xref: /freebsd/sys/kern/vfs_subr.c (revision 6683132d54bd6d589889e43dabdc53d35e38a028)
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_usecount(struct vnode *);
111 static void	v_incr_usecount_locked(struct vnode *);
112 static void	v_incr_devcount(struct vnode *);
113 static void	v_decr_devcount(struct vnode *);
114 static void	vgonel(struct vnode *);
115 static void	vfs_knllock(void *arg);
116 static void	vfs_knlunlock(void *arg);
117 static void	vfs_knl_assert_locked(void *arg);
118 static void	vfs_knl_assert_unlocked(void *arg);
119 static void	vnlru_return_batches(struct vfsops *mnt_op);
120 static void	destroy_vpollinfo(struct vpollinfo *vi);
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 VI_DOOMED vnode.
141  */
142 static unsigned long	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 static u_long mnt_free_list_batch = 128;
152 SYSCTL_ULONG(_vfs, OID_AUTO, mnt_free_list_batch, CTLFLAG_RW,
153     &mnt_free_list_batch, 0, "Limit of vnodes held on mnt's free list");
154 
155 /*
156  * Conversion tables for conversion from vnode types to inode formats
157  * and back.
158  */
159 enum vtype iftovt_tab[16] = {
160 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
161 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
162 };
163 int vttoif_tab[10] = {
164 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
165 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
166 };
167 
168 /*
169  * List of vnodes that are ready for recycling.
170  */
171 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
172 
173 /*
174  * "Free" vnode target.  Free vnodes are rarely completely free, but are
175  * just ones that are cheap to recycle.  Usually they are for files which
176  * have been stat'd but not read; these usually have inode and namecache
177  * data attached to them.  This target is the preferred minimum size of a
178  * sub-cache consisting mostly of such files. The system balances the size
179  * of this sub-cache with its complement to try to prevent either from
180  * thrashing while the other is relatively inactive.  The targets express
181  * a preference for the best balance.
182  *
183  * "Above" this target there are 2 further targets (watermarks) related
184  * to recyling of free vnodes.  In the best-operating case, the cache is
185  * exactly full, the free list has size between vlowat and vhiwat above the
186  * free target, and recycling from it and normal use maintains this state.
187  * Sometimes the free list is below vlowat or even empty, but this state
188  * is even better for immediate use provided the cache is not full.
189  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
190  * ones) to reach one of these states.  The watermarks are currently hard-
191  * coded as 4% and 9% of the available space higher.  These and the default
192  * of 25% for wantfreevnodes are too large if the memory size is large.
193  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
194  * whenever vnlru_proc() becomes active.
195  */
196 static u_long wantfreevnodes;
197 SYSCTL_ULONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW,
198     &wantfreevnodes, 0, "Target for minimum number of \"free\" vnodes");
199 static u_long freevnodes;
200 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
201     &freevnodes, 0, "Number of \"free\" vnodes");
202 
203 static counter_u64_t recycles_count;
204 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
205     "Number of vnodes recycled to meet vnode cache targets");
206 
207 /*
208  * Various variables used for debugging the new implementation of
209  * reassignbuf().
210  * XXX these are probably of (very) limited utility now.
211  */
212 static int reassignbufcalls;
213 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0,
214     "Number of calls to reassignbuf");
215 
216 static counter_u64_t free_owe_inact;
217 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, free_owe_inact, CTLFLAG_RD, &free_owe_inact,
218     "Number of times free vnodes kept on active list due to VFS "
219     "owing inactivation");
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_free_list
227  *	numvnodes
228  *	freevnodes
229  */
230 static struct mtx vnode_free_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 /*
301  * When shutting down the syncer, run it at four times normal speed.
302  */
303 #define SYNCER_SHUTDOWN_SPEEDUP		4
304 static int sync_vnode_count;
305 static int syncer_worklist_len;
306 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
307     syncer_state;
308 
309 /* Target for maximum number of vnodes. */
310 int desiredvnodes;
311 static int gapvnodes;		/* gap between wanted and desired */
312 static int vhiwat;		/* enough extras after expansion */
313 static int vlowat;		/* minimal extras before expansion */
314 static int vstir;		/* nonzero to stir non-free vnodes */
315 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
316 
317 static int
318 sysctl_update_desiredvnodes(SYSCTL_HANDLER_ARGS)
319 {
320 	int error, old_desiredvnodes;
321 
322 	old_desiredvnodes = desiredvnodes;
323 	if ((error = sysctl_handle_int(oidp, arg1, arg2, req)) != 0)
324 		return (error);
325 	if (old_desiredvnodes != desiredvnodes) {
326 		wantfreevnodes = desiredvnodes / 4;
327 		/* XXX locking seems to be incomplete. */
328 		vfs_hash_changesize(desiredvnodes);
329 		cache_changesize(desiredvnodes);
330 	}
331 	return (0);
332 }
333 
334 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
335     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, &desiredvnodes, 0,
336     sysctl_update_desiredvnodes, "I", "Target for maximum number of vnodes");
337 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
338     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
339 static int vnlru_nowhere;
340 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
341     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
342 
343 static int
344 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
345 {
346 	struct vnode *vp;
347 	struct nameidata nd;
348 	char *buf;
349 	unsigned long ndflags;
350 	int error;
351 
352 	if (req->newptr == NULL)
353 		return (EINVAL);
354 	if (req->newlen > PATH_MAX)
355 		return (E2BIG);
356 
357 	buf = malloc(PATH_MAX + 1, M_TEMP, M_WAITOK);
358 	error = SYSCTL_IN(req, buf, req->newlen);
359 	if (error != 0)
360 		goto out;
361 
362 	buf[req->newlen] = '\0';
363 
364 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | NOCACHE | SAVENAME;
365 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread);
366 	if ((error = namei(&nd)) != 0)
367 		goto out;
368 	vp = nd.ni_vp;
369 
370 	if ((vp->v_iflag & VI_DOOMED) != 0) {
371 		/*
372 		 * This vnode is being recycled.  Return != 0 to let the caller
373 		 * know that the sysctl had no effect.  Return EAGAIN because a
374 		 * subsequent call will likely succeed (since namei will create
375 		 * a new vnode if necessary)
376 		 */
377 		error = EAGAIN;
378 		goto putvnode;
379 	}
380 
381 	counter_u64_add(recycles_count, 1);
382 	vgone(vp);
383 putvnode:
384 	NDFREE(&nd, 0);
385 out:
386 	free(buf, M_TEMP);
387 	return (error);
388 }
389 
390 static int
391 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
392 {
393 	struct thread *td = curthread;
394 	struct vnode *vp;
395 	struct file *fp;
396 	int error;
397 	int fd;
398 
399 	if (req->newptr == NULL)
400 		return (EBADF);
401 
402         error = sysctl_handle_int(oidp, &fd, 0, req);
403         if (error != 0)
404                 return (error);
405 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
406 	if (error != 0)
407 		return (error);
408 	vp = fp->f_vnode;
409 
410 	error = vn_lock(vp, LK_EXCLUSIVE);
411 	if (error != 0)
412 		goto drop;
413 
414 	counter_u64_add(recycles_count, 1);
415 	vgone(vp);
416 	VOP_UNLOCK(vp, 0);
417 drop:
418 	fdrop(fp, td);
419 	return (error);
420 }
421 
422 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
423     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
424     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
425 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
426     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
427     sysctl_ftry_reclaim_vnode, "I",
428     "Try to reclaim a vnode by its file descriptor");
429 
430 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
431 static int vnsz2log;
432 
433 /*
434  * Support for the bufobj clean & dirty pctrie.
435  */
436 static void *
437 buf_trie_alloc(struct pctrie *ptree)
438 {
439 
440 	return uma_zalloc(buf_trie_zone, M_NOWAIT);
441 }
442 
443 static void
444 buf_trie_free(struct pctrie *ptree, void *node)
445 {
446 
447 	uma_zfree(buf_trie_zone, node);
448 }
449 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
450 
451 /*
452  * Initialize the vnode management data structures.
453  *
454  * Reevaluate the following cap on the number of vnodes after the physical
455  * memory size exceeds 512GB.  In the limit, as the physical memory size
456  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
457  */
458 #ifndef	MAXVNODES_MAX
459 #define	MAXVNODES_MAX	(512 * 1024 * 1024 / 64)	/* 8M */
460 #endif
461 
462 /*
463  * Initialize a vnode as it first enters the zone.
464  */
465 static int
466 vnode_init(void *mem, int size, int flags)
467 {
468 	struct vnode *vp;
469 
470 	vp = mem;
471 	bzero(vp, size);
472 	/*
473 	 * Setup locks.
474 	 */
475 	vp->v_vnlock = &vp->v_lock;
476 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
477 	/*
478 	 * By default, don't allow shared locks unless filesystems opt-in.
479 	 */
480 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
481 	    LK_NOSHARE | LK_IS_VNODE);
482 	/*
483 	 * Initialize bufobj.
484 	 */
485 	bufobj_init(&vp->v_bufobj, vp);
486 	/*
487 	 * Initialize namecache.
488 	 */
489 	LIST_INIT(&vp->v_cache_src);
490 	TAILQ_INIT(&vp->v_cache_dst);
491 	/*
492 	 * Initialize rangelocks.
493 	 */
494 	rangelock_init(&vp->v_rl);
495 	return (0);
496 }
497 
498 /*
499  * Free a vnode when it is cleared from the zone.
500  */
501 static void
502 vnode_fini(void *mem, int size)
503 {
504 	struct vnode *vp;
505 	struct bufobj *bo;
506 
507 	vp = mem;
508 	rangelock_destroy(&vp->v_rl);
509 	lockdestroy(vp->v_vnlock);
510 	mtx_destroy(&vp->v_interlock);
511 	bo = &vp->v_bufobj;
512 	rw_destroy(BO_LOCKPTR(bo));
513 }
514 
515 /*
516  * Provide the size of NFS nclnode and NFS fh for calculation of the
517  * vnode memory consumption.  The size is specified directly to
518  * eliminate dependency on NFS-private header.
519  *
520  * Other filesystems may use bigger or smaller (like UFS and ZFS)
521  * private inode data, but the NFS-based estimation is ample enough.
522  * Still, we care about differences in the size between 64- and 32-bit
523  * platforms.
524  *
525  * Namecache structure size is heuristically
526  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
527  */
528 #ifdef _LP64
529 #define	NFS_NCLNODE_SZ	(528 + 64)
530 #define	NC_SZ		148
531 #else
532 #define	NFS_NCLNODE_SZ	(360 + 32)
533 #define	NC_SZ		92
534 #endif
535 
536 static void
537 vntblinit(void *dummy __unused)
538 {
539 	u_int i;
540 	int physvnodes, virtvnodes;
541 
542 	/*
543 	 * Desiredvnodes is a function of the physical memory size and the
544 	 * kernel's heap size.  Generally speaking, it scales with the
545 	 * physical memory size.  The ratio of desiredvnodes to the physical
546 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
547 	 * Thereafter, the
548 	 * marginal ratio of desiredvnodes to the physical memory size is
549 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
550 	 * size.  The memory required by desiredvnodes vnodes and vm objects
551 	 * must not exceed 1/10th of the kernel's heap size.
552 	 */
553 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
554 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
555 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
556 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
557 	desiredvnodes = min(physvnodes, virtvnodes);
558 	if (desiredvnodes > MAXVNODES_MAX) {
559 		if (bootverbose)
560 			printf("Reducing kern.maxvnodes %d -> %d\n",
561 			    desiredvnodes, MAXVNODES_MAX);
562 		desiredvnodes = MAXVNODES_MAX;
563 	}
564 	wantfreevnodes = desiredvnodes / 4;
565 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
566 	TAILQ_INIT(&vnode_free_list);
567 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
568 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
569 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
570 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
571 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
572 	/*
573 	 * Preallocate enough nodes to support one-per buf so that
574 	 * we can not fail an insert.  reassignbuf() callers can not
575 	 * tolerate the insertion failure.
576 	 */
577 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
578 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
579 	    UMA_ZONE_NOFREE | UMA_ZONE_VM);
580 	uma_prealloc(buf_trie_zone, nbuf);
581 
582 	vnodes_created = counter_u64_alloc(M_WAITOK);
583 	recycles_count = counter_u64_alloc(M_WAITOK);
584 	free_owe_inact = counter_u64_alloc(M_WAITOK);
585 
586 	/*
587 	 * Initialize the filesystem syncer.
588 	 */
589 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
590 	    &syncer_mask);
591 	syncer_maxdelay = syncer_mask + 1;
592 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
593 	cv_init(&sync_wakeup, "syncer");
594 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
595 		vnsz2log++;
596 	vnsz2log--;
597 }
598 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
599 
600 
601 /*
602  * Mark a mount point as busy. Used to synchronize access and to delay
603  * unmounting. Eventually, mountlist_mtx is not released on failure.
604  *
605  * vfs_busy() is a custom lock, it can block the caller.
606  * vfs_busy() only sleeps if the unmount is active on the mount point.
607  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
608  * vnode belonging to mp.
609  *
610  * Lookup uses vfs_busy() to traverse mount points.
611  * root fs			var fs
612  * / vnode lock		A	/ vnode lock (/var)		D
613  * /var vnode lock	B	/log vnode lock(/var/log)	E
614  * vfs_busy lock	C	vfs_busy lock			F
615  *
616  * Within each file system, the lock order is C->A->B and F->D->E.
617  *
618  * When traversing across mounts, the system follows that lock order:
619  *
620  *        C->A->B
621  *              |
622  *              +->F->D->E
623  *
624  * The lookup() process for namei("/var") illustrates the process:
625  *  VOP_LOOKUP() obtains B while A is held
626  *  vfs_busy() obtains a shared lock on F while A and B are held
627  *  vput() releases lock on B
628  *  vput() releases lock on A
629  *  VFS_ROOT() obtains lock on D while shared lock on F is held
630  *  vfs_unbusy() releases shared lock on F
631  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
632  *    Attempt to lock A (instead of vp_crossmp) while D is held would
633  *    violate the global order, causing deadlocks.
634  *
635  * dounmount() locks B while F is drained.
636  */
637 int
638 vfs_busy(struct mount *mp, int flags)
639 {
640 
641 	MPASS((flags & ~MBF_MASK) == 0);
642 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
643 
644 	MNT_ILOCK(mp);
645 	MNT_REF(mp);
646 	/*
647 	 * If mount point is currently being unmounted, sleep until the
648 	 * mount point fate is decided.  If thread doing the unmounting fails,
649 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
650 	 * that this mount point has survived the unmount attempt and vfs_busy
651 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
652 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
653 	 * about to be really destroyed.  vfs_busy needs to release its
654 	 * reference on the mount point in this case and return with ENOENT,
655 	 * telling the caller that mount mount it tried to busy is no longer
656 	 * valid.
657 	 */
658 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
659 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
660 			MNT_REL(mp);
661 			MNT_IUNLOCK(mp);
662 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
663 			    __func__);
664 			return (ENOENT);
665 		}
666 		if (flags & MBF_MNTLSTLOCK)
667 			mtx_unlock(&mountlist_mtx);
668 		mp->mnt_kern_flag |= MNTK_MWAIT;
669 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
670 		if (flags & MBF_MNTLSTLOCK)
671 			mtx_lock(&mountlist_mtx);
672 		MNT_ILOCK(mp);
673 	}
674 	if (flags & MBF_MNTLSTLOCK)
675 		mtx_unlock(&mountlist_mtx);
676 	mp->mnt_lockref++;
677 	MNT_IUNLOCK(mp);
678 	return (0);
679 }
680 
681 /*
682  * Free a busy filesystem.
683  */
684 void
685 vfs_unbusy(struct mount *mp)
686 {
687 
688 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
689 	MNT_ILOCK(mp);
690 	MNT_REL(mp);
691 	KASSERT(mp->mnt_lockref > 0, ("negative mnt_lockref"));
692 	mp->mnt_lockref--;
693 	if (mp->mnt_lockref == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
694 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
695 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
696 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
697 		wakeup(&mp->mnt_lockref);
698 	}
699 	MNT_IUNLOCK(mp);
700 }
701 
702 /*
703  * Lookup a mount point by filesystem identifier.
704  */
705 struct mount *
706 vfs_getvfs(fsid_t *fsid)
707 {
708 	struct mount *mp;
709 
710 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
711 	mtx_lock(&mountlist_mtx);
712 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
713 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
714 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
715 			vfs_ref(mp);
716 			mtx_unlock(&mountlist_mtx);
717 			return (mp);
718 		}
719 	}
720 	mtx_unlock(&mountlist_mtx);
721 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
722 	return ((struct mount *) 0);
723 }
724 
725 /*
726  * Lookup a mount point by filesystem identifier, busying it before
727  * returning.
728  *
729  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
730  * cache for popular filesystem identifiers.  The cache is lockess, using
731  * the fact that struct mount's are never freed.  In worst case we may
732  * get pointer to unmounted or even different filesystem, so we have to
733  * check what we got, and go slow way if so.
734  */
735 struct mount *
736 vfs_busyfs(fsid_t *fsid)
737 {
738 #define	FSID_CACHE_SIZE	256
739 	typedef struct mount * volatile vmp_t;
740 	static vmp_t cache[FSID_CACHE_SIZE];
741 	struct mount *mp;
742 	int error;
743 	uint32_t hash;
744 
745 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
746 	hash = fsid->val[0] ^ fsid->val[1];
747 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
748 	mp = cache[hash];
749 	if (mp == NULL ||
750 	    mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
751 	    mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
752 		goto slow;
753 	if (vfs_busy(mp, 0) != 0) {
754 		cache[hash] = NULL;
755 		goto slow;
756 	}
757 	if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
758 	    mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
759 		return (mp);
760 	else
761 	    vfs_unbusy(mp);
762 
763 slow:
764 	mtx_lock(&mountlist_mtx);
765 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
766 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
767 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
768 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
769 			if (error) {
770 				cache[hash] = NULL;
771 				mtx_unlock(&mountlist_mtx);
772 				return (NULL);
773 			}
774 			cache[hash] = mp;
775 			return (mp);
776 		}
777 	}
778 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
779 	mtx_unlock(&mountlist_mtx);
780 	return ((struct mount *) 0);
781 }
782 
783 /*
784  * Check if a user can access privileged mount options.
785  */
786 int
787 vfs_suser(struct mount *mp, struct thread *td)
788 {
789 	int error;
790 
791 	if (jailed(td->td_ucred)) {
792 		/*
793 		 * If the jail of the calling thread lacks permission for
794 		 * this type of file system, deny immediately.
795 		 */
796 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
797 			return (EPERM);
798 
799 		/*
800 		 * If the file system was mounted outside the jail of the
801 		 * calling thread, deny immediately.
802 		 */
803 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
804 			return (EPERM);
805 	}
806 
807 	/*
808 	 * If file system supports delegated administration, we don't check
809 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
810 	 * by the file system itself.
811 	 * If this is not the user that did original mount, we check for
812 	 * the PRIV_VFS_MOUNT_OWNER privilege.
813 	 */
814 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
815 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
816 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
817 			return (error);
818 	}
819 	return (0);
820 }
821 
822 /*
823  * Get a new unique fsid.  Try to make its val[0] unique, since this value
824  * will be used to create fake device numbers for stat().  Also try (but
825  * not so hard) make its val[0] unique mod 2^16, since some emulators only
826  * support 16-bit device numbers.  We end up with unique val[0]'s for the
827  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
828  *
829  * Keep in mind that several mounts may be running in parallel.  Starting
830  * the search one past where the previous search terminated is both a
831  * micro-optimization and a defense against returning the same fsid to
832  * different mounts.
833  */
834 void
835 vfs_getnewfsid(struct mount *mp)
836 {
837 	static uint16_t mntid_base;
838 	struct mount *nmp;
839 	fsid_t tfsid;
840 	int mtype;
841 
842 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
843 	mtx_lock(&mntid_mtx);
844 	mtype = mp->mnt_vfc->vfc_typenum;
845 	tfsid.val[1] = mtype;
846 	mtype = (mtype & 0xFF) << 24;
847 	for (;;) {
848 		tfsid.val[0] = makedev(255,
849 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
850 		mntid_base++;
851 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
852 			break;
853 		vfs_rel(nmp);
854 	}
855 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
856 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
857 	mtx_unlock(&mntid_mtx);
858 }
859 
860 /*
861  * Knob to control the precision of file timestamps:
862  *
863  *   0 = seconds only; nanoseconds zeroed.
864  *   1 = seconds and nanoseconds, accurate within 1/HZ.
865  *   2 = seconds and nanoseconds, truncated to microseconds.
866  * >=3 = seconds and nanoseconds, maximum precision.
867  */
868 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
869 
870 static int timestamp_precision = TSP_USEC;
871 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
872     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
873     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
874     "3+: sec + ns (max. precision))");
875 
876 /*
877  * Get a current timestamp.
