xref: /freebsd/sys/kern/vfs_subr.c (revision 05d7dd739c5b89b5cad3153124667d743f3ccb3c)
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/bio.h>
52 #include <sys/buf.h>
53 #include <sys/capsicum.h>
54 #include <sys/condvar.h>
55 #include <sys/conf.h>
56 #include <sys/counter.h>
57 #include <sys/dirent.h>
58 #include <sys/event.h>
59 #include <sys/eventhandler.h>
60 #include <sys/extattr.h>
61 #include <sys/file.h>
62 #include <sys/fcntl.h>
63 #include <sys/jail.h>
64 #include <sys/kdb.h>
65 #include <sys/kernel.h>
66 #include <sys/kthread.h>
67 #include <sys/ktr.h>
68 #include <sys/lockf.h>
69 #include <sys/malloc.h>
70 #include <sys/mount.h>
71 #include <sys/namei.h>
72 #include <sys/pctrie.h>
73 #include <sys/priv.h>
74 #include <sys/reboot.h>
75 #include <sys/refcount.h>
76 #include <sys/rwlock.h>
77 #include <sys/sched.h>
78 #include <sys/sleepqueue.h>
79 #include <sys/smp.h>
80 #include <sys/stat.h>
81 #include <sys/sysctl.h>
82 #include <sys/syslog.h>
83 #include <sys/vmmeter.h>
84 #include <sys/vnode.h>
85 #include <sys/watchdog.h>
86 
87 #include <machine/stdarg.h>
88 
89 #include <security/mac/mac_framework.h>
90 
91 #include <vm/vm.h>
92 #include <vm/vm_object.h>
93 #include <vm/vm_extern.h>
94 #include <vm/pmap.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_page.h>
97 #include <vm/vm_kern.h>
98 #include <vm/uma.h>
99 
100 #ifdef DDB
101 #include <ddb/ddb.h>
102 #endif
103 
104 static void	delmntque(struct vnode *vp);
105 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
106 		    int slpflag, int slptimeo);
107 static void	syncer_shutdown(void *arg, int howto);
108 static int	vtryrecycle(struct vnode *vp);
109 static void	v_init_counters(struct vnode *);
110 static void	v_incr_devcount(struct vnode *);
111 static void	v_decr_devcount(struct vnode *);
112 static void	vgonel(struct vnode *);
113 static void	vfs_knllock(void *arg);
114 static void	vfs_knlunlock(void *arg);
115 static void	vfs_knl_assert_locked(void *arg);
116 static void	vfs_knl_assert_unlocked(void *arg);
117 static void	destroy_vpollinfo(struct vpollinfo *vi);
118 static int	v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
119 		    daddr_t startlbn, daddr_t endlbn);
120 static void	vnlru_recalc(void);
121 
122 /*
123  * These fences are intended for cases where some synchronization is
124  * needed between access of v_iflags and lockless vnode refcount (v_holdcnt
125  * and v_usecount) updates.  Access to v_iflags is generally synchronized
126  * by the interlock, but we have some internal assertions that check vnode
127  * flags without acquiring the lock.  Thus, these fences are INVARIANTS-only
128  * for now.
129  */
130 #ifdef INVARIANTS
131 #define	VNODE_REFCOUNT_FENCE_ACQ()	atomic_thread_fence_acq()
132 #define	VNODE_REFCOUNT_FENCE_REL()	atomic_thread_fence_rel()
133 #else
134 #define	VNODE_REFCOUNT_FENCE_ACQ()
135 #define	VNODE_REFCOUNT_FENCE_REL()
136 #endif
137 
138 /*
139  * Number of vnodes in existence.  Increased whenever getnewvnode()
140  * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
141  */
142 static u_long __exclusive_cache_line numvnodes;
143 
144 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
145     "Number of vnodes in existence");
146 
147 static counter_u64_t vnodes_created;
148 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
149     "Number of vnodes created by getnewvnode");
150 
151 /*
152  * Conversion tables for conversion from vnode types to inode formats
153  * and back.
154  */
155 enum vtype iftovt_tab[16] = {
156 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
157 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
158 };
159 int vttoif_tab[10] = {
160 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
161 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
162 };
163 
164 /*
165  * List of allocates vnodes in the system.
166  */
167 static TAILQ_HEAD(freelst, vnode) vnode_list;
168 static struct vnode *vnode_list_free_marker;
169 static struct vnode *vnode_list_reclaim_marker;
170 
171 /*
172  * "Free" vnode target.  Free vnodes are rarely completely free, but are
173  * just ones that are cheap to recycle.  Usually they are for files which
174  * have been stat'd but not read; these usually have inode and namecache
175  * data attached to them.  This target is the preferred minimum size of a
176  * sub-cache consisting mostly of such files. The system balances the size
177  * of this sub-cache with its complement to try to prevent either from
178  * thrashing while the other is relatively inactive.  The targets express
179  * a preference for the best balance.
180  *
181  * "Above" this target there are 2 further targets (watermarks) related
182  * to recyling of free vnodes.  In the best-operating case, the cache is
183  * exactly full, the free list has size between vlowat and vhiwat above the
184  * free target, and recycling from it and normal use maintains this state.
185  * Sometimes the free list is below vlowat or even empty, but this state
186  * is even better for immediate use provided the cache is not full.
187  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
188  * ones) to reach one of these states.  The watermarks are currently hard-
189  * coded as 4% and 9% of the available space higher.  These and the default
190  * of 25% for wantfreevnodes are too large if the memory size is large.
191  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
192  * whenever vnlru_proc() becomes active.
193  */
194 static long wantfreevnodes;
195 static long __exclusive_cache_line freevnodes;
196 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
197     &freevnodes, 0, "Number of \"free\" vnodes");
198 static long freevnodes_old;
199 
200 static counter_u64_t recycles_count;
201 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
202     "Number of vnodes recycled to meet vnode cache targets");
203 
204 static counter_u64_t recycles_free_count;
205 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count,
206     "Number of free vnodes recycled to meet vnode cache targets");
207 
208 /*
209  * Various variables used for debugging the new implementation of
210  * reassignbuf().
211  * XXX these are probably of (very) limited utility now.
212  */
213 static int reassignbufcalls;
214 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW | CTLFLAG_STATS,
215     &reassignbufcalls, 0, "Number of calls to reassignbuf");
216 
217 static counter_u64_t deferred_inact;
218 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact,
219     "Number of times inactive processing was deferred");
220 
221 /* To keep more than one thread at a time from running vfs_getnewfsid */
222 static struct mtx mntid_mtx;
223 
224 /*
225  * Lock for any access to the following:
226  *	vnode_list
227  *	numvnodes
228  *	freevnodes
229  */
230 static struct mtx __exclusive_cache_line vnode_list_mtx;
231 
232 /* Publicly exported FS */
233 struct nfs_public nfs_pub;
234 
235 static uma_zone_t buf_trie_zone;
236 
237 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
238 static uma_zone_t vnode_zone;
239 static uma_zone_t vnodepoll_zone;
240 
241 /*
242  * The workitem queue.
243  *
244  * It is useful to delay writes of file data and filesystem metadata
245  * for tens of seconds so that quickly created and deleted files need
246  * not waste disk bandwidth being created and removed. To realize this,
247  * we append vnodes to a "workitem" queue. When running with a soft
248  * updates implementation, most pending metadata dependencies should
249  * not wait for more than a few seconds. Thus, mounted on block devices
250  * are delayed only about a half the time that file data is delayed.
251  * Similarly, directory updates are more critical, so are only delayed
252  * about a third the time that file data is delayed. Thus, there are
253  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
254  * one each second (driven off the filesystem syncer process). The
255  * syncer_delayno variable indicates the next queue that is to be processed.
256  * Items that need to be processed soon are placed in this queue:
257  *
258  *	syncer_workitem_pending[syncer_delayno]
259  *
260  * A delay of fifteen seconds is done by placing the request fifteen
261  * entries later in the queue:
262  *
263  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
264  *
265  */
266 static int syncer_delayno;
267 static long syncer_mask;
268 LIST_HEAD(synclist, bufobj);
269 static struct synclist *syncer_workitem_pending;
270 /*
271  * The sync_mtx protects:
272  *	bo->bo_synclist
273  *	sync_vnode_count
274  *	syncer_delayno
275  *	syncer_state
276  *	syncer_workitem_pending
277  *	syncer_worklist_len
278  *	rushjob
279  */
280 static struct mtx sync_mtx;
281 static struct cv sync_wakeup;
282 
283 #define SYNCER_MAXDELAY		32
284 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
285 static int syncdelay = 30;		/* max time to delay syncing data */
286 static int filedelay = 30;		/* time to delay syncing files */
287 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
288     "Time to delay syncing files (in seconds)");
289 static int dirdelay = 29;		/* time to delay syncing directories */
290 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
291     "Time to delay syncing directories (in seconds)");
292 static int metadelay = 28;		/* time to delay syncing metadata */
293 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
294     "Time to delay syncing metadata (in seconds)");
295 static int rushjob;		/* number of slots to run ASAP */
296 static int stat_rush_requests;	/* number of times I/O speeded up */
297 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
298     "Number of times I/O speeded up (rush requests)");
299 
300 #define	VDBATCH_SIZE 8
301 struct vdbatch {
302 	u_int index;
303 	long freevnodes;
304 	struct mtx lock;
305 	struct vnode *tab[VDBATCH_SIZE];
306 };
307 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
308 
309 static void	vdbatch_dequeue(struct vnode *vp);
310 
311 /*
312  * When shutting down the syncer, run it at four times normal speed.
313  */
314 #define SYNCER_SHUTDOWN_SPEEDUP		4
315 static int sync_vnode_count;
316 static int syncer_worklist_len;
317 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
318     syncer_state;
319 
320 /* Target for maximum number of vnodes. */
321 u_long desiredvnodes;
322 static u_long gapvnodes;		/* gap between wanted and desired */
323 static u_long vhiwat;		/* enough extras after expansion */
324 static u_long vlowat;		/* minimal extras before expansion */
325 static u_long vstir;		/* nonzero to stir non-free vnodes */
326 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
327 
328 static u_long vnlru_read_freevnodes(void);
329 
330 /*
331  * Note that no attempt is made to sanitize these parameters.
332  */
333 static int
334 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
335 {
336 	u_long val;
337 	int error;
338 
339 	val = desiredvnodes;
340 	error = sysctl_handle_long(oidp, &val, 0, req);
341 	if (error != 0 || req->newptr == NULL)
342 		return (error);
343 
344 	if (val == desiredvnodes)
345 		return (0);
346 	mtx_lock(&vnode_list_mtx);
347 	desiredvnodes = val;
348 	wantfreevnodes = desiredvnodes / 4;
349 	vnlru_recalc();
350 	mtx_unlock(&vnode_list_mtx);
351 	/*
352 	 * XXX There is no protection against multiple threads changing
353 	 * desiredvnodes at the same time. Locking above only helps vnlru and
354 	 * getnewvnode.
355 	 */
356 	vfs_hash_changesize(desiredvnodes);
357 	cache_changesize(desiredvnodes);
358 	return (0);
359 }
360 
361 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
362     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
363     "UL", "Target for maximum number of vnodes");
364 
365 static int
366 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
367 {
368 	u_long val;
369 	int error;
370 
371 	val = wantfreevnodes;
372 	error = sysctl_handle_long(oidp, &val, 0, req);
373 	if (error != 0 || req->newptr == NULL)
374 		return (error);
375 
376 	if (val == wantfreevnodes)
377 		return (0);
378 	mtx_lock(&vnode_list_mtx);
379 	wantfreevnodes = val;
380 	vnlru_recalc();
381 	mtx_unlock(&vnode_list_mtx);
382 	return (0);
383 }
384 
385 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
386     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
387     "UL", "Target for minimum number of \"free\" vnodes");
388 
389 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
390     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
391 static int vnlru_nowhere;
392 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
393     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
394 
395 static int
396 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
397 {
398 	struct vnode *vp;
399 	struct nameidata nd;
400 	char *buf;
401 	unsigned long ndflags;
402 	int error;
403 
404 	if (req->newptr == NULL)
405 		return (EINVAL);
406 	if (req->newlen >= PATH_MAX)
407 		return (E2BIG);
408 
409 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
410 	error = SYSCTL_IN(req, buf, req->newlen);
411 	if (error != 0)
412 		goto out;
413 
414 	buf[req->newlen] = '\0';
415 
416 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | NOCACHE | SAVENAME;
417 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread);
418 	if ((error = namei(&nd)) != 0)
419 		goto out;
420 	vp = nd.ni_vp;
421 
422 	if (VN_IS_DOOMED(vp)) {
423 		/*
424 		 * This vnode is being recycled.  Return != 0 to let the caller
425 		 * know that the sysctl had no effect.  Return EAGAIN because a
426 		 * subsequent call will likely succeed (since namei will create
427 		 * a new vnode if necessary)
428 		 */
429 		error = EAGAIN;
430 		goto putvnode;
431 	}
432 
433 	counter_u64_add(recycles_count, 1);
434 	vgone(vp);
435 putvnode:
436 	NDFREE(&nd, 0);
437 out:
438 	free(buf, M_TEMP);
439 	return (error);
440 }
441 
442 static int
443 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
444 {
445 	struct thread *td = curthread;
446 	struct vnode *vp;
447 	struct file *fp;
448 	int error;
449 	int fd;
450 
451 	if (req->newptr == NULL)
452 		return (EBADF);
453 
454         error = sysctl_handle_int(oidp, &fd, 0, req);
455         if (error != 0)
456                 return (error);
457 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
458 	if (error != 0)
459 		return (error);
460 	vp = fp->f_vnode;
461 
462 	error = vn_lock(vp, LK_EXCLUSIVE);
463 	if (error != 0)
464 		goto drop;
465 
466 	counter_u64_add(recycles_count, 1);
467 	vgone(vp);
468 	VOP_UNLOCK(vp);
469 drop:
470 	fdrop(fp, td);
471 	return (error);
472 }
473 
474 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
475     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
476     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
477 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
478     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
479     sysctl_ftry_reclaim_vnode, "I",
480     "Try to reclaim a vnode by its file descriptor");
481 
482 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
483 static int vnsz2log;
484 
485 /*
486  * Support for the bufobj clean & dirty pctrie.
487  */
488 static void *
489 buf_trie_alloc(struct pctrie *ptree)
490 {
491 
492 	return uma_zalloc(buf_trie_zone, M_NOWAIT);
493 }
494 
495 static void
496 buf_trie_free(struct pctrie *ptree, void *node)
497 {
498 
499 	uma_zfree(buf_trie_zone, node);
500 }
501 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
502 
503 /*
504  * Initialize the vnode management data structures.
505  *
506  * Reevaluate the following cap on the number of vnodes after the physical
507  * memory size exceeds 512GB.  In the limit, as the physical memory size
508  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
509  */
510 #ifndef	MAXVNODES_MAX
511 #define	MAXVNODES_MAX	(512UL * 1024 * 1024 / 64)	/* 8M */
512 #endif
513 
514 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
515 
516 static struct vnode *
517 vn_alloc_marker(struct mount *mp)
518 {
519 	struct vnode *vp;
520 
521 	vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
522 	vp->v_type = VMARKER;
523 	vp->v_mount = mp;
524 
525 	return (vp);
526 }
527 
528 static void
529 vn_free_marker(struct vnode *vp)
530 {
531 
532 	MPASS(vp->v_type == VMARKER);
533 	free(vp, M_VNODE_MARKER);
534 }
535 
536 /*
537  * Initialize a vnode as it first enters the zone.
538  */
539 static int
540 vnode_init(void *mem, int size, int flags)
541 {
542 	struct vnode *vp;
543 
544 	vp = mem;
545 	bzero(vp, size);
546 	/*
547 	 * Setup locks.
548 	 */
549 	vp->v_vnlock = &vp->v_lock;
550 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
551 	/*
552 	 * By default, don't allow shared locks unless filesystems opt-in.
553 	 */
554 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
555 	    LK_NOSHARE | LK_IS_VNODE);
556 	/*
557 	 * Initialize bufobj.
558 	 */
559 	bufobj_init(&vp->v_bufobj, vp);
560 	/*
561 	 * Initialize namecache.
562 	 */
563 	LIST_INIT(&vp->v_cache_src);
564 	TAILQ_INIT(&vp->v_cache_dst);
565 	/*
566 	 * Initialize rangelocks.
567 	 */
568 	rangelock_init(&vp->v_rl);
569 
570 	vp->v_dbatchcpu = NOCPU;
571 
572 	mtx_lock(&vnode_list_mtx);
573 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
574 	mtx_unlock(&vnode_list_mtx);
575 	return (0);
576 }
577 
578 /*
579  * Free a vnode when it is cleared from the zone.
580  */
581 static void
582 vnode_fini(void *mem, int size)
583 {
584 	struct vnode *vp;
585 	struct bufobj *bo;
586 
587 	vp = mem;
588 	vdbatch_dequeue(vp);
589 	mtx_lock(&vnode_list_mtx);
590 	TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
591 	mtx_unlock(&vnode_list_mtx);
592 	rangelock_destroy(&vp->v_rl);
593 	lockdestroy(vp->v_vnlock);
594 	mtx_destroy(&vp->v_interlock);
595 	bo = &vp->v_bufobj;
596 	rw_destroy(BO_LOCKPTR(bo));
597 }
598 
599 /*
600  * Provide the size of NFS nclnode and NFS fh for calculation of the
601  * vnode memory consumption.  The size is specified directly to
602  * eliminate dependency on NFS-private header.
603  *
604  * Other filesystems may use bigger or smaller (like UFS and ZFS)
605  * private inode data, but the NFS-based estimation is ample enough.
606  * Still, we care about differences in the size between 64- and 32-bit
607  * platforms.
608  *
609  * Namecache structure size is heuristically
610  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
611  */
612 #ifdef _LP64
613 #define	NFS_NCLNODE_SZ	(528 + 64)
614 #define	NC_SZ		148
615 #else
616 #define	NFS_NCLNODE_SZ	(360 + 32)
617 #define	NC_SZ		92
618 #endif
619 
620 static void
621 vntblinit(void *dummy __unused)
622 {
623 	struct vdbatch *vd;
624 	int cpu, physvnodes, virtvnodes;
625 	u_int i;
626 
627 	/*
628 	 * Desiredvnodes is a function of the physical memory size and the
629 	 * kernel's heap size.  Generally speaking, it scales with the
630 	 * physical memory size.  The ratio of desiredvnodes to the physical
631 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
632 	 * Thereafter, the
633 	 * marginal ratio of desiredvnodes to the physical memory size is
634 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
635 	 * size.  The memory required by desiredvnodes vnodes and vm objects
636 	 * must not exceed 1/10th of the kernel's heap size.
637 	 */
638 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
639 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
640 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
641 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
642 	desiredvnodes = min(physvnodes, virtvnodes);
643 	if (desiredvnodes > MAXVNODES_MAX) {
644 		if (bootverbose)
645 			printf("Reducing kern.maxvnodes %lu -> %lu\n",
646 			    desiredvnodes, MAXVNODES_MAX);
647 		desiredvnodes = MAXVNODES_MAX;
648 	}
649 	wantfreevnodes = desiredvnodes / 4;
650 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
651 	TAILQ_INIT(&vnode_list);
652 	mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
653 	/*
654 	 * The lock is taken to appease WITNESS.
655 	 */
656 	mtx_lock(&vnode_list_mtx);
657 	vnlru_recalc();
658 	mtx_unlock(&vnode_list_mtx);
659 	vnode_list_free_marker = vn_alloc_marker(NULL);
660 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
661 	vnode_list_reclaim_marker = vn_alloc_marker(NULL);
662 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
663 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
664 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
665 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
666 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
667 	/*
668 	 * Preallocate enough nodes to support one-per buf so that
669 	 * we can not fail an insert.  reassignbuf() callers can not
670 	 * tolerate the insertion failure.
671 	 */
672 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
673 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
674 	    UMA_ZONE_NOFREE | UMA_ZONE_VM);
675 	uma_prealloc(buf_trie_zone, nbuf);
676 
677 	vnodes_created = counter_u64_alloc(M_WAITOK);
678 	recycles_count = counter_u64_alloc(M_WAITOK);
679 	recycles_free_count = counter_u64_alloc(M_WAITOK);
680 	deferred_inact = counter_u64_alloc(M_WAITOK);
681 
682 	/*
683 	 * Initialize the filesystem syncer.
684 	 */
685 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
686 	    &syncer_mask);
687 	syncer_maxdelay = syncer_mask + 1;
688 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
689 	cv_init(&sync_wakeup, "syncer");
690 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
691 		vnsz2log++;
692 	vnsz2log--;
693 
694 	CPU_FOREACH(cpu) {
695 		vd = DPCPU_ID_PTR((cpu), vd);
696 		bzero(vd, sizeof(*vd));
697 		mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
698 	}
699 }
700 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
701 
702 
703 /*
704  * Mark a mount point as busy. Used to synchronize access and to delay
705  * unmounting. Eventually, mountlist_mtx is not released on failure.
706  *
707  * vfs_busy() is a custom lock, it can block the caller.
708  * vfs_busy() only sleeps if the unmount is active on the mount point.
709  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
710  * vnode belonging to mp.
711  *
712  * Lookup uses vfs_busy() to traverse mount points.
713  * root fs			var fs
714  * / vnode lock		A	/ vnode lock (/var)		D
715  * /var vnode lock	B	/log vnode lock(/var/log)	E
716  * vfs_busy lock	C	vfs_busy lock			F
717  *
718  * Within each file system, the lock order is C->A->B and F->D->E.
719  *
720  * When traversing across mounts, the system follows that lock order:
721  *
722  *        C->A->B
723  *              |
724  *              +->F->D->E
725  *
726  * The lookup() process for namei("/var") illustrates the process:
727  *  VOP_LOOKUP() obtains B while A is held
728  *  vfs_busy() obtains a shared lock on F while A and B are held
729  *  vput() releases lock on B
730  *  vput() releases lock on A
731  *  VFS_ROOT() obtains lock on D while shared lock on F is held
732  *  vfs_unbusy() releases shared lock on F
733  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
734  *    Attempt to lock A (instead of vp_crossmp) while D is held would
735  *    violate the global order, causing deadlocks.
736  *
737  * dounmount() locks B while F is drained.
738  */
739 int
740 vfs_busy(struct mount *mp, int flags)
741 {
742 
743 	MPASS((flags & ~MBF_MASK) == 0);
744 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
745 
746 	if (vfs_op_thread_enter(mp)) {
747 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
748 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
749 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
750 		vfs_mp_count_add_pcpu(mp, ref, 1);
751 		vfs_mp_count_add_pcpu(mp, lockref, 1);
752 		vfs_op_thread_exit(mp);
753 		if (flags & MBF_MNTLSTLOCK)
754 			mtx_unlock(&mountlist_mtx);
755 		return (0);
756 	}
757 
758 	MNT_ILOCK(mp);
759 	vfs_assert_mount_counters(mp);
760 	MNT_REF(mp);
761 	/*
762 	 * If mount point is currently being unmounted, sleep until the
763 	 * mount point fate is decided.  If thread doing the unmounting fails,
764 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
765 	 * that this mount point has survived the unmount attempt and vfs_busy
766 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
767 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
768 	 * about to be really destroyed.  vfs_busy needs to release its
769 	 * reference on the mount point in this case and return with ENOENT,
770 	 * telling the caller that mount mount it tried to busy is no longer
771 	 * valid.
772 	 */
773 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
774 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
775 			MNT_REL(mp);
776 			MNT_IUNLOCK(mp);
777 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
778 			    __func__);
779 			return (ENOENT);
780 		}
781 		if (flags & MBF_MNTLSTLOCK)
782 			mtx_unlock(&mountlist_mtx);
783 		mp->mnt_kern_flag |= MNTK_MWAIT;
784 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
785 		if (flags & MBF_MNTLSTLOCK)
786 			mtx_lock(&mountlist_mtx);
787 		MNT_ILOCK(mp);
788 	}
789 	if (flags & MBF_MNTLSTLOCK)
790 		mtx_unlock(&mountlist_mtx);
791 	mp->mnt_lockref++;
792 	MNT_IUNLOCK(mp);
793 	return (0);
794 }
795 
796 /*
797  * Free a busy filesystem.
