xref: /freebsd/sys/kern/vfs_subr.c (revision 81e0e7b9e36d6a25b3af6482811318e085537d2f)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	@(#)vfs_subr.c	8.31 (Berkeley) 5/26/95
37  */
38 
39 /*
40  * External virtual filesystem routines
41  */
42 
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
45 
46 #include "opt_ddb.h"
47 #include "opt_watchdog.h"
48 
49 #include <sys/param.h>
50 #include <sys/systm.h>
51 #include <sys/asan.h>
52 #include <sys/bio.h>
53 #include <sys/buf.h>
54 #include <sys/capsicum.h>
55 #include <sys/condvar.h>
56 #include <sys/conf.h>
57 #include <sys/counter.h>
58 #include <sys/dirent.h>
59 #include <sys/event.h>
60 #include <sys/eventhandler.h>
61 #include <sys/extattr.h>
62 #include <sys/file.h>
63 #include <sys/fcntl.h>
64 #include <sys/jail.h>
65 #include <sys/kdb.h>
66 #include <sys/kernel.h>
67 #include <sys/kthread.h>
68 #include <sys/ktr.h>
69 #include <sys/lockf.h>
70 #include <sys/malloc.h>
71 #include <sys/mount.h>
72 #include <sys/namei.h>
73 #include <sys/pctrie.h>
74 #include <sys/priv.h>
75 #include <sys/reboot.h>
76 #include <sys/refcount.h>
77 #include <sys/rwlock.h>
78 #include <sys/sched.h>
79 #include <sys/sleepqueue.h>
80 #include <sys/smr.h>
81 #include <sys/smp.h>
82 #include <sys/stat.h>
83 #include <sys/sysctl.h>
84 #include <sys/syslog.h>
85 #include <sys/vmmeter.h>
86 #include <sys/vnode.h>
87 #include <sys/watchdog.h>
88 
89 #include <machine/stdarg.h>
90 
91 #include <security/mac/mac_framework.h>
92 
93 #include <vm/vm.h>
94 #include <vm/vm_object.h>
95 #include <vm/vm_extern.h>
96 #include <vm/pmap.h>
97 #include <vm/vm_map.h>
98 #include <vm/vm_page.h>
99 #include <vm/vm_kern.h>
100 #include <vm/uma.h>
101 
102 #ifdef DDB
103 #include <ddb/ddb.h>
104 #endif
105 
106 static void	delmntque(struct vnode *vp);
107 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
108 		    int slpflag, int slptimeo);
109 static void	syncer_shutdown(void *arg, int howto);
110 static int	vtryrecycle(struct vnode *vp);
111 static void	v_init_counters(struct vnode *);
112 static void	vn_seqc_init(struct vnode *);
113 static void	vn_seqc_write_end_free(struct vnode *vp);
114 static void	vgonel(struct vnode *);
115 static bool	vhold_recycle_free(struct vnode *);
116 static void	vdropl_recycle(struct vnode *vp);
117 static void	vdrop_recycle(struct vnode *vp);
118 static void	vfs_knllock(void *arg);
119 static void	vfs_knlunlock(void *arg);
120 static void	vfs_knl_assert_lock(void *arg, int what);
121 static void	destroy_vpollinfo(struct vpollinfo *vi);
122 static int	v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
123 		    daddr_t startlbn, daddr_t endlbn);
124 static void	vnlru_recalc(void);
125 
126 /*
127  * Number of vnodes in existence.  Increased whenever getnewvnode()
128  * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
129  */
130 static u_long __exclusive_cache_line numvnodes;
131 
132 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
133     "Number of vnodes in existence");
134 
135 static counter_u64_t vnodes_created;
136 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
137     "Number of vnodes created by getnewvnode");
138 
139 /*
140  * Conversion tables for conversion from vnode types to inode formats
141  * and back.
142  */
143 enum vtype iftovt_tab[16] = {
144 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
145 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
146 };
147 int vttoif_tab[10] = {
148 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
149 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
150 };
151 
152 /*
153  * List of allocates vnodes in the system.
154  */
155 static TAILQ_HEAD(freelst, vnode) vnode_list;
156 static struct vnode *vnode_list_free_marker;
157 static struct vnode *vnode_list_reclaim_marker;
158 
159 /*
160  * "Free" vnode target.  Free vnodes are rarely completely free, but are
161  * just ones that are cheap to recycle.  Usually they are for files which
162  * have been stat'd but not read; these usually have inode and namecache
163  * data attached to them.  This target is the preferred minimum size of a
164  * sub-cache consisting mostly of such files. The system balances the size
165  * of this sub-cache with its complement to try to prevent either from
166  * thrashing while the other is relatively inactive.  The targets express
167  * a preference for the best balance.
168  *
169  * "Above" this target there are 2 further targets (watermarks) related
170  * to recyling of free vnodes.  In the best-operating case, the cache is
171  * exactly full, the free list has size between vlowat and vhiwat above the
172  * free target, and recycling from it and normal use maintains this state.
173  * Sometimes the free list is below vlowat or even empty, but this state
174  * is even better for immediate use provided the cache is not full.
175  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
176  * ones) to reach one of these states.  The watermarks are currently hard-
177  * coded as 4% and 9% of the available space higher.  These and the default
178  * of 25% for wantfreevnodes are too large if the memory size is large.
179  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
180  * whenever vnlru_proc() becomes active.
181  */
182 static long wantfreevnodes;
183 static long __exclusive_cache_line freevnodes;
184 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
185     &freevnodes, 0, "Number of \"free\" vnodes");
186 static long freevnodes_old;
187 
188 static counter_u64_t recycles_count;
189 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
190     "Number of vnodes recycled to meet vnode cache targets");
191 
192 static counter_u64_t recycles_free_count;
193 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count,
194     "Number of free vnodes recycled to meet vnode cache targets");
195 
196 static counter_u64_t deferred_inact;
197 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact,
198     "Number of times inactive processing was deferred");
199 
200 /* To keep more than one thread at a time from running vfs_getnewfsid */
201 static struct mtx mntid_mtx;
202 
203 /*
204  * Lock for any access to the following:
205  *	vnode_list
206  *	numvnodes
207  *	freevnodes
208  */
209 static struct mtx __exclusive_cache_line vnode_list_mtx;
210 
211 /* Publicly exported FS */
212 struct nfs_public nfs_pub;
213 
214 static uma_zone_t buf_trie_zone;
215 static smr_t buf_trie_smr;
216 
217 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
218 static uma_zone_t vnode_zone;
219 MALLOC_DEFINE(M_VNODEPOLL, "VN POLL", "vnode poll");
220 
221 __read_frequently smr_t vfs_smr;
222 
223 /*
224  * The workitem queue.
225  *
226  * It is useful to delay writes of file data and filesystem metadata
227  * for tens of seconds so that quickly created and deleted files need
228  * not waste disk bandwidth being created and removed. To realize this,
229  * we append vnodes to a "workitem" queue. When running with a soft
230  * updates implementation, most pending metadata dependencies should
231  * not wait for more than a few seconds. Thus, mounted on block devices
232  * are delayed only about a half the time that file data is delayed.
233  * Similarly, directory updates are more critical, so are only delayed
234  * about a third the time that file data is delayed. Thus, there are
235  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
236  * one each second (driven off the filesystem syncer process). The
237  * syncer_delayno variable indicates the next queue that is to be processed.
238  * Items that need to be processed soon are placed in this queue:
239  *
240  *	syncer_workitem_pending[syncer_delayno]
241  *
242  * A delay of fifteen seconds is done by placing the request fifteen
243  * entries later in the queue:
244  *
245  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
246  *
247  */
248 static int syncer_delayno;
249 static long syncer_mask;
250 LIST_HEAD(synclist, bufobj);
251 static struct synclist *syncer_workitem_pending;
252 /*
253  * The sync_mtx protects:
254  *	bo->bo_synclist
255  *	sync_vnode_count
256  *	syncer_delayno
257  *	syncer_state
258  *	syncer_workitem_pending
259  *	syncer_worklist_len
260  *	rushjob
261  */
262 static struct mtx sync_mtx;
263 static struct cv sync_wakeup;
264 
265 #define SYNCER_MAXDELAY		32
266 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
267 static int syncdelay = 30;		/* max time to delay syncing data */
268 static int filedelay = 30;		/* time to delay syncing files */
269 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
270     "Time to delay syncing files (in seconds)");
271 static int dirdelay = 29;		/* time to delay syncing directories */
272 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
273     "Time to delay syncing directories (in seconds)");
274 static int metadelay = 28;		/* time to delay syncing metadata */
275 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
276     "Time to delay syncing metadata (in seconds)");
277 static int rushjob;		/* number of slots to run ASAP */
278 static int stat_rush_requests;	/* number of times I/O speeded up */
279 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
280     "Number of times I/O speeded up (rush requests)");
281 
282 #define	VDBATCH_SIZE 8
283 struct vdbatch {
284 	u_int index;
285 	long freevnodes;
286 	struct mtx lock;
287 	struct vnode *tab[VDBATCH_SIZE];
288 };
289 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
290 
291 static void	vdbatch_dequeue(struct vnode *vp);
292 
293 /*
294  * When shutting down the syncer, run it at four times normal speed.
295  */
296 #define SYNCER_SHUTDOWN_SPEEDUP		4
297 static int sync_vnode_count;
298 static int syncer_worklist_len;
299 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
300     syncer_state;
301 
302 /* Target for maximum number of vnodes. */
303 u_long desiredvnodes;
304 static u_long gapvnodes;		/* gap between wanted and desired */
305 static u_long vhiwat;		/* enough extras after expansion */
306 static u_long vlowat;		/* minimal extras before expansion */
307 static u_long vstir;		/* nonzero to stir non-free vnodes */
308 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
309 
310 static u_long vnlru_read_freevnodes(void);
311 
312 /*
313  * Note that no attempt is made to sanitize these parameters.
314  */
315 static int
316 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
317 {
318 	u_long val;
319 	int error;
320 
321 	val = desiredvnodes;
322 	error = sysctl_handle_long(oidp, &val, 0, req);
323 	if (error != 0 || req->newptr == NULL)
324 		return (error);
325 
326 	if (val == desiredvnodes)
327 		return (0);
328 	mtx_lock(&vnode_list_mtx);
329 	desiredvnodes = val;
330 	wantfreevnodes = desiredvnodes / 4;
331 	vnlru_recalc();
332 	mtx_unlock(&vnode_list_mtx);
333 	/*
334 	 * XXX There is no protection against multiple threads changing
335 	 * desiredvnodes at the same time. Locking above only helps vnlru and
336 	 * getnewvnode.
337 	 */
338 	vfs_hash_changesize(desiredvnodes);
339 	cache_changesize(desiredvnodes);
340 	return (0);
341 }
342 
343 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
344     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
345     "LU", "Target for maximum number of vnodes");
346 
347 static int
348 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
349 {
350 	u_long val;
351 	int error;
352 
353 	val = wantfreevnodes;
354 	error = sysctl_handle_long(oidp, &val, 0, req);
355 	if (error != 0 || req->newptr == NULL)
356 		return (error);
357 
358 	if (val == wantfreevnodes)
359 		return (0);
360 	mtx_lock(&vnode_list_mtx);
361 	wantfreevnodes = val;
362 	vnlru_recalc();
363 	mtx_unlock(&vnode_list_mtx);
364 	return (0);
365 }
366 
367 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
368     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
369     "LU", "Target for minimum number of \"free\" vnodes");
370 
371 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
372     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
373 static int vnlru_nowhere;
374 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
375     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
376 
377 static int
378 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
379 {
380 	struct vnode *vp;
381 	struct nameidata nd;
382 	char *buf;
383 	unsigned long ndflags;
384 	int error;
385 
386 	if (req->newptr == NULL)
387 		return (EINVAL);
388 	if (req->newlen >= PATH_MAX)
389 		return (E2BIG);
390 
391 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
392 	error = SYSCTL_IN(req, buf, req->newlen);
393 	if (error != 0)
394 		goto out;
395 
396 	buf[req->newlen] = '\0';
397 
398 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | SAVENAME;
399 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf);
400 	if ((error = namei(&nd)) != 0)
401 		goto out;
402 	vp = nd.ni_vp;
403 
404 	if (VN_IS_DOOMED(vp)) {
405 		/*
406 		 * This vnode is being recycled.  Return != 0 to let the caller
407 		 * know that the sysctl had no effect.  Return EAGAIN because a
408 		 * subsequent call will likely succeed (since namei will create
409 		 * a new vnode if necessary)
410 		 */
411 		error = EAGAIN;
412 		goto putvnode;
413 	}
414 
415 	counter_u64_add(recycles_count, 1);
416 	vgone(vp);
417 putvnode:
418 	NDFREE(&nd, 0);
419 out:
420 	free(buf, M_TEMP);
421 	return (error);
422 }
423 
424 static int
425 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
426 {
427 	struct thread *td = curthread;
428 	struct vnode *vp;
429 	struct file *fp;
430 	int error;
431 	int fd;
432 
433 	if (req->newptr == NULL)
434 		return (EBADF);
435 
436         error = sysctl_handle_int(oidp, &fd, 0, req);
437         if (error != 0)
438                 return (error);
439 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
440 	if (error != 0)
441 		return (error);
442 	vp = fp->f_vnode;
443 
444 	error = vn_lock(vp, LK_EXCLUSIVE);
445 	if (error != 0)
446 		goto drop;
447 
448 	counter_u64_add(recycles_count, 1);
449 	vgone(vp);
450 	VOP_UNLOCK(vp);
451 drop:
452 	fdrop(fp, td);
453 	return (error);
454 }
455 
456 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
457     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
458     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
459 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
460     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
461     sysctl_ftry_reclaim_vnode, "I",
462     "Try to reclaim a vnode by its file descriptor");
463 
464 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
465 #define vnsz2log 8
466 #ifndef DEBUG_LOCKS
467 _Static_assert(sizeof(struct vnode) >= 1UL << vnsz2log &&
468     sizeof(struct vnode) < 1UL << (vnsz2log + 1),
469     "vnsz2log needs to be updated");
470 #endif
471 
472 /*
473  * Support for the bufobj clean & dirty pctrie.
474  */
475 static void *
476 buf_trie_alloc(struct pctrie *ptree)
477 {
478 	return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT));
479 }
480 
481 static void
482 buf_trie_free(struct pctrie *ptree, void *node)
483 {
484 	uma_zfree_smr(buf_trie_zone, node);
485 }
486 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free,
487     buf_trie_smr);
488 
489 /*
490  * Initialize the vnode management data structures.
491  *
492  * Reevaluate the following cap on the number of vnodes after the physical
493  * memory size exceeds 512GB.  In the limit, as the physical memory size
494  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
495  */
496 #ifndef	MAXVNODES_MAX
497 #define	MAXVNODES_MAX	(512UL * 1024 * 1024 / 64)	/* 8M */
498 #endif
499 
500 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
501 
502 static struct vnode *
503 vn_alloc_marker(struct mount *mp)
504 {
505 	struct vnode *vp;
506 
507 	vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
508 	vp->v_type = VMARKER;
509 	vp->v_mount = mp;
510 
511 	return (vp);
512 }
513 
514 static void
515 vn_free_marker(struct vnode *vp)
516 {
517 
518 	MPASS(vp->v_type == VMARKER);
519 	free(vp, M_VNODE_MARKER);
520 }
521 
522 #ifdef KASAN
523 static int
524 vnode_ctor(void *mem, int size, void *arg __unused, int flags __unused)
525 {
526 	kasan_mark(mem, size, roundup2(size, UMA_ALIGN_PTR + 1), 0);
527 	return (0);
528 }
529 
530 static void
531 vnode_dtor(void *mem, int size, void *arg __unused)
532 {
533 	size_t end1, end2, off1, off2;
534 
535 	_Static_assert(offsetof(struct vnode, v_vnodelist) <
536 	    offsetof(struct vnode, v_dbatchcpu),
537 	    "KASAN marks require updating");
538 
539 	off1 = offsetof(struct vnode, v_vnodelist);
540 	off2 = offsetof(struct vnode, v_dbatchcpu);
541 	end1 = off1 + sizeof(((struct vnode *)NULL)->v_vnodelist);
542 	end2 = off2 + sizeof(((struct vnode *)NULL)->v_dbatchcpu);
543 
544 	/*
545 	 * Access to the v_vnodelist and v_dbatchcpu fields are permitted even
546 	 * after the vnode has been freed.  Try to get some KASAN coverage by
547 	 * marking everything except those two fields as invalid.  Because
548 	 * KASAN's tracking is not byte-granular, any preceding fields sharing
549 	 * the same 8-byte aligned word must also be marked valid.
550 	 */
551 
552 	/* Handle the area from the start until v_vnodelist... */
553 	off1 = rounddown2(off1, KASAN_SHADOW_SCALE);
554 	kasan_mark(mem, off1, off1, KASAN_UMA_FREED);
555 
556 	/* ... then the area between v_vnodelist and v_dbatchcpu ... */
557 	off1 = roundup2(end1, KASAN_SHADOW_SCALE);
558 	off2 = rounddown2(off2, KASAN_SHADOW_SCALE);
559 	if (off2 > off1)
560 		kasan_mark((void *)((char *)mem + off1), off2 - off1,
561 		    off2 - off1, KASAN_UMA_FREED);
562 
563 	/* ... and finally the area from v_dbatchcpu to the end. */
564 	off2 = roundup2(end2, KASAN_SHADOW_SCALE);
565 	kasan_mark((void *)((char *)mem + off2), size - off2, size - off2,
566 	    KASAN_UMA_FREED);
567 }
568 #endif /* KASAN */
569 
570 /*
571  * Initialize a vnode as it first enters the zone.
572  */
573 static int
574 vnode_init(void *mem, int size, int flags)
575 {
576 	struct vnode *vp;
577 
578 	vp = mem;
579 	bzero(vp, size);
580 	/*
581 	 * Setup locks.
582 	 */
583 	vp->v_vnlock = &vp->v_lock;
584 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
585 	/*
586 	 * By default, don't allow shared locks unless filesystems opt-in.
587 	 */
588 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
589 	    LK_NOSHARE | LK_IS_VNODE);
590 	/*
591 	 * Initialize bufobj.
592 	 */
593 	bufobj_init(&vp->v_bufobj, vp);
594 	/*
595 	 * Initialize namecache.
596 	 */
597 	cache_vnode_init(vp);
598 	/*
599 	 * Initialize rangelocks.
600 	 */
601 	rangelock_init(&vp->v_rl);
602 
603 	vp->v_dbatchcpu = NOCPU;
604 
605 	/*
606 	 * Check vhold_recycle_free for an explanation.
607 	 */
608 	vp->v_holdcnt = VHOLD_NO_SMR;
609 	vp->v_type = VNON;
610 	mtx_lock(&vnode_list_mtx);
611 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
612 	mtx_unlock(&vnode_list_mtx);
613 	return (0);
614 }
615 
616 /*
617  * Free a vnode when it is cleared from the zone.
618  */
619 static void
620 vnode_fini(void *mem, int size)
621 {
622 	struct vnode *vp;
623 	struct bufobj *bo;
624 
625 	vp = mem;
626 	vdbatch_dequeue(vp);
627 	mtx_lock(&vnode_list_mtx);
628 	TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
629 	mtx_unlock(&vnode_list_mtx);
630 	rangelock_destroy(&vp->v_rl);
631 	lockdestroy(vp->v_vnlock);
632 	mtx_destroy(&vp->v_interlock);
633 	bo = &vp->v_bufobj;
634 	rw_destroy(BO_LOCKPTR(bo));
635 
636 	kasan_mark(mem, size, size, 0);
637 }
638 
639 /*
640  * Provide the size of NFS nclnode and NFS fh for calculation of the
641  * vnode memory consumption.  The size is specified directly to
642  * eliminate dependency on NFS-private header.
643  *
644  * Other filesystems may use bigger or smaller (like UFS and ZFS)
645  * private inode data, but the NFS-based estimation is ample enough.
646  * Still, we care about differences in the size between 64- and 32-bit
647  * platforms.
648  *
649  * Namecache structure size is heuristically
650  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
651  */
652 #ifdef _LP64
653 #define	NFS_NCLNODE_SZ	(528 + 64)
654 #define	NC_SZ		148
655 #else
656 #define	NFS_NCLNODE_SZ	(360 + 32)
657 #define	NC_SZ		92
658 #endif
659 
660 static void
661 vntblinit(void *dummy __unused)
662 {
663 	struct vdbatch *vd;
664 	uma_ctor ctor;
665 	uma_dtor dtor;
666 	int cpu, physvnodes, virtvnodes;
667 
668 	/*
669 	 * Desiredvnodes is a function of the physical memory size and the
670 	 * kernel's heap size.  Generally speaking, it scales with the
671 	 * physical memory size.  The ratio of desiredvnodes to the physical
672 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
673 	 * Thereafter, the
674 	 * marginal ratio of desiredvnodes to the physical memory size is
675 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
676 	 * size.  The memory required by desiredvnodes vnodes and vm objects
677 	 * must not exceed 1/10th of the kernel's heap size.
678 	 */
679 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
680 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
681 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
682 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
683 	desiredvnodes = min(physvnodes, virtvnodes);
684 	if (desiredvnodes > MAXVNODES_MAX) {
685 		if (bootverbose)
686 			printf("Reducing kern.maxvnodes %lu -> %lu\n",
687 			    desiredvnodes, MAXVNODES_MAX);
688 		desiredvnodes = MAXVNODES_MAX;
689 	}
690 	wantfreevnodes = desiredvnodes / 4;
691 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
692 	TAILQ_INIT(&vnode_list);
693 	mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
694 	/*
695 	 * The lock is taken to appease WITNESS.
696 	 */
697 	mtx_lock(&vnode_list_mtx);
698 	vnlru_recalc();
699 	mtx_unlock(&vnode_list_mtx);
700 	vnode_list_free_marker = vn_alloc_marker(NULL);
701 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
702 	vnode_list_reclaim_marker = vn_alloc_marker(NULL);
703 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
704 
705 #ifdef KASAN
706 	ctor = vnode_ctor;
707 	dtor = vnode_dtor;
708 #else
709 	ctor = NULL;
710 	dtor = NULL;
711 #endif
712 	vnode_zone = uma_zcreate("VNODE", sizeof(struct vnode), ctor, dtor,
713 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, UMA_ZONE_NOKASAN);
714 	uma_zone_set_smr(vnode_zone, vfs_smr);
715 
716 	/*
717 	 * Preallocate enough nodes to support one-per buf so that
718 	 * we can not fail an insert.  reassignbuf() callers can not
719 	 * tolerate the insertion failure.
720 	 */
721 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
722 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
723 	    UMA_ZONE_NOFREE | UMA_ZONE_SMR);
724 	buf_trie_smr = uma_zone_get_smr(buf_trie_zone);
725 	uma_prealloc(buf_trie_zone, nbuf);
726 
727 	vnodes_created = counter_u64_alloc(M_WAITOK);
728 	recycles_count = counter_u64_alloc(M_WAITOK);
729 	recycles_free_count = counter_u64_alloc(M_WAITOK);
730 	deferred_inact = counter_u64_alloc(M_WAITOK);
731 
732 	/*
733 	 * Initialize the filesystem syncer.
734 	 */
735 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
736 	    &syncer_mask);
737 	syncer_maxdelay = syncer_mask + 1;
738 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
739 	cv_init(&sync_wakeup, "syncer");
740 
741 	CPU_FOREACH(cpu) {
742 		vd = DPCPU_ID_PTR((cpu), vd);
743 		bzero(vd, sizeof(*vd));
744 		mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
745 	}
746 }
747 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
748 
749 /*
750  * Mark a mount point as busy. Used to synchronize access and to delay
751  * unmounting. Eventually, mountlist_mtx is not released on failure.
752  *
753  * vfs_busy() is a custom lock, it can block the caller.
754  * vfs_busy() only sleeps if the unmount is active on the mount point.
755  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
756  * vnode belonging to mp.
757  *
758  * Lookup uses vfs_busy() to traverse mount points.
759  * root fs			var fs
760  * / vnode lock		A	/ vnode lock (/var)		D
761  * /var vnode lock	B	/log vnode lock(/var/log)	E
762  * vfs_busy lock	C	vfs_busy lock			F
763  *
764  * Within each file system, the lock order is C->A->B and F->D->E.
765  *
766  * When traversing across mounts, the system follows that lock order:
767  *
768  *        C->A->B
769  *              |
770  *              +->F->D->E
771  *
772  * The lookup() process for namei("/var") illustrates the process:
773  *  VOP_LOOKUP() obtains B while A is held
774  *  vfs_busy() obtains a shared lock on F while A and B are held
775  *  vput() releases lock on B
776  *  vput() releases lock on A
777  *  VFS_ROOT() obtains lock on D while shared lock on F is held
778  *  vfs_unbusy() releases shared lock on F
779  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
780  *    Attempt to lock A (instead of vp_crossmp) while D is held would
781  *    violate the global order, causing deadlocks.
782  *
783  * dounmount() locks B while F is drained.
784  */
785 int
786 vfs_busy(struct mount *mp, int flags)
787 {
788 	struct mount_pcpu *mpcpu;
789 
790 	MPASS((flags & ~MBF_MASK) == 0);
791 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
792 
793 	if (vfs_op_thread_enter(mp, mpcpu)) {
794 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
795 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
796 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
797 		vfs_mp_count_add_pcpu(mpcpu, ref, 1);
798 		vfs_mp_count_add_pcpu(mpcpu, lockref, 1);
799 		vfs_op_thread_exit(mp, mpcpu);
800 		if (flags & MBF_MNTLSTLOCK)
801 			mtx_unlock(&mountlist_mtx);
802 		return (0);
803 	}
804 
805 	MNT_ILOCK(mp);
806 	vfs_assert_mount_counters(mp);
807 	MNT_REF(mp);
808 	/*
809 	 * If mount point is currently being unmounted, sleep until the
810 	 * mount point fate is decided.  If thread doing the unmounting fails,
811 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
812 	 * that this mount point has survived the unmount attempt and vfs_busy
813 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
814 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
815 	 * about to be really destroyed.  vfs_busy needs to release its
816 	 * reference on the mount point in this case and return with ENOENT,
817 	 * telling the caller that mount mount it tried to busy is no longer
818 	 * valid.
819 	 */
820 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
821 		KASSERT(TAILQ_EMPTY(&mp->mnt_uppers),
822 		    ("%s: non-empty upper mount list with pending unmount",
823 		    __func__));
824 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
825 			MNT_REL(mp);
826 			MNT_IUNLOCK(mp);
827 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
828 			    __func__);
829 			return (ENOENT);
830 		}
831 		if (flags & MBF_MNTLSTLOCK)
832 			mtx_unlock(&mountlist_mtx);
833 		mp->mnt_kern_flag |= MNTK_MWAIT;
834 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
835 		if (flags & MBF_MNTLSTLOCK)
836 			mtx_lock(&mountlist_mtx);
837 		MNT_ILOCK(mp);
838 	}
839 	if (flags & MBF_MNTLSTLOCK)
840 		mtx_unlock(&mountlist_mtx);
841 	mp->mnt_lockref++;
842 	MNT_IUNLOCK(mp);
843 	return (0);
844 }
845 
846 /*
847  * Free a busy filesystem.
