xref: /freebsd/sys/kern/vfs_subr.c (revision b027d6545b60535c1c01508bc6f1254b2a2c3b94)
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/condvar.h>
54 #include <sys/conf.h>
55 #include <sys/counter.h>
56 #include <sys/dirent.h>
57 #include <sys/event.h>
58 #include <sys/eventhandler.h>
59 #include <sys/extattr.h>
60 #include <sys/file.h>
61 #include <sys/fcntl.h>
62 #include <sys/jail.h>
63 #include <sys/kdb.h>
64 #include <sys/kernel.h>
65 #include <sys/kthread.h>
66 #include <sys/lockf.h>
67 #include <sys/malloc.h>
68 #include <sys/mount.h>
69 #include <sys/namei.h>
70 #include <sys/pctrie.h>
71 #include <sys/priv.h>
72 #include <sys/reboot.h>
73 #include <sys/refcount.h>
74 #include <sys/rwlock.h>
75 #include <sys/sched.h>
76 #include <sys/sleepqueue.h>
77 #include <sys/smp.h>
78 #include <sys/stat.h>
79 #include <sys/sysctl.h>
80 #include <sys/syslog.h>
81 #include <sys/vmmeter.h>
82 #include <sys/vnode.h>
83 #include <sys/watchdog.h>
84 
85 #include <machine/stdarg.h>
86 
87 #include <security/mac/mac_framework.h>
88 
89 #include <vm/vm.h>
90 #include <vm/vm_object.h>
91 #include <vm/vm_extern.h>
92 #include <vm/pmap.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_page.h>
95 #include <vm/vm_kern.h>
96 #include <vm/uma.h>
97 
98 #ifdef DDB
99 #include <ddb/ddb.h>
100 #endif
101 
102 static void	delmntque(struct vnode *vp);
103 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
104 		    int slpflag, int slptimeo);
105 static void	syncer_shutdown(void *arg, int howto);
106 static int	vtryrecycle(struct vnode *vp);
107 static void	v_init_counters(struct vnode *);
108 static void	v_incr_usecount(struct vnode *);
109 static void	v_incr_usecount_locked(struct vnode *);
110 static void	v_incr_devcount(struct vnode *);
111 static void	v_decr_devcount(struct vnode *);
112 static void	vgonel(struct vnode *);
113 static void	vfs_knllock(void *arg);
114 static void	vfs_knlunlock(void *arg);
115 static void	vfs_knl_assert_locked(void *arg);
116 static void	vfs_knl_assert_unlocked(void *arg);
117 static void	vnlru_return_batches(struct vfsops *mnt_op);
118 static void	destroy_vpollinfo(struct vpollinfo *vi);
119 
120 /*
121  * Number of vnodes in existence.  Increased whenever getnewvnode()
122  * allocates a new vnode, decreased in vdropl() for VI_DOOMED vnode.
123  */
124 static unsigned long	numvnodes;
125 
126 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
127     "Number of vnodes in existence");
128 
129 static counter_u64_t vnodes_created;
130 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
131     "Number of vnodes created by getnewvnode");
132 
133 static u_long mnt_free_list_batch = 128;
134 SYSCTL_ULONG(_vfs, OID_AUTO, mnt_free_list_batch, CTLFLAG_RW,
135     &mnt_free_list_batch, 0, "Limit of vnodes held on mnt's free list");
136 
137 /*
138  * Conversion tables for conversion from vnode types to inode formats
139  * and back.
140  */
141 enum vtype iftovt_tab[16] = {
142 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
143 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
144 };
145 int vttoif_tab[10] = {
146 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
147 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
148 };
149 
150 /*
151  * List of vnodes that are ready for recycling.
152  */
153 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
154 
155 /*
156  * "Free" vnode target.  Free vnodes are rarely completely free, but are
157  * just ones that are cheap to recycle.  Usually they are for files which
158  * have been stat'd but not read; these usually have inode and namecache
159  * data attached to them.  This target is the preferred minimum size of a
160  * sub-cache consisting mostly of such files. The system balances the size
161  * of this sub-cache with its complement to try to prevent either from
162  * thrashing while the other is relatively inactive.  The targets express
163  * a preference for the best balance.
164  *
165  * "Above" this target there are 2 further targets (watermarks) related
166  * to recyling of free vnodes.  In the best-operating case, the cache is
167  * exactly full, the free list has size between vlowat and vhiwat above the
168  * free target, and recycling from it and normal use maintains this state.
169  * Sometimes the free list is below vlowat or even empty, but this state
170  * is even better for immediate use provided the cache is not full.
171  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
172  * ones) to reach one of these states.  The watermarks are currently hard-
173  * coded as 4% and 9% of the available space higher.  These and the default
174  * of 25% for wantfreevnodes are too large if the memory size is large.
175  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
176  * whenever vnlru_proc() becomes active.
177  */
178 static u_long wantfreevnodes;
179 SYSCTL_ULONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW,
180     &wantfreevnodes, 0, "Target for minimum number of \"free\" vnodes");
181 static u_long freevnodes;
182 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
183     &freevnodes, 0, "Number of \"free\" vnodes");
184 
185 static counter_u64_t recycles_count;
186 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
187     "Number of vnodes recycled to meet vnode cache targets");
188 
189 /*
190  * Various variables used for debugging the new implementation of
191  * reassignbuf().
192  * XXX these are probably of (very) limited utility now.
193  */
194 static int reassignbufcalls;
195 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0,
196     "Number of calls to reassignbuf");
197 
198 static counter_u64_t free_owe_inact;
199 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, free_owe_inact, CTLFLAG_RD, &free_owe_inact,
200     "Number of times free vnodes kept on active list due to VFS "
201     "owing inactivation");
202 
203 /* To keep more than one thread at a time from running vfs_getnewfsid */
204 static struct mtx mntid_mtx;
205 
206 /*
207  * Lock for any access to the following:
208  *	vnode_free_list
209  *	numvnodes
210  *	freevnodes
211  */
212 static struct mtx vnode_free_list_mtx;
213 
214 /* Publicly exported FS */
215 struct nfs_public nfs_pub;
216 
217 static uma_zone_t buf_trie_zone;
218 
219 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
220 static uma_zone_t vnode_zone;
221 static uma_zone_t vnodepoll_zone;
222 
223 /*
224  * The workitem queue.
225  *
226  * It is useful to delay writes of file data and filesystem metadata
227  * for tens of seconds so that quickly created and deleted files need
228  * not waste disk bandwidth being created and removed. To realize this,
229  * we append vnodes to a "workitem" queue. When running with a soft
230  * updates implementation, most pending metadata dependencies should
231  * not wait for more than a few seconds. Thus, mounted on block devices
232  * are delayed only about a half the time that file data is delayed.
233  * Similarly, directory updates are more critical, so are only delayed
234  * about a third the time that file data is delayed. Thus, there are
235  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
236  * one each second (driven off the filesystem syncer process). The
237  * syncer_delayno variable indicates the next queue that is to be processed.
238  * Items that need to be processed soon are placed in this queue:
239  *
240  *	syncer_workitem_pending[syncer_delayno]
241  *
242  * A delay of fifteen seconds is done by placing the request fifteen
243  * entries later in the queue:
244  *
245  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
246  *
247  */
248 static int syncer_delayno;
249 static long syncer_mask;
250 LIST_HEAD(synclist, bufobj);
251 static struct synclist *syncer_workitem_pending;
252 /*
253  * The sync_mtx protects:
254  *	bo->bo_synclist
255  *	sync_vnode_count
256  *	syncer_delayno
257  *	syncer_state
258  *	syncer_workitem_pending
259  *	syncer_worklist_len
260  *	rushjob
261  */
262 static struct mtx sync_mtx;
263 static struct cv sync_wakeup;
264 
265 #define SYNCER_MAXDELAY		32
266 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
267 static int syncdelay = 30;		/* max time to delay syncing data */
268 static int filedelay = 30;		/* time to delay syncing files */
269 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
270     "Time to delay syncing files (in seconds)");
271 static int dirdelay = 29;		/* time to delay syncing directories */
272 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
273     "Time to delay syncing directories (in seconds)");
274 static int metadelay = 28;		/* time to delay syncing metadata */
275 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
276     "Time to delay syncing metadata (in seconds)");
277 static int rushjob;		/* number of slots to run ASAP */
278 static int stat_rush_requests;	/* number of times I/O speeded up */
279 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
280     "Number of times I/O speeded up (rush requests)");
281 
282 /*
283  * When shutting down the syncer, run it at four times normal speed.
284  */
285 #define SYNCER_SHUTDOWN_SPEEDUP		4
286 static int sync_vnode_count;
287 static int syncer_worklist_len;
288 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
289     syncer_state;
290 
291 /* Target for maximum number of vnodes. */
292 int desiredvnodes;
293 static int gapvnodes;		/* gap between wanted and desired */
294 static int vhiwat;		/* enough extras after expansion */
295 static int vlowat;		/* minimal extras before expansion */
296 static int vstir;		/* nonzero to stir non-free vnodes */
297 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
298 
299 static int
300 sysctl_update_desiredvnodes(SYSCTL_HANDLER_ARGS)
301 {
302 	int error, old_desiredvnodes;
303 
304 	old_desiredvnodes = desiredvnodes;
305 	if ((error = sysctl_handle_int(oidp, arg1, arg2, req)) != 0)
306 		return (error);
307 	if (old_desiredvnodes != desiredvnodes) {
308 		wantfreevnodes = desiredvnodes / 4;
309 		/* XXX locking seems to be incomplete. */
310 		vfs_hash_changesize(desiredvnodes);
311 		cache_changesize(desiredvnodes);
312 	}
313 	return (0);
314 }
315 
316 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
317     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, &desiredvnodes, 0,
318     sysctl_update_desiredvnodes, "I", "Target for maximum number of vnodes");
319 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
320     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
321 static int vnlru_nowhere;
322 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
323     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
324 
325 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
326 static int vnsz2log;
327 
328 /*
329  * Support for the bufobj clean & dirty pctrie.
330  */
331 static void *
332 buf_trie_alloc(struct pctrie *ptree)
333 {
334 
335 	return uma_zalloc(buf_trie_zone, M_NOWAIT);
336 }
337 
338 static void
339 buf_trie_free(struct pctrie *ptree, void *node)
340 {
341 
342 	uma_zfree(buf_trie_zone, node);
343 }
344 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
345 
346 /*
347  * Initialize the vnode management data structures.
348  *
349  * Reevaluate the following cap on the number of vnodes after the physical
350  * memory size exceeds 512GB.  In the limit, as the physical memory size
351  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
352  */
353 #ifndef	MAXVNODES_MAX
354 #define	MAXVNODES_MAX	(512 * 1024 * 1024 / 64)	/* 8M */
355 #endif
356 
357 /*
358  * Initialize a vnode as it first enters the zone.
359  */
360 static int
361 vnode_init(void *mem, int size, int flags)
362 {
363 	struct vnode *vp;
364 
365 	vp = mem;
366 	bzero(vp, size);
367 	/*
368 	 * Setup locks.
369 	 */
370 	vp->v_vnlock = &vp->v_lock;
371 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
372 	/*
373 	 * By default, don't allow shared locks unless filesystems opt-in.
374 	 */
375 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
376 	    LK_NOSHARE | LK_IS_VNODE);
377 	/*
378 	 * Initialize bufobj.
379 	 */
380 	bufobj_init(&vp->v_bufobj, vp);
381 	/*
382 	 * Initialize namecache.
383 	 */
384 	LIST_INIT(&vp->v_cache_src);
385 	TAILQ_INIT(&vp->v_cache_dst);
386 	/*
387 	 * Initialize rangelocks.
388 	 */
389 	rangelock_init(&vp->v_rl);
390 	return (0);
391 }
392 
393 /*
394  * Free a vnode when it is cleared from the zone.
395  */
396 static void
397 vnode_fini(void *mem, int size)
398 {
399 	struct vnode *vp;
400 	struct bufobj *bo;
401 
402 	vp = mem;
403 	rangelock_destroy(&vp->v_rl);
404 	lockdestroy(vp->v_vnlock);
405 	mtx_destroy(&vp->v_interlock);
406 	bo = &vp->v_bufobj;
407 	rw_destroy(BO_LOCKPTR(bo));
408 }
409 
410 /*
411  * Provide the size of NFS nclnode and NFS fh for calculation of the
412  * vnode memory consumption.  The size is specified directly to
413  * eliminate dependency on NFS-private header.
414  *
415  * Other filesystems may use bigger or smaller (like UFS and ZFS)
416  * private inode data, but the NFS-based estimation is ample enough.
417  * Still, we care about differences in the size between 64- and 32-bit
418  * platforms.
419  *
420  * Namecache structure size is heuristically
421  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
422  */
423 #ifdef _LP64
424 #define	NFS_NCLNODE_SZ	(528 + 64)
425 #define	NC_SZ		148
426 #else
427 #define	NFS_NCLNODE_SZ	(360 + 32)
428 #define	NC_SZ		92
429 #endif
430 
431 static void
432 vntblinit(void *dummy __unused)
433 {
434 	u_int i;
435 	int physvnodes, virtvnodes;
436 
437 	/*
438 	 * Desiredvnodes is a function of the physical memory size and the
439 	 * kernel's heap size.  Generally speaking, it scales with the
440 	 * physical memory size.  The ratio of desiredvnodes to the physical
441 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
442 	 * Thereafter, the
443 	 * marginal ratio of desiredvnodes to the physical memory size is
444 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
445 	 * size.  The memory required by desiredvnodes vnodes and vm objects
446 	 * must not exceed 1/10th of the kernel's heap size.
447 	 */
448 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
449 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
450 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
451 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
452 	desiredvnodes = min(physvnodes, virtvnodes);
453 	if (desiredvnodes > MAXVNODES_MAX) {
454 		if (bootverbose)
455 			printf("Reducing kern.maxvnodes %d -> %d\n",
456 			    desiredvnodes, MAXVNODES_MAX);
457 		desiredvnodes = MAXVNODES_MAX;
458 	}
459 	wantfreevnodes = desiredvnodes / 4;
460 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
461 	TAILQ_INIT(&vnode_free_list);
462 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
463 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
464 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
465 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
466 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
467 	/*
468 	 * Preallocate enough nodes to support one-per buf so that
469 	 * we can not fail an insert.  reassignbuf() callers can not
470 	 * tolerate the insertion failure.
471 	 */
472 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
473 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
474 	    UMA_ZONE_NOFREE | UMA_ZONE_VM);
475 	uma_prealloc(buf_trie_zone, nbuf);
476 
477 	vnodes_created = counter_u64_alloc(M_WAITOK);
478 	recycles_count = counter_u64_alloc(M_WAITOK);
479 	free_owe_inact = counter_u64_alloc(M_WAITOK);
480 
481 	/*
482 	 * Initialize the filesystem syncer.
483 	 */
484 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
485 	    &syncer_mask);
486 	syncer_maxdelay = syncer_mask + 1;
487 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
488 	cv_init(&sync_wakeup, "syncer");
489 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
490 		vnsz2log++;
491 	vnsz2log--;
492 }
493 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
494 
495 
496 /*
497  * Mark a mount point as busy. Used to synchronize access and to delay
498  * unmounting. Eventually, mountlist_mtx is not released on failure.
499  *
500  * vfs_busy() is a custom lock, it can block the caller.
501  * vfs_busy() only sleeps if the unmount is active on the mount point.
502  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
503  * vnode belonging to mp.
504  *
505  * Lookup uses vfs_busy() to traverse mount points.
506  * root fs			var fs
507  * / vnode lock		A	/ vnode lock (/var)		D
508  * /var vnode lock	B	/log vnode lock(/var/log)	E
509  * vfs_busy lock	C	vfs_busy lock			F
510  *
511  * Within each file system, the lock order is C->A->B and F->D->E.
512  *
513  * When traversing across mounts, the system follows that lock order:
514  *
515  *        C->A->B
516  *              |
517  *              +->F->D->E
518  *
519  * The lookup() process for namei("/var") illustrates the process:
520  *  VOP_LOOKUP() obtains B while A is held
521  *  vfs_busy() obtains a shared lock on F while A and B are held
522  *  vput() releases lock on B
523  *  vput() releases lock on A
524  *  VFS_ROOT() obtains lock on D while shared lock on F is held
525  *  vfs_unbusy() releases shared lock on F
526  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
527  *    Attempt to lock A (instead of vp_crossmp) while D is held would
528  *    violate the global order, causing deadlocks.
529  *
530  * dounmount() locks B while F is drained.
531  */
532 int
533 vfs_busy(struct mount *mp, int flags)
534 {
535 
536 	MPASS((flags & ~MBF_MASK) == 0);
537 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
538 
539 	MNT_ILOCK(mp);
540 	MNT_REF(mp);
541 	/*
542 	 * If mount point is currently being unmounted, sleep until the
543 	 * mount point fate is decided.  If thread doing the unmounting fails,
544 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
545 	 * that this mount point has survived the unmount attempt and vfs_busy
546 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
547 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
548 	 * about to be really destroyed.  vfs_busy needs to release its
549 	 * reference on the mount point in this case and return with ENOENT,
550 	 * telling the caller that mount mount it tried to busy is no longer
551 	 * valid.
552 	 */
553 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
554 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
555 			MNT_REL(mp);
556 			MNT_IUNLOCK(mp);
557 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
558 			    __func__);
559 			return (ENOENT);
560 		}
561 		if (flags & MBF_MNTLSTLOCK)
562 			mtx_unlock(&mountlist_mtx);
563 		mp->mnt_kern_flag |= MNTK_MWAIT;
564 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
565 		if (flags & MBF_MNTLSTLOCK)
566 			mtx_lock(&mountlist_mtx);
567 		MNT_ILOCK(mp);
568 	}
569 	if (flags & MBF_MNTLSTLOCK)
570 		mtx_unlock(&mountlist_mtx);
571 	mp->mnt_lockref++;
572 	MNT_IUNLOCK(mp);
573 	return (0);
574 }
575 
576 /*
577  * Free a busy filesystem.
578  */
579 void
580 vfs_unbusy(struct mount *mp)
581 {
582 
583 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
584 	MNT_ILOCK(mp);
585 	MNT_REL(mp);
586 	KASSERT(mp->mnt_lockref > 0, ("negative mnt_lockref"));
587 	mp->mnt_lockref--;
588 	if (mp->mnt_lockref == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
589 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
590 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
591 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
592 		wakeup(&mp->mnt_lockref);
593 	}
594 	MNT_IUNLOCK(mp);
595 }
596 
597 /*
598  * Lookup a mount point by filesystem identifier.
599  */
600 struct mount *
601 vfs_getvfs(fsid_t *fsid)
602 {
603 	struct mount *mp;
604 
605 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
606 	mtx_lock(&mountlist_mtx);
607 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
608 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
609 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
610 			vfs_ref(mp);
611 			mtx_unlock(&mountlist_mtx);
612 			return (mp);
613 		}
614 	}
615 	mtx_unlock(&mountlist_mtx);
616 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
617 	return ((struct mount *) 0);
618 }
619 
620 /*
621  * Lookup a mount point by filesystem identifier, busying it before
622  * returning.
623  *
624  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
625  * cache for popular filesystem identifiers.  The cache is lockess, using
626  * the fact that struct mount's are never freed.  In worst case we may
627  * get pointer to unmounted or even different filesystem, so we have to
628  * check what we got, and go slow way if so.
629  */
630 struct mount *
631 vfs_busyfs(fsid_t *fsid)
632 {
633 #define	FSID_CACHE_SIZE	256
634 	typedef struct mount * volatile vmp_t;
635 	static vmp_t cache[FSID_CACHE_SIZE];
636 	struct mount *mp;
637 	int error;
638 	uint32_t hash;
639 
640 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
641 	hash = fsid->val[0] ^ fsid->val[1];
642 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
643 	mp = cache[hash];
644 	if (mp == NULL ||
645 	    mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
646 	    mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
647 		goto slow;
648 	if (vfs_busy(mp, 0) != 0) {
649 		cache[hash] = NULL;
650 		goto slow;
651 	}
652 	if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
653 	    mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
654 		return (mp);
655 	else
656 	    vfs_unbusy(mp);
657 
658 slow:
659 	mtx_lock(&mountlist_mtx);
660 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
661 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
662 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
663 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
664 			if (error) {
665 				cache[hash] = NULL;
666 				mtx_unlock(&mountlist_mtx);
667 				return (NULL);
668 			}
669 			cache[hash] = mp;
670 			return (mp);
671 		}
672 	}
673 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
674 	mtx_unlock(&mountlist_mtx);
675 	return ((struct mount *) 0);
676 }
677 
678 /*
679  * Check if a user can access privileged mount options.
680  */
681 int
682 vfs_suser(struct mount *mp, struct thread *td)
683 {
684 	int error;
685 
686 	if (jailed(td->td_ucred)) {
687 		/*
688 		 * If the jail of the calling thread lacks permission for
689 		 * this type of file system, deny immediately.
690 		 */
691 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
692 			return (EPERM);
693 
694 		/*
695 		 * If the file system was mounted outside the jail of the
696 		 * calling thread, deny immediately.
697 		 */
698 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
699 			return (EPERM);
700 	}
701 
702 	/*
703 	 * If file system supports delegated administration, we don't check
704 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
705 	 * by the file system itself.
706 	 * If this is not the user that did original mount, we check for
707 	 * the PRIV_VFS_MOUNT_OWNER privilege.
708 	 */
709 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
710 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
711 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
712 			return (error);
713 	}
714 	return (0);
715 }
716 
717 /*
718  * Get a new unique fsid.  Try to make its val[0] unique, since this value
719  * will be used to create fake device numbers for stat().  Also try (but
720  * not so hard) make its val[0] unique mod 2^16, since some emulators only
721  * support 16-bit device numbers.  We end up with unique val[0]'s for the
722  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
723  *
724  * Keep in mind that several mounts may be running in parallel.  Starting
725  * the search one past where the previous search terminated is both a
726  * micro-optimization and a defense against returning the same fsid to
727  * different mounts.
