xref: /freebsd/sys/kern/vfs_subr.c (revision 56e766af41cd68310f5583bb893b13c006fcb44f)
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 #ifdef WITNESS
1673 	if (lo->lo_name != tag) {
1674 #endif
1675 		lo->lo_name = tag;
1676 #ifdef WITNESS
1677 		WITNESS_DESTROY(lo);
1678 		WITNESS_INIT(lo, tag);
1679 	}
1680 #endif
1681 	/*
1682 	 * By default, don't allow shared locks unless filesystems opt-in.
1683 	 */
1684 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1685 	/*
1686 	 * Finalize various vnode identity bits.
1687 	 */
1688 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1689 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1690 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1691 	vp->v_type = VNON;
1692 	vp->v_op = vops;
1693 	v_init_counters(vp);
1694 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1695 #ifdef DIAGNOSTIC
1696 	if (mp == NULL && vops != &dead_vnodeops)
1697 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1698 #endif
1699 #ifdef MAC
1700 	mac_vnode_init(vp);
1701 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1702 		mac_vnode_associate_singlelabel(mp, vp);
1703 #endif
1704 	if (mp != NULL) {
1705 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1706 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1707 			vp->v_vflag |= VV_NOKNOTE;
1708 	}
1709 
1710 	/*
1711 	 * For the filesystems which do not use vfs_hash_insert(),
1712 	 * still initialize v_hash to have vfs_hash_index() useful.
1713 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1714 	 * its own hashing.
1715 	 */
1716 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1717 
1718 	*vpp = vp;
1719 	return (0);
1720 }
1721 
1722 void
1723 getnewvnode_reserve(void)
1724 {
1725 	struct thread *td;
1726 
1727 	td = curthread;
1728 	MPASS(td->td_vp_reserved == NULL);
1729 	td->td_vp_reserved = vn_alloc(NULL);
1730 }
1731 
1732 void
1733 getnewvnode_drop_reserve(void)
1734 {
1735 	struct thread *td;
1736 
1737 	td = curthread;
1738 	if (td->td_vp_reserved != NULL) {
1739 		vn_free(td->td_vp_reserved);
1740 		td->td_vp_reserved = NULL;
1741 	}
1742 }
1743 
1744 static void
1745 freevnode(struct vnode *vp)
1746 {
1747 	struct bufobj *bo;
1748 
1749 	/*
1750 	 * The vnode has been marked for destruction, so free it.
1751 	 *
1752 	 * The vnode will be returned to the zone where it will
1753 	 * normally remain until it is needed for another vnode. We
1754 	 * need to cleanup (or verify that the cleanup has already
1755 	 * been done) any residual data left from its current use
1756 	 * so as not to contaminate the freshly allocated vnode.
1757 	 */
1758 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1759 	bo = &vp->v_bufobj;
1760 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1761 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
1762 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1763 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1764 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1765 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1766 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1767 	    ("clean blk trie not empty"));
1768 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1769 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1770 	    ("dirty blk trie not empty"));
1771 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1772 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1773 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1774 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1775 	    ("Dangling rangelock waiters"));
1776 	VI_UNLOCK(vp);
1777 #ifdef MAC
1778 	mac_vnode_destroy(vp);
1779 #endif
1780 	if (vp->v_pollinfo != NULL) {
1781 		destroy_vpollinfo(vp->v_pollinfo);
1782 		vp->v_pollinfo = NULL;
1783 	}
1784 #ifdef INVARIANTS
1785 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
1786 	vp->v_op = NULL;
1787 #endif
1788 	vp->v_mountedhere = NULL;
1789 	vp->v_unpcb = NULL;
1790 	vp->v_rdev = NULL;
1791 	vp->v_fifoinfo = NULL;
1792 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
1793 	vp->v_irflag = 0;
1794 	vp->v_iflag = 0;
1795 	vp->v_vflag = 0;
1796 	bo->bo_flag = 0;
1797 	vn_free(vp);
1798 }
1799 
1800 /*
1801  * Delete from old mount point vnode list, if on one.
1802  */
1803 static void
1804 delmntque(struct vnode *vp)
1805 {
1806 	struct mount *mp;
1807 
1808 	mp = vp->v_mount;
1809 	if (mp == NULL)
1810 		return;
1811 	MNT_ILOCK(mp);
1812 	VI_LOCK(vp);
1813 	if (vp->v_mflag & VMP_LAZYLIST) {
1814 		mtx_lock(&mp->mnt_listmtx);
1815 		if (vp->v_mflag & VMP_LAZYLIST) {
1816 			vp->v_mflag &= ~VMP_LAZYLIST;
1817 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
1818 			mp->mnt_lazyvnodelistsize--;
1819 		}
1820 		mtx_unlock(&mp->mnt_listmtx);
1821 	}
1822 	vp->v_mount = NULL;
1823 	VI_UNLOCK(vp);
1824 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1825 		("bad mount point vnode list size"));
1826 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1827 	mp->mnt_nvnodelistsize--;
1828 	MNT_REL(mp);
1829 	MNT_IUNLOCK(mp);
1830 }
1831 
1832 static void
1833 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1834 {
1835 
1836 	vp->v_data = NULL;
1837 	vp->v_op = &dead_vnodeops;
1838 	vgone(vp);
1839 	vput(vp);
1840 }
1841 
1842 /*
1843  * Insert into list of vnodes for the new mount point, if available.
1844  */
1845 int
1846 insmntque1(struct vnode *vp, struct mount *mp,
1847 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1848 {
1849 
1850 	KASSERT(vp->v_mount == NULL,
1851 		("insmntque: vnode already on per mount vnode list"));
1852 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1853 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1854 
1855 	/*
1856 	 * We acquire the vnode interlock early to ensure that the
1857 	 * vnode cannot be recycled by another process releasing a
1858 	 * holdcnt on it before we get it on both the vnode list
1859 	 * and the active vnode list. The mount mutex protects only
1860 	 * manipulation of the vnode list and the vnode freelist
1861 	 * mutex protects only manipulation of the active vnode list.
1862 	 * Hence the need to hold the vnode interlock throughout.
1863 	 */
1864 	MNT_ILOCK(mp);
1865 	VI_LOCK(vp);
1866 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1867 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1868 	    mp->mnt_nvnodelistsize == 0)) &&
1869 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1870 		VI_UNLOCK(vp);
1871 		MNT_IUNLOCK(mp);
1872 		if (dtr != NULL)
1873 			dtr(vp, dtr_arg);
1874 		return (EBUSY);
1875 	}
1876 	vp->v_mount = mp;
1877 	MNT_REF(mp);
1878 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1879 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1880 		("neg mount point vnode list size"));
1881 	mp->mnt_nvnodelistsize++;
1882 	VI_UNLOCK(vp);
1883 	MNT_IUNLOCK(mp);
1884 	return (0);
1885 }
1886 
1887 int
1888 insmntque(struct vnode *vp, struct mount *mp)
1889 {
1890 
1891 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1892 }
1893 
1894 /*
1895  * Flush out and invalidate all buffers associated with a bufobj
1896  * Called with the underlying object locked.
1897  */
1898 int
1899 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1900 {
1901 	int error;
1902 
1903 	BO_LOCK(bo);
1904 	if (flags & V_SAVE) {
1905 		error = bufobj_wwait(bo, slpflag, slptimeo);
1906 		if (error) {
1907 			BO_UNLOCK(bo);
1908 			return (error);
1909 		}
1910 		if (bo->bo_dirty.bv_cnt > 0) {
1911 			BO_UNLOCK(bo);
1912 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1913 				return (error);
1914 			/*
1915 			 * XXX We could save a lock/unlock if this was only
1916 			 * enabled under INVARIANTS
1917 			 */
1918 			BO_LOCK(bo);
1919 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1920 				panic("vinvalbuf: dirty bufs");
1921 		}
1922 	}
1923 	/*
1924 	 * If you alter this loop please notice that interlock is dropped and
1925 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1926 	 * no race conditions occur from this.
1927 	 */
1928 	do {
1929 		error = flushbuflist(&bo->bo_clean,
1930 		    flags, bo, slpflag, slptimeo);
1931 		if (error == 0 && !(flags & V_CLEANONLY))
1932 			error = flushbuflist(&bo->bo_dirty,
1933 			    flags, bo, slpflag, slptimeo);
1934 		if (error != 0 && error != EAGAIN) {
1935 			BO_UNLOCK(bo);
1936 			return (error);
1937 		}
1938 	} while (error != 0);
1939 
1940 	/*
1941 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1942 	 * have write I/O in-progress but if there is a VM object then the
1943 	 * VM object can also have read-I/O in-progress.
1944 	 */
1945 	do {
1946 		bufobj_wwait(bo, 0, 0);
1947 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
1948 			BO_UNLOCK(bo);
1949 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
1950 			BO_LOCK(bo);
1951 		}
1952 	} while (bo->bo_numoutput > 0);
1953 	BO_UNLOCK(bo);
1954 
1955 	/*
1956 	 * Destroy the copy in the VM cache, too.
1957 	 */
1958 	if (bo->bo_object != NULL &&
1959 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1960 		VM_OBJECT_WLOCK(bo->bo_object);
1961 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1962 		    OBJPR_CLEANONLY : 0);
1963 		VM_OBJECT_WUNLOCK(bo->bo_object);
1964 	}
1965 
1966 #ifdef INVARIANTS
1967 	BO_LOCK(bo);
1968 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1969 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1970 	    bo->bo_clean.bv_cnt > 0))
1971 		panic("vinvalbuf: flush failed");
1972 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1973 	    bo->bo_dirty.bv_cnt > 0)
1974 		panic("vinvalbuf: flush dirty failed");
1975 	BO_UNLOCK(bo);
1976 #endif
1977 	return (0);
1978 }
1979 
1980 /*
1981  * Flush out and invalidate all buffers associated with a vnode.
1982  * Called with the underlying object locked.
1983  */
1984 int
1985 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1986 {
1987 
1988 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1989 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1990 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1991 		return (0);
1992 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1993 }
1994 
1995 /*
1996  * Flush out buffers on the specified list.
1997  *
1998  */
1999 static int
2000 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2001     int slptimeo)
2002 {
2003 	struct buf *bp, *nbp;
2004 	int retval, error;
2005 	daddr_t lblkno;
2006 	b_xflags_t xflags;
2007 
2008 	ASSERT_BO_WLOCKED(bo);
2009 
2010 	retval = 0;
2011 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2012 		/*
2013 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2014 		 * do not skip any buffers. If we are flushing only V_NORMAL
2015 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2016 		 * flushing only V_ALT buffers then skip buffers not marked
2017 		 * as BX_ALTDATA.
2018 		 */
2019 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2020 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2021 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2022 			continue;
2023 		}
2024 		if (nbp != NULL) {
2025 			lblkno = nbp->b_lblkno;
2026 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2027 		}
2028 		retval = EAGAIN;
2029 		error = BUF_TIMELOCK(bp,
2030 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2031 		    "flushbuf", slpflag, slptimeo);
2032 		if (error) {
2033 			BO_LOCK(bo);
2034 			return (error != ENOLCK ? error : EAGAIN);
2035 		}
2036 		KASSERT(bp->b_bufobj == bo,
2037 		    ("bp %p wrong b_bufobj %p should be %p",
2038 		    bp, bp->b_bufobj, bo));
2039 		/*
2040 		 * XXX Since there are no node locks for NFS, I
2041 		 * believe there is a slight chance that a delayed
2042 		 * write will occur while sleeping just above, so
2043 		 * check for it.
2044 		 */
2045 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2046 		    (flags & V_SAVE)) {
2047 			bremfree(bp);
2048 			bp->b_flags |= B_ASYNC;
2049 			bwrite(bp);
2050 			BO_LOCK(bo);
2051 			return (EAGAIN);	/* XXX: why not loop ? */
2052 		}
2053 		bremfree(bp);
2054 		bp->b_flags |= (B_INVAL | B_RELBUF);
2055 		bp->b_flags &= ~B_ASYNC;
2056 		brelse(bp);
2057 		BO_LOCK(bo);
2058 		if (nbp == NULL)
2059 			break;
2060 		nbp = gbincore(bo, lblkno);
2061 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2062 		    != xflags)
2063 			break;			/* nbp invalid */
2064 	}
2065 	return (retval);
2066 }
2067 
2068 int
2069 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2070 {
2071 	struct buf *bp;
2072 	int error;
2073 	daddr_t lblkno;
2074 
2075 	ASSERT_BO_LOCKED(bo);
2076 
2077 	for (lblkno = startn;;) {
2078 again:
2079 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2080 		if (bp == NULL || bp->b_lblkno >= endn ||
2081 		    bp->b_lblkno < startn)
2082 			break;
2083 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2084 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2085 		if (error != 0) {
2086 			BO_RLOCK(bo);
2087 			if (error == ENOLCK)
2088 				goto again;
2089 			return (error);
2090 		}
2091 		KASSERT(bp->b_bufobj == bo,
2092 		    ("bp %p wrong b_bufobj %p should be %p",
2093 		    bp, bp->b_bufobj, bo));
2094 		lblkno = bp->b_lblkno + 1;
2095 		if ((bp->b_flags & B_MANAGED) == 0)
2096 			bremfree(bp);
2097 		bp->b_flags |= B_RELBUF;
2098 		/*
2099 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2100 		 * pages backing each buffer in the range are unlikely to be
2101 		 * reused.  Dirty buffers will have the hint applied once
2102 		 * they've been written.
2103 		 */
2104 		if ((bp->b_flags & B_VMIO) != 0)
2105 			bp->b_flags |= B_NOREUSE;
2106 		brelse(bp);
2107 		BO_RLOCK(bo);
2108 	}
2109 	return (0);
2110 }
2111 
2112 /*
2113  * Truncate a file's buffer and pages to a specified length.  This
2114  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2115  * sync activity.
2116  */
2117 int
2118 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2119 {
2120 	struct buf *bp, *nbp;
2121 	struct bufobj *bo;
2122 	daddr_t startlbn;
2123 
2124 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2125 	    vp, blksize, (uintmax_t)length);
2126 
2127 	/*
2128 	 * Round up to the *next* lbn.
2129 	 */
2130 	startlbn = howmany(length, blksize);
2131 
2132 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2133 
2134 	bo = &vp->v_bufobj;
2135 restart_unlocked:
2136 	BO_LOCK(bo);
2137 
2138 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2139 		;
2140 
2141 	if (length > 0) {
2142 restartsync:
2143 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2144 			if (bp->b_lblkno > 0)
2145 				continue;
2146 			/*
2147 			 * Since we hold the vnode lock this should only
2148 			 * fail if we're racing with the buf daemon.
2149 			 */
2150 			if (BUF_LOCK(bp,
2151 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2152 			    BO_LOCKPTR(bo)) == ENOLCK)
2153 				goto restart_unlocked;
2154 
2155 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2156 			    ("buf(%p) on dirty queue without DELWRI", bp));
2157 
2158 			bremfree(bp);
2159 			bawrite(bp);
2160 			BO_LOCK(bo);
2161 			goto restartsync;
2162 		}
2163 	}
2164 
2165 	bufobj_wwait(bo, 0, 0);
2166 	BO_UNLOCK(bo);
2167 	vnode_pager_setsize(vp, length);
2168 
2169 	return (0);
2170 }
2171 
2172 /*
2173  * Invalidate the cached pages of a file's buffer within the range of block
2174  * numbers [startlbn, endlbn).
2175  */
2176 void
2177 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2178     int blksize)
2179 {
2180 	struct bufobj *bo;
2181 	off_t start, end;
2182 
2183 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2184 
2185 	start = blksize * startlbn;
2186 	end = blksize * endlbn;
2187 
2188 	bo = &vp->v_bufobj;
2189 	BO_LOCK(bo);
2190 	MPASS(blksize == bo->bo_bsize);
2191 
2192 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2193 		;
2194 
2195 	BO_UNLOCK(bo);
2196 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2197 }
2198 
2199 static int
2200 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2201     daddr_t startlbn, daddr_t endlbn)
2202 {
2203 	struct buf *bp, *nbp;
2204 	bool anyfreed;
2205 
2206 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2207 	ASSERT_BO_LOCKED(bo);
2208 
2209 	do {
2210 		anyfreed = false;
2211 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2212 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2213 				continue;
2214 			if (BUF_LOCK(bp,
2215 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2216 			    BO_LOCKPTR(bo)) == ENOLCK) {
2217 				BO_LOCK(bo);
2218 				return (EAGAIN);
2219 			}
2220 
2221 			bremfree(bp);
2222 			bp->b_flags |= B_INVAL | B_RELBUF;
2223 			bp->b_flags &= ~B_ASYNC;
2224 			brelse(bp);
2225 			anyfreed = true;
2226 
2227 			BO_LOCK(bo);
2228 			if (nbp != NULL &&
2229 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2230 			    nbp->b_vp != vp ||
2231 			    (nbp->b_flags & B_DELWRI) != 0))
2232 				return (EAGAIN);
2233 		}
2234 
2235 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2236 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2237 				continue;
2238 			if (BUF_LOCK(bp,
2239 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2240 			    BO_LOCKPTR(bo)) == ENOLCK) {
2241 				BO_LOCK(bo);
2242 				return (EAGAIN);
2243 			}
2244 			bremfree(bp);
2245 			bp->b_flags |= B_INVAL | B_RELBUF;
2246 			bp->b_flags &= ~B_ASYNC;
2247 			brelse(bp);
2248 			anyfreed = true;
2249 
2250 			BO_LOCK(bo);
2251 			if (nbp != NULL &&
2252 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2253 			    (nbp->b_vp != vp) ||
2254 			    (nbp->b_flags & B_DELWRI) == 0))
2255 				return (EAGAIN);
2256 		}
2257 	} while (anyfreed);
2258 	return (0);
2259 }
2260 
2261 static void
2262 buf_vlist_remove(struct buf *bp)
2263 {
2264 	struct bufv *bv;
2265 
2266 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2267 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2268 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
2269 	    (BX_VNDIRTY|BX_VNCLEAN),
2270 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
2271 	if (bp->b_xflags & BX_VNDIRTY)
2272 		bv = &bp->b_bufobj->bo_dirty;
2273 	else
2274 		bv = &bp->b_bufobj->bo_clean;
2275 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2276 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2277 	bv->bv_cnt--;
2278 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2279 }
2280 
2281 /*
2282  * Add the buffer to the sorted clean or dirty block list.