798  */
799 void
800 vfs_unbusy(struct mount *mp)
801 {
802 	int c;
803 
804 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
805 
806 	if (vfs_op_thread_enter(mp)) {
807 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
808 		vfs_mp_count_sub_pcpu(mp, lockref, 1);
809 		vfs_mp_count_sub_pcpu(mp, ref, 1);
810 		vfs_op_thread_exit(mp);
811 		return;
812 	}
813 
814 	MNT_ILOCK(mp);
815 	vfs_assert_mount_counters(mp);
816 	MNT_REL(mp);
817 	c = --mp->mnt_lockref;
818 	if (mp->mnt_vfs_ops == 0) {
819 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
820 		MNT_IUNLOCK(mp);
821 		return;
822 	}
823 	if (c < 0)
824 		vfs_dump_mount_counters(mp);
825 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
826 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
827 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
828 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
829 		wakeup(&mp->mnt_lockref);
830 	}
831 	MNT_IUNLOCK(mp);
832 }
833 
834 /*
835  * Lookup a mount point by filesystem identifier.
836  */
837 struct mount *
838 vfs_getvfs(fsid_t *fsid)
839 {
840 	struct mount *mp;
841 
842 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
843 	mtx_lock(&mountlist_mtx);
844 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
845 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
846 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
847 			vfs_ref(mp);
848 			mtx_unlock(&mountlist_mtx);
849 			return (mp);
850 		}
851 	}
852 	mtx_unlock(&mountlist_mtx);
853 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
854 	return ((struct mount *) 0);
855 }
856 
857 /*
858  * Lookup a mount point by filesystem identifier, busying it before
859  * returning.
860  *
861  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
862  * cache for popular filesystem identifiers.  The cache is lockess, using
863  * the fact that struct mount's are never freed.  In worst case we may
864  * get pointer to unmounted or even different filesystem, so we have to
865  * check what we got, and go slow way if so.
866  */
867 struct mount *
868 vfs_busyfs(fsid_t *fsid)
869 {
870 #define	FSID_CACHE_SIZE	256
871 	typedef struct mount * volatile vmp_t;
872 	static vmp_t cache[FSID_CACHE_SIZE];
873 	struct mount *mp;
874 	int error;
875 	uint32_t hash;
876 
877 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
878 	hash = fsid->val[0] ^ fsid->val[1];
879 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
880 	mp = cache[hash];
881 	if (mp == NULL ||
882 	    mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
883 	    mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
884 		goto slow;
885 	if (vfs_busy(mp, 0) != 0) {
886 		cache[hash] = NULL;
887 		goto slow;
888 	}
889 	if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
890 	    mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
891 		return (mp);
892 	else
893 	    vfs_unbusy(mp);
894 
895 slow:
896 	mtx_lock(&mountlist_mtx);
897 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
898 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
899 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
900 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
901 			if (error) {
902 				cache[hash] = NULL;
903 				mtx_unlock(&mountlist_mtx);
904 				return (NULL);
905 			}
906 			cache[hash] = mp;
907 			return (mp);
908 		}
909 	}
910 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
911 	mtx_unlock(&mountlist_mtx);
912 	return ((struct mount *) 0);
913 }
914 
915 /*
916  * Check if a user can access privileged mount options.
917  */
918 int
919 vfs_suser(struct mount *mp, struct thread *td)
920 {
921 	int error;
922 
923 	if (jailed(td->td_ucred)) {
924 		/*
925 		 * If the jail of the calling thread lacks permission for
926 		 * this type of file system, deny immediately.
927 		 */
928 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
929 			return (EPERM);
930 
931 		/*
932 		 * If the file system was mounted outside the jail of the
933 		 * calling thread, deny immediately.
934 		 */
935 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
936 			return (EPERM);
937 	}
938 
939 	/*
940 	 * If file system supports delegated administration, we don't check
941 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
942 	 * by the file system itself.
943 	 * If this is not the user that did original mount, we check for
944 	 * the PRIV_VFS_MOUNT_OWNER privilege.
945 	 */
946 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
947 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
948 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
949 			return (error);
950 	}
951 	return (0);
952 }
953 
954 /*
955  * Get a new unique fsid.  Try to make its val[0] unique, since this value
956  * will be used to create fake device numbers for stat().  Also try (but
957  * not so hard) make its val[0] unique mod 2^16, since some emulators only
958  * support 16-bit device numbers.  We end up with unique val[0]'s for the
959  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
960  *
961  * Keep in mind that several mounts may be running in parallel.  Starting
962  * the search one past where the previous search terminated is both a
963  * micro-optimization and a defense against returning the same fsid to
964  * different mounts.
965  */
966 void
967 vfs_getnewfsid(struct mount *mp)
968 {
969 	static uint16_t mntid_base;
970 	struct mount *nmp;
971 	fsid_t tfsid;
972 	int mtype;
973 
974 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
975 	mtx_lock(&mntid_mtx);
976 	mtype = mp->mnt_vfc->vfc_typenum;
977 	tfsid.val[1] = mtype;
978 	mtype = (mtype & 0xFF) << 24;
979 	for (;;) {
980 		tfsid.val[0] = makedev(255,
981 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
982 		mntid_base++;
983 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
984 			break;
985 		vfs_rel(nmp);
986 	}
987 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
988 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
989 	mtx_unlock(&mntid_mtx);
990 }
991 
992 /*
993  * Knob to control the precision of file timestamps:
994  *
995  *   0 = seconds only; nanoseconds zeroed.
996  *   1 = seconds and nanoseconds, accurate within 1/HZ.
997  *   2 = seconds and nanoseconds, truncated to microseconds.
998  * >=3 = seconds and nanoseconds, maximum precision.
999  */
1000 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
1001 
1002 static int timestamp_precision = TSP_USEC;
1003 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
1004     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
1005     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1006     "3+: sec + ns (max. precision))");
1007 
1008 /*
1009  * Get a current timestamp.
1010  */
1011 void
1012 vfs_timestamp(struct timespec *tsp)
1013 {
1014 	struct timeval tv;
1015 
1016 	switch (timestamp_precision) {
1017 	case TSP_SEC:
1018 		tsp->tv_sec = time_second;
1019 		tsp->tv_nsec = 0;
1020 		break;
1021 	case TSP_HZ:
1022 		getnanotime(tsp);
1023 		break;
1024 	case TSP_USEC:
1025 		microtime(&tv);
1026 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1027 		break;
1028 	case TSP_NSEC:
1029 	default:
1030 		nanotime(tsp);
1031 		break;
1032 	}
1033 }
1034 
1035 /*
1036  * Set vnode attributes to VNOVAL
1037  */
1038 void
1039 vattr_null(struct vattr *vap)
1040 {
1041 
1042 	vap->va_type = VNON;
1043 	vap->va_size = VNOVAL;
1044 	vap->va_bytes = VNOVAL;
1045 	vap->va_mode = VNOVAL;
1046 	vap->va_nlink = VNOVAL;
1047 	vap->va_uid = VNOVAL;
1048 	vap->va_gid = VNOVAL;
1049 	vap->va_fsid = VNOVAL;
1050 	vap->va_fileid = VNOVAL;
1051 	vap->va_blocksize = VNOVAL;
1052 	vap->va_rdev = VNOVAL;
1053 	vap->va_atime.tv_sec = VNOVAL;
1054 	vap->va_atime.tv_nsec = VNOVAL;
1055 	vap->va_mtime.tv_sec = VNOVAL;
1056 	vap->va_mtime.tv_nsec = VNOVAL;
1057 	vap->va_ctime.tv_sec = VNOVAL;
1058 	vap->va_ctime.tv_nsec = VNOVAL;
1059 	vap->va_birthtime.tv_sec = VNOVAL;
1060 	vap->va_birthtime.tv_nsec = VNOVAL;
1061 	vap->va_flags = VNOVAL;
1062 	vap->va_gen = VNOVAL;
1063 	vap->va_vaflags = 0;
1064 }
1065 
1066 /*
1067  * Try to reduce the total number of vnodes.
1068  *
1069  * This routine (and its user) are buggy in at least the following ways:
1070  * - all parameters were picked years ago when RAM sizes were significantly
1071  *   smaller
1072  * - it can pick vnodes based on pages used by the vm object, but filesystems
1073  *   like ZFS don't use it making the pick broken
1074  * - since ZFS has its own aging policy it gets partially combated by this one
1075  * - a dedicated method should be provided for filesystems to let them decide
1076  *   whether the vnode should be recycled
1077  *
1078  * This routine is called when we have too many vnodes.  It attempts
1079  * to free <count> vnodes and will potentially free vnodes that still
1080  * have VM backing store (VM backing store is typically the cause
1081  * of a vnode blowout so we want to do this).  Therefore, this operation
1082  * is not considered cheap.
1083  *
1084  * A number of conditions may prevent a vnode from being reclaimed.
1085  * the buffer cache may have references on the vnode, a directory
1086  * vnode may still have references due to the namei cache representing
1087  * underlying files, or the vnode may be in active use.   It is not
1088  * desirable to reuse such vnodes.  These conditions may cause the
1089  * number of vnodes to reach some minimum value regardless of what
1090  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
1091  *
1092  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1093  * 			 entries if this argument is strue
1094  * @param trigger	 Only reclaim vnodes with fewer than this many resident
1095  *			 pages.
1096  * @param target	 How many vnodes to reclaim.
1097  * @return		 The number of vnodes that were reclaimed.
1098  */
1099 static int
1100 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1101 {
1102 	struct vnode *vp, *mvp;
1103 	struct mount *mp;
1104 	u_long done;
1105 	bool retried;
1106 
1107 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1108 
1109 	retried = false;
1110 	done = 0;
1111 
1112 	mvp = vnode_list_reclaim_marker;
1113 restart:
1114 	vp = mvp;
1115 	while (done < target) {
1116 		vp = TAILQ_NEXT(vp, v_vnodelist);
1117 		if (__predict_false(vp == NULL))
1118 			break;
1119 
1120 		if (__predict_false(vp->v_type == VMARKER))
1121 			continue;
1122 
1123 		/*
1124 		 * If it's been deconstructed already, it's still
1125 		 * referenced, or it exceeds the trigger, skip it.
1126 		 * Also skip free vnodes.  We are trying to make space
1127 		 * to expand the free list, not reduce it.
1128 		 */
1129 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1130 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1131 			goto next_iter;
1132 
1133 		if (vp->v_type == VBAD || vp->v_type == VNON)
1134 			goto next_iter;
1135 
1136 		if (!VI_TRYLOCK(vp))
1137 			goto next_iter;
1138 
1139 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1140 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1141 		    vp->v_type == VBAD || vp->v_type == VNON ||
1142 		    (vp->v_object != NULL &&
1143 		    vp->v_object->resident_page_count > trigger)) {
1144 			VI_UNLOCK(vp);
1145 			goto next_iter;
1146 		}
1147 		vholdl(vp);
1148 		VI_UNLOCK(vp);
1149 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1150 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1151 		mtx_unlock(&vnode_list_mtx);
1152 
1153 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1154 			vdrop(vp);
1155 			goto next_iter_unlocked;
1156 		}
1157 		if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1158 			vdrop(vp);
1159 			vn_finished_write(mp);
1160 			goto next_iter_unlocked;
1161 		}
1162 
1163 		VI_LOCK(vp);
1164 		if (vp->v_usecount > 0 ||
1165 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1166 		    (vp->v_object != NULL &&
1167 		    vp->v_object->resident_page_count > trigger)) {
1168 			VOP_UNLOCK(vp);
1169 			vdropl(vp);
1170 			vn_finished_write(mp);
1171 			goto next_iter_unlocked;
1172 		}
1173 		counter_u64_add(recycles_count, 1);
1174 		vgonel(vp);
1175 		VOP_UNLOCK(vp);
1176 		vdropl(vp);
1177 		vn_finished_write(mp);
1178 		done++;
1179 next_iter_unlocked:
1180 		if (should_yield())
1181 			kern_yield(PRI_USER);
1182 		mtx_lock(&vnode_list_mtx);
1183 		goto restart;
1184 next_iter:
1185 		MPASS(vp->v_type != VMARKER);
1186 		if (!should_yield())
1187 			continue;
1188 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1189 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1190 		mtx_unlock(&vnode_list_mtx);
1191 		kern_yield(PRI_USER);
1192 		mtx_lock(&vnode_list_mtx);
1193 		goto restart;
1194 	}
1195 	if (done == 0 && !retried) {
1196 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1197 		TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1198 		retried = true;
1199 		goto restart;
1200 	}
1201 	return (done);
1202 }
1203 
1204 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1205 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1206     0,
1207     "limit on vnode free requests per call to the vnlru_free routine");
1208 
1209 /*
1210  * Attempt to reduce the free list by the requested amount.
1211  */
1212 static int
1213 vnlru_free_locked(int count, struct vfsops *mnt_op)
1214 {
1215 	struct vnode *vp, *mvp;
1216 	struct mount *mp;
1217 	int ocount;
1218 
1219 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1220 	if (count > max_vnlru_free)
1221 		count = max_vnlru_free;
1222 	ocount = count;
1223 	mvp = vnode_list_free_marker;
1224 restart:
1225 	vp = mvp;
1226 	while (count > 0) {
1227 		vp = TAILQ_NEXT(vp, v_vnodelist);
1228 		if (__predict_false(vp == NULL)) {
1229 			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1230 			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1231 			break;
1232 		}
1233 		if (__predict_false(vp->v_type == VMARKER))
1234 			continue;
1235 
1236 		/*
1237 		 * Don't recycle if our vnode is from different type
1238 		 * of mount point.  Note that mp is type-safe, the
1239 		 * check does not reach unmapped address even if
1240 		 * vnode is reclaimed.
1241 		 * Don't recycle if we can't get the interlock without
1242 		 * blocking.
1243 		 */
1244 		if (vp->v_holdcnt > 0 || (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1245 		    mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) {
1246 			continue;
1247 		}
1248 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1249 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1250 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1251 			VI_UNLOCK(vp);
1252 			continue;
1253 		}
1254 		vholdl(vp);
1255 		count--;
1256 		mtx_unlock(&vnode_list_mtx);
1257 		VI_UNLOCK(vp);
1258 		vtryrecycle(vp);
1259 		vdrop(vp);
1260 		mtx_lock(&vnode_list_mtx);
1261 		goto restart;
1262 	}
1263 	return (ocount - count);
1264 }
1265 
1266 void
1267 vnlru_free(int count, struct vfsops *mnt_op)
1268 {
1269 
1270 	mtx_lock(&vnode_list_mtx);
1271 	vnlru_free_locked(count, mnt_op);
1272 	mtx_unlock(&vnode_list_mtx);
1273 }
1274 
1275 static void
1276 vnlru_recalc(void)
1277 {
1278 
1279 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1280 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1281 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1282 	vlowat = vhiwat / 2;
1283 }
1284 
1285 /*
1286  * Attempt to recycle vnodes in a context that is always safe to block.
1287  * Calling vlrurecycle() from the bowels of filesystem code has some
1288  * interesting deadlock problems.
1289  */
1290 static struct proc *vnlruproc;
1291 static int vnlruproc_sig;
1292 
1293 /*
1294  * The main freevnodes counter is only updated when threads requeue their vnode
1295  * batches. CPUs are conditionally walked to compute a more accurate total.
1296  *
1297  * Limit how much of a slop are we willing to tolerate. Note: the actual value
1298  * at any given moment can still exceed slop, but it should not be by significant
1299  * margin in practice.
1300  */
1301 #define VNLRU_FREEVNODES_SLOP 128
1302 
1303 static u_long
1304 vnlru_read_freevnodes(void)
1305 {
1306 	struct vdbatch *vd;
1307 	long slop;
1308 	int cpu;
1309 
1310 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1311 	if (freevnodes > freevnodes_old)
1312 		slop = freevnodes - freevnodes_old;
1313 	else
1314 		slop = freevnodes_old - freevnodes;
1315 	if (slop < VNLRU_FREEVNODES_SLOP)
1316 		return (freevnodes >= 0 ? freevnodes : 0);
1317 	freevnodes_old = freevnodes;
1318 	CPU_FOREACH(cpu) {
1319 		vd = DPCPU_ID_PTR((cpu), vd);
1320 		freevnodes_old += vd->freevnodes;
1321 	}
1322 	return (freevnodes_old >= 0 ? freevnodes_old : 0);
1323 }
1324 
1325 static bool
1326 vnlru_under(u_long rnumvnodes, u_long limit)
1327 {
1328 	u_long rfreevnodes, space;
1329 
1330 	if (__predict_false(rnumvnodes > desiredvnodes))
1331 		return (true);
1332 
1333 	space = desiredvnodes - rnumvnodes;
1334 	if (space < limit) {
1335 		rfreevnodes = vnlru_read_freevnodes();
1336 		if (rfreevnodes > wantfreevnodes)
1337 			space += rfreevnodes - wantfreevnodes;
1338 	}
1339 	return (space < limit);
1340 }
1341 
1342 static bool
1343 vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1344 {
1345 	long rfreevnodes, space;
1346 
1347 	if (__predict_false(rnumvnodes > desiredvnodes))
1348 		return (true);
1349 
1350 	space = desiredvnodes - rnumvnodes;
1351 	if (space < limit) {
1352 		rfreevnodes = atomic_load_long(&freevnodes);
1353 		if (rfreevnodes > wantfreevnodes)
1354 			space += rfreevnodes - wantfreevnodes;
1355 	}
1356 	return (space < limit);
1357 }
1358 
1359 static void
1360 vnlru_kick(void)
1361 {
1362 
1363 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1364 	if (vnlruproc_sig == 0) {
1365 		vnlruproc_sig = 1;
1366 		wakeup(vnlruproc);
1367 	}
1368 }
1369 
1370 static void
1371 vnlru_proc(void)
1372 {
1373 	u_long rnumvnodes, rfreevnodes, target;
1374 	unsigned long onumvnodes;
1375 	int done, force, trigger, usevnodes;
1376 	bool reclaim_nc_src, want_reread;
1377 
1378 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1379 	    SHUTDOWN_PRI_FIRST);
1380 
1381 	force = 0;
1382 	want_reread = false;
1383 	for (;;) {
1384 		kproc_suspend_check(vnlruproc);
1385 		mtx_lock(&vnode_list_mtx);
1386 		rnumvnodes = atomic_load_long(&numvnodes);
1387 
1388 		if (want_reread) {
1389 			force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1390 			want_reread = false;
1391 		}
1392 
1393 		/*
1394 		 * If numvnodes is too large (due to desiredvnodes being
1395 		 * adjusted using its sysctl, or emergency growth), first
1396 		 * try to reduce it by discarding from the free list.
1397 		 */
1398 		if (rnumvnodes > desiredvnodes) {
1399 			vnlru_free_locked(rnumvnodes - desiredvnodes, NULL);
1400 			rnumvnodes = atomic_load_long(&numvnodes);
1401 		}
1402 		/*
1403 		 * Sleep if the vnode cache is in a good state.  This is
1404 		 * when it is not over-full and has space for about a 4%
1405 		 * or 9% expansion (by growing its size or inexcessively
1406 		 * reducing its free list).  Otherwise, try to reclaim
1407 		 * space for a 10% expansion.
1408 		 */
1409 		if (vstir && force == 0) {
1410 			force = 1;
1411 			vstir = 0;
1412 		}
1413 		if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1414 			vnlruproc_sig = 0;
1415 			wakeup(&vnlruproc_sig);
1416 			msleep(vnlruproc, &vnode_list_mtx,
1417 			    PVFS|PDROP, "vlruwt", hz);
1418 			continue;
1419 		}
1420 		rfreevnodes = vnlru_read_freevnodes();
1421 
1422 		onumvnodes = rnumvnodes;
1423 		/*
1424 		 * Calculate parameters for recycling.  These are the same
1425 		 * throughout the loop to give some semblance of fairness.
1426 		 * The trigger point is to avoid recycling vnodes with lots
1427 		 * of resident pages.  We aren't trying to free memory; we
1428 		 * are trying to recycle or at least free vnodes.
1429 		 */
1430 		if (rnumvnodes <= desiredvnodes)
1431 			usevnodes = rnumvnodes - rfreevnodes;
1432 		else
1433 			usevnodes = rnumvnodes;
1434 		if (usevnodes <= 0)
1435 			usevnodes = 1;
1436 		/*
1437 		 * The trigger value is is chosen to give a conservatively
1438 		 * large value to ensure that it alone doesn't prevent
1439 		 * making progress.  The value can easily be so large that
1440 		 * it is effectively infinite in some congested and
1441 		 * misconfigured cases, and this is necessary.  Normally
1442 		 * it is about 8 to 100 (pages), which is quite large.
1443 		 */
1444 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1445 		if (force < 2)
1446 			trigger = vsmalltrigger;
1447 		reclaim_nc_src = force >= 3;
1448 		target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1449 		target = target / 10 + 1;
1450 		done = vlrureclaim(reclaim_nc_src, trigger, target);
1451 		mtx_unlock(&vnode_list_mtx);
1452 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1453 			uma_reclaim(UMA_RECLAIM_DRAIN);
1454 		if (done == 0) {
1455 			if (force == 0 || force == 1) {
1456 				force = 2;
1457 				continue;
1458 			}
1459 			if (force == 2) {
1460 				force = 3;
1461 				continue;
1462 			}
1463 			want_reread = true;
1464 			force = 0;
1465 			vnlru_nowhere++;
1466 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1467 		} else {
1468 			want_reread = true;
1469 			kern_yield(PRI_USER);
1470 		}
1471 	}
1472 }
1473 
1474 static struct kproc_desc vnlru_kp = {
1475 	"vnlru",
1476 	vnlru_proc,
1477 	&vnlruproc
1478 };
1479 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1480     &vnlru_kp);
1481 
1482 /*
1483  * Routines having to do with the management of the vnode table.
1484  */
1485 
1486 /*
1487  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1488  * before we actually vgone().  This function must be called with the vnode
1489  * held to prevent the vnode from being returned to the free list midway
1490  * through vgone().
1491  */
1492 static int
1493 vtryrecycle(struct vnode *vp)
1494 {
1495 	struct mount *vnmp;
1496 
1497 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1498 	VNASSERT(vp->v_holdcnt, vp,
1499 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1500 	/*
1501 	 * This vnode may found and locked via some other list, if so we
1502 	 * can't recycle it yet.
1503 	 */
1504 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1505 		CTR2(KTR_VFS,
1506 		    "%s: impossible to recycle, vp %p lock is already held",
1507 		    __func__, vp);
1508 		return (EWOULDBLOCK);
1509 	}
1510 	/*
1511 	 * Don't recycle if its filesystem is being suspended.
1512 	 */
1513 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1514 		VOP_UNLOCK(vp);
1515 		CTR2(KTR_VFS,
1516 		    "%s: impossible to recycle, cannot start the write for %p",
1517 		    __func__, vp);
1518 		return (EBUSY);
1519 	}
1520 	/*
1521 	 * If we got this far, we need to acquire the interlock and see if
1522 	 * anyone picked up this vnode from another list.  If not, we will
1523 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1524 	 * will skip over it.
1525 	 */
1526 	VI_LOCK(vp);
1527 	if (vp->v_usecount) {
1528 		VOP_UNLOCK(vp);
1529 		VI_UNLOCK(vp);
1530 		vn_finished_write(vnmp);
1531 		CTR2(KTR_VFS,
1532 		    "%s: impossible to recycle, %p is already referenced",
1533 		    __func__, vp);
1534 		return (EBUSY);
1535 	}
1536 	if (!VN_IS_DOOMED(vp)) {
1537 		counter_u64_add(recycles_free_count, 1);
1538 		vgonel(vp);
1539 	}
1540 	VOP_UNLOCK(vp);
1541 	VI_UNLOCK(vp);
1542 	vn_finished_write(vnmp);
1543 	return (0);
1544 }
1545 
1546 /*
1547  * Allocate a new vnode.
1548  *
1549  * The operation never returns an error. Returning an error was disabled
1550  * in r145385 (dated 2005) with the following comment:
1551  *
1552  * XXX Not all VFS_VGET/ffs_vget callers check returns.
1553  *
1554  * Given the age of this commit (almost 15 years at the time of writing this
1555  * comment) restoring the ability to fail requires a significant audit of
1556  * all codepaths.
