xref: /freebsd/sys/kern/vfs_subr.c (revision 57d7e4cc2c6f50edf9b1d54567e03066b8eb42c5)
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     "LU", "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     "LU", "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  * Mark a mount point as busy. Used to synchronize access and to delay
704  * unmounting. Eventually, mountlist_mtx is not released on failure.
705  *
706  * vfs_busy() is a custom lock, it can block the caller.
707  * vfs_busy() only sleeps if the unmount is active on the mount point.
708  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
709  * vnode belonging to mp.
710  *
711  * Lookup uses vfs_busy() to traverse mount points.
712  * root fs			var fs
713  * / vnode lock		A	/ vnode lock (/var)		D
714  * /var vnode lock	B	/log vnode lock(/var/log)	E
715  * vfs_busy lock	C	vfs_busy lock			F
716  *
717  * Within each file system, the lock order is C->A->B and F->D->E.
718  *
719  * When traversing across mounts, the system follows that lock order:
720  *
721  *        C->A->B
722  *              |
723  *              +->F->D->E
724  *
725  * The lookup() process for namei("/var") illustrates the process:
726  *  VOP_LOOKUP() obtains B while A is held
727  *  vfs_busy() obtains a shared lock on F while A and B are held
728  *  vput() releases lock on B
729  *  vput() releases lock on A
730  *  VFS_ROOT() obtains lock on D while shared lock on F is held
731  *  vfs_unbusy() releases shared lock on F
732  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
733  *    Attempt to lock A (instead of vp_crossmp) while D is held would
734  *    violate the global order, causing deadlocks.
735  *
736  * dounmount() locks B while F is drained.
737  */
738 int
739 vfs_busy(struct mount *mp, int flags)
740 {
741 
742 	MPASS((flags & ~MBF_MASK) == 0);
743 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
744 
745 	if (vfs_op_thread_enter(mp)) {
746 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
747 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
748 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
749 		vfs_mp_count_add_pcpu(mp, ref, 1);
750 		vfs_mp_count_add_pcpu(mp, lockref, 1);
751 		vfs_op_thread_exit(mp);
752 		if (flags & MBF_MNTLSTLOCK)
753 			mtx_unlock(&mountlist_mtx);
754 		return (0);
755 	}
756 
757 	MNT_ILOCK(mp);
758 	vfs_assert_mount_counters(mp);
759 	MNT_REF(mp);
760 	/*
761 	 * If mount point is currently being unmounted, sleep until the
762 	 * mount point fate is decided.  If thread doing the unmounting fails,
763 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
764 	 * that this mount point has survived the unmount attempt and vfs_busy
765 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
766 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
767 	 * about to be really destroyed.  vfs_busy needs to release its
768 	 * reference on the mount point in this case and return with ENOENT,
769 	 * telling the caller that mount mount it tried to busy is no longer
770 	 * valid.
771 	 */
772 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
773 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
774 			MNT_REL(mp);
775 			MNT_IUNLOCK(mp);
776 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
777 			    __func__);
778 			return (ENOENT);
779 		}
780 		if (flags & MBF_MNTLSTLOCK)
781 			mtx_unlock(&mountlist_mtx);
782 		mp->mnt_kern_flag |= MNTK_MWAIT;
783 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
784 		if (flags & MBF_MNTLSTLOCK)
785 			mtx_lock(&mountlist_mtx);
786 		MNT_ILOCK(mp);
787 	}
788 	if (flags & MBF_MNTLSTLOCK)
789 		mtx_unlock(&mountlist_mtx);
790 	mp->mnt_lockref++;
791 	MNT_IUNLOCK(mp);
792 	return (0);
793 }
794 
795 /*
796  * Free a busy filesystem.
797  */
798 void
799 vfs_unbusy(struct mount *mp)
800 {
801 	int c;
802 
803 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
804 
805 	if (vfs_op_thread_enter(mp)) {
806 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
807 		vfs_mp_count_sub_pcpu(mp, lockref, 1);
808 		vfs_mp_count_sub_pcpu(mp, ref, 1);
809 		vfs_op_thread_exit(mp);
810 		return;
811 	}
812 
813 	MNT_ILOCK(mp);
814 	vfs_assert_mount_counters(mp);
815 	MNT_REL(mp);
816 	c = --mp->mnt_lockref;
817 	if (mp->mnt_vfs_ops == 0) {
818 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
819 		MNT_IUNLOCK(mp);
820 		return;
821 	}
822 	if (c < 0)
823 		vfs_dump_mount_counters(mp);
824 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
825 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
826 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
827 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
828 		wakeup(&mp->mnt_lockref);
829 	}
830 	MNT_IUNLOCK(mp);
831 }
832 
833 /*
834  * Lookup a mount point by filesystem identifier.
835  */
836 struct mount *
837 vfs_getvfs(fsid_t *fsid)
838 {
839 	struct mount *mp;
840 
841 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
842 	mtx_lock(&mountlist_mtx);
843 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
844 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
845 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
846 			vfs_ref(mp);
847 			mtx_unlock(&mountlist_mtx);
848 			return (mp);
849 		}
850 	}
851 	mtx_unlock(&mountlist_mtx);
852 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
853 	return ((struct mount *) 0);
854 }
855 
856 /*
857  * Lookup a mount point by filesystem identifier, busying it before
858  * returning.
859  *
860  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
861  * cache for popular filesystem identifiers.  The cache is lockess, using
862  * the fact that struct mount's are never freed.  In worst case we may
863  * get pointer to unmounted or even different filesystem, so we have to
864  * check what we got, and go slow way if so.
865  */
866 struct mount *
867 vfs_busyfs(fsid_t *fsid)
868 {
869 #define	FSID_CACHE_SIZE	256
870 	typedef struct mount * volatile vmp_t;
871 	static vmp_t cache[FSID_CACHE_SIZE];
872 	struct mount *mp;
873 	int error;
874 	uint32_t hash;
875 
876 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
877 	hash = fsid->val[0] ^ fsid->val[1];
878 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
879 	mp = cache[hash];
880 	if (mp == NULL ||
881 	    mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
882 	    mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
883 		goto slow;
884 	if (vfs_busy(mp, 0) != 0) {
885 		cache[hash] = NULL;
886 		goto slow;
887 	}
888 	if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
889 	    mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
890 		return (mp);
891 	else
892 	    vfs_unbusy(mp);
893 
894 slow:
895 	mtx_lock(&mountlist_mtx);
896 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
897 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
898 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
899 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
900 			if (error) {
901 				cache[hash] = NULL;
902 				mtx_unlock(&mountlist_mtx);
903 				return (NULL);
904 			}
905 			cache[hash] = mp;
906 			return (mp);
907 		}
908 	}
909 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
910 	mtx_unlock(&mountlist_mtx);
911 	return ((struct mount *) 0);
912 }
913 
914 /*
915  * Check if a user can access privileged mount options.
916  */
917 int
918 vfs_suser(struct mount *mp, struct thread *td)
919 {
920 	int error;
921 
922 	if (jailed(td->td_ucred)) {
923 		/*
924 		 * If the jail of the calling thread lacks permission for
925 		 * this type of file system, deny immediately.
926 		 */
927 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
928 			return (EPERM);
929 
930 		/*
931 		 * If the file system was mounted outside the jail of the
932 		 * calling thread, deny immediately.
933 		 */
934 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
935 			return (EPERM);
936 	}
937 
938 	/*
939 	 * If file system supports delegated administration, we don't check
940 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
941 	 * by the file system itself.
942 	 * If this is not the user that did original mount, we check for
943 	 * the PRIV_VFS_MOUNT_OWNER privilege.
944 	 */
945 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
946 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
947 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
948 			return (error);
949 	}
950 	return (0);
951 }
952 
953 /*
954  * Get a new unique fsid.  Try to make its val[0] unique, since this value
955  * will be used to create fake device numbers for stat().  Also try (but
956  * not so hard) make its val[0] unique mod 2^16, since some emulators only
957  * support 16-bit device numbers.  We end up with unique val[0]'s for the
958  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
959  *
960  * Keep in mind that several mounts may be running in parallel.  Starting
961  * the search one past where the previous search terminated is both a
962  * micro-optimization and a defense against returning the same fsid to
963  * different mounts.
964  */
965 void
966 vfs_getnewfsid(struct mount *mp)
967 {
968 	static uint16_t mntid_base;
969 	struct mount *nmp;
970 	fsid_t tfsid;
971 	int mtype;
972 
973 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
974 	mtx_lock(&mntid_mtx);
975 	mtype = mp->mnt_vfc->vfc_typenum;
976 	tfsid.val[1] = mtype;
977 	mtype = (mtype & 0xFF) << 24;
978 	for (;;) {
979 		tfsid.val[0] = makedev(255,
980 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
981 		mntid_base++;
982 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
983 			break;
984 		vfs_rel(nmp);
985 	}
986 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
987 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
988 	mtx_unlock(&mntid_mtx);
989 }
990 
991 /*
992  * Knob to control the precision of file timestamps:
993  *
994  *   0 = seconds only; nanoseconds zeroed.
995  *   1 = seconds and nanoseconds, accurate within 1/HZ.
996  *   2 = seconds and nanoseconds, truncated to microseconds.
997  * >=3 = seconds and nanoseconds, maximum precision.
998  */
999 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
1000 
1001 static int timestamp_precision = TSP_USEC;
1002 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
1003     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
1004     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1005     "3+: sec + ns (max. precision))");
1006 
1007 /*
1008  * Get a current timestamp.
1009  */
1010 void
1011 vfs_timestamp(struct timespec *tsp)
1012 {
1013 	struct timeval tv;
1014 
1015 	switch (timestamp_precision) {
1016 	case TSP_SEC:
1017 		tsp->tv_sec = time_second;
1018 		tsp->tv_nsec = 0;
1019 		break;
1020 	case TSP_HZ:
1021 		getnanotime(tsp);
1022 		break;
1023 	case TSP_USEC:
1024 		microtime(&tv);
1025 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1026 		break;
1027 	case TSP_NSEC:
1028 	default:
1029 		nanotime(tsp);
1030 		break;
1031 	}
1032 }
1033 
1034 /*
1035  * Set vnode attributes to VNOVAL
1036  */
1037 void
1038 vattr_null(struct vattr *vap)
1039 {
1040 
1041 	vap->va_type = VNON;
1042 	vap->va_size = VNOVAL;
1043 	vap->va_bytes = VNOVAL;
1044 	vap->va_mode = VNOVAL;
1045 	vap->va_nlink = VNOVAL;
1046 	vap->va_uid = VNOVAL;
1047 	vap->va_gid = VNOVAL;
1048 	vap->va_fsid = VNOVAL;
1049 	vap->va_fileid = VNOVAL;
1050 	vap->va_blocksize = VNOVAL;
1051 	vap->va_rdev = VNOVAL;
1052 	vap->va_atime.tv_sec = VNOVAL;
1053 	vap->va_atime.tv_nsec = VNOVAL;
1054 	vap->va_mtime.tv_sec = VNOVAL;
1055 	vap->va_mtime.tv_nsec = VNOVAL;
1056 	vap->va_ctime.tv_sec = VNOVAL;
1057 	vap->va_ctime.tv_nsec = VNOVAL;
1058 	vap->va_birthtime.tv_sec = VNOVAL;
1059 	vap->va_birthtime.tv_nsec = VNOVAL;
1060 	vap->va_flags = VNOVAL;
1061 	vap->va_gen = VNOVAL;
1062 	vap->va_vaflags = 0;
1063 }
1064 
1065 /*
1066  * Try to reduce the total number of vnodes.
1067  *
1068  * This routine (and its user) are buggy in at least the following ways:
1069  * - all parameters were picked years ago when RAM sizes were significantly
1070  *   smaller
1071  * - it can pick vnodes based on pages used by the vm object, but filesystems
1072  *   like ZFS don't use it making the pick broken
1073  * - since ZFS has its own aging policy it gets partially combated by this one
1074  * - a dedicated method should be provided for filesystems to let them decide
1075  *   whether the vnode should be recycled
1076  *
1077  * This routine is called when we have too many vnodes.  It attempts
1078  * to free <count> vnodes and will potentially free vnodes that still
1079  * have VM backing store (VM backing store is typically the cause
1080  * of a vnode blowout so we want to do this).  Therefore, this operation
1081  * is not considered cheap.
1082  *
1083  * A number of conditions may prevent a vnode from being reclaimed.
1084  * the buffer cache may have references on the vnode, a directory
1085  * vnode may still have references due to the namei cache representing
1086  * underlying files, or the vnode may be in active use.   It is not
1087  * desirable to reuse such vnodes.  These conditions may cause the
1088  * number of vnodes to reach some minimum value regardless of what
1089  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
1090  *
1091  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1092  * 			 entries if this argument is strue
1093  * @param trigger	 Only reclaim vnodes with fewer than this many resident
1094  *			 pages.
1095  * @param target	 How many vnodes to reclaim.
1096  * @return		 The number of vnodes that were reclaimed.
1097  */
1098 static int
1099 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1100 {
1101 	struct vnode *vp, *mvp;
1102 	struct mount *mp;
1103 	u_long done;
1104 	bool retried;
1105 
1106 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1107 
1108 	retried = false;
1109 	done = 0;
1110 
1111 	mvp = vnode_list_reclaim_marker;
1112 restart:
1113 	vp = mvp;
1114 	while (done < target) {
1115 		vp = TAILQ_NEXT(vp, v_vnodelist);
1116 		if (__predict_false(vp == NULL))
1117 			break;
1118 
1119 		if (__predict_false(vp->v_type == VMARKER))
1120 			continue;
1121 
1122 		/*
1123 		 * If it's been deconstructed already, it's still
1124 		 * referenced, or it exceeds the trigger, skip it.
1125 		 * Also skip free vnodes.  We are trying to make space
1126 		 * to expand the free list, not reduce it.
1127 		 */
1128 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1129 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1130 			goto next_iter;
1131 
1132 		if (vp->v_type == VBAD || vp->v_type == VNON)
1133 			goto next_iter;
1134 
1135 		if (!VI_TRYLOCK(vp))
1136 			goto next_iter;
1137 
1138 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1139 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1140 		    vp->v_type == VBAD || vp->v_type == VNON ||
1141 		    (vp->v_object != NULL &&
1142 		    vp->v_object->resident_page_count > trigger)) {
1143 			VI_UNLOCK(vp);
1144 			goto next_iter;
1145 		}
1146 		vholdl(vp);
1147 		VI_UNLOCK(vp);
1148 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1149 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1150 		mtx_unlock(&vnode_list_mtx);
1151 
1152 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1153 			vdrop(vp);
1154 			goto next_iter_unlocked;
1155 		}
1156 		if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1157 			vdrop(vp);
1158 			vn_finished_write(mp);
1159 			goto next_iter_unlocked;
1160 		}
1161 
1162 		VI_LOCK(vp);
1163 		if (vp->v_usecount > 0 ||
1164 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1165 		    (vp->v_object != NULL &&
1166 		    vp->v_object->resident_page_count > trigger)) {
1167 			VOP_UNLOCK(vp);
1168 			vdropl(vp);
1169 			vn_finished_write(mp);
1170 			goto next_iter_unlocked;
1171 		}
1172 		counter_u64_add(recycles_count, 1);
1173 		vgonel(vp);
1174 		VOP_UNLOCK(vp);
1175 		vdropl(vp);
1176 		vn_finished_write(mp);
1177 		done++;
1178 next_iter_unlocked:
1179 		if (should_yield())
1180 			kern_yield(PRI_USER);
1181 		mtx_lock(&vnode_list_mtx);
1182 		goto restart;
1183 next_iter:
1184 		MPASS(vp->v_type != VMARKER);
1185 		if (!should_yield())
1186 			continue;
1187 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1188 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1189 		mtx_unlock(&vnode_list_mtx);
1190 		kern_yield(PRI_USER);
1191 		mtx_lock(&vnode_list_mtx);
1192 		goto restart;
1193 	}
1194 	if (done == 0 && !retried) {
1195 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1196 		TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1197 		retried = true;
1198 		goto restart;
1199 	}
1200 	return (done);
1201 }
1202 
1203 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1204 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1205     0,
1206     "limit on vnode free requests per call to the vnlru_free routine");
1207 
1208 /*
1209  * Attempt to reduce the free list by the requested amount.
1210  */
1211 static int
1212 vnlru_free_locked(int count, struct vfsops *mnt_op)
1213 {
1214 	struct vnode *vp, *mvp;
1215 	struct mount *mp;
1216 	int ocount;
1217 
1218 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1219 	if (count > max_vnlru_free)
1220 		count = max_vnlru_free;
1221 	ocount = count;
1222 	mvp = vnode_list_free_marker;
1223 restart:
1224 	vp = mvp;
1225 	while (count > 0) {
1226 		vp = TAILQ_NEXT(vp, v_vnodelist);
1227 		if (__predict_false(vp == NULL)) {
1228 			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1229 			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1230 			break;
1231 		}
1232 		if (__predict_false(vp->v_type == VMARKER))
1233 			continue;
1234 
1235 		/*
1236 		 * Don't recycle if our vnode is from different type
1237 		 * of mount point.  Note that mp is type-safe, the
1238 		 * check does not reach unmapped address even if
1239 		 * vnode is reclaimed.
1240 		 * Don't recycle if we can't get the interlock without
1241 		 * blocking.
1242 		 */
1243 		if (vp->v_holdcnt > 0 || (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1244 		    mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) {
1245 			continue;
1246 		}
1247 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1248 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1249 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1250 			VI_UNLOCK(vp);
1251 			continue;
1252 		}
1253 		vholdl(vp);
1254 		count--;
1255 		mtx_unlock(&vnode_list_mtx);
1256 		VI_UNLOCK(vp);
1257 		vtryrecycle(vp);
1258 		vdrop(vp);
1259 		mtx_lock(&vnode_list_mtx);
1260 		goto restart;
1261 	}
1262 	return (ocount - count);
1263 }
1264 
1265 void
1266 vnlru_free(int count, struct vfsops *mnt_op)
1267 {
1268 
1269 	mtx_lock(&vnode_list_mtx);
1270 	vnlru_free_locked(count, mnt_op);
1271 	mtx_unlock(&vnode_list_mtx);
1272 }
1273 
1274 static void
1275 vnlru_recalc(void)
1276 {
1277 
1278 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1279 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1280 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1281 	vlowat = vhiwat / 2;
1282 }
1283 
1284 /*
1285  * Attempt to recycle vnodes in a context that is always safe to block.
1286  * Calling vlrurecycle() from the bowels of filesystem code has some
1287  * interesting deadlock problems.
1288  */
1289 static struct proc *vnlruproc;
1290 static int vnlruproc_sig;
1291 
1292 /*
1293  * The main freevnodes counter is only updated when threads requeue their vnode
1294  * batches. CPUs are conditionally walked to compute a more accurate total.
1295  *
1296  * Limit how much of a slop are we willing to tolerate. Note: the actual value
1297  * at any given moment can still exceed slop, but it should not be by significant
1298  * margin in practice.
1299  */
1300 #define VNLRU_FREEVNODES_SLOP 128
1301 
1302 static u_long
1303 vnlru_read_freevnodes(void)
1304 {
1305 	struct vdbatch *vd;
1306 	long slop;
1307 	int cpu;
1308 
1309 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1310 	if (freevnodes > freevnodes_old)
1311 		slop = freevnodes - freevnodes_old;
1312 	else
1313 		slop = freevnodes_old - freevnodes;
1314 	if (slop < VNLRU_FREEVNODES_SLOP)
1315 		return (freevnodes >= 0 ? freevnodes : 0);
1316 	freevnodes_old = freevnodes;
1317 	CPU_FOREACH(cpu) {
1318 		vd = DPCPU_ID_PTR((cpu), vd);
1319 		freevnodes_old += vd->freevnodes;
1320 	}
1321 	return (freevnodes_old >= 0 ? freevnodes_old : 0);
1322 }
1323 
1324 static bool
1325 vnlru_under(u_long rnumvnodes, u_long limit)
1326 {
1327 	u_long rfreevnodes, space;
1328 
1329 	if (__predict_false(rnumvnodes > desiredvnodes))
1330 		return (true);
1331 
1332 	space = desiredvnodes - rnumvnodes;
1333 	if (space < limit) {
1334 		rfreevnodes = vnlru_read_freevnodes();
1335 		if (rfreevnodes > wantfreevnodes)
1336 			space += rfreevnodes - wantfreevnodes;
1337 	}
1338 	return (space < limit);
1339 }
1340 
1341 static bool
1342 vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1343 {
1344 	long rfreevnodes, space;
1345 
1346 	if (__predict_false(rnumvnodes > desiredvnodes))
1347 		return (true);
1348 
1349 	space = desiredvnodes - rnumvnodes;
1350 	if (space < limit) {
1351 		rfreevnodes = atomic_load_long(&freevnodes);
1352 		if (rfreevnodes > wantfreevnodes)
1353 			space += rfreevnodes - wantfreevnodes;
1354 	}
1355 	return (space < limit);
1356 }
1357 
1358 static void
1359 vnlru_kick(void)
1360 {
1361 
1362 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1363 	if (vnlruproc_sig == 0) {
1364 		vnlruproc_sig = 1;
1365 		wakeup(vnlruproc);
1366 	}
1367 }
1368 
1369 static void
1370 vnlru_proc(void)
1371 {
1372 	u_long rnumvnodes, rfreevnodes, target;
1373 	unsigned long onumvnodes;
1374 	int done, force, trigger, usevnodes;
1375 	bool reclaim_nc_src, want_reread;
1376 
1377 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1378 	    SHUTDOWN_PRI_FIRST);
1379 
1380 	force = 0;
1381 	want_reread = false;
1382 	for (;;) {
1383 		kproc_suspend_check(vnlruproc);
1384 		mtx_lock(&vnode_list_mtx);
1385 		rnumvnodes = atomic_load_long(&numvnodes);
1386 
1387 		if (want_reread) {
1388 			force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1389 			want_reread = false;
1390 		}
1391 
1392 		/*
1393 		 * If numvnodes is too large (due to desiredvnodes being
1394 		 * adjusted using its sysctl, or emergency growth), first
1395 		 * try to reduce it by discarding from the free list.