2283  *
2284  * NOTE: xflags is passed as a constant, optimizing this inline function!
2285  */
2286 static void
2287 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2288 {
2289 	struct bufv *bv;
2290 	struct buf *n;
2291 	int error;
2292 
2293 	ASSERT_BO_WLOCKED(bo);
2294 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2295 	    ("dead bo %p", bo));
2296 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2297 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2298 	bp->b_xflags |= xflags;
2299 	if (xflags & BX_VNDIRTY)
2300 		bv = &bo->bo_dirty;
2301 	else
2302 		bv = &bo->bo_clean;
2303 
2304 	/*
2305 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2306 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2307 	 * than _ge.
2308 	 */
2309 	if (bv->bv_cnt == 0 ||
2310 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2311 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2312 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2313 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2314 	else
2315 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2316 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2317 	if (error)
2318 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2319 	bv->bv_cnt++;
2320 }
2321 
2322 /*
2323  * Look up a buffer using the buffer tries.
2324  */
2325 struct buf *
2326 gbincore(struct bufobj *bo, daddr_t lblkno)
2327 {
2328 	struct buf *bp;
2329 
2330 	ASSERT_BO_LOCKED(bo);
2331 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2332 	if (bp != NULL)
2333 		return (bp);
2334 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
2335 }
2336 
2337 /*
2338  * Associate a buffer with a vnode.
2339  */
2340 void
2341 bgetvp(struct vnode *vp, struct buf *bp)
2342 {
2343 	struct bufobj *bo;
2344 
2345 	bo = &vp->v_bufobj;
2346 	ASSERT_BO_WLOCKED(bo);
2347 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2348 
2349 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2350 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2351 	    ("bgetvp: bp already attached! %p", bp));
2352 
2353 	vhold(vp);
2354 	bp->b_vp = vp;
2355 	bp->b_bufobj = bo;
2356 	/*
2357 	 * Insert onto list for new vnode.
2358 	 */
2359 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2360 }
2361 
2362 /*
2363  * Disassociate a buffer from a vnode.
2364  */
2365 void
2366 brelvp(struct buf *bp)
2367 {
2368 	struct bufobj *bo;
2369 	struct vnode *vp;
2370 
2371 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2372 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2373 
2374 	/*
2375 	 * Delete from old vnode list, if on one.
2376 	 */
2377 	vp = bp->b_vp;		/* XXX */
2378 	bo = bp->b_bufobj;
2379 	BO_LOCK(bo);
2380 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2381 		buf_vlist_remove(bp);
2382 	else
2383 		panic("brelvp: Buffer %p not on queue.", bp);
2384 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2385 		bo->bo_flag &= ~BO_ONWORKLST;
2386 		mtx_lock(&sync_mtx);
2387 		LIST_REMOVE(bo, bo_synclist);
2388 		syncer_worklist_len--;
2389 		mtx_unlock(&sync_mtx);
2390 	}
2391 	bp->b_vp = NULL;
2392 	bp->b_bufobj = NULL;
2393 	BO_UNLOCK(bo);
2394 	vdrop(vp);
2395 }
2396 
2397 /*
2398  * Add an item to the syncer work queue.
2399  */
2400 static void
2401 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2402 {
2403 	int slot;
2404 
2405 	ASSERT_BO_WLOCKED(bo);
2406 
2407 	mtx_lock(&sync_mtx);
2408 	if (bo->bo_flag & BO_ONWORKLST)
2409 		LIST_REMOVE(bo, bo_synclist);
2410 	else {
2411 		bo->bo_flag |= BO_ONWORKLST;
2412 		syncer_worklist_len++;
2413 	}
2414 
2415 	if (delay > syncer_maxdelay - 2)
2416 		delay = syncer_maxdelay - 2;
2417 	slot = (syncer_delayno + delay) & syncer_mask;
2418 
2419 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2420 	mtx_unlock(&sync_mtx);
2421 }
2422 
2423 static int
2424 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2425 {
2426 	int error, len;
2427 
2428 	mtx_lock(&sync_mtx);
2429 	len = syncer_worklist_len - sync_vnode_count;
2430 	mtx_unlock(&sync_mtx);
2431 	error = SYSCTL_OUT(req, &len, sizeof(len));
2432 	return (error);
2433 }
2434 
2435 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2436     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2437     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2438 
2439 static struct proc *updateproc;
2440 static void sched_sync(void);
2441 static struct kproc_desc up_kp = {
2442 	"syncer",
2443 	sched_sync,
2444 	&updateproc
2445 };
2446 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2447 
2448 static int
2449 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2450 {
2451 	struct vnode *vp;
2452 	struct mount *mp;
2453 
2454 	*bo = LIST_FIRST(slp);
2455 	if (*bo == NULL)
2456 		return (0);
2457 	vp = bo2vnode(*bo);
2458 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2459 		return (1);
2460 	/*
2461 	 * We use vhold in case the vnode does not
2462 	 * successfully sync.  vhold prevents the vnode from
2463 	 * going away when we unlock the sync_mtx so that
2464 	 * we can acquire the vnode interlock.
2465 	 */
2466 	vholdl(vp);
2467 	mtx_unlock(&sync_mtx);
2468 	VI_UNLOCK(vp);
2469 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2470 		vdrop(vp);
2471 		mtx_lock(&sync_mtx);
2472 		return (*bo == LIST_FIRST(slp));
2473 	}
2474 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2475 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2476 	VOP_UNLOCK(vp);
2477 	vn_finished_write(mp);
2478 	BO_LOCK(*bo);
2479 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2480 		/*
2481 		 * Put us back on the worklist.  The worklist
2482 		 * routine will remove us from our current
2483 		 * position and then add us back in at a later
2484 		 * position.
2485 		 */
2486 		vn_syncer_add_to_worklist(*bo, syncdelay);
2487 	}
2488 	BO_UNLOCK(*bo);
2489 	vdrop(vp);
2490 	mtx_lock(&sync_mtx);
2491 	return (0);
2492 }
2493 
2494 static int first_printf = 1;
2495 
2496 /*
2497  * System filesystem synchronizer daemon.
2498  */
2499 static void
2500 sched_sync(void)
2501 {
2502 	struct synclist *next, *slp;
2503 	struct bufobj *bo;
2504 	long starttime;
2505 	struct thread *td = curthread;
2506 	int last_work_seen;
2507 	int net_worklist_len;
2508 	int syncer_final_iter;
2509 	int error;
2510 
2511 	last_work_seen = 0;
2512 	syncer_final_iter = 0;
2513 	syncer_state = SYNCER_RUNNING;
2514 	starttime = time_uptime;
2515 	td->td_pflags |= TDP_NORUNNINGBUF;
2516 
2517 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2518 	    SHUTDOWN_PRI_LAST);
2519 
2520 	mtx_lock(&sync_mtx);
2521 	for (;;) {
2522 		if (syncer_state == SYNCER_FINAL_DELAY &&
2523 		    syncer_final_iter == 0) {
2524 			mtx_unlock(&sync_mtx);
2525 			kproc_suspend_check(td->td_proc);
2526 			mtx_lock(&sync_mtx);
2527 		}
2528 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2529 		if (syncer_state != SYNCER_RUNNING &&
2530 		    starttime != time_uptime) {
2531 			if (first_printf) {
2532 				printf("\nSyncing disks, vnodes remaining... ");
2533 				first_printf = 0;
2534 			}
2535 			printf("%d ", net_worklist_len);
2536 		}
2537 		starttime = time_uptime;
2538 
2539 		/*
2540 		 * Push files whose dirty time has expired.  Be careful
2541 		 * of interrupt race on slp queue.
2542 		 *
2543 		 * Skip over empty worklist slots when shutting down.
2544 		 */
2545 		do {
2546 			slp = &syncer_workitem_pending[syncer_delayno];
2547 			syncer_delayno += 1;
2548 			if (syncer_delayno == syncer_maxdelay)
2549 				syncer_delayno = 0;
2550 			next = &syncer_workitem_pending[syncer_delayno];
2551 			/*
2552 			 * If the worklist has wrapped since the
2553 			 * it was emptied of all but syncer vnodes,
2554 			 * switch to the FINAL_DELAY state and run
2555 			 * for one more second.
2556 			 */
2557 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2558 			    net_worklist_len == 0 &&
2559 			    last_work_seen == syncer_delayno) {
2560 				syncer_state = SYNCER_FINAL_DELAY;
2561 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2562 			}
2563 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2564 		    syncer_worklist_len > 0);
2565 
2566 		/*
2567 		 * Keep track of the last time there was anything
2568 		 * on the worklist other than syncer vnodes.
2569 		 * Return to the SHUTTING_DOWN state if any
2570 		 * new work appears.
2571 		 */
2572 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2573 			last_work_seen = syncer_delayno;
2574 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2575 			syncer_state = SYNCER_SHUTTING_DOWN;
2576 		while (!LIST_EMPTY(slp)) {
2577 			error = sync_vnode(slp, &bo, td);
2578 			if (error == 1) {
2579 				LIST_REMOVE(bo, bo_synclist);
2580 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2581 				continue;
2582 			}
2583 
2584 			if (first_printf == 0) {
2585 				/*
2586 				 * Drop the sync mutex, because some watchdog
2587 				 * drivers need to sleep while patting
2588 				 */
2589 				mtx_unlock(&sync_mtx);
2590 				wdog_kern_pat(WD_LASTVAL);
2591 				mtx_lock(&sync_mtx);
2592 			}
2593 
2594 		}
2595 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2596 			syncer_final_iter--;
2597 		/*
2598 		 * The variable rushjob allows the kernel to speed up the
2599 		 * processing of the filesystem syncer process. A rushjob
2600 		 * value of N tells the filesystem syncer to process the next
2601 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2602 		 * is used by the soft update code to speed up the filesystem
2603 		 * syncer process when the incore state is getting so far
2604 		 * ahead of the disk that the kernel memory pool is being
2605 		 * threatened with exhaustion.
2606 		 */
2607 		if (rushjob > 0) {
2608 			rushjob -= 1;
2609 			continue;
2610 		}
2611 		/*
2612 		 * Just sleep for a short period of time between
2613 		 * iterations when shutting down to allow some I/O
2614 		 * to happen.
2615 		 *
2616 		 * If it has taken us less than a second to process the
2617 		 * current work, then wait. Otherwise start right over
2618 		 * again. We can still lose time if any single round
2619 		 * takes more than two seconds, but it does not really
2620 		 * matter as we are just trying to generally pace the
2621 		 * filesystem activity.
2622 		 */
2623 		if (syncer_state != SYNCER_RUNNING ||
2624 		    time_uptime == starttime) {
2625 			thread_lock(td);
2626 			sched_prio(td, PPAUSE);
2627 			thread_unlock(td);
2628 		}
2629 		if (syncer_state != SYNCER_RUNNING)
2630 			cv_timedwait(&sync_wakeup, &sync_mtx,
2631 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2632 		else if (time_uptime == starttime)
2633 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2634 	}
2635 }
2636 
2637 /*
2638  * Request the syncer daemon to speed up its work.
2639  * We never push it to speed up more than half of its
2640  * normal turn time, otherwise it could take over the cpu.
2641  */
2642 int
2643 speedup_syncer(void)
2644 {
2645 	int ret = 0;
2646 
2647 	mtx_lock(&sync_mtx);
2648 	if (rushjob < syncdelay / 2) {
2649 		rushjob += 1;
2650 		stat_rush_requests += 1;
2651 		ret = 1;
2652 	}
2653 	mtx_unlock(&sync_mtx);
2654 	cv_broadcast(&sync_wakeup);
2655 	return (ret);
2656 }
2657 
2658 /*
2659  * Tell the syncer to speed up its work and run though its work
2660  * list several times, then tell it to shut down.
2661  */
2662 static void
2663 syncer_shutdown(void *arg, int howto)
2664 {
2665 
2666 	if (howto & RB_NOSYNC)
2667 		return;
2668 	mtx_lock(&sync_mtx);
2669 	syncer_state = SYNCER_SHUTTING_DOWN;
2670 	rushjob = 0;
2671 	mtx_unlock(&sync_mtx);
2672 	cv_broadcast(&sync_wakeup);
2673 	kproc_shutdown(arg, howto);
2674 }
2675 
2676 void
2677 syncer_suspend(void)
2678 {
2679 
2680 	syncer_shutdown(updateproc, 0);
2681 }
2682 
2683 void
2684 syncer_resume(void)
2685 {
2686 
2687 	mtx_lock(&sync_mtx);
2688 	first_printf = 1;
2689 	syncer_state = SYNCER_RUNNING;
2690 	mtx_unlock(&sync_mtx);
2691 	cv_broadcast(&sync_wakeup);
2692 	kproc_resume(updateproc);
2693 }
2694 
2695 /*
2696  * Reassign a buffer from one vnode to another.
2697  * Used to assign file specific control information
2698  * (indirect blocks) to the vnode to which they belong.
2699  */
2700 void
2701 reassignbuf(struct buf *bp)
2702 {
2703 	struct vnode *vp;
2704 	struct bufobj *bo;
2705 	int delay;
2706 #ifdef INVARIANTS
2707 	struct bufv *bv;
2708 #endif
2709 
2710 	vp = bp->b_vp;
2711 	bo = bp->b_bufobj;
2712 	++reassignbufcalls;
2713 
2714 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2715 	    bp, bp->b_vp, bp->b_flags);
2716 	/*
2717 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2718 	 * is not fully linked in.
2719 	 */
2720 	if (bp->b_flags & B_PAGING)
2721 		panic("cannot reassign paging buffer");
2722 
2723 	/*
2724 	 * Delete from old vnode list, if on one.
2725 	 */
2726 	BO_LOCK(bo);
2727 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2728 		buf_vlist_remove(bp);
2729 	else
2730 		panic("reassignbuf: Buffer %p not on queue.", bp);
2731 	/*
2732 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2733 	 * of clean buffers.
2734 	 */
2735 	if (bp->b_flags & B_DELWRI) {
2736 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2737 			switch (vp->v_type) {
2738 			case VDIR:
2739 				delay = dirdelay;
2740 				break;
2741 			case VCHR:
2742 				delay = metadelay;
2743 				break;
2744 			default:
2745 				delay = filedelay;
2746 			}
2747 			vn_syncer_add_to_worklist(bo, delay);
2748 		}
2749 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2750 	} else {
2751 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2752 
2753 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2754 			mtx_lock(&sync_mtx);
2755 			LIST_REMOVE(bo, bo_synclist);
2756 			syncer_worklist_len--;
2757 			mtx_unlock(&sync_mtx);
2758 			bo->bo_flag &= ~BO_ONWORKLST;
2759 		}
2760 	}
2761 #ifdef INVARIANTS
2762 	bv = &bo->bo_clean;
2763 	bp = TAILQ_FIRST(&bv->bv_hd);
2764 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2765 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2766 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2767 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2768 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2769 	bv = &bo->bo_dirty;
2770 	bp = TAILQ_FIRST(&bv->bv_hd);
2771 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2772 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2773 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2774 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2775 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2776 #endif
2777 	BO_UNLOCK(bo);
2778 }
2779 
2780 static void
2781 v_init_counters(struct vnode *vp)
2782 {
2783 
2784 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2785 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2786 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2787 
2788 	refcount_init(&vp->v_holdcnt, 1);
2789 	refcount_init(&vp->v_usecount, 1);
2790 }
2791 
2792 /*
2793  * Increment si_usecount of the associated device, if any.
2794  */
2795 static void
2796 v_incr_devcount(struct vnode *vp)
2797 {
2798 
2799 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2800 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2801 		dev_lock();
2802 		vp->v_rdev->si_usecount++;
2803 		dev_unlock();
2804 	}
2805 }
2806 
2807 /*
2808  * Decrement si_usecount of the associated device, if any.
2809  */
2810 static void
2811 v_decr_devcount(struct vnode *vp)
2812 {
2813 
2814 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2815 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2816 		dev_lock();
2817 		vp->v_rdev->si_usecount--;
2818 		dev_unlock();
2819 	}
2820 }
2821 
2822 /*
2823  * Grab a particular vnode from the free list, increment its
2824  * reference count and lock it.  VIRF_DOOMED is set if the vnode
2825  * is being destroyed.  Only callers who specify LK_RETRY will
2826  * see doomed vnodes.  If inactive processing was delayed in
2827  * vput try to do it here.
2828  *
2829  * usecount is manipulated using atomics without holding any locks.
2830  *
2831  * holdcnt can be manipulated using atomics without holding any locks,
2832  * except when transitioning 1<->0, in which case the interlock is held.
2833  */
2834 enum vgetstate
2835 vget_prep(struct vnode *vp)
2836 {
2837 	enum vgetstate vs;
2838 
2839 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2840 		vs = VGET_USECOUNT;
2841 	} else {
2842 		vhold(vp);
2843 		vs = VGET_HOLDCNT;
2844 	}
2845 	return (vs);
2846 }
2847 
2848 int
2849 vget(struct vnode *vp, int flags, struct thread *td)
2850 {
2851 	enum vgetstate vs;
2852 
2853 	MPASS(td == curthread);
2854 
2855 	vs = vget_prep(vp);
2856 	return (vget_finish(vp, flags, vs));
2857 }
2858 
2859 static int __noinline
2860 vget_finish_vchr(struct vnode *vp)
2861 {
2862 
2863 	VNASSERT(vp->v_type == VCHR, vp, ("type != VCHR)"));
2864 
2865 	/*
2866 	 * See the comment in vget_finish before usecount bump.
