xref: /freebsd/sys/kern/vfs_subr.c (revision 22cf89c938886d14f5796fc49f9f020c23ea8eaf)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	@(#)vfs_subr.c	8.31 (Berkeley) 5/26/95
37  */
38 
39 /*
40  * External virtual filesystem routines
41  */
42 
43 #include <sys/cdefs.h>
44 #include "opt_ddb.h"
45 #include "opt_watchdog.h"
46 
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/asan.h>
50 #include <sys/bio.h>
51 #include <sys/buf.h>
52 #include <sys/capsicum.h>
53 #include <sys/condvar.h>
54 #include <sys/conf.h>
55 #include <sys/counter.h>
56 #include <sys/dirent.h>
57 #include <sys/event.h>
58 #include <sys/eventhandler.h>
59 #include <sys/extattr.h>
60 #include <sys/file.h>
61 #include <sys/fcntl.h>
62 #include <sys/jail.h>
63 #include <sys/kdb.h>
64 #include <sys/kernel.h>
65 #include <sys/kthread.h>
66 #include <sys/ktr.h>
67 #include <sys/limits.h>
68 #include <sys/lockf.h>
69 #include <sys/malloc.h>
70 #include <sys/mount.h>
71 #include <sys/namei.h>
72 #include <sys/pctrie.h>
73 #include <sys/priv.h>
74 #include <sys/reboot.h>
75 #include <sys/refcount.h>
76 #include <sys/rwlock.h>
77 #include <sys/sched.h>
78 #include <sys/sleepqueue.h>
79 #include <sys/smr.h>
80 #include <sys/smp.h>
81 #include <sys/stat.h>
82 #include <sys/sysctl.h>
83 #include <sys/syslog.h>
84 #include <sys/vmmeter.h>
85 #include <sys/vnode.h>
86 #include <sys/watchdog.h>
87 
88 #include <machine/stdarg.h>
89 
90 #include <security/mac/mac_framework.h>
91 
92 #include <vm/vm.h>
93 #include <vm/vm_object.h>
94 #include <vm/vm_extern.h>
95 #include <vm/pmap.h>
96 #include <vm/vm_map.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_kern.h>
99 #include <vm/uma.h>
100 
101 #if defined(DEBUG_VFS_LOCKS) && (!defined(INVARIANTS) || !defined(WITNESS))
102 #error DEBUG_VFS_LOCKS requires INVARIANTS and WITNESS
103 #endif
104 
105 #ifdef DDB
106 #include <ddb/ddb.h>
107 #endif
108 
109 static void	delmntque(struct vnode *vp);
110 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
111 		    int slpflag, int slptimeo);
112 static void	syncer_shutdown(void *arg, int howto);
113 static int	vtryrecycle(struct vnode *vp);
114 static void	v_init_counters(struct vnode *);
115 static void	vn_seqc_init(struct vnode *);
116 static void	vn_seqc_write_end_free(struct vnode *vp);
117 static void	vgonel(struct vnode *);
118 static bool	vhold_recycle_free(struct vnode *);
119 static void	vdropl_recycle(struct vnode *vp);
120 static void	vdrop_recycle(struct vnode *vp);
121 static void	vfs_knllock(void *arg);
122 static void	vfs_knlunlock(void *arg);
123 static void	vfs_knl_assert_lock(void *arg, int what);
124 static void	destroy_vpollinfo(struct vpollinfo *vi);
125 static int	v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
126 		    daddr_t startlbn, daddr_t endlbn);
127 static void	vnlru_recalc(void);
128 
129 /*
130  * Number of vnodes in existence.  Increased whenever getnewvnode()
131  * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
132  */
133 static u_long __exclusive_cache_line numvnodes;
134 
135 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
136     "Number of vnodes in existence");
137 
138 static counter_u64_t vnodes_created;
139 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
140     "Number of vnodes created by getnewvnode");
141 
142 /*
143  * Conversion tables for conversion from vnode types to inode formats
144  * and back.
145  */
146 __enum_uint8(vtype) iftovt_tab[16] = {
147 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
148 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
149 };
150 int vttoif_tab[10] = {
151 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
152 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
153 };
154 
155 /*
156  * List of allocates vnodes in the system.
157  */
158 static TAILQ_HEAD(freelst, vnode) vnode_list;
159 static struct vnode *vnode_list_free_marker;
160 static struct vnode *vnode_list_reclaim_marker;
161 
162 /*
163  * "Free" vnode target.  Free vnodes are rarely completely free, but are
164  * just ones that are cheap to recycle.  Usually they are for files which
165  * have been stat'd but not read; these usually have inode and namecache
166  * data attached to them.  This target is the preferred minimum size of a
167  * sub-cache consisting mostly of such files. The system balances the size
168  * of this sub-cache with its complement to try to prevent either from
169  * thrashing while the other is relatively inactive.  The targets express
170  * a preference for the best balance.
171  *
172  * "Above" this target there are 2 further targets (watermarks) related
173  * to recyling of free vnodes.  In the best-operating case, the cache is
174  * exactly full, the free list has size between vlowat and vhiwat above the
175  * free target, and recycling from it and normal use maintains this state.
176  * Sometimes the free list is below vlowat or even empty, but this state
177  * is even better for immediate use provided the cache is not full.
178  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
179  * ones) to reach one of these states.  The watermarks are currently hard-
180  * coded as 4% and 9% of the available space higher.  These and the default
181  * of 25% for wantfreevnodes are too large if the memory size is large.
182  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
183  * whenever vnlru_proc() becomes active.
184  */
185 static long wantfreevnodes;
186 static long __exclusive_cache_line freevnodes;
187 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
188     &freevnodes, 0, "Number of \"free\" vnodes");
189 static long freevnodes_old;
190 
191 static counter_u64_t recycles_count;
192 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
193     "Number of vnodes recycled to meet vnode cache targets");
194 
195 static counter_u64_t recycles_free_count;
196 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count,
197     "Number of free vnodes recycled to meet vnode cache targets");
198 
199 static counter_u64_t vnode_skipped_requeues;
200 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnode_skipped_requeues, CTLFLAG_RD, &vnode_skipped_requeues,
201     "Number of times LRU requeue was skipped due to lock contention");
202 
203 static u_long deferred_inact;
204 SYSCTL_ULONG(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD,
205     &deferred_inact, 0, "Number of times inactive processing was deferred");
206 
207 /* To keep more than one thread at a time from running vfs_getnewfsid */
208 static struct mtx mntid_mtx;
209 
210 /*
211  * Lock for any access to the following:
212  *	vnode_list
213  *	numvnodes
214  *	freevnodes
215  */
216 static struct mtx __exclusive_cache_line vnode_list_mtx;
217 
218 /* Publicly exported FS */
219 struct nfs_public nfs_pub;
220 
221 static uma_zone_t buf_trie_zone;
222 static smr_t buf_trie_smr;
223 
224 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
225 static uma_zone_t vnode_zone;
226 MALLOC_DEFINE(M_VNODEPOLL, "VN POLL", "vnode poll");
227 
228 __read_frequently smr_t vfs_smr;
229 
230 /*
231  * The workitem queue.
232  *
233  * It is useful to delay writes of file data and filesystem metadata
234  * for tens of seconds so that quickly created and deleted files need
235  * not waste disk bandwidth being created and removed. To realize this,
236  * we append vnodes to a "workitem" queue. When running with a soft
237  * updates implementation, most pending metadata dependencies should
238  * not wait for more than a few seconds. Thus, mounted on block devices
239  * are delayed only about a half the time that file data is delayed.
240  * Similarly, directory updates are more critical, so are only delayed
241  * about a third the time that file data is delayed. Thus, there are
242  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
243  * one each second (driven off the filesystem syncer process). The
244  * syncer_delayno variable indicates the next queue that is to be processed.
245  * Items that need to be processed soon are placed in this queue:
246  *
247  *	syncer_workitem_pending[syncer_delayno]
248  *
249  * A delay of fifteen seconds is done by placing the request fifteen
250  * entries later in the queue:
251  *
252  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
253  *
254  */
255 static int syncer_delayno;
256 static long syncer_mask;
257 LIST_HEAD(synclist, bufobj);
258 static struct synclist *syncer_workitem_pending;
259 /*
260  * The sync_mtx protects:
261  *	bo->bo_synclist
262  *	sync_vnode_count
263  *	syncer_delayno
264  *	syncer_state
265  *	syncer_workitem_pending
266  *	syncer_worklist_len
267  *	rushjob
268  */
269 static struct mtx sync_mtx;
270 static struct cv sync_wakeup;
271 
272 #define SYNCER_MAXDELAY		32
273 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
274 static int syncdelay = 30;		/* max time to delay syncing data */
275 static int filedelay = 30;		/* time to delay syncing files */
276 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
277     "Time to delay syncing files (in seconds)");
278 static int dirdelay = 29;		/* time to delay syncing directories */
279 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
280     "Time to delay syncing directories (in seconds)");
281 static int metadelay = 28;		/* time to delay syncing metadata */
282 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
283     "Time to delay syncing metadata (in seconds)");
284 static int rushjob;		/* number of slots to run ASAP */
285 static int stat_rush_requests;	/* number of times I/O speeded up */
286 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
287     "Number of times I/O speeded up (rush requests)");
288 
289 #define	VDBATCH_SIZE 8
290 struct vdbatch {
291 	u_int index;
292 	struct mtx lock;
293 	struct vnode *tab[VDBATCH_SIZE];
294 };
295 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
296 
297 static void	vdbatch_dequeue(struct vnode *vp);
298 
299 /*
300  * When shutting down the syncer, run it at four times normal speed.
301  */
302 #define SYNCER_SHUTDOWN_SPEEDUP		4
303 static int sync_vnode_count;
304 static int syncer_worklist_len;
305 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
306     syncer_state;
307 
308 /* Target for maximum number of vnodes. */
309 u_long desiredvnodes;
310 static u_long gapvnodes;		/* gap between wanted and desired */
311 static u_long vhiwat;		/* enough extras after expansion */
312 static u_long vlowat;		/* minimal extras before expansion */
313 static u_long vstir;		/* nonzero to stir non-free vnodes */
314 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
315 
316 static u_long vnlru_read_freevnodes(void);
317 
318 /*
319  * Note that no attempt is made to sanitize these parameters.
320  */
321 static int
322 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
323 {
324 	u_long val;
325 	int error;
326 
327 	val = desiredvnodes;
328 	error = sysctl_handle_long(oidp, &val, 0, req);
329 	if (error != 0 || req->newptr == NULL)
330 		return (error);
331 
332 	if (val == desiredvnodes)
333 		return (0);
334 	mtx_lock(&vnode_list_mtx);
335 	desiredvnodes = val;
336 	wantfreevnodes = desiredvnodes / 4;
337 	vnlru_recalc();
338 	mtx_unlock(&vnode_list_mtx);
339 	/*
340 	 * XXX There is no protection against multiple threads changing
341 	 * desiredvnodes at the same time. Locking above only helps vnlru and
342 	 * getnewvnode.
343 	 */
344 	vfs_hash_changesize(desiredvnodes);
345 	cache_changesize(desiredvnodes);
346 	return (0);
347 }
348 
349 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
350     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
351     "LU", "Target for maximum number of vnodes");
352 
353 static int
354 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
355 {
356 	u_long val;
357 	int error;
358 
359 	val = wantfreevnodes;
360 	error = sysctl_handle_long(oidp, &val, 0, req);
361 	if (error != 0 || req->newptr == NULL)
362 		return (error);
363 
364 	if (val == wantfreevnodes)
365 		return (0);
366 	mtx_lock(&vnode_list_mtx);
367 	wantfreevnodes = val;
368 	vnlru_recalc();
369 	mtx_unlock(&vnode_list_mtx);
370 	return (0);
371 }
372 
373 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
374     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
375     "LU", "Target for minimum number of \"free\" vnodes");
376 
377 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
378     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
379 static int vnlru_nowhere;
380 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW | CTLFLAG_STATS,
381     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
382 
383 static int
384 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
385 {
386 	struct vnode *vp;
387 	struct nameidata nd;
388 	char *buf;
389 	unsigned long ndflags;
390 	int error;
391 
392 	if (req->newptr == NULL)
393 		return (EINVAL);
394 	if (req->newlen >= PATH_MAX)
395 		return (E2BIG);
396 
397 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
398 	error = SYSCTL_IN(req, buf, req->newlen);
399 	if (error != 0)
400 		goto out;
401 
402 	buf[req->newlen] = '\0';
403 
404 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1;
405 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf);
406 	if ((error = namei(&nd)) != 0)
407 		goto out;
408 	vp = nd.ni_vp;
409 
410 	if (VN_IS_DOOMED(vp)) {
411 		/*
412 		 * This vnode is being recycled.  Return != 0 to let the caller
413 		 * know that the sysctl had no effect.  Return EAGAIN because a
414 		 * subsequent call will likely succeed (since namei will create
415 		 * a new vnode if necessary)
416 		 */
417 		error = EAGAIN;
418 		goto putvnode;
419 	}
420 
421 	counter_u64_add(recycles_count, 1);
422 	vgone(vp);
423 putvnode:
424 	vput(vp);
425 	NDFREE_PNBUF(&nd);
426 out:
427 	free(buf, M_TEMP);
428 	return (error);
429 }
430 
431 static int
432 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
433 {
434 	struct thread *td = curthread;
435 	struct vnode *vp;
436 	struct file *fp;
437 	int error;
438 	int fd;
439 
440 	if (req->newptr == NULL)
441 		return (EBADF);
442 
443         error = sysctl_handle_int(oidp, &fd, 0, req);
444         if (error != 0)
445                 return (error);
446 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
447 	if (error != 0)
448 		return (error);
449 	vp = fp->f_vnode;
450 
451 	error = vn_lock(vp, LK_EXCLUSIVE);
452 	if (error != 0)
453 		goto drop;
454 
455 	counter_u64_add(recycles_count, 1);
456 	vgone(vp);
457 	VOP_UNLOCK(vp);
458 drop:
459 	fdrop(fp, td);
460 	return (error);
461 }
462 
463 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
464     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
465     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
466 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
467     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
468     sysctl_ftry_reclaim_vnode, "I",
469     "Try to reclaim a vnode by its file descriptor");
470 
471 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
472 #define vnsz2log 8
473 #ifndef DEBUG_LOCKS
474 _Static_assert(sizeof(struct vnode) >= 1UL << vnsz2log &&
475     sizeof(struct vnode) < 1UL << (vnsz2log + 1),
476     "vnsz2log needs to be updated");
477 #endif
478 
479 /*
480  * Support for the bufobj clean & dirty pctrie.
481  */
482 static void *
483 buf_trie_alloc(struct pctrie *ptree)
484 {
485 	return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT));
486 }
487 
488 static void
489 buf_trie_free(struct pctrie *ptree, void *node)
490 {
491 	uma_zfree_smr(buf_trie_zone, node);
492 }
493 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free,
494     buf_trie_smr);
495 
496 /*
497  * Initialize the vnode management data structures.
498  *
499  * Reevaluate the following cap on the number of vnodes after the physical
500  * memory size exceeds 512GB.  In the limit, as the physical memory size
501  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
502  */
503 #ifndef	MAXVNODES_MAX
504 #define	MAXVNODES_MAX	(512UL * 1024 * 1024 / 64)	/* 8M */
505 #endif
506 
507 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
508 
509 static struct vnode *
510 vn_alloc_marker(struct mount *mp)
511 {
512 	struct vnode *vp;
513 
514 	vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
515 	vp->v_type = VMARKER;
516 	vp->v_mount = mp;
517 
518 	return (vp);
519 }
520 
521 static void
522 vn_free_marker(struct vnode *vp)
523 {
524 
525 	MPASS(vp->v_type == VMARKER);
526 	free(vp, M_VNODE_MARKER);
527 }
528 
529 #ifdef KASAN
530 static int
531 vnode_ctor(void *mem, int size, void *arg __unused, int flags __unused)
532 {
533 	kasan_mark(mem, size, roundup2(size, UMA_ALIGN_PTR + 1), 0);
534 	return (0);
535 }
536 
537 static void
538 vnode_dtor(void *mem, int size, void *arg __unused)
539 {
540 	size_t end1, end2, off1, off2;
541 
542 	_Static_assert(offsetof(struct vnode, v_vnodelist) <
543 	    offsetof(struct vnode, v_dbatchcpu),
544 	    "KASAN marks require updating");
545 
546 	off1 = offsetof(struct vnode, v_vnodelist);
547 	off2 = offsetof(struct vnode, v_dbatchcpu);
548 	end1 = off1 + sizeof(((struct vnode *)NULL)->v_vnodelist);
549 	end2 = off2 + sizeof(((struct vnode *)NULL)->v_dbatchcpu);
550 
551 	/*
552 	 * Access to the v_vnodelist and v_dbatchcpu fields are permitted even
553 	 * after the vnode has been freed.  Try to get some KASAN coverage by
554 	 * marking everything except those two fields as invalid.  Because
555 	 * KASAN's tracking is not byte-granular, any preceding fields sharing
556 	 * the same 8-byte aligned word must also be marked valid.
557 	 */
558 
559 	/* Handle the area from the start until v_vnodelist... */
560 	off1 = rounddown2(off1, KASAN_SHADOW_SCALE);
561 	kasan_mark(mem, off1, off1, KASAN_UMA_FREED);
562 
563 	/* ... then the area between v_vnodelist and v_dbatchcpu ... */
564 	off1 = roundup2(end1, KASAN_SHADOW_SCALE);
565 	off2 = rounddown2(off2, KASAN_SHADOW_SCALE);
566 	if (off2 > off1)
567 		kasan_mark((void *)((char *)mem + off1), off2 - off1,
568 		    off2 - off1, KASAN_UMA_FREED);
569 
570 	/* ... and finally the area from v_dbatchcpu to the end. */
571 	off2 = roundup2(end2, KASAN_SHADOW_SCALE);
572 	kasan_mark((void *)((char *)mem + off2), size - off2, size - off2,
573 	    KASAN_UMA_FREED);
574 }
575 #endif /* KASAN */
576 
577 /*
578  * Initialize a vnode as it first enters the zone.
579  */
580 static int
581 vnode_init(void *mem, int size, int flags)
582 {
583 	struct vnode *vp;
584 
585 	vp = mem;
586 	bzero(vp, size);
587 	/*
588 	 * Setup locks.
589 	 */
590 	vp->v_vnlock = &vp->v_lock;
591 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
592 	/*
593 	 * By default, don't allow shared locks unless filesystems opt-in.
594 	 */
595 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
596 	    LK_NOSHARE | LK_IS_VNODE);
597 	/*
598 	 * Initialize bufobj.
599 	 */
600 	bufobj_init(&vp->v_bufobj, vp);
601 	/*
602 	 * Initialize namecache.
603 	 */
604 	cache_vnode_init(vp);
605 	/*
606 	 * Initialize rangelocks.
607 	 */
608 	rangelock_init(&vp->v_rl);
609 
610 	vp->v_dbatchcpu = NOCPU;
611 
612 	vp->v_state = VSTATE_DEAD;
613 
614 	/*
615 	 * Check vhold_recycle_free for an explanation.
616 	 */
617 	vp->v_holdcnt = VHOLD_NO_SMR;
618 	vp->v_type = VNON;
619 	mtx_lock(&vnode_list_mtx);
620 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
621 	mtx_unlock(&vnode_list_mtx);
622 	return (0);
623 }
624 
625 /*
626  * Free a vnode when it is cleared from the zone.
627  */
628 static void
629 vnode_fini(void *mem, int size)
630 {
631 	struct vnode *vp;
632 	struct bufobj *bo;
633 
634 	vp = mem;
635 	vdbatch_dequeue(vp);
636 	mtx_lock(&vnode_list_mtx);
637 	TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
638 	mtx_unlock(&vnode_list_mtx);
639 	rangelock_destroy(&vp->v_rl);
640 	lockdestroy(vp->v_vnlock);
641 	mtx_destroy(&vp->v_interlock);
642 	bo = &vp->v_bufobj;
643 	rw_destroy(BO_LOCKPTR(bo));
644 
645 	kasan_mark(mem, size, size, 0);
646 }
647 
648 /*
649  * Provide the size of NFS nclnode and NFS fh for calculation of the
650  * vnode memory consumption.  The size is specified directly to
651  * eliminate dependency on NFS-private header.
652  *
653  * Other filesystems may use bigger or smaller (like UFS and ZFS)
654  * private inode data, but the NFS-based estimation is ample enough.
655  * Still, we care about differences in the size between 64- and 32-bit
656  * platforms.
657  *
658  * Namecache structure size is heuristically
659  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
660  */
661 #ifdef _LP64
662 #define	NFS_NCLNODE_SZ	(528 + 64)
663 #define	NC_SZ		148
664 #else
665 #define	NFS_NCLNODE_SZ	(360 + 32)
666 #define	NC_SZ		92
667 #endif
668 
669 static void
670 vntblinit(void *dummy __unused)
671 {
672 	struct vdbatch *vd;
673 	uma_ctor ctor;
674 	uma_dtor dtor;
675 	int cpu, physvnodes, virtvnodes;
676 
677 	/*
678 	 * Desiredvnodes is a function of the physical memory size and the
679 	 * kernel's heap size.  Generally speaking, it scales with the
680 	 * physical memory size.  The ratio of desiredvnodes to the physical
681 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
682 	 * Thereafter, the
683 	 * marginal ratio of desiredvnodes to the physical memory size is
684 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
685 	 * size.  The memory required by desiredvnodes vnodes and vm objects
686 	 * must not exceed 1/10th of the kernel's heap size.
687 	 */
688 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
689 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
690 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
691 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
692 	desiredvnodes = min(physvnodes, virtvnodes);
693 	if (desiredvnodes > MAXVNODES_MAX) {
694 		if (bootverbose)
695 			printf("Reducing kern.maxvnodes %lu -> %lu\n",
696 			    desiredvnodes, MAXVNODES_MAX);
697 		desiredvnodes = MAXVNODES_MAX;
698 	}
699 	wantfreevnodes = desiredvnodes / 4;
700 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
701 	TAILQ_INIT(&vnode_list);
702 	mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
703 	/*
704 	 * The lock is taken to appease WITNESS.
705 	 */
706 	mtx_lock(&vnode_list_mtx);
707 	vnlru_recalc();
708 	mtx_unlock(&vnode_list_mtx);
709 	vnode_list_free_marker = vn_alloc_marker(NULL);
710 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
711 	vnode_list_reclaim_marker = vn_alloc_marker(NULL);
712 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
713 
714 #ifdef KASAN
715 	ctor = vnode_ctor;
716 	dtor = vnode_dtor;
717 #else
718 	ctor = NULL;
719 	dtor = NULL;
720 #endif
721 	vnode_zone = uma_zcreate("VNODE", sizeof(struct vnode), ctor, dtor,
722 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, UMA_ZONE_NOKASAN);
723 	uma_zone_set_smr(vnode_zone, vfs_smr);
724 
725 	/*
726 	 * Preallocate enough nodes to support one-per buf so that
727 	 * we can not fail an insert.  reassignbuf() callers can not
728 	 * tolerate the insertion failure.
729 	 */
730 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
731 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
732 	    UMA_ZONE_NOFREE | UMA_ZONE_SMR);
733 	buf_trie_smr = uma_zone_get_smr(buf_trie_zone);
734 	uma_prealloc(buf_trie_zone, nbuf);
735 
736 	vnodes_created = counter_u64_alloc(M_WAITOK);
737 	recycles_count = counter_u64_alloc(M_WAITOK);
738 	recycles_free_count = counter_u64_alloc(M_WAITOK);
739 	vnode_skipped_requeues = counter_u64_alloc(M_WAITOK);
740 
741 	/*
742 	 * Initialize the filesystem syncer.
743 	 */
744 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
745 	    &syncer_mask);
746 	syncer_maxdelay = syncer_mask + 1;
747 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
748 	cv_init(&sync_wakeup, "syncer");
749 
750 	CPU_FOREACH(cpu) {
751 		vd = DPCPU_ID_PTR((cpu), vd);
752 		bzero(vd, sizeof(*vd));
753 		mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
754 	}
755 }
756 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
757 
758 /*
759  * Mark a mount point as busy. Used to synchronize access and to delay
760  * unmounting. Eventually, mountlist_mtx is not released on failure.
761  *
762  * vfs_busy() is a custom lock, it can block the caller.
763  * vfs_busy() only sleeps if the unmount is active on the mount point.
764  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
765  * vnode belonging to mp.
766  *
767  * Lookup uses vfs_busy() to traverse mount points.
768  * root fs			var fs
769  * / vnode lock		A	/ vnode lock (/var)		D
770  * /var vnode lock	B	/log vnode lock(/var/log)	E
771  * vfs_busy lock	C	vfs_busy lock			F
772  *
773  * Within each file system, the lock order is C->A->B and F->D->E.
774  *
775  * When traversing across mounts, the system follows that lock order:
776  *
777  *        C->A->B
778  *              |
779  *              +->F->D->E
780  *
781  * The lookup() process for namei("/var") illustrates the process:
782  *  1. VOP_LOOKUP() obtains B while A is held
783  *  2. vfs_busy() obtains a shared lock on F while A and B are held
784  *  3. vput() releases lock on B
785  *  4. vput() releases lock on A
786  *  5. VFS_ROOT() obtains lock on D while shared lock on F is held
787  *  6. vfs_unbusy() releases shared lock on F
788  *  7. vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
789  *     Attempt to lock A (instead of vp_crossmp) while D is held would
790  *     violate the global order, causing deadlocks.
791  *
792  * dounmount() locks B while F is drained.  Note that for stacked
793  * filesystems, D and B in the example above may be the same lock,
794  * which introdues potential lock order reversal deadlock between
795  * dounmount() and step 5 above.  These filesystems may avoid the LOR
796  * by setting VV_CROSSLOCK on the covered vnode so that lock B will
797  * remain held until after step 5.
798  */
799 int
800 vfs_busy(struct mount *mp, int flags)
801 {
802 	struct mount_pcpu *mpcpu;
803 
804 	MPASS((flags & ~MBF_MASK) == 0);
805 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
806 
807 	if (vfs_op_thread_enter(mp, mpcpu)) {
808 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
809 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
810 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
811 		vfs_mp_count_add_pcpu(mpcpu, ref, 1);
812 		vfs_mp_count_add_pcpu(mpcpu, lockref, 1);
813 		vfs_op_thread_exit(mp, mpcpu);
814 		if (flags & MBF_MNTLSTLOCK)
815 			mtx_unlock(&mountlist_mtx);
816 		return (0);
817 	}
818 
819 	MNT_ILOCK(mp);
820 	vfs_assert_mount_counters(mp);
821 	MNT_REF(mp);
822 	/*
823 	 * If mount point is currently being unmounted, sleep until the
824 	 * mount point fate is decided.  If thread doing the unmounting fails,
825 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
826 	 * that this mount point has survived the unmount attempt and vfs_busy
827 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
828 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
829 	 * about to be really destroyed.  vfs_busy needs to release its
830 	 * reference on the mount point in this case and return with ENOENT,
831 	 * telling the caller the mount it tried to busy is no longer valid.
832 	 */
833 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
834 		KASSERT(TAILQ_EMPTY(&mp->mnt_uppers),
835 		    ("%s: non-empty upper mount list with pending unmount",
836 		    __func__));
837 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
838 			MNT_REL(mp);
839 			MNT_IUNLOCK(mp);
840 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
841 			    __func__);
842 			return (ENOENT);
843 		}
844 		if (flags & MBF_MNTLSTLOCK)
845 			mtx_unlock(&mountlist_mtx);
846 		mp->mnt_kern_flag |= MNTK_MWAIT;
847 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
848 		if (flags & MBF_MNTLSTLOCK)
849 			mtx_lock(&mountlist_mtx);
850 		MNT_ILOCK(mp);
851 	}
852 	if (flags & MBF_MNTLSTLOCK)
853 		mtx_unlock(&mountlist_mtx);
854 	mp->mnt_lockref++;
855 	MNT_IUNLOCK(mp);
856 	return (0);
857 }
858 
859 /*
860  * Free a busy filesystem.
