xref: /freebsd/sys/kern/vfs_subr.c (revision 8e3e3a7ae841ccf6f6ac30a2eeab85df5d7f04bc)
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_compat.h"
47 #include "opt_ddb.h"
48 #include "opt_watchdog.h"
49 
50 #include <sys/param.h>
51 #include <sys/systm.h>
52 #include <sys/bio.h>
53 #include <sys/buf.h>
54 #include <sys/condvar.h>
55 #include <sys/conf.h>
56 #include <sys/counter.h>
57 #include <sys/dirent.h>
58 #include <sys/event.h>
59 #include <sys/eventhandler.h>
60 #include <sys/extattr.h>
61 #include <sys/file.h>
62 #include <sys/fcntl.h>
63 #include <sys/jail.h>
64 #include <sys/kdb.h>
65 #include <sys/kernel.h>
66 #include <sys/kthread.h>
67 #include <sys/lockf.h>
68 #include <sys/malloc.h>
69 #include <sys/mount.h>
70 #include <sys/namei.h>
71 #include <sys/pctrie.h>
72 #include <sys/priv.h>
73 #include <sys/reboot.h>
74 #include <sys/refcount.h>
75 #include <sys/rwlock.h>
76 #include <sys/sched.h>
77 #include <sys/sleepqueue.h>
78 #include <sys/smp.h>
79 #include <sys/stat.h>
80 #include <sys/sysctl.h>
81 #include <sys/syslog.h>
82 #include <sys/vmmeter.h>
83 #include <sys/vnode.h>
84 #include <sys/watchdog.h>
85 
86 #include <machine/stdarg.h>
87 
88 #include <security/mac/mac_framework.h>
89 
90 #include <vm/vm.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_extern.h>
93 #include <vm/pmap.h>
94 #include <vm/vm_map.h>
95 #include <vm/vm_page.h>
96 #include <vm/vm_kern.h>
97 #include <vm/uma.h>
98 
99 #ifdef DDB
100 #include <ddb/ddb.h>
101 #endif
102 
103 static void	delmntque(struct vnode *vp);
104 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
105 		    int slpflag, int slptimeo);
106 static void	syncer_shutdown(void *arg, int howto);
107 static int	vtryrecycle(struct vnode *vp);
108 static void	v_init_counters(struct vnode *);
109 static void	v_incr_usecount(struct vnode *);
110 static void	v_incr_usecount_locked(struct vnode *);
111 static void	v_incr_devcount(struct vnode *);
112 static void	v_decr_devcount(struct vnode *);
113 static void	vgonel(struct vnode *);
114 static void	vfs_knllock(void *arg);
115 static void	vfs_knlunlock(void *arg);
116 static void	vfs_knl_assert_locked(void *arg);
117 static void	vfs_knl_assert_unlocked(void *arg);
118 static void	vnlru_return_batches(struct vfsops *mnt_op);
119 static void	destroy_vpollinfo(struct vpollinfo *vi);
120 
121 /*
122  * Number of vnodes in existence.  Increased whenever getnewvnode()
123  * allocates a new vnode, decreased in vdropl() for VI_DOOMED vnode.
124  */
125 static unsigned long	numvnodes;
126 
127 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
128     "Number of vnodes in existence");
129 
130 static counter_u64_t vnodes_created;
131 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
132     "Number of vnodes created by getnewvnode");
133 
134 static u_long mnt_free_list_batch = 128;
135 SYSCTL_ULONG(_vfs, OID_AUTO, mnt_free_list_batch, CTLFLAG_RW,
136     &mnt_free_list_batch, 0, "Limit of vnodes held on mnt's free list");
137 
138 /*
139  * Conversion tables for conversion from vnode types to inode formats
140  * and back.
141  */
142 enum vtype iftovt_tab[16] = {
143 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
144 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
145 };
146 int vttoif_tab[10] = {
147 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
148 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
149 };
150 
151 /*
152  * List of vnodes that are ready for recycling.
153  */
154 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
155 
156 /*
157  * "Free" vnode target.  Free vnodes are rarely completely free, but are
158  * just ones that are cheap to recycle.  Usually they are for files which
159  * have been stat'd but not read; these usually have inode and namecache
160  * data attached to them.  This target is the preferred minimum size of a
161  * sub-cache consisting mostly of such files. The system balances the size
162  * of this sub-cache with its complement to try to prevent either from
163  * thrashing while the other is relatively inactive.  The targets express
164  * a preference for the best balance.
165  *
166  * "Above" this target there are 2 further targets (watermarks) related
167  * to recyling of free vnodes.  In the best-operating case, the cache is
168  * exactly full, the free list has size between vlowat and vhiwat above the
169  * free target, and recycling from it and normal use maintains this state.
170  * Sometimes the free list is below vlowat or even empty, but this state
171  * is even better for immediate use provided the cache is not full.
172  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
173  * ones) to reach one of these states.  The watermarks are currently hard-
174  * coded as 4% and 9% of the available space higher.  These and the default
175  * of 25% for wantfreevnodes are too large if the memory size is large.
176  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
177  * whenever vnlru_proc() becomes active.
178  */
179 static u_long wantfreevnodes;
180 SYSCTL_ULONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW,
181     &wantfreevnodes, 0, "Target for minimum number of \"free\" vnodes");
182 static u_long freevnodes;
183 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
184     &freevnodes, 0, "Number of \"free\" vnodes");
185 
186 static counter_u64_t recycles_count;
187 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
188     "Number of vnodes recycled to meet vnode cache targets");
189 
190 /*
191  * Various variables used for debugging the new implementation of
192  * reassignbuf().
193  * XXX these are probably of (very) limited utility now.
194  */
195 static int reassignbufcalls;
196 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0,
197     "Number of calls to reassignbuf");
198 
199 static counter_u64_t free_owe_inact;
200 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, free_owe_inact, CTLFLAG_RD, &free_owe_inact,
201     "Number of times free vnodes kept on active list due to VFS "
202     "owing inactivation");
203 
204 /* To keep more than one thread at a time from running vfs_getnewfsid */
205 static struct mtx mntid_mtx;
206 
207 /*
208  * Lock for any access to the following:
209  *	vnode_free_list
210  *	numvnodes
211  *	freevnodes
212  */
213 static struct mtx vnode_free_list_mtx;
214 
215 /* Publicly exported FS */
216 struct nfs_public nfs_pub;
217 
218 static uma_zone_t buf_trie_zone;
219 
220 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
221 static uma_zone_t vnode_zone;
222 static uma_zone_t vnodepoll_zone;
223 
224 /*
225  * The workitem queue.
226  *
227  * It is useful to delay writes of file data and filesystem metadata
228  * for tens of seconds so that quickly created and deleted files need
229  * not waste disk bandwidth being created and removed. To realize this,
230  * we append vnodes to a "workitem" queue. When running with a soft
231  * updates implementation, most pending metadata dependencies should
232  * not wait for more than a few seconds. Thus, mounted on block devices
233  * are delayed only about a half the time that file data is delayed.
234  * Similarly, directory updates are more critical, so are only delayed
235  * about a third the time that file data is delayed. Thus, there are
236  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
237  * one each second (driven off the filesystem syncer process). The
238  * syncer_delayno variable indicates the next queue that is to be processed.
239  * Items that need to be processed soon are placed in this queue:
240  *
241  *	syncer_workitem_pending[syncer_delayno]
242  *
243  * A delay of fifteen seconds is done by placing the request fifteen
244  * entries later in the queue:
245  *
246  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
247  *
248  */
249 static int syncer_delayno;
250 static long syncer_mask;
251 LIST_HEAD(synclist, bufobj);
252 static struct synclist *syncer_workitem_pending;
253 /*
254  * The sync_mtx protects:
255  *	bo->bo_synclist
256  *	sync_vnode_count
257  *	syncer_delayno
258  *	syncer_state
259  *	syncer_workitem_pending
260  *	syncer_worklist_len
261  *	rushjob
262  */
263 static struct mtx sync_mtx;
264 static struct cv sync_wakeup;
265 
266 #define SYNCER_MAXDELAY		32
267 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
268 static int syncdelay = 30;		/* max time to delay syncing data */
269 static int filedelay = 30;		/* time to delay syncing files */
270 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
271     "Time to delay syncing files (in seconds)");
272 static int dirdelay = 29;		/* time to delay syncing directories */
273 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
274     "Time to delay syncing directories (in seconds)");
275 static int metadelay = 28;		/* time to delay syncing metadata */
276 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
277     "Time to delay syncing metadata (in seconds)");
278 static int rushjob;		/* number of slots to run ASAP */
279 static int stat_rush_requests;	/* number of times I/O speeded up */
280 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
281     "Number of times I/O speeded up (rush requests)");
282 
283 /*
284  * When shutting down the syncer, run it at four times normal speed.
285  */
286 #define SYNCER_SHUTDOWN_SPEEDUP		4
287 static int sync_vnode_count;
288 static int syncer_worklist_len;
289 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
290     syncer_state;
291 
292 /* Target for maximum number of vnodes. */
293 int desiredvnodes;
294 static int gapvnodes;		/* gap between wanted and desired */
295 static int vhiwat;		/* enough extras after expansion */
296 static int vlowat;		/* minimal extras before expansion */
297 static int vstir;		/* nonzero to stir non-free vnodes */
298 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
299 
300 static int
301 sysctl_update_desiredvnodes(SYSCTL_HANDLER_ARGS)
302 {
303 	int error, old_desiredvnodes;
304 
305 	old_desiredvnodes = desiredvnodes;
306 	if ((error = sysctl_handle_int(oidp, arg1, arg2, req)) != 0)
307 		return (error);
308 	if (old_desiredvnodes != desiredvnodes) {
309 		wantfreevnodes = desiredvnodes / 4;
310 		/* XXX locking seems to be incomplete. */
311 		vfs_hash_changesize(desiredvnodes);
312 		cache_changesize(desiredvnodes);
313 	}
314 	return (0);
315 }
316 
317 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
318     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, &desiredvnodes, 0,
319     sysctl_update_desiredvnodes, "I", "Target for maximum number of vnodes");
320 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
321     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
322 static int vnlru_nowhere;
323 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
324     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
325 
326 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
327 static int vnsz2log;
328 
329 /*
330  * Support for the bufobj clean & dirty pctrie.
331  */
332 static void *
333 buf_trie_alloc(struct pctrie *ptree)
334 {
335 
336 	return uma_zalloc(buf_trie_zone, M_NOWAIT);
337 }
338 
339 static void
340 buf_trie_free(struct pctrie *ptree, void *node)
341 {
342 
343 	uma_zfree(buf_trie_zone, node);
344 }
345 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
346 
347 /*
348  * Initialize the vnode management data structures.
349  *
350  * Reevaluate the following cap on the number of vnodes after the physical
351  * memory size exceeds 512GB.  In the limit, as the physical memory size
352  * grows, the ratio of the memory size in KB to to vnodes approaches 64:1.
353  */
354 #ifndef	MAXVNODES_MAX
355 #define	MAXVNODES_MAX	(512 * 1024 * 1024 / 64)	/* 8M */
356 #endif
357 
358 /*
359  * Initialize a vnode as it first enters the zone.
360  */
361 static int
362 vnode_init(void *mem, int size, int flags)
363 {
364 	struct vnode *vp;
365 	struct bufobj *bo;
366 
367 	vp = mem;
368 	bzero(vp, size);
369 	/*
370 	 * Setup locks.
371 	 */
372 	vp->v_vnlock = &vp->v_lock;
373 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
374 	/*
375 	 * By default, don't allow shared locks unless filesystems opt-in.
376 	 */
377 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
378 	    LK_NOSHARE | LK_IS_VNODE);
379 	/*
380 	 * Initialize bufobj.
381 	 */
382 	bo = &vp->v_bufobj;
383 	rw_init(BO_LOCKPTR(bo), "bufobj interlock");
384 	bo->bo_private = vp;
385 	TAILQ_INIT(&bo->bo_clean.bv_hd);
386 	TAILQ_INIT(&bo->bo_dirty.bv_hd);
387 	/*
388 	 * Initialize namecache.
389 	 */
390 	LIST_INIT(&vp->v_cache_src);
391 	TAILQ_INIT(&vp->v_cache_dst);
392 	/*
393 	 * Initialize rangelocks.
394 	 */
395 	rangelock_init(&vp->v_rl);
396 	return (0);
397 }
398 
399 /*
400  * Free a vnode when it is cleared from the zone.
401  */
402 static void
403 vnode_fini(void *mem, int size)
404 {
405 	struct vnode *vp;
406 	struct bufobj *bo;
407 
408 	vp = mem;
409 	rangelock_destroy(&vp->v_rl);
410 	lockdestroy(vp->v_vnlock);
411 	mtx_destroy(&vp->v_interlock);
412 	bo = &vp->v_bufobj;
413 	rw_destroy(BO_LOCKPTR(bo));
414 }
415 
416 /*
417  * Provide the size of NFS nclnode and NFS fh for calculation of the
418  * vnode memory consumption.  The size is specified directly to
419  * eliminate dependency on NFS-private header.
420  *
421  * Other filesystems may use bigger or smaller (like UFS and ZFS)
422  * private inode data, but the NFS-based estimation is ample enough.
423  * Still, we care about differences in the size between 64- and 32-bit
424  * platforms.
425  *
426  * Namecache structure size is heuristically
427  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
428  */
429 #ifdef _LP64
430 #define	NFS_NCLNODE_SZ	(528 + 64)
431 #define	NC_SZ		148
432 #else
433 #define	NFS_NCLNODE_SZ	(360 + 32)
434 #define	NC_SZ		92
435 #endif
436 
437 static void
438 vntblinit(void *dummy __unused)
439 {
440 	u_int i;
441 	int physvnodes, virtvnodes;
442 
443 	/*
444 	 * Desiredvnodes is a function of the physical memory size and the
445 	 * kernel's heap size.  Generally speaking, it scales with the
446 	 * physical memory size.  The ratio of desiredvnodes to the physical
447 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
448 	 * Thereafter, the
449 	 * marginal ratio of desiredvnodes to the physical memory size is
450 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
451 	 * size.  The memory required by desiredvnodes vnodes and vm objects
452 	 * must not exceed 1/10th of the kernel's heap size.
453 	 */
454 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
455 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
456 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
457 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
458 	desiredvnodes = min(physvnodes, virtvnodes);
459 	if (desiredvnodes > MAXVNODES_MAX) {
460 		if (bootverbose)
461 			printf("Reducing kern.maxvnodes %d -> %d\n",
462 			    desiredvnodes, MAXVNODES_MAX);
463 		desiredvnodes = MAXVNODES_MAX;
464 	}
465 	wantfreevnodes = desiredvnodes / 4;
466 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
467 	TAILQ_INIT(&vnode_free_list);
468 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
469 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
470 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
471 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
472 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
473 	/*
474 	 * Preallocate enough nodes to support one-per buf so that
475 	 * we can not fail an insert.  reassignbuf() callers can not
476 	 * tolerate the insertion failure.
477 	 */
478 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
479 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
480 	    UMA_ZONE_NOFREE | UMA_ZONE_VM);
481 	uma_prealloc(buf_trie_zone, nbuf);
482 
483 	vnodes_created = counter_u64_alloc(M_WAITOK);
484 	recycles_count = counter_u64_alloc(M_WAITOK);
485 	free_owe_inact = counter_u64_alloc(M_WAITOK);
486 
487 	/*
488 	 * Initialize the filesystem syncer.
489 	 */
490 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
491 	    &syncer_mask);
492 	syncer_maxdelay = syncer_mask + 1;
493 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
494 	cv_init(&sync_wakeup, "syncer");
495 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
496 		vnsz2log++;
497 	vnsz2log--;
498 }
499 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
500 
501 
502 /*
503  * Mark a mount point as busy. Used to synchronize access and to delay
504  * unmounting. Eventually, mountlist_mtx is not released on failure.
505  *
506  * vfs_busy() is a custom lock, it can block the caller.
507  * vfs_busy() only sleeps if the unmount is active on the mount point.
508  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
509  * vnode belonging to mp.
510  *
511  * Lookup uses vfs_busy() to traverse mount points.
512  * root fs			var fs
513  * / vnode lock		A	/ vnode lock (/var)		D
514  * /var vnode lock	B	/log vnode lock(/var/log)	E
515  * vfs_busy lock	C	vfs_busy lock			F
516  *
517  * Within each file system, the lock order is C->A->B and F->D->E.
518  *
519  * When traversing across mounts, the system follows that lock order:
520  *
521  *        C->A->B
522  *              |
523  *              +->F->D->E
524  *
525  * The lookup() process for namei("/var") illustrates the process:
526  *  VOP_LOOKUP() obtains B while A is held
527  *  vfs_busy() obtains a shared lock on F while A and B are held
528  *  vput() releases lock on B
529  *  vput() releases lock on A
530  *  VFS_ROOT() obtains lock on D while shared lock on F is held
531  *  vfs_unbusy() releases shared lock on F
532  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
533  *    Attempt to lock A (instead of vp_crossmp) while D is held would
534  *    violate the global order, causing deadlocks.
535  *
536  * dounmount() locks B while F is drained.
537  */
538 int
539 vfs_busy(struct mount *mp, int flags)
540 {
541 
542 	MPASS((flags & ~MBF_MASK) == 0);
543 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
544 
545 	MNT_ILOCK(mp);
546 	MNT_REF(mp);
547 	/*
548 	 * If mount point is currently being unmounted, sleep until the
549 	 * mount point fate is decided.  If thread doing the unmounting fails,
550 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
551 	 * that this mount point has survived the unmount attempt and vfs_busy
552 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
553 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
554 	 * about to be really destroyed.  vfs_busy needs to release its
555 	 * reference on the mount point in this case and return with ENOENT,
556 	 * telling the caller that mount mount it tried to busy is no longer
557 	 * valid.
558 	 */
559 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
560 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
561 			MNT_REL(mp);
562 			MNT_IUNLOCK(mp);
563 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
564 			    __func__);
565 			return (ENOENT);
566 		}
567 		if (flags & MBF_MNTLSTLOCK)
568 			mtx_unlock(&mountlist_mtx);
569 		mp->mnt_kern_flag |= MNTK_MWAIT;
570 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
571 		if (flags & MBF_MNTLSTLOCK)
572 			mtx_lock(&mountlist_mtx);
573 		MNT_ILOCK(mp);
574 	}
575 	if (flags & MBF_MNTLSTLOCK)
576 		mtx_unlock(&mountlist_mtx);
577 	mp->mnt_lockref++;
578 	MNT_IUNLOCK(mp);
579 	return (0);
580 }
581 
582 /*
583  * Free a busy filesystem.
584  */
585 void
586 vfs_unbusy(struct mount *mp)
587 {
588 
589 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
590 	MNT_ILOCK(mp);
591 	MNT_REL(mp);
592 	KASSERT(mp->mnt_lockref > 0, ("negative mnt_lockref"));
593 	mp->mnt_lockref--;
594 	if (mp->mnt_lockref == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
595 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
596 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
597 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
598 		wakeup(&mp->mnt_lockref);
599 	}
600 	MNT_IUNLOCK(mp);
601 }
602 
603 /*
604  * Lookup a mount point by filesystem identifier.
605  */
606 struct mount *
607 vfs_getvfs(fsid_t *fsid)
608 {
609 	struct mount *mp;
610 
611 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
612 	mtx_lock(&mountlist_mtx);
613 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
614 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
615 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
616 			vfs_ref(mp);
617 			mtx_unlock(&mountlist_mtx);
618 			return (mp);
619 		}
620 	}
621 	mtx_unlock(&mountlist_mtx);
622 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
623 	return ((struct mount *) 0);
624 }
625 
626 /*
627  * Lookup a mount point by filesystem identifier, busying it before
628  * returning.
629  *
630  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
631  * cache for popular filesystem identifiers.  The cache is lockess, using
632  * the fact that struct mount's are never freed.  In worst case we may
633  * get pointer to unmounted or even different filesystem, so we have to
634  * check what we got, and go slow way if so.
635  */
636 struct mount *
637 vfs_busyfs(fsid_t *fsid)
638 {
639 #define	FSID_CACHE_SIZE	256
640 	typedef struct mount * volatile vmp_t;
641 	static vmp_t cache[FSID_CACHE_SIZE];
642 	struct mount *mp;
643 	int error;
644 	uint32_t hash;
645 
646 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
647 	hash = fsid->val[0] ^ fsid->val[1];
648 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
649 	mp = cache[hash];
650 	if (mp == NULL ||
651 	    mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
652 	    mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
653 		goto slow;
654 	if (vfs_busy(mp, 0) != 0) {
655 		cache[hash] = NULL;
656 		goto slow;
657 	}
658 	if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
659 	    mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
660 		return (mp);
661 	else
662 	    vfs_unbusy(mp);
663 
664 slow:
665 	mtx_lock(&mountlist_mtx);
666 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
667 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
668 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
669 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
670 			if (error) {
671 				cache[hash] = NULL;
672 				mtx_unlock(&mountlist_mtx);
673 				return (NULL);
674 			}
675 			cache[hash] = mp;
676 			return (mp);
677 		}
678 	}
679 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
680 	mtx_unlock(&mountlist_mtx);
681 	return ((struct mount *) 0);
682 }
683 
684 /*
685  * Check if a user can access privileged mount options.
686  */
687 int
688 vfs_suser(struct mount *mp, struct thread *td)
689 {
690 	int error;
691 
692 	/*
693 	 * If the thread is jailed, but this is not a jail-friendly file
694 	 * system, deny immediately.
695 	 */
696 	if (!(mp->mnt_vfc->vfc_flags & VFCF_JAIL) && jailed(td->td_ucred))
697 		return (EPERM);
698 
699 	/*
700 	 * If the file system was mounted outside the jail of the calling
701 	 * thread, deny immediately.
702 	 */
703 	if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
704 		return (EPERM);
705 
706 	/*
707 	 * If file system supports delegated administration, we don't check
708 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
709 	 * by the file system itself.
710 	 * If this is not the user that did original mount, we check for
711 	 * the PRIV_VFS_MOUNT_OWNER privilege.