1396 		 */
1397 		if (rnumvnodes > desiredvnodes) {
1398 			vnlru_free_locked(rnumvnodes - desiredvnodes, NULL);
1399 			rnumvnodes = atomic_load_long(&numvnodes);
1400 		}
1401 		/*
1402 		 * Sleep if the vnode cache is in a good state.  This is
1403 		 * when it is not over-full and has space for about a 4%
1404 		 * or 9% expansion (by growing its size or inexcessively
1405 		 * reducing its free list).  Otherwise, try to reclaim
1406 		 * space for a 10% expansion.
1407 		 */
1408 		if (vstir && force == 0) {
1409 			force = 1;
1410 			vstir = 0;
1411 		}
1412 		if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1413 			vnlruproc_sig = 0;
1414 			wakeup(&vnlruproc_sig);
1415 			msleep(vnlruproc, &vnode_list_mtx,
1416 			    PVFS|PDROP, "vlruwt", hz);
1417 			continue;
1418 		}
1419 		rfreevnodes = vnlru_read_freevnodes();
1420 
1421 		onumvnodes = rnumvnodes;
1422 		/*
1423 		 * Calculate parameters for recycling.  These are the same
1424 		 * throughout the loop to give some semblance of fairness.
1425 		 * The trigger point is to avoid recycling vnodes with lots
1426 		 * of resident pages.  We aren't trying to free memory; we
1427 		 * are trying to recycle or at least free vnodes.
1428 		 */
1429 		if (rnumvnodes <= desiredvnodes)
1430 			usevnodes = rnumvnodes - rfreevnodes;
1431 		else
1432 			usevnodes = rnumvnodes;
1433 		if (usevnodes <= 0)
1434 			usevnodes = 1;
1435 		/*
1436 		 * The trigger value is is chosen to give a conservatively
1437 		 * large value to ensure that it alone doesn't prevent
1438 		 * making progress.  The value can easily be so large that
1439 		 * it is effectively infinite in some congested and
1440 		 * misconfigured cases, and this is necessary.  Normally
1441 		 * it is about 8 to 100 (pages), which is quite large.
1442 		 */
1443 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1444 		if (force < 2)
1445 			trigger = vsmalltrigger;
1446 		reclaim_nc_src = force >= 3;
1447 		target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1448 		target = target / 10 + 1;
1449 		done = vlrureclaim(reclaim_nc_src, trigger, target);
1450 		mtx_unlock(&vnode_list_mtx);
1451 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1452 			uma_reclaim(UMA_RECLAIM_DRAIN);
1453 		if (done == 0) {
1454 			if (force == 0 || force == 1) {
1455 				force = 2;
1456 				continue;
1457 			}
1458 			if (force == 2) {
1459 				force = 3;
1460 				continue;
1461 			}
1462 			want_reread = true;
1463 			force = 0;
1464 			vnlru_nowhere++;
1465 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1466 		} else {
1467 			want_reread = true;
1468 			kern_yield(PRI_USER);
1469 		}
1470 	}
1471 }
1472 
1473 static struct kproc_desc vnlru_kp = {
1474 	"vnlru",
1475 	vnlru_proc,
1476 	&vnlruproc
1477 };
1478 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1479     &vnlru_kp);
1480 
1481 /*
1482  * Routines having to do with the management of the vnode table.
1483  */
1484 
1485 /*
1486  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1487  * before we actually vgone().  This function must be called with the vnode
1488  * held to prevent the vnode from being returned to the free list midway
1489  * through vgone().
1490  */
1491 static int
1492 vtryrecycle(struct vnode *vp)
1493 {
1494 	struct mount *vnmp;
1495 
1496 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1497 	VNASSERT(vp->v_holdcnt, vp,
1498 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1499 	/*
1500 	 * This vnode may found and locked via some other list, if so we
1501 	 * can't recycle it yet.
1502 	 */
1503 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1504 		CTR2(KTR_VFS,
1505 		    "%s: impossible to recycle, vp %p lock is already held",
1506 		    __func__, vp);
1507 		return (EWOULDBLOCK);
1508 	}
1509 	/*
1510 	 * Don't recycle if its filesystem is being suspended.
1511 	 */
1512 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1513 		VOP_UNLOCK(vp);
1514 		CTR2(KTR_VFS,
1515 		    "%s: impossible to recycle, cannot start the write for %p",
1516 		    __func__, vp);
1517 		return (EBUSY);
1518 	}
1519 	/*
1520 	 * If we got this far, we need to acquire the interlock and see if
1521 	 * anyone picked up this vnode from another list.  If not, we will
1522 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1523 	 * will skip over it.
1524 	 */
1525 	VI_LOCK(vp);
1526 	if (vp->v_usecount) {
1527 		VOP_UNLOCK(vp);
1528 		VI_UNLOCK(vp);
1529 		vn_finished_write(vnmp);
1530 		CTR2(KTR_VFS,
1531 		    "%s: impossible to recycle, %p is already referenced",
1532 		    __func__, vp);
1533 		return (EBUSY);
1534 	}
1535 	if (!VN_IS_DOOMED(vp)) {
1536 		counter_u64_add(recycles_free_count, 1);
1537 		vgonel(vp);
1538 	}
1539 	VOP_UNLOCK(vp);
1540 	VI_UNLOCK(vp);
1541 	vn_finished_write(vnmp);
1542 	return (0);
1543 }
1544 
1545 /*
1546  * Allocate a new vnode.
1547  *
1548  * The operation never returns an error. Returning an error was disabled
1549  * in r145385 (dated 2005) with the following comment:
1550  *
1551  * XXX Not all VFS_VGET/ffs_vget callers check returns.
1552  *
1553  * Given the age of this commit (almost 15 years at the time of writing this
1554  * comment) restoring the ability to fail requires a significant audit of
1555  * all codepaths.
1556  *
1557  * The routine can try to free a vnode or stall for up to 1 second waiting for
1558  * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1559  */
1560 static u_long vn_alloc_cyclecount;
1561 
1562 static struct vnode * __noinline
1563 vn_alloc_hard(struct mount *mp)
1564 {
1565 	u_long rnumvnodes, rfreevnodes;
1566 
1567 	mtx_lock(&vnode_list_mtx);
1568 	rnumvnodes = atomic_load_long(&numvnodes);
1569 	if (rnumvnodes + 1 < desiredvnodes) {
1570 		vn_alloc_cyclecount = 0;
1571 		goto alloc;
1572 	}
1573 	rfreevnodes = vnlru_read_freevnodes();
1574 	if (vn_alloc_cyclecount++ >= rfreevnodes) {
1575 		vn_alloc_cyclecount = 0;
1576 		vstir = 1;
1577 	}
1578 	/*
1579 	 * Grow the vnode cache if it will not be above its target max
1580 	 * after growing.  Otherwise, if the free list is nonempty, try
1581 	 * to reclaim 1 item from it before growing the cache (possibly
1582 	 * above its target max if the reclamation failed or is delayed).
1583 	 * Otherwise, wait for some space.  In all cases, schedule
1584 	 * vnlru_proc() if we are getting short of space.  The watermarks
1585 	 * should be chosen so that we never wait or even reclaim from
1586 	 * the free list to below its target minimum.
1587 	 */
1588 	if (vnlru_free_locked(1, NULL) > 0)
1589 		goto alloc;
1590 	if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1591 		/*
1592 		 * Wait for space for a new vnode.
1593 		 */
1594 		vnlru_kick();
1595 		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1596 		if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1597 		    vnlru_read_freevnodes() > 1)
1598 			vnlru_free_locked(1, NULL);
1599 	}
1600 alloc:
1601 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1602 	if (vnlru_under(rnumvnodes, vlowat))
1603 		vnlru_kick();
1604 	mtx_unlock(&vnode_list_mtx);
1605 	return (uma_zalloc(vnode_zone, M_WAITOK));
1606 }
1607 
1608 static struct vnode *
1609 vn_alloc(struct mount *mp)
1610 {
1611 	u_long rnumvnodes;
1612 
1613 	if (__predict_false(vn_alloc_cyclecount != 0))
1614 		return (vn_alloc_hard(mp));
1615 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1616 	if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1617 		atomic_subtract_long(&numvnodes, 1);
1618 		return (vn_alloc_hard(mp));
1619 	}
1620 
1621 	return (uma_zalloc(vnode_zone, M_WAITOK));
1622 }
1623 
1624 static void
1625 vn_free(struct vnode *vp)
1626 {
1627 
1628 	atomic_subtract_long(&numvnodes, 1);
1629 	uma_zfree(vnode_zone, vp);
1630 }
1631 
1632 /*
1633  * Return the next vnode from the free list.
1634  */
1635 int
1636 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1637     struct vnode **vpp)
1638 {
1639 	struct vnode *vp;
1640 	struct thread *td;
1641 	struct lock_object *lo;
1642 
1643 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1644 
1645 	KASSERT(vops->registered,
1646 	    ("%s: not registered vector op %p\n", __func__, vops));
1647 
1648 	td = curthread;
1649 	if (td->td_vp_reserved != NULL) {
1650 		vp = td->td_vp_reserved;
1651 		td->td_vp_reserved = NULL;
1652 	} else {
1653 		vp = vn_alloc(mp);
1654 	}
1655 	counter_u64_add(vnodes_created, 1);
1656 	/*
1657 	 * Locks are given the generic name "vnode" when created.
1658 	 * Follow the historic practice of using the filesystem
1659 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1660 	 *
1661 	 * Locks live in a witness group keyed on their name. Thus,
1662 	 * when a lock is renamed, it must also move from the witness
1663 	 * group of its old name to the witness group of its new name.
1664 	 *
1665 	 * The change only needs to be made when the vnode moves
1666 	 * from one filesystem type to another. We ensure that each
1667 	 * filesystem use a single static name pointer for its tag so
1668 	 * that we can compare pointers rather than doing a strcmp().
1669 	 */
1670 	lo = &vp->v_vnlock->lock_object;
1671 #ifdef WITNESS
1672 	if (lo->lo_name != tag) {
1673 #endif
1674 		lo->lo_name = tag;
1675 #ifdef WITNESS
1676 		WITNESS_DESTROY(lo);
1677 		WITNESS_INIT(lo, tag);
1678 	}
1679 #endif
1680 	/*
1681 	 * By default, don't allow shared locks unless filesystems opt-in.
1682 	 */
1683 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1684 	/*
1685 	 * Finalize various vnode identity bits.
1686 	 */
1687 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1688 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1689 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1690 	vp->v_type = VNON;
1691 	vp->v_op = vops;
1692 	v_init_counters(vp);
1693 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1694 #ifdef DIAGNOSTIC
1695 	if (mp == NULL && vops != &dead_vnodeops)
1696 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1697 #endif
1698 #ifdef MAC
1699 	mac_vnode_init(vp);
1700 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1701 		mac_vnode_associate_singlelabel(mp, vp);
1702 #endif
1703 	if (mp != NULL) {
1704 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1705 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1706 			vp->v_vflag |= VV_NOKNOTE;
1707 	}
1708 
1709 	/*
1710 	 * For the filesystems which do not use vfs_hash_insert(),
1711 	 * still initialize v_hash to have vfs_hash_index() useful.
1712 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1713 	 * its own hashing.
1714 	 */
1715 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1716 
1717 	*vpp = vp;
1718 	return (0);
1719 }
1720 
1721 void
1722 getnewvnode_reserve(void)
1723 {
1724 	struct thread *td;
1725 
1726 	td = curthread;
1727 	MPASS(td->td_vp_reserved == NULL);
1728 	td->td_vp_reserved = vn_alloc(NULL);
1729 }
1730 
1731 void
1732 getnewvnode_drop_reserve(void)
1733 {
1734 	struct thread *td;
1735 
1736 	td = curthread;
1737 	if (td->td_vp_reserved != NULL) {
1738 		vn_free(td->td_vp_reserved);
1739 		td->td_vp_reserved = NULL;
1740 	}
1741 }
1742 
1743 static void
1744 freevnode(struct vnode *vp)
1745 {
1746 	struct bufobj *bo;
1747 
1748 	/*
1749 	 * The vnode has been marked for destruction, so free it.
1750 	 *
1751 	 * The vnode will be returned to the zone where it will
1752 	 * normally remain until it is needed for another vnode. We
1753 	 * need to cleanup (or verify that the cleanup has already
1754 	 * been done) any residual data left from its current use
1755 	 * so as not to contaminate the freshly allocated vnode.
1756 	 */
1757 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1758 	bo = &vp->v_bufobj;
1759 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1760 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
1761 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1762 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1763 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1764 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1765 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1766 	    ("clean blk trie not empty"));
1767 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1768 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1769 	    ("dirty blk trie not empty"));
1770 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
1771 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
1772 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
1773 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1774 	    ("Dangling rangelock waiters"));
1775 	VI_UNLOCK(vp);
1776 #ifdef MAC
1777 	mac_vnode_destroy(vp);
1778 #endif
1779 	if (vp->v_pollinfo != NULL) {
1780 		destroy_vpollinfo(vp->v_pollinfo);
1781 		vp->v_pollinfo = NULL;
1782 	}
1783 #ifdef INVARIANTS
1784 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
1785 	vp->v_op = NULL;
1786 #endif
1787 	vp->v_mountedhere = NULL;
1788 	vp->v_unpcb = NULL;
1789 	vp->v_rdev = NULL;
1790 	vp->v_fifoinfo = NULL;
1791 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
1792 	vp->v_irflag = 0;
1793 	vp->v_iflag = 0;
1794 	vp->v_vflag = 0;
1795 	bo->bo_flag = 0;
1796 	vn_free(vp);
1797 }
1798 
1799 /*
1800  * Delete from old mount point vnode list, if on one.
1801  */
1802 static void
1803 delmntque(struct vnode *vp)
1804 {
1805 	struct mount *mp;
1806 
1807 	VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
1808 
1809 	mp = vp->v_mount;
1810 	if (mp == NULL)
1811 		return;
1812 	MNT_ILOCK(mp);
1813 	VI_LOCK(vp);
1814 	vp->v_mount = NULL;
1815 	VI_UNLOCK(vp);
1816 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1817 		("bad mount point vnode list size"));
1818 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1819 	mp->mnt_nvnodelistsize--;
1820 	MNT_REL(mp);
1821 	MNT_IUNLOCK(mp);
1822 }
1823 
1824 static void
1825 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1826 {
1827 
1828 	vp->v_data = NULL;
1829 	vp->v_op = &dead_vnodeops;
1830 	vgone(vp);
1831 	vput(vp);
1832 }
1833 
1834 /*
1835  * Insert into list of vnodes for the new mount point, if available.
1836  */
1837 int
1838 insmntque1(struct vnode *vp, struct mount *mp,
1839 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1840 {
1841 
1842 	KASSERT(vp->v_mount == NULL,
1843 		("insmntque: vnode already on per mount vnode list"));
1844 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1845 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1846 
1847 	/*
1848 	 * We acquire the vnode interlock early to ensure that the
1849 	 * vnode cannot be recycled by another process releasing a
1850 	 * holdcnt on it before we get it on both the vnode list
1851 	 * and the active vnode list. The mount mutex protects only
1852 	 * manipulation of the vnode list and the vnode freelist
1853 	 * mutex protects only manipulation of the active vnode list.
1854 	 * Hence the need to hold the vnode interlock throughout.
1855 	 */
1856 	MNT_ILOCK(mp);
1857 	VI_LOCK(vp);
1858 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1859 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1860 	    mp->mnt_nvnodelistsize == 0)) &&
1861 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1862 		VI_UNLOCK(vp);
1863 		MNT_IUNLOCK(mp);
1864 		if (dtr != NULL)
1865 			dtr(vp, dtr_arg);
1866 		return (EBUSY);
1867 	}
1868 	vp->v_mount = mp;
1869 	MNT_REF(mp);
1870 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1871 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1872 		("neg mount point vnode list size"));
1873 	mp->mnt_nvnodelistsize++;
1874 	VI_UNLOCK(vp);
1875 	MNT_IUNLOCK(mp);
1876 	return (0);
1877 }
1878 
1879 int
1880 insmntque(struct vnode *vp, struct mount *mp)
1881 {
1882 
1883 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1884 }
1885 
1886 /*
1887  * Flush out and invalidate all buffers associated with a bufobj
1888  * Called with the underlying object locked.
1889  */
1890 int
1891 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1892 {
1893 	int error;
1894 
1895 	BO_LOCK(bo);
1896 	if (flags & V_SAVE) {
1897 		error = bufobj_wwait(bo, slpflag, slptimeo);
1898 		if (error) {
1899 			BO_UNLOCK(bo);
1900 			return (error);
1901 		}
1902 		if (bo->bo_dirty.bv_cnt > 0) {
1903 			BO_UNLOCK(bo);
1904 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1905 				return (error);
1906 			/*
1907 			 * XXX We could save a lock/unlock if this was only
1908 			 * enabled under INVARIANTS
1909 			 */
1910 			BO_LOCK(bo);
1911 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1912 				panic("vinvalbuf: dirty bufs");
1913 		}
1914 	}
1915 	/*
1916 	 * If you alter this loop please notice that interlock is dropped and
1917 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1918 	 * no race conditions occur from this.
1919 	 */
1920 	do {
1921 		error = flushbuflist(&bo->bo_clean,
1922 		    flags, bo, slpflag, slptimeo);
1923 		if (error == 0 && !(flags & V_CLEANONLY))
1924 			error = flushbuflist(&bo->bo_dirty,
1925 			    flags, bo, slpflag, slptimeo);
1926 		if (error != 0 && error != EAGAIN) {
1927 			BO_UNLOCK(bo);
1928 			return (error);
1929 		}
1930 	} while (error != 0);
1931 
1932 	/*
1933 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1934 	 * have write I/O in-progress but if there is a VM object then the
1935 	 * VM object can also have read-I/O in-progress.
1936 	 */
1937 	do {
1938 		bufobj_wwait(bo, 0, 0);
1939 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
1940 			BO_UNLOCK(bo);
1941 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
1942 			BO_LOCK(bo);
1943 		}
1944 	} while (bo->bo_numoutput > 0);
1945 	BO_UNLOCK(bo);
1946 
1947 	/*
1948 	 * Destroy the copy in the VM cache, too.
1949 	 */
1950 	if (bo->bo_object != NULL &&
1951 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1952 		VM_OBJECT_WLOCK(bo->bo_object);
1953 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1954 		    OBJPR_CLEANONLY : 0);
1955 		VM_OBJECT_WUNLOCK(bo->bo_object);
1956 	}
1957 
1958 #ifdef INVARIANTS
1959 	BO_LOCK(bo);
1960 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1961 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1962 	    bo->bo_clean.bv_cnt > 0))
1963 		panic("vinvalbuf: flush failed");
1964 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1965 	    bo->bo_dirty.bv_cnt > 0)
1966 		panic("vinvalbuf: flush dirty failed");
1967 	BO_UNLOCK(bo);
1968 #endif
1969 	return (0);
1970 }
1971 
1972 /*
1973  * Flush out and invalidate all buffers associated with a vnode.
1974  * Called with the underlying object locked.
1975  */
1976 int
1977 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1978 {
1979 
1980 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1981 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1982 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1983 		return (0);
1984 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1985 }
1986 
1987 /*
1988  * Flush out buffers on the specified list.
1989  *
1990  */
1991 static int
1992 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1993     int slptimeo)
1994 {
1995 	struct buf *bp, *nbp;
1996 	int retval, error;
1997 	daddr_t lblkno;
1998 	b_xflags_t xflags;
1999 
2000 	ASSERT_BO_WLOCKED(bo);
2001 
2002 	retval = 0;
2003 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2004 		/*
2005 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2006 		 * do not skip any buffers. If we are flushing only V_NORMAL
2007 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2008 		 * flushing only V_ALT buffers then skip buffers not marked
2009 		 * as BX_ALTDATA.
2010 		 */
2011 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2012 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2013 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2014 			continue;
2015 		}
2016 		if (nbp != NULL) {
2017 			lblkno = nbp->b_lblkno;
2018 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2019 		}
2020 		retval = EAGAIN;
2021 		error = BUF_TIMELOCK(bp,
2022 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2023 		    "flushbuf", slpflag, slptimeo);
2024 		if (error) {
2025 			BO_LOCK(bo);
2026 			return (error != ENOLCK ? error : EAGAIN);
2027 		}
2028 		KASSERT(bp->b_bufobj == bo,
2029 		    ("bp %p wrong b_bufobj %p should be %p",
2030 		    bp, bp->b_bufobj, bo));
2031 		/*
2032 		 * XXX Since there are no node locks for NFS, I
2033 		 * believe there is a slight chance that a delayed
2034 		 * write will occur while sleeping just above, so
2035 		 * check for it.
2036 		 */
2037 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2038 		    (flags & V_SAVE)) {
2039 			bremfree(bp);
2040 			bp->b_flags |= B_ASYNC;
2041 			bwrite(bp);
2042 			BO_LOCK(bo);
2043 			return (EAGAIN);	/* XXX: why not loop ? */
2044 		}
2045 		bremfree(bp);
2046 		bp->b_flags |= (B_INVAL | B_RELBUF);
2047 		bp->b_flags &= ~B_ASYNC;
2048 		brelse(bp);
2049 		BO_LOCK(bo);
2050 		if (nbp == NULL)
2051 			break;
2052 		nbp = gbincore(bo, lblkno);
2053 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2054 		    != xflags)
2055 			break;			/* nbp invalid */
2056 	}
2057 	return (retval);
2058 }
2059 
2060 int
2061 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2062 {
2063 	struct buf *bp;
2064 	int error;
2065 	daddr_t lblkno;
2066 
2067 	ASSERT_BO_LOCKED(bo);
2068 
2069 	for (lblkno = startn;;) {
2070 again:
2071 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2072 		if (bp == NULL || bp->b_lblkno >= endn ||
2073 		    bp->b_lblkno < startn)
2074 			break;
2075 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2076 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2077 		if (error != 0) {
2078 			BO_RLOCK(bo);
2079 			if (error == ENOLCK)
2080 				goto again;
2081 			return (error);
2082 		}
2083 		KASSERT(bp->b_bufobj == bo,
2084 		    ("bp %p wrong b_bufobj %p should be %p",
2085 		    bp, bp->b_bufobj, bo));
2086 		lblkno = bp->b_lblkno + 1;
2087 		if ((bp->b_flags & B_MANAGED) == 0)
2088 			bremfree(bp);
2089 		bp->b_flags |= B_RELBUF;
2090 		/*
2091 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2092 		 * pages backing each buffer in the range are unlikely to be
2093 		 * reused.  Dirty buffers will have the hint applied once
2094 		 * they've been written.
