }
6400 
6401 /*
6402  * Returns whether the directory is empty or not.
6403  * If it is empty, the return value is 0; otherwise
6404  * the return value is an error value (which may
6405  * be ENOTEMPTY).
6406  */
6407 int
6408 vfs_emptydir(struct vnode *vp)
6409 {
6410 	struct uio uio;
6411 	struct iovec iov;
6412 	struct dirent *dirent, *dp, *endp;
6413 	int error, eof;
6414 
6415 	error = 0;
6416 	eof = 0;
6417 
6418 	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
6419 	VNASSERT(vp->v_type == VDIR, vp, ("vp is not a directory"));
6420 
6421 	dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
6422 	iov.iov_base = dirent;
6423 	iov.iov_len = sizeof(struct dirent);
6424 
6425 	uio.uio_iov = &iov;
6426 	uio.uio_iovcnt = 1;
6427 	uio.uio_offset = 0;
6428 	uio.uio_resid = sizeof(struct dirent);
6429 	uio.uio_segflg = UIO_SYSSPACE;
6430 	uio.uio_rw = UIO_READ;
6431 	uio.uio_td = curthread;
6432 
6433 	while (eof == 0 && error == 0) {
6434 		error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
6435 		    NULL, NULL);
6436 		if (error != 0)
6437 			break;
6438 		endp = (void *)((uint8_t *)dirent +
6439 		    sizeof(struct dirent) - uio.uio_resid);
6440 		for (dp = dirent; dp < endp;
6441 		     dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
6442 			if (dp->d_type == DT_WHT)
6443 				continue;
6444 			if (dp->d_namlen == 0)
6445 				continue;
6446 			if (dp->d_type != DT_DIR &&
6447 			    dp->d_type != DT_UNKNOWN) {
6448 				error = ENOTEMPTY;
6449 				break;
6450 			}
6451 			if (dp->d_namlen > 2) {
6452 				error = ENOTEMPTY;
6453 				break;
6454 			}
6455 			if (dp->d_namlen == 1 &&
6456 			    dp->d_name[0] != '.') {
6457 				error = ENOTEMPTY;
6458 				break;
6459 			}
6460 			if (dp->d_namlen == 2 &&
6461 			    dp->d_name[1] != '.') {
6462 				error = ENOTEMPTY;
6463 				break;
6464 			}
6465 			uio.uio_resid = sizeof(struct dirent);
6466 		}
6467 	}
6468 	free(dirent, M_TEMP);
6469 	return (error);
6470 }
6471 
6472 int
6473 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6474 {
6475 	int error;
6476 
6477 	if (dp->d_reclen > ap->a_uio->uio_resid)
6478 		return (ENAMETOOLONG);
6479 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
6480 	if (error) {
6481 		if (ap->a_ncookies != NULL) {
6482 			if (ap->a_cookies != NULL)
6483 				free(ap->a_cookies, M_TEMP);
6484 			ap->a_cookies = NULL;
6485 			*ap->a_ncookies = 0;
6486 		}
6487 		return (error);
6488 	}
6489 	if (ap->a_ncookies == NULL)
6490 		return (0);
6491 
6492 	KASSERT(ap->a_cookies,
6493 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6494 
6495 	*ap->a_cookies = realloc(*ap->a_cookies,
6496 	    (*ap->a_ncookies + 1) * sizeof(uint64_t), M_TEMP, M_WAITOK | M_ZERO);
6497 	(*ap->a_cookies)[*ap->a_ncookies] = off;
6498 	*ap->a_ncookies += 1;
6499 	return (0);
6500 }
6501 
6502 /*
6503  * The purpose of this routine is to remove granularity from accmode_t,
6504  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6505  * VADMIN and VAPPEND.
6506  *
6507  * If it returns 0, the caller is supposed to continue with the usual
6508  * access checks using 'accmode' as modified by this routine.  If it
6509  * returns nonzero value, the caller is supposed to return that value
6510  * as errno.
6511  *
6512  * Note that after this routine runs, accmode may be zero.
6513  */
6514 int
6515 vfs_unixify_accmode(accmode_t *accmode)
6516 {
6517 	/*
6518 	 * There is no way to specify explicit "deny" rule using
6519 	 * file mode or POSIX.1e ACLs.
6520 	 */
6521 	if (*accmode & VEXPLICIT_DENY) {
6522 		*accmode = 0;
6523 		return (0);
6524 	}
6525 
6526 	/*
6527 	 * None of these can be translated into usual access bits.