728  */
729 void
730 vfs_getnewfsid(struct mount *mp)
731 {
732 	static uint16_t mntid_base;
733 	struct mount *nmp;
734 	fsid_t tfsid;
735 	int mtype;
736 
737 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
738 	mtx_lock(&mntid_mtx);
739 	mtype = mp->mnt_vfc->vfc_typenum;
740 	tfsid.val[1] = mtype;
741 	mtype = (mtype & 0xFF) << 24;
742 	for (;;) {
743 		tfsid.val[0] = makedev(255,
744 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
745 		mntid_base++;
746 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
747 			break;
748 		vfs_rel(nmp);
749 	}
750 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
751 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
752 	mtx_unlock(&mntid_mtx);
753 }
754 
755 /*
756  * Knob to control the precision of file timestamps:
757  *
758  *   0 = seconds only; nanoseconds zeroed.
759  *   1 = seconds and nanoseconds, accurate within 1/HZ.
760  *   2 = seconds and nanoseconds, truncated to microseconds.
761  * >=3 = seconds and nanoseconds, maximum precision.
762  */
763 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
764 
765 static int timestamp_precision = TSP_USEC;
766 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
767     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
768     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
769     "3+: sec + ns (max. precision))");
770 
771 /*
772  * Get a current timestamp.
773  */
774 void
775 vfs_timestamp(struct timespec *tsp)
776 {
777 	struct timeval tv;
778 
779 	switch (timestamp_precision) {
780 	case TSP_SEC:
781 		tsp->tv_sec = time_second;
782 		tsp->tv_nsec = 0;
783 		break;
784 	case TSP_HZ:
785 		getnanotime(tsp);
786 		break;
787 	case TSP_USEC:
788 		microtime(&tv);
789 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
790 		break;
791 	case TSP_NSEC:
792 	default:
793 		nanotime(tsp);
794 		break;
795 	}
796 }
797 
798 /*
799  * Set vnode attributes to VNOVAL
800  */
801 void
802 vattr_null(struct vattr *vap)
803 {
804 
805 	vap->va_type = VNON;
806 	vap->va_size = VNOVAL;
807 	vap->va_bytes = VNOVAL;
808 	vap->va_mode = VNOVAL;
809 	vap->va_nlink = VNOVAL;
810 	vap->va_uid = VNOVAL;
811 	vap->va_gid = VNOVAL;
812 	vap->va_fsid = VNOVAL;
813 	vap->va_fileid = VNOVAL;
814 	vap->va_blocksize = VNOVAL;
815 	vap->va_rdev = VNOVAL;
816 	vap->va_atime.tv_sec = VNOVAL;
817 	vap->va_atime.tv_nsec = VNOVAL;
818 	vap->va_mtime.tv_sec = VNOVAL;
819 	vap->va_mtime.tv_nsec = VNOVAL;
820 	vap->va_ctime.tv_sec = VNOVAL;
821 	vap->va_ctime.tv_nsec = VNOVAL;
822 	vap->va_birthtime.tv_sec = VNOVAL;
823 	vap->va_birthtime.tv_nsec = VNOVAL;
824 	vap->va_flags = VNOVAL;
825 	vap->va_gen = VNOVAL;
826 	vap->va_vaflags = 0;
827 }
828 
829 /*
830  * This routine is called when we have too many vnodes.  It attempts
831  * to free <count> vnodes and will potentially free vnodes that still
832  * have VM backing store (VM backing store is typically the cause
833  * of a vnode blowout so we want to do this).  Therefore, this operation
834  * is not considered cheap.
835  *
836  * A number of conditions may prevent a vnode from being reclaimed.
837  * the buffer cache may have references on the vnode, a directory
838  * vnode may still have references due to the namei cache representing
839  * underlying files, or the vnode may be in active use.   It is not
840  * desirable to reuse such vnodes.  These conditions may cause the
841  * number of vnodes to reach some minimum value regardless of what
842  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
843  */
844 static int
845 vlrureclaim(struct mount *mp, int reclaim_nc_src, int trigger)
846 {
847 	struct vnode *vp;
848 	int count, done, target;
849 
850 	done = 0;
851 	vn_start_write(NULL, &mp, V_WAIT);
852 	MNT_ILOCK(mp);
853 	count = mp->mnt_nvnodelistsize;
854 	target = count * (int64_t)gapvnodes / imax(desiredvnodes, 1);
855 	target = target / 10 + 1;
856 	while (count != 0 && done < target) {
857 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
858 		while (vp != NULL && vp->v_type == VMARKER)
859 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
860 		if (vp == NULL)
861 			break;
862 		/*
863 		 * XXX LRU is completely broken for non-free vnodes.  First
864 		 * by calling here in mountpoint order, then by moving
865 		 * unselected vnodes to the end here, and most grossly by
866 		 * removing the vlruvp() function that was supposed to
867 		 * maintain the order.  (This function was born broken
868 		 * since syncer problems prevented it doing anything.)  The
869 		 * order is closer to LRC (C = Created).
870 		 *
871 		 * LRU reclaiming of vnodes seems to have last worked in
872 		 * FreeBSD-3 where LRU wasn't mentioned under any spelling.
873 		 * Then there was no hold count, and inactive vnodes were
874 		 * simply put on the free list in LRU order.  The separate
875 		 * lists also break LRU.  We prefer to reclaim from the
876 		 * free list for technical reasons.  This tends to thrash
877 		 * the free list to keep very unrecently used held vnodes.
878 		 * The problem is mitigated by keeping the free list large.
879 		 */
880 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
881 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
882 		--count;
883 		if (!VI_TRYLOCK(vp))
884 			goto next_iter;
885 		/*
886 		 * If it's been deconstructed already, it's still
887 		 * referenced, or it exceeds the trigger, skip it.
888 		 * Also skip free vnodes.  We are trying to make space
889 		 * to expand the free list, not reduce it.
890 		 */
891 		if (vp->v_usecount ||
892 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
893 		    ((vp->v_iflag & VI_FREE) != 0) ||
894 		    (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
895 		    vp->v_object->resident_page_count > trigger)) {
896 			VI_UNLOCK(vp);
897 			goto next_iter;
898 		}
899 		MNT_IUNLOCK(mp);
900 		vholdl(vp);
901 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) {
902 			vdrop(vp);
903 			goto next_iter_mntunlocked;
904 		}
905 		VI_LOCK(vp);
906 		/*
907 		 * v_usecount may have been bumped after VOP_LOCK() dropped
908 		 * the vnode interlock and before it was locked again.
909 		 *
910 		 * It is not necessary to recheck VI_DOOMED because it can
911 		 * only be set by another thread that holds both the vnode
912 		 * lock and vnode interlock.  If another thread has the
913 		 * vnode lock before we get to VOP_LOCK() and obtains the
914 		 * vnode interlock after VOP_LOCK() drops the vnode
915 		 * interlock, the other thread will be unable to drop the
916 		 * vnode lock before our VOP_LOCK() call fails.
917 		 */
918 		if (vp->v_usecount ||
919 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
920 		    (vp->v_iflag & VI_FREE) != 0 ||
921 		    (vp->v_object != NULL &&
922 		    vp->v_object->resident_page_count > trigger)) {
923 			VOP_UNLOCK(vp, LK_INTERLOCK);
924 			vdrop(vp);
925 			goto next_iter_mntunlocked;
926 		}
927 		KASSERT((vp->v_iflag & VI_DOOMED) == 0,
928 		    ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
929 		counter_u64_add(recycles_count, 1);
930 		vgonel(vp);
931 		VOP_UNLOCK(vp, 0);
932 		vdropl(vp);
933 		done++;
934 next_iter_mntunlocked:
935 		if (!should_yield())
936 			goto relock_mnt;
937 		goto yield;
938 next_iter:
939 		if (!should_yield())
940 			continue;
941 		MNT_IUNLOCK(mp);
942 yield:
943 		kern_yield(PRI_USER);
944 relock_mnt:
945 		MNT_ILOCK(mp);
946 	}
947 	MNT_IUNLOCK(mp);
948 	vn_finished_write(mp);
949 	return done;
950 }
951 
952 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
953 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
954     0,
955     "limit on vnode free requests per call to the vnlru_free routine");
956 
957 /*
958  * Attempt to reduce the free list by the requested amount.
959  */
960 static void
961 vnlru_free_locked(int count, struct vfsops *mnt_op)
962 {
963 	struct vnode *vp;
964 	struct mount *mp;
965 	bool tried_batches;
966 
967 	tried_batches = false;
968 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
969 	if (count > max_vnlru_free)
970 		count = max_vnlru_free;
971 	for (; count > 0; count--) {
972 		vp = TAILQ_FIRST(&vnode_free_list);
973 		/*
974 		 * The list can be modified while the free_list_mtx
975 		 * has been dropped and vp could be NULL here.
976 		 */
977 		if (vp == NULL) {
978 			if (tried_batches)
979 				break;
980 			mtx_unlock(&vnode_free_list_mtx);
981 			vnlru_return_batches(mnt_op);
982 			tried_batches = true;
983 			mtx_lock(&vnode_free_list_mtx);
984 			continue;
985 		}
986 
987 		VNASSERT(vp->v_op != NULL, vp,
988 		    ("vnlru_free: vnode already reclaimed."));
989 		KASSERT((vp->v_iflag & VI_FREE) != 0,
990 		    ("Removing vnode not on freelist"));
991 		KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
992 		    ("Mangling active vnode"));
993 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
994 
995 		/*
996 		 * Don't recycle if our vnode is from different type
997 		 * of mount point.  Note that mp is type-safe, the
998 		 * check does not reach unmapped address even if
999 		 * vnode is reclaimed.
1000 		 * Don't recycle if we can't get the interlock without
1001 		 * blocking.
1002 		 */
1003 		if ((mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1004 		    mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) {
1005 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist);
1006 			continue;
1007 		}
1008 		VNASSERT((vp->v_iflag & VI_FREE) != 0 && vp->v_holdcnt == 0,
1009 		    vp, ("vp inconsistent on freelist"));
1010 
1011 		/*
1012 		 * The clear of VI_FREE prevents activation of the
1013 		 * vnode.  There is no sense in putting the vnode on
1014 		 * the mount point active list, only to remove it
1015 		 * later during recycling.  Inline the relevant part
1016 		 * of vholdl(), to avoid triggering assertions or
1017 		 * activating.
1018 		 */
1019 		freevnodes--;
1020 		vp->v_iflag &= ~VI_FREE;
1021 		refcount_acquire(&vp->v_holdcnt);
1022 
1023 		mtx_unlock(&vnode_free_list_mtx);
1024 		VI_UNLOCK(vp);
1025 		vtryrecycle(vp);
1026 		/*
1027 		 * If the recycled succeeded this vdrop will actually free
1028 		 * the vnode.  If not it will simply place it back on
1029 		 * the free list.
1030 		 */
1031 		vdrop(vp);
1032 		mtx_lock(&vnode_free_list_mtx);
1033 	}
1034 }
1035 
1036 void
1037 vnlru_free(int count, struct vfsops *mnt_op)
1038 {
1039 
1040 	mtx_lock(&vnode_free_list_mtx);
1041 	vnlru_free_locked(count, mnt_op);
1042 	mtx_unlock(&vnode_free_list_mtx);
1043 }
1044 
1045 
1046 /* XXX some names and initialization are bad for limits and watermarks. */
1047 static int
1048 vspace(void)
1049 {
1050 	int space;
1051 
1052 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1053 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1054 	vlowat = vhiwat / 2;
1055 	if (numvnodes > desiredvnodes)
1056 		return (0);
1057 	space = desiredvnodes - numvnodes;
1058 	if (freevnodes > wantfreevnodes)
1059 		space += freevnodes - wantfreevnodes;
1060 	return (space);
1061 }
1062 
1063 static void
1064 vnlru_return_batch_locked(struct mount *mp)
1065 {
1066 	struct vnode *vp;
1067 
1068 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
1069 
1070 	if (mp->mnt_tmpfreevnodelistsize == 0)
1071 		return;
1072 
1073 	TAILQ_FOREACH(vp, &mp->mnt_tmpfreevnodelist, v_actfreelist) {
1074 		VNASSERT((vp->v_mflag & VMP_TMPMNTFREELIST) != 0, vp,
1075 		    ("vnode without VMP_TMPMNTFREELIST on mnt_tmpfreevnodelist"));
1076 		vp->v_mflag &= ~VMP_TMPMNTFREELIST;
1077 	}
1078 	mtx_lock(&vnode_free_list_mtx);
1079 	TAILQ_CONCAT(&vnode_free_list, &mp->mnt_tmpfreevnodelist, v_actfreelist);
1080 	freevnodes += mp->mnt_tmpfreevnodelistsize;
1081 	mtx_unlock(&vnode_free_list_mtx);
1082 	mp->mnt_tmpfreevnodelistsize = 0;
1083 }
1084 
1085 static void
1086 vnlru_return_batch(struct mount *mp)
1087 {
1088 
1089 	mtx_lock(&mp->mnt_listmtx);
1090 	vnlru_return_batch_locked(mp);
1091 	mtx_unlock(&mp->mnt_listmtx);
1092 }
1093 
1094 static void
1095 vnlru_return_batches(struct vfsops *mnt_op)
1096 {
1097 	struct mount *mp, *nmp;
1098 	bool need_unbusy;
1099 
1100 	mtx_lock(&mountlist_mtx);
1101 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1102 		need_unbusy = false;
1103 		if (mnt_op != NULL && mp->mnt_op != mnt_op)
1104 			goto next;
1105 		if (mp->mnt_tmpfreevnodelistsize == 0)
1106 			goto next;
1107 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK) == 0) {
1108 			vnlru_return_batch(mp);
1109 			need_unbusy = true;
1110 			mtx_lock(&mountlist_mtx);
1111 		}
1112 next:
1113 		nmp = TAILQ_NEXT(mp, mnt_list);
1114 		if (need_unbusy)
1115 			vfs_unbusy(mp);
1116 	}
1117 	mtx_unlock(&mountlist_mtx);
1118 }
1119 
1120 /*
1121  * Attempt to recycle vnodes in a context that is always safe to block.
1122  * Calling vlrurecycle() from the bowels of filesystem code has some
1123  * interesting deadlock problems.
1124  */
1125 static struct proc *vnlruproc;
1126 static int vnlruproc_sig;
1127 
1128 static void
1129 vnlru_proc(void)
1130 {
1131 	struct mount *mp, *nmp;
1132 	unsigned long onumvnodes;
1133 	int done, force, reclaim_nc_src, trigger, usevnodes;
1134 
1135 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1136 	    SHUTDOWN_PRI_FIRST);
1137 
1138 	force = 0;
1139 	for (;;) {
1140 		kproc_suspend_check(vnlruproc);
1141 		mtx_lock(&vnode_free_list_mtx);
1142 		/*
1143 		 * If numvnodes is too large (due to desiredvnodes being
1144 		 * adjusted using its sysctl, or emergency growth), first
1145 		 * try to reduce it by discarding from the free list.
1146 		 */
1147 		if (numvnodes > desiredvnodes)
1148 			vnlru_free_locked(numvnodes - desiredvnodes, NULL);
1149 		/*
1150 		 * Sleep if the vnode cache is in a good state.  This is
1151 		 * when it is not over-full and has space for about a 4%
1152 		 * or 9% expansion (by growing its size or inexcessively
1153 		 * reducing its free list).  Otherwise, try to reclaim
1154 		 * space for a 10% expansion.
1155 		 */
1156 		if (vstir && force == 0) {
1157 			force = 1;
1158 			vstir = 0;
1159 		}
1160 		if (vspace() >= vlowat && force == 0) {
1161 			vnlruproc_sig = 0;
1162 			wakeup(&vnlruproc_sig);
1163 			msleep(vnlruproc, &vnode_free_list_mtx,
1164 			    PVFS|PDROP, "vlruwt", hz);
1165 			continue;
1166 		}
1167 		mtx_unlock(&vnode_free_list_mtx);
1168 		done = 0;
1169 		onumvnodes = numvnodes;
1170 		/*
1171 		 * Calculate parameters for recycling.  These are the same
1172 		 * throughout the loop to give some semblance of fairness.
1173 		 * The trigger point is to avoid recycling vnodes with lots
1174 		 * of resident pages.  We aren't trying to free memory; we
1175 		 * are trying to recycle or at least free vnodes.
1176 		 */
1177 		if (numvnodes <= desiredvnodes)
1178 			usevnodes = numvnodes - freevnodes;
1179 		else
1180 			usevnodes = numvnodes;
1181 		if (usevnodes <= 0)
1182 			usevnodes = 1;
1183 		/*
1184 		 * The trigger value is is chosen to give a conservatively
1185 		 * large value to ensure that it alone doesn't prevent
1186 		 * making progress.  The value can easily be so large that
1187 		 * it is effectively infinite in some congested and
1188 		 * misconfigured cases, and this is necessary.  Normally
1189 		 * it is about 8 to 100 (pages), which is quite large.
1190 		 */
1191 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1192 		if (force < 2)
1193 			trigger = vsmalltrigger;
1194 		reclaim_nc_src = force >= 3;
1195 		mtx_lock(&mountlist_mtx);
1196 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1197 			if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) {
1198 				nmp = TAILQ_NEXT(mp, mnt_list);
1199 				continue;
1200 			}
1201 			done += vlrureclaim(mp, reclaim_nc_src, trigger);
1202 			mtx_lock(&mountlist_mtx);
1203 			nmp = TAILQ_NEXT(mp, mnt_list);
1204 			vfs_unbusy(mp);
1205 		}
1206 		mtx_unlock(&mountlist_mtx);
1207 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1208 			uma_reclaim();
1209 		if (done == 0) {
1210 			if (force == 0 || force == 1) {
1211 				force = 2;
1212 				continue;
1213 			}
1214 			if (force == 2) {
1215 				force = 3;
1216 				continue;
1217 			}
1218 			force = 0;
1219 			vnlru_nowhere++;
1220 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1221 		} else
1222 			kern_yield(PRI_USER);
1223 		/*
1224 		 * After becoming active to expand above low water, keep
1225 		 * active until above high water.
1226 		 */
1227 		force = vspace() < vhiwat;
1228 	}
1229 }
1230 
1231 static struct kproc_desc vnlru_kp = {
1232 	"vnlru",
1233 	vnlru_proc,
1234 	&vnlruproc
1235 };
1236 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1237     &vnlru_kp);
1238 
1239 /*
1240  * Routines having to do with the management of the vnode table.
1241  */
1242 
1243 /*
1244  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1245  * before we actually vgone().  This function must be called with the vnode
1246  * held to prevent the vnode from being returned to the free list midway
1247  * through vgone().
1248  */
1249 static int
1250 vtryrecycle(struct vnode *vp)
1251 {
1252 	struct mount *vnmp;
1253 
1254 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1255 	VNASSERT(vp->v_holdcnt, vp,
1256 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1257 	/*
1258 	 * This vnode may found and locked via some other list, if so we
1259 	 * can't recycle it yet.
1260 	 */
1261 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1262 		CTR2(KTR_VFS,
1263 		    "%s: impossible to recycle, vp %p lock is already held",
1264 		    __func__, vp);
1265 		return (EWOULDBLOCK);
1266 	}
1267 	/*
1268 	 * Don't recycle if its filesystem is being suspended.
1269 	 */
1270 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1271 		VOP_UNLOCK(vp, 0);
1272 		CTR2(KTR_VFS,
1273 		    "%s: impossible to recycle, cannot start the write for %p",
1274 		    __func__, vp);
1275 		return (EBUSY);
1276 	}
1277 	/*
1278 	 * If we got this far, we need to acquire the interlock and see if
1279 	 * anyone picked up this vnode from another list.  If not, we will
1280 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1281 	 * will skip over it.
1282 	 */
1283 	VI_LOCK(vp);
1284 	if (vp->v_usecount) {
1285 		VOP_UNLOCK(vp, LK_INTERLOCK);
1286 		vn_finished_write(vnmp);
1287 		CTR2(KTR_VFS,
1288 		    "%s: impossible to recycle, %p is already referenced",
1289 		    __func__, vp);
1290 		return (EBUSY);
1291 	}
1292 	if ((vp->v_iflag & VI_DOOMED) == 0) {
1293 		counter_u64_add(recycles_count, 1);
1294 		vgonel(vp);
1295 	}
1296 	VOP_UNLOCK(vp, LK_INTERLOCK);
1297 	vn_finished_write(vnmp);
1298 	return (0);
1299 }
1300 
1301 static void
1302 vcheckspace(void)
1303 {
1304 
1305 	if (vspace() < vlowat && vnlruproc_sig == 0) {
1306 		vnlruproc_sig = 1;
1307 		wakeup(vnlruproc);
1308 	}
1309 }
1310 
1311 /*
1312  * Wait if necessary for space for a new vnode.
1313  */
1314 static int
1315 getnewvnode_wait(int suspended)
1316 {
1317 
1318 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1319 	if (numvnodes >= desiredvnodes) {
1320 		if (suspended) {
1321 			/*
1322 			 * The file system is being suspended.  We cannot
1323 			 * risk a deadlock here, so allow allocation of
1324 			 * another vnode even if this would give too many.
1325 			 */
1326 			return (0);
1327 		}
1328 		if (vnlruproc_sig == 0) {
1329 			vnlruproc_sig = 1;	/* avoid unnecessary wakeups */
1330 			wakeup(vnlruproc);
1331 		}
1332 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
1333 		    "vlruwk", hz);
1334 	}
1335 	/* Post-adjust like the pre-adjust in getnewvnode(). */
1336 	if (numvnodes + 1 > desiredvnodes && freevnodes > 1)
1337 		vnlru_free_locked(1, NULL);
1338 	return (numvnodes >= desiredvnodes ? ENFILE : 0);
1339 }
1340 
1341 /*
1342  * This hack is fragile, and probably not needed any more now that the
1343  * watermark handling works.