2095 		 */
2096 		if ((bp->b_flags & B_VMIO) != 0)
2097 			bp->b_flags |= B_NOREUSE;
2098 		brelse(bp);
2099 		BO_RLOCK(bo);
2100 	}
2101 	return (0);
2102 }
2103 
2104 /*
2105  * Truncate a file's buffer and pages to a specified length.  This
2106  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2107  * sync activity.
2108  */
2109 int
2110 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2111 {
2112 	struct buf *bp, *nbp;
2113 	struct bufobj *bo;
2114 	daddr_t startlbn;
2115 
2116 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2117 	    vp, blksize, (uintmax_t)length);
2118 
2119 	/*
2120 	 * Round up to the *next* lbn.
2121 	 */
2122 	startlbn = howmany(length, blksize);
2123 
2124 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2125 
2126 	bo = &vp->v_bufobj;
2127 restart_unlocked:
2128 	BO_LOCK(bo);
2129 
2130 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2131 		;
2132 
2133 	if (length > 0) {
2134 restartsync:
2135 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2136 			if (bp->b_lblkno > 0)
2137 				continue;
2138 			/*
2139 			 * Since we hold the vnode lock this should only
2140 			 * fail if we're racing with the buf daemon.
2141 			 */
2142 			if (BUF_LOCK(bp,
2143 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2144 			    BO_LOCKPTR(bo)) == ENOLCK)
2145 				goto restart_unlocked;
2146 
2147 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2148 			    ("buf(%p) on dirty queue without DELWRI", bp));
2149 
2150 			bremfree(bp);
2151 			bawrite(bp);
2152 			BO_LOCK(bo);
2153 			goto restartsync;
2154 		}
2155 	}
2156 
2157 	bufobj_wwait(bo, 0, 0);
2158 	BO_UNLOCK(bo);
2159 	vnode_pager_setsize(vp, length);
2160 
2161 	return (0);
2162 }
2163 
2164 /*
2165  * Invalidate the cached pages of a file's buffer within the range of block
2166  * numbers [startlbn, endlbn).
2167  */
2168 void
2169 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2170     int blksize)
2171 {
2172 	struct bufobj *bo;
2173 	off_t start, end;
2174 
2175 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2176 
2177 	start = blksize * startlbn;
2178 	end = blksize * endlbn;
2179 
2180 	bo = &vp->v_bufobj;
2181 	BO_LOCK(bo);
2182 	MPASS(blksize == bo->bo_bsize);
2183 
2184 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2185 		;
2186 
2187 	BO_UNLOCK(bo);
2188 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2189 }
2190 
2191 static int
2192 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2193     daddr_t startlbn, daddr_t endlbn)
2194 {
2195 	struct buf *bp, *nbp;
2196 	bool anyfreed;
2197 
2198 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2199 	ASSERT_BO_LOCKED(bo);
2200 
2201 	do {
2202 		anyfreed = false;
2203 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2204 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2205 				continue;
2206 			if (BUF_LOCK(bp,
2207 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2208 			    BO_LOCKPTR(bo)) == ENOLCK) {
2209 				BO_LOCK(bo);
2210 				return (EAGAIN);
2211 			}
2212 
2213 			bremfree(bp);
2214 			bp->b_flags |= B_INVAL | B_RELBUF;
2215 			bp->b_flags &= ~B_ASYNC;
2216 			brelse(bp);
2217 			anyfreed = true;
2218 
2219 			BO_LOCK(bo);
2220 			if (nbp != NULL &&
2221 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2222 			    nbp->b_vp != vp ||
2223 			    (nbp->b_flags & B_DELWRI) != 0))
2224 				return (EAGAIN);
2225 		}
2226 
2227 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2228 			if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2229 				continue;
2230 			if (BUF_LOCK(bp,
2231 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2232 			    BO_LOCKPTR(bo)) == ENOLCK) {
2233 				BO_LOCK(bo);
2234 				return (EAGAIN);
2235 			}
2236 			bremfree(bp);
2237 			bp->b_flags |= B_INVAL | B_RELBUF;
2238 			bp->b_flags &= ~B_ASYNC;
2239 			brelse(bp);
2240 			anyfreed = true;
2241 
2242 			BO_LOCK(bo);
2243 			if (nbp != NULL &&
2244 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2245 			    (nbp->b_vp != vp) ||
2246 			    (nbp->b_flags & B_DELWRI) == 0))
2247 				return (EAGAIN);
2248 		}
2249 	} while (anyfreed);
2250 	return (0);
2251 }
2252 
2253 static void
2254 buf_vlist_remove(struct buf *bp)
2255 {
2256 	struct bufv *bv;
2257 
2258 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2259 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2260 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
2261 	    (BX_VNDIRTY|BX_VNCLEAN),
2262 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
2263 	if (bp->b_xflags & BX_VNDIRTY)
2264 		bv = &bp->b_bufobj->bo_dirty;
2265 	else
2266 		bv = &bp->b_bufobj->bo_clean;
2267 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2268 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2269 	bv->bv_cnt--;
2270 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2271 }
2272 
2273 /*
2274  * Add the buffer to the sorted clean or dirty block list.
2275  *
2276  * NOTE: xflags is passed as a constant, optimizing this inline function!
2277  */
2278 static void
2279 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2280 {
2281 	struct bufv *bv;
2282 	struct buf *n;
2283 	int error;
2284 
2285 	ASSERT_BO_WLOCKED(bo);
2286 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2287 	    ("dead bo %p", bo));
2288 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2289 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2290 	bp->b_xflags |= xflags;
2291 	if (xflags & BX_VNDIRTY)
2292 		bv = &bo->bo_dirty;
2293 	else
2294 		bv = &bo->bo_clean;
2295 
2296 	/*
2297 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2298 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2299 	 * than _ge.
2300 	 */
2301 	if (bv->bv_cnt == 0 ||
2302 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2303 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2304 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2305 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2306 	else
2307 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2308 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2309 	if (error)
2310 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2311 	bv->bv_cnt++;
2312 }
2313 
2314 /*
2315  * Look up a buffer using the buffer tries.
2316  */
2317 struct buf *
2318 gbincore(struct bufobj *bo, daddr_t lblkno)
2319 {
2320 	struct buf *bp;
2321 
2322 	ASSERT_BO_LOCKED(bo);
2323 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2324 	if (bp != NULL)
2325 		return (bp);
2326 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
2327 }
2328 
2329 /*
2330  * Associate a buffer with a vnode.
2331  */
2332 void
2333 bgetvp(struct vnode *vp, struct buf *bp)
2334 {
2335 	struct bufobj *bo;
2336 
2337 	bo = &vp->v_bufobj;
2338 	ASSERT_BO_WLOCKED(bo);
2339 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2340 
2341 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2342 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2343 	    ("bgetvp: bp already attached! %p", bp));
2344 
2345 	vhold(vp);
2346 	bp->b_vp = vp;
2347 	bp->b_bufobj = bo;
2348 	/*
2349 	 * Insert onto list for new vnode.
2350 	 */
2351 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2352 }
2353 
2354 /*
2355  * Disassociate a buffer from a vnode.
2356  */
2357 void
2358 brelvp(struct buf *bp)
2359 {
2360 	struct bufobj *bo;
2361 	struct vnode *vp;
2362 
2363 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2364 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2365 
2366 	/*
2367 	 * Delete from old vnode list, if on one.
2368 	 */
2369 	vp = bp->b_vp;		/* XXX */
2370 	bo = bp->b_bufobj;
2371 	BO_LOCK(bo);
2372 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2373 		buf_vlist_remove(bp);
2374 	else
2375 		panic("brelvp: Buffer %p not on queue.", bp);
2376 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2377 		bo->bo_flag &= ~BO_ONWORKLST;
2378 		mtx_lock(&sync_mtx);
2379 		LIST_REMOVE(bo, bo_synclist);
2380 		syncer_worklist_len--;
2381 		mtx_unlock(&sync_mtx);
2382 	}
2383 	bp->b_vp = NULL;
2384 	bp->b_bufobj = NULL;
2385 	BO_UNLOCK(bo);
2386 	vdrop(vp);
2387 }
2388 
2389 /*
2390  * Add an item to the syncer work queue.
2391  */
2392 static void
2393 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2394 {
2395 	int slot;
2396 
2397 	ASSERT_BO_WLOCKED(bo);
2398 
2399 	mtx_lock(&sync_mtx);
2400 	if (bo->bo_flag & BO_ONWORKLST)
2401 		LIST_REMOVE(bo, bo_synclist);
2402 	else {
2403 		bo->bo_flag |= BO_ONWORKLST;
2404 		syncer_worklist_len++;
2405 	}
2406 
2407 	if (delay > syncer_maxdelay - 2)
2408 		delay = syncer_maxdelay - 2;
2409 	slot = (syncer_delayno + delay) & syncer_mask;
2410 
2411 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2412 	mtx_unlock(&sync_mtx);
2413 }
2414 
2415 static int
2416 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2417 {
2418 	int error, len;
2419 
2420 	mtx_lock(&sync_mtx);
2421 	len = syncer_worklist_len - sync_vnode_count;
2422 	mtx_unlock(&sync_mtx);
2423 	error = SYSCTL_OUT(req, &len, sizeof(len));
2424 	return (error);
2425 }
2426 
2427 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2428     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2429     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2430 
2431 static struct proc *updateproc;
2432 static void sched_sync(void);
2433 static struct kproc_desc up_kp = {
2434 	"syncer",
2435 	sched_sync,
2436 	&updateproc
2437 };
2438 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2439 
2440 static int
2441 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2442 {
2443 	struct vnode *vp;
2444 	struct mount *mp;
2445 
2446 	*bo = LIST_FIRST(slp);
2447 	if (*bo == NULL)
2448 		return (0);
2449 	vp = bo2vnode(*bo);
2450 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2451 		return (1);
2452 	/*
2453 	 * We use vhold in case the vnode does not
2454 	 * successfully sync.  vhold prevents the vnode from
2455 	 * going away when we unlock the sync_mtx so that
2456 	 * we can acquire the vnode interlock.
2457 	 */
2458 	vholdl(vp);
2459 	mtx_unlock(&sync_mtx);
2460 	VI_UNLOCK(vp);
2461 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2462 		vdrop(vp);
2463 		mtx_lock(&sync_mtx);
2464 		return (*bo == LIST_FIRST(slp));
2465 	}
2466 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2467 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2468 	VOP_UNLOCK(vp);
2469 	vn_finished_write(mp);
2470 	BO_LOCK(*bo);
2471 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2472 		/*
2473 		 * Put us back on the worklist.  The worklist
2474 		 * routine will remove us from our current
2475 		 * position and then add us back in at a later
2476 		 * position.
2477 		 */
2478 		vn_syncer_add_to_worklist(*bo, syncdelay);
2479 	}
2480 	BO_UNLOCK(*bo);
2481 	vdrop(vp);
2482 	mtx_lock(&sync_mtx);
2483 	return (0);
2484 }
2485 
2486 static int first_printf = 1;
2487 
2488 /*
2489  * System filesystem synchronizer daemon.
2490  */
2491 static void
2492 sched_sync(void)
2493 {
2494 	struct synclist *next, *slp;
2495 	struct bufobj *bo;
2496 	long starttime;
2497 	struct thread *td = curthread;
2498 	int last_work_seen;
2499 	int net_worklist_len;
2500 	int syncer_final_iter;
2501 	int error;
2502 
2503 	last_work_seen = 0;
2504 	syncer_final_iter = 0;
2505 	syncer_state = SYNCER_RUNNING;
2506 	starttime = time_uptime;
2507 	td->td_pflags |= TDP_NORUNNINGBUF;
2508 
2509 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2510 	    SHUTDOWN_PRI_LAST);
2511 
2512 	mtx_lock(&sync_mtx);
2513 	for (;;) {
2514 		if (syncer_state == SYNCER_FINAL_DELAY &&
2515 		    syncer_final_iter == 0) {
2516 			mtx_unlock(&sync_mtx);
2517 			kproc_suspend_check(td->td_proc);
2518 			mtx_lock(&sync_mtx);
2519 		}
2520 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2521 		if (syncer_state != SYNCER_RUNNING &&
2522 		    starttime != time_uptime) {
2523 			if (first_printf) {
2524 				printf("\nSyncing disks, vnodes remaining... ");
2525 				first_printf = 0;
2526 			}
2527 			printf("%d ", net_worklist_len);
2528 		}
2529 		starttime = time_uptime;
2530 
2531 		/*
2532 		 * Push files whose dirty time has expired.  Be careful
2533 		 * of interrupt race on slp queue.
2534 		 *
2535 		 * Skip over empty worklist slots when shutting down.
2536 		 */
2537 		do {
2538 			slp = &syncer_workitem_pending[syncer_delayno];
2539 			syncer_delayno += 1;
2540 			if (syncer_delayno == syncer_maxdelay)
2541 				syncer_delayno = 0;
2542 			next = &syncer_workitem_pending[syncer_delayno];
2543 			/*
2544 			 * If the worklist has wrapped since the
2545 			 * it was emptied of all but syncer vnodes,
2546 			 * switch to the FINAL_DELAY state and run
2547 			 * for one more second.
2548 			 */
2549 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2550 			    net_worklist_len == 0 &&
2551 			    last_work_seen == syncer_delayno) {
2552 				syncer_state = SYNCER_FINAL_DELAY;
2553 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2554 			}
2555 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2556 		    syncer_worklist_len > 0);
2557 
2558 		/*
2559 		 * Keep track of the last time there was anything
2560 		 * on the worklist other than syncer vnodes.
2561 		 * Return to the SHUTTING_DOWN state if any
2562 		 * new work appears.
2563 		 */
2564 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2565 			last_work_seen = syncer_delayno;
2566 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2567 			syncer_state = SYNCER_SHUTTING_DOWN;
2568 		while (!LIST_EMPTY(slp)) {
2569 			error = sync_vnode(slp, &bo, td);
2570 			if (error == 1) {
2571 				LIST_REMOVE(bo, bo_synclist);
2572 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2573 				continue;
2574 			}
2575 
2576 			if (first_printf == 0) {
2577 				/*
2578 				 * Drop the sync mutex, because some watchdog
2579 				 * drivers need to sleep while patting
2580 				 */
2581 				mtx_unlock(&sync_mtx);
2582 				wdog_kern_pat(WD_LASTVAL);
2583 				mtx_lock(&sync_mtx);
2584 			}
2585 
2586 		}
2587 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2588 			syncer_final_iter--;
2589 		/*
2590 		 * The variable rushjob allows the kernel to speed up the
2591 		 * processing of the filesystem syncer process. A rushjob
2592 		 * value of N tells the filesystem syncer to process the next
2593 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2594 		 * is used by the soft update code to speed up the filesystem
2595 		 * syncer process when the incore state is getting so far
2596 		 * ahead of the disk that the kernel memory pool is being
2597 		 * threatened with exhaustion.
2598 		 */
2599 		if (rushjob > 0) {
2600 			rushjob -= 1;
2601 			continue;
2602 		}
2603 		/*
2604 		 * Just sleep for a short period of time between
2605 		 * iterations when shutting down to allow some I/O
2606 		 * to happen.
2607 		 *
2608 		 * If it has taken us less than a second to process the
2609 		 * current work, then wait. Otherwise start right over
2610 		 * again. We can still lose time if any single round
2611 		 * takes more than two seconds, but it does not really
2612 		 * matter as we are just trying to generally pace the
2613 		 * filesystem activity.
2614 		 */
2615 		if (syncer_state != SYNCER_RUNNING ||
2616 		    time_uptime == starttime) {
2617 			thread_lock(td);
2618 			sched_prio(td, PPAUSE);
2619 			thread_unlock(td);
2620 		}
2621 		if (syncer_state != SYNCER_RUNNING)
2622 			cv_timedwait(&sync_wakeup, &sync_mtx,
2623 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2624 		else if (time_uptime == starttime)
2625 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2626 	}
2627 }
2628 
2629 /*
2630  * Request the syncer daemon to speed up its work.
2631  * We never push it to speed up more than half of its
2632  * normal turn time, otherwise it could take over the cpu.
2633  */
2634 int
2635 speedup_syncer(void)
2636 {
2637 	int ret = 0;
2638 
2639 	mtx_lock(&sync_mtx);
2640 	if (rushjob < syncdelay / 2) {
2641 		rushjob += 1;
2642 		stat_rush_requests += 1;
2643 		ret = 1;
2644 	}
2645 	mtx_unlock(&sync_mtx);
2646 	cv_broadcast(&sync_wakeup);
2647 	return (ret);
2648 }
2649 
2650 /*
2651  * Tell the syncer to speed up its work and run though its work
2652  * list several times, then tell it to shut down.
2653  */
2654 static void
2655 syncer_shutdown(void *arg, int howto)
2656 {
2657 
2658 	if (howto & RB_NOSYNC)
2659 		return;
2660 	mtx_lock(&sync_mtx);
2661 	syncer_state = SYNCER_SHUTTING_DOWN;
2662 	rushjob = 0;
2663 	mtx_unlock(&sync_mtx);
2664 	cv_broadcast(&sync_wakeup);
2665 	kproc_shutdown(arg, howto);
2666 }
2667 
2668 void
2669 syncer_suspend(void)
2670 {
2671 
2672 	syncer_shutdown(updateproc, 0);
2673 }
2674 
2675 void
2676 syncer_resume(void)
2677 {
2678 
2679 	mtx_lock(&sync_mtx);
2680 	first_printf = 1;
2681 	syncer_state = SYNCER_RUNNING;
2682 	mtx_unlock(&sync_mtx);
2683 	cv_broadcast(&sync_wakeup);
2684 	kproc_resume(updateproc);
2685 }
2686 
2687 /*
2688  * Reassign a buffer from one vnode to another.
2689  * Used to assign file specific control information
2690  * (indirect blocks) to the vnode to which they belong.
2691  */
2692 void
2693 reassignbuf(struct buf *bp)
2694 {
2695 	struct vnode *vp;
2696 	struct bufobj *bo;
2697 	int delay;
2698 #ifdef INVARIANTS
2699 	struct bufv *bv;
2700 #endif
2701 
2702 	vp = bp->b_vp;
2703 	bo = bp->b_bufobj;
2704 	++reassignbufcalls;
2705 
2706 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2707 	    bp, bp->b_vp, bp->b_flags);
2708 	/*
2709 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2710 	 * is not fully linked in.
2711 	 */
2712 	if (bp->b_flags & B_PAGING)
2713 		panic("cannot reassign paging buffer");
2714 
2715 	/*
2716 	 * Delete from old vnode list, if on one.
2717 	 */
2718 	BO_LOCK(bo);
2719 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2720 		buf_vlist_remove(bp);
2721 	else
2722 		panic("reassignbuf: Buffer %p not on queue.", bp);
2723 	/*
2724 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2725 	 * of clean buffers.
2726 	 */
2727 	if (bp->b_flags & B_DELWRI) {
2728 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2729 			switch (vp->v_type) {
2730 			case VDIR:
2731 				delay = dirdelay;
2732 				break;
2733 			case VCHR:
2734 				delay = metadelay;
2735 				break;
2736 			default:
2737 				delay = filedelay;
2738 			}
2739 			vn_syncer_add_to_worklist(bo, delay);
2740 		}
2741 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2742 	} else {
2743 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2744 
2745 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2746 			mtx_lock(&sync_mtx);
2747 			LIST_REMOVE(bo, bo_synclist);
2748 			syncer_worklist_len--;
2749 			mtx_unlock(&sync_mtx);
2750 			bo->bo_flag &= ~BO_ONWORKLST;
2751 		}
2752 	}
2753 #ifdef INVARIANTS
2754 	bv = &bo->bo_clean;
2755 	bp = TAILQ_FIRST(&bv->bv_hd);
2756 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2757 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2758 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2759 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2760 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2761 	bv = &bo->bo_dirty;
2762 	bp = TAILQ_FIRST(&bv->bv_hd);
2763 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2764 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2765 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2766 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2767 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2768 #endif
2769 	BO_UNLOCK(bo);
2770 }
2771 
2772 static void
2773 v_init_counters(struct vnode *vp)
2774 {
2775 
2776 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2777 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2778 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2779 
2780 	refcount_init(&vp->v_holdcnt, 1);
2781 	refcount_init(&vp->v_usecount, 1);
2782 }
2783 
2784 /*
2785  * Increment si_usecount of the associated device, if any.
2786  */
2787 static void
2788 v_incr_devcount(struct vnode *vp)
2789 {
2790 
2791 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2792 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2793 		dev_lock();
2794 		vp->v_rdev->si_usecount++;
2795 		dev_unlock();
2796 	}
2797 }
2798 
2799 /*
2800  * Decrement si_usecount of the associated device, if any.
2801  *
2802  * The caller is required to hold the interlock when transitioning a VCHR use
2803  * count to zero. This prevents a race with devfs_reclaim_vchr() that would
2804  * leak a si_usecount reference. The vnode lock will also prevent this race
2805  * if it is held while dropping the last ref.
2806  *
2807  * The race is:
2808  *
2809  * CPU1					CPU2
2810  *				  	devfs_reclaim_vchr
2811  * make v_usecount == 0
2812  * 				    	  VI_LOCK
2813  * 				    	  sees v_usecount == 0, no updates
2814  * 				    	  vp->v_rdev = NULL;
2815  * 				    	  ...