6528 	 * Also, the common case for NFSv4 ACLs is to not contain
6529 	 * either of these bits. Caller should check for VWRITE
6530 	 * on the containing directory instead.
6531 	 */
6532 	if (*accmode & (VDELETE_CHILD | VDELETE))
6533 		return (EPERM);
6534 
6535 	if (*accmode & VADMIN_PERMS) {
6536 		*accmode &= ~VADMIN_PERMS;
6537 		*accmode |= VADMIN;
6538 	}
6539 
6540 	/*
6541 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6542 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6543 	 */
6544 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6545 
6546 	return (0);
6547 }
6548 
6549 /*
6550  * Clear out a doomed vnode (if any) and replace it with a new one as long
6551  * as the fs is not being unmounted. Return the root vnode to the caller.
6552  */
6553 static int __noinline
6554 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6555 {
6556 	struct vnode *vp;
6557 	int error;
6558 
6559 restart:
6560 	if (mp->mnt_rootvnode != NULL) {
6561 		MNT_ILOCK(mp);
6562 		vp = mp->mnt_rootvnode;
6563 		if (vp != NULL) {
6564 			if (!VN_IS_DOOMED(vp)) {
6565 				vrefact(vp);
6566 				MNT_IUNLOCK(mp);
6567 				error = vn_lock(vp, flags);
6568 				if (error == 0) {
6569 					*vpp = vp;
6570 					return (0);
6571 				}
6572 				vrele(vp);
6573 				goto restart;
6574 			}
6575 			/*
6576 			 * Clear the old one.
6577 			 */
6578 			mp->mnt_rootvnode = NULL;
6579 		}
6580 		MNT_IUNLOCK(mp);
6581 		if (vp != NULL) {
6582 			vfs_op_barrier_wait(mp);
6583 			vrele(vp);
6584 		}
6585 	}
6586 	error = VFS_CACHEDROOT(mp, flags, vpp);
6587 	if (error != 0)
6588 		return (error);
6589 	if (mp->mnt_vfs_ops == 0) {
6590 		MNT_ILOCK(mp);
6591 		if (mp->mnt_vfs_ops != 0) {
6592 			MNT_IUNLOCK(mp);
6593 			return (0);
6594 		}
6595 		if (mp->mnt_rootvnode == NULL) {
6596 			vrefact(*vpp);
6597 			mp->mnt_rootvnode = *vpp;
6598 		} else {
6599 			if (mp->mnt_rootvnode != *vpp) {
6600 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6601 					panic("%s: mismatch between vnode returned "
6602 					    " by VFS_CACHEDROOT and the one cached "
6603 					    " (%p != %p)",
6604 					    __func__, *vpp, mp->mnt_rootvnode);
6605 				}
6606 			}
6607 		}
6608 		MNT_IUNLOCK(mp);
6609 	}
6610 	return (0);
6611 }
6612 
6613 int
6614 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6615 {
6616 	struct mount_pcpu *mpcpu;
6617 	struct vnode *vp;
6618 	int error;
6619 
6620 	if (!vfs_op_thread_enter(mp, mpcpu))
6621 		return (vfs_cache_root_fallback(mp, flags, vpp));
6622 	vp = atomic_load_ptr(&mp->mnt_rootvnode);
6623 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6624 		vfs_op_thread_exit(mp, mpcpu);
6625 		return (vfs_cache_root_fallback(mp, flags, vpp));
6626 	}
6627 	vrefact(vp);
6628 	vfs_op_thread_exit(mp, mpcpu);
6629 	error = vn_lock(vp, flags);
6630 	if (error != 0) {
6631 		vrele(vp);
6632 		return (vfs_cache_root_fallback(mp, flags, vpp));
6633 	}
6634 	*vpp = vp;
6635 	return (0);
6636 }
6637 
6638 struct vnode *
6639 vfs_cache_root_clear(struct mount *mp)
6640 {
6641 	struct vnode *vp;
6642 
6643 	/*
6644 	 * ops > 0 guarantees there is nobody who can see this vnode
6645 	 */
6646 	MPASS(mp->mnt_vfs_ops > 0);
6647 	vp = mp->mnt_rootvnode;
6648 	if (vp != NULL)
6649 		vn_seqc_write_begin(vp);
6650 	mp->mnt_rootvnode = NULL;
6651 	return (vp);
6652 }
6653 
6654 void
6655 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6656 {
6657 
6658 	MPASS(mp->mnt_vfs_ops > 0);
6659 	vrefact(vp);
6660 	mp->mnt_rootvnode = vp;
6661 }
6662 
6663 /*
6664  * These are helper functions for filesystems to traverse all
6665  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6666  *
6667  * This interface replaces MNT_VNODE_FOREACH.