712 	 */
713 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
714 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
715 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
716 			return (error);
717 	}
718 	return (0);
719 }
720 
721 /*
722  * Get a new unique fsid.  Try to make its val[0] unique, since this value
723  * will be used to create fake device numbers for stat().  Also try (but
724  * not so hard) make its val[0] unique mod 2^16, since some emulators only
725  * support 16-bit device numbers.  We end up with unique val[0]'s for the
726  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
727  *
728  * Keep in mind that several mounts may be running in parallel.  Starting
729  * the search one past where the previous search terminated is both a
730  * micro-optimization and a defense against returning the same fsid to
731  * different mounts.
732  */
733 void
734 vfs_getnewfsid(struct mount *mp)
735 {
736 	static uint16_t mntid_base;
737 	struct mount *nmp;
738 	fsid_t tfsid;
739 	int mtype;
740 
741 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
742 	mtx_lock(&mntid_mtx);
743 	mtype = mp->mnt_vfc->vfc_typenum;
744 	tfsid.val[1] = mtype;
745 	mtype = (mtype & 0xFF) << 24;
746 	for (;;) {
747 		tfsid.val[0] = makedev(255,
748 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
749 		mntid_base++;
750 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
751 			break;
752 		vfs_rel(nmp);
753 	}
754 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
755 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
756 	mtx_unlock(&mntid_mtx);
757 }
758 
759 /*
760  * Knob to control the precision of file timestamps:
761  *
762  *   0 = seconds only; nanoseconds zeroed.
763  *   1 = seconds and nanoseconds, accurate within 1/HZ.
764  *   2 = seconds and nanoseconds, truncated to microseconds.
765  * >=3 = seconds and nanoseconds, maximum precision.
766  */
767 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
768 
769 static int timestamp_precision = TSP_USEC;
770 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
771     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
772     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
773     "3+: sec + ns (max. precision))");
774 
775 /*
776  * Get a current timestamp.
777  */
778 void
779 vfs_timestamp(struct timespec *tsp)
780 {
781 	struct timeval tv;
782 
783 	switch (timestamp_precision) {
784 	case TSP_SEC:
785 		tsp->tv_sec = time_second;
786 		tsp->tv_nsec = 0;
787 		break;
788 	case TSP_HZ:
789 		getnanotime(tsp);
790 		break;
791 	case TSP_USEC:
792 		microtime(&tv);
793 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
794 		break;
795 	case TSP_NSEC:
796 	default:
797 		nanotime(tsp);
798 		break;
799 	}
800 }
801 
802 /*
803  * Set vnode attributes to VNOVAL
804  */
805 void
806 vattr_null(struct vattr *vap)
807 {
808 
809 	vap->va_type = VNON;
810 	vap->va_size = VNOVAL;
811 	vap->va_bytes = VNOVAL;
812 	vap->va_mode = VNOVAL;
813 	vap->va_nlink = VNOVAL;
814 	vap->va_uid = VNOVAL;
815 	vap->va_gid = VNOVAL;
816 	vap->va_fsid = VNOVAL;
817 	vap->va_fileid = VNOVAL;
818 	vap->va_blocksize = VNOVAL;
819 	vap->va_rdev = VNOVAL;
820 	vap->va_atime.tv_sec = VNOVAL;
821 	vap->va_atime.tv_nsec = VNOVAL;
822 	vap->va_mtime.tv_sec = VNOVAL;
823 	vap->va_mtime.tv_nsec = VNOVAL;
824 	vap->va_ctime.tv_sec = VNOVAL;
825 	vap->va_ctime.tv_nsec = VNOVAL;
826 	vap->va_birthtime.tv_sec = VNOVAL;
827 	vap->va_birthtime.tv_nsec = VNOVAL;
828 	vap->va_flags = VNOVAL;
829 	vap->va_gen = VNOVAL;
830 	vap->va_vaflags = 0;
831 }
832 
833 /*
834  * This routine is called when we have too many vnodes.  It attempts
835  * to free <count> vnodes and will potentially free vnodes that still
836  * have VM backing store (VM backing store is typically the cause
837  * of a vnode blowout so we want to do this).  Therefore, this operation
838  * is not considered cheap.
839  *
840  * A number of conditions may prevent a vnode from being reclaimed.
841  * the buffer cache may have references on the vnode, a directory
842  * vnode may still have references due to the namei cache representing
843  * underlying files, or the vnode may be in active use.   It is not
844  * desirable to reuse such vnodes.  These conditions may cause the
845  * number of vnodes to reach some minimum value regardless of what
846  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
847  */
848 static int
849 vlrureclaim(struct mount *mp, int reclaim_nc_src, int trigger)
850 {
851 	struct vnode *vp;
852 	int count, done, target;
853 
854 	done = 0;
855 	vn_start_write(NULL, &mp, V_WAIT);
856 	MNT_ILOCK(mp);
857 	count = mp->mnt_nvnodelistsize;
858 	target = count * (int64_t)gapvnodes / imax(desiredvnodes, 1);
859 	target = target / 10 + 1;
860 	while (count != 0 && done < target) {
861 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
862 		while (vp != NULL && vp->v_type == VMARKER)
863 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
864 		if (vp == NULL)
865 			break;
866 		/*
867 		 * XXX LRU is completely broken for non-free vnodes.  First
868 		 * by calling here in mountpoint order, then by moving
869 		 * unselected vnodes to the end here, and most grossly by
870 		 * removing the vlruvp() function that was supposed to
871 		 * maintain the order.  (This function was born broken
872 		 * since syncer problems prevented it doing anything.)  The
873 		 * order is closer to LRC (C = Created).
874 		 *
875 		 * LRU reclaiming of vnodes seems to have last worked in
876 		 * FreeBSD-3 where LRU wasn't mentioned under any spelling.
877 		 * Then there was no hold count, and inactive vnodes were
878 		 * simply put on the free list in LRU order.  The separate
879 		 * lists also break LRU.  We prefer to reclaim from the
880 		 * free list for technical reasons.  This tends to thrash
881 		 * the free list to keep very unrecently used held vnodes.
882 		 * The problem is mitigated by keeping the free list large.
883 		 */
884 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
885 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
886 		--count;
887 		if (!VI_TRYLOCK(vp))
888 			goto next_iter;
889 		/*
890 		 * If it's been deconstructed already, it's still
891 		 * referenced, or it exceeds the trigger, skip it.
892 		 * Also skip free vnodes.  We are trying to make space
893 		 * to expand the free list, not reduce it.
894 		 */
895 		if (vp->v_usecount ||
896 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
897 		    ((vp->v_iflag & VI_FREE) != 0) ||
898 		    (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
899 		    vp->v_object->resident_page_count > trigger)) {
900 			VI_UNLOCK(vp);
901 			goto next_iter;
902 		}
903 		MNT_IUNLOCK(mp);
904 		vholdl(vp);
905 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) {
906 			vdrop(vp);
907 			goto next_iter_mntunlocked;
908 		}
909 		VI_LOCK(vp);
910 		/*
911 		 * v_usecount may have been bumped after VOP_LOCK() dropped
912 		 * the vnode interlock and before it was locked again.
913 		 *
914 		 * It is not necessary to recheck VI_DOOMED because it can
915 		 * only be set by another thread that holds both the vnode
916 		 * lock and vnode interlock.  If another thread has the
917 		 * vnode lock before we get to VOP_LOCK() and obtains the
918 		 * vnode interlock after VOP_LOCK() drops the vnode
919 		 * interlock, the other thread will be unable to drop the
920 		 * vnode lock before our VOP_LOCK() call fails.
921 		 */
922 		if (vp->v_usecount ||
923 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
924 		    (vp->v_iflag & VI_FREE) != 0 ||
925 		    (vp->v_object != NULL &&
926 		    vp->v_object->resident_page_count > trigger)) {
927 			VOP_UNLOCK(vp, LK_INTERLOCK);
928 			vdrop(vp);
929 			goto next_iter_mntunlocked;
930 		}
931 		KASSERT((vp->v_iflag & VI_DOOMED) == 0,
932 		    ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
933 		counter_u64_add(recycles_count, 1);
934 		vgonel(vp);
935 		VOP_UNLOCK(vp, 0);
936 		vdropl(vp);
937 		done++;
938 next_iter_mntunlocked:
939 		if (!should_yield())
940 			goto relock_mnt;
941 		goto yield;
942 next_iter:
943 		if (!should_yield())
944 			continue;
945 		MNT_IUNLOCK(mp);
946 yield:
947 		kern_yield(PRI_USER);
948 relock_mnt:
949 		MNT_ILOCK(mp);
950 	}
951 	MNT_IUNLOCK(mp);
952 	vn_finished_write(mp);
953 	return done;
954 }
955 
956 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
957 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
958     0,
959     "limit on vnode free requests per call to the vnlru_free routine");
960 
961 /*
962  * Attempt to reduce the free list by the requested amount.
963  */
964 static void
965 vnlru_free_locked(int count, struct vfsops *mnt_op)
966 {
967 	struct vnode *vp;
968 	struct mount *mp;
969 	bool tried_batches;
970 
971 	tried_batches = false;
972 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
973 	if (count > max_vnlru_free)
974 		count = max_vnlru_free;
975 	for (; count > 0; count--) {
976 		vp = TAILQ_FIRST(&vnode_free_list);
977 		/*
978 		 * The list can be modified while the free_list_mtx
979 		 * has been dropped and vp could be NULL here.
980 		 */
981 		if (vp == NULL) {
982 			if (tried_batches)
983 				break;
984 			mtx_unlock(&vnode_free_list_mtx);
985 			vnlru_return_batches(mnt_op);
986 			tried_batches = true;
987 			mtx_lock(&vnode_free_list_mtx);
988 			continue;
989 		}
990 
991 		VNASSERT(vp->v_op != NULL, vp,
992 		    ("vnlru_free: vnode already reclaimed."));
993 		KASSERT((vp->v_iflag & VI_FREE) != 0,
994 		    ("Removing vnode not on freelist"));
995 		KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
996 		    ("Mangling active vnode"));
997 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
998 
999 		/*
1000 		 * Don't recycle if our vnode is from different type
1001 		 * of mount point.  Note that mp is type-safe, the
1002 		 * check does not reach unmapped address even if
1003 		 * vnode is reclaimed.
1004 		 * Don't recycle if we can't get the interlock without
1005 		 * blocking.
1006 		 */
1007 		if ((mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1008 		    mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) {
1009 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist);
1010 			continue;
1011 		}
1012 		VNASSERT((vp->v_iflag & VI_FREE) != 0 && vp->v_holdcnt == 0,
1013 		    vp, ("vp inconsistent on freelist"));
1014 
1015 		/*
1016 		 * The clear of VI_FREE prevents activation of the
1017 		 * vnode.  There is no sense in putting the vnode on
1018 		 * the mount point active list, only to remove it
1019 		 * later during recycling.  Inline the relevant part
1020 		 * of vholdl(), to avoid triggering assertions or
1021 		 * activating.
1022 		 */
1023 		freevnodes--;
1024 		vp->v_iflag &= ~VI_FREE;
1025 		refcount_acquire(&vp->v_holdcnt);
1026 
1027 		mtx_unlock(&vnode_free_list_mtx);
1028 		VI_UNLOCK(vp);
1029 		vtryrecycle(vp);
1030 		/*
1031 		 * If the recycled succeeded this vdrop will actually free
1032 		 * the vnode.  If not it will simply place it back on
1033 		 * the free list.
1034 		 */
1035 		vdrop(vp);
1036 		mtx_lock(&vnode_free_list_mtx);
1037 	}
1038 }
1039 
1040 void
1041 vnlru_free(int count, struct vfsops *mnt_op)
1042 {
1043 
1044 	mtx_lock(&vnode_free_list_mtx);
1045 	vnlru_free_locked(count, mnt_op);
1046 	mtx_unlock(&vnode_free_list_mtx);
1047 }
1048 
1049 
1050 /* XXX some names and initialization are bad for limits and watermarks. */
1051 static int
1052 vspace(void)
1053 {
1054 	int space;
1055 
1056 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1057 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1058 	vlowat = vhiwat / 2;
1059 	if (numvnodes > desiredvnodes)
1060 		return (0);
1061 	space = desiredvnodes - numvnodes;
1062 	if (freevnodes > wantfreevnodes)
1063 		space += freevnodes - wantfreevnodes;
1064 	return (space);
1065 }
1066 
1067 static void
1068 vnlru_return_batch_locked(struct mount *mp)
1069 {
1070 	struct vnode *vp;
1071 
1072 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
1073 
1074 	if (mp->mnt_tmpfreevnodelistsize == 0)
1075 		return;
1076 
1077 	TAILQ_FOREACH(vp, &mp->mnt_tmpfreevnodelist, v_actfreelist) {
1078 		VNASSERT((vp->v_mflag & VMP_TMPMNTFREELIST) != 0, vp,
1079 		    ("vnode without VMP_TMPMNTFREELIST on mnt_tmpfreevnodelist"));
1080 		vp->v_mflag &= ~VMP_TMPMNTFREELIST;
1081 	}
1082 	mtx_lock(&vnode_free_list_mtx);
1083 	TAILQ_CONCAT(&vnode_free_list, &mp->mnt_tmpfreevnodelist, v_actfreelist);
1084 	freevnodes += mp->mnt_tmpfreevnodelistsize;
1085 	mtx_unlock(&vnode_free_list_mtx);
1086 	mp->mnt_tmpfreevnodelistsize = 0;
1087 }
1088 
1089 static void
1090 vnlru_return_batch(struct mount *mp)
1091 {
1092 
1093 	mtx_lock(&mp->mnt_listmtx);
1094 	vnlru_return_batch_locked(mp);
1095 	mtx_unlock(&mp->mnt_listmtx);
1096 }
1097 
1098 static void
1099 vnlru_return_batches(struct vfsops *mnt_op)
1100 {
1101 	struct mount *mp, *nmp;
1102 	bool need_unbusy;
1103 
1104 	mtx_lock(&mountlist_mtx);
1105 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1106 		need_unbusy = false;
1107 		if (mnt_op != NULL && mp->mnt_op != mnt_op)
1108 			goto next;
1109 		if (mp->mnt_tmpfreevnodelistsize == 0)
1110 			goto next;
1111 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK) == 0) {
1112 			vnlru_return_batch(mp);
1113 			need_unbusy = true;
1114 			mtx_lock(&mountlist_mtx);
1115 		}
1116 next:
1117 		nmp = TAILQ_NEXT(mp, mnt_list);
1118 		if (need_unbusy)
1119 			vfs_unbusy(mp);
1120 	}
1121 	mtx_unlock(&mountlist_mtx);
1122 }
1123 
1124 /*
1125  * Attempt to recycle vnodes in a context that is always safe to block.
1126  * Calling vlrurecycle() from the bowels of filesystem code has some
1127  * interesting deadlock problems.
1128  */
1129 static struct proc *vnlruproc;
1130 static int vnlruproc_sig;
1131 
1132 static void
1133 vnlru_proc(void)
1134 {
1135 	struct mount *mp, *nmp;
1136 	unsigned long ofreevnodes, onumvnodes;
1137 	int done, force, reclaim_nc_src, trigger, usevnodes;
1138 
1139 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1140 	    SHUTDOWN_PRI_FIRST);
1141 
1142 	force = 0;
1143 	for (;;) {
1144 		kproc_suspend_check(vnlruproc);
1145 		mtx_lock(&vnode_free_list_mtx);
1146 		/*
1147 		 * If numvnodes is too large (due to desiredvnodes being
1148 		 * adjusted using its sysctl, or emergency growth), first
1149 		 * try to reduce it by discarding from the free list.
1150 		 */
1151 		if (numvnodes > desiredvnodes)
1152 			vnlru_free_locked(numvnodes - desiredvnodes, NULL);
1153 		/*
1154 		 * Sleep if the vnode cache is in a good state.  This is
1155 		 * when it is not over-full and has space for about a 4%
1156 		 * or 9% expansion (by growing its size or inexcessively
1157 		 * reducing its free list).  Otherwise, try to reclaim
1158 		 * space for a 10% expansion.
1159 		 */
1160 		if (vstir && force == 0) {
1161 			force = 1;
1162 			vstir = 0;
1163 		}
1164 		if (vspace() >= vlowat && force == 0) {
1165 			vnlruproc_sig = 0;
1166 			wakeup(&vnlruproc_sig);
1167 			msleep(vnlruproc, &vnode_free_list_mtx,
1168 			    PVFS|PDROP, "vlruwt", hz);
1169 			continue;
1170 		}
1171 		mtx_unlock(&vnode_free_list_mtx);
1172 		done = 0;
1173 		ofreevnodes = freevnodes;
1174 		onumvnodes = numvnodes;
1175 		/*
1176 		 * Calculate parameters for recycling.  These are the same
1177 		 * throughout the loop to give some semblance of fairness.
1178 		 * The trigger point is to avoid recycling vnodes with lots
1179 		 * of resident pages.  We aren't trying to free memory; we
1180 		 * are trying to recycle or at least free vnodes.
1181 		 */
1182 		if (numvnodes <= desiredvnodes)
1183 			usevnodes = numvnodes - freevnodes;
1184 		else
1185 			usevnodes = numvnodes;
1186 		if (usevnodes <= 0)
1187 			usevnodes = 1;
1188 		/*
1189 		 * The trigger value is is chosen to give a conservatively
1190 		 * large value to ensure that it alone doesn't prevent
1191 		 * making progress.  The value can easily be so large that
1192 		 * it is effectively infinite in some congested and
1193 		 * misconfigured cases, and this is necessary.  Normally
1194 		 * it is about 8 to 100 (pages), which is quite large.
1195 		 */
1196 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1197 		if (force < 2)
1198 			trigger = vsmalltrigger;
1199 		reclaim_nc_src = force >= 3;
1200 		mtx_lock(&mountlist_mtx);
1201 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1202 			if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) {
1203 				nmp = TAILQ_NEXT(mp, mnt_list);
1204 				continue;
1205 			}
1206 			done += vlrureclaim(mp, reclaim_nc_src, trigger);
1207 			mtx_lock(&mountlist_mtx);
1208 			nmp = TAILQ_NEXT(mp, mnt_list);
1209 			vfs_unbusy(mp);
1210 		}
1211 		mtx_unlock(&mountlist_mtx);
1212 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1213 			uma_reclaim();
1214 		if (done == 0) {
1215 			if (force == 0 || force == 1) {
1216 				force = 2;
1217 				continue;
1218 			}
1219 			if (force == 2) {
1220 				force = 3;
1221 				continue;
1222 			}
1223 			force = 0;
1224 			vnlru_nowhere++;
1225 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1226 		} else
1227 			kern_yield(PRI_USER);
1228 		/*
1229 		 * After becoming active to expand above low water, keep
1230 		 * active until above high water.
1231 		 */
1232 		force = vspace() < vhiwat;
1233 	}
1234 }
1235 
1236 static struct kproc_desc vnlru_kp = {
1237 	"vnlru",
1238 	vnlru_proc,
1239 	&vnlruproc
1240 };
1241 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1242     &vnlru_kp);
1243 
1244 /*
1245  * Routines having to do with the management of the vnode table.
1246  */
1247 
1248 /*
1249  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1250  * before we actually vgone().  This function must be called with the vnode
1251  * held to prevent the vnode from being returned to the free list midway
1252  * through vgone().
1253  */
1254 static int
1255 vtryrecycle(struct vnode *vp)
1256 {
1257 	struct mount *vnmp;
1258 
1259 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1260 	VNASSERT(vp->v_holdcnt, vp,
1261 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1262 	/*
1263 	 * This vnode may found and locked via some other list, if so we
1264 	 * can't recycle it yet.
1265 	 */
1266 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1267 		CTR2(KTR_VFS,
1268 		    "%s: impossible to recycle, vp %p lock is already held",
1269 		    __func__, vp);
1270 		return (EWOULDBLOCK);
1271 	}
1272 	/*
1273 	 * Don't recycle if its filesystem is being suspended.
1274 	 */
1275 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1276 		VOP_UNLOCK(vp, 0);
1277 		CTR2(KTR_VFS,
1278 		    "%s: impossible to recycle, cannot start the write for %p",
1279 		    __func__, vp);
1280 		return (EBUSY);
1281 	}
1282 	/*
1283 	 * If we got this far, we need to acquire the interlock and see if
1284 	 * anyone picked up this vnode from another list.  If not, we will
1285 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1286 	 * will skip over it.
1287 	 */
1288 	VI_LOCK(vp);
1289 	if (vp->v_usecount) {
1290 		VOP_UNLOCK(vp, LK_INTERLOCK);
1291 		vn_finished_write(vnmp);
1292 		CTR2(KTR_VFS,
1293 		    "%s: impossible to recycle, %p is already referenced",
1294 		    __func__, vp);
1295 		return (EBUSY);
1296 	}
1297 	if ((vp->v_iflag & VI_DOOMED) == 0) {
1298 		counter_u64_add(recycles_count, 1);
1299 		vgonel(vp);
1300 	}
1301 	VOP_UNLOCK(vp, LK_INTERLOCK);
1302 	vn_finished_write(vnmp);
1303 	return (0);
1304 }
1305 
1306 static void
1307 vcheckspace(void)
1308 {
1309 
1310 	if (vspace() < vlowat && vnlruproc_sig == 0) {
1311 		vnlruproc_sig = 1;
1312 		wakeup(vnlruproc);
1313 	}
1314 }
1315 
1316 /*
1317  * Wait if necessary for space for a new vnode.
1318  */
1319 static int
1320 getnewvnode_wait(int suspended)
1321 {
1322 
1323 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1324 	if (numvnodes >= desiredvnodes) {
1325 		if (suspended) {
1326 			/*
1327 			 * The file system is being suspended.  We cannot
1328 			 * risk a deadlock here, so allow allocation of
1329 			 * another vnode even if this would give too many.
1330 			 */
1331 			return (0);
1332 		}
1333 		if (vnlruproc_sig == 0) {
1334 			vnlruproc_sig = 1;	/* avoid unnecessary wakeups */
1335 			wakeup(vnlruproc);
1336 		}
1337 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
1338 		    "vlruwk", hz);
1339 	}
1340 	/* Post-adjust like the pre-adjust in getnewvnode(). */
1341 	if (numvnodes + 1 > desiredvnodes && freevnodes > 1)
1342 		vnlru_free_locked(1, NULL);
1343 	return (numvnodes >= desiredvnodes ? ENFILE : 0);
1344 }
1345 
1346 /*
1347  * This hack is fragile, and probably not needed any more now that the
1348  * watermark handling works.