848  */
849 void
850 vfs_unbusy(struct mount *mp)
851 {
852 	struct mount_pcpu *mpcpu;
853 	int c;
854 
855 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
856 
857 	if (vfs_op_thread_enter(mp, mpcpu)) {
858 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
859 		vfs_mp_count_sub_pcpu(mpcpu, lockref, 1);
860 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
861 		vfs_op_thread_exit(mp, mpcpu);
862 		return;
863 	}
864 
865 	MNT_ILOCK(mp);
866 	vfs_assert_mount_counters(mp);
867 	MNT_REL(mp);
868 	c = --mp->mnt_lockref;
869 	if (mp->mnt_vfs_ops == 0) {
870 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
871 		MNT_IUNLOCK(mp);
872 		return;
873 	}
874 	if (c < 0)
875 		vfs_dump_mount_counters(mp);
876 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
877 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
878 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
879 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
880 		wakeup(&mp->mnt_lockref);
881 	}
882 	MNT_IUNLOCK(mp);
883 }
884 
885 /*
886  * Lookup a mount point by filesystem identifier.
887  */
888 struct mount *
889 vfs_getvfs(fsid_t *fsid)
890 {
891 	struct mount *mp;
892 
893 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
894 	mtx_lock(&mountlist_mtx);
895 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
896 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
897 			vfs_ref(mp);
898 			mtx_unlock(&mountlist_mtx);
899 			return (mp);
900 		}
901 	}
902 	mtx_unlock(&mountlist_mtx);
903 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
904 	return ((struct mount *) 0);
905 }
906 
907 /*
908  * Lookup a mount point by filesystem identifier, busying it before
909  * returning.
910  *
911  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
912  * cache for popular filesystem identifiers.  The cache is lockess, using
913  * the fact that struct mount's are never freed.  In worst case we may
914  * get pointer to unmounted or even different filesystem, so we have to
915  * check what we got, and go slow way if so.
916  */
917 struct mount *
918 vfs_busyfs(fsid_t *fsid)
919 {
920 #define	FSID_CACHE_SIZE	256
921 	typedef struct mount * volatile vmp_t;
922 	static vmp_t cache[FSID_CACHE_SIZE];
923 	struct mount *mp;
924 	int error;
925 	uint32_t hash;
926 
927 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
928 	hash = fsid->val[0] ^ fsid->val[1];
929 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
930 	mp = cache[hash];
931 	if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0)
932 		goto slow;
933 	if (vfs_busy(mp, 0) != 0) {
934 		cache[hash] = NULL;
935 		goto slow;
936 	}
937 	if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0)
938 		return (mp);
939 	else
940 	    vfs_unbusy(mp);
941 
942 slow:
943 	mtx_lock(&mountlist_mtx);
944 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
945 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
946 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
947 			if (error) {
948 				cache[hash] = NULL;
949 				mtx_unlock(&mountlist_mtx);
950 				return (NULL);
951 			}
952 			cache[hash] = mp;
953 			return (mp);
954 		}
955 	}
956 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
957 	mtx_unlock(&mountlist_mtx);
958 	return ((struct mount *) 0);
959 }
960 
961 /*
962  * Check if a user can access privileged mount options.
963  */
964 int
965 vfs_suser(struct mount *mp, struct thread *td)
966 {
967 	int error;
968 
969 	if (jailed(td->td_ucred)) {
970 		/*
971 		 * If the jail of the calling thread lacks permission for
972 		 * this type of file system, deny immediately.
973 		 */
974 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
975 			return (EPERM);
976 
977 		/*
978 		 * If the file system was mounted outside the jail of the
979 		 * calling thread, deny immediately.
980 		 */
981 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
982 			return (EPERM);
983 	}
984 
985 	/*
986 	 * If file system supports delegated administration, we don't check
987 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
988 	 * by the file system itself.
989 	 * If this is not the user that did original mount, we check for
990 	 * the PRIV_VFS_MOUNT_OWNER privilege.
991 	 */
992 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
993 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
994 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
995 			return (error);
996 	}
997 	return (0);
998 }
999 
1000 /*
1001  * Get a new unique fsid.  Try to make its val[0] unique, since this value
1002  * will be used to create fake device numbers for stat().  Also try (but
1003  * not so hard) make its val[0] unique mod 2^16, since some emulators only
1004  * support 16-bit device numbers.  We end up with unique val[0]'s for the
1005  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
1006  *
1007  * Keep in mind that several mounts may be running in parallel.  Starting
1008  * the search one past where the previous search terminated is both a
1009  * micro-optimization and a defense against returning the same fsid to
1010  * different mounts.
1011  */
1012 void
1013 vfs_getnewfsid(struct mount *mp)
1014 {
1015 	static uint16_t mntid_base;
1016 	struct mount *nmp;
1017 	fsid_t tfsid;
1018 	int mtype;
1019 
1020 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
1021 	mtx_lock(&mntid_mtx);
1022 	mtype = mp->mnt_vfc->vfc_typenum;
1023 	tfsid.val[1] = mtype;
1024 	mtype = (mtype & 0xFF) << 24;
1025 	for (;;) {
1026 		tfsid.val[0] = makedev(255,
1027 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
1028 		mntid_base++;
1029 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
1030 			break;
1031 		vfs_rel(nmp);
1032 	}
1033 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
1034 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
1035 	mtx_unlock(&mntid_mtx);
1036 }
1037 
1038 /*
1039  * Knob to control the precision of file timestamps:
1040  *
1041  *   0 = seconds only; nanoseconds zeroed.
1042  *   1 = seconds and nanoseconds, accurate within 1/HZ.
1043  *   2 = seconds and nanoseconds, truncated to microseconds.
1044  * >=3 = seconds and nanoseconds, maximum precision.
1045  */
1046 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
1047 
1048 static int timestamp_precision = TSP_USEC;
1049 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
1050     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
1051     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1052     "3+: sec + ns (max. precision))");
1053 
1054 /*
1055  * Get a current timestamp.
1056  */
1057 void
1058 vfs_timestamp(struct timespec *tsp)
1059 {
1060 	struct timeval tv;
1061 
1062 	switch (timestamp_precision) {
1063 	case TSP_SEC:
1064 		tsp->tv_sec = time_second;
1065 		tsp->tv_nsec = 0;
1066 		break;
1067 	case TSP_HZ:
1068 		getnanotime(tsp);
1069 		break;
1070 	case TSP_USEC:
1071 		microtime(&tv);
1072 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1073 		break;
1074 	case TSP_NSEC:
1075 	default:
1076 		nanotime(tsp);
1077 		break;
1078 	}
1079 }
1080 
1081 /*
1082  * Set vnode attributes to VNOVAL
1083  */
1084 void
1085 vattr_null(struct vattr *vap)
1086 {
1087 
1088 	vap->va_type = VNON;
1089 	vap->va_size = VNOVAL;
1090 	vap->va_bytes = VNOVAL;
1091 	vap->va_mode = VNOVAL;
1092 	vap->va_nlink = VNOVAL;
1093 	vap->va_uid = VNOVAL;
1094 	vap->va_gid = VNOVAL;
1095 	vap->va_fsid = VNOVAL;
1096 	vap->va_fileid = VNOVAL;
1097 	vap->va_blocksize = VNOVAL;
1098 	vap->va_rdev = VNOVAL;
1099 	vap->va_atime.tv_sec = VNOVAL;
1100 	vap->va_atime.tv_nsec = VNOVAL;
1101 	vap->va_mtime.tv_sec = VNOVAL;
1102 	vap->va_mtime.tv_nsec = VNOVAL;
1103 	vap->va_ctime.tv_sec = VNOVAL;
1104 	vap->va_ctime.tv_nsec = VNOVAL;
1105 	vap->va_birthtime.tv_sec = VNOVAL;
1106 	vap->va_birthtime.tv_nsec = VNOVAL;
1107 	vap->va_flags = VNOVAL;
1108 	vap->va_gen = VNOVAL;
1109 	vap->va_vaflags = 0;
1110 }
1111 
1112 /*
1113  * Try to reduce the total number of vnodes.
1114  *
1115  * This routine (and its user) are buggy in at least the following ways:
1116  * - all parameters were picked years ago when RAM sizes were significantly
1117  *   smaller
1118  * - it can pick vnodes based on pages used by the vm object, but filesystems
1119  *   like ZFS don't use it making the pick broken
1120  * - since ZFS has its own aging policy it gets partially combated by this one
1121  * - a dedicated method should be provided for filesystems to let them decide
1122  *   whether the vnode should be recycled
1123  *
1124  * This routine is called when we have too many vnodes.  It attempts
1125  * to free <count> vnodes and will potentially free vnodes that still
1126  * have VM backing store (VM backing store is typically the cause
1127  * of a vnode blowout so we want to do this).  Therefore, this operation
1128  * is not considered cheap.
1129  *
1130  * A number of conditions may prevent a vnode from being reclaimed.
1131  * the buffer cache may have references on the vnode, a directory
1132  * vnode may still have references due to the namei cache representing
1133  * underlying files, or the vnode may be in active use.   It is not
1134  * desirable to reuse such vnodes.  These conditions may cause the
1135  * number of vnodes to reach some minimum value regardless of what
1136  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
1137  *
1138  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1139  * 			 entries if this argument is strue
1140  * @param trigger	 Only reclaim vnodes with fewer than this many resident
1141  *			 pages.
1142  * @param target	 How many vnodes to reclaim.
1143  * @return		 The number of vnodes that were reclaimed.
1144  */
1145 static int
1146 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1147 {
1148 	struct vnode *vp, *mvp;
1149 	struct mount *mp;
1150 	struct vm_object *object;
1151 	u_long done;
1152 	bool retried;
1153 
1154 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1155 
1156 	retried = false;
1157 	done = 0;
1158 
1159 	mvp = vnode_list_reclaim_marker;
1160 restart:
1161 	vp = mvp;
1162 	while (done < target) {
1163 		vp = TAILQ_NEXT(vp, v_vnodelist);
1164 		if (__predict_false(vp == NULL))
1165 			break;
1166 
1167 		if (__predict_false(vp->v_type == VMARKER))
1168 			continue;
1169 
1170 		/*
1171 		 * If it's been deconstructed already, it's still
1172 		 * referenced, or it exceeds the trigger, skip it.
1173 		 * Also skip free vnodes.  We are trying to make space
1174 		 * to expand the free list, not reduce it.
1175 		 */
1176 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1177 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1178 			goto next_iter;
1179 
1180 		if (vp->v_type == VBAD || vp->v_type == VNON)
1181 			goto next_iter;
1182 
1183 		object = atomic_load_ptr(&vp->v_object);
1184 		if (object == NULL || object->resident_page_count > trigger) {
1185 			goto next_iter;
1186 		}
1187 
1188 		/*
1189 		 * Handle races against vnode allocation. Filesystems lock the
1190 		 * vnode some time after it gets returned from getnewvnode,
1191 		 * despite type and hold count being manipulated earlier.
1192 		 * Resorting to checking v_mount restores guarantees present
1193 		 * before the global list was reworked to contain all vnodes.
1194 		 */
1195 		if (!VI_TRYLOCK(vp))
1196 			goto next_iter;
1197 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1198 			VI_UNLOCK(vp);
1199 			goto next_iter;
1200 		}
1201 		if (vp->v_mount == NULL) {
1202 			VI_UNLOCK(vp);
1203 			goto next_iter;
1204 		}
1205 		vholdl(vp);
1206 		VI_UNLOCK(vp);
1207 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1208 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1209 		mtx_unlock(&vnode_list_mtx);
1210 
1211 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1212 			vdrop_recycle(vp);
1213 			goto next_iter_unlocked;
1214 		}
1215 		if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1216 			vdrop_recycle(vp);
1217 			vn_finished_write(mp);
1218 			goto next_iter_unlocked;
1219 		}
1220 
1221 		VI_LOCK(vp);
1222 		if (vp->v_usecount > 0 ||
1223 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1224 		    (vp->v_object != NULL && vp->v_object->handle == vp &&
1225 		    vp->v_object->resident_page_count > trigger)) {
1226 			VOP_UNLOCK(vp);
1227 			vdropl_recycle(vp);
1228 			vn_finished_write(mp);
1229 			goto next_iter_unlocked;
1230 		}
1231 		counter_u64_add(recycles_count, 1);
1232 		vgonel(vp);
1233 		VOP_UNLOCK(vp);
1234 		vdropl_recycle(vp);
1235 		vn_finished_write(mp);
1236 		done++;
1237 next_iter_unlocked:
1238 		if (should_yield())
1239 			kern_yield(PRI_USER);
1240 		mtx_lock(&vnode_list_mtx);
1241 		goto restart;
1242 next_iter:
1243 		MPASS(vp->v_type != VMARKER);
1244 		if (!should_yield())
1245 			continue;
1246 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1247 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1248 		mtx_unlock(&vnode_list_mtx);
1249 		kern_yield(PRI_USER);
1250 		mtx_lock(&vnode_list_mtx);
1251 		goto restart;
1252 	}
1253 	if (done == 0 && !retried) {
1254 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1255 		TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1256 		retried = true;
1257 		goto restart;
1258 	}
1259 	return (done);
1260 }
1261 
1262 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1263 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1264     0,
1265     "limit on vnode free requests per call to the vnlru_free routine");
1266 
1267 /*
1268  * Attempt to reduce the free list by the requested amount.
1269  */
1270 static int
1271 vnlru_free_impl(int count, struct vfsops *mnt_op, struct vnode *mvp)
1272 {
1273 	struct vnode *vp;
1274 	struct mount *mp;
1275 	int ocount;
1276 
1277 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1278 	if (count > max_vnlru_free)
1279 		count = max_vnlru_free;
1280 	ocount = count;
1281 	vp = mvp;
1282 	for (;;) {
1283 		if (count == 0) {
1284 			break;
1285 		}
1286 		vp = TAILQ_NEXT(vp, v_vnodelist);
1287 		if (__predict_false(vp == NULL)) {
1288 			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1289 			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1290 			break;
1291 		}
1292 		if (__predict_false(vp->v_type == VMARKER))
1293 			continue;
1294 		if (vp->v_holdcnt > 0)
1295 			continue;
1296 		/*
1297 		 * Don't recycle if our vnode is from different type
1298 		 * of mount point.  Note that mp is type-safe, the
1299 		 * check does not reach unmapped address even if
1300 		 * vnode is reclaimed.
1301 		 */
1302 		if (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1303 		    mp->mnt_op != mnt_op) {
1304 			continue;
1305 		}
1306 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1307 			continue;
1308 		}
1309 		if (!vhold_recycle_free(vp))
1310 			continue;
1311 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1312 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1313 		mtx_unlock(&vnode_list_mtx);
1314 		/*
1315 		 * FIXME: ignores the return value, meaning it may be nothing
1316 		 * got recycled but it claims otherwise to the caller.
1317 		 *
1318 		 * Originally the value started being ignored in 2005 with
1319 		 * 114a1006a8204aa156e1f9ad6476cdff89cada7f .
1320 		 *
1321 		 * Respecting the value can run into significant stalls if most
1322 		 * vnodes belong to one file system and it has writes
1323 		 * suspended.  In presence of many threads and millions of
1324 		 * vnodes they keep contending on the vnode_list_mtx lock only
1325 		 * to find vnodes they can't recycle.
1326 		 *
1327 		 * The solution would be to pre-check if the vnode is likely to
1328 		 * be recycle-able, but it needs to happen with the
1329 		 * vnode_list_mtx lock held. This runs into a problem where
1330 		 * VOP_GETWRITEMOUNT (currently needed to find out about if
1331 		 * writes are frozen) can take locks which LOR against it.
1332 		 *
1333 		 * Check nullfs for one example (null_getwritemount).
1334 		 */
1335 		vtryrecycle(vp);
1336 		count--;
1337 		mtx_lock(&vnode_list_mtx);
1338 		vp = mvp;
1339 	}
1340 	return (ocount - count);
1341 }
1342 
1343 static int
1344 vnlru_free_locked(int count)
1345 {
1346 
1347 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1348 	return (vnlru_free_impl(count, NULL, vnode_list_free_marker));
1349 }
1350 
1351 void
1352 vnlru_free_vfsops(int count, struct vfsops *mnt_op, struct vnode *mvp)
1353 {
1354 
1355 	MPASS(mnt_op != NULL);
1356 	MPASS(mvp != NULL);
1357 	VNPASS(mvp->v_type == VMARKER, mvp);
1358 	mtx_lock(&vnode_list_mtx);
1359 	vnlru_free_impl(count, mnt_op, mvp);
1360 	mtx_unlock(&vnode_list_mtx);
1361 }
1362 
1363 struct vnode *
1364 vnlru_alloc_marker(void)
1365 {
1366 	struct vnode *mvp;
1367 
1368 	mvp = vn_alloc_marker(NULL);
1369 	mtx_lock(&vnode_list_mtx);
1370 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, mvp, v_vnodelist);
1371 	mtx_unlock(&vnode_list_mtx);
1372 	return (mvp);
1373 }
1374 
1375 void
1376 vnlru_free_marker(struct vnode *mvp)
1377 {
1378 	mtx_lock(&vnode_list_mtx);
1379 	TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1380 	mtx_unlock(&vnode_list_mtx);
1381 	vn_free_marker(mvp);
1382 }
1383 
1384 static void
1385 vnlru_recalc(void)
1386 {
1387 
1388 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1389 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1390 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1391 	vlowat = vhiwat / 2;
1392 }
1393 
1394 /*
1395  * Attempt to recycle vnodes in a context that is always safe to block.
1396  * Calling vlrurecycle() from the bowels of filesystem code has some
1397  * interesting deadlock problems.
1398  */
1399 static struct proc *vnlruproc;
1400 static int vnlruproc_sig;
1401 
1402 /*
1403  * The main freevnodes counter is only updated when threads requeue their vnode
1404  * batches. CPUs are conditionally walked to compute a more accurate total.
1405  *
1406  * Limit how much of a slop are we willing to tolerate. Note: the actual value
1407  * at any given moment can still exceed slop, but it should not be by significant
1408  * margin in practice.
1409  */
1410 #define VNLRU_FREEVNODES_SLOP 128
1411 
1412 static __inline void
1413 vfs_freevnodes_inc(void)
1414 {
1415 	struct vdbatch *vd;
1416 
1417 	critical_enter();
1418 	vd = DPCPU_PTR(vd);
1419 	vd->freevnodes++;
1420 	critical_exit();
1421 }
1422 
1423 static __inline void
1424 vfs_freevnodes_dec(void)
1425 {
1426 	struct vdbatch *vd;
1427 
1428 	critical_enter();
1429 	vd = DPCPU_PTR(vd);
1430 	vd->freevnodes--;
1431 	critical_exit();
1432 }
1433 
1434 static u_long
1435 vnlru_read_freevnodes(void)
1436 {
1437 	struct vdbatch *vd;
1438 	long slop;
1439 	int cpu;
1440 
1441 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1442 	if (freevnodes > freevnodes_old)
1443 		slop = freevnodes - freevnodes_old;
1444 	else
1445 		slop = freevnodes_old - freevnodes;
1446 	if (slop < VNLRU_FREEVNODES_SLOP)
1447 		return (freevnodes >= 0 ? freevnodes : 0);
1448 	freevnodes_old = freevnodes;
1449 	CPU_FOREACH(cpu) {
1450 		vd = DPCPU_ID_PTR((cpu), vd);
1451 		freevnodes_old += vd->freevnodes;
1452 	}
1453 	return (freevnodes_old >= 0 ? freevnodes_old : 0);
1454 }
1455 
1456 static bool
1457 vnlru_under(u_long rnumvnodes, u_long limit)
1458 {
1459 	u_long rfreevnodes, space;
1460 
1461 	if (__predict_false(rnumvnodes > desiredvnodes))
1462 		return (true);
1463 
1464 	space = desiredvnodes - rnumvnodes;
1465 	if (space < limit) {
1466 		rfreevnodes = vnlru_read_freevnodes();
1467 		if (rfreevnodes > wantfreevnodes)
1468 			space += rfreevnodes - wantfreevnodes;
1469 	}
1470 	return (space < limit);
1471 }
1472 
1473 static bool
1474 vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1475 {
1476 	long rfreevnodes, space;
1477 
1478 	if (__predict_false(rnumvnodes > desiredvnodes))
1479 		return (true);
1480 
1481 	space = desiredvnodes - rnumvnodes;
1482 	if (space < limit) {
1483 		rfreevnodes = atomic_load_long(&freevnodes);
1484 		if (rfreevnodes > wantfreevnodes)
1485 			space += rfreevnodes - wantfreevnodes;
1486 	}
1487 	return (space < limit);
1488 }
1489 
1490 static void
1491 vnlru_kick(void)
1492 {
1493 
1494 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1495 	if (vnlruproc_sig == 0) {
1496 		vnlruproc_sig = 1;
1497 		wakeup(vnlruproc);
1498 	}
1499 }
1500 
1501 static void
1502 vnlru_proc(void)
1503 {
1504 	u_long rnumvnodes, rfreevnodes, target;
1505 	unsigned long onumvnodes;
1506 	int done, force, trigger, usevnodes;
1507 	bool reclaim_nc_src, want_reread;
1508 
1509 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1510 	    SHUTDOWN_PRI_FIRST);
1511 
1512 	force = 0;
1513 	want_reread = false;
1514 	for (;;) {
1515 		kproc_suspend_check(vnlruproc);
1516 		mtx_lock(&vnode_list_mtx);
1517 		rnumvnodes = atomic_load_long(&numvnodes);
1518 
1519 		if (want_reread) {
1520 			force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1521 			want_reread = false;
1522 		}
1523 
1524 		/*
1525 		 * If numvnodes is too large (due to desiredvnodes being
1526 		 * adjusted using its sysctl, or emergency growth), first
1527 		 * try to reduce it by discarding from the free list.
1528 		 */
1529 		if (rnumvnodes > desiredvnodes) {
1530 			vnlru_free_locked(rnumvnodes - desiredvnodes);
1531 			rnumvnodes = atomic_load_long(&numvnodes);
1532 		}
1533 		/*
1534 		 * Sleep if the vnode cache is in a good state.  This is
1535 		 * when it is not over-full and has space for about a 4%
1536 		 * or 9% expansion (by growing its size or inexcessively
1537 		 * reducing its free list).  Otherwise, try to reclaim
1538 		 * space for a 10% expansion.
1539 		 */
1540 		if (vstir && force == 0) {
1541 			force = 1;
1542 			vstir = 0;
1543 		}
1544 		if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1545 			vnlruproc_sig = 0;
1546 			wakeup(&vnlruproc_sig);
1547 			msleep(vnlruproc, &vnode_list_mtx,
1548 			    PVFS|PDROP, "vlruwt", hz);
1549 			continue;
1550 		}
1551 		rfreevnodes = vnlru_read_freevnodes();
1552 
1553 		onumvnodes = rnumvnodes;
1554 		/*
1555 		 * Calculate parameters for recycling.  These are the same
1556 		 * throughout the loop to give some semblance of fairness.
1557 		 * The trigger point is to avoid recycling vnodes with lots
1558 		 * of resident pages.  We aren't trying to free memory; we
1559 		 * are trying to recycle or at least free vnodes.
1560 		 */
1561 		if (rnumvnodes <= desiredvnodes)
1562 			usevnodes = rnumvnodes - rfreevnodes;
1563 		else
1564 			usevnodes = rnumvnodes;
1565 		if (usevnodes <= 0)
1566 			usevnodes = 1;
1567 		/*
1568 		 * The trigger value is is chosen to give a conservatively
1569 		 * large value to ensure that it alone doesn't prevent
1570 		 * making progress.  The value can easily be so large that
1571 		 * it is effectively infinite in some congested and
1572 		 * misconfigured cases, and this is necessary.  Normally
1573 		 * it is about 8 to 100 (pages), which is quite large.
1574 		 */
1575 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1576 		if (force < 2)
1577 			trigger = vsmalltrigger;
1578 		reclaim_nc_src = force >= 3;
1579 		target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1580 		target = target / 10 + 1;
1581 		done = vlrureclaim(reclaim_nc_src, trigger, target);
1582 		mtx_unlock(&vnode_list_mtx);
1583 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1584 			uma_reclaim(UMA_RECLAIM_DRAIN);
1585 		if (done == 0) {
1586 			if (force == 0 || force == 1) {
1587 				force = 2;
1588 				continue;
1589 			}
1590 			if (force == 2) {
1591 				force = 3;
1592 				continue;
1593 			}
1594 			want_reread = true;
1595 			force = 0;
1596 			vnlru_nowhere++;
1597 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1598 		} else {
1599 			want_reread = true;
1600 			kern_yield(PRI_USER);
1601 		}
1602 	}
1603 }
1604 
1605 static struct kproc_desc vnlru_kp = {
1606 	"vnlru",
1607 	vnlru_proc,
1608 	&vnlruproc
1609 };
1610 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1611     &vnlru_kp);
1612 
1613 /*
1614  * Routines having to do with the management of the vnode table.
1615  */
1616 
1617 /*
1618  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1619  * before we actually vgone().  This function must be called with the vnode
1620  * held to prevent the vnode from being returned to the free list midway
1621  * through vgone().
1622  */
1623 static int
1624 vtryrecycle(struct vnode *vp)
1625 {
1626 	struct mount *vnmp;
1627 
1628 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1629 	VNASSERT(vp->v_holdcnt, vp,
1630 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1631 	/*
1632 	 * This vnode may found and locked via some other list, if so we
1633 	 * can't recycle it yet.
1634 	 */
1635 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1636 		CTR2(KTR_VFS,
1637 		    "%s: impossible to recycle, vp %p lock is already held",
1638 		    __func__, vp);
1639 		vdrop_recycle(vp);
1640 		return (EWOULDBLOCK);
1641 	}
1642 	/*
1643 	 * Don't recycle if its filesystem is being suspended.
1644 	 */
1645 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1646 		VOP_UNLOCK(vp);
1647 		CTR2(KTR_VFS,
1648 		    "%s: impossible to recycle, cannot start the write for %p",
1649 		    __func__, vp);
1650 		vdrop_recycle(vp);
1651 		return (EBUSY);
1652 	}
1653 	/*
1654 	 * If we got this far, we need to acquire the interlock and see if
1655 	 * anyone picked up this vnode from another list.  If not, we will
1656 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1657 	 * will skip over it.
1658 	 */
1659 	VI_LOCK(vp);
1660 	if (vp->v_usecount) {
1661 		VOP_UNLOCK(vp);
1662 		vdropl_recycle(vp);
1663 		vn_finished_write(vnmp);
1664 		CTR2(KTR_VFS,
1665 		    "%s: impossible to recycle, %p is already referenced",
1666 		    __func__, vp);
1667 		return (EBUSY);
1668 	}
1669 	if (!VN_IS_DOOMED(vp)) {
1670 		counter_u64_add(recycles_free_count, 1);
1671 		vgonel(vp);
1672 	}
1673 	VOP_UNLOCK(vp);
1674 	vdropl_recycle(vp);
1675 	vn_finished_write(vnmp);
1676 	return (0);
1677 }
1678 
1679 /*
1680  * Allocate a new vnode.
1681  *
1682  * The operation never returns an error. Returning an error was disabled
1683  * in r145385 (dated 2005) with the following comment:
1684  *
1685  * XXX Not all VFS_VGET/ffs_vget callers check returns.
1686  *
1687  * Given the age of this commit (almost 15 years at the time of writing this
1688  * comment) restoring the ability to fail requires a significant audit of
1689  * all codepaths.