861  */
862 void
863 vfs_unbusy(struct mount *mp)
864 {
865 	struct mount_pcpu *mpcpu;
866 	int c;
867 
868 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
869 
870 	if (vfs_op_thread_enter(mp, mpcpu)) {
871 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
872 		vfs_mp_count_sub_pcpu(mpcpu, lockref, 1);
873 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
874 		vfs_op_thread_exit(mp, mpcpu);
875 		return;
876 	}
877 
878 	MNT_ILOCK(mp);
879 	vfs_assert_mount_counters(mp);
880 	MNT_REL(mp);
881 	c = --mp->mnt_lockref;
882 	if (mp->mnt_vfs_ops == 0) {
883 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
884 		MNT_IUNLOCK(mp);
885 		return;
886 	}
887 	if (c < 0)
888 		vfs_dump_mount_counters(mp);
889 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
890 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
891 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
892 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
893 		wakeup(&mp->mnt_lockref);
894 	}
895 	MNT_IUNLOCK(mp);
896 }
897 
898 /*
899  * Lookup a mount point by filesystem identifier.
900  */
901 struct mount *
902 vfs_getvfs(fsid_t *fsid)
903 {
904 	struct mount *mp;
905 
906 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
907 	mtx_lock(&mountlist_mtx);
908 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
909 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
910 			vfs_ref(mp);
911 			mtx_unlock(&mountlist_mtx);
912 			return (mp);
913 		}
914 	}
915 	mtx_unlock(&mountlist_mtx);
916 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
917 	return ((struct mount *) 0);
918 }
919 
920 /*
921  * Lookup a mount point by filesystem identifier, busying it before
922  * returning.
923  *
924  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
925  * cache for popular filesystem identifiers.  The cache is lockess, using
926  * the fact that struct mount's are never freed.  In worst case we may
927  * get pointer to unmounted or even different filesystem, so we have to
928  * check what we got, and go slow way if so.
929  */
930 struct mount *
931 vfs_busyfs(fsid_t *fsid)
932 {
933 #define	FSID_CACHE_SIZE	256
934 	typedef struct mount * volatile vmp_t;
935 	static vmp_t cache[FSID_CACHE_SIZE];
936 	struct mount *mp;
937 	int error;
938 	uint32_t hash;
939 
940 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
941 	hash = fsid->val[0] ^ fsid->val[1];
942 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
943 	mp = cache[hash];
944 	if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0)
945 		goto slow;
946 	if (vfs_busy(mp, 0) != 0) {
947 		cache[hash] = NULL;
948 		goto slow;
949 	}
950 	if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0)
951 		return (mp);
952 	else
953 	    vfs_unbusy(mp);
954 
955 slow:
956 	mtx_lock(&mountlist_mtx);
957 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
958 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
959 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
960 			if (error) {
961 				cache[hash] = NULL;
962 				mtx_unlock(&mountlist_mtx);
963 				return (NULL);
964 			}
965 			cache[hash] = mp;
966 			return (mp);
967 		}
968 	}
969 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
970 	mtx_unlock(&mountlist_mtx);
971 	return ((struct mount *) 0);
972 }
973 
974 /*
975  * Check if a user can access privileged mount options.
976  */
977 int
978 vfs_suser(struct mount *mp, struct thread *td)
979 {
980 	int error;
981 
982 	if (jailed(td->td_ucred)) {
983 		/*
984 		 * If the jail of the calling thread lacks permission for
985 		 * this type of file system, deny immediately.
986 		 */
987 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
988 			return (EPERM);
989 
990 		/*
991 		 * If the file system was mounted outside the jail of the
992 		 * calling thread, deny immediately.
993 		 */
994 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
995 			return (EPERM);
996 	}
997 
998 	/*
999 	 * If file system supports delegated administration, we don't check
1000 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
1001 	 * by the file system itself.
1002 	 * If this is not the user that did original mount, we check for
1003 	 * the PRIV_VFS_MOUNT_OWNER privilege.
1004 	 */
1005 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
1006 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
1007 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
1008 			return (error);
1009 	}
1010 	return (0);
1011 }
1012 
1013 /*
1014  * Get a new unique fsid.  Try to make its val[0] unique, since this value
1015  * will be used to create fake device numbers for stat().  Also try (but
1016  * not so hard) make its val[0] unique mod 2^16, since some emulators only
1017  * support 16-bit device numbers.  We end up with unique val[0]'s for the
1018  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
1019  *
1020  * Keep in mind that several mounts may be running in parallel.  Starting
1021  * the search one past where the previous search terminated is both a
1022  * micro-optimization and a defense against returning the same fsid to
1023  * different mounts.
1024  */
1025 void
1026 vfs_getnewfsid(struct mount *mp)
1027 {
1028 	static uint16_t mntid_base;
1029 	struct mount *nmp;
1030 	fsid_t tfsid;
1031 	int mtype;
1032 
1033 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
1034 	mtx_lock(&mntid_mtx);
1035 	mtype = mp->mnt_vfc->vfc_typenum;
1036 	tfsid.val[1] = mtype;
1037 	mtype = (mtype & 0xFF) << 24;
1038 	for (;;) {
1039 		tfsid.val[0] = makedev(255,
1040 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
1041 		mntid_base++;
1042 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
1043 			break;
1044 		vfs_rel(nmp);
1045 	}
1046 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
1047 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
1048 	mtx_unlock(&mntid_mtx);
1049 }
1050 
1051 /*
1052  * Knob to control the precision of file timestamps:
1053  *
1054  *   0 = seconds only; nanoseconds zeroed.
1055  *   1 = seconds and nanoseconds, accurate within 1/HZ.
1056  *   2 = seconds and nanoseconds, truncated to microseconds.
1057  * >=3 = seconds and nanoseconds, maximum precision.
1058  */
1059 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
1060 
1061 static int timestamp_precision = TSP_USEC;
1062 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
1063     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
1064     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1065     "3+: sec + ns (max. precision))");
1066 
1067 /*
1068  * Get a current timestamp.
1069  */
1070 void
1071 vfs_timestamp(struct timespec *tsp)
1072 {
1073 	struct timeval tv;
1074 
1075 	switch (timestamp_precision) {
1076 	case TSP_SEC:
1077 		tsp->tv_sec = time_second;
1078 		tsp->tv_nsec = 0;
1079 		break;
1080 	case TSP_HZ:
1081 		getnanotime(tsp);
1082 		break;
1083 	case TSP_USEC:
1084 		microtime(&tv);
1085 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1086 		break;
1087 	case TSP_NSEC:
1088 	default:
1089 		nanotime(tsp);
1090 		break;
1091 	}
1092 }
1093 
1094 /*
1095  * Set vnode attributes to VNOVAL
1096  */
1097 void
1098 vattr_null(struct vattr *vap)
1099 {
1100 
1101 	vap->va_type = VNON;
1102 	vap->va_size = VNOVAL;
1103 	vap->va_bytes = VNOVAL;
1104 	vap->va_mode = VNOVAL;
1105 	vap->va_nlink = VNOVAL;
1106 	vap->va_uid = VNOVAL;
1107 	vap->va_gid = VNOVAL;
1108 	vap->va_fsid = VNOVAL;
1109 	vap->va_fileid = VNOVAL;
1110 	vap->va_blocksize = VNOVAL;
1111 	vap->va_rdev = VNOVAL;
1112 	vap->va_atime.tv_sec = VNOVAL;
1113 	vap->va_atime.tv_nsec = VNOVAL;
1114 	vap->va_mtime.tv_sec = VNOVAL;
1115 	vap->va_mtime.tv_nsec = VNOVAL;
1116 	vap->va_ctime.tv_sec = VNOVAL;
1117 	vap->va_ctime.tv_nsec = VNOVAL;
1118 	vap->va_birthtime.tv_sec = VNOVAL;
1119 	vap->va_birthtime.tv_nsec = VNOVAL;
1120 	vap->va_flags = VNOVAL;
1121 	vap->va_gen = VNOVAL;
1122 	vap->va_vaflags = 0;
1123 }
1124 
1125 /*
1126  * Try to reduce the total number of vnodes.
1127  *
1128  * This routine (and its user) are buggy in at least the following ways:
1129  * - all parameters were picked years ago when RAM sizes were significantly
1130  *   smaller
1131  * - it can pick vnodes based on pages used by the vm object, but filesystems
1132  *   like ZFS don't use it making the pick broken
1133  * - since ZFS has its own aging policy it gets partially combated by this one
1134  * - a dedicated method should be provided for filesystems to let them decide
1135  *   whether the vnode should be recycled
1136  *
1137  * This routine is called when we have too many vnodes.  It attempts
1138  * to free <count> vnodes and will potentially free vnodes that still
1139  * have VM backing store (VM backing store is typically the cause
1140  * of a vnode blowout so we want to do this).  Therefore, this operation
1141  * is not considered cheap.
1142  *
1143  * A number of conditions may prevent a vnode from being reclaimed.
1144  * the buffer cache may have references on the vnode, a directory
1145  * vnode may still have references due to the namei cache representing
1146  * underlying files, or the vnode may be in active use.   It is not
1147  * desirable to reuse such vnodes.  These conditions may cause the
1148  * number of vnodes to reach some minimum value regardless of what
1149  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
1150  *
1151  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1152  * 			 entries if this argument is strue
1153  * @param trigger	 Only reclaim vnodes with fewer than this many resident
1154  *			 pages.
1155  * @param target	 How many vnodes to reclaim.
1156  * @return		 The number of vnodes that were reclaimed.
1157  */
1158 static int
1159 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1160 {
1161 	struct vnode *vp, *mvp;
1162 	struct mount *mp;
1163 	struct vm_object *object;
1164 	u_long done;
1165 	bool retried;
1166 
1167 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1168 
1169 	retried = false;
1170 	done = 0;
1171 
1172 	mvp = vnode_list_reclaim_marker;
1173 restart:
1174 	vp = mvp;
1175 	while (done < target) {
1176 		vp = TAILQ_NEXT(vp, v_vnodelist);
1177 		if (__predict_false(vp == NULL))
1178 			break;
1179 
1180 		if (__predict_false(vp->v_type == VMARKER))
1181 			continue;
1182 
1183 		/*
1184 		 * If it's been deconstructed already, it's still
1185 		 * referenced, or it exceeds the trigger, skip it.
1186 		 * Also skip free vnodes.  We are trying to make space
1187 		 * to expand the free list, not reduce it.
1188 		 */
1189 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1190 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1191 			goto next_iter;
1192 
1193 		if (vp->v_type == VBAD || vp->v_type == VNON)
1194 			goto next_iter;
1195 
1196 		object = atomic_load_ptr(&vp->v_object);
1197 		if (object == NULL || object->resident_page_count > trigger) {
1198 			goto next_iter;
1199 		}
1200 
1201 		/*
1202 		 * Handle races against vnode allocation. Filesystems lock the
1203 		 * vnode some time after it gets returned from getnewvnode,
1204 		 * despite type and hold count being manipulated earlier.
1205 		 * Resorting to checking v_mount restores guarantees present
1206 		 * before the global list was reworked to contain all vnodes.
1207 		 */
1208 		if (!VI_TRYLOCK(vp))
1209 			goto next_iter;
1210 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1211 			VI_UNLOCK(vp);
1212 			goto next_iter;
1213 		}
1214 		if (vp->v_mount == NULL) {
1215 			VI_UNLOCK(vp);
1216 			goto next_iter;
1217 		}
1218 		vholdl(vp);
1219 		VI_UNLOCK(vp);
1220 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1221 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1222 		mtx_unlock(&vnode_list_mtx);
1223 
1224 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1225 			vdrop_recycle(vp);
1226 			goto next_iter_unlocked;
1227 		}
1228 		if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1229 			vdrop_recycle(vp);
1230 			vn_finished_write(mp);
1231 			goto next_iter_unlocked;
1232 		}
1233 
1234 		VI_LOCK(vp);
1235 		if (vp->v_usecount > 0 ||
1236 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1237 		    (vp->v_object != NULL && vp->v_object->handle == vp &&
1238 		    vp->v_object->resident_page_count > trigger)) {
1239 			VOP_UNLOCK(vp);
1240 			vdropl_recycle(vp);
1241 			vn_finished_write(mp);
1242 			goto next_iter_unlocked;
1243 		}
1244 		counter_u64_add(recycles_count, 1);
1245 		vgonel(vp);
1246 		VOP_UNLOCK(vp);
1247 		vdropl_recycle(vp);
1248 		vn_finished_write(mp);
1249 		done++;
1250 next_iter_unlocked:
1251 		maybe_yield();
1252 		mtx_lock(&vnode_list_mtx);
1253 		goto restart;
1254 next_iter:
1255 		MPASS(vp->v_type != VMARKER);
1256 		if (!should_yield())
1257 			continue;
1258 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1259 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1260 		mtx_unlock(&vnode_list_mtx);
1261 		kern_yield(PRI_USER);
1262 		mtx_lock(&vnode_list_mtx);
1263 		goto restart;
1264 	}
1265 	if (done == 0 && !retried) {
1266 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1267 		TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1268 		retried = true;
1269 		goto restart;
1270 	}
1271 	return (done);
1272 }
1273 
1274 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1275 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1276     0,
1277     "limit on vnode free requests per call to the vnlru_free routine");
1278 
1279 /*
1280  * Attempt to reduce the free list by the requested amount.
1281  */
1282 static int
1283 vnlru_free_impl(int count, struct vfsops *mnt_op, struct vnode *mvp)
1284 {
1285 	struct vnode *vp;
1286 	struct mount *mp;
1287 	int ocount;
1288 	bool retried;
1289 
1290 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1291 	if (count > max_vnlru_free)
1292 		count = max_vnlru_free;
1293 	ocount = count;
1294 	retried = false;
1295 	vp = mvp;
1296 	for (;;) {
1297 		if (count == 0) {
1298 			break;
1299 		}
1300 		vp = TAILQ_NEXT(vp, v_vnodelist);
1301 		if (__predict_false(vp == NULL)) {
1302 			/*
1303 			 * The free vnode marker can be past eligible vnodes:
1304 			 * 1. if vdbatch_process trylock failed
1305 			 * 2. if vtryrecycle failed
1306 			 *
1307 			 * If so, start the scan from scratch.
1308 			 */
1309 			if (!retried && vnlru_read_freevnodes() > 0) {
1310 				TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1311 				TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1312 				vp = mvp;
1313 				retried = true;
1314 				continue;
1315 			}
1316 
1317 			/*
1318 			 * Give up
1319 			 */
1320 			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1321 			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1322 			break;
1323 		}
1324 		if (__predict_false(vp->v_type == VMARKER))
1325 			continue;
1326 		if (vp->v_holdcnt > 0)
1327 			continue;
1328 		/*
1329 		 * Don't recycle if our vnode is from different type
1330 		 * of mount point.  Note that mp is type-safe, the
1331 		 * check does not reach unmapped address even if
1332 		 * vnode is reclaimed.
1333 		 */
1334 		if (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1335 		    mp->mnt_op != mnt_op) {
1336 			continue;
1337 		}
1338 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1339 			continue;
1340 		}
1341 		if (!vhold_recycle_free(vp))
1342 			continue;
1343 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1344 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1345 		mtx_unlock(&vnode_list_mtx);
1346 		/*
1347 		 * FIXME: ignores the return value, meaning it may be nothing
1348 		 * got recycled but it claims otherwise to the caller.
1349 		 *
1350 		 * Originally the value started being ignored in 2005 with
1351 		 * 114a1006a8204aa156e1f9ad6476cdff89cada7f .
1352 		 *
1353 		 * Respecting the value can run into significant stalls if most
1354 		 * vnodes belong to one file system and it has writes
1355 		 * suspended.  In presence of many threads and millions of
1356 		 * vnodes they keep contending on the vnode_list_mtx lock only
1357 		 * to find vnodes they can't recycle.
1358 		 *
1359 		 * The solution would be to pre-check if the vnode is likely to
1360 		 * be recycle-able, but it needs to happen with the
1361 		 * vnode_list_mtx lock held. This runs into a problem where
1362 		 * VOP_GETWRITEMOUNT (currently needed to find out about if
1363 		 * writes are frozen) can take locks which LOR against it.
1364 		 *
1365 		 * Check nullfs for one example (null_getwritemount).
1366 		 */
1367 		vtryrecycle(vp);
1368 		count--;
1369 		mtx_lock(&vnode_list_mtx);
1370 		vp = mvp;
1371 	}
1372 	return (ocount - count);
1373 }
1374 
1375 static int
1376 vnlru_free_locked(int count)
1377 {
1378 
1379 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1380 	return (vnlru_free_impl(count, NULL, vnode_list_free_marker));
1381 }
1382 
1383 void
1384 vnlru_free_vfsops(int count, struct vfsops *mnt_op, struct vnode *mvp)
1385 {
1386 
1387 	MPASS(mnt_op != NULL);
1388 	MPASS(mvp != NULL);
1389 	VNPASS(mvp->v_type == VMARKER, mvp);
1390 	mtx_lock(&vnode_list_mtx);
1391 	vnlru_free_impl(count, mnt_op, mvp);
1392 	mtx_unlock(&vnode_list_mtx);
1393 }
1394 
1395 struct vnode *
1396 vnlru_alloc_marker(void)
1397 {
1398 	struct vnode *mvp;
1399 
1400 	mvp = vn_alloc_marker(NULL);
1401 	mtx_lock(&vnode_list_mtx);
1402 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, mvp, v_vnodelist);
1403 	mtx_unlock(&vnode_list_mtx);
1404 	return (mvp);
1405 }
1406 
1407 void
1408 vnlru_free_marker(struct vnode *mvp)
1409 {
1410 	mtx_lock(&vnode_list_mtx);
1411 	TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1412 	mtx_unlock(&vnode_list_mtx);
1413 	vn_free_marker(mvp);
1414 }
1415 
1416 static void
1417 vnlru_recalc(void)
1418 {
1419 
1420 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1421 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1422 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1423 	vlowat = vhiwat / 2;
1424 }
1425 
1426 /*
1427  * Attempt to recycle vnodes in a context that is always safe to block.
1428  * Calling vlrurecycle() from the bowels of filesystem code has some
1429  * interesting deadlock problems.
1430  */
1431 static struct proc *vnlruproc;
1432 static int vnlruproc_sig;
1433 
1434 /*
1435  * The main freevnodes counter is only updated when threads requeue their vnode
1436  * batches. CPUs are conditionally walked to compute a more accurate total.
1437  *
1438  * Limit how much of a slop are we willing to tolerate. Note: the actual value
1439  * at any given moment can still exceed slop, but it should not be by significant
1440  * margin in practice.
1441  */
1442 #define VNLRU_FREEVNODES_SLOP 126
1443 
1444 static void __noinline
1445 vfs_freevnodes_rollup(int8_t *lfreevnodes)
1446 {
1447 
1448 	atomic_add_long(&freevnodes, *lfreevnodes);
1449 	*lfreevnodes = 0;
1450 	critical_exit();
1451 }
1452 
1453 static __inline void
1454 vfs_freevnodes_inc(void)
1455 {
1456 	int8_t *lfreevnodes;
1457 
1458 	critical_enter();
1459 	lfreevnodes = PCPU_PTR(vfs_freevnodes);
1460 	(*lfreevnodes)++;
1461 	if (__predict_false(*lfreevnodes == VNLRU_FREEVNODES_SLOP))
1462 		vfs_freevnodes_rollup(lfreevnodes);
1463 	else
1464 		critical_exit();
1465 }
1466 
1467 static __inline void
1468 vfs_freevnodes_dec(void)
1469 {
1470 	int8_t *lfreevnodes;
1471 
1472 	critical_enter();
1473 	lfreevnodes = PCPU_PTR(vfs_freevnodes);
1474 	(*lfreevnodes)--;
1475 	if (__predict_false(*lfreevnodes == -VNLRU_FREEVNODES_SLOP))
1476 		vfs_freevnodes_rollup(lfreevnodes);
1477 	else
1478 		critical_exit();
1479 }
1480 
1481 static u_long
1482 vnlru_read_freevnodes(void)
1483 {
1484 	long slop, rfreevnodes;
1485 	int cpu;
1486 
1487 	rfreevnodes = atomic_load_long(&freevnodes);
1488 
1489 	if (rfreevnodes > freevnodes_old)
1490 		slop = rfreevnodes - freevnodes_old;
1491 	else
1492 		slop = freevnodes_old - rfreevnodes;
1493 	if (slop < VNLRU_FREEVNODES_SLOP)
1494 		return (rfreevnodes >= 0 ? rfreevnodes : 0);
1495 	freevnodes_old = rfreevnodes;
1496 	CPU_FOREACH(cpu) {
1497 		freevnodes_old += cpuid_to_pcpu[cpu]->pc_vfs_freevnodes;
1498 	}
1499 	return (freevnodes_old >= 0 ? freevnodes_old : 0);
1500 }
1501 
1502 static bool
1503 vnlru_under(u_long rnumvnodes, u_long limit)
1504 {
1505 	u_long rfreevnodes, space;
1506 
1507 	if (__predict_false(rnumvnodes > desiredvnodes))
1508 		return (true);
1509 
1510 	space = desiredvnodes - rnumvnodes;
1511 	if (space < limit) {
1512 		rfreevnodes = vnlru_read_freevnodes();
1513 		if (rfreevnodes > wantfreevnodes)
1514 			space += rfreevnodes - wantfreevnodes;
1515 	}
1516 	return (space < limit);
1517 }
1518 
1519 static bool
1520 vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1521 {
1522 	long rfreevnodes, space;
1523 
1524 	if (__predict_false(rnumvnodes > desiredvnodes))
1525 		return (true);
1526 
1527 	space = desiredvnodes - rnumvnodes;
1528 	if (space < limit) {
1529 		rfreevnodes = atomic_load_long(&freevnodes);
1530 		if (rfreevnodes > wantfreevnodes)
1531 			space += rfreevnodes - wantfreevnodes;
1532 	}
1533 	return (space < limit);
1534 }
1535 
1536 static void
1537 vnlru_kick(void)
1538 {
1539 
1540 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1541 	if (vnlruproc_sig == 0) {
1542 		vnlruproc_sig = 1;
1543 		wakeup(vnlruproc);
1544 	}
1545 }
1546 
1547 static void
1548 vnlru_proc(void)
1549 {
1550 	u_long rnumvnodes, rfreevnodes, target;
1551 	unsigned long onumvnodes;
1552 	int done, force, trigger, usevnodes;
1553 	bool reclaim_nc_src, want_reread;
1554 
1555 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1556 	    SHUTDOWN_PRI_FIRST);
1557 
1558 	force = 0;
1559 	want_reread = false;
1560 	for (;;) {
1561 		kproc_suspend_check(vnlruproc);
1562 		mtx_lock(&vnode_list_mtx);
1563 		rnumvnodes = atomic_load_long(&numvnodes);
1564 
1565 		if (want_reread) {
1566 			force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1567 			want_reread = false;
1568 		}
1569 
1570 		/*
1571 		 * If numvnodes is too large (due to desiredvnodes being
1572 		 * adjusted using its sysctl, or emergency growth), first
1573 		 * try to reduce it by discarding from the free list.
1574 		 */
1575 		if (rnumvnodes > desiredvnodes) {
1576 			vnlru_free_locked(rnumvnodes - desiredvnodes);
1577 			rnumvnodes = atomic_load_long(&numvnodes);
1578 		}
1579 		/*
1580 		 * Sleep if the vnode cache is in a good state.  This is
1581 		 * when it is not over-full and has space for about a 4%
1582 		 * or 9% expansion (by growing its size or inexcessively
1583 		 * reducing its free list).  Otherwise, try to reclaim
1584 		 * space for a 10% expansion.
1585 		 */
1586 		if (vstir && force == 0) {
1587 			force = 1;
1588 			vstir = 0;
1589 		}
1590 		if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1591 			vnlruproc_sig = 0;
1592 			wakeup(&vnlruproc_sig);
1593 			msleep(vnlruproc, &vnode_list_mtx,
1594 			    PVFS|PDROP, "vlruwt", hz);
1595 			continue;
1596 		}
1597 		rfreevnodes = vnlru_read_freevnodes();
1598 
1599 		onumvnodes = rnumvnodes;
1600 		/*
1601 		 * Calculate parameters for recycling.  These are the same
1602 		 * throughout the loop to give some semblance of fairness.
1603 		 * The trigger point is to avoid recycling vnodes with lots
1604 		 * of resident pages.  We aren't trying to free memory; we
1605 		 * are trying to recycle or at least free vnodes.
1606 		 */
1607 		if (rnumvnodes <= desiredvnodes)
1608 			usevnodes = rnumvnodes - rfreevnodes;
1609 		else
1610 			usevnodes = rnumvnodes;
1611 		if (usevnodes <= 0)
1612 			usevnodes = 1;
1613 		/*
1614 		 * The trigger value is chosen to give a conservatively
1615 		 * large value to ensure that it alone doesn't prevent
1616 		 * making progress.  The value can easily be so large that
1617 		 * it is effectively infinite in some congested and
1618 		 * misconfigured cases, and this is necessary.  Normally
1619 		 * it is about 8 to 100 (pages), which is quite large.
1620 		 */
1621 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1622 		if (force < 2)
1623 			trigger = vsmalltrigger;
1624 		reclaim_nc_src = force >= 3;
1625 		target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1626 		target = target / 10 + 1;
1627 		done = vlrureclaim(reclaim_nc_src, trigger, target);
1628 		mtx_unlock(&vnode_list_mtx);
1629 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1630 			uma_reclaim(UMA_RECLAIM_DRAIN);
1631 		if (done == 0) {
1632 			if (force == 0 || force == 1) {
1633 				force = 2;
1634 				continue;
1635 			}
1636 			if (force == 2) {
1637 				force = 3;
1638 				continue;
1639 			}
1640 			want_reread = true;
1641 			force = 0;
1642 			vnlru_nowhere++;
1643 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1644 		} else {
1645 			want_reread = true;
1646 			kern_yield(PRI_USER);
1647 		}
1648 	}
1649 }
1650 
1651 static struct kproc_desc vnlru_kp = {
1652 	"vnlru",
1653 	vnlru_proc,
1654 	&vnlruproc
1655 };
1656 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1657     &vnlru_kp);
1658 
1659 /*
1660  * Routines having to do with the management of the vnode table.
1661  */
1662 
1663 /*
1664  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1665  * before we actually vgone().  This function must be called with the vnode
1666  * held to prevent the vnode from being returned to the free list midway
1667  * through vgone().
1668  */
1669 static int
1670 vtryrecycle(struct vnode *vp)
1671 {
1672 	struct mount *vnmp;
1673 
1674 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1675 	VNPASS(vp->v_holdcnt > 0, vp);
1676 	/*
1677 	 * This vnode may found and locked via some other list, if so we
1678 	 * can't recycle it yet.
1679 	 */
1680 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1681 		CTR2(KTR_VFS,
1682 		    "%s: impossible to recycle, vp %p lock is already held",
1683 		    __func__, vp);
1684 		vdrop_recycle(vp);
1685 		return (EWOULDBLOCK);
1686 	}
1687 	/*
1688 	 * Don't recycle if its filesystem is being suspended.
1689 	 */
1690 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1691 		VOP_UNLOCK(vp);
1692 		CTR2(KTR_VFS,
1693 		    "%s: impossible to recycle, cannot start the write for %p",
1694 		    __func__, vp);
1695 		vdrop_recycle(vp);
1696 		return (EBUSY);
1697 	}
1698 	/*
1699 	 * If we got this far, we need to acquire the interlock and see if
1700 	 * anyone picked up this vnode from another list.  If not, we will
1701 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1702 	 * will skip over it.