1344  */
1345 void
1346 getnewvnode_reserve(u_int count)
1347 {
1348 	struct thread *td;
1349 
1350 	/* Pre-adjust like the pre-adjust in getnewvnode(), with any count. */
1351 	/* XXX no longer so quick, but this part is not racy. */
1352 	mtx_lock(&vnode_free_list_mtx);
1353 	if (numvnodes + count > desiredvnodes && freevnodes > wantfreevnodes)
1354 		vnlru_free_locked(ulmin(numvnodes + count - desiredvnodes,
1355 		    freevnodes - wantfreevnodes), NULL);
1356 	mtx_unlock(&vnode_free_list_mtx);
1357 
1358 	td = curthread;
1359 	/* First try to be quick and racy. */
1360 	if (atomic_fetchadd_long(&numvnodes, count) + count <= desiredvnodes) {
1361 		td->td_vp_reserv += count;
1362 		vcheckspace();	/* XXX no longer so quick, but more racy */
1363 		return;
1364 	} else
1365 		atomic_subtract_long(&numvnodes, count);
1366 
1367 	mtx_lock(&vnode_free_list_mtx);
1368 	while (count > 0) {
1369 		if (getnewvnode_wait(0) == 0) {
1370 			count--;
1371 			td->td_vp_reserv++;
1372 			atomic_add_long(&numvnodes, 1);
1373 		}
1374 	}
1375 	vcheckspace();
1376 	mtx_unlock(&vnode_free_list_mtx);
1377 }
1378 
1379 /*
1380  * This hack is fragile, especially if desiredvnodes or wantvnodes are
1381  * misconfgured or changed significantly.  Reducing desiredvnodes below
1382  * the reserved amount should cause bizarre behaviour like reducing it
1383  * below the number of active vnodes -- the system will try to reduce
1384  * numvnodes to match, but should fail, so the subtraction below should
1385  * not overflow.
1386  */
1387 void
1388 getnewvnode_drop_reserve(void)
1389 {
1390 	struct thread *td;
1391 
1392 	td = curthread;
1393 	atomic_subtract_long(&numvnodes, td->td_vp_reserv);
1394 	td->td_vp_reserv = 0;
1395 }
1396 
1397 /*
1398  * Return the next vnode from the free list.
1399  */
1400 int
1401 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1402     struct vnode **vpp)
1403 {
1404 	struct vnode *vp;
1405 	struct thread *td;
1406 	struct lock_object *lo;
1407 	static int cyclecount;
1408 	int error __unused;
1409 
1410 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1411 	vp = NULL;
1412 	td = curthread;
1413 	if (td->td_vp_reserv > 0) {
1414 		td->td_vp_reserv -= 1;
1415 		goto alloc;
1416 	}
1417 	mtx_lock(&vnode_free_list_mtx);
1418 	if (numvnodes < desiredvnodes)
1419 		cyclecount = 0;
1420 	else if (cyclecount++ >= freevnodes) {
1421 		cyclecount = 0;
1422 		vstir = 1;
1423 	}
1424 	/*
1425 	 * Grow the vnode cache if it will not be above its target max
1426 	 * after growing.  Otherwise, if the free list is nonempty, try
1427 	 * to reclaim 1 item from it before growing the cache (possibly
1428 	 * above its target max if the reclamation failed or is delayed).
1429 	 * Otherwise, wait for some space.  In all cases, schedule
1430 	 * vnlru_proc() if we are getting short of space.  The watermarks
1431 	 * should be chosen so that we never wait or even reclaim from
1432 	 * the free list to below its target minimum.
1433 	 */
1434 	if (numvnodes + 1 <= desiredvnodes)
1435 		;
1436 	else if (freevnodes > 0)
1437 		vnlru_free_locked(1, NULL);
1438 	else {
1439 		error = getnewvnode_wait(mp != NULL && (mp->mnt_kern_flag &
1440 		    MNTK_SUSPEND));
1441 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
1442 		if (error != 0) {
1443 			mtx_unlock(&vnode_free_list_mtx);
1444 			return (error);
1445 		}
1446 #endif
1447 	}
1448 	vcheckspace();
1449 	atomic_add_long(&numvnodes, 1);
1450 	mtx_unlock(&vnode_free_list_mtx);
1451 alloc:
1452 	counter_u64_add(vnodes_created, 1);
1453 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK);
1454 	/*
1455 	 * Locks are given the generic name "vnode" when created.
1456 	 * Follow the historic practice of using the filesystem
1457 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1458 	 *
1459 	 * Locks live in a witness group keyed on their name. Thus,
1460 	 * when a lock is renamed, it must also move from the witness
1461 	 * group of its old name to the witness group of its new name.
1462 	 *
1463 	 * The change only needs to be made when the vnode moves
1464 	 * from one filesystem type to another. We ensure that each
1465 	 * filesystem use a single static name pointer for its tag so
1466 	 * that we can compare pointers rather than doing a strcmp().
1467 	 */
1468 	lo = &vp->v_vnlock->lock_object;
1469 	if (lo->lo_name != tag) {
1470 		lo->lo_name = tag;
1471 		WITNESS_DESTROY(lo);
1472 		WITNESS_INIT(lo, tag);
1473 	}
1474 	/*
1475 	 * By default, don't allow shared locks unless filesystems opt-in.
1476 	 */
1477 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1478 	/*
1479 	 * Finalize various vnode identity bits.
1480 	 */
1481 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1482 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1483 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1484 	vp->v_type = VNON;
1485 	vp->v_tag = tag;
1486 	vp->v_op = vops;
1487 	v_init_counters(vp);
1488 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1489 #ifdef DIAGNOSTIC
1490 	if (mp == NULL && vops != &dead_vnodeops)
1491 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1492 #endif
1493 #ifdef MAC
1494 	mac_vnode_init(vp);
1495 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1496 		mac_vnode_associate_singlelabel(mp, vp);
1497 #endif
1498 	if (mp != NULL) {
1499 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1500 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1501 			vp->v_vflag |= VV_NOKNOTE;
1502 	}
1503 
1504 	/*
1505 	 * For the filesystems which do not use vfs_hash_insert(),
1506 	 * still initialize v_hash to have vfs_hash_index() useful.
1507 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1508 	 * its own hashing.
1509 	 */
1510 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1511 
1512 	*vpp = vp;
1513 	return (0);
1514 }
1515 
1516 /*
1517  * Delete from old mount point vnode list, if on one.
1518  */
1519 static void
1520 delmntque(struct vnode *vp)
1521 {
1522 	struct mount *mp;
1523 	int active;
1524 
1525 	mp = vp->v_mount;
1526 	if (mp == NULL)
1527 		return;
1528 	MNT_ILOCK(mp);
1529 	VI_LOCK(vp);
1530 	KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize,
1531 	    ("Active vnode list size %d > Vnode list size %d",
1532 	     mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize));
1533 	active = vp->v_iflag & VI_ACTIVE;
1534 	vp->v_iflag &= ~VI_ACTIVE;
1535 	if (active) {
1536 		mtx_lock(&mp->mnt_listmtx);
1537 		TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist);
1538 		mp->mnt_activevnodelistsize--;
1539 		mtx_unlock(&mp->mnt_listmtx);
1540 	}
1541 	vp->v_mount = NULL;
1542 	VI_UNLOCK(vp);
1543 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1544 		("bad mount point vnode list size"));
1545 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1546 	mp->mnt_nvnodelistsize--;
1547 	MNT_REL(mp);
1548 	MNT_IUNLOCK(mp);
1549 }
1550 
1551 static void
1552 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1553 {
1554 
1555 	vp->v_data = NULL;
1556 	vp->v_op = &dead_vnodeops;
1557 	vgone(vp);
1558 	vput(vp);
1559 }
1560 
1561 /*
1562  * Insert into list of vnodes for the new mount point, if available.
1563  */
1564 int
1565 insmntque1(struct vnode *vp, struct mount *mp,
1566 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1567 {
1568 
1569 	KASSERT(vp->v_mount == NULL,
1570 		("insmntque: vnode already on per mount vnode list"));
1571 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1572 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1573 
1574 	/*
1575 	 * We acquire the vnode interlock early to ensure that the
1576 	 * vnode cannot be recycled by another process releasing a
1577 	 * holdcnt on it before we get it on both the vnode list
1578 	 * and the active vnode list. The mount mutex protects only
1579 	 * manipulation of the vnode list and the vnode freelist
1580 	 * mutex protects only manipulation of the active vnode list.
1581 	 * Hence the need to hold the vnode interlock throughout.
1582 	 */
1583 	MNT_ILOCK(mp);
1584 	VI_LOCK(vp);
1585 	if (((mp->mnt_kern_flag & MNTK_NOINSMNTQ) != 0 &&
1586 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1587 	    mp->mnt_nvnodelistsize == 0)) &&
1588 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1589 		VI_UNLOCK(vp);
1590 		MNT_IUNLOCK(mp);
1591 		if (dtr != NULL)
1592 			dtr(vp, dtr_arg);
1593 		return (EBUSY);
1594 	}
1595 	vp->v_mount = mp;
1596 	MNT_REF(mp);
1597 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1598 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1599 		("neg mount point vnode list size"));
1600 	mp->mnt_nvnodelistsize++;
1601 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1602 	    ("Activating already active vnode"));
1603 	vp->v_iflag |= VI_ACTIVE;
1604 	mtx_lock(&mp->mnt_listmtx);
1605 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
1606 	mp->mnt_activevnodelistsize++;
1607 	mtx_unlock(&mp->mnt_listmtx);
1608 	VI_UNLOCK(vp);
1609 	MNT_IUNLOCK(mp);
1610 	return (0);
1611 }
1612 
1613 int
1614 insmntque(struct vnode *vp, struct mount *mp)
1615 {
1616 
1617 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1618 }
1619 
1620 /*
1621  * Flush out and invalidate all buffers associated with a bufobj
1622  * Called with the underlying object locked.
1623  */
1624 int
1625 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1626 {
1627 	int error;
1628 
1629 	BO_LOCK(bo);
1630 	if (flags & V_SAVE) {
1631 		error = bufobj_wwait(bo, slpflag, slptimeo);
1632 		if (error) {
1633 			BO_UNLOCK(bo);
1634 			return (error);
1635 		}
1636 		if (bo->bo_dirty.bv_cnt > 0) {
1637 			BO_UNLOCK(bo);
1638 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1639 				return (error);
1640 			/*
1641 			 * XXX We could save a lock/unlock if this was only
1642 			 * enabled under INVARIANTS
1643 			 */
1644 			BO_LOCK(bo);
1645 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1646 				panic("vinvalbuf: dirty bufs");
1647 		}
1648 	}
1649 	/*
1650 	 * If you alter this loop please notice that interlock is dropped and
1651 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1652 	 * no race conditions occur from this.
1653 	 */
1654 	do {
1655 		error = flushbuflist(&bo->bo_clean,
1656 		    flags, bo, slpflag, slptimeo);
1657 		if (error == 0 && !(flags & V_CLEANONLY))
1658 			error = flushbuflist(&bo->bo_dirty,
1659 			    flags, bo, slpflag, slptimeo);
1660 		if (error != 0 && error != EAGAIN) {
1661 			BO_UNLOCK(bo);
1662 			return (error);
1663 		}
1664 	} while (error != 0);
1665 
1666 	/*
1667 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1668 	 * have write I/O in-progress but if there is a VM object then the
1669 	 * VM object can also have read-I/O in-progress.
1670 	 */
1671 	do {
1672 		bufobj_wwait(bo, 0, 0);
1673 		if ((flags & V_VMIO) == 0) {
1674 			BO_UNLOCK(bo);
1675 			if (bo->bo_object != NULL) {
1676 				VM_OBJECT_WLOCK(bo->bo_object);
1677 				vm_object_pip_wait(bo->bo_object, "bovlbx");
1678 				VM_OBJECT_WUNLOCK(bo->bo_object);
1679 			}
1680 			BO_LOCK(bo);
1681 		}
1682 	} while (bo->bo_numoutput > 0);
1683 	BO_UNLOCK(bo);
1684 
1685 	/*
1686 	 * Destroy the copy in the VM cache, too.
1687 	 */
1688 	if (bo->bo_object != NULL &&
1689 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1690 		VM_OBJECT_WLOCK(bo->bo_object);
1691 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1692 		    OBJPR_CLEANONLY : 0);
1693 		VM_OBJECT_WUNLOCK(bo->bo_object);
1694 	}
1695 
1696 #ifdef INVARIANTS
1697 	BO_LOCK(bo);
1698 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1699 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1700 	    bo->bo_clean.bv_cnt > 0))
1701 		panic("vinvalbuf: flush failed");
1702 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1703 	    bo->bo_dirty.bv_cnt > 0)
1704 		panic("vinvalbuf: flush dirty failed");
1705 	BO_UNLOCK(bo);
1706 #endif
1707 	return (0);
1708 }
1709 
1710 /*
1711  * Flush out and invalidate all buffers associated with a vnode.
1712  * Called with the underlying object locked.
1713  */
1714 int
1715 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1716 {
1717 
1718 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1719 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1720 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1721 		return (0);
1722 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1723 }
1724 
1725 /*
1726  * Flush out buffers on the specified list.
1727  *
1728  */
1729 static int
1730 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1731     int slptimeo)
1732 {
1733 	struct buf *bp, *nbp;
1734 	int retval, error;
1735 	daddr_t lblkno;
1736 	b_xflags_t xflags;
1737 
1738 	ASSERT_BO_WLOCKED(bo);
1739 
1740 	retval = 0;
1741 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1742 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1743 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1744 			continue;
1745 		}
1746 		if (nbp != NULL) {
1747 			lblkno = nbp->b_lblkno;
1748 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
1749 		}
1750 		retval = EAGAIN;
1751 		error = BUF_TIMELOCK(bp,
1752 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
1753 		    "flushbuf", slpflag, slptimeo);
1754 		if (error) {
1755 			BO_LOCK(bo);
1756 			return (error != ENOLCK ? error : EAGAIN);
1757 		}
1758 		KASSERT(bp->b_bufobj == bo,
1759 		    ("bp %p wrong b_bufobj %p should be %p",
1760 		    bp, bp->b_bufobj, bo));
1761 		/*
1762 		 * XXX Since there are no node locks for NFS, I
1763 		 * believe there is a slight chance that a delayed
1764 		 * write will occur while sleeping just above, so
1765 		 * check for it.
1766 		 */
1767 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1768 		    (flags & V_SAVE)) {
1769 			bremfree(bp);
1770 			bp->b_flags |= B_ASYNC;
1771 			bwrite(bp);
1772 			BO_LOCK(bo);
1773 			return (EAGAIN);	/* XXX: why not loop ? */
1774 		}
1775 		bremfree(bp);
1776 		bp->b_flags |= (B_INVAL | B_RELBUF);
1777 		bp->b_flags &= ~B_ASYNC;
1778 		brelse(bp);
1779 		BO_LOCK(bo);
1780 		if (nbp == NULL)
1781 			break;
1782 		nbp = gbincore(bo, lblkno);
1783 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1784 		    != xflags)
1785 			break;			/* nbp invalid */
1786 	}
1787 	return (retval);
1788 }
1789 
1790 int
1791 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
1792 {
1793 	struct buf *bp;
1794 	int error;
1795 	daddr_t lblkno;
1796 
1797 	ASSERT_BO_LOCKED(bo);
1798 
1799 	for (lblkno = startn;;) {
1800 again:
1801 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
1802 		if (bp == NULL || bp->b_lblkno >= endn ||
1803 		    bp->b_lblkno < startn)
1804 			break;
1805 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
1806 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
1807 		if (error != 0) {
1808 			BO_RLOCK(bo);
1809 			if (error == ENOLCK)
1810 				goto again;
1811 			return (error);
1812 		}
1813 		KASSERT(bp->b_bufobj == bo,
1814 		    ("bp %p wrong b_bufobj %p should be %p",
1815 		    bp, bp->b_bufobj, bo));
1816 		lblkno = bp->b_lblkno + 1;
1817 		if ((bp->b_flags & B_MANAGED) == 0)
1818 			bremfree(bp);
1819 		bp->b_flags |= B_RELBUF;
1820 		/*
1821 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
1822 		 * pages backing each buffer in the range are unlikely to be
1823 		 * reused.  Dirty buffers will have the hint applied once
1824 		 * they've been written.
1825 		 */
1826 		if (bp->b_vp->v_object != NULL)
1827 			bp->b_flags |= B_NOREUSE;
1828 		brelse(bp);
1829 		BO_RLOCK(bo);
1830 	}
1831 	return (0);
1832 }
1833 
1834 /*
1835  * Truncate a file's buffer and pages to a specified length.  This
1836  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1837  * sync activity.
1838  */
1839 int
1840 vtruncbuf(struct vnode *vp, struct ucred *cred, off_t length, int blksize)
1841 {
1842 	struct buf *bp, *nbp;
1843 	int anyfreed;
1844 	int trunclbn;
1845 	struct bufobj *bo;
1846 
1847 	CTR5(KTR_VFS, "%s: vp %p with cred %p and block %d:%ju", __func__,
1848 	    vp, cred, blksize, (uintmax_t)length);
1849 
1850 	/*
1851 	 * Round up to the *next* lbn.
1852 	 */
1853 	trunclbn = howmany(length, blksize);
1854 
1855 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1856 restart:
1857 	bo = &vp->v_bufobj;
1858 	BO_LOCK(bo);
1859 	anyfreed = 1;
1860 	for (;anyfreed;) {
1861 		anyfreed = 0;
1862 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1863 			if (bp->b_lblkno < trunclbn)
1864 				continue;
1865 			if (BUF_LOCK(bp,
1866 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1867 			    BO_LOCKPTR(bo)) == ENOLCK)
1868 				goto restart;
1869 
1870 			bremfree(bp);
1871 			bp->b_flags |= (B_INVAL | B_RELBUF);
1872 			bp->b_flags &= ~B_ASYNC;
1873 			brelse(bp);
1874 			anyfreed = 1;
1875 
1876 			BO_LOCK(bo);
1877 			if (nbp != NULL &&
1878 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1879 			    (nbp->b_vp != vp) ||
1880 			    (nbp->b_flags & B_DELWRI))) {
1881 				BO_UNLOCK(bo);
1882 				goto restart;
1883 			}
1884 		}
1885 
1886 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1887 			if (bp->b_lblkno < trunclbn)
1888 				continue;
1889 			if (BUF_LOCK(bp,
1890 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1891 			    BO_LOCKPTR(bo)) == ENOLCK)
1892 				goto restart;
1893 			bremfree(bp);
1894 			bp->b_flags |= (B_INVAL | B_RELBUF);
1895 			bp->b_flags &= ~B_ASYNC;
1896 			brelse(bp);
1897 			anyfreed = 1;
1898 
1899 			BO_LOCK(bo);
1900 			if (nbp != NULL &&
1901 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1902 			    (nbp->b_vp != vp) ||
1903 			    (nbp->b_flags & B_DELWRI) == 0)) {
1904 				BO_UNLOCK(bo);
1905 				goto restart;
1906 			}
1907 		}
1908 	}
1909 
1910 	if (length > 0) {
1911 restartsync:
1912 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1913 			if (bp->b_lblkno > 0)
1914 				continue;
1915 			/*
1916 			 * Since we hold the vnode lock this should only
1917 			 * fail if we're racing with the buf daemon.
1918 			 */
1919 			if (BUF_LOCK(bp,
1920 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1921 			    BO_LOCKPTR(bo)) == ENOLCK) {
1922 				goto restart;
1923 			}
1924 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1925 			    ("buf(%p) on dirty queue without DELWRI", bp));
1926 
1927 			bremfree(bp);
1928 			bawrite(bp);
1929 			BO_LOCK(bo);
1930 			goto restartsync;
1931 		}
1932 	}
1933 
1934 	bufobj_wwait(bo, 0, 0);
1935 	BO_UNLOCK(bo);
1936 	vnode_pager_setsize(vp, length);
1937 
1938 	return (0);
1939 }
1940 
1941 static void
1942 buf_vlist_remove(struct buf *bp)
1943 {
1944 	struct bufv *bv;
1945 
1946 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1947 	ASSERT_BO_WLOCKED(bp->b_bufobj);
1948 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
1949 	    (BX_VNDIRTY|BX_VNCLEAN),
1950 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
1951 	if (bp->b_xflags & BX_VNDIRTY)
1952 		bv = &bp->b_bufobj->bo_dirty;
1953 	else
1954 		bv = &bp->b_bufobj->bo_clean;
1955 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
1956 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1957 	bv->bv_cnt--;
1958 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1959 }
1960 
1961 /*
1962  * Add the buffer to the sorted clean or dirty block list.
1963  *
1964  * NOTE: xflags is passed as a constant, optimizing this inline function!
1965  */
1966 static void
1967 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1968 {
1969 	struct bufv *bv;
1970 	struct buf *n;
1971 	int error;
1972 
1973 	ASSERT_BO_WLOCKED(bo);
1974 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
1975 	    ("dead bo %p", bo));
1976 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1977 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
1978 	bp->b_xflags |= xflags;
1979 	if (xflags & BX_VNDIRTY)
1980 		bv = &bo->bo_dirty;
1981 	else
1982 		bv = &bo->bo_clean;
1983 
1984 	/*
1985 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
1986 	 * we tend to grow at the tail so lookup_le should usually be cheaper
1987 	 * than _ge.
1988 	 */
1989 	if (bv->bv_cnt == 0 ||
1990 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
1991 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1992 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
1993 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
1994 	else
1995 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
1996 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
1997 	if (error)
1998 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
1999 	bv->bv_cnt++;
2000 }
2001 
2002 /*
2003  * Look up a buffer using the buffer tries.
2004  */
2005 struct buf *
2006 gbincore(struct bufobj *bo, daddr_t lblkno)
2007 {
2008 	struct buf *bp;
2009 
2010 	ASSERT_BO_LOCKED(bo);
2011 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2012 	if (bp != NULL)
2013 		return (bp);
2014 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
2015 }
2016 
2017 /*
2018  * Associate a buffer with a vnode.
2019  */
2020 void
2021 bgetvp(struct vnode *vp, struct buf *bp)
2022 {
2023 	struct bufobj *bo;
2024 
2025 	bo = &vp->v_bufobj;
2026 	ASSERT_BO_WLOCKED(bo);
2027 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2028 
2029 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2030 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2031 	    ("bgetvp: bp already attached! %p", bp));
2032 
2033 	vhold(vp);
2034 	bp->b_vp = vp;
2035 	bp->b_bufobj = bo;
2036 	/*
2037 	 * Insert onto list for new vnode.