878  */
879 void
880 vfs_timestamp(struct timespec *tsp)
881 {
882 	struct timeval tv;
883 
884 	switch (timestamp_precision) {
885 	case TSP_SEC:
886 		tsp->tv_sec = time_second;
887 		tsp->tv_nsec = 0;
888 		break;
889 	case TSP_HZ:
890 		getnanotime(tsp);
891 		break;
892 	case TSP_USEC:
893 		microtime(&tv);
894 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
895 		break;
896 	case TSP_NSEC:
897 	default:
898 		nanotime(tsp);
899 		break;
900 	}
901 }
902 
903 /*
904  * Set vnode attributes to VNOVAL
905  */
906 void
907 vattr_null(struct vattr *vap)
908 {
909 
910 	vap->va_type = VNON;
911 	vap->va_size = VNOVAL;
912 	vap->va_bytes = VNOVAL;
913 	vap->va_mode = VNOVAL;
914 	vap->va_nlink = VNOVAL;
915 	vap->va_uid = VNOVAL;
916 	vap->va_gid = VNOVAL;
917 	vap->va_fsid = VNOVAL;
918 	vap->va_fileid = VNOVAL;
919 	vap->va_blocksize = VNOVAL;
920 	vap->va_rdev = VNOVAL;
921 	vap->va_atime.tv_sec = VNOVAL;
922 	vap->va_atime.tv_nsec = VNOVAL;
923 	vap->va_mtime.tv_sec = VNOVAL;
924 	vap->va_mtime.tv_nsec = VNOVAL;
925 	vap->va_ctime.tv_sec = VNOVAL;
926 	vap->va_ctime.tv_nsec = VNOVAL;
927 	vap->va_birthtime.tv_sec = VNOVAL;
928 	vap->va_birthtime.tv_nsec = VNOVAL;
929 	vap->va_flags = VNOVAL;
930 	vap->va_gen = VNOVAL;
931 	vap->va_vaflags = 0;
932 }
933 
934 /*
935  * This routine is called when we have too many vnodes.  It attempts
936  * to free <count> vnodes and will potentially free vnodes that still
937  * have VM backing store (VM backing store is typically the cause
938  * of a vnode blowout so we want to do this).  Therefore, this operation
939  * is not considered cheap.
940  *
941  * A number of conditions may prevent a vnode from being reclaimed.
942  * the buffer cache may have references on the vnode, a directory
943  * vnode may still have references due to the namei cache representing
944  * underlying files, or the vnode may be in active use.   It is not
945  * desirable to reuse such vnodes.  These conditions may cause the
946  * number of vnodes to reach some minimum value regardless of what
947  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
948  */
949 static int
950 vlrureclaim(struct mount *mp, int reclaim_nc_src, int trigger)
951 {
952 	struct vnode *vp;
953 	int count, done, target;
954 
955 	done = 0;
956 	vn_start_write(NULL, &mp, V_WAIT);
957 	MNT_ILOCK(mp);
958 	count = mp->mnt_nvnodelistsize;
959 	target = count * (int64_t)gapvnodes / imax(desiredvnodes, 1);
960 	target = target / 10 + 1;
961 	while (count != 0 && done < target) {
962 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
963 		while (vp != NULL && vp->v_type == VMARKER)
964 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
965 		if (vp == NULL)
966 			break;
967 		/*
968 		 * XXX LRU is completely broken for non-free vnodes.  First
969 		 * by calling here in mountpoint order, then by moving
970 		 * unselected vnodes to the end here, and most grossly by
971 		 * removing the vlruvp() function that was supposed to
972 		 * maintain the order.  (This function was born broken
973 		 * since syncer problems prevented it doing anything.)  The
974 		 * order is closer to LRC (C = Created).
975 		 *
976 		 * LRU reclaiming of vnodes seems to have last worked in
977 		 * FreeBSD-3 where LRU wasn't mentioned under any spelling.
978 		 * Then there was no hold count, and inactive vnodes were
979 		 * simply put on the free list in LRU order.  The separate
980 		 * lists also break LRU.  We prefer to reclaim from the
981 		 * free list for technical reasons.  This tends to thrash
982 		 * the free list to keep very unrecently used held vnodes.
983 		 * The problem is mitigated by keeping the free list large.
984 		 */
985 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
986 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
987 		--count;
988 		if (!VI_TRYLOCK(vp))
989 			goto next_iter;
990 		/*
991 		 * If it's been deconstructed already, it's still
992 		 * referenced, or it exceeds the trigger, skip it.
993 		 * Also skip free vnodes.  We are trying to make space
994 		 * to expand the free list, not reduce it.
995 		 */
996 		if (vp->v_usecount ||
997 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
998 		    ((vp->v_iflag & VI_FREE) != 0) ||
999 		    (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
1000 		    vp->v_object->resident_page_count > trigger)) {
1001 			VI_UNLOCK(vp);
1002 			goto next_iter;
1003 		}
1004 		MNT_IUNLOCK(mp);
1005 		vholdl(vp);
1006 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) {
1007 			vdrop(vp);
1008 			goto next_iter_mntunlocked;
1009 		}
1010 		VI_LOCK(vp);
1011 		/*
1012 		 * v_usecount may have been bumped after VOP_LOCK() dropped
1013 		 * the vnode interlock and before it was locked again.
1014 		 *
1015 		 * It is not necessary to recheck VI_DOOMED because it can
1016 		 * only be set by another thread that holds both the vnode
1017 		 * lock and vnode interlock.  If another thread has the
1018 		 * vnode lock before we get to VOP_LOCK() and obtains the
1019 		 * vnode interlock after VOP_LOCK() drops the vnode
1020 		 * interlock, the other thread will be unable to drop the
1021 		 * vnode lock before our VOP_LOCK() call fails.
1022 		 */
1023 		if (vp->v_usecount ||
1024 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1025 		    (vp->v_iflag & VI_FREE) != 0 ||
1026 		    (vp->v_object != NULL &&
1027 		    vp->v_object->resident_page_count > trigger)) {
1028 			VOP_UNLOCK(vp, LK_INTERLOCK);
1029 			vdrop(vp);
1030 			goto next_iter_mntunlocked;
1031 		}
1032 		KASSERT((vp->v_iflag & VI_DOOMED) == 0,
1033 		    ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
1034 		counter_u64_add(recycles_count, 1);
1035 		vgonel(vp);
1036 		VOP_UNLOCK(vp, 0);
1037 		vdropl(vp);
1038 		done++;
1039 next_iter_mntunlocked:
1040 		if (!should_yield())
1041 			goto relock_mnt;
1042 		goto yield;
1043 next_iter:
1044 		if (!should_yield())
1045 			continue;
1046 		MNT_IUNLOCK(mp);
1047 yield:
1048 		kern_yield(PRI_USER);
1049 relock_mnt:
1050 		MNT_ILOCK(mp);
1051 	}
1052 	MNT_IUNLOCK(mp);
1053 	vn_finished_write(mp);
1054 	return done;
1055 }
1056 
1057 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1058 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1059     0,
1060     "limit on vnode free requests per call to the vnlru_free routine");
1061 
1062 /*
1063  * Attempt to reduce the free list by the requested amount.
1064  */
1065 static void
1066 vnlru_free_locked(int count, struct vfsops *mnt_op)
1067 {
1068 	struct vnode *vp;
1069 	struct mount *mp;
1070 	bool tried_batches;
1071 
1072 	tried_batches = false;
1073 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1074 	if (count > max_vnlru_free)
1075 		count = max_vnlru_free;
1076 	for (; count > 0; count--) {
1077 		vp = TAILQ_FIRST(&vnode_free_list);
1078 		/*
1079 		 * The list can be modified while the free_list_mtx
1080 		 * has been dropped and vp could be NULL here.
1081 		 */
1082 		if (vp == NULL) {
1083 			if (tried_batches)
1084 				break;
1085 			mtx_unlock(&vnode_free_list_mtx);
1086 			vnlru_return_batches(mnt_op);
1087 			tried_batches = true;
1088 			mtx_lock(&vnode_free_list_mtx);
1089 			continue;
1090 		}
1091 
1092 		VNASSERT(vp->v_op != NULL, vp,
1093 		    ("vnlru_free: vnode already reclaimed."));
1094 		KASSERT((vp->v_iflag & VI_FREE) != 0,
1095 		    ("Removing vnode not on freelist"));
1096 		KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1097 		    ("Mangling active vnode"));
1098 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
1099 
1100 		/*
1101 		 * Don't recycle if our vnode is from different type
1102 		 * of mount point.  Note that mp is type-safe, the
1103 		 * check does not reach unmapped address even if
1104 		 * vnode is reclaimed.
1105 		 * Don't recycle if we can't get the interlock without
1106 		 * blocking.
1107 		 */
1108 		if ((mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1109 		    mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) {
1110 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist);
1111 			continue;
1112 		}
1113 		VNASSERT((vp->v_iflag & VI_FREE) != 0 && vp->v_holdcnt == 0,
1114 		    vp, ("vp inconsistent on freelist"));
1115 
1116 		/*
1117 		 * The clear of VI_FREE prevents activation of the
1118 		 * vnode.  There is no sense in putting the vnode on
1119 		 * the mount point active list, only to remove it
1120 		 * later during recycling.  Inline the relevant part
1121 		 * of vholdl(), to avoid triggering assertions or
1122 		 * activating.
1123 		 */
1124 		freevnodes--;
1125 		vp->v_iflag &= ~VI_FREE;
1126 		VNODE_REFCOUNT_FENCE_REL();
1127 		refcount_acquire(&vp->v_holdcnt);
1128 
1129 		mtx_unlock(&vnode_free_list_mtx);
1130 		VI_UNLOCK(vp);
1131 		vtryrecycle(vp);
1132 		/*
1133 		 * If the recycled succeeded this vdrop will actually free
1134 		 * the vnode.  If not it will simply place it back on
1135 		 * the free list.
1136 		 */
1137 		vdrop(vp);
1138 		mtx_lock(&vnode_free_list_mtx);
1139 	}
1140 }
1141 
1142 void
1143 vnlru_free(int count, struct vfsops *mnt_op)
1144 {
1145 
1146 	mtx_lock(&vnode_free_list_mtx);
1147 	vnlru_free_locked(count, mnt_op);
1148 	mtx_unlock(&vnode_free_list_mtx);
1149 }
1150 
1151 
1152 /* XXX some names and initialization are bad for limits and watermarks. */
1153 static int
1154 vspace(void)
1155 {
1156 	int space;
1157 
1158 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1159 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1160 	vlowat = vhiwat / 2;
1161 	if (numvnodes > desiredvnodes)
1162 		return (0);
1163 	space = desiredvnodes - numvnodes;
1164 	if (freevnodes > wantfreevnodes)
1165 		space += freevnodes - wantfreevnodes;
1166 	return (space);
1167 }
1168 
1169 static void
1170 vnlru_return_batch_locked(struct mount *mp)
1171 {
1172 	struct vnode *vp;
1173 
1174 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
1175 
1176 	if (mp->mnt_tmpfreevnodelistsize == 0)
1177 		return;
1178 
1179 	TAILQ_FOREACH(vp, &mp->mnt_tmpfreevnodelist, v_actfreelist) {
1180 		VNASSERT((vp->v_mflag & VMP_TMPMNTFREELIST) != 0, vp,
1181 		    ("vnode without VMP_TMPMNTFREELIST on mnt_tmpfreevnodelist"));
1182 		vp->v_mflag &= ~VMP_TMPMNTFREELIST;
1183 	}
1184 	mtx_lock(&vnode_free_list_mtx);
1185 	TAILQ_CONCAT(&vnode_free_list, &mp->mnt_tmpfreevnodelist, v_actfreelist);
1186 	freevnodes += mp->mnt_tmpfreevnodelistsize;
1187 	mtx_unlock(&vnode_free_list_mtx);
1188 	mp->mnt_tmpfreevnodelistsize = 0;
1189 }
1190 
1191 static void
1192 vnlru_return_batch(struct mount *mp)
1193 {
1194 
1195 	mtx_lock(&mp->mnt_listmtx);
1196 	vnlru_return_batch_locked(mp);
1197 	mtx_unlock(&mp->mnt_listmtx);
1198 }
1199 
1200 static void
1201 vnlru_return_batches(struct vfsops *mnt_op)
1202 {
1203 	struct mount *mp, *nmp;
1204 	bool need_unbusy;
1205 
1206 	mtx_lock(&mountlist_mtx);
1207 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1208 		need_unbusy = false;
1209 		if (mnt_op != NULL && mp->mnt_op != mnt_op)
1210 			goto next;
1211 		if (mp->mnt_tmpfreevnodelistsize == 0)
1212 			goto next;
1213 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK) == 0) {
1214 			vnlru_return_batch(mp);
1215 			need_unbusy = true;
1216 			mtx_lock(&mountlist_mtx);
1217 		}
1218 next:
1219 		nmp = TAILQ_NEXT(mp, mnt_list);
1220 		if (need_unbusy)
1221 			vfs_unbusy(mp);
1222 	}
1223 	mtx_unlock(&mountlist_mtx);
1224 }
1225 
1226 /*
1227  * Attempt to recycle vnodes in a context that is always safe to block.
1228  * Calling vlrurecycle() from the bowels of filesystem code has some
1229  * interesting deadlock problems.
1230  */
1231 static struct proc *vnlruproc;
1232 static int vnlruproc_sig;
1233 
1234 static void
1235 vnlru_proc(void)
1236 {
1237 	struct mount *mp, *nmp;
1238 	unsigned long onumvnodes;
1239 	int done, force, reclaim_nc_src, trigger, usevnodes;
1240 
1241 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1242 	    SHUTDOWN_PRI_FIRST);
1243 
1244 	force = 0;
1245 	for (;;) {
1246 		kproc_suspend_check(vnlruproc);
1247 		mtx_lock(&vnode_free_list_mtx);
1248 		/*
1249 		 * If numvnodes is too large (due to desiredvnodes being
1250 		 * adjusted using its sysctl, or emergency growth), first
1251 		 * try to reduce it by discarding from the free list.
1252 		 */
1253 		if (numvnodes > desiredvnodes)
1254 			vnlru_free_locked(numvnodes - desiredvnodes, NULL);
1255 		/*
1256 		 * Sleep if the vnode cache is in a good state.  This is
1257 		 * when it is not over-full and has space for about a 4%
1258 		 * or 9% expansion (by growing its size or inexcessively
1259 		 * reducing its free list).  Otherwise, try to reclaim
1260 		 * space for a 10% expansion.
1261 		 */
1262 		if (vstir && force == 0) {
1263 			force = 1;
1264 			vstir = 0;
1265 		}
1266 		if (vspace() >= vlowat && force == 0) {
1267 			vnlruproc_sig = 0;
1268 			wakeup(&vnlruproc_sig);
1269 			msleep(vnlruproc, &vnode_free_list_mtx,
1270 			    PVFS|PDROP, "vlruwt", hz);
1271 			continue;
1272 		}
1273 		mtx_unlock(&vnode_free_list_mtx);
1274 		done = 0;
1275 		onumvnodes = numvnodes;
1276 		/*
1277 		 * Calculate parameters for recycling.  These are the same
1278 		 * throughout the loop to give some semblance of fairness.
1279 		 * The trigger point is to avoid recycling vnodes with lots
1280 		 * of resident pages.  We aren't trying to free memory; we
1281 		 * are trying to recycle or at least free vnodes.
1282 		 */
1283 		if (numvnodes <= desiredvnodes)
1284 			usevnodes = numvnodes - freevnodes;
1285 		else
1286 			usevnodes = numvnodes;
1287 		if (usevnodes <= 0)
1288 			usevnodes = 1;
1289 		/*
1290 		 * The trigger value is is chosen to give a conservatively
1291 		 * large value to ensure that it alone doesn't prevent
1292 		 * making progress.  The value can easily be so large that
1293 		 * it is effectively infinite in some congested and
1294 		 * misconfigured cases, and this is necessary.  Normally
1295 		 * it is about 8 to 100 (pages), which is quite large.
1296 		 */
1297 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1298 		if (force < 2)
1299 			trigger = vsmalltrigger;
1300 		reclaim_nc_src = force >= 3;
1301 		mtx_lock(&mountlist_mtx);
1302 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1303 			if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) {
1304 				nmp = TAILQ_NEXT(mp, mnt_list);
1305 				continue;
1306 			}
1307 			done += vlrureclaim(mp, reclaim_nc_src, trigger);
1308 			mtx_lock(&mountlist_mtx);
1309 			nmp = TAILQ_NEXT(mp, mnt_list);
1310 			vfs_unbusy(mp);
1311 		}
1312 		mtx_unlock(&mountlist_mtx);
1313 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1314 			uma_reclaim();
1315 		if (done == 0) {
1316 			if (force == 0 || force == 1) {
1317 				force = 2;
1318 				continue;
1319 			}
1320 			if (force == 2) {
1321 				force = 3;
1322 				continue;
1323 			}
1324 			force = 0;
1325 			vnlru_nowhere++;
1326 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1327 		} else
1328 			kern_yield(PRI_USER);
1329 		/*
1330 		 * After becoming active to expand above low water, keep
1331 		 * active until above high water.
1332 		 */
1333 		force = vspace() < vhiwat;
1334 	}
1335 }
1336 
1337 static struct kproc_desc vnlru_kp = {
1338 	"vnlru",
1339 	vnlru_proc,
1340 	&vnlruproc
1341 };
1342 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1343     &vnlru_kp);
1344 
1345 /*
1346  * Routines having to do with the management of the vnode table.
1347  */
1348 
1349 /*
1350  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1351  * before we actually vgone().  This function must be called with the vnode
1352  * held to prevent the vnode from being returned to the free list midway
1353  * through vgone().
1354  */
1355 static int
1356 vtryrecycle(struct vnode *vp)
1357 {
1358 	struct mount *vnmp;
1359 
1360 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1361 	VNASSERT(vp->v_holdcnt, vp,
1362 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1363 	/*
1364 	 * This vnode may found and locked via some other list, if so we
1365 	 * can't recycle it yet.
1366 	 */
1367 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1368 		CTR2(KTR_VFS,
1369 		    "%s: impossible to recycle, vp %p lock is already held",
1370 		    __func__, vp);
1371 		return (EWOULDBLOCK);
1372 	}
1373 	/*
1374 	 * Don't recycle if its filesystem is being suspended.
1375 	 */
1376 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1377 		VOP_UNLOCK(vp, 0);
1378 		CTR2(KTR_VFS,
1379 		    "%s: impossible to recycle, cannot start the write for %p",
1380 		    __func__, vp);
1381 		return (EBUSY);
1382 	}
1383 	/*
1384 	 * If we got this far, we need to acquire the interlock and see if
1385 	 * anyone picked up this vnode from another list.  If not, we will
1386 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1387 	 * will skip over it.
1388 	 */
1389 	VI_LOCK(vp);
1390 	if (vp->v_usecount) {
1391 		VOP_UNLOCK(vp, LK_INTERLOCK);
1392 		vn_finished_write(vnmp);
1393 		CTR2(KTR_VFS,
1394 		    "%s: impossible to recycle, %p is already referenced",
1395 		    __func__, vp);
1396 		return (EBUSY);
1397 	}
1398 	if ((vp->v_iflag & VI_DOOMED) == 0) {
1399 		counter_u64_add(recycles_count, 1);
1400 		vgonel(vp);
1401 	}
1402 	VOP_UNLOCK(vp, LK_INTERLOCK);
1403 	vn_finished_write(vnmp);
1404 	return (0);
1405 }
1406 
1407 static void
1408 vcheckspace(void)
1409 {
1410 
1411 	if (vspace() < vlowat && vnlruproc_sig == 0) {
1412 		vnlruproc_sig = 1;
1413 		wakeup(vnlruproc);
1414 	}
1415 }
1416 
1417 /*
1418  * Wait if necessary for space for a new vnode.
1419  */
1420 static int
1421 getnewvnode_wait(int suspended)
1422 {
1423 
1424 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1425 	if (numvnodes >= desiredvnodes) {
1426 		if (suspended) {
1427 			/*
1428 			 * The file system is being suspended.  We cannot
1429 			 * risk a deadlock here, so allow allocation of
1430 			 * another vnode even if this would give too many.
1431 			 */
1432 			return (0);
1433 		}
1434 		if (vnlruproc_sig == 0) {
1435 			vnlruproc_sig = 1;	/* avoid unnecessary wakeups */
1436 			wakeup(vnlruproc);
1437 		}
1438 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
1439 		    "vlruwk", hz);
1440 	}
1441 	/* Post-adjust like the pre-adjust in getnewvnode(). */
1442 	if (numvnodes + 1 > desiredvnodes && freevnodes > 1)
1443 		vnlru_free_locked(1, NULL);
1444 	return (numvnodes >= desiredvnodes ? ENFILE : 0);
1445 }
1446 
1447 /*
1448  * This hack is fragile, and probably not needed any more now that the
1449  * watermark handling works.
1450  */
1451 void
1452 getnewvnode_reserve(u_int count)
1453 {
1454 	struct thread *td;
1455 
1456 	/* Pre-adjust like the pre-adjust in getnewvnode(), with any count. */
1457 	/* XXX no longer so quick, but this part is not racy. */
1458 	mtx_lock(&vnode_free_list_mtx);
1459 	if (numvnodes + count > desiredvnodes && freevnodes > wantfreevnodes)
1460 		vnlru_free_locked(ulmin(numvnodes + count - desiredvnodes,
1461 		    freevnodes - wantfreevnodes), NULL);
1462 	mtx_unlock(&vnode_free_list_mtx);
1463 
1464 	td = curthread;
1465 	/* First try to be quick and racy. */
1466 	if (atomic_fetchadd_long(&numvnodes, count) + count <= desiredvnodes) {
1467 		td->td_vp_reserv += count;
1468 		vcheckspace();	/* XXX no longer so quick, but more racy */
1469 		return;
1470 	} else
1471 		atomic_subtract_long(&numvnodes, count);
1472 
1473 	mtx_lock(&vnode_free_list_mtx);
1474 	while (count > 0) {
1475 		if (getnewvnode_wait(0) == 0) {
1476 			count--;
1477 			td->td_vp_reserv++;
1478 			atomic_add_long(&numvnodes, 1);
1479 		}
1480 	}
1481 	vcheckspace();
1482 	mtx_unlock(&vnode_free_list_mtx);
1483 }
1484 
1485 /*
1486  * This hack is fragile, especially if desiredvnodes or wantvnodes are
1487  * misconfgured or changed significantly.  Reducing desiredvnodes below
1488  * the reserved amount should cause bizarre behaviour like reducing it
1489  * below the number of active vnodes -- the system will try to reduce
1490  * numvnodes to match, but should fail, so the subtraction below should
1491  * not overflow.