2867 	 */
2868 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2869 #ifdef INVARIANTS
2870 		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2871 		VNASSERT(old > 0, vp, ("%s: wrong hold count %d", __func__, old));
2872 #else
2873 		refcount_release(&vp->v_holdcnt);
2874 #endif
2875 		return (0);
2876 	}
2877 
2878 	VI_LOCK(vp);
2879 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2880 #ifdef INVARIANTS
2881 		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2882 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2883 #else
2884 		refcount_release(&vp->v_holdcnt);
2885 #endif
2886 		VI_UNLOCK(vp);
2887 		return (0);
2888 	}
2889 	v_incr_devcount(vp);
2890 	refcount_acquire(&vp->v_usecount);
2891 	VI_UNLOCK(vp);
2892 	return (0);
2893 }
2894 
2895 int
2896 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2897 {
2898 	int error, old;
2899 
2900 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2901 	    ("%s: invalid lock operation", __func__));
2902 
2903 	if ((flags & LK_INTERLOCK) != 0)
2904 		ASSERT_VI_LOCKED(vp, __func__);
2905 	else
2906 		ASSERT_VI_UNLOCKED(vp, __func__);
2907 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
2908 	if (vs == VGET_USECOUNT) {
2909 		VNASSERT(vp->v_usecount > 0, vp,
2910 		    ("%s: vnode without usecount when VGET_USECOUNT was passed",
2911 		    __func__));
2912 	}
2913 
2914 	if ((error = vn_lock(vp, flags)) != 0) {
2915 		if (vs == VGET_USECOUNT)
2916 			vrele(vp);
2917 		else
2918 			vdrop(vp);
2919 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2920 		    vp);
2921 		return (error);
2922 	}
2923 
2924 	if (vs == VGET_USECOUNT) {
2925 		return (0);
2926 	}
2927 
2928 	if (__predict_false(vp->v_type == VCHR))
2929 		return (vget_finish_vchr(vp));
2930 
2931 	/*
2932 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
2933 	 * the vnode around. Otherwise someone else lended their hold count and
2934 	 * we have to drop ours.
2935 	 */
2936 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
2937 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
2938 	if (old != 0) {
2939 #ifdef INVARIANTS
2940 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2941 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2942 #else
2943 		refcount_release(&vp->v_holdcnt);
2944 #endif
2945 	}
2946 	return (0);
2947 }
2948 
2949 /*
2950  * Increase the reference (use) and hold count of a vnode.
2951  * This will also remove the vnode from the free list if it is presently free.
2952  */
2953 static void __noinline
2954 vref_vchr(struct vnode *vp, bool interlock)
2955 {
2956 
2957 	/*
2958 	 * See the comment in vget_finish before usecount bump.
2959 	 */
2960 	if (!interlock) {
2961 		if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2962 			VNODE_REFCOUNT_FENCE_ACQ();
2963 			VNASSERT(vp->v_holdcnt > 0, vp,
2964 			    ("%s: active vnode not held", __func__));
2965 			return;
2966 		}
2967 		VI_LOCK(vp);
2968 		/*
2969 		 * By the time we get here the vnode might have been doomed, at
2970 		 * which point the 0->1 use count transition is no longer
2971 		 * protected by the interlock. Since it can't bounce back to
2972 		 * VCHR and requires vref semantics, punt it back
2973 		 */
2974 		if (__predict_false(vp->v_type == VBAD)) {
2975 			VI_UNLOCK(vp);
2976 			vref(vp);
2977 			return;
2978 		}
2979 	}
2980 	VNASSERT(vp->v_type == VCHR, vp, ("type != VCHR)"));
2981 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2982 		VNODE_REFCOUNT_FENCE_ACQ();
2983 		VNASSERT(vp->v_holdcnt > 0, vp,
2984 		    ("%s: active vnode not held", __func__));
2985 		if (!interlock)
2986 			VI_UNLOCK(vp);
2987 		return;
2988 	}
2989 	vhold(vp);
2990 	v_incr_devcount(vp);
2991 	refcount_acquire(&vp->v_usecount);
2992 	if (!interlock)
2993 		VI_UNLOCK(vp);
2994 	return;
2995 }
2996 
2997 void
2998 vref(struct vnode *vp)
2999 {
3000 	int old;
3001 
3002 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3003 	if (__predict_false(vp->v_type == VCHR)) {
3004 		 vref_vchr(vp, false);
3005 		 return;
3006 	}
3007 
3008 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
3009 		VNODE_REFCOUNT_FENCE_ACQ();
3010 		VNASSERT(vp->v_holdcnt > 0, vp,
3011 		    ("%s: active vnode not held", __func__));
3012 		return;
3013 	}
3014 	vhold(vp);
3015 	/*
3016 	 * See the comment in vget_finish.
3017 	 */
3018 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
3019 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3020 	if (old != 0) {
3021 #ifdef INVARIANTS
3022 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3023 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3024 #else
3025 		refcount_release(&vp->v_holdcnt);
3026 #endif
3027 	}
3028 }
3029 
3030 void
3031 vrefl(struct vnode *vp)
3032 {
3033 
3034 	ASSERT_VI_LOCKED(vp, __func__);
3035 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3036 	if (__predict_false(vp->v_type == VCHR)) {
3037 		vref_vchr(vp, true);
3038 		return;
3039 	}
3040 	vref(vp);
3041 }
3042 
3043 void
3044 vrefact(struct vnode *vp)
3045 {
3046 
3047 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3048 #ifdef INVARIANTS
3049 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3050 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3051 #else
3052 	refcount_acquire(&vp->v_usecount);
3053 #endif
3054 }
3055 
3056 /*
3057  * Return reference count of a vnode.
3058  *
3059  * The results of this call are only guaranteed when some mechanism is used to
3060  * stop other processes from gaining references to the vnode.  This may be the
3061  * case if the caller holds the only reference.  This is also useful when stale
3062  * data is acceptable as race conditions may be accounted for by some other
3063  * means.
3064  */
3065 int
3066 vrefcnt(struct vnode *vp)
3067 {
3068 
3069 	return (vp->v_usecount);
3070 }
3071 
3072 void
3073 vlazy(struct vnode *vp)
3074 {
3075 	struct mount *mp;
3076 
3077 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3078 
3079 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3080 		return;
3081 	mp = vp->v_mount;
3082 	mtx_lock(&mp->mnt_listmtx);
3083 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3084 		vp->v_mflag |= VMP_LAZYLIST;
3085 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3086 		mp->mnt_lazyvnodelistsize++;
3087 	}
3088 	mtx_unlock(&mp->mnt_listmtx);
3089 }
3090 
3091 static void
3092 vdefer_inactive(struct vnode *vp)
3093 {
3094 
3095 	ASSERT_VI_LOCKED(vp, __func__);
3096 	VNASSERT(vp->v_holdcnt > 0, vp,
3097 	    ("%s: vnode without hold count", __func__));
3098 	if (VN_IS_DOOMED(vp)) {
3099 		vdropl(vp);
3100 		return;
3101 	}
3102 	if (vp->v_iflag & VI_DEFINACT) {
3103 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3104 		vdropl(vp);
3105 		return;
3106 	}
3107 	if (vp->v_usecount > 0) {
3108 		vp->v_iflag &= ~VI_OWEINACT;
3109 		vdropl(vp);
3110 		return;
3111 	}
3112 	vlazy(vp);
3113 	vp->v_iflag |= VI_DEFINACT;
3114 	VI_UNLOCK(vp);
3115 	counter_u64_add(deferred_inact, 1);
3116 }
3117 
3118 static void
3119 vdefer_inactive_unlocked(struct vnode *vp)
3120 {
3121 
3122 	VI_LOCK(vp);
3123 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3124 		vdropl(vp);
3125 		return;
3126 	}
3127 	vdefer_inactive(vp);
3128 }
3129 
3130 enum vputx_op { VPUTX_VRELE, VPUTX_VPUT, VPUTX_VUNREF };
3131 
3132 /*
3133  * Decrement the use and hold counts for a vnode.
3134  *
3135  * See an explanation near vget() as to why atomic operation is safe.
3136  *
3137  * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3138  * on the lock being held all the way until VOP_INACTIVE. This in particular
3139  * happens with UFS which adds half-constructed vnodes to the hash, where they
3140  * can be found by other code.
3141  */
3142 static void
3143 vputx(struct vnode *vp, enum vputx_op func)
3144 {
3145 	int error;
3146 
3147 	KASSERT(vp != NULL, ("vputx: null vp"));
3148 	if (func == VPUTX_VUNREF)
3149 		ASSERT_VOP_LOCKED(vp, "vunref");
3150 	else if (func == VPUTX_VPUT)
3151 		ASSERT_VOP_LOCKED(vp, "vput");
3152 	ASSERT_VI_UNLOCKED(vp, __func__);
3153 	VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp,
3154 	    ("%s: wrong ref counts", __func__));
3155 
3156 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3157 
3158 	/*
3159 	 * We want to hold the vnode until the inactive finishes to
3160 	 * prevent vgone() races.  We drop the use count here and the
3161 	 * hold count below when we're done.
3162 	 *
3163 	 * If we release the last usecount we take ownership of the hold
3164 	 * count which provides liveness of the vnode, in which case we
3165 	 * have to vdrop.
3166 	 */
3167 	if (!refcount_release(&vp->v_usecount)) {
3168 		if (func == VPUTX_VPUT)
3169 			VOP_UNLOCK(vp);
3170 		return;
3171 	}
3172 	VI_LOCK(vp);
3173 	v_decr_devcount(vp);
3174 	/*
3175 	 * By the time we got here someone else might have transitioned
3176 	 * the count back to > 0.
3177 	 */
3178 	if (vp->v_usecount > 0 || vp->v_iflag & VI_DOINGINACT)
3179 		goto out;
3180 
3181 	/*
3182 	 * Check if the fs wants to perform inactive processing. Note we
3183 	 * may be only holding the interlock, in which case it is possible
3184 	 * someone else called vgone on the vnode and ->v_data is now NULL.
3185 	 * Since vgone performs inactive on its own there is nothing to do
3186 	 * here but to drop our hold count.
3187 	 */
3188 	if (__predict_false(VN_IS_DOOMED(vp)) ||
3189 	    VOP_NEED_INACTIVE(vp) == 0)
3190 		goto out;
3191 
3192 	/*
3193 	 * We must call VOP_INACTIVE with the node locked. Mark
3194 	 * as VI_DOINGINACT to avoid recursion.
3195 	 */
3196 	vp->v_iflag |= VI_OWEINACT;
3197 	switch (func) {
3198 	case VPUTX_VRELE:
3199 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3200 		VI_LOCK(vp);
3201 		break;
3202 	case VPUTX_VPUT:
3203 		error = 0;
3204 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3205 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3206 			    LK_NOWAIT);
3207 			VI_LOCK(vp);
3208 		}
3209 		break;
3210 	case VPUTX_VUNREF:
3211 		error = 0;
3212 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3213 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3214 			VI_LOCK(vp);
3215 		}
3216 		break;
3217 	}
3218 	if (error == 0) {
3219 		vinactive(vp);
3220 		if (func != VPUTX_VUNREF)
3221 			VOP_UNLOCK(vp);
3222 		vdropl(vp);
3223 	} else {
3224 		vdefer_inactive(vp);
3225 	}
3226 	return;
3227 out:
3228 	if (func == VPUTX_VPUT)
3229 		VOP_UNLOCK(vp);
3230 	vdropl(vp);
3231 }
3232 
3233 /*
3234  * Vnode put/release.
3235  * If count drops to zero, call inactive routine and return to freelist.
3236  */
3237 void
3238 vrele(struct vnode *vp)
3239 {
3240 
3241 	vputx(vp, VPUTX_VRELE);
3242 }
3243 
3244 /*
3245  * Release an already locked vnode.  This give the same effects as
3246  * unlock+vrele(), but takes less time and avoids releasing and
3247  * re-aquiring the lock (as vrele() acquires the lock internally.)
3248  */
3249 void
3250 vput(struct vnode *vp)
3251 {
3252 
3253 	vputx(vp, VPUTX_VPUT);
3254 }
3255 
3256 /*
3257  * Release an exclusively locked vnode. Do not unlock the vnode lock.
3258  */
3259 void
3260 vunref(struct vnode *vp)
3261 {
3262 
3263 	vputx(vp, VPUTX_VUNREF);
3264 }
3265 
3266 void
3267 vhold(struct vnode *vp)
3268 {
3269 	struct vdbatch *vd;
3270 	int old;
3271 
3272 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3273 	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3274 	VNASSERT(old >= 0, vp, ("%s: wrong hold count %d", __func__, old));
3275 	if (old != 0)
3276 		return;
3277 	critical_enter();
3278 	vd = DPCPU_PTR(vd);
3279 	vd->freevnodes--;
3280 	critical_exit();
3281 }
3282 
3283 void
3284 vholdl(struct vnode *vp)
3285 {
3286 
3287 	ASSERT_VI_LOCKED(vp, __func__);
3288 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3289 	vhold(vp);
3290 }
3291 
3292 void
3293 vholdnz(struct vnode *vp)
3294 {
3295 
3296 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3297 #ifdef INVARIANTS
3298 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3299 	VNASSERT(old > 0, vp, ("%s: wrong hold count %d", __func__, old));
3300 #else
3301 	atomic_add_int(&vp->v_holdcnt, 1);
3302 #endif
3303 }
3304 
3305 static void __noinline
3306 vdbatch_process(struct vdbatch *vd)
3307 {
3308 	struct vnode *vp;
3309 	int i;
3310 
3311 	mtx_assert(&vd->lock, MA_OWNED);
3312 	MPASS(curthread->td_pinned > 0);
3313 	MPASS(vd->index == VDBATCH_SIZE);
3314 
3315 	mtx_lock(&vnode_list_mtx);
3316 	critical_enter();
3317 	freevnodes += vd->freevnodes;
3318 	for (i = 0; i < VDBATCH_SIZE; i++) {
3319 		vp = vd->tab[i];
3320 		TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3321 		TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3322 		MPASS(vp->v_dbatchcpu != NOCPU);
3323 		vp->v_dbatchcpu = NOCPU;
3324 	}
3325 	mtx_unlock(&vnode_list_mtx);
3326 	critical_exit();
3327 	vd->freevnodes = 0;
3328 	bzero(vd->tab, sizeof(vd->tab));
3329 	vd->index = 0;
3330 }
3331 
3332 static void
3333 vdbatch_enqueue(struct vnode *vp)
3334 {
3335 	struct vdbatch *vd;
3336 
3337 	ASSERT_VI_LOCKED(vp, __func__);
3338 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3339 	    ("%s: deferring requeue of a doomed vnode", __func__));
3340 
3341 	critical_enter();
3342 	vd = DPCPU_PTR(vd);
3343 	vd->freevnodes++;
3344 	if (vp->v_dbatchcpu != NOCPU) {
3345 		VI_UNLOCK(vp);
3346 		critical_exit();
3347 		return;
3348 	}
3349 
3350 	sched_pin();
3351 	critical_exit();
3352 	mtx_lock(&vd->lock);
3353 	MPASS(vd->index < VDBATCH_SIZE);
3354 	MPASS(vd->tab[vd->index] == NULL);
3355 	/*
3356 	 * A hack: we depend on being pinned so that we know what to put in
3357 	 * ->v_dbatchcpu.
3358 	 */
3359 	vp->v_dbatchcpu = curcpu;
3360 	vd->tab[vd->index] = vp;
3361 	vd->index++;
3362 	VI_UNLOCK(vp);
3363 	if (vd->index == VDBATCH_SIZE)
3364 		vdbatch_process(vd);
3365 	mtx_unlock(&vd->lock);
3366 	sched_unpin();
3367 }
3368 
3369 /*
3370  * This routine must only be called for vnodes which are about to be
3371  * deallocated. Supporting dequeue for arbitrary vndoes would require
3372  * validating that the locked batch matches.
3373  */
3374 static void
3375 vdbatch_dequeue(struct vnode *vp)
3376 {
3377 	struct vdbatch *vd;
3378 	int i;
3379 	short cpu;
3380 
3381 	VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3382 	    ("%s: called for a used vnode\n", __func__));
3383 
3384 	cpu = vp->v_dbatchcpu;
3385 	if (cpu == NOCPU)
3386 		return;
3387 
3388 	vd = DPCPU_ID_PTR(cpu, vd);
3389 	mtx_lock(&vd->lock);
3390 	for (i = 0; i < vd->index; i++) {
3391 		if (vd->tab[i] != vp)
3392 			continue;
3393 		vp->v_dbatchcpu = NOCPU;
3394 		vd->index--;
3395 		vd->tab[i] = vd->tab[vd->index];
3396 		vd->tab[vd->index] = NULL;
3397 		break;
3398 	}
3399 	mtx_unlock(&vd->lock);
3400 	/*
3401 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3402 	 */
3403 	MPASS(vp->v_dbatchcpu == NOCPU);
3404 }
3405 
3406 /*
3407  * Drop the hold count of the vnode.  If this is the last reference to
3408  * the vnode we place it on the free list unless it has been vgone'd
3409  * (marked VIRF_DOOMED) in which case we will free it.
3410  *
3411  * Because the vnode vm object keeps a hold reference on the vnode if
3412  * there is at least one resident non-cached page, the vnode cannot
3413  * leave the active list without the page cleanup done.
3414  */
3415 static void
3416 vdrop_deactivate(struct vnode *vp)
3417 {
3418 	struct mount *mp;
3419 
3420 	ASSERT_VI_LOCKED(vp, __func__);
3421 	/*
3422 	 * Mark a vnode as free: remove it from its active list
3423 	 * and put it up for recycling on the freelist.