2038 	 */
2039 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2040 }
2041 
2042 /*
2043  * Disassociate a buffer from a vnode.
2044  */
2045 void
2046 brelvp(struct buf *bp)
2047 {
2048 	struct bufobj *bo;
2049 	struct vnode *vp;
2050 
2051 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2052 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2053 
2054 	/*
2055 	 * Delete from old vnode list, if on one.
2056 	 */
2057 	vp = bp->b_vp;		/* XXX */
2058 	bo = bp->b_bufobj;
2059 	BO_LOCK(bo);
2060 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2061 		buf_vlist_remove(bp);
2062 	else
2063 		panic("brelvp: Buffer %p not on queue.", bp);
2064 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2065 		bo->bo_flag &= ~BO_ONWORKLST;
2066 		mtx_lock(&sync_mtx);
2067 		LIST_REMOVE(bo, bo_synclist);
2068 		syncer_worklist_len--;
2069 		mtx_unlock(&sync_mtx);
2070 	}
2071 	bp->b_vp = NULL;
2072 	bp->b_bufobj = NULL;
2073 	BO_UNLOCK(bo);
2074 	vdrop(vp);
2075 }
2076 
2077 /*
2078  * Add an item to the syncer work queue.
2079  */
2080 static void
2081 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2082 {
2083 	int slot;
2084 
2085 	ASSERT_BO_WLOCKED(bo);
2086 
2087 	mtx_lock(&sync_mtx);
2088 	if (bo->bo_flag & BO_ONWORKLST)
2089 		LIST_REMOVE(bo, bo_synclist);
2090 	else {
2091 		bo->bo_flag |= BO_ONWORKLST;
2092 		syncer_worklist_len++;
2093 	}
2094 
2095 	if (delay > syncer_maxdelay - 2)
2096 		delay = syncer_maxdelay - 2;
2097 	slot = (syncer_delayno + delay) & syncer_mask;
2098 
2099 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2100 	mtx_unlock(&sync_mtx);
2101 }
2102 
2103 static int
2104 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2105 {
2106 	int error, len;
2107 
2108 	mtx_lock(&sync_mtx);
2109 	len = syncer_worklist_len - sync_vnode_count;
2110 	mtx_unlock(&sync_mtx);
2111 	error = SYSCTL_OUT(req, &len, sizeof(len));
2112 	return (error);
2113 }
2114 
2115 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
2116     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2117 
2118 static struct proc *updateproc;
2119 static void sched_sync(void);
2120 static struct kproc_desc up_kp = {
2121 	"syncer",
2122 	sched_sync,
2123 	&updateproc
2124 };
2125 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2126 
2127 static int
2128 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2129 {
2130 	struct vnode *vp;
2131 	struct mount *mp;
2132 
2133 	*bo = LIST_FIRST(slp);
2134 	if (*bo == NULL)
2135 		return (0);
2136 	vp = bo2vnode(*bo);
2137 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2138 		return (1);
2139 	/*
2140 	 * We use vhold in case the vnode does not
2141 	 * successfully sync.  vhold prevents the vnode from
2142 	 * going away when we unlock the sync_mtx so that
2143 	 * we can acquire the vnode interlock.
2144 	 */
2145 	vholdl(vp);
2146 	mtx_unlock(&sync_mtx);
2147 	VI_UNLOCK(vp);
2148 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2149 		vdrop(vp);
2150 		mtx_lock(&sync_mtx);
2151 		return (*bo == LIST_FIRST(slp));
2152 	}
2153 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2154 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2155 	VOP_UNLOCK(vp, 0);
2156 	vn_finished_write(mp);
2157 	BO_LOCK(*bo);
2158 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2159 		/*
2160 		 * Put us back on the worklist.  The worklist
2161 		 * routine will remove us from our current
2162 		 * position and then add us back in at a later
2163 		 * position.
2164 		 */
2165 		vn_syncer_add_to_worklist(*bo, syncdelay);
2166 	}
2167 	BO_UNLOCK(*bo);
2168 	vdrop(vp);
2169 	mtx_lock(&sync_mtx);
2170 	return (0);
2171 }
2172 
2173 static int first_printf = 1;
2174 
2175 /*
2176  * System filesystem synchronizer daemon.
2177  */
2178 static void
2179 sched_sync(void)
2180 {
2181 	struct synclist *next, *slp;
2182 	struct bufobj *bo;
2183 	long starttime;
2184 	struct thread *td = curthread;
2185 	int last_work_seen;
2186 	int net_worklist_len;
2187 	int syncer_final_iter;
2188 	int error;
2189 
2190 	last_work_seen = 0;
2191 	syncer_final_iter = 0;
2192 	syncer_state = SYNCER_RUNNING;
2193 	starttime = time_uptime;
2194 	td->td_pflags |= TDP_NORUNNINGBUF;
2195 
2196 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2197 	    SHUTDOWN_PRI_LAST);
2198 
2199 	mtx_lock(&sync_mtx);
2200 	for (;;) {
2201 		if (syncer_state == SYNCER_FINAL_DELAY &&
2202 		    syncer_final_iter == 0) {
2203 			mtx_unlock(&sync_mtx);
2204 			kproc_suspend_check(td->td_proc);
2205 			mtx_lock(&sync_mtx);
2206 		}
2207 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2208 		if (syncer_state != SYNCER_RUNNING &&
2209 		    starttime != time_uptime) {
2210 			if (first_printf) {
2211 				printf("\nSyncing disks, vnodes remaining... ");
2212 				first_printf = 0;
2213 			}
2214 			printf("%d ", net_worklist_len);
2215 		}
2216 		starttime = time_uptime;
2217 
2218 		/*
2219 		 * Push files whose dirty time has expired.  Be careful
2220 		 * of interrupt race on slp queue.
2221 		 *
2222 		 * Skip over empty worklist slots when shutting down.
2223 		 */
2224 		do {
2225 			slp = &syncer_workitem_pending[syncer_delayno];
2226 			syncer_delayno += 1;
2227 			if (syncer_delayno == syncer_maxdelay)
2228 				syncer_delayno = 0;
2229 			next = &syncer_workitem_pending[syncer_delayno];
2230 			/*
2231 			 * If the worklist has wrapped since the
2232 			 * it was emptied of all but syncer vnodes,
2233 			 * switch to the FINAL_DELAY state and run
2234 			 * for one more second.
2235 			 */
2236 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2237 			    net_worklist_len == 0 &&
2238 			    last_work_seen == syncer_delayno) {
2239 				syncer_state = SYNCER_FINAL_DELAY;
2240 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2241 			}
2242 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2243 		    syncer_worklist_len > 0);
2244 
2245 		/*
2246 		 * Keep track of the last time there was anything
2247 		 * on the worklist other than syncer vnodes.
2248 		 * Return to the SHUTTING_DOWN state if any
2249 		 * new work appears.
2250 		 */
2251 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2252 			last_work_seen = syncer_delayno;
2253 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2254 			syncer_state = SYNCER_SHUTTING_DOWN;
2255 		while (!LIST_EMPTY(slp)) {
2256 			error = sync_vnode(slp, &bo, td);
2257 			if (error == 1) {
2258 				LIST_REMOVE(bo, bo_synclist);
2259 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2260 				continue;
2261 			}
2262 
2263 			if (first_printf == 0) {
2264 				/*
2265 				 * Drop the sync mutex, because some watchdog
2266 				 * drivers need to sleep while patting
2267 				 */
2268 				mtx_unlock(&sync_mtx);
2269 				wdog_kern_pat(WD_LASTVAL);
2270 				mtx_lock(&sync_mtx);
2271 			}
2272 
2273 		}
2274 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2275 			syncer_final_iter--;
2276 		/*
2277 		 * The variable rushjob allows the kernel to speed up the
2278 		 * processing of the filesystem syncer process. A rushjob
2279 		 * value of N tells the filesystem syncer to process the next
2280 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2281 		 * is used by the soft update code to speed up the filesystem
2282 		 * syncer process when the incore state is getting so far
2283 		 * ahead of the disk that the kernel memory pool is being
2284 		 * threatened with exhaustion.
2285 		 */
2286 		if (rushjob > 0) {
2287 			rushjob -= 1;
2288 			continue;
2289 		}
2290 		/*
2291 		 * Just sleep for a short period of time between
2292 		 * iterations when shutting down to allow some I/O
2293 		 * to happen.
2294 		 *
2295 		 * If it has taken us less than a second to process the
2296 		 * current work, then wait. Otherwise start right over
2297 		 * again. We can still lose time if any single round
2298 		 * takes more than two seconds, but it does not really
2299 		 * matter as we are just trying to generally pace the
2300 		 * filesystem activity.
2301 		 */
2302 		if (syncer_state != SYNCER_RUNNING ||
2303 		    time_uptime == starttime) {
2304 			thread_lock(td);
2305 			sched_prio(td, PPAUSE);
2306 			thread_unlock(td);
2307 		}
2308 		if (syncer_state != SYNCER_RUNNING)
2309 			cv_timedwait(&sync_wakeup, &sync_mtx,
2310 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2311 		else if (time_uptime == starttime)
2312 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2313 	}
2314 }
2315 
2316 /*
2317  * Request the syncer daemon to speed up its work.
2318  * We never push it to speed up more than half of its
2319  * normal turn time, otherwise it could take over the cpu.
2320  */
2321 int
2322 speedup_syncer(void)
2323 {
2324 	int ret = 0;
2325 
2326 	mtx_lock(&sync_mtx);
2327 	if (rushjob < syncdelay / 2) {
2328 		rushjob += 1;
2329 		stat_rush_requests += 1;
2330 		ret = 1;
2331 	}
2332 	mtx_unlock(&sync_mtx);
2333 	cv_broadcast(&sync_wakeup);
2334 	return (ret);
2335 }
2336 
2337 /*
2338  * Tell the syncer to speed up its work and run though its work
2339  * list several times, then tell it to shut down.
2340  */
2341 static void
2342 syncer_shutdown(void *arg, int howto)
2343 {
2344 
2345 	if (howto & RB_NOSYNC)
2346 		return;
2347 	mtx_lock(&sync_mtx);
2348 	syncer_state = SYNCER_SHUTTING_DOWN;
2349 	rushjob = 0;
2350 	mtx_unlock(&sync_mtx);
2351 	cv_broadcast(&sync_wakeup);
2352 	kproc_shutdown(arg, howto);
2353 }
2354 
2355 void
2356 syncer_suspend(void)
2357 {
2358 
2359 	syncer_shutdown(updateproc, 0);
2360 }
2361 
2362 void
2363 syncer_resume(void)
2364 {
2365 
2366 	mtx_lock(&sync_mtx);
2367 	first_printf = 1;
2368 	syncer_state = SYNCER_RUNNING;
2369 	mtx_unlock(&sync_mtx);
2370 	cv_broadcast(&sync_wakeup);
2371 	kproc_resume(updateproc);
2372 }
2373 
2374 /*
2375  * Reassign a buffer from one vnode to another.
2376  * Used to assign file specific control information
2377  * (indirect blocks) to the vnode to which they belong.
2378  */
2379 void
2380 reassignbuf(struct buf *bp)
2381 {
2382 	struct vnode *vp;
2383 	struct bufobj *bo;
2384 	int delay;
2385 #ifdef INVARIANTS
2386 	struct bufv *bv;
2387 #endif
2388 
2389 	vp = bp->b_vp;
2390 	bo = bp->b_bufobj;
2391 	++reassignbufcalls;
2392 
2393 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2394 	    bp, bp->b_vp, bp->b_flags);
2395 	/*
2396 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2397 	 * is not fully linked in.
2398 	 */
2399 	if (bp->b_flags & B_PAGING)
2400 		panic("cannot reassign paging buffer");
2401 
2402 	/*
2403 	 * Delete from old vnode list, if on one.
2404 	 */
2405 	BO_LOCK(bo);
2406 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2407 		buf_vlist_remove(bp);
2408 	else
2409 		panic("reassignbuf: Buffer %p not on queue.", bp);
2410 	/*
2411 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2412 	 * of clean buffers.
2413 	 */
2414 	if (bp->b_flags & B_DELWRI) {
2415 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2416 			switch (vp->v_type) {
2417 			case VDIR:
2418 				delay = dirdelay;
2419 				break;
2420 			case VCHR:
2421 				delay = metadelay;
2422 				break;
2423 			default:
2424 				delay = filedelay;
2425 			}
2426 			vn_syncer_add_to_worklist(bo, delay);
2427 		}
2428 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2429 	} else {
2430 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2431 
2432 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2433 			mtx_lock(&sync_mtx);
2434 			LIST_REMOVE(bo, bo_synclist);
2435 			syncer_worklist_len--;
2436 			mtx_unlock(&sync_mtx);
2437 			bo->bo_flag &= ~BO_ONWORKLST;
2438 		}
2439 	}
2440 #ifdef INVARIANTS
2441 	bv = &bo->bo_clean;
2442 	bp = TAILQ_FIRST(&bv->bv_hd);
2443 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2444 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2445 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2446 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2447 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2448 	bv = &bo->bo_dirty;
2449 	bp = TAILQ_FIRST(&bv->bv_hd);
2450 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2451 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2452 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2453 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2454 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2455 #endif
2456 	BO_UNLOCK(bo);
2457 }
2458 
2459 static void
2460 v_init_counters(struct vnode *vp)
2461 {
2462 
2463 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2464 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2465 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2466 
2467 	refcount_init(&vp->v_holdcnt, 1);
2468 	refcount_init(&vp->v_usecount, 1);
2469 }
2470 
2471 static void
2472 v_incr_usecount_locked(struct vnode *vp)
2473 {
2474 
2475 	ASSERT_VI_LOCKED(vp, __func__);
2476 	if ((vp->v_iflag & VI_OWEINACT) != 0) {
2477 		VNASSERT(vp->v_usecount == 0, vp,
2478 		    ("vnode with usecount and VI_OWEINACT set"));
2479 		vp->v_iflag &= ~VI_OWEINACT;
2480 	}
2481 	refcount_acquire(&vp->v_usecount);
2482 	v_incr_devcount(vp);
2483 }
2484 
2485 /*
2486  * Increment the use count on the vnode, taking care to reference
2487  * the driver's usecount if this is a chardev.
2488  */
2489 static void
2490 v_incr_usecount(struct vnode *vp)
2491 {
2492 
2493 	ASSERT_VI_UNLOCKED(vp, __func__);
2494 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2495 
2496 	if (vp->v_type != VCHR &&
2497 	    refcount_acquire_if_not_zero(&vp->v_usecount)) {
2498 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2499 		    ("vnode with usecount and VI_OWEINACT set"));
2500 	} else {
2501 		VI_LOCK(vp);
2502 		v_incr_usecount_locked(vp);
2503 		VI_UNLOCK(vp);
2504 	}
2505 }
2506 
2507 /*
2508  * Increment si_usecount of the associated device, if any.
2509  */
2510 static void
2511 v_incr_devcount(struct vnode *vp)
2512 {
2513 
2514 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2515 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2516 		dev_lock();
2517 		vp->v_rdev->si_usecount++;
2518 		dev_unlock();
2519 	}
2520 }
2521 
2522 /*
2523  * Decrement si_usecount of the associated device, if any.
2524  */
2525 static void
2526 v_decr_devcount(struct vnode *vp)
2527 {
2528 
2529 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2530 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2531 		dev_lock();
2532 		vp->v_rdev->si_usecount--;
2533 		dev_unlock();
2534 	}
2535 }
2536 
2537 /*
2538  * Grab a particular vnode from the free list, increment its
2539  * reference count and lock it.  VI_DOOMED is set if the vnode
2540  * is being destroyed.  Only callers who specify LK_RETRY will
2541  * see doomed vnodes.  If inactive processing was delayed in
2542  * vput try to do it here.
2543  *
2544  * Notes on lockless counter manipulation:
2545  * _vhold, vputx and other routines make various decisions based
2546  * on either holdcnt or usecount being 0. As long as either counter
2547  * is not transitioning 0->1 nor 1->0, the manipulation can be done
2548  * with atomic operations. Otherwise the interlock is taken covering
2549  * both the atomic and additional actions.
2550  */
2551 int
2552 vget(struct vnode *vp, int flags, struct thread *td)
2553 {
2554 	int error, oweinact;
2555 
2556 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2557 	    ("vget: invalid lock operation"));
2558 
2559 	if ((flags & LK_INTERLOCK) != 0)
2560 		ASSERT_VI_LOCKED(vp, __func__);
2561 	else
2562 		ASSERT_VI_UNLOCKED(vp, __func__);
2563 	if ((flags & LK_VNHELD) != 0)
2564 		VNASSERT((vp->v_holdcnt > 0), vp,
2565 		    ("vget: LK_VNHELD passed but vnode not held"));
2566 
2567 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2568 
2569 	if ((flags & LK_VNHELD) == 0)
2570 		_vhold(vp, (flags & LK_INTERLOCK) != 0);
2571 
2572 	if ((error = vn_lock(vp, flags)) != 0) {
2573 		vdrop(vp);
2574 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2575 		    vp);
2576 		return (error);
2577 	}
2578 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
2579 		panic("vget: vn_lock failed to return ENOENT\n");
2580 	/*
2581 	 * We don't guarantee that any particular close will
2582 	 * trigger inactive processing so just make a best effort
2583 	 * here at preventing a reference to a removed file.  If
2584 	 * we don't succeed no harm is done.
2585 	 *
2586 	 * Upgrade our holdcnt to a usecount.
2587 	 */
2588 	if (vp->v_type == VCHR ||
2589 	    !refcount_acquire_if_not_zero(&vp->v_usecount)) {
2590 		VI_LOCK(vp);
2591 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2592 			oweinact = 0;
2593 		} else {
2594 			oweinact = 1;
2595 			vp->v_iflag &= ~VI_OWEINACT;
2596 		}
2597 		refcount_acquire(&vp->v_usecount);
2598 		v_incr_devcount(vp);
2599 		if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2600 		    (flags & LK_NOWAIT) == 0)
2601 			vinactive(vp, td);
2602 		VI_UNLOCK(vp);
2603 	}
2604 	return (0);
2605 }
2606 
2607 /*
2608  * Increase the reference (use) and hold count of a vnode.
2609  * This will also remove the vnode from the free list if it is presently free.
2610  */
2611 void
2612 vref(struct vnode *vp)
2613 {
2614 
2615 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2616 	_vhold(vp, false);
2617 	v_incr_usecount(vp);
2618 }
2619 
2620 void
2621 vrefl(struct vnode *vp)
2622 {
2623 
2624 	ASSERT_VI_LOCKED(vp, __func__);
2625 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2626 	_vhold(vp, true);
2627 	v_incr_usecount_locked(vp);
2628 }
2629 
2630 void
2631 vrefact(struct vnode *vp)
2632 {
2633 
2634 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2635 	if (__predict_false(vp->v_type == VCHR)) {
2636 		VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp,
2637 		    ("%s: wrong ref counts", __func__));
2638 		vref(vp);
2639 		return;
2640 	}
2641 #ifdef INVARIANTS
2642 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
2643 	VNASSERT(old > 0, vp, ("%s: wrong hold count", __func__));
2644 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
2645 	VNASSERT(old > 0, vp, ("%s: wrong use count", __func__));
2646 #else
2647 	refcount_acquire(&vp->v_holdcnt);
2648 	refcount_acquire(&vp->v_usecount);
2649 #endif
2650 }
2651 
2652 /*
2653  * Return reference count of a vnode.
2654  *
2655  * The results of this call are only guaranteed when some mechanism is used to
2656  * stop other processes from gaining references to the vnode.  This may be the
2657  * case if the caller holds the only reference.  This is also useful when stale
2658  * data is acceptable as race conditions may be accounted for by some other
2659  * means.
2660  */
2661 int
2662 vrefcnt(struct vnode *vp)
2663 {
2664 
2665 	return (vp->v_usecount);
2666 }
2667 
2668 #define	VPUTX_VRELE	1
2669 #define	VPUTX_VPUT	2
2670 #define	VPUTX_VUNREF	3
2671 
2672 /*
2673  * Decrement the use and hold counts for a vnode.
2674  *
2675  * See an explanation near vget() as to why atomic operation is safe.
2676  */
2677 static void
2678 vputx(struct vnode *vp, int func)
2679 {
2680 	int error;
2681 
2682 	KASSERT(vp != NULL, ("vputx: null vp"));
2683 	if (func == VPUTX_VUNREF)
2684 		ASSERT_VOP_LOCKED(vp, "vunref");
2685 	else if (func == VPUTX_VPUT)
2686 		ASSERT_VOP_LOCKED(vp, "vput");
2687 	else
2688 		KASSERT(func == VPUTX_VRELE, ("vputx: wrong func"));
2689 	ASSERT_VI_UNLOCKED(vp, __func__);
2690 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2691 
2692 	if (vp->v_type != VCHR &&
2693 	    refcount_release_if_not_last(&vp->v_usecount)) {
2694 		if (func == VPUTX_VPUT)
2695 			VOP_UNLOCK(vp, 0);
2696 		vdrop(vp);
2697 		return;
2698 	}
2699 
2700 	VI_LOCK(vp);
2701 
2702 	/*
2703 	 * We want to hold the vnode until the inactive finishes to
2704 	 * prevent vgone() races.  We drop the use count here and the
2705 	 * hold count below when we're done.
2706 	 */
2707 	if (!refcount_release(&vp->v_usecount) ||
2708 	    (vp->v_iflag & VI_DOINGINACT)) {
2709 		if (func == VPUTX_VPUT)
2710 			VOP_UNLOCK(vp, 0);
2711 		v_decr_devcount(vp);
2712 		vdropl(vp);
2713 		return;
2714 	}
2715 
2716 	v_decr_devcount(vp);
2717 
2718 	error = 0;
2719 
2720 	if (vp->v_usecount != 0) {
2721 		vn_printf(vp, "vputx: usecount not zero for vnode ");
2722 		panic("vputx: usecount not zero");
2723 	}
2724 
2725 	CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp);
2726 
2727 	/*
2728 	 * We must call VOP_INACTIVE with the node locked. Mark
2729 	 * as VI_DOINGINACT to avoid recursion.