1690  *
1691  * The routine can try to free a vnode or stall for up to 1 second waiting for
1692  * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1693  */
1694 static u_long vn_alloc_cyclecount;
1695 
1696 static struct vnode * __noinline
1697 vn_alloc_hard(struct mount *mp)
1698 {
1699 	u_long rnumvnodes, rfreevnodes;
1700 
1701 	mtx_lock(&vnode_list_mtx);
1702 	rnumvnodes = atomic_load_long(&numvnodes);
1703 	if (rnumvnodes + 1 < desiredvnodes) {
1704 		vn_alloc_cyclecount = 0;
1705 		goto alloc;
1706 	}
1707 	rfreevnodes = vnlru_read_freevnodes();
1708 	if (vn_alloc_cyclecount++ >= rfreevnodes) {
1709 		vn_alloc_cyclecount = 0;
1710 		vstir = 1;
1711 	}
1712 	/*
1713 	 * Grow the vnode cache if it will not be above its target max
1714 	 * after growing.  Otherwise, if the free list is nonempty, try
1715 	 * to reclaim 1 item from it before growing the cache (possibly
1716 	 * above its target max if the reclamation failed or is delayed).
1717 	 * Otherwise, wait for some space.  In all cases, schedule
1718 	 * vnlru_proc() if we are getting short of space.  The watermarks
1719 	 * should be chosen so that we never wait or even reclaim from
1720 	 * the free list to below its target minimum.
1721 	 */
1722 	if (vnlru_free_locked(1) > 0)
1723 		goto alloc;
1724 	if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1725 		/*
1726 		 * Wait for space for a new vnode.
1727 		 */
1728 		vnlru_kick();
1729 		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1730 		if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1731 		    vnlru_read_freevnodes() > 1)
1732 			vnlru_free_locked(1);
1733 	}
1734 alloc:
1735 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1736 	if (vnlru_under(rnumvnodes, vlowat))
1737 		vnlru_kick();
1738 	mtx_unlock(&vnode_list_mtx);
1739 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1740 }
1741 
1742 static struct vnode *
1743 vn_alloc(struct mount *mp)
1744 {
1745 	u_long rnumvnodes;
1746 
1747 	if (__predict_false(vn_alloc_cyclecount != 0))
1748 		return (vn_alloc_hard(mp));
1749 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1750 	if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1751 		atomic_subtract_long(&numvnodes, 1);
1752 		return (vn_alloc_hard(mp));
1753 	}
1754 
1755 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1756 }
1757 
1758 static void
1759 vn_free(struct vnode *vp)
1760 {
1761 
1762 	atomic_subtract_long(&numvnodes, 1);
1763 	uma_zfree_smr(vnode_zone, vp);
1764 }
1765 
1766 /*
1767  * Return the next vnode from the free list.
1768  */
1769 int
1770 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1771     struct vnode **vpp)
1772 {
1773 	struct vnode *vp;
1774 	struct thread *td;
1775 	struct lock_object *lo;
1776 
1777 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1778 
1779 	KASSERT(vops->registered,
1780 	    ("%s: not registered vector op %p\n", __func__, vops));
1781 
1782 	td = curthread;
1783 	if (td->td_vp_reserved != NULL) {
1784 		vp = td->td_vp_reserved;
1785 		td->td_vp_reserved = NULL;
1786 	} else {
1787 		vp = vn_alloc(mp);
1788 	}
1789 	counter_u64_add(vnodes_created, 1);
1790 	/*
1791 	 * Locks are given the generic name "vnode" when created.
1792 	 * Follow the historic practice of using the filesystem
1793 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1794 	 *
1795 	 * Locks live in a witness group keyed on their name. Thus,
1796 	 * when a lock is renamed, it must also move from the witness
1797 	 * group of its old name to the witness group of its new name.
1798 	 *
1799 	 * The change only needs to be made when the vnode moves
1800 	 * from one filesystem type to another. We ensure that each
1801 	 * filesystem use a single static name pointer for its tag so
1802 	 * that we can compare pointers rather than doing a strcmp().
1803 	 */
1804 	lo = &vp->v_vnlock->lock_object;
1805 #ifdef WITNESS
1806 	if (lo->lo_name != tag) {
1807 #endif
1808 		lo->lo_name = tag;
1809 #ifdef WITNESS
1810 		WITNESS_DESTROY(lo);
1811 		WITNESS_INIT(lo, tag);
1812 	}
1813 #endif
1814 	/*
1815 	 * By default, don't allow shared locks unless filesystems opt-in.
1816 	 */
1817 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1818 	/*
1819 	 * Finalize various vnode identity bits.
1820 	 */
1821 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1822 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1823 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1824 	vp->v_type = VNON;
1825 	vp->v_op = vops;
1826 	vp->v_irflag = 0;
1827 	v_init_counters(vp);
1828 	vn_seqc_init(vp);
1829 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1830 #ifdef DIAGNOSTIC
1831 	if (mp == NULL && vops != &dead_vnodeops)
1832 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1833 #endif
1834 #ifdef MAC
1835 	mac_vnode_init(vp);
1836 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1837 		mac_vnode_associate_singlelabel(mp, vp);
1838 #endif
1839 	if (mp != NULL) {
1840 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1841 	}
1842 
1843 	/*
1844 	 * For the filesystems which do not use vfs_hash_insert(),
1845 	 * still initialize v_hash to have vfs_hash_index() useful.
1846 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1847 	 * its own hashing.
1848 	 */
1849 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1850 
1851 	*vpp = vp;
1852 	return (0);
1853 }
1854 
1855 void
1856 getnewvnode_reserve(void)
1857 {
1858 	struct thread *td;
1859 
1860 	td = curthread;
1861 	MPASS(td->td_vp_reserved == NULL);
1862 	td->td_vp_reserved = vn_alloc(NULL);
1863 }
1864 
1865 void
1866 getnewvnode_drop_reserve(void)
1867 {
1868 	struct thread *td;
1869 
1870 	td = curthread;
1871 	if (td->td_vp_reserved != NULL) {
1872 		vn_free(td->td_vp_reserved);
1873 		td->td_vp_reserved = NULL;
1874 	}
1875 }
1876 
1877 static void __noinline
1878 freevnode(struct vnode *vp)
1879 {
1880 	struct bufobj *bo;
1881 
1882 	/*
1883 	 * The vnode has been marked for destruction, so free it.
1884 	 *
1885 	 * The vnode will be returned to the zone where it will
1886 	 * normally remain until it is needed for another vnode. We
1887 	 * need to cleanup (or verify that the cleanup has already
1888 	 * been done) any residual data left from its current use
1889 	 * so as not to contaminate the freshly allocated vnode.
1890 	 */
1891 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1892 	/*
1893 	 * Paired with vgone.
1894 	 */
1895 	vn_seqc_write_end_free(vp);
1896 
1897 	bo = &vp->v_bufobj;
1898 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1899 	VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp);
1900 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1901 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1902 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1903 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1904 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1905 	    ("clean blk trie not empty"));
1906 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1907 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1908 	    ("dirty blk trie not empty"));
1909 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1910 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1911 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1912 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1913 	    ("Dangling rangelock waiters"));
1914 	VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp,
1915 	    ("Leaked inactivation"));
1916 	VI_UNLOCK(vp);
1917 #ifdef MAC
1918 	mac_vnode_destroy(vp);
1919 #endif
1920 	if (vp->v_pollinfo != NULL) {
1921 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1922 		destroy_vpollinfo(vp->v_pollinfo);
1923 		VOP_UNLOCK(vp);
1924 		vp->v_pollinfo = NULL;
1925 	}
1926 	vp->v_mountedhere = NULL;
1927 	vp->v_unpcb = NULL;
1928 	vp->v_rdev = NULL;
1929 	vp->v_fifoinfo = NULL;
1930 	vp->v_iflag = 0;
1931 	vp->v_vflag = 0;
1932 	bo->bo_flag = 0;
1933 	vn_free(vp);
1934 }
1935 
1936 /*
1937  * Delete from old mount point vnode list, if on one.
1938  */
1939 static void
1940 delmntque(struct vnode *vp)
1941 {
1942 	struct mount *mp;
1943 
1944 	VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
1945 
1946 	mp = vp->v_mount;
1947 	if (mp == NULL)
1948 		return;
1949 	MNT_ILOCK(mp);
1950 	VI_LOCK(vp);
1951 	vp->v_mount = NULL;
1952 	VI_UNLOCK(vp);
1953 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1954 		("bad mount point vnode list size"));
1955 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1956 	mp->mnt_nvnodelistsize--;
1957 	MNT_REL(mp);
1958 	MNT_IUNLOCK(mp);
1959 }
1960 
1961 static int
1962 insmntque1_int(struct vnode *vp, struct mount *mp, bool dtr)
1963 {
1964 
1965 	KASSERT(vp->v_mount == NULL,
1966 		("insmntque: vnode already on per mount vnode list"));
1967 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1968 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1969 
1970 	/*
1971 	 * We acquire the vnode interlock early to ensure that the
1972 	 * vnode cannot be recycled by another process releasing a
1973 	 * holdcnt on it before we get it on both the vnode list
1974 	 * and the active vnode list. The mount mutex protects only
1975 	 * manipulation of the vnode list and the vnode freelist
1976 	 * mutex protects only manipulation of the active vnode list.
1977 	 * Hence the need to hold the vnode interlock throughout.
1978 	 */
1979 	MNT_ILOCK(mp);
1980 	VI_LOCK(vp);
1981 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1982 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1983 	    mp->mnt_nvnodelistsize == 0)) &&
1984 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1985 		VI_UNLOCK(vp);
1986 		MNT_IUNLOCK(mp);
1987 		if (dtr) {
1988 			vp->v_data = NULL;
1989 			vp->v_op = &dead_vnodeops;
1990 			vgone(vp);
1991 			vput(vp);
1992 		}
1993 		return (EBUSY);
1994 	}
1995 	vp->v_mount = mp;
1996 	MNT_REF(mp);
1997 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1998 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1999 		("neg mount point vnode list size"));
2000 	mp->mnt_nvnodelistsize++;
2001 	VI_UNLOCK(vp);
2002 	MNT_IUNLOCK(mp);
2003 	return (0);
2004 }
2005 
2006 /*
2007  * Insert into list of vnodes for the new mount point, if available.
2008  * insmntque() reclaims the vnode on insertion failure, insmntque1()
2009  * leaves handling of the vnode to the caller.
2010  */
2011 int
2012 insmntque(struct vnode *vp, struct mount *mp)
2013 {
2014 	return (insmntque1_int(vp, mp, true));
2015 }
2016 
2017 int
2018 insmntque1(struct vnode *vp, struct mount *mp)
2019 {
2020 	return (insmntque1_int(vp, mp, false));
2021 }
2022 
2023 /*
2024  * Flush out and invalidate all buffers associated with a bufobj
2025  * Called with the underlying object locked.
2026  */
2027 int
2028 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
2029 {
2030 	int error;
2031 
2032 	BO_LOCK(bo);
2033 	if (flags & V_SAVE) {
2034 		error = bufobj_wwait(bo, slpflag, slptimeo);
2035 		if (error) {
2036 			BO_UNLOCK(bo);
2037 			return (error);
2038 		}
2039 		if (bo->bo_dirty.bv_cnt > 0) {
2040 			BO_UNLOCK(bo);
2041 			do {
2042 				error = BO_SYNC(bo, MNT_WAIT);
2043 			} while (error == ERELOOKUP);
2044 			if (error != 0)
2045 				return (error);
2046 			BO_LOCK(bo);
2047 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) {
2048 				BO_UNLOCK(bo);
2049 				return (EBUSY);
2050 			}
2051 		}
2052 	}
2053 	/*
2054 	 * If you alter this loop please notice that interlock is dropped and
2055 	 * reacquired in flushbuflist.  Special care is needed to ensure that
2056 	 * no race conditions occur from this.
2057 	 */
2058 	do {
2059 		error = flushbuflist(&bo->bo_clean,
2060 		    flags, bo, slpflag, slptimeo);
2061 		if (error == 0 && !(flags & V_CLEANONLY))
2062 			error = flushbuflist(&bo->bo_dirty,
2063 			    flags, bo, slpflag, slptimeo);
2064 		if (error != 0 && error != EAGAIN) {
2065 			BO_UNLOCK(bo);
2066 			return (error);
2067 		}
2068 	} while (error != 0);
2069 
2070 	/*
2071 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
2072 	 * have write I/O in-progress but if there is a VM object then the
2073 	 * VM object can also have read-I/O in-progress.
2074 	 */
2075 	do {
2076 		bufobj_wwait(bo, 0, 0);
2077 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
2078 			BO_UNLOCK(bo);
2079 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
2080 			BO_LOCK(bo);
2081 		}
2082 	} while (bo->bo_numoutput > 0);
2083 	BO_UNLOCK(bo);
2084 
2085 	/*
2086 	 * Destroy the copy in the VM cache, too.
2087 	 */
2088 	if (bo->bo_object != NULL &&
2089 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
2090 		VM_OBJECT_WLOCK(bo->bo_object);
2091 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
2092 		    OBJPR_CLEANONLY : 0);
2093 		VM_OBJECT_WUNLOCK(bo->bo_object);
2094 	}
2095 
2096 #ifdef INVARIANTS
2097 	BO_LOCK(bo);
2098 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
2099 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
2100 	    bo->bo_clean.bv_cnt > 0))
2101 		panic("vinvalbuf: flush failed");
2102 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
2103 	    bo->bo_dirty.bv_cnt > 0)
2104 		panic("vinvalbuf: flush dirty failed");
2105 	BO_UNLOCK(bo);
2106 #endif
2107 	return (0);
2108 }
2109 
2110 /*
2111  * Flush out and invalidate all buffers associated with a vnode.
2112  * Called with the underlying object locked.
2113  */
2114 int
2115 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
2116 {
2117 
2118 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2119 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
2120 	if (vp->v_object != NULL && vp->v_object->handle != vp)
2121 		return (0);
2122 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
2123 }
2124 
2125 /*
2126  * Flush out buffers on the specified list.
2127  *
2128  */
2129 static int
2130 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2131     int slptimeo)
2132 {
2133 	struct buf *bp, *nbp;
2134 	int retval, error;
2135 	daddr_t lblkno;
2136 	b_xflags_t xflags;
2137 
2138 	ASSERT_BO_WLOCKED(bo);
2139 
2140 	retval = 0;
2141 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2142 		/*
2143 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2144 		 * do not skip any buffers. If we are flushing only V_NORMAL
2145 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2146 		 * flushing only V_ALT buffers then skip buffers not marked
2147 		 * as BX_ALTDATA.
2148 		 */
2149 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2150 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2151 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2152 			continue;
2153 		}
2154 		if (nbp != NULL) {
2155 			lblkno = nbp->b_lblkno;
2156 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2157 		}
2158 		retval = EAGAIN;
2159 		error = BUF_TIMELOCK(bp,
2160 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2161 		    "flushbuf", slpflag, slptimeo);
2162 		if (error) {
2163 			BO_LOCK(bo);
2164 			return (error != ENOLCK ? error : EAGAIN);
2165 		}
2166 		KASSERT(bp->b_bufobj == bo,
2167 		    ("bp %p wrong b_bufobj %p should be %p",
2168 		    bp, bp->b_bufobj, bo));
2169 		/*
2170 		 * XXX Since there are no node locks for NFS, I
2171 		 * believe there is a slight chance that a delayed
2172 		 * write will occur while sleeping just above, so
2173 		 * check for it.
2174 		 */
2175 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2176 		    (flags & V_SAVE)) {
2177 			bremfree(bp);
2178 			bp->b_flags |= B_ASYNC;
2179 			bwrite(bp);
2180 			BO_LOCK(bo);
2181 			return (EAGAIN);	/* XXX: why not loop ? */
2182 		}
2183 		bremfree(bp);
2184 		bp->b_flags |= (B_INVAL | B_RELBUF);
2185 		bp->b_flags &= ~B_ASYNC;
2186 		brelse(bp);
2187 		BO_LOCK(bo);
2188 		if (nbp == NULL)
2189 			break;
2190 		nbp = gbincore(bo, lblkno);
2191 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2192 		    != xflags)
2193 			break;			/* nbp invalid */
2194 	}
2195 	return (retval);
2196 }
2197 
2198 int
2199 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2200 {
2201 	struct buf *bp;
2202 	int error;
2203 	daddr_t lblkno;
2204 
2205 	ASSERT_BO_LOCKED(bo);
2206 
2207 	for (lblkno = startn;;) {
2208 again:
2209 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2210 		if (bp == NULL || bp->b_lblkno >= endn ||
2211 		    bp->b_lblkno < startn)
2212 			break;
2213 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2214 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2215 		if (error != 0) {
2216 			BO_RLOCK(bo);
2217 			if (error == ENOLCK)
2218 				goto again;
2219 			return (error);
2220 		}
2221 		KASSERT(bp->b_bufobj == bo,
2222 		    ("bp %p wrong b_bufobj %p should be %p",
2223 		    bp, bp->b_bufobj, bo));
2224 		lblkno = bp->b_lblkno + 1;
2225 		if ((bp->b_flags & B_MANAGED) == 0)
2226 			bremfree(bp);
2227 		bp->b_flags |= B_RELBUF;
2228 		/*
2229 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2230 		 * pages backing each buffer in the range are unlikely to be
2231 		 * reused.  Dirty buffers will have the hint applied once
2232 		 * they've been written.
2233 		 */
2234 		if ((bp->b_flags & B_VMIO) != 0)
2235 			bp->b_flags |= B_NOREUSE;
2236 		brelse(bp);
2237 		BO_RLOCK(bo);
2238 	}
2239 	return (0);
2240 }
2241 
2242 /*
2243  * Truncate a file's buffer and pages to a specified length.  This
2244  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2245  * sync activity.
2246  */
2247 int
2248 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2249 {
2250 	struct buf *bp, *nbp;
2251 	struct bufobj *bo;
2252 	daddr_t startlbn;
2253 
2254 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2255 	    vp, blksize, (uintmax_t)length);
2256 
2257 	/*
2258 	 * Round up to the *next* lbn.
2259 	 */
2260 	startlbn = howmany(length, blksize);
2261 
2262 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2263 
2264 	bo = &vp->v_bufobj;
2265 restart_unlocked:
2266 	BO_LOCK(bo);
2267 
2268 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2269 		;
2270 
2271 	if (length > 0) {
2272 restartsync:
2273 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2274 			if (bp->b_lblkno > 0)
2275 				continue;
2276 			/*
2277 			 * Since we hold the vnode lock this should only
2278 			 * fail if we're racing with the buf daemon.
2279 			 */
2280 			if (BUF_LOCK(bp,
2281 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2282 			    BO_LOCKPTR(bo)) == ENOLCK)
2283 				goto restart_unlocked;
2284 
2285 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2286 			    ("buf(%p) on dirty queue without DELWRI", bp));
2287 
2288 			bremfree(bp);
2289 			bawrite(bp);
2290 			BO_LOCK(bo);
2291 			goto restartsync;
2292 		}
2293 	}
2294 
2295 	bufobj_wwait(bo, 0, 0);
2296 	BO_UNLOCK(bo);
2297 	vnode_pager_setsize(vp, length);
2298 
2299 	return (0);
2300 }
2301 
2302 /*
2303  * Invalidate the cached pages of a file's buffer within the range of block
2304  * numbers [startlbn, endlbn).
2305  */
2306 void
2307 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2308     int blksize)
2309 {
2310 	struct bufobj *bo;
2311 	off_t start, end;
2312 
2313 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2314 
2315 	start = blksize * startlbn;
2316 	end = blksize * endlbn;
2317 
2318 	bo = &vp->v_bufobj;
2319 	BO_LOCK(bo);
2320 	MPASS(blksize == bo->bo_bsize);
2321 
2322 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2323 		;
2324 
2325 	BO_UNLOCK(bo);
2326 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2327 }
2328 
2329 static int
2330 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2331     daddr_t startlbn, daddr_t endlbn)
2332 {
2333 	struct buf *bp, *nbp;
2334 	bool anyfreed;
2335 
2336 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2337 	ASSERT_BO_LOCKED(bo);
2338 
2339 	do {
2340 		anyfreed = false;
2341 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2342 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2343 				continue;
2344 			if (BUF_LOCK(bp,
2345 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2346 			    BO_LOCKPTR(bo)) == ENOLCK) {
2347 				BO_LOCK(bo);
2348 				return (EAGAIN);
2349 			}
2350 
2351 			bremfree(bp);
2352 			bp->b_flags |= B_INVAL | B_RELBUF;
2353 			bp->b_flags &= ~B_ASYNC;
2354 			brelse(bp);
2355 			anyfreed = true;
2356 
2357 			BO_LOCK(bo);
2358 			if (nbp != NULL &&
2359 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2360 			    nbp->b_vp != vp ||
2361 			    (nbp->b_flags & B_DELWRI) != 0))
2362 				return (EAGAIN);
2363 		}
2364 
2365 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2366 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2367 				continue;
2368 			if (BUF_LOCK(bp,
2369 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2370 			    BO_LOCKPTR(bo)) == ENOLCK) {
2371 				BO_LOCK(bo);
2372 				return (EAGAIN);
2373 			}
2374 			bremfree(bp);
2375 			bp->b_flags |= B_INVAL | B_RELBUF;
2376 			bp->b_flags &= ~B_ASYNC;
2377 			brelse(bp);
2378 			anyfreed = true;
2379 
2380 			BO_LOCK(bo);
2381 			if (nbp != NULL &&
2382 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2383 			    (nbp->b_vp != vp) ||
2384 			    (nbp->b_flags & B_DELWRI) == 0))
2385 				return (EAGAIN);
2386 		}
2387 	} while (anyfreed);
2388 	return (0);
2389 }
2390 
2391 static void
2392 buf_vlist_remove(struct buf *bp)
2393 {
2394 	struct bufv *bv;
2395 	b_xflags_t flags;
2396 
2397 	flags = bp->b_xflags;
2398 
2399 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2400 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2401 	KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 &&
2402 	    (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN),
2403 	    ("%s: buffer %p has invalid queue state", __func__, bp));
2404 
2405 	if ((flags & BX_VNDIRTY) != 0)
2406 		bv = &bp->b_bufobj->bo_dirty;
2407 	else
2408 		bv = &bp->b_bufobj->bo_clean;
2409 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2410 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2411 	bv->bv_cnt--;
2412 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2413 }
2414 
2415 /*
2416  * Add the buffer to the sorted clean or dirty block list.
2417  *
2418  * NOTE: xflags is passed as a constant, optimizing this inline function!
2419  */
2420 static void
2421 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2422 {
2423 	struct bufv *bv;
2424 	struct buf *n;
2425 	int error;
2426 
2427 	ASSERT_BO_WLOCKED(bo);
2428 	KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2429 	    ("buf_vlist_add: bo %p does not allow bufs", bo));
2430 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2431 	    ("dead bo %p", bo));
2432 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2433 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2434 	bp->b_xflags |= xflags;
2435 	if (xflags & BX_VNDIRTY)
2436 		bv = &bo->bo_dirty;
2437 	else
2438 		bv = &bo->bo_clean;
2439 
2440 	/*
2441 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2442 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2443 	 * than _ge.
2444 	 */
2445 	if (bv->bv_cnt == 0 ||
2446 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2447 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2448 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2449 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2450 	else
2451 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2452 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2453 	if (error)
2454 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2455 	bv->bv_cnt++;
2456 }
2457 
2458 /*
2459  * Look up a buffer using the buffer tries.
2460  */
2461 struct buf *
2462 gbincore(struct bufobj *bo, daddr_t lblkno)
2463 {
2464 	struct buf *bp;
2465 
2466 	ASSERT_BO_LOCKED(bo);
2467 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2468 	if (bp != NULL)
2469 		return (bp);
2470 	return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno));
2471 }
2472 
2473 /*
2474  * Look up a buf using the buffer tries, without the bufobj lock.  This relies
2475  * on SMR for safe lookup, and bufs being in a no-free zone to provide type
2476  * stability of the result.  Like other lockless lookups, the found buf may
2477  * already be invalid by the time this function returns.
2478  */
2479 struct buf *
2480 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno)
2481 {
2482 	struct buf *bp;
2483 
2484 	ASSERT_BO_UNLOCKED(bo);
2485 	bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno);
2486 	if (bp != NULL)
2487 		return (bp);
2488 	return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno));
2489 }
2490 
2491 /*
2492  * Associate a buffer with a vnode.
2493  */
2494 void
2495 bgetvp(struct vnode *vp, struct buf *bp)
2496 {
2497 	struct bufobj *bo;
2498 
2499 	bo = &vp->v_bufobj;
2500 	ASSERT_BO_WLOCKED(bo);
2501 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2502 
2503 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2504 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2505 	    ("bgetvp: bp already attached! %p", bp));
2506 
2507 	vhold(vp);
2508 	bp->b_vp = vp;
2509 	bp->b_bufobj = bo;
2510 	/*
2511 	 * Insert onto list for new vnode.
2512 	 */
2513 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2514 }
2515 
2516 /*
2517  * Disassociate a buffer from a vnode.
2518  */
2519 void
2520 brelvp(struct buf *bp)
2521 {
2522 	struct bufobj *bo;
2523 	struct vnode *vp;
2524 
2525 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2526 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2527 
2528 	/*
2529 	 * Delete from old vnode list, if on one.
2530 	 */
2531 	vp = bp->b_vp;		/* XXX */
2532 	bo = bp->b_bufobj;
2533 	BO_LOCK(bo);
2534 	buf_vlist_remove(bp);
2535 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2536 		bo->bo_flag &= ~BO_ONWORKLST;
2537 		mtx_lock(&sync_mtx);
2538 		LIST_REMOVE(bo, bo_synclist);
2539 		syncer_worklist_len--;
2540 		mtx_unlock(&sync_mtx);
2541 	}
2542 	bp->b_vp = NULL;
2543 	bp->b_bufobj = NULL;
2544 	BO_UNLOCK(bo);
2545 	vdrop(vp);
2546 }
2547 
2548 /*
2549  * Add an item to the syncer work queue.
2550  */
2551 static void
2552 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2553 {
2554 	int slot;
2555 
2556 	ASSERT_BO_WLOCKED(bo);
2557 
2558 	mtx_lock(&sync_mtx);
2559 	if (bo->bo_flag & BO_ONWORKLST)
2560 		LIST_REMOVE(bo, bo_synclist);
2561 	else {
2562 		bo->bo_flag |= BO_ONWORKLST;
2563 		syncer_worklist_len++;
2564 	}
2565 
2566 	if (delay > syncer_maxdelay - 2)
2567 		delay = syncer_maxdelay - 2;
2568 	slot = (syncer_delayno + delay) & syncer_mask;
2569 
2570 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2571 	mtx_unlock(&sync_mtx);
2572 }
2573 
2574 static int
2575 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2576 {
2577 	int error, len;
2578 
2579 	mtx_lock(&sync_mtx);
2580 	len = syncer_worklist_len - sync_vnode_count;
2581 	mtx_unlock(&sync_mtx);
2582 	error = SYSCTL_OUT(req, &len, sizeof(len));
2583 	return (error);
2584 }
2585 
2586 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2587     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2588     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2589 
2590 static struct proc *updateproc;
2591 static void sched_sync(void);
2592 static struct kproc_desc up_kp = {
2593 	"syncer",
2594 	sched_sync,
2595 	&updateproc
2596 };
2597 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2598 
2599 static int
2600 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2601 {
2602 	struct vnode *vp;
2603 	struct mount *mp;
2604 
2605 	*bo = LIST_FIRST(slp);
2606 	if (*bo == NULL)
2607 		return (0);
2608 	vp = bo2vnode(*bo);
2609 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2610 		return (1);
2611 	/*
2612 	 * We use vhold in case the vnode does not
2613 	 * successfully sync.  vhold prevents the vnode from
2614 	 * going away when we unlock the sync_mtx so that
2615 	 * we can acquire the vnode interlock.