1349  */
1350 void
1351 getnewvnode_reserve(u_int count)
1352 {
1353 	struct thread *td;
1354 
1355 	/* Pre-adjust like the pre-adjust in getnewvnode(), with any count. */
1356 	/* XXX no longer so quick, but this part is not racy. */
1357 	mtx_lock(&vnode_free_list_mtx);
1358 	if (numvnodes + count > desiredvnodes && freevnodes > wantfreevnodes)
1359 		vnlru_free_locked(ulmin(numvnodes + count - desiredvnodes,
1360 		    freevnodes - wantfreevnodes), NULL);
1361 	mtx_unlock(&vnode_free_list_mtx);
1362 
1363 	td = curthread;
1364 	/* First try to be quick and racy. */
1365 	if (atomic_fetchadd_long(&numvnodes, count) + count <= desiredvnodes) {
1366 		td->td_vp_reserv += count;
1367 		vcheckspace();	/* XXX no longer so quick, but more racy */
1368 		return;
1369 	} else
1370 		atomic_subtract_long(&numvnodes, count);
1371 
1372 	mtx_lock(&vnode_free_list_mtx);
1373 	while (count > 0) {
1374 		if (getnewvnode_wait(0) == 0) {
1375 			count--;
1376 			td->td_vp_reserv++;
1377 			atomic_add_long(&numvnodes, 1);
1378 		}
1379 	}
1380 	vcheckspace();
1381 	mtx_unlock(&vnode_free_list_mtx);
1382 }
1383 
1384 /*
1385  * This hack is fragile, especially if desiredvnodes or wantvnodes are
1386  * misconfgured or changed significantly.  Reducing desiredvnodes below
1387  * the reserved amount should cause bizarre behaviour like reducing it
1388  * below the number of active vnodes -- the system will try to reduce
1389  * numvnodes to match, but should fail, so the subtraction below should
1390  * not overflow.
1391  */
1392 void
1393 getnewvnode_drop_reserve(void)
1394 {
1395 	struct thread *td;
1396 
1397 	td = curthread;
1398 	atomic_subtract_long(&numvnodes, td->td_vp_reserv);
1399 	td->td_vp_reserv = 0;
1400 }
1401 
1402 /*
1403  * Return the next vnode from the free list.
1404  */
1405 int
1406 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1407     struct vnode **vpp)
1408 {
1409 	struct vnode *vp;
1410 	struct thread *td;
1411 	struct lock_object *lo;
1412 	static int cyclecount;
1413 	int error;
1414 
1415 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1416 	vp = NULL;
1417 	td = curthread;
1418 	if (td->td_vp_reserv > 0) {
1419 		td->td_vp_reserv -= 1;
1420 		goto alloc;
1421 	}
1422 	mtx_lock(&vnode_free_list_mtx);
1423 	if (numvnodes < desiredvnodes)
1424 		cyclecount = 0;
1425 	else if (cyclecount++ >= freevnodes) {
1426 		cyclecount = 0;
1427 		vstir = 1;
1428 	}
1429 	/*
1430 	 * Grow the vnode cache if it will not be above its target max
1431 	 * after growing.  Otherwise, if the free list is nonempty, try
1432 	 * to reclaim 1 item from it before growing the cache (possibly
1433 	 * above its target max if the reclamation failed or is delayed).
1434 	 * Otherwise, wait for some space.  In all cases, schedule
1435 	 * vnlru_proc() if we are getting short of space.  The watermarks
1436 	 * should be chosen so that we never wait or even reclaim from
1437 	 * the free list to below its target minimum.
1438 	 */
1439 	if (numvnodes + 1 <= desiredvnodes)
1440 		;
1441 	else if (freevnodes > 0)
1442 		vnlru_free_locked(1, NULL);
1443 	else {
1444 		error = getnewvnode_wait(mp != NULL && (mp->mnt_kern_flag &
1445 		    MNTK_SUSPEND));
1446 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
1447 		if (error != 0) {
1448 			mtx_unlock(&vnode_free_list_mtx);
1449 			return (error);
1450 		}
1451 #endif
1452 	}
1453 	vcheckspace();
1454 	atomic_add_long(&numvnodes, 1);
1455 	mtx_unlock(&vnode_free_list_mtx);
1456 alloc:
1457 	counter_u64_add(vnodes_created, 1);
1458 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK);
1459 	/*
1460 	 * Locks are given the generic name "vnode" when created.
1461 	 * Follow the historic practice of using the filesystem
1462 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1463 	 *
1464 	 * Locks live in a witness group keyed on their name. Thus,
1465 	 * when a lock is renamed, it must also move from the witness
1466 	 * group of its old name to the witness group of its new name.
1467 	 *
1468 	 * The change only needs to be made when the vnode moves
1469 	 * from one filesystem type to another. We ensure that each
1470 	 * filesystem use a single static name pointer for its tag so
1471 	 * that we can compare pointers rather than doing a strcmp().
1472 	 */
1473 	lo = &vp->v_vnlock->lock_object;
1474 	if (lo->lo_name != tag) {
1475 		lo->lo_name = tag;
1476 		WITNESS_DESTROY(lo);
1477 		WITNESS_INIT(lo, tag);
1478 	}
1479 	/*
1480 	 * By default, don't allow shared locks unless filesystems opt-in.
1481 	 */
1482 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1483 	/*
1484 	 * Finalize various vnode identity bits.
1485 	 */
1486 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1487 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1488 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1489 	vp->v_type = VNON;
1490 	vp->v_tag = tag;
1491 	vp->v_op = vops;
1492 	v_init_counters(vp);
1493 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1494 #ifdef DIAGNOSTIC
1495 	if (mp == NULL && vops != &dead_vnodeops)
1496 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1497 #endif
1498 #ifdef MAC
1499 	mac_vnode_init(vp);
1500 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1501 		mac_vnode_associate_singlelabel(mp, vp);
1502 #endif
1503 	if (mp != NULL) {
1504 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1505 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1506 			vp->v_vflag |= VV_NOKNOTE;
1507 	}
1508 
1509 	/*
1510 	 * For the filesystems which do not use vfs_hash_insert(),
1511 	 * still initialize v_hash to have vfs_hash_index() useful.
1512 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1513 	 * its own hashing.
1514 	 */
1515 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1516 
1517 	*vpp = vp;
1518 	return (0);
1519 }
1520 
1521 /*
1522  * Delete from old mount point vnode list, if on one.
1523  */
1524 static void
1525 delmntque(struct vnode *vp)
1526 {
1527 	struct mount *mp;
1528 	int active;
1529 
1530 	mp = vp->v_mount;
1531 	if (mp == NULL)
1532 		return;
1533 	MNT_ILOCK(mp);
1534 	VI_LOCK(vp);
1535 	KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize,
1536 	    ("Active vnode list size %d > Vnode list size %d",
1537 	     mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize));
1538 	active = vp->v_iflag & VI_ACTIVE;
1539 	vp->v_iflag &= ~VI_ACTIVE;
1540 	if (active) {
1541 		mtx_lock(&mp->mnt_listmtx);
1542 		TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist);
1543 		mp->mnt_activevnodelistsize--;
1544 		mtx_unlock(&mp->mnt_listmtx);
1545 	}
1546 	vp->v_mount = NULL;
1547 	VI_UNLOCK(vp);
1548 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1549 		("bad mount point vnode list size"));
1550 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1551 	mp->mnt_nvnodelistsize--;
1552 	MNT_REL(mp);
1553 	MNT_IUNLOCK(mp);
1554 }
1555 
1556 static void
1557 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1558 {
1559 
1560 	vp->v_data = NULL;
1561 	vp->v_op = &dead_vnodeops;
1562 	vgone(vp);
1563 	vput(vp);
1564 }
1565 
1566 /*
1567  * Insert into list of vnodes for the new mount point, if available.
1568  */
1569 int
1570 insmntque1(struct vnode *vp, struct mount *mp,
1571 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1572 {
1573 
1574 	KASSERT(vp->v_mount == NULL,
1575 		("insmntque: vnode already on per mount vnode list"));
1576 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1577 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1578 
1579 	/*
1580 	 * We acquire the vnode interlock early to ensure that the
1581 	 * vnode cannot be recycled by another process releasing a
1582 	 * holdcnt on it before we get it on both the vnode list
1583 	 * and the active vnode list. The mount mutex protects only
1584 	 * manipulation of the vnode list and the vnode freelist
1585 	 * mutex protects only manipulation of the active vnode list.
1586 	 * Hence the need to hold the vnode interlock throughout.
1587 	 */
1588 	MNT_ILOCK(mp);
1589 	VI_LOCK(vp);
1590 	if (((mp->mnt_kern_flag & MNTK_NOINSMNTQ) != 0 &&
1591 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1592 	    mp->mnt_nvnodelistsize == 0)) &&
1593 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1594 		VI_UNLOCK(vp);
1595 		MNT_IUNLOCK(mp);
1596 		if (dtr != NULL)
1597 			dtr(vp, dtr_arg);
1598 		return (EBUSY);
1599 	}
1600 	vp->v_mount = mp;
1601 	MNT_REF(mp);
1602 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1603 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1604 		("neg mount point vnode list size"));
1605 	mp->mnt_nvnodelistsize++;
1606 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1607 	    ("Activating already active vnode"));
1608 	vp->v_iflag |= VI_ACTIVE;
1609 	mtx_lock(&mp->mnt_listmtx);
1610 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
1611 	mp->mnt_activevnodelistsize++;
1612 	mtx_unlock(&mp->mnt_listmtx);
1613 	VI_UNLOCK(vp);
1614 	MNT_IUNLOCK(mp);
1615 	return (0);
1616 }
1617 
1618 int
1619 insmntque(struct vnode *vp, struct mount *mp)
1620 {
1621 
1622 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1623 }
1624 
1625 /*
1626  * Flush out and invalidate all buffers associated with a bufobj
1627  * Called with the underlying object locked.
1628  */
1629 int
1630 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1631 {
1632 	int error;
1633 
1634 	BO_LOCK(bo);
1635 	if (flags & V_SAVE) {
1636 		error = bufobj_wwait(bo, slpflag, slptimeo);
1637 		if (error) {
1638 			BO_UNLOCK(bo);
1639 			return (error);
1640 		}
1641 		if (bo->bo_dirty.bv_cnt > 0) {
1642 			BO_UNLOCK(bo);
1643 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1644 				return (error);
1645 			/*
1646 			 * XXX We could save a lock/unlock if this was only
1647 			 * enabled under INVARIANTS
1648 			 */
1649 			BO_LOCK(bo);
1650 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1651 				panic("vinvalbuf: dirty bufs");
1652 		}
1653 	}
1654 	/*
1655 	 * If you alter this loop please notice that interlock is dropped and
1656 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1657 	 * no race conditions occur from this.
1658 	 */
1659 	do {
1660 		error = flushbuflist(&bo->bo_clean,
1661 		    flags, bo, slpflag, slptimeo);
1662 		if (error == 0 && !(flags & V_CLEANONLY))
1663 			error = flushbuflist(&bo->bo_dirty,
1664 			    flags, bo, slpflag, slptimeo);
1665 		if (error != 0 && error != EAGAIN) {
1666 			BO_UNLOCK(bo);
1667 			return (error);
1668 		}
1669 	} while (error != 0);
1670 
1671 	/*
1672 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1673 	 * have write I/O in-progress but if there is a VM object then the
1674 	 * VM object can also have read-I/O in-progress.
1675 	 */
1676 	do {
1677 		bufobj_wwait(bo, 0, 0);
1678 		if ((flags & V_VMIO) == 0) {
1679 			BO_UNLOCK(bo);
1680 			if (bo->bo_object != NULL) {
1681 				VM_OBJECT_WLOCK(bo->bo_object);
1682 				vm_object_pip_wait(bo->bo_object, "bovlbx");
1683 				VM_OBJECT_WUNLOCK(bo->bo_object);
1684 			}
1685 			BO_LOCK(bo);
1686 		}
1687 	} while (bo->bo_numoutput > 0);
1688 	BO_UNLOCK(bo);
1689 
1690 	/*
1691 	 * Destroy the copy in the VM cache, too.
1692 	 */
1693 	if (bo->bo_object != NULL &&
1694 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1695 		VM_OBJECT_WLOCK(bo->bo_object);
1696 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1697 		    OBJPR_CLEANONLY : 0);
1698 		VM_OBJECT_WUNLOCK(bo->bo_object);
1699 	}
1700 
1701 #ifdef INVARIANTS
1702 	BO_LOCK(bo);
1703 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1704 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1705 	    bo->bo_clean.bv_cnt > 0))
1706 		panic("vinvalbuf: flush failed");
1707 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1708 	    bo->bo_dirty.bv_cnt > 0)
1709 		panic("vinvalbuf: flush dirty failed");
1710 	BO_UNLOCK(bo);
1711 #endif
1712 	return (0);
1713 }
1714 
1715 /*
1716  * Flush out and invalidate all buffers associated with a vnode.
1717  * Called with the underlying object locked.
1718  */
1719 int
1720 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1721 {
1722 
1723 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1724 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1725 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1726 		return (0);
1727 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1728 }
1729 
1730 /*
1731  * Flush out buffers on the specified list.
1732  *
1733  */
1734 static int
1735 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1736     int slptimeo)
1737 {
1738 	struct buf *bp, *nbp;
1739 	int retval, error;
1740 	daddr_t lblkno;
1741 	b_xflags_t xflags;
1742 
1743 	ASSERT_BO_WLOCKED(bo);
1744 
1745 	retval = 0;
1746 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1747 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1748 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1749 			continue;
1750 		}
1751 		if (nbp != NULL) {
1752 			lblkno = nbp->b_lblkno;
1753 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
1754 		}
1755 		retval = EAGAIN;
1756 		error = BUF_TIMELOCK(bp,
1757 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
1758 		    "flushbuf", slpflag, slptimeo);
1759 		if (error) {
1760 			BO_LOCK(bo);
1761 			return (error != ENOLCK ? error : EAGAIN);
1762 		}
1763 		KASSERT(bp->b_bufobj == bo,
1764 		    ("bp %p wrong b_bufobj %p should be %p",
1765 		    bp, bp->b_bufobj, bo));
1766 		/*
1767 		 * XXX Since there are no node locks for NFS, I
1768 		 * believe there is a slight chance that a delayed
1769 		 * write will occur while sleeping just above, so
1770 		 * check for it.
1771 		 */
1772 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1773 		    (flags & V_SAVE)) {
1774 			bremfree(bp);
1775 			bp->b_flags |= B_ASYNC;
1776 			bwrite(bp);
1777 			BO_LOCK(bo);
1778 			return (EAGAIN);	/* XXX: why not loop ? */
1779 		}
1780 		bremfree(bp);
1781 		bp->b_flags |= (B_INVAL | B_RELBUF);
1782 		bp->b_flags &= ~B_ASYNC;
1783 		brelse(bp);
1784 		BO_LOCK(bo);
1785 		if (nbp == NULL)
1786 			break;
1787 		nbp = gbincore(bo, lblkno);
1788 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1789 		    != xflags)
1790 			break;			/* nbp invalid */
1791 	}
1792 	return (retval);
1793 }
1794 
1795 int
1796 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
1797 {
1798 	struct buf *bp;
1799 	int error;
1800 	daddr_t lblkno;
1801 
1802 	ASSERT_BO_LOCKED(bo);
1803 
1804 	for (lblkno = startn;;) {
1805 again:
1806 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
1807 		if (bp == NULL || bp->b_lblkno >= endn ||
1808 		    bp->b_lblkno < startn)
1809 			break;
1810 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
1811 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
1812 		if (error != 0) {
1813 			BO_RLOCK(bo);
1814 			if (error == ENOLCK)
1815 				goto again;
1816 			return (error);
1817 		}
1818 		KASSERT(bp->b_bufobj == bo,
1819 		    ("bp %p wrong b_bufobj %p should be %p",
1820 		    bp, bp->b_bufobj, bo));
1821 		lblkno = bp->b_lblkno + 1;
1822 		if ((bp->b_flags & B_MANAGED) == 0)
1823 			bremfree(bp);
1824 		bp->b_flags |= B_RELBUF;
1825 		/*
1826 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
1827 		 * pages backing each buffer in the range are unlikely to be
1828 		 * reused.  Dirty buffers will have the hint applied once
1829 		 * they've been written.
1830 		 */
1831 		if (bp->b_vp->v_object != NULL)
1832 			bp->b_flags |= B_NOREUSE;
1833 		brelse(bp);
1834 		BO_RLOCK(bo);
1835 	}
1836 	return (0);
1837 }
1838 
1839 /*
1840  * Truncate a file's buffer and pages to a specified length.  This
1841  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1842  * sync activity.
1843  */
1844 int
1845 vtruncbuf(struct vnode *vp, struct ucred *cred, off_t length, int blksize)
1846 {
1847 	struct buf *bp, *nbp;
1848 	int anyfreed;
1849 	int trunclbn;
1850 	struct bufobj *bo;
1851 
1852 	CTR5(KTR_VFS, "%s: vp %p with cred %p and block %d:%ju", __func__,
1853 	    vp, cred, blksize, (uintmax_t)length);
1854 
1855 	/*
1856 	 * Round up to the *next* lbn.
1857 	 */
1858 	trunclbn = howmany(length, blksize);
1859 
1860 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1861 restart:
1862 	bo = &vp->v_bufobj;
1863 	BO_LOCK(bo);
1864 	anyfreed = 1;
1865 	for (;anyfreed;) {
1866 		anyfreed = 0;
1867 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1868 			if (bp->b_lblkno < trunclbn)
1869 				continue;
1870 			if (BUF_LOCK(bp,
1871 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1872 			    BO_LOCKPTR(bo)) == ENOLCK)
1873 				goto restart;
1874 
1875 			bremfree(bp);
1876 			bp->b_flags |= (B_INVAL | B_RELBUF);
1877 			bp->b_flags &= ~B_ASYNC;
1878 			brelse(bp);
1879 			anyfreed = 1;
1880 
1881 			BO_LOCK(bo);
1882 			if (nbp != NULL &&
1883 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1884 			    (nbp->b_vp != vp) ||
1885 			    (nbp->b_flags & B_DELWRI))) {
1886 				BO_UNLOCK(bo);
1887 				goto restart;
1888 			}
1889 		}
1890 
1891 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1892 			if (bp->b_lblkno < trunclbn)
1893 				continue;
1894 			if (BUF_LOCK(bp,
1895 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1896 			    BO_LOCKPTR(bo)) == ENOLCK)
1897 				goto restart;
1898 			bremfree(bp);
1899 			bp->b_flags |= (B_INVAL | B_RELBUF);
1900 			bp->b_flags &= ~B_ASYNC;
1901 			brelse(bp);
1902 			anyfreed = 1;
1903 
1904 			BO_LOCK(bo);
1905 			if (nbp != NULL &&
1906 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1907 			    (nbp->b_vp != vp) ||
1908 			    (nbp->b_flags & B_DELWRI) == 0)) {
1909 				BO_UNLOCK(bo);
1910 				goto restart;
1911 			}
1912 		}
1913 	}
1914 
1915 	if (length > 0) {
1916 restartsync:
1917 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1918 			if (bp->b_lblkno > 0)
1919 				continue;
1920 			/*
1921 			 * Since we hold the vnode lock this should only
1922 			 * fail if we're racing with the buf daemon.
1923 			 */
1924 			if (BUF_LOCK(bp,
1925 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1926 			    BO_LOCKPTR(bo)) == ENOLCK) {
1927 				goto restart;
1928 			}
1929 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1930 			    ("buf(%p) on dirty queue without DELWRI", bp));
1931 
1932 			bremfree(bp);
1933 			bawrite(bp);
1934 			BO_LOCK(bo);
1935 			goto restartsync;
1936 		}
1937 	}
1938 
1939 	bufobj_wwait(bo, 0, 0);
1940 	BO_UNLOCK(bo);
1941 	vnode_pager_setsize(vp, length);
1942 
1943 	return (0);
1944 }
1945 
1946 static void
1947 buf_vlist_remove(struct buf *bp)
1948 {
1949 	struct bufv *bv;
1950 
1951 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1952 	ASSERT_BO_WLOCKED(bp->b_bufobj);
1953 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
1954 	    (BX_VNDIRTY|BX_VNCLEAN),
1955 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
1956 	if (bp->b_xflags & BX_VNDIRTY)
1957 		bv = &bp->b_bufobj->bo_dirty;
1958 	else
1959 		bv = &bp->b_bufobj->bo_clean;
1960 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
1961 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1962 	bv->bv_cnt--;
1963 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1964 }
1965 
1966 /*
1967  * Add the buffer to the sorted clean or dirty block list.
1968  *
1969  * NOTE: xflags is passed as a constant, optimizing this inline function!
1970  */
1971 static void
1972 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1973 {
1974 	struct bufv *bv;
1975 	struct buf *n;
1976 	int error;
1977 
1978 	ASSERT_BO_WLOCKED(bo);
1979 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
1980 	    ("dead bo %p", bo));
1981 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1982 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
1983 	bp->b_xflags |= xflags;
1984 	if (xflags & BX_VNDIRTY)
1985 		bv = &bo->bo_dirty;
1986 	else
1987 		bv = &bo->bo_clean;
1988 
1989 	/*
1990 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
1991 	 * we tend to grow at the tail so lookup_le should usually be cheaper
1992 	 * than _ge.
1993 	 */
1994 	if (bv->bv_cnt == 0 ||
1995 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
1996 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1997 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
1998 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
1999 	else
2000 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2001 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2002 	if (error)
2003 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2004 	bv->bv_cnt++;
2005 }
2006 
2007 /*
2008  * Look up a buffer using the buffer tries.
2009  */
2010 struct buf *
2011 gbincore(struct bufobj *bo, daddr_t lblkno)
2012 {
2013 	struct buf *bp;
2014 
2015 	ASSERT_BO_LOCKED(bo);
2016 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2017 	if (bp != NULL)
2018 		return (bp);
2019 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
2020 }
2021 
2022 /*
2023  * Associate a buffer with a vnode.