2816  * 				    	  VI_UNLOCK
2817  * VI_LOCK
2818  * v_decr_devcount
2819  *   sees v_rdev == NULL, no updates
2820  *
2821  * In this scenario si_devcount decrement is not performed.
2822  */
2823 static void
2824 v_decr_devcount(struct vnode *vp)
2825 {
2826 
2827 	ASSERT_VOP_LOCKED(vp, __func__);
2828 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2829 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2830 		dev_lock();
2831 		VNPASS(vp->v_rdev->si_usecount > 0, vp);
2832 		vp->v_rdev->si_usecount--;
2833 		dev_unlock();
2834 	}
2835 }
2836 
2837 /*
2838  * Grab a particular vnode from the free list, increment its
2839  * reference count and lock it.  VIRF_DOOMED is set if the vnode
2840  * is being destroyed.  Only callers who specify LK_RETRY will
2841  * see doomed vnodes.  If inactive processing was delayed in
2842  * vput try to do it here.
2843  *
2844  * usecount is manipulated using atomics without holding any locks.
2845  *
2846  * holdcnt can be manipulated using atomics without holding any locks,
2847  * except when transitioning 1<->0, in which case the interlock is held.
2848  */
2849 enum vgetstate
2850 vget_prep(struct vnode *vp)
2851 {
2852 	enum vgetstate vs;
2853 
2854 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2855 		vs = VGET_USECOUNT;
2856 	} else {
2857 		vhold(vp);
2858 		vs = VGET_HOLDCNT;
2859 	}
2860 	return (vs);
2861 }
2862 
2863 int
2864 vget(struct vnode *vp, int flags, struct thread *td)
2865 {
2866 	enum vgetstate vs;
2867 
2868 	MPASS(td == curthread);
2869 
2870 	vs = vget_prep(vp);
2871 	return (vget_finish(vp, flags, vs));
2872 }
2873 
2874 static int __noinline
2875 vget_finish_vchr(struct vnode *vp)
2876 {
2877 
2878 	VNASSERT(vp->v_type == VCHR, vp, ("type != VCHR)"));
2879 
2880 	/*
2881 	 * See the comment in vget_finish before usecount bump.
2882 	 */
2883 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2884 #ifdef INVARIANTS
2885 		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2886 		VNASSERT(old > 0, vp, ("%s: wrong hold count %d", __func__, old));
2887 #else
2888 		refcount_release(&vp->v_holdcnt);
2889 #endif
2890 		return (0);
2891 	}
2892 
2893 	VI_LOCK(vp);
2894 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2895 #ifdef INVARIANTS
2896 		int old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2897 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2898 #else
2899 		refcount_release(&vp->v_holdcnt);
2900 #endif
2901 		VI_UNLOCK(vp);
2902 		return (0);
2903 	}
2904 	v_incr_devcount(vp);
2905 	refcount_acquire(&vp->v_usecount);
2906 	VI_UNLOCK(vp);
2907 	return (0);
2908 }
2909 
2910 int
2911 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
2912 {
2913 	int error, old;
2914 
2915 	if ((flags & LK_INTERLOCK) != 0)
2916 		ASSERT_VI_LOCKED(vp, __func__);
2917 	else
2918 		ASSERT_VI_UNLOCKED(vp, __func__);
2919 	VNPASS(vp->v_holdcnt > 0, vp);
2920 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
2921 
2922 	error = vn_lock(vp, flags);
2923 	if (__predict_false(error != 0)) {
2924 		if (vs == VGET_USECOUNT)
2925 			vrele(vp);
2926 		else
2927 			vdrop(vp);
2928 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2929 		    vp);
2930 		return (error);
2931 	}
2932 
2933 	if (vs == VGET_USECOUNT)
2934 		return (0);
2935 
2936 	if (__predict_false(vp->v_type == VCHR))
2937 		return (vget_finish_vchr(vp));
2938 
2939 	/*
2940 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
2941 	 * the vnode around. Otherwise someone else lended their hold count and
2942 	 * we have to drop ours.
2943 	 */
2944 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
2945 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
2946 	if (old != 0) {
2947 #ifdef INVARIANTS
2948 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
2949 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
2950 #else
2951 		refcount_release(&vp->v_holdcnt);
2952 #endif
2953 	}
2954 	return (0);
2955 }
2956 
2957 /*
2958  * Increase the reference (use) and hold count of a vnode.
2959  * This will also remove the vnode from the free list if it is presently free.
2960  */
2961 static void __noinline
2962 vref_vchr(struct vnode *vp, bool interlock)
2963 {
2964 
2965 	/*
2966 	 * See the comment in vget_finish before usecount bump.
2967 	 */
2968 	if (!interlock) {
2969 		if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2970 			VNODE_REFCOUNT_FENCE_ACQ();
2971 			VNASSERT(vp->v_holdcnt > 0, vp,
2972 			    ("%s: active vnode not held", __func__));
2973 			return;
2974 		}
2975 		VI_LOCK(vp);
2976 		/*
2977 		 * By the time we get here the vnode might have been doomed, at
2978 		 * which point the 0->1 use count transition is no longer
2979 		 * protected by the interlock. Since it can't bounce back to
2980 		 * VCHR and requires vref semantics, punt it back
2981 		 */
2982 		if (__predict_false(vp->v_type == VBAD)) {
2983 			VI_UNLOCK(vp);
2984 			vref(vp);
2985 			return;
2986 		}
2987 	}
2988 	VNASSERT(vp->v_type == VCHR, vp, ("type != VCHR)"));
2989 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2990 		VNODE_REFCOUNT_FENCE_ACQ();
2991 		VNASSERT(vp->v_holdcnt > 0, vp,
2992 		    ("%s: active vnode not held", __func__));
2993 		if (!interlock)
2994 			VI_UNLOCK(vp);
2995 		return;
2996 	}
2997 	vhold(vp);
2998 	v_incr_devcount(vp);
2999 	refcount_acquire(&vp->v_usecount);
3000 	if (!interlock)
3001 		VI_UNLOCK(vp);
3002 	return;
3003 }
3004 
3005 void
3006 vref(struct vnode *vp)
3007 {
3008 	int old;
3009 
3010 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3011 	if (__predict_false(vp->v_type == VCHR)) {
3012 		 vref_vchr(vp, false);
3013 		 return;
3014 	}
3015 
3016 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
3017 		VNODE_REFCOUNT_FENCE_ACQ();
3018 		VNASSERT(vp->v_holdcnt > 0, vp,
3019 		    ("%s: active vnode not held", __func__));
3020 		return;
3021 	}
3022 	vhold(vp);
3023 	/*
3024 	 * See the comment in vget_finish.
3025 	 */
3026 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
3027 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3028 	if (old != 0) {
3029 #ifdef INVARIANTS
3030 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3031 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3032 #else
3033 		refcount_release(&vp->v_holdcnt);
3034 #endif
3035 	}
3036 }
3037 
3038 void
3039 vrefl(struct vnode *vp)
3040 {
3041 
3042 	ASSERT_VI_LOCKED(vp, __func__);
3043 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3044 	if (__predict_false(vp->v_type == VCHR)) {
3045 		vref_vchr(vp, true);
3046 		return;
3047 	}
3048 	vref(vp);
3049 }
3050 
3051 void
3052 vrefact(struct vnode *vp)
3053 {
3054 
3055 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3056 #ifdef INVARIANTS
3057 	int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3058 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3059 #else
3060 	refcount_acquire(&vp->v_usecount);
3061 #endif
3062 }
3063 
3064 void
3065 vrefactn(struct vnode *vp, u_int n)
3066 {
3067 
3068 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3069 #ifdef INVARIANTS
3070 	int old = atomic_fetchadd_int(&vp->v_usecount, n);
3071 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3072 #else
3073 	atomic_add_int(&vp->v_usecount, n);
3074 #endif
3075 }
3076 
3077 /*
3078  * Return reference count of a vnode.
3079  *
3080  * The results of this call are only guaranteed when some mechanism is used to
3081  * stop other processes from gaining references to the vnode.  This may be the
3082  * case if the caller holds the only reference.  This is also useful when stale
3083  * data is acceptable as race conditions may be accounted for by some other
3084  * means.
3085  */
3086 int
3087 vrefcnt(struct vnode *vp)
3088 {
3089 
3090 	return (vp->v_usecount);
3091 }
3092 
3093 void
3094 vlazy(struct vnode *vp)
3095 {
3096 	struct mount *mp;
3097 
3098 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3099 
3100 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3101 		return;
3102 	/*
3103 	 * We may get here for inactive routines after the vnode got doomed.
3104 	 */
3105 	if (VN_IS_DOOMED(vp))
3106 		return;
3107 	mp = vp->v_mount;
3108 	mtx_lock(&mp->mnt_listmtx);
3109 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3110 		vp->v_mflag |= VMP_LAZYLIST;
3111 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3112 		mp->mnt_lazyvnodelistsize++;
3113 	}
3114 	mtx_unlock(&mp->mnt_listmtx);
3115 }
3116 
3117 /*
3118  * This routine is only meant to be called from vgonel prior to dooming
3119  * the vnode.
3120  */
3121 static void
3122 vunlazy_gone(struct vnode *vp)
3123 {
3124 	struct mount *mp;
3125 
3126 	ASSERT_VOP_ELOCKED(vp, __func__);
3127 	ASSERT_VI_LOCKED(vp, __func__);
3128 	VNPASS(!VN_IS_DOOMED(vp), vp);
3129 
3130 	if (vp->v_mflag & VMP_LAZYLIST) {
3131 		mp = vp->v_mount;
3132 		mtx_lock(&mp->mnt_listmtx);
3133 		VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3134 		vp->v_mflag &= ~VMP_LAZYLIST;
3135 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3136 		mp->mnt_lazyvnodelistsize--;
3137 		mtx_unlock(&mp->mnt_listmtx);
3138 	}
3139 }
3140 
3141 static void
3142 vdefer_inactive(struct vnode *vp)
3143 {
3144 
3145 	ASSERT_VI_LOCKED(vp, __func__);
3146 	VNASSERT(vp->v_holdcnt > 0, vp,
3147 	    ("%s: vnode without hold count", __func__));
3148 	if (VN_IS_DOOMED(vp)) {
3149 		vdropl(vp);
3150 		return;
3151 	}
3152 	if (vp->v_iflag & VI_DEFINACT) {
3153 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3154 		vdropl(vp);
3155 		return;
3156 	}
3157 	if (vp->v_usecount > 0) {
3158 		vp->v_iflag &= ~VI_OWEINACT;
3159 		vdropl(vp);
3160 		return;
3161 	}
3162 	vlazy(vp);
3163 	vp->v_iflag |= VI_DEFINACT;
3164 	VI_UNLOCK(vp);
3165 	counter_u64_add(deferred_inact, 1);
3166 }
3167 
3168 static void
3169 vdefer_inactive_unlocked(struct vnode *vp)
3170 {
3171 
3172 	VI_LOCK(vp);
3173 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3174 		vdropl(vp);
3175 		return;
3176 	}
3177 	vdefer_inactive(vp);
3178 }
3179 
3180 enum vput_op { VRELE, VPUT, VUNREF };
3181 
3182 /*
3183  * Handle ->v_usecount transitioning to 0.
3184  *
3185  * By releasing the last usecount we take ownership of the hold count which
3186  * provides liveness of the vnode, meaning we have to vdrop.
3187  *
3188  * If the vnode is of type VCHR we may need to decrement si_usecount, see
3189  * v_decr_devcount for details.
3190  *
3191  * For all vnodes we may need to perform inactive processing. It requires an
3192  * exclusive lock on the vnode, while it is legal to call here with only a
3193  * shared lock (or no locks). If locking the vnode in an expected manner fails,
3194  * inactive processing gets deferred to the syncer.
3195  *
3196  * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3197  * on the lock being held all the way until VOP_INACTIVE. This in particular
3198  * happens with UFS which adds half-constructed vnodes to the hash, where they
3199  * can be found by other code.
3200  */
3201 static void
3202 vput_final(struct vnode *vp, enum vput_op func)
3203 {
3204 	int error;
3205 	bool want_unlock;
3206 
3207 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3208 	VNPASS(vp->v_holdcnt > 0, vp);
3209 
3210 	VI_LOCK(vp);
3211 	if (func != VRELE)
3212 		v_decr_devcount(vp);
3213 
3214 	/*
3215 	 * By the time we got here someone else might have transitioned
3216 	 * the count back to > 0.
3217 	 */
3218 	if (vp->v_usecount > 0)
3219 		goto out;
3220 
3221 	/*
3222 	 * If the vnode is doomed vgone already performed inactive processing
3223 	 * (if needed).
3224 	 */
3225 	if (VN_IS_DOOMED(vp))
3226 		goto out;
3227 
3228 	if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3229 		goto out;
3230 
3231 	if (vp->v_iflag & VI_DOINGINACT)
3232 		goto out;
3233 
3234 	/*
3235 	 * Locking operations here will drop the interlock and possibly the
3236 	 * vnode lock, opening a window where the vnode can get doomed all the
3237 	 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3238 	 * perform inactive.
3239 	 */
3240 	vp->v_iflag |= VI_OWEINACT;
3241 	want_unlock = false;
3242 	error = 0;
3243 	switch (func) {
3244 	case VRELE:
3245 		switch (VOP_ISLOCKED(vp)) {
3246 		case LK_EXCLUSIVE:
3247 			break;
3248 		case LK_EXCLOTHER:
3249 		case 0:
3250 			want_unlock = true;
3251 			error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3252 			VI_LOCK(vp);
3253 			break;
3254 		default:
3255 			/*
3256 			 * The lock has at least one sharer, but we have no way
3257 			 * to conclude whether this is us. Play it safe and
3258 			 * defer processing.
3259 			 */
3260 			error = EAGAIN;
3261 			break;
3262 		}
3263 		break;
3264 	case VPUT:
3265 		want_unlock = true;
3266 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3267 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3268 			    LK_NOWAIT);
3269 			VI_LOCK(vp);
3270 		}
3271 		break;
3272 	case VUNREF:
3273 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3274 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3275 			VI_LOCK(vp);
3276 		}
3277 		break;
3278 	}
3279 	if (error == 0) {
3280 		vinactive(vp);
3281 		if (want_unlock)
3282 			VOP_UNLOCK(vp);
3283 		vdropl(vp);
3284 	} else {
3285 		vdefer_inactive(vp);
3286 	}
3287 	return;
3288 out:
3289 	if (func == VPUT)
3290 		VOP_UNLOCK(vp);
3291 	vdropl(vp);
3292 }
3293 
3294 /*
3295  * Decrement ->v_usecount for a vnode.
3296  *
3297  * Releasing the last use count requires additional processing, see vput_final
3298  * above for details.
3299  *
3300  * Note that releasing use count without the vnode lock requires special casing
3301  * for VCHR, see v_decr_devcount for details.
3302  *
3303  * Comment above each variant denotes lock state on entry and exit.
3304  */
3305 
3306 static void __noinline
3307 vrele_vchr(struct vnode *vp)
3308 {
3309 
3310 	if (refcount_release_if_not_last(&vp->v_usecount))
3311 		return;
3312 	VI_LOCK(vp);
3313 	if (!refcount_release(&vp->v_usecount)) {
3314 		VI_UNLOCK(vp);
3315 		return;
3316 	}
3317 	v_decr_devcount(vp);
3318 	VI_UNLOCK(vp);
3319 	vput_final(vp, VRELE);
3320 }
3321 
3322 /*
3323  * in: any
3324  * out: same as passed in
3325  */
3326 void
3327 vrele(struct vnode *vp)
3328 {
3329 
3330 	ASSERT_VI_UNLOCKED(vp, __func__);
3331 	if (__predict_false(vp->v_type == VCHR)) {
3332 		vrele_vchr(vp);
3333 		return;
3334 	}
3335 	if (!refcount_release(&vp->v_usecount))
3336 		return;
3337 	vput_final(vp, VRELE);
3338 }
3339 
3340 /*
3341  * in: locked
3342  * out: unlocked
3343  */
3344 void
3345 vput(struct vnode *vp)
3346 {
3347 
3348 	ASSERT_VOP_LOCKED(vp, __func__);
3349 	ASSERT_VI_UNLOCKED(vp, __func__);
3350 	if (!refcount_release(&vp->v_usecount)) {
3351 		VOP_UNLOCK(vp);
3352 		return;
3353 	}
3354 	vput_final(vp, VPUT);
3355 }
3356 
3357 /*
3358  * in: locked
3359  * out: locked
3360  */
3361 void
3362 vunref(struct vnode *vp)
3363 {
3364 
3365 	ASSERT_VOP_LOCKED(vp, __func__);
3366 	ASSERT_VI_UNLOCKED(vp, __func__);
3367 	if (!refcount_release(&vp->v_usecount))
3368 		return;
3369 	vput_final(vp, VUNREF);
3370 }
3371 
3372 void
3373 vhold(struct vnode *vp)
3374 {
3375 	struct vdbatch *vd;
3376 	int old;
3377 
3378 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3379 	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3380 	VNASSERT(old >= 0, vp, ("%s: wrong hold count %d", __func__, old));
3381 	if (old != 0)
3382 		return;
3383 	critical_enter();
3384 	vd = DPCPU_PTR(vd);
3385 	vd->freevnodes--;
3386 	critical_exit();
3387 }
3388 
3389 void
3390 vholdl(struct vnode *vp)
3391 {
3392 
3393 	ASSERT_VI_LOCKED(vp, __func__);
3394 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3395 	vhold(vp);
3396 }
3397 
3398 void
3399 vholdnz(struct vnode *vp)
3400 {
3401 
3402 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3403 #ifdef INVARIANTS
3404 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3405 	VNASSERT(old > 0, vp, ("%s: wrong hold count %d", __func__, old));
3406 #else
3407 	atomic_add_int(&vp->v_holdcnt, 1);
3408 #endif
3409 }
3410 
3411 static void __noinline
3412 vdbatch_process(struct vdbatch *vd)
3413 {
3414 	struct vnode *vp;
3415 	int i;
3416 
3417 	mtx_assert(&vd->lock, MA_OWNED);
3418 	MPASS(curthread->td_pinned > 0);
3419 	MPASS(vd->index == VDBATCH_SIZE);
3420 
3421 	mtx_lock(&vnode_list_mtx);
3422 	critical_enter();
3423 	freevnodes += vd->freevnodes;
3424 	for (i = 0; i < VDBATCH_SIZE; i++) {
3425 		vp = vd->tab[i];
3426 		TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3427 		TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3428 		MPASS(vp->v_dbatchcpu != NOCPU);
3429 		vp->v_dbatchcpu = NOCPU;
3430 	}
3431 	mtx_unlock(&vnode_list_mtx);
3432 	vd->freevnodes = 0;
3433 	bzero(vd->tab, sizeof(vd->tab));
3434 	vd->index = 0;
3435 	critical_exit();
3436 }
3437 
3438 static void
3439 vdbatch_enqueue(struct vnode *vp)
3440 {
3441 	struct vdbatch *vd;
3442 
3443 	ASSERT_VI_LOCKED(vp, __func__);
3444 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3445 	    ("%s: deferring requeue of a doomed vnode", __func__));
3446 
3447 	critical_enter();
3448 	vd = DPCPU_PTR(vd);
3449 	vd->freevnodes++;
3450 	if (vp->v_dbatchcpu != NOCPU) {
3451 		VI_UNLOCK(vp);
3452 		critical_exit();
3453 		return;
3454 	}
3455 
3456 	sched_pin();
3457 	critical_exit();
3458 	mtx_lock(&vd->lock);
3459 	MPASS(vd->index < VDBATCH_SIZE);
3460 	MPASS(vd->tab[vd->index] == NULL);
3461 	/*
3462 	 * A hack: we depend on being pinned so that we know what to put in
3463 	 * ->v_dbatchcpu.
3464 	 */
3465 	vp->v_dbatchcpu = curcpu;
3466 	vd->tab[vd->index] = vp;
3467 	vd->index++;
3468 	VI_UNLOCK(vp);
3469 	if (vd->index == VDBATCH_SIZE)
3470 		vdbatch_process(vd);
3471 	mtx_unlock(&vd->lock);
3472 	sched_unpin();
3473 }
3474 
3475 /*
3476  * This routine must only be called for vnodes which are about to be
3477  * deallocated. Supporting dequeue for arbitrary vndoes would require
3478  * validating that the locked batch matches.
3479  */
3480 static void
3481 vdbatch_dequeue(struct vnode *vp)
3482 {
3483 	struct vdbatch *vd;
3484 	int i;
3485 	short cpu;
3486 
3487 	VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3488 	    ("%s: called for a used vnode\n", __func__));
3489 
3490 	cpu = vp->v_dbatchcpu;
3491 	if (cpu == NOCPU)
3492 		return;
3493 
3494 	vd = DPCPU_ID_PTR(cpu, vd);
3495 	mtx_lock(&vd->lock);
3496 	for (i = 0; i < vd->index; i++) {
3497 		if (vd->tab[i] != vp)
3498 			continue;
3499 		vp->v_dbatchcpu = NOCPU;
3500 		vd->index--;
3501 		vd->tab[i] = vd->tab[vd->index];
3502 		vd->tab[vd->index] = NULL;
3503 		break;
3504 	}
3505 	mtx_unlock(&vd->lock);
3506 	/*
3507 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3508 	 */
3509 	MPASS(vp->v_dbatchcpu == NOCPU);
3510 }
3511 
3512 /*
3513  * Drop the hold count of the vnode.  If this is the last reference to
3514  * the vnode we place it on the free list unless it has been vgone'd
3515  * (marked VIRF_DOOMED) in which case we will free it.
3516  *
3517  * Because the vnode vm object keeps a hold reference on the vnode if
3518  * there is at least one resident non-cached page, the vnode cannot
3519  * leave the active list without the page cleanup done.
3520  */
3521 static void
3522 vdrop_deactivate(struct vnode *vp)
3523 {
3524 	struct mount *mp;
3525 
3526 	ASSERT_VI_LOCKED(vp, __func__);
3527 	/*
3528 	 * Mark a vnode as free: remove it from its active list
3529 	 * and put it up for recycling on the freelist.