6668  */
6669 
6670 struct vnode *
6671 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6672 {
6673 	struct vnode *vp;
6674 
6675 	if (should_yield())
6676 		kern_yield(PRI_USER);
6677 	MNT_ILOCK(mp);
6678 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6679 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6680 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6681 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6682 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6683 			continue;
6684 		VI_LOCK(vp);
6685 		if (VN_IS_DOOMED(vp)) {
6686 			VI_UNLOCK(vp);
6687 			continue;
6688 		}
6689 		break;
6690 	}
6691 	if (vp == NULL) {
6692 		__mnt_vnode_markerfree_all(mvp, mp);
6693 		/* MNT_IUNLOCK(mp); -- done in above function */
6694 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6695 		return (NULL);
6696 	}
6697 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6698 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6699 	MNT_IUNLOCK(mp);
6700 	return (vp);
6701 }
6702 
6703 struct vnode *
6704 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6705 {
6706 	struct vnode *vp;
6707 
6708 	*mvp = vn_alloc_marker(mp);
6709 	MNT_ILOCK(mp);
6710 	MNT_REF(mp);
6711 
6712 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6713 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6714 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6715 			continue;
6716 		VI_LOCK(vp);
6717 		if (VN_IS_DOOMED(vp)) {
6718 			VI_UNLOCK(vp);
6719 			continue;
6720 		}
6721 		break;
6722 	}
6723 	if (vp == NULL) {
6724 		MNT_REL(mp);
6725 		MNT_IUNLOCK(mp);
6726 		vn_free_marker(*mvp);
6727 		*mvp = NULL;
6728 		return (NULL);
6729 	}
6730 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6731 	MNT_IUNLOCK(mp);
6732 	return (vp);
6733 }
6734 
6735 void
6736 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6737 {
6738 
6739 	if (*mvp == NULL) {
6740 		MNT_IUNLOCK(mp);
6741 		return;
6742 	}
6743 
6744 	mtx_assert(MNT_MTX(mp), MA_OWNED);
6745 
6746 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6747 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6748 	MNT_REL(mp);
6749 	MNT_IUNLOCK(mp);
6750 	vn_free_marker(*mvp);
6751 	*mvp = NULL;
6752 }
6753 
6754 /*
6755  * These are helper functions for filesystems to traverse their
6756  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6757  */
6758 static void
6759 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6760 {
6761 
6762 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6763 
6764 	MNT_ILOCK(mp);
6765 	MNT_REL(mp);
6766 	MNT_IUNLOCK(mp);
6767 	vn_free_marker(*mvp);
6768 	*mvp = NULL;
6769 }
6770 
6771 /*
6772  * Relock the mp mount vnode list lock with the vp vnode interlock in the
6773  * conventional lock order during mnt_vnode_next_lazy iteration.
6774  *
6775  * On entry, the mount vnode list lock is held and the vnode interlock is not.
6776  * The list lock is dropped and reacquired.  On success, both locks are held.
6777  * On failure, the mount vnode list lock is held but the vnode interlock is
6778  * not, and the procedure may have yielded.
6779  */
6780 static bool
6781 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6782     struct vnode *vp)
6783 {
6784 
6785 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6786 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6787 	    ("%s: bad marker", __func__));
6788 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6789 	    ("%s: inappropriate vnode", __func__));
6790 	ASSERT_VI_UNLOCKED(vp, __func__);
6791 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6792 
6793 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6794 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6795 
6796 	/*
6797 	 * Note we may be racing against vdrop which transitioned the hold
6798 	 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6799 	 * if we are the only user after we get the interlock we will just
6800 	 * vdrop.
6801 	 */
6802 	vhold(vp);
6803 	mtx_unlock(&mp->mnt_listmtx);
6804 	VI_LOCK(vp);
6805 	if (VN_IS_DOOMED(vp)) {
6806 		VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6807 		goto out_lost;
6808 	}
6809 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6810 	/*
6811 	 * There is nothing to do if we are the last user.