1492  */
1493 void
1494 getnewvnode_drop_reserve(void)
1495 {
1496 	struct thread *td;
1497 
1498 	td = curthread;
1499 	atomic_subtract_long(&numvnodes, td->td_vp_reserv);
1500 	td->td_vp_reserv = 0;
1501 }
1502 
1503 /*
1504  * Return the next vnode from the free list.
1505  */
1506 int
1507 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1508     struct vnode **vpp)
1509 {
1510 	struct vnode *vp;
1511 	struct thread *td;
1512 	struct lock_object *lo;
1513 	static int cyclecount;
1514 	int error __unused;
1515 
1516 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1517 	vp = NULL;
1518 	td = curthread;
1519 	if (td->td_vp_reserv > 0) {
1520 		td->td_vp_reserv -= 1;
1521 		goto alloc;
1522 	}
1523 	mtx_lock(&vnode_free_list_mtx);
1524 	if (numvnodes < desiredvnodes)
1525 		cyclecount = 0;
1526 	else if (cyclecount++ >= freevnodes) {
1527 		cyclecount = 0;
1528 		vstir = 1;
1529 	}
1530 	/*
1531 	 * Grow the vnode cache if it will not be above its target max
1532 	 * after growing.  Otherwise, if the free list is nonempty, try
1533 	 * to reclaim 1 item from it before growing the cache (possibly
1534 	 * above its target max if the reclamation failed or is delayed).
1535 	 * Otherwise, wait for some space.  In all cases, schedule
1536 	 * vnlru_proc() if we are getting short of space.  The watermarks
1537 	 * should be chosen so that we never wait or even reclaim from
1538 	 * the free list to below its target minimum.
1539 	 */
1540 	if (numvnodes + 1 <= desiredvnodes)
1541 		;
1542 	else if (freevnodes > 0)
1543 		vnlru_free_locked(1, NULL);
1544 	else {
1545 		error = getnewvnode_wait(mp != NULL && (mp->mnt_kern_flag &
1546 		    MNTK_SUSPEND));
1547 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
1548 		if (error != 0) {
1549 			mtx_unlock(&vnode_free_list_mtx);
1550 			return (error);
1551 		}
1552 #endif
1553 	}
1554 	vcheckspace();
1555 	atomic_add_long(&numvnodes, 1);
1556 	mtx_unlock(&vnode_free_list_mtx);
1557 alloc:
1558 	counter_u64_add(vnodes_created, 1);
1559 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK);
1560 	/*
1561 	 * Locks are given the generic name "vnode" when created.
1562 	 * Follow the historic practice of using the filesystem
1563 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1564 	 *
1565 	 * Locks live in a witness group keyed on their name. Thus,
1566 	 * when a lock is renamed, it must also move from the witness
1567 	 * group of its old name to the witness group of its new name.
1568 	 *
1569 	 * The change only needs to be made when the vnode moves
1570 	 * from one filesystem type to another. We ensure that each
1571 	 * filesystem use a single static name pointer for its tag so
1572 	 * that we can compare pointers rather than doing a strcmp().
1573 	 */
1574 	lo = &vp->v_vnlock->lock_object;
1575 	if (lo->lo_name != tag) {
1576 		lo->lo_name = tag;
1577 		WITNESS_DESTROY(lo);
1578 		WITNESS_INIT(lo, tag);
1579 	}
1580 	/*
1581 	 * By default, don't allow shared locks unless filesystems opt-in.
1582 	 */
1583 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1584 	/*
1585 	 * Finalize various vnode identity bits.
1586 	 */
1587 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1588 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1589 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1590 	vp->v_type = VNON;
1591 	vp->v_tag = tag;
1592 	vp->v_op = vops;
1593 	v_init_counters(vp);
1594 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1595 #ifdef DIAGNOSTIC
1596 	if (mp == NULL && vops != &dead_vnodeops)
1597 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1598 #endif
1599 #ifdef MAC
1600 	mac_vnode_init(vp);
1601 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1602 		mac_vnode_associate_singlelabel(mp, vp);
1603 #endif
1604 	if (mp != NULL) {
1605 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1606 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1607 			vp->v_vflag |= VV_NOKNOTE;
1608 	}
1609 
1610 	/*
1611 	 * For the filesystems which do not use vfs_hash_insert(),
1612 	 * still initialize v_hash to have vfs_hash_index() useful.
1613 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1614 	 * its own hashing.
1615 	 */
1616 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1617 
1618 	*vpp = vp;
1619 	return (0);
1620 }
1621 
1622 /*
1623  * Delete from old mount point vnode list, if on one.
1624  */
1625 static void
1626 delmntque(struct vnode *vp)
1627 {
1628 	struct mount *mp;
1629 	int active;
1630 
1631 	mp = vp->v_mount;
1632 	if (mp == NULL)
1633 		return;
1634 	MNT_ILOCK(mp);
1635 	VI_LOCK(vp);
1636 	KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize,
1637 	    ("Active vnode list size %d > Vnode list size %d",
1638 	     mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize));
1639 	active = vp->v_iflag & VI_ACTIVE;
1640 	vp->v_iflag &= ~VI_ACTIVE;
1641 	if (active) {
1642 		mtx_lock(&mp->mnt_listmtx);
1643 		TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist);
1644 		mp->mnt_activevnodelistsize--;
1645 		mtx_unlock(&mp->mnt_listmtx);
1646 	}
1647 	vp->v_mount = NULL;
1648 	VI_UNLOCK(vp);
1649 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1650 		("bad mount point vnode list size"));
1651 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1652 	mp->mnt_nvnodelistsize--;
1653 	MNT_REL(mp);
1654 	MNT_IUNLOCK(mp);
1655 }
1656 
1657 static void
1658 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1659 {
1660 
1661 	vp->v_data = NULL;
1662 	vp->v_op = &dead_vnodeops;
1663 	vgone(vp);
1664 	vput(vp);
1665 }
1666 
1667 /*
1668  * Insert into list of vnodes for the new mount point, if available.
1669  */
1670 int
1671 insmntque1(struct vnode *vp, struct mount *mp,
1672 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1673 {
1674 
1675 	KASSERT(vp->v_mount == NULL,
1676 		("insmntque: vnode already on per mount vnode list"));
1677 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1678 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1679 
1680 	/*
1681 	 * We acquire the vnode interlock early to ensure that the
1682 	 * vnode cannot be recycled by another process releasing a
1683 	 * holdcnt on it before we get it on both the vnode list
1684 	 * and the active vnode list. The mount mutex protects only
1685 	 * manipulation of the vnode list and the vnode freelist
1686 	 * mutex protects only manipulation of the active vnode list.
1687 	 * Hence the need to hold the vnode interlock throughout.
1688 	 */
1689 	MNT_ILOCK(mp);
1690 	VI_LOCK(vp);
1691 	if (((mp->mnt_kern_flag & MNTK_NOINSMNTQ) != 0 &&
1692 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1693 	    mp->mnt_nvnodelistsize == 0)) &&
1694 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1695 		VI_UNLOCK(vp);
1696 		MNT_IUNLOCK(mp);
1697 		if (dtr != NULL)
1698 			dtr(vp, dtr_arg);
1699 		return (EBUSY);
1700 	}
1701 	vp->v_mount = mp;
1702 	MNT_REF(mp);
1703 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1704 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1705 		("neg mount point vnode list size"));
1706 	mp->mnt_nvnodelistsize++;
1707 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1708 	    ("Activating already active vnode"));
1709 	vp->v_iflag |= VI_ACTIVE;
1710 	mtx_lock(&mp->mnt_listmtx);
1711 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
1712 	mp->mnt_activevnodelistsize++;
1713 	mtx_unlock(&mp->mnt_listmtx);
1714 	VI_UNLOCK(vp);
1715 	MNT_IUNLOCK(mp);
1716 	return (0);
1717 }
1718 
1719 int
1720 insmntque(struct vnode *vp, struct mount *mp)
1721 {
1722 
1723 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1724 }
1725 
1726 /*
1727  * Flush out and invalidate all buffers associated with a bufobj
1728  * Called with the underlying object locked.
1729  */
1730 int
1731 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1732 {
1733 	int error;
1734 
1735 	BO_LOCK(bo);
1736 	if (flags & V_SAVE) {
1737 		error = bufobj_wwait(bo, slpflag, slptimeo);
1738 		if (error) {
1739 			BO_UNLOCK(bo);
1740 			return (error);
1741 		}
1742 		if (bo->bo_dirty.bv_cnt > 0) {
1743 			BO_UNLOCK(bo);
1744 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1745 				return (error);
1746 			/*
1747 			 * XXX We could save a lock/unlock if this was only
1748 			 * enabled under INVARIANTS
1749 			 */
1750 			BO_LOCK(bo);
1751 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1752 				panic("vinvalbuf: dirty bufs");
1753 		}
1754 	}
1755 	/*
1756 	 * If you alter this loop please notice that interlock is dropped and
1757 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1758 	 * no race conditions occur from this.
1759 	 */
1760 	do {
1761 		error = flushbuflist(&bo->bo_clean,
1762 		    flags, bo, slpflag, slptimeo);
1763 		if (error == 0 && !(flags & V_CLEANONLY))
1764 			error = flushbuflist(&bo->bo_dirty,
1765 			    flags, bo, slpflag, slptimeo);
1766 		if (error != 0 && error != EAGAIN) {
1767 			BO_UNLOCK(bo);
1768 			return (error);
1769 		}
1770 	} while (error != 0);
1771 
1772 	/*
1773 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1774 	 * have write I/O in-progress but if there is a VM object then the
1775 	 * VM object can also have read-I/O in-progress.
1776 	 */
1777 	do {
1778 		bufobj_wwait(bo, 0, 0);
1779 		if ((flags & V_VMIO) == 0) {
1780 			BO_UNLOCK(bo);
1781 			if (bo->bo_object != NULL) {
1782 				VM_OBJECT_WLOCK(bo->bo_object);
1783 				vm_object_pip_wait(bo->bo_object, "bovlbx");
1784 				VM_OBJECT_WUNLOCK(bo->bo_object);
1785 			}
1786 			BO_LOCK(bo);
1787 		}
1788 	} while (bo->bo_numoutput > 0);
1789 	BO_UNLOCK(bo);
1790 
1791 	/*
1792 	 * Destroy the copy in the VM cache, too.
1793 	 */
1794 	if (bo->bo_object != NULL &&
1795 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1796 		VM_OBJECT_WLOCK(bo->bo_object);
1797 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1798 		    OBJPR_CLEANONLY : 0);
1799 		VM_OBJECT_WUNLOCK(bo->bo_object);
1800 	}
1801 
1802 #ifdef INVARIANTS
1803 	BO_LOCK(bo);
1804 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1805 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1806 	    bo->bo_clean.bv_cnt > 0))
1807 		panic("vinvalbuf: flush failed");
1808 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1809 	    bo->bo_dirty.bv_cnt > 0)
1810 		panic("vinvalbuf: flush dirty failed");
1811 	BO_UNLOCK(bo);
1812 #endif
1813 	return (0);
1814 }
1815 
1816 /*
1817  * Flush out and invalidate all buffers associated with a vnode.
1818  * Called with the underlying object locked.
1819  */
1820 int
1821 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1822 {
1823 
1824 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1825 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1826 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1827 		return (0);
1828 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1829 }
1830 
1831 /*
1832  * Flush out buffers on the specified list.
1833  *
1834  */
1835 static int
1836 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1837     int slptimeo)
1838 {
1839 	struct buf *bp, *nbp;
1840 	int retval, error;
1841 	daddr_t lblkno;
1842 	b_xflags_t xflags;
1843 
1844 	ASSERT_BO_WLOCKED(bo);
1845 
1846 	retval = 0;
1847 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1848 		/*
1849 		 * If we are flushing both V_NORMAL and V_ALT buffers then
1850 		 * do not skip any buffers. If we are flushing only V_NORMAL
1851 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
1852 		 * flushing only V_ALT buffers then skip buffers not marked
1853 		 * as BX_ALTDATA.
1854 		 */
1855 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
1856 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
1857 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
1858 			continue;
1859 		}
1860 		if (nbp != NULL) {
1861 			lblkno = nbp->b_lblkno;
1862 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
1863 		}
1864 		retval = EAGAIN;
1865 		error = BUF_TIMELOCK(bp,
1866 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
1867 		    "flushbuf", slpflag, slptimeo);
1868 		if (error) {
1869 			BO_LOCK(bo);
1870 			return (error != ENOLCK ? error : EAGAIN);
1871 		}
1872 		KASSERT(bp->b_bufobj == bo,
1873 		    ("bp %p wrong b_bufobj %p should be %p",
1874 		    bp, bp->b_bufobj, bo));
1875 		/*
1876 		 * XXX Since there are no node locks for NFS, I
1877 		 * believe there is a slight chance that a delayed
1878 		 * write will occur while sleeping just above, so
1879 		 * check for it.
1880 		 */
1881 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1882 		    (flags & V_SAVE)) {
1883 			bremfree(bp);
1884 			bp->b_flags |= B_ASYNC;
1885 			bwrite(bp);
1886 			BO_LOCK(bo);
1887 			return (EAGAIN);	/* XXX: why not loop ? */
1888 		}
1889 		bremfree(bp);
1890 		bp->b_flags |= (B_INVAL | B_RELBUF);
1891 		bp->b_flags &= ~B_ASYNC;
1892 		brelse(bp);
1893 		BO_LOCK(bo);
1894 		if (nbp == NULL)
1895 			break;
1896 		nbp = gbincore(bo, lblkno);
1897 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1898 		    != xflags)
1899 			break;			/* nbp invalid */
1900 	}
1901 	return (retval);
1902 }
1903 
1904 int
1905 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
1906 {
1907 	struct buf *bp;
1908 	int error;
1909 	daddr_t lblkno;
1910 
1911 	ASSERT_BO_LOCKED(bo);
1912 
1913 	for (lblkno = startn;;) {
1914 again:
1915 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
1916 		if (bp == NULL || bp->b_lblkno >= endn ||
1917 		    bp->b_lblkno < startn)
1918 			break;
1919 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
1920 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
1921 		if (error != 0) {
1922 			BO_RLOCK(bo);
1923 			if (error == ENOLCK)
1924 				goto again;
1925 			return (error);
1926 		}
1927 		KASSERT(bp->b_bufobj == bo,
1928 		    ("bp %p wrong b_bufobj %p should be %p",
1929 		    bp, bp->b_bufobj, bo));
1930 		lblkno = bp->b_lblkno + 1;
1931 		if ((bp->b_flags & B_MANAGED) == 0)
1932 			bremfree(bp);
1933 		bp->b_flags |= B_RELBUF;
1934 		/*
1935 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
1936 		 * pages backing each buffer in the range are unlikely to be
1937 		 * reused.  Dirty buffers will have the hint applied once
1938 		 * they've been written.
1939 		 */
1940 		if ((bp->b_flags & B_VMIO) != 0)
1941 			bp->b_flags |= B_NOREUSE;
1942 		brelse(bp);
1943 		BO_RLOCK(bo);
1944 	}
1945 	return (0);
1946 }
1947 
1948 /*
1949  * Truncate a file's buffer and pages to a specified length.  This
1950  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1951  * sync activity.
1952  */
1953 int
1954 vtruncbuf(struct vnode *vp, off_t length, int blksize)
1955 {
1956 	struct buf *bp, *nbp;
1957 	int anyfreed;
1958 	daddr_t trunclbn;
1959 	struct bufobj *bo;
1960 
1961 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
1962 	    vp, blksize, (uintmax_t)length);
1963 
1964 	/*
1965 	 * Round up to the *next* lbn.
1966 	 */
1967 	trunclbn = howmany(length, blksize);
1968 
1969 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1970 restart:
1971 	bo = &vp->v_bufobj;
1972 	BO_LOCK(bo);
1973 	anyfreed = 1;
1974 	for (;anyfreed;) {
1975 		anyfreed = 0;
1976 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1977 			if (bp->b_lblkno < trunclbn)
1978 				continue;
1979 			if (BUF_LOCK(bp,
1980 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1981 			    BO_LOCKPTR(bo)) == ENOLCK)
1982 				goto restart;
1983 
1984 			bremfree(bp);
1985 			bp->b_flags |= (B_INVAL | B_RELBUF);
1986 			bp->b_flags &= ~B_ASYNC;
1987 			brelse(bp);
1988 			anyfreed = 1;
1989 
1990 			BO_LOCK(bo);
1991 			if (nbp != NULL &&
1992 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1993 			    (nbp->b_vp != vp) ||
1994 			    (nbp->b_flags & B_DELWRI))) {
1995 				BO_UNLOCK(bo);
1996 				goto restart;
1997 			}
1998 		}
1999 
2000 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2001 			if (bp->b_lblkno < trunclbn)
2002 				continue;
2003 			if (BUF_LOCK(bp,
2004 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2005 			    BO_LOCKPTR(bo)) == ENOLCK)
2006 				goto restart;
2007 			bremfree(bp);
2008 			bp->b_flags |= (B_INVAL | B_RELBUF);
2009 			bp->b_flags &= ~B_ASYNC;
2010 			brelse(bp);
2011 			anyfreed = 1;
2012 
2013 			BO_LOCK(bo);
2014 			if (nbp != NULL &&
2015 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2016 			    (nbp->b_vp != vp) ||
2017 			    (nbp->b_flags & B_DELWRI) == 0)) {
2018 				BO_UNLOCK(bo);
2019 				goto restart;
2020 			}
2021 		}
2022 	}
2023 
2024 	if (length > 0) {
2025 restartsync:
2026 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2027 			if (bp->b_lblkno > 0)
2028 				continue;
2029 			/*
2030 			 * Since we hold the vnode lock this should only
2031 			 * fail if we're racing with the buf daemon.
2032 			 */
2033 			if (BUF_LOCK(bp,
2034 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2035 			    BO_LOCKPTR(bo)) == ENOLCK) {
2036 				goto restart;
2037 			}
2038 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2039 			    ("buf(%p) on dirty queue without DELWRI", bp));
2040 
2041 			bremfree(bp);
2042 			bawrite(bp);
2043 			BO_LOCK(bo);
2044 			goto restartsync;
2045 		}
2046 	}
2047 
2048 	bufobj_wwait(bo, 0, 0);
2049 	BO_UNLOCK(bo);
2050 	vnode_pager_setsize(vp, length);
2051 
2052 	return (0);
2053 }
2054 
2055 static void
2056 buf_vlist_remove(struct buf *bp)
2057 {
2058 	struct bufv *bv;
2059 
2060 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2061 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2062 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
2063 	    (BX_VNDIRTY|BX_VNCLEAN),
2064 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
2065 	if (bp->b_xflags & BX_VNDIRTY)
2066 		bv = &bp->b_bufobj->bo_dirty;
2067 	else
2068 		bv = &bp->b_bufobj->bo_clean;
2069 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2070 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2071 	bv->bv_cnt--;
2072 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2073 }
2074 
2075 /*
2076  * Add the buffer to the sorted clean or dirty block list.
2077  *
2078  * NOTE: xflags is passed as a constant, optimizing this inline function!
2079  */
2080 static void
2081 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2082 {
2083 	struct bufv *bv;
2084 	struct buf *n;
2085 	int error;
2086 
2087 	ASSERT_BO_WLOCKED(bo);
2088 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2089 	    ("dead bo %p", bo));
2090 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2091 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2092 	bp->b_xflags |= xflags;
2093 	if (xflags & BX_VNDIRTY)
2094 		bv = &bo->bo_dirty;
2095 	else
2096 		bv = &bo->bo_clean;
2097 
2098 	/*
2099 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
2100 	 * we tend to grow at the tail so lookup_le should usually be cheaper
2101 	 * than _ge.
2102 	 */
2103 	if (bv->bv_cnt == 0 ||
2104 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2105 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2106 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2107 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2108 	else
2109 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2110 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2111 	if (error)
2112 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2113 	bv->bv_cnt++;
2114 }
2115 
2116 /*
2117  * Look up a buffer using the buffer tries.
2118  */
2119 struct buf *
2120 gbincore(struct bufobj *bo, daddr_t lblkno)
2121 {
2122 	struct buf *bp;
2123 
2124 	ASSERT_BO_LOCKED(bo);
2125 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2126 	if (bp != NULL)
2127 		return (bp);
2128 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
2129 }
2130 
2131 /*
2132  * Associate a buffer with a vnode.
2133  */
2134 void
2135 bgetvp(struct vnode *vp, struct buf *bp)
2136 {
2137 	struct bufobj *bo;
2138 
2139 	bo = &vp->v_bufobj;
2140 	ASSERT_BO_WLOCKED(bo);
2141 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2142 
2143 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2144 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2145 	    ("bgetvp: bp already attached! %p", bp));
2146 
2147 	vhold(vp);
2148 	bp->b_vp = vp;
2149 	bp->b_bufobj = bo;
2150 	/*
2151 	 * Insert onto list for new vnode.
2152 	 */
2153 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2154 }
2155 
2156 /*
2157  * Disassociate a buffer from a vnode.
2158  */
2159 void
2160 brelvp(struct buf *bp)
2161 {
2162 	struct bufobj *bo;
2163 	struct vnode *vp;
2164 
2165 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2166 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2167 
2168 	/*
2169 	 * Delete from old vnode list, if on one.
2170 	 */
2171 	vp = bp->b_vp;		/* XXX */
2172 	bo = bp->b_bufobj;
2173 	BO_LOCK(bo);
2174 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2175 		buf_vlist_remove(bp);
2176 	else
2177 		panic("brelvp: Buffer %p not on queue.", bp);
2178 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2179 		bo->bo_flag &= ~BO_ONWORKLST;
2180 		mtx_lock(&sync_mtx);
2181 		LIST_REMOVE(bo, bo_synclist);
2182 		syncer_worklist_len--;
2183 		mtx_unlock(&sync_mtx);
2184 	}
2185 	bp->b_vp = NULL;
2186 	bp->b_bufobj = NULL;
2187 	BO_UNLOCK(bo);
2188 	vdrop(vp);
2189 }
2190 
2191 /*
2192  * Add an item to the syncer work queue.