1703 	 */
1704 	VI_LOCK(vp);
1705 	if (vp->v_usecount) {
1706 		VOP_UNLOCK(vp);
1707 		vdropl_recycle(vp);
1708 		vn_finished_write(vnmp);
1709 		CTR2(KTR_VFS,
1710 		    "%s: impossible to recycle, %p is already referenced",
1711 		    __func__, vp);
1712 		return (EBUSY);
1713 	}
1714 	if (!VN_IS_DOOMED(vp)) {
1715 		counter_u64_add(recycles_free_count, 1);
1716 		vgonel(vp);
1717 	}
1718 	VOP_UNLOCK(vp);
1719 	vdropl_recycle(vp);
1720 	vn_finished_write(vnmp);
1721 	return (0);
1722 }
1723 
1724 /*
1725  * Allocate a new vnode.
1726  *
1727  * The operation never returns an error. Returning an error was disabled
1728  * in r145385 (dated 2005) with the following comment:
1729  *
1730  * XXX Not all VFS_VGET/ffs_vget callers check returns.
1731  *
1732  * Given the age of this commit (almost 15 years at the time of writing this
1733  * comment) restoring the ability to fail requires a significant audit of
1734  * all codepaths.
1735  *
1736  * The routine can try to free a vnode or stall for up to 1 second waiting for
1737  * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1738  */
1739 static u_long vn_alloc_cyclecount;
1740 static u_long vn_alloc_sleeps;
1741 
1742 SYSCTL_ULONG(_vfs, OID_AUTO, vnode_alloc_sleeps, CTLFLAG_RD, &vn_alloc_sleeps, 0,
1743     "Number of times vnode allocation blocked waiting on vnlru");
1744 
1745 static struct vnode * __noinline
1746 vn_alloc_hard(struct mount *mp)
1747 {
1748 	u_long rnumvnodes, rfreevnodes;
1749 
1750 	mtx_lock(&vnode_list_mtx);
1751 	rnumvnodes = atomic_load_long(&numvnodes);
1752 	if (rnumvnodes + 1 < desiredvnodes) {
1753 		vn_alloc_cyclecount = 0;
1754 		goto alloc;
1755 	}
1756 	rfreevnodes = vnlru_read_freevnodes();
1757 	if (vn_alloc_cyclecount++ >= rfreevnodes) {
1758 		vn_alloc_cyclecount = 0;
1759 		vstir = 1;
1760 	}
1761 	/*
1762 	 * Grow the vnode cache if it will not be above its target max
1763 	 * after growing.  Otherwise, if the free list is nonempty, try
1764 	 * to reclaim 1 item from it before growing the cache (possibly
1765 	 * above its target max if the reclamation failed or is delayed).
1766 	 * Otherwise, wait for some space.  In all cases, schedule
1767 	 * vnlru_proc() if we are getting short of space.  The watermarks
1768 	 * should be chosen so that we never wait or even reclaim from
1769 	 * the free list to below its target minimum.
1770 	 */
1771 	if (vnlru_free_locked(1) > 0)
1772 		goto alloc;
1773 	if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1774 		/*
1775 		 * Wait for space for a new vnode.
1776 		 */
1777 		vnlru_kick();
1778 		vn_alloc_sleeps++;
1779 		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1780 		if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1781 		    vnlru_read_freevnodes() > 1)
1782 			vnlru_free_locked(1);
1783 	}
1784 alloc:
1785 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1786 	if (vnlru_under(rnumvnodes, vlowat))
1787 		vnlru_kick();
1788 	mtx_unlock(&vnode_list_mtx);
1789 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1790 }
1791 
1792 static struct vnode *
1793 vn_alloc(struct mount *mp)
1794 {
1795 	u_long rnumvnodes;
1796 
1797 	if (__predict_false(vn_alloc_cyclecount != 0))
1798 		return (vn_alloc_hard(mp));
1799 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1800 	if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1801 		atomic_subtract_long(&numvnodes, 1);
1802 		return (vn_alloc_hard(mp));
1803 	}
1804 
1805 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1806 }
1807 
1808 static void
1809 vn_free(struct vnode *vp)
1810 {
1811 
1812 	atomic_subtract_long(&numvnodes, 1);
1813 	uma_zfree_smr(vnode_zone, vp);
1814 }
1815 
1816 /*
1817  * Return the next vnode from the free list.
1818  */
1819 int
1820 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1821     struct vnode **vpp)
1822 {
1823 	struct vnode *vp;
1824 	struct thread *td;
1825 	struct lock_object *lo;
1826 
1827 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1828 
1829 	KASSERT(vops->registered,
1830 	    ("%s: not registered vector op %p\n", __func__, vops));
1831 	cache_validate_vop_vector(mp, vops);
1832 
1833 	td = curthread;
1834 	if (td->td_vp_reserved != NULL) {
1835 		vp = td->td_vp_reserved;
1836 		td->td_vp_reserved = NULL;
1837 	} else {
1838 		vp = vn_alloc(mp);
1839 	}
1840 	counter_u64_add(vnodes_created, 1);
1841 
1842 	vn_set_state(vp, VSTATE_UNINITIALIZED);
1843 
1844 	/*
1845 	 * Locks are given the generic name "vnode" when created.
1846 	 * Follow the historic practice of using the filesystem
1847 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1848 	 *
1849 	 * Locks live in a witness group keyed on their name. Thus,
1850 	 * when a lock is renamed, it must also move from the witness
1851 	 * group of its old name to the witness group of its new name.
1852 	 *
1853 	 * The change only needs to be made when the vnode moves
1854 	 * from one filesystem type to another. We ensure that each
1855 	 * filesystem use a single static name pointer for its tag so
1856 	 * that we can compare pointers rather than doing a strcmp().
1857 	 */
1858 	lo = &vp->v_vnlock->lock_object;
1859 #ifdef WITNESS
1860 	if (lo->lo_name != tag) {
1861 #endif
1862 		lo->lo_name = tag;
1863 #ifdef WITNESS
1864 		WITNESS_DESTROY(lo);
1865 		WITNESS_INIT(lo, tag);
1866 	}
1867 #endif
1868 	/*
1869 	 * By default, don't allow shared locks unless filesystems opt-in.
1870 	 */
1871 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1872 	/*
1873 	 * Finalize various vnode identity bits.
1874 	 */
1875 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1876 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1877 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1878 	vp->v_type = VNON;
1879 	vp->v_op = vops;
1880 	vp->v_irflag = 0;
1881 	v_init_counters(vp);
1882 	vn_seqc_init(vp);
1883 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1884 #ifdef DIAGNOSTIC
1885 	if (mp == NULL && vops != &dead_vnodeops)
1886 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1887 #endif
1888 #ifdef MAC
1889 	mac_vnode_init(vp);
1890 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1891 		mac_vnode_associate_singlelabel(mp, vp);
1892 #endif
1893 	if (mp != NULL) {
1894 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1895 	}
1896 
1897 	/*
1898 	 * For the filesystems which do not use vfs_hash_insert(),
1899 	 * still initialize v_hash to have vfs_hash_index() useful.
1900 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1901 	 * its own hashing.
1902 	 */
1903 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1904 
1905 	*vpp = vp;
1906 	return (0);
1907 }
1908 
1909 void
1910 getnewvnode_reserve(void)
1911 {
1912 	struct thread *td;
1913 
1914 	td = curthread;
1915 	MPASS(td->td_vp_reserved == NULL);
1916 	td->td_vp_reserved = vn_alloc(NULL);
1917 }
1918 
1919 void
1920 getnewvnode_drop_reserve(void)
1921 {
1922 	struct thread *td;
1923 
1924 	td = curthread;
1925 	if (td->td_vp_reserved != NULL) {
1926 		vn_free(td->td_vp_reserved);
1927 		td->td_vp_reserved = NULL;
1928 	}
1929 }
1930 
1931 static void __noinline
1932 freevnode(struct vnode *vp)
1933 {
1934 	struct bufobj *bo;
1935 
1936 	/*
1937 	 * The vnode has been marked for destruction, so free it.
1938 	 *
1939 	 * The vnode will be returned to the zone where it will
1940 	 * normally remain until it is needed for another vnode. We
1941 	 * need to cleanup (or verify that the cleanup has already
1942 	 * been done) any residual data left from its current use
1943 	 * so as not to contaminate the freshly allocated vnode.
1944 	 */
1945 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1946 	/*
1947 	 * Paired with vgone.
1948 	 */
1949 	vn_seqc_write_end_free(vp);
1950 
1951 	bo = &vp->v_bufobj;
1952 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1953 	VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp);
1954 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1955 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1956 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1957 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1958 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1959 	    ("clean blk trie not empty"));
1960 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1961 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1962 	    ("dirty blk trie not empty"));
1963 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1964 	    ("Dangling rangelock waiters"));
1965 	VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp,
1966 	    ("Leaked inactivation"));
1967 	VI_UNLOCK(vp);
1968 	cache_assert_no_entries(vp);
1969 
1970 #ifdef MAC
1971 	mac_vnode_destroy(vp);
1972 #endif
1973 	if (vp->v_pollinfo != NULL) {
1974 		/*
1975 		 * Use LK_NOWAIT to shut up witness about the lock. We may get
1976 		 * here while having another vnode locked when trying to
1977 		 * satisfy a lookup and needing to recycle.
1978 		 */
1979 		VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT);
1980 		destroy_vpollinfo(vp->v_pollinfo);
1981 		VOP_UNLOCK(vp);
1982 		vp->v_pollinfo = NULL;
1983 	}
1984 	vp->v_mountedhere = NULL;
1985 	vp->v_unpcb = NULL;
1986 	vp->v_rdev = NULL;
1987 	vp->v_fifoinfo = NULL;
1988 	vp->v_iflag = 0;
1989 	vp->v_vflag = 0;
1990 	bo->bo_flag = 0;
1991 	vn_free(vp);
1992 }
1993 
1994 /*
1995  * Delete from old mount point vnode list, if on one.
1996  */
1997 static void
1998 delmntque(struct vnode *vp)
1999 {
2000 	struct mount *mp;
2001 
2002 	VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
2003 
2004 	mp = vp->v_mount;
2005 	MNT_ILOCK(mp);
2006 	VI_LOCK(vp);
2007 	vp->v_mount = NULL;
2008 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
2009 		("bad mount point vnode list size"));
2010 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2011 	mp->mnt_nvnodelistsize--;
2012 	MNT_REL(mp);
2013 	MNT_IUNLOCK(mp);
2014 	/*
2015 	 * The caller expects the interlock to be still held.
2016 	 */
2017 	ASSERT_VI_LOCKED(vp, __func__);
2018 }
2019 
2020 static int
2021 insmntque1_int(struct vnode *vp, struct mount *mp, bool dtr)
2022 {
2023 
2024 	KASSERT(vp->v_mount == NULL,
2025 		("insmntque: vnode already on per mount vnode list"));
2026 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
2027 	if ((mp->mnt_kern_flag & MNTK_UNLOCKED_INSMNTQUE) == 0) {
2028 		ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
2029 	} else {
2030 		KASSERT(!dtr,
2031 		    ("%s: can't have MNTK_UNLOCKED_INSMNTQUE and cleanup",
2032 		    __func__));
2033 	}
2034 
2035 	/*
2036 	 * We acquire the vnode interlock early to ensure that the
2037 	 * vnode cannot be recycled by another process releasing a
2038 	 * holdcnt on it before we get it on both the vnode list
2039 	 * and the active vnode list. The mount mutex protects only
2040 	 * manipulation of the vnode list and the vnode freelist
2041 	 * mutex protects only manipulation of the active vnode list.
2042 	 * Hence the need to hold the vnode interlock throughout.
2043 	 */
2044 	MNT_ILOCK(mp);
2045 	VI_LOCK(vp);
2046 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
2047 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
2048 	    mp->mnt_nvnodelistsize == 0)) &&
2049 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
2050 		VI_UNLOCK(vp);
2051 		MNT_IUNLOCK(mp);
2052 		if (dtr) {
2053 			vp->v_data = NULL;
2054 			vp->v_op = &dead_vnodeops;
2055 			vgone(vp);
2056 			vput(vp);
2057 		}
2058 		return (EBUSY);
2059 	}
2060 	vp->v_mount = mp;
2061 	MNT_REF(mp);
2062 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2063 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
2064 		("neg mount point vnode list size"));
2065 	mp->mnt_nvnodelistsize++;
2066 	VI_UNLOCK(vp);
2067 	MNT_IUNLOCK(mp);
2068 	return (0);
2069 }
2070 
2071 /*
2072  * Insert into list of vnodes for the new mount point, if available.
2073  * insmntque() reclaims the vnode on insertion failure, insmntque1()
2074  * leaves handling of the vnode to the caller.
2075  */
2076 int
2077 insmntque(struct vnode *vp, struct mount *mp)
2078 {
2079 	return (insmntque1_int(vp, mp, true));
2080 }
2081 
2082 int
2083 insmntque1(struct vnode *vp, struct mount *mp)
2084 {
2085 	return (insmntque1_int(vp, mp, false));
2086 }
2087 
2088 /*
2089  * Flush out and invalidate all buffers associated with a bufobj
2090  * Called with the underlying object locked.
2091  */
2092 int
2093 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
2094 {
2095 	int error;
2096 
2097 	BO_LOCK(bo);
2098 	if (flags & V_SAVE) {
2099 		error = bufobj_wwait(bo, slpflag, slptimeo);
2100 		if (error) {
2101 			BO_UNLOCK(bo);
2102 			return (error);
2103 		}
2104 		if (bo->bo_dirty.bv_cnt > 0) {
2105 			BO_UNLOCK(bo);
2106 			do {
2107 				error = BO_SYNC(bo, MNT_WAIT);
2108 			} while (error == ERELOOKUP);
2109 			if (error != 0)
2110 				return (error);
2111 			BO_LOCK(bo);
2112 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) {
2113 				BO_UNLOCK(bo);
2114 				return (EBUSY);
2115 			}
2116 		}
2117 	}
2118 	/*
2119 	 * If you alter this loop please notice that interlock is dropped and
2120 	 * reacquired in flushbuflist.  Special care is needed to ensure that
2121 	 * no race conditions occur from this.
2122 	 */
2123 	do {
2124 		error = flushbuflist(&bo->bo_clean,
2125 		    flags, bo, slpflag, slptimeo);
2126 		if (error == 0 && !(flags & V_CLEANONLY))
2127 			error = flushbuflist(&bo->bo_dirty,
2128 			    flags, bo, slpflag, slptimeo);
2129 		if (error != 0 && error != EAGAIN) {
2130 			BO_UNLOCK(bo);
2131 			return (error);
2132 		}
2133 	} while (error != 0);
2134 
2135 	/*
2136 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
2137 	 * have write I/O in-progress but if there is a VM object then the
2138 	 * VM object can also have read-I/O in-progress.
2139 	 */
2140 	do {
2141 		bufobj_wwait(bo, 0, 0);
2142 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
2143 			BO_UNLOCK(bo);
2144 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
2145 			BO_LOCK(bo);
2146 		}
2147 	} while (bo->bo_numoutput > 0);
2148 	BO_UNLOCK(bo);
2149 
2150 	/*
2151 	 * Destroy the copy in the VM cache, too.
2152 	 */
2153 	if (bo->bo_object != NULL &&
2154 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
2155 		VM_OBJECT_WLOCK(bo->bo_object);
2156 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
2157 		    OBJPR_CLEANONLY : 0);
2158 		VM_OBJECT_WUNLOCK(bo->bo_object);
2159 	}
2160 
2161 #ifdef INVARIANTS
2162 	BO_LOCK(bo);
2163 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
2164 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
2165 	    bo->bo_clean.bv_cnt > 0))
2166 		panic("vinvalbuf: flush failed");
2167 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
2168 	    bo->bo_dirty.bv_cnt > 0)
2169 		panic("vinvalbuf: flush dirty failed");
2170 	BO_UNLOCK(bo);
2171 #endif
2172 	return (0);
2173 }
2174 
2175 /*
2176  * Flush out and invalidate all buffers associated with a vnode.
2177  * Called with the underlying object locked.
2178  */
2179 int
2180 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
2181 {
2182 
2183 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2184 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
2185 	if (vp->v_object != NULL && vp->v_object->handle != vp)
2186 		return (0);
2187 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
2188 }
2189 
2190 /*
2191  * Flush out buffers on the specified list.
2192  *
2193  */
2194 static int
2195 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2196     int slptimeo)
2197 {
2198 	struct buf *bp, *nbp;
2199 	int retval, error;
2200 	daddr_t lblkno;
2201 	b_xflags_t xflags;
2202 
2203 	ASSERT_BO_WLOCKED(bo);
2204 
2205 	retval = 0;
2206 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2207 		/*
2208 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2209 		 * do not skip any buffers. If we are flushing only V_NORMAL
2210 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2211 		 * flushing only V_ALT buffers then skip buffers not marked
2212 		 * as BX_ALTDATA.
2213 		 */
2214 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2215 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2216 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2217 			continue;
2218 		}
2219 		if (nbp != NULL) {
2220 			lblkno = nbp->b_lblkno;
2221 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2222 		}
2223 		retval = EAGAIN;
2224 		error = BUF_TIMELOCK(bp,
2225 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2226 		    "flushbuf", slpflag, slptimeo);
2227 		if (error) {
2228 			BO_LOCK(bo);
2229 			return (error != ENOLCK ? error : EAGAIN);
2230 		}
2231 		KASSERT(bp->b_bufobj == bo,
2232 		    ("bp %p wrong b_bufobj %p should be %p",
2233 		    bp, bp->b_bufobj, bo));
2234 		/*
2235 		 * XXX Since there are no node locks for NFS, I
2236 		 * believe there is a slight chance that a delayed
2237 		 * write will occur while sleeping just above, so
2238 		 * check for it.
2239 		 */
2240 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2241 		    (flags & V_SAVE)) {
2242 			bremfree(bp);
2243 			bp->b_flags |= B_ASYNC;
2244 			bwrite(bp);
2245 			BO_LOCK(bo);
2246 			return (EAGAIN);	/* XXX: why not loop ? */
2247 		}
2248 		bremfree(bp);
2249 		bp->b_flags |= (B_INVAL | B_RELBUF);
2250 		bp->b_flags &= ~B_ASYNC;
2251 		brelse(bp);
2252 		BO_LOCK(bo);
2253 		if (nbp == NULL)
2254 			break;
2255 		nbp = gbincore(bo, lblkno);
2256 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2257 		    != xflags)
2258 			break;			/* nbp invalid */
2259 	}
2260 	return (retval);
2261 }
2262 
2263 int
2264 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2265 {
2266 	struct buf *bp;
2267 	int error;
2268 	daddr_t lblkno;
2269 
2270 	ASSERT_BO_LOCKED(bo);
2271 
2272 	for (lblkno = startn;;) {
2273 again:
2274 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2275 		if (bp == NULL || bp->b_lblkno >= endn ||
2276 		    bp->b_lblkno < startn)
2277 			break;
2278 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2279 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2280 		if (error != 0) {
2281 			BO_RLOCK(bo);
2282 			if (error == ENOLCK)
2283 				goto again;
2284 			return (error);
2285 		}
2286 		KASSERT(bp->b_bufobj == bo,
2287 		    ("bp %p wrong b_bufobj %p should be %p",
2288 		    bp, bp->b_bufobj, bo));
2289 		lblkno = bp->b_lblkno + 1;
2290 		if ((bp->b_flags & B_MANAGED) == 0)
2291 			bremfree(bp);
2292 		bp->b_flags |= B_RELBUF;
2293 		/*
2294 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2295 		 * pages backing each buffer in the range are unlikely to be
2296 		 * reused.  Dirty buffers will have the hint applied once
2297 		 * they've been written.
2298 		 */
2299 		if ((bp->b_flags & B_VMIO) != 0)
2300 			bp->b_flags |= B_NOREUSE;
2301 		brelse(bp);
2302 		BO_RLOCK(bo);
2303 	}
2304 	return (0);
2305 }
2306 
2307 /*
2308  * Truncate a file's buffer and pages to a specified length.  This
2309  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2310  * sync activity.
2311  */
2312 int
2313 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2314 {
2315 	struct buf *bp, *nbp;
2316 	struct bufobj *bo;
2317 	daddr_t startlbn;
2318 
2319 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2320 	    vp, blksize, (uintmax_t)length);
2321 
2322 	/*
2323 	 * Round up to the *next* lbn.
2324 	 */
2325 	startlbn = howmany(length, blksize);
2326 
2327 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2328 
2329 	bo = &vp->v_bufobj;
2330 restart_unlocked:
2331 	BO_LOCK(bo);
2332 
2333 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2334 		;
2335 
2336 	if (length > 0) {
2337 restartsync:
2338 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2339 			if (bp->b_lblkno > 0)
2340 				continue;
2341 			/*
2342 			 * Since we hold the vnode lock this should only
2343 			 * fail if we're racing with the buf daemon.
2344 			 */
2345 			if (BUF_LOCK(bp,
2346 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2347 			    BO_LOCKPTR(bo)) == ENOLCK)
2348 				goto restart_unlocked;
2349 
2350 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2351 			    ("buf(%p) on dirty queue without DELWRI", bp));
2352 
2353 			bremfree(bp);
2354 			bawrite(bp);
2355 			BO_LOCK(bo);
2356 			goto restartsync;
2357 		}
2358 	}
2359 
2360 	bufobj_wwait(bo, 0, 0);
2361 	BO_UNLOCK(bo);
2362 	vnode_pager_setsize(vp, length);
2363 
2364 	return (0);
2365 }
2366 
2367 /*
2368  * Invalidate the cached pages of a file's buffer within the range of block
2369  * numbers [startlbn, endlbn).
2370  */
2371 void
2372 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2373     int blksize)
2374 {
2375 	struct bufobj *bo;
2376 	off_t start, end;
2377 
2378 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2379 
2380 	start = blksize * startlbn;
2381 	end = blksize * endlbn;
2382 
2383 	bo = &vp->v_bufobj;
2384 	BO_LOCK(bo);
2385 	MPASS(blksize == bo->bo_bsize);
2386 
2387 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2388 		;
2389 
2390 	BO_UNLOCK(bo);
2391 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2392 }
2393 
2394 static int
2395 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2396     daddr_t startlbn, daddr_t endlbn)
2397 {
2398 	struct buf *bp, *nbp;
2399 	bool anyfreed;
2400 
2401 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2402 	ASSERT_BO_LOCKED(bo);
2403 
2404 	do {
2405 		anyfreed = false;
2406 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2407 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2408 				continue;
2409 			if (BUF_LOCK(bp,
2410 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2411 			    BO_LOCKPTR(bo)) == ENOLCK) {
2412 				BO_LOCK(bo);
2413 				return (EAGAIN);
2414 			}
2415 
2416 			bremfree(bp);
2417 			bp->b_flags |= B_INVAL | B_RELBUF;
2418 			bp->b_flags &= ~B_ASYNC;
2419 			brelse(bp);
2420 			anyfreed = true;
2421 
2422 			BO_LOCK(bo);
2423 			if (nbp != NULL &&
2424 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2425 			    nbp->b_vp != vp ||
2426 			    (nbp->b_flags & B_DELWRI) != 0))
2427 				return (EAGAIN);
2428 		}
2429 
2430 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2431 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2432 				continue;
2433 			if (BUF_LOCK(bp,
2434 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2435 			    BO_LOCKPTR(bo)) == ENOLCK) {
2436 				BO_LOCK(bo);
2437 				return (EAGAIN);
2438 			}
2439 			bremfree(bp);
2440 			bp->b_flags |= B_INVAL | B_RELBUF;
2441 			bp->b_flags &= ~B_ASYNC;
2442 			brelse(bp);
2443 			anyfreed = true;
2444 
2445 			BO_LOCK(bo);
2446 			if (nbp != NULL &&
2447 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2448 			    (nbp->b_vp != vp) ||
2449 			    (nbp->b_flags & B_DELWRI) == 0))
2450 				return (EAGAIN);
2451 		}
2452 	} while (anyfreed);
2453 	return (0);
2454 }
2455 
2456 static void
2457 buf_vlist_remove(struct buf *bp)
2458 {
2459 	struct bufv *bv;
2460 	b_xflags_t flags;
2461 
2462 	flags = bp->b_xflags;
2463 
2464 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2465 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2466 	KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 &&
2467 	    (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN),
2468 	    ("%s: buffer %p has invalid queue state", __func__, bp));
2469 
2470 	if ((flags & BX_VNDIRTY) != 0)
2471 		bv = &bp->b_bufobj->bo_dirty;
2472 	else
2473 		bv = &bp->b_bufobj->bo_clean;
2474 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2475 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2476 	bv->bv_cnt--;
2477 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2478 }
2479 
2480 /*
2481  * Add the buffer to the sorted clean or dirty block list.
2482  *
2483  * NOTE: xflags is passed as a constant, optimizing this inline function!
2484  */
2485 static void
2486 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2487 {
2488 	struct bufv *bv;
2489 	struct buf *n;
2490 	int error;
2491 
2492 	ASSERT_BO_WLOCKED(bo);
2493 	KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2494 	    ("buf_vlist_add: bo %p does not allow bufs", bo));
2495 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2496 	    ("dead bo %p", bo));
2497 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2498 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2499 	bp->b_xflags |= xflags;
2500 	if (xflags & BX_VNDIRTY)
2501 		bv = &bo->bo_dirty;
2502 	else
2503 		bv = &bo->bo_clean;
2504 
2505 	/*
2506 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2507 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2508 	 * than _ge.
2509 	 */
2510 	if (bv->bv_cnt == 0 ||
2511 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2512 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2513 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2514 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2515 	else
2516 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2517 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2518 	if (error)
2519 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2520 	bv->bv_cnt++;
2521 }
2522 
2523 /*
2524  * Look up a buffer using the buffer tries.
2525  */
2526 struct buf *
2527 gbincore(struct bufobj *bo, daddr_t lblkno)
2528 {
2529 	struct buf *bp;
2530 
2531 	ASSERT_BO_LOCKED(bo);
2532 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2533 	if (bp != NULL)
2534 		return (bp);
2535 	return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno));
2536 }
2537 
2538 /*
2539  * Look up a buf using the buffer tries, without the bufobj lock.  This relies
2540  * on SMR for safe lookup, and bufs being in a no-free zone to provide type
2541  * stability of the result.  Like other lockless lookups, the found buf may
2542  * already be invalid by the time this function returns.
2543  */
2544 struct buf *
2545 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno)
2546 {
2547 	struct buf *bp;
2548 
2549 	ASSERT_BO_UNLOCKED(bo);
2550 	bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno);
2551 	if (bp != NULL)
2552 		return (bp);
2553 	return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno));
2554 }
2555 
2556 /*
2557  * Associate a buffer with a vnode.
2558  */
2559 void
2560 bgetvp(struct vnode *vp, struct buf *bp)
2561 {
2562 	struct bufobj *bo;
2563 
2564 	bo = &vp->v_bufobj;
2565 	ASSERT_BO_WLOCKED(bo);
2566 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2567 
2568 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2569 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2570 	    ("bgetvp: bp already attached! %p", bp));
2571 
2572 	vhold(vp);
2573 	bp->b_vp = vp;
2574 	bp->b_bufobj = bo;
2575 	/*
2576 	 * Insert onto list for new vnode.
2577 	 */
2578 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2579 }
2580 
2581 /*
2582  * Disassociate a buffer from a vnode.
2583  */
2584 void
2585 brelvp(struct buf *bp)
2586 {
2587 	struct bufobj *bo;
2588 	struct vnode *vp;
2589 
2590 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2591 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2592 
2593 	/*
2594 	 * Delete from old vnode list, if on one.
2595 	 */
2596 	vp = bp->b_vp;		/* XXX */
2597 	bo = bp->b_bufobj;
2598 	BO_LOCK(bo);
2599 	buf_vlist_remove(bp);
2600 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2601 		bo->bo_flag &= ~BO_ONWORKLST;
2602 		mtx_lock(&sync_mtx);
2603 		LIST_REMOVE(bo, bo_synclist);
2604 		syncer_worklist_len--;
2605 		mtx_unlock(&sync_mtx);
2606 	}
2607 	bp->b_vp = NULL;
2608 	bp->b_bufobj = NULL;
2609 	BO_UNLOCK(bo);
2610 	vdrop(vp);
2611 }
2612 
2613 /*
2614  * Add an item to the syncer work queue.