2730 	 */
2731 	vp->v_iflag |= VI_OWEINACT;
2732 	switch (func) {
2733 	case VPUTX_VRELE:
2734 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
2735 		VI_LOCK(vp);
2736 		break;
2737 	case VPUTX_VPUT:
2738 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2739 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
2740 			    LK_NOWAIT);
2741 			VI_LOCK(vp);
2742 		}
2743 		break;
2744 	case VPUTX_VUNREF:
2745 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2746 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
2747 			VI_LOCK(vp);
2748 		}
2749 		break;
2750 	}
2751 	VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
2752 	    ("vnode with usecount and VI_OWEINACT set"));
2753 	if (error == 0) {
2754 		if (vp->v_iflag & VI_OWEINACT)
2755 			vinactive(vp, curthread);
2756 		if (func != VPUTX_VUNREF)
2757 			VOP_UNLOCK(vp, 0);
2758 	}
2759 	vdropl(vp);
2760 }
2761 
2762 /*
2763  * Vnode put/release.
2764  * If count drops to zero, call inactive routine and return to freelist.
2765  */
2766 void
2767 vrele(struct vnode *vp)
2768 {
2769 
2770 	vputx(vp, VPUTX_VRELE);
2771 }
2772 
2773 /*
2774  * Release an already locked vnode.  This give the same effects as
2775  * unlock+vrele(), but takes less time and avoids releasing and
2776  * re-aquiring the lock (as vrele() acquires the lock internally.)
2777  */
2778 void
2779 vput(struct vnode *vp)
2780 {
2781 
2782 	vputx(vp, VPUTX_VPUT);
2783 }
2784 
2785 /*
2786  * Release an exclusively locked vnode. Do not unlock the vnode lock.
2787  */
2788 void
2789 vunref(struct vnode *vp)
2790 {
2791 
2792 	vputx(vp, VPUTX_VUNREF);
2793 }
2794 
2795 /*
2796  * Increase the hold count and activate if this is the first reference.
2797  */
2798 void
2799 _vhold(struct vnode *vp, bool locked)
2800 {
2801 	struct mount *mp;
2802 
2803 	if (locked)
2804 		ASSERT_VI_LOCKED(vp, __func__);
2805 	else
2806 		ASSERT_VI_UNLOCKED(vp, __func__);
2807 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2808 	if (!locked) {
2809 		if (refcount_acquire_if_not_zero(&vp->v_holdcnt)) {
2810 			VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2811 			    ("_vhold: vnode with holdcnt is free"));
2812 			return;
2813 		}
2814 		VI_LOCK(vp);
2815 	}
2816 	if ((vp->v_iflag & VI_FREE) == 0) {
2817 		refcount_acquire(&vp->v_holdcnt);
2818 		if (!locked)
2819 			VI_UNLOCK(vp);
2820 		return;
2821 	}
2822 	VNASSERT(vp->v_holdcnt == 0, vp,
2823 	    ("%s: wrong hold count", __func__));
2824 	VNASSERT(vp->v_op != NULL, vp,
2825 	    ("%s: vnode already reclaimed.", __func__));
2826 	/*
2827 	 * Remove a vnode from the free list, mark it as in use,
2828 	 * and put it on the active list.
2829 	 */
2830 	VNASSERT(vp->v_mount != NULL, vp,
2831 	    ("_vhold: vnode not on per mount vnode list"));
2832 	mp = vp->v_mount;
2833 	mtx_lock(&mp->mnt_listmtx);
2834 	if ((vp->v_mflag & VMP_TMPMNTFREELIST) != 0) {
2835 		TAILQ_REMOVE(&mp->mnt_tmpfreevnodelist, vp, v_actfreelist);
2836 		mp->mnt_tmpfreevnodelistsize--;
2837 		vp->v_mflag &= ~VMP_TMPMNTFREELIST;
2838 	} else {
2839 		mtx_lock(&vnode_free_list_mtx);
2840 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
2841 		freevnodes--;
2842 		mtx_unlock(&vnode_free_list_mtx);
2843 	}
2844 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
2845 	    ("Activating already active vnode"));
2846 	vp->v_iflag &= ~VI_FREE;
2847 	vp->v_iflag |= VI_ACTIVE;
2848 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
2849 	mp->mnt_activevnodelistsize++;
2850 	mtx_unlock(&mp->mnt_listmtx);
2851 	refcount_acquire(&vp->v_holdcnt);
2852 	if (!locked)
2853 		VI_UNLOCK(vp);
2854 }
2855 
2856 /*
2857  * Drop the hold count of the vnode.  If this is the last reference to
2858  * the vnode we place it on the free list unless it has been vgone'd
2859  * (marked VI_DOOMED) in which case we will free it.
2860  *
2861  * Because the vnode vm object keeps a hold reference on the vnode if
2862  * there is at least one resident non-cached page, the vnode cannot
2863  * leave the active list without the page cleanup done.
2864  */
2865 void
2866 _vdrop(struct vnode *vp, bool locked)
2867 {
2868 	struct bufobj *bo;
2869 	struct mount *mp;
2870 	int active;
2871 
2872 	if (locked)
2873 		ASSERT_VI_LOCKED(vp, __func__);
2874 	else
2875 		ASSERT_VI_UNLOCKED(vp, __func__);
2876 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2877 	if ((int)vp->v_holdcnt <= 0)
2878 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
2879 	if (!locked) {
2880 		if (refcount_release_if_not_last(&vp->v_holdcnt))
2881 			return;
2882 		VI_LOCK(vp);
2883 	}
2884 	if (refcount_release(&vp->v_holdcnt) == 0) {
2885 		VI_UNLOCK(vp);
2886 		return;
2887 	}
2888 	if ((vp->v_iflag & VI_DOOMED) == 0) {
2889 		/*
2890 		 * Mark a vnode as free: remove it from its active list
2891 		 * and put it up for recycling on the freelist.
2892 		 */
2893 		VNASSERT(vp->v_op != NULL, vp,
2894 		    ("vdropl: vnode already reclaimed."));
2895 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2896 		    ("vnode already free"));
2897 		VNASSERT(vp->v_holdcnt == 0, vp,
2898 		    ("vdropl: freeing when we shouldn't"));
2899 		active = vp->v_iflag & VI_ACTIVE;
2900 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2901 			vp->v_iflag &= ~VI_ACTIVE;
2902 			mp = vp->v_mount;
2903 			if (mp != NULL) {
2904 				mtx_lock(&mp->mnt_listmtx);
2905 				if (active) {
2906 					TAILQ_REMOVE(&mp->mnt_activevnodelist,
2907 					    vp, v_actfreelist);
2908 					mp->mnt_activevnodelistsize--;
2909 				}
2910 				TAILQ_INSERT_TAIL(&mp->mnt_tmpfreevnodelist,
2911 				    vp, v_actfreelist);
2912 				mp->mnt_tmpfreevnodelistsize++;
2913 				vp->v_iflag |= VI_FREE;
2914 				vp->v_mflag |= VMP_TMPMNTFREELIST;
2915 				VI_UNLOCK(vp);
2916 				if (mp->mnt_tmpfreevnodelistsize >=
2917 				    mnt_free_list_batch)
2918 					vnlru_return_batch_locked(mp);
2919 				mtx_unlock(&mp->mnt_listmtx);
2920 			} else {
2921 				VNASSERT(active == 0, vp,
2922 				    ("vdropl: active vnode not on per mount "
2923 				    "vnode list"));
2924 				mtx_lock(&vnode_free_list_mtx);
2925 				TAILQ_INSERT_TAIL(&vnode_free_list, vp,
2926 				    v_actfreelist);
2927 				freevnodes++;
2928 				vp->v_iflag |= VI_FREE;
2929 				VI_UNLOCK(vp);
2930 				mtx_unlock(&vnode_free_list_mtx);
2931 			}
2932 		} else {
2933 			VI_UNLOCK(vp);
2934 			counter_u64_add(free_owe_inact, 1);
2935 		}
2936 		return;
2937 	}
2938 	/*
2939 	 * The vnode has been marked for destruction, so free it.
2940 	 *
2941 	 * The vnode will be returned to the zone where it will
2942 	 * normally remain until it is needed for another vnode. We
2943 	 * need to cleanup (or verify that the cleanup has already
2944 	 * been done) any residual data left from its current use
2945 	 * so as not to contaminate the freshly allocated vnode.
2946 	 */
2947 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
2948 	atomic_subtract_long(&numvnodes, 1);
2949 	bo = &vp->v_bufobj;
2950 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2951 	    ("cleaned vnode still on the free list."));
2952 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
2953 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
2954 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
2955 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
2956 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
2957 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
2958 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
2959 	    ("clean blk trie not empty"));
2960 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
2961 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
2962 	    ("dirty blk trie not empty"));
2963 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
2964 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
2965 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
2966 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
2967 	    ("Dangling rangelock waiters"));
2968 	VI_UNLOCK(vp);
2969 #ifdef MAC
2970 	mac_vnode_destroy(vp);
2971 #endif
2972 	if (vp->v_pollinfo != NULL) {
2973 		destroy_vpollinfo(vp->v_pollinfo);
2974 		vp->v_pollinfo = NULL;
2975 	}
2976 #ifdef INVARIANTS
2977 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
2978 	vp->v_op = NULL;
2979 #endif
2980 	vp->v_mountedhere = NULL;
2981 	vp->v_unpcb = NULL;
2982 	vp->v_rdev = NULL;
2983 	vp->v_fifoinfo = NULL;
2984 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
2985 	vp->v_iflag = 0;
2986 	vp->v_vflag = 0;
2987 	bo->bo_flag = 0;
2988 	uma_zfree(vnode_zone, vp);
2989 }
2990 
2991 /*
2992  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
2993  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
2994  * OWEINACT tracks whether a vnode missed a call to inactive due to a
2995  * failed lock upgrade.
2996  */
2997 void
2998 vinactive(struct vnode *vp, struct thread *td)
2999 {
3000 	struct vm_object *obj;
3001 
3002 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3003 	ASSERT_VI_LOCKED(vp, "vinactive");
3004 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3005 	    ("vinactive: recursed on VI_DOINGINACT"));
3006 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3007 	vp->v_iflag |= VI_DOINGINACT;
3008 	vp->v_iflag &= ~VI_OWEINACT;
3009 	VI_UNLOCK(vp);
3010 	/*
3011 	 * Before moving off the active list, we must be sure that any
3012 	 * modified pages are converted into the vnode's dirty
3013 	 * buffers, since these will no longer be checked once the
3014 	 * vnode is on the inactive list.
3015 	 *
3016 	 * The write-out of the dirty pages is asynchronous.  At the
3017 	 * point that VOP_INACTIVE() is called, there could still be
3018 	 * pending I/O and dirty pages in the object.
3019 	 */
3020 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3021 	    (obj->flags & OBJ_MIGHTBEDIRTY) != 0) {
3022 		VM_OBJECT_WLOCK(obj);
3023 		vm_object_page_clean(obj, 0, 0, 0);
3024 		VM_OBJECT_WUNLOCK(obj);
3025 	}
3026 	VOP_INACTIVE(vp, td);
3027 	VI_LOCK(vp);
3028 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3029 	    ("vinactive: lost VI_DOINGINACT"));
3030 	vp->v_iflag &= ~VI_DOINGINACT;
3031 }
3032 
3033 /*
3034  * Remove any vnodes in the vnode table belonging to mount point mp.
3035  *
3036  * If FORCECLOSE is not specified, there should not be any active ones,
3037  * return error if any are found (nb: this is a user error, not a
3038  * system error). If FORCECLOSE is specified, detach any active vnodes
3039  * that are found.
3040  *
3041  * If WRITECLOSE is set, only flush out regular file vnodes open for
3042  * writing.
3043  *
3044  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3045  *
3046  * `rootrefs' specifies the base reference count for the root vnode
3047  * of this filesystem. The root vnode is considered busy if its
3048  * v_usecount exceeds this value. On a successful return, vflush(, td)
3049  * will call vrele() on the root vnode exactly rootrefs times.
3050  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3051  * be zero.
3052  */
3053 #ifdef DIAGNOSTIC
3054 static int busyprt = 0;		/* print out busy vnodes */
3055 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3056 #endif
3057 
3058 int
3059 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3060 {
3061 	struct vnode *vp, *mvp, *rootvp = NULL;
3062 	struct vattr vattr;
3063 	int busy = 0, error;
3064 
3065 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3066 	    rootrefs, flags);
3067 	if (rootrefs > 0) {
3068 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3069 		    ("vflush: bad args"));
3070 		/*
3071 		 * Get the filesystem root vnode. We can vput() it
3072 		 * immediately, since with rootrefs > 0, it won't go away.
3073 		 */
3074 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3075 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3076 			    __func__, error);
3077 			return (error);
3078 		}
3079 		vput(rootvp);
3080 	}
3081 loop:
3082 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3083 		vholdl(vp);
3084 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3085 		if (error) {
3086 			vdrop(vp);
3087 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3088 			goto loop;
3089 		}
3090 		/*
3091 		 * Skip over a vnodes marked VV_SYSTEM.
3092 		 */
3093 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3094 			VOP_UNLOCK(vp, 0);
3095 			vdrop(vp);
3096 			continue;
3097 		}
3098 		/*
3099 		 * If WRITECLOSE is set, flush out unlinked but still open
3100 		 * files (even if open only for reading) and regular file
3101 		 * vnodes open for writing.
3102 		 */
3103 		if (flags & WRITECLOSE) {
3104 			if (vp->v_object != NULL) {
3105 				VM_OBJECT_WLOCK(vp->v_object);
3106 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3107 				VM_OBJECT_WUNLOCK(vp->v_object);
3108 			}
3109 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3110 			if (error != 0) {
3111 				VOP_UNLOCK(vp, 0);
3112 				vdrop(vp);
3113 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3114 				return (error);
3115 			}
3116 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3117 			VI_LOCK(vp);
3118 
3119 			if ((vp->v_type == VNON ||
3120 			    (error == 0 && vattr.va_nlink > 0)) &&
3121 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
3122 				VOP_UNLOCK(vp, 0);
3123 				vdropl(vp);
3124 				continue;
3125 			}
3126 		} else
3127 			VI_LOCK(vp);
3128 		/*
3129 		 * With v_usecount == 0, all we need to do is clear out the
3130 		 * vnode data structures and we are done.
3131 		 *
3132 		 * If FORCECLOSE is set, forcibly close the vnode.
3133 		 */
3134 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3135 			vgonel(vp);
3136 		} else {
3137 			busy++;
3138 #ifdef DIAGNOSTIC
3139 			if (busyprt)
3140 				vn_printf(vp, "vflush: busy vnode ");
3141 #endif
3142 		}
3143 		VOP_UNLOCK(vp, 0);
3144 		vdropl(vp);
3145 	}
3146 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3147 		/*
3148 		 * If just the root vnode is busy, and if its refcount
3149 		 * is equal to `rootrefs', then go ahead and kill it.
3150 		 */
3151 		VI_LOCK(rootvp);
3152 		KASSERT(busy > 0, ("vflush: not busy"));
3153 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3154 		    ("vflush: usecount %d < rootrefs %d",
3155 		     rootvp->v_usecount, rootrefs));
3156 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3157 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3158 			vgone(rootvp);
3159 			VOP_UNLOCK(rootvp, 0);
3160 			busy = 0;
3161 		} else
3162 			VI_UNLOCK(rootvp);
3163 	}
3164 	if (busy) {
3165 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3166 		    busy);
3167 		return (EBUSY);
3168 	}
3169 	for (; rootrefs > 0; rootrefs--)
3170 		vrele(rootvp);
3171 	return (0);
3172 }
3173 
3174 /*
3175  * Recycle an unused vnode to the front of the free list.
3176  */
3177 int
3178 vrecycle(struct vnode *vp)
3179 {
3180 	int recycled;
3181 
3182 	VI_LOCK(vp);
3183 	recycled = vrecyclel(vp);
3184 	VI_UNLOCK(vp);
3185 	return (recycled);
3186 }
3187 
3188 /*
3189  * vrecycle, with the vp interlock held.
3190  */
3191 int
3192 vrecyclel(struct vnode *vp)
3193 {
3194 	int recycled;
3195 
3196 	ASSERT_VOP_ELOCKED(vp, __func__);
3197 	ASSERT_VI_LOCKED(vp, __func__);
3198 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3199 	recycled = 0;
3200 	if (vp->v_usecount == 0) {
3201 		recycled = 1;
3202 		vgonel(vp);
3203 	}
3204 	return (recycled);
3205 }
3206 
3207 /*
3208  * Eliminate all activity associated with a vnode
3209  * in preparation for reuse.
3210  */
3211 void
3212 vgone(struct vnode *vp)
3213 {
3214 	VI_LOCK(vp);
3215 	vgonel(vp);
3216 	VI_UNLOCK(vp);
3217 }
3218 
3219 static void
3220 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3221     struct vnode *lowervp __unused)
3222 {
3223 }
3224 
3225 /*
3226  * Notify upper mounts about reclaimed or unlinked vnode.
3227  */
3228 void
3229 vfs_notify_upper(struct vnode *vp, int event)
3230 {
3231 	static struct vfsops vgonel_vfsops = {
3232 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3233 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3234 	};
3235 	struct mount *mp, *ump, *mmp;
3236 
3237 	mp = vp->v_mount;
3238 	if (mp == NULL)
3239 		return;
3240 
3241 	MNT_ILOCK(mp);
3242 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3243 		goto unlock;
3244 	MNT_IUNLOCK(mp);
3245 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3246 	mmp->mnt_op = &vgonel_vfsops;
3247 	mmp->mnt_kern_flag |= MNTK_MARKER;
3248 	MNT_ILOCK(mp);
3249 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3250 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3251 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3252 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3253 			continue;
3254 		}
3255 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3256 		MNT_IUNLOCK(mp);
3257 		switch (event) {
3258 		case VFS_NOTIFY_UPPER_RECLAIM:
3259 			VFS_RECLAIM_LOWERVP(ump, vp);
3260 			break;
3261 		case VFS_NOTIFY_UPPER_UNLINK:
3262 			VFS_UNLINK_LOWERVP(ump, vp);
3263 			break;
3264 		default:
3265 			KASSERT(0, ("invalid event %d", event));
3266 			break;
3267 		}
3268 		MNT_ILOCK(mp);
3269 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3270 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3271 	}
3272 	free(mmp, M_TEMP);
3273 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3274 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3275 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3276 		wakeup(&mp->mnt_uppers);
3277 	}
3278 unlock:
3279 	MNT_IUNLOCK(mp);
3280 }
3281 
3282 /*
3283  * vgone, with the vp interlock held.
3284  */
3285 static void
3286 vgonel(struct vnode *vp)
3287 {
3288 	struct thread *td;
3289 	int oweinact;
3290 	int active;
3291 	struct mount *mp;
3292 
3293 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3294 	ASSERT_VI_LOCKED(vp, "vgonel");
3295 	VNASSERT(vp->v_holdcnt, vp,
3296 	    ("vgonel: vp %p has no reference.", vp));
3297 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3298 	td = curthread;
3299 
3300 	/*
3301 	 * Don't vgonel if we're already doomed.
3302 	 */
3303 	if (vp->v_iflag & VI_DOOMED)
3304 		return;
3305 	vp->v_iflag |= VI_DOOMED;
3306 
3307 	/*
3308 	 * Check to see if the vnode is in use.  If so, we have to call
3309 	 * VOP_CLOSE() and VOP_INACTIVE().
3310 	 */
3311 	active = vp->v_usecount;
3312 	oweinact = (vp->v_iflag & VI_OWEINACT);
3313 	VI_UNLOCK(vp);
3314 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3315 
3316 	/*
3317 	 * If purging an active vnode, it must be closed and
3318 	 * deactivated before being reclaimed.
3319 	 */
3320 	if (active)
3321 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3322 	if (oweinact || active) {
3323 		VI_LOCK(vp);
3324 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
3325 			vinactive(vp, td);
3326 		VI_UNLOCK(vp);
3327 	}
3328 	if (vp->v_type == VSOCK)
3329 		vfs_unp_reclaim(vp);
3330 
3331 	/*
3332 	 * Clean out any buffers associated with the vnode.
3333 	 * If the flush fails, just toss the buffers.
3334 	 */
3335 	mp = NULL;
3336 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3337 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3338 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3339 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3340 			;
3341 	}
3342 
3343 	BO_LOCK(&vp->v_bufobj);
3344 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3345 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3346 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3347 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3348 	    ("vp %p bufobj not invalidated", vp));
3349 
3350 	/*
3351 	 * For VMIO bufobj, BO_DEAD is set in vm_object_terminate()
3352 	 * after the object's page queue is flushed.
3353 	 */
3354 	if (vp->v_bufobj.bo_object == NULL)
3355 		vp->v_bufobj.bo_flag |= BO_DEAD;
3356 	BO_UNLOCK(&vp->v_bufobj);
3357 
3358 	/*
3359 	 * Reclaim the vnode.
3360 	 */
3361 	if (VOP_RECLAIM(vp, td))
3362 		panic("vgone: cannot reclaim");
3363 	if (mp != NULL)
3364 		vn_finished_secondary_write(mp);
3365 	VNASSERT(vp->v_object == NULL, vp,
3366 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
3367 	/*
3368 	 * Clear the advisory locks and wake up waiting threads.
3369 	 */
3370 	(void)VOP_ADVLOCKPURGE(vp);
3371 	vp->v_lockf = NULL;
3372 	/*
3373 	 * Delete from old mount point vnode list.
3374 	 */
3375 	delmntque(vp);
3376 	cache_purge(vp);
3377 	/*
3378 	 * Done with purge, reset to the standard lock and invalidate
3379 	 * the vnode.
3380 	 */
3381 	VI_LOCK(vp);
3382 	vp->v_vnlock = &vp->v_lock;
3383 	vp->v_op = &dead_vnodeops;
3384 	vp->v_tag = "none";
3385 	vp->v_type = VBAD;
3386 }
3387 
3388 /*
3389  * Calculate the total number of references to a special device.