3424 	 */
3425 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3426 	    ("vdrop: returning doomed vnode"));
3427 	VNASSERT(vp->v_op != NULL, vp,
3428 	    ("vdrop: vnode already reclaimed."));
3429 	VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
3430 	    ("vnode with VI_OWEINACT set"));
3431 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp,
3432 	    ("vnode with VI_DEFINACT set"));
3433 	if (vp->v_mflag & VMP_LAZYLIST) {
3434 		mp = vp->v_mount;
3435 		mtx_lock(&mp->mnt_listmtx);
3436 		VNASSERT(vp->v_mflag & VMP_LAZYLIST, vp, ("lost VMP_LAZYLIST"));
3437 		/*
3438 		 * Don't remove the vnode from the lazy list if another thread
3439 		 * has increased the hold count. It may have re-enqueued the
3440 		 * vnode to the lazy list and is now responsible for its
3441 		 * removal.
3442 		 */
3443 		if (vp->v_holdcnt == 0) {
3444 			vp->v_mflag &= ~VMP_LAZYLIST;
3445 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3446 			mp->mnt_lazyvnodelistsize--;
3447 		}
3448 		mtx_unlock(&mp->mnt_listmtx);
3449 	}
3450 	vdbatch_enqueue(vp);
3451 }
3452 
3453 void
3454 vdrop(struct vnode *vp)
3455 {
3456 
3457 	ASSERT_VI_UNLOCKED(vp, __func__);
3458 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3459 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3460 		return;
3461 	VI_LOCK(vp);
3462 	vdropl(vp);
3463 }
3464 
3465 void
3466 vdropl(struct vnode *vp)
3467 {
3468 
3469 	ASSERT_VI_LOCKED(vp, __func__);
3470 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3471 	if (!refcount_release(&vp->v_holdcnt)) {
3472 		VI_UNLOCK(vp);
3473 		return;
3474 	}
3475 	if (VN_IS_DOOMED(vp)) {
3476 		freevnode(vp);
3477 		return;
3478 	}
3479 	vdrop_deactivate(vp);
3480 }
3481 
3482 /*
3483  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3484  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3485  */
3486 static void
3487 vinactivef(struct vnode *vp)
3488 {
3489 	struct vm_object *obj;
3490 
3491 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3492 	ASSERT_VI_LOCKED(vp, "vinactive");
3493 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3494 	    ("vinactive: recursed on VI_DOINGINACT"));
3495 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3496 	vp->v_iflag |= VI_DOINGINACT;
3497 	vp->v_iflag &= ~VI_OWEINACT;
3498 	VI_UNLOCK(vp);
3499 	/*
3500 	 * Before moving off the active list, we must be sure that any
3501 	 * modified pages are converted into the vnode's dirty
3502 	 * buffers, since these will no longer be checked once the
3503 	 * vnode is on the inactive list.
3504 	 *
3505 	 * The write-out of the dirty pages is asynchronous.  At the
3506 	 * point that VOP_INACTIVE() is called, there could still be
3507 	 * pending I/O and dirty pages in the object.
3508 	 */
3509 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3510 	    vm_object_mightbedirty(obj)) {
3511 		VM_OBJECT_WLOCK(obj);
3512 		vm_object_page_clean(obj, 0, 0, 0);
3513 		VM_OBJECT_WUNLOCK(obj);
3514 	}
3515 	VOP_INACTIVE(vp, curthread);
3516 	VI_LOCK(vp);
3517 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3518 	    ("vinactive: lost VI_DOINGINACT"));
3519 	vp->v_iflag &= ~VI_DOINGINACT;
3520 }
3521 
3522 void
3523 vinactive(struct vnode *vp)
3524 {
3525 
3526 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3527 	ASSERT_VI_LOCKED(vp, "vinactive");
3528 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3529 
3530 	if ((vp->v_iflag & VI_OWEINACT) == 0)
3531 		return;
3532 	if (vp->v_iflag & VI_DOINGINACT)
3533 		return;
3534 	if (vp->v_usecount > 0) {
3535 		vp->v_iflag &= ~VI_OWEINACT;
3536 		return;
3537 	}
3538 	vinactivef(vp);
3539 }
3540 
3541 /*
3542  * Remove any vnodes in the vnode table belonging to mount point mp.
3543  *
3544  * If FORCECLOSE is not specified, there should not be any active ones,
3545  * return error if any are found (nb: this is a user error, not a
3546  * system error). If FORCECLOSE is specified, detach any active vnodes
3547  * that are found.
3548  *
3549  * If WRITECLOSE is set, only flush out regular file vnodes open for
3550  * writing.
3551  *
3552  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3553  *
3554  * `rootrefs' specifies the base reference count for the root vnode
3555  * of this filesystem. The root vnode is considered busy if its
3556  * v_usecount exceeds this value. On a successful return, vflush(, td)
3557  * will call vrele() on the root vnode exactly rootrefs times.
3558  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3559  * be zero.
3560  */
3561 #ifdef DIAGNOSTIC
3562 static int busyprt = 0;		/* print out busy vnodes */
3563 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3564 #endif
3565 
3566 int
3567 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3568 {
3569 	struct vnode *vp, *mvp, *rootvp = NULL;
3570 	struct vattr vattr;
3571 	int busy = 0, error;
3572 
3573 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3574 	    rootrefs, flags);
3575 	if (rootrefs > 0) {
3576 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3577 		    ("vflush: bad args"));
3578 		/*
3579 		 * Get the filesystem root vnode. We can vput() it
3580 		 * immediately, since with rootrefs > 0, it won't go away.
3581 		 */
3582 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3583 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3584 			    __func__, error);
3585 			return (error);
3586 		}
3587 		vput(rootvp);
3588 	}
3589 loop:
3590 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3591 		vholdl(vp);
3592 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3593 		if (error) {
3594 			vdrop(vp);
3595 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3596 			goto loop;
3597 		}
3598 		/*
3599 		 * Skip over a vnodes marked VV_SYSTEM.
3600 		 */
3601 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3602 			VOP_UNLOCK(vp);
3603 			vdrop(vp);
3604 			continue;
3605 		}
3606 		/*
3607 		 * If WRITECLOSE is set, flush out unlinked but still open
3608 		 * files (even if open only for reading) and regular file
3609 		 * vnodes open for writing.
3610 		 */
3611 		if (flags & WRITECLOSE) {
3612 			if (vp->v_object != NULL) {
3613 				VM_OBJECT_WLOCK(vp->v_object);
3614 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3615 				VM_OBJECT_WUNLOCK(vp->v_object);
3616 			}
3617 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3618 			if (error != 0) {
3619 				VOP_UNLOCK(vp);
3620 				vdrop(vp);
3621 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3622 				return (error);
3623 			}
3624 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3625 			VI_LOCK(vp);
3626 
3627 			if ((vp->v_type == VNON ||
3628 			    (error == 0 && vattr.va_nlink > 0)) &&
3629 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3630 				VOP_UNLOCK(vp);
3631 				vdropl(vp);
3632 				continue;
3633 			}
3634 		} else
3635 			VI_LOCK(vp);
3636 		/*
3637 		 * With v_usecount == 0, all we need to do is clear out the
3638 		 * vnode data structures and we are done.
3639 		 *
3640 		 * If FORCECLOSE is set, forcibly close the vnode.
3641 		 */
3642 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3643 			vgonel(vp);
3644 		} else {
3645 			busy++;
3646 #ifdef DIAGNOSTIC
3647 			if (busyprt)
3648 				vn_printf(vp, "vflush: busy vnode ");
3649 #endif
3650 		}
3651 		VOP_UNLOCK(vp);
3652 		vdropl(vp);
3653 	}
3654 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3655 		/*
3656 		 * If just the root vnode is busy, and if its refcount
3657 		 * is equal to `rootrefs', then go ahead and kill it.
3658 		 */
3659 		VI_LOCK(rootvp);
3660 		KASSERT(busy > 0, ("vflush: not busy"));
3661 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3662 		    ("vflush: usecount %d < rootrefs %d",
3663 		     rootvp->v_usecount, rootrefs));
3664 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3665 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3666 			vgone(rootvp);
3667 			VOP_UNLOCK(rootvp);
3668 			busy = 0;
3669 		} else
3670 			VI_UNLOCK(rootvp);
3671 	}
3672 	if (busy) {
3673 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3674 		    busy);
3675 		return (EBUSY);
3676 	}
3677 	for (; rootrefs > 0; rootrefs--)
3678 		vrele(rootvp);
3679 	return (0);
3680 }
3681 
3682 /*
3683  * Recycle an unused vnode to the front of the free list.
3684  */
3685 int
3686 vrecycle(struct vnode *vp)
3687 {
3688 	int recycled;
3689 
3690 	VI_LOCK(vp);
3691 	recycled = vrecyclel(vp);
3692 	VI_UNLOCK(vp);
3693 	return (recycled);
3694 }
3695 
3696 /*
3697  * vrecycle, with the vp interlock held.
3698  */
3699 int
3700 vrecyclel(struct vnode *vp)
3701 {
3702 	int recycled;
3703 
3704 	ASSERT_VOP_ELOCKED(vp, __func__);
3705 	ASSERT_VI_LOCKED(vp, __func__);
3706 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3707 	recycled = 0;
3708 	if (vp->v_usecount == 0) {
3709 		recycled = 1;
3710 		vgonel(vp);
3711 	}
3712 	return (recycled);
3713 }
3714 
3715 /*
3716  * Eliminate all activity associated with a vnode
3717  * in preparation for reuse.
3718  */
3719 void
3720 vgone(struct vnode *vp)
3721 {
3722 	VI_LOCK(vp);
3723 	vgonel(vp);
3724 	VI_UNLOCK(vp);
3725 }
3726 
3727 static void
3728 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3729     struct vnode *lowervp __unused)
3730 {
3731 }
3732 
3733 /*
3734  * Notify upper mounts about reclaimed or unlinked vnode.
3735  */
3736 void
3737 vfs_notify_upper(struct vnode *vp, int event)
3738 {
3739 	static struct vfsops vgonel_vfsops = {
3740 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3741 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3742 	};
3743 	struct mount *mp, *ump, *mmp;
3744 
3745 	mp = vp->v_mount;
3746 	if (mp == NULL)
3747 		return;
3748 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3749 		return;
3750 
3751 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3752 	mmp->mnt_op = &vgonel_vfsops;
3753 	mmp->mnt_kern_flag |= MNTK_MARKER;
3754 	MNT_ILOCK(mp);
3755 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3756 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3757 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3758 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3759 			continue;
3760 		}
3761 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3762 		MNT_IUNLOCK(mp);
3763 		switch (event) {
3764 		case VFS_NOTIFY_UPPER_RECLAIM:
3765 			VFS_RECLAIM_LOWERVP(ump, vp);
3766 			break;
3767 		case VFS_NOTIFY_UPPER_UNLINK:
3768 			VFS_UNLINK_LOWERVP(ump, vp);
3769 			break;
3770 		default:
3771 			KASSERT(0, ("invalid event %d", event));
3772 			break;
3773 		}
3774 		MNT_ILOCK(mp);
3775 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3776 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3777 	}
3778 	free(mmp, M_TEMP);
3779 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3780 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3781 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3782 		wakeup(&mp->mnt_uppers);
3783 	}
3784 	MNT_IUNLOCK(mp);
3785 }
3786 
3787 /*
3788  * vgone, with the vp interlock held.
3789  */
3790 static void
3791 vgonel(struct vnode *vp)
3792 {
3793 	struct thread *td;
3794 	struct mount *mp;
3795 	vm_object_t object;
3796 	bool active, oweinact;
3797 
3798 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3799 	ASSERT_VI_LOCKED(vp, "vgonel");
3800 	VNASSERT(vp->v_holdcnt, vp,
3801 	    ("vgonel: vp %p has no reference.", vp));
3802 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3803 	td = curthread;
3804 
3805 	/*
3806 	 * Don't vgonel if we're already doomed.
3807 	 */
3808 	if (vp->v_irflag & VIRF_DOOMED)
3809 		return;
3810 	vp->v_irflag |= VIRF_DOOMED;
3811 
3812 	/*
3813 	 * Check to see if the vnode is in use.  If so, we have to call
3814 	 * VOP_CLOSE() and VOP_INACTIVE().
3815 	 */
3816 	active = vp->v_usecount > 0;
3817 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3818 	/*
3819 	 * If we need to do inactive VI_OWEINACT will be set.
3820 	 */
3821 	if (vp->v_iflag & VI_DEFINACT) {
3822 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3823 		vp->v_iflag &= ~VI_DEFINACT;
3824 		vdropl(vp);
3825 	} else {
3826 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
3827 		VI_UNLOCK(vp);
3828 	}
3829 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3830 
3831 	/*
3832 	 * If purging an active vnode, it must be closed and
3833 	 * deactivated before being reclaimed.
3834 	 */
3835 	if (active)
3836 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3837 	if (oweinact || active) {
3838 		VI_LOCK(vp);
3839 		vinactivef(vp);
3840 		VI_UNLOCK(vp);
3841 	}
3842 	if (vp->v_type == VSOCK)
3843 		vfs_unp_reclaim(vp);
3844 
3845 	/*
3846 	 * Clean out any buffers associated with the vnode.
3847 	 * If the flush fails, just toss the buffers.
3848 	 */
3849 	mp = NULL;
3850 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3851 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3852 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3853 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3854 			;
3855 	}
3856 
3857 	BO_LOCK(&vp->v_bufobj);
3858 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3859 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3860 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3861 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3862 	    ("vp %p bufobj not invalidated", vp));
3863 
3864 	/*
3865 	 * For VMIO bufobj, BO_DEAD is set later, or in
3866 	 * vm_object_terminate() after the object's page queue is
3867 	 * flushed.
3868 	 */
3869 	object = vp->v_bufobj.bo_object;
3870 	if (object == NULL)
3871 		vp->v_bufobj.bo_flag |= BO_DEAD;
3872 	BO_UNLOCK(&vp->v_bufobj);
3873 
3874 	/*
3875 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
3876 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
3877 	 * should not touch the object borrowed from the lower vnode
3878 	 * (the handle check).
3879 	 */
3880 	if (object != NULL && object->type == OBJT_VNODE &&
3881 	    object->handle == vp)
3882 		vnode_destroy_vobject(vp);
3883 
3884 	/*
3885 	 * Reclaim the vnode.
3886 	 */
3887 	if (VOP_RECLAIM(vp, td))
3888 		panic("vgone: cannot reclaim");
3889 	if (mp != NULL)
3890 		vn_finished_secondary_write(mp);
3891 	VNASSERT(vp->v_object == NULL, vp,
3892 	    ("vop_reclaim left v_object vp=%p", vp));
3893 	/*
3894 	 * Clear the advisory locks and wake up waiting threads.
3895 	 */
3896 	(void)VOP_ADVLOCKPURGE(vp);
3897 	vp->v_lockf = NULL;
3898 	/*
3899 	 * Delete from old mount point vnode list.
3900 	 */
3901 	delmntque(vp);
3902 	cache_purge(vp);
3903 	/*
3904 	 * Done with purge, reset to the standard lock and invalidate
3905 	 * the vnode.
3906 	 */
3907 	VI_LOCK(vp);
3908 	vp->v_vnlock = &vp->v_lock;
3909 	vp->v_op = &dead_vnodeops;
3910 	vp->v_type = VBAD;
3911 }
3912 
3913 /*
3914  * Calculate the total number of references to a special device.
3915  */
3916 int
3917 vcount(struct vnode *vp)
3918 {
3919 	int count;
3920 
3921 	dev_lock();
3922 	count = vp->v_rdev->si_usecount;
3923 	dev_unlock();
3924 	return (count);
3925 }
3926 
3927 /*
3928  * Print out a description of a vnode.
3929  */
3930 static char *typename[] =
3931 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3932  "VMARKER"};
3933 
3934 void
3935 vn_printf(struct vnode *vp, const char *fmt, ...)
3936 {
3937 	va_list ap;
3938 	char buf[256], buf2[16];
3939 	u_long flags;
3940 
3941 	va_start(ap, fmt);
3942 	vprintf(fmt, ap);
3943 	va_end(ap);
3944 	printf("%p: ", (void *)vp);
3945 	printf("type %s\n", typename[vp->v_type]);
3946 	printf("    usecount %d, writecount %d, refcount %d",
3947 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
3948 	switch (vp->v_type) {
3949 	case VDIR:
3950 		printf(" mountedhere %p\n", vp->v_mountedhere);
3951 		break;
3952 	case VCHR:
3953 		printf(" rdev %p\n", vp->v_rdev);
3954 		break;
3955 	case VSOCK:
3956 		printf(" socket %p\n", vp->v_unpcb);
3957 		break;
3958 	case VFIFO:
3959 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
3960 		break;
3961 	default:
3962 		printf("\n");
3963 		break;
3964 	}
3965 	buf[0] = '\0';
3966 	buf[1] = '\0';
3967 	if (vp->v_irflag & VIRF_DOOMED)
3968 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
3969 	flags = vp->v_irflag & ~(VIRF_DOOMED);
3970 	if (flags != 0) {
3971 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
3972 		strlcat(buf, buf2, sizeof(buf));
3973 	}
3974 	if (vp->v_vflag & VV_ROOT)
3975 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3976 	if (vp->v_vflag & VV_ISTTY)
3977 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3978 	if (vp->v_vflag & VV_NOSYNC)
3979 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3980 	if (vp->v_vflag & VV_ETERNALDEV)
3981 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3982 	if (vp->v_vflag & VV_CACHEDLABEL)
3983 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3984 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
3985 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
3986 	if (vp->v_vflag & VV_COPYONWRITE)
3987 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3988 	if (vp->v_vflag & VV_SYSTEM)
3989 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3990 	if (vp->v_vflag & VV_PROCDEP)
3991 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3992 	if (vp->v_vflag & VV_NOKNOTE)
3993 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3994 	if (vp->v_vflag & VV_DELETED)
3995 		strlcat(buf, "|VV_DELETED", sizeof(buf));
3996 	if (vp->v_vflag & VV_MD)
3997 		strlcat(buf, "|VV_MD", sizeof(buf));
3998 	if (vp->v_vflag & VV_FORCEINSMQ)
3999 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4000 	if (vp->v_vflag & VV_READLINK)
4001 		strlcat(buf, "|VV_READLINK", sizeof(buf));
4002 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4003 	    VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
4004 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
4005 	if (flags != 0) {
4006 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4007 		strlcat(buf, buf2, sizeof(buf));
4008 	}
4009 	if (vp->v_iflag & VI_TEXT_REF)
4010 		strlcat(buf, "|VI_TEXT_REF", sizeof(buf));
4011 	if (vp->v_iflag & VI_MOUNT)
4012 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
4013 	if (vp->v_iflag & VI_DOINGINACT)
4014 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4015 	if (vp->v_iflag & VI_OWEINACT)
4016 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4017 	if (vp->v_iflag & VI_DEFINACT)
4018 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4019 	flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_DOINGINACT |
4020 	    VI_OWEINACT | VI_DEFINACT);
4021 	if (flags != 0) {
4022 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4023 		strlcat(buf, buf2, sizeof(buf));
4024 	}
4025 	if (vp->v_mflag & VMP_LAZYLIST)
4026 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4027 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
4028 	if (flags != 0) {
4029 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4030 		strlcat(buf, buf2, sizeof(buf));
4031 	}
4032 	printf("    flags (%s)\n", buf + 1);
4033 	if (mtx_owned(VI_MTX(vp)))
4034 		printf(" VI_LOCKed");
4035 	if (vp->v_object != NULL)
4036 		printf("    v_object %p ref %d pages %d "
4037 		    "cleanbuf %d dirtybuf %d\n",
4038 		    vp->v_object, vp->v_object->ref_count,
4039 		    vp->v_object->resident_page_count,
4040 		    vp->v_bufobj.bo_clean.bv_cnt,
4041 		    vp->v_bufobj.bo_dirty.bv_cnt);
4042 	printf("    ");
4043 	lockmgr_printinfo(vp->v_vnlock);
4044 	if (vp->v_data != NULL)
4045 		VOP_PRINT(vp);
4046 }
4047 
4048 #ifdef DDB
4049 /*
4050  * List all of the locked vnodes in the system.
