xref: /freebsd/sys/kern/vfs_subr.c (revision eb69d1f144a6fcc765d1b9d44a5ae8082353e70b)
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 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 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 		onumvnodes = numvnodes;
1174 		/*
1175 		 * Calculate parameters for recycling.  These are the same
1176 		 * throughout the loop to give some semblance of fairness.
1177 		 * The trigger point is to avoid recycling vnodes with lots
1178 		 * of resident pages.  We aren't trying to free memory; we
1179 		 * are trying to recycle or at least free vnodes.
1180 		 */
1181 		if (numvnodes <= desiredvnodes)
1182 			usevnodes = numvnodes - freevnodes;
1183 		else
1184 			usevnodes = numvnodes;
1185 		if (usevnodes <= 0)
1186 			usevnodes = 1;
1187 		/*
1188 		 * The trigger value is is chosen to give a conservatively
1189 		 * large value to ensure that it alone doesn't prevent
1190 		 * making progress.  The value can easily be so large that
1191 		 * it is effectively infinite in some congested and
1192 		 * misconfigured cases, and this is necessary.  Normally
1193 		 * it is about 8 to 100 (pages), which is quite large.
1194 		 */
1195 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1196 		if (force < 2)
1197 			trigger = vsmalltrigger;
1198 		reclaim_nc_src = force >= 3;
1199 		mtx_lock(&mountlist_mtx);
1200 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1201 			if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) {
1202 				nmp = TAILQ_NEXT(mp, mnt_list);
1203 				continue;
1204 			}
1205 			done += vlrureclaim(mp, reclaim_nc_src, trigger);
1206 			mtx_lock(&mountlist_mtx);
1207 			nmp = TAILQ_NEXT(mp, mnt_list);
1208 			vfs_unbusy(mp);
1209 		}
1210 		mtx_unlock(&mountlist_mtx);
1211 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1212 			uma_reclaim();
1213 		if (done == 0) {
1214 			if (force == 0 || force == 1) {
1215 				force = 2;
1216 				continue;
1217 			}
1218 			if (force == 2) {
1219 				force = 3;
1220 				continue;
1221 			}
1222 			force = 0;
1223 			vnlru_nowhere++;
1224 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1225 		} else
1226 			kern_yield(PRI_USER);
1227 		/*
1228 		 * After becoming active to expand above low water, keep
1229 		 * active until above high water.
1230 		 */
1231 		force = vspace() < vhiwat;
1232 	}
1233 }
1234 
1235 static struct kproc_desc vnlru_kp = {
1236 	"vnlru",
1237 	vnlru_proc,
1238 	&vnlruproc
1239 };
1240 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1241     &vnlru_kp);
1242 
1243 /*
1244  * Routines having to do with the management of the vnode table.
1245  */
1246 
1247 /*
1248  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1249  * before we actually vgone().  This function must be called with the vnode
1250  * held to prevent the vnode from being returned to the free list midway
1251  * through vgone().
1252  */
1253 static int
1254 vtryrecycle(struct vnode *vp)
1255 {
1256 	struct mount *vnmp;
1257 
1258 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1259 	VNASSERT(vp->v_holdcnt, vp,
1260 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1261 	/*
1262 	 * This vnode may found and locked via some other list, if so we
1263 	 * can't recycle it yet.
1264 	 */
1265 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1266 		CTR2(KTR_VFS,
1267 		    "%s: impossible to recycle, vp %p lock is already held",
1268 		    __func__, vp);
1269 		return (EWOULDBLOCK);
1270 	}
1271 	/*
1272 	 * Don't recycle if its filesystem is being suspended.
1273 	 */
1274 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1275 		VOP_UNLOCK(vp, 0);
1276 		CTR2(KTR_VFS,
1277 		    "%s: impossible to recycle, cannot start the write for %p",
1278 		    __func__, vp);
1279 		return (EBUSY);
1280 	}
1281 	/*
1282 	 * If we got this far, we need to acquire the interlock and see if
1283 	 * anyone picked up this vnode from another list.  If not, we will
1284 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1285 	 * will skip over it.