2193  */
2194 static void
2195 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2196 {
2197 	int slot;
2198 
2199 	ASSERT_BO_WLOCKED(bo);
2200 
2201 	mtx_lock(&sync_mtx);
2202 	if (bo->bo_flag & BO_ONWORKLST)
2203 		LIST_REMOVE(bo, bo_synclist);
2204 	else {
2205 		bo->bo_flag |= BO_ONWORKLST;
2206 		syncer_worklist_len++;
2207 	}
2208 
2209 	if (delay > syncer_maxdelay - 2)
2210 		delay = syncer_maxdelay - 2;
2211 	slot = (syncer_delayno + delay) & syncer_mask;
2212 
2213 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2214 	mtx_unlock(&sync_mtx);
2215 }
2216 
2217 static int
2218 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2219 {
2220 	int error, len;
2221 
2222 	mtx_lock(&sync_mtx);
2223 	len = syncer_worklist_len - sync_vnode_count;
2224 	mtx_unlock(&sync_mtx);
2225 	error = SYSCTL_OUT(req, &len, sizeof(len));
2226 	return (error);
2227 }
2228 
2229 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
2230     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2231 
2232 static struct proc *updateproc;
2233 static void sched_sync(void);
2234 static struct kproc_desc up_kp = {
2235 	"syncer",
2236 	sched_sync,
2237 	&updateproc
2238 };
2239 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2240 
2241 static int
2242 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2243 {
2244 	struct vnode *vp;
2245 	struct mount *mp;
2246 
2247 	*bo = LIST_FIRST(slp);
2248 	if (*bo == NULL)
2249 		return (0);
2250 	vp = bo2vnode(*bo);
2251 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2252 		return (1);
2253 	/*
2254 	 * We use vhold in case the vnode does not
2255 	 * successfully sync.  vhold prevents the vnode from
2256 	 * going away when we unlock the sync_mtx so that
2257 	 * we can acquire the vnode interlock.
2258 	 */
2259 	vholdl(vp);
2260 	mtx_unlock(&sync_mtx);
2261 	VI_UNLOCK(vp);
2262 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2263 		vdrop(vp);
2264 		mtx_lock(&sync_mtx);
2265 		return (*bo == LIST_FIRST(slp));
2266 	}
2267 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2268 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2269 	VOP_UNLOCK(vp, 0);
2270 	vn_finished_write(mp);
2271 	BO_LOCK(*bo);
2272 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2273 		/*
2274 		 * Put us back on the worklist.  The worklist
2275 		 * routine will remove us from our current
2276 		 * position and then add us back in at a later
2277 		 * position.
2278 		 */
2279 		vn_syncer_add_to_worklist(*bo, syncdelay);
2280 	}
2281 	BO_UNLOCK(*bo);
2282 	vdrop(vp);
2283 	mtx_lock(&sync_mtx);
2284 	return (0);
2285 }
2286 
2287 static int first_printf = 1;
2288 
2289 /*
2290  * System filesystem synchronizer daemon.
2291  */
2292 static void
2293 sched_sync(void)
2294 {
2295 	struct synclist *next, *slp;
2296 	struct bufobj *bo;
2297 	long starttime;
2298 	struct thread *td = curthread;
2299 	int last_work_seen;
2300 	int net_worklist_len;
2301 	int syncer_final_iter;
2302 	int error;
2303 
2304 	last_work_seen = 0;
2305 	syncer_final_iter = 0;
2306 	syncer_state = SYNCER_RUNNING;
2307 	starttime = time_uptime;
2308 	td->td_pflags |= TDP_NORUNNINGBUF;
2309 
2310 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2311 	    SHUTDOWN_PRI_LAST);
2312 
2313 	mtx_lock(&sync_mtx);
2314 	for (;;) {
2315 		if (syncer_state == SYNCER_FINAL_DELAY &&
2316 		    syncer_final_iter == 0) {
2317 			mtx_unlock(&sync_mtx);
2318 			kproc_suspend_check(td->td_proc);
2319 			mtx_lock(&sync_mtx);
2320 		}
2321 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2322 		if (syncer_state != SYNCER_RUNNING &&
2323 		    starttime != time_uptime) {
2324 			if (first_printf) {
2325 				printf("\nSyncing disks, vnodes remaining... ");
2326 				first_printf = 0;
2327 			}
2328 			printf("%d ", net_worklist_len);
2329 		}
2330 		starttime = time_uptime;
2331 
2332 		/*
2333 		 * Push files whose dirty time has expired.  Be careful
2334 		 * of interrupt race on slp queue.
2335 		 *
2336 		 * Skip over empty worklist slots when shutting down.
2337 		 */
2338 		do {
2339 			slp = &syncer_workitem_pending[syncer_delayno];
2340 			syncer_delayno += 1;
2341 			if (syncer_delayno == syncer_maxdelay)
2342 				syncer_delayno = 0;
2343 			next = &syncer_workitem_pending[syncer_delayno];
2344 			/*
2345 			 * If the worklist has wrapped since the
2346 			 * it was emptied of all but syncer vnodes,
2347 			 * switch to the FINAL_DELAY state and run
2348 			 * for one more second.
2349 			 */
2350 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2351 			    net_worklist_len == 0 &&
2352 			    last_work_seen == syncer_delayno) {
2353 				syncer_state = SYNCER_FINAL_DELAY;
2354 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2355 			}
2356 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2357 		    syncer_worklist_len > 0);
2358 
2359 		/*
2360 		 * Keep track of the last time there was anything
2361 		 * on the worklist other than syncer vnodes.
2362 		 * Return to the SHUTTING_DOWN state if any
2363 		 * new work appears.
2364 		 */
2365 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2366 			last_work_seen = syncer_delayno;
2367 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2368 			syncer_state = SYNCER_SHUTTING_DOWN;
2369 		while (!LIST_EMPTY(slp)) {
2370 			error = sync_vnode(slp, &bo, td);
2371 			if (error == 1) {
2372 				LIST_REMOVE(bo, bo_synclist);
2373 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2374 				continue;
2375 			}
2376 
2377 			if (first_printf == 0) {
2378 				/*
2379 				 * Drop the sync mutex, because some watchdog
2380 				 * drivers need to sleep while patting
2381 				 */
2382 				mtx_unlock(&sync_mtx);
2383 				wdog_kern_pat(WD_LASTVAL);
2384 				mtx_lock(&sync_mtx);
2385 			}
2386 
2387 		}
2388 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2389 			syncer_final_iter--;
2390 		/*
2391 		 * The variable rushjob allows the kernel to speed up the
2392 		 * processing of the filesystem syncer process. A rushjob
2393 		 * value of N tells the filesystem syncer to process the next
2394 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2395 		 * is used by the soft update code to speed up the filesystem
2396 		 * syncer process when the incore state is getting so far
2397 		 * ahead of the disk that the kernel memory pool is being
2398 		 * threatened with exhaustion.
2399 		 */
2400 		if (rushjob > 0) {
2401 			rushjob -= 1;
2402 			continue;
2403 		}
2404 		/*
2405 		 * Just sleep for a short period of time between
2406 		 * iterations when shutting down to allow some I/O
2407 		 * to happen.
2408 		 *
2409 		 * If it has taken us less than a second to process the
2410 		 * current work, then wait. Otherwise start right over
2411 		 * again. We can still lose time if any single round
2412 		 * takes more than two seconds, but it does not really
2413 		 * matter as we are just trying to generally pace the
2414 		 * filesystem activity.
2415 		 */
2416 		if (syncer_state != SYNCER_RUNNING ||
2417 		    time_uptime == starttime) {
2418 			thread_lock(td);
2419 			sched_prio(td, PPAUSE);
2420 			thread_unlock(td);
2421 		}
2422 		if (syncer_state != SYNCER_RUNNING)
2423 			cv_timedwait(&sync_wakeup, &sync_mtx,
2424 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2425 		else if (time_uptime == starttime)
2426 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2427 	}
2428 }
2429 
2430 /*
2431  * Request the syncer daemon to speed up its work.
2432  * We never push it to speed up more than half of its
2433  * normal turn time, otherwise it could take over the cpu.
2434  */
2435 int
2436 speedup_syncer(void)
2437 {
2438 	int ret = 0;
2439 
2440 	mtx_lock(&sync_mtx);
2441 	if (rushjob < syncdelay / 2) {
2442 		rushjob += 1;
2443 		stat_rush_requests += 1;
2444 		ret = 1;
2445 	}
2446 	mtx_unlock(&sync_mtx);
2447 	cv_broadcast(&sync_wakeup);
2448 	return (ret);
2449 }
2450 
2451 /*
2452  * Tell the syncer to speed up its work and run though its work
2453  * list several times, then tell it to shut down.
2454  */
2455 static void
2456 syncer_shutdown(void *arg, int howto)
2457 {
2458 
2459 	if (howto & RB_NOSYNC)
2460 		return;
2461 	mtx_lock(&sync_mtx);
2462 	syncer_state = SYNCER_SHUTTING_DOWN;
2463 	rushjob = 0;
2464 	mtx_unlock(&sync_mtx);
2465 	cv_broadcast(&sync_wakeup);
2466 	kproc_shutdown(arg, howto);
2467 }
2468 
2469 void
2470 syncer_suspend(void)
2471 {
2472 
2473 	syncer_shutdown(updateproc, 0);
2474 }
2475 
2476 void
2477 syncer_resume(void)
2478 {
2479 
2480 	mtx_lock(&sync_mtx);
2481 	first_printf = 1;
2482 	syncer_state = SYNCER_RUNNING;
2483 	mtx_unlock(&sync_mtx);
2484 	cv_broadcast(&sync_wakeup);
2485 	kproc_resume(updateproc);
2486 }
2487 
2488 /*
2489  * Reassign a buffer from one vnode to another.
2490  * Used to assign file specific control information
2491  * (indirect blocks) to the vnode to which they belong.
2492  */
2493 void
2494 reassignbuf(struct buf *bp)
2495 {
2496 	struct vnode *vp;
2497 	struct bufobj *bo;
2498 	int delay;
2499 #ifdef INVARIANTS
2500 	struct bufv *bv;
2501 #endif
2502 
2503 	vp = bp->b_vp;
2504 	bo = bp->b_bufobj;
2505 	++reassignbufcalls;
2506 
2507 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2508 	    bp, bp->b_vp, bp->b_flags);
2509 	/*
2510 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2511 	 * is not fully linked in.
2512 	 */
2513 	if (bp->b_flags & B_PAGING)
2514 		panic("cannot reassign paging buffer");
2515 
2516 	/*
2517 	 * Delete from old vnode list, if on one.
2518 	 */
2519 	BO_LOCK(bo);
2520 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2521 		buf_vlist_remove(bp);
2522 	else
2523 		panic("reassignbuf: Buffer %p not on queue.", bp);
2524 	/*
2525 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2526 	 * of clean buffers.
2527 	 */
2528 	if (bp->b_flags & B_DELWRI) {
2529 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2530 			switch (vp->v_type) {
2531 			case VDIR:
2532 				delay = dirdelay;
2533 				break;
2534 			case VCHR:
2535 				delay = metadelay;
2536 				break;
2537 			default:
2538 				delay = filedelay;
2539 			}
2540 			vn_syncer_add_to_worklist(bo, delay);
2541 		}
2542 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2543 	} else {
2544 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2545 
2546 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2547 			mtx_lock(&sync_mtx);
2548 			LIST_REMOVE(bo, bo_synclist);
2549 			syncer_worklist_len--;
2550 			mtx_unlock(&sync_mtx);
2551 			bo->bo_flag &= ~BO_ONWORKLST;
2552 		}
2553 	}
2554 #ifdef INVARIANTS
2555 	bv = &bo->bo_clean;
2556 	bp = TAILQ_FIRST(&bv->bv_hd);
2557 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2558 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2559 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2560 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2561 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2562 	bv = &bo->bo_dirty;
2563 	bp = TAILQ_FIRST(&bv->bv_hd);
2564 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2565 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2566 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2567 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2568 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2569 #endif
2570 	BO_UNLOCK(bo);
2571 }
2572 
2573 static void
2574 v_init_counters(struct vnode *vp)
2575 {
2576 
2577 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2578 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2579 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2580 
2581 	refcount_init(&vp->v_holdcnt, 1);
2582 	refcount_init(&vp->v_usecount, 1);
2583 }
2584 
2585 static void
2586 v_incr_usecount_locked(struct vnode *vp)
2587 {
2588 
2589 	ASSERT_VI_LOCKED(vp, __func__);
2590 	if ((vp->v_iflag & VI_OWEINACT) != 0) {
2591 		VNASSERT(vp->v_usecount == 0, vp,
2592 		    ("vnode with usecount and VI_OWEINACT set"));
2593 		vp->v_iflag &= ~VI_OWEINACT;
2594 	}
2595 	refcount_acquire(&vp->v_usecount);
2596 	v_incr_devcount(vp);
2597 }
2598 
2599 /*
2600  * Increment the use count on the vnode, taking care to reference
2601  * the driver's usecount if this is a chardev.
2602  */
2603 static void
2604 v_incr_usecount(struct vnode *vp)
2605 {
2606 
2607 	ASSERT_VI_UNLOCKED(vp, __func__);
2608 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2609 
2610 	if (vp->v_type != VCHR &&
2611 	    refcount_acquire_if_not_zero(&vp->v_usecount)) {
2612 		VNODE_REFCOUNT_FENCE_ACQ();
2613 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2614 		    ("vnode with usecount and VI_OWEINACT set"));
2615 	} else {
2616 		VI_LOCK(vp);
2617 		v_incr_usecount_locked(vp);
2618 		VI_UNLOCK(vp);
2619 	}
2620 }
2621 
2622 /*
2623  * Increment si_usecount of the associated device, if any.
2624  */
2625 static void
2626 v_incr_devcount(struct vnode *vp)
2627 {
2628 
2629 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2630 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2631 		dev_lock();
2632 		vp->v_rdev->si_usecount++;
2633 		dev_unlock();
2634 	}
2635 }
2636 
2637 /*
2638  * Decrement si_usecount of the associated device, if any.
2639  */
2640 static void
2641 v_decr_devcount(struct vnode *vp)
2642 {
2643 
2644 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2645 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2646 		dev_lock();
2647 		vp->v_rdev->si_usecount--;
2648 		dev_unlock();
2649 	}
2650 }
2651 
2652 /*
2653  * Grab a particular vnode from the free list, increment its
2654  * reference count and lock it.  VI_DOOMED is set if the vnode
2655  * is being destroyed.  Only callers who specify LK_RETRY will
2656  * see doomed vnodes.  If inactive processing was delayed in
2657  * vput try to do it here.
2658  *
2659  * Notes on lockless counter manipulation:
2660  * _vhold, vputx and other routines make various decisions based
2661  * on either holdcnt or usecount being 0. As long as either counter
2662  * is not transitioning 0->1 nor 1->0, the manipulation can be done
2663  * with atomic operations. Otherwise the interlock is taken covering
2664  * both the atomic and additional actions.
2665  */
2666 int
2667 vget(struct vnode *vp, int flags, struct thread *td)
2668 {
2669 	int error, oweinact;
2670 
2671 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2672 	    ("vget: invalid lock operation"));
2673 
2674 	if ((flags & LK_INTERLOCK) != 0)
2675 		ASSERT_VI_LOCKED(vp, __func__);
2676 	else
2677 		ASSERT_VI_UNLOCKED(vp, __func__);
2678 	if ((flags & LK_VNHELD) != 0)
2679 		VNASSERT((vp->v_holdcnt > 0), vp,
2680 		    ("vget: LK_VNHELD passed but vnode not held"));
2681 
2682 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2683 
2684 	if ((flags & LK_VNHELD) == 0)
2685 		_vhold(vp, (flags & LK_INTERLOCK) != 0);
2686 
2687 	if ((error = vn_lock(vp, flags)) != 0) {
2688 		vdrop(vp);
2689 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2690 		    vp);
2691 		return (error);
2692 	}
2693 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
2694 		panic("vget: vn_lock failed to return ENOENT\n");
2695 	/*
2696 	 * We don't guarantee that any particular close will
2697 	 * trigger inactive processing so just make a best effort
2698 	 * here at preventing a reference to a removed file.  If
2699 	 * we don't succeed no harm is done.
2700 	 *
2701 	 * Upgrade our holdcnt to a usecount.
2702 	 */
2703 	if (vp->v_type == VCHR ||
2704 	    !refcount_acquire_if_not_zero(&vp->v_usecount)) {
2705 		VI_LOCK(vp);
2706 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2707 			oweinact = 0;
2708 		} else {
2709 			oweinact = 1;
2710 			vp->v_iflag &= ~VI_OWEINACT;
2711 			VNODE_REFCOUNT_FENCE_REL();
2712 		}
2713 		refcount_acquire(&vp->v_usecount);
2714 		v_incr_devcount(vp);
2715 		if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2716 		    (flags & LK_NOWAIT) == 0)
2717 			vinactive(vp, td);
2718 		VI_UNLOCK(vp);
2719 	}
2720 	return (0);
2721 }
2722 
2723 /*
2724  * Increase the reference (use) and hold count of a vnode.
2725  * This will also remove the vnode from the free list if it is presently free.
2726  */
2727 void
2728 vref(struct vnode *vp)
2729 {
2730 
2731 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2732 	_vhold(vp, false);
2733 	v_incr_usecount(vp);
2734 }
2735 
2736 void
2737 vrefl(struct vnode *vp)
2738 {
2739 
2740 	ASSERT_VI_LOCKED(vp, __func__);
2741 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2742 	_vhold(vp, true);
2743 	v_incr_usecount_locked(vp);
2744 }
2745 
2746 void
2747 vrefact(struct vnode *vp)
2748 {
2749 
2750 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2751 	if (__predict_false(vp->v_type == VCHR)) {
2752 		VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp,
2753 		    ("%s: wrong ref counts", __func__));
2754 		vref(vp);
2755 		return;
2756 	}
2757 #ifdef INVARIANTS
2758 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
2759 	VNASSERT(old > 0, vp, ("%s: wrong hold count", __func__));
2760 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
2761 	VNASSERT(old > 0, vp, ("%s: wrong use count", __func__));
2762 #else
2763 	refcount_acquire(&vp->v_holdcnt);
2764 	refcount_acquire(&vp->v_usecount);
2765 #endif
2766 }
2767 
2768 /*
2769  * Return reference count of a vnode.
2770  *
2771  * The results of this call are only guaranteed when some mechanism is used to
2772  * stop other processes from gaining references to the vnode.  This may be the
2773  * case if the caller holds the only reference.  This is also useful when stale
2774  * data is acceptable as race conditions may be accounted for by some other
2775  * means.
2776  */
2777 int
2778 vrefcnt(struct vnode *vp)
2779 {
2780 
2781 	return (vp->v_usecount);
2782 }
2783 
2784 #define	VPUTX_VRELE	1
2785 #define	VPUTX_VPUT	2
2786 #define	VPUTX_VUNREF	3
2787 
2788 /*
2789  * Decrement the use and hold counts for a vnode.
2790  *
2791  * See an explanation near vget() as to why atomic operation is safe.
2792  */
2793 static void
2794 vputx(struct vnode *vp, int func)
2795 {
2796 	int error;
2797 
2798 	KASSERT(vp != NULL, ("vputx: null vp"));
2799 	if (func == VPUTX_VUNREF)
2800 		ASSERT_VOP_LOCKED(vp, "vunref");
2801 	else if (func == VPUTX_VPUT)
2802 		ASSERT_VOP_LOCKED(vp, "vput");
2803 	else
2804 		KASSERT(func == VPUTX_VRELE, ("vputx: wrong func"));
2805 	ASSERT_VI_UNLOCKED(vp, __func__);
2806 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2807 
2808 	if (vp->v_type != VCHR &&
2809 	    refcount_release_if_not_last(&vp->v_usecount)) {
2810 		if (func == VPUTX_VPUT)
2811 			VOP_UNLOCK(vp, 0);
2812 		vdrop(vp);
2813 		return;
2814 	}
2815 
2816 	VI_LOCK(vp);
2817 
2818 	/*
2819 	 * We want to hold the vnode until the inactive finishes to
2820 	 * prevent vgone() races.  We drop the use count here and the
2821 	 * hold count below when we're done.
2822 	 */
2823 	if (!refcount_release(&vp->v_usecount) ||
2824 	    (vp->v_iflag & VI_DOINGINACT)) {
2825 		if (func == VPUTX_VPUT)
2826 			VOP_UNLOCK(vp, 0);
2827 		v_decr_devcount(vp);
2828 		vdropl(vp);
2829 		return;
2830 	}
2831 
2832 	v_decr_devcount(vp);
2833 
2834 	error = 0;
2835 
2836 	if (vp->v_usecount != 0) {
2837 		vn_printf(vp, "vputx: usecount not zero for vnode ");
2838 		panic("vputx: usecount not zero");
2839 	}
2840 
2841 	CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp);
2842 
2843 	/*
2844 	 * We must call VOP_INACTIVE with the node locked. Mark
2845 	 * as VI_DOINGINACT to avoid recursion.
2846 	 */
2847 	vp->v_iflag |= VI_OWEINACT;
2848 	switch (func) {
2849 	case VPUTX_VRELE:
2850 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
2851 		VI_LOCK(vp);
2852 		break;
2853 	case VPUTX_VPUT:
2854 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2855 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
2856 			    LK_NOWAIT);
2857 			VI_LOCK(vp);
2858 		}
2859 		break;
2860 	case VPUTX_VUNREF:
2861 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2862 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
2863 			VI_LOCK(vp);
2864 		}
2865 		break;
2866 	}
2867 	VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
2868 	    ("vnode with usecount and VI_OWEINACT set"));
2869 	if (error == 0) {
2870 		if (vp->v_iflag & VI_OWEINACT)
2871 			vinactive(vp, curthread);
2872 		if (func != VPUTX_VUNREF)
2873 			VOP_UNLOCK(vp, 0);
2874 	}
2875 	vdropl(vp);
2876 }
2877 
2878 /*
2879  * Vnode put/release.