3530 	 */
3531 	VNASSERT(!VN_IS_DOOMED(vp), vp,
3532 	    ("vdrop: returning doomed vnode"));
3533 	VNASSERT(vp->v_op != NULL, vp,
3534 	    ("vdrop: vnode already reclaimed."));
3535 	VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
3536 	    ("vnode with VI_OWEINACT set"));
3537 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp,
3538 	    ("vnode with VI_DEFINACT set"));
3539 	if (vp->v_mflag & VMP_LAZYLIST) {
3540 		mp = vp->v_mount;
3541 		mtx_lock(&mp->mnt_listmtx);
3542 		VNASSERT(vp->v_mflag & VMP_LAZYLIST, vp, ("lost VMP_LAZYLIST"));
3543 		/*
3544 		 * Don't remove the vnode from the lazy list if another thread
3545 		 * has increased the hold count. It may have re-enqueued the
3546 		 * vnode to the lazy list and is now responsible for its
3547 		 * removal.
3548 		 */
3549 		if (vp->v_holdcnt == 0) {
3550 			vp->v_mflag &= ~VMP_LAZYLIST;
3551 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3552 			mp->mnt_lazyvnodelistsize--;
3553 		}
3554 		mtx_unlock(&mp->mnt_listmtx);
3555 	}
3556 	vdbatch_enqueue(vp);
3557 }
3558 
3559 void
3560 vdrop(struct vnode *vp)
3561 {
3562 
3563 	ASSERT_VI_UNLOCKED(vp, __func__);
3564 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3565 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3566 		return;
3567 	VI_LOCK(vp);
3568 	vdropl(vp);
3569 }
3570 
3571 void
3572 vdropl(struct vnode *vp)
3573 {
3574 
3575 	ASSERT_VI_LOCKED(vp, __func__);
3576 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3577 	if (!refcount_release(&vp->v_holdcnt)) {
3578 		VI_UNLOCK(vp);
3579 		return;
3580 	}
3581 	if (VN_IS_DOOMED(vp)) {
3582 		freevnode(vp);
3583 		return;
3584 	}
3585 	vdrop_deactivate(vp);
3586 }
3587 
3588 /*
3589  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3590  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3591  */
3592 static void
3593 vinactivef(struct vnode *vp)
3594 {
3595 	struct vm_object *obj;
3596 
3597 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3598 	ASSERT_VI_LOCKED(vp, "vinactive");
3599 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3600 	    ("vinactive: recursed on VI_DOINGINACT"));
3601 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3602 	vp->v_iflag |= VI_DOINGINACT;
3603 	vp->v_iflag &= ~VI_OWEINACT;
3604 	VI_UNLOCK(vp);
3605 	/*
3606 	 * Before moving off the active list, we must be sure that any
3607 	 * modified pages are converted into the vnode's dirty
3608 	 * buffers, since these will no longer be checked once the
3609 	 * vnode is on the inactive list.
3610 	 *
3611 	 * The write-out of the dirty pages is asynchronous.  At the
3612 	 * point that VOP_INACTIVE() is called, there could still be
3613 	 * pending I/O and dirty pages in the object.
3614 	 */
3615 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3616 	    vm_object_mightbedirty(obj)) {
3617 		VM_OBJECT_WLOCK(obj);
3618 		vm_object_page_clean(obj, 0, 0, 0);
3619 		VM_OBJECT_WUNLOCK(obj);
3620 	}
3621 	VOP_INACTIVE(vp, curthread);
3622 	VI_LOCK(vp);
3623 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3624 	    ("vinactive: lost VI_DOINGINACT"));
3625 	vp->v_iflag &= ~VI_DOINGINACT;
3626 }
3627 
3628 void
3629 vinactive(struct vnode *vp)
3630 {
3631 
3632 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3633 	ASSERT_VI_LOCKED(vp, "vinactive");
3634 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3635 
3636 	if ((vp->v_iflag & VI_OWEINACT) == 0)
3637 		return;
3638 	if (vp->v_iflag & VI_DOINGINACT)
3639 		return;
3640 	if (vp->v_usecount > 0) {
3641 		vp->v_iflag &= ~VI_OWEINACT;
3642 		return;
3643 	}
3644 	vinactivef(vp);
3645 }
3646 
3647 /*
3648  * Remove any vnodes in the vnode table belonging to mount point mp.
3649  *
3650  * If FORCECLOSE is not specified, there should not be any active ones,
3651  * return error if any are found (nb: this is a user error, not a
3652  * system error). If FORCECLOSE is specified, detach any active vnodes
3653  * that are found.
3654  *
3655  * If WRITECLOSE is set, only flush out regular file vnodes open for
3656  * writing.
3657  *
3658  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3659  *
3660  * `rootrefs' specifies the base reference count for the root vnode
3661  * of this filesystem. The root vnode is considered busy if its
3662  * v_usecount exceeds this value. On a successful return, vflush(, td)
3663  * will call vrele() on the root vnode exactly rootrefs times.
3664  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3665  * be zero.
3666  */
3667 #ifdef DIAGNOSTIC
3668 static int busyprt = 0;		/* print out busy vnodes */
3669 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3670 #endif
3671 
3672 int
3673 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3674 {
3675 	struct vnode *vp, *mvp, *rootvp = NULL;
3676 	struct vattr vattr;
3677 	int busy = 0, error;
3678 
3679 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3680 	    rootrefs, flags);
3681 	if (rootrefs > 0) {
3682 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3683 		    ("vflush: bad args"));
3684 		/*
3685 		 * Get the filesystem root vnode. We can vput() it
3686 		 * immediately, since with rootrefs > 0, it won't go away.
3687 		 */
3688 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3689 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3690 			    __func__, error);
3691 			return (error);
3692 		}
3693 		vput(rootvp);
3694 	}
3695 loop:
3696 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3697 		vholdl(vp);
3698 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3699 		if (error) {
3700 			vdrop(vp);
3701 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3702 			goto loop;
3703 		}
3704 		/*
3705 		 * Skip over a vnodes marked VV_SYSTEM.
3706 		 */
3707 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3708 			VOP_UNLOCK(vp);
3709 			vdrop(vp);
3710 			continue;
3711 		}
3712 		/*
3713 		 * If WRITECLOSE is set, flush out unlinked but still open
3714 		 * files (even if open only for reading) and regular file
3715 		 * vnodes open for writing.
3716 		 */
3717 		if (flags & WRITECLOSE) {
3718 			if (vp->v_object != NULL) {
3719 				VM_OBJECT_WLOCK(vp->v_object);
3720 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3721 				VM_OBJECT_WUNLOCK(vp->v_object);
3722 			}
3723 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3724 			if (error != 0) {
3725 				VOP_UNLOCK(vp);
3726 				vdrop(vp);
3727 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3728 				return (error);
3729 			}
3730 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3731 			VI_LOCK(vp);
3732 
3733 			if ((vp->v_type == VNON ||
3734 			    (error == 0 && vattr.va_nlink > 0)) &&
3735 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3736 				VOP_UNLOCK(vp);
3737 				vdropl(vp);
3738 				continue;
3739 			}
3740 		} else
3741 			VI_LOCK(vp);
3742 		/*
3743 		 * With v_usecount == 0, all we need to do is clear out the
3744 		 * vnode data structures and we are done.
3745 		 *
3746 		 * If FORCECLOSE is set, forcibly close the vnode.
3747 		 */
3748 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3749 			vgonel(vp);
3750 		} else {
3751 			busy++;
3752 #ifdef DIAGNOSTIC
3753 			if (busyprt)
3754 				vn_printf(vp, "vflush: busy vnode ");
3755 #endif
3756 		}
3757 		VOP_UNLOCK(vp);
3758 		vdropl(vp);
3759 	}
3760 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3761 		/*
3762 		 * If just the root vnode is busy, and if its refcount
3763 		 * is equal to `rootrefs', then go ahead and kill it.
3764 		 */
3765 		VI_LOCK(rootvp);
3766 		KASSERT(busy > 0, ("vflush: not busy"));
3767 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3768 		    ("vflush: usecount %d < rootrefs %d",
3769 		     rootvp->v_usecount, rootrefs));
3770 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3771 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3772 			vgone(rootvp);
3773 			VOP_UNLOCK(rootvp);
3774 			busy = 0;
3775 		} else
3776 			VI_UNLOCK(rootvp);
3777 	}
3778 	if (busy) {
3779 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3780 		    busy);
3781 		return (EBUSY);
3782 	}
3783 	for (; rootrefs > 0; rootrefs--)
3784 		vrele(rootvp);
3785 	return (0);
3786 }
3787 
3788 /*
3789  * Recycle an unused vnode to the front of the free list.
3790  */
3791 int
3792 vrecycle(struct vnode *vp)
3793 {
3794 	int recycled;
3795 
3796 	VI_LOCK(vp);
3797 	recycled = vrecyclel(vp);
3798 	VI_UNLOCK(vp);
3799 	return (recycled);
3800 }
3801 
3802 /*
3803  * vrecycle, with the vp interlock held.
3804  */
3805 int
3806 vrecyclel(struct vnode *vp)
3807 {
3808 	int recycled;
3809 
3810 	ASSERT_VOP_ELOCKED(vp, __func__);
3811 	ASSERT_VI_LOCKED(vp, __func__);
3812 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3813 	recycled = 0;
3814 	if (vp->v_usecount == 0) {
3815 		recycled = 1;
3816 		vgonel(vp);
3817 	}
3818 	return (recycled);
3819 }
3820 
3821 /*
3822  * Eliminate all activity associated with a vnode
3823  * in preparation for reuse.
3824  */
3825 void
3826 vgone(struct vnode *vp)
3827 {
3828 	VI_LOCK(vp);
3829 	vgonel(vp);
3830 	VI_UNLOCK(vp);
3831 }
3832 
3833 static void
3834 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3835     struct vnode *lowervp __unused)
3836 {
3837 }
3838 
3839 /*
3840  * Notify upper mounts about reclaimed or unlinked vnode.
3841  */
3842 void
3843 vfs_notify_upper(struct vnode *vp, int event)
3844 {
3845 	static struct vfsops vgonel_vfsops = {
3846 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3847 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3848 	};
3849 	struct mount *mp, *ump, *mmp;
3850 
3851 	mp = vp->v_mount;
3852 	if (mp == NULL)
3853 		return;
3854 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3855 		return;
3856 
3857 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3858 	mmp->mnt_op = &vgonel_vfsops;
3859 	mmp->mnt_kern_flag |= MNTK_MARKER;
3860 	MNT_ILOCK(mp);
3861 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3862 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3863 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3864 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3865 			continue;
3866 		}
3867 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3868 		MNT_IUNLOCK(mp);
3869 		switch (event) {
3870 		case VFS_NOTIFY_UPPER_RECLAIM:
3871 			VFS_RECLAIM_LOWERVP(ump, vp);
3872 			break;
3873 		case VFS_NOTIFY_UPPER_UNLINK:
3874 			VFS_UNLINK_LOWERVP(ump, vp);
3875 			break;
3876 		default:
3877 			KASSERT(0, ("invalid event %d", event));
3878 			break;
3879 		}
3880 		MNT_ILOCK(mp);
3881 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3882 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3883 	}
3884 	free(mmp, M_TEMP);
3885 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3886 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3887 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3888 		wakeup(&mp->mnt_uppers);
3889 	}
3890 	MNT_IUNLOCK(mp);
3891 }
3892 
3893 /*
3894  * vgone, with the vp interlock held.
3895  */
3896 static void
3897 vgonel(struct vnode *vp)
3898 {
3899 	struct thread *td;
3900 	struct mount *mp;
3901 	vm_object_t object;
3902 	bool active, oweinact;
3903 
3904 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3905 	ASSERT_VI_LOCKED(vp, "vgonel");
3906 	VNASSERT(vp->v_holdcnt, vp,
3907 	    ("vgonel: vp %p has no reference.", vp));
3908 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3909 	td = curthread;
3910 
3911 	/*
3912 	 * Don't vgonel if we're already doomed.
3913 	 */
3914 	if (vp->v_irflag & VIRF_DOOMED)
3915 		return;
3916 	vunlazy_gone(vp);
3917 	vp->v_irflag |= VIRF_DOOMED;
3918 
3919 	/*
3920 	 * Check to see if the vnode is in use.  If so, we have to call
3921 	 * VOP_CLOSE() and VOP_INACTIVE().
3922 	 */
3923 	active = vp->v_usecount > 0;
3924 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
3925 	/*
3926 	 * If we need to do inactive VI_OWEINACT will be set.
3927 	 */
3928 	if (vp->v_iflag & VI_DEFINACT) {
3929 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3930 		vp->v_iflag &= ~VI_DEFINACT;
3931 		vdropl(vp);
3932 	} else {
3933 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
3934 		VI_UNLOCK(vp);
3935 	}
3936 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3937 
3938 	/*
3939 	 * If purging an active vnode, it must be closed and
3940 	 * deactivated before being reclaimed.
3941 	 */
3942 	if (active)
3943 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3944 	if (oweinact || active) {
3945 		VI_LOCK(vp);
3946 		vinactivef(vp);
3947 		VI_UNLOCK(vp);
3948 	}
3949 	if (vp->v_type == VSOCK)
3950 		vfs_unp_reclaim(vp);
3951 
3952 	/*
3953 	 * Clean out any buffers associated with the vnode.
3954 	 * If the flush fails, just toss the buffers.
3955 	 */
3956 	mp = NULL;
3957 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3958 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3959 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3960 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3961 			;
3962 	}
3963 
3964 	BO_LOCK(&vp->v_bufobj);
3965 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3966 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3967 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3968 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3969 	    ("vp %p bufobj not invalidated", vp));
3970 
3971 	/*
3972 	 * For VMIO bufobj, BO_DEAD is set later, or in
3973 	 * vm_object_terminate() after the object's page queue is
3974 	 * flushed.
3975 	 */
3976 	object = vp->v_bufobj.bo_object;
3977 	if (object == NULL)
3978 		vp->v_bufobj.bo_flag |= BO_DEAD;
3979 	BO_UNLOCK(&vp->v_bufobj);
3980 
3981 	/*
3982 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
3983 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
3984 	 * should not touch the object borrowed from the lower vnode
3985 	 * (the handle check).
3986 	 */
3987 	if (object != NULL && object->type == OBJT_VNODE &&
3988 	    object->handle == vp)
3989 		vnode_destroy_vobject(vp);
3990 
3991 	/*
3992 	 * Reclaim the vnode.
3993 	 */
3994 	if (VOP_RECLAIM(vp, td))
3995 		panic("vgone: cannot reclaim");
3996 	if (mp != NULL)
3997 		vn_finished_secondary_write(mp);
3998 	VNASSERT(vp->v_object == NULL, vp,
3999 	    ("vop_reclaim left v_object vp=%p", vp));
4000 	/*
4001 	 * Clear the advisory locks and wake up waiting threads.
4002 	 */
4003 	(void)VOP_ADVLOCKPURGE(vp);
4004 	vp->v_lockf = NULL;
4005 	/*
4006 	 * Delete from old mount point vnode list.
4007 	 */
4008 	delmntque(vp);
4009 	cache_purge(vp);
4010 	/*
4011 	 * Done with purge, reset to the standard lock and invalidate
4012 	 * the vnode.
4013 	 */
4014 	VI_LOCK(vp);
4015 	vp->v_vnlock = &vp->v_lock;
4016 	vp->v_op = &dead_vnodeops;
4017 	vp->v_type = VBAD;
4018 }
4019 
4020 /*
4021  * Calculate the total number of references to a special device.
4022  */
4023 int
4024 vcount(struct vnode *vp)
4025 {
4026 	int count;
4027 
4028 	dev_lock();
4029 	count = vp->v_rdev->si_usecount;
4030 	dev_unlock();
4031 	return (count);
4032 }
4033 
4034 /*
4035  * Print out a description of a vnode.
4036  */
4037 static char *typename[] =
4038 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
4039  "VMARKER"};
4040 
4041 void
4042 vn_printf(struct vnode *vp, const char *fmt, ...)
4043 {
4044 	va_list ap;
4045 	char buf[256], buf2[16];
4046 	u_long flags;
4047 
4048 	va_start(ap, fmt);
4049 	vprintf(fmt, ap);
4050 	va_end(ap);
4051 	printf("%p: ", (void *)vp);
4052 	printf("type %s\n", typename[vp->v_type]);
4053 	printf("    usecount %d, writecount %d, refcount %d",
4054 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
4055 	switch (vp->v_type) {
4056 	case VDIR:
4057 		printf(" mountedhere %p\n", vp->v_mountedhere);
4058 		break;
4059 	case VCHR:
4060 		printf(" rdev %p\n", vp->v_rdev);
4061 		break;
4062 	case VSOCK:
4063 		printf(" socket %p\n", vp->v_unpcb);
4064 		break;
4065 	case VFIFO:
4066 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
4067 		break;
4068 	default:
4069 		printf("\n");
4070 		break;
4071 	}
4072 	buf[0] = '\0';
4073 	buf[1] = '\0';
4074 	if (vp->v_irflag & VIRF_DOOMED)
4075 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4076 	flags = vp->v_irflag & ~(VIRF_DOOMED);
4077 	if (flags != 0) {
4078 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4079 		strlcat(buf, buf2, sizeof(buf));
4080 	}
4081 	if (vp->v_vflag & VV_ROOT)
4082 		strlcat(buf, "|VV_ROOT", sizeof(buf));
4083 	if (vp->v_vflag & VV_ISTTY)
4084 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
4085 	if (vp->v_vflag & VV_NOSYNC)
4086 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4087 	if (vp->v_vflag & VV_ETERNALDEV)
4088 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4089 	if (vp->v_vflag & VV_CACHEDLABEL)
4090 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4091 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
4092 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4093 	if (vp->v_vflag & VV_COPYONWRITE)
4094 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4095 	if (vp->v_vflag & VV_SYSTEM)
4096 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4097 	if (vp->v_vflag & VV_PROCDEP)
4098 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4099 	if (vp->v_vflag & VV_NOKNOTE)
4100 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
4101 	if (vp->v_vflag & VV_DELETED)
4102 		strlcat(buf, "|VV_DELETED", sizeof(buf));
4103 	if (vp->v_vflag & VV_MD)
4104 		strlcat(buf, "|VV_MD", sizeof(buf));
4105 	if (vp->v_vflag & VV_FORCEINSMQ)
4106 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4107 	if (vp->v_vflag & VV_READLINK)
4108 		strlcat(buf, "|VV_READLINK", sizeof(buf));
4109 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4110 	    VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
4111 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
4112 	if (flags != 0) {
4113 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4114 		strlcat(buf, buf2, sizeof(buf));
4115 	}
4116 	if (vp->v_iflag & VI_TEXT_REF)
4117 		strlcat(buf, "|VI_TEXT_REF", sizeof(buf));
4118 	if (vp->v_iflag & VI_MOUNT)
4119 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
4120 	if (vp->v_iflag & VI_DOINGINACT)
4121 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4122 	if (vp->v_iflag & VI_OWEINACT)
4123 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4124 	if (vp->v_iflag & VI_DEFINACT)
4125 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4126 	flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_DOINGINACT |
4127 	    VI_OWEINACT | VI_DEFINACT);
4128 	if (flags != 0) {
4129 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4130 		strlcat(buf, buf2, sizeof(buf));
4131 	}
4132 	if (vp->v_mflag & VMP_LAZYLIST)
4133 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4134 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
4135 	if (flags != 0) {
4136 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4137 		strlcat(buf, buf2, sizeof(buf));
4138 	}
4139 	printf("    flags (%s)\n", buf + 1);
4140 	if (mtx_owned(VI_MTX(vp)))
4141 		printf(" VI_LOCKed");
4142 	if (vp->v_object != NULL)
4143 		printf("    v_object %p ref %d pages %d "
4144 		    "cleanbuf %d dirtybuf %d\n",
4145 		    vp->v_object, vp->v_object->ref_count,
4146 		    vp->v_object->resident_page_count,
4147 		    vp->v_bufobj.bo_clean.bv_cnt,
4148 		    vp->v_bufobj.bo_dirty.bv_cnt);
4149 	printf("    ");
4150 	lockmgr_printinfo(vp->v_vnlock);
4151 	if (vp->v_data != NULL)
4152 		VOP_PRINT(vp);
4153 }
4154 
4155 #ifdef DDB
4156 /*
4157  * List all of the locked vnodes in the system.
4158  * Called when debugging the kernel.
4159  */
4160 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
4161 {
4162 	struct mount *mp;
4163 	struct vnode *vp;
4164 
4165 	/*
4166 	 * Note: because this is DDB, we can't obey the locking semantics
4167 	 * for these structures, which means we could catch an inconsistent
4168 	 * state and dereference a nasty pointer.  Not much to be done
4169 	 * about that.
4170 	 */
4171 	db_printf("Locked vnodes\n");
4172 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4173 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4174 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4175 				vn_printf(vp, "vnode ");
4176 		}
4177 	}
4178 }
4179 
4180 /*
4181  * Show details about the given vnode.
4182  */
4183 DB_SHOW_COMMAND(vnode, db_show_vnode)
4184 {
4185 	struct vnode *vp;
4186 
4187 	if (!have_addr)
4188 		return;
4189 	vp = (struct vnode *)addr;
4190 	vn_printf(vp, "vnode ");
4191 }
4192 
4193 /*
4194  * Show details about the given mount point.