4051  * Called when debugging the kernel.
4052  */
4053 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
4054 {
4055 	struct mount *mp;
4056 	struct vnode *vp;
4057 
4058 	/*
4059 	 * Note: because this is DDB, we can't obey the locking semantics
4060 	 * for these structures, which means we could catch an inconsistent
4061 	 * state and dereference a nasty pointer.  Not much to be done
4062 	 * about that.
4063 	 */
4064 	db_printf("Locked vnodes\n");
4065 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4066 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4067 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4068 				vn_printf(vp, "vnode ");
4069 		}
4070 	}
4071 }
4072 
4073 /*
4074  * Show details about the given vnode.
4075  */
4076 DB_SHOW_COMMAND(vnode, db_show_vnode)
4077 {
4078 	struct vnode *vp;
4079 
4080 	if (!have_addr)
4081 		return;
4082 	vp = (struct vnode *)addr;
4083 	vn_printf(vp, "vnode ");
4084 }
4085 
4086 /*
4087  * Show details about the given mount point.
4088  */
4089 DB_SHOW_COMMAND(mount, db_show_mount)
4090 {
4091 	struct mount *mp;
4092 	struct vfsopt *opt;
4093 	struct statfs *sp;
4094 	struct vnode *vp;
4095 	char buf[512];
4096 	uint64_t mflags;
4097 	u_int flags;
4098 
4099 	if (!have_addr) {
4100 		/* No address given, print short info about all mount points. */
4101 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4102 			db_printf("%p %s on %s (%s)\n", mp,
4103 			    mp->mnt_stat.f_mntfromname,
4104 			    mp->mnt_stat.f_mntonname,
4105 			    mp->mnt_stat.f_fstypename);
4106 			if (db_pager_quit)
4107 				break;
4108 		}
4109 		db_printf("\nMore info: show mount <addr>\n");
4110 		return;
4111 	}
4112 
4113 	mp = (struct mount *)addr;
4114 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4115 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4116 
4117 	buf[0] = '\0';
4118 	mflags = mp->mnt_flag;
4119 #define	MNT_FLAG(flag)	do {						\
4120 	if (mflags & (flag)) {						\
4121 		if (buf[0] != '\0')					\
4122 			strlcat(buf, ", ", sizeof(buf));		\
4123 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4124 		mflags &= ~(flag);					\
4125 	}								\
4126 } while (0)
4127 	MNT_FLAG(MNT_RDONLY);
4128 	MNT_FLAG(MNT_SYNCHRONOUS);
4129 	MNT_FLAG(MNT_NOEXEC);
4130 	MNT_FLAG(MNT_NOSUID);
4131 	MNT_FLAG(MNT_NFS4ACLS);
4132 	MNT_FLAG(MNT_UNION);
4133 	MNT_FLAG(MNT_ASYNC);
4134 	MNT_FLAG(MNT_SUIDDIR);
4135 	MNT_FLAG(MNT_SOFTDEP);
4136 	MNT_FLAG(MNT_NOSYMFOLLOW);
4137 	MNT_FLAG(MNT_GJOURNAL);
4138 	MNT_FLAG(MNT_MULTILABEL);
4139 	MNT_FLAG(MNT_ACLS);
4140 	MNT_FLAG(MNT_NOATIME);
4141 	MNT_FLAG(MNT_NOCLUSTERR);
4142 	MNT_FLAG(MNT_NOCLUSTERW);
4143 	MNT_FLAG(MNT_SUJ);
4144 	MNT_FLAG(MNT_EXRDONLY);
4145 	MNT_FLAG(MNT_EXPORTED);
4146 	MNT_FLAG(MNT_DEFEXPORTED);
4147 	MNT_FLAG(MNT_EXPORTANON);
4148 	MNT_FLAG(MNT_EXKERB);
4149 	MNT_FLAG(MNT_EXPUBLIC);
4150 	MNT_FLAG(MNT_LOCAL);
4151 	MNT_FLAG(MNT_QUOTA);
4152 	MNT_FLAG(MNT_ROOTFS);
4153 	MNT_FLAG(MNT_USER);
4154 	MNT_FLAG(MNT_IGNORE);
4155 	MNT_FLAG(MNT_UPDATE);
4156 	MNT_FLAG(MNT_DELEXPORT);
4157 	MNT_FLAG(MNT_RELOAD);
4158 	MNT_FLAG(MNT_FORCE);
4159 	MNT_FLAG(MNT_SNAPSHOT);
4160 	MNT_FLAG(MNT_BYFSID);
4161 #undef MNT_FLAG
4162 	if (mflags != 0) {
4163 		if (buf[0] != '\0')
4164 			strlcat(buf, ", ", sizeof(buf));
4165 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4166 		    "0x%016jx", mflags);
4167 	}
4168 	db_printf("    mnt_flag = %s\n", buf);
4169 
4170 	buf[0] = '\0';
4171 	flags = mp->mnt_kern_flag;
4172 #define	MNT_KERN_FLAG(flag)	do {					\
4173 	if (flags & (flag)) {						\
4174 		if (buf[0] != '\0')					\
4175 			strlcat(buf, ", ", sizeof(buf));		\
4176 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4177 		flags &= ~(flag);					\
4178 	}								\
4179 } while (0)
4180 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4181 	MNT_KERN_FLAG(MNTK_ASYNC);
4182 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4183 	MNT_KERN_FLAG(MNTK_DRAINING);
4184 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4185 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4186 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4187 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4188 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
4189 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
4190 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
4191 	MNT_KERN_FLAG(MNTK_MARKER);
4192 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4193 	MNT_KERN_FLAG(MNTK_NOASYNC);
4194 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4195 	MNT_KERN_FLAG(MNTK_MWAIT);
4196 	MNT_KERN_FLAG(MNTK_SUSPEND);
4197 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4198 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4199 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4200 	MNT_KERN_FLAG(MNTK_NOKNOTE);
4201 #undef MNT_KERN_FLAG
4202 	if (flags != 0) {
4203 		if (buf[0] != '\0')
4204 			strlcat(buf, ", ", sizeof(buf));
4205 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4206 		    "0x%08x", flags);
4207 	}
4208 	db_printf("    mnt_kern_flag = %s\n", buf);
4209 
4210 	db_printf("    mnt_opt = ");
4211 	opt = TAILQ_FIRST(mp->mnt_opt);
4212 	if (opt != NULL) {
4213 		db_printf("%s", opt->name);
4214 		opt = TAILQ_NEXT(opt, link);
4215 		while (opt != NULL) {
4216 			db_printf(", %s", opt->name);
4217 			opt = TAILQ_NEXT(opt, link);
4218 		}
4219 	}
4220 	db_printf("\n");
4221 
4222 	sp = &mp->mnt_stat;
4223 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4224 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4225 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4226 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4227 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4228 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4229 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4230 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4231 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4232 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4233 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4234 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4235 
4236 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4237 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4238 	if (jailed(mp->mnt_cred))
4239 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4240 	db_printf(" }\n");
4241 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4242 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4243 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4244 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4245 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4246 	    mp->mnt_lazyvnodelistsize);
4247 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4248 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4249 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
4250 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4251 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4252 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4253 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4254 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4255 	db_printf("    mnt_secondary_accwrites = %d\n",
4256 	    mp->mnt_secondary_accwrites);
4257 	db_printf("    mnt_gjprovider = %s\n",
4258 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4259 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4260 
4261 	db_printf("\n\nList of active vnodes\n");
4262 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4263 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4264 			vn_printf(vp, "vnode ");
4265 			if (db_pager_quit)
4266 				break;
4267 		}
4268 	}
4269 	db_printf("\n\nList of inactive vnodes\n");
4270 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4271 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4272 			vn_printf(vp, "vnode ");
4273 			if (db_pager_quit)
4274 				break;
4275 		}
4276 	}
4277 }
4278 #endif	/* DDB */
4279 
4280 /*
4281  * Fill in a struct xvfsconf based on a struct vfsconf.
4282  */
4283 static int
4284 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4285 {
4286 	struct xvfsconf xvfsp;
4287 
4288 	bzero(&xvfsp, sizeof(xvfsp));
4289 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4290 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4291 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4292 	xvfsp.vfc_flags = vfsp->vfc_flags;
4293 	/*
4294 	 * These are unused in userland, we keep them
4295 	 * to not break binary compatibility.
4296 	 */
4297 	xvfsp.vfc_vfsops = NULL;
4298 	xvfsp.vfc_next = NULL;
4299 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4300 }
4301 
4302 #ifdef COMPAT_FREEBSD32
4303 struct xvfsconf32 {
4304 	uint32_t	vfc_vfsops;
4305 	char		vfc_name[MFSNAMELEN];
4306 	int32_t		vfc_typenum;
4307 	int32_t		vfc_refcount;
4308 	int32_t		vfc_flags;
4309 	uint32_t	vfc_next;
4310 };
4311 
4312 static int
4313 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4314 {
4315 	struct xvfsconf32 xvfsp;
4316 
4317 	bzero(&xvfsp, sizeof(xvfsp));
4318 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4319 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4320 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4321 	xvfsp.vfc_flags = vfsp->vfc_flags;
4322 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4323 }
4324 #endif
4325 
4326 /*
4327  * Top level filesystem related information gathering.
4328  */
4329 static int
4330 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4331 {
4332 	struct vfsconf *vfsp;
4333 	int error;
4334 
4335 	error = 0;
4336 	vfsconf_slock();
4337 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4338 #ifdef COMPAT_FREEBSD32
4339 		if (req->flags & SCTL_MASK32)
4340 			error = vfsconf2x32(req, vfsp);
4341 		else
4342 #endif
4343 			error = vfsconf2x(req, vfsp);
4344 		if (error)
4345 			break;
4346 	}
4347 	vfsconf_sunlock();
4348 	return (error);
4349 }
4350 
4351 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4352     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4353     "S,xvfsconf", "List of all configured filesystems");
4354 
4355 #ifndef BURN_BRIDGES
4356 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4357 
4358 static int
4359 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4360 {
4361 	int *name = (int *)arg1 - 1;	/* XXX */
4362 	u_int namelen = arg2 + 1;	/* XXX */
4363 	struct vfsconf *vfsp;
4364 
4365 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4366 	    "please rebuild world\n");
4367 
4368 #if 1 || defined(COMPAT_PRELITE2)
4369 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4370 	if (namelen == 1)
4371 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4372 #endif
4373 
4374 	switch (name[1]) {
4375 	case VFS_MAXTYPENUM:
4376 		if (namelen != 2)
4377 			return (ENOTDIR);
4378 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4379 	case VFS_CONF:
4380 		if (namelen != 3)
4381 			return (ENOTDIR);	/* overloaded */
4382 		vfsconf_slock();
4383 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4384 			if (vfsp->vfc_typenum == name[2])
4385 				break;
4386 		}
4387 		vfsconf_sunlock();
4388 		if (vfsp == NULL)
4389 			return (EOPNOTSUPP);
4390 #ifdef COMPAT_FREEBSD32
4391 		if (req->flags & SCTL_MASK32)
4392 			return (vfsconf2x32(req, vfsp));
4393 		else
4394 #endif
4395 			return (vfsconf2x(req, vfsp));
4396 	}
4397 	return (EOPNOTSUPP);
4398 }
4399 
4400 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4401     CTLFLAG_MPSAFE, vfs_sysctl,
4402     "Generic filesystem");
4403 
4404 #if 1 || defined(COMPAT_PRELITE2)
4405 
4406 static int
4407 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4408 {
4409 	int error;
4410 	struct vfsconf *vfsp;
4411 	struct ovfsconf ovfs;
4412 
4413 	vfsconf_slock();
4414 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4415 		bzero(&ovfs, sizeof(ovfs));
4416 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4417 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4418 		ovfs.vfc_index = vfsp->vfc_typenum;
4419 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4420 		ovfs.vfc_flags = vfsp->vfc_flags;
4421 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4422 		if (error != 0) {
4423 			vfsconf_sunlock();
4424 			return (error);
4425 		}
4426 	}
4427 	vfsconf_sunlock();
4428 	return (0);
4429 }
4430 
4431 #endif /* 1 || COMPAT_PRELITE2 */
4432 #endif /* !BURN_BRIDGES */
4433 
4434 #define KINFO_VNODESLOP		10
4435 #ifdef notyet
4436 /*
4437  * Dump vnode list (via sysctl).
4438  */
4439 /* ARGSUSED */
4440 static int
4441 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4442 {
4443 	struct xvnode *xvn;
4444 	struct mount *mp;
4445 	struct vnode *vp;
4446 	int error, len, n;
4447 
4448 	/*
4449 	 * Stale numvnodes access is not fatal here.
4450 	 */
4451 	req->lock = 0;
4452 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4453 	if (!req->oldptr)
4454 		/* Make an estimate */
4455 		return (SYSCTL_OUT(req, 0, len));
4456 
4457 	error = sysctl_wire_old_buffer(req, 0);
4458 	if (error != 0)
4459 		return (error);
4460 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4461 	n = 0;
4462 	mtx_lock(&mountlist_mtx);
4463 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4464 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4465 			continue;
4466 		MNT_ILOCK(mp);
4467 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4468 			if (n == len)
4469 				break;
4470 			vref(vp);
4471 			xvn[n].xv_size = sizeof *xvn;
4472 			xvn[n].xv_vnode = vp;
4473 			xvn[n].xv_id = 0;	/* XXX compat */
4474 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4475 			XV_COPY(usecount);
4476 			XV_COPY(writecount);
4477 			XV_COPY(holdcnt);
4478 			XV_COPY(mount);
4479 			XV_COPY(numoutput);
4480 			XV_COPY(type);
4481 #undef XV_COPY
4482 			xvn[n].xv_flag = vp->v_vflag;
4483 
4484 			switch (vp->v_type) {
4485 			case VREG:
4486 			case VDIR:
4487 			case VLNK:
4488 				break;
4489 			case VBLK:
4490 			case VCHR:
4491 				if (vp->v_rdev == NULL) {
4492 					vrele(vp);
4493 					continue;
4494 				}
4495 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4496 				break;
4497 			case VSOCK:
4498 				xvn[n].xv_socket = vp->v_socket;
4499 				break;
4500 			case VFIFO:
4501 				xvn[n].xv_fifo = vp->v_fifoinfo;
4502 				break;
4503 			case VNON:
4504 			case VBAD:
4505 			default:
4506 				/* shouldn't happen? */
4507 				vrele(vp);
4508 				continue;
4509 			}
4510 			vrele(vp);
4511 			++n;
4512 		}
4513 		MNT_IUNLOCK(mp);
4514 		mtx_lock(&mountlist_mtx);
4515 		vfs_unbusy(mp);
4516 		if (n == len)
4517 			break;
4518 	}
4519 	mtx_unlock(&mountlist_mtx);
4520 
4521 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4522 	free(xvn, M_TEMP);
4523 	return (error);
4524 }
4525 
4526 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4527     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4528     "");
4529 #endif
4530 
4531 static void
4532 unmount_or_warn(struct mount *mp)
4533 {
4534 	int error;
4535 
4536 	error = dounmount(mp, MNT_FORCE, curthread);
4537 	if (error != 0) {
4538 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4539 		if (error == EBUSY)
4540 			printf("BUSY)\n");
4541 		else
4542 			printf("%d)\n", error);
4543 	}
4544 }
4545 
4546 /*
4547  * Unmount all filesystems. The list is traversed in reverse order
4548  * of mounting to avoid dependencies.
4549  */
4550 void
4551 vfs_unmountall(void)
4552 {
4553 	struct mount *mp, *tmp;
4554 
4555 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4556 
4557 	/*
4558 	 * Since this only runs when rebooting, it is not interlocked.
4559 	 */
4560 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4561 		vfs_ref(mp);
4562 
4563 		/*
4564 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4565 		 * unmount of the latter.