2615  */
2616 static void
2617 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2618 {
2619 	int slot;
2620 
2621 	ASSERT_BO_WLOCKED(bo);
2622 
2623 	mtx_lock(&sync_mtx);
2624 	if (bo->bo_flag & BO_ONWORKLST)
2625 		LIST_REMOVE(bo, bo_synclist);
2626 	else {
2627 		bo->bo_flag |= BO_ONWORKLST;
2628 		syncer_worklist_len++;
2629 	}
2630 
2631 	if (delay > syncer_maxdelay - 2)
2632 		delay = syncer_maxdelay - 2;
2633 	slot = (syncer_delayno + delay) & syncer_mask;
2634 
2635 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2636 	mtx_unlock(&sync_mtx);
2637 }
2638 
2639 static int
2640 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2641 {
2642 	int error, len;
2643 
2644 	mtx_lock(&sync_mtx);
2645 	len = syncer_worklist_len - sync_vnode_count;
2646 	mtx_unlock(&sync_mtx);
2647 	error = SYSCTL_OUT(req, &len, sizeof(len));
2648 	return (error);
2649 }
2650 
2651 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2652     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2653     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2654 
2655 static struct proc *updateproc;
2656 static void sched_sync(void);
2657 static struct kproc_desc up_kp = {
2658 	"syncer",
2659 	sched_sync,
2660 	&updateproc
2661 };
2662 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2663 
2664 static int
2665 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2666 {
2667 	struct vnode *vp;
2668 	struct mount *mp;
2669 
2670 	*bo = LIST_FIRST(slp);
2671 	if (*bo == NULL)
2672 		return (0);
2673 	vp = bo2vnode(*bo);
2674 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2675 		return (1);
2676 	/*
2677 	 * We use vhold in case the vnode does not
2678 	 * successfully sync.  vhold prevents the vnode from
2679 	 * going away when we unlock the sync_mtx so that
2680 	 * we can acquire the vnode interlock.
2681 	 */
2682 	vholdl(vp);
2683 	mtx_unlock(&sync_mtx);
2684 	VI_UNLOCK(vp);
2685 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2686 		vdrop(vp);
2687 		mtx_lock(&sync_mtx);
2688 		return (*bo == LIST_FIRST(slp));
2689 	}
2690 	MPASSERT(mp == NULL || (curthread->td_pflags & TDP_IGNSUSP) != 0 ||
2691 	    (mp->mnt_kern_flag & MNTK_SUSPENDED) == 0, mp,
2692 	    ("suspended mp syncing vp %p", vp));
2693 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2694 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2695 	VOP_UNLOCK(vp);
2696 	vn_finished_write(mp);
2697 	BO_LOCK(*bo);
2698 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2699 		/*
2700 		 * Put us back on the worklist.  The worklist
2701 		 * routine will remove us from our current
2702 		 * position and then add us back in at a later
2703 		 * position.
2704 		 */
2705 		vn_syncer_add_to_worklist(*bo, syncdelay);
2706 	}
2707 	BO_UNLOCK(*bo);
2708 	vdrop(vp);
2709 	mtx_lock(&sync_mtx);
2710 	return (0);
2711 }
2712 
2713 static int first_printf = 1;
2714 
2715 /*
2716  * System filesystem synchronizer daemon.
2717  */
2718 static void
2719 sched_sync(void)
2720 {
2721 	struct synclist *next, *slp;
2722 	struct bufobj *bo;
2723 	long starttime;
2724 	struct thread *td = curthread;
2725 	int last_work_seen;
2726 	int net_worklist_len;
2727 	int syncer_final_iter;
2728 	int error;
2729 
2730 	last_work_seen = 0;
2731 	syncer_final_iter = 0;
2732 	syncer_state = SYNCER_RUNNING;
2733 	starttime = time_uptime;
2734 	td->td_pflags |= TDP_NORUNNINGBUF;
2735 
2736 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2737 	    SHUTDOWN_PRI_LAST);
2738 
2739 	mtx_lock(&sync_mtx);
2740 	for (;;) {
2741 		if (syncer_state == SYNCER_FINAL_DELAY &&
2742 		    syncer_final_iter == 0) {
2743 			mtx_unlock(&sync_mtx);
2744 			kproc_suspend_check(td->td_proc);
2745 			mtx_lock(&sync_mtx);
2746 		}
2747 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2748 		if (syncer_state != SYNCER_RUNNING &&
2749 		    starttime != time_uptime) {
2750 			if (first_printf) {
2751 				printf("\nSyncing disks, vnodes remaining... ");
2752 				first_printf = 0;
2753 			}
2754 			printf("%d ", net_worklist_len);
2755 		}
2756 		starttime = time_uptime;
2757 
2758 		/*
2759 		 * Push files whose dirty time has expired.  Be careful
2760 		 * of interrupt race on slp queue.
2761 		 *
2762 		 * Skip over empty worklist slots when shutting down.
2763 		 */
2764 		do {
2765 			slp = &syncer_workitem_pending[syncer_delayno];
2766 			syncer_delayno += 1;
2767 			if (syncer_delayno == syncer_maxdelay)
2768 				syncer_delayno = 0;
2769 			next = &syncer_workitem_pending[syncer_delayno];
2770 			/*
2771 			 * If the worklist has wrapped since the
2772 			 * it was emptied of all but syncer vnodes,
2773 			 * switch to the FINAL_DELAY state and run
2774 			 * for one more second.
2775 			 */
2776 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2777 			    net_worklist_len == 0 &&
2778 			    last_work_seen == syncer_delayno) {
2779 				syncer_state = SYNCER_FINAL_DELAY;
2780 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2781 			}
2782 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2783 		    syncer_worklist_len > 0);
2784 
2785 		/*
2786 		 * Keep track of the last time there was anything
2787 		 * on the worklist other than syncer vnodes.
2788 		 * Return to the SHUTTING_DOWN state if any
2789 		 * new work appears.
2790 		 */
2791 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2792 			last_work_seen = syncer_delayno;
2793 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2794 			syncer_state = SYNCER_SHUTTING_DOWN;
2795 		while (!LIST_EMPTY(slp)) {
2796 			error = sync_vnode(slp, &bo, td);
2797 			if (error == 1) {
2798 				LIST_REMOVE(bo, bo_synclist);
2799 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2800 				continue;
2801 			}
2802 
2803 			if (first_printf == 0) {
2804 				/*
2805 				 * Drop the sync mutex, because some watchdog
2806 				 * drivers need to sleep while patting
2807 				 */
2808 				mtx_unlock(&sync_mtx);
2809 				wdog_kern_pat(WD_LASTVAL);
2810 				mtx_lock(&sync_mtx);
2811 			}
2812 		}
2813 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2814 			syncer_final_iter--;
2815 		/*
2816 		 * The variable rushjob allows the kernel to speed up the
2817 		 * processing of the filesystem syncer process. A rushjob
2818 		 * value of N tells the filesystem syncer to process the next
2819 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2820 		 * is used by the soft update code to speed up the filesystem
2821 		 * syncer process when the incore state is getting so far
2822 		 * ahead of the disk that the kernel memory pool is being
2823 		 * threatened with exhaustion.
2824 		 */
2825 		if (rushjob > 0) {
2826 			rushjob -= 1;
2827 			continue;
2828 		}
2829 		/*
2830 		 * Just sleep for a short period of time between
2831 		 * iterations when shutting down to allow some I/O
2832 		 * to happen.
2833 		 *
2834 		 * If it has taken us less than a second to process the
2835 		 * current work, then wait. Otherwise start right over
2836 		 * again. We can still lose time if any single round
2837 		 * takes more than two seconds, but it does not really
2838 		 * matter as we are just trying to generally pace the
2839 		 * filesystem activity.
2840 		 */
2841 		if (syncer_state != SYNCER_RUNNING ||
2842 		    time_uptime == starttime) {
2843 			thread_lock(td);
2844 			sched_prio(td, PPAUSE);
2845 			thread_unlock(td);
2846 		}
2847 		if (syncer_state != SYNCER_RUNNING)
2848 			cv_timedwait(&sync_wakeup, &sync_mtx,
2849 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2850 		else if (time_uptime == starttime)
2851 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2852 	}
2853 }
2854 
2855 /*
2856  * Request the syncer daemon to speed up its work.
2857  * We never push it to speed up more than half of its
2858  * normal turn time, otherwise it could take over the cpu.
2859  */
2860 int
2861 speedup_syncer(void)
2862 {
2863 	int ret = 0;
2864 
2865 	mtx_lock(&sync_mtx);
2866 	if (rushjob < syncdelay / 2) {
2867 		rushjob += 1;
2868 		stat_rush_requests += 1;
2869 		ret = 1;
2870 	}
2871 	mtx_unlock(&sync_mtx);
2872 	cv_broadcast(&sync_wakeup);
2873 	return (ret);
2874 }
2875 
2876 /*
2877  * Tell the syncer to speed up its work and run though its work
2878  * list several times, then tell it to shut down.
2879  */
2880 static void
2881 syncer_shutdown(void *arg, int howto)
2882 {
2883 
2884 	if (howto & RB_NOSYNC)
2885 		return;
2886 	mtx_lock(&sync_mtx);
2887 	syncer_state = SYNCER_SHUTTING_DOWN;
2888 	rushjob = 0;
2889 	mtx_unlock(&sync_mtx);
2890 	cv_broadcast(&sync_wakeup);
2891 	kproc_shutdown(arg, howto);
2892 }
2893 
2894 void
2895 syncer_suspend(void)
2896 {
2897 
2898 	syncer_shutdown(updateproc, 0);
2899 }
2900 
2901 void
2902 syncer_resume(void)
2903 {
2904 
2905 	mtx_lock(&sync_mtx);
2906 	first_printf = 1;
2907 	syncer_state = SYNCER_RUNNING;
2908 	mtx_unlock(&sync_mtx);
2909 	cv_broadcast(&sync_wakeup);
2910 	kproc_resume(updateproc);
2911 }
2912 
2913 /*
2914  * Move the buffer between the clean and dirty lists of its vnode.
2915  */
2916 void
2917 reassignbuf(struct buf *bp)
2918 {
2919 	struct vnode *vp;
2920 	struct bufobj *bo;
2921 	int delay;
2922 #ifdef INVARIANTS
2923 	struct bufv *bv;
2924 #endif
2925 
2926 	vp = bp->b_vp;
2927 	bo = bp->b_bufobj;
2928 
2929 	KASSERT((bp->b_flags & B_PAGING) == 0,
2930 	    ("%s: cannot reassign paging buffer %p", __func__, bp));
2931 
2932 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2933 	    bp, bp->b_vp, bp->b_flags);
2934 
2935 	BO_LOCK(bo);
2936 	buf_vlist_remove(bp);
2937 
2938 	/*
2939 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2940 	 * of clean buffers.
2941 	 */
2942 	if (bp->b_flags & B_DELWRI) {
2943 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2944 			switch (vp->v_type) {
2945 			case VDIR:
2946 				delay = dirdelay;
2947 				break;
2948 			case VCHR:
2949 				delay = metadelay;
2950 				break;
2951 			default:
2952 				delay = filedelay;
2953 			}
2954 			vn_syncer_add_to_worklist(bo, delay);
2955 		}
2956 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2957 	} else {
2958 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2959 
2960 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2961 			mtx_lock(&sync_mtx);
2962 			LIST_REMOVE(bo, bo_synclist);
2963 			syncer_worklist_len--;
2964 			mtx_unlock(&sync_mtx);
2965 			bo->bo_flag &= ~BO_ONWORKLST;
2966 		}
2967 	}
2968 #ifdef INVARIANTS
2969 	bv = &bo->bo_clean;
2970 	bp = TAILQ_FIRST(&bv->bv_hd);
2971 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2972 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2973 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2974 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2975 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2976 	bv = &bo->bo_dirty;
2977 	bp = TAILQ_FIRST(&bv->bv_hd);
2978 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2979 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2980 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2981 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2982 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2983 #endif
2984 	BO_UNLOCK(bo);
2985 }
2986 
2987 static void
2988 v_init_counters(struct vnode *vp)
2989 {
2990 
2991 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2992 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2993 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2994 
2995 	refcount_init(&vp->v_holdcnt, 1);
2996 	refcount_init(&vp->v_usecount, 1);
2997 }
2998 
2999 /*
3000  * Grab a particular vnode from the free list, increment its
3001  * reference count and lock it.  VIRF_DOOMED is set if the vnode
3002  * is being destroyed.  Only callers who specify LK_RETRY will
3003  * see doomed vnodes.  If inactive processing was delayed in
3004  * vput try to do it here.
3005  *
3006  * usecount is manipulated using atomics without holding any locks.
3007  *
3008  * holdcnt can be manipulated using atomics without holding any locks,
3009  * except when transitioning 1<->0, in which case the interlock is held.
3010  *
3011  * Consumers which don't guarantee liveness of the vnode can use SMR to
3012  * try to get a reference. Note this operation can fail since the vnode
3013  * may be awaiting getting freed by the time they get to it.
3014  */
3015 enum vgetstate
3016 vget_prep_smr(struct vnode *vp)
3017 {
3018 	enum vgetstate vs;
3019 
3020 	VFS_SMR_ASSERT_ENTERED();
3021 
3022 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
3023 		vs = VGET_USECOUNT;
3024 	} else {
3025 		if (vhold_smr(vp))
3026 			vs = VGET_HOLDCNT;
3027 		else
3028 			vs = VGET_NONE;
3029 	}
3030 	return (vs);
3031 }
3032 
3033 enum vgetstate
3034 vget_prep(struct vnode *vp)
3035 {
3036 	enum vgetstate vs;
3037 
3038 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
3039 		vs = VGET_USECOUNT;
3040 	} else {
3041 		vhold(vp);
3042 		vs = VGET_HOLDCNT;
3043 	}
3044 	return (vs);
3045 }
3046 
3047 void
3048 vget_abort(struct vnode *vp, enum vgetstate vs)
3049 {
3050 
3051 	switch (vs) {
3052 	case VGET_USECOUNT:
3053 		vrele(vp);
3054 		break;
3055 	case VGET_HOLDCNT:
3056 		vdrop(vp);
3057 		break;
3058 	default:
3059 		__assert_unreachable();
3060 	}
3061 }
3062 
3063 int
3064 vget(struct vnode *vp, int flags)
3065 {
3066 	enum vgetstate vs;
3067 
3068 	vs = vget_prep(vp);
3069 	return (vget_finish(vp, flags, vs));
3070 }
3071 
3072 int
3073 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
3074 {
3075 	int error;
3076 
3077 	if ((flags & LK_INTERLOCK) != 0)
3078 		ASSERT_VI_LOCKED(vp, __func__);
3079 	else
3080 		ASSERT_VI_UNLOCKED(vp, __func__);
3081 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3082 	VNPASS(vp->v_holdcnt > 0, vp);
3083 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3084 
3085 	error = vn_lock(vp, flags);
3086 	if (__predict_false(error != 0)) {
3087 		vget_abort(vp, vs);
3088 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
3089 		    vp);
3090 		return (error);
3091 	}
3092 
3093 	vget_finish_ref(vp, vs);
3094 	return (0);
3095 }
3096 
3097 void
3098 vget_finish_ref(struct vnode *vp, enum vgetstate vs)
3099 {
3100 	int old;
3101 
3102 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3103 	VNPASS(vp->v_holdcnt > 0, vp);
3104 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3105 
3106 	if (vs == VGET_USECOUNT)
3107 		return;
3108 
3109 	/*
3110 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
3111 	 * the vnode around. Otherwise someone else lended their hold count and
3112 	 * we have to drop ours.
3113 	 */
3114 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
3115 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3116 	if (old != 0) {
3117 #ifdef INVARIANTS
3118 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3119 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3120 #else
3121 		refcount_release(&vp->v_holdcnt);
3122 #endif
3123 	}
3124 }
3125 
3126 void
3127 vref(struct vnode *vp)
3128 {
3129 	enum vgetstate vs;
3130 
3131 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3132 	vs = vget_prep(vp);
3133 	vget_finish_ref(vp, vs);
3134 }
3135 
3136 void
3137 vrefact(struct vnode *vp)
3138 {
3139 
3140 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3141 #ifdef INVARIANTS
3142 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3143 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3144 #else
3145 	refcount_acquire(&vp->v_usecount);
3146 #endif
3147 }
3148 
3149 void
3150 vlazy(struct vnode *vp)
3151 {
3152 	struct mount *mp;
3153 
3154 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3155 
3156 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3157 		return;
3158 	/*
3159 	 * We may get here for inactive routines after the vnode got doomed.
3160 	 */
3161 	if (VN_IS_DOOMED(vp))
3162 		return;
3163 	mp = vp->v_mount;
3164 	mtx_lock(&mp->mnt_listmtx);
3165 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3166 		vp->v_mflag |= VMP_LAZYLIST;
3167 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3168 		mp->mnt_lazyvnodelistsize++;
3169 	}
3170 	mtx_unlock(&mp->mnt_listmtx);
3171 }
3172 
3173 static void
3174 vunlazy(struct vnode *vp)
3175 {
3176 	struct mount *mp;
3177 
3178 	ASSERT_VI_LOCKED(vp, __func__);
3179 	VNPASS(!VN_IS_DOOMED(vp), vp);
3180 
3181 	mp = vp->v_mount;
3182 	mtx_lock(&mp->mnt_listmtx);
3183 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3184 	/*
3185 	 * Don't remove the vnode from the lazy list if another thread
3186 	 * has increased the hold count. It may have re-enqueued the
3187 	 * vnode to the lazy list and is now responsible for its
3188 	 * removal.
3189 	 */
3190 	if (vp->v_holdcnt == 0) {
3191 		vp->v_mflag &= ~VMP_LAZYLIST;
3192 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3193 		mp->mnt_lazyvnodelistsize--;
3194 	}
3195 	mtx_unlock(&mp->mnt_listmtx);
3196 }
3197 
3198 /*
3199  * This routine is only meant to be called from vgonel prior to dooming
3200  * the vnode.
3201  */
3202 static void
3203 vunlazy_gone(struct vnode *vp)
3204 {
3205 	struct mount *mp;
3206 
3207 	ASSERT_VOP_ELOCKED(vp, __func__);
3208 	ASSERT_VI_LOCKED(vp, __func__);
3209 	VNPASS(!VN_IS_DOOMED(vp), vp);
3210 
3211 	if (vp->v_mflag & VMP_LAZYLIST) {
3212 		mp = vp->v_mount;
3213 		mtx_lock(&mp->mnt_listmtx);
3214 		VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3215 		vp->v_mflag &= ~VMP_LAZYLIST;
3216 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3217 		mp->mnt_lazyvnodelistsize--;
3218 		mtx_unlock(&mp->mnt_listmtx);
3219 	}
3220 }
3221 
3222 static void
3223 vdefer_inactive(struct vnode *vp)
3224 {
3225 
3226 	ASSERT_VI_LOCKED(vp, __func__);
3227 	VNPASS(vp->v_holdcnt > 0, vp);
3228 	if (VN_IS_DOOMED(vp)) {
3229 		vdropl(vp);
3230 		return;
3231 	}
3232 	if (vp->v_iflag & VI_DEFINACT) {
3233 		VNPASS(vp->v_holdcnt > 1, vp);
3234 		vdropl(vp);
3235 		return;
3236 	}
3237 	if (vp->v_usecount > 0) {
3238 		vp->v_iflag &= ~VI_OWEINACT;
3239 		vdropl(vp);
3240 		return;
3241 	}
3242 	vlazy(vp);
3243 	vp->v_iflag |= VI_DEFINACT;
3244 	VI_UNLOCK(vp);
3245 	atomic_add_long(&deferred_inact, 1);
3246 }
3247 
3248 static void
3249 vdefer_inactive_unlocked(struct vnode *vp)
3250 {
3251 
3252 	VI_LOCK(vp);
3253 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3254 		vdropl(vp);
3255 		return;
3256 	}
3257 	vdefer_inactive(vp);
3258 }
3259 
3260 enum vput_op { VRELE, VPUT, VUNREF };
3261 
3262 /*
3263  * Handle ->v_usecount transitioning to 0.
3264  *
3265  * By releasing the last usecount we take ownership of the hold count which
3266  * provides liveness of the vnode, meaning we have to vdrop.
3267  *
3268  * For all vnodes we may need to perform inactive processing. It requires an
3269  * exclusive lock on the vnode, while it is legal to call here with only a
3270  * shared lock (or no locks). If locking the vnode in an expected manner fails,
3271  * inactive processing gets deferred to the syncer.
3272  *
3273  * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3274  * on the lock being held all the way until VOP_INACTIVE. This in particular
3275  * happens with UFS which adds half-constructed vnodes to the hash, where they
3276  * can be found by other code.
3277  */
3278 static void
3279 vput_final(struct vnode *vp, enum vput_op func)
3280 {
3281 	int error;
3282 	bool want_unlock;
3283 
3284 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3285 	VNPASS(vp->v_holdcnt > 0, vp);
3286 
3287 	VI_LOCK(vp);
3288 
3289 	/*
3290 	 * By the time we got here someone else might have transitioned
3291 	 * the count back to > 0.
3292 	 */
3293 	if (vp->v_usecount > 0)
3294 		goto out;
3295 
3296 	/*
3297 	 * If the vnode is doomed vgone already performed inactive processing
3298 	 * (if needed).
3299 	 */
3300 	if (VN_IS_DOOMED(vp))
3301 		goto out;
3302 
3303 	if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3304 		goto out;
3305 
3306 	if (vp->v_iflag & VI_DOINGINACT)
3307 		goto out;
3308 
3309 	/*
3310 	 * Locking operations here will drop the interlock and possibly the
3311 	 * vnode lock, opening a window where the vnode can get doomed all the
3312 	 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3313 	 * perform inactive.
3314 	 */
3315 	vp->v_iflag |= VI_OWEINACT;
3316 	want_unlock = false;
3317 	error = 0;
3318 	switch (func) {
3319 	case VRELE:
3320 		switch (VOP_ISLOCKED(vp)) {
3321 		case LK_EXCLUSIVE:
3322 			break;
3323 		case LK_EXCLOTHER:
3324 		case 0:
3325 			want_unlock = true;
3326 			error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3327 			VI_LOCK(vp);
3328 			break;
3329 		default:
3330 			/*
3331 			 * The lock has at least one sharer, but we have no way
3332 			 * to conclude whether this is us. Play it safe and
3333 			 * defer processing.
3334 			 */
3335 			error = EAGAIN;
3336 			break;
3337 		}
3338 		break;
3339 	case VPUT:
3340 		want_unlock = true;
3341 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3342 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3343 			    LK_NOWAIT);
3344 			VI_LOCK(vp);
3345 		}
3346 		break;
3347 	case VUNREF:
3348 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3349 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3350 			VI_LOCK(vp);
3351 		}
3352 		break;
3353 	}
3354 	if (error == 0) {
3355 		if (func == VUNREF) {
3356 			VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp,
3357 			    ("recursive vunref"));
3358 			vp->v_vflag |= VV_UNREF;
3359 		}
3360 		for (;;) {
3361 			error = vinactive(vp);
3362 			if (want_unlock)
3363 				VOP_UNLOCK(vp);
3364 			if (error != ERELOOKUP || !want_unlock)
3365 				break;
3366 			VOP_LOCK(vp, LK_EXCLUSIVE);
3367 		}
3368 		if (func == VUNREF)
3369 			vp->v_vflag &= ~VV_UNREF;
3370 		vdropl(vp);
3371 	} else {
3372 		vdefer_inactive(vp);
3373 	}
3374 	return;
3375 out:
3376 	if (func == VPUT)
3377 		VOP_UNLOCK(vp);
3378 	vdropl(vp);
3379 }
3380 
3381 /*
3382  * Decrement ->v_usecount for a vnode.
3383  *
3384  * Releasing the last use count requires additional processing, see vput_final
3385  * above for details.
3386  *
3387  * Comment above each variant denotes lock state on entry and exit.
3388  */
3389 
3390 /*
3391  * in: any
3392  * out: same as passed in
3393  */
3394 void
3395 vrele(struct vnode *vp)
3396 {
3397 
3398 	ASSERT_VI_UNLOCKED(vp, __func__);
3399 	if (!refcount_release(&vp->v_usecount))
3400 		return;
3401 	vput_final(vp, VRELE);
3402 }
3403 
3404 /*
3405  * in: locked
3406  * out: unlocked
3407  */
3408 void
3409 vput(struct vnode *vp)
3410 {
3411 
3412 	ASSERT_VOP_LOCKED(vp, __func__);
3413 	ASSERT_VI_UNLOCKED(vp, __func__);
3414 	if (!refcount_release(&vp->v_usecount)) {
3415 		VOP_UNLOCK(vp);
3416 		return;
3417 	}
3418 	vput_final(vp, VPUT);
3419 }
3420 
3421 /*
3422  * in: locked
3423  * out: locked
3424  */
3425 void
3426 vunref(struct vnode *vp)
3427 {
3428 
3429 	ASSERT_VOP_LOCKED(vp, __func__);
3430 	ASSERT_VI_UNLOCKED(vp, __func__);
3431 	if (!refcount_release(&vp->v_usecount))
3432 		return;
3433 	vput_final(vp, VUNREF);
3434 }
3435 
3436 void
3437 vhold(struct vnode *vp)
3438 {
3439 	int old;
3440 
3441 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3442 	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3443 	VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3444 	    ("%s: wrong hold count %d", __func__, old));
3445 	if (old == 0)
3446 		vfs_freevnodes_dec();
3447 }
3448 
3449 void
3450 vholdnz(struct vnode *vp)
3451 {
3452 
3453 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3454 #ifdef INVARIANTS
3455 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3456 	VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3457 	    ("%s: wrong hold count %d", __func__, old));
3458 #else
3459 	atomic_add_int(&vp->v_holdcnt, 1);
3460 #endif
3461 }
3462 
3463 /*
3464  * Grab a hold count unless the vnode is freed.
3465  *
3466  * Only use this routine if vfs smr is the only protection you have against
3467  * freeing the vnode.
3468  *
3469  * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag
3470  * is not set.  After the flag is set the vnode becomes immutable to anyone but
3471  * the thread which managed to set the flag.
3472  *
3473  * It may be tempting to replace the loop with:
3474  * count = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3475  * if (count & VHOLD_NO_SMR) {
3476  *     backpedal and error out;
3477  * }
3478  *
3479  * However, while this is more performant, it hinders debugging by eliminating
3480  * the previously mentioned invariant.
3481  */
3482 bool
3483 vhold_smr(struct vnode *vp)
3484 {
3485 	int count;
3486 
3487 	VFS_SMR_ASSERT_ENTERED();
3488 
3489 	count = atomic_load_int(&vp->v_holdcnt);
3490 	for (;;) {
3491 		if (count & VHOLD_NO_SMR) {
3492 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3493 			    ("non-zero hold count with flags %d\n", count));
3494 			return (false);
3495 		}
3496 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3497 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3498 			if (count == 0)
3499 				vfs_freevnodes_dec();
3500 			return (true);
3501 		}
3502 	}
3503 }
3504 
3505 /*
3506  * Hold a free vnode for recycling.
3507  *
3508  * Note: vnode_init references this comment.
3509  *
3510  * Attempts to recycle only need the global vnode list lock and have no use for
3511  * SMR.
3512  *
3513  * However, vnodes get inserted into the global list before they get fully
3514  * initialized and stay there until UMA decides to free the memory. This in
3515  * particular means the target can be found before it becomes usable and after
3516  * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to
3517  * VHOLD_NO_SMR.
3518  *
3519  * Note: the vnode may gain more references after we transition the count 0->1.