2024  */
2025 void
2026 bgetvp(struct vnode *vp, struct buf *bp)
2027 {
2028 	struct bufobj *bo;
2029 
2030 	bo = &vp->v_bufobj;
2031 	ASSERT_BO_WLOCKED(bo);
2032 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2033 
2034 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2035 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2036 	    ("bgetvp: bp already attached! %p", bp));
2037 
2038 	vhold(vp);
2039 	bp->b_vp = vp;
2040 	bp->b_bufobj = bo;
2041 	/*
2042 	 * Insert onto list for new vnode.
2043 	 */
2044 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2045 }
2046 
2047 /*
2048  * Disassociate a buffer from a vnode.
2049  */
2050 void
2051 brelvp(struct buf *bp)
2052 {
2053 	struct bufobj *bo;
2054 	struct vnode *vp;
2055 
2056 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2057 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2058 
2059 	/*
2060 	 * Delete from old vnode list, if on one.
2061 	 */
2062 	vp = bp->b_vp;		/* XXX */
2063 	bo = bp->b_bufobj;
2064 	BO_LOCK(bo);
2065 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2066 		buf_vlist_remove(bp);
2067 	else
2068 		panic("brelvp: Buffer %p not on queue.", bp);
2069 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2070 		bo->bo_flag &= ~BO_ONWORKLST;
2071 		mtx_lock(&sync_mtx);
2072 		LIST_REMOVE(bo, bo_synclist);
2073 		syncer_worklist_len--;
2074 		mtx_unlock(&sync_mtx);
2075 	}
2076 	bp->b_vp = NULL;
2077 	bp->b_bufobj = NULL;
2078 	BO_UNLOCK(bo);
2079 	vdrop(vp);
2080 }
2081 
2082 /*
2083  * Add an item to the syncer work queue.
2084  */
2085 static void
2086 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2087 {
2088 	int slot;
2089 
2090 	ASSERT_BO_WLOCKED(bo);
2091 
2092 	mtx_lock(&sync_mtx);
2093 	if (bo->bo_flag & BO_ONWORKLST)
2094 		LIST_REMOVE(bo, bo_synclist);
2095 	else {
2096 		bo->bo_flag |= BO_ONWORKLST;
2097 		syncer_worklist_len++;
2098 	}
2099 
2100 	if (delay > syncer_maxdelay - 2)
2101 		delay = syncer_maxdelay - 2;
2102 	slot = (syncer_delayno + delay) & syncer_mask;
2103 
2104 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2105 	mtx_unlock(&sync_mtx);
2106 }
2107 
2108 static int
2109 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2110 {
2111 	int error, len;
2112 
2113 	mtx_lock(&sync_mtx);
2114 	len = syncer_worklist_len - sync_vnode_count;
2115 	mtx_unlock(&sync_mtx);
2116 	error = SYSCTL_OUT(req, &len, sizeof(len));
2117 	return (error);
2118 }
2119 
2120 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
2121     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2122 
2123 static struct proc *updateproc;
2124 static void sched_sync(void);
2125 static struct kproc_desc up_kp = {
2126 	"syncer",
2127 	sched_sync,
2128 	&updateproc
2129 };
2130 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2131 
2132 static int
2133 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2134 {
2135 	struct vnode *vp;
2136 	struct mount *mp;
2137 
2138 	*bo = LIST_FIRST(slp);
2139 	if (*bo == NULL)
2140 		return (0);
2141 	vp = bo2vnode(*bo);
2142 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2143 		return (1);
2144 	/*
2145 	 * We use vhold in case the vnode does not
2146 	 * successfully sync.  vhold prevents the vnode from
2147 	 * going away when we unlock the sync_mtx so that
2148 	 * we can acquire the vnode interlock.
2149 	 */
2150 	vholdl(vp);
2151 	mtx_unlock(&sync_mtx);
2152 	VI_UNLOCK(vp);
2153 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2154 		vdrop(vp);
2155 		mtx_lock(&sync_mtx);
2156 		return (*bo == LIST_FIRST(slp));
2157 	}
2158 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2159 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2160 	VOP_UNLOCK(vp, 0);
2161 	vn_finished_write(mp);
2162 	BO_LOCK(*bo);
2163 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2164 		/*
2165 		 * Put us back on the worklist.  The worklist
2166 		 * routine will remove us from our current
2167 		 * position and then add us back in at a later
2168 		 * position.
2169 		 */
2170 		vn_syncer_add_to_worklist(*bo, syncdelay);
2171 	}
2172 	BO_UNLOCK(*bo);
2173 	vdrop(vp);
2174 	mtx_lock(&sync_mtx);
2175 	return (0);
2176 }
2177 
2178 static int first_printf = 1;
2179 
2180 /*
2181  * System filesystem synchronizer daemon.
2182  */
2183 static void
2184 sched_sync(void)
2185 {
2186 	struct synclist *next, *slp;
2187 	struct bufobj *bo;
2188 	long starttime;
2189 	struct thread *td = curthread;
2190 	int last_work_seen;
2191 	int net_worklist_len;
2192 	int syncer_final_iter;
2193 	int error;
2194 
2195 	last_work_seen = 0;
2196 	syncer_final_iter = 0;
2197 	syncer_state = SYNCER_RUNNING;
2198 	starttime = time_uptime;
2199 	td->td_pflags |= TDP_NORUNNINGBUF;
2200 
2201 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2202 	    SHUTDOWN_PRI_LAST);
2203 
2204 	mtx_lock(&sync_mtx);
2205 	for (;;) {
2206 		if (syncer_state == SYNCER_FINAL_DELAY &&
2207 		    syncer_final_iter == 0) {
2208 			mtx_unlock(&sync_mtx);
2209 			kproc_suspend_check(td->td_proc);
2210 			mtx_lock(&sync_mtx);
2211 		}
2212 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2213 		if (syncer_state != SYNCER_RUNNING &&
2214 		    starttime != time_uptime) {
2215 			if (first_printf) {
2216 				printf("\nSyncing disks, vnodes remaining... ");
2217 				first_printf = 0;
2218 			}
2219 			printf("%d ", net_worklist_len);
2220 		}
2221 		starttime = time_uptime;
2222 
2223 		/*
2224 		 * Push files whose dirty time has expired.  Be careful
2225 		 * of interrupt race on slp queue.
2226 		 *
2227 		 * Skip over empty worklist slots when shutting down.
2228 		 */
2229 		do {
2230 			slp = &syncer_workitem_pending[syncer_delayno];
2231 			syncer_delayno += 1;
2232 			if (syncer_delayno == syncer_maxdelay)
2233 				syncer_delayno = 0;
2234 			next = &syncer_workitem_pending[syncer_delayno];
2235 			/*
2236 			 * If the worklist has wrapped since the
2237 			 * it was emptied of all but syncer vnodes,
2238 			 * switch to the FINAL_DELAY state and run
2239 			 * for one more second.
2240 			 */
2241 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2242 			    net_worklist_len == 0 &&
2243 			    last_work_seen == syncer_delayno) {
2244 				syncer_state = SYNCER_FINAL_DELAY;
2245 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2246 			}
2247 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2248 		    syncer_worklist_len > 0);
2249 
2250 		/*
2251 		 * Keep track of the last time there was anything
2252 		 * on the worklist other than syncer vnodes.
2253 		 * Return to the SHUTTING_DOWN state if any
2254 		 * new work appears.
2255 		 */
2256 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2257 			last_work_seen = syncer_delayno;
2258 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2259 			syncer_state = SYNCER_SHUTTING_DOWN;
2260 		while (!LIST_EMPTY(slp)) {
2261 			error = sync_vnode(slp, &bo, td);
2262 			if (error == 1) {
2263 				LIST_REMOVE(bo, bo_synclist);
2264 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2265 				continue;
2266 			}
2267 
2268 			if (first_printf == 0) {
2269 				/*
2270 				 * Drop the sync mutex, because some watchdog
2271 				 * drivers need to sleep while patting
2272 				 */
2273 				mtx_unlock(&sync_mtx);
2274 				wdog_kern_pat(WD_LASTVAL);
2275 				mtx_lock(&sync_mtx);
2276 			}
2277 
2278 		}
2279 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2280 			syncer_final_iter--;
2281 		/*
2282 		 * The variable rushjob allows the kernel to speed up the
2283 		 * processing of the filesystem syncer process. A rushjob
2284 		 * value of N tells the filesystem syncer to process the next
2285 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2286 		 * is used by the soft update code to speed up the filesystem
2287 		 * syncer process when the incore state is getting so far
2288 		 * ahead of the disk that the kernel memory pool is being
2289 		 * threatened with exhaustion.
2290 		 */
2291 		if (rushjob > 0) {
2292 			rushjob -= 1;
2293 			continue;
2294 		}
2295 		/*
2296 		 * Just sleep for a short period of time between
2297 		 * iterations when shutting down to allow some I/O
2298 		 * to happen.
2299 		 *
2300 		 * If it has taken us less than a second to process the
2301 		 * current work, then wait. Otherwise start right over
2302 		 * again. We can still lose time if any single round
2303 		 * takes more than two seconds, but it does not really
2304 		 * matter as we are just trying to generally pace the
2305 		 * filesystem activity.
2306 		 */
2307 		if (syncer_state != SYNCER_RUNNING ||
2308 		    time_uptime == starttime) {
2309 			thread_lock(td);
2310 			sched_prio(td, PPAUSE);
2311 			thread_unlock(td);
2312 		}
2313 		if (syncer_state != SYNCER_RUNNING)
2314 			cv_timedwait(&sync_wakeup, &sync_mtx,
2315 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2316 		else if (time_uptime == starttime)
2317 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2318 	}
2319 }
2320 
2321 /*
2322  * Request the syncer daemon to speed up its work.
2323  * We never push it to speed up more than half of its
2324  * normal turn time, otherwise it could take over the cpu.
2325  */
2326 int
2327 speedup_syncer(void)
2328 {
2329 	int ret = 0;
2330 
2331 	mtx_lock(&sync_mtx);
2332 	if (rushjob < syncdelay / 2) {
2333 		rushjob += 1;
2334 		stat_rush_requests += 1;
2335 		ret = 1;
2336 	}
2337 	mtx_unlock(&sync_mtx);
2338 	cv_broadcast(&sync_wakeup);
2339 	return (ret);
2340 }
2341 
2342 /*
2343  * Tell the syncer to speed up its work and run though its work
2344  * list several times, then tell it to shut down.
2345  */
2346 static void
2347 syncer_shutdown(void *arg, int howto)
2348 {
2349 
2350 	if (howto & RB_NOSYNC)
2351 		return;
2352 	mtx_lock(&sync_mtx);
2353 	syncer_state = SYNCER_SHUTTING_DOWN;
2354 	rushjob = 0;
2355 	mtx_unlock(&sync_mtx);
2356 	cv_broadcast(&sync_wakeup);
2357 	kproc_shutdown(arg, howto);
2358 }
2359 
2360 void
2361 syncer_suspend(void)
2362 {
2363 
2364 	syncer_shutdown(updateproc, 0);
2365 }
2366 
2367 void
2368 syncer_resume(void)
2369 {
2370 
2371 	mtx_lock(&sync_mtx);
2372 	first_printf = 1;
2373 	syncer_state = SYNCER_RUNNING;
2374 	mtx_unlock(&sync_mtx);
2375 	cv_broadcast(&sync_wakeup);
2376 	kproc_resume(updateproc);
2377 }
2378 
2379 /*
2380  * Reassign a buffer from one vnode to another.
2381  * Used to assign file specific control information
2382  * (indirect blocks) to the vnode to which they belong.
2383  */
2384 void
2385 reassignbuf(struct buf *bp)
2386 {
2387 	struct vnode *vp;
2388 	struct bufobj *bo;
2389 	int delay;
2390 #ifdef INVARIANTS
2391 	struct bufv *bv;
2392 #endif
2393 
2394 	vp = bp->b_vp;
2395 	bo = bp->b_bufobj;
2396 	++reassignbufcalls;
2397 
2398 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2399 	    bp, bp->b_vp, bp->b_flags);
2400 	/*
2401 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2402 	 * is not fully linked in.
2403 	 */
2404 	if (bp->b_flags & B_PAGING)
2405 		panic("cannot reassign paging buffer");
2406 
2407 	/*
2408 	 * Delete from old vnode list, if on one.
2409 	 */
2410 	BO_LOCK(bo);
2411 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2412 		buf_vlist_remove(bp);
2413 	else
2414 		panic("reassignbuf: Buffer %p not on queue.", bp);
2415 	/*
2416 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2417 	 * of clean buffers.
2418 	 */
2419 	if (bp->b_flags & B_DELWRI) {
2420 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2421 			switch (vp->v_type) {
2422 			case VDIR:
2423 				delay = dirdelay;
2424 				break;
2425 			case VCHR:
2426 				delay = metadelay;
2427 				break;
2428 			default:
2429 				delay = filedelay;
2430 			}
2431 			vn_syncer_add_to_worklist(bo, delay);
2432 		}
2433 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2434 	} else {
2435 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2436 
2437 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2438 			mtx_lock(&sync_mtx);
2439 			LIST_REMOVE(bo, bo_synclist);
2440 			syncer_worklist_len--;
2441 			mtx_unlock(&sync_mtx);
2442 			bo->bo_flag &= ~BO_ONWORKLST;
2443 		}
2444 	}
2445 #ifdef INVARIANTS
2446 	bv = &bo->bo_clean;
2447 	bp = TAILQ_FIRST(&bv->bv_hd);
2448 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2449 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2450 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2451 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2452 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2453 	bv = &bo->bo_dirty;
2454 	bp = TAILQ_FIRST(&bv->bv_hd);
2455 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2456 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2457 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2458 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2459 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2460 #endif
2461 	BO_UNLOCK(bo);
2462 }
2463 
2464 /*
2465  * A temporary hack until refcount_* APIs are sorted out.
2466  */
2467 static __inline int
2468 vfs_refcount_acquire_if_not_zero(volatile u_int *count)
2469 {
2470 	u_int old;
2471 
2472 	old = *count;
2473 	for (;;) {
2474 		if (old == 0)
2475 			return (0);
2476 		if (atomic_fcmpset_int(count, &old, old + 1))
2477 			return (1);
2478 	}
2479 }
2480 
2481 static __inline int
2482 vfs_refcount_release_if_not_last(volatile u_int *count)
2483 {
2484 	u_int old;
2485 
2486 	old = *count;
2487 	for (;;) {
2488 		if (old == 1)
2489 			return (0);
2490 		if (atomic_fcmpset_int(count, &old, old - 1))
2491 			return (1);
2492 	}
2493 }
2494 
2495 static void
2496 v_init_counters(struct vnode *vp)
2497 {
2498 
2499 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2500 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2501 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2502 
2503 	refcount_init(&vp->v_holdcnt, 1);
2504 	refcount_init(&vp->v_usecount, 1);
2505 }
2506 
2507 static void
2508 v_incr_usecount_locked(struct vnode *vp)
2509 {
2510 
2511 	ASSERT_VI_LOCKED(vp, __func__);
2512 	if ((vp->v_iflag & VI_OWEINACT) != 0) {
2513 		VNASSERT(vp->v_usecount == 0, vp,
2514 		    ("vnode with usecount and VI_OWEINACT set"));
2515 		vp->v_iflag &= ~VI_OWEINACT;
2516 	}
2517 	refcount_acquire(&vp->v_usecount);
2518 	v_incr_devcount(vp);
2519 }
2520 
2521 /*
2522  * Increment the use count on the vnode, taking care to reference
2523  * the driver's usecount if this is a chardev.
2524  */
2525 static void
2526 v_incr_usecount(struct vnode *vp)
2527 {
2528 
2529 	ASSERT_VI_UNLOCKED(vp, __func__);
2530 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2531 
2532 	if (vp->v_type != VCHR &&
2533 	    vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) {
2534 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2535 		    ("vnode with usecount and VI_OWEINACT set"));
2536 	} else {
2537 		VI_LOCK(vp);
2538 		v_incr_usecount_locked(vp);
2539 		VI_UNLOCK(vp);
2540 	}
2541 }
2542 
2543 /*
2544  * Increment si_usecount of the associated device, if any.
2545  */
2546 static void
2547 v_incr_devcount(struct vnode *vp)
2548 {
2549 
2550 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2551 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2552 		dev_lock();
2553 		vp->v_rdev->si_usecount++;
2554 		dev_unlock();
2555 	}
2556 }
2557 
2558 /*
2559  * Decrement si_usecount of the associated device, if any.
2560  */
2561 static void
2562 v_decr_devcount(struct vnode *vp)
2563 {
2564 
2565 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2566 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2567 		dev_lock();
2568 		vp->v_rdev->si_usecount--;
2569 		dev_unlock();
2570 	}
2571 }
2572 
2573 /*
2574  * Grab a particular vnode from the free list, increment its
2575  * reference count and lock it.  VI_DOOMED is set if the vnode
2576  * is being destroyed.  Only callers who specify LK_RETRY will
2577  * see doomed vnodes.  If inactive processing was delayed in
2578  * vput try to do it here.
2579  *
2580  * Notes on lockless counter manipulation:
2581  * _vhold, vputx and other routines make various decisions based
2582  * on either holdcnt or usecount being 0. As long as either counter
2583  * is not transitioning 0->1 nor 1->0, the manipulation can be done
2584  * with atomic operations. Otherwise the interlock is taken covering
2585  * both the atomic and additional actions.
2586  */
2587 int
2588 vget(struct vnode *vp, int flags, struct thread *td)
2589 {
2590 	int error, oweinact;
2591 
2592 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2593 	    ("vget: invalid lock operation"));
2594 
2595 	if ((flags & LK_INTERLOCK) != 0)
2596 		ASSERT_VI_LOCKED(vp, __func__);
2597 	else
2598 		ASSERT_VI_UNLOCKED(vp, __func__);
2599 	if ((flags & LK_VNHELD) != 0)
2600 		VNASSERT((vp->v_holdcnt > 0), vp,
2601 		    ("vget: LK_VNHELD passed but vnode not held"));
2602 
2603 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2604 
2605 	if ((flags & LK_VNHELD) == 0)
2606 		_vhold(vp, (flags & LK_INTERLOCK) != 0);
2607 
2608 	if ((error = vn_lock(vp, flags)) != 0) {
2609 		vdrop(vp);
2610 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2611 		    vp);
2612 		return (error);
2613 	}
2614 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
2615 		panic("vget: vn_lock failed to return ENOENT\n");
2616 	/*
2617 	 * We don't guarantee that any particular close will
2618 	 * trigger inactive processing so just make a best effort
2619 	 * here at preventing a reference to a removed file.  If
2620 	 * we don't succeed no harm is done.
2621 	 *
2622 	 * Upgrade our holdcnt to a usecount.
2623 	 */
2624 	if (vp->v_type == VCHR ||
2625 	    !vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) {
2626 		VI_LOCK(vp);
2627 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2628 			oweinact = 0;
2629 		} else {
2630 			oweinact = 1;
2631 			vp->v_iflag &= ~VI_OWEINACT;
2632 		}
2633 		refcount_acquire(&vp->v_usecount);
2634 		v_incr_devcount(vp);
2635 		if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2636 		    (flags & LK_NOWAIT) == 0)
2637 			vinactive(vp, td);
2638 		VI_UNLOCK(vp);
2639 	}
2640 	return (0);
2641 }
2642 
2643 /*
2644  * Increase the reference (use) and hold count of a vnode.
2645  * This will also remove the vnode from the free list if it is presently free.
2646  */
2647 void
2648 vref(struct vnode *vp)
2649 {
2650 
2651 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2652 	_vhold(vp, false);
2653 	v_incr_usecount(vp);
2654 }
2655 
2656 void
2657 vrefl(struct vnode *vp)
2658 {
2659 
2660 	ASSERT_VI_LOCKED(vp, __func__);
2661 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2662 	_vhold(vp, true);
2663 	v_incr_usecount_locked(vp);
2664 }
2665 
2666 void
2667 vrefact(struct vnode *vp)
2668 {
2669 
2670 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2671 	if (__predict_false(vp->v_type == VCHR)) {
2672 		VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp,
2673 		    ("%s: wrong ref counts", __func__));
2674 		vref(vp);
2675 		return;
2676 	}
2677 #ifdef INVARIANTS
2678 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
2679 	VNASSERT(old > 0, vp, ("%s: wrong hold count", __func__));
2680 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
2681 	VNASSERT(old > 0, vp, ("%s: wrong use count", __func__));
2682 #else
2683 	refcount_acquire(&vp->v_holdcnt);
2684 	refcount_acquire(&vp->v_usecount);
2685 #endif
2686 }
2687 
2688 /*
2689  * Return reference count of a vnode.
2690  *
2691  * The results of this call are only guaranteed when some mechanism is used to
2692  * stop other processes from gaining references to the vnode.  This may be the
2693  * case if the caller holds the only reference.  This is also useful when stale
2694  * data is acceptable as race conditions may be accounted for by some other
2695  * means.
2696  */
2697 int
2698 vrefcnt(struct vnode *vp)
2699 {
2700 
2701 	return (vp->v_usecount);
2702 }
2703 
2704 #define	VPUTX_VRELE	1
2705 #define	VPUTX_VPUT	2
2706 #define	VPUTX_VUNREF	3
2707 
2708 /*
2709  * Decrement the use and hold counts for a vnode.
2710  *
2711  * See an explanation near vget() as to why atomic operation is safe.
2712  */
2713 static void
2714 vputx(struct vnode *vp, int func)
2715 {
2716 	int error;
2717 
2718 	KASSERT(vp != NULL, ("vputx: null vp"));
2719 	if (func == VPUTX_VUNREF)
2720 		ASSERT_VOP_LOCKED(vp, "vunref");
2721 	else if (func == VPUTX_VPUT)
2722 		ASSERT_VOP_LOCKED(vp, "vput");
2723 	else
2724 		KASSERT(func == VPUTX_VRELE, ("vputx: wrong func"));
2725 	ASSERT_VI_UNLOCKED(vp, __func__);
2726 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2727 
2728 	if (vp->v_type != VCHR &&
2729 	    vfs_refcount_release_if_not_last(&vp->v_usecount)) {
2730 		if (func == VPUTX_VPUT)
2731 			VOP_UNLOCK(vp, 0);
2732 		vdrop(vp);
2733 		return;
2734 	}
2735 
2736 	VI_LOCK(vp);
2737 
2738 	/*
2739 	 * We want to hold the vnode until the inactive finishes to
2740 	 * prevent vgone() races.  We drop the use count here and the
2741 	 * hold count below when we're done.