4195  */
4196 DB_SHOW_COMMAND(mount, db_show_mount)
4197 {
4198 	struct mount *mp;
4199 	struct vfsopt *opt;
4200 	struct statfs *sp;
4201 	struct vnode *vp;
4202 	char buf[512];
4203 	uint64_t mflags;
4204 	u_int flags;
4205 
4206 	if (!have_addr) {
4207 		/* No address given, print short info about all mount points. */
4208 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4209 			db_printf("%p %s on %s (%s)\n", mp,
4210 			    mp->mnt_stat.f_mntfromname,
4211 			    mp->mnt_stat.f_mntonname,
4212 			    mp->mnt_stat.f_fstypename);
4213 			if (db_pager_quit)
4214 				break;
4215 		}
4216 		db_printf("\nMore info: show mount <addr>\n");
4217 		return;
4218 	}
4219 
4220 	mp = (struct mount *)addr;
4221 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4222 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4223 
4224 	buf[0] = '\0';
4225 	mflags = mp->mnt_flag;
4226 #define	MNT_FLAG(flag)	do {						\
4227 	if (mflags & (flag)) {						\
4228 		if (buf[0] != '\0')					\
4229 			strlcat(buf, ", ", sizeof(buf));		\
4230 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4231 		mflags &= ~(flag);					\
4232 	}								\
4233 } while (0)
4234 	MNT_FLAG(MNT_RDONLY);
4235 	MNT_FLAG(MNT_SYNCHRONOUS);
4236 	MNT_FLAG(MNT_NOEXEC);
4237 	MNT_FLAG(MNT_NOSUID);
4238 	MNT_FLAG(MNT_NFS4ACLS);
4239 	MNT_FLAG(MNT_UNION);
4240 	MNT_FLAG(MNT_ASYNC);
4241 	MNT_FLAG(MNT_SUIDDIR);
4242 	MNT_FLAG(MNT_SOFTDEP);
4243 	MNT_FLAG(MNT_NOSYMFOLLOW);
4244 	MNT_FLAG(MNT_GJOURNAL);
4245 	MNT_FLAG(MNT_MULTILABEL);
4246 	MNT_FLAG(MNT_ACLS);
4247 	MNT_FLAG(MNT_NOATIME);
4248 	MNT_FLAG(MNT_NOCLUSTERR);
4249 	MNT_FLAG(MNT_NOCLUSTERW);
4250 	MNT_FLAG(MNT_SUJ);
4251 	MNT_FLAG(MNT_EXRDONLY);
4252 	MNT_FLAG(MNT_EXPORTED);
4253 	MNT_FLAG(MNT_DEFEXPORTED);
4254 	MNT_FLAG(MNT_EXPORTANON);
4255 	MNT_FLAG(MNT_EXKERB);
4256 	MNT_FLAG(MNT_EXPUBLIC);
4257 	MNT_FLAG(MNT_LOCAL);
4258 	MNT_FLAG(MNT_QUOTA);
4259 	MNT_FLAG(MNT_ROOTFS);
4260 	MNT_FLAG(MNT_USER);
4261 	MNT_FLAG(MNT_IGNORE);
4262 	MNT_FLAG(MNT_UPDATE);
4263 	MNT_FLAG(MNT_DELEXPORT);
4264 	MNT_FLAG(MNT_RELOAD);
4265 	MNT_FLAG(MNT_FORCE);
4266 	MNT_FLAG(MNT_SNAPSHOT);
4267 	MNT_FLAG(MNT_BYFSID);
4268 #undef MNT_FLAG
4269 	if (mflags != 0) {
4270 		if (buf[0] != '\0')
4271 			strlcat(buf, ", ", sizeof(buf));
4272 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4273 		    "0x%016jx", mflags);
4274 	}
4275 	db_printf("    mnt_flag = %s\n", buf);
4276 
4277 	buf[0] = '\0';
4278 	flags = mp->mnt_kern_flag;
4279 #define	MNT_KERN_FLAG(flag)	do {					\
4280 	if (flags & (flag)) {						\
4281 		if (buf[0] != '\0')					\
4282 			strlcat(buf, ", ", sizeof(buf));		\
4283 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4284 		flags &= ~(flag);					\
4285 	}								\
4286 } while (0)
4287 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4288 	MNT_KERN_FLAG(MNTK_ASYNC);
4289 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4290 	MNT_KERN_FLAG(MNTK_DRAINING);
4291 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4292 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4293 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4294 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4295 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
4296 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
4297 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
4298 	MNT_KERN_FLAG(MNTK_MARKER);
4299 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4300 	MNT_KERN_FLAG(MNTK_NOASYNC);
4301 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4302 	MNT_KERN_FLAG(MNTK_MWAIT);
4303 	MNT_KERN_FLAG(MNTK_SUSPEND);
4304 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4305 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4306 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4307 	MNT_KERN_FLAG(MNTK_NOKNOTE);
4308 #undef MNT_KERN_FLAG
4309 	if (flags != 0) {
4310 		if (buf[0] != '\0')
4311 			strlcat(buf, ", ", sizeof(buf));
4312 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4313 		    "0x%08x", flags);
4314 	}
4315 	db_printf("    mnt_kern_flag = %s\n", buf);
4316 
4317 	db_printf("    mnt_opt = ");
4318 	opt = TAILQ_FIRST(mp->mnt_opt);
4319 	if (opt != NULL) {
4320 		db_printf("%s", opt->name);
4321 		opt = TAILQ_NEXT(opt, link);
4322 		while (opt != NULL) {
4323 			db_printf(", %s", opt->name);
4324 			opt = TAILQ_NEXT(opt, link);
4325 		}
4326 	}
4327 	db_printf("\n");
4328 
4329 	sp = &mp->mnt_stat;
4330 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4331 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4332 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4333 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4334 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4335 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4336 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4337 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4338 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4339 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4340 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4341 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4342 
4343 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4344 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4345 	if (jailed(mp->mnt_cred))
4346 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4347 	db_printf(" }\n");
4348 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4349 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4350 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4351 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4352 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4353 	    mp->mnt_lazyvnodelistsize);
4354 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4355 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4356 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
4357 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4358 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4359 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4360 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4361 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4362 	db_printf("    mnt_secondary_accwrites = %d\n",
4363 	    mp->mnt_secondary_accwrites);
4364 	db_printf("    mnt_gjprovider = %s\n",
4365 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4366 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4367 
4368 	db_printf("\n\nList of active vnodes\n");
4369 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4370 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4371 			vn_printf(vp, "vnode ");
4372 			if (db_pager_quit)
4373 				break;
4374 		}
4375 	}
4376 	db_printf("\n\nList of inactive vnodes\n");
4377 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4378 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4379 			vn_printf(vp, "vnode ");
4380 			if (db_pager_quit)
4381 				break;
4382 		}
4383 	}
4384 }
4385 #endif	/* DDB */
4386 
4387 /*
4388  * Fill in a struct xvfsconf based on a struct vfsconf.
4389  */
4390 static int
4391 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4392 {
4393 	struct xvfsconf xvfsp;
4394 
4395 	bzero(&xvfsp, sizeof(xvfsp));
4396 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4397 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4398 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4399 	xvfsp.vfc_flags = vfsp->vfc_flags;
4400 	/*
4401 	 * These are unused in userland, we keep them
4402 	 * to not break binary compatibility.
4403 	 */
4404 	xvfsp.vfc_vfsops = NULL;
4405 	xvfsp.vfc_next = NULL;
4406 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4407 }
4408 
4409 #ifdef COMPAT_FREEBSD32
4410 struct xvfsconf32 {
4411 	uint32_t	vfc_vfsops;
4412 	char		vfc_name[MFSNAMELEN];
4413 	int32_t		vfc_typenum;
4414 	int32_t		vfc_refcount;
4415 	int32_t		vfc_flags;
4416 	uint32_t	vfc_next;
4417 };
4418 
4419 static int
4420 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4421 {
4422 	struct xvfsconf32 xvfsp;
4423 
4424 	bzero(&xvfsp, sizeof(xvfsp));
4425 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4426 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4427 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4428 	xvfsp.vfc_flags = vfsp->vfc_flags;
4429 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4430 }
4431 #endif
4432 
4433 /*
4434  * Top level filesystem related information gathering.
4435  */
4436 static int
4437 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4438 {
4439 	struct vfsconf *vfsp;
4440 	int error;
4441 
4442 	error = 0;
4443 	vfsconf_slock();
4444 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4445 #ifdef COMPAT_FREEBSD32
4446 		if (req->flags & SCTL_MASK32)
4447 			error = vfsconf2x32(req, vfsp);
4448 		else
4449 #endif
4450 			error = vfsconf2x(req, vfsp);
4451 		if (error)
4452 			break;
4453 	}
4454 	vfsconf_sunlock();
4455 	return (error);
4456 }
4457 
4458 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4459     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4460     "S,xvfsconf", "List of all configured filesystems");
4461 
4462 #ifndef BURN_BRIDGES
4463 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4464 
4465 static int
4466 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4467 {
4468 	int *name = (int *)arg1 - 1;	/* XXX */
4469 	u_int namelen = arg2 + 1;	/* XXX */
4470 	struct vfsconf *vfsp;
4471 
4472 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4473 	    "please rebuild world\n");
4474 
4475 #if 1 || defined(COMPAT_PRELITE2)
4476 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4477 	if (namelen == 1)
4478 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4479 #endif
4480 
4481 	switch (name[1]) {
4482 	case VFS_MAXTYPENUM:
4483 		if (namelen != 2)
4484 			return (ENOTDIR);
4485 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4486 	case VFS_CONF:
4487 		if (namelen != 3)
4488 			return (ENOTDIR);	/* overloaded */
4489 		vfsconf_slock();
4490 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4491 			if (vfsp->vfc_typenum == name[2])
4492 				break;
4493 		}
4494 		vfsconf_sunlock();
4495 		if (vfsp == NULL)
4496 			return (EOPNOTSUPP);
4497 #ifdef COMPAT_FREEBSD32
4498 		if (req->flags & SCTL_MASK32)
4499 			return (vfsconf2x32(req, vfsp));
4500 		else
4501 #endif
4502 			return (vfsconf2x(req, vfsp));
4503 	}
4504 	return (EOPNOTSUPP);
4505 }
4506 
4507 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4508     CTLFLAG_MPSAFE, vfs_sysctl,
4509     "Generic filesystem");
4510 
4511 #if 1 || defined(COMPAT_PRELITE2)
4512 
4513 static int
4514 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4515 {
4516 	int error;
4517 	struct vfsconf *vfsp;
4518 	struct ovfsconf ovfs;
4519 
4520 	vfsconf_slock();
4521 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4522 		bzero(&ovfs, sizeof(ovfs));
4523 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4524 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4525 		ovfs.vfc_index = vfsp->vfc_typenum;
4526 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4527 		ovfs.vfc_flags = vfsp->vfc_flags;
4528 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4529 		if (error != 0) {
4530 			vfsconf_sunlock();
4531 			return (error);
4532 		}
4533 	}
4534 	vfsconf_sunlock();
4535 	return (0);
4536 }
4537 
4538 #endif /* 1 || COMPAT_PRELITE2 */
4539 #endif /* !BURN_BRIDGES */
4540 
4541 #define KINFO_VNODESLOP		10
4542 #ifdef notyet
4543 /*
4544  * Dump vnode list (via sysctl).
4545  */
4546 /* ARGSUSED */
4547 static int
4548 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4549 {
4550 	struct xvnode *xvn;
4551 	struct mount *mp;
4552 	struct vnode *vp;
4553 	int error, len, n;
4554 
4555 	/*
4556 	 * Stale numvnodes access is not fatal here.
4557 	 */
4558 	req->lock = 0;
4559 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4560 	if (!req->oldptr)
4561 		/* Make an estimate */
4562 		return (SYSCTL_OUT(req, 0, len));
4563 
4564 	error = sysctl_wire_old_buffer(req, 0);
4565 	if (error != 0)
4566 		return (error);
4567 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4568 	n = 0;
4569 	mtx_lock(&mountlist_mtx);
4570 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4571 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4572 			continue;
4573 		MNT_ILOCK(mp);
4574 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4575 			if (n == len)
4576 				break;
4577 			vref(vp);
4578 			xvn[n].xv_size = sizeof *xvn;
4579 			xvn[n].xv_vnode = vp;
4580 			xvn[n].xv_id = 0;	/* XXX compat */
4581 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4582 			XV_COPY(usecount);
4583 			XV_COPY(writecount);
4584 			XV_COPY(holdcnt);
4585 			XV_COPY(mount);
4586 			XV_COPY(numoutput);
4587 			XV_COPY(type);
4588 #undef XV_COPY
4589 			xvn[n].xv_flag = vp->v_vflag;
4590 
4591 			switch (vp->v_type) {
4592 			case VREG:
4593 			case VDIR:
4594 			case VLNK:
4595 				break;
4596 			case VBLK:
4597 			case VCHR:
4598 				if (vp->v_rdev == NULL) {
4599 					vrele(vp);
4600 					continue;
4601 				}
4602 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4603 				break;
4604 			case VSOCK:
4605 				xvn[n].xv_socket = vp->v_socket;
4606 				break;
4607 			case VFIFO:
4608 				xvn[n].xv_fifo = vp->v_fifoinfo;
4609 				break;
4610 			case VNON:
4611 			case VBAD:
4612 			default:
4613 				/* shouldn't happen? */
4614 				vrele(vp);
4615 				continue;
4616 			}
4617 			vrele(vp);
4618 			++n;
4619 		}
4620 		MNT_IUNLOCK(mp);
4621 		mtx_lock(&mountlist_mtx);
4622 		vfs_unbusy(mp);
4623 		if (n == len)
4624 			break;
4625 	}
4626 	mtx_unlock(&mountlist_mtx);
4627 
4628 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4629 	free(xvn, M_TEMP);
4630 	return (error);
4631 }
4632 
4633 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4634     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4635     "");
4636 #endif
4637 
4638 static void
4639 unmount_or_warn(struct mount *mp)
4640 {
4641 	int error;
4642 
4643 	error = dounmount(mp, MNT_FORCE, curthread);
4644 	if (error != 0) {
4645 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4646 		if (error == EBUSY)
4647 			printf("BUSY)\n");
4648 		else
4649 			printf("%d)\n", error);
4650 	}
4651 }
4652 
4653 /*
4654  * Unmount all filesystems. The list is traversed in reverse order
4655  * of mounting to avoid dependencies.
4656  */
4657 void
4658 vfs_unmountall(void)
4659 {
4660 	struct mount *mp, *tmp;
4661 
4662 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4663 
4664 	/*
4665 	 * Since this only runs when rebooting, it is not interlocked.
4666 	 */
4667 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4668 		vfs_ref(mp);
4669 
4670 		/*
4671 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4672 		 * unmount of the latter.
4673 		 */
4674 		if (mp == rootdevmp)
4675 			continue;
4676 
4677 		unmount_or_warn(mp);
4678 	}
4679 
4680 	if (rootdevmp != NULL)
4681 		unmount_or_warn(rootdevmp);
4682 }
4683 
4684 static void
4685 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4686 {
4687 
4688 	ASSERT_VI_LOCKED(vp, __func__);
4689 	VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4690 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
4691 		vdropl(vp);
4692 		return;
4693 	}
4694 	if (vn_lock(vp, lkflags) == 0) {
4695 		VI_LOCK(vp);
4696 		vinactive(vp);
4697 		VOP_UNLOCK(vp);
4698 		vdropl(vp);
4699 		return;
4700 	}
4701 	vdefer_inactive_unlocked(vp);
4702 }
4703 
4704 static int
4705 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4706 {
4707 
4708 	return (vp->v_iflag & VI_DEFINACT);
4709 }
4710 
4711 static void __noinline
4712 vfs_periodic_inactive(struct mount *mp, int flags)
4713 {
4714 	struct vnode *vp, *mvp;
4715 	int lkflags;
4716 
4717 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4718 	if (flags != MNT_WAIT)
4719 		lkflags |= LK_NOWAIT;
4720 
4721 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4722 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
4723 			VI_UNLOCK(vp);
4724 			continue;
4725 		}
4726 		vp->v_iflag &= ~VI_DEFINACT;
4727 		vfs_deferred_inactive(vp, lkflags);
4728 	}
4729 }
4730 
4731 static inline bool
4732 vfs_want_msync(struct vnode *vp)
4733 {
4734 	struct vm_object *obj;
4735 
4736 	/*
4737 	 * This test may be performed without any locks held.
4738 	 * We rely on vm_object's type stability.
4739 	 */
4740 	if (vp->v_vflag & VV_NOSYNC)
4741 		return (false);
4742 	obj = vp->v_object;
4743 	return (obj != NULL && vm_object_mightbedirty(obj));
4744 }
4745 
4746 static int
4747 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4748 {
4749 
4750 	if (vp->v_vflag & VV_NOSYNC)
4751 		return (false);
4752 	if (vp->v_iflag & VI_DEFINACT)
4753 		return (true);
4754 	return (vfs_want_msync(vp));
4755 }
4756 
4757 static void __noinline
4758 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4759 {
4760 	struct vnode *vp, *mvp;
4761 	struct vm_object *obj;
4762 	struct thread *td;
4763 	int lkflags, objflags;
4764 	bool seen_defer;
4765 
4766 	td = curthread;
4767 
4768 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4769 	if (flags != MNT_WAIT) {
4770 		lkflags |= LK_NOWAIT;
4771 		objflags = OBJPC_NOSYNC;
4772 	} else {
4773 		objflags = OBJPC_SYNC;
4774 	}
4775 
4776 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4777 		seen_defer = false;
4778 		if (vp->v_iflag & VI_DEFINACT) {
4779 			vp->v_iflag &= ~VI_DEFINACT;
4780 			seen_defer = true;
4781 		}
4782 		if (!vfs_want_msync(vp)) {
4783 			if (seen_defer)
4784 				vfs_deferred_inactive(vp, lkflags);
4785 			else
4786 				VI_UNLOCK(vp);
4787 			continue;
4788 		}
4789 		if (vget(vp, lkflags, td) == 0) {
4790 			obj = vp->v_object;
4791 			if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4792 				VM_OBJECT_WLOCK(obj);
4793 				vm_object_page_clean(obj, 0, 0, objflags);
4794 				VM_OBJECT_WUNLOCK(obj);
4795 			}
4796 			vput(vp);
4797 			if (seen_defer)
4798 				vdrop(vp);
4799 		} else {
4800 			if (seen_defer)
4801 				vdefer_inactive_unlocked(vp);
4802 		}
4803 	}
4804 }
4805 
4806 void
4807 vfs_periodic(struct mount *mp, int flags)
4808 {
4809 
4810 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4811 
4812 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4813 		vfs_periodic_inactive(mp, flags);
4814 	else
4815 		vfs_periodic_msync_inactive(mp, flags);
4816 }
4817 
4818 static void
4819 destroy_vpollinfo_free(struct vpollinfo *vi)
4820 {
4821 
4822 	knlist_destroy(&vi->vpi_selinfo.si_note);
4823 	mtx_destroy(&vi->vpi_lock);
4824 	uma_zfree(vnodepoll_zone, vi);
4825 }
4826 
4827 static void
4828 destroy_vpollinfo(struct vpollinfo *vi)
4829 {
4830 
4831 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4832 	seldrain(&vi->vpi_selinfo);
4833 	destroy_vpollinfo_free(vi);
4834 }
4835 
4836 /*
4837  * Initialize per-vnode helper structure to hold poll-related state.
4838  */
4839 void
4840 v_addpollinfo(struct vnode *vp)
4841 {
4842 	struct vpollinfo *vi;
4843 
4844 	if (vp->v_pollinfo != NULL)
4845 		return;
4846 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4847 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4848 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4849 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4850 	VI_LOCK(vp);
4851 	if (vp->v_pollinfo != NULL) {
4852 		VI_UNLOCK(vp);
4853 		destroy_vpollinfo_free(vi);
4854 		return;
4855 	}
4856 	vp->v_pollinfo = vi;
4857 	VI_UNLOCK(vp);
4858 }
4859 
4860 /*
4861  * Record a process's interest in events which might happen to
4862  * a vnode.  Because poll uses the historic select-style interface
4863  * internally, this routine serves as both the ``check for any
4864  * pending events'' and the ``record my interest in future events''
4865  * functions.  (These are done together, while the lock is held,
4866  * to avoid race conditions.)
4867  */
4868 int
4869 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4870 {
4871 
4872 	v_addpollinfo(vp);
4873 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4874 	if (vp->v_pollinfo->vpi_revents & events) {
4875 		/*
4876 		 * This leaves events we are not interested
4877 		 * in available for the other process which
4878 		 * which presumably had requested them
4879 		 * (otherwise they would never have been
4880 		 * recorded).
4881 		 */
4882 		events &= vp->v_pollinfo->vpi_revents;
4883 		vp->v_pollinfo->vpi_revents &= ~events;
4884 
4885 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4886 		return (events);
4887 	}
4888 	vp->v_pollinfo->vpi_events |= events;
4889 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4890 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4891 	return (0);
4892 }
4893 
4894 /*
4895  * Routine to create and manage a filesystem syncer vnode.
4896  */
4897 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4898 static int	sync_fsync(struct  vop_fsync_args *);
4899 static int	sync_inactive(struct  vop_inactive_args *);
4900 static int	sync_reclaim(struct  vop_reclaim_args *);
4901 
4902 static struct vop_vector sync_vnodeops = {
4903 	.vop_bypass =	VOP_EOPNOTSUPP,
4904 	.vop_close =	sync_close,		/* close */
4905 	.vop_fsync =	sync_fsync,		/* fsync */
4906 	.vop_inactive =	sync_inactive,	/* inactive */
4907 	.vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
4908 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4909 	.vop_lock1 =	vop_stdlock,	/* lock */
4910 	.vop_unlock =	vop_stdunlock,	/* unlock */
4911 	.vop_islocked =	vop_stdislocked,	/* islocked */
4912 };
4913 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
4914 
4915 /*
4916  * Create a new filesystem syncer vnode for the specified mount point.
4917  */
4918 void
4919 vfs_allocate_syncvnode(struct mount *mp)
4920 {
4921 	struct vnode *vp;
4922 	struct bufobj *bo;
4923 	static long start, incr, next;
4924 	int error;
4925 
4926 	/* Allocate a new vnode */
4927 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4928 	if (error != 0)
4929 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4930 	vp->v_type = VNON;
4931 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4932 	vp->v_vflag |= VV_FORCEINSMQ;
4933 	error = insmntque(vp, mp);
4934 	if (error != 0)
4935 		panic("vfs_allocate_syncvnode: insmntque() failed");
4936 	vp->v_vflag &= ~VV_FORCEINSMQ;
4937 	VOP_UNLOCK(vp);
4938 	/*
4939 	 * Place the vnode onto the syncer worklist. We attempt to
4940 	 * scatter them about on the list so that they will go off
4941 	 * at evenly distributed times even if all the filesystems
4942 	 * are mounted at once.