3520  */
3521 static bool
3522 vhold_recycle_free(struct vnode *vp)
3523 {
3524 	int count;
3525 
3526 	mtx_assert(&vnode_list_mtx, MA_OWNED);
3527 
3528 	count = atomic_load_int(&vp->v_holdcnt);
3529 	for (;;) {
3530 		if (count & VHOLD_NO_SMR) {
3531 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3532 			    ("non-zero hold count with flags %d\n", count));
3533 			return (false);
3534 		}
3535 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3536 		if (count > 0) {
3537 			return (false);
3538 		}
3539 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3540 			vfs_freevnodes_dec();
3541 			return (true);
3542 		}
3543 	}
3544 }
3545 
3546 static void __noinline
3547 vdbatch_process(struct vdbatch *vd)
3548 {
3549 	struct vnode *vp;
3550 	int i;
3551 
3552 	mtx_assert(&vd->lock, MA_OWNED);
3553 	MPASS(curthread->td_pinned > 0);
3554 	MPASS(vd->index == VDBATCH_SIZE);
3555 
3556 	/*
3557 	 * Attempt to requeue the passed batch, but give up easily.
3558 	 *
3559 	 * Despite batching the mechanism is prone to transient *significant*
3560 	 * lock contention, where vnode_list_mtx becomes the primary bottleneck
3561 	 * if multiple CPUs get here (one real-world example is highly parallel
3562 	 * do-nothing make , which will stat *tons* of vnodes). Since it is
3563 	 * quasi-LRU (read: not that great even if fully honoured) just dodge
3564 	 * the problem. Parties which don't like it are welcome to implement
3565 	 * something better.
3566 	 */
3567 	critical_enter();
3568 	if (mtx_trylock(&vnode_list_mtx)) {
3569 		for (i = 0; i < VDBATCH_SIZE; i++) {
3570 			vp = vd->tab[i];
3571 			vd->tab[i] = NULL;
3572 			TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3573 			TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3574 			MPASS(vp->v_dbatchcpu != NOCPU);
3575 			vp->v_dbatchcpu = NOCPU;
3576 		}
3577 		mtx_unlock(&vnode_list_mtx);
3578 	} else {
3579 		counter_u64_add(vnode_skipped_requeues, 1);
3580 
3581 		for (i = 0; i < VDBATCH_SIZE; i++) {
3582 			vp = vd->tab[i];
3583 			vd->tab[i] = NULL;
3584 			MPASS(vp->v_dbatchcpu != NOCPU);
3585 			vp->v_dbatchcpu = NOCPU;
3586 		}
3587 	}
3588 	vd->index = 0;
3589 	critical_exit();
3590 }
3591 
3592 static void
3593 vdbatch_enqueue(struct vnode *vp)
3594 {
3595 	struct vdbatch *vd;
3596 
3597 	ASSERT_VI_LOCKED(vp, __func__);
3598 	VNPASS(!VN_IS_DOOMED(vp), vp);
3599 
3600 	if (vp->v_dbatchcpu != NOCPU) {
3601 		VI_UNLOCK(vp);
3602 		return;
3603 	}
3604 
3605 	sched_pin();
3606 	vd = DPCPU_PTR(vd);
3607 	mtx_lock(&vd->lock);
3608 	MPASS(vd->index < VDBATCH_SIZE);
3609 	MPASS(vd->tab[vd->index] == NULL);
3610 	/*
3611 	 * A hack: we depend on being pinned so that we know what to put in
3612 	 * ->v_dbatchcpu.
3613 	 */
3614 	vp->v_dbatchcpu = curcpu;
3615 	vd->tab[vd->index] = vp;
3616 	vd->index++;
3617 	VI_UNLOCK(vp);
3618 	if (vd->index == VDBATCH_SIZE)
3619 		vdbatch_process(vd);
3620 	mtx_unlock(&vd->lock);
3621 	sched_unpin();
3622 }
3623 
3624 /*
3625  * This routine must only be called for vnodes which are about to be
3626  * deallocated. Supporting dequeue for arbitrary vndoes would require
3627  * validating that the locked batch matches.
3628  */
3629 static void
3630 vdbatch_dequeue(struct vnode *vp)
3631 {
3632 	struct vdbatch *vd;
3633 	int i;
3634 	short cpu;
3635 
3636 	VNPASS(vp->v_type == VBAD || vp->v_type == VNON, vp);
3637 
3638 	cpu = vp->v_dbatchcpu;
3639 	if (cpu == NOCPU)
3640 		return;
3641 
3642 	vd = DPCPU_ID_PTR(cpu, vd);
3643 	mtx_lock(&vd->lock);
3644 	for (i = 0; i < vd->index; i++) {
3645 		if (vd->tab[i] != vp)
3646 			continue;
3647 		vp->v_dbatchcpu = NOCPU;
3648 		vd->index--;
3649 		vd->tab[i] = vd->tab[vd->index];
3650 		vd->tab[vd->index] = NULL;
3651 		break;
3652 	}
3653 	mtx_unlock(&vd->lock);
3654 	/*
3655 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3656 	 */
3657 	MPASS(vp->v_dbatchcpu == NOCPU);
3658 }
3659 
3660 /*
3661  * Drop the hold count of the vnode.  If this is the last reference to
3662  * the vnode we place it on the free list unless it has been vgone'd
3663  * (marked VIRF_DOOMED) in which case we will free it.
3664  *
3665  * Because the vnode vm object keeps a hold reference on the vnode if
3666  * there is at least one resident non-cached page, the vnode cannot
3667  * leave the active list without the page cleanup done.
3668  */
3669 static void __noinline
3670 vdropl_final(struct vnode *vp)
3671 {
3672 
3673 	ASSERT_VI_LOCKED(vp, __func__);
3674 	VNPASS(VN_IS_DOOMED(vp), vp);
3675 	/*
3676 	 * Set the VHOLD_NO_SMR flag.
3677 	 *
3678 	 * We may be racing against vhold_smr. If they win we can just pretend
3679 	 * we never got this far, they will vdrop later.
3680 	 */
3681 	if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) {
3682 		vfs_freevnodes_inc();
3683 		VI_UNLOCK(vp);
3684 		/*
3685 		 * We lost the aforementioned race. Any subsequent access is
3686 		 * invalid as they might have managed to vdropl on their own.
3687 		 */
3688 		return;
3689 	}
3690 	/*
3691 	 * Don't bump freevnodes as this one is going away.
3692 	 */
3693 	freevnode(vp);
3694 }
3695 
3696 void
3697 vdrop(struct vnode *vp)
3698 {
3699 
3700 	ASSERT_VI_UNLOCKED(vp, __func__);
3701 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3702 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3703 		return;
3704 	VI_LOCK(vp);
3705 	vdropl(vp);
3706 }
3707 
3708 static void __always_inline
3709 vdropl_impl(struct vnode *vp, bool enqueue)
3710 {
3711 
3712 	ASSERT_VI_LOCKED(vp, __func__);
3713 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3714 	if (!refcount_release(&vp->v_holdcnt)) {
3715 		VI_UNLOCK(vp);
3716 		return;
3717 	}
3718 	VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp);
3719 	VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
3720 	if (VN_IS_DOOMED(vp)) {
3721 		vdropl_final(vp);
3722 		return;
3723 	}
3724 
3725 	vfs_freevnodes_inc();
3726 	if (vp->v_mflag & VMP_LAZYLIST) {
3727 		vunlazy(vp);
3728 	}
3729 
3730 	if (!enqueue) {
3731 		VI_UNLOCK(vp);
3732 		return;
3733 	}
3734 
3735 	/*
3736 	 * Also unlocks the interlock. We can't assert on it as we
3737 	 * released our hold and by now the vnode might have been
3738 	 * freed.
3739 	 */
3740 	vdbatch_enqueue(vp);
3741 }
3742 
3743 void
3744 vdropl(struct vnode *vp)
3745 {
3746 
3747 	vdropl_impl(vp, true);
3748 }
3749 
3750 /*
3751  * vdrop a vnode when recycling
3752  *
3753  * This is a special case routine only to be used when recycling, differs from
3754  * regular vdrop by not requeieing the vnode on LRU.
3755  *
3756  * Consider a case where vtryrecycle continuously fails with all vnodes (due to
3757  * e.g., frozen writes on the filesystem), filling the batch and causing it to
3758  * be requeued. Then vnlru will end up revisiting the same vnodes. This is a
3759  * loop which can last for as long as writes are frozen.
3760  */
3761 static void
3762 vdropl_recycle(struct vnode *vp)
3763 {
3764 
3765 	vdropl_impl(vp, false);
3766 }
3767 
3768 static void
3769 vdrop_recycle(struct vnode *vp)
3770 {
3771 
3772 	VI_LOCK(vp);
3773 	vdropl_recycle(vp);
3774 }
3775 
3776 /*
3777  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3778  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3779  */
3780 static int
3781 vinactivef(struct vnode *vp)
3782 {
3783 	struct vm_object *obj;
3784 	int error;
3785 
3786 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3787 	ASSERT_VI_LOCKED(vp, "vinactive");
3788 	VNPASS((vp->v_iflag & VI_DOINGINACT) == 0, vp);
3789 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3790 	vp->v_iflag |= VI_DOINGINACT;
3791 	vp->v_iflag &= ~VI_OWEINACT;
3792 	VI_UNLOCK(vp);
3793 	/*
3794 	 * Before moving off the active list, we must be sure that any
3795 	 * modified pages are converted into the vnode's dirty
3796 	 * buffers, since these will no longer be checked once the
3797 	 * vnode is on the inactive list.
3798 	 *
3799 	 * The write-out of the dirty pages is asynchronous.  At the
3800 	 * point that VOP_INACTIVE() is called, there could still be
3801 	 * pending I/O and dirty pages in the object.
3802 	 */
3803 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3804 	    vm_object_mightbedirty(obj)) {
3805 		VM_OBJECT_WLOCK(obj);
3806 		vm_object_page_clean(obj, 0, 0, 0);
3807 		VM_OBJECT_WUNLOCK(obj);
3808 	}
3809 	error = VOP_INACTIVE(vp);
3810 	VI_LOCK(vp);
3811 	VNPASS(vp->v_iflag & VI_DOINGINACT, vp);
3812 	vp->v_iflag &= ~VI_DOINGINACT;
3813 	return (error);
3814 }
3815 
3816 int
3817 vinactive(struct vnode *vp)
3818 {
3819 
3820 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3821 	ASSERT_VI_LOCKED(vp, "vinactive");
3822 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3823 
3824 	if ((vp->v_iflag & VI_OWEINACT) == 0)
3825 		return (0);
3826 	if (vp->v_iflag & VI_DOINGINACT)
3827 		return (0);
3828 	if (vp->v_usecount > 0) {
3829 		vp->v_iflag &= ~VI_OWEINACT;
3830 		return (0);
3831 	}
3832 	return (vinactivef(vp));
3833 }
3834 
3835 /*
3836  * Remove any vnodes in the vnode table belonging to mount point mp.
3837  *
3838  * If FORCECLOSE is not specified, there should not be any active ones,
3839  * return error if any are found (nb: this is a user error, not a
3840  * system error). If FORCECLOSE is specified, detach any active vnodes
3841  * that are found.
3842  *
3843  * If WRITECLOSE is set, only flush out regular file vnodes open for
3844  * writing.
3845  *
3846  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3847  *
3848  * `rootrefs' specifies the base reference count for the root vnode
3849  * of this filesystem. The root vnode is considered busy if its
3850  * v_usecount exceeds this value. On a successful return, vflush(, td)
3851  * will call vrele() on the root vnode exactly rootrefs times.
3852  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3853  * be zero.
3854  */
3855 #ifdef DIAGNOSTIC
3856 static int busyprt = 0;		/* print out busy vnodes */
3857 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3858 #endif
3859 
3860 int
3861 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3862 {
3863 	struct vnode *vp, *mvp, *rootvp = NULL;
3864 	struct vattr vattr;
3865 	int busy = 0, error;
3866 
3867 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3868 	    rootrefs, flags);
3869 	if (rootrefs > 0) {
3870 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3871 		    ("vflush: bad args"));
3872 		/*
3873 		 * Get the filesystem root vnode. We can vput() it
3874 		 * immediately, since with rootrefs > 0, it won't go away.
3875 		 */
3876 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3877 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3878 			    __func__, error);
3879 			return (error);
3880 		}
3881 		vput(rootvp);
3882 	}
3883 loop:
3884 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3885 		vholdl(vp);
3886 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3887 		if (error) {
3888 			vdrop(vp);
3889 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3890 			goto loop;
3891 		}
3892 		/*
3893 		 * Skip over a vnodes marked VV_SYSTEM.
3894 		 */
3895 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3896 			VOP_UNLOCK(vp);
3897 			vdrop(vp);
3898 			continue;
3899 		}
3900 		/*
3901 		 * If WRITECLOSE is set, flush out unlinked but still open
3902 		 * files (even if open only for reading) and regular file
3903 		 * vnodes open for writing.
3904 		 */
3905 		if (flags & WRITECLOSE) {
3906 			if (vp->v_object != NULL) {
3907 				VM_OBJECT_WLOCK(vp->v_object);
3908 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3909 				VM_OBJECT_WUNLOCK(vp->v_object);
3910 			}
3911 			do {
3912 				error = VOP_FSYNC(vp, MNT_WAIT, td);
3913 			} while (error == ERELOOKUP);
3914 			if (error != 0) {
3915 				VOP_UNLOCK(vp);
3916 				vdrop(vp);
3917 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3918 				return (error);
3919 			}
3920 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3921 			VI_LOCK(vp);
3922 
3923 			if ((vp->v_type == VNON ||
3924 			    (error == 0 && vattr.va_nlink > 0)) &&
3925 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3926 				VOP_UNLOCK(vp);
3927 				vdropl(vp);
3928 				continue;
3929 			}
3930 		} else
3931 			VI_LOCK(vp);
3932 		/*
3933 		 * With v_usecount == 0, all we need to do is clear out the
3934 		 * vnode data structures and we are done.
3935 		 *
3936 		 * If FORCECLOSE is set, forcibly close the vnode.
3937 		 */
3938 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3939 			vgonel(vp);
3940 		} else {
3941 			busy++;
3942 #ifdef DIAGNOSTIC
3943 			if (busyprt)
3944 				vn_printf(vp, "vflush: busy vnode ");
3945 #endif
3946 		}
3947 		VOP_UNLOCK(vp);
3948 		vdropl(vp);
3949 	}
3950 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3951 		/*
3952 		 * If just the root vnode is busy, and if its refcount
3953 		 * is equal to `rootrefs', then go ahead and kill it.
3954 		 */
3955 		VI_LOCK(rootvp);
3956 		KASSERT(busy > 0, ("vflush: not busy"));
3957 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3958 		    ("vflush: usecount %d < rootrefs %d",
3959 		     rootvp->v_usecount, rootrefs));
3960 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3961 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3962 			vgone(rootvp);
3963 			VOP_UNLOCK(rootvp);
3964 			busy = 0;
3965 		} else
3966 			VI_UNLOCK(rootvp);
3967 	}
3968 	if (busy) {
3969 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3970 		    busy);
3971 		return (EBUSY);
3972 	}
3973 	for (; rootrefs > 0; rootrefs--)
3974 		vrele(rootvp);
3975 	return (0);
3976 }
3977 
3978 /*
3979  * Recycle an unused vnode to the front of the free list.
3980  */
3981 int
3982 vrecycle(struct vnode *vp)
3983 {
3984 	int recycled;
3985 
3986 	VI_LOCK(vp);
3987 	recycled = vrecyclel(vp);
3988 	VI_UNLOCK(vp);
3989 	return (recycled);
3990 }
3991 
3992 /*
3993  * vrecycle, with the vp interlock held.
3994  */
3995 int
3996 vrecyclel(struct vnode *vp)
3997 {
3998 	int recycled;
3999 
4000 	ASSERT_VOP_ELOCKED(vp, __func__);
4001 	ASSERT_VI_LOCKED(vp, __func__);
4002 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4003 	recycled = 0;
4004 	if (vp->v_usecount == 0) {
4005 		recycled = 1;
4006 		vgonel(vp);
4007 	}
4008 	return (recycled);
4009 }
4010 
4011 /*
4012  * Eliminate all activity associated with a vnode
4013  * in preparation for reuse.
4014  */
4015 void
4016 vgone(struct vnode *vp)
4017 {
4018 	VI_LOCK(vp);
4019 	vgonel(vp);
4020 	VI_UNLOCK(vp);
4021 }
4022 
4023 /*
4024  * Notify upper mounts about reclaimed or unlinked vnode.
4025  */
4026 void
4027 vfs_notify_upper(struct vnode *vp, enum vfs_notify_upper_type event)
4028 {
4029 	struct mount *mp;
4030 	struct mount_upper_node *ump;
4031 
4032 	mp = atomic_load_ptr(&vp->v_mount);
4033 	if (mp == NULL)
4034 		return;
4035 	if (TAILQ_EMPTY(&mp->mnt_notify))
4036 		return;
4037 
4038 	MNT_ILOCK(mp);
4039 	mp->mnt_upper_pending++;
4040 	KASSERT(mp->mnt_upper_pending > 0,
4041 	    ("%s: mnt_upper_pending %d", __func__, mp->mnt_upper_pending));
4042 	TAILQ_FOREACH(ump, &mp->mnt_notify, mnt_upper_link) {
4043 		MNT_IUNLOCK(mp);
4044 		switch (event) {
4045 		case VFS_NOTIFY_UPPER_RECLAIM:
4046 			VFS_RECLAIM_LOWERVP(ump->mp, vp);
4047 			break;
4048 		case VFS_NOTIFY_UPPER_UNLINK:
4049 			VFS_UNLINK_LOWERVP(ump->mp, vp);
4050 			break;
4051 		}
4052 		MNT_ILOCK(mp);
4053 	}
4054 	mp->mnt_upper_pending--;
4055 	if ((mp->mnt_kern_flag & MNTK_UPPER_WAITER) != 0 &&
4056 	    mp->mnt_upper_pending == 0) {
4057 		mp->mnt_kern_flag &= ~MNTK_UPPER_WAITER;
4058 		wakeup(&mp->mnt_uppers);
4059 	}
4060 	MNT_IUNLOCK(mp);
4061 }
4062 
4063 /*
4064  * vgone, with the vp interlock held.
4065  */
4066 static void
4067 vgonel(struct vnode *vp)
4068 {
4069 	struct thread *td;
4070 	struct mount *mp;
4071 	vm_object_t object;
4072 	bool active, doinginact, oweinact;
4073 
4074 	ASSERT_VOP_ELOCKED(vp, "vgonel");
4075 	ASSERT_VI_LOCKED(vp, "vgonel");
4076 	VNASSERT(vp->v_holdcnt, vp,
4077 	    ("vgonel: vp %p has no reference.", vp));
4078 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4079 	td = curthread;
4080 
4081 	/*
4082 	 * Don't vgonel if we're already doomed.
4083 	 */
4084 	if (VN_IS_DOOMED(vp)) {
4085 		VNPASS(vn_get_state(vp) == VSTATE_DESTROYING || \
4086 		    vn_get_state(vp) == VSTATE_DEAD, vp);
4087 		return;
4088 	}
4089 	/*
4090 	 * Paired with freevnode.
4091 	 */
4092 	vn_seqc_write_begin_locked(vp);
4093 	vunlazy_gone(vp);
4094 	vn_irflag_set_locked(vp, VIRF_DOOMED);
4095 	vn_set_state(vp, VSTATE_DESTROYING);
4096 
4097 	/*
4098 	 * Check to see if the vnode is in use.  If so, we have to
4099 	 * call VOP_CLOSE() and VOP_INACTIVE().
4100 	 *
4101 	 * It could be that VOP_INACTIVE() requested reclamation, in
4102 	 * which case we should avoid recursion, so check
4103 	 * VI_DOINGINACT.  This is not precise but good enough.
4104 	 */
4105 	active = vp->v_usecount > 0;
4106 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4107 	doinginact = (vp->v_iflag & VI_DOINGINACT) != 0;
4108 
4109 	/*
4110 	 * If we need to do inactive VI_OWEINACT will be set.
4111 	 */
4112 	if (vp->v_iflag & VI_DEFINACT) {
4113 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
4114 		vp->v_iflag &= ~VI_DEFINACT;
4115 		vdropl(vp);
4116 	} else {
4117 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
4118 		VI_UNLOCK(vp);
4119 	}
4120 	cache_purge_vgone(vp);
4121 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
4122 
4123 	/*
4124 	 * If purging an active vnode, it must be closed and
4125 	 * deactivated before being reclaimed.
4126 	 */
4127 	if (active)
4128 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
4129 	if (!doinginact) {
4130 		do {
4131 			if (oweinact || active) {
4132 				VI_LOCK(vp);
4133 				vinactivef(vp);
4134 				oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4135 				VI_UNLOCK(vp);
4136 			}
4137 		} while (oweinact);
4138 	}
4139 	if (vp->v_type == VSOCK)
4140 		vfs_unp_reclaim(vp);
4141 
4142 	/*
4143 	 * Clean out any buffers associated with the vnode.
4144 	 * If the flush fails, just toss the buffers.
4145 	 */
4146 	mp = NULL;
4147 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
4148 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
4149 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
4150 		while (vinvalbuf(vp, 0, 0, 0) != 0)
4151 			;
4152 	}
4153 
4154 	BO_LOCK(&vp->v_bufobj);
4155 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
4156 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
4157 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
4158 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
4159 	    ("vp %p bufobj not invalidated", vp));
4160 
4161 	/*
4162 	 * For VMIO bufobj, BO_DEAD is set later, or in
4163 	 * vm_object_terminate() after the object's page queue is
4164 	 * flushed.
4165 	 */
4166 	object = vp->v_bufobj.bo_object;
4167 	if (object == NULL)
4168 		vp->v_bufobj.bo_flag |= BO_DEAD;
4169 	BO_UNLOCK(&vp->v_bufobj);
4170 
4171 	/*
4172 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
4173 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
4174 	 * should not touch the object borrowed from the lower vnode
4175 	 * (the handle check).
4176 	 */
4177 	if (object != NULL && object->type == OBJT_VNODE &&
4178 	    object->handle == vp)
4179 		vnode_destroy_vobject(vp);
4180 
4181 	/*
4182 	 * Reclaim the vnode.
4183 	 */
4184 	if (VOP_RECLAIM(vp))
4185 		panic("vgone: cannot reclaim");
4186 	if (mp != NULL)
4187 		vn_finished_secondary_write(mp);
4188 	VNASSERT(vp->v_object == NULL, vp,
4189 	    ("vop_reclaim left v_object vp=%p", vp));
4190 	/*
4191 	 * Clear the advisory locks and wake up waiting threads.
4192 	 */
4193 	if (vp->v_lockf != NULL) {
4194 		(void)VOP_ADVLOCKPURGE(vp);
4195 		vp->v_lockf = NULL;
4196 	}
4197 	/*
4198 	 * Delete from old mount point vnode list.
4199 	 */
4200 	if (vp->v_mount == NULL) {
4201 		VI_LOCK(vp);
4202 	} else {
4203 		delmntque(vp);
4204 		ASSERT_VI_LOCKED(vp, "vgonel 2");
4205 	}
4206 	/*
4207 	 * Done with purge, reset to the standard lock and invalidate
4208 	 * the vnode.
4209 	 */
4210 	vp->v_vnlock = &vp->v_lock;
4211 	vp->v_op = &dead_vnodeops;
4212 	vp->v_type = VBAD;
4213 	vn_set_state(vp, VSTATE_DEAD);
4214 }
4215 
4216 /*
4217  * Print out a description of a vnode.
4218  */
4219 static const char *const vtypename[] = {
4220 	[VNON] = "VNON",
4221 	[VREG] = "VREG",
4222 	[VDIR] = "VDIR",
4223 	[VBLK] = "VBLK",
4224 	[VCHR] = "VCHR",
4225 	[VLNK] = "VLNK",
4226 	[VSOCK] = "VSOCK",
4227 	[VFIFO] = "VFIFO",
4228 	[VBAD] = "VBAD",
4229 	[VMARKER] = "VMARKER",
4230 };
4231 _Static_assert(nitems(vtypename) == VLASTTYPE + 1,
4232     "vnode type name not added to vtypename");
4233 
4234 static const char *const vstatename[] = {
4235 	[VSTATE_UNINITIALIZED] = "VSTATE_UNINITIALIZED",
4236 	[VSTATE_CONSTRUCTED] = "VSTATE_CONSTRUCTED",
4237 	[VSTATE_DESTROYING] = "VSTATE_DESTROYING",
4238 	[VSTATE_DEAD] = "VSTATE_DEAD",
4239 };
4240 _Static_assert(nitems(vstatename) == VLASTSTATE + 1,
4241     "vnode state name not added to vstatename");
4242 
4243 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0,
4244     "new hold count flag not added to vn_printf");
4245 
4246 void
4247 vn_printf(struct vnode *vp, const char *fmt, ...)
4248 {
4249 	va_list ap;
4250 	char buf[256], buf2[16];
4251 	u_long flags;
4252 	u_int holdcnt;
4253 	short irflag;
4254 
4255 	va_start(ap, fmt);
4256 	vprintf(fmt, ap);
4257 	va_end(ap);
4258 	printf("%p: ", (void *)vp);
4259 	printf("type %s state %s op %p\n", vtypename[vp->v_type],
4260 	    vstatename[vp->v_state], vp->v_op);
4261 	holdcnt = atomic_load_int(&vp->v_holdcnt);
4262 	printf("    usecount %d, writecount %d, refcount %d seqc users %d",
4263 	    vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS,
4264 	    vp->v_seqc_users);
4265 	switch (vp->v_type) {
4266 	case VDIR:
4267 		printf(" mountedhere %p\n", vp->v_mountedhere);
4268 		break;
4269 	case VCHR:
4270 		printf(" rdev %p\n", vp->v_rdev);
4271 		break;
4272 	case VSOCK:
4273 		printf(" socket %p\n", vp->v_unpcb);
4274 		break;
4275 	case VFIFO:
4276 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
4277 		break;
4278 	default:
4279 		printf("\n");
4280 		break;
4281 	}
4282 	buf[0] = '\0';
4283 	buf[1] = '\0';
4284 	if (holdcnt & VHOLD_NO_SMR)
4285 		strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf));
4286 	printf("    hold count flags (%s)\n", buf + 1);
4287 
4288 	buf[0] = '\0';
4289 	buf[1] = '\0';
4290 	irflag = vn_irflag_read(vp);
4291 	if (irflag & VIRF_DOOMED)
4292 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4293 	if (irflag & VIRF_PGREAD)
4294 		strlcat(buf, "|VIRF_PGREAD", sizeof(buf));
4295 	if (irflag & VIRF_MOUNTPOINT)
4296 		strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf));
4297 	if (irflag & VIRF_TEXT_REF)
4298 		strlcat(buf, "|VIRF_TEXT_REF", sizeof(buf));
4299 	flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT | VIRF_TEXT_REF);
4300 	if (flags != 0) {
4301 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4302 		strlcat(buf, buf2, sizeof(buf));
4303 	}
4304 	if (vp->v_vflag & VV_ROOT)
4305 		strlcat(buf, "|VV_ROOT", sizeof(buf));
4306 	if (vp->v_vflag & VV_ISTTY)
4307 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
4308 	if (vp->v_vflag & VV_NOSYNC)
4309 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4310 	if (vp->v_vflag & VV_ETERNALDEV)
4311 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4312 	if (vp->v_vflag & VV_CACHEDLABEL)
4313 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4314 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
4315 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4316 	if (vp->v_vflag & VV_COPYONWRITE)
4317 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4318 	if (vp->v_vflag & VV_SYSTEM)
4319 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4320 	if (vp->v_vflag & VV_PROCDEP)
4321 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4322 	if (vp->v_vflag & VV_DELETED)
4323 		strlcat(buf, "|VV_DELETED", sizeof(buf));
4324 	if (vp->v_vflag & VV_MD)
4325 		strlcat(buf, "|VV_MD", sizeof(buf));
4326 	if (vp->v_vflag & VV_FORCEINSMQ)
4327 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4328 	if (vp->v_vflag & VV_READLINK)
4329 		strlcat(buf, "|VV_READLINK", sizeof(buf));
4330 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4331 	    VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM |
4332 	    VV_PROCDEP | VV_DELETED | VV_MD | VV_FORCEINSMQ | VV_READLINK);
4333 	if (flags != 0) {
4334 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4335 		strlcat(buf, buf2, sizeof(buf));
4336 	}
4337 	if (vp->v_iflag & VI_MOUNT)
4338 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
4339 	if (vp->v_iflag & VI_DOINGINACT)
4340 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4341 	if (vp->v_iflag & VI_OWEINACT)
4342 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4343 	if (vp->v_iflag & VI_DEFINACT)
4344 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4345 	if (vp->v_iflag & VI_FOPENING)
4346 		strlcat(buf, "|VI_FOPENING", sizeof(buf));
4347 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOINGINACT |
4348 	    VI_OWEINACT | VI_DEFINACT | VI_FOPENING);
4349 	if (flags != 0) {
4350 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4351 		strlcat(buf, buf2, sizeof(buf));
4352 	}
4353 	if (vp->v_mflag & VMP_LAZYLIST)
4354 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4355 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
4356 	if (flags != 0) {
4357 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4358 		strlcat(buf, buf2, sizeof(buf));
4359 	}
4360 	printf("    flags (%s)", buf + 1);
4361 	if (mtx_owned(VI_MTX(vp)))
4362 		printf(" VI_LOCKed");
4363 	printf("\n");
4364 	if (vp->v_object != NULL)
4365 		printf("    v_object %p ref %d pages %d "
4366 		    "cleanbuf %d dirtybuf %d\n",
4367 		    vp->v_object, vp->v_object->ref_count,
4368 		    vp->v_object->resident_page_count,
4369 		    vp->v_bufobj.bo_clean.bv_cnt,
4370 		    vp->v_bufobj.bo_dirty.bv_cnt);
4371 	printf("    ");
4372 	lockmgr_printinfo(vp->v_vnlock);
4373 	if (vp->v_data != NULL)
4374 		VOP_PRINT(vp);
4375 }
4376 
4377 #ifdef DDB
4378 /*
4379  * List all of the locked vnodes in the system.