1557  *
1558  * The routine can try to free a vnode or stall for up to 1 second waiting for
1559  * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1560  */
1561 static u_long vn_alloc_cyclecount;
1562 
1563 static struct vnode * __noinline
1564 vn_alloc_hard(struct mount *mp)
1565 {
1566 	u_long rnumvnodes, rfreevnodes;
1567 
1568 	mtx_lock(&vnode_list_mtx);
1569 	rnumvnodes = atomic_load_long(&numvnodes);
1570 	if (rnumvnodes + 1 < desiredvnodes) {
1571 		vn_alloc_cyclecount = 0;
1572 		goto alloc;
1573 	}
1574 	rfreevnodes = vnlru_read_freevnodes();
1575 	if (vn_alloc_cyclecount++ >= rfreevnodes) {
1576 		vn_alloc_cyclecount = 0;
1577 		vstir = 1;
1578 	}
1579 	/*
1580 	 * Grow the vnode cache if it will not be above its target max
1581 	 * after growing.  Otherwise, if the free list is nonempty, try
1582 	 * to reclaim 1 item from it before growing the cache (possibly
1583 	 * above its target max if the reclamation failed or is delayed).
1584 	 * Otherwise, wait for some space.  In all cases, schedule
1585 	 * vnlru_proc() if we are getting short of space.  The watermarks
1586 	 * should be chosen so that we never wait or even reclaim from
1587 	 * the free list to below its target minimum.
1588 	 */
1589 	if (vnlru_free_locked(1, NULL) > 0)
1590 		goto alloc;
1591 	if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1592 		/*
1593 		 * Wait for space for a new vnode.
1594 		 */
1595 		vnlru_kick();
1596 		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1597 		if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1598 		    vnlru_read_freevnodes() > 1)
1599 			vnlru_free_locked(1, NULL);
1600 	}
1601 alloc:
1602 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1603 	if (vnlru_under(rnumvnodes, vlowat))
1604 		vnlru_kick();
1605 	mtx_unlock(&vnode_list_mtx);
1606 	return (uma_zalloc(vnode_zone, M_WAITOK));
1607 }
1608 
1609 static struct vnode *
1610 vn_alloc(struct mount *mp)
1611 {
1612 	u_long rnumvnodes;
1613 
1614 	if (__predict_false(vn_alloc_cyclecount != 0))
1615 		return (vn_alloc_hard(mp));
1616 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1617 	if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1618 		atomic_subtract_long(&numvnodes, 1);
1619 		return (vn_alloc_hard(mp));
1620 	}
1621 
1622 	return (uma_zalloc(vnode_zone, M_WAITOK));
1623 }
1624 
1625 static void
1626 vn_free(struct vnode *vp)
1627 {
1628 
1629 	atomic_subtract_long(&numvnodes, 1);
1630 	uma_zfree(vnode_zone, vp);
1631 }
1632 
1633 /*
1634  * Return the next vnode from the free list.
1635  */
1636 int
1637 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1638     struct vnode **vpp)
1639 {
1640 	struct vnode *vp;
1641 	struct thread *td;
1642 	struct lock_object *lo;
1643 
1644 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1645 
1646 	KASSERT(vops->registered,
1647 	    ("%s: not registered vector op %p\n", __func__, vops));
1648 
1649 	td = curthread;
1650 	if (td->td_vp_reserved != NULL) {
1651 		vp = td->td_vp_reserved;
1652 		td->td_vp_reserved = NULL;
1653 	} else {
1654 		vp = vn_alloc(mp);
1655 	}
1656 	counter_u64_add(vnodes_created, 1);
1657 	/*
1658 	 * Locks are given the generic name "vnode" when created.
1659 	 * Follow the historic practice of using the filesystem
1660 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1661 	 *
1662 	 * Locks live in a witness group keyed on their name. Thus,
1663 	 * when a lock is renamed, it must also move from the witness
1664 	 * group of its old name to the witness group of its new name.
1665 	 *
1666 	 * The change only needs to be made when the vnode moves
1667 	 * from one filesystem type to another. We ensure that each
1668 	 * filesystem use a single static name pointer for its tag so
1669 	 * that we can compare pointers rather than doing a strcmp().
1670 	 */
1671 	lo = &vp->v_vnlock->lock_object;
1672 	if (lo->lo_name != tag) {
1673 		lo->lo_name = tag;
1674 		WITNESS_DESTROY(lo);
1675 		WITNESS_INIT(lo, tag);
1676 	}
1677 	/*
1678 	 * By default, don't allow shared locks unless filesystems opt-in.
1679 	 */
1680 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1681 	/*
1682 	 * Finalize various vnode identity bits.
1683 	 */
1684 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1685 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1686 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1687 	vp->v_type = VNON;
1688 	vp->v_op = vops;
1689 	v_init_counters(vp);
1690 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1691 #ifdef DIAGNOSTIC
1692 	if (mp == NULL && vops != &dead_vnodeops)
1693 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1694 #endif
1695 #ifdef MAC
1696 	mac_vnode_init(vp);
1697 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1698 		mac_vnode_associate_singlelabel(mp, vp);
1699 #endif
1700 	if (mp != NULL) {
1701 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1702 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1703 			vp->v_vflag |= VV_NOKNOTE;
1704 	}
1705 
1706 	/*
1707 	 * For the filesystems which do not use vfs_hash_insert(),
1708 	 * still initialize v_hash to have vfs_hash_index() useful.
1709 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1710 	 * its own hashing.
1711 	 */
1712 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1713 
1714 	*vpp = vp;
1715 	return (0);
1716 }
1717 
1718 void
1719 getnewvnode_reserve(void)
1720 {
1721 	struct thread *td;
1722 
1723 	td = curthread;
1724 	MPASS(td->td_vp_reserved == NULL);
1725 	td->td_vp_reserved = vn_alloc(NULL);
1726 }
1727 
1728 void
1729 getnewvnode_drop_reserve(void)
1730 {
1731 	struct thread *td;
1732 
1733 	td = curthread;
1734 	if (td->td_vp_reserved != NULL) {
1735 		vn_free(td->td_vp_reserved);
1736 		td->td_vp_reserved = NULL;
1737 	}
1738 }
1739 
1740 static void
1741 freevnode(struct vnode *vp)
1742 {
1743 	struct bufobj *bo;
1744 
1745 	/*
1746 	 * The vnode has been marked for destruction, so free it.
1747 	 *
1748 	 * The vnode will be returned to the zone where it will
1749 	 * normally remain until it is needed for another vnode. We
1750 	 * need to cleanup (or verify that the cleanup has already
1751 	 * been done) any residual data left from its current use
1752 	 * so as not to contaminate the freshly allocated vnode.
1753 	 */
1754 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1755 	bo = &vp->v_bufobj;
1756 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1757 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
1758 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1759 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1760 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1761 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1762 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1763 	    ("clean blk trie not empty"));
1764 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1765 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1766 	    ("dirty blk trie not empty"));
1767 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1768 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1769 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1770 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1771 	    ("Dangling rangelock waiters"));
1772 	VI_UNLOCK(vp);
1773 #ifdef MAC
1774 	mac_vnode_destroy(vp);
1775 #endif
1776 	if (vp->v_pollinfo != NULL) {
1777 		destroy_vpollinfo(vp->v_pollinfo);
1778 		vp->v_pollinfo = NULL;
1779 	}
1780 #ifdef INVARIANTS
1781 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
1782 	vp->v_op = NULL;
1783 #endif
1784 	vp->v_mountedhere = NULL;
1785 	vp->v_unpcb = NULL;
1786 	vp->v_rdev = NULL;
1787 	vp->v_fifoinfo = NULL;
1788 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
1789 	vp->v_irflag = 0;
1790 	vp->v_iflag = 0;
1791 	vp->v_vflag = 0;
1792 	bo->bo_flag = 0;
1793 	vn_free(vp);
1794 }
1795 
1796 /*
1797  * Delete from old mount point vnode list, if on one.
1798  */
1799 static void
1800 delmntque(struct vnode *vp)
1801 {
1802 	struct mount *mp;
1803 
1804 	mp = vp->v_mount;
1805 	if (mp == NULL)
1806 		return;
1807 	MNT_ILOCK(mp);
1808 	VI_LOCK(vp);
1809 	if (vp->v_mflag & VMP_LAZYLIST) {
1810 		mtx_lock(&mp->mnt_listmtx);
1811 		if (vp->v_mflag & VMP_LAZYLIST) {
1812 			vp->v_mflag &= ~VMP_LAZYLIST;
1813 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
1814 			mp->mnt_lazyvnodelistsize--;
1815 		}
1816 		mtx_unlock(&mp->mnt_listmtx);
1817 	}
1818 	vp->v_mount = NULL;
1819 	VI_UNLOCK(vp);
1820 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1821 		("bad mount point vnode list size"));
1822 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1823 	mp->mnt_nvnodelistsize--;
1824 	MNT_REL(mp);
1825 	MNT_IUNLOCK(mp);
1826 }
1827 
1828 static void
1829 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1830 {
1831 
1832 	vp->v_data = NULL;
1833 	vp->v_op = &dead_vnodeops;
1834 	vgone(vp);
1835 	vput(vp);
1836 }
1837 
1838 /*
1839  * Insert into list of vnodes for the new mount point, if available.
1840  */
1841 int
1842 insmntque1(struct vnode *vp, struct mount *mp,
1843 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1844 {
1845 
1846 	KASSERT(vp->v_mount == NULL,
1847 		("insmntque: vnode already on per mount vnode list"));
1848 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1849 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1850 
1851 	/*
1852 	 * We acquire the vnode interlock early to ensure that the
1853 	 * vnode cannot be recycled by another process releasing a
1854 	 * holdcnt on it before we get it on both the vnode list
1855 	 * and the active vnode list. The mount mutex protects only
1856 	 * manipulation of the vnode list and the vnode freelist
1857 	 * mutex protects only manipulation of the active vnode list.
1858 	 * Hence the need to hold the vnode interlock throughout.
1859 	 */
1860 	MNT_ILOCK(mp);
1861 	VI_LOCK(vp);
1862 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1863 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1864 	    mp->mnt_nvnodelistsize == 0)) &&
1865 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1866 		VI_UNLOCK(vp);
1867 		MNT_IUNLOCK(mp);
1868 		if (dtr != NULL)
1869 			dtr(vp, dtr_arg);
1870 		return (EBUSY);
1871 	}
1872 	vp->v_mount = mp;
1873 	MNT_REF(mp);
1874 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1875 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1876 		("neg mount point vnode list size"));
1877 	mp->mnt_nvnodelistsize++;
1878 	VI_UNLOCK(vp);
1879 	MNT_IUNLOCK(mp);
1880 	return (0);
1881 }
1882 
1883 int
1884 insmntque(struct vnode *vp, struct mount *mp)
1885 {
1886 
1887 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1888 }
1889 
1890 /*
1891  * Flush out and invalidate all buffers associated with a bufobj
1892  * Called with the underlying object locked.
1893  */
1894 int
1895 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1896 {
1897 	int error;
1898 
1899 	BO_LOCK(bo);
1900 	if (flags & V_SAVE) {
1901 		error = bufobj_wwait(bo, slpflag, slptimeo);
1902 		if (error) {
1903 			BO_UNLOCK(bo);
1904 			return (error);
1905 		}
1906 		if (bo->bo_dirty.bv_cnt > 0) {
1907 			BO_UNLOCK(bo);
1908 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1909 				return (error);
1910 			/*
1911 			 * XXX We could save a lock/unlock if this was only
1912 			 * enabled under INVARIANTS
1913 			 */
1914 			BO_LOCK(bo);
1915 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1916 				panic("vinvalbuf: dirty bufs");
1917 		}
1918 	}
1919 	/*
1920 	 * If you alter this loop please notice that interlock is dropped and
1921 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1922 	 * no race conditions occur from this.
1923 	 */
1924 	do {
1925 		error = flushbuflist(&bo->bo_clean,
1926 		    flags, bo, slpflag, slptimeo);
1927 		if (error == 0 && !(flags & V_CLEANONLY))
1928 			error = flushbuflist(&bo->bo_dirty,
1929 			    flags, bo, slpflag, slptimeo);
1930 		if (error != 0 && error != EAGAIN) {
1931 			BO_UNLOCK(bo);
1932 			return (error);
1933 		}
1934 	} while (error != 0);
1935 
1936 	/*
1937 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1938 	 * have write I/O in-progress but if there is a VM object then the
1939 	 * VM object can also have read-I/O in-progress.
1940 	 */
1941 	do {
1942 		bufobj_wwait(bo, 0, 0);
1943 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
1944 			BO_UNLOCK(bo);
1945 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
1946 			BO_LOCK(bo);
1947 		}
1948 	} while (bo->bo_numoutput > 0);
1949 	BO_UNLOCK(bo);
1950 
1951 	/*
1952 	 * Destroy the copy in the VM cache, too.
1953 	 */
1954 	if (bo->bo_object != NULL &&
1955 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1956 		VM_OBJECT_WLOCK(bo->bo_object);
1957 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1958 		    OBJPR_CLEANONLY : 0);
1959 		VM_OBJECT_WUNLOCK(bo->bo_object);
1960 	}
1961 
1962 #ifdef INVARIANTS
1963 	BO_LOCK(bo);
1964 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1965 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1966 	    bo->bo_clean.bv_cnt > 0))
1967 		panic("vinvalbuf: flush failed");
1968 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1969 	    bo->bo_dirty.bv_cnt > 0)
1970 		panic("vinvalbuf: flush dirty failed");
1971 	BO_UNLOCK(bo);
1972 #endif
1973 	return (0);
1974 }
1975 
1976 /*
1977  * Flush out and invalidate all buffers associated with a vnode.
1978  * Called with the underlying object locked.
1979  */
1980 int
1981 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1982 {
1983 
1984 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1985 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1986 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1987 		return (0);
1988 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1989 }
1990 
1991 /*
1992  * Flush out buffers on the specified list.
1993  *
1994  */
1995 static int
1996 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1997     int slptimeo)
1998 {
1999 	struct buf *bp, *nbp;
2000 	int retval, error;
2001 	daddr_t lblkno;
2002 	b_xflags_t xflags;
2003 
2004 	ASSERT_BO_WLOCKED(bo);
2005 
2006 	retval = 0;
2007 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2008 		/*
2009 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2010 		 * do not skip any buffers. If we are flushing only V_NORMAL
2011 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2012 		 * flushing only V_ALT buffers then skip buffers not marked
2013 		 * as BX_ALTDATA.
2014 		 */
2015 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2016 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2017 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2018 			continue;
2019 		}
2020 		if (nbp != NULL) {
2021 			lblkno = nbp->b_lblkno;
2022 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2023 		}
2024 		retval = EAGAIN;
2025 		error = BUF_TIMELOCK(bp,
2026 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2027 		    "flushbuf", slpflag, slptimeo);
2028 		if (error) {
2029 			BO_LOCK(bo);
2030 			return (error != ENOLCK ? error : EAGAIN);
2031 		}
2032 		KASSERT(bp->b_bufobj == bo,
2033 		    ("bp %p wrong b_bufobj %p should be %p",
2034 		    bp, bp->b_bufobj, bo));
2035 		/*
2036 		 * XXX Since there are no node locks for NFS, I
2037 		 * believe there is a slight chance that a delayed
2038 		 * write will occur while sleeping just above, so
2039 		 * check for it.
2040 		 */
2041 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2042 		    (flags & V_SAVE)) {
2043 			bremfree(bp);
2044 			bp->b_flags |= B_ASYNC;
2045 			bwrite(bp);
2046 			BO_LOCK(bo);
2047 			return (EAGAIN);	/* XXX: why not loop ? */
2048 		}
2049 		bremfree(bp);
2050 		bp->b_flags |= (B_INVAL | B_RELBUF);
2051 		bp->b_flags &= ~B_ASYNC;
2052 		brelse(bp);
2053 		BO_LOCK(bo);
2054 		if (nbp == NULL)
2055 			break;
2056 		nbp = gbincore(bo, lblkno);
2057 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2058 		    != xflags)
2059 			break;			/* nbp invalid */
2060 	}
2061 	return (retval);
2062 }
2063 
2064 int
2065 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2066 {
2067 	struct buf *bp;
2068 	int error;
2069 	daddr_t lblkno;
2070 
2071 	ASSERT_BO_LOCKED(bo);
2072 
2073 	for (lblkno = startn;;) {
2074 again:
2075 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2076 		if (bp == NULL || bp->b_lblkno >= endn ||
2077 		    bp->b_lblkno < startn)
2078 			break;
2079 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2080 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2081 		if (error != 0) {
2082 			BO_RLOCK(bo);
2083 			if (error == ENOLCK)
2084 				goto again;
2085 			return (error);
2086 		}
2087 		KASSERT(bp->b_bufobj == bo,
2088 		    ("bp %p wrong b_bufobj %p should be %p",
2089 		    bp, bp->b_bufobj, bo));
2090 		lblkno = bp->b_lblkno + 1;
2091 		if ((bp->b_flags & B_MANAGED) == 0)
2092 			bremfree(bp);
2093 		bp->b_flags |= B_RELBUF;
2094 		/*
2095 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2096 		 * pages backing each buffer in the range are unlikely to be
2097 		 * reused.  Dirty buffers will have the hint applied once
2098 		 * they've been written.
2099 		 */
2100 		if ((bp->b_flags & B_VMIO) != 0)
2101 			bp->b_flags |= B_NOREUSE;
2102 		brelse(bp);
2103 		BO_RLOCK(bo);
2104 	}
2105 	return (0);
2106 }
2107 
2108 /*
2109  * Truncate a file's buffer and pages to a specified length.  This
2110  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2111  * sync activity.
2112  */
2113 int
2114 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2115 {
2116 	struct buf *bp, *nbp;
2117 	struct bufobj *bo;
2118 	daddr_t startlbn;
2119 
2120 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2121 	    vp, blksize, (uintmax_t)length);
2122 
2123 	/*
2124 	 * Round up to the *next* lbn.
2125 	 */
2126 	startlbn = howmany(length, blksize);
2127 
2128 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2129 
2130 	bo = &vp->v_bufobj;
2131 restart_unlocked:
2132 	BO_LOCK(bo);
2133 
2134 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2135 		;
2136 
2137 	if (length > 0) {
2138 restartsync:
2139 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2140 			if (bp->b_lblkno > 0)
2141 				continue;
2142 			/*
2143 			 * Since we hold the vnode lock this should only
2144 			 * fail if we're racing with the buf daemon.
2145 			 */
2146 			if (BUF_LOCK(bp,
2147 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2148 			    BO_LOCKPTR(bo)) == ENOLCK)
2149 				goto restart_unlocked;
2150 
2151 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2152 			    ("buf(%p) on dirty queue without DELWRI", bp));
2153 
2154 			bremfree(bp);
2155 			bawrite(bp);
2156 			BO_LOCK(bo);
2157 			goto restartsync;
2158 		}
2159 	}
2160 
2161 	bufobj_wwait(bo, 0, 0);
2162 	BO_UNLOCK(bo);
2163 	vnode_pager_setsize(vp, length);
2164 
2165 	return (0);
2166 }
2167 
2168 /*
2169  * Invalidate the cached pages of a file's buffer within the range of block
2170  * numbers [startlbn, endlbn).
2171  */
2172 void
2173 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2174     int blksize)
2175 {
2176 	struct bufobj *bo;
2177 	off_t start, end;
2178 
2179 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2180 
2181 	start = blksize * startlbn;
2182 	end = blksize * endlbn;
2183 
2184 	bo = &vp->v_bufobj;
2185 	BO_LOCK(bo);
2186 	MPASS(blksize == bo->bo_bsize);
2187 
2188 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2189 		;
2190 
2191 	BO_UNLOCK(bo);
2192 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2193 }
2194 
2195 static int
2196 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2197     daddr_t startlbn, daddr_t endlbn)
2198 {
2199 	struct buf *bp, *nbp;
2200 	bool anyfreed;
2201 
2202 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2203 	ASSERT_BO_LOCKED(bo);
2204 
2205 	do {
2206 		anyfreed = false;
2207 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2208 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2209 				continue;
2210 			if (BUF_LOCK(bp,
2211 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2212 			    BO_LOCKPTR(bo)) == ENOLCK) {
2213 				BO_LOCK(bo);
2214 				return (EAGAIN);
2215 			}
2216 
2217 			bremfree(bp);
2218 			bp->b_flags |= B_INVAL | B_RELBUF;
2219 			bp->b_flags &= ~B_ASYNC;
2220 			brelse(bp);
2221 			anyfreed = true;
2222 
2223 			BO_LOCK(bo);
2224 			if (nbp != NULL &&
2225 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2226 			    nbp->b_vp != vp ||
2227 			    (nbp->b_flags & B_DELWRI) != 0))
2228 				return (EAGAIN);
2229 		}
2230 
2231 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2232 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2233 				continue;
2234 			if (BUF_LOCK(bp,
2235 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2236 			    BO_LOCKPTR(bo)) == ENOLCK) {
2237 				BO_LOCK(bo);
2238 				return (EAGAIN);
2239 			}
2240 			bremfree(bp);
2241 			bp->b_flags |= B_INVAL | B_RELBUF;
2242 			bp->b_flags &= ~B_ASYNC;
2243 			brelse(bp);
2244 			anyfreed = true;
2245 
2246 			BO_LOCK(bo);
2247 			if (nbp != NULL &&
2248 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2249 			    (nbp->b_vp != vp) ||
2250 			    (nbp->b_flags & B_DELWRI) == 0))
2251 				return (EAGAIN);
2252 		}
2253 	} while (anyfreed);
2254 	return (0);
2255 }
2256 
2257 static void
2258 buf_vlist_remove(struct buf *bp)
2259 {
2260 	struct bufv *bv;
2261 
2262 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2263 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2264 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
2265 	    (BX_VNDIRTY|BX_VNCLEAN),
2266 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
2267 	if (bp->b_xflags & BX_VNDIRTY)
2268 		bv = &bp->b_bufobj->bo_dirty;
2269 	else
2270 		bv = &bp->b_bufobj->bo_clean;
2271 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2272 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2273 	bv->bv_cnt--;
2274 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2275 }
2276 
2277 /*
2278  * Add the buffer to the sorted clean or dirty block list.
2279  *
2280  * NOTE: xflags is passed as a constant, optimizing this inline function!
2281  */
2282 static void
2283 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2284 {
2285 	struct bufv *bv;
2286 	struct buf *n;
2287 	int error;
2288 
2289 	ASSERT_BO_WLOCKED(bo);
2290 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2291 	    ("dead bo %p", bo));
2292 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2293 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2294 	bp->b_xflags |= xflags;
2295 	if (xflags & BX_VNDIRTY)
2296 		bv = &bo->bo_dirty;
2297 	else
2298 		bv = &bo->bo_clean;
2299 
2300 	/*
2301 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2302 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2303 	 * than _ge.
2304 	 */
2305 	if (bv->bv_cnt == 0 ||
2306 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2307 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2308 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2309 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2310 	else
2311 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2312 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2313 	if (error)
2314 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2315 	bv->bv_cnt++;
2316 }
2317 
2318 /*
2319  * Look up a buffer using the buffer tries.
2320  */
2321 struct buf *
2322 gbincore(struct bufobj *bo, daddr_t lblkno)
2323 {
2324 	struct buf *bp;
2325 
2326 	ASSERT_BO_LOCKED(bo);
2327 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2328 	if (bp != NULL)
2329 		return (bp);
2330 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
2331 }
2332 
2333 /*
2334  * Associate a buffer with a vnode.
2335  */
2336 void
2337 bgetvp(struct vnode *vp, struct buf *bp)
2338 {
2339 	struct bufobj *bo;
2340 
2341 	bo = &vp->v_bufobj;
2342 	ASSERT_BO_WLOCKED(bo);
2343 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2344 
2345 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2346 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2347 	    ("bgetvp: bp already attached! %p", bp));
2348 
2349 	vhold(vp);
2350 	bp->b_vp = vp;
2351 	bp->b_bufobj = bo;
2352 	/*
2353 	 * Insert onto list for new vnode.
2354 	 */
2355 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2356 }
2357 
2358 /*
2359  * Disassociate a buffer from a vnode.
2360  */
2361 void
2362 brelvp(struct buf *bp)
2363 {
2364 	struct bufobj *bo;
2365 	struct vnode *vp;
2366 
2367 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2368 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2369 
2370 	/*
2371 	 * Delete from old vnode list, if on one.