6812 	 */
6813 	if (!refcount_release_if_not_last(&vp->v_holdcnt))
6814 		goto out_lost;
6815 	mtx_lock(&mp->mnt_listmtx);
6816 	return (true);
6817 out_lost:
6818 	vdropl(vp);
6819 	maybe_yield();
6820 	mtx_lock(&mp->mnt_listmtx);
6821 	return (false);
6822 }
6823 
6824 static struct vnode *
6825 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6826     void *cbarg)
6827 {
6828 	struct vnode *vp;
6829 
6830 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6831 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6832 restart:
6833 	vp = TAILQ_NEXT(*mvp, v_lazylist);
6834 	while (vp != NULL) {
6835 		if (vp->v_type == VMARKER) {
6836 			vp = TAILQ_NEXT(vp, v_lazylist);
6837 			continue;
6838 		}
6839 		/*
6840 		 * See if we want to process the vnode. Note we may encounter a
6841 		 * long string of vnodes we don't care about and hog the list
6842 		 * as a result. Check for it and requeue the marker.
6843 		 */
6844 		VNPASS(!VN_IS_DOOMED(vp), vp);
6845 		if (!cb(vp, cbarg)) {
6846 			if (!should_yield()) {
6847 				vp = TAILQ_NEXT(vp, v_lazylist);
6848 				continue;
6849 			}
6850 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6851 			    v_lazylist);
6852 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6853 			    v_lazylist);
6854 			mtx_unlock(&mp->mnt_listmtx);
6855 			kern_yield(PRI_USER);
6856 			mtx_lock(&mp->mnt_listmtx);
6857 			goto restart;
6858 		}
6859 		/*
6860 		 * Try-lock because this is the wrong lock order.
6861 		 */
6862 		if (!VI_TRYLOCK(vp) &&
6863 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6864 			goto restart;
6865 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6866 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6867 		    ("alien vnode on the lazy list %p %p", vp, mp));
6868 		VNPASS(vp->v_mount == mp, vp);
6869 		VNPASS(!VN_IS_DOOMED(vp), vp);
6870 		break;
6871 	}
6872 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6873 
6874 	/* Check if we are done */
6875 	if (vp == NULL) {
6876 		mtx_unlock(&mp->mnt_listmtx);
6877 		mnt_vnode_markerfree_lazy(mvp, mp);
6878 		return (NULL);
6879 	}
6880 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6881 	mtx_unlock(&mp->mnt_listmtx);
6882 	ASSERT_VI_LOCKED(vp, "lazy iter");
6883 	return (vp);
6884 }
6885 
6886 struct vnode *
6887 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6888     void *cbarg)
6889 {
6890 
6891 	if (should_yield())
6892 		kern_yield(PRI_USER);
6893 	mtx_lock(&mp->mnt_listmtx);
6894 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6895 }
6896 
6897 struct vnode *
6898 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6899     void *cbarg)
6900 {
6901 	struct vnode *vp;
6902 
6903 	if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6904 		return (NULL);
6905 
6906 	*mvp = vn_alloc_marker(mp);
6907 	MNT_ILOCK(mp);
6908 	MNT_REF(mp);
6909 	MNT_IUNLOCK(mp);
6910 
6911 	mtx_lock(&mp->mnt_listmtx);
6912 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6913 	if (vp == NULL) {
6914 		mtx_unlock(&mp->mnt_listmtx);
6915 		mnt_vnode_markerfree_lazy(mvp, mp);
6916 		return (NULL);
6917 	}
6918 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6919 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6920 }
6921 
6922 void
6923 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6924 {
6925 
6926 	if (*mvp == NULL)
6927 		return;
6928 
6929 	mtx_lock(&mp->mnt_listmtx);
6930 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6931 	mtx_unlock(&mp->mnt_listmtx);
6932 	mnt_vnode_markerfree_lazy(mvp, mp);
6933 }
6934 
6935 int
6936 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6937 {
6938 
6939 	if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6940 		cnp->cn_flags &= ~NOEXECCHECK;
6941 		return (0);
6942 	}
6943 
6944 	return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, curthread));
6945 }
6946 
6947 /*
6948  * Do not use this variant unless you have means other than the hold count
6949  * to prevent the vnode from getting freed.