2742 	 */
2743 	if (!refcount_release(&vp->v_usecount) ||
2744 	    (vp->v_iflag & VI_DOINGINACT)) {
2745 		if (func == VPUTX_VPUT)
2746 			VOP_UNLOCK(vp, 0);
2747 		v_decr_devcount(vp);
2748 		vdropl(vp);
2749 		return;
2750 	}
2751 
2752 	v_decr_devcount(vp);
2753 
2754 	error = 0;
2755 
2756 	if (vp->v_usecount != 0) {
2757 		vn_printf(vp, "vputx: usecount not zero for vnode ");
2758 		panic("vputx: usecount not zero");
2759 	}
2760 
2761 	CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp);
2762 
2763 	/*
2764 	 * We must call VOP_INACTIVE with the node locked. Mark
2765 	 * as VI_DOINGINACT to avoid recursion.
2766 	 */
2767 	vp->v_iflag |= VI_OWEINACT;
2768 	switch (func) {
2769 	case VPUTX_VRELE:
2770 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
2771 		VI_LOCK(vp);
2772 		break;
2773 	case VPUTX_VPUT:
2774 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2775 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
2776 			    LK_NOWAIT);
2777 			VI_LOCK(vp);
2778 		}
2779 		break;
2780 	case VPUTX_VUNREF:
2781 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2782 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
2783 			VI_LOCK(vp);
2784 		}
2785 		break;
2786 	}
2787 	VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
2788 	    ("vnode with usecount and VI_OWEINACT set"));
2789 	if (error == 0) {
2790 		if (vp->v_iflag & VI_OWEINACT)
2791 			vinactive(vp, curthread);
2792 		if (func != VPUTX_VUNREF)
2793 			VOP_UNLOCK(vp, 0);
2794 	}
2795 	vdropl(vp);
2796 }
2797 
2798 /*
2799  * Vnode put/release.
2800  * If count drops to zero, call inactive routine and return to freelist.
2801  */
2802 void
2803 vrele(struct vnode *vp)
2804 {
2805 
2806 	vputx(vp, VPUTX_VRELE);
2807 }
2808 
2809 /*
2810  * Release an already locked vnode.  This give the same effects as
2811  * unlock+vrele(), but takes less time and avoids releasing and
2812  * re-aquiring the lock (as vrele() acquires the lock internally.)
2813  */
2814 void
2815 vput(struct vnode *vp)
2816 {
2817 
2818 	vputx(vp, VPUTX_VPUT);
2819 }
2820 
2821 /*
2822  * Release an exclusively locked vnode. Do not unlock the vnode lock.
2823  */
2824 void
2825 vunref(struct vnode *vp)
2826 {
2827 
2828 	vputx(vp, VPUTX_VUNREF);
2829 }
2830 
2831 /*
2832  * Increase the hold count and activate if this is the first reference.
2833  */
2834 void
2835 _vhold(struct vnode *vp, bool locked)
2836 {
2837 	struct mount *mp;
2838 
2839 	if (locked)
2840 		ASSERT_VI_LOCKED(vp, __func__);
2841 	else
2842 		ASSERT_VI_UNLOCKED(vp, __func__);
2843 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2844 	if (!locked && vfs_refcount_acquire_if_not_zero(&vp->v_holdcnt)) {
2845 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2846 		    ("_vhold: vnode with holdcnt is free"));
2847 		return;
2848 	}
2849 
2850 	if (!locked)
2851 		VI_LOCK(vp);
2852 	if ((vp->v_iflag & VI_FREE) == 0) {
2853 		refcount_acquire(&vp->v_holdcnt);
2854 		if (!locked)
2855 			VI_UNLOCK(vp);
2856 		return;
2857 	}
2858 	VNASSERT(vp->v_holdcnt == 0, vp,
2859 	    ("%s: wrong hold count", __func__));
2860 	VNASSERT(vp->v_op != NULL, vp,
2861 	    ("%s: vnode already reclaimed.", __func__));
2862 	/*
2863 	 * Remove a vnode from the free list, mark it as in use,
2864 	 * and put it on the active list.
2865 	 */
2866 	VNASSERT(vp->v_mount != NULL, vp,
2867 	    ("_vhold: vnode not on per mount vnode list"));
2868 	mp = vp->v_mount;
2869 	mtx_lock(&mp->mnt_listmtx);
2870 	if ((vp->v_mflag & VMP_TMPMNTFREELIST) != 0) {
2871 		TAILQ_REMOVE(&mp->mnt_tmpfreevnodelist, vp, v_actfreelist);
2872 		mp->mnt_tmpfreevnodelistsize--;
2873 		vp->v_mflag &= ~VMP_TMPMNTFREELIST;
2874 	} else {
2875 		mtx_lock(&vnode_free_list_mtx);
2876 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
2877 		freevnodes--;
2878 		mtx_unlock(&vnode_free_list_mtx);
2879 	}
2880 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
2881 	    ("Activating already active vnode"));
2882 	vp->v_iflag &= ~VI_FREE;
2883 	vp->v_iflag |= VI_ACTIVE;
2884 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
2885 	mp->mnt_activevnodelistsize++;
2886 	mtx_unlock(&mp->mnt_listmtx);
2887 	refcount_acquire(&vp->v_holdcnt);
2888 	if (!locked)
2889 		VI_UNLOCK(vp);
2890 }
2891 
2892 /*
2893  * Drop the hold count of the vnode.  If this is the last reference to
2894  * the vnode we place it on the free list unless it has been vgone'd
2895  * (marked VI_DOOMED) in which case we will free it.
2896  *
2897  * Because the vnode vm object keeps a hold reference on the vnode if
2898  * there is at least one resident non-cached page, the vnode cannot
2899  * leave the active list without the page cleanup done.
2900  */
2901 void
2902 _vdrop(struct vnode *vp, bool locked)
2903 {
2904 	struct bufobj *bo;
2905 	struct mount *mp;
2906 	int active;
2907 
2908 	if (locked)
2909 		ASSERT_VI_LOCKED(vp, __func__);
2910 	else
2911 		ASSERT_VI_UNLOCKED(vp, __func__);
2912 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2913 	if ((int)vp->v_holdcnt <= 0)
2914 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
2915 	if (vfs_refcount_release_if_not_last(&vp->v_holdcnt)) {
2916 		if (locked)
2917 			VI_UNLOCK(vp);
2918 		return;
2919 	}
2920 
2921 	if (!locked)
2922 		VI_LOCK(vp);
2923 	if (refcount_release(&vp->v_holdcnt) == 0) {
2924 		VI_UNLOCK(vp);
2925 		return;
2926 	}
2927 	if ((vp->v_iflag & VI_DOOMED) == 0) {
2928 		/*
2929 		 * Mark a vnode as free: remove it from its active list
2930 		 * and put it up for recycling on the freelist.
2931 		 */
2932 		VNASSERT(vp->v_op != NULL, vp,
2933 		    ("vdropl: vnode already reclaimed."));
2934 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2935 		    ("vnode already free"));
2936 		VNASSERT(vp->v_holdcnt == 0, vp,
2937 		    ("vdropl: freeing when we shouldn't"));
2938 		active = vp->v_iflag & VI_ACTIVE;
2939 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2940 			vp->v_iflag &= ~VI_ACTIVE;
2941 			mp = vp->v_mount;
2942 			if (mp != NULL) {
2943 				mtx_lock(&mp->mnt_listmtx);
2944 				if (active) {
2945 					TAILQ_REMOVE(&mp->mnt_activevnodelist,
2946 					    vp, v_actfreelist);
2947 					mp->mnt_activevnodelistsize--;
2948 				}
2949 				TAILQ_INSERT_TAIL(&mp->mnt_tmpfreevnodelist,
2950 				    vp, v_actfreelist);
2951 				mp->mnt_tmpfreevnodelistsize++;
2952 				vp->v_iflag |= VI_FREE;
2953 				vp->v_mflag |= VMP_TMPMNTFREELIST;
2954 				VI_UNLOCK(vp);
2955 				if (mp->mnt_tmpfreevnodelistsize >=
2956 				    mnt_free_list_batch)
2957 					vnlru_return_batch_locked(mp);
2958 				mtx_unlock(&mp->mnt_listmtx);
2959 			} else {
2960 				VNASSERT(active == 0, vp,
2961 				    ("vdropl: active vnode not on per mount "
2962 				    "vnode list"));
2963 				mtx_lock(&vnode_free_list_mtx);
2964 				TAILQ_INSERT_TAIL(&vnode_free_list, vp,
2965 				    v_actfreelist);
2966 				freevnodes++;
2967 				vp->v_iflag |= VI_FREE;
2968 				VI_UNLOCK(vp);
2969 				mtx_unlock(&vnode_free_list_mtx);
2970 			}
2971 		} else {
2972 			VI_UNLOCK(vp);
2973 			counter_u64_add(free_owe_inact, 1);
2974 		}
2975 		return;
2976 	}
2977 	/*
2978 	 * The vnode has been marked for destruction, so free it.
2979 	 *
2980 	 * The vnode will be returned to the zone where it will
2981 	 * normally remain until it is needed for another vnode. We
2982 	 * need to cleanup (or verify that the cleanup has already
2983 	 * been done) any residual data left from its current use
2984 	 * so as not to contaminate the freshly allocated vnode.
2985 	 */
2986 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
2987 	atomic_subtract_long(&numvnodes, 1);
2988 	bo = &vp->v_bufobj;
2989 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2990 	    ("cleaned vnode still on the free list."));
2991 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
2992 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
2993 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
2994 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
2995 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
2996 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
2997 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
2998 	    ("clean blk trie not empty"));
2999 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
3000 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
3001 	    ("dirty blk trie not empty"));
3002 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
3003 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
3004 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
3005 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
3006 	    ("Dangling rangelock waiters"));
3007 	VI_UNLOCK(vp);
3008 #ifdef MAC
3009 	mac_vnode_destroy(vp);
3010 #endif
3011 	if (vp->v_pollinfo != NULL) {
3012 		destroy_vpollinfo(vp->v_pollinfo);
3013 		vp->v_pollinfo = NULL;
3014 	}
3015 #ifdef INVARIANTS
3016 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
3017 	vp->v_op = NULL;
3018 #endif
3019 	vp->v_mountedhere = NULL;
3020 	vp->v_unpcb = NULL;
3021 	vp->v_rdev = NULL;
3022 	vp->v_fifoinfo = NULL;
3023 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
3024 	vp->v_iflag = 0;
3025 	vp->v_vflag = 0;
3026 	bo->bo_flag = 0;
3027 	uma_zfree(vnode_zone, vp);
3028 }
3029 
3030 /*
3031  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3032  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3033  * OWEINACT tracks whether a vnode missed a call to inactive due to a
3034  * failed lock upgrade.
3035  */
3036 void
3037 vinactive(struct vnode *vp, struct thread *td)
3038 {
3039 	struct vm_object *obj;
3040 
3041 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3042 	ASSERT_VI_LOCKED(vp, "vinactive");
3043 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3044 	    ("vinactive: recursed on VI_DOINGINACT"));
3045 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3046 	vp->v_iflag |= VI_DOINGINACT;
3047 	vp->v_iflag &= ~VI_OWEINACT;
3048 	VI_UNLOCK(vp);
3049 	/*
3050 	 * Before moving off the active list, we must be sure that any
3051 	 * modified pages are converted into the vnode's dirty
3052 	 * buffers, since these will no longer be checked once the
3053 	 * vnode is on the inactive list.
3054 	 *
3055 	 * The write-out of the dirty pages is asynchronous.  At the
3056 	 * point that VOP_INACTIVE() is called, there could still be
3057 	 * pending I/O and dirty pages in the object.
3058 	 */
3059 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3060 	    (obj->flags & OBJ_MIGHTBEDIRTY) != 0) {
3061 		VM_OBJECT_WLOCK(obj);
3062 		vm_object_page_clean(obj, 0, 0, 0);
3063 		VM_OBJECT_WUNLOCK(obj);
3064 	}
3065 	VOP_INACTIVE(vp, td);
3066 	VI_LOCK(vp);
3067 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3068 	    ("vinactive: lost VI_DOINGINACT"));
3069 	vp->v_iflag &= ~VI_DOINGINACT;
3070 }
3071 
3072 /*
3073  * Remove any vnodes in the vnode table belonging to mount point mp.
3074  *
3075  * If FORCECLOSE is not specified, there should not be any active ones,
3076  * return error if any are found (nb: this is a user error, not a
3077  * system error). If FORCECLOSE is specified, detach any active vnodes
3078  * that are found.
3079  *
3080  * If WRITECLOSE is set, only flush out regular file vnodes open for
3081  * writing.
3082  *
3083  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3084  *
3085  * `rootrefs' specifies the base reference count for the root vnode
3086  * of this filesystem. The root vnode is considered busy if its
3087  * v_usecount exceeds this value. On a successful return, vflush(, td)
3088  * will call vrele() on the root vnode exactly rootrefs times.
3089  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3090  * be zero.
3091  */
3092 #ifdef DIAGNOSTIC
3093 static int busyprt = 0;		/* print out busy vnodes */
3094 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3095 #endif
3096 
3097 int
3098 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3099 {
3100 	struct vnode *vp, *mvp, *rootvp = NULL;
3101 	struct vattr vattr;
3102 	int busy = 0, error;
3103 
3104 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3105 	    rootrefs, flags);
3106 	if (rootrefs > 0) {
3107 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3108 		    ("vflush: bad args"));
3109 		/*
3110 		 * Get the filesystem root vnode. We can vput() it
3111 		 * immediately, since with rootrefs > 0, it won't go away.
3112 		 */
3113 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3114 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3115 			    __func__, error);
3116 			return (error);
3117 		}
3118 		vput(rootvp);
3119 	}
3120 loop:
3121 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3122 		vholdl(vp);
3123 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3124 		if (error) {
3125 			vdrop(vp);
3126 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3127 			goto loop;
3128 		}
3129 		/*
3130 		 * Skip over a vnodes marked VV_SYSTEM.
3131 		 */
3132 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3133 			VOP_UNLOCK(vp, 0);
3134 			vdrop(vp);
3135 			continue;
3136 		}
3137 		/*
3138 		 * If WRITECLOSE is set, flush out unlinked but still open
3139 		 * files (even if open only for reading) and regular file
3140 		 * vnodes open for writing.
3141 		 */
3142 		if (flags & WRITECLOSE) {
3143 			if (vp->v_object != NULL) {
3144 				VM_OBJECT_WLOCK(vp->v_object);
3145 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3146 				VM_OBJECT_WUNLOCK(vp->v_object);
3147 			}
3148 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3149 			if (error != 0) {
3150 				VOP_UNLOCK(vp, 0);
3151 				vdrop(vp);
3152 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3153 				return (error);
3154 			}
3155 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3156 			VI_LOCK(vp);
3157 
3158 			if ((vp->v_type == VNON ||
3159 			    (error == 0 && vattr.va_nlink > 0)) &&
3160 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
3161 				VOP_UNLOCK(vp, 0);
3162 				vdropl(vp);
3163 				continue;
3164 			}
3165 		} else
3166 			VI_LOCK(vp);
3167 		/*
3168 		 * With v_usecount == 0, all we need to do is clear out the
3169 		 * vnode data structures and we are done.
3170 		 *
3171 		 * If FORCECLOSE is set, forcibly close the vnode.
3172 		 */
3173 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3174 			vgonel(vp);
3175 		} else {
3176 			busy++;
3177 #ifdef DIAGNOSTIC
3178 			if (busyprt)
3179 				vn_printf(vp, "vflush: busy vnode ");
3180 #endif
3181 		}
3182 		VOP_UNLOCK(vp, 0);
3183 		vdropl(vp);
3184 	}
3185 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3186 		/*
3187 		 * If just the root vnode is busy, and if its refcount
3188 		 * is equal to `rootrefs', then go ahead and kill it.
3189 		 */
3190 		VI_LOCK(rootvp);
3191 		KASSERT(busy > 0, ("vflush: not busy"));
3192 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3193 		    ("vflush: usecount %d < rootrefs %d",
3194 		     rootvp->v_usecount, rootrefs));
3195 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3196 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3197 			vgone(rootvp);
3198 			VOP_UNLOCK(rootvp, 0);
3199 			busy = 0;
3200 		} else
3201 			VI_UNLOCK(rootvp);
3202 	}
3203 	if (busy) {
3204 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3205 		    busy);
3206 		return (EBUSY);
3207 	}
3208 	for (; rootrefs > 0; rootrefs--)
3209 		vrele(rootvp);
3210 	return (0);
3211 }
3212 
3213 /*
3214  * Recycle an unused vnode to the front of the free list.
3215  */
3216 int
3217 vrecycle(struct vnode *vp)
3218 {
3219 	int recycled;
3220 
3221 	VI_LOCK(vp);
3222 	recycled = vrecyclel(vp);
3223 	VI_UNLOCK(vp);
3224 	return (recycled);
3225 }
3226 
3227 /*
3228  * vrecycle, with the vp interlock held.
3229  */
3230 int
3231 vrecyclel(struct vnode *vp)
3232 {
3233 	int recycled;
3234 
3235 	ASSERT_VOP_ELOCKED(vp, __func__);
3236 	ASSERT_VI_LOCKED(vp, __func__);
3237 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3238 	recycled = 0;
3239 	if (vp->v_usecount == 0) {
3240 		recycled = 1;
3241 		vgonel(vp);
3242 	}
3243 	return (recycled);
3244 }
3245 
3246 /*
3247  * Eliminate all activity associated with a vnode
3248  * in preparation for reuse.
3249  */
3250 void
3251 vgone(struct vnode *vp)
3252 {
3253 	VI_LOCK(vp);
3254 	vgonel(vp);
3255 	VI_UNLOCK(vp);
3256 }
3257 
3258 static void
3259 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3260     struct vnode *lowervp __unused)
3261 {
3262 }
3263 
3264 /*
3265  * Notify upper mounts about reclaimed or unlinked vnode.
3266  */
3267 void
3268 vfs_notify_upper(struct vnode *vp, int event)
3269 {
3270 	static struct vfsops vgonel_vfsops = {
3271 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3272 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3273 	};
3274 	struct mount *mp, *ump, *mmp;
3275 
3276 	mp = vp->v_mount;
3277 	if (mp == NULL)
3278 		return;
3279 
3280 	MNT_ILOCK(mp);
3281 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3282 		goto unlock;
3283 	MNT_IUNLOCK(mp);
3284 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3285 	mmp->mnt_op = &vgonel_vfsops;
3286 	mmp->mnt_kern_flag |= MNTK_MARKER;
3287 	MNT_ILOCK(mp);
3288 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3289 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3290 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3291 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3292 			continue;
3293 		}
3294 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3295 		MNT_IUNLOCK(mp);
3296 		switch (event) {
3297 		case VFS_NOTIFY_UPPER_RECLAIM:
3298 			VFS_RECLAIM_LOWERVP(ump, vp);
3299 			break;
3300 		case VFS_NOTIFY_UPPER_UNLINK:
3301 			VFS_UNLINK_LOWERVP(ump, vp);
3302 			break;
3303 		default:
3304 			KASSERT(0, ("invalid event %d", event));
3305 			break;
3306 		}
3307 		MNT_ILOCK(mp);
3308 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3309 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3310 	}
3311 	free(mmp, M_TEMP);
3312 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3313 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3314 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3315 		wakeup(&mp->mnt_uppers);
3316 	}
3317 unlock:
3318 	MNT_IUNLOCK(mp);
3319 }
3320 
3321 /*
3322  * vgone, with the vp interlock held.
3323  */
3324 static void
3325 vgonel(struct vnode *vp)
3326 {
3327 	struct thread *td;
3328 	int oweinact;
3329 	int active;
3330 	struct mount *mp;
3331 
3332 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3333 	ASSERT_VI_LOCKED(vp, "vgonel");
3334 	VNASSERT(vp->v_holdcnt, vp,
3335 	    ("vgonel: vp %p has no reference.", vp));
3336 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3337 	td = curthread;
3338 
3339 	/*
3340 	 * Don't vgonel if we're already doomed.
3341 	 */
3342 	if (vp->v_iflag & VI_DOOMED)
3343 		return;
3344 	vp->v_iflag |= VI_DOOMED;
3345 
3346 	/*
3347 	 * Check to see if the vnode is in use.  If so, we have to call
3348 	 * VOP_CLOSE() and VOP_INACTIVE().
3349 	 */
3350 	active = vp->v_usecount;
3351 	oweinact = (vp->v_iflag & VI_OWEINACT);
3352 	VI_UNLOCK(vp);
3353 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3354 
3355 	/*
3356 	 * If purging an active vnode, it must be closed and
3357 	 * deactivated before being reclaimed.
3358 	 */
3359 	if (active)
3360 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3361 	if (oweinact || active) {
3362 		VI_LOCK(vp);
3363 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
3364 			vinactive(vp, td);
3365 		VI_UNLOCK(vp);
3366 	}
3367 	if (vp->v_type == VSOCK)
3368 		vfs_unp_reclaim(vp);
3369 
3370 	/*
3371 	 * Clean out any buffers associated with the vnode.
3372 	 * If the flush fails, just toss the buffers.
3373 	 */
3374 	mp = NULL;
3375 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3376 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3377 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3378 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3379 			;
3380 	}
3381 
3382 	BO_LOCK(&vp->v_bufobj);
3383 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3384 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3385 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3386 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3387 	    ("vp %p bufobj not invalidated", vp));
3388 
3389 	/*
3390 	 * For VMIO bufobj, BO_DEAD is set in vm_object_terminate()
3391 	 * after the object's page queue is flushed.
3392 	 */
3393 	if (vp->v_bufobj.bo_object == NULL)
3394 		vp->v_bufobj.bo_flag |= BO_DEAD;
3395 	BO_UNLOCK(&vp->v_bufobj);
3396 
3397 	/*
3398 	 * Reclaim the vnode.
3399 	 */
3400 	if (VOP_RECLAIM(vp, td))
3401 		panic("vgone: cannot reclaim");
3402 	if (mp != NULL)
3403 		vn_finished_secondary_write(mp);
3404 	VNASSERT(vp->v_object == NULL, vp,
3405 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
3406 	/*
3407 	 * Clear the advisory locks and wake up waiting threads.