2372 	 */
2373 	vp = bp->b_vp;		/* XXX */
2374 	bo = bp->b_bufobj;
2375 	BO_LOCK(bo);
2376 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2377 		buf_vlist_remove(bp);
2378 	else
2379 		panic("brelvp: Buffer %p not on queue.", bp);
2380 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2381 		bo->bo_flag &= ~BO_ONWORKLST;
2382 		mtx_lock(&sync_mtx);
2383 		LIST_REMOVE(bo, bo_synclist);
2384 		syncer_worklist_len--;
2385 		mtx_unlock(&sync_mtx);
2386 	}
2387 	bp->b_vp = NULL;
2388 	bp->b_bufobj = NULL;
2389 	BO_UNLOCK(bo);
2390 	vdrop(vp);
2391 }
2392 
2393 /*
2394  * Add an item to the syncer work queue.
2395  */
2396 static void
2397 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2398 {
2399 	int slot;
2400 
2401 	ASSERT_BO_WLOCKED(bo);
2402 
2403 	mtx_lock(&sync_mtx);
2404 	if (bo->bo_flag & BO_ONWORKLST)
2405 		LIST_REMOVE(bo, bo_synclist);
2406 	else {
2407 		bo->bo_flag |= BO_ONWORKLST;
2408 		syncer_worklist_len++;
2409 	}
2410 
2411 	if (delay > syncer_maxdelay - 2)
2412 		delay = syncer_maxdelay - 2;
2413 	slot = (syncer_delayno + delay) & syncer_mask;
2414 
2415 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2416 	mtx_unlock(&sync_mtx);
2417 }
2418 
2419 static int
2420 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2421 {
2422 	int error, len;
2423 
2424 	mtx_lock(&sync_mtx);
2425 	len = syncer_worklist_len - sync_vnode_count;
2426 	mtx_unlock(&sync_mtx);
2427 	error = SYSCTL_OUT(req, &len, sizeof(len));
2428 	return (error);
2429 }
2430 
2431 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2432     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2433     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2434 
2435 static struct proc *updateproc;
2436 static void sched_sync(void);
2437 static struct kproc_desc up_kp = {
2438 	"syncer",
2439 	sched_sync,
2440 	&updateproc
2441 };
2442 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2443 
2444 static int
2445 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2446 {
2447 	struct vnode *vp;
2448 	struct mount *mp;
2449 
2450 	*bo = LIST_FIRST(slp);
2451 	if (*bo == NULL)
2452 		return (0);
2453 	vp = bo2vnode(*bo);
2454 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2455 		return (1);
2456 	/*
2457 	 * We use vhold in case the vnode does not
2458 	 * successfully sync.  vhold prevents the vnode from
2459 	 * going away when we unlock the sync_mtx so that
2460 	 * we can acquire the vnode interlock.
2461 	 */
2462 	vholdl(vp);
2463 	mtx_unlock(&sync_mtx);
2464 	VI_UNLOCK(vp);
2465 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2466 		vdrop(vp);
2467 		mtx_lock(&sync_mtx);
2468 		return (*bo == LIST_FIRST(slp));
2469 	}
2470 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2471 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2472 	VOP_UNLOCK(vp);
2473 	vn_finished_write(mp);
2474 	BO_LOCK(*bo);
2475 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2476 		/*
2477 		 * Put us back on the worklist.  The worklist
2478 		 * routine will remove us from our current
2479 		 * position and then add us back in at a later
2480 		 * position.
2481 		 */
2482 		vn_syncer_add_to_worklist(*bo, syncdelay);
2483 	}
2484 	BO_UNLOCK(*bo);
2485 	vdrop(vp);
2486 	mtx_lock(&sync_mtx);
2487 	return (0);
2488 }
2489 
2490 static int first_printf = 1;
2491 
2492 /*
2493  * System filesystem synchronizer daemon.
2494  */
2495 static void
2496 sched_sync(void)
2497 {
2498 	struct synclist *next, *slp;
2499 	struct bufobj *bo;
2500 	long starttime;
2501 	struct thread *td = curthread;
2502 	int last_work_seen;
2503 	int net_worklist_len;
2504 	int syncer_final_iter;
2505 	int error;
2506 
2507 	last_work_seen = 0;
2508 	syncer_final_iter = 0;
2509 	syncer_state = SYNCER_RUNNING;
2510 	starttime = time_uptime;
2511 	td->td_pflags |= TDP_NORUNNINGBUF;
2512 
2513 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2514 	    SHUTDOWN_PRI_LAST);
2515 
2516 	mtx_lock(&sync_mtx);
2517 	for (;;) {
2518 		if (syncer_state == SYNCER_FINAL_DELAY &&
2519 		    syncer_final_iter == 0) {
2520 			mtx_unlock(&sync_mtx);
2521 			kproc_suspend_check(td->td_proc);
2522 			mtx_lock(&sync_mtx);
2523 		}
2524 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2525 		if (syncer_state != SYNCER_RUNNING &&
2526 		    starttime != time_uptime) {
2527 			if (first_printf) {
2528 				printf("\nSyncing disks, vnodes remaining... ");
2529 				first_printf = 0;
2530 			}
2531 			printf("%d ", net_worklist_len);
2532 		}
2533 		starttime = time_uptime;
2534 
2535 		/*
2536 		 * Push files whose dirty time has expired.  Be careful
2537 		 * of interrupt race on slp queue.
2538 		 *
2539 		 * Skip over empty worklist slots when shutting down.
2540 		 */
2541 		do {
2542 			slp = &syncer_workitem_pending[syncer_delayno];
2543 			syncer_delayno += 1;
2544 			if (syncer_delayno == syncer_maxdelay)
2545 				syncer_delayno = 0;
2546 			next = &syncer_workitem_pending[syncer_delayno];
2547 			/*
2548 			 * If the worklist has wrapped since the
2549 			 * it was emptied of all but syncer vnodes,
2550 			 * switch to the FINAL_DELAY state and run
2551 			 * for one more second.
2552 			 */
2553 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2554 			    net_worklist_len == 0 &&
2555 			    last_work_seen == syncer_delayno) {
2556 				syncer_state = SYNCER_FINAL_DELAY;
2557 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2558 			}
2559 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2560 		    syncer_worklist_len > 0);
2561 
2562 		/*
2563 		 * Keep track of the last time there was anything
2564 		 * on the worklist other than syncer vnodes.
2565 		 * Return to the SHUTTING_DOWN state if any
2566 		 * new work appears.
2567 		 */
2568 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2569 			last_work_seen = syncer_delayno;
2570 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2571 			syncer_state = SYNCER_SHUTTING_DOWN;
2572 		while (!LIST_EMPTY(slp)) {
2573 			error = sync_vnode(slp, &bo, td);
2574 			if (error == 1) {
2575 				LIST_REMOVE(bo, bo_synclist);
2576 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2577 				continue;
2578 			}
2579 
2580 			if (first_printf == 0) {
2581 				/*
2582 				 * Drop the sync mutex, because some watchdog
2583 				 * drivers need to sleep while patting
2584 				 */
2585 				mtx_unlock(&sync_mtx);
2586 				wdog_kern_pat(WD_LASTVAL);
2587 				mtx_lock(&sync_mtx);
2588 			}
2589 
2590 		}
2591 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2592 			syncer_final_iter--;
2593 		/*
2594 		 * The variable rushjob allows the kernel to speed up the
2595 		 * processing of the filesystem syncer process. A rushjob
2596 		 * value of N tells the filesystem syncer to process the next
2597 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2598 		 * is used by the soft update code to speed up the filesystem
2599 		 * syncer process when the incore state is getting so far
2600 		 * ahead of the disk that the kernel memory pool is being
2601 		 * threatened with exhaustion.
2602 		 */
2603 		if (rushjob > 0) {
2604 			rushjob -= 1;
2605 			continue;
2606 		}
2607 		/*
2608 		 * Just sleep for a short period of time between
2609 		 * iterations when shutting down to allow some I/O
2610 		 * to happen.
2611 		 *
2612 		 * If it has taken us less than a second to process the
2613 		 * current work, then wait. Otherwise start right over
2614 		 * again. We can still lose time if any single round
2615 		 * takes more than two seconds, but it does not really
2616 		 * matter as we are just trying to generally pace the
2617 		 * filesystem activity.
2618 		 */
2619 		if (syncer_state != SYNCER_RUNNING ||
2620 		    time_uptime == starttime) {
2621 			thread_lock(td);
2622 			sched_prio(td, PPAUSE);
2623 			thread_unlock(td);
2624 		}
2625 		if (syncer_state != SYNCER_RUNNING)
2626 			cv_timedwait(&sync_wakeup, &sync_mtx,
2627 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2628 		else if (time_uptime == starttime)
2629 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2630 	}
2631 }
2632 
2633 /*
2634  * Request the syncer daemon to speed up its work.
2635  * We never push it to speed up more than half of its
2636  * normal turn time, otherwise it could take over the cpu.
2637  */
2638 int
2639 speedup_syncer(void)
2640 {
2641 	int ret = 0;
2642 
2643 	mtx_lock(&sync_mtx);
2644 	if (rushjob < syncdelay / 2) {
2645 		rushjob += 1;
2646 		stat_rush_requests += 1;
2647 		ret = 1;
2648 	}
2649 	mtx_unlock(&sync_mtx);
2650 	cv_broadcast(&sync_wakeup);
2651 	return (ret);
2652 }
2653 
2654 /*
2655  * Tell the syncer to speed up its work and run though its work
2656  * list several times, then tell it to shut down.
2657  */
2658 static void
2659 syncer_shutdown(void *arg, int howto)
2660 {
2661 
2662 	if (howto & RB_NOSYNC)
2663 		return;
2664 	mtx_lock(&sync_mtx);
2665 	syncer_state = SYNCER_SHUTTING_DOWN;
2666 	rushjob = 0;
2667 	mtx_unlock(&sync_mtx);
2668 	cv_broadcast(&sync_wakeup);
2669 	kproc_shutdown(arg, howto);
2670 }
2671 
2672 void
2673 syncer_suspend(void)
2674 {
2675 
2676 	syncer_shutdown(updateproc, 0);
2677 }
2678 
2679 void
2680 syncer_resume(void)
2681 {
2682 
2683 	mtx_lock(&sync_mtx);
2684 	first_printf = 1;
2685 	syncer_state = SYNCER_RUNNING;
2686 	mtx_unlock(&sync_mtx);
2687 	cv_broadcast(&sync_wakeup);
2688 	kproc_resume(updateproc);
2689 }
2690 
2691 /*
2692  * Reassign a buffer from one vnode to another.
2693  * Used to assign file specific control information
2694  * (indirect blocks) to the vnode to which they belong.
2695  */
2696 void
2697 reassignbuf(struct buf *bp)
2698 {
2699 	struct vnode *vp;
2700 	struct bufobj *bo;
2701 	int delay;
2702 #ifdef INVARIANTS
2703 	struct bufv *bv;
2704 #endif
2705 
2706 	vp = bp->b_vp;
2707 	bo = bp->b_bufobj;
2708 	++reassignbufcalls;
2709 
2710 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2711 	    bp, bp->b_vp, bp->b_flags);
2712 	/*
2713 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2714 	 * is not fully linked in.
2715 	 */
2716 	if (bp->b_flags & B_PAGING)
2717 		panic("cannot reassign paging buffer");
2718 
2719 	/*
2720 	 * Delete from old vnode list, if on one.
2721 	 */
2722 	BO_LOCK(bo);
2723 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2724 		buf_vlist_remove(bp);
2725 	else
2726 		panic("reassignbuf: Buffer %p not on queue.", bp);
2727 	/*
2728 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2729 	 * of clean buffers.
2730 	 */
2731 	if (bp->b_flags & B_DELWRI) {
2732 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2733 			switch (vp->v_type) {
2734 			case VDIR:
2735 				delay = dirdelay;
2736 				break;
2737 			case VCHR:
2738 				delay = metadelay;
2739 				break;
2740 			default:
2741 				delay = filedelay;
2742 			}
2743 			vn_syncer_add_to_worklist(bo, delay);
2744 		}
2745 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2746 	} else {
2747 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2748 
2749 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2750 			mtx_lock(&sync_mtx);
2751 			LIST_REMOVE(bo, bo_synclist);
2752 			syncer_worklist_len--;
2753 			mtx_unlock(&sync_mtx);
2754 			bo->bo_flag &= ~BO_ONWORKLST;
2755 		}
2756 	}
2757 #ifdef INVARIANTS
2758 	bv = &bo->bo_clean;
2759 	bp = TAILQ_FIRST(&bv->bv_hd);
2760 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2761 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2762 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2763 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2764 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2765 	bv = &bo->bo_dirty;
2766 	bp = TAILQ_FIRST(&bv->bv_hd);
2767 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2768 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2769 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2770 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2771 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2772 #endif
2773 	BO_UNLOCK(bo);
2774 }
2775 
2776 static void
2777 v_init_counters(struct vnode *vp)
2778 {
2779 
2780 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2781 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2782 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2783 
2784 	refcount_init(&vp->v_holdcnt, 1);
2785 	refcount_init(&vp->v_usecount, 1);
2786 }
2787 
2788 /*
2789  * Increment si_usecount of the associated device, if any.
2790  */
2791 static void
2792 v_incr_devcount(struct vnode *vp)
2793 {
2794 
2795 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2796 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2797 		dev_lock();
2798 		vp->v_rdev->si_usecount++;
2799 		dev_unlock();
2800 	}
2801 }
2802 
2803 /*
2804  * Decrement si_usecount of the associated device, if any.
2805  */
2806 static void
2807 v_decr_devcount(struct vnode *vp)
2808 {
2809 
2810 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2811 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2812 		dev_lock();
2813 		vp->v_rdev->si_usecount--;
2814 		dev_unlock();
2815 	}
2816 }
2817 
2818 /*
2819  * Grab a particular vnode from the free list, increment its
2820  * reference count and lock it.  VIRF_DOOMED is set if the vnode
2821  * is being destroyed.  Only callers who specify LK_RETRY will
2822  * see doomed vnodes.  If inactive processing was delayed in
2823  * vput try to do it here.
2824  *
2825  * Both holdcnt and usecount can be manipulated using atomics without holding
2826  * any locks except in these cases which require the vnode interlock:
2827  * holdcnt: 1->0 and 0->1
2828  * usecount: 0->1
2829  *
2830  * usecount is permitted to transition 1->0 without the interlock because
2831  * vnode is kept live by holdcnt.
2832  */
2833 static enum vgetstate __always_inline
2834 _vget_prep(struct vnode *vp, bool interlock)
2835 {
2836 	enum vgetstate vs;
2837 
2838 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2839 		vs = VGET_USECOUNT;
2840 	} else {
2841 		if (interlock)
2842 			vholdl(vp);
2843 		else
2844 			vhold(vp);
2845 		vs = VGET_HOLDCNT;
2846 	}
2847 	return (vs);
2848 }
2849 
2850 enum vgetstate
2851 vget_prep(struct vnode *vp)
2852 {
2853 
2854 	return (_vget_prep(vp, false));
2855 }
2856 
2857 int
2858 vget(struct vnode *vp, int flags, struct thread *td)
2859 {
2860 	enum vgetstate vs;
2861 
2862 	MPASS(td == curthread);
2863 
2864 	vs = _vget_prep(vp, (flags & LK_INTERLOCK) != 0);
2865 	return (vget_finish(vp, flags, vs));
2866 }
2867 
2868 int
2869 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2870 {
2871 	int error, oweinact;
2872 
2873 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2874 	    ("%s: invalid lock operation", __func__));
2875 
2876 	if ((flags & LK_INTERLOCK) != 0)
2877 		ASSERT_VI_LOCKED(vp, __func__);
2878 	else
2879 		ASSERT_VI_UNLOCKED(vp, __func__);
2880 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
2881 	if (vs == VGET_USECOUNT) {
2882 		VNASSERT(vp->v_usecount > 0, vp,
2883 		    ("%s: vnode without usecount when VGET_USECOUNT was passed",
2884 		    __func__));
2885 	}
2886 
2887 	if ((error = vn_lock(vp, flags)) != 0) {
2888 		if (vs == VGET_USECOUNT)
2889 			vrele(vp);
2890 		else
2891 			vdrop(vp);
2892 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2893 		    vp);
2894 		return (error);
2895 	}
2896 
2897 	if (vs == VGET_USECOUNT) {
2898 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2899 		    ("%s: vnode with usecount and VI_OWEINACT set", __func__));
2900 		return (0);
2901 	}
2902 
2903 	/*
2904 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
2905 	 * the vnode around. Otherwise someone else lended their hold count and
2906 	 * we have to drop ours.
2907 	 */
2908 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2909 #ifdef INVARIANTS
2910 		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2911 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2912 #else
2913 		refcount_release(&vp->v_holdcnt);
2914 #endif
2915 		VNODE_REFCOUNT_FENCE_ACQ();
2916 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2917 		    ("%s: vnode with usecount and VI_OWEINACT set", __func__));
2918 		return (0);
2919 	}
2920 
2921 	/*
2922 	 * We don't guarantee that any particular close will
2923 	 * trigger inactive processing so just make a best effort
2924 	 * here at preventing a reference to a removed file.  If
2925 	 * we don't succeed no harm is done.
2926 	 *
2927 	 * Upgrade our holdcnt to a usecount.
2928 	 */
2929 	VI_LOCK(vp);
2930 	/*
2931 	 * See the previous section. By the time we get here we may find
2932 	 * ourselves in the same spot.
2933 	 */
2934 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2935 #ifdef INVARIANTS
2936 		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2937 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2938 #else
2939 		refcount_release(&vp->v_holdcnt);
2940 #endif
2941 		VNODE_REFCOUNT_FENCE_ACQ();
2942 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2943 		    ("%s: vnode with usecount and VI_OWEINACT set",
2944 		    __func__));
2945 		VI_UNLOCK(vp);
2946 		return (0);
2947 	}
2948 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
2949 		oweinact = 0;
2950 	} else {
2951 		oweinact = 1;
2952 		vp->v_iflag &= ~VI_OWEINACT;
2953 		VNODE_REFCOUNT_FENCE_REL();
2954 	}
2955 	v_incr_devcount(vp);
2956 	refcount_acquire(&vp->v_usecount);
2957 	if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2958 	    (flags & LK_NOWAIT) == 0)
2959 		vinactive(vp);
2960 	VI_UNLOCK(vp);
2961 	return (0);
2962 }
2963 
2964 /*
2965  * Increase the reference (use) and hold count of a vnode.
2966  * This will also remove the vnode from the free list if it is presently free.
2967  */
2968 void
2969 vref(struct vnode *vp)
2970 {
2971 
2972 	ASSERT_VI_UNLOCKED(vp, __func__);
2973 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2974 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2975 		VNODE_REFCOUNT_FENCE_ACQ();
2976 		VNASSERT(vp->v_holdcnt > 0, vp,
2977 		    ("%s: active vnode not held", __func__));
2978 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2979 		    ("%s: vnode with usecount and VI_OWEINACT set", __func__));
2980 		return;
2981 	}
2982 	VI_LOCK(vp);
2983 	vrefl(vp);
2984 	VI_UNLOCK(vp);
2985 }
2986 
2987 void
2988 vrefl(struct vnode *vp)
2989 {
2990 
2991 	ASSERT_VI_LOCKED(vp, __func__);
2992 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2993 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2994 		VNODE_REFCOUNT_FENCE_ACQ();
2995 		VNASSERT(vp->v_holdcnt > 0, vp,
2996 		    ("%s: active vnode not held", __func__));
2997 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2998 		    ("%s: vnode with usecount and VI_OWEINACT set", __func__));
2999 		return;
3000 	}
3001 	vholdl(vp);
3002 	if ((vp->v_iflag & VI_OWEINACT) != 0) {
3003 		vp->v_iflag &= ~VI_OWEINACT;
3004 		VNODE_REFCOUNT_FENCE_REL();
3005 	}
3006 	v_incr_devcount(vp);
3007 	refcount_acquire(&vp->v_usecount);
3008 }
3009 
3010 void
3011 vrefact(struct vnode *vp)
3012 {
3013 
3014 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3015 #ifdef INVARIANTS
3016 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3017 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3018 #else
3019 	refcount_acquire(&vp->v_usecount);
3020 #endif
3021 }
3022 
3023 /*
3024  * Return reference count of a vnode.
3025  *
3026  * The results of this call are only guaranteed when some mechanism is used to
3027  * stop other processes from gaining references to the vnode.  This may be the
3028  * case if the caller holds the only reference.  This is also useful when stale
3029  * data is acceptable as race conditions may be accounted for by some other
3030  * means.
3031  */
3032 int
3033 vrefcnt(struct vnode *vp)
3034 {
3035 
3036 	return (vp->v_usecount);
3037 }
3038 
3039 void
3040 vlazy(struct vnode *vp)
3041 {
3042 	struct mount *mp;
3043 
3044 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3045 
3046 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3047 		return;
3048 	mp = vp->v_mount;
3049 	mtx_lock(&mp->mnt_listmtx);
3050 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3051 		vp->v_mflag |= VMP_LAZYLIST;
3052 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3053 		mp->mnt_lazyvnodelistsize++;
3054 	}
3055 	mtx_unlock(&mp->mnt_listmtx);
3056 }
3057 
3058 static void
3059 vdefer_inactive(struct vnode *vp)
3060 {
3061 
3062 	ASSERT_VI_LOCKED(vp, __func__);
3063 	VNASSERT(vp->v_iflag & VI_OWEINACT, vp,
3064 	    ("%s: vnode without VI_OWEINACT", __func__));
3065 	if (VN_IS_DOOMED(vp)) {
3066 		vdropl(vp);
3067 		return;
3068 	}
3069 	if (vp->v_iflag & VI_DEFINACT) {
3070 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3071 		vdropl(vp);
3072 		return;
3073 	}
3074 	vlazy(vp);
3075 	vp->v_iflag |= VI_DEFINACT;
3076 	VI_UNLOCK(vp);
3077 	counter_u64_add(deferred_inact, 1);
3078 }
3079 
3080 static void
3081 vdefer_inactive_cond(struct vnode *vp)
3082 {
3083 
3084 	VI_LOCK(vp);
3085 	VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
3086 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3087 		vdropl(vp);
3088 		return;
3089 	}
3090 	vdefer_inactive(vp);
3091 }
3092 
3093 enum vputx_op { VPUTX_VRELE, VPUTX_VPUT, VPUTX_VUNREF };
3094 
3095 /*
3096  * Decrement the use and hold counts for a vnode.
3097  *
3098  * See an explanation near vget() as to why atomic operation is safe.
3099  */
3100 static void
3101 vputx(struct vnode *vp, enum vputx_op func)
3102 {
3103 	int error;
3104 
3105 	KASSERT(vp != NULL, ("vputx: null vp"));
3106 	if (func == VPUTX_VUNREF)
3107 		ASSERT_VOP_LOCKED(vp, "vunref");
3108 	ASSERT_VI_UNLOCKED(vp, __func__);
3109 	VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp,
3110 	    ("%s: wrong ref counts", __func__));
3111 
3112 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3113 
3114 	/*
3115 	 * We want to hold the vnode until the inactive finishes to
3116 	 * prevent vgone() races.  We drop the use count here and the
3117 	 * hold count below when we're done.
3118 	 *
3119 	 * If we release the last usecount we take ownership of the hold
3120 	 * count which provides liveness of the vnode, in which case we
3121 	 * have to vdrop.
3122 	 */
3123 	if (!refcount_release(&vp->v_usecount))
3124 		return;
3125 	VI_LOCK(vp);
3126 	v_decr_devcount(vp);
3127 	/*
3128 	 * By the time we got here someone else might have transitioned
3129 	 * the count back to > 0.
3130 	 */
3131 	if (vp->v_usecount > 0) {
3132 		vdropl(vp);
3133 		return;
3134 	}
3135 	if (vp->v_iflag & VI_DOINGINACT) {
3136 		vdropl(vp);
3137 		return;
3138 	}
3139 
3140 	/*
3141 	 * Check if the fs wants to perform inactive processing. Note we
3142 	 * may be only holding the interlock, in which case it is possible
3143 	 * someone else called vgone on the vnode and ->v_data is now NULL.
3144 	 * Since vgone performs inactive on its own there is nothing to do
3145 	 * here but to drop our hold count.
3146 	 */
3147 	if (__predict_false(VN_IS_DOOMED(vp)) ||
3148 	    VOP_NEED_INACTIVE(vp) == 0) {
3149 		vdropl(vp);
3150 		return;
3151 	}
3152 
3153 	/*
3154 	 * We must call VOP_INACTIVE with the node locked. Mark
3155 	 * as VI_DOINGINACT to avoid recursion.