2880  * If count drops to zero, call inactive routine and return to freelist.
2881  */
2882 void
2883 vrele(struct vnode *vp)
2884 {
2885 
2886 	vputx(vp, VPUTX_VRELE);
2887 }
2888 
2889 /*
2890  * Release an already locked vnode.  This give the same effects as
2891  * unlock+vrele(), but takes less time and avoids releasing and
2892  * re-aquiring the lock (as vrele() acquires the lock internally.)
2893  */
2894 void
2895 vput(struct vnode *vp)
2896 {
2897 
2898 	vputx(vp, VPUTX_VPUT);
2899 }
2900 
2901 /*
2902  * Release an exclusively locked vnode. Do not unlock the vnode lock.
2903  */
2904 void
2905 vunref(struct vnode *vp)
2906 {
2907 
2908 	vputx(vp, VPUTX_VUNREF);
2909 }
2910 
2911 /*
2912  * Increase the hold count and activate if this is the first reference.
2913  */
2914 void
2915 _vhold(struct vnode *vp, bool locked)
2916 {
2917 	struct mount *mp;
2918 
2919 	if (locked)
2920 		ASSERT_VI_LOCKED(vp, __func__);
2921 	else
2922 		ASSERT_VI_UNLOCKED(vp, __func__);
2923 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2924 	if (!locked) {
2925 		if (refcount_acquire_if_not_zero(&vp->v_holdcnt)) {
2926 			VNODE_REFCOUNT_FENCE_ACQ();
2927 			VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2928 			    ("_vhold: vnode with holdcnt is free"));
2929 			return;
2930 		}
2931 		VI_LOCK(vp);
2932 	}
2933 	if ((vp->v_iflag & VI_FREE) == 0) {
2934 		refcount_acquire(&vp->v_holdcnt);
2935 		if (!locked)
2936 			VI_UNLOCK(vp);
2937 		return;
2938 	}
2939 	VNASSERT(vp->v_holdcnt == 0, vp,
2940 	    ("%s: wrong hold count", __func__));
2941 	VNASSERT(vp->v_op != NULL, vp,
2942 	    ("%s: vnode already reclaimed.", __func__));
2943 	/*
2944 	 * Remove a vnode from the free list, mark it as in use,
2945 	 * and put it on the active list.
2946 	 */
2947 	VNASSERT(vp->v_mount != NULL, vp,
2948 	    ("_vhold: vnode not on per mount vnode list"));
2949 	mp = vp->v_mount;
2950 	mtx_lock(&mp->mnt_listmtx);
2951 	if ((vp->v_mflag & VMP_TMPMNTFREELIST) != 0) {
2952 		TAILQ_REMOVE(&mp->mnt_tmpfreevnodelist, vp, v_actfreelist);
2953 		mp->mnt_tmpfreevnodelistsize--;
2954 		vp->v_mflag &= ~VMP_TMPMNTFREELIST;
2955 	} else {
2956 		mtx_lock(&vnode_free_list_mtx);
2957 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
2958 		freevnodes--;
2959 		mtx_unlock(&vnode_free_list_mtx);
2960 	}
2961 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
2962 	    ("Activating already active vnode"));
2963 	vp->v_iflag &= ~VI_FREE;
2964 	vp->v_iflag |= VI_ACTIVE;
2965 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
2966 	mp->mnt_activevnodelistsize++;
2967 	mtx_unlock(&mp->mnt_listmtx);
2968 	refcount_acquire(&vp->v_holdcnt);
2969 	if (!locked)
2970 		VI_UNLOCK(vp);
2971 }
2972 
2973 /*
2974  * Drop the hold count of the vnode.  If this is the last reference to
2975  * the vnode we place it on the free list unless it has been vgone'd
2976  * (marked VI_DOOMED) in which case we will free it.
2977  *
2978  * Because the vnode vm object keeps a hold reference on the vnode if
2979  * there is at least one resident non-cached page, the vnode cannot
2980  * leave the active list without the page cleanup done.
2981  */
2982 void
2983 _vdrop(struct vnode *vp, bool locked)
2984 {
2985 	struct bufobj *bo;
2986 	struct mount *mp;
2987 	int active;
2988 
2989 	if (locked)
2990 		ASSERT_VI_LOCKED(vp, __func__);
2991 	else
2992 		ASSERT_VI_UNLOCKED(vp, __func__);
2993 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2994 	if ((int)vp->v_holdcnt <= 0)
2995 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
2996 	if (!locked) {
2997 		if (refcount_release_if_not_last(&vp->v_holdcnt))
2998 			return;
2999 		VI_LOCK(vp);
3000 	}
3001 	if (refcount_release(&vp->v_holdcnt) == 0) {
3002 		VI_UNLOCK(vp);
3003 		return;
3004 	}
3005 	if ((vp->v_iflag & VI_DOOMED) == 0) {
3006 		/*
3007 		 * Mark a vnode as free: remove it from its active list
3008 		 * and put it up for recycling on the freelist.
3009 		 */
3010 		VNASSERT(vp->v_op != NULL, vp,
3011 		    ("vdropl: vnode already reclaimed."));
3012 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
3013 		    ("vnode already free"));
3014 		VNASSERT(vp->v_holdcnt == 0, vp,
3015 		    ("vdropl: freeing when we shouldn't"));
3016 		active = vp->v_iflag & VI_ACTIVE;
3017 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
3018 			vp->v_iflag &= ~VI_ACTIVE;
3019 			mp = vp->v_mount;
3020 			if (mp != NULL) {
3021 				mtx_lock(&mp->mnt_listmtx);
3022 				if (active) {
3023 					TAILQ_REMOVE(&mp->mnt_activevnodelist,
3024 					    vp, v_actfreelist);
3025 					mp->mnt_activevnodelistsize--;
3026 				}
3027 				TAILQ_INSERT_TAIL(&mp->mnt_tmpfreevnodelist,
3028 				    vp, v_actfreelist);
3029 				mp->mnt_tmpfreevnodelistsize++;
3030 				vp->v_iflag |= VI_FREE;
3031 				vp->v_mflag |= VMP_TMPMNTFREELIST;
3032 				VI_UNLOCK(vp);
3033 				if (mp->mnt_tmpfreevnodelistsize >=
3034 				    mnt_free_list_batch)
3035 					vnlru_return_batch_locked(mp);
3036 				mtx_unlock(&mp->mnt_listmtx);
3037 			} else {
3038 				VNASSERT(active == 0, vp,
3039 				    ("vdropl: active vnode not on per mount "
3040 				    "vnode list"));
3041 				mtx_lock(&vnode_free_list_mtx);
3042 				TAILQ_INSERT_TAIL(&vnode_free_list, vp,
3043 				    v_actfreelist);
3044 				freevnodes++;
3045 				vp->v_iflag |= VI_FREE;
3046 				VI_UNLOCK(vp);
3047 				mtx_unlock(&vnode_free_list_mtx);
3048 			}
3049 		} else {
3050 			VI_UNLOCK(vp);
3051 			counter_u64_add(free_owe_inact, 1);
3052 		}
3053 		return;
3054 	}
3055 	/*
3056 	 * The vnode has been marked for destruction, so free it.
3057 	 *
3058 	 * The vnode will be returned to the zone where it will
3059 	 * normally remain until it is needed for another vnode. We
3060 	 * need to cleanup (or verify that the cleanup has already
3061 	 * been done) any residual data left from its current use
3062 	 * so as not to contaminate the freshly allocated vnode.
3063 	 */
3064 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
3065 	atomic_subtract_long(&numvnodes, 1);
3066 	bo = &vp->v_bufobj;
3067 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
3068 	    ("cleaned vnode still on the free list."));
3069 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
3070 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
3071 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
3072 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
3073 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
3074 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
3075 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
3076 	    ("clean blk trie not empty"));
3077 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
3078 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
3079 	    ("dirty blk trie not empty"));
3080 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
3081 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
3082 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
3083 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
3084 	    ("Dangling rangelock waiters"));
3085 	VI_UNLOCK(vp);
3086 #ifdef MAC
3087 	mac_vnode_destroy(vp);
3088 #endif
3089 	if (vp->v_pollinfo != NULL) {
3090 		destroy_vpollinfo(vp->v_pollinfo);
3091 		vp->v_pollinfo = NULL;
3092 	}
3093 #ifdef INVARIANTS
3094 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
3095 	vp->v_op = NULL;
3096 #endif
3097 	vp->v_mountedhere = NULL;
3098 	vp->v_unpcb = NULL;
3099 	vp->v_rdev = NULL;
3100 	vp->v_fifoinfo = NULL;
3101 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
3102 	vp->v_iflag = 0;
3103 	vp->v_vflag = 0;
3104 	bo->bo_flag = 0;
3105 	uma_zfree(vnode_zone, vp);
3106 }
3107 
3108 /*
3109  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3110  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3111  * OWEINACT tracks whether a vnode missed a call to inactive due to a
3112  * failed lock upgrade.
3113  */
3114 void
3115 vinactive(struct vnode *vp, struct thread *td)
3116 {
3117 	struct vm_object *obj;
3118 
3119 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3120 	ASSERT_VI_LOCKED(vp, "vinactive");
3121 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3122 	    ("vinactive: recursed on VI_DOINGINACT"));
3123 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3124 	vp->v_iflag |= VI_DOINGINACT;
3125 	vp->v_iflag &= ~VI_OWEINACT;
3126 	VI_UNLOCK(vp);
3127 	/*
3128 	 * Before moving off the active list, we must be sure that any
3129 	 * modified pages are converted into the vnode's dirty
3130 	 * buffers, since these will no longer be checked once the
3131 	 * vnode is on the inactive list.
3132 	 *
3133 	 * The write-out of the dirty pages is asynchronous.  At the
3134 	 * point that VOP_INACTIVE() is called, there could still be
3135 	 * pending I/O and dirty pages in the object.
3136 	 */
3137 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3138 	    (obj->flags & OBJ_MIGHTBEDIRTY) != 0) {
3139 		VM_OBJECT_WLOCK(obj);
3140 		vm_object_page_clean(obj, 0, 0, 0);
3141 		VM_OBJECT_WUNLOCK(obj);
3142 	}
3143 	VOP_INACTIVE(vp, td);
3144 	VI_LOCK(vp);
3145 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3146 	    ("vinactive: lost VI_DOINGINACT"));
3147 	vp->v_iflag &= ~VI_DOINGINACT;
3148 }
3149 
3150 /*
3151  * Remove any vnodes in the vnode table belonging to mount point mp.
3152  *
3153  * If FORCECLOSE is not specified, there should not be any active ones,
3154  * return error if any are found (nb: this is a user error, not a
3155  * system error). If FORCECLOSE is specified, detach any active vnodes
3156  * that are found.
3157  *
3158  * If WRITECLOSE is set, only flush out regular file vnodes open for
3159  * writing.
3160  *
3161  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3162  *
3163  * `rootrefs' specifies the base reference count for the root vnode
3164  * of this filesystem. The root vnode is considered busy if its
3165  * v_usecount exceeds this value. On a successful return, vflush(, td)
3166  * will call vrele() on the root vnode exactly rootrefs times.
3167  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3168  * be zero.
3169  */
3170 #ifdef DIAGNOSTIC
3171 static int busyprt = 0;		/* print out busy vnodes */
3172 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3173 #endif
3174 
3175 int
3176 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3177 {
3178 	struct vnode *vp, *mvp, *rootvp = NULL;
3179 	struct vattr vattr;
3180 	int busy = 0, error;
3181 
3182 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3183 	    rootrefs, flags);
3184 	if (rootrefs > 0) {
3185 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3186 		    ("vflush: bad args"));
3187 		/*
3188 		 * Get the filesystem root vnode. We can vput() it
3189 		 * immediately, since with rootrefs > 0, it won't go away.
3190 		 */
3191 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3192 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3193 			    __func__, error);
3194 			return (error);
3195 		}
3196 		vput(rootvp);
3197 	}
3198 loop:
3199 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3200 		vholdl(vp);
3201 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3202 		if (error) {
3203 			vdrop(vp);
3204 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3205 			goto loop;
3206 		}
3207 		/*
3208 		 * Skip over a vnodes marked VV_SYSTEM.
3209 		 */
3210 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3211 			VOP_UNLOCK(vp, 0);
3212 			vdrop(vp);
3213 			continue;
3214 		}
3215 		/*
3216 		 * If WRITECLOSE is set, flush out unlinked but still open
3217 		 * files (even if open only for reading) and regular file
3218 		 * vnodes open for writing.
3219 		 */
3220 		if (flags & WRITECLOSE) {
3221 			if (vp->v_object != NULL) {
3222 				VM_OBJECT_WLOCK(vp->v_object);
3223 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3224 				VM_OBJECT_WUNLOCK(vp->v_object);
3225 			}
3226 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3227 			if (error != 0) {
3228 				VOP_UNLOCK(vp, 0);
3229 				vdrop(vp);
3230 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3231 				return (error);
3232 			}
3233 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3234 			VI_LOCK(vp);
3235 
3236 			if ((vp->v_type == VNON ||
3237 			    (error == 0 && vattr.va_nlink > 0)) &&
3238 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3239 				VOP_UNLOCK(vp, 0);
3240 				vdropl(vp);
3241 				continue;
3242 			}
3243 		} else
3244 			VI_LOCK(vp);
3245 		/*
3246 		 * With v_usecount == 0, all we need to do is clear out the
3247 		 * vnode data structures and we are done.
3248 		 *
3249 		 * If FORCECLOSE is set, forcibly close the vnode.
3250 		 */
3251 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3252 			vgonel(vp);
3253 		} else {
3254 			busy++;
3255 #ifdef DIAGNOSTIC
3256 			if (busyprt)
3257 				vn_printf(vp, "vflush: busy vnode ");
3258 #endif
3259 		}
3260 		VOP_UNLOCK(vp, 0);
3261 		vdropl(vp);
3262 	}
3263 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3264 		/*
3265 		 * If just the root vnode is busy, and if its refcount
3266 		 * is equal to `rootrefs', then go ahead and kill it.
3267 		 */
3268 		VI_LOCK(rootvp);
3269 		KASSERT(busy > 0, ("vflush: not busy"));
3270 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3271 		    ("vflush: usecount %d < rootrefs %d",
3272 		     rootvp->v_usecount, rootrefs));
3273 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3274 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3275 			vgone(rootvp);
3276 			VOP_UNLOCK(rootvp, 0);
3277 			busy = 0;
3278 		} else
3279 			VI_UNLOCK(rootvp);
3280 	}
3281 	if (busy) {
3282 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3283 		    busy);
3284 		return (EBUSY);
3285 	}
3286 	for (; rootrefs > 0; rootrefs--)
3287 		vrele(rootvp);
3288 	return (0);
3289 }
3290 
3291 /*
3292  * Recycle an unused vnode to the front of the free list.
3293  */
3294 int
3295 vrecycle(struct vnode *vp)
3296 {
3297 	int recycled;
3298 
3299 	VI_LOCK(vp);
3300 	recycled = vrecyclel(vp);
3301 	VI_UNLOCK(vp);
3302 	return (recycled);
3303 }
3304 
3305 /*
3306  * vrecycle, with the vp interlock held.
3307  */
3308 int
3309 vrecyclel(struct vnode *vp)
3310 {
3311 	int recycled;
3312 
3313 	ASSERT_VOP_ELOCKED(vp, __func__);
3314 	ASSERT_VI_LOCKED(vp, __func__);
3315 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3316 	recycled = 0;
3317 	if (vp->v_usecount == 0) {
3318 		recycled = 1;
3319 		vgonel(vp);
3320 	}
3321 	return (recycled);
3322 }
3323 
3324 /*
3325  * Eliminate all activity associated with a vnode
3326  * in preparation for reuse.
3327  */
3328 void
3329 vgone(struct vnode *vp)
3330 {
3331 	VI_LOCK(vp);
3332 	vgonel(vp);
3333 	VI_UNLOCK(vp);
3334 }
3335 
3336 static void
3337 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3338     struct vnode *lowervp __unused)
3339 {
3340 }
3341 
3342 /*
3343  * Notify upper mounts about reclaimed or unlinked vnode.
3344  */
3345 void
3346 vfs_notify_upper(struct vnode *vp, int event)
3347 {
3348 	static struct vfsops vgonel_vfsops = {
3349 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3350 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3351 	};
3352 	struct mount *mp, *ump, *mmp;
3353 
3354 	mp = vp->v_mount;
3355 	if (mp == NULL)
3356 		return;
3357 
3358 	MNT_ILOCK(mp);
3359 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3360 		goto unlock;
3361 	MNT_IUNLOCK(mp);
3362 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3363 	mmp->mnt_op = &vgonel_vfsops;
3364 	mmp->mnt_kern_flag |= MNTK_MARKER;
3365 	MNT_ILOCK(mp);
3366 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3367 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3368 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3369 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3370 			continue;
3371 		}
3372 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3373 		MNT_IUNLOCK(mp);
3374 		switch (event) {
3375 		case VFS_NOTIFY_UPPER_RECLAIM:
3376 			VFS_RECLAIM_LOWERVP(ump, vp);
3377 			break;
3378 		case VFS_NOTIFY_UPPER_UNLINK:
3379 			VFS_UNLINK_LOWERVP(ump, vp);
3380 			break;
3381 		default:
3382 			KASSERT(0, ("invalid event %d", event));
3383 			break;
3384 		}
3385 		MNT_ILOCK(mp);
3386 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3387 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3388 	}
3389 	free(mmp, M_TEMP);
3390 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3391 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3392 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3393 		wakeup(&mp->mnt_uppers);
3394 	}
3395 unlock:
3396 	MNT_IUNLOCK(mp);
3397 }
3398 
3399 /*
3400  * vgone, with the vp interlock held.
3401  */
3402 static void
3403 vgonel(struct vnode *vp)
3404 {
3405 	struct thread *td;
3406 	int oweinact;
3407 	int active;
3408 	struct mount *mp;
3409 
3410 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3411 	ASSERT_VI_LOCKED(vp, "vgonel");
3412 	VNASSERT(vp->v_holdcnt, vp,
3413 	    ("vgonel: vp %p has no reference.", vp));
3414 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3415 	td = curthread;
3416 
3417 	/*
3418 	 * Don't vgonel if we're already doomed.
3419 	 */
3420 	if (vp->v_iflag & VI_DOOMED)
3421 		return;
3422 	vp->v_iflag |= VI_DOOMED;
3423 
3424 	/*
3425 	 * Check to see if the vnode is in use.  If so, we have to call
3426 	 * VOP_CLOSE() and VOP_INACTIVE().
3427 	 */
3428 	active = vp->v_usecount;
3429 	oweinact = (vp->v_iflag & VI_OWEINACT);
3430 	VI_UNLOCK(vp);
3431 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3432 
3433 	/*
3434 	 * If purging an active vnode, it must be closed and
3435 	 * deactivated before being reclaimed.
3436 	 */
3437 	if (active)
3438 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3439 	if (oweinact || active) {
3440 		VI_LOCK(vp);
3441 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
3442 			vinactive(vp, td);
3443 		VI_UNLOCK(vp);
3444 	}
3445 	if (vp->v_type == VSOCK)
3446 		vfs_unp_reclaim(vp);
3447 
3448 	/*
3449 	 * Clean out any buffers associated with the vnode.
3450 	 * If the flush fails, just toss the buffers.
3451 	 */
3452 	mp = NULL;
3453 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3454 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3455 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3456 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3457 			;
3458 	}
3459 
3460 	BO_LOCK(&vp->v_bufobj);
3461 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3462 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3463 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3464 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3465 	    ("vp %p bufobj not invalidated", vp));
3466 
3467 	/*
3468 	 * For VMIO bufobj, BO_DEAD is set in vm_object_terminate()
3469 	 * after the object's page queue is flushed.
3470 	 */
3471 	if (vp->v_bufobj.bo_object == NULL)
3472 		vp->v_bufobj.bo_flag |= BO_DEAD;
3473 	BO_UNLOCK(&vp->v_bufobj);
3474 
3475 	/*
3476 	 * Reclaim the vnode.
3477 	 */
3478 	if (VOP_RECLAIM(vp, td))
3479 		panic("vgone: cannot reclaim");
3480 	if (mp != NULL)
3481 		vn_finished_secondary_write(mp);
3482 	VNASSERT(vp->v_object == NULL, vp,
3483 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
3484 	/*
3485 	 * Clear the advisory locks and wake up waiting threads.
3486 	 */
3487 	(void)VOP_ADVLOCKPURGE(vp);
3488 	vp->v_lockf = NULL;
3489 	/*
3490 	 * Delete from old mount point vnode list.
3491 	 */
3492 	delmntque(vp);
3493 	cache_purge(vp);
3494 	/*
3495 	 * Done with purge, reset to the standard lock and invalidate
3496 	 * the vnode.
3497 	 */
3498 	VI_LOCK(vp);
3499 	vp->v_vnlock = &vp->v_lock;
3500 	vp->v_op = &dead_vnodeops;
3501 	vp->v_tag = "none";
3502 	vp->v_type = VBAD;
3503 }
3504 
3505 /*
3506  * Calculate the total number of references to a special device.
3507  */
3508 int
3509 vcount(struct vnode *vp)
3510 {
3511 	int count;
3512 
3513 	dev_lock();
3514 	count = vp->v_rdev->si_usecount;
3515 	dev_unlock();
3516 	return (count);
3517 }
3518 
3519 /*
3520  * Same as above, but using the struct cdev *as argument
3521  */
3522 int
3523 count_dev(struct cdev *dev)
3524 {
3525 	int count;
3526 
3527 	dev_lock();
3528 	count = dev->si_usecount;
3529 	dev_unlock();
3530 	return(count);
3531 }
3532 
3533 /*
3534  * Print out a description of a vnode.