3408 	 */
3409 	(void)VOP_ADVLOCKPURGE(vp);
3410 	vp->v_lockf = NULL;
3411 	/*
3412 	 * Delete from old mount point vnode list.
3413 	 */
3414 	delmntque(vp);
3415 	cache_purge(vp);
3416 	/*
3417 	 * Done with purge, reset to the standard lock and invalidate
3418 	 * the vnode.
3419 	 */
3420 	VI_LOCK(vp);
3421 	vp->v_vnlock = &vp->v_lock;
3422 	vp->v_op = &dead_vnodeops;
3423 	vp->v_tag = "none";
3424 	vp->v_type = VBAD;
3425 }
3426 
3427 /*
3428  * Calculate the total number of references to a special device.
3429  */
3430 int
3431 vcount(struct vnode *vp)
3432 {
3433 	int count;
3434 
3435 	dev_lock();
3436 	count = vp->v_rdev->si_usecount;
3437 	dev_unlock();
3438 	return (count);
3439 }
3440 
3441 /*
3442  * Same as above, but using the struct cdev *as argument
3443  */
3444 int
3445 count_dev(struct cdev *dev)
3446 {
3447 	int count;
3448 
3449 	dev_lock();
3450 	count = dev->si_usecount;
3451 	dev_unlock();
3452 	return(count);
3453 }
3454 
3455 /*
3456  * Print out a description of a vnode.
3457  */
3458 static char *typename[] =
3459 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3460  "VMARKER"};
3461 
3462 void
3463 vn_printf(struct vnode *vp, const char *fmt, ...)
3464 {
3465 	va_list ap;
3466 	char buf[256], buf2[16];
3467 	u_long flags;
3468 
3469 	va_start(ap, fmt);
3470 	vprintf(fmt, ap);
3471 	va_end(ap);
3472 	printf("%p: ", (void *)vp);
3473 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
3474 	printf("    usecount %d, writecount %d, refcount %d mountedhere %p\n",
3475 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere);
3476 	buf[0] = '\0';
3477 	buf[1] = '\0';
3478 	if (vp->v_vflag & VV_ROOT)
3479 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3480 	if (vp->v_vflag & VV_ISTTY)
3481 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3482 	if (vp->v_vflag & VV_NOSYNC)
3483 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3484 	if (vp->v_vflag & VV_ETERNALDEV)
3485 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3486 	if (vp->v_vflag & VV_CACHEDLABEL)
3487 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3488 	if (vp->v_vflag & VV_TEXT)
3489 		strlcat(buf, "|VV_TEXT", sizeof(buf));
3490 	if (vp->v_vflag & VV_COPYONWRITE)
3491 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3492 	if (vp->v_vflag & VV_SYSTEM)
3493 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3494 	if (vp->v_vflag & VV_PROCDEP)
3495 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3496 	if (vp->v_vflag & VV_NOKNOTE)
3497 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3498 	if (vp->v_vflag & VV_DELETED)
3499 		strlcat(buf, "|VV_DELETED", sizeof(buf));
3500 	if (vp->v_vflag & VV_MD)
3501 		strlcat(buf, "|VV_MD", sizeof(buf));
3502 	if (vp->v_vflag & VV_FORCEINSMQ)
3503 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
3504 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
3505 	    VV_CACHEDLABEL | VV_TEXT | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
3506 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
3507 	if (flags != 0) {
3508 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
3509 		strlcat(buf, buf2, sizeof(buf));
3510 	}
3511 	if (vp->v_iflag & VI_MOUNT)
3512 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
3513 	if (vp->v_iflag & VI_DOOMED)
3514 		strlcat(buf, "|VI_DOOMED", sizeof(buf));
3515 	if (vp->v_iflag & VI_FREE)
3516 		strlcat(buf, "|VI_FREE", sizeof(buf));
3517 	if (vp->v_iflag & VI_ACTIVE)
3518 		strlcat(buf, "|VI_ACTIVE", sizeof(buf));
3519 	if (vp->v_iflag & VI_DOINGINACT)
3520 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
3521 	if (vp->v_iflag & VI_OWEINACT)
3522 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
3523 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOOMED | VI_FREE |
3524 	    VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT);
3525 	if (flags != 0) {
3526 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
3527 		strlcat(buf, buf2, sizeof(buf));
3528 	}
3529 	printf("    flags (%s)\n", buf + 1);
3530 	if (mtx_owned(VI_MTX(vp)))
3531 		printf(" VI_LOCKed");
3532 	if (vp->v_object != NULL)
3533 		printf("    v_object %p ref %d pages %d "
3534 		    "cleanbuf %d dirtybuf %d\n",
3535 		    vp->v_object, vp->v_object->ref_count,
3536 		    vp->v_object->resident_page_count,
3537 		    vp->v_bufobj.bo_clean.bv_cnt,
3538 		    vp->v_bufobj.bo_dirty.bv_cnt);
3539 	printf("    ");
3540 	lockmgr_printinfo(vp->v_vnlock);
3541 	if (vp->v_data != NULL)
3542 		VOP_PRINT(vp);
3543 }
3544 
3545 #ifdef DDB
3546 /*
3547  * List all of the locked vnodes in the system.
3548  * Called when debugging the kernel.
3549  */
3550 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
3551 {
3552 	struct mount *mp;
3553 	struct vnode *vp;
3554 
3555 	/*
3556 	 * Note: because this is DDB, we can't obey the locking semantics
3557 	 * for these structures, which means we could catch an inconsistent
3558 	 * state and dereference a nasty pointer.  Not much to be done
3559 	 * about that.
3560 	 */
3561 	db_printf("Locked vnodes\n");
3562 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3563 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3564 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
3565 				vn_printf(vp, "vnode ");
3566 		}
3567 	}
3568 }
3569 
3570 /*
3571  * Show details about the given vnode.
3572  */
3573 DB_SHOW_COMMAND(vnode, db_show_vnode)
3574 {
3575 	struct vnode *vp;
3576 
3577 	if (!have_addr)
3578 		return;
3579 	vp = (struct vnode *)addr;
3580 	vn_printf(vp, "vnode ");
3581 }
3582 
3583 /*
3584  * Show details about the given mount point.
3585  */
3586 DB_SHOW_COMMAND(mount, db_show_mount)
3587 {
3588 	struct mount *mp;
3589 	struct vfsopt *opt;
3590 	struct statfs *sp;
3591 	struct vnode *vp;
3592 	char buf[512];
3593 	uint64_t mflags;
3594 	u_int flags;
3595 
3596 	if (!have_addr) {
3597 		/* No address given, print short info about all mount points. */
3598 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3599 			db_printf("%p %s on %s (%s)\n", mp,
3600 			    mp->mnt_stat.f_mntfromname,
3601 			    mp->mnt_stat.f_mntonname,
3602 			    mp->mnt_stat.f_fstypename);
3603 			if (db_pager_quit)
3604 				break;
3605 		}
3606 		db_printf("\nMore info: show mount <addr>\n");
3607 		return;
3608 	}
3609 
3610 	mp = (struct mount *)addr;
3611 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
3612 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
3613 
3614 	buf[0] = '\0';
3615 	mflags = mp->mnt_flag;
3616 #define	MNT_FLAG(flag)	do {						\
3617 	if (mflags & (flag)) {						\
3618 		if (buf[0] != '\0')					\
3619 			strlcat(buf, ", ", sizeof(buf));		\
3620 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
3621 		mflags &= ~(flag);					\
3622 	}								\
3623 } while (0)
3624 	MNT_FLAG(MNT_RDONLY);
3625 	MNT_FLAG(MNT_SYNCHRONOUS);
3626 	MNT_FLAG(MNT_NOEXEC);
3627 	MNT_FLAG(MNT_NOSUID);
3628 	MNT_FLAG(MNT_NFS4ACLS);
3629 	MNT_FLAG(MNT_UNION);
3630 	MNT_FLAG(MNT_ASYNC);
3631 	MNT_FLAG(MNT_SUIDDIR);
3632 	MNT_FLAG(MNT_SOFTDEP);
3633 	MNT_FLAG(MNT_NOSYMFOLLOW);
3634 	MNT_FLAG(MNT_GJOURNAL);
3635 	MNT_FLAG(MNT_MULTILABEL);
3636 	MNT_FLAG(MNT_ACLS);
3637 	MNT_FLAG(MNT_NOATIME);
3638 	MNT_FLAG(MNT_NOCLUSTERR);
3639 	MNT_FLAG(MNT_NOCLUSTERW);
3640 	MNT_FLAG(MNT_SUJ);
3641 	MNT_FLAG(MNT_EXRDONLY);
3642 	MNT_FLAG(MNT_EXPORTED);
3643 	MNT_FLAG(MNT_DEFEXPORTED);
3644 	MNT_FLAG(MNT_EXPORTANON);
3645 	MNT_FLAG(MNT_EXKERB);
3646 	MNT_FLAG(MNT_EXPUBLIC);
3647 	MNT_FLAG(MNT_LOCAL);
3648 	MNT_FLAG(MNT_QUOTA);
3649 	MNT_FLAG(MNT_ROOTFS);
3650 	MNT_FLAG(MNT_USER);
3651 	MNT_FLAG(MNT_IGNORE);
3652 	MNT_FLAG(MNT_UPDATE);
3653 	MNT_FLAG(MNT_DELEXPORT);
3654 	MNT_FLAG(MNT_RELOAD);
3655 	MNT_FLAG(MNT_FORCE);
3656 	MNT_FLAG(MNT_SNAPSHOT);
3657 	MNT_FLAG(MNT_BYFSID);
3658 #undef MNT_FLAG
3659 	if (mflags != 0) {
3660 		if (buf[0] != '\0')
3661 			strlcat(buf, ", ", sizeof(buf));
3662 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3663 		    "0x%016jx", mflags);
3664 	}
3665 	db_printf("    mnt_flag = %s\n", buf);
3666 
3667 	buf[0] = '\0';
3668 	flags = mp->mnt_kern_flag;
3669 #define	MNT_KERN_FLAG(flag)	do {					\
3670 	if (flags & (flag)) {						\
3671 		if (buf[0] != '\0')					\
3672 			strlcat(buf, ", ", sizeof(buf));		\
3673 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
3674 		flags &= ~(flag);					\
3675 	}								\
3676 } while (0)
3677 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
3678 	MNT_KERN_FLAG(MNTK_ASYNC);
3679 	MNT_KERN_FLAG(MNTK_SOFTDEP);
3680 	MNT_KERN_FLAG(MNTK_NOINSMNTQ);
3681 	MNT_KERN_FLAG(MNTK_DRAINING);
3682 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
3683 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
3684 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
3685 	MNT_KERN_FLAG(MNTK_NO_IOPF);
3686 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
3687 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
3688 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
3689 	MNT_KERN_FLAG(MNTK_MARKER);
3690 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
3691 	MNT_KERN_FLAG(MNTK_NOASYNC);
3692 	MNT_KERN_FLAG(MNTK_UNMOUNT);
3693 	MNT_KERN_FLAG(MNTK_MWAIT);
3694 	MNT_KERN_FLAG(MNTK_SUSPEND);
3695 	MNT_KERN_FLAG(MNTK_SUSPEND2);
3696 	MNT_KERN_FLAG(MNTK_SUSPENDED);
3697 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
3698 	MNT_KERN_FLAG(MNTK_NOKNOTE);
3699 #undef MNT_KERN_FLAG
3700 	if (flags != 0) {
3701 		if (buf[0] != '\0')
3702 			strlcat(buf, ", ", sizeof(buf));
3703 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3704 		    "0x%08x", flags);
3705 	}
3706 	db_printf("    mnt_kern_flag = %s\n", buf);
3707 
3708 	db_printf("    mnt_opt = ");
3709 	opt = TAILQ_FIRST(mp->mnt_opt);
3710 	if (opt != NULL) {
3711 		db_printf("%s", opt->name);
3712 		opt = TAILQ_NEXT(opt, link);
3713 		while (opt != NULL) {
3714 			db_printf(", %s", opt->name);
3715 			opt = TAILQ_NEXT(opt, link);
3716 		}
3717 	}
3718 	db_printf("\n");
3719 
3720 	sp = &mp->mnt_stat;
3721 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
3722 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
3723 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
3724 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
3725 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
3726 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
3727 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
3728 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
3729 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
3730 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
3731 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
3732 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
3733 
3734 	db_printf("    mnt_cred = { uid=%u ruid=%u",
3735 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
3736 	if (jailed(mp->mnt_cred))
3737 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
3738 	db_printf(" }\n");
3739 	db_printf("    mnt_ref = %d\n", mp->mnt_ref);
3740 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
3741 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
3742 	db_printf("    mnt_activevnodelistsize = %d\n",
3743 	    mp->mnt_activevnodelistsize);
3744 	db_printf("    mnt_writeopcount = %d\n", mp->mnt_writeopcount);
3745 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
3746 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
3747 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
3748 	db_printf("    mnt_lockref = %d\n", mp->mnt_lockref);
3749 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
3750 	db_printf("    mnt_secondary_accwrites = %d\n",
3751 	    mp->mnt_secondary_accwrites);
3752 	db_printf("    mnt_gjprovider = %s\n",
3753 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
3754 
3755 	db_printf("\n\nList of active vnodes\n");
3756 	TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) {
3757 		if (vp->v_type != VMARKER) {
3758 			vn_printf(vp, "vnode ");
3759 			if (db_pager_quit)
3760 				break;
3761 		}
3762 	}
3763 	db_printf("\n\nList of inactive vnodes\n");
3764 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3765 		if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) {
3766 			vn_printf(vp, "vnode ");
3767 			if (db_pager_quit)
3768 				break;
3769 		}
3770 	}
3771 }
3772 #endif	/* DDB */
3773 
3774 /*
3775  * Fill in a struct xvfsconf based on a struct vfsconf.
3776  */
3777 static int
3778 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
3779 {
3780 	struct xvfsconf xvfsp;
3781 
3782 	bzero(&xvfsp, sizeof(xvfsp));
3783 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3784 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3785 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3786 	xvfsp.vfc_flags = vfsp->vfc_flags;
3787 	/*
3788 	 * These are unused in userland, we keep them
3789 	 * to not break binary compatibility.
3790 	 */
3791 	xvfsp.vfc_vfsops = NULL;
3792 	xvfsp.vfc_next = NULL;
3793 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3794 }
3795 
3796 #ifdef COMPAT_FREEBSD32
3797 struct xvfsconf32 {
3798 	uint32_t	vfc_vfsops;
3799 	char		vfc_name[MFSNAMELEN];
3800 	int32_t		vfc_typenum;
3801 	int32_t		vfc_refcount;
3802 	int32_t		vfc_flags;
3803 	uint32_t	vfc_next;
3804 };
3805 
3806 static int
3807 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
3808 {
3809 	struct xvfsconf32 xvfsp;
3810 
3811 	bzero(&xvfsp, sizeof(xvfsp));
3812 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3813 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3814 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3815 	xvfsp.vfc_flags = vfsp->vfc_flags;
3816 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3817 }
3818 #endif
3819 
3820 /*
3821  * Top level filesystem related information gathering.
3822  */
3823 static int
3824 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
3825 {
3826 	struct vfsconf *vfsp;
3827 	int error;
3828 
3829 	error = 0;
3830 	vfsconf_slock();
3831 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3832 #ifdef COMPAT_FREEBSD32
3833 		if (req->flags & SCTL_MASK32)
3834 			error = vfsconf2x32(req, vfsp);
3835 		else
3836 #endif
3837 			error = vfsconf2x(req, vfsp);
3838 		if (error)
3839 			break;
3840 	}
3841 	vfsconf_sunlock();
3842 	return (error);
3843 }
3844 
3845 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
3846     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
3847     "S,xvfsconf", "List of all configured filesystems");
3848 
3849 #ifndef BURN_BRIDGES
3850 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
3851 
3852 static int
3853 vfs_sysctl(SYSCTL_HANDLER_ARGS)
3854 {
3855 	int *name = (int *)arg1 - 1;	/* XXX */
3856 	u_int namelen = arg2 + 1;	/* XXX */
3857 	struct vfsconf *vfsp;
3858 
3859 	log(LOG_WARNING, "userland calling deprecated sysctl, "
3860 	    "please rebuild world\n");
3861 
3862 #if 1 || defined(COMPAT_PRELITE2)
3863 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
3864 	if (namelen == 1)
3865 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
3866 #endif
3867 
3868 	switch (name[1]) {
3869 	case VFS_MAXTYPENUM:
3870 		if (namelen != 2)
3871 			return (ENOTDIR);
3872 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
3873 	case VFS_CONF:
3874 		if (namelen != 3)
3875 			return (ENOTDIR);	/* overloaded */
3876 		vfsconf_slock();
3877 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3878 			if (vfsp->vfc_typenum == name[2])
3879 				break;
3880 		}
3881 		vfsconf_sunlock();
3882 		if (vfsp == NULL)
3883 			return (EOPNOTSUPP);
3884 #ifdef COMPAT_FREEBSD32
3885 		if (req->flags & SCTL_MASK32)
3886 			return (vfsconf2x32(req, vfsp));
3887 		else
3888 #endif
3889 			return (vfsconf2x(req, vfsp));
3890 	}
3891 	return (EOPNOTSUPP);
3892 }
3893 
3894 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
3895     CTLFLAG_MPSAFE, vfs_sysctl,
3896     "Generic filesystem");
3897 
3898 #if 1 || defined(COMPAT_PRELITE2)
3899 
3900 static int
3901 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
3902 {
3903 	int error;
3904 	struct vfsconf *vfsp;
3905 	struct ovfsconf ovfs;
3906 
3907 	vfsconf_slock();
3908 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3909 		bzero(&ovfs, sizeof(ovfs));
3910 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
3911 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
3912 		ovfs.vfc_index = vfsp->vfc_typenum;
3913 		ovfs.vfc_refcount = vfsp->vfc_refcount;
3914 		ovfs.vfc_flags = vfsp->vfc_flags;
3915 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
3916 		if (error != 0) {
3917 			vfsconf_sunlock();
3918 			return (error);
3919 		}
3920 	}
3921 	vfsconf_sunlock();
3922 	return (0);
3923 }
3924 
3925 #endif /* 1 || COMPAT_PRELITE2 */
3926 #endif /* !BURN_BRIDGES */
3927 
3928 #define KINFO_VNODESLOP		10
3929 #ifdef notyet
3930 /*
3931  * Dump vnode list (via sysctl).
3932  */
3933 /* ARGSUSED */
3934 static int
3935 sysctl_vnode(SYSCTL_HANDLER_ARGS)
3936 {
3937 	struct xvnode *xvn;
3938 	struct mount *mp;
3939 	struct vnode *vp;
3940 	int error, len, n;
3941 
3942 	/*
3943 	 * Stale numvnodes access is not fatal here.
3944 	 */
3945 	req->lock = 0;
3946 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
3947 	if (!req->oldptr)
3948 		/* Make an estimate */
3949 		return (SYSCTL_OUT(req, 0, len));
3950 
3951 	error = sysctl_wire_old_buffer(req, 0);
3952 	if (error != 0)
3953 		return (error);
3954 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
3955 	n = 0;
3956 	mtx_lock(&mountlist_mtx);
3957 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3958 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
3959 			continue;
3960 		MNT_ILOCK(mp);
3961 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3962 			if (n == len)
3963 				break;
3964 			vref(vp);
3965 			xvn[n].xv_size = sizeof *xvn;
3966 			xvn[n].xv_vnode = vp;
3967 			xvn[n].xv_id = 0;	/* XXX compat */
3968 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
3969 			XV_COPY(usecount);
3970 			XV_COPY(writecount);
3971 			XV_COPY(holdcnt);
3972 			XV_COPY(mount);
3973 			XV_COPY(numoutput);
3974 			XV_COPY(type);
3975 #undef XV_COPY
3976 			xvn[n].xv_flag = vp->v_vflag;
3977 
3978 			switch (vp->v_type) {
3979 			case VREG:
3980 			case VDIR:
3981 			case VLNK:
3982 				break;
3983 			case VBLK:
3984 			case VCHR:
3985 				if (vp->v_rdev == NULL) {
3986 					vrele(vp);
3987 					continue;
3988 				}
3989 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
3990 				break;
3991 			case VSOCK:
3992 				xvn[n].xv_socket = vp->v_socket;
3993 				break;
3994 			case VFIFO:
3995 				xvn[n].xv_fifo = vp->v_fifoinfo;
3996 				break;
3997 			case VNON:
3998 			case VBAD:
3999 			default:
4000 				/* shouldn't happen? */
4001 				vrele(vp);
4002 				continue;
4003 			}
4004 			vrele(vp);
4005 			++n;
4006 		}
4007 		MNT_IUNLOCK(mp);
4008 		mtx_lock(&mountlist_mtx);
4009 		vfs_unbusy(mp);
4010 		if (n == len)
4011 			break;
4012 	}
4013 	mtx_unlock(&mountlist_mtx);
4014 
4015 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4016 	free(xvn, M_TEMP);
4017 	return (error);
4018 }
4019 
4020 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4021     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4022     "");
4023 #endif
4024 
4025 static void
4026 unmount_or_warn(struct mount *mp)
4027 {
4028 	int error;
4029 
4030 	error = dounmount(mp, MNT_FORCE, curthread);
4031 	if (error != 0) {
4032 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4033 		if (error == EBUSY)
4034 			printf("BUSY)\n");
4035 		else
4036 			printf("%d)\n", error);
4037 	}
4038 }
4039 
4040 /*
4041  * Unmount all filesystems. The list is traversed in reverse order
4042  * of mounting to avoid dependencies.
4043  */
4044 void
4045 vfs_unmountall(void)
4046 {
4047 	struct mount *mp, *tmp;
4048 
4049 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4050 
4051 	/*
4052 	 * Since this only runs when rebooting, it is not interlocked.
4053 	 */
4054 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4055 		vfs_ref(mp);
4056 
4057 		/*
4058 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4059 		 * unmount of the latter.
4060 		 */
4061 		if (mp == rootdevmp)
4062 			continue;
4063 
4064 		unmount_or_warn(mp);
4065 	}
4066 
4067 	if (rootdevmp != NULL)
4068 		unmount_or_warn(rootdevmp);
4069 }
4070 
4071 /*
4072  * perform msync on all vnodes under a mount point
4073  * the mount point must be locked.