4566 		 */
4567 		if (mp == rootdevmp)
4568 			continue;
4569 
4570 		unmount_or_warn(mp);
4571 	}
4572 
4573 	if (rootdevmp != NULL)
4574 		unmount_or_warn(rootdevmp);
4575 }
4576 
4577 static void
4578 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4579 {
4580 
4581 	ASSERT_VI_LOCKED(vp, __func__);
4582 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4583 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
4584 		vdropl(vp);
4585 		return;
4586 	}
4587 	if (vn_lock(vp, lkflags) == 0) {
4588 		VI_LOCK(vp);
4589 		vinactive(vp);
4590 		VOP_UNLOCK(vp);
4591 		vdropl(vp);
4592 		return;
4593 	}
4594 	vdefer_inactive_unlocked(vp);
4595 }
4596 
4597 static int
4598 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4599 {
4600 
4601 	return (vp->v_iflag & VI_DEFINACT);
4602 }
4603 
4604 static void __noinline
4605 vfs_periodic_inactive(struct mount *mp, int flags)
4606 {
4607 	struct vnode *vp, *mvp;
4608 	int lkflags;
4609 
4610 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4611 	if (flags != MNT_WAIT)
4612 		lkflags |= LK_NOWAIT;
4613 
4614 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4615 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
4616 			VI_UNLOCK(vp);
4617 			continue;
4618 		}
4619 		vp->v_iflag &= ~VI_DEFINACT;
4620 		vfs_deferred_inactive(vp, lkflags);
4621 	}
4622 }
4623 
4624 static inline bool
4625 vfs_want_msync(struct vnode *vp)
4626 {
4627 	struct vm_object *obj;
4628 
4629 	/*
4630 	 * This test may be performed without any locks held.
4631 	 * We rely on vm_object's type stability.
4632 	 */
4633 	if (vp->v_vflag & VV_NOSYNC)
4634 		return (false);
4635 	obj = vp->v_object;
4636 	return (obj != NULL && vm_object_mightbedirty(obj));
4637 }
4638 
4639 static int
4640 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4641 {
4642 
4643 	if (vp->v_vflag & VV_NOSYNC)
4644 		return (false);
4645 	if (vp->v_iflag & VI_DEFINACT)
4646 		return (true);
4647 	return (vfs_want_msync(vp));
4648 }
4649 
4650 static void __noinline
4651 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4652 {
4653 	struct vnode *vp, *mvp;
4654 	struct vm_object *obj;
4655 	struct thread *td;
4656 	int lkflags, objflags;
4657 	bool seen_defer;
4658 
4659 	td = curthread;
4660 
4661 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4662 	if (flags != MNT_WAIT) {
4663 		lkflags |= LK_NOWAIT;
4664 		objflags = OBJPC_NOSYNC;
4665 	} else {
4666 		objflags = OBJPC_SYNC;
4667 	}
4668 
4669 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4670 		seen_defer = false;
4671 		if (vp->v_iflag & VI_DEFINACT) {
4672 			vp->v_iflag &= ~VI_DEFINACT;
4673 			seen_defer = true;
4674 		}
4675 		if (!vfs_want_msync(vp)) {
4676 			if (seen_defer)
4677 				vfs_deferred_inactive(vp, lkflags);
4678 			else
4679 				VI_UNLOCK(vp);
4680 			continue;
4681 		}
4682 		if (vget(vp, lkflags, td) == 0) {
4683 			obj = vp->v_object;
4684 			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4685 				VM_OBJECT_WLOCK(obj);
4686 				vm_object_page_clean(obj, 0, 0, objflags);
4687 				VM_OBJECT_WUNLOCK(obj);
4688 			}
4689 			vput(vp);
4690 			if (seen_defer)
4691 				vdrop(vp);
4692 		} else {
4693 			if (seen_defer)
4694 				vdefer_inactive_unlocked(vp);
4695 		}
4696 	}
4697 }
4698 
4699 void
4700 vfs_periodic(struct mount *mp, int flags)
4701 {
4702 
4703 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4704 
4705 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4706 		vfs_periodic_inactive(mp, flags);
4707 	else
4708 		vfs_periodic_msync_inactive(mp, flags);
4709 }
4710 
4711 static void
4712 destroy_vpollinfo_free(struct vpollinfo *vi)
4713 {
4714 
4715 	knlist_destroy(&vi->vpi_selinfo.si_note);
4716 	mtx_destroy(&vi->vpi_lock);
4717 	uma_zfree(vnodepoll_zone, vi);
4718 }
4719 
4720 static void
4721 destroy_vpollinfo(struct vpollinfo *vi)
4722 {
4723 
4724 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4725 	seldrain(&vi->vpi_selinfo);
4726 	destroy_vpollinfo_free(vi);
4727 }
4728 
4729 /*
4730  * Initialize per-vnode helper structure to hold poll-related state.
4731  */
4732 void
4733 v_addpollinfo(struct vnode *vp)
4734 {
4735 	struct vpollinfo *vi;
4736 
4737 	if (vp->v_pollinfo != NULL)
4738 		return;
4739 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4740 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4741 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4742 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4743 	VI_LOCK(vp);
4744 	if (vp->v_pollinfo != NULL) {
4745 		VI_UNLOCK(vp);
4746 		destroy_vpollinfo_free(vi);
4747 		return;
4748 	}
4749 	vp->v_pollinfo = vi;
4750 	VI_UNLOCK(vp);
4751 }
4752 
4753 /*
4754  * Record a process's interest in events which might happen to
4755  * a vnode.  Because poll uses the historic select-style interface
4756  * internally, this routine serves as both the ``check for any
4757  * pending events'' and the ``record my interest in future events''
4758  * functions.  (These are done together, while the lock is held,
4759  * to avoid race conditions.)
4760  */
4761 int
4762 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4763 {
4764 
4765 	v_addpollinfo(vp);
4766 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4767 	if (vp->v_pollinfo->vpi_revents & events) {
4768 		/*
4769 		 * This leaves events we are not interested
4770 		 * in available for the other process which
4771 		 * which presumably had requested them
4772 		 * (otherwise they would never have been
4773 		 * recorded).
4774 		 */
4775 		events &= vp->v_pollinfo->vpi_revents;
4776 		vp->v_pollinfo->vpi_revents &= ~events;
4777 
4778 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4779 		return (events);
4780 	}
4781 	vp->v_pollinfo->vpi_events |= events;
4782 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4783 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4784 	return (0);
4785 }
4786 
4787 /*
4788  * Routine to create and manage a filesystem syncer vnode.
4789  */
4790 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4791 static int	sync_fsync(struct  vop_fsync_args *);
4792 static int	sync_inactive(struct  vop_inactive_args *);
4793 static int	sync_reclaim(struct  vop_reclaim_args *);
4794 
4795 static struct vop_vector sync_vnodeops = {
4796 	.vop_bypass =	VOP_EOPNOTSUPP,
4797 	.vop_close =	sync_close,		/* close */
4798 	.vop_fsync =	sync_fsync,		/* fsync */
4799 	.vop_inactive =	sync_inactive,	/* inactive */
4800 	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4801 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4802 	.vop_lock1 =	vop_stdlock,	/* lock */
4803 	.vop_unlock =	vop_stdunlock,	/* unlock */
4804 	.vop_islocked =	vop_stdislocked,	/* islocked */
4805 };
4806 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4807 
4808 /*
4809  * Create a new filesystem syncer vnode for the specified mount point.
4810  */
4811 void
4812 vfs_allocate_syncvnode(struct mount *mp)
4813 {
4814 	struct vnode *vp;
4815 	struct bufobj *bo;
4816 	static long start, incr, next;
4817 	int error;
4818 
4819 	/* Allocate a new vnode */
4820 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4821 	if (error != 0)
4822 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4823 	vp->v_type = VNON;
4824 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4825 	vp->v_vflag |= VV_FORCEINSMQ;
4826 	error = insmntque(vp, mp);
4827 	if (error != 0)
4828 		panic("vfs_allocate_syncvnode: insmntque() failed");
4829 	vp->v_vflag &= ~VV_FORCEINSMQ;
4830 	VOP_UNLOCK(vp);
4831 	/*
4832 	 * Place the vnode onto the syncer worklist. We attempt to
4833 	 * scatter them about on the list so that they will go off
4834 	 * at evenly distributed times even if all the filesystems
4835 	 * are mounted at once.
4836 	 */
4837 	next += incr;
4838 	if (next == 0 || next > syncer_maxdelay) {
4839 		start /= 2;
4840 		incr /= 2;
4841 		if (start == 0) {
4842 			start = syncer_maxdelay / 2;
4843 			incr = syncer_maxdelay;
4844 		}
4845 		next = start;
4846 	}
4847 	bo = &vp->v_bufobj;
4848 	BO_LOCK(bo);
4849 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4850 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4851 	mtx_lock(&sync_mtx);
4852 	sync_vnode_count++;
4853 	if (mp->mnt_syncer == NULL) {
4854 		mp->mnt_syncer = vp;
4855 		vp = NULL;
4856 	}
4857 	mtx_unlock(&sync_mtx);
4858 	BO_UNLOCK(bo);
4859 	if (vp != NULL) {
4860 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4861 		vgone(vp);
4862 		vput(vp);
4863 	}
4864 }
4865 
4866 void
4867 vfs_deallocate_syncvnode(struct mount *mp)
4868 {
4869 	struct vnode *vp;
4870 
4871 	mtx_lock(&sync_mtx);
4872 	vp = mp->mnt_syncer;
4873 	if (vp != NULL)
4874 		mp->mnt_syncer = NULL;
4875 	mtx_unlock(&sync_mtx);
4876 	if (vp != NULL)
4877 		vrele(vp);
4878 }
4879 
4880 /*
4881  * Do a lazy sync of the filesystem.
4882  */
4883 static int
4884 sync_fsync(struct vop_fsync_args *ap)
4885 {
4886 	struct vnode *syncvp = ap->a_vp;
4887 	struct mount *mp = syncvp->v_mount;
4888 	int error, save;
4889 	struct bufobj *bo;
4890 
4891 	/*
4892 	 * We only need to do something if this is a lazy evaluation.
4893 	 */
4894 	if (ap->a_waitfor != MNT_LAZY)
4895 		return (0);
4896 
4897 	/*
4898 	 * Move ourselves to the back of the sync list.
4899 	 */
4900 	bo = &syncvp->v_bufobj;
4901 	BO_LOCK(bo);
4902 	vn_syncer_add_to_worklist(bo, syncdelay);
4903 	BO_UNLOCK(bo);
4904 
4905 	/*
4906 	 * Walk the list of vnodes pushing all that are dirty and
4907 	 * not already on the sync list.
4908 	 */
4909 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4910 		return (0);
4911 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4912 		vfs_unbusy(mp);
4913 		return (0);
4914 	}
4915 	save = curthread_pflags_set(TDP_SYNCIO);
4916 	/*
4917 	 * The filesystem at hand may be idle with free vnodes stored in the
4918 	 * batch.  Return them instead of letting them stay there indefinitely.
4919 	 */
4920 	vfs_periodic(mp, MNT_NOWAIT);
4921 	error = VFS_SYNC(mp, MNT_LAZY);
4922 	curthread_pflags_restore(save);
4923 	vn_finished_write(mp);
4924 	vfs_unbusy(mp);
4925 	return (error);
4926 }
4927 
4928 /*
4929  * The syncer vnode is no referenced.
4930  */
4931 static int
4932 sync_inactive(struct vop_inactive_args *ap)
4933 {
4934 
4935 	vgone(ap->a_vp);
4936 	return (0);
4937 }
4938 
4939 /*
4940  * The syncer vnode is no longer needed and is being decommissioned.
4941  *
4942  * Modifications to the worklist must be protected by sync_mtx.
4943  */
4944 static int
4945 sync_reclaim(struct vop_reclaim_args *ap)
4946 {
4947 	struct vnode *vp = ap->a_vp;
4948 	struct bufobj *bo;
4949 
4950 	bo = &vp->v_bufobj;
4951 	BO_LOCK(bo);
4952 	mtx_lock(&sync_mtx);
4953 	if (vp->v_mount->mnt_syncer == vp)
4954 		vp->v_mount->mnt_syncer = NULL;
4955 	if (bo->bo_flag & BO_ONWORKLST) {
4956 		LIST_REMOVE(bo, bo_synclist);
4957 		syncer_worklist_len--;
4958 		sync_vnode_count--;
4959 		bo->bo_flag &= ~BO_ONWORKLST;
4960 	}
4961 	mtx_unlock(&sync_mtx);
4962 	BO_UNLOCK(bo);
4963 
4964 	return (0);
4965 }
4966 
4967 int
4968 vn_need_pageq_flush(struct vnode *vp)
4969 {
4970 	struct vm_object *obj;
4971 	int need;
4972 
4973 	MPASS(mtx_owned(VI_MTX(vp)));
4974 	need = 0;
4975 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
4976 	    vm_object_mightbedirty(obj))
4977 		need = 1;
4978 	return (need);
4979 }
4980 
4981 /*
4982  * Check if vnode represents a disk device
4983  */
4984 int
4985 vn_isdisk(struct vnode *vp, int *errp)
4986 {
4987 	int error;
4988 
4989 	if (vp->v_type != VCHR) {
4990 		error = ENOTBLK;
4991 		goto out;
4992 	}
4993 	error = 0;
4994 	dev_lock();
4995 	if (vp->v_rdev == NULL)
4996 		error = ENXIO;
4997 	else if (vp->v_rdev->si_devsw == NULL)
4998 		error = ENXIO;
4999 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5000 		error = ENOTBLK;
5001 	dev_unlock();
5002 out:
5003 	if (errp != NULL)
5004 		*errp = error;
5005 	return (error == 0);
5006 }
5007 
5008 /*
5009  * Common filesystem object access control check routine.  Accepts a
5010  * vnode's type, "mode", uid and gid, requested access mode, credentials,
5011  * and optional call-by-reference privused argument allowing vaccess()
5012  * to indicate to the caller whether privilege was used to satisfy the
5013  * request (obsoleted).  Returns 0 on success, or an errno on failure.
5014  */
5015 int
5016 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5017     accmode_t accmode, struct ucred *cred, int *privused)
5018 {
5019 	accmode_t dac_granted;
5020 	accmode_t priv_granted;
5021 
5022 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5023 	    ("invalid bit in accmode"));
5024 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5025 	    ("VAPPEND without VWRITE"));
5026 
5027 	/*
5028 	 * Look for a normal, non-privileged way to access the file/directory
5029 	 * as requested.  If it exists, go with that.
5030 	 */
5031 
5032 	if (privused != NULL)
5033 		*privused = 0;
5034 
5035 	dac_granted = 0;
5036 
5037 	/* Check the owner. */
5038 	if (cred->cr_uid == file_uid) {
5039 		dac_granted |= VADMIN;
5040 		if (file_mode & S_IXUSR)
5041 			dac_granted |= VEXEC;
5042 		if (file_mode & S_IRUSR)
5043 			dac_granted |= VREAD;
5044 		if (file_mode & S_IWUSR)
5045 			dac_granted |= (VWRITE | VAPPEND);
5046 
5047 		if ((accmode & dac_granted) == accmode)
5048 			return (0);
5049 
5050 		goto privcheck;
5051 	}
5052 
5053 	/* Otherwise, check the groups (first match) */
5054 	if (groupmember(file_gid, cred)) {
5055 		if (file_mode & S_IXGRP)
5056 			dac_granted |= VEXEC;
5057 		if (file_mode & S_IRGRP)
5058 			dac_granted |= VREAD;
5059 		if (file_mode & S_IWGRP)
5060 			dac_granted |= (VWRITE | VAPPEND);
5061 
5062 		if ((accmode & dac_granted) == accmode)
5063 			return (0);
5064 
5065 		goto privcheck;
5066 	}
5067 
5068 	/* Otherwise, check everyone else. */
5069 	if (file_mode & S_IXOTH)
5070 		dac_granted |= VEXEC;
5071 	if (file_mode & S_IROTH)
5072 		dac_granted |= VREAD;
5073 	if (file_mode & S_IWOTH)
5074 		dac_granted |= (VWRITE | VAPPEND);
5075 	if ((accmode & dac_granted) == accmode)
5076 		return (0);
5077 
5078 privcheck:
5079 	/*
5080 	 * Build a privilege mask to determine if the set of privileges
5081 	 * satisfies the requirements when combined with the granted mask
5082 	 * from above.  For each privilege, if the privilege is required,
5083 	 * bitwise or the request type onto the priv_granted mask.
5084 	 */
5085 	priv_granted = 0;
5086 
5087 	if (type == VDIR) {
5088 		/*
5089 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5090 		 * requests, instead of PRIV_VFS_EXEC.
5091 		 */
5092 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5093 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5094 			priv_granted |= VEXEC;
5095 	} else {
5096 		/*
5097 		 * Ensure that at least one execute bit is on. Otherwise,
5098 		 * a privileged user will always succeed, and we don't want
5099 		 * this to happen unless the file really is executable.
5100 		 */
5101 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5102 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5103 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5104 			priv_granted |= VEXEC;
5105 	}
5106 
5107 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5108 	    !priv_check_cred(cred, PRIV_VFS_READ))
5109 		priv_granted |= VREAD;
5110 
5111 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5112 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5113 		priv_granted |= (VWRITE | VAPPEND);
5114 
5115 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5116 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5117 		priv_granted |= VADMIN;
5118 
5119 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5120 		/* XXX audit: privilege used */
5121 		if (privused != NULL)
5122 			*privused = 1;
5123 		return (0);
5124 	}
5125 
5126 	return ((accmode & VADMIN) ? EPERM : EACCES);
5127 }
5128 
5129 /*
5130  * Credential check based on process requesting service, and per-attribute
5131  * permissions.
5132  */
5133 int
5134 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5135     struct thread *td, accmode_t accmode)
5136 {
5137 
5138 	/*
5139 	 * Kernel-invoked always succeeds.
5140 	 */
5141 	if (cred == NOCRED)
5142 		return (0);
5143 
5144 	/*
5145 	 * Do not allow privileged processes in jail to directly manipulate
5146 	 * system attributes.