4943 	 */
4944 	next += incr;
4945 	if (next == 0 || next > syncer_maxdelay) {
4946 		start /= 2;
4947 		incr /= 2;
4948 		if (start == 0) {
4949 			start = syncer_maxdelay / 2;
4950 			incr = syncer_maxdelay;
4951 		}
4952 		next = start;
4953 	}
4954 	bo = &vp->v_bufobj;
4955 	BO_LOCK(bo);
4956 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4957 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4958 	mtx_lock(&sync_mtx);
4959 	sync_vnode_count++;
4960 	if (mp->mnt_syncer == NULL) {
4961 		mp->mnt_syncer = vp;
4962 		vp = NULL;
4963 	}
4964 	mtx_unlock(&sync_mtx);
4965 	BO_UNLOCK(bo);
4966 	if (vp != NULL) {
4967 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4968 		vgone(vp);
4969 		vput(vp);
4970 	}
4971 }
4972 
4973 void
4974 vfs_deallocate_syncvnode(struct mount *mp)
4975 {
4976 	struct vnode *vp;
4977 
4978 	mtx_lock(&sync_mtx);
4979 	vp = mp->mnt_syncer;
4980 	if (vp != NULL)
4981 		mp->mnt_syncer = NULL;
4982 	mtx_unlock(&sync_mtx);
4983 	if (vp != NULL)
4984 		vrele(vp);
4985 }
4986 
4987 /*
4988  * Do a lazy sync of the filesystem.
4989  */
4990 static int
4991 sync_fsync(struct vop_fsync_args *ap)
4992 {
4993 	struct vnode *syncvp = ap->a_vp;
4994 	struct mount *mp = syncvp->v_mount;
4995 	int error, save;
4996 	struct bufobj *bo;
4997 
4998 	/*
4999 	 * We only need to do something if this is a lazy evaluation.
5000 	 */
5001 	if (ap->a_waitfor != MNT_LAZY)
5002 		return (0);
5003 
5004 	/*
5005 	 * Move ourselves to the back of the sync list.
5006 	 */
5007 	bo = &syncvp->v_bufobj;
5008 	BO_LOCK(bo);
5009 	vn_syncer_add_to_worklist(bo, syncdelay);
5010 	BO_UNLOCK(bo);
5011 
5012 	/*
5013 	 * Walk the list of vnodes pushing all that are dirty and
5014 	 * not already on the sync list.
5015 	 */
5016 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
5017 		return (0);
5018 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
5019 		vfs_unbusy(mp);
5020 		return (0);
5021 	}
5022 	save = curthread_pflags_set(TDP_SYNCIO);
5023 	/*
5024 	 * The filesystem at hand may be idle with free vnodes stored in the
5025 	 * batch.  Return them instead of letting them stay there indefinitely.
5026 	 */
5027 	vfs_periodic(mp, MNT_NOWAIT);
5028 	error = VFS_SYNC(mp, MNT_LAZY);
5029 	curthread_pflags_restore(save);
5030 	vn_finished_write(mp);
5031 	vfs_unbusy(mp);
5032 	return (error);
5033 }
5034 
5035 /*
5036  * The syncer vnode is no referenced.
5037  */
5038 static int
5039 sync_inactive(struct vop_inactive_args *ap)
5040 {
5041 
5042 	vgone(ap->a_vp);
5043 	return (0);
5044 }
5045 
5046 /*
5047  * The syncer vnode is no longer needed and is being decommissioned.
5048  *
5049  * Modifications to the worklist must be protected by sync_mtx.
5050  */
5051 static int
5052 sync_reclaim(struct vop_reclaim_args *ap)
5053 {
5054 	struct vnode *vp = ap->a_vp;
5055 	struct bufobj *bo;
5056 
5057 	bo = &vp->v_bufobj;
5058 	BO_LOCK(bo);
5059 	mtx_lock(&sync_mtx);
5060 	if (vp->v_mount->mnt_syncer == vp)
5061 		vp->v_mount->mnt_syncer = NULL;
5062 	if (bo->bo_flag & BO_ONWORKLST) {
5063 		LIST_REMOVE(bo, bo_synclist);
5064 		syncer_worklist_len--;
5065 		sync_vnode_count--;
5066 		bo->bo_flag &= ~BO_ONWORKLST;
5067 	}
5068 	mtx_unlock(&sync_mtx);
5069 	BO_UNLOCK(bo);
5070 
5071 	return (0);
5072 }
5073 
5074 int
5075 vn_need_pageq_flush(struct vnode *vp)
5076 {
5077 	struct vm_object *obj;
5078 	int need;
5079 
5080 	MPASS(mtx_owned(VI_MTX(vp)));
5081 	need = 0;
5082 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5083 	    vm_object_mightbedirty(obj))
5084 		need = 1;
5085 	return (need);
5086 }
5087 
5088 /*
5089  * Check if vnode represents a disk device
5090  */
5091 int
5092 vn_isdisk(struct vnode *vp, int *errp)
5093 {
5094 	int error;
5095 
5096 	if (vp->v_type != VCHR) {
5097 		error = ENOTBLK;
5098 		goto out;
5099 	}
5100 	error = 0;
5101 	dev_lock();
5102 	if (vp->v_rdev == NULL)
5103 		error = ENXIO;
5104 	else if (vp->v_rdev->si_devsw == NULL)
5105 		error = ENXIO;
5106 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5107 		error = ENOTBLK;
5108 	dev_unlock();
5109 out:
5110 	if (errp != NULL)
5111 		*errp = error;
5112 	return (error == 0);
5113 }
5114 
5115 /*
5116  * Common filesystem object access control check routine.  Accepts a
5117  * vnode's type, "mode", uid and gid, requested access mode, credentials,
5118  * and optional call-by-reference privused argument allowing vaccess()
5119  * to indicate to the caller whether privilege was used to satisfy the
5120  * request (obsoleted).  Returns 0 on success, or an errno on failure.
5121  */
5122 int
5123 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5124     accmode_t accmode, struct ucred *cred, int *privused)
5125 {
5126 	accmode_t dac_granted;
5127 	accmode_t priv_granted;
5128 
5129 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5130 	    ("invalid bit in accmode"));
5131 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5132 	    ("VAPPEND without VWRITE"));
5133 
5134 	/*
5135 	 * Look for a normal, non-privileged way to access the file/directory
5136 	 * as requested.  If it exists, go with that.
5137 	 */
5138 
5139 	if (privused != NULL)
5140 		*privused = 0;
5141 
5142 	dac_granted = 0;
5143 
5144 	/* Check the owner. */
5145 	if (cred->cr_uid == file_uid) {
5146 		dac_granted |= VADMIN;
5147 		if (file_mode & S_IXUSR)
5148 			dac_granted |= VEXEC;
5149 		if (file_mode & S_IRUSR)
5150 			dac_granted |= VREAD;
5151 		if (file_mode & S_IWUSR)
5152 			dac_granted |= (VWRITE | VAPPEND);
5153 
5154 		if ((accmode & dac_granted) == accmode)
5155 			return (0);
5156 
5157 		goto privcheck;
5158 	}
5159 
5160 	/* Otherwise, check the groups (first match) */
5161 	if (groupmember(file_gid, cred)) {
5162 		if (file_mode & S_IXGRP)
5163 			dac_granted |= VEXEC;
5164 		if (file_mode & S_IRGRP)
5165 			dac_granted |= VREAD;
5166 		if (file_mode & S_IWGRP)
5167 			dac_granted |= (VWRITE | VAPPEND);
5168 
5169 		if ((accmode & dac_granted) == accmode)
5170 			return (0);
5171 
5172 		goto privcheck;
5173 	}
5174 
5175 	/* Otherwise, check everyone else. */
5176 	if (file_mode & S_IXOTH)
5177 		dac_granted |= VEXEC;
5178 	if (file_mode & S_IROTH)
5179 		dac_granted |= VREAD;
5180 	if (file_mode & S_IWOTH)
5181 		dac_granted |= (VWRITE | VAPPEND);
5182 	if ((accmode & dac_granted) == accmode)
5183 		return (0);
5184 
5185 privcheck:
5186 	/*
5187 	 * Build a privilege mask to determine if the set of privileges
5188 	 * satisfies the requirements when combined with the granted mask
5189 	 * from above.  For each privilege, if the privilege is required,
5190 	 * bitwise or the request type onto the priv_granted mask.
5191 	 */
5192 	priv_granted = 0;
5193 
5194 	if (type == VDIR) {
5195 		/*
5196 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5197 		 * requests, instead of PRIV_VFS_EXEC.
5198 		 */
5199 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5200 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5201 			priv_granted |= VEXEC;
5202 	} else {
5203 		/*
5204 		 * Ensure that at least one execute bit is on. Otherwise,
5205 		 * a privileged user will always succeed, and we don't want
5206 		 * this to happen unless the file really is executable.
5207 		 */
5208 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5209 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5210 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5211 			priv_granted |= VEXEC;
5212 	}
5213 
5214 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5215 	    !priv_check_cred(cred, PRIV_VFS_READ))
5216 		priv_granted |= VREAD;
5217 
5218 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5219 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5220 		priv_granted |= (VWRITE | VAPPEND);
5221 
5222 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5223 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5224 		priv_granted |= VADMIN;
5225 
5226 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5227 		/* XXX audit: privilege used */
5228 		if (privused != NULL)
5229 			*privused = 1;
5230 		return (0);
5231 	}
5232 
5233 	return ((accmode & VADMIN) ? EPERM : EACCES);
5234 }
5235 
5236 /*
5237  * Credential check based on process requesting service, and per-attribute
5238  * permissions.
5239  */
5240 int
5241 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5242     struct thread *td, accmode_t accmode)
5243 {
5244 
5245 	/*
5246 	 * Kernel-invoked always succeeds.
5247 	 */
5248 	if (cred == NOCRED)
5249 		return (0);
5250 
5251 	/*
5252 	 * Do not allow privileged processes in jail to directly manipulate
5253 	 * system attributes.
5254 	 */
5255 	switch (attrnamespace) {
5256 	case EXTATTR_NAMESPACE_SYSTEM:
5257 		/* Potentially should be: return (EPERM); */
5258 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5259 	case EXTATTR_NAMESPACE_USER:
5260 		return (VOP_ACCESS(vp, accmode, cred, td));
5261 	default:
5262 		return (EPERM);
5263 	}
5264 }
5265 
5266 #ifdef DEBUG_VFS_LOCKS
5267 /*
5268  * This only exists to suppress warnings from unlocked specfs accesses.  It is
5269  * no longer ok to have an unlocked VFS.
5270  */
5271 #define	IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL ||		\
5272 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
5273 
5274 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5275 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5276     "Drop into debugger on lock violation");
5277 
5278 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5279 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5280     0, "Check for interlock across VOPs");
5281 
5282 int vfs_badlock_print = 1;	/* Print lock violations. */
5283 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5284     0, "Print lock violations");
5285 
5286 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5287 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5288     0, "Print vnode details on lock violations");
5289 
5290 #ifdef KDB
5291 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5292 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5293     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5294 #endif
5295 
5296 static void
5297 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5298 {
5299 
5300 #ifdef KDB
5301 	if (vfs_badlock_backtrace)
5302 		kdb_backtrace();
5303 #endif
5304 	if (vfs_badlock_vnode)
5305 		vn_printf(vp, "vnode ");
5306 	if (vfs_badlock_print)
5307 		printf("%s: %p %s\n", str, (void *)vp, msg);
5308 	if (vfs_badlock_ddb)
5309 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5310 }
5311 
5312 void
5313 assert_vi_locked(struct vnode *vp, const char *str)
5314 {
5315 
5316 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5317 		vfs_badlock("interlock is not locked but should be", str, vp);
5318 }
5319 
5320 void
5321 assert_vi_unlocked(struct vnode *vp, const char *str)
5322 {
5323 
5324 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5325 		vfs_badlock("interlock is locked but should not be", str, vp);
5326 }
5327 
5328 void
5329 assert_vop_locked(struct vnode *vp, const char *str)
5330 {
5331 	int locked;
5332 
5333 	if (!IGNORE_LOCK(vp)) {
5334 		locked = VOP_ISLOCKED(vp);
5335 		if (locked == 0 || locked == LK_EXCLOTHER)
5336 			vfs_badlock("is not locked but should be", str, vp);
5337 	}
5338 }
5339 
5340 void
5341 assert_vop_unlocked(struct vnode *vp, const char *str)
5342 {
5343 
5344 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5345 		vfs_badlock("is locked but should not be", str, vp);
5346 }
5347 
5348 void
5349 assert_vop_elocked(struct vnode *vp, const char *str)
5350 {
5351 
5352 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5353 		vfs_badlock("is not exclusive locked but should be", str, vp);
5354 }
5355 #endif /* DEBUG_VFS_LOCKS */
5356 
5357 void
5358 vop_rename_fail(struct vop_rename_args *ap)
5359 {
5360 
5361 	if (ap->a_tvp != NULL)
5362 		vput(ap->a_tvp);
5363 	if (ap->a_tdvp == ap->a_tvp)
5364 		vrele(ap->a_tdvp);
5365 	else
5366 		vput(ap->a_tdvp);
5367 	vrele(ap->a_fdvp);
5368 	vrele(ap->a_fvp);
5369 }
5370 
5371 void
5372 vop_rename_pre(void *ap)
5373 {
5374 	struct vop_rename_args *a = ap;
5375 
5376 #ifdef DEBUG_VFS_LOCKS
5377 	if (a->a_tvp)
5378 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5379 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5380 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5381 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5382 
5383 	/* Check the source (from). */
5384 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5385 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5386 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5387 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5388 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5389 
5390 	/* Check the target. */
5391 	if (a->a_tvp)
5392 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5393 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5394 #endif
5395 	if (a->a_tdvp != a->a_fdvp)
5396 		vhold(a->a_fdvp);
5397 	if (a->a_tvp != a->a_fvp)
5398 		vhold(a->a_fvp);
5399 	vhold(a->a_tdvp);
5400 	if (a->a_tvp)
5401 		vhold(a->a_tvp);
5402 }
5403 
5404 #ifdef DEBUG_VFS_LOCKS
5405 void
5406 vop_strategy_pre(void *ap)
5407 {
5408 	struct vop_strategy_args *a;
5409 	struct buf *bp;
5410 
5411 	a = ap;
5412 	bp = a->a_bp;
5413 
5414 	/*
5415 	 * Cluster ops lock their component buffers but not the IO container.
5416 	 */
5417 	if ((bp->b_flags & B_CLUSTER) != 0)
5418 		return;
5419 
5420 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5421 		if (vfs_badlock_print)
5422 			printf(
5423 			    "VOP_STRATEGY: bp is not locked but should be\n");
5424 		if (vfs_badlock_ddb)
5425 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5426 	}
5427 }
5428 
5429 void
5430 vop_lock_pre(void *ap)
5431 {
5432 	struct vop_lock1_args *a = ap;
5433 
5434 	if ((a->a_flags & LK_INTERLOCK) == 0)
5435 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5436 	else
5437 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5438 }
5439 
5440 void
5441 vop_lock_post(void *ap, int rc)
5442 {
5443 	struct vop_lock1_args *a = ap;
5444 
5445 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5446 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5447 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5448 }
5449 
5450 void
5451 vop_unlock_pre(void *ap)
5452 {
5453 	struct vop_unlock_args *a = ap;
5454 
5455 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
5456 }
5457 
5458 void
5459 vop_need_inactive_pre(void *ap)
5460 {
5461 	struct vop_need_inactive_args *a = ap;
5462 
5463 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5464 }
5465 
5466 void
5467 vop_need_inactive_post(void *ap, int rc)
5468 {
5469 	struct vop_need_inactive_args *a = ap;
5470 
5471 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5472 }
5473 #endif
5474 
5475 void
5476 vop_create_post(void *ap, int rc)
5477 {
5478 	struct vop_create_args *a = ap;
5479 
5480 	if (!rc)
5481 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5482 }
5483 
5484 void
5485 vop_deleteextattr_post(void *ap, int rc)
5486 {
5487 	struct vop_deleteextattr_args *a = ap;
5488 
5489 	if (!rc)
5490 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5491 }
5492 
5493 void
5494 vop_link_post(void *ap, int rc)
5495 {
5496 	struct vop_link_args *a = ap;
5497 
5498 	if (!rc) {
5499 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
5500 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
5501 	}
5502 }
5503 
5504 void
5505 vop_mkdir_post(void *ap, int rc)
5506 {
5507 	struct vop_mkdir_args *a = ap;
5508 
5509 	if (!rc)
5510 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5511 }
5512 
5513 void
5514 vop_mknod_post(void *ap, int rc)
5515 {
5516 	struct vop_mknod_args *a = ap;
5517 
5518 	if (!rc)
5519 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5520 }
5521 
5522 void
5523 vop_reclaim_post(void *ap, int rc)
5524 {
5525 	struct vop_reclaim_args *a = ap;
5526 
5527 	if (!rc)
5528 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
5529 }
5530 
5531 void
5532 vop_remove_post(void *ap, int rc)
5533 {
5534 	struct vop_remove_args *a = ap;
5535 
5536 	if (!rc) {
5537 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5538 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5539 	}
5540 }
5541 
5542 void
5543 vop_rename_post(void *ap, int rc)
5544 {
5545 	struct vop_rename_args *a = ap;
5546 	long hint;
5547 
5548 	if (!rc) {
5549 		hint = NOTE_WRITE;
5550 		if (a->a_fdvp == a->a_tdvp) {
5551 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5552 				hint |= NOTE_LINK;
5553 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5554 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5555 		} else {
5556 			hint |= NOTE_EXTEND;
5557 			if (a->a_fvp->v_type == VDIR)
5558 				hint |= NOTE_LINK;
5559 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5560 
5561 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5562 			    a->a_tvp->v_type == VDIR)
5563 				hint &= ~NOTE_LINK;
5564 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5565 		}
5566 
5567 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
5568 		if (a->a_tvp)
5569 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
5570 	}
5571 	if (a->a_tdvp != a->a_fdvp)
5572 		vdrop(a->a_fdvp);
5573 	if (a->a_tvp != a->a_fvp)
5574 		vdrop(a->a_fvp);
5575 	vdrop(a->a_tdvp);
5576 	if (a->a_tvp)
5577 		vdrop(a->a_tvp);
5578 }
5579 
5580 void
5581 vop_rmdir_post(void *ap, int rc)
5582 {
5583 	struct vop_rmdir_args *a = ap;
5584 
5585 	if (!rc) {
5586 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
5587 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
5588 	}
5589 }
5590 
5591 void
5592 vop_setattr_post(void *ap, int rc)
5593 {
5594 	struct vop_setattr_args *a = ap;
5595 
5596 	if (!rc)
5597 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5598 }
5599 
5600 void
5601 vop_setextattr_post(void *ap, int rc)
5602 {
5603 	struct vop_setextattr_args *a = ap;
5604 
5605 	if (!rc)
5606 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5607 }
5608 
5609 void
5610 vop_symlink_post(void *ap, int rc)
5611 {
5612 	struct vop_symlink_args *a = ap;
5613 
5614 	if (!rc)
5615 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
5616 }
5617 
5618 void
5619 vop_open_post(void *ap, int rc)
5620 {
5621 	struct vop_open_args *a = ap;
5622 
5623 	if (!rc)
5624 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
5625 }
5626 
5627 void
5628 vop_close_post(void *ap, int rc)
5629 {
5630 	struct vop_close_args *a = ap;
5631 
5632 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
5633 	    !VN_IS_DOOMED(a->a_vp))) {
5634 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
5635 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
5636 	}
5637 }
5638 
5639 void
5640 vop_read_post(void *ap, int rc)
5641 {
5642 	struct vop_read_args *a = ap;
5643 
5644 	if (!rc)
5645 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5646 }
5647 
5648 void
5649 vop_readdir_post(void *ap, int rc)
5650 {
5651 	struct vop_readdir_args *a = ap;
5652 
5653 	if (!rc)
5654 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5655 }
5656 
5657 static struct knlist fs_knlist;
5658 
5659 static void
5660 vfs_event_init(void *arg)
5661 {
5662 	knlist_init_mtx(&fs_knlist, NULL);
5663 }
5664 /* XXX - correct order? */
5665 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5666 
5667 void
5668 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5669 {
5670 
5671 	KNOTE_UNLOCKED(&fs_knlist, event);
5672 }
5673 
5674 static int	filt_fsattach(struct knote *kn);
5675 static void	filt_fsdetach(struct knote *kn);
5676 static int	filt_fsevent(struct knote *kn, long hint);
5677 
5678 struct filterops fs_filtops = {
5679 	.f_isfd = 0,
5680 	.f_attach = filt_fsattach,
5681 	.f_detach = filt_fsdetach,
5682 	.f_event = filt_fsevent
5683 };
5684 
5685 static int
5686 filt_fsattach(struct knote *kn)
5687 {
5688 
5689 	kn->kn_flags |= EV_CLEAR;
5690 	knlist_add(&fs_knlist, kn, 0);
5691 	return (0);
5692 }
5693 
5694 static void
5695 filt_fsdetach(struct knote *kn)
5696 {
5697 
5698 	knlist_remove(&fs_knlist, kn, 0);
5699 }
5700 
5701 static int
5702 filt_fsevent(struct knote *kn, long hint)
5703 {
5704 
5705 	kn->kn_fflags |= hint;
5706 	return (kn->kn_fflags != 0);
5707 }
5708 
5709 static int
5710 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5711 {
5712 	struct vfsidctl vc;
5713 	int error;
5714 	struct mount *mp;
5715 
5716 	error = SYSCTL_IN(req, &vc, sizeof(vc));
5717 	if (error)
5718 		return (error);
5719 	if (vc.vc_vers != VFS_CTL_VERS1)
5720 		return (EINVAL);
5721 	mp = vfs_getvfs(&vc.vc_fsid);
5722 	if (mp == NULL)
5723 		return (ENOENT);
5724 	/* ensure that a specific sysctl goes to the right filesystem. */
5725 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
5726 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
5727 		vfs_rel(mp);
5728 		return (EINVAL);
5729 	}
5730 	VCTLTOREQ(&vc, req);
5731 	error = VFS_SYSCTL(mp, vc.vc_op, req);
5732 	vfs_rel(mp);
5733 	return (error);
5734 }
5735 
5736 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
5737     NULL, 0, sysctl_vfs_ctl, "",
5738     "Sysctl by fsid");
5739 
5740 /*
5741  * Function to initialize a va_filerev field sensibly.