4074  */
4075 void
4076 vfs_msync(struct mount *mp, int flags)
4077 {
4078 	struct vnode *vp, *mvp;
4079 	struct vm_object *obj;
4080 
4081 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4082 
4083 	vnlru_return_batch(mp);
4084 
4085 	MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
4086 		obj = vp->v_object;
4087 		if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 &&
4088 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
4089 			if (!vget(vp,
4090 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
4091 			    curthread)) {
4092 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
4093 					vput(vp);
4094 					continue;
4095 				}
4096 
4097 				obj = vp->v_object;
4098 				if (obj != NULL) {
4099 					VM_OBJECT_WLOCK(obj);
4100 					vm_object_page_clean(obj, 0, 0,
4101 					    flags == MNT_WAIT ?
4102 					    OBJPC_SYNC : OBJPC_NOSYNC);
4103 					VM_OBJECT_WUNLOCK(obj);
4104 				}
4105 				vput(vp);
4106 			}
4107 		} else
4108 			VI_UNLOCK(vp);
4109 	}
4110 }
4111 
4112 static void
4113 destroy_vpollinfo_free(struct vpollinfo *vi)
4114 {
4115 
4116 	knlist_destroy(&vi->vpi_selinfo.si_note);
4117 	mtx_destroy(&vi->vpi_lock);
4118 	uma_zfree(vnodepoll_zone, vi);
4119 }
4120 
4121 static void
4122 destroy_vpollinfo(struct vpollinfo *vi)
4123 {
4124 
4125 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4126 	seldrain(&vi->vpi_selinfo);
4127 	destroy_vpollinfo_free(vi);
4128 }
4129 
4130 /*
4131  * Initialize per-vnode helper structure to hold poll-related state.
4132  */
4133 void
4134 v_addpollinfo(struct vnode *vp)
4135 {
4136 	struct vpollinfo *vi;
4137 
4138 	if (vp->v_pollinfo != NULL)
4139 		return;
4140 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4141 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4142 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4143 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4144 	VI_LOCK(vp);
4145 	if (vp->v_pollinfo != NULL) {
4146 		VI_UNLOCK(vp);
4147 		destroy_vpollinfo_free(vi);
4148 		return;
4149 	}
4150 	vp->v_pollinfo = vi;
4151 	VI_UNLOCK(vp);
4152 }
4153 
4154 /*
4155  * Record a process's interest in events which might happen to
4156  * a vnode.  Because poll uses the historic select-style interface
4157  * internally, this routine serves as both the ``check for any
4158  * pending events'' and the ``record my interest in future events''
4159  * functions.  (These are done together, while the lock is held,
4160  * to avoid race conditions.)
4161  */
4162 int
4163 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4164 {
4165 
4166 	v_addpollinfo(vp);
4167 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4168 	if (vp->v_pollinfo->vpi_revents & events) {
4169 		/*
4170 		 * This leaves events we are not interested
4171 		 * in available for the other process which
4172 		 * which presumably had requested them
4173 		 * (otherwise they would never have been
4174 		 * recorded).
4175 		 */
4176 		events &= vp->v_pollinfo->vpi_revents;
4177 		vp->v_pollinfo->vpi_revents &= ~events;
4178 
4179 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4180 		return (events);
4181 	}
4182 	vp->v_pollinfo->vpi_events |= events;
4183 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4184 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4185 	return (0);
4186 }
4187 
4188 /*
4189  * Routine to create and manage a filesystem syncer vnode.
4190  */
4191 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4192 static int	sync_fsync(struct  vop_fsync_args *);
4193 static int	sync_inactive(struct  vop_inactive_args *);
4194 static int	sync_reclaim(struct  vop_reclaim_args *);
4195 
4196 static struct vop_vector sync_vnodeops = {
4197 	.vop_bypass =	VOP_EOPNOTSUPP,
4198 	.vop_close =	sync_close,		/* close */
4199 	.vop_fsync =	sync_fsync,		/* fsync */
4200 	.vop_inactive =	sync_inactive,	/* inactive */
4201 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4202 	.vop_lock1 =	vop_stdlock,	/* lock */
4203 	.vop_unlock =	vop_stdunlock,	/* unlock */
4204 	.vop_islocked =	vop_stdislocked,	/* islocked */
4205 };
4206 
4207 /*
4208  * Create a new filesystem syncer vnode for the specified mount point.
4209  */
4210 void
4211 vfs_allocate_syncvnode(struct mount *mp)
4212 {
4213 	struct vnode *vp;
4214 	struct bufobj *bo;
4215 	static long start, incr, next;
4216 	int error;
4217 
4218 	/* Allocate a new vnode */
4219 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4220 	if (error != 0)
4221 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4222 	vp->v_type = VNON;
4223 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4224 	vp->v_vflag |= VV_FORCEINSMQ;
4225 	error = insmntque(vp, mp);
4226 	if (error != 0)
4227 		panic("vfs_allocate_syncvnode: insmntque() failed");
4228 	vp->v_vflag &= ~VV_FORCEINSMQ;
4229 	VOP_UNLOCK(vp, 0);
4230 	/*
4231 	 * Place the vnode onto the syncer worklist. We attempt to
4232 	 * scatter them about on the list so that they will go off
4233 	 * at evenly distributed times even if all the filesystems
4234 	 * are mounted at once.
4235 	 */
4236 	next += incr;
4237 	if (next == 0 || next > syncer_maxdelay) {
4238 		start /= 2;
4239 		incr /= 2;
4240 		if (start == 0) {
4241 			start = syncer_maxdelay / 2;
4242 			incr = syncer_maxdelay;
4243 		}
4244 		next = start;
4245 	}
4246 	bo = &vp->v_bufobj;
4247 	BO_LOCK(bo);
4248 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4249 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4250 	mtx_lock(&sync_mtx);
4251 	sync_vnode_count++;
4252 	if (mp->mnt_syncer == NULL) {
4253 		mp->mnt_syncer = vp;
4254 		vp = NULL;
4255 	}
4256 	mtx_unlock(&sync_mtx);
4257 	BO_UNLOCK(bo);
4258 	if (vp != NULL) {
4259 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4260 		vgone(vp);
4261 		vput(vp);
4262 	}
4263 }
4264 
4265 void
4266 vfs_deallocate_syncvnode(struct mount *mp)
4267 {
4268 	struct vnode *vp;
4269 
4270 	mtx_lock(&sync_mtx);
4271 	vp = mp->mnt_syncer;
4272 	if (vp != NULL)
4273 		mp->mnt_syncer = NULL;
4274 	mtx_unlock(&sync_mtx);
4275 	if (vp != NULL)
4276 		vrele(vp);
4277 }
4278 
4279 /*
4280  * Do a lazy sync of the filesystem.
4281  */
4282 static int
4283 sync_fsync(struct vop_fsync_args *ap)
4284 {
4285 	struct vnode *syncvp = ap->a_vp;
4286 	struct mount *mp = syncvp->v_mount;
4287 	int error, save;
4288 	struct bufobj *bo;
4289 
4290 	/*
4291 	 * We only need to do something if this is a lazy evaluation.
4292 	 */
4293 	if (ap->a_waitfor != MNT_LAZY)
4294 		return (0);
4295 
4296 	/*
4297 	 * Move ourselves to the back of the sync list.
4298 	 */
4299 	bo = &syncvp->v_bufobj;
4300 	BO_LOCK(bo);
4301 	vn_syncer_add_to_worklist(bo, syncdelay);
4302 	BO_UNLOCK(bo);
4303 
4304 	/*
4305 	 * Walk the list of vnodes pushing all that are dirty and
4306 	 * not already on the sync list.
4307 	 */
4308 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4309 		return (0);
4310 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4311 		vfs_unbusy(mp);
4312 		return (0);
4313 	}
4314 	save = curthread_pflags_set(TDP_SYNCIO);
4315 	vfs_msync(mp, MNT_NOWAIT);
4316 	error = VFS_SYNC(mp, MNT_LAZY);
4317 	curthread_pflags_restore(save);
4318 	vn_finished_write(mp);
4319 	vfs_unbusy(mp);
4320 	return (error);
4321 }
4322 
4323 /*
4324  * The syncer vnode is no referenced.
4325  */
4326 static int
4327 sync_inactive(struct vop_inactive_args *ap)
4328 {
4329 
4330 	vgone(ap->a_vp);
4331 	return (0);
4332 }
4333 
4334 /*
4335  * The syncer vnode is no longer needed and is being decommissioned.
4336  *
4337  * Modifications to the worklist must be protected by sync_mtx.
4338  */
4339 static int
4340 sync_reclaim(struct vop_reclaim_args *ap)
4341 {
4342 	struct vnode *vp = ap->a_vp;
4343 	struct bufobj *bo;
4344 
4345 	bo = &vp->v_bufobj;
4346 	BO_LOCK(bo);
4347 	mtx_lock(&sync_mtx);
4348 	if (vp->v_mount->mnt_syncer == vp)
4349 		vp->v_mount->mnt_syncer = NULL;
4350 	if (bo->bo_flag & BO_ONWORKLST) {
4351 		LIST_REMOVE(bo, bo_synclist);
4352 		syncer_worklist_len--;
4353 		sync_vnode_count--;
4354 		bo->bo_flag &= ~BO_ONWORKLST;
4355 	}
4356 	mtx_unlock(&sync_mtx);
4357 	BO_UNLOCK(bo);
4358 
4359 	return (0);
4360 }
4361 
4362 /*
4363  * Check if vnode represents a disk device
4364  */
4365 int
4366 vn_isdisk(struct vnode *vp, int *errp)
4367 {
4368 	int error;
4369 
4370 	if (vp->v_type != VCHR) {
4371 		error = ENOTBLK;
4372 		goto out;
4373 	}
4374 	error = 0;
4375 	dev_lock();
4376 	if (vp->v_rdev == NULL)
4377 		error = ENXIO;
4378 	else if (vp->v_rdev->si_devsw == NULL)
4379 		error = ENXIO;
4380 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
4381 		error = ENOTBLK;
4382 	dev_unlock();
4383 out:
4384 	if (errp != NULL)
4385 		*errp = error;
4386 	return (error == 0);
4387 }
4388 
4389 /*
4390  * Common filesystem object access control check routine.  Accepts a
4391  * vnode's type, "mode", uid and gid, requested access mode, credentials,
4392  * and optional call-by-reference privused argument allowing vaccess()
4393  * to indicate to the caller whether privilege was used to satisfy the
4394  * request (obsoleted).  Returns 0 on success, or an errno on failure.
4395  */
4396 int
4397 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
4398     accmode_t accmode, struct ucred *cred, int *privused)
4399 {
4400 	accmode_t dac_granted;
4401 	accmode_t priv_granted;
4402 
4403 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
4404 	    ("invalid bit in accmode"));
4405 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
4406 	    ("VAPPEND without VWRITE"));
4407 
4408 	/*
4409 	 * Look for a normal, non-privileged way to access the file/directory
4410 	 * as requested.  If it exists, go with that.
4411 	 */
4412 
4413 	if (privused != NULL)
4414 		*privused = 0;
4415 
4416 	dac_granted = 0;
4417 
4418 	/* Check the owner. */
4419 	if (cred->cr_uid == file_uid) {
4420 		dac_granted |= VADMIN;
4421 		if (file_mode & S_IXUSR)
4422 			dac_granted |= VEXEC;
4423 		if (file_mode & S_IRUSR)
4424 			dac_granted |= VREAD;
4425 		if (file_mode & S_IWUSR)
4426 			dac_granted |= (VWRITE | VAPPEND);
4427 
4428 		if ((accmode & dac_granted) == accmode)
4429 			return (0);
4430 
4431 		goto privcheck;
4432 	}
4433 
4434 	/* Otherwise, check the groups (first match) */
4435 	if (groupmember(file_gid, cred)) {
4436 		if (file_mode & S_IXGRP)
4437 			dac_granted |= VEXEC;
4438 		if (file_mode & S_IRGRP)
4439 			dac_granted |= VREAD;
4440 		if (file_mode & S_IWGRP)
4441 			dac_granted |= (VWRITE | VAPPEND);
4442 
4443 		if ((accmode & dac_granted) == accmode)
4444 			return (0);
4445 
4446 		goto privcheck;
4447 	}
4448 
4449 	/* Otherwise, check everyone else. */
4450 	if (file_mode & S_IXOTH)
4451 		dac_granted |= VEXEC;
4452 	if (file_mode & S_IROTH)
4453 		dac_granted |= VREAD;
4454 	if (file_mode & S_IWOTH)
4455 		dac_granted |= (VWRITE | VAPPEND);
4456 	if ((accmode & dac_granted) == accmode)
4457 		return (0);
4458 
4459 privcheck:
4460 	/*
4461 	 * Build a privilege mask to determine if the set of privileges
4462 	 * satisfies the requirements when combined with the granted mask
4463 	 * from above.  For each privilege, if the privilege is required,
4464 	 * bitwise or the request type onto the priv_granted mask.
4465 	 */
4466 	priv_granted = 0;
4467 
4468 	if (type == VDIR) {
4469 		/*
4470 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
4471 		 * requests, instead of PRIV_VFS_EXEC.
4472 		 */
4473 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4474 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP, 0))
4475 			priv_granted |= VEXEC;
4476 	} else {
4477 		/*
4478 		 * Ensure that at least one execute bit is on. Otherwise,
4479 		 * a privileged user will always succeed, and we don't want
4480 		 * this to happen unless the file really is executable.
4481 		 */
4482 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4483 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
4484 		    !priv_check_cred(cred, PRIV_VFS_EXEC, 0))
4485 			priv_granted |= VEXEC;
4486 	}
4487 
4488 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
4489 	    !priv_check_cred(cred, PRIV_VFS_READ, 0))
4490 		priv_granted |= VREAD;
4491 
4492 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
4493 	    !priv_check_cred(cred, PRIV_VFS_WRITE, 0))
4494 		priv_granted |= (VWRITE | VAPPEND);
4495 
4496 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
4497 	    !priv_check_cred(cred, PRIV_VFS_ADMIN, 0))
4498 		priv_granted |= VADMIN;
4499 
4500 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
4501 		/* XXX audit: privilege used */
4502 		if (privused != NULL)
4503 			*privused = 1;
4504 		return (0);
4505 	}
4506 
4507 	return ((accmode & VADMIN) ? EPERM : EACCES);
4508 }
4509 
4510 /*
4511  * Credential check based on process requesting service, and per-attribute
4512  * permissions.
4513  */
4514 int
4515 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
4516     struct thread *td, accmode_t accmode)
4517 {
4518 
4519 	/*
4520 	 * Kernel-invoked always succeeds.
4521 	 */
4522 	if (cred == NOCRED)
4523 		return (0);
4524 
4525 	/*
4526 	 * Do not allow privileged processes in jail to directly manipulate
4527 	 * system attributes.
4528 	 */
4529 	switch (attrnamespace) {
4530 	case EXTATTR_NAMESPACE_SYSTEM:
4531 		/* Potentially should be: return (EPERM); */
4532 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0));
4533 	case EXTATTR_NAMESPACE_USER:
4534 		return (VOP_ACCESS(vp, accmode, cred, td));
4535 	default:
4536 		return (EPERM);
4537 	}
4538 }
4539 
4540 #ifdef DEBUG_VFS_LOCKS
4541 /*
4542  * This only exists to suppress warnings from unlocked specfs accesses.  It is
4543  * no longer ok to have an unlocked VFS.
4544  */
4545 #define	IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL ||		\
4546 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
4547 
4548 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
4549 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
4550     "Drop into debugger on lock violation");
4551 
4552 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
4553 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
4554     0, "Check for interlock across VOPs");
4555 
4556 int vfs_badlock_print = 1;	/* Print lock violations. */
4557 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
4558     0, "Print lock violations");
4559 
4560 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
4561 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
4562     0, "Print vnode details on lock violations");
4563 
4564 #ifdef KDB
4565 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
4566 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
4567     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
4568 #endif
4569 
4570 static void
4571 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
4572 {
4573 
4574 #ifdef KDB
4575 	if (vfs_badlock_backtrace)
4576 		kdb_backtrace();
4577 #endif
4578 	if (vfs_badlock_vnode)
4579 		vn_printf(vp, "vnode ");
4580 	if (vfs_badlock_print)
4581 		printf("%s: %p %s\n", str, (void *)vp, msg);
4582 	if (vfs_badlock_ddb)
4583 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4584 }
4585 
4586 void
4587 assert_vi_locked(struct vnode *vp, const char *str)
4588 {
4589 
4590 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
4591 		vfs_badlock("interlock is not locked but should be", str, vp);
4592 }
4593 
4594 void
4595 assert_vi_unlocked(struct vnode *vp, const char *str)
4596 {
4597 
4598 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
4599 		vfs_badlock("interlock is locked but should not be", str, vp);
4600 }
4601 
4602 void
4603 assert_vop_locked(struct vnode *vp, const char *str)
4604 {
4605 	int locked;
4606 
4607 	if (!IGNORE_LOCK(vp)) {
4608 		locked = VOP_ISLOCKED(vp);
4609 		if (locked == 0 || locked == LK_EXCLOTHER)
4610 			vfs_badlock("is not locked but should be", str, vp);
4611 	}
4612 }
4613 
4614 void
4615 assert_vop_unlocked(struct vnode *vp, const char *str)
4616 {
4617 
4618 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
4619 		vfs_badlock("is locked but should not be", str, vp);
4620 }
4621 
4622 void
4623 assert_vop_elocked(struct vnode *vp, const char *str)
4624 {
4625 
4626 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
4627 		vfs_badlock("is not exclusive locked but should be", str, vp);
4628 }
4629 #endif /* DEBUG_VFS_LOCKS */
4630 
4631 void
4632 vop_rename_fail(struct vop_rename_args *ap)
4633 {
4634 
4635 	if (ap->a_tvp != NULL)
4636 		vput(ap->a_tvp);
4637 	if (ap->a_tdvp == ap->a_tvp)
4638 		vrele(ap->a_tdvp);
4639 	else
4640 		vput(ap->a_tdvp);
4641 	vrele(ap->a_fdvp);
4642 	vrele(ap->a_fvp);
4643 }
4644 
4645 void
4646 vop_rename_pre(void *ap)
4647 {
4648 	struct vop_rename_args *a = ap;
4649 
4650 #ifdef DEBUG_VFS_LOCKS
4651 	if (a->a_tvp)
4652 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
4653 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
4654 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
4655 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
4656 
4657 	/* Check the source (from). */
4658 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
4659 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
4660 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
4661 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
4662 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
4663 
4664 	/* Check the target. */
4665 	if (a->a_tvp)
4666 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
4667 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
4668 #endif
4669 	if (a->a_tdvp != a->a_fdvp)
4670 		vhold(a->a_fdvp);
4671 	if (a->a_tvp != a->a_fvp)
4672 		vhold(a->a_fvp);
4673 	vhold(a->a_tdvp);
4674 	if (a->a_tvp)
4675 		vhold(a->a_tvp);
4676 }
4677 
4678 #ifdef DEBUG_VFS_LOCKS
4679 void
4680 vop_strategy_pre(void *ap)
4681 {
4682 	struct vop_strategy_args *a;
4683 	struct buf *bp;
4684 
4685 	a = ap;
4686 	bp = a->a_bp;
4687 
4688 	/*
4689 	 * Cluster ops lock their component buffers but not the IO container.