4380  * Called when debugging the kernel.
4381  */
4382 DB_SHOW_COMMAND_FLAGS(lockedvnods, lockedvnodes, DB_CMD_MEMSAFE)
4383 {
4384 	struct mount *mp;
4385 	struct vnode *vp;
4386 
4387 	/*
4388 	 * Note: because this is DDB, we can't obey the locking semantics
4389 	 * for these structures, which means we could catch an inconsistent
4390 	 * state and dereference a nasty pointer.  Not much to be done
4391 	 * about that.
4392 	 */
4393 	db_printf("Locked vnodes\n");
4394 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4395 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4396 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4397 				vn_printf(vp, "vnode ");
4398 		}
4399 	}
4400 }
4401 
4402 /*
4403  * Show details about the given vnode.
4404  */
4405 DB_SHOW_COMMAND(vnode, db_show_vnode)
4406 {
4407 	struct vnode *vp;
4408 
4409 	if (!have_addr)
4410 		return;
4411 	vp = (struct vnode *)addr;
4412 	vn_printf(vp, "vnode ");
4413 }
4414 
4415 /*
4416  * Show details about the given mount point.
4417  */
4418 DB_SHOW_COMMAND(mount, db_show_mount)
4419 {
4420 	struct mount *mp;
4421 	struct vfsopt *opt;
4422 	struct statfs *sp;
4423 	struct vnode *vp;
4424 	char buf[512];
4425 	uint64_t mflags;
4426 	u_int flags;
4427 
4428 	if (!have_addr) {
4429 		/* No address given, print short info about all mount points. */
4430 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4431 			db_printf("%p %s on %s (%s)\n", mp,
4432 			    mp->mnt_stat.f_mntfromname,
4433 			    mp->mnt_stat.f_mntonname,
4434 			    mp->mnt_stat.f_fstypename);
4435 			if (db_pager_quit)
4436 				break;
4437 		}
4438 		db_printf("\nMore info: show mount <addr>\n");
4439 		return;
4440 	}
4441 
4442 	mp = (struct mount *)addr;
4443 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4444 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4445 
4446 	buf[0] = '\0';
4447 	mflags = mp->mnt_flag;
4448 #define	MNT_FLAG(flag)	do {						\
4449 	if (mflags & (flag)) {						\
4450 		if (buf[0] != '\0')					\
4451 			strlcat(buf, ", ", sizeof(buf));		\
4452 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4453 		mflags &= ~(flag);					\
4454 	}								\
4455 } while (0)
4456 	MNT_FLAG(MNT_RDONLY);
4457 	MNT_FLAG(MNT_SYNCHRONOUS);
4458 	MNT_FLAG(MNT_NOEXEC);
4459 	MNT_FLAG(MNT_NOSUID);
4460 	MNT_FLAG(MNT_NFS4ACLS);
4461 	MNT_FLAG(MNT_UNION);
4462 	MNT_FLAG(MNT_ASYNC);
4463 	MNT_FLAG(MNT_SUIDDIR);
4464 	MNT_FLAG(MNT_SOFTDEP);
4465 	MNT_FLAG(MNT_NOSYMFOLLOW);
4466 	MNT_FLAG(MNT_GJOURNAL);
4467 	MNT_FLAG(MNT_MULTILABEL);
4468 	MNT_FLAG(MNT_ACLS);
4469 	MNT_FLAG(MNT_NOATIME);
4470 	MNT_FLAG(MNT_NOCLUSTERR);
4471 	MNT_FLAG(MNT_NOCLUSTERW);
4472 	MNT_FLAG(MNT_SUJ);
4473 	MNT_FLAG(MNT_EXRDONLY);
4474 	MNT_FLAG(MNT_EXPORTED);
4475 	MNT_FLAG(MNT_DEFEXPORTED);
4476 	MNT_FLAG(MNT_EXPORTANON);
4477 	MNT_FLAG(MNT_EXKERB);
4478 	MNT_FLAG(MNT_EXPUBLIC);
4479 	MNT_FLAG(MNT_LOCAL);
4480 	MNT_FLAG(MNT_QUOTA);
4481 	MNT_FLAG(MNT_ROOTFS);
4482 	MNT_FLAG(MNT_USER);
4483 	MNT_FLAG(MNT_IGNORE);
4484 	MNT_FLAG(MNT_UPDATE);
4485 	MNT_FLAG(MNT_DELEXPORT);
4486 	MNT_FLAG(MNT_RELOAD);
4487 	MNT_FLAG(MNT_FORCE);
4488 	MNT_FLAG(MNT_SNAPSHOT);
4489 	MNT_FLAG(MNT_BYFSID);
4490 #undef MNT_FLAG
4491 	if (mflags != 0) {
4492 		if (buf[0] != '\0')
4493 			strlcat(buf, ", ", sizeof(buf));
4494 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4495 		    "0x%016jx", mflags);
4496 	}
4497 	db_printf("    mnt_flag = %s\n", buf);
4498 
4499 	buf[0] = '\0';
4500 	flags = mp->mnt_kern_flag;
4501 #define	MNT_KERN_FLAG(flag)	do {					\
4502 	if (flags & (flag)) {						\
4503 		if (buf[0] != '\0')					\
4504 			strlcat(buf, ", ", sizeof(buf));		\
4505 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4506 		flags &= ~(flag);					\
4507 	}								\
4508 } while (0)
4509 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4510 	MNT_KERN_FLAG(MNTK_ASYNC);
4511 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4512 	MNT_KERN_FLAG(MNTK_NOMSYNC);
4513 	MNT_KERN_FLAG(MNTK_DRAINING);
4514 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4515 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4516 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4517 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4518 	MNT_KERN_FLAG(MNTK_RECURSE);
4519 	MNT_KERN_FLAG(MNTK_UPPER_WAITER);
4520 	MNT_KERN_FLAG(MNTK_UNLOCKED_INSMNTQUE);
4521 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4522 	MNT_KERN_FLAG(MNTK_VMSETSIZE_BUG);
4523 	MNT_KERN_FLAG(MNTK_FPLOOKUP);
4524 	MNT_KERN_FLAG(MNTK_TASKQUEUE_WAITER);
4525 	MNT_KERN_FLAG(MNTK_NOASYNC);
4526 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4527 	MNT_KERN_FLAG(MNTK_MWAIT);
4528 	MNT_KERN_FLAG(MNTK_SUSPEND);
4529 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4530 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4531 	MNT_KERN_FLAG(MNTK_NULL_NOCACHE);
4532 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4533 #undef MNT_KERN_FLAG
4534 	if (flags != 0) {
4535 		if (buf[0] != '\0')
4536 			strlcat(buf, ", ", sizeof(buf));
4537 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4538 		    "0x%08x", flags);
4539 	}
4540 	db_printf("    mnt_kern_flag = %s\n", buf);
4541 
4542 	db_printf("    mnt_opt = ");
4543 	opt = TAILQ_FIRST(mp->mnt_opt);
4544 	if (opt != NULL) {
4545 		db_printf("%s", opt->name);
4546 		opt = TAILQ_NEXT(opt, link);
4547 		while (opt != NULL) {
4548 			db_printf(", %s", opt->name);
4549 			opt = TAILQ_NEXT(opt, link);
4550 		}
4551 	}
4552 	db_printf("\n");
4553 
4554 	sp = &mp->mnt_stat;
4555 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4556 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4557 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4558 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4559 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4560 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4561 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4562 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4563 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4564 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4565 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4566 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4567 
4568 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4569 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4570 	if (jailed(mp->mnt_cred))
4571 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4572 	db_printf(" }\n");
4573 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4574 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4575 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4576 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4577 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4578 	    mp->mnt_lazyvnodelistsize);
4579 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4580 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4581 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4582 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4583 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4584 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4585 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4586 	db_printf("    mnt_secondary_accwrites = %d\n",
4587 	    mp->mnt_secondary_accwrites);
4588 	db_printf("    mnt_gjprovider = %s\n",
4589 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4590 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4591 
4592 	db_printf("\n\nList of active vnodes\n");
4593 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4594 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4595 			vn_printf(vp, "vnode ");
4596 			if (db_pager_quit)
4597 				break;
4598 		}
4599 	}
4600 	db_printf("\n\nList of inactive vnodes\n");
4601 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4602 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4603 			vn_printf(vp, "vnode ");
4604 			if (db_pager_quit)
4605 				break;
4606 		}
4607 	}
4608 }
4609 #endif	/* DDB */
4610 
4611 /*
4612  * Fill in a struct xvfsconf based on a struct vfsconf.
4613  */
4614 static int
4615 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4616 {
4617 	struct xvfsconf xvfsp;
4618 
4619 	bzero(&xvfsp, sizeof(xvfsp));
4620 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4621 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4622 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4623 	xvfsp.vfc_flags = vfsp->vfc_flags;
4624 	/*
4625 	 * These are unused in userland, we keep them
4626 	 * to not break binary compatibility.
4627 	 */
4628 	xvfsp.vfc_vfsops = NULL;
4629 	xvfsp.vfc_next = NULL;
4630 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4631 }
4632 
4633 #ifdef COMPAT_FREEBSD32
4634 struct xvfsconf32 {
4635 	uint32_t	vfc_vfsops;
4636 	char		vfc_name[MFSNAMELEN];
4637 	int32_t		vfc_typenum;
4638 	int32_t		vfc_refcount;
4639 	int32_t		vfc_flags;
4640 	uint32_t	vfc_next;
4641 };
4642 
4643 static int
4644 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4645 {
4646 	struct xvfsconf32 xvfsp;
4647 
4648 	bzero(&xvfsp, sizeof(xvfsp));
4649 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4650 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4651 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4652 	xvfsp.vfc_flags = vfsp->vfc_flags;
4653 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4654 }
4655 #endif
4656 
4657 /*
4658  * Top level filesystem related information gathering.
4659  */
4660 static int
4661 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4662 {
4663 	struct vfsconf *vfsp;
4664 	int error;
4665 
4666 	error = 0;
4667 	vfsconf_slock();
4668 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4669 #ifdef COMPAT_FREEBSD32
4670 		if (req->flags & SCTL_MASK32)
4671 			error = vfsconf2x32(req, vfsp);
4672 		else
4673 #endif
4674 			error = vfsconf2x(req, vfsp);
4675 		if (error)
4676 			break;
4677 	}
4678 	vfsconf_sunlock();
4679 	return (error);
4680 }
4681 
4682 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4683     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4684     "S,xvfsconf", "List of all configured filesystems");
4685 
4686 #ifndef BURN_BRIDGES
4687 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4688 
4689 static int
4690 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4691 {
4692 	int *name = (int *)arg1 - 1;	/* XXX */
4693 	u_int namelen = arg2 + 1;	/* XXX */
4694 	struct vfsconf *vfsp;
4695 
4696 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4697 	    "please rebuild world\n");
4698 
4699 #if 1 || defined(COMPAT_PRELITE2)
4700 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4701 	if (namelen == 1)
4702 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4703 #endif
4704 
4705 	switch (name[1]) {
4706 	case VFS_MAXTYPENUM:
4707 		if (namelen != 2)
4708 			return (ENOTDIR);
4709 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4710 	case VFS_CONF:
4711 		if (namelen != 3)
4712 			return (ENOTDIR);	/* overloaded */
4713 		vfsconf_slock();
4714 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4715 			if (vfsp->vfc_typenum == name[2])
4716 				break;
4717 		}
4718 		vfsconf_sunlock();
4719 		if (vfsp == NULL)
4720 			return (EOPNOTSUPP);
4721 #ifdef COMPAT_FREEBSD32
4722 		if (req->flags & SCTL_MASK32)
4723 			return (vfsconf2x32(req, vfsp));
4724 		else
4725 #endif
4726 			return (vfsconf2x(req, vfsp));
4727 	}
4728 	return (EOPNOTSUPP);
4729 }
4730 
4731 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4732     CTLFLAG_MPSAFE, vfs_sysctl,
4733     "Generic filesystem");
4734 
4735 #if 1 || defined(COMPAT_PRELITE2)
4736 
4737 static int
4738 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4739 {
4740 	int error;
4741 	struct vfsconf *vfsp;
4742 	struct ovfsconf ovfs;
4743 
4744 	vfsconf_slock();
4745 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4746 		bzero(&ovfs, sizeof(ovfs));
4747 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4748 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4749 		ovfs.vfc_index = vfsp->vfc_typenum;
4750 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4751 		ovfs.vfc_flags = vfsp->vfc_flags;
4752 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4753 		if (error != 0) {
4754 			vfsconf_sunlock();
4755 			return (error);
4756 		}
4757 	}
4758 	vfsconf_sunlock();
4759 	return (0);
4760 }
4761 
4762 #endif /* 1 || COMPAT_PRELITE2 */
4763 #endif /* !BURN_BRIDGES */
4764 
4765 static void
4766 unmount_or_warn(struct mount *mp)
4767 {
4768 	int error;
4769 
4770 	error = dounmount(mp, MNT_FORCE, curthread);
4771 	if (error != 0) {
4772 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4773 		if (error == EBUSY)
4774 			printf("BUSY)\n");
4775 		else
4776 			printf("%d)\n", error);
4777 	}
4778 }
4779 
4780 /*
4781  * Unmount all filesystems. The list is traversed in reverse order
4782  * of mounting to avoid dependencies.
4783  */
4784 void
4785 vfs_unmountall(void)
4786 {
4787 	struct mount *mp, *tmp;
4788 
4789 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4790 
4791 	/*
4792 	 * Since this only runs when rebooting, it is not interlocked.
4793 	 */
4794 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4795 		vfs_ref(mp);
4796 
4797 		/*
4798 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4799 		 * unmount of the latter.
4800 		 */
4801 		if (mp == rootdevmp)
4802 			continue;
4803 
4804 		unmount_or_warn(mp);
4805 	}
4806 
4807 	if (rootdevmp != NULL)
4808 		unmount_or_warn(rootdevmp);
4809 }
4810 
4811 static void
4812 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4813 {
4814 
4815 	ASSERT_VI_LOCKED(vp, __func__);
4816 	VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
4817 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
4818 		vdropl(vp);
4819 		return;
4820 	}
4821 	if (vn_lock(vp, lkflags) == 0) {
4822 		VI_LOCK(vp);
4823 		vinactive(vp);
4824 		VOP_UNLOCK(vp);
4825 		vdropl(vp);
4826 		return;
4827 	}
4828 	vdefer_inactive_unlocked(vp);
4829 }
4830 
4831 static int
4832 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4833 {
4834 
4835 	return (vp->v_iflag & VI_DEFINACT);
4836 }
4837 
4838 static void __noinline
4839 vfs_periodic_inactive(struct mount *mp, int flags)
4840 {
4841 	struct vnode *vp, *mvp;
4842 	int lkflags;
4843 
4844 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4845 	if (flags != MNT_WAIT)
4846 		lkflags |= LK_NOWAIT;
4847 
4848 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4849 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
4850 			VI_UNLOCK(vp);
4851 			continue;
4852 		}
4853 		vp->v_iflag &= ~VI_DEFINACT;
4854 		vfs_deferred_inactive(vp, lkflags);
4855 	}
4856 }
4857 
4858 static inline bool
4859 vfs_want_msync(struct vnode *vp)
4860 {
4861 	struct vm_object *obj;
4862 
4863 	/*
4864 	 * This test may be performed without any locks held.
4865 	 * We rely on vm_object's type stability.
4866 	 */
4867 	if (vp->v_vflag & VV_NOSYNC)
4868 		return (false);
4869 	obj = vp->v_object;
4870 	return (obj != NULL && vm_object_mightbedirty(obj));
4871 }
4872 
4873 static int
4874 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4875 {
4876 
4877 	if (vp->v_vflag & VV_NOSYNC)
4878 		return (false);
4879 	if (vp->v_iflag & VI_DEFINACT)
4880 		return (true);
4881 	return (vfs_want_msync(vp));
4882 }
4883 
4884 static void __noinline
4885 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4886 {
4887 	struct vnode *vp, *mvp;
4888 	struct vm_object *obj;
4889 	int lkflags, objflags;
4890 	bool seen_defer;
4891 
4892 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4893 	if (flags != MNT_WAIT) {
4894 		lkflags |= LK_NOWAIT;
4895 		objflags = OBJPC_NOSYNC;
4896 	} else {
4897 		objflags = OBJPC_SYNC;
4898 	}
4899 
4900 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4901 		seen_defer = false;
4902 		if (vp->v_iflag & VI_DEFINACT) {
4903 			vp->v_iflag &= ~VI_DEFINACT;
4904 			seen_defer = true;
4905 		}
4906 		if (!vfs_want_msync(vp)) {
4907 			if (seen_defer)
4908 				vfs_deferred_inactive(vp, lkflags);
4909 			else
4910 				VI_UNLOCK(vp);
4911 			continue;
4912 		}
4913 		if (vget(vp, lkflags) == 0) {
4914 			obj = vp->v_object;
4915 			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4916 				VM_OBJECT_WLOCK(obj);
4917 				vm_object_page_clean(obj, 0, 0, objflags);
4918 				VM_OBJECT_WUNLOCK(obj);
4919 			}
4920 			vput(vp);
4921 			if (seen_defer)
4922 				vdrop(vp);
4923 		} else {
4924 			if (seen_defer)
4925 				vdefer_inactive_unlocked(vp);
4926 		}
4927 	}
4928 }
4929 
4930 void
4931 vfs_periodic(struct mount *mp, int flags)
4932 {
4933 
4934 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4935 
4936 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4937 		vfs_periodic_inactive(mp, flags);
4938 	else
4939 		vfs_periodic_msync_inactive(mp, flags);
4940 }
4941 
4942 static void
4943 destroy_vpollinfo_free(struct vpollinfo *vi)
4944 {
4945 
4946 	knlist_destroy(&vi->vpi_selinfo.si_note);
4947 	mtx_destroy(&vi->vpi_lock);
4948 	free(vi, M_VNODEPOLL);
4949 }
4950 
4951 static void
4952 destroy_vpollinfo(struct vpollinfo *vi)
4953 {
4954 
4955 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4956 	seldrain(&vi->vpi_selinfo);
4957 	destroy_vpollinfo_free(vi);
4958 }
4959 
4960 /*
4961  * Initialize per-vnode helper structure to hold poll-related state.
4962  */
4963 void
4964 v_addpollinfo(struct vnode *vp)
4965 {
4966 	struct vpollinfo *vi;
4967 
4968 	if (vp->v_pollinfo != NULL)
4969 		return;
4970 	vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO);
4971 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4972 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4973 	    vfs_knlunlock, vfs_knl_assert_lock);
4974 	VI_LOCK(vp);
4975 	if (vp->v_pollinfo != NULL) {
4976 		VI_UNLOCK(vp);
4977 		destroy_vpollinfo_free(vi);
4978 		return;
4979 	}
4980 	vp->v_pollinfo = vi;
4981 	VI_UNLOCK(vp);
4982 }
4983 
4984 /*
4985  * Record a process's interest in events which might happen to
4986  * a vnode.  Because poll uses the historic select-style interface
4987  * internally, this routine serves as both the ``check for any
4988  * pending events'' and the ``record my interest in future events''
4989  * functions.  (These are done together, while the lock is held,
4990  * to avoid race conditions.)
4991  */
4992 int
4993 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4994 {
4995 
4996 	v_addpollinfo(vp);
4997 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4998 	if (vp->v_pollinfo->vpi_revents & events) {
4999 		/*
5000 		 * This leaves events we are not interested
5001 		 * in available for the other process which
5002 		 * which presumably had requested them
5003 		 * (otherwise they would never have been
5004 		 * recorded).
5005 		 */
5006 		events &= vp->v_pollinfo->vpi_revents;
5007 		vp->v_pollinfo->vpi_revents &= ~events;
5008 
5009 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
5010 		return (events);
5011 	}
5012 	vp->v_pollinfo->vpi_events |= events;
5013 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
5014 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
5015 	return (0);
5016 }
5017 
5018 /*
5019  * Routine to create and manage a filesystem syncer vnode.
5020  */
5021 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
5022 static int	sync_fsync(struct  vop_fsync_args *);
5023 static int	sync_inactive(struct  vop_inactive_args *);
5024 static int	sync_reclaim(struct  vop_reclaim_args *);
5025 
5026 static struct vop_vector sync_vnodeops = {
5027 	.vop_bypass =	VOP_EOPNOTSUPP,
5028 	.vop_close =	sync_close,
5029 	.vop_fsync =	sync_fsync,
5030 	.vop_getwritemount = vop_stdgetwritemount,
5031 	.vop_inactive =	sync_inactive,
5032 	.vop_need_inactive = vop_stdneed_inactive,
5033 	.vop_reclaim =	sync_reclaim,
5034 	.vop_lock1 =	vop_stdlock,
5035 	.vop_unlock =	vop_stdunlock,
5036 	.vop_islocked =	vop_stdislocked,
5037 	.vop_fplookup_vexec = VOP_EAGAIN,
5038 	.vop_fplookup_symlink = VOP_EAGAIN,
5039 };
5040 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
5041 
5042 /*
5043  * Create a new filesystem syncer vnode for the specified mount point.
5044  */
5045 void
5046 vfs_allocate_syncvnode(struct mount *mp)
5047 {
5048 	struct vnode *vp;
5049 	struct bufobj *bo;
5050 	static long start, incr, next;
5051 	int error;
5052 
5053 	/* Allocate a new vnode */
5054 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
5055 	if (error != 0)
5056 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
5057 	vp->v_type = VNON;
5058 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5059 	vp->v_vflag |= VV_FORCEINSMQ;
5060 	error = insmntque1(vp, mp);
5061 	if (error != 0)
5062 		panic("vfs_allocate_syncvnode: insmntque() failed");
5063 	vp->v_vflag &= ~VV_FORCEINSMQ;
5064 	vn_set_state(vp, VSTATE_CONSTRUCTED);
5065 	VOP_UNLOCK(vp);
5066 	/*
5067 	 * Place the vnode onto the syncer worklist. We attempt to
5068 	 * scatter them about on the list so that they will go off
5069 	 * at evenly distributed times even if all the filesystems
5070 	 * are mounted at once.
5071 	 */
5072 	next += incr;
5073 	if (next == 0 || next > syncer_maxdelay) {
5074 		start /= 2;
5075 		incr /= 2;
5076 		if (start == 0) {
5077 			start = syncer_maxdelay / 2;
5078 			incr = syncer_maxdelay;
5079 		}
5080 		next = start;
5081 	}
5082 	bo = &vp->v_bufobj;
5083 	BO_LOCK(bo);
5084 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
5085 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
5086 	mtx_lock(&sync_mtx);
5087 	sync_vnode_count++;
5088 	if (mp->mnt_syncer == NULL) {
5089 		mp->mnt_syncer = vp;
5090 		vp = NULL;
5091 	}
5092 	mtx_unlock(&sync_mtx);
5093 	BO_UNLOCK(bo);
5094 	if (vp != NULL) {
5095 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5096 		vgone(vp);
5097 		vput(vp);
5098 	}
5099 }
5100 
5101 void
5102 vfs_deallocate_syncvnode(struct mount *mp)
5103 {
5104 	struct vnode *vp;
5105 
5106 	mtx_lock(&sync_mtx);
5107 	vp = mp->mnt_syncer;
5108 	if (vp != NULL)
5109 		mp->mnt_syncer = NULL;
5110 	mtx_unlock(&sync_mtx);
5111 	if (vp != NULL)
5112 		vrele(vp);
5113 }
5114 
5115 /*
5116  * Do a lazy sync of the filesystem.
5117  */
5118 static int
5119 sync_fsync(struct vop_fsync_args *ap)
5120 {
5121 	struct vnode *syncvp = ap->a_vp;
5122 	struct mount *mp = syncvp->v_mount;
5123 	int error, save;
5124 	struct bufobj *bo;
5125 
5126 	/*
5127 	 * We only need to do something if this is a lazy evaluation.
5128 	 */
5129 	if (ap->a_waitfor != MNT_LAZY)
5130 		return (0);
5131 
5132 	/*
5133 	 * Move ourselves to the back of the sync list.
5134 	 */
5135 	bo = &syncvp->v_bufobj;
5136 	BO_LOCK(bo);
5137 	vn_syncer_add_to_worklist(bo, syncdelay);
5138 	BO_UNLOCK(bo);
5139 
5140 	/*
5141 	 * Walk the list of vnodes pushing all that are dirty and
5142 	 * not already on the sync list.
5143 	 */
5144 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
5145 		return (0);
5146 	VOP_UNLOCK(syncvp);
5147 	save = curthread_pflags_set(TDP_SYNCIO);
5148 	/*
5149 	 * The filesystem at hand may be idle with free vnodes stored in the
5150 	 * batch.  Return them instead of letting them stay there indefinitely.
5151 	 */
5152 	vfs_periodic(mp, MNT_NOWAIT);
5153 	error = VFS_SYNC(mp, MNT_LAZY);
5154 	curthread_pflags_restore(save);
5155 	vn_lock(syncvp, LK_EXCLUSIVE | LK_RETRY);
5156 	vfs_unbusy(mp);
5157 	return (error);
5158 }
5159 
5160 /*
5161  * The syncer vnode is no referenced.
5162  */
5163 static int
5164 sync_inactive(struct vop_inactive_args *ap)
5165 {
5166 
5167 	vgone(ap->a_vp);
5168 	return (0);
5169 }
5170 
5171 /*
5172  * The syncer vnode is no longer needed and is being decommissioned.
5173  *
5174  * Modifications to the worklist must be protected by sync_mtx.