3535  */
3536 static char *typename[] =
3537 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3538  "VMARKER"};
3539 
3540 void
3541 vn_printf(struct vnode *vp, const char *fmt, ...)
3542 {
3543 	va_list ap;
3544 	char buf[256], buf2[16];
3545 	u_long flags;
3546 
3547 	va_start(ap, fmt);
3548 	vprintf(fmt, ap);
3549 	va_end(ap);
3550 	printf("%p: ", (void *)vp);
3551 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
3552 	printf("    usecount %d, writecount %d, refcount %d",
3553 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
3554 	switch (vp->v_type) {
3555 	case VDIR:
3556 		printf(" mountedhere %p\n", vp->v_mountedhere);
3557 		break;
3558 	case VCHR:
3559 		printf(" rdev %p\n", vp->v_rdev);
3560 		break;
3561 	case VSOCK:
3562 		printf(" socket %p\n", vp->v_unpcb);
3563 		break;
3564 	case VFIFO:
3565 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
3566 		break;
3567 	default:
3568 		printf("\n");
3569 		break;
3570 	}
3571 	buf[0] = '\0';
3572 	buf[1] = '\0';
3573 	if (vp->v_vflag & VV_ROOT)
3574 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3575 	if (vp->v_vflag & VV_ISTTY)
3576 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3577 	if (vp->v_vflag & VV_NOSYNC)
3578 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3579 	if (vp->v_vflag & VV_ETERNALDEV)
3580 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3581 	if (vp->v_vflag & VV_CACHEDLABEL)
3582 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3583 	if (vp->v_vflag & VV_COPYONWRITE)
3584 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3585 	if (vp->v_vflag & VV_SYSTEM)
3586 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3587 	if (vp->v_vflag & VV_PROCDEP)
3588 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3589 	if (vp->v_vflag & VV_NOKNOTE)
3590 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3591 	if (vp->v_vflag & VV_DELETED)
3592 		strlcat(buf, "|VV_DELETED", sizeof(buf));
3593 	if (vp->v_vflag & VV_MD)
3594 		strlcat(buf, "|VV_MD", sizeof(buf));
3595 	if (vp->v_vflag & VV_FORCEINSMQ)
3596 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
3597 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
3598 	    VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
3599 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
3600 	if (flags != 0) {
3601 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
3602 		strlcat(buf, buf2, sizeof(buf));
3603 	}
3604 	if (vp->v_iflag & VI_MOUNT)
3605 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
3606 	if (vp->v_iflag & VI_DOOMED)
3607 		strlcat(buf, "|VI_DOOMED", sizeof(buf));
3608 	if (vp->v_iflag & VI_FREE)
3609 		strlcat(buf, "|VI_FREE", sizeof(buf));
3610 	if (vp->v_iflag & VI_ACTIVE)
3611 		strlcat(buf, "|VI_ACTIVE", sizeof(buf));
3612 	if (vp->v_iflag & VI_DOINGINACT)
3613 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
3614 	if (vp->v_iflag & VI_OWEINACT)
3615 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
3616 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOOMED | VI_FREE |
3617 	    VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT);
3618 	if (flags != 0) {
3619 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
3620 		strlcat(buf, buf2, sizeof(buf));
3621 	}
3622 	printf("    flags (%s)\n", buf + 1);
3623 	if (mtx_owned(VI_MTX(vp)))
3624 		printf(" VI_LOCKed");
3625 	if (vp->v_object != NULL)
3626 		printf("    v_object %p ref %d pages %d "
3627 		    "cleanbuf %d dirtybuf %d\n",
3628 		    vp->v_object, vp->v_object->ref_count,
3629 		    vp->v_object->resident_page_count,
3630 		    vp->v_bufobj.bo_clean.bv_cnt,
3631 		    vp->v_bufobj.bo_dirty.bv_cnt);
3632 	printf("    ");
3633 	lockmgr_printinfo(vp->v_vnlock);
3634 	if (vp->v_data != NULL)
3635 		VOP_PRINT(vp);
3636 }
3637 
3638 #ifdef DDB
3639 /*
3640  * List all of the locked vnodes in the system.
3641  * Called when debugging the kernel.
3642  */
3643 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
3644 {
3645 	struct mount *mp;
3646 	struct vnode *vp;
3647 
3648 	/*
3649 	 * Note: because this is DDB, we can't obey the locking semantics
3650 	 * for these structures, which means we could catch an inconsistent
3651 	 * state and dereference a nasty pointer.  Not much to be done
3652 	 * about that.
3653 	 */
3654 	db_printf("Locked vnodes\n");
3655 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3656 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3657 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
3658 				vn_printf(vp, "vnode ");
3659 		}
3660 	}
3661 }
3662 
3663 /*
3664  * Show details about the given vnode.
3665  */
3666 DB_SHOW_COMMAND(vnode, db_show_vnode)
3667 {
3668 	struct vnode *vp;
3669 
3670 	if (!have_addr)
3671 		return;
3672 	vp = (struct vnode *)addr;
3673 	vn_printf(vp, "vnode ");
3674 }
3675 
3676 /*
3677  * Show details about the given mount point.
3678  */
3679 DB_SHOW_COMMAND(mount, db_show_mount)
3680 {
3681 	struct mount *mp;
3682 	struct vfsopt *opt;
3683 	struct statfs *sp;
3684 	struct vnode *vp;
3685 	char buf[512];
3686 	uint64_t mflags;
3687 	u_int flags;
3688 
3689 	if (!have_addr) {
3690 		/* No address given, print short info about all mount points. */
3691 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3692 			db_printf("%p %s on %s (%s)\n", mp,
3693 			    mp->mnt_stat.f_mntfromname,
3694 			    mp->mnt_stat.f_mntonname,
3695 			    mp->mnt_stat.f_fstypename);
3696 			if (db_pager_quit)
3697 				break;
3698 		}
3699 		db_printf("\nMore info: show mount <addr>\n");
3700 		return;
3701 	}
3702 
3703 	mp = (struct mount *)addr;
3704 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
3705 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
3706 
3707 	buf[0] = '\0';
3708 	mflags = mp->mnt_flag;
3709 #define	MNT_FLAG(flag)	do {						\
3710 	if (mflags & (flag)) {						\
3711 		if (buf[0] != '\0')					\
3712 			strlcat(buf, ", ", sizeof(buf));		\
3713 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
3714 		mflags &= ~(flag);					\
3715 	}								\
3716 } while (0)
3717 	MNT_FLAG(MNT_RDONLY);
3718 	MNT_FLAG(MNT_SYNCHRONOUS);
3719 	MNT_FLAG(MNT_NOEXEC);
3720 	MNT_FLAG(MNT_NOSUID);
3721 	MNT_FLAG(MNT_NFS4ACLS);
3722 	MNT_FLAG(MNT_UNION);
3723 	MNT_FLAG(MNT_ASYNC);
3724 	MNT_FLAG(MNT_SUIDDIR);
3725 	MNT_FLAG(MNT_SOFTDEP);
3726 	MNT_FLAG(MNT_NOSYMFOLLOW);
3727 	MNT_FLAG(MNT_GJOURNAL);
3728 	MNT_FLAG(MNT_MULTILABEL);
3729 	MNT_FLAG(MNT_ACLS);
3730 	MNT_FLAG(MNT_NOATIME);
3731 	MNT_FLAG(MNT_NOCLUSTERR);
3732 	MNT_FLAG(MNT_NOCLUSTERW);
3733 	MNT_FLAG(MNT_SUJ);
3734 	MNT_FLAG(MNT_EXRDONLY);
3735 	MNT_FLAG(MNT_EXPORTED);
3736 	MNT_FLAG(MNT_DEFEXPORTED);
3737 	MNT_FLAG(MNT_EXPORTANON);
3738 	MNT_FLAG(MNT_EXKERB);
3739 	MNT_FLAG(MNT_EXPUBLIC);
3740 	MNT_FLAG(MNT_LOCAL);
3741 	MNT_FLAG(MNT_QUOTA);
3742 	MNT_FLAG(MNT_ROOTFS);
3743 	MNT_FLAG(MNT_USER);
3744 	MNT_FLAG(MNT_IGNORE);
3745 	MNT_FLAG(MNT_UPDATE);
3746 	MNT_FLAG(MNT_DELEXPORT);
3747 	MNT_FLAG(MNT_RELOAD);
3748 	MNT_FLAG(MNT_FORCE);
3749 	MNT_FLAG(MNT_SNAPSHOT);
3750 	MNT_FLAG(MNT_BYFSID);
3751 #undef MNT_FLAG
3752 	if (mflags != 0) {
3753 		if (buf[0] != '\0')
3754 			strlcat(buf, ", ", sizeof(buf));
3755 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3756 		    "0x%016jx", mflags);
3757 	}
3758 	db_printf("    mnt_flag = %s\n", buf);
3759 
3760 	buf[0] = '\0';
3761 	flags = mp->mnt_kern_flag;
3762 #define	MNT_KERN_FLAG(flag)	do {					\
3763 	if (flags & (flag)) {						\
3764 		if (buf[0] != '\0')					\
3765 			strlcat(buf, ", ", sizeof(buf));		\
3766 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
3767 		flags &= ~(flag);					\
3768 	}								\
3769 } while (0)
3770 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
3771 	MNT_KERN_FLAG(MNTK_ASYNC);
3772 	MNT_KERN_FLAG(MNTK_SOFTDEP);
3773 	MNT_KERN_FLAG(MNTK_NOINSMNTQ);
3774 	MNT_KERN_FLAG(MNTK_DRAINING);
3775 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
3776 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
3777 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
3778 	MNT_KERN_FLAG(MNTK_NO_IOPF);
3779 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
3780 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
3781 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
3782 	MNT_KERN_FLAG(MNTK_MARKER);
3783 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
3784 	MNT_KERN_FLAG(MNTK_NOASYNC);
3785 	MNT_KERN_FLAG(MNTK_UNMOUNT);
3786 	MNT_KERN_FLAG(MNTK_MWAIT);
3787 	MNT_KERN_FLAG(MNTK_SUSPEND);
3788 	MNT_KERN_FLAG(MNTK_SUSPEND2);
3789 	MNT_KERN_FLAG(MNTK_SUSPENDED);
3790 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
3791 	MNT_KERN_FLAG(MNTK_NOKNOTE);
3792 #undef MNT_KERN_FLAG
3793 	if (flags != 0) {
3794 		if (buf[0] != '\0')
3795 			strlcat(buf, ", ", sizeof(buf));
3796 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3797 		    "0x%08x", flags);
3798 	}
3799 	db_printf("    mnt_kern_flag = %s\n", buf);
3800 
3801 	db_printf("    mnt_opt = ");
3802 	opt = TAILQ_FIRST(mp->mnt_opt);
3803 	if (opt != NULL) {
3804 		db_printf("%s", opt->name);
3805 		opt = TAILQ_NEXT(opt, link);
3806 		while (opt != NULL) {
3807 			db_printf(", %s", opt->name);
3808 			opt = TAILQ_NEXT(opt, link);
3809 		}
3810 	}
3811 	db_printf("\n");
3812 
3813 	sp = &mp->mnt_stat;
3814 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
3815 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
3816 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
3817 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
3818 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
3819 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
3820 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
3821 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
3822 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
3823 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
3824 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
3825 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
3826 
3827 	db_printf("    mnt_cred = { uid=%u ruid=%u",
3828 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
3829 	if (jailed(mp->mnt_cred))
3830 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
3831 	db_printf(" }\n");
3832 	db_printf("    mnt_ref = %d\n", mp->mnt_ref);
3833 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
3834 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
3835 	db_printf("    mnt_activevnodelistsize = %d\n",
3836 	    mp->mnt_activevnodelistsize);
3837 	db_printf("    mnt_writeopcount = %d\n", mp->mnt_writeopcount);
3838 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
3839 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
3840 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
3841 	db_printf("    mnt_lockref = %d\n", mp->mnt_lockref);
3842 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
3843 	db_printf("    mnt_secondary_accwrites = %d\n",
3844 	    mp->mnt_secondary_accwrites);
3845 	db_printf("    mnt_gjprovider = %s\n",
3846 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
3847 
3848 	db_printf("\n\nList of active vnodes\n");
3849 	TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) {
3850 		if (vp->v_type != VMARKER) {
3851 			vn_printf(vp, "vnode ");
3852 			if (db_pager_quit)
3853 				break;
3854 		}
3855 	}
3856 	db_printf("\n\nList of inactive vnodes\n");
3857 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3858 		if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) {
3859 			vn_printf(vp, "vnode ");
3860 			if (db_pager_quit)
3861 				break;
3862 		}
3863 	}
3864 }
3865 #endif	/* DDB */
3866 
3867 /*
3868  * Fill in a struct xvfsconf based on a struct vfsconf.
3869  */
3870 static int
3871 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
3872 {
3873 	struct xvfsconf xvfsp;
3874 
3875 	bzero(&xvfsp, sizeof(xvfsp));
3876 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3877 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3878 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3879 	xvfsp.vfc_flags = vfsp->vfc_flags;
3880 	/*
3881 	 * These are unused in userland, we keep them
3882 	 * to not break binary compatibility.
3883 	 */
3884 	xvfsp.vfc_vfsops = NULL;
3885 	xvfsp.vfc_next = NULL;
3886 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3887 }
3888 
3889 #ifdef COMPAT_FREEBSD32
3890 struct xvfsconf32 {
3891 	uint32_t	vfc_vfsops;
3892 	char		vfc_name[MFSNAMELEN];
3893 	int32_t		vfc_typenum;
3894 	int32_t		vfc_refcount;
3895 	int32_t		vfc_flags;
3896 	uint32_t	vfc_next;
3897 };
3898 
3899 static int
3900 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
3901 {
3902 	struct xvfsconf32 xvfsp;
3903 
3904 	bzero(&xvfsp, sizeof(xvfsp));
3905 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3906 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3907 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3908 	xvfsp.vfc_flags = vfsp->vfc_flags;
3909 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3910 }
3911 #endif
3912 
3913 /*
3914  * Top level filesystem related information gathering.
3915  */
3916 static int
3917 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
3918 {
3919 	struct vfsconf *vfsp;
3920 	int error;
3921 
3922 	error = 0;
3923 	vfsconf_slock();
3924 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3925 #ifdef COMPAT_FREEBSD32
3926 		if (req->flags & SCTL_MASK32)
3927 			error = vfsconf2x32(req, vfsp);
3928 		else
3929 #endif
3930 			error = vfsconf2x(req, vfsp);
3931 		if (error)
3932 			break;
3933 	}
3934 	vfsconf_sunlock();
3935 	return (error);
3936 }
3937 
3938 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
3939     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
3940     "S,xvfsconf", "List of all configured filesystems");
3941 
3942 #ifndef BURN_BRIDGES
3943 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
3944 
3945 static int
3946 vfs_sysctl(SYSCTL_HANDLER_ARGS)
3947 {
3948 	int *name = (int *)arg1 - 1;	/* XXX */
3949 	u_int namelen = arg2 + 1;	/* XXX */
3950 	struct vfsconf *vfsp;
3951 
3952 	log(LOG_WARNING, "userland calling deprecated sysctl, "
3953 	    "please rebuild world\n");
3954 
3955 #if 1 || defined(COMPAT_PRELITE2)
3956 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
3957 	if (namelen == 1)
3958 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
3959 #endif
3960 
3961 	switch (name[1]) {
3962 	case VFS_MAXTYPENUM:
3963 		if (namelen != 2)
3964 			return (ENOTDIR);
3965 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
3966 	case VFS_CONF:
3967 		if (namelen != 3)
3968 			return (ENOTDIR);	/* overloaded */
3969 		vfsconf_slock();
3970 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3971 			if (vfsp->vfc_typenum == name[2])
3972 				break;
3973 		}
3974 		vfsconf_sunlock();
3975 		if (vfsp == NULL)
3976 			return (EOPNOTSUPP);
3977 #ifdef COMPAT_FREEBSD32
3978 		if (req->flags & SCTL_MASK32)
3979 			return (vfsconf2x32(req, vfsp));
3980 		else
3981 #endif
3982 			return (vfsconf2x(req, vfsp));
3983 	}
3984 	return (EOPNOTSUPP);
3985 }
3986 
3987 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
3988     CTLFLAG_MPSAFE, vfs_sysctl,
3989     "Generic filesystem");
3990 
3991 #if 1 || defined(COMPAT_PRELITE2)
3992 
3993 static int
3994 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
3995 {
3996 	int error;
3997 	struct vfsconf *vfsp;
3998 	struct ovfsconf ovfs;
3999 
4000 	vfsconf_slock();
4001 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4002 		bzero(&ovfs, sizeof(ovfs));
4003 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
4004 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
4005 		ovfs.vfc_index = vfsp->vfc_typenum;
4006 		ovfs.vfc_refcount = vfsp->vfc_refcount;
4007 		ovfs.vfc_flags = vfsp->vfc_flags;
4008 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4009 		if (error != 0) {
4010 			vfsconf_sunlock();
4011 			return (error);
4012 		}
4013 	}
4014 	vfsconf_sunlock();
4015 	return (0);
4016 }
4017 
4018 #endif /* 1 || COMPAT_PRELITE2 */
4019 #endif /* !BURN_BRIDGES */
4020 
4021 #define KINFO_VNODESLOP		10
4022 #ifdef notyet
4023 /*
4024  * Dump vnode list (via sysctl).
4025  */
4026 /* ARGSUSED */
4027 static int
4028 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4029 {
4030 	struct xvnode *xvn;
4031 	struct mount *mp;
4032 	struct vnode *vp;
4033 	int error, len, n;
4034 
4035 	/*
4036 	 * Stale numvnodes access is not fatal here.
4037 	 */
4038 	req->lock = 0;
4039 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4040 	if (!req->oldptr)
4041 		/* Make an estimate */
4042 		return (SYSCTL_OUT(req, 0, len));
4043 
4044 	error = sysctl_wire_old_buffer(req, 0);
4045 	if (error != 0)
4046 		return (error);
4047 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4048 	n = 0;
4049 	mtx_lock(&mountlist_mtx);
4050 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4051 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4052 			continue;
4053 		MNT_ILOCK(mp);
4054 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4055 			if (n == len)
4056 				break;
4057 			vref(vp);
4058 			xvn[n].xv_size = sizeof *xvn;
4059 			xvn[n].xv_vnode = vp;
4060 			xvn[n].xv_id = 0;	/* XXX compat */
4061 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4062 			XV_COPY(usecount);
4063 			XV_COPY(writecount);
4064 			XV_COPY(holdcnt);
4065 			XV_COPY(mount);
4066 			XV_COPY(numoutput);
4067 			XV_COPY(type);
4068 #undef XV_COPY
4069 			xvn[n].xv_flag = vp->v_vflag;
4070 
4071 			switch (vp->v_type) {
4072 			case VREG:
4073 			case VDIR:
4074 			case VLNK:
4075 				break;
4076 			case VBLK:
4077 			case VCHR:
4078 				if (vp->v_rdev == NULL) {
4079 					vrele(vp);
4080 					continue;
4081 				}
4082 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4083 				break;
4084 			case VSOCK:
4085 				xvn[n].xv_socket = vp->v_socket;
4086 				break;
4087 			case VFIFO:
4088 				xvn[n].xv_fifo = vp->v_fifoinfo;
4089 				break;
4090 			case VNON:
4091 			case VBAD:
4092 			default:
4093 				/* shouldn't happen? */
4094 				vrele(vp);
4095 				continue;
4096 			}
4097 			vrele(vp);
4098 			++n;
4099 		}
4100 		MNT_IUNLOCK(mp);
4101 		mtx_lock(&mountlist_mtx);
4102 		vfs_unbusy(mp);
4103 		if (n == len)
4104 			break;
4105 	}
4106 	mtx_unlock(&mountlist_mtx);
4107 
4108 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4109 	free(xvn, M_TEMP);
4110 	return (error);
4111 }
4112 
4113 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4114     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4115     "");
4116 #endif
4117 
4118 static void
4119 unmount_or_warn(struct mount *mp)
4120 {
4121 	int error;
4122 
4123 	error = dounmount(mp, MNT_FORCE, curthread);
4124 	if (error != 0) {
4125 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4126 		if (error == EBUSY)
4127 			printf("BUSY)\n");
4128 		else
4129 			printf("%d)\n", error);
4130 	}
4131 }
4132 
4133 /*
4134  * Unmount all filesystems. The list is traversed in reverse order
4135  * of mounting to avoid dependencies.
4136  */
4137 void
4138 vfs_unmountall(void)
4139 {
4140 	struct mount *mp, *tmp;
4141 
4142 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4143 
4144 	/*
4145 	 * Since this only runs when rebooting, it is not interlocked.
4146 	 */
4147 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4148 		vfs_ref(mp);
4149 
4150 		/*
4151 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4152 		 * unmount of the latter.
4153 		 */
4154 		if (mp == rootdevmp)
4155 			continue;
4156 
4157 		unmount_or_warn(mp);
4158 	}
4159 
4160 	if (rootdevmp != NULL)
4161 		unmount_or_warn(rootdevmp);
4162 }
4163 
4164 /*
4165  * perform msync on all vnodes under a mount point
4166  * the mount point must be locked.
4167  */
4168 void
4169 vfs_msync(struct mount *mp, int flags)
4170 {
4171 	struct vnode *vp, *mvp;
4172 	struct vm_object *obj;
4173 
4174 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4175 
4176 	vnlru_return_batch(mp);
4177 
4178 	MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
4179 		obj = vp->v_object;
4180 		if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 &&
4181 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
4182 			if (!vget(vp,
4183 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
4184 			    curthread)) {
4185 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
4186 					vput(vp);
4187 					continue;
4188 				}
4189 
4190 				obj = vp->v_object;
4191 				if (obj != NULL) {
4192 					VM_OBJECT_WLOCK(obj);
4193 					vm_object_page_clean(obj, 0, 0,
4194 					    flags == MNT_WAIT ?