6950  */
6951 void
6952 vn_seqc_write_begin_locked(struct vnode *vp)
6953 {
6954 
6955 	ASSERT_VI_LOCKED(vp, __func__);
6956 	VNPASS(vp->v_holdcnt > 0, vp);
6957 	VNPASS(vp->v_seqc_users >= 0, vp);
6958 	vp->v_seqc_users++;
6959 	if (vp->v_seqc_users == 1)
6960 		seqc_sleepable_write_begin(&vp->v_seqc);
6961 }
6962 
6963 void
6964 vn_seqc_write_begin(struct vnode *vp)
6965 {
6966 
6967 	VI_LOCK(vp);
6968 	vn_seqc_write_begin_locked(vp);
6969 	VI_UNLOCK(vp);
6970 }
6971 
6972 void
6973 vn_seqc_write_end_locked(struct vnode *vp)
6974 {
6975 
6976 	ASSERT_VI_LOCKED(vp, __func__);
6977 	VNPASS(vp->v_seqc_users > 0, vp);
6978 	vp->v_seqc_users--;
6979 	if (vp->v_seqc_users == 0)
6980 		seqc_sleepable_write_end(&vp->v_seqc);
6981 }
6982 
6983 void
6984 vn_seqc_write_end(struct vnode *vp)
6985 {
6986 
6987 	VI_LOCK(vp);
6988 	vn_seqc_write_end_locked(vp);
6989 	VI_UNLOCK(vp);
6990 }
6991 
6992 /*
6993  * Special case handling for allocating and freeing vnodes.
6994  *
6995  * The counter remains unchanged on free so that a doomed vnode will
6996  * keep testing as in modify as long as it is accessible with SMR.
6997  */
6998 static void
6999 vn_seqc_init(struct vnode *vp)
7000 {
7001 
7002 	vp->v_seqc = 0;
7003 	vp->v_seqc_users = 0;
7004 }
7005 
7006 static void
7007 vn_seqc_write_end_free(struct vnode *vp)
7008 {
7009 
7010 	VNPASS(seqc_in_modify(vp->v_seqc), vp);
7011 	VNPASS(vp->v_seqc_users == 1, vp);
7012 }
7013 
7014 void
7015 vn_irflag_set_locked(struct vnode *vp, short toset)
7016 {
7017 	short flags;
7018 
7019 	ASSERT_VI_LOCKED(vp, __func__);
7020 	flags = vn_irflag_read(vp);
7021 	VNASSERT((flags & toset) == 0, vp,
7022 	    ("%s: some of the passed flags already set (have %d, passed %d)\n",
7023 	    __func__, flags, toset));
7024 	atomic_store_short(&vp->v_irflag, flags | toset);
7025 }
7026 
7027 void
7028 vn_irflag_set(struct vnode *vp, short toset)
7029 {
7030 
7031 	VI_LOCK(vp);
7032 	vn_irflag_set_locked(vp, toset);
7033 	VI_UNLOCK(vp);
7034 }
7035 
7036 void
7037 vn_irflag_set_cond_locked(struct vnode *vp, short toset)
7038 {
7039 	short flags;
7040 
7041 	ASSERT_VI_LOCKED(vp, __func__);
7042 	flags = vn_irflag_read(vp);
7043 	atomic_store_short(&vp->v_irflag, flags | toset);
7044 }
7045 
7046 void
7047 vn_irflag_set_cond(struct vnode *vp, short toset)
7048 {
7049 
7050 	VI_LOCK(vp);
7051 	vn_irflag_set_cond_locked(vp, toset);
7052 	VI_UNLOCK(vp);
7053 }
7054 
7055 void
7056 vn_irflag_unset_locked(struct vnode *vp, short tounset)
7057 {
7058 	short flags;
7059 
7060 	ASSERT_VI_LOCKED(vp, __func__);
7061 	flags = vn_irflag_read(vp);
7062 	VNASSERT((flags & tounset) == tounset, vp,
7063 	    ("%s: some of the passed flags not set (have %d, passed %d)\n",
7064 	    __func__, flags, tounset));
7065 	atomic_store_short(&vp->v_irflag, flags & ~tounset);
7066 }
7067 
7068 void
7069 vn_irflag_unset(struct vnode *vp, short tounset)
7070 {
7071 
7072 	VI_LOCK(vp);
7073 	vn_irflag_unset_locked(vp, tounset);
7074 	VI_UNLOCK(vp);
7075 }
7076