3156 	 */
3157 	vp->v_iflag |= VI_OWEINACT;
3158 	switch (func) {
3159 	case VPUTX_VRELE:
3160 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3161 		VI_LOCK(vp);
3162 		break;
3163 	case VPUTX_VPUT:
3164 		error = VOP_LOCK(vp, LK_EXCLUSIVE | LK_INTERLOCK | LK_NOWAIT);
3165 		VI_LOCK(vp);
3166 		break;
3167 	case VPUTX_VUNREF:
3168 		error = 0;
3169 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3170 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3171 			VI_LOCK(vp);
3172 		}
3173 		break;
3174 	}
3175 	VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
3176 	    ("vnode with usecount and VI_OWEINACT set"));
3177 	if (error == 0) {
3178 		if (vp->v_iflag & VI_OWEINACT)
3179 			vinactive(vp);
3180 		if (func != VPUTX_VUNREF)
3181 			VOP_UNLOCK(vp);
3182 		vdropl(vp);
3183 	} else if (vp->v_iflag & VI_OWEINACT) {
3184 		vdefer_inactive(vp);
3185 	} else {
3186 		vdropl(vp);
3187 	}
3188 }
3189 
3190 /*
3191  * Vnode put/release.
3192  * If count drops to zero, call inactive routine and return to freelist.
3193  */
3194 void
3195 vrele(struct vnode *vp)
3196 {
3197 
3198 	vputx(vp, VPUTX_VRELE);
3199 }
3200 
3201 /*
3202  * Release an already locked vnode.  This give the same effects as
3203  * unlock+vrele(), but takes less time and avoids releasing and
3204  * re-aquiring the lock (as vrele() acquires the lock internally.)
3205  *
3206  * It is an invariant that all VOP_* calls operate on a held vnode.
3207  * We may be only having an implicit hold stemming from our usecount,
3208  * which we are about to release. If we unlock the vnode afterwards we
3209  * open a time window where someone else dropped the last usecount and
3210  * proceeded to free the vnode before our unlock finished. For this
3211  * reason we unlock the vnode early. This is a little bit wasteful as
3212  * it may be the vnode is exclusively locked and inactive processing is
3213  * needed, in which case we are adding work.
3214  */
3215 void
3216 vput(struct vnode *vp)
3217 {
3218 
3219 	VOP_UNLOCK(vp);
3220 	vputx(vp, VPUTX_VPUT);
3221 }
3222 
3223 /*
3224  * Release an exclusively locked vnode. Do not unlock the vnode lock.
3225  */
3226 void
3227 vunref(struct vnode *vp)
3228 {
3229 
3230 	vputx(vp, VPUTX_VUNREF);
3231 }
3232 
3233 /*
3234  * Increase the hold count and activate if this is the first reference.
3235  */
3236 static void
3237 vhold_activate(struct vnode *vp)
3238 {
3239 	struct vdbatch *vd;
3240 
3241 	ASSERT_VI_LOCKED(vp, __func__);
3242 	VNASSERT(vp->v_holdcnt == 0, vp,
3243 	    ("%s: wrong hold count", __func__));
3244 	VNASSERT(vp->v_op != NULL, vp,
3245 	    ("%s: vnode already reclaimed.", __func__));
3246 	critical_enter();
3247 	vd = DPCPU_PTR(vd);
3248 	vd->freevnodes--;
3249 	critical_exit();
3250 	refcount_acquire(&vp->v_holdcnt);
3251 }
3252 
3253 void
3254 vhold(struct vnode *vp)
3255 {
3256 
3257 	ASSERT_VI_UNLOCKED(vp, __func__);
3258 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3259 	if (refcount_acquire_if_not_zero(&vp->v_holdcnt))
3260 		return;
3261 	VI_LOCK(vp);
3262 	vholdl(vp);
3263 	VI_UNLOCK(vp);
3264 }
3265 
3266 void
3267 vholdl(struct vnode *vp)
3268 {
3269 
3270 	ASSERT_VI_LOCKED(vp, __func__);
3271 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3272 	if (vp->v_holdcnt > 0) {
3273 		refcount_acquire(&vp->v_holdcnt);
3274 		return;
3275 	}
3276 	vhold_activate(vp);
3277 }
3278 
3279 void
3280 vholdnz(struct vnode *vp)
3281 {
3282 
3283 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3284 #ifdef INVARIANTS
3285 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3286 	VNASSERT(old > 0, vp, ("%s: wrong hold count %d", __func__, old));
3287 #else
3288 	atomic_add_int(&vp->v_holdcnt, 1);
3289 #endif
3290 }
3291 
3292 static void __noinline
3293 vdbatch_process(struct vdbatch *vd)
3294 {
3295 	struct vnode *vp;
3296 	int i;
3297 
3298 	mtx_assert(&vd->lock, MA_OWNED);
3299 	MPASS(curthread->td_pinned > 0);
3300 	MPASS(vd->index == VDBATCH_SIZE);
3301 
3302 	mtx_lock(&vnode_list_mtx);
3303 	critical_enter();
3304 	freevnodes += vd->freevnodes;
3305 	for (i = 0; i < VDBATCH_SIZE; i++) {
3306 		vp = vd->tab[i];
3307 		TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3308 		TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3309 		MPASS(vp->v_dbatchcpu != NOCPU);
3310 		vp->v_dbatchcpu = NOCPU;
3311 	}
3312 	mtx_unlock(&vnode_list_mtx);
3313 	critical_exit();
3314 	vd->freevnodes = 0;
3315 	bzero(vd->tab, sizeof(vd->tab));
3316 	vd->index = 0;
3317 }
3318 
3319 static void
3320 vdbatch_enqueue(struct vnode *vp)
3321 {
3322 	struct vdbatch *vd;
3323 
3324 	ASSERT_VI_LOCKED(vp, __func__);
3325 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3326 	    ("%s: deferring requeue of a doomed vnode", __func__));
3327 
3328 	critical_enter();
3329 	vd = DPCPU_PTR(vd);
3330 	vd->freevnodes++;
3331 	if (vp->v_dbatchcpu != NOCPU) {
3332 		VI_UNLOCK(vp);
3333 		critical_exit();
3334 		return;
3335 	}
3336 
3337 	sched_pin();
3338 	critical_exit();
3339 	mtx_lock(&vd->lock);
3340 	MPASS(vd->index < VDBATCH_SIZE);
3341 	MPASS(vd->tab[vd->index] == NULL);
3342 	/*
3343 	 * A hack: we depend on being pinned so that we know what to put in
3344 	 * ->v_dbatchcpu.
3345 	 */
3346 	vp->v_dbatchcpu = curcpu;
3347 	vd->tab[vd->index] = vp;
3348 	vd->index++;
3349 	VI_UNLOCK(vp);
3350 	if (vd->index == VDBATCH_SIZE)
3351 		vdbatch_process(vd);
3352 	mtx_unlock(&vd->lock);
3353 	sched_unpin();
3354 }
3355 
3356 /*
3357  * This routine must only be called for vnodes which are about to be
3358  * deallocated. Supporting dequeue for arbitrary vndoes would require
3359  * validating that the locked batch matches.
3360  */
3361 static void
3362 vdbatch_dequeue(struct vnode *vp)
3363 {
3364 	struct vdbatch *vd;
3365 	int i;
3366 	short cpu;
3367 
3368 	VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3369 	    ("%s: called for a used vnode\n", __func__));
3370 
3371 	cpu = vp->v_dbatchcpu;
3372 	if (cpu == NOCPU)
3373 		return;
3374 
3375 	vd = DPCPU_ID_PTR(cpu, vd);
3376 	mtx_lock(&vd->lock);
3377 	for (i = 0; i < vd->index; i++) {
3378 		if (vd->tab[i] != vp)
3379 			continue;
3380 		vp->v_dbatchcpu = NOCPU;
3381 		vd->index--;
3382 		vd->tab[i] = vd->tab[vd->index];
3383 		vd->tab[vd->index] = NULL;
3384 		break;
3385 	}
3386 	mtx_unlock(&vd->lock);
3387 	/*
3388 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3389 	 */
3390 	MPASS(vp->v_dbatchcpu == NOCPU);
3391 }
3392 
3393 /*
3394  * Drop the hold count of the vnode.  If this is the last reference to
3395  * the vnode we place it on the free list unless it has been vgone'd
3396  * (marked VIRF_DOOMED) in which case we will free it.
3397  *
3398  * Because the vnode vm object keeps a hold reference on the vnode if
3399  * there is at least one resident non-cached page, the vnode cannot
3400  * leave the active list without the page cleanup done.
3401  */
3402 static void
3403 vdrop_deactivate(struct vnode *vp)
3404 {
3405 	struct mount *mp;
3406 
3407 	ASSERT_VI_LOCKED(vp, __func__);
3408 	/*
3409 	 * Mark a vnode as free: remove it from its active list
3410 	 * and put it up for recycling on the freelist.
3411 	 */
3412 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3413 	    ("vdrop: returning doomed vnode"));
3414 	VNASSERT(vp->v_op != NULL, vp,
3415 	    ("vdrop: vnode already reclaimed."));
3416 	VNASSERT(vp->v_holdcnt == 0, vp,
3417 	    ("vdrop: freeing when we shouldn't"));
3418 	VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
3419 	    ("vnode with VI_OWEINACT set"));
3420 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp,
3421 	    ("vnode with VI_DEFINACT set"));
3422 	if (vp->v_mflag & VMP_LAZYLIST) {
3423 		mp = vp->v_mount;
3424 		mtx_lock(&mp->mnt_listmtx);
3425 		vp->v_mflag &= ~VMP_LAZYLIST;
3426 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3427 		mp->mnt_lazyvnodelistsize--;
3428 		mtx_unlock(&mp->mnt_listmtx);
3429 	}
3430 	vdbatch_enqueue(vp);
3431 }
3432 
3433 void
3434 vdrop(struct vnode *vp)
3435 {
3436 
3437 	ASSERT_VI_UNLOCKED(vp, __func__);
3438 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3439 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3440 		return;
3441 	VI_LOCK(vp);
3442 	vdropl(vp);
3443 }
3444 
3445 void
3446 vdropl(struct vnode *vp)
3447 {
3448 
3449 	ASSERT_VI_LOCKED(vp, __func__);
3450 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3451 	if (!refcount_release(&vp->v_holdcnt)) {
3452 		VI_UNLOCK(vp);
3453 		return;
3454 	}
3455 	if (VN_IS_DOOMED(vp)) {
3456 		freevnode(vp);
3457 		return;
3458 	}
3459 	vdrop_deactivate(vp);
3460 }
3461 
3462 /*
3463  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3464  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3465  * OWEINACT tracks whether a vnode missed a call to inactive due to a
3466  * failed lock upgrade.
3467  */
3468 void
3469 vinactive(struct vnode *vp)
3470 {
3471 	struct vm_object *obj;
3472 
3473 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3474 	ASSERT_VI_LOCKED(vp, "vinactive");
3475 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3476 	    ("vinactive: recursed on VI_DOINGINACT"));
3477 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3478 	vp->v_iflag |= VI_DOINGINACT;
3479 	vp->v_iflag &= ~VI_OWEINACT;
3480 	VI_UNLOCK(vp);
3481 	/*
3482 	 * Before moving off the active list, we must be sure that any
3483 	 * modified pages are converted into the vnode's dirty
3484 	 * buffers, since these will no longer be checked once the
3485 	 * vnode is on the inactive list.
3486 	 *
3487 	 * The write-out of the dirty pages is asynchronous.  At the
3488 	 * point that VOP_INACTIVE() is called, there could still be
3489 	 * pending I/O and dirty pages in the object.
3490 	 */
3491 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3492 	    vm_object_mightbedirty(obj)) {
3493 		VM_OBJECT_WLOCK(obj);
3494 		vm_object_page_clean(obj, 0, 0, 0);
3495 		VM_OBJECT_WUNLOCK(obj);
3496 	}
3497 	VOP_INACTIVE(vp, curthread);
3498 	VI_LOCK(vp);
3499 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3500 	    ("vinactive: lost VI_DOINGINACT"));
3501 	vp->v_iflag &= ~VI_DOINGINACT;
3502 }
3503 
3504 /*
3505  * Remove any vnodes in the vnode table belonging to mount point mp.
3506  *
3507  * If FORCECLOSE is not specified, there should not be any active ones,
3508  * return error if any are found (nb: this is a user error, not a
3509  * system error). If FORCECLOSE is specified, detach any active vnodes
3510  * that are found.
3511  *
3512  * If WRITECLOSE is set, only flush out regular file vnodes open for
3513  * writing.
3514  *
3515  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3516  *
3517  * `rootrefs' specifies the base reference count for the root vnode
3518  * of this filesystem. The root vnode is considered busy if its
3519  * v_usecount exceeds this value. On a successful return, vflush(, td)
3520  * will call vrele() on the root vnode exactly rootrefs times.
3521  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3522  * be zero.
3523  */
3524 #ifdef DIAGNOSTIC
3525 static int busyprt = 0;		/* print out busy vnodes */
3526 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3527 #endif
3528 
3529 int
3530 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3531 {
3532 	struct vnode *vp, *mvp, *rootvp = NULL;
3533 	struct vattr vattr;
3534 	int busy = 0, error;
3535 
3536 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3537 	    rootrefs, flags);
3538 	if (rootrefs > 0) {
3539 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3540 		    ("vflush: bad args"));
3541 		/*
3542 		 * Get the filesystem root vnode. We can vput() it
3543 		 * immediately, since with rootrefs > 0, it won't go away.
3544 		 */
3545 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3546 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3547 			    __func__, error);
3548 			return (error);
3549 		}
3550 		vput(rootvp);
3551 	}
3552 loop:
3553 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3554 		vholdl(vp);
3555 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3556 		if (error) {
3557 			vdrop(vp);
3558 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3559 			goto loop;
3560 		}
3561 		/*
3562 		 * Skip over a vnodes marked VV_SYSTEM.
3563 		 */
3564 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3565 			VOP_UNLOCK(vp);
3566 			vdrop(vp);
3567 			continue;
3568 		}
3569 		/*
3570 		 * If WRITECLOSE is set, flush out unlinked but still open
3571 		 * files (even if open only for reading) and regular file
3572 		 * vnodes open for writing.
3573 		 */
3574 		if (flags & WRITECLOSE) {
3575 			if (vp->v_object != NULL) {
3576 				VM_OBJECT_WLOCK(vp->v_object);
3577 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3578 				VM_OBJECT_WUNLOCK(vp->v_object);
3579 			}
3580 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3581 			if (error != 0) {
3582 				VOP_UNLOCK(vp);
3583 				vdrop(vp);
3584 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3585 				return (error);
3586 			}
3587 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3588 			VI_LOCK(vp);
3589 
3590 			if ((vp->v_type == VNON ||
3591 			    (error == 0 && vattr.va_nlink > 0)) &&
3592 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3593 				VOP_UNLOCK(vp);
3594 				vdropl(vp);
3595 				continue;
3596 			}
3597 		} else
3598 			VI_LOCK(vp);
3599 		/*
3600 		 * With v_usecount == 0, all we need to do is clear out the
3601 		 * vnode data structures and we are done.
3602 		 *
3603 		 * If FORCECLOSE is set, forcibly close the vnode.
3604 		 */
3605 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3606 			vgonel(vp);
3607 		} else {
3608 			busy++;
3609 #ifdef DIAGNOSTIC
3610 			if (busyprt)
3611 				vn_printf(vp, "vflush: busy vnode ");
3612 #endif
3613 		}
3614 		VOP_UNLOCK(vp);
3615 		vdropl(vp);
3616 	}
3617 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3618 		/*
3619 		 * If just the root vnode is busy, and if its refcount
3620 		 * is equal to `rootrefs', then go ahead and kill it.
3621 		 */
3622 		VI_LOCK(rootvp);
3623 		KASSERT(busy > 0, ("vflush: not busy"));
3624 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3625 		    ("vflush: usecount %d < rootrefs %d",
3626 		     rootvp->v_usecount, rootrefs));
3627 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3628 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3629 			vgone(rootvp);
3630 			VOP_UNLOCK(rootvp);
3631 			busy = 0;
3632 		} else
3633 			VI_UNLOCK(rootvp);
3634 	}
3635 	if (busy) {
3636 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3637 		    busy);
3638 		return (EBUSY);
3639 	}
3640 	for (; rootrefs > 0; rootrefs--)
3641 		vrele(rootvp);
3642 	return (0);
3643 }
3644 
3645 /*
3646  * Recycle an unused vnode to the front of the free list.
3647  */
3648 int
3649 vrecycle(struct vnode *vp)
3650 {
3651 	int recycled;
3652 
3653 	VI_LOCK(vp);
3654 	recycled = vrecyclel(vp);
3655 	VI_UNLOCK(vp);
3656 	return (recycled);
3657 }
3658 
3659 /*
3660  * vrecycle, with the vp interlock held.
3661  */
3662 int
3663 vrecyclel(struct vnode *vp)
3664 {
3665 	int recycled;
3666 
3667 	ASSERT_VOP_ELOCKED(vp, __func__);
3668 	ASSERT_VI_LOCKED(vp, __func__);
3669 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3670 	recycled = 0;
3671 	if (vp->v_usecount == 0) {
3672 		recycled = 1;
3673 		vgonel(vp);
3674 	}
3675 	return (recycled);
3676 }
3677 
3678 /*
3679  * Eliminate all activity associated with a vnode
3680  * in preparation for reuse.
3681  */
3682 void
3683 vgone(struct vnode *vp)
3684 {
3685 	VI_LOCK(vp);
3686 	vgonel(vp);
3687 	VI_UNLOCK(vp);
3688 }
3689 
3690 static void
3691 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3692     struct vnode *lowervp __unused)
3693 {
3694 }
3695 
3696 /*
3697  * Notify upper mounts about reclaimed or unlinked vnode.
3698  */
3699 void
3700 vfs_notify_upper(struct vnode *vp, int event)
3701 {
3702 	static struct vfsops vgonel_vfsops = {
3703 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3704 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3705 	};
3706 	struct mount *mp, *ump, *mmp;
3707 
3708 	mp = vp->v_mount;
3709 	if (mp == NULL)
3710 		return;
3711 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3712 		return;
3713 
3714 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3715 	mmp->mnt_op = &vgonel_vfsops;
3716 	mmp->mnt_kern_flag |= MNTK_MARKER;
3717 	MNT_ILOCK(mp);
3718 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3719 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3720 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3721 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3722 			continue;
3723 		}
3724 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3725 		MNT_IUNLOCK(mp);
3726 		switch (event) {
3727 		case VFS_NOTIFY_UPPER_RECLAIM:
3728 			VFS_RECLAIM_LOWERVP(ump, vp);
3729 			break;
3730 		case VFS_NOTIFY_UPPER_UNLINK:
3731 			VFS_UNLINK_LOWERVP(ump, vp);
3732 			break;
3733 		default:
3734 			KASSERT(0, ("invalid event %d", event));
3735 			break;
3736 		}
3737 		MNT_ILOCK(mp);
3738 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3739 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3740 	}
3741 	free(mmp, M_TEMP);
3742 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3743 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3744 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3745 		wakeup(&mp->mnt_uppers);
3746 	}
3747 	MNT_IUNLOCK(mp);
3748 }
3749 
3750 /*
3751  * vgone, with the vp interlock held.
3752  */
3753 static void
3754 vgonel(struct vnode *vp)
3755 {
3756 	struct thread *td;
3757 	struct mount *mp;
3758 	vm_object_t object;
3759 	bool active, oweinact;
3760 
3761 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3762 	ASSERT_VI_LOCKED(vp, "vgonel");
3763 	VNASSERT(vp->v_holdcnt, vp,
3764 	    ("vgonel: vp %p has no reference.", vp));
3765 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3766 	td = curthread;
3767 
3768 	/*
3769 	 * Don't vgonel if we're already doomed.
3770 	 */
3771 	if (vp->v_irflag & VIRF_DOOMED)
3772 		return;
3773 	vp->v_irflag |= VIRF_DOOMED;
3774 
3775 	/*
3776 	 * Check to see if the vnode is in use.  If so, we have to call
3777 	 * VOP_CLOSE() and VOP_INACTIVE().
3778 	 */
3779 	active = vp->v_usecount > 0;
3780 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3781 	/*
3782 	 * If we need to do inactive VI_OWEINACT will be set.
3783 	 */
3784 	if (vp->v_iflag & VI_DEFINACT) {
3785 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3786 		vp->v_iflag &= ~VI_DEFINACT;
3787 		vdropl(vp);
3788 	} else {
3789 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
3790 		VI_UNLOCK(vp);
3791 	}
3792 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3793 
3794 	/*
3795 	 * If purging an active vnode, it must be closed and
3796 	 * deactivated before being reclaimed.
3797 	 */
3798 	if (active)
3799 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3800 	if (oweinact || active) {
3801 		VI_LOCK(vp);
3802 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
3803 			vinactive(vp);
3804 		VI_UNLOCK(vp);
3805 	}
3806 	if (vp->v_type == VSOCK)
3807 		vfs_unp_reclaim(vp);
3808 
3809 	/*
3810 	 * Clean out any buffers associated with the vnode.
3811 	 * If the flush fails, just toss the buffers.
3812 	 */
3813 	mp = NULL;
3814 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3815 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3816 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3817 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3818 			;
3819 	}
3820 
3821 	BO_LOCK(&vp->v_bufobj);
3822 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3823 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3824 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3825 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3826 	    ("vp %p bufobj not invalidated", vp));
3827 
3828 	/*
3829 	 * For VMIO bufobj, BO_DEAD is set later, or in
3830 	 * vm_object_terminate() after the object's page queue is
3831 	 * flushed.
3832 	 */
3833 	object = vp->v_bufobj.bo_object;
3834 	if (object == NULL)
3835 		vp->v_bufobj.bo_flag |= BO_DEAD;
3836 	BO_UNLOCK(&vp->v_bufobj);
3837 
3838 	/*
3839 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
3840 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
3841 	 * should not touch the object borrowed from the lower vnode
3842 	 * (the handle check).
3843 	 */
3844 	if (object != NULL && object->type == OBJT_VNODE &&
3845 	    object->handle == vp)
3846 		vnode_destroy_vobject(vp);
3847 
3848 	/*
3849 	 * Reclaim the vnode.
3850 	 */
3851 	if (VOP_RECLAIM(vp, td))
3852 		panic("vgone: cannot reclaim");
3853 	if (mp != NULL)
3854 		vn_finished_secondary_write(mp);
3855 	VNASSERT(vp->v_object == NULL, vp,
3856 	    ("vop_reclaim left v_object vp=%p", vp));
3857 	/*
3858 	 * Clear the advisory locks and wake up waiting threads.
3859 	 */
3860 	(void)VOP_ADVLOCKPURGE(vp);
3861 	vp->v_lockf = NULL;
3862 	/*
3863 	 * Delete from old mount point vnode list.
3864 	 */
3865 	delmntque(vp);
3866 	cache_purge(vp);
3867 	/*
3868 	 * Done with purge, reset to the standard lock and invalidate
3869 	 * the vnode.
3870 	 */
3871 	VI_LOCK(vp);
3872 	vp->v_vnlock = &vp->v_lock;
3873 	vp->v_op = &dead_vnodeops;
3874 	vp->v_type = VBAD;
3875 }
3876 
3877 /*
3878  * Calculate the total number of references to a special device.
3879  */
3880 int
3881 vcount(struct vnode *vp)
3882 {
3883 	int count;
3884 
3885 	dev_lock();
3886 	count = vp->v_rdev->si_usecount;
3887 	dev_unlock();
3888 	return (count);
3889 }
3890 
3891 /*
3892  * Print out a description of a vnode.
3893  */
3894 static char *typename[] =
3895 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3896  "VMARKER"};
3897 
3898 void
3899 vn_printf(struct vnode *vp, const char *fmt, ...)