5147 	 */
5148 	switch (attrnamespace) {
5149 	case EXTATTR_NAMESPACE_SYSTEM:
5150 		/* Potentially should be: return (EPERM); */
5151 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5152 	case EXTATTR_NAMESPACE_USER:
5153 		return (VOP_ACCESS(vp, accmode, cred, td));
5154 	default:
5155 		return (EPERM);
5156 	}
5157 }
5158 
5159 #ifdef DEBUG_VFS_LOCKS
5160 /*
5161  * This only exists to suppress warnings from unlocked specfs accesses.  It is
5162  * no longer ok to have an unlocked VFS.
5163  */
5164 #define	IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL ||		\
5165 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
5166 
5167 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5168 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5169     "Drop into debugger on lock violation");
5170 
5171 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5172 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5173     0, "Check for interlock across VOPs");
5174 
5175 int vfs_badlock_print = 1;	/* Print lock violations. */
5176 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5177     0, "Print lock violations");
5178 
5179 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5180 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5181     0, "Print vnode details on lock violations");
5182 
5183 #ifdef KDB
5184 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5185 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5186     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5187 #endif
5188 
5189 static void
5190 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5191 {
5192 
5193 #ifdef KDB
5194 	if (vfs_badlock_backtrace)
5195 		kdb_backtrace();
5196 #endif
5197 	if (vfs_badlock_vnode)
5198 		vn_printf(vp, "vnode ");
5199 	if (vfs_badlock_print)
5200 		printf("%s: %p %s\n", str, (void *)vp, msg);
5201 	if (vfs_badlock_ddb)
5202 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5203 }
5204 
5205 void
5206 assert_vi_locked(struct vnode *vp, const char *str)
5207 {
5208 
5209 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5210 		vfs_badlock("interlock is not locked but should be", str, vp);
5211 }
5212 
5213 void
5214 assert_vi_unlocked(struct vnode *vp, const char *str)
5215 {
5216 
5217 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5218 		vfs_badlock("interlock is locked but should not be", str, vp);
5219 }
5220 
5221 void
5222 assert_vop_locked(struct vnode *vp, const char *str)
5223 {
5224 	int locked;
5225 
5226 	if (!IGNORE_LOCK(vp)) {
5227 		locked = VOP_ISLOCKED(vp);
5228 		if (locked == 0 || locked == LK_EXCLOTHER)
5229 			vfs_badlock("is not locked but should be", str, vp);
5230 	}
5231 }
5232 
5233 void
5234 assert_vop_unlocked(struct vnode *vp, const char *str)
5235 {
5236 
5237 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5238 		vfs_badlock("is locked but should not be", str, vp);
5239 }
5240 
5241 void
5242 assert_vop_elocked(struct vnode *vp, const char *str)
5243 {
5244 
5245 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5246 		vfs_badlock("is not exclusive locked but should be", str, vp);
5247 }
5248 #endif /* DEBUG_VFS_LOCKS */
5249 
5250 void
5251 vop_rename_fail(struct vop_rename_args *ap)
5252 {
5253 
5254 	if (ap->a_tvp != NULL)
5255 		vput(ap->a_tvp);
5256 	if (ap->a_tdvp == ap->a_tvp)
5257 		vrele(ap->a_tdvp);
5258 	else
5259 		vput(ap->a_tdvp);
5260 	vrele(ap->a_fdvp);
5261 	vrele(ap->a_fvp);
5262 }
5263 
5264 void
5265 vop_rename_pre(void *ap)
5266 {
5267 	struct vop_rename_args *a = ap;
5268 
5269 #ifdef DEBUG_VFS_LOCKS
5270 	if (a->a_tvp)
5271 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5272 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5273 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5274 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5275 
5276 	/* Check the source (from). */
5277 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5278 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5279 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5280 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5281 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5282 
5283 	/* Check the target. */
5284 	if (a->a_tvp)
5285 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5286 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5287 #endif
5288 	if (a->a_tdvp != a->a_fdvp)
5289 		vhold(a->a_fdvp);
5290 	if (a->a_tvp != a->a_fvp)
5291 		vhold(a->a_fvp);
5292 	vhold(a->a_tdvp);
5293 	if (a->a_tvp)
5294 		vhold(a->a_tvp);
5295 }
5296 
5297 #ifdef DEBUG_VFS_LOCKS
5298 void
5299 vop_strategy_pre(void *ap)
5300 {
5301 	struct vop_strategy_args *a;
5302 	struct buf *bp;
5303 
5304 	a = ap;
5305 	bp = a->a_bp;
5306 
5307 	/*
5308 	 * Cluster ops lock their component buffers but not the IO container.
5309 	 */
5310 	if ((bp->b_flags & B_CLUSTER) != 0)
5311 		return;
5312 
5313 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5314 		if (vfs_badlock_print)
5315 			printf(
5316 			    "VOP_STRATEGY: bp is not locked but should be\n");
5317 		if (vfs_badlock_ddb)
5318 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5319 	}
5320 }
5321 
5322 void
5323 vop_lock_pre(void *ap)
5324 {
5325 	struct vop_lock1_args *a = ap;
5326 
5327 	if ((a->a_flags & LK_INTERLOCK) == 0)
5328 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5329 	else
5330 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5331 }
5332 
5333 void
5334 vop_lock_post(void *ap, int rc)
5335 {
5336 	struct vop_lock1_args *a = ap;
5337 
5338 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5339 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5340 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5341 }
5342 
5343 void
5344 vop_unlock_pre(void *ap)
5345 {
5346 	struct vop_unlock_args *a = ap;
5347 
5348 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5349 }
5350 
5351 void
5352 vop_unlock_post(void *ap, int rc)
5353 {
5354 	return;
5355 }
5356 
5357 void
5358 vop_need_inactive_pre(void *ap)
5359 {
5360 	struct vop_need_inactive_args *a = ap;
5361 
5362 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5363 }
5364 
5365 void
5366 vop_need_inactive_post(void *ap, int rc)
5367 {
5368 	struct vop_need_inactive_args *a = ap;
5369 
5370 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5371 }
5372 #endif
5373 
5374 void
5375 vop_create_post(void *ap, int rc)
5376 {
5377 	struct vop_create_args *a = ap;
5378 
5379 	if (!rc)
5380 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5381 }
5382 
5383 void
5384 vop_deleteextattr_post(void *ap, int rc)
5385 {
5386 	struct vop_deleteextattr_args *a = ap;
5387 
5388 	if (!rc)
5389 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5390 }
5391 
5392 void
5393 vop_link_post(void *ap, int rc)
5394 {
5395 	struct vop_link_args *a = ap;
5396 
5397 	if (!rc) {
5398 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
5399 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
5400 	}
5401 }
5402 
5403 void
5404 vop_mkdir_post(void *ap, int rc)
5405 {
5406 	struct vop_mkdir_args *a = ap;
5407 
5408 	if (!rc)
5409 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5410 }
5411 
5412 void
5413 vop_mknod_post(void *ap, int rc)
5414 {
5415 	struct vop_mknod_args *a = ap;
5416 
5417 	if (!rc)
5418 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5419 }
5420 
5421 void
5422 vop_reclaim_post(void *ap, int rc)
5423 {
5424 	struct vop_reclaim_args *a = ap;
5425 
5426 	if (!rc)
5427 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
5428 }
5429 
5430 void
5431 vop_remove_post(void *ap, int rc)
5432 {
5433 	struct vop_remove_args *a = ap;
5434 
5435 	if (!rc) {
5436 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5437 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5438 	}
5439 }
5440 
5441 void
5442 vop_rename_post(void *ap, int rc)
5443 {
5444 	struct vop_rename_args *a = ap;
5445 	long hint;
5446 
5447 	if (!rc) {
5448 		hint = NOTE_WRITE;
5449 		if (a->a_fdvp == a->a_tdvp) {
5450 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5451 				hint |= NOTE_LINK;
5452 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5453 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5454 		} else {
5455 			hint |= NOTE_EXTEND;
5456 			if (a->a_fvp->v_type == VDIR)
5457 				hint |= NOTE_LINK;
5458 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5459 
5460 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5461 			    a->a_tvp->v_type == VDIR)
5462 				hint &= ~NOTE_LINK;
5463 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5464 		}
5465 
5466 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5467 		if (a->a_tvp)
5468 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5469 	}
5470 	if (a->a_tdvp != a->a_fdvp)
5471 		vdrop(a->a_fdvp);
5472 	if (a->a_tvp != a->a_fvp)
5473 		vdrop(a->a_fvp);
5474 	vdrop(a->a_tdvp);
5475 	if (a->a_tvp)
5476 		vdrop(a->a_tvp);
5477 }
5478 
5479 void
5480 vop_rmdir_post(void *ap, int rc)
5481 {
5482 	struct vop_rmdir_args *a = ap;
5483 
5484 	if (!rc) {
5485 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5486 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5487 	}
5488 }
5489 
5490 void
5491 vop_setattr_post(void *ap, int rc)
5492 {
5493 	struct vop_setattr_args *a = ap;
5494 
5495 	if (!rc)
5496 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5497 }
5498 
5499 void
5500 vop_setextattr_post(void *ap, int rc)
5501 {
5502 	struct vop_setextattr_args *a = ap;
5503 
5504 	if (!rc)
5505 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5506 }
5507 
5508 void
5509 vop_symlink_post(void *ap, int rc)
5510 {
5511 	struct vop_symlink_args *a = ap;
5512 
5513 	if (!rc)
5514 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5515 }
5516 
5517 void
5518 vop_open_post(void *ap, int rc)
5519 {
5520 	struct vop_open_args *a = ap;
5521 
5522 	if (!rc)
5523 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5524 }
5525 
5526 void
5527 vop_close_post(void *ap, int rc)
5528 {
5529 	struct vop_close_args *a = ap;
5530 
5531 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5532 	    !VN_IS_DOOMED(a->a_vp))) {
5533 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5534 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
5535 	}
5536 }
5537 
5538 void
5539 vop_read_post(void *ap, int rc)
5540 {
5541 	struct vop_read_args *a = ap;
5542 
5543 	if (!rc)
5544 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5545 }
5546 
5547 void
5548 vop_readdir_post(void *ap, int rc)
5549 {
5550 	struct vop_readdir_args *a = ap;
5551 
5552 	if (!rc)
5553 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5554 }
5555 
5556 static struct knlist fs_knlist;
5557 
5558 static void
5559 vfs_event_init(void *arg)
5560 {
5561 	knlist_init_mtx(&fs_knlist, NULL);
5562 }
5563 /* XXX - correct order? */
5564 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5565 
5566 void
5567 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5568 {
5569 
5570 	KNOTE_UNLOCKED(&fs_knlist, event);
5571 }
5572 
5573 static int	filt_fsattach(struct knote *kn);
5574 static void	filt_fsdetach(struct knote *kn);
5575 static int	filt_fsevent(struct knote *kn, long hint);
5576 
5577 struct filterops fs_filtops = {
5578 	.f_isfd = 0,
5579 	.f_attach = filt_fsattach,
5580 	.f_detach = filt_fsdetach,
5581 	.f_event = filt_fsevent
5582 };
5583 
5584 static int
5585 filt_fsattach(struct knote *kn)
5586 {
5587 
5588 	kn->kn_flags |= EV_CLEAR;
5589 	knlist_add(&fs_knlist, kn, 0);
5590 	return (0);
5591 }
5592 
5593 static void
5594 filt_fsdetach(struct knote *kn)
5595 {
5596 
5597 	knlist_remove(&fs_knlist, kn, 0);
5598 }
5599 
5600 static int
5601 filt_fsevent(struct knote *kn, long hint)
5602 {
5603 
5604 	kn->kn_fflags |= hint;
5605 	return (kn->kn_fflags != 0);
5606 }
5607 
5608 static int
5609 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5610 {
5611 	struct vfsidctl vc;
5612 	int error;
5613 	struct mount *mp;
5614 
5615 	error = SYSCTL_IN(req, &vc, sizeof(vc));
5616 	if (error)
5617 		return (error);
5618 	if (vc.vc_vers != VFS_CTL_VERS1)
5619 		return (EINVAL);
5620 	mp = vfs_getvfs(&vc.vc_fsid);
5621 	if (mp == NULL)
5622 		return (ENOENT);
5623 	/* ensure that a specific sysctl goes to the right filesystem. */
5624 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
5625 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
5626 		vfs_rel(mp);
5627 		return (EINVAL);
5628 	}
5629 	VCTLTOREQ(&vc, req);
5630 	error = VFS_SYSCTL(mp, vc.vc_op, req);
5631 	vfs_rel(mp);
5632 	return (error);
5633 }
5634 
5635 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
5636     NULL, 0, sysctl_vfs_ctl, "",
5637     "Sysctl by fsid");
5638 
5639 /*
5640  * Function to initialize a va_filerev field sensibly.
5641  * XXX: Wouldn't a random number make a lot more sense ??
5642  */
5643 u_quad_t
5644 init_va_filerev(void)
5645 {
5646 	struct bintime bt;
5647 
5648 	getbinuptime(&bt);
5649 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5650 }
5651 
5652 static int	filt_vfsread(struct knote *kn, long hint);
5653 static int	filt_vfswrite(struct knote *kn, long hint);
5654 static int	filt_vfsvnode(struct knote *kn, long hint);
5655 static void	filt_vfsdetach(struct knote *kn);
5656 static struct filterops vfsread_filtops = {
5657 	.f_isfd = 1,
5658 	.f_detach = filt_vfsdetach,
5659 	.f_event = filt_vfsread
5660 };
5661 static struct filterops vfswrite_filtops = {
5662 	.f_isfd = 1,
5663 	.f_detach = filt_vfsdetach,
5664 	.f_event = filt_vfswrite
5665 };
5666 static struct filterops vfsvnode_filtops = {
5667 	.f_isfd = 1,
5668 	.f_detach = filt_vfsdetach,
5669 	.f_event = filt_vfsvnode
5670 };
5671 
5672 static void
5673 vfs_knllock(void *arg)
5674 {
5675 	struct vnode *vp = arg;
5676 
5677 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5678 }
5679 
5680 static void
5681 vfs_knlunlock(void *arg)
5682 {
5683 	struct vnode *vp = arg;
5684 
5685 	VOP_UNLOCK(vp);
5686 }
5687 
5688 static void
5689 vfs_knl_assert_locked(void *arg)
5690 {
5691 #ifdef DEBUG_VFS_LOCKS
5692 	struct vnode *vp = arg;
5693 
5694 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
5695 #endif
5696 }
5697 
5698 static void
5699 vfs_knl_assert_unlocked(void *arg)
5700 {
5701 #ifdef DEBUG_VFS_LOCKS
5702 	struct vnode *vp = arg;
5703 
5704 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
5705 #endif
5706 }
5707 
5708 int
5709 vfs_kqfilter(struct vop_kqfilter_args *ap)
5710 {
5711 	struct vnode *vp = ap->a_vp;
5712 	struct knote *kn = ap->a_kn;
5713 	struct knlist *knl;
5714 
5715 	switch (kn->kn_filter) {
5716 	case EVFILT_READ:
5717 		kn->kn_fop = &vfsread_filtops;
5718 		break;
5719 	case EVFILT_WRITE:
5720 		kn->kn_fop = &vfswrite_filtops;
5721 		break;
5722 	case EVFILT_VNODE:
5723 		kn->kn_fop = &vfsvnode_filtops;
5724 		break;
5725 	default:
5726 		return (EINVAL);
5727 	}
5728 
5729 	kn->kn_hook = (caddr_t)vp;
5730 
5731 	v_addpollinfo(vp);
5732 	if (vp->v_pollinfo == NULL)
5733 		return (ENOMEM);
5734 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
5735 	vhold(vp);
5736 	knlist_add(knl, kn, 0);
5737 
5738 	return (0);
5739 }
5740 
5741 /*
5742  * Detach knote from vnode
5743  */
5744 static void
5745 filt_vfsdetach(struct knote *kn)
5746 {
5747 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5748 
5749 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5750 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5751 	vdrop(vp);
5752 }
5753 
5754 /*ARGSUSED*/
5755 static int
5756 filt_vfsread(struct knote *kn, long hint)
5757 {
5758 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5759 	struct vattr va;
5760 	int res;
5761 
5762 	/*
5763 	 * filesystem is gone, so set the EOF flag and schedule
5764 	 * the knote for deletion.
5765 	 */
5766 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5767 		VI_LOCK(vp);
5768 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5769 		VI_UNLOCK(vp);
5770 		return (1);
5771 	}
5772 
5773 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5774 		return (0);
5775 
5776 	VI_LOCK(vp);
5777 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5778 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5779 	VI_UNLOCK(vp);
5780 	return (res);
5781 }
5782 
5783 /*ARGSUSED*/
5784 static int
5785 filt_vfswrite(struct knote *kn, long hint)
5786 {
5787 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5788 
5789 	VI_LOCK(vp);
5790 
5791 	/*
5792 	 * filesystem is gone, so set the EOF flag and schedule
5793 	 * the knote for deletion.
5794 	 */
5795 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5796 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5797 
5798 	kn->kn_data = 0;
5799 	VI_UNLOCK(vp);
5800 	return (1);
5801 }
5802 
5803 static int
5804 filt_vfsvnode(struct knote *kn, long hint)
5805 {
5806 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5807 	int res;
5808 
5809 	VI_LOCK(vp);
5810 	if (kn->kn_sfflags & hint)
5811 		kn->kn_fflags |= hint;
5812 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5813 		kn->kn_flags |= EV_EOF;
5814 		VI_UNLOCK(vp);
5815 		return (1);
5816 	}
5817 	res = (kn->kn_fflags != 0);
5818 	VI_UNLOCK(vp);
5819 	return (res);
5820 }
5821 
5822 /*
5823  * Returns whether the directory is empty or not.
5824  * If it is empty, the return value is 0; otherwise
5825  * the return value is an error value (which may
5826  * be ENOTEMPTY).