4195 					    OBJPC_SYNC : OBJPC_NOSYNC);
4196 					VM_OBJECT_WUNLOCK(obj);
4197 				}
4198 				vput(vp);
4199 			}
4200 		} else
4201 			VI_UNLOCK(vp);
4202 	}
4203 }
4204 
4205 static void
4206 destroy_vpollinfo_free(struct vpollinfo *vi)
4207 {
4208 
4209 	knlist_destroy(&vi->vpi_selinfo.si_note);
4210 	mtx_destroy(&vi->vpi_lock);
4211 	uma_zfree(vnodepoll_zone, vi);
4212 }
4213 
4214 static void
4215 destroy_vpollinfo(struct vpollinfo *vi)
4216 {
4217 
4218 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4219 	seldrain(&vi->vpi_selinfo);
4220 	destroy_vpollinfo_free(vi);
4221 }
4222 
4223 /*
4224  * Initialize per-vnode helper structure to hold poll-related state.
4225  */
4226 void
4227 v_addpollinfo(struct vnode *vp)
4228 {
4229 	struct vpollinfo *vi;
4230 
4231 	if (vp->v_pollinfo != NULL)
4232 		return;
4233 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4234 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4235 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4236 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4237 	VI_LOCK(vp);
4238 	if (vp->v_pollinfo != NULL) {
4239 		VI_UNLOCK(vp);
4240 		destroy_vpollinfo_free(vi);
4241 		return;
4242 	}
4243 	vp->v_pollinfo = vi;
4244 	VI_UNLOCK(vp);
4245 }
4246 
4247 /*
4248  * Record a process's interest in events which might happen to
4249  * a vnode.  Because poll uses the historic select-style interface
4250  * internally, this routine serves as both the ``check for any
4251  * pending events'' and the ``record my interest in future events''
4252  * functions.  (These are done together, while the lock is held,
4253  * to avoid race conditions.)
4254  */
4255 int
4256 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4257 {
4258 
4259 	v_addpollinfo(vp);
4260 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4261 	if (vp->v_pollinfo->vpi_revents & events) {
4262 		/*
4263 		 * This leaves events we are not interested
4264 		 * in available for the other process which
4265 		 * which presumably had requested them
4266 		 * (otherwise they would never have been
4267 		 * recorded).
4268 		 */
4269 		events &= vp->v_pollinfo->vpi_revents;
4270 		vp->v_pollinfo->vpi_revents &= ~events;
4271 
4272 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4273 		return (events);
4274 	}
4275 	vp->v_pollinfo->vpi_events |= events;
4276 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4277 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4278 	return (0);
4279 }
4280 
4281 /*
4282  * Routine to create and manage a filesystem syncer vnode.
4283  */
4284 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4285 static int	sync_fsync(struct  vop_fsync_args *);
4286 static int	sync_inactive(struct  vop_inactive_args *);
4287 static int	sync_reclaim(struct  vop_reclaim_args *);
4288 
4289 static struct vop_vector sync_vnodeops = {
4290 	.vop_bypass =	VOP_EOPNOTSUPP,
4291 	.vop_close =	sync_close,		/* close */
4292 	.vop_fsync =	sync_fsync,		/* fsync */
4293 	.vop_inactive =	sync_inactive,	/* inactive */
4294 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4295 	.vop_lock1 =	vop_stdlock,	/* lock */
4296 	.vop_unlock =	vop_stdunlock,	/* unlock */
4297 	.vop_islocked =	vop_stdislocked,	/* islocked */
4298 };
4299 
4300 /*
4301  * Create a new filesystem syncer vnode for the specified mount point.
4302  */
4303 void
4304 vfs_allocate_syncvnode(struct mount *mp)
4305 {
4306 	struct vnode *vp;
4307 	struct bufobj *bo;
4308 	static long start, incr, next;
4309 	int error;
4310 
4311 	/* Allocate a new vnode */
4312 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4313 	if (error != 0)
4314 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4315 	vp->v_type = VNON;
4316 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4317 	vp->v_vflag |= VV_FORCEINSMQ;
4318 	error = insmntque(vp, mp);
4319 	if (error != 0)
4320 		panic("vfs_allocate_syncvnode: insmntque() failed");
4321 	vp->v_vflag &= ~VV_FORCEINSMQ;
4322 	VOP_UNLOCK(vp, 0);
4323 	/*
4324 	 * Place the vnode onto the syncer worklist. We attempt to
4325 	 * scatter them about on the list so that they will go off
4326 	 * at evenly distributed times even if all the filesystems
4327 	 * are mounted at once.
4328 	 */
4329 	next += incr;
4330 	if (next == 0 || next > syncer_maxdelay) {
4331 		start /= 2;
4332 		incr /= 2;
4333 		if (start == 0) {
4334 			start = syncer_maxdelay / 2;
4335 			incr = syncer_maxdelay;
4336 		}
4337 		next = start;
4338 	}
4339 	bo = &vp->v_bufobj;
4340 	BO_LOCK(bo);
4341 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4342 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4343 	mtx_lock(&sync_mtx);
4344 	sync_vnode_count++;
4345 	if (mp->mnt_syncer == NULL) {
4346 		mp->mnt_syncer = vp;
4347 		vp = NULL;
4348 	}
4349 	mtx_unlock(&sync_mtx);
4350 	BO_UNLOCK(bo);
4351 	if (vp != NULL) {
4352 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4353 		vgone(vp);
4354 		vput(vp);
4355 	}
4356 }
4357 
4358 void
4359 vfs_deallocate_syncvnode(struct mount *mp)
4360 {
4361 	struct vnode *vp;
4362 
4363 	mtx_lock(&sync_mtx);
4364 	vp = mp->mnt_syncer;
4365 	if (vp != NULL)
4366 		mp->mnt_syncer = NULL;
4367 	mtx_unlock(&sync_mtx);
4368 	if (vp != NULL)
4369 		vrele(vp);
4370 }
4371 
4372 /*
4373  * Do a lazy sync of the filesystem.
4374  */
4375 static int
4376 sync_fsync(struct vop_fsync_args *ap)
4377 {
4378 	struct vnode *syncvp = ap->a_vp;
4379 	struct mount *mp = syncvp->v_mount;
4380 	int error, save;
4381 	struct bufobj *bo;
4382 
4383 	/*
4384 	 * We only need to do something if this is a lazy evaluation.
4385 	 */
4386 	if (ap->a_waitfor != MNT_LAZY)
4387 		return (0);
4388 
4389 	/*
4390 	 * Move ourselves to the back of the sync list.
4391 	 */
4392 	bo = &syncvp->v_bufobj;
4393 	BO_LOCK(bo);
4394 	vn_syncer_add_to_worklist(bo, syncdelay);
4395 	BO_UNLOCK(bo);
4396 
4397 	/*
4398 	 * Walk the list of vnodes pushing all that are dirty and
4399 	 * not already on the sync list.
4400 	 */
4401 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4402 		return (0);
4403 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4404 		vfs_unbusy(mp);
4405 		return (0);
4406 	}
4407 	save = curthread_pflags_set(TDP_SYNCIO);
4408 	vfs_msync(mp, MNT_NOWAIT);
4409 	error = VFS_SYNC(mp, MNT_LAZY);
4410 	curthread_pflags_restore(save);
4411 	vn_finished_write(mp);
4412 	vfs_unbusy(mp);
4413 	return (error);
4414 }
4415 
4416 /*
4417  * The syncer vnode is no referenced.
4418  */
4419 static int
4420 sync_inactive(struct vop_inactive_args *ap)
4421 {
4422 
4423 	vgone(ap->a_vp);
4424 	return (0);
4425 }
4426 
4427 /*
4428  * The syncer vnode is no longer needed and is being decommissioned.
4429  *
4430  * Modifications to the worklist must be protected by sync_mtx.
4431  */
4432 static int
4433 sync_reclaim(struct vop_reclaim_args *ap)
4434 {
4435 	struct vnode *vp = ap->a_vp;
4436 	struct bufobj *bo;
4437 
4438 	bo = &vp->v_bufobj;
4439 	BO_LOCK(bo);
4440 	mtx_lock(&sync_mtx);
4441 	if (vp->v_mount->mnt_syncer == vp)
4442 		vp->v_mount->mnt_syncer = NULL;
4443 	if (bo->bo_flag & BO_ONWORKLST) {
4444 		LIST_REMOVE(bo, bo_synclist);
4445 		syncer_worklist_len--;
4446 		sync_vnode_count--;
4447 		bo->bo_flag &= ~BO_ONWORKLST;
4448 	}
4449 	mtx_unlock(&sync_mtx);
4450 	BO_UNLOCK(bo);
4451 
4452 	return (0);
4453 }
4454 
4455 /*
4456  * Check if vnode represents a disk device
4457  */
4458 int
4459 vn_isdisk(struct vnode *vp, int *errp)
4460 {
4461 	int error;
4462 
4463 	if (vp->v_type != VCHR) {
4464 		error = ENOTBLK;
4465 		goto out;
4466 	}
4467 	error = 0;
4468 	dev_lock();
4469 	if (vp->v_rdev == NULL)
4470 		error = ENXIO;
4471 	else if (vp->v_rdev->si_devsw == NULL)
4472 		error = ENXIO;
4473 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
4474 		error = ENOTBLK;
4475 	dev_unlock();
4476 out:
4477 	if (errp != NULL)
4478 		*errp = error;
4479 	return (error == 0);
4480 }
4481 
4482 /*
4483  * Common filesystem object access control check routine.  Accepts a
4484  * vnode's type, "mode", uid and gid, requested access mode, credentials,
4485  * and optional call-by-reference privused argument allowing vaccess()
4486  * to indicate to the caller whether privilege was used to satisfy the
4487  * request (obsoleted).  Returns 0 on success, or an errno on failure.
4488  */
4489 int
4490 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
4491     accmode_t accmode, struct ucred *cred, int *privused)
4492 {
4493 	accmode_t dac_granted;
4494 	accmode_t priv_granted;
4495 
4496 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
4497 	    ("invalid bit in accmode"));
4498 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
4499 	    ("VAPPEND without VWRITE"));
4500 
4501 	/*
4502 	 * Look for a normal, non-privileged way to access the file/directory
4503 	 * as requested.  If it exists, go with that.
4504 	 */
4505 
4506 	if (privused != NULL)
4507 		*privused = 0;
4508 
4509 	dac_granted = 0;
4510 
4511 	/* Check the owner. */
4512 	if (cred->cr_uid == file_uid) {
4513 		dac_granted |= VADMIN;
4514 		if (file_mode & S_IXUSR)
4515 			dac_granted |= VEXEC;
4516 		if (file_mode & S_IRUSR)
4517 			dac_granted |= VREAD;
4518 		if (file_mode & S_IWUSR)
4519 			dac_granted |= (VWRITE | VAPPEND);
4520 
4521 		if ((accmode & dac_granted) == accmode)
4522 			return (0);
4523 
4524 		goto privcheck;
4525 	}
4526 
4527 	/* Otherwise, check the groups (first match) */
4528 	if (groupmember(file_gid, cred)) {
4529 		if (file_mode & S_IXGRP)
4530 			dac_granted |= VEXEC;
4531 		if (file_mode & S_IRGRP)
4532 			dac_granted |= VREAD;
4533 		if (file_mode & S_IWGRP)
4534 			dac_granted |= (VWRITE | VAPPEND);
4535 
4536 		if ((accmode & dac_granted) == accmode)
4537 			return (0);
4538 
4539 		goto privcheck;
4540 	}
4541 
4542 	/* Otherwise, check everyone else. */
4543 	if (file_mode & S_IXOTH)
4544 		dac_granted |= VEXEC;
4545 	if (file_mode & S_IROTH)
4546 		dac_granted |= VREAD;
4547 	if (file_mode & S_IWOTH)
4548 		dac_granted |= (VWRITE | VAPPEND);
4549 	if ((accmode & dac_granted) == accmode)
4550 		return (0);
4551 
4552 privcheck:
4553 	/*
4554 	 * Build a privilege mask to determine if the set of privileges
4555 	 * satisfies the requirements when combined with the granted mask
4556 	 * from above.  For each privilege, if the privilege is required,
4557 	 * bitwise or the request type onto the priv_granted mask.
4558 	 */
4559 	priv_granted = 0;
4560 
4561 	if (type == VDIR) {
4562 		/*
4563 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
4564 		 * requests, instead of PRIV_VFS_EXEC.
4565 		 */
4566 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4567 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
4568 			priv_granted |= VEXEC;
4569 	} else {
4570 		/*
4571 		 * Ensure that at least one execute bit is on. Otherwise,
4572 		 * a privileged user will always succeed, and we don't want
4573 		 * this to happen unless the file really is executable.
4574 		 */
4575 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4576 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
4577 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
4578 			priv_granted |= VEXEC;
4579 	}
4580 
4581 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
4582 	    !priv_check_cred(cred, PRIV_VFS_READ))
4583 		priv_granted |= VREAD;
4584 
4585 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
4586 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
4587 		priv_granted |= (VWRITE | VAPPEND);
4588 
4589 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
4590 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
4591 		priv_granted |= VADMIN;
4592 
4593 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
4594 		/* XXX audit: privilege used */
4595 		if (privused != NULL)
4596 			*privused = 1;
4597 		return (0);
4598 	}
4599 
4600 	return ((accmode & VADMIN) ? EPERM : EACCES);
4601 }
4602 
4603 /*
4604  * Credential check based on process requesting service, and per-attribute
4605  * permissions.
4606  */
4607 int
4608 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
4609     struct thread *td, accmode_t accmode)
4610 {
4611 
4612 	/*
4613 	 * Kernel-invoked always succeeds.
4614 	 */
4615 	if (cred == NOCRED)
4616 		return (0);
4617 
4618 	/*
4619 	 * Do not allow privileged processes in jail to directly manipulate
4620 	 * system attributes.
4621 	 */
4622 	switch (attrnamespace) {
4623 	case EXTATTR_NAMESPACE_SYSTEM:
4624 		/* Potentially should be: return (EPERM); */
4625 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
4626 	case EXTATTR_NAMESPACE_USER:
4627 		return (VOP_ACCESS(vp, accmode, cred, td));
4628 	default:
4629 		return (EPERM);
4630 	}
4631 }
4632 
4633 #ifdef DEBUG_VFS_LOCKS
4634 /*
4635  * This only exists to suppress warnings from unlocked specfs accesses.  It is
4636  * no longer ok to have an unlocked VFS.
4637  */
4638 #define	IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL ||		\
4639 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
4640 
4641 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
4642 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
4643     "Drop into debugger on lock violation");
4644 
4645 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
4646 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
4647     0, "Check for interlock across VOPs");
4648 
4649 int vfs_badlock_print = 1;	/* Print lock violations. */
4650 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
4651     0, "Print lock violations");
4652 
4653 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
4654 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
4655     0, "Print vnode details on lock violations");
4656 
4657 #ifdef KDB
4658 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
4659 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
4660     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
4661 #endif
4662 
4663 static void
4664 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
4665 {
4666 
4667 #ifdef KDB
4668 	if (vfs_badlock_backtrace)
4669 		kdb_backtrace();
4670 #endif
4671 	if (vfs_badlock_vnode)
4672 		vn_printf(vp, "vnode ");
4673 	if (vfs_badlock_print)
4674 		printf("%s: %p %s\n", str, (void *)vp, msg);
4675 	if (vfs_badlock_ddb)
4676 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4677 }
4678 
4679 void
4680 assert_vi_locked(struct vnode *vp, const char *str)
4681 {
4682 
4683 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
4684 		vfs_badlock("interlock is not locked but should be", str, vp);
4685 }
4686 
4687 void
4688 assert_vi_unlocked(struct vnode *vp, const char *str)
4689 {
4690 
4691 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
4692 		vfs_badlock("interlock is locked but should not be", str, vp);
4693 }
4694 
4695 void
4696 assert_vop_locked(struct vnode *vp, const char *str)
4697 {
4698 	int locked;
4699 
4700 	if (!IGNORE_LOCK(vp)) {
4701 		locked = VOP_ISLOCKED(vp);
4702 		if (locked == 0 || locked == LK_EXCLOTHER)
4703 			vfs_badlock("is not locked but should be", str, vp);
4704 	}
4705 }
4706 
4707 void
4708 assert_vop_unlocked(struct vnode *vp, const char *str)
4709 {
4710 
4711 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
4712 		vfs_badlock("is locked but should not be", str, vp);
4713 }
4714 
4715 void
4716 assert_vop_elocked(struct vnode *vp, const char *str)
4717 {
4718 
4719 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
4720 		vfs_badlock("is not exclusive locked but should be", str, vp);
4721 }
4722 #endif /* DEBUG_VFS_LOCKS */
4723 
4724 void
4725 vop_rename_fail(struct vop_rename_args *ap)
4726 {
4727 
4728 	if (ap->a_tvp != NULL)
4729 		vput(ap->a_tvp);
4730 	if (ap->a_tdvp == ap->a_tvp)
4731 		vrele(ap->a_tdvp);
4732 	else
4733 		vput(ap->a_tdvp);
4734 	vrele(ap->a_fdvp);
4735 	vrele(ap->a_fvp);
4736 }
4737 
4738 void
4739 vop_rename_pre(void *ap)
4740 {
4741 	struct vop_rename_args *a = ap;
4742 
4743 #ifdef DEBUG_VFS_LOCKS
4744 	if (a->a_tvp)
4745 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
4746 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
4747 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
4748 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
4749 
4750 	/* Check the source (from). */
4751 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
4752 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
4753 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
4754 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
4755 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
4756 
4757 	/* Check the target. */
4758 	if (a->a_tvp)
4759 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
4760 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
4761 #endif
4762 	if (a->a_tdvp != a->a_fdvp)
4763 		vhold(a->a_fdvp);
4764 	if (a->a_tvp != a->a_fvp)
4765 		vhold(a->a_fvp);
4766 	vhold(a->a_tdvp);
4767 	if (a->a_tvp)
4768 		vhold(a->a_tvp);
4769 }
4770 
4771 #ifdef DEBUG_VFS_LOCKS
4772 void
4773 vop_strategy_pre(void *ap)
4774 {
4775 	struct vop_strategy_args *a;
4776 	struct buf *bp;
4777 
4778 	a = ap;
4779 	bp = a->a_bp;
4780 
4781 	/*
4782 	 * Cluster ops lock their component buffers but not the IO container.