3900 {
3901 	va_list ap;
3902 	char buf[256], buf2[16];
3903 	u_long flags;
3904 
3905 	va_start(ap, fmt);
3906 	vprintf(fmt, ap);
3907 	va_end(ap);
3908 	printf("%p: ", (void *)vp);
3909 	printf("type %s\n", typename[vp->v_type]);
3910 	printf("    usecount %d, writecount %d, refcount %d",
3911 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
3912 	switch (vp->v_type) {
3913 	case VDIR:
3914 		printf(" mountedhere %p\n", vp->v_mountedhere);
3915 		break;
3916 	case VCHR:
3917 		printf(" rdev %p\n", vp->v_rdev);
3918 		break;
3919 	case VSOCK:
3920 		printf(" socket %p\n", vp->v_unpcb);
3921 		break;
3922 	case VFIFO:
3923 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
3924 		break;
3925 	default:
3926 		printf("\n");
3927 		break;
3928 	}
3929 	buf[0] = '\0';
3930 	buf[1] = '\0';
3931 	if (vp->v_irflag & VIRF_DOOMED)
3932 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
3933 	flags = vp->v_irflag & ~(VIRF_DOOMED);
3934 	if (flags != 0) {
3935 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
3936 		strlcat(buf, buf2, sizeof(buf));
3937 	}
3938 	if (vp->v_vflag & VV_ROOT)
3939 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3940 	if (vp->v_vflag & VV_ISTTY)
3941 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3942 	if (vp->v_vflag & VV_NOSYNC)
3943 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3944 	if (vp->v_vflag & VV_ETERNALDEV)
3945 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3946 	if (vp->v_vflag & VV_CACHEDLABEL)
3947 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3948 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
3949 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
3950 	if (vp->v_vflag & VV_COPYONWRITE)
3951 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3952 	if (vp->v_vflag & VV_SYSTEM)
3953 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3954 	if (vp->v_vflag & VV_PROCDEP)
3955 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3956 	if (vp->v_vflag & VV_NOKNOTE)
3957 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3958 	if (vp->v_vflag & VV_DELETED)
3959 		strlcat(buf, "|VV_DELETED", sizeof(buf));
3960 	if (vp->v_vflag & VV_MD)
3961 		strlcat(buf, "|VV_MD", sizeof(buf));
3962 	if (vp->v_vflag & VV_FORCEINSMQ)
3963 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
3964 	if (vp->v_vflag & VV_READLINK)
3965 		strlcat(buf, "|VV_READLINK", sizeof(buf));
3966 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
3967 	    VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
3968 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
3969 	if (flags != 0) {
3970 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
3971 		strlcat(buf, buf2, sizeof(buf));
3972 	}
3973 	if (vp->v_iflag & VI_TEXT_REF)
3974 		strlcat(buf, "|VI_TEXT_REF", sizeof(buf));
3975 	if (vp->v_iflag & VI_MOUNT)
3976 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
3977 	if (vp->v_iflag & VI_DOINGINACT)
3978 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
3979 	if (vp->v_iflag & VI_OWEINACT)
3980 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
3981 	if (vp->v_iflag & VI_DEFINACT)
3982 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
3983 	flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_DOINGINACT |
3984 	    VI_OWEINACT | VI_DEFINACT);
3985 	if (flags != 0) {
3986 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
3987 		strlcat(buf, buf2, sizeof(buf));
3988 	}
3989 	if (vp->v_mflag & VMP_LAZYLIST)
3990 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
3991 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
3992 	if (flags != 0) {
3993 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
3994 		strlcat(buf, buf2, sizeof(buf));
3995 	}
3996 	printf("    flags (%s)\n", buf + 1);
3997 	if (mtx_owned(VI_MTX(vp)))
3998 		printf(" VI_LOCKed");
3999 	if (vp->v_object != NULL)
4000 		printf("    v_object %p ref %d pages %d "
4001 		    "cleanbuf %d dirtybuf %d\n",
4002 		    vp->v_object, vp->v_object->ref_count,
4003 		    vp->v_object->resident_page_count,
4004 		    vp->v_bufobj.bo_clean.bv_cnt,
4005 		    vp->v_bufobj.bo_dirty.bv_cnt);
4006 	printf("    ");
4007 	lockmgr_printinfo(vp->v_vnlock);
4008 	if (vp->v_data != NULL)
4009 		VOP_PRINT(vp);
4010 }
4011 
4012 #ifdef DDB
4013 /*
4014  * List all of the locked vnodes in the system.
4015  * Called when debugging the kernel.
4016  */
4017 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
4018 {
4019 	struct mount *mp;
4020 	struct vnode *vp;
4021 
4022 	/*
4023 	 * Note: because this is DDB, we can't obey the locking semantics
4024 	 * for these structures, which means we could catch an inconsistent
4025 	 * state and dereference a nasty pointer.  Not much to be done
4026 	 * about that.
4027 	 */
4028 	db_printf("Locked vnodes\n");
4029 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4030 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4031 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4032 				vn_printf(vp, "vnode ");
4033 		}
4034 	}
4035 }
4036 
4037 /*
4038  * Show details about the given vnode.
4039  */
4040 DB_SHOW_COMMAND(vnode, db_show_vnode)
4041 {
4042 	struct vnode *vp;
4043 
4044 	if (!have_addr)
4045 		return;
4046 	vp = (struct vnode *)addr;
4047 	vn_printf(vp, "vnode ");
4048 }
4049 
4050 /*
4051  * Show details about the given mount point.
4052  */
4053 DB_SHOW_COMMAND(mount, db_show_mount)
4054 {
4055 	struct mount *mp;
4056 	struct vfsopt *opt;
4057 	struct statfs *sp;
4058 	struct vnode *vp;
4059 	char buf[512];
4060 	uint64_t mflags;
4061 	u_int flags;
4062 
4063 	if (!have_addr) {
4064 		/* No address given, print short info about all mount points. */
4065 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4066 			db_printf("%p %s on %s (%s)\n", mp,
4067 			    mp->mnt_stat.f_mntfromname,
4068 			    mp->mnt_stat.f_mntonname,
4069 			    mp->mnt_stat.f_fstypename);
4070 			if (db_pager_quit)
4071 				break;
4072 		}
4073 		db_printf("\nMore info: show mount <addr>\n");
4074 		return;
4075 	}
4076 
4077 	mp = (struct mount *)addr;
4078 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4079 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4080 
4081 	buf[0] = '\0';
4082 	mflags = mp->mnt_flag;
4083 #define	MNT_FLAG(flag)	do {						\
4084 	if (mflags & (flag)) {						\
4085 		if (buf[0] != '\0')					\
4086 			strlcat(buf, ", ", sizeof(buf));		\
4087 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4088 		mflags &= ~(flag);					\
4089 	}								\
4090 } while (0)
4091 	MNT_FLAG(MNT_RDONLY);
4092 	MNT_FLAG(MNT_SYNCHRONOUS);
4093 	MNT_FLAG(MNT_NOEXEC);
4094 	MNT_FLAG(MNT_NOSUID);
4095 	MNT_FLAG(MNT_NFS4ACLS);
4096 	MNT_FLAG(MNT_UNION);
4097 	MNT_FLAG(MNT_ASYNC);
4098 	MNT_FLAG(MNT_SUIDDIR);
4099 	MNT_FLAG(MNT_SOFTDEP);
4100 	MNT_FLAG(MNT_NOSYMFOLLOW);
4101 	MNT_FLAG(MNT_GJOURNAL);
4102 	MNT_FLAG(MNT_MULTILABEL);
4103 	MNT_FLAG(MNT_ACLS);
4104 	MNT_FLAG(MNT_NOATIME);
4105 	MNT_FLAG(MNT_NOCLUSTERR);
4106 	MNT_FLAG(MNT_NOCLUSTERW);
4107 	MNT_FLAG(MNT_SUJ);
4108 	MNT_FLAG(MNT_EXRDONLY);
4109 	MNT_FLAG(MNT_EXPORTED);
4110 	MNT_FLAG(MNT_DEFEXPORTED);
4111 	MNT_FLAG(MNT_EXPORTANON);
4112 	MNT_FLAG(MNT_EXKERB);
4113 	MNT_FLAG(MNT_EXPUBLIC);
4114 	MNT_FLAG(MNT_LOCAL);
4115 	MNT_FLAG(MNT_QUOTA);
4116 	MNT_FLAG(MNT_ROOTFS);
4117 	MNT_FLAG(MNT_USER);
4118 	MNT_FLAG(MNT_IGNORE);
4119 	MNT_FLAG(MNT_UPDATE);
4120 	MNT_FLAG(MNT_DELEXPORT);
4121 	MNT_FLAG(MNT_RELOAD);
4122 	MNT_FLAG(MNT_FORCE);
4123 	MNT_FLAG(MNT_SNAPSHOT);
4124 	MNT_FLAG(MNT_BYFSID);
4125 #undef MNT_FLAG
4126 	if (mflags != 0) {
4127 		if (buf[0] != '\0')
4128 			strlcat(buf, ", ", sizeof(buf));
4129 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4130 		    "0x%016jx", mflags);
4131 	}
4132 	db_printf("    mnt_flag = %s\n", buf);
4133 
4134 	buf[0] = '\0';
4135 	flags = mp->mnt_kern_flag;
4136 #define	MNT_KERN_FLAG(flag)	do {					\
4137 	if (flags & (flag)) {						\
4138 		if (buf[0] != '\0')					\
4139 			strlcat(buf, ", ", sizeof(buf));		\
4140 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4141 		flags &= ~(flag);					\
4142 	}								\
4143 } while (0)
4144 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4145 	MNT_KERN_FLAG(MNTK_ASYNC);
4146 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4147 	MNT_KERN_FLAG(MNTK_DRAINING);
4148 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4149 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4150 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4151 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4152 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
4153 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
4154 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
4155 	MNT_KERN_FLAG(MNTK_MARKER);
4156 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4157 	MNT_KERN_FLAG(MNTK_NOASYNC);
4158 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4159 	MNT_KERN_FLAG(MNTK_MWAIT);
4160 	MNT_KERN_FLAG(MNTK_SUSPEND);
4161 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4162 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4163 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4164 	MNT_KERN_FLAG(MNTK_NOKNOTE);
4165 #undef MNT_KERN_FLAG
4166 	if (flags != 0) {
4167 		if (buf[0] != '\0')
4168 			strlcat(buf, ", ", sizeof(buf));
4169 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4170 		    "0x%08x", flags);
4171 	}
4172 	db_printf("    mnt_kern_flag = %s\n", buf);
4173 
4174 	db_printf("    mnt_opt = ");
4175 	opt = TAILQ_FIRST(mp->mnt_opt);
4176 	if (opt != NULL) {
4177 		db_printf("%s", opt->name);
4178 		opt = TAILQ_NEXT(opt, link);
4179 		while (opt != NULL) {
4180 			db_printf(", %s", opt->name);
4181 			opt = TAILQ_NEXT(opt, link);
4182 		}
4183 	}
4184 	db_printf("\n");
4185 
4186 	sp = &mp->mnt_stat;
4187 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4188 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4189 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4190 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4191 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4192 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4193 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4194 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4195 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4196 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4197 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4198 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4199 
4200 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4201 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4202 	if (jailed(mp->mnt_cred))
4203 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4204 	db_printf(" }\n");
4205 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4206 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4207 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4208 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4209 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4210 	    mp->mnt_lazyvnodelistsize);
4211 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4212 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4213 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
4214 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4215 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4216 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4217 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4218 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4219 	db_printf("    mnt_secondary_accwrites = %d\n",
4220 	    mp->mnt_secondary_accwrites);
4221 	db_printf("    mnt_gjprovider = %s\n",
4222 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4223 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4224 
4225 	db_printf("\n\nList of active vnodes\n");
4226 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4227 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4228 			vn_printf(vp, "vnode ");
4229 			if (db_pager_quit)
4230 				break;
4231 		}
4232 	}
4233 	db_printf("\n\nList of inactive vnodes\n");
4234 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4235 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4236 			vn_printf(vp, "vnode ");
4237 			if (db_pager_quit)
4238 				break;
4239 		}
4240 	}
4241 }
4242 #endif	/* DDB */
4243 
4244 /*
4245  * Fill in a struct xvfsconf based on a struct vfsconf.
4246  */
4247 static int
4248 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4249 {
4250 	struct xvfsconf xvfsp;
4251 
4252 	bzero(&xvfsp, sizeof(xvfsp));
4253 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4254 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4255 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4256 	xvfsp.vfc_flags = vfsp->vfc_flags;
4257 	/*
4258 	 * These are unused in userland, we keep them
4259 	 * to not break binary compatibility.
4260 	 */
4261 	xvfsp.vfc_vfsops = NULL;
4262 	xvfsp.vfc_next = NULL;
4263 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4264 }
4265 
4266 #ifdef COMPAT_FREEBSD32
4267 struct xvfsconf32 {
4268 	uint32_t	vfc_vfsops;
4269 	char		vfc_name[MFSNAMELEN];
4270 	int32_t		vfc_typenum;
4271 	int32_t		vfc_refcount;
4272 	int32_t		vfc_flags;
4273 	uint32_t	vfc_next;
4274 };
4275 
4276 static int
4277 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4278 {
4279 	struct xvfsconf32 xvfsp;
4280 
4281 	bzero(&xvfsp, sizeof(xvfsp));
4282 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4283 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4284 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4285 	xvfsp.vfc_flags = vfsp->vfc_flags;
4286 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4287 }
4288 #endif
4289 
4290 /*
4291  * Top level filesystem related information gathering.
4292  */
4293 static int
4294 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4295 {
4296 	struct vfsconf *vfsp;
4297 	int error;
4298 
4299 	error = 0;
4300 	vfsconf_slock();
4301 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4302 #ifdef COMPAT_FREEBSD32
4303 		if (req->flags & SCTL_MASK32)
4304 			error = vfsconf2x32(req, vfsp);
4305 		else
4306 #endif
4307 			error = vfsconf2x(req, vfsp);
4308 		if (error)
4309 			break;
4310 	}
4311 	vfsconf_sunlock();
4312 	return (error);
4313 }
4314 
4315 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4316     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4317     "S,xvfsconf", "List of all configured filesystems");
4318 
4319 #ifndef BURN_BRIDGES
4320 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4321 
4322 static int
4323 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4324 {
4325 	int *name = (int *)arg1 - 1;	/* XXX */
4326 	u_int namelen = arg2 + 1;	/* XXX */
4327 	struct vfsconf *vfsp;
4328 
4329 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4330 	    "please rebuild world\n");
4331 
4332 #if 1 || defined(COMPAT_PRELITE2)
4333 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4334 	if (namelen == 1)
4335 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4336 #endif
4337 
4338 	switch (name[1]) {
4339 	case VFS_MAXTYPENUM:
4340 		if (namelen != 2)
4341 			return (ENOTDIR);
4342 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4343 	case VFS_CONF:
4344 		if (namelen != 3)
4345 			return (ENOTDIR);	/* overloaded */
4346 		vfsconf_slock();
4347 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4348 			if (vfsp->vfc_typenum == name[2])
4349 				break;
4350 		}
4351 		vfsconf_sunlock();
4352 		if (vfsp == NULL)
4353 			return (EOPNOTSUPP);
4354 #ifdef COMPAT_FREEBSD32
4355 		if (req->flags & SCTL_MASK32)
4356 			return (vfsconf2x32(req, vfsp));
4357 		else
4358 #endif
4359 			return (vfsconf2x(req, vfsp));
4360 	}
4361 	return (EOPNOTSUPP);
4362 }
4363 
4364 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4365     CTLFLAG_MPSAFE, vfs_sysctl,
4366     "Generic filesystem");
4367 
4368 #if 1 || defined(COMPAT_PRELITE2)
4369 
4370 static int
4371 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4372 {
4373 	int error;
4374 	struct vfsconf *vfsp;
4375 	struct ovfsconf ovfs;
4376 
4377 	vfsconf_slock();
4378 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4379 		bzero(&ovfs, sizeof(ovfs));
4380 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4381 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4382 		ovfs.vfc_index = vfsp->vfc_typenum;
4383 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4384 		ovfs.vfc_flags = vfsp->vfc_flags;
4385 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4386 		if (error != 0) {
4387 			vfsconf_sunlock();
4388 			return (error);
4389 		}
4390 	}
4391 	vfsconf_sunlock();
4392 	return (0);
4393 }
4394 
4395 #endif /* 1 || COMPAT_PRELITE2 */
4396 #endif /* !BURN_BRIDGES */
4397 
4398 #define KINFO_VNODESLOP		10
4399 #ifdef notyet
4400 /*
4401  * Dump vnode list (via sysctl).
4402  */
4403 /* ARGSUSED */
4404 static int
4405 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4406 {
4407 	struct xvnode *xvn;
4408 	struct mount *mp;
4409 	struct vnode *vp;
4410 	int error, len, n;
4411 
4412 	/*
4413 	 * Stale numvnodes access is not fatal here.
4414 	 */
4415 	req->lock = 0;
4416 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4417 	if (!req->oldptr)
4418 		/* Make an estimate */
4419 		return (SYSCTL_OUT(req, 0, len));
4420 
4421 	error = sysctl_wire_old_buffer(req, 0);
4422 	if (error != 0)
4423 		return (error);
4424 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4425 	n = 0;
4426 	mtx_lock(&mountlist_mtx);
4427 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4428 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4429 			continue;
4430 		MNT_ILOCK(mp);
4431 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4432 			if (n == len)
4433 				break;
4434 			vref(vp);
4435 			xvn[n].xv_size = sizeof *xvn;
4436 			xvn[n].xv_vnode = vp;
4437 			xvn[n].xv_id = 0;	/* XXX compat */
4438 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4439 			XV_COPY(usecount);
4440 			XV_COPY(writecount);
4441 			XV_COPY(holdcnt);
4442 			XV_COPY(mount);
4443 			XV_COPY(numoutput);
4444 			XV_COPY(type);
4445 #undef XV_COPY
4446 			xvn[n].xv_flag = vp->v_vflag;
4447 
4448 			switch (vp->v_type) {
4449 			case VREG:
4450 			case VDIR:
4451 			case VLNK:
4452 				break;
4453 			case VBLK:
4454 			case VCHR:
4455 				if (vp->v_rdev == NULL) {
4456 					vrele(vp);
4457 					continue;
4458 				}
4459 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4460 				break;
4461 			case VSOCK:
4462 				xvn[n].xv_socket = vp->v_socket;
4463 				break;
4464 			case VFIFO:
4465 				xvn[n].xv_fifo = vp->v_fifoinfo;
4466 				break;
4467 			case VNON:
4468 			case VBAD:
4469 			default:
4470 				/* shouldn't happen? */
4471 				vrele(vp);
4472 				continue;
4473 			}
4474 			vrele(vp);
4475 			++n;
4476 		}
4477 		MNT_IUNLOCK(mp);
4478 		mtx_lock(&mountlist_mtx);
4479 		vfs_unbusy(mp);
4480 		if (n == len)
4481 			break;
4482 	}
4483 	mtx_unlock(&mountlist_mtx);
4484 
4485 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4486 	free(xvn, M_TEMP);
4487 	return (error);
4488 }
4489 
4490 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4491     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4492     "");
4493 #endif
4494 
4495 static void
4496 unmount_or_warn(struct mount *mp)
4497 {
4498 	int error;
4499 
4500 	error = dounmount(mp, MNT_FORCE, curthread);
4501 	if (error != 0) {
4502 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4503 		if (error == EBUSY)
4504 			printf("BUSY)\n");
4505 		else
4506 			printf("%d)\n", error);
4507 	}
4508 }
4509 
4510 /*
4511  * Unmount all filesystems. The list is traversed in reverse order
4512  * of mounting to avoid dependencies.
4513  */
4514 void
4515 vfs_unmountall(void)
4516 {
4517 	struct mount *mp, *tmp;
4518 
4519 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4520 
4521 	/*
4522 	 * Since this only runs when rebooting, it is not interlocked.
4523 	 */
4524 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4525 		vfs_ref(mp);
4526 
4527 		/*
4528 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4529 		 * unmount of the latter.
4530 		 */
4531 		if (mp == rootdevmp)
4532 			continue;
4533 
4534 		unmount_or_warn(mp);
4535 	}
4536 
4537 	if (rootdevmp != NULL)
4538 		unmount_or_warn(rootdevmp);
4539 }
4540 
4541 static void
4542 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4543 {
4544 
4545 	ASSERT_VI_LOCKED(vp, __func__);
4546 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4547 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
4548 		vdropl(vp);
4549 		return;
4550 	}
4551 	if (vn_lock(vp, lkflags) == 0) {
4552 		VI_LOCK(vp);
4553 		if ((vp->v_iflag & (VI_OWEINACT | VI_DOINGINACT)) == VI_OWEINACT)
4554 			vinactive(vp);
4555 		VOP_UNLOCK(vp);
4556 		vdropl(vp);
4557 		return;
4558 	}
4559 	vdefer_inactive_cond(vp);
4560 }
4561 
4562 static int
4563 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4564 {
4565 
4566 	return (vp->v_iflag & VI_DEFINACT);
4567 }
4568 
4569 static void __noinline
4570 vfs_periodic_inactive(struct mount *mp, int flags)
4571 {
4572 	struct vnode *vp, *mvp;
4573 	int lkflags;
4574 
4575 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4576 	if (flags != MNT_WAIT)
4577 		lkflags |= LK_NOWAIT;
4578 
4579 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4580 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
4581 			VI_UNLOCK(vp);
4582 			continue;
4583 		}
4584 		vp->v_iflag &= ~VI_DEFINACT;
4585 		vfs_deferred_inactive(vp, lkflags);
4586 	}
4587 }
4588 
4589 static inline bool
4590 vfs_want_msync(struct vnode *vp)
4591 {
4592 	struct vm_object *obj;
4593 
4594 	/*
4595 	 * This test may be performed without any locks held.
4596 	 * We rely on vm_object's type stability.
4597 	 */
4598 	if (vp->v_vflag & VV_NOSYNC)
4599 		return (false);
4600 	obj = vp->v_object;
4601 	return (obj != NULL && vm_object_mightbedirty(obj));
4602 }
4603 
4604 static int
4605 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4606 {
4607 
4608 	if (vp->v_vflag & VV_NOSYNC)
4609 		return (false);
4610 	if (vp->v_iflag & VI_DEFINACT)
4611 		return (true);
4612 	return (vfs_want_msync(vp));
4613 }
4614 
4615 static void __noinline
4616 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4617 {
4618 	struct vnode *vp, *mvp;
4619 	struct vm_object *obj;
4620 	struct thread *td;
4621 	int lkflags, objflags;
4622 	bool seen_defer;
4623 
4624 	td = curthread;
4625 
4626 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4627 	if (flags != MNT_WAIT) {
4628 		lkflags |= LK_NOWAIT;
4629 		objflags = OBJPC_NOSYNC;
4630 	} else {
4631 		objflags = OBJPC_SYNC;
4632 	}
4633 
4634 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4635 		seen_defer = false;
4636 		if (vp->v_iflag & VI_DEFINACT) {
4637 			vp->v_iflag &= ~VI_DEFINACT;
4638 			seen_defer = true;
4639 		}
4640 		if (!vfs_want_msync(vp)) {
4641 			if (seen_defer)
4642 				vfs_deferred_inactive(vp, lkflags);
4643 			else
4644 				VI_UNLOCK(vp);
4645 			continue;
4646 		}
4647 		if (vget(vp, lkflags, td) == 0) {
4648 			obj = vp->v_object;
4649 			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4650 				VM_OBJECT_WLOCK(obj);
4651 				vm_object_page_clean(obj, 0, 0, objflags);
4652 				VM_OBJECT_WUNLOCK(obj);
4653 			}
4654 			vput(vp);
4655 			if (seen_defer)
4656 				vdrop(vp);
4657 		} else {
4658 			if (seen_defer)
4659 				vdefer_inactive_cond(vp);
4660 		}
4661 	}
4662 }
4663 
4664 void
4665 vfs_periodic(struct mount *mp, int flags)
4666 {
4667 
4668 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4669 
4670 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4671 		vfs_periodic_inactive(mp, flags);
4672 	else
4673 		vfs_periodic_msync_inactive(mp, flags);
4674 }
4675 
4676 static void
4677 destroy_vpollinfo_free(struct vpollinfo *vi)
4678 {
4679 
4680 	knlist_destroy(&vi->vpi_selinfo.si_note);
4681 	mtx_destroy(&vi->vpi_lock);
4682 	uma_zfree(vnodepoll_zone, vi);
4683 }
4684 
4685 static void
4686 destroy_vpollinfo(struct vpollinfo *vi)
4687 {
4688 
4689 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4690 	seldrain(&vi->vpi_selinfo);
4691 	destroy_vpollinfo_free(vi);
4692 }
4693 
4694 /*
4695  * Initialize per-vnode helper structure to hold poll-related state.