5175  */
5176 static int
5177 sync_reclaim(struct vop_reclaim_args *ap)
5178 {
5179 	struct vnode *vp = ap->a_vp;
5180 	struct bufobj *bo;
5181 
5182 	bo = &vp->v_bufobj;
5183 	BO_LOCK(bo);
5184 	mtx_lock(&sync_mtx);
5185 	if (vp->v_mount->mnt_syncer == vp)
5186 		vp->v_mount->mnt_syncer = NULL;
5187 	if (bo->bo_flag & BO_ONWORKLST) {
5188 		LIST_REMOVE(bo, bo_synclist);
5189 		syncer_worklist_len--;
5190 		sync_vnode_count--;
5191 		bo->bo_flag &= ~BO_ONWORKLST;
5192 	}
5193 	mtx_unlock(&sync_mtx);
5194 	BO_UNLOCK(bo);
5195 
5196 	return (0);
5197 }
5198 
5199 int
5200 vn_need_pageq_flush(struct vnode *vp)
5201 {
5202 	struct vm_object *obj;
5203 
5204 	obj = vp->v_object;
5205 	return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5206 	    vm_object_mightbedirty(obj));
5207 }
5208 
5209 /*
5210  * Check if vnode represents a disk device
5211  */
5212 bool
5213 vn_isdisk_error(struct vnode *vp, int *errp)
5214 {
5215 	int error;
5216 
5217 	if (vp->v_type != VCHR) {
5218 		error = ENOTBLK;
5219 		goto out;
5220 	}
5221 	error = 0;
5222 	dev_lock();
5223 	if (vp->v_rdev == NULL)
5224 		error = ENXIO;
5225 	else if (vp->v_rdev->si_devsw == NULL)
5226 		error = ENXIO;
5227 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5228 		error = ENOTBLK;
5229 	dev_unlock();
5230 out:
5231 	*errp = error;
5232 	return (error == 0);
5233 }
5234 
5235 bool
5236 vn_isdisk(struct vnode *vp)
5237 {
5238 	int error;
5239 
5240 	return (vn_isdisk_error(vp, &error));
5241 }
5242 
5243 /*
5244  * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
5245  * the comment above cache_fplookup for details.
5246  */
5247 int
5248 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred)
5249 {
5250 	int error;
5251 
5252 	VFS_SMR_ASSERT_ENTERED();
5253 
5254 	/* Check the owner. */
5255 	if (cred->cr_uid == file_uid) {
5256 		if (file_mode & S_IXUSR)
5257 			return (0);
5258 		goto out_error;
5259 	}
5260 
5261 	/* Otherwise, check the groups (first match) */
5262 	if (groupmember(file_gid, cred)) {
5263 		if (file_mode & S_IXGRP)
5264 			return (0);
5265 		goto out_error;
5266 	}
5267 
5268 	/* Otherwise, check everyone else. */
5269 	if (file_mode & S_IXOTH)
5270 		return (0);
5271 out_error:
5272 	/*
5273 	 * Permission check failed, but it is possible denial will get overwritten
5274 	 * (e.g., when root is traversing through a 700 directory owned by someone
5275 	 * else).
5276 	 *
5277 	 * vaccess() calls priv_check_cred which in turn can descent into MAC
5278 	 * modules overriding this result. It's quite unclear what semantics
5279 	 * are allowed for them to operate, thus for safety we don't call them
5280 	 * from within the SMR section. This also means if any such modules
5281 	 * are present, we have to let the regular lookup decide.
5282 	 */
5283 	error = priv_check_cred_vfs_lookup_nomac(cred);
5284 	switch (error) {
5285 	case 0:
5286 		return (0);
5287 	case EAGAIN:
5288 		/*
5289 		 * MAC modules present.
5290 		 */
5291 		return (EAGAIN);
5292 	case EPERM:
5293 		return (EACCES);
5294 	default:
5295 		return (error);
5296 	}
5297 }
5298 
5299 /*
5300  * Common filesystem object access control check routine.  Accepts a
5301  * vnode's type, "mode", uid and gid, requested access mode, and credentials.
5302  * Returns 0 on success, or an errno on failure.
5303  */
5304 int
5305 vaccess(__enum_uint8(vtype) type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5306     accmode_t accmode, struct ucred *cred)
5307 {
5308 	accmode_t dac_granted;
5309 	accmode_t priv_granted;
5310 
5311 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5312 	    ("invalid bit in accmode"));
5313 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5314 	    ("VAPPEND without VWRITE"));
5315 
5316 	/*
5317 	 * Look for a normal, non-privileged way to access the file/directory
5318 	 * as requested.  If it exists, go with that.
5319 	 */
5320 
5321 	dac_granted = 0;
5322 
5323 	/* Check the owner. */
5324 	if (cred->cr_uid == file_uid) {
5325 		dac_granted |= VADMIN;
5326 		if (file_mode & S_IXUSR)
5327 			dac_granted |= VEXEC;
5328 		if (file_mode & S_IRUSR)
5329 			dac_granted |= VREAD;
5330 		if (file_mode & S_IWUSR)
5331 			dac_granted |= (VWRITE | VAPPEND);
5332 
5333 		if ((accmode & dac_granted) == accmode)
5334 			return (0);
5335 
5336 		goto privcheck;
5337 	}
5338 
5339 	/* Otherwise, check the groups (first match) */
5340 	if (groupmember(file_gid, cred)) {
5341 		if (file_mode & S_IXGRP)
5342 			dac_granted |= VEXEC;
5343 		if (file_mode & S_IRGRP)
5344 			dac_granted |= VREAD;
5345 		if (file_mode & S_IWGRP)
5346 			dac_granted |= (VWRITE | VAPPEND);
5347 
5348 		if ((accmode & dac_granted) == accmode)
5349 			return (0);
5350 
5351 		goto privcheck;
5352 	}
5353 
5354 	/* Otherwise, check everyone else. */
5355 	if (file_mode & S_IXOTH)
5356 		dac_granted |= VEXEC;
5357 	if (file_mode & S_IROTH)
5358 		dac_granted |= VREAD;
5359 	if (file_mode & S_IWOTH)
5360 		dac_granted |= (VWRITE | VAPPEND);
5361 	if ((accmode & dac_granted) == accmode)
5362 		return (0);
5363 
5364 privcheck:
5365 	/*
5366 	 * Build a privilege mask to determine if the set of privileges
5367 	 * satisfies the requirements when combined with the granted mask
5368 	 * from above.  For each privilege, if the privilege is required,
5369 	 * bitwise or the request type onto the priv_granted mask.
5370 	 */
5371 	priv_granted = 0;
5372 
5373 	if (type == VDIR) {
5374 		/*
5375 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5376 		 * requests, instead of PRIV_VFS_EXEC.
5377 		 */
5378 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5379 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5380 			priv_granted |= VEXEC;
5381 	} else {
5382 		/*
5383 		 * Ensure that at least one execute bit is on. Otherwise,
5384 		 * a privileged user will always succeed, and we don't want
5385 		 * this to happen unless the file really is executable.
5386 		 */
5387 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5388 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5389 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5390 			priv_granted |= VEXEC;
5391 	}
5392 
5393 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5394 	    !priv_check_cred(cred, PRIV_VFS_READ))
5395 		priv_granted |= VREAD;
5396 
5397 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5398 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5399 		priv_granted |= (VWRITE | VAPPEND);
5400 
5401 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5402 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5403 		priv_granted |= VADMIN;
5404 
5405 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5406 		return (0);
5407 	}
5408 
5409 	return ((accmode & VADMIN) ? EPERM : EACCES);
5410 }
5411 
5412 /*
5413  * Credential check based on process requesting service, and per-attribute
5414  * permissions.
5415  */
5416 int
5417 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5418     struct thread *td, accmode_t accmode)
5419 {
5420 
5421 	/*
5422 	 * Kernel-invoked always succeeds.
5423 	 */
5424 	if (cred == NOCRED)
5425 		return (0);
5426 
5427 	/*
5428 	 * Do not allow privileged processes in jail to directly manipulate
5429 	 * system attributes.
5430 	 */
5431 	switch (attrnamespace) {
5432 	case EXTATTR_NAMESPACE_SYSTEM:
5433 		/* Potentially should be: return (EPERM); */
5434 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5435 	case EXTATTR_NAMESPACE_USER:
5436 		return (VOP_ACCESS(vp, accmode, cred, td));
5437 	default:
5438 		return (EPERM);
5439 	}
5440 }
5441 
5442 #ifdef DEBUG_VFS_LOCKS
5443 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5444 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5445     "Drop into debugger on lock violation");
5446 
5447 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5448 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5449     0, "Check for interlock across VOPs");
5450 
5451 int vfs_badlock_print = 1;	/* Print lock violations. */
5452 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5453     0, "Print lock violations");
5454 
5455 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5456 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5457     0, "Print vnode details on lock violations");
5458 
5459 #ifdef KDB
5460 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5461 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5462     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5463 #endif
5464 
5465 static void
5466 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5467 {
5468 
5469 #ifdef KDB
5470 	if (vfs_badlock_backtrace)
5471 		kdb_backtrace();
5472 #endif
5473 	if (vfs_badlock_vnode)
5474 		vn_printf(vp, "vnode ");
5475 	if (vfs_badlock_print)
5476 		printf("%s: %p %s\n", str, (void *)vp, msg);
5477 	if (vfs_badlock_ddb)
5478 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5479 }
5480 
5481 void
5482 assert_vi_locked(struct vnode *vp, const char *str)
5483 {
5484 
5485 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5486 		vfs_badlock("interlock is not locked but should be", str, vp);
5487 }
5488 
5489 void
5490 assert_vi_unlocked(struct vnode *vp, const char *str)
5491 {
5492 
5493 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5494 		vfs_badlock("interlock is locked but should not be", str, vp);
5495 }
5496 
5497 void
5498 assert_vop_locked(struct vnode *vp, const char *str)
5499 {
5500 	if (KERNEL_PANICKED() || vp == NULL)
5501 		return;
5502 
5503 #ifdef WITNESS
5504 	if ((vp->v_irflag & VIRF_CROSSMP) == 0 &&
5505 	    witness_is_owned(&vp->v_vnlock->lock_object) == -1)
5506 #else
5507 	int locked = VOP_ISLOCKED(vp);
5508 	if (locked == 0 || locked == LK_EXCLOTHER)
5509 #endif
5510 		vfs_badlock("is not locked but should be", str, vp);
5511 }
5512 
5513 void
5514 assert_vop_unlocked(struct vnode *vp, const char *str)
5515 {
5516 	if (KERNEL_PANICKED() || vp == NULL)
5517 		return;
5518 
5519 #ifdef WITNESS
5520 	if ((vp->v_irflag & VIRF_CROSSMP) == 0 &&
5521 	    witness_is_owned(&vp->v_vnlock->lock_object) == 1)
5522 #else
5523 	if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5524 #endif
5525 		vfs_badlock("is locked but should not be", str, vp);
5526 }
5527 
5528 void
5529 assert_vop_elocked(struct vnode *vp, const char *str)
5530 {
5531 	if (KERNEL_PANICKED() || vp == NULL)
5532 		return;
5533 
5534 	if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5535 		vfs_badlock("is not exclusive locked but should be", str, vp);
5536 }
5537 #endif /* DEBUG_VFS_LOCKS */
5538 
5539 void
5540 vop_rename_fail(struct vop_rename_args *ap)
5541 {
5542 
5543 	if (ap->a_tvp != NULL)
5544 		vput(ap->a_tvp);
5545 	if (ap->a_tdvp == ap->a_tvp)
5546 		vrele(ap->a_tdvp);
5547 	else
5548 		vput(ap->a_tdvp);
5549 	vrele(ap->a_fdvp);
5550 	vrele(ap->a_fvp);
5551 }
5552 
5553 void
5554 vop_rename_pre(void *ap)
5555 {
5556 	struct vop_rename_args *a = ap;
5557 
5558 #ifdef DEBUG_VFS_LOCKS
5559 	if (a->a_tvp)
5560 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5561 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5562 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5563 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5564 
5565 	/* Check the source (from). */
5566 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5567 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5568 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5569 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5570 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5571 
5572 	/* Check the target. */
5573 	if (a->a_tvp)
5574 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5575 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5576 #endif
5577 	/*
5578 	 * It may be tempting to add vn_seqc_write_begin/end calls here and
5579 	 * in vop_rename_post but that's not going to work out since some
5580 	 * filesystems relookup vnodes mid-rename. This is probably a bug.
5581 	 *
5582 	 * For now filesystems are expected to do the relevant calls after they
5583 	 * decide what vnodes to operate on.
5584 	 */
5585 	if (a->a_tdvp != a->a_fdvp)
5586 		vhold(a->a_fdvp);
5587 	if (a->a_tvp != a->a_fvp)
5588 		vhold(a->a_fvp);
5589 	vhold(a->a_tdvp);
5590 	if (a->a_tvp)
5591 		vhold(a->a_tvp);
5592 }
5593 
5594 #ifdef DEBUG_VFS_LOCKS
5595 void
5596 vop_fplookup_vexec_debugpre(void *ap __unused)
5597 {
5598 
5599 	VFS_SMR_ASSERT_ENTERED();
5600 }
5601 
5602 void
5603 vop_fplookup_vexec_debugpost(void *ap, int rc)
5604 {
5605 	struct vop_fplookup_vexec_args *a;
5606 	struct vnode *vp;
5607 
5608 	a = ap;
5609 	vp = a->a_vp;
5610 
5611 	VFS_SMR_ASSERT_ENTERED();
5612 	if (rc == EOPNOTSUPP)
5613 		VNPASS(VN_IS_DOOMED(vp), vp);
5614 }
5615 
5616 void
5617 vop_fplookup_symlink_debugpre(void *ap __unused)
5618 {
5619 
5620 	VFS_SMR_ASSERT_ENTERED();
5621 }
5622 
5623 void
5624 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused)
5625 {
5626 
5627 	VFS_SMR_ASSERT_ENTERED();
5628 }
5629 
5630 static void
5631 vop_fsync_debugprepost(struct vnode *vp, const char *name)
5632 {
5633 	if (vp->v_type == VCHR)
5634 		;
5635 	else if (MNT_EXTENDED_SHARED(vp->v_mount))
5636 		ASSERT_VOP_LOCKED(vp, name);
5637 	else
5638 		ASSERT_VOP_ELOCKED(vp, name);
5639 }
5640 
5641 void
5642 vop_fsync_debugpre(void *a)
5643 {
5644 	struct vop_fsync_args *ap;
5645 
5646 	ap = a;
5647 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5648 }
5649 
5650 void
5651 vop_fsync_debugpost(void *a, int rc __unused)
5652 {
5653 	struct vop_fsync_args *ap;
5654 
5655 	ap = a;
5656 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5657 }
5658 
5659 void
5660 vop_fdatasync_debugpre(void *a)
5661 {
5662 	struct vop_fdatasync_args *ap;
5663 
5664 	ap = a;
5665 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5666 }
5667 
5668 void
5669 vop_fdatasync_debugpost(void *a, int rc __unused)
5670 {
5671 	struct vop_fdatasync_args *ap;
5672 
5673 	ap = a;
5674 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5675 }
5676 
5677 void
5678 vop_strategy_debugpre(void *ap)
5679 {
5680 	struct vop_strategy_args *a;
5681 	struct buf *bp;
5682 
5683 	a = ap;
5684 	bp = a->a_bp;
5685 
5686 	/*
5687 	 * Cluster ops lock their component buffers but not the IO container.
5688 	 */
5689 	if ((bp->b_flags & B_CLUSTER) != 0)
5690 		return;
5691 
5692 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5693 		if (vfs_badlock_print)
5694 			printf(
5695 			    "VOP_STRATEGY: bp is not locked but should be\n");
5696 		if (vfs_badlock_ddb)
5697 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5698 	}
5699 }
5700 
5701 void
5702 vop_lock_debugpre(void *ap)
5703 {
5704 	struct vop_lock1_args *a = ap;
5705 
5706 	if ((a->a_flags & LK_INTERLOCK) == 0)
5707 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5708 	else
5709 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5710 }
5711 
5712 void
5713 vop_lock_debugpost(void *ap, int rc)
5714 {
5715 	struct vop_lock1_args *a = ap;
5716 
5717 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5718 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5719 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5720 }
5721 
5722 void
5723 vop_unlock_debugpre(void *ap)
5724 {
5725 	struct vop_unlock_args *a = ap;
5726 	struct vnode *vp = a->a_vp;
5727 
5728 	VNPASS(vn_get_state(vp) != VSTATE_UNINITIALIZED, vp);
5729 	ASSERT_VOP_LOCKED(vp, "VOP_UNLOCK");
5730 }
5731 
5732 void
5733 vop_need_inactive_debugpre(void *ap)
5734 {
5735 	struct vop_need_inactive_args *a = ap;
5736 
5737 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5738 }
5739 
5740 void
5741 vop_need_inactive_debugpost(void *ap, int rc)
5742 {
5743 	struct vop_need_inactive_args *a = ap;
5744 
5745 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5746 }
5747 #endif
5748 
5749 void
5750 vop_create_pre(void *ap)
5751 {
5752 	struct vop_create_args *a;
5753 	struct vnode *dvp;
5754 
5755 	a = ap;
5756 	dvp = a->a_dvp;
5757 	vn_seqc_write_begin(dvp);
5758 }
5759 
5760 void
5761 vop_create_post(void *ap, int rc)
5762 {
5763 	struct vop_create_args *a;
5764 	struct vnode *dvp;
5765 
5766 	a = ap;
5767 	dvp = a->a_dvp;
5768 	vn_seqc_write_end(dvp);
5769 	if (!rc)
5770 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5771 }
5772 
5773 void
5774 vop_whiteout_pre(void *ap)
5775 {
5776 	struct vop_whiteout_args *a;
5777 	struct vnode *dvp;
5778 
5779 	a = ap;
5780 	dvp = a->a_dvp;
5781 	vn_seqc_write_begin(dvp);
5782 }
5783 
5784 void
5785 vop_whiteout_post(void *ap, int rc)
5786 {
5787 	struct vop_whiteout_args *a;
5788 	struct vnode *dvp;
5789 
5790 	a = ap;
5791 	dvp = a->a_dvp;
5792 	vn_seqc_write_end(dvp);
5793 }
5794 
5795 void
5796 vop_deleteextattr_pre(void *ap)
5797 {
5798 	struct vop_deleteextattr_args *a;
5799 	struct vnode *vp;
5800 
5801 	a = ap;
5802 	vp = a->a_vp;
5803 	vn_seqc_write_begin(vp);
5804 }
5805 
5806 void
5807 vop_deleteextattr_post(void *ap, int rc)
5808 {
5809 	struct vop_deleteextattr_args *a;
5810 	struct vnode *vp;
5811 
5812 	a = ap;
5813 	vp = a->a_vp;
5814 	vn_seqc_write_end(vp);
5815 	if (!rc)
5816 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5817 }
5818 
5819 void
5820 vop_link_pre(void *ap)
5821 {
5822 	struct vop_link_args *a;
5823 	struct vnode *vp, *tdvp;
5824 
5825 	a = ap;
5826 	vp = a->a_vp;
5827 	tdvp = a->a_tdvp;
5828 	vn_seqc_write_begin(vp);
5829 	vn_seqc_write_begin(tdvp);
5830 }
5831 
5832 void
5833 vop_link_post(void *ap, int rc)
5834 {
5835 	struct vop_link_args *a;
5836 	struct vnode *vp, *tdvp;
5837 
5838 	a = ap;
5839 	vp = a->a_vp;
5840 	tdvp = a->a_tdvp;
5841 	vn_seqc_write_end(vp);
5842 	vn_seqc_write_end(tdvp);
5843 	if (!rc) {
5844 		VFS_KNOTE_LOCKED(vp, NOTE_LINK);
5845 		VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE);
5846 	}
5847 }
5848 
5849 void
5850 vop_mkdir_pre(void *ap)
5851 {
5852 	struct vop_mkdir_args *a;
5853 	struct vnode *dvp;
5854 
5855 	a = ap;
5856 	dvp = a->a_dvp;
5857 	vn_seqc_write_begin(dvp);
5858 }
5859 
5860 void
5861 vop_mkdir_post(void *ap, int rc)
5862 {
5863 	struct vop_mkdir_args *a;
5864 	struct vnode *dvp;
5865 
5866 	a = ap;
5867 	dvp = a->a_dvp;
5868 	vn_seqc_write_end(dvp);
5869 	if (!rc)
5870 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5871 }
5872 
5873 #ifdef DEBUG_VFS_LOCKS
5874 void
5875 vop_mkdir_debugpost(void *ap, int rc)
5876 {
5877 	struct vop_mkdir_args *a;
5878 
5879 	a = ap;
5880 	if (!rc)
5881 		cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp);
5882 }
5883 #endif
5884 
5885 void
5886 vop_mknod_pre(void *ap)
5887 {
5888 	struct vop_mknod_args *a;
5889 	struct vnode *dvp;
5890 
5891 	a = ap;
5892 	dvp = a->a_dvp;
5893 	vn_seqc_write_begin(dvp);
5894 }
5895 
5896 void
5897 vop_mknod_post(void *ap, int rc)
5898 {
5899 	struct vop_mknod_args *a;
5900 	struct vnode *dvp;
5901 
5902 	a = ap;
5903 	dvp = a->a_dvp;
5904 	vn_seqc_write_end(dvp);
5905 	if (!rc)
5906 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5907 }
5908 
5909 void
5910 vop_reclaim_post(void *ap, int rc)
5911 {
5912 	struct vop_reclaim_args *a;
5913 	struct vnode *vp;
5914 
5915 	a = ap;
5916 	vp = a->a_vp;
5917 	ASSERT_VOP_IN_SEQC(vp);
5918 	if (!rc)
5919 		VFS_KNOTE_LOCKED(vp, NOTE_REVOKE);
5920 }
5921 
5922 void
5923 vop_remove_pre(void *ap)
5924 {
5925 	struct vop_remove_args *a;
5926 	struct vnode *dvp, *vp;
5927 
5928 	a = ap;
5929 	dvp = a->a_dvp;
5930 	vp = a->a_vp;
5931 	vn_seqc_write_begin(dvp);
5932 	vn_seqc_write_begin(vp);
5933 }
5934 
5935 void
5936 vop_remove_post(void *ap, int rc)
5937 {
5938 	struct vop_remove_args *a;
5939 	struct vnode *dvp, *vp;
5940 
5941 	a = ap;
5942 	dvp = a->a_dvp;
5943 	vp = a->a_vp;
5944 	vn_seqc_write_end(dvp);
5945 	vn_seqc_write_end(vp);
5946 	if (!rc) {
5947 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5948 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5949 	}
5950 }
5951 
5952 void
5953 vop_rename_post(void *ap, int rc)
5954 {
5955 	struct vop_rename_args *a = ap;
5956 	long hint;
5957 
5958 	if (!rc) {
5959 		hint = NOTE_WRITE;
5960 		if (a->a_fdvp == a->a_tdvp) {
5961 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5962 				hint |= NOTE_LINK;
5963 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5964 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5965 		} else {
5966 			hint |= NOTE_EXTEND;
5967 			if (a->a_fvp->v_type == VDIR)
5968 				hint |= NOTE_LINK;
5969 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5970 
5971 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5972 			    a->a_tvp->v_type == VDIR)
5973 				hint &= ~NOTE_LINK;
5974 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5975 		}
5976 
5977 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5978 		if (a->a_tvp)
5979 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5980 	}
5981 	if (a->a_tdvp != a->a_fdvp)
5982 		vdrop(a->a_fdvp);
5983 	if (a->a_tvp != a->a_fvp)
5984 		vdrop(a->a_fvp);
5985 	vdrop(a->a_tdvp);
5986 	if (a->a_tvp)
5987 		vdrop(a->a_tvp);
5988 }
5989 
5990 void
5991 vop_rmdir_pre(void *ap)
5992 {
5993 	struct vop_rmdir_args *a;
5994 	struct vnode *dvp, *vp;
5995 
5996 	a = ap;
5997 	dvp = a->a_dvp;
5998 	vp = a->a_vp;
5999 	vn_seqc_write_begin(dvp);
6000 	vn_seqc_write_begin(vp);
6001 }
6002 
6003 void
6004 vop_rmdir_post(void *ap, int rc)
6005 {
6006 	struct vop_rmdir_args *a;
6007 	struct vnode *dvp, *vp;
6008 
6009 	a = ap;
6010 	dvp = a->a_dvp;
6011 	vp = a->a_vp;
6012 	vn_seqc_write_end(dvp);
6013 	vn_seqc_write_end(vp);
6014 	if (!rc) {
6015 		vp->v_vflag |= VV_UNLINKED;
6016 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
6017 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
6018 	}
6019 }
6020 
6021 void
6022 vop_setattr_pre(void *ap)
6023 {
6024 	struct vop_setattr_args *a;
6025 	struct vnode *vp;
6026 
6027 	a = ap;
6028 	vp = a->a_vp;
6029 	vn_seqc_write_begin(vp);
6030 }
6031 
6032 void
6033 vop_setattr_post(void *ap, int rc)
6034 {
6035 	struct vop_setattr_args *a;
6036 	struct vnode *vp;
6037 
6038 	a = ap;
6039 	vp = a->a_vp;
6040 	vn_seqc_write_end(vp);
6041 	if (!rc)
6042 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
6043 }
6044 
6045 void
6046 vop_setacl_pre(void *ap)
6047 {
6048 	struct vop_setacl_args *a;
6049 	struct vnode *vp;
6050 
6051 	a = ap;
6052 	vp = a->a_vp;
6053 	vn_seqc_write_begin(vp);
6054 }
6055 
6056 void
6057 vop_setacl_post(void *ap, int rc __unused)
6058 {
6059 	struct vop_setacl_args *a;
6060 	struct vnode *vp;
6061 
6062 	a = ap;
6063 	vp = a->a_vp;
6064 	vn_seqc_write_end(vp);
6065 }
6066 
6067 void
6068 vop_setextattr_pre(void *ap)
6069 {
6070 	struct vop_setextattr_args *a;
6071 	struct vnode *vp;
6072 
6073 	a = ap;
6074 	vp = a->a_vp;
6075 	vn_seqc_write_begin(vp);
6076 }
6077 
6078 void
6079 vop_setextattr_post(void *ap, int rc)
6080 {
6081 	struct vop_setextattr_args *a;
6082 	struct vnode *vp;
6083 
6084 	a = ap;
6085 	vp = a->a_vp;
6086 	vn_seqc_write_end(vp);
6087 	if (!rc)
6088 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
6089 }
6090 
6091 void
6092 vop_symlink_pre(void *ap)
6093 {
6094 	struct vop_symlink_args *a;
6095 	struct vnode *dvp;
6096 
6097 	a = ap;
6098 	dvp = a->a_dvp;
6099 	vn_seqc_write_begin(dvp);
6100 }
6101 
6102 void
6103 vop_symlink_post(void *ap, int rc)
6104 {
6105 	struct vop_symlink_args *a;
6106 	struct vnode *dvp;
6107 
6108 	a = ap;
6109 	dvp = a->a_dvp;
6110 	vn_seqc_write_end(dvp);
6111 	if (!rc)
6112 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6113 }
6114 
6115 void
6116 vop_open_post(void *ap, int rc)
6117 {
6118 	struct vop_open_args *a = ap;
6119 
6120 	if (!rc)
6121 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
6122 }
6123 
6124 void
6125 vop_close_post(void *ap, int rc)
6126 {
6127 	struct vop_close_args *a = ap;
6128 
6129 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
6130 	    !VN_IS_DOOMED(a->a_vp))) {
6131 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
6132 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
6133 	}
6134 }
6135 
6136 void
6137 vop_read_post(void *ap, int rc)
6138 {
6139 	struct vop_read_args *a = ap;
6140 
6141 	if (!rc)
6142 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6143 }
6144 
6145 void
6146 vop_read_pgcache_post(void *ap, int rc)
6147 {
6148 	struct vop_read_pgcache_args *a = ap;
6149 
6150 	if (!rc)
6151 		VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ);
6152 }
6153 
6154 void
6155 vop_readdir_post(void *ap, int rc)
6156 {
6157 	struct vop_readdir_args *a = ap;
6158 
6159 	if (!rc)
6160 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6161 }
6162 
6163 static struct knlist fs_knlist;
6164 
6165 static void
6166 vfs_event_init(void *arg)
6167 {
6168 	knlist_init_mtx(&fs_knlist, NULL);
6169 }
6170 /* XXX - correct order? */
6171 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
6172 
6173 void
6174 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
6175 {
6176 
6177 	KNOTE_UNLOCKED(&fs_knlist, event);
6178 }
6179 
6180 static int	filt_fsattach(struct knote *kn);
6181 static void	filt_fsdetach(struct knote *kn);
6182 static int	filt_fsevent(struct knote *kn, long hint);
6183 
6184 struct filterops fs_filtops = {
6185 	.f_isfd = 0,
6186 	.f_attach = filt_fsattach,
6187 	.f_detach = filt_fsdetach,
6188 	.f_event = filt_fsevent
6189 };
6190 
6191 static int
6192 filt_fsattach(struct knote *kn)
6193 {
6194 
6195 	kn->kn_flags |= EV_CLEAR;
6196 	knlist_add(&fs_knlist, kn, 0);
6197 	return (0);
6198 }
6199 
6200 static void
6201 filt_fsdetach(struct knote *kn)
6202 {
6203 
6204 	knlist_remove(&fs_knlist, kn, 0);
6205 }
6206 
6207 static int
6208 filt_fsevent(struct knote *kn, long hint)
6209 {
6210 
6211 	kn->kn_fflags |= kn->kn_sfflags & hint;
6212 
6213 	return (kn->kn_fflags != 0);
6214 }
6215 
6216 static int
6217 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
6218 {
6219 	struct vfsidctl vc;
6220 	int error;
6221 	struct mount *mp;
6222 
6223 	error = SYSCTL_IN(req, &vc, sizeof(vc));
6224 	if (error)
6225 		return (error);
6226 	if (vc.vc_vers != VFS_CTL_VERS1)
6227 		return (EINVAL);
6228 	mp = vfs_getvfs(&vc.vc_fsid);
6229 	if (mp == NULL)
6230 		return (ENOENT);
6231 	/* ensure that a specific sysctl goes to the right filesystem. */
6232 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
6233 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
6234 		vfs_rel(mp);
6235 		return (EINVAL);
6236 	}
6237 	VCTLTOREQ(&vc, req);
6238 	error = VFS_SYSCTL(mp, vc.vc_op, req);
6239 	vfs_rel(mp);
6240 	return (error);
6241 }
6242 
6243 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
6244     NULL, 0, sysctl_vfs_ctl, "",
6245     "Sysctl by fsid");
6246 
6247 /*
6248  * Function to initialize a va_filerev field sensibly.