4783 	 */
4784 	if ((bp->b_flags & B_CLUSTER) != 0)
4785 		return;
4786 
4787 	if (panicstr == NULL && !BUF_ISLOCKED(bp)) {
4788 		if (vfs_badlock_print)
4789 			printf(
4790 			    "VOP_STRATEGY: bp is not locked but should be\n");
4791 		if (vfs_badlock_ddb)
4792 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4793 	}
4794 }
4795 
4796 void
4797 vop_lock_pre(void *ap)
4798 {
4799 	struct vop_lock1_args *a = ap;
4800 
4801 	if ((a->a_flags & LK_INTERLOCK) == 0)
4802 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4803 	else
4804 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
4805 }
4806 
4807 void
4808 vop_lock_post(void *ap, int rc)
4809 {
4810 	struct vop_lock1_args *a = ap;
4811 
4812 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4813 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
4814 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
4815 }
4816 
4817 void
4818 vop_unlock_pre(void *ap)
4819 {
4820 	struct vop_unlock_args *a = ap;
4821 
4822 	if (a->a_flags & LK_INTERLOCK)
4823 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
4824 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
4825 }
4826 
4827 void
4828 vop_unlock_post(void *ap, int rc)
4829 {
4830 	struct vop_unlock_args *a = ap;
4831 
4832 	if (a->a_flags & LK_INTERLOCK)
4833 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
4834 }
4835 #endif
4836 
4837 void
4838 vop_create_post(void *ap, int rc)
4839 {
4840 	struct vop_create_args *a = ap;
4841 
4842 	if (!rc)
4843 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4844 }
4845 
4846 void
4847 vop_deleteextattr_post(void *ap, int rc)
4848 {
4849 	struct vop_deleteextattr_args *a = ap;
4850 
4851 	if (!rc)
4852 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4853 }
4854 
4855 void
4856 vop_link_post(void *ap, int rc)
4857 {
4858 	struct vop_link_args *a = ap;
4859 
4860 	if (!rc) {
4861 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
4862 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
4863 	}
4864 }
4865 
4866 void
4867 vop_mkdir_post(void *ap, int rc)
4868 {
4869 	struct vop_mkdir_args *a = ap;
4870 
4871 	if (!rc)
4872 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4873 }
4874 
4875 void
4876 vop_mknod_post(void *ap, int rc)
4877 {
4878 	struct vop_mknod_args *a = ap;
4879 
4880 	if (!rc)
4881 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4882 }
4883 
4884 void
4885 vop_reclaim_post(void *ap, int rc)
4886 {
4887 	struct vop_reclaim_args *a = ap;
4888 
4889 	if (!rc)
4890 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
4891 }
4892 
4893 void
4894 vop_remove_post(void *ap, int rc)
4895 {
4896 	struct vop_remove_args *a = ap;
4897 
4898 	if (!rc) {
4899 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4900 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4901 	}
4902 }
4903 
4904 void
4905 vop_rename_post(void *ap, int rc)
4906 {
4907 	struct vop_rename_args *a = ap;
4908 	long hint;
4909 
4910 	if (!rc) {
4911 		hint = NOTE_WRITE;
4912 		if (a->a_fdvp == a->a_tdvp) {
4913 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
4914 				hint |= NOTE_LINK;
4915 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
4916 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
4917 		} else {
4918 			hint |= NOTE_EXTEND;
4919 			if (a->a_fvp->v_type == VDIR)
4920 				hint |= NOTE_LINK;
4921 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
4922 
4923 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
4924 			    a->a_tvp->v_type == VDIR)
4925 				hint &= ~NOTE_LINK;
4926 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
4927 		}
4928 
4929 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
4930 		if (a->a_tvp)
4931 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
4932 	}
4933 	if (a->a_tdvp != a->a_fdvp)
4934 		vdrop(a->a_fdvp);
4935 	if (a->a_tvp != a->a_fvp)
4936 		vdrop(a->a_fvp);
4937 	vdrop(a->a_tdvp);
4938 	if (a->a_tvp)
4939 		vdrop(a->a_tvp);
4940 }
4941 
4942 void
4943 vop_rmdir_post(void *ap, int rc)
4944 {
4945 	struct vop_rmdir_args *a = ap;
4946 
4947 	if (!rc) {
4948 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4949 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4950 	}
4951 }
4952 
4953 void
4954 vop_setattr_post(void *ap, int rc)
4955 {
4956 	struct vop_setattr_args *a = ap;
4957 
4958 	if (!rc)
4959 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4960 }
4961 
4962 void
4963 vop_setextattr_post(void *ap, int rc)
4964 {
4965 	struct vop_setextattr_args *a = ap;
4966 
4967 	if (!rc)
4968 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4969 }
4970 
4971 void
4972 vop_symlink_post(void *ap, int rc)
4973 {
4974 	struct vop_symlink_args *a = ap;
4975 
4976 	if (!rc)
4977 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4978 }
4979 
4980 void
4981 vop_open_post(void *ap, int rc)
4982 {
4983 	struct vop_open_args *a = ap;
4984 
4985 	if (!rc)
4986 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
4987 }
4988 
4989 void
4990 vop_close_post(void *ap, int rc)
4991 {
4992 	struct vop_close_args *a = ap;
4993 
4994 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
4995 	    (a->a_vp->v_iflag & VI_DOOMED) == 0)) {
4996 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
4997 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
4998 	}
4999 }
5000 
5001 void
5002 vop_read_post(void *ap, int rc)
5003 {
5004 	struct vop_read_args *a = ap;
5005 
5006 	if (!rc)
5007 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5008 }
5009 
5010 void
5011 vop_readdir_post(void *ap, int rc)
5012 {
5013 	struct vop_readdir_args *a = ap;
5014 
5015 	if (!rc)
5016 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
5017 }
5018 
5019 static struct knlist fs_knlist;
5020 
5021 static void
5022 vfs_event_init(void *arg)
5023 {
5024 	knlist_init_mtx(&fs_knlist, NULL);
5025 }
5026 /* XXX - correct order? */
5027 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
5028 
5029 void
5030 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
5031 {
5032 
5033 	KNOTE_UNLOCKED(&fs_knlist, event);
5034 }
5035 
5036 static int	filt_fsattach(struct knote *kn);
5037 static void	filt_fsdetach(struct knote *kn);
5038 static int	filt_fsevent(struct knote *kn, long hint);
5039 
5040 struct filterops fs_filtops = {
5041 	.f_isfd = 0,
5042 	.f_attach = filt_fsattach,
5043 	.f_detach = filt_fsdetach,
5044 	.f_event = filt_fsevent
5045 };
5046 
5047 static int
5048 filt_fsattach(struct knote *kn)
5049 {
5050 
5051 	kn->kn_flags |= EV_CLEAR;
5052 	knlist_add(&fs_knlist, kn, 0);
5053 	return (0);
5054 }
5055 
5056 static void
5057 filt_fsdetach(struct knote *kn)
5058 {
5059 
5060 	knlist_remove(&fs_knlist, kn, 0);
5061 }
5062 
5063 static int
5064 filt_fsevent(struct knote *kn, long hint)
5065 {
5066 
5067 	kn->kn_fflags |= hint;
5068 	return (kn->kn_fflags != 0);
5069 }
5070 
5071 static int
5072 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
5073 {
5074 	struct vfsidctl vc;
5075 	int error;
5076 	struct mount *mp;
5077 
5078 	error = SYSCTL_IN(req, &vc, sizeof(vc));
5079 	if (error)
5080 		return (error);
5081 	if (vc.vc_vers != VFS_CTL_VERS1)
5082 		return (EINVAL);
5083 	mp = vfs_getvfs(&vc.vc_fsid);
5084 	if (mp == NULL)
5085 		return (ENOENT);
5086 	/* ensure that a specific sysctl goes to the right filesystem. */
5087 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
5088 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
5089 		vfs_rel(mp);
5090 		return (EINVAL);
5091 	}
5092 	VCTLTOREQ(&vc, req);
5093 	error = VFS_SYSCTL(mp, vc.vc_op, req);
5094 	vfs_rel(mp);
5095 	return (error);
5096 }
5097 
5098 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR,
5099     NULL, 0, sysctl_vfs_ctl, "",
5100     "Sysctl by fsid");
5101 
5102 /*
5103  * Function to initialize a va_filerev field sensibly.
5104  * XXX: Wouldn't a random number make a lot more sense ??
5105  */
5106 u_quad_t
5107 init_va_filerev(void)
5108 {
5109 	struct bintime bt;
5110 
5111 	getbinuptime(&bt);
5112 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5113 }
5114 
5115 static int	filt_vfsread(struct knote *kn, long hint);
5116 static int	filt_vfswrite(struct knote *kn, long hint);
5117 static int	filt_vfsvnode(struct knote *kn, long hint);
5118 static void	filt_vfsdetach(struct knote *kn);
5119 static struct filterops vfsread_filtops = {
5120 	.f_isfd = 1,
5121 	.f_detach = filt_vfsdetach,
5122 	.f_event = filt_vfsread
5123 };
5124 static struct filterops vfswrite_filtops = {
5125 	.f_isfd = 1,
5126 	.f_detach = filt_vfsdetach,
5127 	.f_event = filt_vfswrite
5128 };
5129 static struct filterops vfsvnode_filtops = {
5130 	.f_isfd = 1,
5131 	.f_detach = filt_vfsdetach,
5132 	.f_event = filt_vfsvnode
5133 };
5134 
5135 static void
5136 vfs_knllock(void *arg)
5137 {
5138 	struct vnode *vp = arg;
5139 
5140 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5141 }
5142 
5143 static void
5144 vfs_knlunlock(void *arg)
5145 {
5146 	struct vnode *vp = arg;
5147 
5148 	VOP_UNLOCK(vp, 0);
5149 }
5150 
5151 static void
5152 vfs_knl_assert_locked(void *arg)
5153 {
5154 #ifdef DEBUG_VFS_LOCKS
5155 	struct vnode *vp = arg;
5156 
5157 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
5158 #endif
5159 }
5160 
5161 static void
5162 vfs_knl_assert_unlocked(void *arg)
5163 {
5164 #ifdef DEBUG_VFS_LOCKS
5165 	struct vnode *vp = arg;
5166 
5167 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
5168 #endif
5169 }
5170 
5171 int
5172 vfs_kqfilter(struct vop_kqfilter_args *ap)
5173 {
5174 	struct vnode *vp = ap->a_vp;
5175 	struct knote *kn = ap->a_kn;
5176 	struct knlist *knl;
5177 
5178 	switch (kn->kn_filter) {
5179 	case EVFILT_READ:
5180 		kn->kn_fop = &vfsread_filtops;
5181 		break;
5182 	case EVFILT_WRITE:
5183 		kn->kn_fop = &vfswrite_filtops;
5184 		break;
5185 	case EVFILT_VNODE:
5186 		kn->kn_fop = &vfsvnode_filtops;
5187 		break;
5188 	default:
5189 		return (EINVAL);
5190 	}
5191 
5192 	kn->kn_hook = (caddr_t)vp;
5193 
5194 	v_addpollinfo(vp);
5195 	if (vp->v_pollinfo == NULL)
5196 		return (ENOMEM);
5197 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
5198 	vhold(vp);
5199 	knlist_add(knl, kn, 0);
5200 
5201 	return (0);
5202 }
5203 
5204 /*
5205  * Detach knote from vnode
5206  */
5207 static void
5208 filt_vfsdetach(struct knote *kn)
5209 {
5210 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5211 
5212 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5213 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5214 	vdrop(vp);
5215 }
5216 
5217 /*ARGSUSED*/
5218 static int
5219 filt_vfsread(struct knote *kn, long hint)
5220 {
5221 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5222 	struct vattr va;
5223 	int res;
5224 
5225 	/*
5226 	 * filesystem is gone, so set the EOF flag and schedule
5227 	 * the knote for deletion.
5228 	 */
5229 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5230 		VI_LOCK(vp);
5231 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5232 		VI_UNLOCK(vp);
5233 		return (1);
5234 	}
5235 
5236 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5237 		return (0);
5238 
5239 	VI_LOCK(vp);
5240 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5241 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5242 	VI_UNLOCK(vp);
5243 	return (res);
5244 }
5245 
5246 /*ARGSUSED*/
5247 static int
5248 filt_vfswrite(struct knote *kn, long hint)
5249 {
5250 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5251 
5252 	VI_LOCK(vp);
5253 
5254 	/*
5255 	 * filesystem is gone, so set the EOF flag and schedule
5256 	 * the knote for deletion.
5257 	 */
5258 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5259 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5260 
5261 	kn->kn_data = 0;
5262 	VI_UNLOCK(vp);
5263 	return (1);
5264 }
5265 
5266 static int
5267 filt_vfsvnode(struct knote *kn, long hint)
5268 {
5269 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5270 	int res;
5271 
5272 	VI_LOCK(vp);
5273 	if (kn->kn_sfflags & hint)
5274 		kn->kn_fflags |= hint;
5275 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5276 		kn->kn_flags |= EV_EOF;
5277 		VI_UNLOCK(vp);
5278 		return (1);
5279 	}
5280 	res = (kn->kn_fflags != 0);
5281 	VI_UNLOCK(vp);
5282 	return (res);
5283 }
5284 
5285 int
5286 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5287 {
5288 	int error;
5289 
5290 	if (dp->d_reclen > ap->a_uio->uio_resid)
5291 		return (ENAMETOOLONG);
5292 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
5293 	if (error) {
5294 		if (ap->a_ncookies != NULL) {
5295 			if (ap->a_cookies != NULL)
5296 				free(ap->a_cookies, M_TEMP);
5297 			ap->a_cookies = NULL;
5298 			*ap->a_ncookies = 0;
5299 		}
5300 		return (error);
5301 	}
5302 	if (ap->a_ncookies == NULL)
5303 		return (0);
5304 
5305 	KASSERT(ap->a_cookies,
5306 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
5307 
5308 	*ap->a_cookies = realloc(*ap->a_cookies,
5309 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
5310 	(*ap->a_cookies)[*ap->a_ncookies] = off;
5311 	*ap->a_ncookies += 1;
5312 	return (0);
5313 }
5314 
5315 /*
5316  * Mark for update the access time of the file if the filesystem
5317  * supports VOP_MARKATIME.  This functionality is used by execve and
5318  * mmap, so we want to avoid the I/O implied by directly setting
5319  * va_atime for the sake of efficiency.
5320  */
5321 void
5322 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
5323 {
5324 	struct mount *mp;
5325 
5326 	mp = vp->v_mount;
5327 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
5328 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
5329 		(void)VOP_MARKATIME(vp);
5330 }
5331 
5332 /*
5333  * The purpose of this routine is to remove granularity from accmode_t,
5334  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
5335  * VADMIN and VAPPEND.
5336  *
5337  * If it returns 0, the caller is supposed to continue with the usual
5338  * access checks using 'accmode' as modified by this routine.  If it
5339  * returns nonzero value, the caller is supposed to return that value
5340  * as errno.
5341  *
5342  * Note that after this routine runs, accmode may be zero.
5343  */
5344 int
5345 vfs_unixify_accmode(accmode_t *accmode)
5346 {
5347 	/*
5348 	 * There is no way to specify explicit "deny" rule using
5349 	 * file mode or POSIX.1e ACLs.
5350 	 */
5351 	if (*accmode & VEXPLICIT_DENY) {
5352 		*accmode = 0;
5353 		return (0);
5354 	}
5355 
5356 	/*
5357 	 * None of these can be translated into usual access bits.
5358 	 * Also, the common case for NFSv4 ACLs is to not contain
5359 	 * either of these bits. Caller should check for VWRITE
5360 	 * on the containing directory instead.
5361 	 */
5362 	if (*accmode & (VDELETE_CHILD | VDELETE))
5363 		return (EPERM);
5364 
5365 	if (*accmode & VADMIN_PERMS) {
5366 		*accmode &= ~VADMIN_PERMS;
5367 		*accmode |= VADMIN;
5368 	}
5369 
5370 	/*
5371 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5372 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5373 	 */
5374 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5375 
5376 	return (0);
5377 }
5378 
5379 /*
5380  * These are helper functions for filesystems to traverse all
5381  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
5382  *
5383  * This interface replaces MNT_VNODE_FOREACH.
5384  */
5385 
5386 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
5387 
5388 struct vnode *
5389 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
5390 {
5391 	struct vnode *vp;
5392 
5393 	if (should_yield())
5394 		kern_yield(PRI_USER);
5395 	MNT_ILOCK(mp);
5396 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5397 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
5398 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
5399 		/* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
5400 		if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
5401 			continue;
5402 		VI_LOCK(vp);
5403 		if ((vp->v_iflag & VI_DOOMED) != 0) {
5404 			VI_UNLOCK(vp);
5405 			continue;
5406 		}
5407 		break;
5408 	}
5409 	if (vp == NULL) {
5410 		__mnt_vnode_markerfree_all(mvp, mp);
5411 		/* MNT_IUNLOCK(mp); -- done in above function */
5412 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
5413 		return (NULL);
5414 	}
5415 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5416 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5417 	MNT_IUNLOCK(mp);
5418 	return (vp);
5419 }
5420 
5421 struct vnode *
5422 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
5423 {
5424 	struct vnode *vp;
5425 
5426 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5427 	MNT_ILOCK(mp);
5428 	MNT_REF(mp);
5429 	(*mvp)->v_mount = mp;
5430 	(*mvp)->v_type = VMARKER;
5431 
5432 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
5433 		/* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
5434 		if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
5435 			continue;
5436 		VI_LOCK(vp);
5437 		if ((vp->v_iflag & VI_DOOMED) != 0) {
5438 			VI_UNLOCK(vp);
5439 			continue;
5440 		}
5441 		break;
5442 	}
5443 	if (vp == NULL) {
5444 		MNT_REL(mp);
5445 		MNT_IUNLOCK(mp);
5446 		free(*mvp, M_VNODE_MARKER);
5447 		*mvp = NULL;
5448 		return (NULL);
5449 	}
5450 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5451 	MNT_IUNLOCK(mp);
5452 	return (vp);
5453 }
5454 
5455 void
5456 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
5457 {
5458 
5459 	if (*mvp == NULL) {
5460 		MNT_IUNLOCK(mp);
5461 		return;
5462 	}
5463 
5464 	mtx_assert(MNT_MTX(mp), MA_OWNED);
5465 
5466 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5467 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5468 	MNT_REL(mp);
5469 	MNT_IUNLOCK(mp);
5470 	free(*mvp, M_VNODE_MARKER);
5471 	*mvp = NULL;
5472 }
5473 
5474 /*
5475  * These are helper functions for filesystems to traverse their
5476  * active vnodes.  See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h
5477  */
5478 static void
5479 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5480 {
5481 
5482 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5483 
5484 	MNT_ILOCK(mp);
5485 	MNT_REL(mp);
5486 	MNT_IUNLOCK(mp);
5487 	free(*mvp, M_VNODE_MARKER);
5488 	*mvp = NULL;
5489 }
5490 
5491 /*
5492  * Relock the mp mount vnode list lock with the vp vnode interlock in the
5493  * conventional lock order during mnt_vnode_next_active iteration.
5494  *
5495  * On entry, the mount vnode list lock is held and the vnode interlock is not.
5496  * The list lock is dropped and reacquired.  On success, both locks are held.
5497  * On failure, the mount vnode list lock is held but the vnode interlock is
5498  * not, and the procedure may have yielded.
5499  */
5500 static bool
5501 mnt_vnode_next_active_relock(struct vnode *mvp, struct mount *mp,
5502     struct vnode *vp)
5503 {
5504 	const struct vnode *tmp;
5505 	bool held, ret;
5506 
5507 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
5508 	    TAILQ_NEXT(mvp, v_actfreelist) != NULL, mvp,
5509 	    ("%s: bad marker", __func__));
5510 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
5511 	    ("%s: inappropriate vnode", __func__));
5512 	ASSERT_VI_UNLOCKED(vp, __func__);
5513 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5514 
5515 	ret = false;
5516 
5517 	TAILQ_REMOVE(&mp->mnt_activevnodelist, mvp, v_actfreelist);
5518 	TAILQ_INSERT_BEFORE(vp, mvp, v_actfreelist);
5519 
5520 	/*
5521 	 * Use a hold to prevent vp from disappearing while the mount vnode
5522 	 * list lock is dropped and reacquired.  Normally a hold would be
5523 	 * acquired with vhold(), but that might try to acquire the vnode
5524 	 * interlock, which would be a LOR with the mount vnode list lock.
5525 	 */
5526 	held = refcount_acquire_if_not_zero(&vp->v_holdcnt);
5527 	mtx_unlock(&mp->mnt_listmtx);
5528 	if (!held)
5529 		goto abort;
5530 	VI_LOCK(vp);
5531 	if (!refcount_release_if_not_last(&vp->v_holdcnt)) {
5532 		vdropl(vp);
5533 		goto abort;
5534 	}
5535 	mtx_lock(&mp->mnt_listmtx);
5536 
5537 	/*
5538 	 * Determine whether the vnode is still the next one after the marker,
5539 	 * excepting any other markers.  If the vnode has not been doomed by
5540 	 * vgone() then the hold should have ensured that it remained on the
5541 	 * active list.  If it has been doomed but is still on the active list,
5542 	 * don't abort, but rather skip over it (avoid spinning on doomed
5543 	 * vnodes).
5544 	 */
5545 	tmp = mvp;
5546 	do {
5547 		tmp = TAILQ_NEXT(tmp, v_actfreelist);
5548 	} while (tmp != NULL && tmp->v_type == VMARKER);
5549 	if (tmp != vp) {
5550 		mtx_unlock(&mp->mnt_listmtx);
5551 		VI_UNLOCK(vp);
5552 		goto abort;
5553 	}
5554 
5555 	ret = true;
5556 	goto out;
5557 abort:
5558 	maybe_yield();
5559 	mtx_lock(&mp->mnt_listmtx);
5560 out:
5561 	if (ret)
5562 		ASSERT_VI_LOCKED(vp, __func__);
5563 	else
5564 		ASSERT_VI_UNLOCKED(vp, __func__);
5565 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5566 	return (ret);
5567 }
5568 
5569 static struct vnode *
5570 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5571 {
5572 	struct vnode *vp, *nvp;
5573 
5574 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5575 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5576 restart:
5577 	vp = TAILQ_NEXT(*mvp, v_actfreelist);
5578 	while (vp != NULL) {
5579 		if (vp->v_type == VMARKER) {
5580 			vp = TAILQ_NEXT(vp, v_actfreelist);
5581 			continue;
5582 		}
5583 		/*
5584 		 * Try-lock because this is the wrong lock order.  If that does
5585 		 * not succeed, drop the mount vnode list lock and try to
5586 		 * reacquire it and the vnode interlock in the right order.
5587 		 */
5588 		if (!VI_TRYLOCK(vp) &&
5589 		    !mnt_vnode_next_active_relock(*mvp, mp, vp))
5590 			goto restart;
5591 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
5592 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
5593 		    ("alien vnode on the active list %p %p", vp, mp));
5594 		if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0)
5595 			break;
5596 		nvp = TAILQ_NEXT(vp, v_actfreelist);
5597 		VI_UNLOCK(vp);
5598 		vp = nvp;
5599 	}
5600 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5601 
5602 	/* Check if we are done */
5603 	if (vp == NULL) {
5604 		mtx_unlock(&mp->mnt_listmtx);
5605 		mnt_vnode_markerfree_active(mvp, mp);
5606 		return (NULL);
5607 	}
5608 	TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist);
5609 	mtx_unlock(&mp->mnt_listmtx);
5610 	ASSERT_VI_LOCKED(vp, "active iter");
5611 	KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp));
5612 	return (vp);
5613 }
5614 
5615 struct vnode *
5616 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5617 {
5618 
5619 	if (should_yield())
5620 		kern_yield(PRI_USER);
5621 	mtx_lock(&mp->mnt_listmtx);
5622 	return (mnt_vnode_next_active(mvp, mp));
5623 }
5624 
5625 struct vnode *
5626 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp)
5627 {
5628 	struct vnode *vp;
5629 
5630 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5631 	MNT_ILOCK(mp);
5632 	MNT_REF(mp);
5633 	MNT_IUNLOCK(mp);
5634 	(*mvp)->v_type = VMARKER;
5635 	(*mvp)->v_mount = mp;
5636 
5637 	mtx_lock(&mp->mnt_listmtx);
5638 	vp = TAILQ_FIRST(&mp->mnt_activevnodelist);
5639 	if (vp == NULL) {
5640 		mtx_unlock(&mp->mnt_listmtx);
5641 		mnt_vnode_markerfree_active(mvp, mp);
5642 		return (NULL);
5643 	}
5644 	TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5645 	return (mnt_vnode_next_active(mvp, mp));
5646 }
5647 
5648 void
5649 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5650 {
5651 
5652 	if (*mvp == NULL)
5653 		return;
5654 
5655 	mtx_lock(&mp->mnt_listmtx);
5656 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5657 	mtx_unlock(&mp->mnt_listmtx);
5658 	mnt_vnode_markerfree_active(mvp, mp);
5659 }
5660