4696  */
4697 void
4698 v_addpollinfo(struct vnode *vp)
4699 {
4700 	struct vpollinfo *vi;
4701 
4702 	if (vp->v_pollinfo != NULL)
4703 		return;
4704 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4705 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4706 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4707 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4708 	VI_LOCK(vp);
4709 	if (vp->v_pollinfo != NULL) {
4710 		VI_UNLOCK(vp);
4711 		destroy_vpollinfo_free(vi);
4712 		return;
4713 	}
4714 	vp->v_pollinfo = vi;
4715 	VI_UNLOCK(vp);
4716 }
4717 
4718 /*
4719  * Record a process's interest in events which might happen to
4720  * a vnode.  Because poll uses the historic select-style interface
4721  * internally, this routine serves as both the ``check for any
4722  * pending events'' and the ``record my interest in future events''
4723  * functions.  (These are done together, while the lock is held,
4724  * to avoid race conditions.)
4725  */
4726 int
4727 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4728 {
4729 
4730 	v_addpollinfo(vp);
4731 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4732 	if (vp->v_pollinfo->vpi_revents & events) {
4733 		/*
4734 		 * This leaves events we are not interested
4735 		 * in available for the other process which
4736 		 * which presumably had requested them
4737 		 * (otherwise they would never have been
4738 		 * recorded).
4739 		 */
4740 		events &= vp->v_pollinfo->vpi_revents;
4741 		vp->v_pollinfo->vpi_revents &= ~events;
4742 
4743 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4744 		return (events);
4745 	}
4746 	vp->v_pollinfo->vpi_events |= events;
4747 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4748 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4749 	return (0);
4750 }
4751 
4752 /*
4753  * Routine to create and manage a filesystem syncer vnode.
4754  */
4755 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4756 static int	sync_fsync(struct  vop_fsync_args *);
4757 static int	sync_inactive(struct  vop_inactive_args *);
4758 static int	sync_reclaim(struct  vop_reclaim_args *);
4759 
4760 static struct vop_vector sync_vnodeops = {
4761 	.vop_bypass =	VOP_EOPNOTSUPP,
4762 	.vop_close =	sync_close,		/* close */
4763 	.vop_fsync =	sync_fsync,		/* fsync */
4764 	.vop_inactive =	sync_inactive,	/* inactive */
4765 	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4766 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4767 	.vop_lock1 =	vop_stdlock,	/* lock */
4768 	.vop_unlock =	vop_stdunlock,	/* unlock */
4769 	.vop_islocked =	vop_stdislocked,	/* islocked */
4770 };
4771 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4772 
4773 /*
4774  * Create a new filesystem syncer vnode for the specified mount point.
4775  */
4776 void
4777 vfs_allocate_syncvnode(struct mount *mp)
4778 {
4779 	struct vnode *vp;
4780 	struct bufobj *bo;
4781 	static long start, incr, next;
4782 	int error;
4783 
4784 	/* Allocate a new vnode */
4785 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4786 	if (error != 0)
4787 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4788 	vp->v_type = VNON;
4789 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4790 	vp->v_vflag |= VV_FORCEINSMQ;
4791 	error = insmntque(vp, mp);
4792 	if (error != 0)
4793 		panic("vfs_allocate_syncvnode: insmntque() failed");
4794 	vp->v_vflag &= ~VV_FORCEINSMQ;
4795 	VOP_UNLOCK(vp);
4796 	/*
4797 	 * Place the vnode onto the syncer worklist. We attempt to
4798 	 * scatter them about on the list so that they will go off
4799 	 * at evenly distributed times even if all the filesystems
4800 	 * are mounted at once.
4801 	 */
4802 	next += incr;
4803 	if (next == 0 || next > syncer_maxdelay) {
4804 		start /= 2;
4805 		incr /= 2;
4806 		if (start == 0) {
4807 			start = syncer_maxdelay / 2;
4808 			incr = syncer_maxdelay;
4809 		}
4810 		next = start;
4811 	}
4812 	bo = &vp->v_bufobj;
4813 	BO_LOCK(bo);
4814 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4815 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4816 	mtx_lock(&sync_mtx);
4817 	sync_vnode_count++;
4818 	if (mp->mnt_syncer == NULL) {
4819 		mp->mnt_syncer = vp;
4820 		vp = NULL;
4821 	}
4822 	mtx_unlock(&sync_mtx);
4823 	BO_UNLOCK(bo);
4824 	if (vp != NULL) {
4825 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4826 		vgone(vp);
4827 		vput(vp);
4828 	}
4829 }
4830 
4831 void
4832 vfs_deallocate_syncvnode(struct mount *mp)
4833 {
4834 	struct vnode *vp;
4835 
4836 	mtx_lock(&sync_mtx);
4837 	vp = mp->mnt_syncer;
4838 	if (vp != NULL)
4839 		mp->mnt_syncer = NULL;
4840 	mtx_unlock(&sync_mtx);
4841 	if (vp != NULL)
4842 		vrele(vp);
4843 }
4844 
4845 /*
4846  * Do a lazy sync of the filesystem.
4847  */
4848 static int
4849 sync_fsync(struct vop_fsync_args *ap)
4850 {
4851 	struct vnode *syncvp = ap->a_vp;
4852 	struct mount *mp = syncvp->v_mount;
4853 	int error, save;
4854 	struct bufobj *bo;
4855 
4856 	/*
4857 	 * We only need to do something if this is a lazy evaluation.
4858 	 */
4859 	if (ap->a_waitfor != MNT_LAZY)
4860 		return (0);
4861 
4862 	/*
4863 	 * Move ourselves to the back of the sync list.
4864 	 */
4865 	bo = &syncvp->v_bufobj;
4866 	BO_LOCK(bo);
4867 	vn_syncer_add_to_worklist(bo, syncdelay);
4868 	BO_UNLOCK(bo);
4869 
4870 	/*
4871 	 * Walk the list of vnodes pushing all that are dirty and
4872 	 * not already on the sync list.
4873 	 */
4874 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4875 		return (0);
4876 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4877 		vfs_unbusy(mp);
4878 		return (0);
4879 	}
4880 	save = curthread_pflags_set(TDP_SYNCIO);
4881 	/*
4882 	 * The filesystem at hand may be idle with free vnodes stored in the
4883 	 * batch.  Return them instead of letting them stay there indefinitely.
4884 	 */
4885 	vfs_periodic(mp, MNT_NOWAIT);
4886 	error = VFS_SYNC(mp, MNT_LAZY);
4887 	curthread_pflags_restore(save);
4888 	vn_finished_write(mp);
4889 	vfs_unbusy(mp);
4890 	return (error);
4891 }
4892 
4893 /*
4894  * The syncer vnode is no referenced.
4895  */
4896 static int
4897 sync_inactive(struct vop_inactive_args *ap)
4898 {
4899 
4900 	vgone(ap->a_vp);
4901 	return (0);
4902 }
4903 
4904 /*
4905  * The syncer vnode is no longer needed and is being decommissioned.
4906  *
4907  * Modifications to the worklist must be protected by sync_mtx.
4908  */
4909 static int
4910 sync_reclaim(struct vop_reclaim_args *ap)
4911 {
4912 	struct vnode *vp = ap->a_vp;
4913 	struct bufobj *bo;
4914 
4915 	bo = &vp->v_bufobj;
4916 	BO_LOCK(bo);
4917 	mtx_lock(&sync_mtx);
4918 	if (vp->v_mount->mnt_syncer == vp)
4919 		vp->v_mount->mnt_syncer = NULL;
4920 	if (bo->bo_flag & BO_ONWORKLST) {
4921 		LIST_REMOVE(bo, bo_synclist);
4922 		syncer_worklist_len--;
4923 		sync_vnode_count--;
4924 		bo->bo_flag &= ~BO_ONWORKLST;
4925 	}
4926 	mtx_unlock(&sync_mtx);
4927 	BO_UNLOCK(bo);
4928 
4929 	return (0);
4930 }
4931 
4932 int
4933 vn_need_pageq_flush(struct vnode *vp)
4934 {
4935 	struct vm_object *obj;
4936 	int need;
4937 
4938 	MPASS(mtx_owned(VI_MTX(vp)));
4939 	need = 0;
4940 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
4941 	    vm_object_mightbedirty(obj))
4942 		need = 1;
4943 	return (need);
4944 }
4945 
4946 /*
4947  * Check if vnode represents a disk device
4948  */
4949 int
4950 vn_isdisk(struct vnode *vp, int *errp)
4951 {
4952 	int error;
4953 
4954 	if (vp->v_type != VCHR) {
4955 		error = ENOTBLK;
4956 		goto out;
4957 	}
4958 	error = 0;
4959 	dev_lock();
4960 	if (vp->v_rdev == NULL)
4961 		error = ENXIO;
4962 	else if (vp->v_rdev->si_devsw == NULL)
4963 		error = ENXIO;
4964 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
4965 		error = ENOTBLK;
4966 	dev_unlock();
4967 out:
4968 	if (errp != NULL)
4969 		*errp = error;
4970 	return (error == 0);
4971 }
4972 
4973 /*
4974  * Common filesystem object access control check routine.  Accepts a
4975  * vnode's type, "mode", uid and gid, requested access mode, credentials,
4976  * and optional call-by-reference privused argument allowing vaccess()
4977  * to indicate to the caller whether privilege was used to satisfy the
4978  * request (obsoleted).  Returns 0 on success, or an errno on failure.
4979  */
4980 int
4981 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
4982     accmode_t accmode, struct ucred *cred, int *privused)
4983 {
4984 	accmode_t dac_granted;
4985 	accmode_t priv_granted;
4986 
4987 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
4988 	    ("invalid bit in accmode"));
4989 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
4990 	    ("VAPPEND without VWRITE"));
4991 
4992 	/*
4993 	 * Look for a normal, non-privileged way to access the file/directory
4994 	 * as requested.  If it exists, go with that.
4995 	 */
4996 
4997 	if (privused != NULL)
4998 		*privused = 0;
4999 
5000 	dac_granted = 0;
5001 
5002 	/* Check the owner. */
5003 	if (cred->cr_uid == file_uid) {
5004 		dac_granted |= VADMIN;
5005 		if (file_mode & S_IXUSR)
5006 			dac_granted |= VEXEC;
5007 		if (file_mode & S_IRUSR)
5008 			dac_granted |= VREAD;
5009 		if (file_mode & S_IWUSR)
5010 			dac_granted |= (VWRITE | VAPPEND);
5011 
5012 		if ((accmode & dac_granted) == accmode)
5013 			return (0);
5014 
5015 		goto privcheck;
5016 	}
5017 
5018 	/* Otherwise, check the groups (first match) */
5019 	if (groupmember(file_gid, cred)) {
5020 		if (file_mode & S_IXGRP)
5021 			dac_granted |= VEXEC;
5022 		if (file_mode & S_IRGRP)
5023 			dac_granted |= VREAD;
5024 		if (file_mode & S_IWGRP)
5025 			dac_granted |= (VWRITE | VAPPEND);
5026 
5027 		if ((accmode & dac_granted) == accmode)
5028 			return (0);
5029 
5030 		goto privcheck;
5031 	}
5032 
5033 	/* Otherwise, check everyone else. */
5034 	if (file_mode & S_IXOTH)
5035 		dac_granted |= VEXEC;
5036 	if (file_mode & S_IROTH)
5037 		dac_granted |= VREAD;
5038 	if (file_mode & S_IWOTH)
5039 		dac_granted |= (VWRITE | VAPPEND);
5040 	if ((accmode & dac_granted) == accmode)
5041 		return (0);
5042 
5043 privcheck:
5044 	/*
5045 	 * Build a privilege mask to determine if the set of privileges
5046 	 * satisfies the requirements when combined with the granted mask
5047 	 * from above.  For each privilege, if the privilege is required,
5048 	 * bitwise or the request type onto the priv_granted mask.
5049 	 */
5050 	priv_granted = 0;
5051 
5052 	if (type == VDIR) {
5053 		/*
5054 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5055 		 * requests, instead of PRIV_VFS_EXEC.
5056 		 */
5057 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5058 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5059 			priv_granted |= VEXEC;
5060 	} else {
5061 		/*
5062 		 * Ensure that at least one execute bit is on. Otherwise,
5063 		 * a privileged user will always succeed, and we don't want
5064 		 * this to happen unless the file really is executable.
5065 		 */
5066 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5067 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5068 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5069 			priv_granted |= VEXEC;
5070 	}
5071 
5072 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5073 	    !priv_check_cred(cred, PRIV_VFS_READ))
5074 		priv_granted |= VREAD;
5075 
5076 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5077 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5078 		priv_granted |= (VWRITE | VAPPEND);
5079 
5080 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5081 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5082 		priv_granted |= VADMIN;
5083 
5084 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5085 		/* XXX audit: privilege used */
5086 		if (privused != NULL)
5087 			*privused = 1;
5088 		return (0);
5089 	}
5090 
5091 	return ((accmode & VADMIN) ? EPERM : EACCES);
5092 }
5093 
5094 /*
5095  * Credential check based on process requesting service, and per-attribute
5096  * permissions.
5097  */
5098 int
5099 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5100     struct thread *td, accmode_t accmode)
5101 {
5102 
5103 	/*
5104 	 * Kernel-invoked always succeeds.
5105 	 */
5106 	if (cred == NOCRED)
5107 		return (0);
5108 
5109 	/*
5110 	 * Do not allow privileged processes in jail to directly manipulate
5111 	 * system attributes.
5112 	 */
5113 	switch (attrnamespace) {
5114 	case EXTATTR_NAMESPACE_SYSTEM:
5115 		/* Potentially should be: return (EPERM); */
5116 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5117 	case EXTATTR_NAMESPACE_USER:
5118 		return (VOP_ACCESS(vp, accmode, cred, td));
5119 	default:
5120 		return (EPERM);
5121 	}
5122 }
5123 
5124 #ifdef DEBUG_VFS_LOCKS
5125 /*
5126  * This only exists to suppress warnings from unlocked specfs accesses.  It is
5127  * no longer ok to have an unlocked VFS.
5128  */
5129 #define	IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL ||		\
5130 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
5131 
5132 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5133 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5134     "Drop into debugger on lock violation");
5135 
5136 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5137 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5138     0, "Check for interlock across VOPs");
5139 
5140 int vfs_badlock_print = 1;	/* Print lock violations. */
5141 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5142     0, "Print lock violations");
5143 
5144 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5145 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5146     0, "Print vnode details on lock violations");
5147 
5148 #ifdef KDB
5149 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5150 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5151     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5152 #endif
5153 
5154 static void
5155 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5156 {
5157 
5158 #ifdef KDB
5159 	if (vfs_badlock_backtrace)
5160 		kdb_backtrace();
5161 #endif
5162 	if (vfs_badlock_vnode)
5163 		vn_printf(vp, "vnode ");
5164 	if (vfs_badlock_print)
5165 		printf("%s: %p %s\n", str, (void *)vp, msg);
5166 	if (vfs_badlock_ddb)
5167 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5168 }
5169 
5170 void
5171 assert_vi_locked(struct vnode *vp, const char *str)
5172 {
5173 
5174 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5175 		vfs_badlock("interlock is not locked but should be", str, vp);
5176 }
5177 
5178 void
5179 assert_vi_unlocked(struct vnode *vp, const char *str)
5180 {
5181 
5182 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5183 		vfs_badlock("interlock is locked but should not be", str, vp);
5184 }
5185 
5186 void
5187 assert_vop_locked(struct vnode *vp, const char *str)
5188 {
5189 	int locked;
5190 
5191 	if (!IGNORE_LOCK(vp)) {
5192 		locked = VOP_ISLOCKED(vp);
5193 		if (locked == 0 || locked == LK_EXCLOTHER)
5194 			vfs_badlock("is not locked but should be", str, vp);
5195 	}
5196 }
5197 
5198 void
5199 assert_vop_unlocked(struct vnode *vp, const char *str)
5200 {
5201 
5202 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5203 		vfs_badlock("is locked but should not be", str, vp);
5204 }
5205 
5206 void
5207 assert_vop_elocked(struct vnode *vp, const char *str)
5208 {
5209 
5210 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5211 		vfs_badlock("is not exclusive locked but should be", str, vp);
5212 }
5213 #endif /* DEBUG_VFS_LOCKS */
5214 
5215 void
5216 vop_rename_fail(struct vop_rename_args *ap)
5217 {
5218 
5219 	if (ap->a_tvp != NULL)
5220 		vput(ap->a_tvp);
5221 	if (ap->a_tdvp == ap->a_tvp)
5222 		vrele(ap->a_tdvp);
5223 	else
5224 		vput(ap->a_tdvp);
5225 	vrele(ap->a_fdvp);
5226 	vrele(ap->a_fvp);
5227 }
5228 
5229 void
5230 vop_rename_pre(void *ap)
5231 {
5232 	struct vop_rename_args *a = ap;
5233 
5234 #ifdef DEBUG_VFS_LOCKS
5235 	if (a->a_tvp)
5236 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5237 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5238 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5239 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5240 
5241 	/* Check the source (from). */
5242 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5243 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5244 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5245 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5246 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5247 
5248 	/* Check the target. */
5249 	if (a->a_tvp)
5250 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5251 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5252 #endif
5253 	if (a->a_tdvp != a->a_fdvp)
5254 		vhold(a->a_fdvp);
5255 	if (a->a_tvp != a->a_fvp)
5256 		vhold(a->a_fvp);
5257 	vhold(a->a_tdvp);
5258 	if (a->a_tvp)
5259 		vhold(a->a_tvp);
5260 }
5261 
5262 #ifdef DEBUG_VFS_LOCKS
5263 void
5264 vop_strategy_pre(void *ap)
5265 {
5266 	struct vop_strategy_args *a;
5267 	struct buf *bp;
5268 
5269 	a = ap;
5270 	bp = a->a_bp;
5271 
5272 	/*
5273 	 * Cluster ops lock their component buffers but not the IO container.
5274 	 */
5275 	if ((bp->b_flags & B_CLUSTER) != 0)
5276 		return;
5277 
5278 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5279 		if (vfs_badlock_print)
5280 			printf(
5281 			    "VOP_STRATEGY: bp is not locked but should be\n");
5282 		if (vfs_badlock_ddb)
5283 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5284 	}
5285 }
5286 
5287 void
5288 vop_lock_pre(void *ap)
5289 {
5290 	struct vop_lock1_args *a = ap;
5291 
5292 	if ((a->a_flags & LK_INTERLOCK) == 0)
5293 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5294 	else
5295 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5296 }
5297 
5298 void
5299 vop_lock_post(void *ap, int rc)
5300 {
5301 	struct vop_lock1_args *a = ap;
5302 
5303 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5304 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5305 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5306 }
5307 
5308 void
5309 vop_unlock_pre(void *ap)
5310 {
5311 	struct vop_unlock_args *a = ap;
5312 
5313 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5314 }
5315 
5316 void
5317 vop_unlock_post(void *ap, int rc)
5318 {
5319 	return;
5320 }
5321 
5322 void
5323 vop_need_inactive_pre(void *ap)
5324 {
5325 	struct vop_need_inactive_args *a = ap;
5326 
5327 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5328 }
5329 
5330 void
5331 vop_need_inactive_post(void *ap, int rc)
5332 {
5333 	struct vop_need_inactive_args *a = ap;
5334 
5335 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5336 }
5337 #endif
5338 
5339 void
5340 vop_create_post(void *ap, int rc)
5341 {
5342 	struct vop_create_args *a = ap;
5343 
5344 	if (!rc)
5345 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5346 }
5347 
5348 void
5349 vop_deleteextattr_post(void *ap, int rc)
5350 {
5351 	struct vop_deleteextattr_args *a = ap;
5352 
5353 	if (!rc)
5354 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5355 }
5356 
5357 void
5358 vop_link_post(void *ap, int rc)
5359 {
5360 	struct vop_link_args *a = ap;
5361 
5362 	if (!rc) {
5363 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
5364 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
5365 	}
5366 }
5367 
5368 void
5369 vop_mkdir_post(void *ap, int rc)
5370 {
5371 	struct vop_mkdir_args *a = ap;
5372 
5373 	if (!rc)
5374 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5375 }
5376 
5377 void
5378 vop_mknod_post(void *ap, int rc)
5379 {
5380 	struct vop_mknod_args *a = ap;
5381 
5382 	if (!rc)
5383 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5384 }
5385 
5386 void
5387 vop_reclaim_post(void *ap, int rc)
5388 {
5389 	struct vop_reclaim_args *a = ap;
5390 
5391 	if (!rc)
5392 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
5393 }
5394 
5395 void
5396 vop_remove_post(void *ap, int rc)
5397 {
5398 	struct vop_remove_args *a = ap;
5399 
5400 	if (!rc) {
5401 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5402 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5403 	}
5404 }
5405 
5406 void
5407 vop_rename_post(void *ap, int rc)
5408 {
5409 	struct vop_rename_args *a = ap;
5410 	long hint;
5411 
5412 	if (!rc) {
5413 		hint = NOTE_WRITE;
5414 		if (a->a_fdvp == a->a_tdvp) {
5415 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5416 				hint |= NOTE_LINK;
5417 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5418 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5419 		} else {
5420 			hint |= NOTE_EXTEND;
5421 			if (a->a_fvp->v_type == VDIR)
5422 				hint |= NOTE_LINK;
5423 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5424 
5425 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5426 			    a->a_tvp->v_type == VDIR)
5427 				hint &= ~NOTE_LINK;
5428 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5429 		}
5430 
5431 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5432 		if (a->a_tvp)
5433 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5434 	}
5435 	if (a->a_tdvp != a->a_fdvp)
5436 		vdrop(a->a_fdvp);
5437 	if (a->a_tvp != a->a_fvp)
5438 		vdrop(a->a_fvp);
5439 	vdrop(a->a_tdvp);
5440 	if (a->a_tvp)
5441 		vdrop(a->a_tvp);
5442 }
5443 
5444 void
5445 vop_rmdir_post(void *ap, int rc)
5446 {
5447 	struct vop_rmdir_args *a = ap;
5448 
5449 	if (!rc) {
5450 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5451 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5452 	}
5453 }
5454 
5455 void
5456 vop_setattr_post(void *ap, int rc)
5457 {
5458 	struct vop_setattr_args *a = ap;
5459 
5460 	if (!rc)
5461 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5462 }
5463 
5464 void
5465 vop_setextattr_post(void *ap, int rc)
5466 {
5467 	struct vop_setextattr_args *a = ap;
5468 
5469 	if (!rc)
5470 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5471 }
5472 
5473 void
5474 vop_symlink_post(void *ap, int rc)
5475 {
5476 	struct vop_symlink_args *a = ap;
5477 
5478 	if (!rc)
5479 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5480 }
5481 
5482 void
5483 vop_open_post(void *ap, int rc)
5484 {
5485 	struct vop_open_args *a = ap;
5486 
5487 	if (!rc)
5488 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5489 }
5490 
5491 void
5492 vop_close_post(void *ap, int rc)
5493 {
5494 	struct vop_close_args *a = ap;
5495 
5496 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5497 	    !VN_IS_DOOMED(a->a_vp))) {
5498 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5499 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
5500 	}
5501 }
5502 
5503 void
5504 vop_read_post(void *ap, int rc)
5505 {
5506 	struct vop_read_args *a = ap;
5507 
5508 	if (!rc)
5509 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5510 }
5511 
5512 void
5513 vop_readdir_post(void *ap, int rc)
5514 {
5515 	struct vop_readdir_args *a = ap;
5516 
5517 	if (!rc)
5518 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5519 }
5520 
5521 static struct knlist fs_knlist;
5522 
5523 static void
5524 vfs_event_init(void *arg)
5525 {
5526 	knlist_init_mtx(&fs_knlist, NULL);
5527 }
5528 /* XXX - correct order? */
5529 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5530 
5531 void
5532 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5533 {
5534 
5535 	KNOTE_UNLOCKED(&fs_knlist, event);
5536 }
5537 
5538 static int	filt_fsattach(struct knote *kn);
5539 static void	filt_fsdetach(struct knote *kn);
5540 static int	filt_fsevent(struct knote *kn, long hint);
5541 
5542 struct filterops fs_filtops = {
5543 	.f_isfd = 0,
5544 	.f_attach = filt_fsattach,
5545 	.f_detach = filt_fsdetach,
5546 	.f_event = filt_fsevent
5547 };
5548 
5549 static int
5550 filt_fsattach(struct knote *kn)
5551 {
5552 
5553 	kn->kn_flags |= EV_CLEAR;
5554 	knlist_add(&fs_knlist, kn, 0);
5555 	return (0);
5556 }
5557 
5558 static void
5559 filt_fsdetach(struct knote *kn)
5560 {
5561 
5562 	knlist_remove(&fs_knlist, kn, 0);
5563 }
5564 
5565 static int
5566 filt_fsevent(struct knote *kn, long hint)
5567 {
5568 
5569 	kn->kn_fflags |= hint;
5570 	return (kn->kn_fflags != 0);
5571 }
5572 
5573 static int
5574 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5575 {
5576 	struct vfsidctl vc;
5577 	int error;
5578 	struct mount *mp;
5579 
5580 	error = SYSCTL_IN(req, &vc, sizeof(vc));
5581 	if (error)
5582 		return (error);
5583 	if (vc.vc_vers != VFS_CTL_VERS1)
5584 		return (EINVAL);
5585 	mp = vfs_getvfs(&vc.vc_fsid);
5586 	if (mp == NULL)
5587 		return (ENOENT);
5588 	/* ensure that a specific sysctl goes to the right filesystem. */
5589 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
5590 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
5591 		vfs_rel(mp);
5592 		return (EINVAL);
5593 	}
5594 	VCTLTOREQ(&vc, req);
5595 	error = VFS_SYSCTL(mp, vc.vc_op, req);
5596 	vfs_rel(mp);
5597 	return (error);
5598 }
5599 
5600 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
5601     NULL, 0, sysctl_vfs_ctl, "",
5602     "Sysctl by fsid");
5603 
5604 /*
5605  * Function to initialize a va_filerev field sensibly.