3390  */
3391 int
3392 vcount(struct vnode *vp)
3393 {
3394 	int count;
3395 
3396 	dev_lock();
3397 	count = vp->v_rdev->si_usecount;
3398 	dev_unlock();
3399 	return (count);
3400 }
3401 
3402 /*
3403  * Same as above, but using the struct cdev *as argument
3404  */
3405 int
3406 count_dev(struct cdev *dev)
3407 {
3408 	int count;
3409 
3410 	dev_lock();
3411 	count = dev->si_usecount;
3412 	dev_unlock();
3413 	return(count);
3414 }
3415 
3416 /*
3417  * Print out a description of a vnode.
3418  */
3419 static char *typename[] =
3420 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3421  "VMARKER"};
3422 
3423 void
3424 vn_printf(struct vnode *vp, const char *fmt, ...)
3425 {
3426 	va_list ap;
3427 	char buf[256], buf2[16];
3428 	u_long flags;
3429 
3430 	va_start(ap, fmt);
3431 	vprintf(fmt, ap);
3432 	va_end(ap);
3433 	printf("%p: ", (void *)vp);
3434 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
3435 	printf("    usecount %d, writecount %d, refcount %d",
3436 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
3437 	switch (vp->v_type) {
3438 	case VDIR:
3439 		printf(" mountedhere %p\n", vp->v_mountedhere);
3440 		break;
3441 	case VCHR:
3442 		printf(" rdev %p\n", vp->v_rdev);
3443 		break;
3444 	case VSOCK:
3445 		printf(" socket %p\n", vp->v_unpcb);
3446 		break;
3447 	case VFIFO:
3448 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
3449 		break;
3450 	default:
3451 		printf("\n");
3452 		break;
3453 	}
3454 	buf[0] = '\0';
3455 	buf[1] = '\0';
3456 	if (vp->v_vflag & VV_ROOT)
3457 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3458 	if (vp->v_vflag & VV_ISTTY)
3459 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3460 	if (vp->v_vflag & VV_NOSYNC)
3461 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3462 	if (vp->v_vflag & VV_ETERNALDEV)
3463 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3464 	if (vp->v_vflag & VV_CACHEDLABEL)
3465 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3466 	if (vp->v_vflag & VV_TEXT)
3467 		strlcat(buf, "|VV_TEXT", sizeof(buf));
3468 	if (vp->v_vflag & VV_COPYONWRITE)
3469 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3470 	if (vp->v_vflag & VV_SYSTEM)
3471 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3472 	if (vp->v_vflag & VV_PROCDEP)
3473 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3474 	if (vp->v_vflag & VV_NOKNOTE)
3475 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3476 	if (vp->v_vflag & VV_DELETED)
3477 		strlcat(buf, "|VV_DELETED", sizeof(buf));
3478 	if (vp->v_vflag & VV_MD)
3479 		strlcat(buf, "|VV_MD", sizeof(buf));
3480 	if (vp->v_vflag & VV_FORCEINSMQ)
3481 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
3482 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
3483 	    VV_CACHEDLABEL | VV_TEXT | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
3484 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
3485 	if (flags != 0) {
3486 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
3487 		strlcat(buf, buf2, sizeof(buf));
3488 	}
3489 	if (vp->v_iflag & VI_MOUNT)
3490 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
3491 	if (vp->v_iflag & VI_DOOMED)
3492 		strlcat(buf, "|VI_DOOMED", sizeof(buf));
3493 	if (vp->v_iflag & VI_FREE)
3494 		strlcat(buf, "|VI_FREE", sizeof(buf));
3495 	if (vp->v_iflag & VI_ACTIVE)
3496 		strlcat(buf, "|VI_ACTIVE", sizeof(buf));
3497 	if (vp->v_iflag & VI_DOINGINACT)
3498 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
3499 	if (vp->v_iflag & VI_OWEINACT)
3500 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
3501 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOOMED | VI_FREE |
3502 	    VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT);
3503 	if (flags != 0) {
3504 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
3505 		strlcat(buf, buf2, sizeof(buf));
3506 	}
3507 	printf("    flags (%s)\n", buf + 1);
3508 	if (mtx_owned(VI_MTX(vp)))
3509 		printf(" VI_LOCKed");
3510 	if (vp->v_object != NULL)
3511 		printf("    v_object %p ref %d pages %d "
3512 		    "cleanbuf %d dirtybuf %d\n",
3513 		    vp->v_object, vp->v_object->ref_count,
3514 		    vp->v_object->resident_page_count,
3515 		    vp->v_bufobj.bo_clean.bv_cnt,
3516 		    vp->v_bufobj.bo_dirty.bv_cnt);
3517 	printf("    ");
3518 	lockmgr_printinfo(vp->v_vnlock);
3519 	if (vp->v_data != NULL)
3520 		VOP_PRINT(vp);
3521 }
3522 
3523 #ifdef DDB
3524 /*
3525  * List all of the locked vnodes in the system.
3526  * Called when debugging the kernel.
3527  */
3528 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
3529 {
3530 	struct mount *mp;
3531 	struct vnode *vp;
3532 
3533 	/*
3534 	 * Note: because this is DDB, we can't obey the locking semantics
3535 	 * for these structures, which means we could catch an inconsistent
3536 	 * state and dereference a nasty pointer.  Not much to be done
3537 	 * about that.
3538 	 */
3539 	db_printf("Locked vnodes\n");
3540 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3541 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3542 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
3543 				vn_printf(vp, "vnode ");
3544 		}
3545 	}
3546 }
3547 
3548 /*
3549  * Show details about the given vnode.
3550  */
3551 DB_SHOW_COMMAND(vnode, db_show_vnode)
3552 {
3553 	struct vnode *vp;
3554 
3555 	if (!have_addr)
3556 		return;
3557 	vp = (struct vnode *)addr;
3558 	vn_printf(vp, "vnode ");
3559 }
3560 
3561 /*
3562  * Show details about the given mount point.
3563  */
3564 DB_SHOW_COMMAND(mount, db_show_mount)
3565 {
3566 	struct mount *mp;
3567 	struct vfsopt *opt;
3568 	struct statfs *sp;
3569 	struct vnode *vp;
3570 	char buf[512];
3571 	uint64_t mflags;
3572 	u_int flags;
3573 
3574 	if (!have_addr) {
3575 		/* No address given, print short info about all mount points. */
3576 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3577 			db_printf("%p %s on %s (%s)\n", mp,
3578 			    mp->mnt_stat.f_mntfromname,
3579 			    mp->mnt_stat.f_mntonname,
3580 			    mp->mnt_stat.f_fstypename);
3581 			if (db_pager_quit)
3582 				break;
3583 		}
3584 		db_printf("\nMore info: show mount <addr>\n");
3585 		return;
3586 	}
3587 
3588 	mp = (struct mount *)addr;
3589 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
3590 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
3591 
3592 	buf[0] = '\0';
3593 	mflags = mp->mnt_flag;
3594 #define	MNT_FLAG(flag)	do {						\
3595 	if (mflags & (flag)) {						\
3596 		if (buf[0] != '\0')					\
3597 			strlcat(buf, ", ", sizeof(buf));		\
3598 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
3599 		mflags &= ~(flag);					\
3600 	}								\
3601 } while (0)
3602 	MNT_FLAG(MNT_RDONLY);
3603 	MNT_FLAG(MNT_SYNCHRONOUS);
3604 	MNT_FLAG(MNT_NOEXEC);
3605 	MNT_FLAG(MNT_NOSUID);
3606 	MNT_FLAG(MNT_NFS4ACLS);
3607 	MNT_FLAG(MNT_UNION);
3608 	MNT_FLAG(MNT_ASYNC);
3609 	MNT_FLAG(MNT_SUIDDIR);
3610 	MNT_FLAG(MNT_SOFTDEP);
3611 	MNT_FLAG(MNT_NOSYMFOLLOW);
3612 	MNT_FLAG(MNT_GJOURNAL);
3613 	MNT_FLAG(MNT_MULTILABEL);
3614 	MNT_FLAG(MNT_ACLS);
3615 	MNT_FLAG(MNT_NOATIME);
3616 	MNT_FLAG(MNT_NOCLUSTERR);
3617 	MNT_FLAG(MNT_NOCLUSTERW);
3618 	MNT_FLAG(MNT_SUJ);
3619 	MNT_FLAG(MNT_EXRDONLY);
3620 	MNT_FLAG(MNT_EXPORTED);
3621 	MNT_FLAG(MNT_DEFEXPORTED);
3622 	MNT_FLAG(MNT_EXPORTANON);
3623 	MNT_FLAG(MNT_EXKERB);
3624 	MNT_FLAG(MNT_EXPUBLIC);
3625 	MNT_FLAG(MNT_LOCAL);
3626 	MNT_FLAG(MNT_QUOTA);
3627 	MNT_FLAG(MNT_ROOTFS);
3628 	MNT_FLAG(MNT_USER);
3629 	MNT_FLAG(MNT_IGNORE);
3630 	MNT_FLAG(MNT_UPDATE);
3631 	MNT_FLAG(MNT_DELEXPORT);
3632 	MNT_FLAG(MNT_RELOAD);
3633 	MNT_FLAG(MNT_FORCE);
3634 	MNT_FLAG(MNT_SNAPSHOT);
3635 	MNT_FLAG(MNT_BYFSID);
3636 #undef MNT_FLAG
3637 	if (mflags != 0) {
3638 		if (buf[0] != '\0')
3639 			strlcat(buf, ", ", sizeof(buf));
3640 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3641 		    "0x%016jx", mflags);
3642 	}
3643 	db_printf("    mnt_flag = %s\n", buf);
3644 
3645 	buf[0] = '\0';
3646 	flags = mp->mnt_kern_flag;
3647 #define	MNT_KERN_FLAG(flag)	do {					\
3648 	if (flags & (flag)) {						\
3649 		if (buf[0] != '\0')					\
3650 			strlcat(buf, ", ", sizeof(buf));		\
3651 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
3652 		flags &= ~(flag);					\
3653 	}								\
3654 } while (0)
3655 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
3656 	MNT_KERN_FLAG(MNTK_ASYNC);
3657 	MNT_KERN_FLAG(MNTK_SOFTDEP);
3658 	MNT_KERN_FLAG(MNTK_NOINSMNTQ);
3659 	MNT_KERN_FLAG(MNTK_DRAINING);
3660 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
3661 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
3662 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
3663 	MNT_KERN_FLAG(MNTK_NO_IOPF);
3664 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
3665 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
3666 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
3667 	MNT_KERN_FLAG(MNTK_MARKER);
3668 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
3669 	MNT_KERN_FLAG(MNTK_NOASYNC);
3670 	MNT_KERN_FLAG(MNTK_UNMOUNT);
3671 	MNT_KERN_FLAG(MNTK_MWAIT);
3672 	MNT_KERN_FLAG(MNTK_SUSPEND);
3673 	MNT_KERN_FLAG(MNTK_SUSPEND2);
3674 	MNT_KERN_FLAG(MNTK_SUSPENDED);
3675 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
3676 	MNT_KERN_FLAG(MNTK_NOKNOTE);
3677 #undef MNT_KERN_FLAG
3678 	if (flags != 0) {
3679 		if (buf[0] != '\0')
3680 			strlcat(buf, ", ", sizeof(buf));
3681 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3682 		    "0x%08x", flags);
3683 	}
3684 	db_printf("    mnt_kern_flag = %s\n", buf);
3685 
3686 	db_printf("    mnt_opt = ");
3687 	opt = TAILQ_FIRST(mp->mnt_opt);
3688 	if (opt != NULL) {
3689 		db_printf("%s", opt->name);
3690 		opt = TAILQ_NEXT(opt, link);
3691 		while (opt != NULL) {
3692 			db_printf(", %s", opt->name);
3693 			opt = TAILQ_NEXT(opt, link);
3694 		}
3695 	}
3696 	db_printf("\n");
3697 
3698 	sp = &mp->mnt_stat;
3699 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
3700 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
3701 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
3702 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
3703 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
3704 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
3705 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
3706 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
3707 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
3708 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
3709 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
3710 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
3711 
3712 	db_printf("    mnt_cred = { uid=%u ruid=%u",
3713 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
3714 	if (jailed(mp->mnt_cred))
3715 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
3716 	db_printf(" }\n");
3717 	db_printf("    mnt_ref = %d\n", mp->mnt_ref);
3718 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
3719 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
3720 	db_printf("    mnt_activevnodelistsize = %d\n",
3721 	    mp->mnt_activevnodelistsize);
3722 	db_printf("    mnt_writeopcount = %d\n", mp->mnt_writeopcount);
3723 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
3724 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
3725 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
3726 	db_printf("    mnt_lockref = %d\n", mp->mnt_lockref);
3727 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
3728 	db_printf("    mnt_secondary_accwrites = %d\n",
3729 	    mp->mnt_secondary_accwrites);
3730 	db_printf("    mnt_gjprovider = %s\n",
3731 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
3732 
3733 	db_printf("\n\nList of active vnodes\n");
3734 	TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) {
3735 		if (vp->v_type != VMARKER) {
3736 			vn_printf(vp, "vnode ");
3737 			if (db_pager_quit)
3738 				break;
3739 		}
3740 	}
3741 	db_printf("\n\nList of inactive vnodes\n");
3742 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3743 		if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) {
3744 			vn_printf(vp, "vnode ");
3745 			if (db_pager_quit)
3746 				break;
3747 		}
3748 	}
3749 }
3750 #endif	/* DDB */
3751 
3752 /*
3753  * Fill in a struct xvfsconf based on a struct vfsconf.
3754  */
3755 static int
3756 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
3757 {
3758 	struct xvfsconf xvfsp;
3759 
3760 	bzero(&xvfsp, sizeof(xvfsp));
3761 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3762 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3763 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3764 	xvfsp.vfc_flags = vfsp->vfc_flags;
3765 	/*
3766 	 * These are unused in userland, we keep them
3767 	 * to not break binary compatibility.
3768 	 */
3769 	xvfsp.vfc_vfsops = NULL;
3770 	xvfsp.vfc_next = NULL;
3771 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3772 }
3773 
3774 #ifdef COMPAT_FREEBSD32
3775 struct xvfsconf32 {
3776 	uint32_t	vfc_vfsops;
3777 	char		vfc_name[MFSNAMELEN];
3778 	int32_t		vfc_typenum;
3779 	int32_t		vfc_refcount;
3780 	int32_t		vfc_flags;
3781 	uint32_t	vfc_next;
3782 };
3783 
3784 static int
3785 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
3786 {
3787 	struct xvfsconf32 xvfsp;
3788 
3789 	bzero(&xvfsp, sizeof(xvfsp));
3790 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3791 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3792 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3793 	xvfsp.vfc_flags = vfsp->vfc_flags;
3794 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3795 }
3796 #endif
3797 
3798 /*
3799  * Top level filesystem related information gathering.
3800  */
3801 static int
3802 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
3803 {
3804 	struct vfsconf *vfsp;
3805 	int error;
3806 
3807 	error = 0;
3808 	vfsconf_slock();
3809 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3810 #ifdef COMPAT_FREEBSD32
3811 		if (req->flags & SCTL_MASK32)
3812 			error = vfsconf2x32(req, vfsp);
3813 		else
3814 #endif
3815 			error = vfsconf2x(req, vfsp);
3816 		if (error)
3817 			break;
3818 	}
3819 	vfsconf_sunlock();
3820 	return (error);
3821 }
3822 
3823 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
3824     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
3825     "S,xvfsconf", "List of all configured filesystems");
3826 
3827 #ifndef BURN_BRIDGES
3828 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
3829 
3830 static int
3831 vfs_sysctl(SYSCTL_HANDLER_ARGS)
3832 {
3833 	int *name = (int *)arg1 - 1;	/* XXX */
3834 	u_int namelen = arg2 + 1;	/* XXX */
3835 	struct vfsconf *vfsp;
3836 
3837 	log(LOG_WARNING, "userland calling deprecated sysctl, "
3838 	    "please rebuild world\n");
3839 
3840 #if 1 || defined(COMPAT_PRELITE2)
3841 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
3842 	if (namelen == 1)
3843 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
3844 #endif
3845 
3846 	switch (name[1]) {
3847 	case VFS_MAXTYPENUM:
3848 		if (namelen != 2)
3849 			return (ENOTDIR);
3850 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
3851 	case VFS_CONF:
3852 		if (namelen != 3)
3853 			return (ENOTDIR);	/* overloaded */
3854 		vfsconf_slock();
3855 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3856 			if (vfsp->vfc_typenum == name[2])
3857 				break;
3858 		}
3859 		vfsconf_sunlock();
3860 		if (vfsp == NULL)
3861 			return (EOPNOTSUPP);
3862 #ifdef COMPAT_FREEBSD32
3863 		if (req->flags & SCTL_MASK32)
3864 			return (vfsconf2x32(req, vfsp));
3865 		else
3866 #endif
3867 			return (vfsconf2x(req, vfsp));
3868 	}
3869 	return (EOPNOTSUPP);
3870 }
3871 
3872 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
3873     CTLFLAG_MPSAFE, vfs_sysctl,
3874     "Generic filesystem");
3875 
3876 #if 1 || defined(COMPAT_PRELITE2)
3877 
3878 static int
3879 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
3880 {
3881 	int error;
3882 	struct vfsconf *vfsp;
3883 	struct ovfsconf ovfs;
3884 
3885 	vfsconf_slock();
3886 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3887 		bzero(&ovfs, sizeof(ovfs));
3888 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
3889 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
3890 		ovfs.vfc_index = vfsp->vfc_typenum;
3891 		ovfs.vfc_refcount = vfsp->vfc_refcount;
3892 		ovfs.vfc_flags = vfsp->vfc_flags;
3893 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
3894 		if (error != 0) {
3895 			vfsconf_sunlock();
3896 			return (error);
3897 		}
3898 	}
3899 	vfsconf_sunlock();
3900 	return (0);
3901 }
3902 
3903 #endif /* 1 || COMPAT_PRELITE2 */
3904 #endif /* !BURN_BRIDGES */
3905 
3906 #define KINFO_VNODESLOP		10
3907 #ifdef notyet
3908 /*
3909  * Dump vnode list (via sysctl).
3910  */
3911 /* ARGSUSED */
3912 static int
3913 sysctl_vnode(SYSCTL_HANDLER_ARGS)
3914 {
3915 	struct xvnode *xvn;
3916 	struct mount *mp;
3917 	struct vnode *vp;
3918 	int error, len, n;
3919 
3920 	/*
3921 	 * Stale numvnodes access is not fatal here.
3922 	 */
3923 	req->lock = 0;
3924 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
3925 	if (!req->oldptr)
3926 		/* Make an estimate */
3927 		return (SYSCTL_OUT(req, 0, len));
3928 
3929 	error = sysctl_wire_old_buffer(req, 0);
3930 	if (error != 0)
3931 		return (error);
3932 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
3933 	n = 0;
3934 	mtx_lock(&mountlist_mtx);
3935 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3936 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
3937 			continue;
3938 		MNT_ILOCK(mp);
3939 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3940 			if (n == len)
3941 				break;
3942 			vref(vp);
3943 			xvn[n].xv_size = sizeof *xvn;
3944 			xvn[n].xv_vnode = vp;
3945 			xvn[n].xv_id = 0;	/* XXX compat */
3946 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
3947 			XV_COPY(usecount);
3948 			XV_COPY(writecount);
3949 			XV_COPY(holdcnt);
3950 			XV_COPY(mount);
3951 			XV_COPY(numoutput);
3952 			XV_COPY(type);
3953 #undef XV_COPY
3954 			xvn[n].xv_flag = vp->v_vflag;
3955 
3956 			switch (vp->v_type) {
3957 			case VREG:
3958 			case VDIR:
3959 			case VLNK:
3960 				break;
3961 			case VBLK:
3962 			case VCHR:
3963 				if (vp->v_rdev == NULL) {
3964 					vrele(vp);
3965 					continue;
3966 				}
3967 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
3968 				break;
3969 			case VSOCK:
3970 				xvn[n].xv_socket = vp->v_socket;
3971 				break;
3972 			case VFIFO:
3973 				xvn[n].xv_fifo = vp->v_fifoinfo;
3974 				break;
3975 			case VNON:
3976 			case VBAD:
3977 			default:
3978 				/* shouldn't happen? */
3979 				vrele(vp);
3980 				continue;
3981 			}
3982 			vrele(vp);
3983 			++n;
3984 		}
3985 		MNT_IUNLOCK(mp);
3986 		mtx_lock(&mountlist_mtx);
3987 		vfs_unbusy(mp);
3988 		if (n == len)
3989 			break;
3990 	}
3991 	mtx_unlock(&mountlist_mtx);
3992 
3993 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
3994 	free(xvn, M_TEMP);
3995 	return (error);
3996 }
3997 
3998 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
3999     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4000     "");
4001 #endif
4002 
4003 static void
4004 unmount_or_warn(struct mount *mp)
4005 {
4006 	int error;
4007 
4008 	error = dounmount(mp, MNT_FORCE, curthread);
4009 	if (error != 0) {
4010 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4011 		if (error == EBUSY)
4012 			printf("BUSY)\n");
4013 		else
4014 			printf("%d)\n", error);
4015 	}
4016 }
4017 
4018 /*
4019  * Unmount all filesystems. The list is traversed in reverse order
4020  * of mounting to avoid dependencies.
4021  */
4022 void
4023 vfs_unmountall(void)
4024 {
4025 	struct mount *mp, *tmp;
4026 
4027 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4028 
4029 	/*
4030 	 * Since this only runs when rebooting, it is not interlocked.
4031 	 */
4032 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4033 		vfs_ref(mp);
4034 
4035 		/*
4036 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4037 		 * unmount of the latter.
4038 		 */
4039 		if (mp == rootdevmp)
4040 			continue;
4041 
4042 		unmount_or_warn(mp);
4043 	}
4044 
4045 	if (rootdevmp != NULL)
4046 		unmount_or_warn(rootdevmp);
4047 }
4048 
4049 /*
4050  * perform msync on all vnodes under a mount point
4051  * the mount point must be locked.