5827  */
5828 int
5829 vfs_emptydir(struct vnode *vp)
5830 {
5831 	struct uio uio;
5832 	struct iovec iov;
5833 	struct dirent *dirent, *dp, *endp;
5834 	int error, eof;
5835 
5836 	error = 0;
5837 	eof = 0;
5838 
5839 	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
5840 
5841 	dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
5842 	iov.iov_base = dirent;
5843 	iov.iov_len = sizeof(struct dirent);
5844 
5845 	uio.uio_iov = &iov;
5846 	uio.uio_iovcnt = 1;
5847 	uio.uio_offset = 0;
5848 	uio.uio_resid = sizeof(struct dirent);
5849 	uio.uio_segflg = UIO_SYSSPACE;
5850 	uio.uio_rw = UIO_READ;
5851 	uio.uio_td = curthread;
5852 
5853 	while (eof == 0 && error == 0) {
5854 		error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
5855 		    NULL, NULL);
5856 		if (error != 0)
5857 			break;
5858 		endp = (void *)((uint8_t *)dirent +
5859 		    sizeof(struct dirent) - uio.uio_resid);
5860 		for (dp = dirent; dp < endp;
5861 		     dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
5862 			if (dp->d_type == DT_WHT)
5863 				continue;
5864 			if (dp->d_namlen == 0)
5865 				continue;
5866 			if (dp->d_type != DT_DIR &&
5867 			    dp->d_type != DT_UNKNOWN) {
5868 				error = ENOTEMPTY;
5869 				break;
5870 			}
5871 			if (dp->d_namlen > 2) {
5872 				error = ENOTEMPTY;
5873 				break;
5874 			}
5875 			if (dp->d_namlen == 1 &&
5876 			    dp->d_name[0] != '.') {
5877 				error = ENOTEMPTY;
5878 				break;
5879 			}
5880 			if (dp->d_namlen == 2 &&
5881 			    dp->d_name[1] != '.') {
5882 				error = ENOTEMPTY;
5883 				break;
5884 			}
5885 			uio.uio_resid = sizeof(struct dirent);
5886 		}
5887 	}
5888 	free(dirent, M_TEMP);
5889 	return (error);
5890 }
5891 
5892 int
5893 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5894 {
5895 	int error;
5896 
5897 	if (dp->d_reclen > ap->a_uio->uio_resid)
5898 		return (ENAMETOOLONG);
5899 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
5900 	if (error) {
5901 		if (ap->a_ncookies != NULL) {
5902 			if (ap->a_cookies != NULL)
5903 				free(ap->a_cookies, M_TEMP);
5904 			ap->a_cookies = NULL;
5905 			*ap->a_ncookies = 0;
5906 		}
5907 		return (error);
5908 	}
5909 	if (ap->a_ncookies == NULL)
5910 		return (0);
5911 
5912 	KASSERT(ap->a_cookies,
5913 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
5914 
5915 	*ap->a_cookies = realloc(*ap->a_cookies,
5916 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
5917 	(*ap->a_cookies)[*ap->a_ncookies] = off;
5918 	*ap->a_ncookies += 1;
5919 	return (0);
5920 }
5921 
5922 /*
5923  * Mark for update the access time of the file if the filesystem
5924  * supports VOP_MARKATIME.  This functionality is used by execve and
5925  * mmap, so we want to avoid the I/O implied by directly setting
5926  * va_atime for the sake of efficiency.
5927  */
5928 void
5929 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
5930 {
5931 	struct mount *mp;
5932 
5933 	mp = vp->v_mount;
5934 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
5935 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
5936 		(void)VOP_MARKATIME(vp);
5937 }
5938 
5939 /*
5940  * The purpose of this routine is to remove granularity from accmode_t,
5941  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
5942  * VADMIN and VAPPEND.
5943  *
5944  * If it returns 0, the caller is supposed to continue with the usual
5945  * access checks using 'accmode' as modified by this routine.  If it
5946  * returns nonzero value, the caller is supposed to return that value
5947  * as errno.
5948  *
5949  * Note that after this routine runs, accmode may be zero.
5950  */
5951 int
5952 vfs_unixify_accmode(accmode_t *accmode)
5953 {
5954 	/*
5955 	 * There is no way to specify explicit "deny" rule using
5956 	 * file mode or POSIX.1e ACLs.
5957 	 */
5958 	if (*accmode & VEXPLICIT_DENY) {
5959 		*accmode = 0;
5960 		return (0);
5961 	}
5962 
5963 	/*
5964 	 * None of these can be translated into usual access bits.
5965 	 * Also, the common case for NFSv4 ACLs is to not contain
5966 	 * either of these bits. Caller should check for VWRITE
5967 	 * on the containing directory instead.
5968 	 */
5969 	if (*accmode & (VDELETE_CHILD | VDELETE))
5970 		return (EPERM);
5971 
5972 	if (*accmode & VADMIN_PERMS) {
5973 		*accmode &= ~VADMIN_PERMS;
5974 		*accmode |= VADMIN;
5975 	}
5976 
5977 	/*
5978 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5979 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5980 	 */
5981 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5982 
5983 	return (0);
5984 }
5985 
5986 /*
5987  * Clear out a doomed vnode (if any) and replace it with a new one as long
5988  * as the fs is not being unmounted. Return the root vnode to the caller.
5989  */
5990 static int __noinline
5991 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
5992 {
5993 	struct vnode *vp;
5994 	int error;
5995 
5996 restart:
5997 	if (mp->mnt_rootvnode != NULL) {
5998 		MNT_ILOCK(mp);
5999 		vp = mp->mnt_rootvnode;
6000 		if (vp != NULL) {
6001 			if (!VN_IS_DOOMED(vp)) {
6002 				vrefact(vp);
6003 				MNT_IUNLOCK(mp);
6004 				error = vn_lock(vp, flags);
6005 				if (error == 0) {
6006 					*vpp = vp;
6007 					return (0);
6008 				}
6009 				vrele(vp);
6010 				goto restart;
6011 			}
6012 			/*
6013 			 * Clear the old one.
6014 			 */
6015 			mp->mnt_rootvnode = NULL;
6016 		}
6017 		MNT_IUNLOCK(mp);
6018 		if (vp != NULL) {
6019 			/*
6020 			 * Paired with a fence in vfs_op_thread_exit().
6021 			 */
6022 			atomic_thread_fence_acq();
6023 			vfs_op_barrier_wait(mp);
6024 			vrele(vp);
6025 		}
6026 	}
6027 	error = VFS_CACHEDROOT(mp, flags, vpp);
6028 	if (error != 0)
6029 		return (error);
6030 	if (mp->mnt_vfs_ops == 0) {
6031 		MNT_ILOCK(mp);
6032 		if (mp->mnt_vfs_ops != 0) {
6033 			MNT_IUNLOCK(mp);
6034 			return (0);
6035 		}
6036 		if (mp->mnt_rootvnode == NULL) {
6037 			vrefact(*vpp);
6038 			mp->mnt_rootvnode = *vpp;
6039 		} else {
6040 			if (mp->mnt_rootvnode != *vpp) {
6041 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6042 					panic("%s: mismatch between vnode returned "
6043 					    " by VFS_CACHEDROOT and the one cached "
6044 					    " (%p != %p)",
6045 					    __func__, *vpp, mp->mnt_rootvnode);
6046 				}
6047 			}
6048 		}
6049 		MNT_IUNLOCK(mp);
6050 	}
6051 	return (0);
6052 }
6053 
6054 int
6055 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6056 {
6057 	struct vnode *vp;
6058 	int error;
6059 
6060 	if (!vfs_op_thread_enter(mp))
6061 		return (vfs_cache_root_fallback(mp, flags, vpp));
6062 	vp = (struct vnode *)atomic_load_ptr(&mp->mnt_rootvnode);
6063 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6064 		vfs_op_thread_exit(mp);
6065 		return (vfs_cache_root_fallback(mp, flags, vpp));
6066 	}
6067 	vrefact(vp);
6068 	vfs_op_thread_exit(mp);
6069 	error = vn_lock(vp, flags);
6070 	if (error != 0) {
6071 		vrele(vp);
6072 		return (vfs_cache_root_fallback(mp, flags, vpp));
6073 	}
6074 	*vpp = vp;
6075 	return (0);
6076 }
6077 
6078 struct vnode *
6079 vfs_cache_root_clear(struct mount *mp)
6080 {
6081 	struct vnode *vp;
6082 
6083 	/*
6084 	 * ops > 0 guarantees there is nobody who can see this vnode
6085 	 */
6086 	MPASS(mp->mnt_vfs_ops > 0);
6087 	vp = mp->mnt_rootvnode;
6088 	mp->mnt_rootvnode = NULL;
6089 	return (vp);
6090 }
6091 
6092 void
6093 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6094 {
6095 
6096 	MPASS(mp->mnt_vfs_ops > 0);
6097 	vrefact(vp);
6098 	mp->mnt_rootvnode = vp;
6099 }
6100 
6101 /*
6102  * These are helper functions for filesystems to traverse all
6103  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6104  *
6105  * This interface replaces MNT_VNODE_FOREACH.
6106  */
6107 
6108 
6109 struct vnode *
6110 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6111 {
6112 	struct vnode *vp;
6113 
6114 	if (should_yield())
6115 		kern_yield(PRI_USER);
6116 	MNT_ILOCK(mp);
6117 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6118 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6119 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6120 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6121 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6122 			continue;
6123 		VI_LOCK(vp);
6124 		if (VN_IS_DOOMED(vp)) {
6125 			VI_UNLOCK(vp);
6126 			continue;
6127 		}
6128 		break;
6129 	}
6130 	if (vp == NULL) {
6131 		__mnt_vnode_markerfree_all(mvp, mp);
6132 		/* MNT_IUNLOCK(mp); -- done in above function */
6133 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6134 		return (NULL);
6135 	}
6136 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6137 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6138 	MNT_IUNLOCK(mp);
6139 	return (vp);
6140 }
6141 
6142 struct vnode *
6143 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6144 {
6145 	struct vnode *vp;
6146 
6147 	*mvp = vn_alloc_marker(mp);
6148 	MNT_ILOCK(mp);
6149 	MNT_REF(mp);
6150 
6151 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6152 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6153 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6154 			continue;
6155 		VI_LOCK(vp);
6156 		if (VN_IS_DOOMED(vp)) {
6157 			VI_UNLOCK(vp);
6158 			continue;
6159 		}
6160 		break;
6161 	}
6162 	if (vp == NULL) {
6163 		MNT_REL(mp);
6164 		MNT_IUNLOCK(mp);
6165 		vn_free_marker(*mvp);
6166 		*mvp = NULL;
6167 		return (NULL);
6168 	}
6169 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6170 	MNT_IUNLOCK(mp);
6171 	return (vp);
6172 }
6173 
6174 void
6175 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6176 {
6177 
6178 	if (*mvp == NULL) {
6179 		MNT_IUNLOCK(mp);
6180 		return;
6181 	}
6182 
6183 	mtx_assert(MNT_MTX(mp), MA_OWNED);
6184 
6185 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6186 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6187 	MNT_REL(mp);
6188 	MNT_IUNLOCK(mp);
6189 	vn_free_marker(*mvp);
6190 	*mvp = NULL;
6191 }
6192 
6193 /*
6194  * These are helper functions for filesystems to traverse their
6195  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6196  */
6197 static void
6198 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6199 {
6200 
6201 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6202 
6203 	MNT_ILOCK(mp);
6204 	MNT_REL(mp);
6205 	MNT_IUNLOCK(mp);
6206 	vn_free_marker(*mvp);
6207 	*mvp = NULL;
6208 }
6209 
6210 /*
6211  * Relock the mp mount vnode list lock with the vp vnode interlock in the
6212  * conventional lock order during mnt_vnode_next_lazy iteration.
6213  *
6214  * On entry, the mount vnode list lock is held and the vnode interlock is not.
6215  * The list lock is dropped and reacquired.  On success, both locks are held.
6216  * On failure, the mount vnode list lock is held but the vnode interlock is
6217  * not, and the procedure may have yielded.
6218  */
6219 static bool
6220 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6221     struct vnode *vp)
6222 {
6223 	const struct vnode *tmp;
6224 	bool held, ret;
6225 
6226 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6227 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6228 	    ("%s: bad marker", __func__));
6229 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6230 	    ("%s: inappropriate vnode", __func__));
6231 	ASSERT_VI_UNLOCKED(vp, __func__);
6232 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6233 
6234 	ret = false;
6235 
6236 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6237 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6238 
6239 	/*
6240 	 * Use a hold to prevent vp from disappearing while the mount vnode
6241 	 * list lock is dropped and reacquired.  Normally a hold would be
6242 	 * acquired with vhold(), but that might try to acquire the vnode
6243 	 * interlock, which would be a LOR with the mount vnode list lock.
6244 	 */
6245 	held = refcount_acquire_if_not_zero(&vp->v_holdcnt);
6246 	mtx_unlock(&mp->mnt_listmtx);
6247 	if (!held)
6248 		goto abort;
6249 	VI_LOCK(vp);
6250 	if (!refcount_release_if_not_last(&vp->v_holdcnt)) {
6251 		vdropl(vp);
6252 		goto abort;
6253 	}
6254 	mtx_lock(&mp->mnt_listmtx);
6255 
6256 	/*
6257 	 * Determine whether the vnode is still the next one after the marker,
6258 	 * excepting any other markers.  If the vnode has not been doomed by
6259 	 * vgone() then the hold should have ensured that it remained on the
6260 	 * lazy list.  If it has been doomed but is still on the lazy list,
6261 	 * don't abort, but rather skip over it (avoid spinning on doomed
6262 	 * vnodes).
6263 	 */
6264 	tmp = mvp;
6265 	do {
6266 		tmp = TAILQ_NEXT(tmp, v_lazylist);
6267 	} while (tmp != NULL && tmp->v_type == VMARKER);
6268 	if (tmp != vp) {
6269 		mtx_unlock(&mp->mnt_listmtx);
6270 		VI_UNLOCK(vp);
6271 		goto abort;
6272 	}
6273 
6274 	ret = true;
6275 	goto out;
6276 abort:
6277 	maybe_yield();
6278 	mtx_lock(&mp->mnt_listmtx);
6279 out:
6280 	if (ret)
6281 		ASSERT_VI_LOCKED(vp, __func__);
6282 	else
6283 		ASSERT_VI_UNLOCKED(vp, __func__);
6284 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6285 	return (ret);
6286 }
6287 
6288 static struct vnode *
6289 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6290     void *cbarg)
6291 {
6292 	struct vnode *vp, *nvp;
6293 
6294 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6295 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6296 restart:
6297 	vp = TAILQ_NEXT(*mvp, v_lazylist);
6298 	while (vp != NULL) {
6299 		if (vp->v_type == VMARKER) {
6300 			vp = TAILQ_NEXT(vp, v_lazylist);
6301 			continue;
6302 		}
6303 		/*
6304 		 * See if we want to process the vnode. Note we may encounter a
6305 		 * long string of vnodes we don't care about and hog the list
6306 		 * as a result. Check for it and requeue the marker.
6307 		 */
6308 		if (VN_IS_DOOMED(vp) || !cb(vp, cbarg)) {
6309 			if (!should_yield()) {
6310 				vp = TAILQ_NEXT(vp, v_lazylist);
6311 				continue;
6312 			}
6313 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6314 			    v_lazylist);
6315 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6316 			    v_lazylist);
6317 			mtx_unlock(&mp->mnt_listmtx);
6318 			kern_yield(PRI_USER);
6319 			mtx_lock(&mp->mnt_listmtx);
6320 			goto restart;
6321 		}
6322 		/*
6323 		 * Try-lock because this is the wrong lock order.  If that does
6324 		 * not succeed, drop the mount vnode list lock and try to
6325 		 * reacquire it and the vnode interlock in the right order.
6326 		 */
6327 		if (!VI_TRYLOCK(vp) &&
6328 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6329 			goto restart;
6330 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6331 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6332 		    ("alien vnode on the lazy list %p %p", vp, mp));
6333 		if (vp->v_mount == mp && !VN_IS_DOOMED(vp))
6334 			break;
6335 		nvp = TAILQ_NEXT(vp, v_lazylist);
6336 		VI_UNLOCK(vp);
6337 		vp = nvp;
6338 	}
6339 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6340 
6341 	/* Check if we are done */
6342 	if (vp == NULL) {
6343 		mtx_unlock(&mp->mnt_listmtx);
6344 		mnt_vnode_markerfree_lazy(mvp, mp);
6345 		return (NULL);
6346 	}
6347 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6348 	mtx_unlock(&mp->mnt_listmtx);
6349 	ASSERT_VI_LOCKED(vp, "lazy iter");
6350 	return (vp);
6351 }
6352 
6353 struct vnode *
6354 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6355     void *cbarg)
6356 {
6357 
6358 	if (should_yield())
6359 		kern_yield(PRI_USER);
6360 	mtx_lock(&mp->mnt_listmtx);
6361 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6362 }
6363 
6364 struct vnode *
6365 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6366     void *cbarg)
6367 {
6368 	struct vnode *vp;
6369 
6370 	*mvp = vn_alloc_marker(mp);
6371 	MNT_ILOCK(mp);
6372 	MNT_REF(mp);
6373 	MNT_IUNLOCK(mp);
6374 
6375 	mtx_lock(&mp->mnt_listmtx);
6376 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6377 	if (vp == NULL) {
6378 		mtx_unlock(&mp->mnt_listmtx);
6379 		mnt_vnode_markerfree_lazy(mvp, mp);
6380 		return (NULL);
6381 	}
6382 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6383 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6384 }
6385 
6386 void
6387 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6388 {
6389 
6390 	if (*mvp == NULL)
6391 		return;
6392 
6393 	mtx_lock(&mp->mnt_listmtx);
6394 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6395 	mtx_unlock(&mp->mnt_listmtx);
6396 	mnt_vnode_markerfree_lazy(mvp, mp);
6397 }
6398