4690 	 */
4691 	if ((bp->b_flags & B_CLUSTER) != 0)
4692 		return;
4693 
4694 	if (panicstr == NULL && !BUF_ISLOCKED(bp)) {
4695 		if (vfs_badlock_print)
4696 			printf(
4697 			    "VOP_STRATEGY: bp is not locked but should be\n");
4698 		if (vfs_badlock_ddb)
4699 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4700 	}
4701 }
4702 
4703 void
4704 vop_lock_pre(void *ap)
4705 {
4706 	struct vop_lock1_args *a = ap;
4707 
4708 	if ((a->a_flags & LK_INTERLOCK) == 0)
4709 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4710 	else
4711 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
4712 }
4713 
4714 void
4715 vop_lock_post(void *ap, int rc)
4716 {
4717 	struct vop_lock1_args *a = ap;
4718 
4719 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4720 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
4721 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
4722 }
4723 
4724 void
4725 vop_unlock_pre(void *ap)
4726 {
4727 	struct vop_unlock_args *a = ap;
4728 
4729 	if (a->a_flags & LK_INTERLOCK)
4730 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
4731 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
4732 }
4733 
4734 void
4735 vop_unlock_post(void *ap, int rc)
4736 {
4737 	struct vop_unlock_args *a = ap;
4738 
4739 	if (a->a_flags & LK_INTERLOCK)
4740 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
4741 }
4742 #endif
4743 
4744 void
4745 vop_create_post(void *ap, int rc)
4746 {
4747 	struct vop_create_args *a = ap;
4748 
4749 	if (!rc)
4750 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4751 }
4752 
4753 void
4754 vop_deleteextattr_post(void *ap, int rc)
4755 {
4756 	struct vop_deleteextattr_args *a = ap;
4757 
4758 	if (!rc)
4759 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4760 }
4761 
4762 void
4763 vop_link_post(void *ap, int rc)
4764 {
4765 	struct vop_link_args *a = ap;
4766 
4767 	if (!rc) {
4768 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
4769 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
4770 	}
4771 }
4772 
4773 void
4774 vop_mkdir_post(void *ap, int rc)
4775 {
4776 	struct vop_mkdir_args *a = ap;
4777 
4778 	if (!rc)
4779 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4780 }
4781 
4782 void
4783 vop_mknod_post(void *ap, int rc)
4784 {
4785 	struct vop_mknod_args *a = ap;
4786 
4787 	if (!rc)
4788 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4789 }
4790 
4791 void
4792 vop_reclaim_post(void *ap, int rc)
4793 {
4794 	struct vop_reclaim_args *a = ap;
4795 
4796 	if (!rc)
4797 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
4798 }
4799 
4800 void
4801 vop_remove_post(void *ap, int rc)
4802 {
4803 	struct vop_remove_args *a = ap;
4804 
4805 	if (!rc) {
4806 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4807 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4808 	}
4809 }
4810 
4811 void
4812 vop_rename_post(void *ap, int rc)
4813 {
4814 	struct vop_rename_args *a = ap;
4815 	long hint;
4816 
4817 	if (!rc) {
4818 		hint = NOTE_WRITE;
4819 		if (a->a_fdvp == a->a_tdvp) {
4820 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
4821 				hint |= NOTE_LINK;
4822 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
4823 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
4824 		} else {
4825 			hint |= NOTE_EXTEND;
4826 			if (a->a_fvp->v_type == VDIR)
4827 				hint |= NOTE_LINK;
4828 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
4829 
4830 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
4831 			    a->a_tvp->v_type == VDIR)
4832 				hint &= ~NOTE_LINK;
4833 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
4834 		}
4835 
4836 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
4837 		if (a->a_tvp)
4838 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
4839 	}
4840 	if (a->a_tdvp != a->a_fdvp)
4841 		vdrop(a->a_fdvp);
4842 	if (a->a_tvp != a->a_fvp)
4843 		vdrop(a->a_fvp);
4844 	vdrop(a->a_tdvp);
4845 	if (a->a_tvp)
4846 		vdrop(a->a_tvp);
4847 }
4848 
4849 void
4850 vop_rmdir_post(void *ap, int rc)
4851 {
4852 	struct vop_rmdir_args *a = ap;
4853 
4854 	if (!rc) {
4855 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4856 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4857 	}
4858 }
4859 
4860 void
4861 vop_setattr_post(void *ap, int rc)
4862 {
4863 	struct vop_setattr_args *a = ap;
4864 
4865 	if (!rc)
4866 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4867 }
4868 
4869 void
4870 vop_setextattr_post(void *ap, int rc)
4871 {
4872 	struct vop_setextattr_args *a = ap;
4873 
4874 	if (!rc)
4875 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4876 }
4877 
4878 void
4879 vop_symlink_post(void *ap, int rc)
4880 {
4881 	struct vop_symlink_args *a = ap;
4882 
4883 	if (!rc)
4884 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4885 }
4886 
4887 void
4888 vop_open_post(void *ap, int rc)
4889 {
4890 	struct vop_open_args *a = ap;
4891 
4892 	if (!rc)
4893 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
4894 }
4895 
4896 void
4897 vop_close_post(void *ap, int rc)
4898 {
4899 	struct vop_close_args *a = ap;
4900 
4901 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
4902 	    (a->a_vp->v_iflag & VI_DOOMED) == 0)) {
4903 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
4904 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
4905 	}
4906 }
4907 
4908 void
4909 vop_read_post(void *ap, int rc)
4910 {
4911 	struct vop_read_args *a = ap;
4912 
4913 	if (!rc)
4914 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
4915 }
4916 
4917 void
4918 vop_readdir_post(void *ap, int rc)
4919 {
4920 	struct vop_readdir_args *a = ap;
4921 
4922 	if (!rc)
4923 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
4924 }
4925 
4926 static struct knlist fs_knlist;
4927 
4928 static void
4929 vfs_event_init(void *arg)
4930 {
4931 	knlist_init_mtx(&fs_knlist, NULL);
4932 }
4933 /* XXX - correct order? */
4934 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
4935 
4936 void
4937 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
4938 {
4939 
4940 	KNOTE_UNLOCKED(&fs_knlist, event);
4941 }
4942 
4943 static int	filt_fsattach(struct knote *kn);
4944 static void	filt_fsdetach(struct knote *kn);
4945 static int	filt_fsevent(struct knote *kn, long hint);
4946 
4947 struct filterops fs_filtops = {
4948 	.f_isfd = 0,
4949 	.f_attach = filt_fsattach,
4950 	.f_detach = filt_fsdetach,
4951 	.f_event = filt_fsevent
4952 };
4953 
4954 static int
4955 filt_fsattach(struct knote *kn)
4956 {
4957 
4958 	kn->kn_flags |= EV_CLEAR;
4959 	knlist_add(&fs_knlist, kn, 0);
4960 	return (0);
4961 }
4962 
4963 static void
4964 filt_fsdetach(struct knote *kn)
4965 {
4966 
4967 	knlist_remove(&fs_knlist, kn, 0);
4968 }
4969 
4970 static int
4971 filt_fsevent(struct knote *kn, long hint)
4972 {
4973 
4974 	kn->kn_fflags |= hint;
4975 	return (kn->kn_fflags != 0);
4976 }
4977 
4978 static int
4979 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
4980 {
4981 	struct vfsidctl vc;
4982 	int error;
4983 	struct mount *mp;
4984 
4985 	error = SYSCTL_IN(req, &vc, sizeof(vc));
4986 	if (error)
4987 		return (error);
4988 	if (vc.vc_vers != VFS_CTL_VERS1)
4989 		return (EINVAL);
4990 	mp = vfs_getvfs(&vc.vc_fsid);
4991 	if (mp == NULL)
4992 		return (ENOENT);
4993 	/* ensure that a specific sysctl goes to the right filesystem. */
4994 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
4995 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
4996 		vfs_rel(mp);
4997 		return (EINVAL);
4998 	}
4999 	VCTLTOREQ(&vc, req);
5000 	error = VFS_SYSCTL(mp, vc.vc_op, req);
5001 	vfs_rel(mp);
5002 	return (error);
5003 }
5004 
5005 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR,
5006     NULL, 0, sysctl_vfs_ctl, "",
5007     "Sysctl by fsid");
5008 
5009 /*
5010  * Function to initialize a va_filerev field sensibly.
5011  * XXX: Wouldn't a random number make a lot more sense ??
5012  */
5013 u_quad_t
5014 init_va_filerev(void)
5015 {
5016 	struct bintime bt;
5017 
5018 	getbinuptime(&bt);
5019 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5020 }
5021 
5022 static int	filt_vfsread(struct knote *kn, long hint);
5023 static int	filt_vfswrite(struct knote *kn, long hint);
5024 static int	filt_vfsvnode(struct knote *kn, long hint);
5025 static void	filt_vfsdetach(struct knote *kn);
5026 static struct filterops vfsread_filtops = {
5027 	.f_isfd = 1,
5028 	.f_detach = filt_vfsdetach,
5029 	.f_event = filt_vfsread
5030 };
5031 static struct filterops vfswrite_filtops = {
5032 	.f_isfd = 1,
5033 	.f_detach = filt_vfsdetach,
5034 	.f_event = filt_vfswrite
5035 };
5036 static struct filterops vfsvnode_filtops = {
5037 	.f_isfd = 1,
5038 	.f_detach = filt_vfsdetach,
5039 	.f_event = filt_vfsvnode
5040 };
5041 
5042 static void
5043 vfs_knllock(void *arg)
5044 {
5045 	struct vnode *vp = arg;
5046 
5047 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5048 }
5049 
5050 static void
5051 vfs_knlunlock(void *arg)
5052 {
5053 	struct vnode *vp = arg;
5054 
5055 	VOP_UNLOCK(vp, 0);
5056 }
5057 
5058 static void
5059 vfs_knl_assert_locked(void *arg)
5060 {
5061 #ifdef DEBUG_VFS_LOCKS
5062 	struct vnode *vp = arg;
5063 
5064 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
5065 #endif
5066 }
5067 
5068 static void
5069 vfs_knl_assert_unlocked(void *arg)
5070 {
5071 #ifdef DEBUG_VFS_LOCKS
5072 	struct vnode *vp = arg;
5073 
5074 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
5075 #endif
5076 }
5077 
5078 int
5079 vfs_kqfilter(struct vop_kqfilter_args *ap)
5080 {
5081 	struct vnode *vp = ap->a_vp;
5082 	struct knote *kn = ap->a_kn;
5083 	struct knlist *knl;
5084 
5085 	switch (kn->kn_filter) {
5086 	case EVFILT_READ:
5087 		kn->kn_fop = &vfsread_filtops;
5088 		break;
5089 	case EVFILT_WRITE:
5090 		kn->kn_fop = &vfswrite_filtops;
5091 		break;
5092 	case EVFILT_VNODE:
5093 		kn->kn_fop = &vfsvnode_filtops;
5094 		break;
5095 	default:
5096 		return (EINVAL);
5097 	}
5098 
5099 	kn->kn_hook = (caddr_t)vp;
5100 
5101 	v_addpollinfo(vp);
5102 	if (vp->v_pollinfo == NULL)
5103 		return (ENOMEM);
5104 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
5105 	vhold(vp);
5106 	knlist_add(knl, kn, 0);
5107 
5108 	return (0);
5109 }
5110 
5111 /*
5112  * Detach knote from vnode
5113  */
5114 static void
5115 filt_vfsdetach(struct knote *kn)
5116 {
5117 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5118 
5119 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5120 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5121 	vdrop(vp);
5122 }
5123 
5124 /*ARGSUSED*/
5125 static int
5126 filt_vfsread(struct knote *kn, long hint)
5127 {
5128 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5129 	struct vattr va;
5130 	int res;
5131 
5132 	/*
5133 	 * filesystem is gone, so set the EOF flag and schedule
5134 	 * the knote for deletion.
5135 	 */
5136 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5137 		VI_LOCK(vp);
5138 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5139 		VI_UNLOCK(vp);
5140 		return (1);
5141 	}
5142 
5143 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5144 		return (0);
5145 
5146 	VI_LOCK(vp);
5147 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5148 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5149 	VI_UNLOCK(vp);
5150 	return (res);
5151 }
5152 
5153 /*ARGSUSED*/
5154 static int
5155 filt_vfswrite(struct knote *kn, long hint)
5156 {
5157 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5158 
5159 	VI_LOCK(vp);
5160 
5161 	/*
5162 	 * filesystem is gone, so set the EOF flag and schedule
5163 	 * the knote for deletion.
5164 	 */
5165 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5166 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5167 
5168 	kn->kn_data = 0;
5169 	VI_UNLOCK(vp);
5170 	return (1);
5171 }
5172 
5173 static int
5174 filt_vfsvnode(struct knote *kn, long hint)
5175 {
5176 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5177 	int res;
5178 
5179 	VI_LOCK(vp);
5180 	if (kn->kn_sfflags & hint)
5181 		kn->kn_fflags |= hint;
5182 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5183 		kn->kn_flags |= EV_EOF;
5184 		VI_UNLOCK(vp);
5185 		return (1);
5186 	}
5187 	res = (kn->kn_fflags != 0);
5188 	VI_UNLOCK(vp);
5189 	return (res);
5190 }
5191 
5192 int
5193 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5194 {
5195 	int error;
5196 
5197 	if (dp->d_reclen > ap->a_uio->uio_resid)
5198 		return (ENAMETOOLONG);
5199 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
5200 	if (error) {
5201 		if (ap->a_ncookies != NULL) {
5202 			if (ap->a_cookies != NULL)
5203 				free(ap->a_cookies, M_TEMP);
5204 			ap->a_cookies = NULL;
5205 			*ap->a_ncookies = 0;
5206 		}
5207 		return (error);
5208 	}
5209 	if (ap->a_ncookies == NULL)
5210 		return (0);
5211 
5212 	KASSERT(ap->a_cookies,
5213 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
5214 
5215 	*ap->a_cookies = realloc(*ap->a_cookies,
5216 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
5217 	(*ap->a_cookies)[*ap->a_ncookies] = off;
5218 	*ap->a_ncookies += 1;
5219 	return (0);
5220 }
5221 
5222 /*
5223  * Mark for update the access time of the file if the filesystem
5224  * supports VOP_MARKATIME.  This functionality is used by execve and
5225  * mmap, so we want to avoid the I/O implied by directly setting
5226  * va_atime for the sake of efficiency.
5227  */
5228 void
5229 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
5230 {
5231 	struct mount *mp;
5232 
5233 	mp = vp->v_mount;
5234 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
5235 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
5236 		(void)VOP_MARKATIME(vp);
5237 }
5238 
5239 /*
5240  * The purpose of this routine is to remove granularity from accmode_t,
5241  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
5242  * VADMIN and VAPPEND.
5243  *
5244  * If it returns 0, the caller is supposed to continue with the usual
5245  * access checks using 'accmode' as modified by this routine.  If it
5246  * returns nonzero value, the caller is supposed to return that value
5247  * as errno.
5248  *
5249  * Note that after this routine runs, accmode may be zero.
5250  */
5251 int
5252 vfs_unixify_accmode(accmode_t *accmode)
5253 {
5254 	/*
5255 	 * There is no way to specify explicit "deny" rule using
5256 	 * file mode or POSIX.1e ACLs.
5257 	 */
5258 	if (*accmode & VEXPLICIT_DENY) {
5259 		*accmode = 0;
5260 		return (0);
5261 	}
5262 
5263 	/*
5264 	 * None of these can be translated into usual access bits.
5265 	 * Also, the common case for NFSv4 ACLs is to not contain
5266 	 * either of these bits. Caller should check for VWRITE
5267 	 * on the containing directory instead.
5268 	 */
5269 	if (*accmode & (VDELETE_CHILD | VDELETE))
5270 		return (EPERM);
5271 
5272 	if (*accmode & VADMIN_PERMS) {
5273 		*accmode &= ~VADMIN_PERMS;
5274 		*accmode |= VADMIN;
5275 	}
5276 
5277 	/*
5278 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5279 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5280 	 */
5281 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5282 
5283 	return (0);
5284 }
5285 
5286 /*
5287  * These are helper functions for filesystems to traverse all
5288  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
5289  *
5290  * This interface replaces MNT_VNODE_FOREACH.
5291  */
5292 
5293 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
5294 
5295 struct vnode *
5296 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
5297 {
5298 	struct vnode *vp;
5299 
5300 	if (should_yield())
5301 		kern_yield(PRI_USER);
5302 	MNT_ILOCK(mp);
5303 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5304 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
5305 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
5306 		/* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
5307 		if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
5308 			continue;
5309 		VI_LOCK(vp);
5310 		if ((vp->v_iflag & VI_DOOMED) != 0) {
5311 			VI_UNLOCK(vp);
5312 			continue;
5313 		}
5314 		break;
5315 	}
5316 	if (vp == NULL) {
5317 		__mnt_vnode_markerfree_all(mvp, mp);
5318 		/* MNT_IUNLOCK(mp); -- done in above function */
5319 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
5320 		return (NULL);
5321 	}
5322 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5323 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5324 	MNT_IUNLOCK(mp);
5325 	return (vp);
5326 }
5327 
5328 struct vnode *
5329 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
5330 {
5331 	struct vnode *vp;
5332 
5333 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5334 	MNT_ILOCK(mp);
5335 	MNT_REF(mp);
5336 	(*mvp)->v_mount = mp;
5337 	(*mvp)->v_type = VMARKER;
5338 
5339 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
5340 		/* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
5341 		if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
5342 			continue;
5343 		VI_LOCK(vp);
5344 		if ((vp->v_iflag & VI_DOOMED) != 0) {
5345 			VI_UNLOCK(vp);
5346 			continue;
5347 		}
5348 		break;
5349 	}
5350 	if (vp == NULL) {
5351 		MNT_REL(mp);
5352 		MNT_IUNLOCK(mp);
5353 		free(*mvp, M_VNODE_MARKER);
5354 		*mvp = NULL;
5355 		return (NULL);
5356 	}
5357 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5358 	MNT_IUNLOCK(mp);
5359 	return (vp);
5360 }
5361 
5362 void
5363 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
5364 {
5365 
5366 	if (*mvp == NULL) {
5367 		MNT_IUNLOCK(mp);
5368 		return;
5369 	}
5370 
5371 	mtx_assert(MNT_MTX(mp), MA_OWNED);
5372 
5373 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5374 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5375 	MNT_REL(mp);
5376 	MNT_IUNLOCK(mp);
5377 	free(*mvp, M_VNODE_MARKER);
5378 	*mvp = NULL;
5379 }
5380 
5381 /*
5382  * These are helper functions for filesystems to traverse their
5383  * active vnodes.  See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h
5384  */
5385 static void
5386 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5387 {
5388 
5389 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5390 
5391 	MNT_ILOCK(mp);
5392 	MNT_REL(mp);
5393 	MNT_IUNLOCK(mp);
5394 	free(*mvp, M_VNODE_MARKER);
5395 	*mvp = NULL;
5396 }
5397 
5398 /*
5399  * Relock the mp mount vnode list lock with the vp vnode interlock in the
5400  * conventional lock order during mnt_vnode_next_active iteration.
5401  *
5402  * On entry, the mount vnode list lock is held and the vnode interlock is not.
5403  * The list lock is dropped and reacquired.  On success, both locks are held.
5404  * On failure, the mount vnode list lock is held but the vnode interlock is
5405  * not, and the procedure may have yielded.
5406  */
5407 static bool
5408 mnt_vnode_next_active_relock(struct vnode *mvp, struct mount *mp,
5409     struct vnode *vp)
5410 {
5411 	const struct vnode *tmp;
5412 	bool held, ret;
5413 
5414 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
5415 	    TAILQ_NEXT(mvp, v_actfreelist) != NULL, mvp,
5416 	    ("%s: bad marker", __func__));
5417 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
5418 	    ("%s: inappropriate vnode", __func__));
5419 	ASSERT_VI_UNLOCKED(vp, __func__);
5420 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5421 
5422 	ret = false;
5423 
5424 	TAILQ_REMOVE(&mp->mnt_activevnodelist, mvp, v_actfreelist);
5425 	TAILQ_INSERT_BEFORE(vp, mvp, v_actfreelist);
5426 
5427 	/*
5428 	 * Use a hold to prevent vp from disappearing while the mount vnode
5429 	 * list lock is dropped and reacquired.  Normally a hold would be
5430 	 * acquired with vhold(), but that might try to acquire the vnode
5431 	 * interlock, which would be a LOR with the mount vnode list lock.
5432 	 */
5433 	held = vfs_refcount_acquire_if_not_zero(&vp->v_holdcnt);
5434 	mtx_unlock(&mp->mnt_listmtx);
5435 	if (!held)
5436 		goto abort;
5437 	VI_LOCK(vp);
5438 	if (!vfs_refcount_release_if_not_last(&vp->v_holdcnt)) {
5439 		vdropl(vp);
5440 		goto abort;
5441 	}
5442 	mtx_lock(&mp->mnt_listmtx);
5443 
5444 	/*
5445 	 * Determine whether the vnode is still the next one after the marker,
5446 	 * excepting any other markers.  If the vnode has not been doomed by
5447 	 * vgone() then the hold should have ensured that it remained on the
5448 	 * active list.  If it has been doomed but is still on the active list,
5449 	 * don't abort, but rather skip over it (avoid spinning on doomed
5450 	 * vnodes).
5451 	 */
5452 	tmp = mvp;
5453 	do {
5454 		tmp = TAILQ_NEXT(tmp, v_actfreelist);
5455 	} while (tmp != NULL && tmp->v_type == VMARKER);
5456 	if (tmp != vp) {
5457 		mtx_unlock(&mp->mnt_listmtx);
5458 		VI_UNLOCK(vp);
5459 		goto abort;
5460 	}
5461 
5462 	ret = true;
5463 	goto out;
5464 abort:
5465 	maybe_yield();
5466 	mtx_lock(&mp->mnt_listmtx);
5467 out:
5468 	if (ret)
5469 		ASSERT_VI_LOCKED(vp, __func__);
5470 	else
5471 		ASSERT_VI_UNLOCKED(vp, __func__);
5472 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5473 	return (ret);
5474 }
5475 
5476 static struct vnode *
5477 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5478 {
5479 	struct vnode *vp, *nvp;
5480 
5481 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5482 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5483 restart:
5484 	vp = TAILQ_NEXT(*mvp, v_actfreelist);
5485 	while (vp != NULL) {
5486 		if (vp->v_type == VMARKER) {
5487 			vp = TAILQ_NEXT(vp, v_actfreelist);
5488 			continue;
5489 		}
5490 		/*
5491 		 * Try-lock because this is the wrong lock order.  If that does
5492 		 * not succeed, drop the mount vnode list lock and try to
5493 		 * reacquire it and the vnode interlock in the right order.
5494 		 */
5495 		if (!VI_TRYLOCK(vp) &&
5496 		    !mnt_vnode_next_active_relock(*mvp, mp, vp))
5497 			goto restart;
5498 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
5499 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
5500 		    ("alien vnode on the active list %p %p", vp, mp));
5501 		if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0)
5502 			break;
5503 		nvp = TAILQ_NEXT(vp, v_actfreelist);
5504 		VI_UNLOCK(vp);
5505 		vp = nvp;
5506 	}
5507 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5508 
5509 	/* Check if we are done */
5510 	if (vp == NULL) {
5511 		mtx_unlock(&mp->mnt_listmtx);
5512 		mnt_vnode_markerfree_active(mvp, mp);
5513 		return (NULL);
5514 	}
5515 	TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist);
5516 	mtx_unlock(&mp->mnt_listmtx);
5517 	ASSERT_VI_LOCKED(vp, "active iter");
5518 	KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp));
5519 	return (vp);
5520 }
5521 
5522 struct vnode *
5523 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5524 {
5525 
5526 	if (should_yield())
5527 		kern_yield(PRI_USER);
5528 	mtx_lock(&mp->mnt_listmtx);
5529 	return (mnt_vnode_next_active(mvp, mp));
5530 }
5531 
5532 struct vnode *
5533 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp)
5534 {
5535 	struct vnode *vp;
5536 
5537 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5538 	MNT_ILOCK(mp);
5539 	MNT_REF(mp);
5540 	MNT_IUNLOCK(mp);
5541 	(*mvp)->v_type = VMARKER;
5542 	(*mvp)->v_mount = mp;
5543 
5544 	mtx_lock(&mp->mnt_listmtx);
5545 	vp = TAILQ_FIRST(&mp->mnt_activevnodelist);
5546 	if (vp == NULL) {
5547 		mtx_unlock(&mp->mnt_listmtx);
5548 		mnt_vnode_markerfree_active(mvp, mp);
5549 		return (NULL);
5550 	}
5551 	TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5552 	return (mnt_vnode_next_active(mvp, mp));
5553 }
5554 
5555 void
5556 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5557 {
5558 
5559 	if (*mvp == NULL)
5560 		return;
5561 
5562 	mtx_lock(&mp->mnt_listmtx);
5563 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5564 	mtx_unlock(&mp->mnt_listmtx);
5565 	mnt_vnode_markerfree_active(mvp, mp);
5566 }
5567