4052  */
4053 void
4054 vfs_msync(struct mount *mp, int flags)
4055 {
4056 	struct vnode *vp, *mvp;
4057 	struct vm_object *obj;
4058 
4059 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4060 
4061 	vnlru_return_batch(mp);
4062 
4063 	MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
4064 		obj = vp->v_object;
4065 		if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 &&
4066 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
4067 			if (!vget(vp,
4068 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
4069 			    curthread)) {
4070 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
4071 					vput(vp);
4072 					continue;
4073 				}
4074 
4075 				obj = vp->v_object;
4076 				if (obj != NULL) {
4077 					VM_OBJECT_WLOCK(obj);
4078 					vm_object_page_clean(obj, 0, 0,
4079 					    flags == MNT_WAIT ?
4080 					    OBJPC_SYNC : OBJPC_NOSYNC);
4081 					VM_OBJECT_WUNLOCK(obj);
4082 				}
4083 				vput(vp);
4084 			}
4085 		} else
4086 			VI_UNLOCK(vp);
4087 	}
4088 }
4089 
4090 static void
4091 destroy_vpollinfo_free(struct vpollinfo *vi)
4092 {
4093 
4094 	knlist_destroy(&vi->vpi_selinfo.si_note);
4095 	mtx_destroy(&vi->vpi_lock);
4096 	uma_zfree(vnodepoll_zone, vi);
4097 }
4098 
4099 static void
4100 destroy_vpollinfo(struct vpollinfo *vi)
4101 {
4102 
4103 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4104 	seldrain(&vi->vpi_selinfo);
4105 	destroy_vpollinfo_free(vi);
4106 }
4107 
4108 /*
4109  * Initialize per-vnode helper structure to hold poll-related state.
4110  */
4111 void
4112 v_addpollinfo(struct vnode *vp)
4113 {
4114 	struct vpollinfo *vi;
4115 
4116 	if (vp->v_pollinfo != NULL)
4117 		return;
4118 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4119 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4120 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4121 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4122 	VI_LOCK(vp);
4123 	if (vp->v_pollinfo != NULL) {
4124 		VI_UNLOCK(vp);
4125 		destroy_vpollinfo_free(vi);
4126 		return;
4127 	}
4128 	vp->v_pollinfo = vi;
4129 	VI_UNLOCK(vp);
4130 }
4131 
4132 /*
4133  * Record a process's interest in events which might happen to
4134  * a vnode.  Because poll uses the historic select-style interface
4135  * internally, this routine serves as both the ``check for any
4136  * pending events'' and the ``record my interest in future events''
4137  * functions.  (These are done together, while the lock is held,
4138  * to avoid race conditions.)
4139  */
4140 int
4141 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4142 {
4143 
4144 	v_addpollinfo(vp);
4145 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4146 	if (vp->v_pollinfo->vpi_revents & events) {
4147 		/*
4148 		 * This leaves events we are not interested
4149 		 * in available for the other process which
4150 		 * which presumably had requested them
4151 		 * (otherwise they would never have been
4152 		 * recorded).
4153 		 */
4154 		events &= vp->v_pollinfo->vpi_revents;
4155 		vp->v_pollinfo->vpi_revents &= ~events;
4156 
4157 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4158 		return (events);
4159 	}
4160 	vp->v_pollinfo->vpi_events |= events;
4161 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4162 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4163 	return (0);
4164 }
4165 
4166 /*
4167  * Routine to create and manage a filesystem syncer vnode.
4168  */
4169 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4170 static int	sync_fsync(struct  vop_fsync_args *);
4171 static int	sync_inactive(struct  vop_inactive_args *);
4172 static int	sync_reclaim(struct  vop_reclaim_args *);
4173 
4174 static struct vop_vector sync_vnodeops = {
4175 	.vop_bypass =	VOP_EOPNOTSUPP,
4176 	.vop_close =	sync_close,		/* close */
4177 	.vop_fsync =	sync_fsync,		/* fsync */
4178 	.vop_inactive =	sync_inactive,	/* inactive */
4179 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4180 	.vop_lock1 =	vop_stdlock,	/* lock */
4181 	.vop_unlock =	vop_stdunlock,	/* unlock */
4182 	.vop_islocked =	vop_stdislocked,	/* islocked */
4183 };
4184 
4185 /*
4186  * Create a new filesystem syncer vnode for the specified mount point.
4187  */
4188 void
4189 vfs_allocate_syncvnode(struct mount *mp)
4190 {
4191 	struct vnode *vp;
4192 	struct bufobj *bo;
4193 	static long start, incr, next;
4194 	int error;
4195 
4196 	/* Allocate a new vnode */
4197 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4198 	if (error != 0)
4199 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4200 	vp->v_type = VNON;
4201 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4202 	vp->v_vflag |= VV_FORCEINSMQ;
4203 	error = insmntque(vp, mp);
4204 	if (error != 0)
4205 		panic("vfs_allocate_syncvnode: insmntque() failed");
4206 	vp->v_vflag &= ~VV_FORCEINSMQ;
4207 	VOP_UNLOCK(vp, 0);
4208 	/*
4209 	 * Place the vnode onto the syncer worklist. We attempt to
4210 	 * scatter them about on the list so that they will go off
4211 	 * at evenly distributed times even if all the filesystems
4212 	 * are mounted at once.
4213 	 */
4214 	next += incr;
4215 	if (next == 0 || next > syncer_maxdelay) {
4216 		start /= 2;
4217 		incr /= 2;
4218 		if (start == 0) {
4219 			start = syncer_maxdelay / 2;
4220 			incr = syncer_maxdelay;
4221 		}
4222 		next = start;
4223 	}
4224 	bo = &vp->v_bufobj;
4225 	BO_LOCK(bo);
4226 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4227 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4228 	mtx_lock(&sync_mtx);
4229 	sync_vnode_count++;
4230 	if (mp->mnt_syncer == NULL) {
4231 		mp->mnt_syncer = vp;
4232 		vp = NULL;
4233 	}
4234 	mtx_unlock(&sync_mtx);
4235 	BO_UNLOCK(bo);
4236 	if (vp != NULL) {
4237 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4238 		vgone(vp);
4239 		vput(vp);
4240 	}
4241 }
4242 
4243 void
4244 vfs_deallocate_syncvnode(struct mount *mp)
4245 {
4246 	struct vnode *vp;
4247 
4248 	mtx_lock(&sync_mtx);
4249 	vp = mp->mnt_syncer;
4250 	if (vp != NULL)
4251 		mp->mnt_syncer = NULL;
4252 	mtx_unlock(&sync_mtx);
4253 	if (vp != NULL)
4254 		vrele(vp);
4255 }
4256 
4257 /*
4258  * Do a lazy sync of the filesystem.
4259  */
4260 static int
4261 sync_fsync(struct vop_fsync_args *ap)
4262 {
4263 	struct vnode *syncvp = ap->a_vp;
4264 	struct mount *mp = syncvp->v_mount;
4265 	int error, save;
4266 	struct bufobj *bo;
4267 
4268 	/*
4269 	 * We only need to do something if this is a lazy evaluation.
4270 	 */
4271 	if (ap->a_waitfor != MNT_LAZY)
4272 		return (0);
4273 
4274 	/*
4275 	 * Move ourselves to the back of the sync list.
4276 	 */
4277 	bo = &syncvp->v_bufobj;
4278 	BO_LOCK(bo);
4279 	vn_syncer_add_to_worklist(bo, syncdelay);
4280 	BO_UNLOCK(bo);
4281 
4282 	/*
4283 	 * Walk the list of vnodes pushing all that are dirty and
4284 	 * not already on the sync list.
4285 	 */
4286 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4287 		return (0);
4288 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4289 		vfs_unbusy(mp);
4290 		return (0);
4291 	}
4292 	save = curthread_pflags_set(TDP_SYNCIO);
4293 	vfs_msync(mp, MNT_NOWAIT);
4294 	error = VFS_SYNC(mp, MNT_LAZY);
4295 	curthread_pflags_restore(save);
4296 	vn_finished_write(mp);
4297 	vfs_unbusy(mp);
4298 	return (error);
4299 }
4300 
4301 /*
4302  * The syncer vnode is no referenced.
4303  */
4304 static int
4305 sync_inactive(struct vop_inactive_args *ap)
4306 {
4307 
4308 	vgone(ap->a_vp);
4309 	return (0);
4310 }
4311 
4312 /*
4313  * The syncer vnode is no longer needed and is being decommissioned.
4314  *
4315  * Modifications to the worklist must be protected by sync_mtx.
4316  */
4317 static int
4318 sync_reclaim(struct vop_reclaim_args *ap)
4319 {
4320 	struct vnode *vp = ap->a_vp;
4321 	struct bufobj *bo;
4322 
4323 	bo = &vp->v_bufobj;
4324 	BO_LOCK(bo);
4325 	mtx_lock(&sync_mtx);
4326 	if (vp->v_mount->mnt_syncer == vp)
4327 		vp->v_mount->mnt_syncer = NULL;
4328 	if (bo->bo_flag & BO_ONWORKLST) {
4329 		LIST_REMOVE(bo, bo_synclist);
4330 		syncer_worklist_len--;
4331 		sync_vnode_count--;
4332 		bo->bo_flag &= ~BO_ONWORKLST;
4333 	}
4334 	mtx_unlock(&sync_mtx);
4335 	BO_UNLOCK(bo);
4336 
4337 	return (0);
4338 }
4339 
4340 /*
4341  * Check if vnode represents a disk device
4342  */
4343 int
4344 vn_isdisk(struct vnode *vp, int *errp)
4345 {
4346 	int error;
4347 
4348 	if (vp->v_type != VCHR) {
4349 		error = ENOTBLK;
4350 		goto out;
4351 	}
4352 	error = 0;
4353 	dev_lock();
4354 	if (vp->v_rdev == NULL)
4355 		error = ENXIO;
4356 	else if (vp->v_rdev->si_devsw == NULL)
4357 		error = ENXIO;
4358 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
4359 		error = ENOTBLK;
4360 	dev_unlock();
4361 out:
4362 	if (errp != NULL)
4363 		*errp = error;
4364 	return (error == 0);
4365 }
4366 
4367 /*
4368  * Common filesystem object access control check routine.  Accepts a
4369  * vnode's type, "mode", uid and gid, requested access mode, credentials,
4370  * and optional call-by-reference privused argument allowing vaccess()
4371  * to indicate to the caller whether privilege was used to satisfy the
4372  * request (obsoleted).  Returns 0 on success, or an errno on failure.
4373  */
4374 int
4375 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
4376     accmode_t accmode, struct ucred *cred, int *privused)
4377 {
4378 	accmode_t dac_granted;
4379 	accmode_t priv_granted;
4380 
4381 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
4382 	    ("invalid bit in accmode"));
4383 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
4384 	    ("VAPPEND without VWRITE"));
4385 
4386 	/*
4387 	 * Look for a normal, non-privileged way to access the file/directory
4388 	 * as requested.  If it exists, go with that.
4389 	 */
4390 
4391 	if (privused != NULL)
4392 		*privused = 0;
4393 
4394 	dac_granted = 0;
4395 
4396 	/* Check the owner. */
4397 	if (cred->cr_uid == file_uid) {
4398 		dac_granted |= VADMIN;
4399 		if (file_mode & S_IXUSR)
4400 			dac_granted |= VEXEC;
4401 		if (file_mode & S_IRUSR)
4402 			dac_granted |= VREAD;
4403 		if (file_mode & S_IWUSR)
4404 			dac_granted |= (VWRITE | VAPPEND);
4405 
4406 		if ((accmode & dac_granted) == accmode)
4407 			return (0);
4408 
4409 		goto privcheck;
4410 	}
4411 
4412 	/* Otherwise, check the groups (first match) */
4413 	if (groupmember(file_gid, cred)) {
4414 		if (file_mode & S_IXGRP)
4415 			dac_granted |= VEXEC;
4416 		if (file_mode & S_IRGRP)
4417 			dac_granted |= VREAD;
4418 		if (file_mode & S_IWGRP)
4419 			dac_granted |= (VWRITE | VAPPEND);
4420 
4421 		if ((accmode & dac_granted) == accmode)
4422 			return (0);
4423 
4424 		goto privcheck;
4425 	}
4426 
4427 	/* Otherwise, check everyone else. */
4428 	if (file_mode & S_IXOTH)
4429 		dac_granted |= VEXEC;
4430 	if (file_mode & S_IROTH)
4431 		dac_granted |= VREAD;
4432 	if (file_mode & S_IWOTH)
4433 		dac_granted |= (VWRITE | VAPPEND);
4434 	if ((accmode & dac_granted) == accmode)
4435 		return (0);
4436 
4437 privcheck:
4438 	/*
4439 	 * Build a privilege mask to determine if the set of privileges
4440 	 * satisfies the requirements when combined with the granted mask
4441 	 * from above.  For each privilege, if the privilege is required,
4442 	 * bitwise or the request type onto the priv_granted mask.
4443 	 */
4444 	priv_granted = 0;
4445 
4446 	if (type == VDIR) {
4447 		/*
4448 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
4449 		 * requests, instead of PRIV_VFS_EXEC.
4450 		 */
4451 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4452 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP, 0))
4453 			priv_granted |= VEXEC;
4454 	} else {
4455 		/*
4456 		 * Ensure that at least one execute bit is on. Otherwise,
4457 		 * a privileged user will always succeed, and we don't want
4458 		 * this to happen unless the file really is executable.
4459 		 */
4460 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4461 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
4462 		    !priv_check_cred(cred, PRIV_VFS_EXEC, 0))
4463 			priv_granted |= VEXEC;
4464 	}
4465 
4466 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
4467 	    !priv_check_cred(cred, PRIV_VFS_READ, 0))
4468 		priv_granted |= VREAD;
4469 
4470 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
4471 	    !priv_check_cred(cred, PRIV_VFS_WRITE, 0))
4472 		priv_granted |= (VWRITE | VAPPEND);
4473 
4474 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
4475 	    !priv_check_cred(cred, PRIV_VFS_ADMIN, 0))
4476 		priv_granted |= VADMIN;
4477 
4478 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
4479 		/* XXX audit: privilege used */
4480 		if (privused != NULL)
4481 			*privused = 1;
4482 		return (0);
4483 	}
4484 
4485 	return ((accmode & VADMIN) ? EPERM : EACCES);
4486 }
4487 
4488 /*
4489  * Credential check based on process requesting service, and per-attribute
4490  * permissions.
4491  */
4492 int
4493 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
4494     struct thread *td, accmode_t accmode)
4495 {
4496 
4497 	/*
4498 	 * Kernel-invoked always succeeds.
4499 	 */
4500 	if (cred == NOCRED)
4501 		return (0);
4502 
4503 	/*
4504 	 * Do not allow privileged processes in jail to directly manipulate
4505 	 * system attributes.
4506 	 */
4507 	switch (attrnamespace) {
4508 	case EXTATTR_NAMESPACE_SYSTEM:
4509 		/* Potentially should be: return (EPERM); */
4510 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0));
4511 	case EXTATTR_NAMESPACE_USER:
4512 		return (VOP_ACCESS(vp, accmode, cred, td));
4513 	default:
4514 		return (EPERM);
4515 	}
4516 }
4517 
4518 #ifdef DEBUG_VFS_LOCKS
4519 /*
4520  * This only exists to suppress warnings from unlocked specfs accesses.  It is
4521  * no longer ok to have an unlocked VFS.
4522  */
4523 #define	IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL ||		\
4524 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
4525 
4526 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
4527 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
4528     "Drop into debugger on lock violation");
4529 
4530 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
4531 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
4532     0, "Check for interlock across VOPs");
4533 
4534 int vfs_badlock_print = 1;	/* Print lock violations. */
4535 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
4536     0, "Print lock violations");
4537 
4538 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
4539 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
4540     0, "Print vnode details on lock violations");
4541 
4542 #ifdef KDB
4543 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
4544 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
4545     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
4546 #endif
4547 
4548 static void
4549 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
4550 {
4551 
4552 #ifdef KDB
4553 	if (vfs_badlock_backtrace)
4554 		kdb_backtrace();
4555 #endif
4556 	if (vfs_badlock_vnode)
4557 		vn_printf(vp, "vnode ");
4558 	if (vfs_badlock_print)
4559 		printf("%s: %p %s\n", str, (void *)vp, msg);
4560 	if (vfs_badlock_ddb)
4561 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4562 }
4563 
4564 void
4565 assert_vi_locked(struct vnode *vp, const char *str)
4566 {
4567 
4568 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
4569 		vfs_badlock("interlock is not locked but should be", str, vp);
4570 }
4571 
4572 void
4573 assert_vi_unlocked(struct vnode *vp, const char *str)
4574 {
4575 
4576 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
4577 		vfs_badlock("interlock is locked but should not be", str, vp);
4578 }
4579 
4580 void
4581 assert_vop_locked(struct vnode *vp, const char *str)
4582 {
4583 	int locked;
4584 
4585 	if (!IGNORE_LOCK(vp)) {
4586 		locked = VOP_ISLOCKED(vp);
4587 		if (locked == 0 || locked == LK_EXCLOTHER)
4588 			vfs_badlock("is not locked but should be", str, vp);
4589 	}
4590 }
4591 
4592 void
4593 assert_vop_unlocked(struct vnode *vp, const char *str)
4594 {
4595 
4596 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
4597 		vfs_badlock("is locked but should not be", str, vp);
4598 }
4599 
4600 void
4601 assert_vop_elocked(struct vnode *vp, const char *str)
4602 {
4603 
4604 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
4605 		vfs_badlock("is not exclusive locked but should be", str, vp);
4606 }
4607 #endif /* DEBUG_VFS_LOCKS */
4608 
4609 void
4610 vop_rename_fail(struct vop_rename_args *ap)
4611 {
4612 
4613 	if (ap->a_tvp != NULL)
4614 		vput(ap->a_tvp);
4615 	if (ap->a_tdvp == ap->a_tvp)
4616 		vrele(ap->a_tdvp);
4617 	else
4618 		vput(ap->a_tdvp);
4619 	vrele(ap->a_fdvp);
4620 	vrele(ap->a_fvp);
4621 }
4622 
4623 void
4624 vop_rename_pre(void *ap)
4625 {
4626 	struct vop_rename_args *a = ap;
4627 
4628 #ifdef DEBUG_VFS_LOCKS
4629 	if (a->a_tvp)
4630 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
4631 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
4632 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
4633 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
4634 
4635 	/* Check the source (from). */
4636 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
4637 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
4638 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
4639 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
4640 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
4641 
4642 	/* Check the target. */
4643 	if (a->a_tvp)
4644 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
4645 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
4646 #endif
4647 	if (a->a_tdvp != a->a_fdvp)
4648 		vhold(a->a_fdvp);
4649 	if (a->a_tvp != a->a_fvp)
4650 		vhold(a->a_fvp);
4651 	vhold(a->a_tdvp);
4652 	if (a->a_tvp)
4653 		vhold(a->a_tvp);
4654 }
4655 
4656 #ifdef DEBUG_VFS_LOCKS
4657 void
4658 vop_strategy_pre(void *ap)
4659 {
4660 	struct vop_strategy_args *a;
4661 	struct buf *bp;
4662 
4663 	a = ap;
4664 	bp = a->a_bp;
4665 
4666 	/*
4667 	 * Cluster ops lock their component buffers but not the IO container.