2616 	 */
2617 	vholdl(vp);
2618 	mtx_unlock(&sync_mtx);
2619 	VI_UNLOCK(vp);
2620 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2621 		vdrop(vp);
2622 		mtx_lock(&sync_mtx);
2623 		return (*bo == LIST_FIRST(slp));
2624 	}
2625 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2626 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2627 	VOP_UNLOCK(vp);
2628 	vn_finished_write(mp);
2629 	BO_LOCK(*bo);
2630 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2631 		/*
2632 		 * Put us back on the worklist.  The worklist
2633 		 * routine will remove us from our current
2634 		 * position and then add us back in at a later
2635 		 * position.
2636 		 */
2637 		vn_syncer_add_to_worklist(*bo, syncdelay);
2638 	}
2639 	BO_UNLOCK(*bo);
2640 	vdrop(vp);
2641 	mtx_lock(&sync_mtx);
2642 	return (0);
2643 }
2644 
2645 static int first_printf = 1;
2646 
2647 /*
2648  * System filesystem synchronizer daemon.
2649  */
2650 static void
2651 sched_sync(void)
2652 {
2653 	struct synclist *next, *slp;
2654 	struct bufobj *bo;
2655 	long starttime;
2656 	struct thread *td = curthread;
2657 	int last_work_seen;
2658 	int net_worklist_len;
2659 	int syncer_final_iter;
2660 	int error;
2661 
2662 	last_work_seen = 0;
2663 	syncer_final_iter = 0;
2664 	syncer_state = SYNCER_RUNNING;
2665 	starttime = time_uptime;
2666 	td->td_pflags |= TDP_NORUNNINGBUF;
2667 
2668 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2669 	    SHUTDOWN_PRI_LAST);
2670 
2671 	mtx_lock(&sync_mtx);
2672 	for (;;) {
2673 		if (syncer_state == SYNCER_FINAL_DELAY &&
2674 		    syncer_final_iter == 0) {
2675 			mtx_unlock(&sync_mtx);
2676 			kproc_suspend_check(td->td_proc);
2677 			mtx_lock(&sync_mtx);
2678 		}
2679 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2680 		if (syncer_state != SYNCER_RUNNING &&
2681 		    starttime != time_uptime) {
2682 			if (first_printf) {
2683 				printf("\nSyncing disks, vnodes remaining... ");
2684 				first_printf = 0;
2685 			}
2686 			printf("%d ", net_worklist_len);
2687 		}
2688 		starttime = time_uptime;
2689 
2690 		/*
2691 		 * Push files whose dirty time has expired.  Be careful
2692 		 * of interrupt race on slp queue.
2693 		 *
2694 		 * Skip over empty worklist slots when shutting down.
2695 		 */
2696 		do {
2697 			slp = &syncer_workitem_pending[syncer_delayno];
2698 			syncer_delayno += 1;
2699 			if (syncer_delayno == syncer_maxdelay)
2700 				syncer_delayno = 0;
2701 			next = &syncer_workitem_pending[syncer_delayno];
2702 			/*
2703 			 * If the worklist has wrapped since the
2704 			 * it was emptied of all but syncer vnodes,
2705 			 * switch to the FINAL_DELAY state and run
2706 			 * for one more second.
2707 			 */
2708 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2709 			    net_worklist_len == 0 &&
2710 			    last_work_seen == syncer_delayno) {
2711 				syncer_state = SYNCER_FINAL_DELAY;
2712 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2713 			}
2714 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2715 		    syncer_worklist_len > 0);
2716 
2717 		/*
2718 		 * Keep track of the last time there was anything
2719 		 * on the worklist other than syncer vnodes.
2720 		 * Return to the SHUTTING_DOWN state if any
2721 		 * new work appears.
2722 		 */
2723 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2724 			last_work_seen = syncer_delayno;
2725 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2726 			syncer_state = SYNCER_SHUTTING_DOWN;
2727 		while (!LIST_EMPTY(slp)) {
2728 			error = sync_vnode(slp, &bo, td);
2729 			if (error == 1) {
2730 				LIST_REMOVE(bo, bo_synclist);
2731 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2732 				continue;
2733 			}
2734 
2735 			if (first_printf == 0) {
2736 				/*
2737 				 * Drop the sync mutex, because some watchdog
2738 				 * drivers need to sleep while patting
2739 				 */
2740 				mtx_unlock(&sync_mtx);
2741 				wdog_kern_pat(WD_LASTVAL);
2742 				mtx_lock(&sync_mtx);
2743 			}
2744 		}
2745 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2746 			syncer_final_iter--;
2747 		/*
2748 		 * The variable rushjob allows the kernel to speed up the
2749 		 * processing of the filesystem syncer process. A rushjob
2750 		 * value of N tells the filesystem syncer to process the next
2751 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2752 		 * is used by the soft update code to speed up the filesystem
2753 		 * syncer process when the incore state is getting so far
2754 		 * ahead of the disk that the kernel memory pool is being
2755 		 * threatened with exhaustion.
2756 		 */
2757 		if (rushjob > 0) {
2758 			rushjob -= 1;
2759 			continue;
2760 		}
2761 		/*
2762 		 * Just sleep for a short period of time between
2763 		 * iterations when shutting down to allow some I/O
2764 		 * to happen.
2765 		 *
2766 		 * If it has taken us less than a second to process the
2767 		 * current work, then wait. Otherwise start right over
2768 		 * again. We can still lose time if any single round
2769 		 * takes more than two seconds, but it does not really
2770 		 * matter as we are just trying to generally pace the
2771 		 * filesystem activity.
2772 		 */
2773 		if (syncer_state != SYNCER_RUNNING ||
2774 		    time_uptime == starttime) {
2775 			thread_lock(td);
2776 			sched_prio(td, PPAUSE);
2777 			thread_unlock(td);
2778 		}
2779 		if (syncer_state != SYNCER_RUNNING)
2780 			cv_timedwait(&sync_wakeup, &sync_mtx,
2781 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2782 		else if (time_uptime == starttime)
2783 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2784 	}
2785 }
2786 
2787 /*
2788  * Request the syncer daemon to speed up its work.
2789  * We never push it to speed up more than half of its
2790  * normal turn time, otherwise it could take over the cpu.
2791  */
2792 int
2793 speedup_syncer(void)
2794 {
2795 	int ret = 0;
2796 
2797 	mtx_lock(&sync_mtx);
2798 	if (rushjob < syncdelay / 2) {
2799 		rushjob += 1;
2800 		stat_rush_requests += 1;
2801 		ret = 1;
2802 	}
2803 	mtx_unlock(&sync_mtx);
2804 	cv_broadcast(&sync_wakeup);
2805 	return (ret);
2806 }
2807 
2808 /*
2809  * Tell the syncer to speed up its work and run though its work
2810  * list several times, then tell it to shut down.
2811  */
2812 static void
2813 syncer_shutdown(void *arg, int howto)
2814 {
2815 
2816 	if (howto & RB_NOSYNC)
2817 		return;
2818 	mtx_lock(&sync_mtx);
2819 	syncer_state = SYNCER_SHUTTING_DOWN;
2820 	rushjob = 0;
2821 	mtx_unlock(&sync_mtx);
2822 	cv_broadcast(&sync_wakeup);
2823 	kproc_shutdown(arg, howto);
2824 }
2825 
2826 void
2827 syncer_suspend(void)
2828 {
2829 
2830 	syncer_shutdown(updateproc, 0);
2831 }
2832 
2833 void
2834 syncer_resume(void)
2835 {
2836 
2837 	mtx_lock(&sync_mtx);
2838 	first_printf = 1;
2839 	syncer_state = SYNCER_RUNNING;
2840 	mtx_unlock(&sync_mtx);
2841 	cv_broadcast(&sync_wakeup);
2842 	kproc_resume(updateproc);
2843 }
2844 
2845 /*
2846  * Move the buffer between the clean and dirty lists of its vnode.
2847  */
2848 void
2849 reassignbuf(struct buf *bp)
2850 {
2851 	struct vnode *vp;
2852 	struct bufobj *bo;
2853 	int delay;
2854 #ifdef INVARIANTS
2855 	struct bufv *bv;
2856 #endif
2857 
2858 	vp = bp->b_vp;
2859 	bo = bp->b_bufobj;
2860 
2861 	KASSERT((bp->b_flags & B_PAGING) == 0,
2862 	    ("%s: cannot reassign paging buffer %p", __func__, bp));
2863 
2864 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2865 	    bp, bp->b_vp, bp->b_flags);
2866 
2867 	BO_LOCK(bo);
2868 	buf_vlist_remove(bp);
2869 
2870 	/*
2871 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2872 	 * of clean buffers.
2873 	 */
2874 	if (bp->b_flags & B_DELWRI) {
2875 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2876 			switch (vp->v_type) {
2877 			case VDIR:
2878 				delay = dirdelay;
2879 				break;
2880 			case VCHR:
2881 				delay = metadelay;
2882 				break;
2883 			default:
2884 				delay = filedelay;
2885 			}
2886 			vn_syncer_add_to_worklist(bo, delay);
2887 		}
2888 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2889 	} else {
2890 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2891 
2892 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2893 			mtx_lock(&sync_mtx);
2894 			LIST_REMOVE(bo, bo_synclist);
2895 			syncer_worklist_len--;
2896 			mtx_unlock(&sync_mtx);
2897 			bo->bo_flag &= ~BO_ONWORKLST;
2898 		}
2899 	}
2900 #ifdef INVARIANTS
2901 	bv = &bo->bo_clean;
2902 	bp = TAILQ_FIRST(&bv->bv_hd);
2903 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2904 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2905 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2906 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2907 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2908 	bv = &bo->bo_dirty;
2909 	bp = TAILQ_FIRST(&bv->bv_hd);
2910 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2911 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2912 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2913 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2914 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2915 #endif
2916 	BO_UNLOCK(bo);
2917 }
2918 
2919 static void
2920 v_init_counters(struct vnode *vp)
2921 {
2922 
2923 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2924 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2925 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2926 
2927 	refcount_init(&vp->v_holdcnt, 1);
2928 	refcount_init(&vp->v_usecount, 1);
2929 }
2930 
2931 /*
2932  * Grab a particular vnode from the free list, increment its
2933  * reference count and lock it.  VIRF_DOOMED is set if the vnode
2934  * is being destroyed.  Only callers who specify LK_RETRY will
2935  * see doomed vnodes.  If inactive processing was delayed in
2936  * vput try to do it here.
2937  *
2938  * usecount is manipulated using atomics without holding any locks.
2939  *
2940  * holdcnt can be manipulated using atomics without holding any locks,
2941  * except when transitioning 1<->0, in which case the interlock is held.
2942  *
2943  * Consumers which don't guarantee liveness of the vnode can use SMR to
2944  * try to get a reference. Note this operation can fail since the vnode
2945  * may be awaiting getting freed by the time they get to it.
2946  */
2947 enum vgetstate
2948 vget_prep_smr(struct vnode *vp)
2949 {
2950 	enum vgetstate vs;
2951 
2952 	VFS_SMR_ASSERT_ENTERED();
2953 
2954 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2955 		vs = VGET_USECOUNT;
2956 	} else {
2957 		if (vhold_smr(vp))
2958 			vs = VGET_HOLDCNT;
2959 		else
2960 			vs = VGET_NONE;
2961 	}
2962 	return (vs);
2963 }
2964 
2965 enum vgetstate
2966 vget_prep(struct vnode *vp)
2967 {
2968 	enum vgetstate vs;
2969 
2970 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2971 		vs = VGET_USECOUNT;
2972 	} else {
2973 		vhold(vp);
2974 		vs = VGET_HOLDCNT;
2975 	}
2976 	return (vs);
2977 }
2978 
2979 void
2980 vget_abort(struct vnode *vp, enum vgetstate vs)
2981 {
2982 
2983 	switch (vs) {
2984 	case VGET_USECOUNT:
2985 		vrele(vp);
2986 		break;
2987 	case VGET_HOLDCNT:
2988 		vdrop(vp);
2989 		break;
2990 	default:
2991 		__assert_unreachable();
2992 	}
2993 }
2994 
2995 int
2996 vget(struct vnode *vp, int flags)
2997 {
2998 	enum vgetstate vs;
2999 
3000 	vs = vget_prep(vp);
3001 	return (vget_finish(vp, flags, vs));
3002 }
3003 
3004 int
3005 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
3006 {
3007 	int error;
3008 
3009 	if ((flags & LK_INTERLOCK) != 0)
3010 		ASSERT_VI_LOCKED(vp, __func__);
3011 	else
3012 		ASSERT_VI_UNLOCKED(vp, __func__);
3013 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3014 	VNPASS(vp->v_holdcnt > 0, vp);
3015 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3016 
3017 	error = vn_lock(vp, flags);
3018 	if (__predict_false(error != 0)) {
3019 		vget_abort(vp, vs);
3020 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
3021 		    vp);
3022 		return (error);
3023 	}
3024 
3025 	vget_finish_ref(vp, vs);
3026 	return (0);
3027 }
3028 
3029 void
3030 vget_finish_ref(struct vnode *vp, enum vgetstate vs)
3031 {
3032 	int old;
3033 
3034 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3035 	VNPASS(vp->v_holdcnt > 0, vp);
3036 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3037 
3038 	if (vs == VGET_USECOUNT)
3039 		return;
3040 
3041 	/*
3042 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
3043 	 * the vnode around. Otherwise someone else lended their hold count and
3044 	 * we have to drop ours.
3045 	 */
3046 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
3047 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3048 	if (old != 0) {
3049 #ifdef INVARIANTS
3050 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3051 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3052 #else
3053 		refcount_release(&vp->v_holdcnt);
3054 #endif
3055 	}
3056 }
3057 
3058 void
3059 vref(struct vnode *vp)
3060 {
3061 	enum vgetstate vs;
3062 
3063 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3064 	vs = vget_prep(vp);
3065 	vget_finish_ref(vp, vs);
3066 }
3067 
3068 void
3069 vrefact(struct vnode *vp)
3070 {
3071 
3072 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3073 #ifdef INVARIANTS
3074 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3075 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3076 #else
3077 	refcount_acquire(&vp->v_usecount);
3078 #endif
3079 }
3080 
3081 void
3082 vlazy(struct vnode *vp)
3083 {
3084 	struct mount *mp;
3085 
3086 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3087 
3088 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3089 		return;
3090 	/*
3091 	 * We may get here for inactive routines after the vnode got doomed.
3092 	 */
3093 	if (VN_IS_DOOMED(vp))
3094 		return;
3095 	mp = vp->v_mount;
3096 	mtx_lock(&mp->mnt_listmtx);
3097 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3098 		vp->v_mflag |= VMP_LAZYLIST;
3099 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3100 		mp->mnt_lazyvnodelistsize++;
3101 	}
3102 	mtx_unlock(&mp->mnt_listmtx);
3103 }
3104 
3105 static void
3106 vunlazy(struct vnode *vp)
3107 {
3108 	struct mount *mp;
3109 
3110 	ASSERT_VI_LOCKED(vp, __func__);
3111 	VNPASS(!VN_IS_DOOMED(vp), vp);
3112 
3113 	mp = vp->v_mount;
3114 	mtx_lock(&mp->mnt_listmtx);
3115 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3116 	/*
3117 	 * Don't remove the vnode from the lazy list if another thread
3118 	 * has increased the hold count. It may have re-enqueued the
3119 	 * vnode to the lazy list and is now responsible for its
3120 	 * removal.
3121 	 */
3122 	if (vp->v_holdcnt == 0) {
3123 		vp->v_mflag &= ~VMP_LAZYLIST;
3124 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3125 		mp->mnt_lazyvnodelistsize--;
3126 	}
3127 	mtx_unlock(&mp->mnt_listmtx);
3128 }
3129 
3130 /*
3131  * This routine is only meant to be called from vgonel prior to dooming
3132  * the vnode.
3133  */
3134 static void
3135 vunlazy_gone(struct vnode *vp)
3136 {
3137 	struct mount *mp;
3138 
3139 	ASSERT_VOP_ELOCKED(vp, __func__);
3140 	ASSERT_VI_LOCKED(vp, __func__);
3141 	VNPASS(!VN_IS_DOOMED(vp), vp);
3142 
3143 	if (vp->v_mflag & VMP_LAZYLIST) {
3144 		mp = vp->v_mount;
3145 		mtx_lock(&mp->mnt_listmtx);
3146 		VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3147 		vp->v_mflag &= ~VMP_LAZYLIST;
3148 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3149 		mp->mnt_lazyvnodelistsize--;
3150 		mtx_unlock(&mp->mnt_listmtx);
3151 	}
3152 }
3153 
3154 static void
3155 vdefer_inactive(struct vnode *vp)
3156 {
3157 
3158 	ASSERT_VI_LOCKED(vp, __func__);
3159 	VNASSERT(vp->v_holdcnt > 0, vp,
3160 	    ("%s: vnode without hold count", __func__));
3161 	if (VN_IS_DOOMED(vp)) {
3162 		vdropl(vp);
3163 		return;
3164 	}
3165 	if (vp->v_iflag & VI_DEFINACT) {
3166 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3167 		vdropl(vp);
3168 		return;
3169 	}
3170 	if (vp->v_usecount > 0) {
3171 		vp->v_iflag &= ~VI_OWEINACT;
3172 		vdropl(vp);
3173 		return;
3174 	}
3175 	vlazy(vp);
3176 	vp->v_iflag |= VI_DEFINACT;
3177 	VI_UNLOCK(vp);
3178 	counter_u64_add(deferred_inact, 1);
3179 }
3180 
3181 static void
3182 vdefer_inactive_unlocked(struct vnode *vp)
3183 {
3184 
3185 	VI_LOCK(vp);
3186 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3187 		vdropl(vp);
3188 		return;
3189 	}
3190 	vdefer_inactive(vp);
3191 }
3192 
3193 enum vput_op { VRELE, VPUT, VUNREF };
3194 
3195 /*
3196  * Handle ->v_usecount transitioning to 0.
3197  *
3198  * By releasing the last usecount we take ownership of the hold count which
3199  * provides liveness of the vnode, meaning we have to vdrop.
3200  *
3201  * For all vnodes we may need to perform inactive processing. It requires an
3202  * exclusive lock on the vnode, while it is legal to call here with only a
3203  * shared lock (or no locks). If locking the vnode in an expected manner fails,
3204  * inactive processing gets deferred to the syncer.
3205  *
3206  * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3207  * on the lock being held all the way until VOP_INACTIVE. This in particular
3208  * happens with UFS which adds half-constructed vnodes to the hash, where they
3209  * can be found by other code.
3210  */
3211 static void
3212 vput_final(struct vnode *vp, enum vput_op func)
3213 {
3214 	int error;
3215 	bool want_unlock;
3216 
3217 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3218 	VNPASS(vp->v_holdcnt > 0, vp);
3219 
3220 	VI_LOCK(vp);
3221 
3222 	/*
3223 	 * By the time we got here someone else might have transitioned
3224 	 * the count back to > 0.
3225 	 */
3226 	if (vp->v_usecount > 0)
3227 		goto out;
3228 
3229 	/*
3230 	 * If the vnode is doomed vgone already performed inactive processing
3231 	 * (if needed).
3232 	 */
3233 	if (VN_IS_DOOMED(vp))
3234 		goto out;
3235 
3236 	if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3237 		goto out;
3238 
3239 	if (vp->v_iflag & VI_DOINGINACT)
3240 		goto out;
3241 
3242 	/*
3243 	 * Locking operations here will drop the interlock and possibly the
3244 	 * vnode lock, opening a window where the vnode can get doomed all the
3245 	 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3246 	 * perform inactive.
3247 	 */
3248 	vp->v_iflag |= VI_OWEINACT;
3249 	want_unlock = false;
3250 	error = 0;
3251 	switch (func) {
3252 	case VRELE:
3253 		switch (VOP_ISLOCKED(vp)) {
3254 		case LK_EXCLUSIVE:
3255 			break;
3256 		case LK_EXCLOTHER:
3257 		case 0:
3258 			want_unlock = true;
3259 			error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3260 			VI_LOCK(vp);
3261 			break;
3262 		default:
3263 			/*
3264 			 * The lock has at least one sharer, but we have no way
3265 			 * to conclude whether this is us. Play it safe and
3266 			 * defer processing.
3267 			 */
3268 			error = EAGAIN;
3269 			break;
3270 		}
3271 		break;
3272 	case VPUT:
3273 		want_unlock = true;
3274 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3275 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3276 			    LK_NOWAIT);
3277 			VI_LOCK(vp);
3278 		}
3279 		break;
3280 	case VUNREF:
3281 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3282 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3283 			VI_LOCK(vp);
3284 		}
3285 		break;
3286 	}
3287 	if (error == 0) {
3288 		if (func == VUNREF) {
3289 			VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp,
3290 			    ("recursive vunref"));
3291 			vp->v_vflag |= VV_UNREF;
3292 		}
3293 		for (;;) {
3294 			error = vinactive(vp);
3295 			if (want_unlock)
3296 				VOP_UNLOCK(vp);
3297 			if (error != ERELOOKUP || !want_unlock)
3298 				break;
3299 			VOP_LOCK(vp, LK_EXCLUSIVE);
3300 		}
3301 		if (func == VUNREF)
3302 			vp->v_vflag &= ~VV_UNREF;
3303 		vdropl(vp);
3304 	} else {
3305 		vdefer_inactive(vp);
3306 	}
3307 	return;
3308 out:
3309 	if (func == VPUT)
3310 		VOP_UNLOCK(vp);
3311 	vdropl(vp);
3312 }
3313 
3314 /*
3315  * Decrement ->v_usecount for a vnode.
3316  *
3317  * Releasing the last use count requires additional processing, see vput_final
3318  * above for details.
3319  *
3320  * Comment above each variant denotes lock state on entry and exit.
3321  */
3322 
3323 /*
3324  * in: any
3325  * out: same as passed in
3326  */
3327 void
3328 vrele(struct vnode *vp)
3329 {
3330 
3331 	ASSERT_VI_UNLOCKED(vp, __func__);
3332 	if (!refcount_release(&vp->v_usecount))
3333 		return;
3334 	vput_final(vp, VRELE);
3335 }
3336 
3337 /*
3338  * in: locked
3339  * out: unlocked
3340  */
3341 void
3342 vput(struct vnode *vp)
3343 {
3344 
3345 	ASSERT_VOP_LOCKED(vp, __func__);
3346 	ASSERT_VI_UNLOCKED(vp, __func__);
3347 	if (!refcount_release(&vp->v_usecount)) {
3348 		VOP_UNLOCK(vp);
3349 		return;
3350 	}
3351 	vput_final(vp, VPUT);
3352 }
3353 
3354 /*
3355  * in: locked
3356  * out: locked
3357  */
3358 void
3359 vunref(struct vnode *vp)
3360 {
3361 
3362 	ASSERT_VOP_LOCKED(vp, __func__);
3363 	ASSERT_VI_UNLOCKED(vp, __func__);
3364 	if (!refcount_release(&vp->v_usecount))
3365 		return;
3366 	vput_final(vp, VUNREF);
3367 }
3368 
3369 void
3370 vhold(struct vnode *vp)
3371 {
3372 	int old;
3373 
3374 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3375 	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3376 	VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3377 	    ("%s: wrong hold count %d", __func__, old));
3378 	if (old == 0)
3379 		vfs_freevnodes_dec();
3380 }
3381 
3382 void
3383 vholdnz(struct vnode *vp)
3384 {
3385 
3386 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3387 #ifdef INVARIANTS
3388 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3389 	VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3390 	    ("%s: wrong hold count %d", __func__, old));
3391 #else
3392 	atomic_add_int(&vp->v_holdcnt, 1);
3393 #endif
3394 }
3395 
3396 /*
3397  * Grab a hold count unless the vnode is freed.
3398  *
3399  * Only use this routine if vfs smr is the only protection you have against
3400  * freeing the vnode.
3401  *
3402  * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag
3403  * is not set.  After the flag is set the vnode becomes immutable to anyone but
3404  * the thread which managed to set the flag.
3405  *
3406  * It may be tempting to replace the loop with:
3407  * count = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3408  * if (count & VHOLD_NO_SMR) {
3409  *     backpedal and error out;
3410  * }
3411  *
3412  * However, while this is more performant, it hinders debugging by eliminating
3413  * the previously mentioned invariant.
3414  */
3415 bool
3416 vhold_smr(struct vnode *vp)
3417 {
3418 	int count;
3419 
3420 	VFS_SMR_ASSERT_ENTERED();
3421 
3422 	count = atomic_load_int(&vp->v_holdcnt);
3423 	for (;;) {
3424 		if (count & VHOLD_NO_SMR) {
3425 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3426 			    ("non-zero hold count with flags %d\n", count));
3427 			return (false);
3428 		}
3429 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3430 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3431 			if (count == 0)
3432 				vfs_freevnodes_dec();
3433 			return (true);
3434 		}
3435 	}
3436 }
3437 
3438 /*
3439  * Hold a free vnode for recycling.
3440  *
3441  * Note: vnode_init references this comment.
3442  *
3443  * Attempts to recycle only need the global vnode list lock and have no use for
3444  * SMR.
3445  *
3446  * However, vnodes get inserted into the global list before they get fully
3447  * initialized and stay there until UMA decides to free the memory. This in
3448  * particular means the target can be found before it becomes usable and after
3449  * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to
3450  * VHOLD_NO_SMR.
3451  *
3452  * Note: the vnode may gain more references after we transition the count 0->1.
3453  */
3454 static bool
3455 vhold_recycle_free(struct vnode *vp)
3456 {
3457 	int count;
3458 
3459 	mtx_assert(&vnode_list_mtx, MA_OWNED);
3460 
3461 	count = atomic_load_int(&vp->v_holdcnt);
3462 	for (;;) {
3463 		if (count & VHOLD_NO_SMR) {
3464 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3465 			    ("non-zero hold count with flags %d\n", count));
3466 			return (false);
3467 		}
3468 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3469 		if (count > 0) {
3470 			return (false);
3471 		}
3472 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3473 			vfs_freevnodes_dec();
3474 			return (true);
3475 		}
3476 	}
3477 }
3478 
3479 static void __noinline
3480 vdbatch_process(struct vdbatch *vd)
3481 {
3482 	struct vnode *vp;
3483 	int i;
3484 
3485 	mtx_assert(&vd->lock, MA_OWNED);
3486 	MPASS(curthread->td_pinned > 0);
3487 	MPASS(vd->index == VDBATCH_SIZE);
3488 
3489 	mtx_lock(&vnode_list_mtx);
3490 	critical_enter();
3491 	freevnodes += vd->freevnodes;
3492 	for (i = 0; i < VDBATCH_SIZE; i++) {
3493 		vp = vd->tab[i];
3494 		TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3495 		TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3496 		MPASS(vp->v_dbatchcpu != NOCPU);
3497 		vp->v_dbatchcpu = NOCPU;
3498 	}
3499 	mtx_unlock(&vnode_list_mtx);
3500 	vd->freevnodes = 0;
3501 	bzero(vd->tab, sizeof(vd->tab));
3502 	vd->index = 0;
3503 	critical_exit();
3504 }
3505 
3506 static void
3507 vdbatch_enqueue(struct vnode *vp)
3508 {
3509 	struct vdbatch *vd;
3510 
3511 	ASSERT_VI_LOCKED(vp, __func__);
3512 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3513 	    ("%s: deferring requeue of a doomed vnode", __func__));
3514 
3515 	if (vp->v_dbatchcpu != NOCPU) {
3516 		VI_UNLOCK(vp);
3517 		return;
3518 	}
3519 
3520 	sched_pin();
3521 	vd = DPCPU_PTR(vd);
3522 	mtx_lock(&vd->lock);
3523 	MPASS(vd->index < VDBATCH_SIZE);
3524 	MPASS(vd->tab[vd->index] == NULL);
3525 	/*
3526 	 * A hack: we depend on being pinned so that we know what to put in
3527 	 * ->v_dbatchcpu.