5606  * XXX: Wouldn't a random number make a lot more sense ??
5607  */
5608 u_quad_t
5609 init_va_filerev(void)
5610 {
5611 	struct bintime bt;
5612 
5613 	getbinuptime(&bt);
5614 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5615 }
5616 
5617 static int	filt_vfsread(struct knote *kn, long hint);
5618 static int	filt_vfswrite(struct knote *kn, long hint);
5619 static int	filt_vfsvnode(struct knote *kn, long hint);
5620 static void	filt_vfsdetach(struct knote *kn);
5621 static struct filterops vfsread_filtops = {
5622 	.f_isfd = 1,
5623 	.f_detach = filt_vfsdetach,
5624 	.f_event = filt_vfsread
5625 };
5626 static struct filterops vfswrite_filtops = {
5627 	.f_isfd = 1,
5628 	.f_detach = filt_vfsdetach,
5629 	.f_event = filt_vfswrite
5630 };
5631 static struct filterops vfsvnode_filtops = {
5632 	.f_isfd = 1,
5633 	.f_detach = filt_vfsdetach,
5634 	.f_event = filt_vfsvnode
5635 };
5636 
5637 static void
5638 vfs_knllock(void *arg)
5639 {
5640 	struct vnode *vp = arg;
5641 
5642 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5643 }
5644 
5645 static void
5646 vfs_knlunlock(void *arg)
5647 {
5648 	struct vnode *vp = arg;
5649 
5650 	VOP_UNLOCK(vp);
5651 }
5652 
5653 static void
5654 vfs_knl_assert_locked(void *arg)
5655 {
5656 #ifdef DEBUG_VFS_LOCKS
5657 	struct vnode *vp = arg;
5658 
5659 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
5660 #endif
5661 }
5662 
5663 static void
5664 vfs_knl_assert_unlocked(void *arg)
5665 {
5666 #ifdef DEBUG_VFS_LOCKS
5667 	struct vnode *vp = arg;
5668 
5669 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
5670 #endif
5671 }
5672 
5673 int
5674 vfs_kqfilter(struct vop_kqfilter_args *ap)
5675 {
5676 	struct vnode *vp = ap->a_vp;
5677 	struct knote *kn = ap->a_kn;
5678 	struct knlist *knl;
5679 
5680 	switch (kn->kn_filter) {
5681 	case EVFILT_READ:
5682 		kn->kn_fop = &vfsread_filtops;
5683 		break;
5684 	case EVFILT_WRITE:
5685 		kn->kn_fop = &vfswrite_filtops;
5686 		break;
5687 	case EVFILT_VNODE:
5688 		kn->kn_fop = &vfsvnode_filtops;
5689 		break;
5690 	default:
5691 		return (EINVAL);
5692 	}
5693 
5694 	kn->kn_hook = (caddr_t)vp;
5695 
5696 	v_addpollinfo(vp);
5697 	if (vp->v_pollinfo == NULL)
5698 		return (ENOMEM);
5699 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
5700 	vhold(vp);
5701 	knlist_add(knl, kn, 0);
5702 
5703 	return (0);
5704 }
5705 
5706 /*
5707  * Detach knote from vnode
5708  */
5709 static void
5710 filt_vfsdetach(struct knote *kn)
5711 {
5712 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5713 
5714 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5715 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5716 	vdrop(vp);
5717 }
5718 
5719 /*ARGSUSED*/
5720 static int
5721 filt_vfsread(struct knote *kn, long hint)
5722 {
5723 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5724 	struct vattr va;
5725 	int res;
5726 
5727 	/*
5728 	 * filesystem is gone, so set the EOF flag and schedule
5729 	 * the knote for deletion.
5730 	 */
5731 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5732 		VI_LOCK(vp);
5733 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5734 		VI_UNLOCK(vp);
5735 		return (1);
5736 	}
5737 
5738 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5739 		return (0);
5740 
5741 	VI_LOCK(vp);
5742 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5743 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5744 	VI_UNLOCK(vp);
5745 	return (res);
5746 }
5747 
5748 /*ARGSUSED*/
5749 static int
5750 filt_vfswrite(struct knote *kn, long hint)
5751 {
5752 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5753 
5754 	VI_LOCK(vp);
5755 
5756 	/*
5757 	 * filesystem is gone, so set the EOF flag and schedule
5758 	 * the knote for deletion.
5759 	 */
5760 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5761 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5762 
5763 	kn->kn_data = 0;
5764 	VI_UNLOCK(vp);
5765 	return (1);
5766 }
5767 
5768 static int
5769 filt_vfsvnode(struct knote *kn, long hint)
5770 {
5771 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5772 	int res;
5773 
5774 	VI_LOCK(vp);
5775 	if (kn->kn_sfflags & hint)
5776 		kn->kn_fflags |= hint;
5777 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5778 		kn->kn_flags |= EV_EOF;
5779 		VI_UNLOCK(vp);
5780 		return (1);
5781 	}
5782 	res = (kn->kn_fflags != 0);
5783 	VI_UNLOCK(vp);
5784 	return (res);
5785 }
5786 
5787 /*
5788  * Returns whether the directory is empty or not.
5789  * If it is empty, the return value is 0; otherwise
5790  * the return value is an error value (which may
5791  * be ENOTEMPTY).
5792  */
5793 int
5794 vfs_emptydir(struct vnode *vp)
5795 {
5796 	struct uio uio;
5797 	struct iovec iov;
5798 	struct dirent *dirent, *dp, *endp;
5799 	int error, eof;
5800 
5801 	error = 0;
5802 	eof = 0;
5803 
5804 	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
5805 
5806 	dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
5807 	iov.iov_base = dirent;
5808 	iov.iov_len = sizeof(struct dirent);
5809 
5810 	uio.uio_iov = &iov;
5811 	uio.uio_iovcnt = 1;
5812 	uio.uio_offset = 0;
5813 	uio.uio_resid = sizeof(struct dirent);
5814 	uio.uio_segflg = UIO_SYSSPACE;
5815 	uio.uio_rw = UIO_READ;
5816 	uio.uio_td = curthread;
5817 
5818 	while (eof == 0 && error == 0) {
5819 		error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
5820 		    NULL, NULL);
5821 		if (error != 0)
5822 			break;
5823 		endp = (void *)((uint8_t *)dirent +
5824 		    sizeof(struct dirent) - uio.uio_resid);
5825 		for (dp = dirent; dp < endp;
5826 		     dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
5827 			if (dp->d_type == DT_WHT)
5828 				continue;
5829 			if (dp->d_namlen == 0)
5830 				continue;
5831 			if (dp->d_type != DT_DIR &&
5832 			    dp->d_type != DT_UNKNOWN) {
5833 				error = ENOTEMPTY;
5834 				break;
5835 			}
5836 			if (dp->d_namlen > 2) {
5837 				error = ENOTEMPTY;
5838 				break;
5839 			}
5840 			if (dp->d_namlen == 1 &&
5841 			    dp->d_name[0] != '.') {
5842 				error = ENOTEMPTY;
5843 				break;
5844 			}
5845 			if (dp->d_namlen == 2 &&
5846 			    dp->d_name[1] != '.') {
5847 				error = ENOTEMPTY;
5848 				break;
5849 			}
5850 			uio.uio_resid = sizeof(struct dirent);
5851 		}
5852 	}
5853 	free(dirent, M_TEMP);
5854 	return (error);
5855 }
5856 
5857 int
5858 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5859 {
5860 	int error;
5861 
5862 	if (dp->d_reclen > ap->a_uio->uio_resid)
5863 		return (ENAMETOOLONG);
5864 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
5865 	if (error) {
5866 		if (ap->a_ncookies != NULL) {
5867 			if (ap->a_cookies != NULL)
5868 				free(ap->a_cookies, M_TEMP);
5869 			ap->a_cookies = NULL;
5870 			*ap->a_ncookies = 0;
5871 		}
5872 		return (error);
5873 	}
5874 	if (ap->a_ncookies == NULL)
5875 		return (0);
5876 
5877 	KASSERT(ap->a_cookies,
5878 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
5879 
5880 	*ap->a_cookies = realloc(*ap->a_cookies,
5881 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
5882 	(*ap->a_cookies)[*ap->a_ncookies] = off;
5883 	*ap->a_ncookies += 1;
5884 	return (0);
5885 }
5886 
5887 /*
5888  * Mark for update the access time of the file if the filesystem
5889  * supports VOP_MARKATIME.  This functionality is used by execve and
5890  * mmap, so we want to avoid the I/O implied by directly setting
5891  * va_atime for the sake of efficiency.
5892  */
5893 void
5894 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
5895 {
5896 	struct mount *mp;
5897 
5898 	mp = vp->v_mount;
5899 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
5900 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
5901 		(void)VOP_MARKATIME(vp);
5902 }
5903 
5904 /*
5905  * The purpose of this routine is to remove granularity from accmode_t,
5906  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
5907  * VADMIN and VAPPEND.
5908  *
5909  * If it returns 0, the caller is supposed to continue with the usual
5910  * access checks using 'accmode' as modified by this routine.  If it
5911  * returns nonzero value, the caller is supposed to return that value
5912  * as errno.
5913  *
5914  * Note that after this routine runs, accmode may be zero.
5915  */
5916 int
5917 vfs_unixify_accmode(accmode_t *accmode)
5918 {
5919 	/*
5920 	 * There is no way to specify explicit "deny" rule using
5921 	 * file mode or POSIX.1e ACLs.
5922 	 */
5923 	if (*accmode & VEXPLICIT_DENY) {
5924 		*accmode = 0;
5925 		return (0);
5926 	}
5927 
5928 	/*
5929 	 * None of these can be translated into usual access bits.
5930 	 * Also, the common case for NFSv4 ACLs is to not contain
5931 	 * either of these bits. Caller should check for VWRITE
5932 	 * on the containing directory instead.
5933 	 */
5934 	if (*accmode & (VDELETE_CHILD | VDELETE))
5935 		return (EPERM);
5936 
5937 	if (*accmode & VADMIN_PERMS) {
5938 		*accmode &= ~VADMIN_PERMS;
5939 		*accmode |= VADMIN;
5940 	}
5941 
5942 	/*
5943 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5944 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5945 	 */
5946 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5947 
5948 	return (0);
5949 }
5950 
5951 /*
5952  * Clear out a doomed vnode (if any) and replace it with a new one as long
5953  * as the fs is not being unmounted. Return the root vnode to the caller.
5954  */
5955 static int __noinline
5956 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
5957 {
5958 	struct vnode *vp;
5959 	int error;
5960 
5961 restart:
5962 	if (mp->mnt_rootvnode != NULL) {
5963 		MNT_ILOCK(mp);
5964 		vp = mp->mnt_rootvnode;
5965 		if (vp != NULL) {
5966 			if (!VN_IS_DOOMED(vp)) {
5967 				vrefact(vp);
5968 				MNT_IUNLOCK(mp);
5969 				error = vn_lock(vp, flags);
5970 				if (error == 0) {
5971 					*vpp = vp;
5972 					return (0);
5973 				}
5974 				vrele(vp);
5975 				goto restart;
5976 			}
5977 			/*
5978 			 * Clear the old one.
5979 			 */
5980 			mp->mnt_rootvnode = NULL;
5981 		}
5982 		MNT_IUNLOCK(mp);
5983 		if (vp != NULL) {
5984 			/*
5985 			 * Paired with a fence in vfs_op_thread_exit().
5986 			 */
5987 			atomic_thread_fence_acq();
5988 			vfs_op_barrier_wait(mp);
5989 			vrele(vp);
5990 		}
5991 	}
5992 	error = VFS_CACHEDROOT(mp, flags, vpp);
5993 	if (error != 0)
5994 		return (error);
5995 	if (mp->mnt_vfs_ops == 0) {
5996 		MNT_ILOCK(mp);
5997 		if (mp->mnt_vfs_ops != 0) {
5998 			MNT_IUNLOCK(mp);
5999 			return (0);
6000 		}
6001 		if (mp->mnt_rootvnode == NULL) {
6002 			vrefact(*vpp);
6003 			mp->mnt_rootvnode = *vpp;
6004 		} else {
6005 			if (mp->mnt_rootvnode != *vpp) {
6006 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6007 					panic("%s: mismatch between vnode returned "
6008 					    " by VFS_CACHEDROOT and the one cached "
6009 					    " (%p != %p)",
6010 					    __func__, *vpp, mp->mnt_rootvnode);
6011 				}
6012 			}
6013 		}
6014 		MNT_IUNLOCK(mp);
6015 	}
6016 	return (0);
6017 }
6018 
6019 int
6020 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6021 {
6022 	struct vnode *vp;
6023 	int error;
6024 
6025 	if (!vfs_op_thread_enter(mp))
6026 		return (vfs_cache_root_fallback(mp, flags, vpp));
6027 	vp = (struct vnode *)atomic_load_ptr(&mp->mnt_rootvnode);
6028 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6029 		vfs_op_thread_exit(mp);
6030 		return (vfs_cache_root_fallback(mp, flags, vpp));
6031 	}
6032 	vrefact(vp);
6033 	vfs_op_thread_exit(mp);
6034 	error = vn_lock(vp, flags);
6035 	if (error != 0) {
6036 		vrele(vp);
6037 		return (vfs_cache_root_fallback(mp, flags, vpp));
6038 	}
6039 	*vpp = vp;
6040 	return (0);
6041 }
6042 
6043 struct vnode *
6044 vfs_cache_root_clear(struct mount *mp)
6045 {
6046 	struct vnode *vp;
6047 
6048 	/*
6049 	 * ops > 0 guarantees there is nobody who can see this vnode
6050 	 */
6051 	MPASS(mp->mnt_vfs_ops > 0);
6052 	vp = mp->mnt_rootvnode;
6053 	mp->mnt_rootvnode = NULL;
6054 	return (vp);
6055 }
6056 
6057 void
6058 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6059 {
6060 
6061 	MPASS(mp->mnt_vfs_ops > 0);
6062 	vrefact(vp);
6063 	mp->mnt_rootvnode = vp;
6064 }
6065 
6066 /*
6067  * These are helper functions for filesystems to traverse all
6068  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6069  *
6070  * This interface replaces MNT_VNODE_FOREACH.
6071  */
6072 
6073 
6074 struct vnode *
6075 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6076 {
6077 	struct vnode *vp;
6078 
6079 	if (should_yield())
6080 		kern_yield(PRI_USER);
6081 	MNT_ILOCK(mp);
6082 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6083 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6084 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6085 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6086 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6087 			continue;
6088 		VI_LOCK(vp);
6089 		if (VN_IS_DOOMED(vp)) {
6090 			VI_UNLOCK(vp);
6091 			continue;
6092 		}
6093 		break;
6094 	}
6095 	if (vp == NULL) {
6096 		__mnt_vnode_markerfree_all(mvp, mp);
6097 		/* MNT_IUNLOCK(mp); -- done in above function */
6098 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6099 		return (NULL);
6100 	}
6101 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6102 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6103 	MNT_IUNLOCK(mp);
6104 	return (vp);
6105 }
6106 
6107 struct vnode *
6108 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6109 {
6110 	struct vnode *vp;
6111 
6112 	*mvp = vn_alloc_marker(mp);
6113 	MNT_ILOCK(mp);
6114 	MNT_REF(mp);
6115 
6116 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6117 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6118 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6119 			continue;
6120 		VI_LOCK(vp);
6121 		if (VN_IS_DOOMED(vp)) {
6122 			VI_UNLOCK(vp);
6123 			continue;
6124 		}
6125 		break;
6126 	}
6127 	if (vp == NULL) {
6128 		MNT_REL(mp);
6129 		MNT_IUNLOCK(mp);
6130 		vn_free_marker(*mvp);
6131 		*mvp = NULL;
6132 		return (NULL);
6133 	}
6134 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6135 	MNT_IUNLOCK(mp);
6136 	return (vp);
6137 }
6138 
6139 void
6140 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6141 {
6142 
6143 	if (*mvp == NULL) {
6144 		MNT_IUNLOCK(mp);
6145 		return;
6146 	}
6147 
6148 	mtx_assert(MNT_MTX(mp), MA_OWNED);
6149 
6150 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6151 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6152 	MNT_REL(mp);
6153 	MNT_IUNLOCK(mp);
6154 	vn_free_marker(*mvp);
6155 	*mvp = NULL;
6156 }
6157 
6158 /*
6159  * These are helper functions for filesystems to traverse their
6160  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6161  */
6162 static void
6163 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6164 {
6165 
6166 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6167 
6168 	MNT_ILOCK(mp);
6169 	MNT_REL(mp);
6170 	MNT_IUNLOCK(mp);
6171 	vn_free_marker(*mvp);
6172 	*mvp = NULL;
6173 }
6174 
6175 /*
6176  * Relock the mp mount vnode list lock with the vp vnode interlock in the
6177  * conventional lock order during mnt_vnode_next_lazy iteration.
6178  *
6179  * On entry, the mount vnode list lock is held and the vnode interlock is not.
6180  * The list lock is dropped and reacquired.  On success, both locks are held.
6181  * On failure, the mount vnode list lock is held but the vnode interlock is
6182  * not, and the procedure may have yielded.
6183  */
6184 static bool
6185 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6186     struct vnode *vp)
6187 {
6188 	const struct vnode *tmp;
6189 	bool held, ret;
6190 
6191 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6192 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6193 	    ("%s: bad marker", __func__));
6194 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6195 	    ("%s: inappropriate vnode", __func__));
6196 	ASSERT_VI_UNLOCKED(vp, __func__);
6197 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6198 
6199 	ret = false;
6200 
6201 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6202 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6203 
6204 	/*
6205 	 * Use a hold to prevent vp from disappearing while the mount vnode
6206 	 * list lock is dropped and reacquired.  Normally a hold would be
6207 	 * acquired with vhold(), but that might try to acquire the vnode
6208 	 * interlock, which would be a LOR with the mount vnode list lock.
6209 	 */
6210 	held = refcount_acquire_if_not_zero(&vp->v_holdcnt);
6211 	mtx_unlock(&mp->mnt_listmtx);
6212 	if (!held)
6213 		goto abort;
6214 	VI_LOCK(vp);
6215 	if (!refcount_release_if_not_last(&vp->v_holdcnt)) {
6216 		vdropl(vp);
6217 		goto abort;
6218 	}
6219 	mtx_lock(&mp->mnt_listmtx);
6220 
6221 	/*
6222 	 * Determine whether the vnode is still the next one after the marker,
6223 	 * excepting any other markers.  If the vnode has not been doomed by
6224 	 * vgone() then the hold should have ensured that it remained on the
6225 	 * lazy list.  If it has been doomed but is still on the lazy list,
6226 	 * don't abort, but rather skip over it (avoid spinning on doomed
6227 	 * vnodes).
6228 	 */
6229 	tmp = mvp;
6230 	do {
6231 		tmp = TAILQ_NEXT(tmp, v_lazylist);
6232 	} while (tmp != NULL && tmp->v_type == VMARKER);
6233 	if (tmp != vp) {
6234 		mtx_unlock(&mp->mnt_listmtx);
6235 		VI_UNLOCK(vp);
6236 		goto abort;
6237 	}
6238 
6239 	ret = true;
6240 	goto out;
6241 abort:
6242 	maybe_yield();
6243 	mtx_lock(&mp->mnt_listmtx);
6244 out:
6245 	if (ret)
6246 		ASSERT_VI_LOCKED(vp, __func__);
6247 	else
6248 		ASSERT_VI_UNLOCKED(vp, __func__);
6249 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6250 	return (ret);
6251 }
6252 
6253 static struct vnode *
6254 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6255     void *cbarg)
6256 {
6257 	struct vnode *vp, *nvp;
6258 
6259 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6260 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6261 restart:
6262 	vp = TAILQ_NEXT(*mvp, v_lazylist);
6263 	while (vp != NULL) {
6264 		if (vp->v_type == VMARKER) {
6265 			vp = TAILQ_NEXT(vp, v_lazylist);
6266 			continue;
6267 		}
6268 		/*
6269 		 * See if we want to process the vnode. Note we may encounter a
6270 		 * long string of vnodes we don't care about and hog the list
6271 		 * as a result. Check for it and requeue the marker.
6272 		 */
6273 		if (VN_IS_DOOMED(vp) || !cb(vp, cbarg)) {
6274 			if (!should_yield()) {
6275 				vp = TAILQ_NEXT(vp, v_lazylist);
6276 				continue;
6277 			}
6278 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6279 			    v_lazylist);
6280 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6281 			    v_lazylist);
6282 			mtx_unlock(&mp->mnt_listmtx);
6283 			kern_yield(PRI_USER);
6284 			mtx_lock(&mp->mnt_listmtx);
6285 			goto restart;
6286 		}
6287 		/*
6288 		 * Try-lock because this is the wrong lock order.  If that does
6289 		 * not succeed, drop the mount vnode list lock and try to
6290 		 * reacquire it and the vnode interlock in the right order.
6291 		 */
6292 		if (!VI_TRYLOCK(vp) &&
6293 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6294 			goto restart;
6295 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6296 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6297 		    ("alien vnode on the lazy list %p %p", vp, mp));
6298 		if (vp->v_mount == mp && !VN_IS_DOOMED(vp))
6299 			break;
6300 		nvp = TAILQ_NEXT(vp, v_lazylist);
6301 		VI_UNLOCK(vp);
6302 		vp = nvp;
6303 	}
6304 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6305 
6306 	/* Check if we are done */
6307 	if (vp == NULL) {
6308 		mtx_unlock(&mp->mnt_listmtx);
6309 		mnt_vnode_markerfree_lazy(mvp, mp);
6310 		return (NULL);
6311 	}
6312 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6313 	mtx_unlock(&mp->mnt_listmtx);
6314 	ASSERT_VI_LOCKED(vp, "lazy iter");
6315 	return (vp);
6316 }
6317 
6318 struct vnode *
6319 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6320     void *cbarg)
6321 {
6322 
6323 	if (should_yield())
6324 		kern_yield(PRI_USER);
6325 	mtx_lock(&mp->mnt_listmtx);
6326 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6327 }
6328 
6329 struct vnode *
6330 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6331     void *cbarg)
6332 {
6333 	struct vnode *vp;
6334 
6335 	*mvp = vn_alloc_marker(mp);
6336 	MNT_ILOCK(mp);
6337 	MNT_REF(mp);
6338 	MNT_IUNLOCK(mp);
6339 
6340 	mtx_lock(&mp->mnt_listmtx);
6341 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6342 	if (vp == NULL) {
6343 		mtx_unlock(&mp->mnt_listmtx);
6344 		mnt_vnode_markerfree_lazy(mvp, mp);
6345 		return (NULL);
6346 	}
6347 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6348 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6349 }
6350 
6351 void
6352 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6353 {
6354 
6355 	if (*mvp == NULL)
6356 		return;
6357 
6358 	mtx_lock(&mp->mnt_listmtx);
6359 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6360 	mtx_unlock(&mp->mnt_listmtx);
6361 	mnt_vnode_markerfree_lazy(mvp, mp);
6362 }
6363