6249  * XXX: Wouldn't a random number make a lot more sense ??
6250  */
6251 u_quad_t
6252 init_va_filerev(void)
6253 {
6254 	struct bintime bt;
6255 
6256 	getbinuptime(&bt);
6257 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
6258 }
6259 
6260 static int	filt_vfsread(struct knote *kn, long hint);
6261 static int	filt_vfswrite(struct knote *kn, long hint);
6262 static int	filt_vfsvnode(struct knote *kn, long hint);
6263 static void	filt_vfsdetach(struct knote *kn);
6264 static struct filterops vfsread_filtops = {
6265 	.f_isfd = 1,
6266 	.f_detach = filt_vfsdetach,
6267 	.f_event = filt_vfsread
6268 };
6269 static struct filterops vfswrite_filtops = {
6270 	.f_isfd = 1,
6271 	.f_detach = filt_vfsdetach,
6272 	.f_event = filt_vfswrite
6273 };
6274 static struct filterops vfsvnode_filtops = {
6275 	.f_isfd = 1,
6276 	.f_detach = filt_vfsdetach,
6277 	.f_event = filt_vfsvnode
6278 };
6279 
6280 static void
6281 vfs_knllock(void *arg)
6282 {
6283 	struct vnode *vp = arg;
6284 
6285 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
6286 }
6287 
6288 static void
6289 vfs_knlunlock(void *arg)
6290 {
6291 	struct vnode *vp = arg;
6292 
6293 	VOP_UNLOCK(vp);
6294 }
6295 
6296 static void
6297 vfs_knl_assert_lock(void *arg, int what)
6298 {
6299 #ifdef DEBUG_VFS_LOCKS
6300 	struct vnode *vp = arg;
6301 
6302 	if (what == LA_LOCKED)
6303 		ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
6304 	else
6305 		ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
6306 #endif
6307 }
6308 
6309 int
6310 vfs_kqfilter(struct vop_kqfilter_args *ap)
6311 {
6312 	struct vnode *vp = ap->a_vp;
6313 	struct knote *kn = ap->a_kn;
6314 	struct knlist *knl;
6315 
6316 	KASSERT(vp->v_type != VFIFO || (kn->kn_filter != EVFILT_READ &&
6317 	    kn->kn_filter != EVFILT_WRITE),
6318 	    ("READ/WRITE filter on a FIFO leaked through"));
6319 	switch (kn->kn_filter) {
6320 	case EVFILT_READ:
6321 		kn->kn_fop = &vfsread_filtops;
6322 		break;
6323 	case EVFILT_WRITE:
6324 		kn->kn_fop = &vfswrite_filtops;
6325 		break;
6326 	case EVFILT_VNODE:
6327 		kn->kn_fop = &vfsvnode_filtops;
6328 		break;
6329 	default:
6330 		return (EINVAL);
6331 	}
6332 
6333 	kn->kn_hook = (caddr_t)vp;
6334 
6335 	v_addpollinfo(vp);
6336 	if (vp->v_pollinfo == NULL)
6337 		return (ENOMEM);
6338 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
6339 	vhold(vp);
6340 	knlist_add(knl, kn, 0);
6341 
6342 	return (0);
6343 }
6344 
6345 /*
6346  * Detach knote from vnode
6347  */
6348 static void
6349 filt_vfsdetach(struct knote *kn)
6350 {
6351 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6352 
6353 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
6354 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
6355 	vdrop(vp);
6356 }
6357 
6358 /*ARGSUSED*/
6359 static int
6360 filt_vfsread(struct knote *kn, long hint)
6361 {
6362 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6363 	off_t size;
6364 	int res;
6365 
6366 	/*
6367 	 * filesystem is gone, so set the EOF flag and schedule
6368 	 * the knote for deletion.
6369 	 */
6370 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6371 		VI_LOCK(vp);
6372 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6373 		VI_UNLOCK(vp);
6374 		return (1);
6375 	}
6376 
6377 	if (vn_getsize_locked(vp, &size, curthread->td_ucred) != 0)
6378 		return (0);
6379 
6380 	VI_LOCK(vp);
6381 	kn->kn_data = size - kn->kn_fp->f_offset;
6382 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
6383 	VI_UNLOCK(vp);
6384 	return (res);
6385 }
6386 
6387 /*ARGSUSED*/
6388 static int
6389 filt_vfswrite(struct knote *kn, long hint)
6390 {
6391 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6392 
6393 	VI_LOCK(vp);
6394 
6395 	/*
6396 	 * filesystem is gone, so set the EOF flag and schedule
6397 	 * the knote for deletion.
6398 	 */
6399 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
6400 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6401 
6402 	kn->kn_data = 0;
6403 	VI_UNLOCK(vp);
6404 	return (1);
6405 }
6406 
6407 static int
6408 filt_vfsvnode(struct knote *kn, long hint)
6409 {
6410 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6411 	int res;
6412 
6413 	VI_LOCK(vp);
6414 	if (kn->kn_sfflags & hint)
6415 		kn->kn_fflags |= hint;
6416 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6417 		kn->kn_flags |= EV_EOF;
6418 		VI_UNLOCK(vp);
6419 		return (1);
6420 	}
6421 	res = (kn->kn_fflags != 0);
6422 	VI_UNLOCK(vp);
6423 	return (res);
6424 }
6425 
6426 int
6427 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6428 {
6429 	int error;
6430 
6431 	if (dp->d_reclen > ap->a_uio->uio_resid)
6432 		return (ENAMETOOLONG);
6433 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
6434 	if (error) {
6435 		if (ap->a_ncookies != NULL) {
6436 			if (ap->a_cookies != NULL)
6437 				free(ap->a_cookies, M_TEMP);
6438 			ap->a_cookies = NULL;
6439 			*ap->a_ncookies = 0;
6440 		}
6441 		return (error);
6442 	}
6443 	if (ap->a_ncookies == NULL)
6444 		return (0);
6445 
6446 	KASSERT(ap->a_cookies,
6447 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6448 
6449 	*ap->a_cookies = realloc(*ap->a_cookies,
6450 	    (*ap->a_ncookies + 1) * sizeof(uint64_t), M_TEMP, M_WAITOK | M_ZERO);
6451 	(*ap->a_cookies)[*ap->a_ncookies] = off;
6452 	*ap->a_ncookies += 1;
6453 	return (0);
6454 }
6455 
6456 /*
6457  * The purpose of this routine is to remove granularity from accmode_t,
6458  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6459  * VADMIN and VAPPEND.
6460  *
6461  * If it returns 0, the caller is supposed to continue with the usual
6462  * access checks using 'accmode' as modified by this routine.  If it
6463  * returns nonzero value, the caller is supposed to return that value
6464  * as errno.
6465  *
6466  * Note that after this routine runs, accmode may be zero.
6467  */
6468 int
6469 vfs_unixify_accmode(accmode_t *accmode)
6470 {
6471 	/*
6472 	 * There is no way to specify explicit "deny" rule using
6473 	 * file mode or POSIX.1e ACLs.
6474 	 */
6475 	if (*accmode & VEXPLICIT_DENY) {
6476 		*accmode = 0;
6477 		return (0);
6478 	}
6479 
6480 	/*
6481 	 * None of these can be translated into usual access bits.
6482 	 * Also, the common case for NFSv4 ACLs is to not contain
6483 	 * either of these bits. Caller should check for VWRITE
6484 	 * on the containing directory instead.
6485 	 */
6486 	if (*accmode & (VDELETE_CHILD | VDELETE))
6487 		return (EPERM);
6488 
6489 	if (*accmode & VADMIN_PERMS) {
6490 		*accmode &= ~VADMIN_PERMS;
6491 		*accmode |= VADMIN;
6492 	}
6493 
6494 	/*
6495 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6496 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6497 	 */
6498 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6499 
6500 	return (0);
6501 }
6502 
6503 /*
6504  * Clear out a doomed vnode (if any) and replace it with a new one as long
6505  * as the fs is not being unmounted. Return the root vnode to the caller.
6506  */
6507 static int __noinline
6508 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6509 {
6510 	struct vnode *vp;
6511 	int error;
6512 
6513 restart:
6514 	if (mp->mnt_rootvnode != NULL) {
6515 		MNT_ILOCK(mp);
6516 		vp = mp->mnt_rootvnode;
6517 		if (vp != NULL) {
6518 			if (!VN_IS_DOOMED(vp)) {
6519 				vrefact(vp);
6520 				MNT_IUNLOCK(mp);
6521 				error = vn_lock(vp, flags);
6522 				if (error == 0) {
6523 					*vpp = vp;
6524 					return (0);
6525 				}
6526 				vrele(vp);
6527 				goto restart;
6528 			}
6529 			/*
6530 			 * Clear the old one.
6531 			 */
6532 			mp->mnt_rootvnode = NULL;
6533 		}
6534 		MNT_IUNLOCK(mp);
6535 		if (vp != NULL) {
6536 			vfs_op_barrier_wait(mp);
6537 			vrele(vp);
6538 		}
6539 	}
6540 	error = VFS_CACHEDROOT(mp, flags, vpp);
6541 	if (error != 0)
6542 		return (error);
6543 	if (mp->mnt_vfs_ops == 0) {
6544 		MNT_ILOCK(mp);
6545 		if (mp->mnt_vfs_ops != 0) {
6546 			MNT_IUNLOCK(mp);
6547 			return (0);
6548 		}
6549 		if (mp->mnt_rootvnode == NULL) {
6550 			vrefact(*vpp);
6551 			mp->mnt_rootvnode = *vpp;
6552 		} else {
6553 			if (mp->mnt_rootvnode != *vpp) {
6554 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6555 					panic("%s: mismatch between vnode returned "
6556 					    " by VFS_CACHEDROOT and the one cached "
6557 					    " (%p != %p)",
6558 					    __func__, *vpp, mp->mnt_rootvnode);
6559 				}
6560 			}
6561 		}
6562 		MNT_IUNLOCK(mp);
6563 	}
6564 	return (0);
6565 }
6566 
6567 int
6568 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6569 {
6570 	struct mount_pcpu *mpcpu;
6571 	struct vnode *vp;
6572 	int error;
6573 
6574 	if (!vfs_op_thread_enter(mp, mpcpu))
6575 		return (vfs_cache_root_fallback(mp, flags, vpp));
6576 	vp = atomic_load_ptr(&mp->mnt_rootvnode);
6577 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6578 		vfs_op_thread_exit(mp, mpcpu);
6579 		return (vfs_cache_root_fallback(mp, flags, vpp));
6580 	}
6581 	vrefact(vp);
6582 	vfs_op_thread_exit(mp, mpcpu);
6583 	error = vn_lock(vp, flags);
6584 	if (error != 0) {
6585 		vrele(vp);
6586 		return (vfs_cache_root_fallback(mp, flags, vpp));
6587 	}
6588 	*vpp = vp;
6589 	return (0);
6590 }
6591 
6592 struct vnode *
6593 vfs_cache_root_clear(struct mount *mp)
6594 {
6595 	struct vnode *vp;
6596 
6597 	/*
6598 	 * ops > 0 guarantees there is nobody who can see this vnode
6599 	 */
6600 	MPASS(mp->mnt_vfs_ops > 0);
6601 	vp = mp->mnt_rootvnode;
6602 	if (vp != NULL)
6603 		vn_seqc_write_begin(vp);
6604 	mp->mnt_rootvnode = NULL;
6605 	return (vp);
6606 }
6607 
6608 void
6609 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6610 {
6611 
6612 	MPASS(mp->mnt_vfs_ops > 0);
6613 	vrefact(vp);
6614 	mp->mnt_rootvnode = vp;
6615 }
6616 
6617 /*
6618  * These are helper functions for filesystems to traverse all
6619  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6620  *
6621  * This interface replaces MNT_VNODE_FOREACH.
6622  */
6623 
6624 struct vnode *
6625 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6626 {
6627 	struct vnode *vp;
6628 
6629 	maybe_yield();
6630 	MNT_ILOCK(mp);
6631 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6632 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6633 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6634 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6635 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6636 			continue;
6637 		VI_LOCK(vp);
6638 		if (VN_IS_DOOMED(vp)) {
6639 			VI_UNLOCK(vp);
6640 			continue;
6641 		}
6642 		break;
6643 	}
6644 	if (vp == NULL) {
6645 		__mnt_vnode_markerfree_all(mvp, mp);
6646 		/* MNT_IUNLOCK(mp); -- done in above function */
6647 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6648 		return (NULL);
6649 	}
6650 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6651 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6652 	MNT_IUNLOCK(mp);
6653 	return (vp);
6654 }
6655 
6656 struct vnode *
6657 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6658 {
6659 	struct vnode *vp;
6660 
6661 	*mvp = vn_alloc_marker(mp);
6662 	MNT_ILOCK(mp);
6663 	MNT_REF(mp);
6664 
6665 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6666 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6667 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6668 			continue;
6669 		VI_LOCK(vp);
6670 		if (VN_IS_DOOMED(vp)) {
6671 			VI_UNLOCK(vp);
6672 			continue;
6673 		}
6674 		break;
6675 	}
6676 	if (vp == NULL) {
6677 		MNT_REL(mp);
6678 		MNT_IUNLOCK(mp);
6679 		vn_free_marker(*mvp);
6680 		*mvp = NULL;
6681 		return (NULL);
6682 	}
6683 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6684 	MNT_IUNLOCK(mp);
6685 	return (vp);
6686 }
6687 
6688 void
6689 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6690 {
6691 
6692 	if (*mvp == NULL) {
6693 		MNT_IUNLOCK(mp);
6694 		return;
6695 	}
6696 
6697 	mtx_assert(MNT_MTX(mp), MA_OWNED);
6698 
6699 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6700 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6701 	MNT_REL(mp);
6702 	MNT_IUNLOCK(mp);
6703 	vn_free_marker(*mvp);
6704 	*mvp = NULL;
6705 }
6706 
6707 /*
6708  * These are helper functions for filesystems to traverse their
6709  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6710  */
6711 static void
6712 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6713 {
6714 
6715 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6716 
6717 	MNT_ILOCK(mp);
6718 	MNT_REL(mp);
6719 	MNT_IUNLOCK(mp);
6720 	vn_free_marker(*mvp);
6721 	*mvp = NULL;
6722 }
6723 
6724 /*
6725  * Relock the mp mount vnode list lock with the vp vnode interlock in the
6726  * conventional lock order during mnt_vnode_next_lazy iteration.
6727  *
6728  * On entry, the mount vnode list lock is held and the vnode interlock is not.
6729  * The list lock is dropped and reacquired.  On success, both locks are held.
6730  * On failure, the mount vnode list lock is held but the vnode interlock is
6731  * not, and the procedure may have yielded.
6732  */
6733 static bool
6734 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6735     struct vnode *vp)
6736 {
6737 
6738 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6739 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6740 	    ("%s: bad marker", __func__));
6741 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6742 	    ("%s: inappropriate vnode", __func__));
6743 	ASSERT_VI_UNLOCKED(vp, __func__);
6744 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6745 
6746 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6747 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6748 
6749 	/*
6750 	 * Note we may be racing against vdrop which transitioned the hold
6751 	 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6752 	 * if we are the only user after we get the interlock we will just
6753 	 * vdrop.
6754 	 */
6755 	vhold(vp);
6756 	mtx_unlock(&mp->mnt_listmtx);
6757 	VI_LOCK(vp);
6758 	if (VN_IS_DOOMED(vp)) {
6759 		VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6760 		goto out_lost;
6761 	}
6762 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6763 	/*
6764 	 * There is nothing to do if we are the last user.
6765 	 */
6766 	if (!refcount_release_if_not_last(&vp->v_holdcnt))
6767 		goto out_lost;
6768 	mtx_lock(&mp->mnt_listmtx);
6769 	return (true);
6770 out_lost:
6771 	vdropl(vp);
6772 	maybe_yield();
6773 	mtx_lock(&mp->mnt_listmtx);
6774 	return (false);
6775 }
6776 
6777 static struct vnode *
6778 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6779     void *cbarg)
6780 {
6781 	struct vnode *vp;
6782 
6783 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6784 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6785 restart:
6786 	vp = TAILQ_NEXT(*mvp, v_lazylist);
6787 	while (vp != NULL) {
6788 		if (vp->v_type == VMARKER) {
6789 			vp = TAILQ_NEXT(vp, v_lazylist);
6790 			continue;
6791 		}
6792 		/*
6793 		 * See if we want to process the vnode. Note we may encounter a
6794 		 * long string of vnodes we don't care about and hog the list
6795 		 * as a result. Check for it and requeue the marker.
6796 		 */
6797 		VNPASS(!VN_IS_DOOMED(vp), vp);
6798 		if (!cb(vp, cbarg)) {
6799 			if (!should_yield()) {
6800 				vp = TAILQ_NEXT(vp, v_lazylist);
6801 				continue;
6802 			}
6803 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6804 			    v_lazylist);
6805 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6806 			    v_lazylist);
6807 			mtx_unlock(&mp->mnt_listmtx);
6808 			kern_yield(PRI_USER);
6809 			mtx_lock(&mp->mnt_listmtx);
6810 			goto restart;
6811 		}
6812 		/*
6813 		 * Try-lock because this is the wrong lock order.
6814 		 */
6815 		if (!VI_TRYLOCK(vp) &&
6816 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6817 			goto restart;
6818 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6819 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6820 		    ("alien vnode on the lazy list %p %p", vp, mp));
6821 		VNPASS(vp->v_mount == mp, vp);
6822 		VNPASS(!VN_IS_DOOMED(vp), vp);
6823 		break;
6824 	}
6825 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6826 
6827 	/* Check if we are done */
6828 	if (vp == NULL) {
6829 		mtx_unlock(&mp->mnt_listmtx);
6830 		mnt_vnode_markerfree_lazy(mvp, mp);
6831 		return (NULL);
6832 	}
6833 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6834 	mtx_unlock(&mp->mnt_listmtx);
6835 	ASSERT_VI_LOCKED(vp, "lazy iter");
6836 	return (vp);
6837 }
6838 
6839 struct vnode *
6840 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6841     void *cbarg)
6842 {
6843 
6844 	maybe_yield();
6845 	mtx_lock(&mp->mnt_listmtx);
6846 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6847 }
6848 
6849 struct vnode *
6850 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6851     void *cbarg)
6852 {
6853 	struct vnode *vp;
6854 
6855 	if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6856 		return (NULL);
6857 
6858 	*mvp = vn_alloc_marker(mp);
6859 	MNT_ILOCK(mp);
6860 	MNT_REF(mp);
6861 	MNT_IUNLOCK(mp);
6862 
6863 	mtx_lock(&mp->mnt_listmtx);
6864 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6865 	if (vp == NULL) {
6866 		mtx_unlock(&mp->mnt_listmtx);
6867 		mnt_vnode_markerfree_lazy(mvp, mp);
6868 		return (NULL);
6869 	}
6870 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6871 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6872 }
6873 
6874 void
6875 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6876 {
6877 
6878 	if (*mvp == NULL)
6879 		return;
6880 
6881 	mtx_lock(&mp->mnt_listmtx);
6882 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6883 	mtx_unlock(&mp->mnt_listmtx);
6884 	mnt_vnode_markerfree_lazy(mvp, mp);
6885 }
6886 
6887 int
6888 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6889 {
6890 
6891 	if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6892 		cnp->cn_flags &= ~NOEXECCHECK;
6893 		return (0);
6894 	}
6895 
6896 	return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, curthread));
6897 }
6898 
6899 /*
6900  * Do not use this variant unless you have means other than the hold count
6901  * to prevent the vnode from getting freed.
6902  */
6903 void
6904 vn_seqc_write_begin_locked(struct vnode *vp)
6905 {
6906 
6907 	ASSERT_VI_LOCKED(vp, __func__);
6908 	VNPASS(vp->v_holdcnt > 0, vp);
6909 	VNPASS(vp->v_seqc_users >= 0, vp);
6910 	vp->v_seqc_users++;
6911 	if (vp->v_seqc_users == 1)
6912 		seqc_sleepable_write_begin(&vp->v_seqc);
6913 }
6914 
6915 void
6916 vn_seqc_write_begin(struct vnode *vp)
6917 {
6918 
6919 	VI_LOCK(vp);
6920 	vn_seqc_write_begin_locked(vp);
6921 	VI_UNLOCK(vp);
6922 }
6923 
6924 void
6925 vn_seqc_write_end_locked(struct vnode *vp)
6926 {
6927 
6928 	ASSERT_VI_LOCKED(vp, __func__);
6929 	VNPASS(vp->v_seqc_users > 0, vp);
6930 	vp->v_seqc_users--;
6931 	if (vp->v_seqc_users == 0)
6932 		seqc_sleepable_write_end(&vp->v_seqc);
6933 }
6934 
6935 void
6936 vn_seqc_write_end(struct vnode *vp)
6937 {
6938 
6939 	VI_LOCK(vp);
6940 	vn_seqc_write_end_locked(vp);
6941 	VI_UNLOCK(vp);
6942 }
6943 
6944 /*
6945  * Special case handling for allocating and freeing vnodes.
6946  *
6947  * The counter remains unchanged on free so that a doomed vnode will
6948  * keep testing as in modify as long as it is accessible with SMR.
6949  */
6950 static void
6951 vn_seqc_init(struct vnode *vp)
6952 {
6953 
6954 	vp->v_seqc = 0;
6955 	vp->v_seqc_users = 0;
6956 }
6957 
6958 static void
6959 vn_seqc_write_end_free(struct vnode *vp)
6960 {
6961 
6962 	VNPASS(seqc_in_modify(vp->v_seqc), vp);
6963 	VNPASS(vp->v_seqc_users == 1, vp);
6964 }
6965 
6966 void
6967 vn_irflag_set_locked(struct vnode *vp, short toset)
6968 {
6969 	short flags;
6970 
6971 	ASSERT_VI_LOCKED(vp, __func__);
6972 	flags = vn_irflag_read(vp);
6973 	VNASSERT((flags & toset) == 0, vp,
6974 	    ("%s: some of the passed flags already set (have %d, passed %d)\n",
6975 	    __func__, flags, toset));
6976 	atomic_store_short(&vp->v_irflag, flags | toset);
6977 }
6978 
6979 void
6980 vn_irflag_set(struct vnode *vp, short toset)
6981 {
6982 
6983 	VI_LOCK(vp);
6984 	vn_irflag_set_locked(vp, toset);
6985 	VI_UNLOCK(vp);
6986 }
6987 
6988 void
6989 vn_irflag_set_cond_locked(struct vnode *vp, short toset)
6990 {
6991 	short flags;
6992 
6993 	ASSERT_VI_LOCKED(vp, __func__);
6994 	flags = vn_irflag_read(vp);
6995 	atomic_store_short(&vp->v_irflag, flags | toset);
6996 }
6997 
6998 void
6999 vn_irflag_set_cond(struct vnode *vp, short toset)
7000 {
7001 
7002 	VI_LOCK(vp);
7003 	vn_irflag_set_cond_locked(vp, toset);
7004 	VI_UNLOCK(vp);
7005 }
7006 
7007 void
7008 vn_irflag_unset_locked(struct vnode *vp, short tounset)
7009 {
7010 	short flags;
7011 
7012 	ASSERT_VI_LOCKED(vp, __func__);
7013 	flags = vn_irflag_read(vp);
7014 	VNASSERT((flags & tounset) == tounset, vp,
7015 	    ("%s: some of the passed flags not set (have %d, passed %d)\n",
7016 	    __func__, flags, tounset));
7017 	atomic_store_short(&vp->v_irflag, flags & ~tounset);
7018 }
7019 
7020 void
7021 vn_irflag_unset(struct vnode *vp, short tounset)
7022 {
7023 
7024 	VI_LOCK(vp);
7025 	vn_irflag_unset_locked(vp, tounset);
7026 	VI_UNLOCK(vp);
7027 }
7028 
7029 int
7030 vn_getsize_locked(struct vnode *vp, off_t *size, struct ucred *cred)
7031 {
7032 	struct vattr vattr;
7033 	int error;
7034 
7035 	ASSERT_VOP_LOCKED(vp, __func__);
7036 	error = VOP_GETATTR(vp, &vattr, cred);
7037 	if (__predict_true(error == 0)) {
7038 		if (vattr.va_size <= OFF_MAX)
7039 			*size = vattr.va_size;
7040 		else
7041 			error = EFBIG;
7042 	}
7043 	return (error);
7044 }
7045 
7046 int
7047 vn_getsize(struct vnode *vp, off_t *size, struct ucred *cred)
7048 {
7049 	int error;
7050 
7051 	VOP_LOCK(vp, LK_SHARED);
7052 	error = vn_getsize_locked(vp, size, cred);
7053 	VOP_UNLOCK(vp);
7054 	return (error);
7055 }
7056 
7057 #ifdef INVARIANTS
7058 void
7059 vn_set_state_validate(struct vnode *vp, __enum_uint8(vstate) state)
7060 {
7061 
7062 	switch (vp->v_state) {
7063 	case VSTATE_UNINITIALIZED:
7064 		switch (state) {
7065 		case VSTATE_CONSTRUCTED:
7066 		case VSTATE_DESTROYING:
7067 			return;
7068 		default:
7069 			break;
7070 		}
7071 		break;
7072 	case VSTATE_CONSTRUCTED:
7073 		ASSERT_VOP_ELOCKED(vp, __func__);
7074 		switch (state) {
7075 		case VSTATE_DESTROYING:
7076 			return;
7077 		default:
7078 			break;
7079 		}
7080 		break;
7081 	case VSTATE_DESTROYING:
7082 		ASSERT_VOP_ELOCKED(vp, __func__);
7083 		switch (state) {
7084 		case VSTATE_DEAD:
7085 			return;
7086 		default:
7087 			break;
7088 		}
7089 		break;
7090 	case VSTATE_DEAD:
7091 		switch (state) {
7092 		case VSTATE_UNINITIALIZED:
7093 			return;
7094 		default:
7095 			break;
7096 		}
7097 		break;
7098 	}
7099 
7100 	vn_printf(vp, "invalid state transition %d -> %d\n", vp->v_state, state);
7101 	panic("invalid state transition %d -> %d\n", vp->v_state, state);
7102 }
7103 #endif
7104