4668 	 */
4669 	if ((bp->b_flags & B_CLUSTER) != 0)
4670 		return;
4671 
4672 	if (panicstr == NULL && !BUF_ISLOCKED(bp)) {
4673 		if (vfs_badlock_print)
4674 			printf(
4675 			    "VOP_STRATEGY: bp is not locked but should be\n");
4676 		if (vfs_badlock_ddb)
4677 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4678 	}
4679 }
4680 
4681 void
4682 vop_lock_pre(void *ap)
4683 {
4684 	struct vop_lock1_args *a = ap;
4685 
4686 	if ((a->a_flags & LK_INTERLOCK) == 0)
4687 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4688 	else
4689 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
4690 }
4691 
4692 void
4693 vop_lock_post(void *ap, int rc)
4694 {
4695 	struct vop_lock1_args *a = ap;
4696 
4697 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4698 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
4699 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
4700 }
4701 
4702 void
4703 vop_unlock_pre(void *ap)
4704 {
4705 	struct vop_unlock_args *a = ap;
4706 
4707 	if (a->a_flags & LK_INTERLOCK)
4708 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
4709 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
4710 }
4711 
4712 void
4713 vop_unlock_post(void *ap, int rc)
4714 {
4715 	struct vop_unlock_args *a = ap;
4716 
4717 	if (a->a_flags & LK_INTERLOCK)
4718 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
4719 }
4720 #endif
4721 
4722 void
4723 vop_create_post(void *ap, int rc)
4724 {
4725 	struct vop_create_args *a = ap;
4726 
4727 	if (!rc)
4728 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4729 }
4730 
4731 void
4732 vop_deleteextattr_post(void *ap, int rc)
4733 {
4734 	struct vop_deleteextattr_args *a = ap;
4735 
4736 	if (!rc)
4737 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4738 }
4739 
4740 void
4741 vop_link_post(void *ap, int rc)
4742 {
4743 	struct vop_link_args *a = ap;
4744 
4745 	if (!rc) {
4746 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
4747 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
4748 	}
4749 }
4750 
4751 void
4752 vop_mkdir_post(void *ap, int rc)
4753 {
4754 	struct vop_mkdir_args *a = ap;
4755 
4756 	if (!rc)
4757 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4758 }
4759 
4760 void
4761 vop_mknod_post(void *ap, int rc)
4762 {
4763 	struct vop_mknod_args *a = ap;
4764 
4765 	if (!rc)
4766 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4767 }
4768 
4769 void
4770 vop_reclaim_post(void *ap, int rc)
4771 {
4772 	struct vop_reclaim_args *a = ap;
4773 
4774 	if (!rc)
4775 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
4776 }
4777 
4778 void
4779 vop_remove_post(void *ap, int rc)
4780 {
4781 	struct vop_remove_args *a = ap;
4782 
4783 	if (!rc) {
4784 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4785 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4786 	}
4787 }
4788 
4789 void
4790 vop_rename_post(void *ap, int rc)
4791 {
4792 	struct vop_rename_args *a = ap;
4793 	long hint;
4794 
4795 	if (!rc) {
4796 		hint = NOTE_WRITE;
4797 		if (a->a_fdvp == a->a_tdvp) {
4798 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
4799 				hint |= NOTE_LINK;
4800 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
4801 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
4802 		} else {
4803 			hint |= NOTE_EXTEND;
4804 			if (a->a_fvp->v_type == VDIR)
4805 				hint |= NOTE_LINK;
4806 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
4807 
4808 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
4809 			    a->a_tvp->v_type == VDIR)
4810 				hint &= ~NOTE_LINK;
4811 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
4812 		}
4813 
4814 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
4815 		if (a->a_tvp)
4816 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
4817 	}
4818 	if (a->a_tdvp != a->a_fdvp)
4819 		vdrop(a->a_fdvp);
4820 	if (a->a_tvp != a->a_fvp)
4821 		vdrop(a->a_fvp);
4822 	vdrop(a->a_tdvp);
4823 	if (a->a_tvp)
4824 		vdrop(a->a_tvp);
4825 }
4826 
4827 void
4828 vop_rmdir_post(void *ap, int rc)
4829 {
4830 	struct vop_rmdir_args *a = ap;
4831 
4832 	if (!rc) {
4833 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4834 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4835 	}
4836 }
4837 
4838 void
4839 vop_setattr_post(void *ap, int rc)
4840 {
4841 	struct vop_setattr_args *a = ap;
4842 
4843 	if (!rc)
4844 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4845 }
4846 
4847 void
4848 vop_setextattr_post(void *ap, int rc)
4849 {
4850 	struct vop_setextattr_args *a = ap;
4851 
4852 	if (!rc)
4853 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4854 }
4855 
4856 void
4857 vop_symlink_post(void *ap, int rc)
4858 {
4859 	struct vop_symlink_args *a = ap;
4860 
4861 	if (!rc)
4862 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4863 }
4864 
4865 void
4866 vop_open_post(void *ap, int rc)
4867 {
4868 	struct vop_open_args *a = ap;
4869 
4870 	if (!rc)
4871 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
4872 }
4873 
4874 void
4875 vop_close_post(void *ap, int rc)
4876 {
4877 	struct vop_close_args *a = ap;
4878 
4879 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
4880 	    (a->a_vp->v_iflag & VI_DOOMED) == 0)) {
4881 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
4882 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
4883 	}
4884 }
4885 
4886 void
4887 vop_read_post(void *ap, int rc)
4888 {
4889 	struct vop_read_args *a = ap;
4890 
4891 	if (!rc)
4892 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
4893 }
4894 
4895 void
4896 vop_readdir_post(void *ap, int rc)
4897 {
4898 	struct vop_readdir_args *a = ap;
4899 
4900 	if (!rc)
4901 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
4902 }
4903 
4904 static struct knlist fs_knlist;
4905 
4906 static void
4907 vfs_event_init(void *arg)
4908 {
4909 	knlist_init_mtx(&fs_knlist, NULL);
4910 }
4911 /* XXX - correct order? */
4912 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
4913 
4914 void
4915 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
4916 {
4917 
4918 	KNOTE_UNLOCKED(&fs_knlist, event);
4919 }
4920 
4921 static int	filt_fsattach(struct knote *kn);
4922 static void	filt_fsdetach(struct knote *kn);
4923 static int	filt_fsevent(struct knote *kn, long hint);
4924 
4925 struct filterops fs_filtops = {
4926 	.f_isfd = 0,
4927 	.f_attach = filt_fsattach,
4928 	.f_detach = filt_fsdetach,
4929 	.f_event = filt_fsevent
4930 };
4931 
4932 static int
4933 filt_fsattach(struct knote *kn)
4934 {
4935 
4936 	kn->kn_flags |= EV_CLEAR;
4937 	knlist_add(&fs_knlist, kn, 0);
4938 	return (0);
4939 }
4940 
4941 static void
4942 filt_fsdetach(struct knote *kn)
4943 {
4944 
4945 	knlist_remove(&fs_knlist, kn, 0);
4946 }
4947 
4948 static int
4949 filt_fsevent(struct knote *kn, long hint)
4950 {
4951 
4952 	kn->kn_fflags |= hint;
4953 	return (kn->kn_fflags != 0);
4954 }
4955 
4956 static int
4957 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
4958 {
4959 	struct vfsidctl vc;
4960 	int error;
4961 	struct mount *mp;
4962 
4963 	error = SYSCTL_IN(req, &vc, sizeof(vc));
4964 	if (error)
4965 		return (error);
4966 	if (vc.vc_vers != VFS_CTL_VERS1)
4967 		return (EINVAL);
4968 	mp = vfs_getvfs(&vc.vc_fsid);
4969 	if (mp == NULL)
4970 		return (ENOENT);
4971 	/* ensure that a specific sysctl goes to the right filesystem. */
4972 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
4973 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
4974 		vfs_rel(mp);
4975 		return (EINVAL);
4976 	}
4977 	VCTLTOREQ(&vc, req);
4978 	error = VFS_SYSCTL(mp, vc.vc_op, req);
4979 	vfs_rel(mp);
4980 	return (error);
4981 }
4982 
4983 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR,
4984     NULL, 0, sysctl_vfs_ctl, "",
4985     "Sysctl by fsid");
4986 
4987 /*
4988  * Function to initialize a va_filerev field sensibly.
4989  * XXX: Wouldn't a random number make a lot more sense ??
4990  */
4991 u_quad_t
4992 init_va_filerev(void)
4993 {
4994 	struct bintime bt;
4995 
4996 	getbinuptime(&bt);
4997 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
4998 }
4999 
5000 static int	filt_vfsread(struct knote *kn, long hint);
5001 static int	filt_vfswrite(struct knote *kn, long hint);
5002 static int	filt_vfsvnode(struct knote *kn, long hint);
5003 static void	filt_vfsdetach(struct knote *kn);
5004 static struct filterops vfsread_filtops = {
5005 	.f_isfd = 1,
5006 	.f_detach = filt_vfsdetach,
5007 	.f_event = filt_vfsread
5008 };
5009 static struct filterops vfswrite_filtops = {
5010 	.f_isfd = 1,
5011 	.f_detach = filt_vfsdetach,
5012 	.f_event = filt_vfswrite
5013 };
5014 static struct filterops vfsvnode_filtops = {
5015 	.f_isfd = 1,
5016 	.f_detach = filt_vfsdetach,
5017 	.f_event = filt_vfsvnode
5018 };
5019 
5020 static void
5021 vfs_knllock(void *arg)
5022 {
5023 	struct vnode *vp = arg;
5024 
5025 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5026 }
5027 
5028 static void
5029 vfs_knlunlock(void *arg)
5030 {
5031 	struct vnode *vp = arg;
5032 
5033 	VOP_UNLOCK(vp, 0);
5034 }
5035 
5036 static void
5037 vfs_knl_assert_locked(void *arg)
5038 {
5039 #ifdef DEBUG_VFS_LOCKS
5040 	struct vnode *vp = arg;
5041 
5042 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
5043 #endif
5044 }
5045 
5046 static void
5047 vfs_knl_assert_unlocked(void *arg)
5048 {
5049 #ifdef DEBUG_VFS_LOCKS
5050 	struct vnode *vp = arg;
5051 
5052 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
5053 #endif
5054 }
5055 
5056 int
5057 vfs_kqfilter(struct vop_kqfilter_args *ap)
5058 {
5059 	struct vnode *vp = ap->a_vp;
5060 	struct knote *kn = ap->a_kn;
5061 	struct knlist *knl;
5062 
5063 	switch (kn->kn_filter) {
5064 	case EVFILT_READ:
5065 		kn->kn_fop = &vfsread_filtops;
5066 		break;
5067 	case EVFILT_WRITE:
5068 		kn->kn_fop = &vfswrite_filtops;
5069 		break;
5070 	case EVFILT_VNODE:
5071 		kn->kn_fop = &vfsvnode_filtops;
5072 		break;
5073 	default:
5074 		return (EINVAL);
5075 	}
5076 
5077 	kn->kn_hook = (caddr_t)vp;
5078 
5079 	v_addpollinfo(vp);
5080 	if (vp->v_pollinfo == NULL)
5081 		return (ENOMEM);
5082 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
5083 	vhold(vp);
5084 	knlist_add(knl, kn, 0);
5085 
5086 	return (0);
5087 }
5088 
5089 /*
5090  * Detach knote from vnode
5091  */
5092 static void
5093 filt_vfsdetach(struct knote *kn)
5094 {
5095 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5096 
5097 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5098 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5099 	vdrop(vp);
5100 }
5101 
5102 /*ARGSUSED*/
5103 static int
5104 filt_vfsread(struct knote *kn, long hint)
5105 {
5106 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5107 	struct vattr va;
5108 	int res;
5109 
5110 	/*
5111 	 * filesystem is gone, so set the EOF flag and schedule
5112 	 * the knote for deletion.
5113 	 */
5114 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5115 		VI_LOCK(vp);
5116 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5117 		VI_UNLOCK(vp);
5118 		return (1);
5119 	}
5120 
5121 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5122 		return (0);
5123 
5124 	VI_LOCK(vp);
5125 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5126 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5127 	VI_UNLOCK(vp);
5128 	return (res);
5129 }
5130 
5131 /*ARGSUSED*/
5132 static int
5133 filt_vfswrite(struct knote *kn, long hint)
5134 {
5135 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5136 
5137 	VI_LOCK(vp);
5138 
5139 	/*
5140 	 * filesystem is gone, so set the EOF flag and schedule
5141 	 * the knote for deletion.
5142 	 */
5143 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5144 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5145 
5146 	kn->kn_data = 0;
5147 	VI_UNLOCK(vp);
5148 	return (1);
5149 }
5150 
5151 static int
5152 filt_vfsvnode(struct knote *kn, long hint)
5153 {
5154 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5155 	int res;
5156 
5157 	VI_LOCK(vp);
5158 	if (kn->kn_sfflags & hint)
5159 		kn->kn_fflags |= hint;
5160 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5161 		kn->kn_flags |= EV_EOF;
5162 		VI_UNLOCK(vp);
5163 		return (1);
5164 	}
5165 	res = (kn->kn_fflags != 0);
5166 	VI_UNLOCK(vp);
5167 	return (res);
5168 }
5169 
5170 int
5171 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5172 {
5173 	int error;
5174 
5175 	if (dp->d_reclen > ap->a_uio->uio_resid)
5176 		return (ENAMETOOLONG);
5177 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
5178 	if (error) {
5179 		if (ap->a_ncookies != NULL) {
5180 			if (ap->a_cookies != NULL)
5181 				free(ap->a_cookies, M_TEMP);
5182 			ap->a_cookies = NULL;
5183 			*ap->a_ncookies = 0;
5184 		}
5185 		return (error);
5186 	}
5187 	if (ap->a_ncookies == NULL)
5188 		return (0);
5189 
5190 	KASSERT(ap->a_cookies,
5191 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
5192 
5193 	*ap->a_cookies = realloc(*ap->a_cookies,
5194 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
5195 	(*ap->a_cookies)[*ap->a_ncookies] = off;
5196 	*ap->a_ncookies += 1;
5197 	return (0);
5198 }
5199 
5200 /*
5201  * Mark for update the access time of the file if the filesystem
5202  * supports VOP_MARKATIME.  This functionality is used by execve and
5203  * mmap, so we want to avoid the I/O implied by directly setting
5204  * va_atime for the sake of efficiency.
5205  */
5206 void
5207 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
5208 {
5209 	struct mount *mp;
5210 
5211 	mp = vp->v_mount;
5212 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
5213 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
5214 		(void)VOP_MARKATIME(vp);
5215 }
5216 
5217 /*
5218  * The purpose of this routine is to remove granularity from accmode_t,
5219  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
5220  * VADMIN and VAPPEND.
5221  *
5222  * If it returns 0, the caller is supposed to continue with the usual
5223  * access checks using 'accmode' as modified by this routine.  If it
5224  * returns nonzero value, the caller is supposed to return that value
5225  * as errno.
5226  *
5227  * Note that after this routine runs, accmode may be zero.
5228  */
5229 int
5230 vfs_unixify_accmode(accmode_t *accmode)
5231 {
5232 	/*
5233 	 * There is no way to specify explicit "deny" rule using
5234 	 * file mode or POSIX.1e ACLs.
5235 	 */
5236 	if (*accmode & VEXPLICIT_DENY) {
5237 		*accmode = 0;
5238 		return (0);
5239 	}
5240 
5241 	/*
5242 	 * None of these can be translated into usual access bits.
5243 	 * Also, the common case for NFSv4 ACLs is to not contain
5244 	 * either of these bits. Caller should check for VWRITE
5245 	 * on the containing directory instead.
5246 	 */
5247 	if (*accmode & (VDELETE_CHILD | VDELETE))
5248 		return (EPERM);
5249 
5250 	if (*accmode & VADMIN_PERMS) {
5251 		*accmode &= ~VADMIN_PERMS;
5252 		*accmode |= VADMIN;
5253 	}
5254 
5255 	/*
5256 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5257 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5258 	 */
5259 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5260 
5261 	return (0);
5262 }
5263 
5264 /*
5265  * These are helper functions for filesystems to traverse all
5266  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
5267  *
5268  * This interface replaces MNT_VNODE_FOREACH.
5269  */
5270 
5271 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
5272 
5273 struct vnode *
5274 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
5275 {
5276 	struct vnode *vp;
5277 
5278 	if (should_yield())
5279 		kern_yield(PRI_USER);
5280 	MNT_ILOCK(mp);
5281 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5282 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
5283 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
5284 		/* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
5285 		if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
5286 			continue;
5287 		VI_LOCK(vp);
5288 		if ((vp->v_iflag & VI_DOOMED) != 0) {
5289 			VI_UNLOCK(vp);
5290 			continue;
5291 		}
5292 		break;
5293 	}
5294 	if (vp == NULL) {
5295 		__mnt_vnode_markerfree_all(mvp, mp);
5296 		/* MNT_IUNLOCK(mp); -- done in above function */
5297 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
5298 		return (NULL);
5299 	}
5300 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5301 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5302 	MNT_IUNLOCK(mp);
5303 	return (vp);
5304 }
5305 
5306 struct vnode *
5307 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
5308 {
5309 	struct vnode *vp;
5310 
5311 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5312 	MNT_ILOCK(mp);
5313 	MNT_REF(mp);
5314 	(*mvp)->v_mount = mp;
5315 	(*mvp)->v_type = VMARKER;
5316 
5317 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
5318 		/* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
5319 		if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
5320 			continue;
5321 		VI_LOCK(vp);
5322 		if ((vp->v_iflag & VI_DOOMED) != 0) {
5323 			VI_UNLOCK(vp);
5324 			continue;
5325 		}
5326 		break;
5327 	}
5328 	if (vp == NULL) {
5329 		MNT_REL(mp);
5330 		MNT_IUNLOCK(mp);
5331 		free(*mvp, M_VNODE_MARKER);
5332 		*mvp = NULL;
5333 		return (NULL);
5334 	}
5335 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5336 	MNT_IUNLOCK(mp);
5337 	return (vp);
5338 }
5339 
5340 void
5341 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
5342 {
5343 
5344 	if (*mvp == NULL) {
5345 		MNT_IUNLOCK(mp);
5346 		return;
5347 	}
5348 
5349 	mtx_assert(MNT_MTX(mp), MA_OWNED);
5350 
5351 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5352 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5353 	MNT_REL(mp);
5354 	MNT_IUNLOCK(mp);
5355 	free(*mvp, M_VNODE_MARKER);
5356 	*mvp = NULL;
5357 }
5358 
5359 /*
5360  * These are helper functions for filesystems to traverse their
5361  * active vnodes.  See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h
5362  */
5363 static void
5364 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5365 {
5366 
5367 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5368 
5369 	MNT_ILOCK(mp);
5370 	MNT_REL(mp);
5371 	MNT_IUNLOCK(mp);
5372 	free(*mvp, M_VNODE_MARKER);
5373 	*mvp = NULL;
5374 }
5375 
5376 /*
5377  * Relock the mp mount vnode list lock with the vp vnode interlock in the
5378  * conventional lock order during mnt_vnode_next_active iteration.
5379  *
5380  * On entry, the mount vnode list lock is held and the vnode interlock is not.
5381  * The list lock is dropped and reacquired.  On success, both locks are held.
5382  * On failure, the mount vnode list lock is held but the vnode interlock is
5383  * not, and the procedure may have yielded.
5384  */
5385 static bool
5386 mnt_vnode_next_active_relock(struct vnode *mvp, struct mount *mp,
5387     struct vnode *vp)
5388 {
5389 	const struct vnode *tmp;
5390 	bool held, ret;
5391 
5392 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
5393 	    TAILQ_NEXT(mvp, v_actfreelist) != NULL, mvp,
5394 	    ("%s: bad marker", __func__));
5395 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
5396 	    ("%s: inappropriate vnode", __func__));
5397 	ASSERT_VI_UNLOCKED(vp, __func__);
5398 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5399 
5400 	ret = false;
5401 
5402 	TAILQ_REMOVE(&mp->mnt_activevnodelist, mvp, v_actfreelist);
5403 	TAILQ_INSERT_BEFORE(vp, mvp, v_actfreelist);
5404 
5405 	/*
5406 	 * Use a hold to prevent vp from disappearing while the mount vnode
5407 	 * list lock is dropped and reacquired.  Normally a hold would be
5408 	 * acquired with vhold(), but that might try to acquire the vnode
5409 	 * interlock, which would be a LOR with the mount vnode list lock.
5410 	 */
5411 	held = refcount_acquire_if_not_zero(&vp->v_holdcnt);
5412 	mtx_unlock(&mp->mnt_listmtx);
5413 	if (!held)
5414 		goto abort;
5415 	VI_LOCK(vp);
5416 	if (!refcount_release_if_not_last(&vp->v_holdcnt)) {
5417 		vdropl(vp);
5418 		goto abort;
5419 	}
5420 	mtx_lock(&mp->mnt_listmtx);
5421 
5422 	/*
5423 	 * Determine whether the vnode is still the next one after the marker,
5424 	 * excepting any other markers.  If the vnode has not been doomed by
5425 	 * vgone() then the hold should have ensured that it remained on the
5426 	 * active list.  If it has been doomed but is still on the active list,
5427 	 * don't abort, but rather skip over it (avoid spinning on doomed
5428 	 * vnodes).
5429 	 */
5430 	tmp = mvp;
5431 	do {
5432 		tmp = TAILQ_NEXT(tmp, v_actfreelist);
5433 	} while (tmp != NULL && tmp->v_type == VMARKER);
5434 	if (tmp != vp) {
5435 		mtx_unlock(&mp->mnt_listmtx);
5436 		VI_UNLOCK(vp);
5437 		goto abort;
5438 	}
5439 
5440 	ret = true;
5441 	goto out;
5442 abort:
5443 	maybe_yield();
5444 	mtx_lock(&mp->mnt_listmtx);
5445 out:
5446 	if (ret)
5447 		ASSERT_VI_LOCKED(vp, __func__);
5448 	else
5449 		ASSERT_VI_UNLOCKED(vp, __func__);
5450 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5451 	return (ret);
5452 }
5453 
5454 static struct vnode *
5455 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5456 {
5457 	struct vnode *vp, *nvp;
5458 
5459 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5460 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5461 restart:
5462 	vp = TAILQ_NEXT(*mvp, v_actfreelist);
5463 	while (vp != NULL) {
5464 		if (vp->v_type == VMARKER) {
5465 			vp = TAILQ_NEXT(vp, v_actfreelist);
5466 			continue;
5467 		}
5468 		/*
5469 		 * Try-lock because this is the wrong lock order.  If that does
5470 		 * not succeed, drop the mount vnode list lock and try to
5471 		 * reacquire it and the vnode interlock in the right order.
5472 		 */
5473 		if (!VI_TRYLOCK(vp) &&
5474 		    !mnt_vnode_next_active_relock(*mvp, mp, vp))
5475 			goto restart;
5476 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
5477 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
5478 		    ("alien vnode on the active list %p %p", vp, mp));
5479 		if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0)
5480 			break;
5481 		nvp = TAILQ_NEXT(vp, v_actfreelist);
5482 		VI_UNLOCK(vp);
5483 		vp = nvp;
5484 	}
5485 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5486 
5487 	/* Check if we are done */
5488 	if (vp == NULL) {
5489 		mtx_unlock(&mp->mnt_listmtx);
5490 		mnt_vnode_markerfree_active(mvp, mp);
5491 		return (NULL);
5492 	}
5493 	TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist);
5494 	mtx_unlock(&mp->mnt_listmtx);
5495 	ASSERT_VI_LOCKED(vp, "active iter");
5496 	KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp));
5497 	return (vp);
5498 }
5499 
5500 struct vnode *
5501 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5502 {
5503 
5504 	if (should_yield())
5505 		kern_yield(PRI_USER);
5506 	mtx_lock(&mp->mnt_listmtx);
5507 	return (mnt_vnode_next_active(mvp, mp));
5508 }
5509 
5510 struct vnode *
5511 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp)
5512 {
5513 	struct vnode *vp;
5514 
5515 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5516 	MNT_ILOCK(mp);
5517 	MNT_REF(mp);
5518 	MNT_IUNLOCK(mp);
5519 	(*mvp)->v_type = VMARKER;
5520 	(*mvp)->v_mount = mp;
5521 
5522 	mtx_lock(&mp->mnt_listmtx);
5523 	vp = TAILQ_FIRST(&mp->mnt_activevnodelist);
5524 	if (vp == NULL) {
5525 		mtx_unlock(&mp->mnt_listmtx);
5526 		mnt_vnode_markerfree_active(mvp, mp);
5527 		return (NULL);
5528 	}
5529 	TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5530 	return (mnt_vnode_next_active(mvp, mp));
5531 }
5532 
5533 void
5534 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5535 {
5536 
5537 	if (*mvp == NULL)
5538 		return;
5539 
5540 	mtx_lock(&mp->mnt_listmtx);
5541 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5542 	mtx_unlock(&mp->mnt_listmtx);
5543 	mnt_vnode_markerfree_active(mvp, mp);
5544 }
5545