3528 	 */
3529 	vp->v_dbatchcpu = curcpu;
3530 	vd->tab[vd->index] = vp;
3531 	vd->index++;
3532 	VI_UNLOCK(vp);
3533 	if (vd->index == VDBATCH_SIZE)
3534 		vdbatch_process(vd);
3535 	mtx_unlock(&vd->lock);
3536 	sched_unpin();
3537 }
3538 
3539 /*
3540  * This routine must only be called for vnodes which are about to be
3541  * deallocated. Supporting dequeue for arbitrary vndoes would require
3542  * validating that the locked batch matches.
3543  */
3544 static void
3545 vdbatch_dequeue(struct vnode *vp)
3546 {
3547 	struct vdbatch *vd;
3548 	int i;
3549 	short cpu;
3550 
3551 	VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3552 	    ("%s: called for a used vnode\n", __func__));
3553 
3554 	cpu = vp->v_dbatchcpu;
3555 	if (cpu == NOCPU)
3556 		return;
3557 
3558 	vd = DPCPU_ID_PTR(cpu, vd);
3559 	mtx_lock(&vd->lock);
3560 	for (i = 0; i < vd->index; i++) {
3561 		if (vd->tab[i] != vp)
3562 			continue;
3563 		vp->v_dbatchcpu = NOCPU;
3564 		vd->index--;
3565 		vd->tab[i] = vd->tab[vd->index];
3566 		vd->tab[vd->index] = NULL;
3567 		break;
3568 	}
3569 	mtx_unlock(&vd->lock);
3570 	/*
3571 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3572 	 */
3573 	MPASS(vp->v_dbatchcpu == NOCPU);
3574 }
3575 
3576 /*
3577  * Drop the hold count of the vnode.  If this is the last reference to
3578  * the vnode we place it on the free list unless it has been vgone'd
3579  * (marked VIRF_DOOMED) in which case we will free it.
3580  *
3581  * Because the vnode vm object keeps a hold reference on the vnode if
3582  * there is at least one resident non-cached page, the vnode cannot
3583  * leave the active list without the page cleanup done.
3584  */
3585 static void __noinline
3586 vdropl_final(struct vnode *vp)
3587 {
3588 
3589 	ASSERT_VI_LOCKED(vp, __func__);
3590 	VNPASS(VN_IS_DOOMED(vp), vp);
3591 	/*
3592 	 * Set the VHOLD_NO_SMR flag.
3593 	 *
3594 	 * We may be racing against vhold_smr. If they win we can just pretend
3595 	 * we never got this far, they will vdrop later.
3596 	 */
3597 	if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) {
3598 		vfs_freevnodes_inc();
3599 		VI_UNLOCK(vp);
3600 		/*
3601 		 * We lost the aforementioned race. Any subsequent access is
3602 		 * invalid as they might have managed to vdropl on their own.
3603 		 */
3604 		return;
3605 	}
3606 	/*
3607 	 * Don't bump freevnodes as this one is going away.
3608 	 */
3609 	freevnode(vp);
3610 }
3611 
3612 void
3613 vdrop(struct vnode *vp)
3614 {
3615 
3616 	ASSERT_VI_UNLOCKED(vp, __func__);
3617 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3618 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3619 		return;
3620 	VI_LOCK(vp);
3621 	vdropl(vp);
3622 }
3623 
3624 static void __always_inline
3625 vdropl_impl(struct vnode *vp, bool enqueue)
3626 {
3627 
3628 	ASSERT_VI_LOCKED(vp, __func__);
3629 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3630 	if (!refcount_release(&vp->v_holdcnt)) {
3631 		VI_UNLOCK(vp);
3632 		return;
3633 	}
3634 	VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp);
3635 	VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
3636 	if (VN_IS_DOOMED(vp)) {
3637 		vdropl_final(vp);
3638 		return;
3639 	}
3640 
3641 	vfs_freevnodes_inc();
3642 	if (vp->v_mflag & VMP_LAZYLIST) {
3643 		vunlazy(vp);
3644 	}
3645 	/*
3646 	 * Also unlocks the interlock. We can't assert on it as we
3647 	 * released our hold and by now the vnode might have been
3648 	 * freed.
3649 	 */
3650 	vdbatch_enqueue(vp);
3651 }
3652 
3653 void
3654 vdropl(struct vnode *vp)
3655 {
3656 
3657 	vdropl_impl(vp, true);
3658 }
3659 
3660 /*
3661  * vdrop a vnode when recycling
3662  *
3663  * This is a special case routine only to be used when recycling, differs from
3664  * regular vdrop by not requeieing the vnode on LRU.
3665  *
3666  * Consider a case where vtryrecycle continuously fails with all vnodes (due to
3667  * e.g., frozen writes on the filesystem), filling the batch and causing it to
3668  * be requeued. Then vnlru will end up revisiting the same vnodes. This is a
3669  * loop which can last for as long as writes are frozen.
3670  */
3671 static void
3672 vdropl_recycle(struct vnode *vp)
3673 {
3674 
3675 	vdropl_impl(vp, false);
3676 }
3677 
3678 static void
3679 vdrop_recycle(struct vnode *vp)
3680 {
3681 
3682 	VI_LOCK(vp);
3683 	vdropl_recycle(vp);
3684 }
3685 
3686 /*
3687  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3688  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3689  */
3690 static int
3691 vinactivef(struct vnode *vp)
3692 {
3693 	struct vm_object *obj;
3694 	int error;
3695 
3696 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3697 	ASSERT_VI_LOCKED(vp, "vinactive");
3698 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3699 	    ("vinactive: recursed on VI_DOINGINACT"));
3700 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3701 	vp->v_iflag |= VI_DOINGINACT;
3702 	vp->v_iflag &= ~VI_OWEINACT;
3703 	VI_UNLOCK(vp);
3704 	/*
3705 	 * Before moving off the active list, we must be sure that any
3706 	 * modified pages are converted into the vnode's dirty
3707 	 * buffers, since these will no longer be checked once the
3708 	 * vnode is on the inactive list.
3709 	 *
3710 	 * The write-out of the dirty pages is asynchronous.  At the
3711 	 * point that VOP_INACTIVE() is called, there could still be
3712 	 * pending I/O and dirty pages in the object.
3713 	 */
3714 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3715 	    vm_object_mightbedirty(obj)) {
3716 		VM_OBJECT_WLOCK(obj);
3717 		vm_object_page_clean(obj, 0, 0, 0);
3718 		VM_OBJECT_WUNLOCK(obj);
3719 	}
3720 	error = VOP_INACTIVE(vp);
3721 	VI_LOCK(vp);
3722 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3723 	    ("vinactive: lost VI_DOINGINACT"));
3724 	vp->v_iflag &= ~VI_DOINGINACT;
3725 	return (error);
3726 }
3727 
3728 int
3729 vinactive(struct vnode *vp)
3730 {
3731 
3732 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3733 	ASSERT_VI_LOCKED(vp, "vinactive");
3734 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3735 
3736 	if ((vp->v_iflag & VI_OWEINACT) == 0)
3737 		return (0);
3738 	if (vp->v_iflag & VI_DOINGINACT)
3739 		return (0);
3740 	if (vp->v_usecount > 0) {
3741 		vp->v_iflag &= ~VI_OWEINACT;
3742 		return (0);
3743 	}
3744 	return (vinactivef(vp));
3745 }
3746 
3747 /*
3748  * Remove any vnodes in the vnode table belonging to mount point mp.
3749  *
3750  * If FORCECLOSE is not specified, there should not be any active ones,
3751  * return error if any are found (nb: this is a user error, not a
3752  * system error). If FORCECLOSE is specified, detach any active vnodes
3753  * that are found.
3754  *
3755  * If WRITECLOSE is set, only flush out regular file vnodes open for
3756  * writing.
3757  *
3758  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3759  *
3760  * `rootrefs' specifies the base reference count for the root vnode
3761  * of this filesystem. The root vnode is considered busy if its
3762  * v_usecount exceeds this value. On a successful return, vflush(, td)
3763  * will call vrele() on the root vnode exactly rootrefs times.
3764  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3765  * be zero.
3766  */
3767 #ifdef DIAGNOSTIC
3768 static int busyprt = 0;		/* print out busy vnodes */
3769 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3770 #endif
3771 
3772 int
3773 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3774 {
3775 	struct vnode *vp, *mvp, *rootvp = NULL;
3776 	struct vattr vattr;
3777 	int busy = 0, error;
3778 
3779 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3780 	    rootrefs, flags);
3781 	if (rootrefs > 0) {
3782 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3783 		    ("vflush: bad args"));
3784 		/*
3785 		 * Get the filesystem root vnode. We can vput() it
3786 		 * immediately, since with rootrefs > 0, it won't go away.
3787 		 */
3788 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3789 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3790 			    __func__, error);
3791 			return (error);
3792 		}
3793 		vput(rootvp);
3794 	}
3795 loop:
3796 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3797 		vholdl(vp);
3798 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3799 		if (error) {
3800 			vdrop(vp);
3801 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3802 			goto loop;
3803 		}
3804 		/*
3805 		 * Skip over a vnodes marked VV_SYSTEM.
3806 		 */
3807 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3808 			VOP_UNLOCK(vp);
3809 			vdrop(vp);
3810 			continue;
3811 		}
3812 		/*
3813 		 * If WRITECLOSE is set, flush out unlinked but still open
3814 		 * files (even if open only for reading) and regular file
3815 		 * vnodes open for writing.
3816 		 */
3817 		if (flags & WRITECLOSE) {
3818 			if (vp->v_object != NULL) {
3819 				VM_OBJECT_WLOCK(vp->v_object);
3820 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3821 				VM_OBJECT_WUNLOCK(vp->v_object);
3822 			}
3823 			do {
3824 				error = VOP_FSYNC(vp, MNT_WAIT, td);
3825 			} while (error == ERELOOKUP);
3826 			if (error != 0) {
3827 				VOP_UNLOCK(vp);
3828 				vdrop(vp);
3829 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3830 				return (error);
3831 			}
3832 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3833 			VI_LOCK(vp);
3834 
3835 			if ((vp->v_type == VNON ||
3836 			    (error == 0 && vattr.va_nlink > 0)) &&
3837 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3838 				VOP_UNLOCK(vp);
3839 				vdropl(vp);
3840 				continue;
3841 			}
3842 		} else
3843 			VI_LOCK(vp);
3844 		/*
3845 		 * With v_usecount == 0, all we need to do is clear out the
3846 		 * vnode data structures and we are done.
3847 		 *
3848 		 * If FORCECLOSE is set, forcibly close the vnode.
3849 		 */
3850 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3851 			vgonel(vp);
3852 		} else {
3853 			busy++;
3854 #ifdef DIAGNOSTIC
3855 			if (busyprt)
3856 				vn_printf(vp, "vflush: busy vnode ");
3857 #endif
3858 		}
3859 		VOP_UNLOCK(vp);
3860 		vdropl(vp);
3861 	}
3862 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3863 		/*
3864 		 * If just the root vnode is busy, and if its refcount
3865 		 * is equal to `rootrefs', then go ahead and kill it.
3866 		 */
3867 		VI_LOCK(rootvp);
3868 		KASSERT(busy > 0, ("vflush: not busy"));
3869 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3870 		    ("vflush: usecount %d < rootrefs %d",
3871 		     rootvp->v_usecount, rootrefs));
3872 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3873 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3874 			vgone(rootvp);
3875 			VOP_UNLOCK(rootvp);
3876 			busy = 0;
3877 		} else
3878 			VI_UNLOCK(rootvp);
3879 	}
3880 	if (busy) {
3881 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3882 		    busy);
3883 		return (EBUSY);
3884 	}
3885 	for (; rootrefs > 0; rootrefs--)
3886 		vrele(rootvp);
3887 	return (0);
3888 }
3889 
3890 /*
3891  * Recycle an unused vnode to the front of the free list.
3892  */
3893 int
3894 vrecycle(struct vnode *vp)
3895 {
3896 	int recycled;
3897 
3898 	VI_LOCK(vp);
3899 	recycled = vrecyclel(vp);
3900 	VI_UNLOCK(vp);
3901 	return (recycled);
3902 }
3903 
3904 /*
3905  * vrecycle, with the vp interlock held.
3906  */
3907 int
3908 vrecyclel(struct vnode *vp)
3909 {
3910 	int recycled;
3911 
3912 	ASSERT_VOP_ELOCKED(vp, __func__);
3913 	ASSERT_VI_LOCKED(vp, __func__);
3914 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3915 	recycled = 0;
3916 	if (vp->v_usecount == 0) {
3917 		recycled = 1;
3918 		vgonel(vp);
3919 	}
3920 	return (recycled);
3921 }
3922 
3923 /*
3924  * Eliminate all activity associated with a vnode
3925  * in preparation for reuse.
3926  */
3927 void
3928 vgone(struct vnode *vp)
3929 {
3930 	VI_LOCK(vp);
3931 	vgonel(vp);
3932 	VI_UNLOCK(vp);
3933 }
3934 
3935 /*
3936  * Notify upper mounts about reclaimed or unlinked vnode.
3937  */
3938 void
3939 vfs_notify_upper(struct vnode *vp, int event)
3940 {
3941 	struct mount *mp;
3942 	struct mount_upper_node *ump;
3943 
3944 	mp = atomic_load_ptr(&vp->v_mount);
3945 	if (mp == NULL)
3946 		return;
3947 	if (TAILQ_EMPTY(&mp->mnt_notify))
3948 		return;
3949 
3950 	MNT_ILOCK(mp);
3951 	mp->mnt_upper_pending++;
3952 	KASSERT(mp->mnt_upper_pending > 0,
3953 	    ("%s: mnt_upper_pending %d", __func__, mp->mnt_upper_pending));
3954 	TAILQ_FOREACH(ump, &mp->mnt_notify, mnt_upper_link) {
3955 		MNT_IUNLOCK(mp);
3956 		switch (event) {
3957 		case VFS_NOTIFY_UPPER_RECLAIM:
3958 			VFS_RECLAIM_LOWERVP(ump->mp, vp);
3959 			break;
3960 		case VFS_NOTIFY_UPPER_UNLINK:
3961 			VFS_UNLINK_LOWERVP(ump->mp, vp);
3962 			break;
3963 		default:
3964 			KASSERT(0, ("invalid event %d", event));
3965 			break;
3966 		}
3967 		MNT_ILOCK(mp);
3968 	}
3969 	mp->mnt_upper_pending--;
3970 	if ((mp->mnt_kern_flag & MNTK_UPPER_WAITER) != 0 &&
3971 	    mp->mnt_upper_pending == 0) {
3972 		mp->mnt_kern_flag &= ~MNTK_UPPER_WAITER;
3973 		wakeup(&mp->mnt_uppers);
3974 	}
3975 	MNT_IUNLOCK(mp);
3976 }
3977 
3978 /*
3979  * vgone, with the vp interlock held.
3980  */
3981 static void
3982 vgonel(struct vnode *vp)
3983 {
3984 	struct thread *td;
3985 	struct mount *mp;
3986 	vm_object_t object;
3987 	bool active, doinginact, oweinact;
3988 
3989 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3990 	ASSERT_VI_LOCKED(vp, "vgonel");
3991 	VNASSERT(vp->v_holdcnt, vp,
3992 	    ("vgonel: vp %p has no reference.", vp));
3993 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3994 	td = curthread;
3995 
3996 	/*
3997 	 * Don't vgonel if we're already doomed.
3998 	 */
3999 	if (VN_IS_DOOMED(vp))
4000 		return;
4001 	/*
4002 	 * Paired with freevnode.
4003 	 */
4004 	vn_seqc_write_begin_locked(vp);
4005 	vunlazy_gone(vp);
4006 	vn_irflag_set_locked(vp, VIRF_DOOMED);
4007 
4008 	/*
4009 	 * Check to see if the vnode is in use.  If so, we have to
4010 	 * call VOP_CLOSE() and VOP_INACTIVE().
4011 	 *
4012 	 * It could be that VOP_INACTIVE() requested reclamation, in
4013 	 * which case we should avoid recursion, so check
4014 	 * VI_DOINGINACT.  This is not precise but good enough.
4015 	 */
4016 	active = vp->v_usecount > 0;
4017 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4018 	doinginact = (vp->v_iflag & VI_DOINGINACT) != 0;
4019 
4020 	/*
4021 	 * If we need to do inactive VI_OWEINACT will be set.
4022 	 */
4023 	if (vp->v_iflag & VI_DEFINACT) {
4024 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
4025 		vp->v_iflag &= ~VI_DEFINACT;
4026 		vdropl(vp);
4027 	} else {
4028 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
4029 		VI_UNLOCK(vp);
4030 	}
4031 	cache_purge_vgone(vp);
4032 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
4033 
4034 	/*
4035 	 * If purging an active vnode, it must be closed and
4036 	 * deactivated before being reclaimed.
4037 	 */
4038 	if (active)
4039 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
4040 	if (!doinginact) {
4041 		do {
4042 			if (oweinact || active) {
4043 				VI_LOCK(vp);
4044 				vinactivef(vp);
4045 				oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4046 				VI_UNLOCK(vp);
4047 			}
4048 		} while (oweinact);
4049 	}
4050 	if (vp->v_type == VSOCK)
4051 		vfs_unp_reclaim(vp);
4052 
4053 	/*
4054 	 * Clean out any buffers associated with the vnode.
4055 	 * If the flush fails, just toss the buffers.
4056 	 */
4057 	mp = NULL;
4058 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
4059 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
4060 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
4061 		while (vinvalbuf(vp, 0, 0, 0) != 0)
4062 			;
4063 	}
4064 
4065 	BO_LOCK(&vp->v_bufobj);
4066 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
4067 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
4068 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
4069 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
4070 	    ("vp %p bufobj not invalidated", vp));
4071 
4072 	/*
4073 	 * For VMIO bufobj, BO_DEAD is set later, or in
4074 	 * vm_object_terminate() after the object's page queue is
4075 	 * flushed.
4076 	 */
4077 	object = vp->v_bufobj.bo_object;
4078 	if (object == NULL)
4079 		vp->v_bufobj.bo_flag |= BO_DEAD;
4080 	BO_UNLOCK(&vp->v_bufobj);
4081 
4082 	/*
4083 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
4084 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
4085 	 * should not touch the object borrowed from the lower vnode
4086 	 * (the handle check).
4087 	 */
4088 	if (object != NULL && object->type == OBJT_VNODE &&
4089 	    object->handle == vp)
4090 		vnode_destroy_vobject(vp);
4091 
4092 	/*
4093 	 * Reclaim the vnode.
4094 	 */
4095 	if (VOP_RECLAIM(vp))
4096 		panic("vgone: cannot reclaim");
4097 	if (mp != NULL)
4098 		vn_finished_secondary_write(mp);
4099 	VNASSERT(vp->v_object == NULL, vp,
4100 	    ("vop_reclaim left v_object vp=%p", vp));
4101 	/*
4102 	 * Clear the advisory locks and wake up waiting threads.
4103 	 */
4104 	(void)VOP_ADVLOCKPURGE(vp);
4105 	vp->v_lockf = NULL;
4106 	/*
4107 	 * Delete from old mount point vnode list.
4108 	 */
4109 	delmntque(vp);
4110 	/*
4111 	 * Done with purge, reset to the standard lock and invalidate
4112 	 * the vnode.
4113 	 */
4114 	VI_LOCK(vp);
4115 	vp->v_vnlock = &vp->v_lock;
4116 	vp->v_op = &dead_vnodeops;
4117 	vp->v_type = VBAD;
4118 }
4119 
4120 /*
4121  * Print out a description of a vnode.
4122  */
4123 static const char * const typename[] =
4124 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
4125  "VMARKER"};
4126 
4127 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0,
4128     "new hold count flag not added to vn_printf");
4129 
4130 void
4131 vn_printf(struct vnode *vp, const char *fmt, ...)
4132 {
4133 	va_list ap;
4134 	char buf[256], buf2[16];
4135 	u_long flags;
4136 	u_int holdcnt;
4137 	short irflag;
4138 
4139 	va_start(ap, fmt);
4140 	vprintf(fmt, ap);
4141 	va_end(ap);
4142 	printf("%p: ", (void *)vp);
4143 	printf("type %s\n", typename[vp->v_type]);
4144 	holdcnt = atomic_load_int(&vp->v_holdcnt);
4145 	printf("    usecount %d, writecount %d, refcount %d seqc users %d",
4146 	    vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS,
4147 	    vp->v_seqc_users);
4148 	switch (vp->v_type) {
4149 	case VDIR:
4150 		printf(" mountedhere %p\n", vp->v_mountedhere);
4151 		break;
4152 	case VCHR:
4153 		printf(" rdev %p\n", vp->v_rdev);
4154 		break;
4155 	case VSOCK:
4156 		printf(" socket %p\n", vp->v_unpcb);
4157 		break;
4158 	case VFIFO:
4159 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
4160 		break;
4161 	default:
4162 		printf("\n");
4163 		break;
4164 	}
4165 	buf[0] = '\0';
4166 	buf[1] = '\0';
4167 	if (holdcnt & VHOLD_NO_SMR)
4168 		strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf));
4169 	printf("    hold count flags (%s)\n", buf + 1);
4170 
4171 	buf[0] = '\0';
4172 	buf[1] = '\0';
4173 	irflag = vn_irflag_read(vp);
4174 	if (irflag & VIRF_DOOMED)
4175 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4176 	if (irflag & VIRF_PGREAD)
4177 		strlcat(buf, "|VIRF_PGREAD", sizeof(buf));
4178 	if (irflag & VIRF_MOUNTPOINT)
4179 		strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf));
4180 	if (irflag & VIRF_TEXT_REF)
4181 		strlcat(buf, "|VIRF_TEXT_REF", sizeof(buf));
4182 	flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT | VIRF_TEXT_REF);
4183 	if (flags != 0) {
4184 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4185 		strlcat(buf, buf2, sizeof(buf));
4186 	}
4187 	if (vp->v_vflag & VV_ROOT)
4188 		strlcat(buf, "|VV_ROOT", sizeof(buf));
4189 	if (vp->v_vflag & VV_ISTTY)
4190 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
4191 	if (vp->v_vflag & VV_NOSYNC)
4192 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4193 	if (vp->v_vflag & VV_ETERNALDEV)
4194 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4195 	if (vp->v_vflag & VV_CACHEDLABEL)
4196 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4197 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
4198 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4199 	if (vp->v_vflag & VV_COPYONWRITE)
4200 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4201 	if (vp->v_vflag & VV_SYSTEM)
4202 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4203 	if (vp->v_vflag & VV_PROCDEP)
4204 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4205 	if (vp->v_vflag & VV_DELETED)
4206 		strlcat(buf, "|VV_DELETED", sizeof(buf));
4207 	if (vp->v_vflag & VV_MD)
4208 		strlcat(buf, "|VV_MD", sizeof(buf));
4209 	if (vp->v_vflag & VV_FORCEINSMQ)
4210 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4211 	if (vp->v_vflag & VV_READLINK)
4212 		strlcat(buf, "|VV_READLINK", sizeof(buf));
4213 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4214 	    VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM |
4215 	    VV_PROCDEP | VV_DELETED | VV_MD | VV_FORCEINSMQ | VV_READLINK);
4216 	if (flags != 0) {
4217 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4218 		strlcat(buf, buf2, sizeof(buf));
4219 	}
4220 	if (vp->v_iflag & VI_MOUNT)
4221 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
4222 	if (vp->v_iflag & VI_DOINGINACT)
4223 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4224 	if (vp->v_iflag & VI_OWEINACT)
4225 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4226 	if (vp->v_iflag & VI_DEFINACT)
4227 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4228 	if (vp->v_iflag & VI_FOPENING)
4229 		strlcat(buf, "|VI_FOPENING", sizeof(buf));
4230 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOINGINACT |
4231 	    VI_OWEINACT | VI_DEFINACT | VI_FOPENING);
4232 	if (flags != 0) {
4233 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4234 		strlcat(buf, buf2, sizeof(buf));
4235 	}
4236 	if (vp->v_mflag & VMP_LAZYLIST)
4237 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4238 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
4239 	if (flags != 0) {
4240 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4241 		strlcat(buf, buf2, sizeof(buf));
4242 	}
4243 	printf("    flags (%s)", buf + 1);
4244 	if (mtx_owned(VI_MTX(vp)))
4245 		printf(" VI_LOCKed");
4246 	printf("\n");
4247 	if (vp->v_object != NULL)
4248 		printf("    v_object %p ref %d pages %d "
4249 		    "cleanbuf %d dirtybuf %d\n",
4250 		    vp->v_object, vp->v_object->ref_count,
4251 		    vp->v_object->resident_page_count,
4252 		    vp->v_bufobj.bo_clean.bv_cnt,
4253 		    vp->v_bufobj.bo_dirty.bv_cnt);
4254 	printf("    ");
4255 	lockmgr_printinfo(vp->v_vnlock);
4256 	if (vp->v_data != NULL)
4257 		VOP_PRINT(vp);
4258 }
4259 
4260 #ifdef DDB
4261 /*
4262  * List all of the locked vnodes in the system.
4263  * Called when debugging the kernel.
4264  */
4265 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
4266 {
4267 	struct mount *mp;
4268 	struct vnode *vp;
4269 
4270 	/*
4271 	 * Note: because this is DDB, we can't obey the locking semantics
4272 	 * for these structures, which means we could catch an inconsistent
4273 	 * state and dereference a nasty pointer.  Not much to be done
4274 	 * about that.
4275 	 */
4276 	db_printf("Locked vnodes\n");
4277 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4278 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4279 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4280 				vn_printf(vp, "vnode ");
4281 		}
4282 	}
4283 }
4284 
4285 /*
4286  * Show details about the given vnode.
4287  */
4288 DB_SHOW_COMMAND(vnode, db_show_vnode)
4289 {
4290 	struct vnode *vp;
4291 
4292 	if (!have_addr)
4293 		return;
4294 	vp = (struct vnode *)addr;
4295 	vn_printf(vp, "vnode ");
4296 }
4297 
4298 /*
4299  * Show details about the given mount point.