1286 	 */
1287 	VI_LOCK(vp);
1288 	if (vp->v_usecount) {
1289 		VOP_UNLOCK(vp, LK_INTERLOCK);
1290 		vn_finished_write(vnmp);
1291 		CTR2(KTR_VFS,
1292 		    "%s: impossible to recycle, %p is already referenced",
1293 		    __func__, vp);
1294 		return (EBUSY);
1295 	}
1296 	if ((vp->v_iflag & VI_DOOMED) == 0) {
1297 		counter_u64_add(recycles_count, 1);
1298 		vgonel(vp);
1299 	}
1300 	VOP_UNLOCK(vp, LK_INTERLOCK);
1301 	vn_finished_write(vnmp);
1302 	return (0);
1303 }
1304 
1305 static void
1306 vcheckspace(void)
1307 {
1308 
1309 	if (vspace() < vlowat && vnlruproc_sig == 0) {
1310 		vnlruproc_sig = 1;
1311 		wakeup(vnlruproc);
1312 	}
1313 }
1314 
1315 /*
1316  * Wait if necessary for space for a new vnode.
1317  */
1318 static int
1319 getnewvnode_wait(int suspended)
1320 {
1321 
1322 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1323 	if (numvnodes >= desiredvnodes) {
1324 		if (suspended) {
1325 			/*
1326 			 * The file system is being suspended.  We cannot
1327 			 * risk a deadlock here, so allow allocation of
1328 			 * another vnode even if this would give too many.
1329 			 */
1330 			return (0);
1331 		}
1332 		if (vnlruproc_sig == 0) {
1333 			vnlruproc_sig = 1;	/* avoid unnecessary wakeups */
1334 			wakeup(vnlruproc);
1335 		}
1336 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
1337 		    "vlruwk", hz);
1338 	}
1339 	/* Post-adjust like the pre-adjust in getnewvnode(). */
1340 	if (numvnodes + 1 > desiredvnodes && freevnodes > 1)
1341 		vnlru_free_locked(1, NULL);
1342 	return (numvnodes >= desiredvnodes ? ENFILE : 0);
1343 }
1344 
1345 /*
1346  * This hack is fragile, and probably not needed any more now that the
1347  * watermark handling works.
1348  */
1349 void
1350 getnewvnode_reserve(u_int count)
1351 {
1352 	struct thread *td;
1353 
1354 	/* Pre-adjust like the pre-adjust in getnewvnode(), with any count. */
1355 	/* XXX no longer so quick, but this part is not racy. */
1356 	mtx_lock(&vnode_free_list_mtx);
1357 	if (numvnodes + count > desiredvnodes && freevnodes > wantfreevnodes)
1358 		vnlru_free_locked(ulmin(numvnodes + count - desiredvnodes,
1359 		    freevnodes - wantfreevnodes), NULL);
1360 	mtx_unlock(&vnode_free_list_mtx);
1361 
1362 	td = curthread;
1363 	/* First try to be quick and racy. */
1364 	if (atomic_fetchadd_long(&numvnodes, count) + count <= desiredvnodes) {
1365 		td->td_vp_reserv += count;
1366 		vcheckspace();	/* XXX no longer so quick, but more racy */
1367 		return;
1368 	} else
1369 		atomic_subtract_long(&numvnodes, count);
1370 
1371 	mtx_lock(&vnode_free_list_mtx);
1372 	while (count > 0) {
1373 		if (getnewvnode_wait(0) == 0) {
1374 			count--;
1375 			td->td_vp_reserv++;
1376 			atomic_add_long(&numvnodes, 1);
1377 		}
1378 	}
1379 	vcheckspace();
1380 	mtx_unlock(&vnode_free_list_mtx);
1381 }
1382 
1383 /*
1384  * This hack is fragile, especially if desiredvnodes or wantvnodes are
1385  * misconfgured or changed significantly.  Reducing desiredvnodes below
1386  * the reserved amount should cause bizarre behaviour like reducing it
1387  * below the number of active vnodes -- the system will try to reduce
1388  * numvnodes to match, but should fail, so the subtraction below should
1389  * not overflow.
1390  */
1391 void
1392 getnewvnode_drop_reserve(void)
1393 {
1394 	struct thread *td;
1395 
1396 	td = curthread;
1397 	atomic_subtract_long(&numvnodes, td->td_vp_reserv);
1398 	td->td_vp_reserv = 0;
1399 }
1400 
1401 /*
1402  * Return the next vnode from the free list.