5742  * XXX: Wouldn't a random number make a lot more sense ??
5743  */
5744 u_quad_t
5745 init_va_filerev(void)
5746 {
5747 	struct bintime bt;
5748 
5749 	getbinuptime(&bt);
5750 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5751 }
5752 
5753 static int	filt_vfsread(struct knote *kn, long hint);
5754 static int	filt_vfswrite(struct knote *kn, long hint);
5755 static int	filt_vfsvnode(struct knote *kn, long hint);
5756 static void	filt_vfsdetach(struct knote *kn);
5757 static struct filterops vfsread_filtops = {
5758 	.f_isfd = 1,
5759 	.f_detach = filt_vfsdetach,
5760 	.f_event = filt_vfsread
5761 };
5762 static struct filterops vfswrite_filtops = {
5763 	.f_isfd = 1,
5764 	.f_detach = filt_vfsdetach,
5765 	.f_event = filt_vfswrite
5766 };
5767 static struct filterops vfsvnode_filtops = {
5768 	.f_isfd = 1,
5769 	.f_detach = filt_vfsdetach,
5770 	.f_event = filt_vfsvnode
5771 };
5772 
5773 static void
5774 vfs_knllock(void *arg)
5775 {
5776 	struct vnode *vp = arg;
5777 
5778 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5779 }
5780 
5781 static void
5782 vfs_knlunlock(void *arg)
5783 {
5784 	struct vnode *vp = arg;
5785 
5786 	VOP_UNLOCK(vp);
5787 }
5788 
5789 static void
5790 vfs_knl_assert_locked(void *arg)
5791 {
5792 #ifdef DEBUG_VFS_LOCKS
5793 	struct vnode *vp = arg;
5794 
5795 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
5796 #endif
5797 }
5798 
5799 static void
5800 vfs_knl_assert_unlocked(void *arg)
5801 {
5802 #ifdef DEBUG_VFS_LOCKS
5803 	struct vnode *vp = arg;
5804 
5805 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
5806 #endif
5807 }
5808 
5809 int
5810 vfs_kqfilter(struct vop_kqfilter_args *ap)
5811 {
5812 	struct vnode *vp = ap->a_vp;
5813 	struct knote *kn = ap->a_kn;
5814 	struct knlist *knl;
5815 
5816 	switch (kn->kn_filter) {
5817 	case EVFILT_READ:
5818 		kn->kn_fop = &vfsread_filtops;
5819 		break;
5820 	case EVFILT_WRITE:
5821 		kn->kn_fop = &vfswrite_filtops;
5822 		break;
5823 	case EVFILT_VNODE:
5824 		kn->kn_fop = &vfsvnode_filtops;
5825 		break;
5826 	default:
5827 		return (EINVAL);
5828 	}
5829 
5830 	kn->kn_hook = (caddr_t)vp;
5831 
5832 	v_addpollinfo(vp);
5833 	if (vp->v_pollinfo == NULL)
5834 		return (ENOMEM);
5835 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
5836 	vhold(vp);
5837 	knlist_add(knl, kn, 0);
5838 
5839 	return (0);
5840 }
5841 
5842 /*
5843  * Detach knote from vnode
5844  */
5845 static void
5846 filt_vfsdetach(struct knote *kn)
5847 {
5848 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5849 
5850 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5851 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5852 	vdrop(vp);
5853 }
5854 
5855 /*ARGSUSED*/
5856 static int
5857 filt_vfsread(struct knote *kn, long hint)
5858 {
5859 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5860 	struct vattr va;
5861 	int res;
5862 
5863 	/*
5864 	 * filesystem is gone, so set the EOF flag and schedule
5865 	 * the knote for deletion.
5866 	 */
5867 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5868 		VI_LOCK(vp);
5869 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5870 		VI_UNLOCK(vp);
5871 		return (1);
5872 	}
5873 
5874 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5875 		return (0);
5876 
5877 	VI_LOCK(vp);
5878 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5879 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5880 	VI_UNLOCK(vp);
5881 	return (res);
5882 }
5883 
5884 /*ARGSUSED*/
5885 static int
5886 filt_vfswrite(struct knote *kn, long hint)
5887 {
5888 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5889 
5890 	VI_LOCK(vp);
5891 
5892 	/*
5893 	 * filesystem is gone, so set the EOF flag and schedule
5894 	 * the knote for deletion.
5895 	 */
5896 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5897 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5898 
5899 	kn->kn_data = 0;
5900 	VI_UNLOCK(vp);
5901 	return (1);
5902 }
5903 
5904 static int
5905 filt_vfsvnode(struct knote *kn, long hint)
5906 {
5907 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5908 	int res;
5909 
5910 	VI_LOCK(vp);
5911 	if (kn->kn_sfflags & hint)
5912 		kn->kn_fflags |= hint;
5913 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5914 		kn->kn_flags |= EV_EOF;
5915 		VI_UNLOCK(vp);
5916 		return (1);
5917 	}
5918 	res = (kn->kn_fflags != 0);
5919 	VI_UNLOCK(vp);
5920 	return (res);
5921 }
5922 
5923 /*
5924  * Returns whether the directory is empty or not.
5925  * If it is empty, the return value is 0; otherwise
5926  * the return value is an error value (which may
5927  * be ENOTEMPTY).
5928  */
5929 int
5930 vfs_emptydir(struct vnode *vp)
5931 {
5932 	struct uio uio;
5933 	struct iovec iov;
5934 	struct dirent *dirent, *dp, *endp;
5935 	int error, eof;
5936 
5937 	error = 0;
5938 	eof = 0;
5939 
5940 	ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
5941 
5942 	dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
5943 	iov.iov_base = dirent;
5944 	iov.iov_len = sizeof(struct dirent);
5945 
5946 	uio.uio_iov = &iov;
5947 	uio.uio_iovcnt = 1;
5948 	uio.uio_offset = 0;
5949 	uio.uio_resid = sizeof(struct dirent);
5950 	uio.uio_segflg = UIO_SYSSPACE;
5951 	uio.uio_rw = UIO_READ;
5952 	uio.uio_td = curthread;
5953 
5954 	while (eof == 0 && error == 0) {
5955 		error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
5956 		    NULL, NULL);
5957 		if (error != 0)
5958 			break;
5959 		endp = (void *)((uint8_t *)dirent +
5960 		    sizeof(struct dirent) - uio.uio_resid);
5961 		for (dp = dirent; dp < endp;
5962 		     dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
5963 			if (dp->d_type == DT_WHT)
5964 				continue;
5965 			if (dp->d_namlen == 0)
5966 				continue;
5967 			if (dp->d_type != DT_DIR &&
5968 			    dp->d_type != DT_UNKNOWN) {
5969 				error = ENOTEMPTY;
5970 				break;
5971 			}
5972 			if (dp->d_namlen > 2) {
5973 				error = ENOTEMPTY;
5974 				break;
5975 			}
5976 			if (dp->d_namlen == 1 &&
5977 			    dp->d_name[0] != '.') {
5978 				error = ENOTEMPTY;
5979 				break;
5980 			}
5981 			if (dp->d_namlen == 2 &&
5982 			    dp->d_name[1] != '.') {
5983 				error = ENOTEMPTY;
5984 				break;
5985 			}
5986 			uio.uio_resid = sizeof(struct dirent);
5987 		}
5988 	}
5989 	free(dirent, M_TEMP);
5990 	return (error);
5991 }
5992 
5993 int
5994 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5995 {
5996 	int error;
5997 
5998 	if (dp->d_reclen > ap->a_uio->uio_resid)
5999 		return (ENAMETOOLONG);
6000 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
6001 	if (error) {
6002 		if (ap->a_ncookies != NULL) {
6003 			if (ap->a_cookies != NULL)
6004 				free(ap->a_cookies, M_TEMP);
6005 			ap->a_cookies = NULL;
6006 			*ap->a_ncookies = 0;
6007 		}
6008 		return (error);
6009 	}
6010 	if (ap->a_ncookies == NULL)
6011 		return (0);
6012 
6013 	KASSERT(ap->a_cookies,
6014 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6015 
6016 	*ap->a_cookies = realloc(*ap->a_cookies,
6017 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
6018 	(*ap->a_cookies)[*ap->a_ncookies] = off;
6019 	*ap->a_ncookies += 1;
6020 	return (0);
6021 }
6022 
6023 /*
6024  * The purpose of this routine is to remove granularity from accmode_t,
6025  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6026  * VADMIN and VAPPEND.
6027  *
6028  * If it returns 0, the caller is supposed to continue with the usual
6029  * access checks using 'accmode' as modified by this routine.  If it
6030  * returns nonzero value, the caller is supposed to return that value
6031  * as errno.
6032  *
6033  * Note that after this routine runs, accmode may be zero.
6034  */
6035 int
6036 vfs_unixify_accmode(accmode_t *accmode)
6037 {
6038 	/*
6039 	 * There is no way to specify explicit "deny" rule using
6040 	 * file mode or POSIX.1e ACLs.
6041 	 */
6042 	if (*accmode & VEXPLICIT_DENY) {
6043 		*accmode = 0;
6044 		return (0);
6045 	}
6046 
6047 	/*
6048 	 * None of these can be translated into usual access bits.
6049 	 * Also, the common case for NFSv4 ACLs is to not contain
6050 	 * either of these bits. Caller should check for VWRITE
6051 	 * on the containing directory instead.
6052 	 */
6053 	if (*accmode & (VDELETE_CHILD | VDELETE))
6054 		return (EPERM);
6055 
6056 	if (*accmode & VADMIN_PERMS) {
6057 		*accmode &= ~VADMIN_PERMS;
6058 		*accmode |= VADMIN;
6059 	}
6060 
6061 	/*
6062 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6063 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6064 	 */
6065 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6066 
6067 	return (0);
6068 }
6069 
6070 /*
6071  * Clear out a doomed vnode (if any) and replace it with a new one as long
6072  * as the fs is not being unmounted. Return the root vnode to the caller.
6073  */
6074 static int __noinline
6075 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6076 {
6077 	struct vnode *vp;
6078 	int error;
6079 
6080 restart:
6081 	if (mp->mnt_rootvnode != NULL) {
6082 		MNT_ILOCK(mp);
6083 		vp = mp->mnt_rootvnode;
6084 		if (vp != NULL) {
6085 			if (!VN_IS_DOOMED(vp)) {
6086 				vrefact(vp);
6087 				MNT_IUNLOCK(mp);
6088 				error = vn_lock(vp, flags);
6089 				if (error == 0) {
6090 					*vpp = vp;
6091 					return (0);
6092 				}
6093 				vrele(vp);
6094 				goto restart;
6095 			}
6096 			/*
6097 			 * Clear the old one.
6098 			 */
6099 			mp->mnt_rootvnode = NULL;
6100 		}
6101 		MNT_IUNLOCK(mp);
6102 		if (vp != NULL) {
6103 			vfs_op_barrier_wait(mp);
6104 			vrele(vp);
6105 		}
6106 	}
6107 	error = VFS_CACHEDROOT(mp, flags, vpp);
6108 	if (error != 0)
6109 		return (error);
6110 	if (mp->mnt_vfs_ops == 0) {
6111 		MNT_ILOCK(mp);
6112 		if (mp->mnt_vfs_ops != 0) {
6113 			MNT_IUNLOCK(mp);
6114 			return (0);
6115 		}
6116 		if (mp->mnt_rootvnode == NULL) {
6117 			vrefact(*vpp);
6118 			mp->mnt_rootvnode = *vpp;
6119 		} else {
6120 			if (mp->mnt_rootvnode != *vpp) {
6121 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6122 					panic("%s: mismatch between vnode returned "
6123 					    " by VFS_CACHEDROOT and the one cached "
6124 					    " (%p != %p)",
6125 					    __func__, *vpp, mp->mnt_rootvnode);
6126 				}
6127 			}
6128 		}
6129 		MNT_IUNLOCK(mp);
6130 	}
6131 	return (0);
6132 }
6133 
6134 int
6135 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6136 {
6137 	struct vnode *vp;
6138 	int error;
6139 
6140 	if (!vfs_op_thread_enter(mp))
6141 		return (vfs_cache_root_fallback(mp, flags, vpp));
6142 	vp = (struct vnode *)atomic_load_ptr(&mp->mnt_rootvnode);
6143 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6144 		vfs_op_thread_exit(mp);
6145 		return (vfs_cache_root_fallback(mp, flags, vpp));
6146 	}
6147 	vrefact(vp);
6148 	vfs_op_thread_exit(mp);
6149 	error = vn_lock(vp, flags);
6150 	if (error != 0) {
6151 		vrele(vp);
6152 		return (vfs_cache_root_fallback(mp, flags, vpp));
6153 	}
6154 	*vpp = vp;
6155 	return (0);
6156 }
6157 
6158 struct vnode *
6159 vfs_cache_root_clear(struct mount *mp)
6160 {
6161 	struct vnode *vp;
6162 
6163 	/*
6164 	 * ops > 0 guarantees there is nobody who can see this vnode
6165 	 */
6166 	MPASS(mp->mnt_vfs_ops > 0);
6167 	vp = mp->mnt_rootvnode;
6168 	mp->mnt_rootvnode = NULL;
6169 	return (vp);
6170 }
6171 
6172 void
6173 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6174 {
6175 
6176 	MPASS(mp->mnt_vfs_ops > 0);
6177 	vrefact(vp);
6178 	mp->mnt_rootvnode = vp;
6179 }
6180 
6181 /*
6182  * These are helper functions for filesystems to traverse all
6183  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6184  *
6185  * This interface replaces MNT_VNODE_FOREACH.
6186  */
6187 
6188 struct vnode *
6189 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6190 {
6191 	struct vnode *vp;
6192 
6193 	if (should_yield())
6194 		kern_yield(PRI_USER);
6195 	MNT_ILOCK(mp);
6196 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6197 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6198 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6199 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6200 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6201 			continue;
6202 		VI_LOCK(vp);
6203 		if (VN_IS_DOOMED(vp)) {
6204 			VI_UNLOCK(vp);
6205 			continue;
6206 		}
6207 		break;
6208 	}
6209 	if (vp == NULL) {
6210 		__mnt_vnode_markerfree_all(mvp, mp);
6211 		/* MNT_IUNLOCK(mp); -- done in above function */
6212 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6213 		return (NULL);
6214 	}
6215 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6216 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6217 	MNT_IUNLOCK(mp);
6218 	return (vp);
6219 }
6220 
6221 struct vnode *
6222 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6223 {
6224 	struct vnode *vp;
6225 
6226 	*mvp = vn_alloc_marker(mp);
6227 	MNT_ILOCK(mp);
6228 	MNT_REF(mp);
6229 
6230 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6231 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6232 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6233 			continue;
6234 		VI_LOCK(vp);
6235 		if (VN_IS_DOOMED(vp)) {
6236 			VI_UNLOCK(vp);
6237 			continue;
6238 		}
6239 		break;
6240 	}
6241 	if (vp == NULL) {
6242 		MNT_REL(mp);
6243 		MNT_IUNLOCK(mp);
6244 		vn_free_marker(*mvp);
6245 		*mvp = NULL;
6246 		return (NULL);
6247 	}
6248 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6249 	MNT_IUNLOCK(mp);
6250 	return (vp);
6251 }
6252 
6253 void
6254 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6255 {
6256 
6257 	if (*mvp == NULL) {
6258 		MNT_IUNLOCK(mp);
6259 		return;
6260 	}
6261 
6262 	mtx_assert(MNT_MTX(mp), MA_OWNED);
6263 
6264 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6265 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6266 	MNT_REL(mp);
6267 	MNT_IUNLOCK(mp);
6268 	vn_free_marker(*mvp);
6269 	*mvp = NULL;
6270 }
6271 
6272 /*
6273  * These are helper functions for filesystems to traverse their
6274  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6275  */
6276 static void
6277 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6278 {
6279 
6280 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6281 
6282 	MNT_ILOCK(mp);
6283 	MNT_REL(mp);
6284 	MNT_IUNLOCK(mp);
6285 	vn_free_marker(*mvp);
6286 	*mvp = NULL;
6287 }
6288 
6289 /*
6290  * Relock the mp mount vnode list lock with the vp vnode interlock in the
6291  * conventional lock order during mnt_vnode_next_lazy iteration.
6292  *
6293  * On entry, the mount vnode list lock is held and the vnode interlock is not.
6294  * The list lock is dropped and reacquired.  On success, both locks are held.
6295  * On failure, the mount vnode list lock is held but the vnode interlock is
6296  * not, and the procedure may have yielded.
6297  */
6298 static bool
6299 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6300     struct vnode *vp)
6301 {
6302 
6303 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6304 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6305 	    ("%s: bad marker", __func__));
6306 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6307 	    ("%s: inappropriate vnode", __func__));
6308 	ASSERT_VI_UNLOCKED(vp, __func__);
6309 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6310 
6311 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6312 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6313 
6314 	/*
6315 	 * Note we may be racing against vdrop which transitioned the hold
6316 	 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6317 	 * if we are the only user after we get the interlock we will just
6318 	 * vdrop.
6319 	 */
6320 	vhold(vp);
6321 	mtx_unlock(&mp->mnt_listmtx);
6322 	VI_LOCK(vp);
6323 	if (VN_IS_DOOMED(vp)) {
6324 		VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6325 		goto out_lost;
6326 	}
6327 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6328 	/*
6329 	 * There is nothing to do if we are the last user.
6330 	 */
6331 	if (!refcount_release_if_not_last(&vp->v_holdcnt))
6332 		goto out_lost;
6333 	mtx_lock(&mp->mnt_listmtx);
6334 	return (true);
6335 out_lost:
6336 	vdropl(vp);
6337 	maybe_yield();
6338 	mtx_lock(&mp->mnt_listmtx);
6339 	return (false);
6340 }
6341 
6342 static struct vnode *
6343 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6344     void *cbarg)
6345 {
6346 	struct vnode *vp;
6347 
6348 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6349 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6350 restart:
6351 	vp = TAILQ_NEXT(*mvp, v_lazylist);
6352 	while (vp != NULL) {
6353 		if (vp->v_type == VMARKER) {
6354 			vp = TAILQ_NEXT(vp, v_lazylist);
6355 			continue;
6356 		}
6357 		/*
6358 		 * See if we want to process the vnode. Note we may encounter a
6359 		 * long string of vnodes we don't care about and hog the list
6360 		 * as a result. Check for it and requeue the marker.
6361 		 */
6362 		VNPASS(!VN_IS_DOOMED(vp), vp);
6363 		if (!cb(vp, cbarg)) {
6364 			if (!should_yield()) {
6365 				vp = TAILQ_NEXT(vp, v_lazylist);
6366 				continue;
6367 			}
6368 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6369 			    v_lazylist);
6370 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6371 			    v_lazylist);
6372 			mtx_unlock(&mp->mnt_listmtx);
6373 			kern_yield(PRI_USER);
6374 			mtx_lock(&mp->mnt_listmtx);
6375 			goto restart;
6376 		}
6377 		/*
6378 		 * Try-lock because this is the wrong lock order.
6379 		 */
6380 		if (!VI_TRYLOCK(vp) &&
6381 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6382 			goto restart;
6383 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6384 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6385 		    ("alien vnode on the lazy list %p %p", vp, mp));
6386 		VNPASS(vp->v_mount == mp, vp);
6387 		VNPASS(!VN_IS_DOOMED(vp), vp);
6388 		break;
6389 	}
6390 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6391 
6392 	/* Check if we are done */
6393 	if (vp == NULL) {
6394 		mtx_unlock(&mp->mnt_listmtx);
6395 		mnt_vnode_markerfree_lazy(mvp, mp);
6396 		return (NULL);
6397 	}
6398 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6399 	mtx_unlock(&mp->mnt_listmtx);
6400 	ASSERT_VI_LOCKED(vp, "lazy iter");
6401 	return (vp);
6402 }
6403 
6404 struct vnode *
6405 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6406     void *cbarg)
6407 {
6408 
6409 	if (should_yield())
6410 		kern_yield(PRI_USER);
6411 	mtx_lock(&mp->mnt_listmtx);
6412 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6413 }
6414 
6415 struct vnode *
6416 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6417     void *cbarg)
6418 {
6419 	struct vnode *vp;
6420 
6421 	if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6422 		return (NULL);
6423 
6424 	*mvp = vn_alloc_marker(mp);
6425 	MNT_ILOCK(mp);
6426 	MNT_REF(mp);
6427 	MNT_IUNLOCK(mp);
6428 
6429 	mtx_lock(&mp->mnt_listmtx);
6430 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6431 	if (vp == NULL) {
6432 		mtx_unlock(&mp->mnt_listmtx);
6433 		mnt_vnode_markerfree_lazy(mvp, mp);
6434 		return (NULL);
6435 	}
6436 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6437 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6438 }
6439 
6440 void
6441 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6442 {
6443 
6444 	if (*mvp == NULL)
6445 		return;
6446 
6447 	mtx_lock(&mp->mnt_listmtx);
6448 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6449 	mtx_unlock(&mp->mnt_listmtx);
6450 	mnt_vnode_markerfree_lazy(mvp, mp);
6451 }
6452 
6453 int
6454 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6455 {
6456 
6457 	if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6458 		cnp->cn_flags &= ~NOEXECCHECK;
6459 		return (0);
6460 	}
6461 
6462 	return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, cnp->cn_thread));
6463 }
6464