4300  */
4301 DB_SHOW_COMMAND(mount, db_show_mount)
4302 {
4303 	struct mount *mp;
4304 	struct vfsopt *opt;
4305 	struct statfs *sp;
4306 	struct vnode *vp;
4307 	char buf[512];
4308 	uint64_t mflags;
4309 	u_int flags;
4310 
4311 	if (!have_addr) {
4312 		/* No address given, print short info about all mount points. */
4313 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4314 			db_printf("%p %s on %s (%s)\n", mp,
4315 			    mp->mnt_stat.f_mntfromname,
4316 			    mp->mnt_stat.f_mntonname,
4317 			    mp->mnt_stat.f_fstypename);
4318 			if (db_pager_quit)
4319 				break;
4320 		}
4321 		db_printf("\nMore info: show mount <addr>\n");
4322 		return;
4323 	}
4324 
4325 	mp = (struct mount *)addr;
4326 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4327 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4328 
4329 	buf[0] = '\0';
4330 	mflags = mp->mnt_flag;
4331 #define	MNT_FLAG(flag)	do {						\
4332 	if (mflags & (flag)) {						\
4333 		if (buf[0] != '\0')					\
4334 			strlcat(buf, ", ", sizeof(buf));		\
4335 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4336 		mflags &= ~(flag);					\
4337 	}								\
4338 } while (0)
4339 	MNT_FLAG(MNT_RDONLY);
4340 	MNT_FLAG(MNT_SYNCHRONOUS);
4341 	MNT_FLAG(MNT_NOEXEC);
4342 	MNT_FLAG(MNT_NOSUID);
4343 	MNT_FLAG(MNT_NFS4ACLS);
4344 	MNT_FLAG(MNT_UNION);
4345 	MNT_FLAG(MNT_ASYNC);
4346 	MNT_FLAG(MNT_SUIDDIR);
4347 	MNT_FLAG(MNT_SOFTDEP);
4348 	MNT_FLAG(MNT_NOSYMFOLLOW);
4349 	MNT_FLAG(MNT_GJOURNAL);
4350 	MNT_FLAG(MNT_MULTILABEL);
4351 	MNT_FLAG(MNT_ACLS);
4352 	MNT_FLAG(MNT_NOATIME);
4353 	MNT_FLAG(MNT_NOCLUSTERR);
4354 	MNT_FLAG(MNT_NOCLUSTERW);
4355 	MNT_FLAG(MNT_SUJ);
4356 	MNT_FLAG(MNT_EXRDONLY);
4357 	MNT_FLAG(MNT_EXPORTED);
4358 	MNT_FLAG(MNT_DEFEXPORTED);
4359 	MNT_FLAG(MNT_EXPORTANON);
4360 	MNT_FLAG(MNT_EXKERB);
4361 	MNT_FLAG(MNT_EXPUBLIC);
4362 	MNT_FLAG(MNT_LOCAL);
4363 	MNT_FLAG(MNT_QUOTA);
4364 	MNT_FLAG(MNT_ROOTFS);
4365 	MNT_FLAG(MNT_USER);
4366 	MNT_FLAG(MNT_IGNORE);
4367 	MNT_FLAG(MNT_UPDATE);
4368 	MNT_FLAG(MNT_DELEXPORT);
4369 	MNT_FLAG(MNT_RELOAD);
4370 	MNT_FLAG(MNT_FORCE);
4371 	MNT_FLAG(MNT_SNAPSHOT);
4372 	MNT_FLAG(MNT_BYFSID);
4373 #undef MNT_FLAG
4374 	if (mflags != 0) {
4375 		if (buf[0] != '\0')
4376 			strlcat(buf, ", ", sizeof(buf));
4377 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4378 		    "0x%016jx", mflags);
4379 	}
4380 	db_printf("    mnt_flag = %s\n", buf);
4381 
4382 	buf[0] = '\0';
4383 	flags = mp->mnt_kern_flag;
4384 #define	MNT_KERN_FLAG(flag)	do {					\
4385 	if (flags & (flag)) {						\
4386 		if (buf[0] != '\0')					\
4387 			strlcat(buf, ", ", sizeof(buf));		\
4388 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4389 		flags &= ~(flag);					\
4390 	}								\
4391 } while (0)
4392 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4393 	MNT_KERN_FLAG(MNTK_ASYNC);
4394 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4395 	MNT_KERN_FLAG(MNTK_DRAINING);
4396 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4397 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4398 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4399 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4400 	MNT_KERN_FLAG(MNTK_RECURSE);
4401 	MNT_KERN_FLAG(MNTK_UPPER_WAITER);
4402 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
4403 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4404 	MNT_KERN_FLAG(MNTK_FPLOOKUP);
4405 	MNT_KERN_FLAG(MNTK_TASKQUEUE_WAITER);
4406 	MNT_KERN_FLAG(MNTK_NOASYNC);
4407 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4408 	MNT_KERN_FLAG(MNTK_MWAIT);
4409 	MNT_KERN_FLAG(MNTK_SUSPEND);
4410 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4411 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4412 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4413 #undef MNT_KERN_FLAG
4414 	if (flags != 0) {
4415 		if (buf[0] != '\0')
4416 			strlcat(buf, ", ", sizeof(buf));
4417 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4418 		    "0x%08x", flags);
4419 	}
4420 	db_printf("    mnt_kern_flag = %s\n", buf);
4421 
4422 	db_printf("    mnt_opt = ");
4423 	opt = TAILQ_FIRST(mp->mnt_opt);
4424 	if (opt != NULL) {
4425 		db_printf("%s", opt->name);
4426 		opt = TAILQ_NEXT(opt, link);
4427 		while (opt != NULL) {
4428 			db_printf(", %s", opt->name);
4429 			opt = TAILQ_NEXT(opt, link);
4430 		}
4431 	}
4432 	db_printf("\n");
4433 
4434 	sp = &mp->mnt_stat;
4435 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4436 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4437 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4438 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4439 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4440 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4441 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4442 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4443 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4444 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4445 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4446 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4447 
4448 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4449 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4450 	if (jailed(mp->mnt_cred))
4451 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4452 	db_printf(" }\n");
4453 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4454 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4455 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4456 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4457 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4458 	    mp->mnt_lazyvnodelistsize);
4459 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4460 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4461 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4462 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4463 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4464 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4465 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4466 	db_printf("    mnt_secondary_accwrites = %d\n",
4467 	    mp->mnt_secondary_accwrites);
4468 	db_printf("    mnt_gjprovider = %s\n",
4469 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4470 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4471 
4472 	db_printf("\n\nList of active vnodes\n");
4473 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4474 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4475 			vn_printf(vp, "vnode ");
4476 			if (db_pager_quit)
4477 				break;
4478 		}
4479 	}
4480 	db_printf("\n\nList of inactive vnodes\n");
4481 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4482 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4483 			vn_printf(vp, "vnode ");
4484 			if (db_pager_quit)
4485 				break;
4486 		}
4487 	}
4488 }
4489 #endif	/* DDB */
4490 
4491 /*
4492  * Fill in a struct xvfsconf based on a struct vfsconf.
4493  */
4494 static int
4495 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4496 {
4497 	struct xvfsconf xvfsp;
4498 
4499 	bzero(&xvfsp, sizeof(xvfsp));
4500 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4501 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4502 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4503 	xvfsp.vfc_flags = vfsp->vfc_flags;
4504 	/*
4505 	 * These are unused in userland, we keep them
4506 	 * to not break binary compatibility.
4507 	 */
4508 	xvfsp.vfc_vfsops = NULL;
4509 	xvfsp.vfc_next = NULL;
4510 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4511 }
4512 
4513 #ifdef COMPAT_FREEBSD32
4514 struct xvfsconf32 {
4515 	uint32_t	vfc_vfsops;
4516 	char		vfc_name[MFSNAMELEN];
4517 	int32_t		vfc_typenum;
4518 	int32_t		vfc_refcount;
4519 	int32_t		vfc_flags;
4520 	uint32_t	vfc_next;
4521 };
4522 
4523 static int
4524 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4525 {
4526 	struct xvfsconf32 xvfsp;
4527 
4528 	bzero(&xvfsp, sizeof(xvfsp));
4529 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4530 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4531 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4532 	xvfsp.vfc_flags = vfsp->vfc_flags;
4533 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4534 }
4535 #endif
4536 
4537 /*
4538  * Top level filesystem related information gathering.
4539  */
4540 static int
4541 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4542 {
4543 	struct vfsconf *vfsp;
4544 	int error;
4545 
4546 	error = 0;
4547 	vfsconf_slock();
4548 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4549 #ifdef COMPAT_FREEBSD32
4550 		if (req->flags & SCTL_MASK32)
4551 			error = vfsconf2x32(req, vfsp);
4552 		else
4553 #endif
4554 			error = vfsconf2x(req, vfsp);
4555 		if (error)
4556 			break;
4557 	}
4558 	vfsconf_sunlock();
4559 	return (error);
4560 }
4561 
4562 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4563     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4564     "S,xvfsconf", "List of all configured filesystems");
4565 
4566 #ifndef BURN_BRIDGES
4567 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4568 
4569 static int
4570 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4571 {
4572 	int *name = (int *)arg1 - 1;	/* XXX */
4573 	u_int namelen = arg2 + 1;	/* XXX */
4574 	struct vfsconf *vfsp;
4575 
4576 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4577 	    "please rebuild world\n");
4578 
4579 #if 1 || defined(COMPAT_PRELITE2)
4580 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4581 	if (namelen == 1)
4582 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4583 #endif
4584 
4585 	switch (name[1]) {
4586 	case VFS_MAXTYPENUM:
4587 		if (namelen != 2)
4588 			return (ENOTDIR);
4589 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4590 	case VFS_CONF:
4591 		if (namelen != 3)
4592 			return (ENOTDIR);	/* overloaded */
4593 		vfsconf_slock();
4594 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4595 			if (vfsp->vfc_typenum == name[2])
4596 				break;
4597 		}
4598 		vfsconf_sunlock();
4599 		if (vfsp == NULL)
4600 			return (EOPNOTSUPP);
4601 #ifdef COMPAT_FREEBSD32
4602 		if (req->flags & SCTL_MASK32)
4603 			return (vfsconf2x32(req, vfsp));
4604 		else
4605 #endif
4606 			return (vfsconf2x(req, vfsp));
4607 	}
4608 	return (EOPNOTSUPP);
4609 }
4610 
4611 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4612     CTLFLAG_MPSAFE, vfs_sysctl,
4613     "Generic filesystem");
4614 
4615 #if 1 || defined(COMPAT_PRELITE2)
4616 
4617 static int
4618 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4619 {
4620 	int error;
4621 	struct vfsconf *vfsp;
4622 	struct ovfsconf ovfs;
4623 
4624 	vfsconf_slock();
4625 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4626 		bzero(&ovfs, sizeof(ovfs));
4627 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4628 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4629 		ovfs.vfc_index = vfsp->vfc_typenum;
4630 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4631 		ovfs.vfc_flags = vfsp->vfc_flags;
4632 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4633 		if (error != 0) {
4634 			vfsconf_sunlock();
4635 			return (error);
4636 		}
4637 	}
4638 	vfsconf_sunlock();
4639 	return (0);
4640 }
4641 
4642 #endif /* 1 || COMPAT_PRELITE2 */
4643 #endif /* !BURN_BRIDGES */
4644 
4645 #define KINFO_VNODESLOP		10
4646 #ifdef notyet
4647 /*
4648  * Dump vnode list (via sysctl).
4649  */
4650 /* ARGSUSED */
4651 static int
4652 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4653 {
4654 	struct xvnode *xvn;
4655 	struct mount *mp;
4656 	struct vnode *vp;
4657 	int error, len, n;
4658 
4659 	/*
4660 	 * Stale numvnodes access is not fatal here.
4661 	 */
4662 	req->lock = 0;
4663 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4664 	if (!req->oldptr)
4665 		/* Make an estimate */
4666 		return (SYSCTL_OUT(req, 0, len));
4667 
4668 	error = sysctl_wire_old_buffer(req, 0);
4669 	if (error != 0)
4670 		return (error);
4671 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4672 	n = 0;
4673 	mtx_lock(&mountlist_mtx);
4674 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4675 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4676 			continue;
4677 		MNT_ILOCK(mp);
4678 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4679 			if (n == len)
4680 				break;
4681 			vref(vp);
4682 			xvn[n].xv_size = sizeof *xvn;
4683 			xvn[n].xv_vnode = vp;
4684 			xvn[n].xv_id = 0;	/* XXX compat */
4685 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4686 			XV_COPY(usecount);
4687 			XV_COPY(writecount);
4688 			XV_COPY(holdcnt);
4689 			XV_COPY(mount);
4690 			XV_COPY(numoutput);
4691 			XV_COPY(type);
4692 #undef XV_COPY
4693 			xvn[n].xv_flag = vp->v_vflag;
4694 
4695 			switch (vp->v_type) {
4696 			case VREG:
4697 			case VDIR:
4698 			case VLNK:
4699 				break;
4700 			case VBLK:
4701 			case VCHR:
4702 				if (vp->v_rdev == NULL) {
4703 					vrele(vp);
4704 					continue;
4705 				}
4706 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4707 				break;
4708 			case VSOCK:
4709 				xvn[n].xv_socket = vp->v_socket;
4710 				break;
4711 			case VFIFO:
4712 				xvn[n].xv_fifo = vp->v_fifoinfo;
4713 				break;
4714 			case VNON:
4715 			case VBAD:
4716 			default:
4717 				/* shouldn't happen? */
4718 				vrele(vp);
4719 				continue;
4720 			}
4721 			vrele(vp);
4722 			++n;
4723 		}
4724 		MNT_IUNLOCK(mp);
4725 		mtx_lock(&mountlist_mtx);
4726 		vfs_unbusy(mp);
4727 		if (n == len)
4728 			break;
4729 	}
4730 	mtx_unlock(&mountlist_mtx);
4731 
4732 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4733 	free(xvn, M_TEMP);
4734 	return (error);
4735 }
4736 
4737 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4738     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4739     "");
4740 #endif
4741 
4742 static void
4743 unmount_or_warn(struct mount *mp)
4744 {
4745 	int error;
4746 
4747 	error = dounmount(mp, MNT_FORCE, curthread);
4748 	if (error != 0) {
4749 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4750 		if (error == EBUSY)
4751 			printf("BUSY)\n");
4752 		else
4753 			printf("%d)\n", error);
4754 	}
4755 }
4756 
4757 /*
4758  * Unmount all filesystems. The list is traversed in reverse order
4759  * of mounting to avoid dependencies.
4760  */
4761 void
4762 vfs_unmountall(void)
4763 {
4764 	struct mount *mp, *tmp;
4765 
4766 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4767 
4768 	/*
4769 	 * Since this only runs when rebooting, it is not interlocked.
4770 	 */
4771 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4772 		vfs_ref(mp);
4773 
4774 		/*
4775 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4776 		 * unmount of the latter.
4777 		 */
4778 		if (mp == rootdevmp)
4779 			continue;
4780 
4781 		unmount_or_warn(mp);
4782 	}
4783 
4784 	if (rootdevmp != NULL)
4785 		unmount_or_warn(rootdevmp);
4786 }
4787 
4788 static void
4789 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4790 {
4791 
4792 	ASSERT_VI_LOCKED(vp, __func__);
4793 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4794 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
4795 		vdropl(vp);
4796 		return;
4797 	}
4798 	if (vn_lock(vp, lkflags) == 0) {
4799 		VI_LOCK(vp);
4800 		vinactive(vp);
4801 		VOP_UNLOCK(vp);
4802 		vdropl(vp);
4803 		return;
4804 	}
4805 	vdefer_inactive_unlocked(vp);
4806 }
4807 
4808 static int
4809 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4810 {
4811 
4812 	return (vp->v_iflag & VI_DEFINACT);
4813 }
4814 
4815 static void __noinline
4816 vfs_periodic_inactive(struct mount *mp, int flags)
4817 {
4818 	struct vnode *vp, *mvp;
4819 	int lkflags;
4820 
4821 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4822 	if (flags != MNT_WAIT)
4823 		lkflags |= LK_NOWAIT;
4824 
4825 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4826 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
4827 			VI_UNLOCK(vp);
4828 			continue;
4829 		}
4830 		vp->v_iflag &= ~VI_DEFINACT;
4831 		vfs_deferred_inactive(vp, lkflags);
4832 	}
4833 }
4834 
4835 static inline bool
4836 vfs_want_msync(struct vnode *vp)
4837 {
4838 	struct vm_object *obj;
4839 
4840 	/*
4841 	 * This test may be performed without any locks held.
4842 	 * We rely on vm_object's type stability.
4843 	 */
4844 	if (vp->v_vflag & VV_NOSYNC)
4845 		return (false);
4846 	obj = vp->v_object;
4847 	return (obj != NULL && vm_object_mightbedirty(obj));
4848 }
4849 
4850 static int
4851 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4852 {
4853 
4854 	if (vp->v_vflag & VV_NOSYNC)
4855 		return (false);
4856 	if (vp->v_iflag & VI_DEFINACT)
4857 		return (true);
4858 	return (vfs_want_msync(vp));
4859 }
4860 
4861 static void __noinline
4862 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4863 {
4864 	struct vnode *vp, *mvp;
4865 	struct vm_object *obj;
4866 	int lkflags, objflags;
4867 	bool seen_defer;
4868 
4869 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4870 	if (flags != MNT_WAIT) {
4871 		lkflags |= LK_NOWAIT;
4872 		objflags = OBJPC_NOSYNC;
4873 	} else {
4874 		objflags = OBJPC_SYNC;
4875 	}
4876 
4877 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4878 		seen_defer = false;
4879 		if (vp->v_iflag & VI_DEFINACT) {
4880 			vp->v_iflag &= ~VI_DEFINACT;
4881 			seen_defer = true;
4882 		}
4883 		if (!vfs_want_msync(vp)) {
4884 			if (seen_defer)
4885 				vfs_deferred_inactive(vp, lkflags);
4886 			else
4887 				VI_UNLOCK(vp);
4888 			continue;
4889 		}
4890 		if (vget(vp, lkflags) == 0) {
4891 			obj = vp->v_object;
4892 			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4893 				VM_OBJECT_WLOCK(obj);
4894 				vm_object_page_clean(obj, 0, 0, objflags);
4895 				VM_OBJECT_WUNLOCK(obj);
4896 			}
4897 			vput(vp);
4898 			if (seen_defer)
4899 				vdrop(vp);
4900 		} else {
4901 			if (seen_defer)
4902 				vdefer_inactive_unlocked(vp);
4903 		}
4904 	}
4905 }
4906 
4907 void
4908 vfs_periodic(struct mount *mp, int flags)
4909 {
4910 
4911 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4912 
4913 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4914 		vfs_periodic_inactive(mp, flags);
4915 	else
4916 		vfs_periodic_msync_inactive(mp, flags);
4917 }
4918 
4919 static void
4920 destroy_vpollinfo_free(struct vpollinfo *vi)
4921 {
4922 
4923 	knlist_destroy(&vi->vpi_selinfo.si_note);
4924 	mtx_destroy(&vi->vpi_lock);
4925 	free(vi, M_VNODEPOLL);
4926 }
4927 
4928 static void
4929 destroy_vpollinfo(struct vpollinfo *vi)
4930 {
4931 
4932 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4933 	seldrain(&vi->vpi_selinfo);
4934 	destroy_vpollinfo_free(vi);
4935 }
4936 
4937 /*
4938  * Initialize per-vnode helper structure to hold poll-related state.
4939  */
4940 void
4941 v_addpollinfo(struct vnode *vp)
4942 {
4943 	struct vpollinfo *vi;
4944 
4945 	if (vp->v_pollinfo != NULL)
4946 		return;
4947 	vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO);
4948 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4949 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4950 	    vfs_knlunlock, vfs_knl_assert_lock);
4951 	VI_LOCK(vp);
4952 	if (vp->v_pollinfo != NULL) {
4953 		VI_UNLOCK(vp);
4954 		destroy_vpollinfo_free(vi);
4955 		return;
4956 	}
4957 	vp->v_pollinfo = vi;
4958 	VI_UNLOCK(vp);
4959 }
4960 
4961 /*
4962  * Record a process's interest in events which might happen to
4963  * a vnode.  Because poll uses the historic select-style interface
4964  * internally, this routine serves as both the ``check for any
4965  * pending events'' and the ``record my interest in future events''
4966  * functions.  (These are done together, while the lock is held,
4967  * to avoid race conditions.)
4968  */
4969 int
4970 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4971 {
4972 
4973 	v_addpollinfo(vp);
4974 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4975 	if (vp->v_pollinfo->vpi_revents & events) {
4976 		/*
4977 		 * This leaves events we are not interested
4978 		 * in available for the other process which
4979 		 * which presumably had requested them
4980 		 * (otherwise they would never have been
4981 		 * recorded).
4982 		 */
4983 		events &= vp->v_pollinfo->vpi_revents;
4984 		vp->v_pollinfo->vpi_revents &= ~events;
4985 
4986 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4987 		return (events);
4988 	}
4989 	vp->v_pollinfo->vpi_events |= events;
4990 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4991 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4992 	return (0);
4993 }
4994 
4995 /*
4996  * Routine to create and manage a filesystem syncer vnode.
4997  */
4998 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4999 static int	sync_fsync(struct  vop_fsync_args *);
5000 static int	sync_inactive(struct  vop_inactive_args *);
5001 static int	sync_reclaim(struct  vop_reclaim_args *);
5002 
5003 static struct vop_vector sync_vnodeops = {
5004 	.vop_bypass =	VOP_EOPNOTSUPP,
5005 	.vop_close =	sync_close,		/* close */
5006 	.vop_fsync =	sync_fsync,		/* fsync */
5007 	.vop_inactive =	sync_inactive,	/* inactive */
5008 	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
5009 	.vop_reclaim =	sync_reclaim,	/* reclaim */
5010 	.vop_lock1 =	vop_stdlock,	/* lock */
5011 	.vop_unlock =	vop_stdunlock,	/* unlock */
5012 	.vop_islocked =	vop_stdislocked,	/* islocked */
5013 };
5014 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
5015 
5016 /*
5017  * Create a new filesystem syncer vnode for the specified mount point.
5018  */
5019 void
5020 vfs_allocate_syncvnode(struct mount *mp)
5021 {
5022 	struct vnode *vp;
5023 	struct bufobj *bo;
5024 	static long start, incr, next;
5025 	int error;
5026 
5027 	/* Allocate a new vnode */
5028 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
5029 	if (error != 0)
5030 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
5031 	vp->v_type = VNON;
5032 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5033 	vp->v_vflag |= VV_FORCEINSMQ;
5034 	error = insmntque(vp, mp);
5035 	if (error != 0)
5036 		panic("vfs_allocate_syncvnode: insmntque() failed");
5037 	vp->v_vflag &= ~VV_FORCEINSMQ;
5038 	VOP_UNLOCK(vp);
5039 	/*
5040 	 * Place the vnode onto the syncer worklist. We attempt to
5041 	 * scatter them about on the list so that they will go off
5042 	 * at evenly distributed times even if all the filesystems
5043 	 * are mounted at once.
5044 	 */
5045 	next += incr;
5046 	if (next == 0 || next > syncer_maxdelay) {
5047 		start /= 2;
5048 		incr /= 2;
5049 		if (start == 0) {
5050 			start = syncer_maxdelay / 2;
5051 			incr = syncer_maxdelay;
5052 		}
5053 		next = start;
5054 	}
5055 	bo = &vp->v_bufobj;
5056 	BO_LOCK(bo);
5057 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
5058 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
5059 	mtx_lock(&sync_mtx);
5060 	sync_vnode_count++;
5061 	if (mp->mnt_syncer == NULL) {
5062 		mp->mnt_syncer = vp;
5063 		vp = NULL;
5064 	}
5065 	mtx_unlock(&sync_mtx);
5066 	BO_UNLOCK(bo);
5067 	if (vp != NULL) {
5068 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5069 		vgone(vp);
5070 		vput(vp);
5071 	}
5072 }
5073 
5074 void
5075 vfs_deallocate_syncvnode(struct mount *mp)
5076 {
5077 	struct vnode *vp;
5078 
5079 	mtx_lock(&sync_mtx);
5080 	vp = mp->mnt_syncer;
5081 	if (vp != NULL)
5082 		mp->mnt_syncer = NULL;
5083 	mtx_unlock(&sync_mtx);
5084 	if (vp != NULL)
5085 		vrele(vp);
5086 }
5087 
5088 /*
5089  * Do a lazy sync of the filesystem.
5090  */
5091 static int
5092 sync_fsync(struct vop_fsync_args *ap)
5093 {
5094 	struct vnode *syncvp = ap->a_vp;
5095 	struct mount *mp = syncvp->v_mount;
5096 	int error, save;
5097 	struct bufobj *bo;
5098 
5099 	/*
5100 	 * We only need to do something if this is a lazy evaluation.
5101 	 */
5102 	if (ap->a_waitfor != MNT_LAZY)
5103 		return (0);
5104 
5105 	/*
5106 	 * Move ourselves to the back of the sync list.
5107 	 */
5108 	bo = &syncvp->v_bufobj;
5109 	BO_LOCK(bo);
5110 	vn_syncer_add_to_worklist(bo, syncdelay);
5111 	BO_UNLOCK(bo);
5112 
5113 	/*
5114 	 * Walk the list of vnodes pushing all that are dirty and
5115 	 * not already on the sync list.
5116 	 */
5117 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
5118 		return (0);
5119 	VOP_UNLOCK(syncvp);
5120 	save = curthread_pflags_set(TDP_SYNCIO);
5121 	/*
5122 	 * The filesystem at hand may be idle with free vnodes stored in the
5123 	 * batch.  Return them instead of letting them stay there indefinitely.
5124 	 */
5125 	vfs_periodic(mp, MNT_NOWAIT);
5126 	error = VFS_SYNC(mp, MNT_LAZY);
5127 	curthread_pflags_restore(save);
5128 	vn_lock(syncvp, LK_EXCLUSIVE | LK_RETRY);
5129 	vfs_unbusy(mp);
5130 	return (error);
5131 }
5132 
5133 /*
5134  * The syncer vnode is no referenced.
5135  */
5136 static int
5137 sync_inactive(struct vop_inactive_args *ap)
5138 {
5139 
5140 	vgone(ap->a_vp);
5141 	return (0);
5142 }
5143 
5144 /*
5145  * The syncer vnode is no longer needed and is being decommissioned.
5146  *
5147  * Modifications to the worklist must be protected by sync_mtx.
5148  */
5149 static int
5150 sync_reclaim(struct vop_reclaim_args *ap)
5151 {
5152 	struct vnode *vp = ap->a_vp;
5153 	struct bufobj *bo;
5154 
5155 	bo = &vp->v_bufobj;
5156 	BO_LOCK(bo);
5157 	mtx_lock(&sync_mtx);
5158 	if (vp->v_mount->mnt_syncer == vp)
5159 		vp->v_mount->mnt_syncer = NULL;
5160 	if (bo->bo_flag & BO_ONWORKLST) {
5161 		LIST_REMOVE(bo, bo_synclist);
5162 		syncer_worklist_len--;
5163 		sync_vnode_count--;
5164 		bo->bo_flag &= ~BO_ONWORKLST;
5165 	}
5166 	mtx_unlock(&sync_mtx);
5167 	BO_UNLOCK(bo);
5168 
5169 	return (0);
5170 }
5171 
5172 int
5173 vn_need_pageq_flush(struct vnode *vp)
5174 {
5175 	struct vm_object *obj;
5176 
5177 	obj = vp->v_object;
5178 	return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5179 	    vm_object_mightbedirty(obj));
5180 }
5181 
5182 /*
5183  * Check if vnode represents a disk device
5184  */
5185 bool
5186 vn_isdisk_error(struct vnode *vp, int *errp)
5187 {
5188 	int error;
5189 
5190 	if (vp->v_type != VCHR) {
5191 		error = ENOTBLK;
5192 		goto out;
5193 	}
5194 	error = 0;
5195 	dev_lock();
5196 	if (vp->v_rdev == NULL)
5197 		error = ENXIO;
5198 	else if (vp->v_rdev->si_devsw == NULL)
5199 		error = ENXIO;
5200 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5201 		error = ENOTBLK;
5202 	dev_unlock();
5203 out:
5204 	*errp = error;
5205 	return (error == 0);
5206 }
5207 
5208 bool
5209 vn_isdisk(struct vnode *vp)
5210 {
5211 	int error;
5212 
5213 	return (vn_isdisk_error(vp, &error));
5214 }
5215 
5216 /*
5217  * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
5218  * the comment above cache_fplookup for details.
5219  */
5220 int
5221 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred)
5222 {
5223 	int error;
5224 
5225 	VFS_SMR_ASSERT_ENTERED();
5226 
5227 	/* Check the owner. */
5228 	if (cred->cr_uid == file_uid) {
5229 		if (file_mode & S_IXUSR)
5230 			return (0);
5231 		goto out_error;
5232 	}
5233 
5234 	/* Otherwise, check the groups (first match) */
5235 	if (groupmember(file_gid, cred)) {
5236 		if (file_mode & S_IXGRP)
5237 			return (0);
5238 		goto out_error;
5239 	}
5240 
5241 	/* Otherwise, check everyone else. */
5242 	if (file_mode & S_IXOTH)
5243 		return (0);
5244 out_error:
5245 	/*
5246 	 * Permission check failed, but it is possible denial will get overwritten
5247 	 * (e.g., when root is traversing through a 700 directory owned by someone
5248 	 * else).
5249 	 *
5250 	 * vaccess() calls priv_check_cred which in turn can descent into MAC
5251 	 * modules overriding this result. It's quite unclear what semantics
5252 	 * are allowed for them to operate, thus for safety we don't call them
5253 	 * from within the SMR section. This also means if any such modules
5254 	 * are present, we have to let the regular lookup decide.
5255 	 */
5256 	error = priv_check_cred_vfs_lookup_nomac(cred);
5257 	switch (error) {
5258 	case 0:
5259 		return (0);
5260 	case EAGAIN:
5261 		/*
5262 		 * MAC modules present.
5263 		 */
5264 		return (EAGAIN);
5265 	case EPERM:
5266 		return (EACCES);
5267 	default:
5268 		return (error);
5269 	}
5270 }
5271 
5272 /*
5273  * Common filesystem object access control check routine.  Accepts a
5274  * vnode's type, "mode", uid and gid, requested access mode, and credentials.
5275  * Returns 0 on success, or an errno on failure.
5276  */
5277 int
5278 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5279     accmode_t accmode, struct ucred *cred)
5280 {
5281 	accmode_t dac_granted;
5282 	accmode_t priv_granted;
5283 
5284 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5285 	    ("invalid bit in accmode"));
5286 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5287 	    ("VAPPEND without VWRITE"));
5288 
5289 	/*
5290 	 * Look for a normal, non-privileged way to access the file/directory
5291 	 * as requested.  If it exists, go with that.
5292 	 */
5293 
5294 	dac_granted = 0;
5295 
5296 	/* Check the owner. */
5297 	if (cred->cr_uid == file_uid) {
5298 		dac_granted |= VADMIN;
5299 		if (file_mode & S_IXUSR)
5300 			dac_granted |= VEXEC;
5301 		if (file_mode & S_IRUSR)
5302 			dac_granted |= VREAD;
5303 		if (file_mode & S_IWUSR)
5304 			dac_granted |= (VWRITE | VAPPEND);
5305 
5306 		if ((accmode & dac_granted) == accmode)
5307 			return (0);
5308 
5309 		goto privcheck;
5310 	}
5311 
5312 	/* Otherwise, check the groups (first match) */
5313 	if (groupmember(file_gid, cred)) {
5314 		if (file_mode & S_IXGRP)
5315 			dac_granted |= VEXEC;
5316 		if (file_mode & S_IRGRP)
5317 			dac_granted |= VREAD;
5318 		if (file_mode & S_IWGRP)
5319 			dac_granted |= (VWRITE | VAPPEND);
5320 
5321 		if ((accmode & dac_granted) == accmode)
5322 			return (0);
5323 
5324 		goto privcheck;
5325 	}
5326 
5327 	/* Otherwise, check everyone else. */
5328 	if (file_mode & S_IXOTH)
5329 		dac_granted |= VEXEC;
5330 	if (file_mode & S_IROTH)
5331 		dac_granted |= VREAD;
5332 	if (file_mode & S_IWOTH)
5333 		dac_granted |= (VWRITE | VAPPEND);
5334 	if ((accmode & dac_granted) == accmode)
5335 		return (0);
5336 
5337 privcheck:
5338 	/*
5339 	 * Build a privilege mask to determine if the set of privileges
5340 	 * satisfies the requirements when combined with the granted mask
5341 	 * from above.  For each privilege, if the privilege is required,
5342 	 * bitwise or the request type onto the priv_granted mask.
5343 	 */
5344 	priv_granted = 0;
5345 
5346 	if (type == VDIR) {
5347 		/*
5348 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5349 		 * requests, instead of PRIV_VFS_EXEC.
5350 		 */
5351 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5352 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5353 			priv_granted |= VEXEC;
5354 	} else {
5355 		/*
5356 		 * Ensure that at least one execute bit is on. Otherwise,
5357 		 * a privileged user will always succeed, and we don't want
5358 		 * this to happen unless the file really is executable.
5359 		 */
5360 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5361 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5362 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5363 			priv_granted |= VEXEC;
5364 	}
5365 
5366 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5367 	    !priv_check_cred(cred, PRIV_VFS_READ))
5368 		priv_granted |= VREAD;
5369 
5370 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5371 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5372 		priv_granted |= (VWRITE | VAPPEND);
5373 
5374 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5375 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5376 		priv_granted |= VADMIN;
5377 
5378 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5379 		return (0);
5380 	}
5381 
5382 	return ((accmode & VADMIN) ? EPERM : EACCES);
5383 }
5384 
5385 /*
5386  * Credential check based on process requesting service, and per-attribute
5387  * permissions.
5388  */
5389 int
5390 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5391     struct thread *td, accmode_t accmode)
5392 {
5393 
5394 	/*
5395 	 * Kernel-invoked always succeeds.
5396 	 */
5397 	if (cred == NOCRED)
5398 		return (0);
5399 
5400 	/*
5401 	 * Do not allow privileged processes in jail to directly manipulate
5402 	 * system attributes.
5403 	 */
5404 	switch (attrnamespace) {
5405 	case EXTATTR_NAMESPACE_SYSTEM:
5406 		/* Potentially should be: return (EPERM); */
5407 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5408 	case EXTATTR_NAMESPACE_USER:
5409 		return (VOP_ACCESS(vp, accmode, cred, td));
5410 	default:
5411 		return (EPERM);
5412 	}
5413 }
5414 
5415 #ifdef DEBUG_VFS_LOCKS
5416 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5417 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5418     "Drop into debugger on lock violation");
5419 
5420 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5421 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5422     0, "Check for interlock across VOPs");
5423 
5424 int vfs_badlock_print = 1;	/* Print lock violations. */
5425 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5426     0, "Print lock violations");
5427 
5428 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5429 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5430     0, "Print vnode details on lock violations");
5431 
5432 #ifdef KDB
5433 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5434 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5435     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5436 #endif
5437 
5438 static void
5439 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5440 {
5441 
5442 #ifdef KDB
5443 	if (vfs_badlock_backtrace)
5444 		kdb_backtrace();
5445 #endif
5446 	if (vfs_badlock_vnode)
5447 		vn_printf(vp, "vnode ");
5448 	if (vfs_badlock_print)
5449 		printf("%s: %p %s\n", str, (void *)vp, msg);
5450 	if (vfs_badlock_ddb)
5451 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5452 }
5453 
5454 void
5455 assert_vi_locked(struct vnode *vp, const char *str)
5456 {
5457 
5458 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5459 		vfs_badlock("interlock is not locked but should be", str, vp);
5460 }
5461 
5462 void
5463 assert_vi_unlocked(struct vnode *vp, const char *str)
5464 {
5465 
5466 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5467 		vfs_badlock("interlock is locked but should not be", str, vp);
5468 }
5469 
5470 void
5471 assert_vop_locked(struct vnode *vp, const char *str)
5472 {
5473 	int locked;
5474 
5475 	if (KERNEL_PANICKED() || vp == NULL)
5476 		return;
5477 
5478 	locked = VOP_ISLOCKED(vp);
5479 	if (locked == 0 || locked == LK_EXCLOTHER)
5480 		vfs_badlock("is not locked but should be", str, vp);
5481 }
5482 
5483 void
5484 assert_vop_unlocked(struct vnode *vp, const char *str)
5485 {
5486 	if (KERNEL_PANICKED() || vp == NULL)
5487 		return;
5488 
5489 	if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5490 		vfs_badlock("is locked but should not be", str, vp);
5491 }
5492 
5493 void
5494 assert_vop_elocked(struct vnode *vp, const char *str)
5495 {
5496 	if (KERNEL_PANICKED() || vp == NULL)
5497 		return;
5498 
5499 	if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5500 		vfs_badlock("is not exclusive locked but should be", str, vp);
5501 }
5502 #endif /* DEBUG_VFS_LOCKS */
5503 
5504 void
5505 vop_rename_fail(struct vop_rename_args *ap)
5506 {
5507 
5508 	if (ap->a_tvp != NULL)
5509 		vput(ap->a_tvp);
5510 	if (ap->a_tdvp == ap->a_tvp)
5511 		vrele(ap->a_tdvp);
5512 	else
5513 		vput(ap->a_tdvp);
5514 	vrele(ap->a_fdvp);
5515 	vrele(ap->a_fvp);
5516 }
5517 
5518 void
5519 vop_rename_pre(void *ap)
5520 {
5521 	struct vop_rename_args *a = ap;
5522 
5523 #ifdef DEBUG_VFS_LOCKS
5524 	if (a->a_tvp)
5525 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5526 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5527 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5528 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5529 
5530 	/* Check the source (from). */
5531 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5532 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5533 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5534 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5535 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5536 
5537 	/* Check the target. */
5538 	if (a->a_tvp)
5539 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5540 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5541 #endif
5542 	/*
5543 	 * It may be tempting to add vn_seqc_write_begin/end calls here and
5544 	 * in vop_rename_post but that's not going to work out since some
5545 	 * filesystems relookup vnodes mid-rename. This is probably a bug.
5546 	 *
5547 	 * For now filesystems are expected to do the relevant calls after they
5548 	 * decide what vnodes to operate on.
5549 	 */
5550 	if (a->a_tdvp != a->a_fdvp)
5551 		vhold(a->a_fdvp);
5552 	if (a->a_tvp != a->a_fvp)
5553 		vhold(a->a_fvp);
5554 	vhold(a->a_tdvp);
5555 	if (a->a_tvp)
5556 		vhold(a->a_tvp);
5557 }
5558 
5559 #ifdef DEBUG_VFS_LOCKS
5560 void
5561 vop_fplookup_vexec_debugpre(void *ap __unused)
5562 {
5563 
5564 	VFS_SMR_ASSERT_ENTERED();
5565 }
5566 
5567 void
5568 vop_fplookup_vexec_debugpost(void *ap __unused, int rc __unused)
5569 {
5570 
5571 	VFS_SMR_ASSERT_ENTERED();
5572 }
5573 
5574 void
5575 vop_fplookup_symlink_debugpre(void *ap __unused)
5576 {
5577 
5578 	VFS_SMR_ASSERT_ENTERED();
5579 }
5580 
5581 void
5582 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused)
5583 {
5584 
5585 	VFS_SMR_ASSERT_ENTERED();
5586 }
5587 
5588 static void
5589 vop_fsync_debugprepost(struct vnode *vp, const char *name)
5590 {
5591 	if (vp->v_type == VCHR)
5592 		;
5593 	else if (MNT_EXTENDED_SHARED(vp->v_mount))
5594 		ASSERT_VOP_LOCKED(vp, name);
5595 	else
5596 		ASSERT_VOP_ELOCKED(vp, name);
5597 }
5598 
5599 void
5600 vop_fsync_debugpre(void *a)
5601 {
5602 	struct vop_fsync_args *ap;
5603 
5604 	ap = a;
5605 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5606 }
5607 
5608 void
5609 vop_fsync_debugpost(void *a, int rc __unused)
5610 {
5611 	struct vop_fsync_args *ap;
5612 
5613 	ap = a;
5614 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5615 }
5616 
5617 void
5618 vop_fdatasync_debugpre(void *a)
5619 {
5620 	struct vop_fdatasync_args *ap;
5621 
5622 	ap = a;
5623 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5624 }
5625 
5626 void
5627 vop_fdatasync_debugpost(void *a, int rc __unused)
5628 {
5629 	struct vop_fdatasync_args *ap;
5630 
5631 	ap = a;
5632 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5633 }
5634 
5635 void
5636 vop_strategy_debugpre(void *ap)
5637 {
5638 	struct vop_strategy_args *a;
5639 	struct buf *bp;
5640 
5641 	a = ap;
5642 	bp = a->a_bp;
5643 
5644 	/*
5645 	 * Cluster ops lock their component buffers but not the IO container.
5646 	 */
5647 	if ((bp->b_flags & B_CLUSTER) != 0)
5648 		return;
5649 
5650 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5651 		if (vfs_badlock_print)
5652 			printf(
5653 			    "VOP_STRATEGY: bp is not locked but should be\n");
5654 		if (vfs_badlock_ddb)
5655 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5656 	}
5657 }
5658 
5659 void
5660 vop_lock_debugpre(void *ap)
5661 {
5662 	struct vop_lock1_args *a = ap;
5663 
5664 	if ((a->a_flags & LK_INTERLOCK) == 0)
5665 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5666 	else
5667 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5668 }
5669 
5670 void
5671 vop_lock_debugpost(void *ap, int rc)
5672 {
5673 	struct vop_lock1_args *a = ap;
5674 
5675 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5676 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5677 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5678 }
5679 
5680 void
5681 vop_unlock_debugpre(void *ap)
5682 {
5683 	struct vop_unlock_args *a = ap;
5684 
5685 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5686 }
5687 
5688 void
5689 vop_need_inactive_debugpre(void *ap)
5690 {
5691 	struct vop_need_inactive_args *a = ap;
5692 
5693 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5694 }
5695 
5696 void
5697 vop_need_inactive_debugpost(void *ap, int rc)
5698 {
5699 	struct vop_need_inactive_args *a = ap;
5700 
5701 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5702 }
5703 #endif
5704 
5705 void
5706 vop_create_pre(void *ap)
5707 {
5708 	struct vop_create_args *a;
5709 	struct vnode *dvp;
5710 
5711 	a = ap;
5712 	dvp = a->a_dvp;
5713 	vn_seqc_write_begin(dvp);
5714 }
5715 
5716 void
5717 vop_create_post(void *ap, int rc)
5718 {
5719 	struct vop_create_args *a;
5720 	struct vnode *dvp;
5721 
5722 	a = ap;
5723 	dvp = a->a_dvp;
5724 	vn_seqc_write_end(dvp);
5725 	if (!rc)
5726 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5727 }
5728 
5729 void
5730 vop_whiteout_pre(void *ap)
5731 {
5732 	struct vop_whiteout_args *a;
5733 	struct vnode *dvp;
5734 
5735 	a = ap;
5736 	dvp = a->a_dvp;
5737 	vn_seqc_write_begin(dvp);
5738 }
5739 
5740 void
5741 vop_whiteout_post(void *ap, int rc)
5742 {
5743 	struct vop_whiteout_args *a;
5744 	struct vnode *dvp;
5745 
5746 	a = ap;
5747 	dvp = a->a_dvp;
5748 	vn_seqc_write_end(dvp);
5749 }
5750 
5751 void
5752 vop_deleteextattr_pre(void *ap)
5753 {
5754 	struct vop_deleteextattr_args *a;
5755 	struct vnode *vp;
5756 
5757 	a = ap;
5758 	vp = a->a_vp;
5759 	vn_seqc_write_begin(vp);
5760 }
5761 
5762 void
5763 vop_deleteextattr_post(void *ap, int rc)
5764 {
5765 	struct vop_deleteextattr_args *a;
5766 	struct vnode *vp;
5767 
5768 	a = ap;
5769 	vp = a->a_vp;
5770 	vn_seqc_write_end(vp);
5771 	if (!rc)
5772 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5773 }
5774 
5775 void
5776 vop_link_pre(void *ap)
5777 {
5778 	struct vop_link_args *a;
5779 	struct vnode *vp, *tdvp;
5780 
5781 	a = ap;
5782 	vp = a->a_vp;
5783 	tdvp = a->a_tdvp;
5784 	vn_seqc_write_begin(vp);
5785 	vn_seqc_write_begin(tdvp);
5786 }
5787 
5788 void
5789 vop_link_post(void *ap, int rc)
5790 {
5791 	struct vop_link_args *a;
5792 	struct vnode *vp, *tdvp;
5793 
5794 	a = ap;
5795 	vp = a->a_vp;
5796 	tdvp = a->a_tdvp;
5797 	vn_seqc_write_end(vp);
5798 	vn_seqc_write_end(tdvp);
5799 	if (!rc) {
5800 		VFS_KNOTE_LOCKED(vp, NOTE_LINK);
5801 		VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE);
5802 	}
5803 }
5804 
5805 void
5806 vop_mkdir_pre(void *ap)
5807 {
5808 	struct vop_mkdir_args *a;
5809 	struct vnode *dvp;
5810 
5811 	a = ap;
5812 	dvp = a->a_dvp;
5813 	vn_seqc_write_begin(dvp);
5814 }
5815 
5816 void
5817 vop_mkdir_post(void *ap, int rc)
5818 {
5819 	struct vop_mkdir_args *a;
5820 	struct vnode *dvp;
5821 
5822 	a = ap;
5823 	dvp = a->a_dvp;
5824 	vn_seqc_write_end(dvp);
5825 	if (!rc)
5826 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5827 }
5828 
5829 #ifdef DEBUG_VFS_LOCKS
5830 void
5831 vop_mkdir_debugpost(void *ap, int rc)
5832 {
5833 	struct vop_mkdir_args *a;
5834 
5835 	a = ap;
5836 	if (!rc)
5837 		cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp);
5838 }
5839 #endif
5840 
5841 void
5842 vop_mknod_pre(void *ap)
5843 {
5844 	struct vop_mknod_args *a;
5845 	struct vnode *dvp;
5846 
5847 	a = ap;
5848 	dvp = a->a_dvp;
5849 	vn_seqc_write_begin(dvp);
5850 }
5851 
5852 void
5853 vop_mknod_post(void *ap, int rc)
5854 {
5855 	struct vop_mknod_args *a;
5856 	struct vnode *dvp;
5857 
5858 	a = ap;
5859 	dvp = a->a_dvp;
5860 	vn_seqc_write_end(dvp);
5861 	if (!rc)
5862 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5863 }
5864 
5865 void
5866 vop_reclaim_post(void *ap, int rc)
5867 {
5868 	struct vop_reclaim_args *a;
5869 	struct vnode *vp;
5870 
5871 	a = ap;
5872 	vp = a->a_vp;
5873 	ASSERT_VOP_IN_SEQC(vp);
5874 	if (!rc)
5875 		VFS_KNOTE_LOCKED(vp, NOTE_REVOKE);
5876 }
5877 
5878 void
5879 vop_remove_pre(void *ap)
5880 {
5881 	struct vop_remove_args *a;
5882 	struct vnode *dvp, *vp;
5883 
5884 	a = ap;
5885 	dvp = a->a_dvp;
5886 	vp = a->a_vp;
5887 	vn_seqc_write_begin(dvp);
5888 	vn_seqc_write_begin(vp);
5889 }
5890 
5891 void
5892 vop_remove_post(void *ap, int rc)
5893 {
5894 	struct vop_remove_args *a;
5895 	struct vnode *dvp, *vp;
5896 
5897 	a = ap;
5898 	dvp = a->a_dvp;
5899 	vp = a->a_vp;
5900 	vn_seqc_write_end(dvp);
5901 	vn_seqc_write_end(vp);
5902 	if (!rc) {
5903 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5904 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5905 	}
5906 }
5907 
5908 void
5909 vop_rename_post(void *ap, int rc)
5910 {
5911 	struct vop_rename_args *a = ap;
5912 	long hint;
5913 
5914 	if (!rc) {
5915 		hint = NOTE_WRITE;
5916 		if (a->a_fdvp == a->a_tdvp) {
5917 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5918 				hint |= NOTE_LINK;
5919 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5920 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5921 		} else {
5922 			hint |= NOTE_EXTEND;
5923 			if (a->a_fvp->v_type == VDIR)
5924 				hint |= NOTE_LINK;
5925 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5926 
5927 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5928 			    a->a_tvp->v_type == VDIR)
5929 				hint &= ~NOTE_LINK;
5930 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5931 		}
5932 
5933 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5934 		if (a->a_tvp)
5935 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5936 	}
5937 	if (a->a_tdvp != a->a_fdvp)
5938 		vdrop(a->a_fdvp);
5939 	if (a->a_tvp != a->a_fvp)
5940 		vdrop(a->a_fvp);
5941 	vdrop(a->a_tdvp);
5942 	if (a->a_tvp)
5943 		vdrop(a->a_tvp);
5944 }
5945 
5946 void
5947 vop_rmdir_pre(void *ap)
5948 {
5949 	struct vop_rmdir_args *a;
5950 	struct vnode *dvp, *vp;
5951 
5952 	a = ap;
5953 	dvp = a->a_dvp;
5954 	vp = a->a_vp;
5955 	vn_seqc_write_begin(dvp);
5956 	vn_seqc_write_begin(vp);
5957 }
5958 
5959 void
5960 vop_rmdir_post(void *ap, int rc)
5961 {
5962 	struct vop_rmdir_args *a;
5963 	struct vnode *dvp, *vp;
5964 
5965 	a = ap;
5966 	dvp = a->a_dvp;
5967 	vp = a->a_vp;
5968 	vn_seqc_write_end(dvp);
5969 	vn_seqc_write_end(vp);
5970 	if (!rc) {
5971 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5972 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5973 	}
5974 }
5975 
5976 void
5977 vop_setattr_pre(void *ap)
5978 {
5979 	struct vop_setattr_args *a;
5980 	struct vnode *vp;
5981 
5982 	a = ap;
5983 	vp = a->a_vp;
5984 	vn_seqc_write_begin(vp);
5985 }
5986 
5987 void
5988 vop_setattr_post(void *ap, int rc)
5989 {
5990 	struct vop_setattr_args *a;
5991 	struct vnode *vp;
5992 
5993 	a = ap;
5994 	vp = a->a_vp;
5995 	vn_seqc_write_end(vp);
5996 	if (!rc)
5997 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
5998 }
5999 
6000 void
6001 vop_setacl_pre(void *ap)
6002 {
6003 	struct vop_setacl_args *a;
6004 	struct vnode *vp;
6005 
6006 	a = ap;
6007 	vp = a->a_vp;
6008 	vn_seqc_write_begin(vp);
6009 }
6010 
6011 void
6012 vop_setacl_post(void *ap, int rc __unused)
6013 {
6014 	struct vop_setacl_args *a;
6015 	struct vnode *vp;
6016 
6017 	a = ap;
6018 	vp = a->a_vp;
6019 	vn_seqc_write_end(vp);
6020 }
6021 
6022 void
6023 vop_setextattr_pre(void *ap)
6024 {
6025 	struct vop_setextattr_args *a;
6026 	struct vnode *vp;
6027 
6028 	a = ap;
6029 	vp = a->a_vp;
6030 	vn_seqc_write_begin(vp);
6031 }
6032 
6033 void
6034 vop_setextattr_post(void *ap, int rc)
6035 {
6036 	struct vop_setextattr_args *a;
6037 	struct vnode *vp;
6038 
6039 	a = ap;
6040 	vp = a->a_vp;
6041 	vn_seqc_write_end(vp);
6042 	if (!rc)
6043 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
6044 }
6045 
6046 void
6047 vop_symlink_pre(void *ap)
6048 {
6049 	struct vop_symlink_args *a;
6050 	struct vnode *dvp;
6051 
6052 	a = ap;
6053 	dvp = a->a_dvp;
6054 	vn_seqc_write_begin(dvp);
6055 }
6056 
6057 void
6058 vop_symlink_post(void *ap, int rc)
6059 {
6060 	struct vop_symlink_args *a;
6061 	struct vnode *dvp;
6062 
6063 	a = ap;
6064 	dvp = a->a_dvp;
6065 	vn_seqc_write_end(dvp);
6066 	if (!rc)
6067 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6068 }
6069 
6070 void
6071 vop_open_post(void *ap, int rc)
6072 {
6073 	struct vop_open_args *a = ap;
6074 
6075 	if (!rc)
6076 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
6077 }
6078 
6079 void
6080 vop_close_post(void *ap, int rc)
6081 {
6082 	struct vop_close_args *a = ap;
6083 
6084 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
6085 	    !VN_IS_DOOMED(a->a_vp))) {
6086 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
6087 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
6088 	}
6089 }
6090 
6091 void
6092 vop_read_post(void *ap, int rc)
6093 {
6094 	struct vop_read_args *a = ap;
6095 
6096 	if (!rc)
6097 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6098 }
6099 
6100 void
6101 vop_read_pgcache_post(void *ap, int rc)
6102 {
6103 	struct vop_read_pgcache_args *a = ap;
6104 
6105 	if (!rc)
6106 		VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ);
6107 }
6108 
6109 void
6110 vop_readdir_post(void *ap, int rc)
6111 {
6112 	struct vop_readdir_args *a = ap;
6113 
6114 	if (!rc)
6115 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6116 }
6117 
6118 static struct knlist fs_knlist;
6119 
6120 static void
6121 vfs_event_init(void *arg)
6122 {
6123 	knlist_init_mtx(&fs_knlist, NULL);
6124 }
6125 /* XXX - correct order? */
6126 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
6127 
6128 void
6129 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
6130 {
6131 
6132 	KNOTE_UNLOCKED(&fs_knlist, event);
6133 }
6134 
6135 static int	filt_fsattach(struct knote *kn);
6136 static void	filt_fsdetach(struct knote *kn);
6137 static int	filt_fsevent(struct knote *kn, long hint);
6138 
6139 struct filterops fs_filtops = {
6140 	.f_isfd = 0,
6141 	.f_attach = filt_fsattach,
6142 	.f_detach = filt_fsdetach,
6143 	.f_event = filt_fsevent
6144 };
6145 
6146 static int
6147 filt_fsattach(struct knote *kn)
6148 {
6149 
6150 	kn->kn_flags |= EV_CLEAR;
6151 	knlist_add(&fs_knlist, kn, 0);
6152 	return (0);
6153 }
6154 
6155 static void
6156 filt_fsdetach(struct knote *kn)
6157 {
6158 
6159 	knlist_remove(&fs_knlist, kn, 0);
6160 }
6161 
6162 static int
6163 filt_fsevent(struct knote *kn, long hint)
6164 {
6165 
6166 	kn->kn_fflags |= kn->kn_sfflags & hint;
6167 
6168 	return (kn->kn_fflags != 0);
6169 }
6170 
6171 static int
6172 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
6173 {
6174 	struct vfsidctl vc;
6175 	int error;
6176 	struct mount *mp;
6177 
6178 	error = SYSCTL_IN(req, &vc, sizeof(vc));
6179 	if (error)
6180 		return (error);
6181 	if (vc.vc_vers != VFS_CTL_VERS1)
6182 		return (EINVAL);
6183 	mp = vfs_getvfs(&vc.vc_fsid);
6184 	if (mp == NULL)
6185 		return (ENOENT);
6186 	/* ensure that a specific sysctl goes to the right filesystem. */
6187 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
6188 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
6189 		vfs_rel(mp);
6190 		return (EINVAL);
6191 	}
6192 	VCTLTOREQ(&vc, req);
6193 	error = VFS_SYSCTL(mp, vc.vc_op, req);
6194 	vfs_rel(mp);
6195 	return (error);
6196 }
6197 
6198 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
6199     NULL, 0, sysctl_vfs_ctl, "",
6200     "Sysctl by fsid");
6201 
6202 /*
6203  * Function to initialize a va_filerev field sensibly.
6204  * XXX: Wouldn't a random number make a lot more sense ??
6205  */
6206 u_quad_t
6207 init_va_filerev(void)
6208 {
6209 	struct bintime bt;
6210 
6211 	getbinuptime(&bt);
6212 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
6213 }
6214 
6215 static int	filt_vfsread(struct knote *kn, long hint);
6216 static int	filt_vfswrite(struct knote *kn, long hint);
6217 static int	filt_vfsvnode(struct knote *kn, long hint);
6218 static void	filt_vfsdetach(struct knote *kn);
6219 static struct filterops vfsread_filtops = {
6220 	.f_isfd = 1,
6221 	.f_detach = filt_vfsdetach,
6222 	.f_event = filt_vfsread
6223 };
6224 static struct filterops vfswrite_filtops = {
6225 	.f_isfd = 1,
6226 	.f_detach = filt_vfsdetach,
6227 	.f_event = filt_vfswrite
6228 };
6229 static struct filterops vfsvnode_filtops = {
6230 	.f_isfd = 1,
6231 	.f_detach = filt_vfsdetach,
6232 	.f_event = filt_vfsvnode
6233 };
6234 
6235 static void
6236 vfs_knllock(void *arg)
6237 {
6238 	struct vnode *vp = arg;
6239 
6240 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
6241 }
6242 
6243 static void
6244 vfs_knlunlock(void *arg)
6245 {
6246 	struct vnode *vp = arg;
6247 
6248 	VOP_UNLOCK(vp);
6249 }
6250 
6251 static void
6252 vfs_knl_assert_lock(void *arg, int what)
6253 {
6254 #ifdef DEBUG_VFS_LOCKS
6255 	struct vnode *vp = arg;
6256 
6257 	if (what == LA_LOCKED)
6258 		ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
6259 	else
6260 		ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
6261 #endif
6262 }
6263 
6264 int
6265 vfs_kqfilter(struct vop_kqfilter_args *ap)
6266 {
6267 	struct vnode *vp = ap->a_vp;
6268 	struct knote *kn = ap->a_kn;
6269 	struct knlist *knl;
6270 
6271 	KASSERT(vp->v_type != VFIFO || (kn->kn_filter != EVFILT_READ &&
6272 	    kn->kn_filter != EVFILT_WRITE),
6273 	    ("READ/WRITE filter on a FIFO leaked through"));
6274 	switch (kn->kn_filter) {
6275 	case EVFILT_READ:
6276 		kn->kn_fop = &vfsread_filtops;
6277 		break;
6278 	case EVFILT_WRITE:
6279 		kn->kn_fop = &vfswrite_filtops;
6280 		break;
6281 	case EVFILT_VNODE:
6282 		kn->kn_fop = &vfsvnode_filtops;
6283 		break;
6284 	default:
6285 		return (EINVAL);
6286 	}
6287 
6288 	kn->kn_hook = (caddr_t)vp;
6289 
6290 	v_addpollinfo(vp);
6291 	if (vp->v_pollinfo == NULL)
6292 		return (ENOMEM);
6293 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
6294 	vhold(vp);
6295 	knlist_add(knl, kn, 0);
6296 
6297 	return (0);
6298 }
6299 
6300 /*
6301  * Detach knote from vnode
6302  */
6303 static void
6304 filt_vfsdetach(struct knote *kn)
6305 {
6306 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6307 
6308 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
6309 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
6310 	vdrop(vp);
6311 }
6312 
6313 /*ARGSUSED*/
6314 static int
6315 filt_vfsread(struct knote *kn, long hint)
6316 {
6317 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6318 	struct vattr va;
6319 	int res;
6320 
6321 	/*
6322 	 * filesystem is gone, so set the EOF flag and schedule
6323 	 * the knote for deletion.
6324 	 */
6325 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6326 		VI_LOCK(vp);
6327 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6328 		VI_UNLOCK(vp);
6329 		return (1);
6330 	}
6331 
6332 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
6333 		return (0);
6334 
6335 	VI_LOCK(vp);
6336 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
6337 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
6338 	VI_UNLOCK(vp);
6339 	return (res);
6340 }
6341 
6342 /*ARGSUSED*/
6343 static int
6344 filt_vfswrite(struct knote *kn, long hint)
6345 {
6346 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6347 
6348 	VI_LOCK(vp);
6349 
6350 	/*
6351 	 * filesystem is gone, so set the EOF flag and schedule
6352 	 * the knote for deletion.
6353 	 */
6354 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
6355 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6356 
6357 	kn->kn_data = 0;
6358 	VI_UNLOCK(vp);
6359 	return (1);
6360 }
6361 
6362 static int
6363 filt_vfsvnode(struct knote *kn, long hint)
6364 {
6365 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6366 	int res;
6367 
6368 	VI_LOCK(vp);
6369 	if (kn->kn_sfflags & hint)
6370 		kn->kn_fflags |= hint;
6371 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6372 		kn->kn_flags |= EV_EOF;
6373 		VI_UNLOCK(vp);
6374 		return (1);
6375 	}
6376 	res = (kn->kn_fflags != 0);
6377 	VI_UNLOCK(vp);
6378 	return (res);
6379 }
6380 
6381 /*
6382  * Returns whether the directory is empty or not.
6383  * If it is empty, the return value is 0; otherwise
6384  * the return value is an error value (which may
6385  * be ENOTEMPTY).
6386  */
6387 int
6388 vfs_emptydir(struct vnode *vp)
6389 {
6390 	struct uio uio;
6391 	struct iovec iov;
6392 	struct dirent *dirent, *dp, *endp;
6393 	int error, eof;
6394 
6395 	error = 0;
6396 	eof = 0;
6397 
6398 	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
6399 	VNASSERT(vp->v_type == VDIR, vp, ("vp is not a directory"));
6400 
6401 	dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
6402 	iov.iov_base = dirent;
6403 	iov.iov_len = sizeof(struct dirent);
6404 
6405 	uio.uio_iov = &iov;
6406 	uio.uio_iovcnt = 1;
6407 	uio.uio_offset = 0;
6408 	uio.uio_resid = sizeof(struct dirent);
6409 	uio.uio_segflg = UIO_SYSSPACE;
6410 	uio.uio_rw = UIO_READ;
6411 	uio.uio_td = curthread;
6412 
6413 	while (eof == 0 && error == 0) {
6414 		error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
6415 		    NULL, NULL);
6416 		if (error != 0)
6417 			break;
6418 		endp = (void *)((uint8_t *)dirent +
6419 		    sizeof(struct dirent) - uio.uio_resid);
6420 		for (dp = dirent; dp < endp;
6421 		     dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
6422 			if (dp->d_type == DT_WHT)
6423 				continue;
6424 			if (dp->d_namlen == 0)
6425 				continue;
6426 			if (dp->d_type != DT_DIR &&
6427 			    dp->d_type != DT_UNKNOWN) {
6428 				error = ENOTEMPTY;
6429 				break;
6430 			}
6431 			if (dp->d_namlen > 2) {
6432 				error = ENOTEMPTY;
6433 				break;
6434 			}
6435 			if (dp->d_namlen == 1 &&
6436 			    dp->d_name[0] != '.') {
6437 				error = ENOTEMPTY;
6438 				break;
6439 			}
6440 			if (dp->d_namlen == 2 &&
6441 			    dp->d_name[1] != '.') {
6442 				error = ENOTEMPTY;
6443 				break;
6444 			}
6445 			uio.uio_resid = sizeof(struct dirent);
6446 		}
6447 	}
6448 	free(dirent, M_TEMP);
6449 	return (error);
6450 }
6451 
6452 int
6453 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6454 {
6455 	int error;
6456 
6457 	if (dp->d_reclen > ap->a_uio->uio_resid)
6458 		return (ENAMETOOLONG);
6459 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
6460 	if (error) {
6461 		if (ap->a_ncookies != NULL) {
6462 			if (ap->a_cookies != NULL)
6463 				free(ap->a_cookies, M_TEMP);
6464 			ap->a_cookies = NULL;
6465 			*ap->a_ncookies = 0;
6466 		}
6467 		return (error);
6468 	}
6469 	if (ap->a_ncookies == NULL)
6470 		return (0);
6471 
6472 	KASSERT(ap->a_cookies,
6473 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6474 
6475 	*ap->a_cookies = realloc(*ap->a_cookies,
6476 	    (*ap->a_ncookies + 1) * sizeof(uint64_t), M_TEMP, M_WAITOK | M_ZERO);
6477 	(*ap->a_cookies)[*ap->a_ncookies] = off;
6478 	*ap->a_ncookies += 1;
6479 	return (0);
6480 }
6481 
6482 /*
6483  * The purpose of this routine is to remove granularity from accmode_t,
6484  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6485  * VADMIN and VAPPEND.
6486  *
6487  * If it returns 0, the caller is supposed to continue with the usual
6488  * access checks using 'accmode' as modified by this routine.  If it
6489  * returns nonzero value, the caller is supposed to return that value
6490  * as errno.
6491  *
6492  * Note that after this routine runs, accmode may be zero.
6493  */
6494 int
6495 vfs_unixify_accmode(accmode_t *accmode)
6496 {
6497 	/*
6498 	 * There is no way to specify explicit "deny" rule using
6499 	 * file mode or POSIX.1e ACLs.
6500 	 */
6501 	if (*accmode & VEXPLICIT_DENY) {
6502 		*accmode = 0;
6503 		return (0);
6504 	}
6505 
6506 	/*
6507 	 * None of these can be translated into usual access bits.
6508 	 * Also, the common case for NFSv4 ACLs is to not contain
6509 	 * either of these bits. Caller should check for VWRITE
6510 	 * on the containing directory instead.
6511 	 */
6512 	if (*accmode & (VDELETE_CHILD | VDELETE))
6513 		return (EPERM);
6514 
6515 	if (*accmode & VADMIN_PERMS) {
6516 		*accmode &= ~VADMIN_PERMS;
6517 		*accmode |= VADMIN;
6518 	}
6519 
6520 	/*
6521 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6522 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6523 	 */
6524 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6525 
6526 	return (0);
6527 }
6528 
6529 /*
6530  * Clear out a doomed vnode (if any) and replace it with a new one as long
6531  * as the fs is not being unmounted. Return the root vnode to the caller.
6532  */
6533 static int __noinline
6534 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6535 {
6536 	struct vnode *vp;
6537 	int error;
6538 
6539 restart:
6540 	if (mp->mnt_rootvnode != NULL) {
6541 		MNT_ILOCK(mp);
6542 		vp = mp->mnt_rootvnode;
6543 		if (vp != NULL) {
6544 			if (!VN_IS_DOOMED(vp)) {
6545 				vrefact(vp);
6546 				MNT_IUNLOCK(mp);
6547 				error = vn_lock(vp, flags);
6548 				if (error == 0) {
6549 					*vpp = vp;
6550 					return (0);
6551 				}
6552 				vrele(vp);
6553 				goto restart;
6554 			}
6555 			/*
6556 			 * Clear the old one.
6557 			 */
6558 			mp->mnt_rootvnode = NULL;
6559 		}
6560 		MNT_IUNLOCK(mp);
6561 		if (vp != NULL) {
6562 			vfs_op_barrier_wait(mp);
6563 			vrele(vp);
6564 		}
6565 	}
6566 	error = VFS_CACHEDROOT(mp, flags, vpp);
6567 	if (error != 0)
6568 		return (error);
6569 	if (mp->mnt_vfs_ops == 0) {
6570 		MNT_ILOCK(mp);
6571 		if (mp->mnt_vfs_ops != 0) {
6572 			MNT_IUNLOCK(mp);
6573 			return (0);
6574 		}
6575 		if (mp->mnt_rootvnode == NULL) {
6576 			vrefact(*vpp);
6577 			mp->mnt_rootvnode = *vpp;
6578 		} else {
6579 			if (mp->mnt_rootvnode != *vpp) {
6580 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6581 					panic("%s: mismatch between vnode returned "
6582 					    " by VFS_CACHEDROOT and the one cached "
6583 					    " (%p != %p)",
6584 					    __func__, *vpp, mp->mnt_rootvnode);
6585 				}
6586 			}
6587 		}
6588 		MNT_IUNLOCK(mp);
6589 	}
6590 	return (0);
6591 }
6592 
6593 int
6594 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6595 {
6596 	struct mount_pcpu *mpcpu;
6597 	struct vnode *vp;
6598 	int error;
6599 
6600 	if (!vfs_op_thread_enter(mp, mpcpu))
6601 		return (vfs_cache_root_fallback(mp, flags, vpp));
6602 	vp = atomic_load_ptr(&mp->mnt_rootvnode);
6603 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6604 		vfs_op_thread_exit(mp, mpcpu);
6605 		return (vfs_cache_root_fallback(mp, flags, vpp));
6606 	}
6607 	vrefact(vp);
6608 	vfs_op_thread_exit(mp, mpcpu);
6609 	error = vn_lock(vp, flags);
6610 	if (error != 0) {
6611 		vrele(vp);
6612 		return (vfs_cache_root_fallback(mp, flags, vpp));
6613 	}
6614 	*vpp = vp;
6615 	return (0);
6616 }
6617 
6618 struct vnode *
6619 vfs_cache_root_clear(struct mount *mp)
6620 {
6621 	struct vnode *vp;
6622 
6623 	/*
6624 	 * ops > 0 guarantees there is nobody who can see this vnode
6625 	 */
6626 	MPASS(mp->mnt_vfs_ops > 0);
6627 	vp = mp->mnt_rootvnode;
6628 	if (vp != NULL)
6629 		vn_seqc_write_begin(vp);
6630 	mp->mnt_rootvnode = NULL;
6631 	return (vp);
6632 }
6633 
6634 void
6635 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6636 {
6637 
6638 	MPASS(mp->mnt_vfs_ops > 0);
6639 	vrefact(vp);
6640 	mp->mnt_rootvnode = vp;
6641 }
6642 
6643 /*
6644  * These are helper functions for filesystems to traverse all
6645  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6646  *
6647  * This interface replaces MNT_VNODE_FOREACH.
6648  */
6649 
6650 struct vnode *
6651 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6652 {
6653 	struct vnode *vp;
6654 
6655 	if (should_yield())
6656 		kern_yield(PRI_USER);
6657 	MNT_ILOCK(mp);
6658 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6659 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6660 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6661 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6662 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6663 			continue;
6664 		VI_LOCK(vp);
6665 		if (VN_IS_DOOMED(vp)) {
6666 			VI_UNLOCK(vp);
6667 			continue;
6668 		}
6669 		break;
6670 	}
6671 	if (vp == NULL) {
6672 		__mnt_vnode_markerfree_all(mvp, mp);
6673 		/* MNT_IUNLOCK(mp); -- done in above function */
6674 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6675 		return (NULL);
6676 	}
6677 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6678 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6679 	MNT_IUNLOCK(mp);
6680 	return (vp);
6681 }
6682 
6683 struct vnode *
6684 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6685 {
6686 	struct vnode *vp;
6687 
6688 	*mvp = vn_alloc_marker(mp);
6689 	MNT_ILOCK(mp);
6690 	MNT_REF(mp);
6691 
6692 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6693 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6694 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6695 			continue;
6696 		VI_LOCK(vp);
6697 		if (VN_IS_DOOMED(vp)) {
6698 			VI_UNLOCK(vp);
6699 			continue;
6700 		}
6701 		break;
6702 	}
6703 	if (vp == NULL) {
6704 		MNT_REL(mp);
6705 		MNT_IUNLOCK(mp);
6706 		vn_free_marker(*mvp);
6707 		*mvp = NULL;
6708 		return (NULL);
6709 	}
6710 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6711 	MNT_IUNLOCK(mp);
6712 	return (vp);
6713 }
6714 
6715 void
6716 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6717 {
6718 
6719 	if (*mvp == NULL) {
6720 		MNT_IUNLOCK(mp);
6721 		return;
6722 	}
6723 
6724 	mtx_assert(MNT_MTX(mp), MA_OWNED);
6725 
6726 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6727 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6728 	MNT_REL(mp);
6729 	MNT_IUNLOCK(mp);
6730 	vn_free_marker(*mvp);
6731 	*mvp = NULL;
6732 }
6733 
6734 /*
6735  * These are helper functions for filesystems to traverse their
6736  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6737  */
6738 static void
6739 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6740 {
6741 
6742 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6743 
6744 	MNT_ILOCK(mp);
6745 	MNT_REL(mp);
6746 	MNT_IUNLOCK(mp);
6747 	vn_free_marker(*mvp);
6748 	*mvp = NULL;
6749 }
6750 
6751 /*
6752  * Relock the mp mount vnode list lock with the vp vnode interlock in the
6753  * conventional lock order during mnt_vnode_next_lazy iteration.
6754  *
6755  * On entry, the mount vnode list lock is held and the vnode interlock is not.
6756  * The list lock is dropped and reacquired.  On success, both locks are held.
6757  * On failure, the mount vnode list lock is held but the vnode interlock is
6758  * not, and the procedure may have yielded.
6759  */
6760 static bool
6761 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6762     struct vnode *vp)
6763 {
6764 
6765 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6766 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6767 	    ("%s: bad marker", __func__));
6768 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6769 	    ("%s: inappropriate vnode", __func__));
6770 	ASSERT_VI_UNLOCKED(vp, __func__);
6771 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6772 
6773 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6774 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6775 
6776 	/*
6777 	 * Note we may be racing against vdrop which transitioned the hold
6778 	 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6779 	 * if we are the only user after we get the interlock we will just
6780 	 * vdrop.
6781 	 */
6782 	vhold(vp);
6783 	mtx_unlock(&mp->mnt_listmtx);
6784 	VI_LOCK(vp);
6785 	if (VN_IS_DOOMED(vp)) {
6786 		VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6787 		goto out_lost;
6788 	}
6789 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6790 	/*
6791 	 * There is nothing to do if we are the last user.
6792 	 */
6793 	if (!refcount_release_if_not_last(&vp->v_holdcnt))
6794 		goto out_lost;
6795 	mtx_lock(&mp->mnt_listmtx);
6796 	return (true);
6797 out_lost:
6798 	vdropl(vp);
6799 	maybe_yield();
6800 	mtx_lock(&mp->mnt_listmtx);
6801 	return (false);
6802 }
6803 
6804 static struct vnode *
6805 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6806     void *cbarg)
6807 {
6808 	struct vnode *vp;
6809 
6810 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6811 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6812 restart:
6813 	vp = TAILQ_NEXT(*mvp, v_lazylist);
6814 	while (vp != NULL) {
6815 		if (vp->v_type == VMARKER) {
6816 			vp = TAILQ_NEXT(vp, v_lazylist);
6817 			continue;
6818 		}
6819 		/*
6820 		 * See if we want to process the vnode. Note we may encounter a
6821 		 * long string of vnodes we don't care about and hog the list
6822 		 * as a result. Check for it and requeue the marker.
6823 		 */
6824 		VNPASS(!VN_IS_DOOMED(vp), vp);
6825 		if (!cb(vp, cbarg)) {
6826 			if (!should_yield()) {
6827 				vp = TAILQ_NEXT(vp, v_lazylist);
6828 				continue;
6829 			}
6830 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6831 			    v_lazylist);
6832 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6833 			    v_lazylist);
6834 			mtx_unlock(&mp->mnt_listmtx);
6835 			kern_yield(PRI_USER);
6836 			mtx_lock(&mp->mnt_listmtx);
6837 			goto restart;
6838 		}
6839 		/*
6840 		 * Try-lock because this is the wrong lock order.
6841 		 */
6842 		if (!VI_TRYLOCK(vp) &&
6843 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6844 			goto restart;
6845 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6846 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6847 		    ("alien vnode on the lazy list %p %p", vp, mp));
6848 		VNPASS(vp->v_mount == mp, vp);
6849 		VNPASS(!VN_IS_DOOMED(vp), vp);
6850 		break;
6851 	}
6852 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6853 
6854 	/* Check if we are done */
6855 	if (vp == NULL) {
6856 		mtx_unlock(&mp->mnt_listmtx);
6857 		mnt_vnode_markerfree_lazy(mvp, mp);
6858 		return (NULL);
6859 	}
6860 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6861 	mtx_unlock(&mp->mnt_listmtx);
6862 	ASSERT_VI_LOCKED(vp, "lazy iter");
6863 	return (vp);
6864 }
6865 
6866 struct vnode *
6867 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6868     void *cbarg)
6869 {
6870 
6871 	if (should_yield())
6872 		kern_yield(PRI_USER);
6873 	mtx_lock(&mp->mnt_listmtx);
6874 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6875 }
6876 
6877 struct vnode *
6878 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6879     void *cbarg)
6880 {
6881 	struct vnode *vp;
6882 
6883 	if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6884 		return (NULL);
6885 
6886 	*mvp = vn_alloc_marker(mp);
6887 	MNT_ILOCK(mp);
6888 	MNT_REF(mp);
6889 	MNT_IUNLOCK(mp);
6890 
6891 	mtx_lock(&mp->mnt_listmtx);
6892 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6893 	if (vp == NULL) {
6894 		mtx_unlock(&mp->mnt_listmtx);
6895 		mnt_vnode_markerfree_lazy(mvp, mp);
6896 		return (NULL);
6897 	}
6898 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6899 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6900 }
6901 
6902 void
6903 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6904 {
6905 
6906 	if (*mvp == NULL)
6907 		return;
6908 
6909 	mtx_lock(&mp->mnt_listmtx);
6910 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6911 	mtx_unlock(&mp->mnt_listmtx);
6912 	mnt_vnode_markerfree_lazy(mvp, mp);
6913 }
6914 
6915 int
6916 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6917 {
6918 
6919 	if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6920 		cnp->cn_flags &= ~NOEXECCHECK;
6921 		return (0);
6922 	}
6923 
6924 	return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, curthread));
6925 }
6926 
6927 /*
6928  * Do not use this variant unless you have means other than the hold count
6929  * to prevent the vnode from getting freed.
6930  */
6931 void
6932 vn_seqc_write_begin_locked(struct vnode *vp)
6933 {
6934 
6935 	ASSERT_VI_LOCKED(vp, __func__);
6936 	VNPASS(vp->v_holdcnt > 0, vp);
6937 	VNPASS(vp->v_seqc_users >= 0, vp);
6938 	vp->v_seqc_users++;
6939 	if (vp->v_seqc_users == 1)
6940 		seqc_sleepable_write_begin(&vp->v_seqc);
6941 }
6942 
6943 void
6944 vn_seqc_write_begin(struct vnode *vp)
6945 {
6946 
6947 	VI_LOCK(vp);
6948 	vn_seqc_write_begin_locked(vp);
6949 	VI_UNLOCK(vp);
6950 }
6951 
6952 void
6953 vn_seqc_write_end_locked(struct vnode *vp)
6954 {
6955 
6956 	ASSERT_VI_LOCKED(vp, __func__);
6957 	VNPASS(vp->v_seqc_users > 0, vp);
6958 	vp->v_seqc_users--;
6959 	if (vp->v_seqc_users == 0)
6960 		seqc_sleepable_write_end(&vp->v_seqc);
6961 }
6962 
6963 void
6964 vn_seqc_write_end(struct vnode *vp)
6965 {
6966 
6967 	VI_LOCK(vp);
6968 	vn_seqc_write_end_locked(vp);
6969 	VI_UNLOCK(vp);
6970 }
6971 
6972 /*
6973  * Special case handling for allocating and freeing vnodes.
6974  *
6975  * The counter remains unchanged on free so that a doomed vnode will
6976  * keep testing as in modify as long as it is accessible with SMR.
6977  */
6978 static void
6979 vn_seqc_init(struct vnode *vp)
6980 {
6981 
6982 	vp->v_seqc = 0;
6983 	vp->v_seqc_users = 0;
6984 }
6985 
6986 static void
6987 vn_seqc_write_end_free(struct vnode *vp)
6988 {
6989 
6990 	VNPASS(seqc_in_modify(vp->v_seqc), vp);
6991 	VNPASS(vp->v_seqc_users == 1, vp);
6992 }
6993 
6994 void
6995 vn_irflag_set_locked(struct vnode *vp, short toset)
6996 {
6997 	short flags;
6998 
6999 	ASSERT_VI_LOCKED(vp, __func__);
7000 	flags = vn_irflag_read(vp);
7001 	VNASSERT((flags & toset) == 0, vp,
7002 	    ("%s: some of the passed flags already set (have %d, passed %d)\n",
7003 	    __func__, flags, toset));
7004 	atomic_store_short(&vp->v_irflag, flags | toset);
7005 }
7006 
7007 void
7008 vn_irflag_set(struct vnode *vp, short toset)
7009 {
7010 
7011 	VI_LOCK(vp);
7012 	vn_irflag_set_locked(vp, toset);
7013 	VI_UNLOCK(vp);
7014 }
7015 
7016 void
7017 vn_irflag_set_cond_locked(struct vnode *vp, short toset)
7018 {
7019 	short flags;
7020 
7021 	ASSERT_VI_LOCKED(vp, __func__);
7022 	flags = vn_irflag_read(vp);
7023 	atomic_store_short(&vp->v_irflag, flags | toset);
7024 }
7025 
7026 void
7027 vn_irflag_set_cond(struct vnode *vp, short toset)
7028 {
7029 
7030 	VI_LOCK(vp);
7031 	vn_irflag_set_cond_locked(vp, toset);
7032 	VI_UNLOCK(vp);
7033 }
7034 
7035 void
7036 vn_irflag_unset_locked(struct vnode *vp, short tounset)
7037 {
7038 	short flags;
7039 
7040 	ASSERT_VI_LOCKED(vp, __func__);
7041 	flags = vn_irflag_read(vp);
7042 	VNASSERT((flags & tounset) == tounset, vp,
7043 	    ("%s: some of the passed flags not set (have %d, passed %d)\n",
7044 	    __func__, flags, tounset));
7045 	atomic_store_short(&vp->v_irflag, flags & ~tounset);
7046 }
7047 
7048 void
7049 vn_irflag_unset(struct vnode *vp, short tounset)
7050 {
7051 
7052 	VI_LOCK(vp);
7053 	vn_irflag_unset_locked(vp, tounset);
7054 	VI_UNLOCK(vp);
7055 }
7056