1403  */
1404 int
1405 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1406     struct vnode **vpp)
1407 {
1408 	struct vnode *vp;
1409 	struct thread *td;
1410 	struct lock_object *lo;
1411 	static int cyclecount;
1412 	int error;
1413 
1414 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1415 	vp = NULL;
1416 	td = curthread;
1417 	if (td->td_vp_reserv > 0) {
1418 		td->td_vp_reserv -= 1;
1419 		goto alloc;
1420 	}
1421 	mtx_lock(&vnode_free_list_mtx);
1422 	if (numvnodes < desiredvnodes)
1423 		cyclecount = 0;
1424 	else if (cyclecount++ >= freevnodes) {
1425 		cyclecount = 0;
1426 		vstir = 1;
1427 	}
1428 	/*
1429 	 * Grow the vnode cache if it will not be above its target max
1430 	 * after growing.  Otherwise, if the free list is nonempty, try
1431 	 * to reclaim 1 item from it before growing the cache (possibly
1432 	 * above its target max if the reclamation failed or is delayed).
1433 	 * Otherwise, wait for some space.  In all cases, schedule
1434 	 * vnlru_proc() if we are getting short of space.  The watermarks
1435 	 * should be chosen so that we never wait or even reclaim from
1436 	 * the free list to below its target minimum.
1437 	 */
1438 	if (numvnodes + 1 <= desiredvnodes)
1439 		;
1440 	else if (freevnodes > 0)
1441 		vnlru_free_locked(1, NULL);
1442 	else {
1443 		error = getnewvnode_wait(mp != NULL && (mp->mnt_kern_flag &
1444 		    MNTK_SUSPEND));
1445 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
1446 		if (error != 0) {
1447 			mtx_unlock(&vnode_free_list_mtx);
1448 			return (error);
1449 		}
1450 #endif
1451 	}
1452 	vcheckspace();
1453 	atomic_add_long(&numvnodes, 1);
1454 	mtx_unlock(&vnode_free_list_mtx);
1455 alloc:
1456 	counter_u64_add(vnodes_created, 1);
1457 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK);
1458 	/*
1459 	 * Locks are given the generic name "vnode" when created.
1460 	 * Follow the historic practice of using the filesystem
1461 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1462 	 *
1463 	 * Locks live in a witness group keyed on their name. Thus,
1464 	 * when a lock is renamed, it must also move from the witness
1465 	 * group of its old name to the witness group of its new name.
1466 	 *
1467 	 * The change only needs to be made when the vnode moves
1468 	 * from one filesystem type to another. We ensure that each
1469 	 * filesystem use a single static name pointer for its tag so
1470 	 * that we can compare pointers rather than doing a strcmp().
1471 	 */
1472 	lo = &vp->v_vnlock->lock_object;
1473 	if (lo->lo_name != tag) {
1474 		lo->lo_name = tag;
1475 		WITNESS_DESTROY(lo);
1476 		WITNESS_INIT(lo, tag);
1477 	}
1478 	/*
1479 	 * By default, don't allow shared locks unless filesystems opt-in.
1480 	 */
1481 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1482 	/*
1483 	 * Finalize various vnode identity bits.
1484 	 */
1485 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1486 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1487 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1488 	vp->v_type = VNON;
1489 	vp->v_tag = tag;
1490 	vp->v_op = vops;
1491 	v_init_counters(vp);
1492 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1493 #ifdef DIAGNOSTIC
1494 	if (mp == NULL && vops != &dead_vnodeops)
1495 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1496 #endif
1497 #ifdef MAC
1498 	mac_vnode_init(vp);
1499 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1500 		mac_vnode_associate_singlelabel(mp, vp);
1501 #endif
1502 	if (mp != NULL) {
1503 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1504 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1505 			vp->v_vflag |= VV_NOKNOTE;
1506 	}
1507 
1508 	/*
1509 	 * For the filesystems which do not use vfs_hash_insert(),
1510 	 * still initialize v_hash to have vfs_hash_index() useful.
1511 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1512 	 * its own hashing.
1513 	 */
1514 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1515 
1516 	*vpp = vp;
1517 	return (0);
1518 }
1519 
1520 /*
1521  * Delete from old mount point vnode list, if on one.
1522  */
1523 static void
1524 delmntque(struct vnode *vp)
1525 {
1526 	struct mount *mp;
1527 	int active;
1528 
1529 	mp = vp->v_mount;
1530 	if (mp == NULL)
1531 		return;
1532 	MNT_ILOCK(mp);
1533 	VI_LOCK(vp);
1534 	KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize,
1535 	    ("Active vnode list size %d > Vnode list size %d",
1536 	     mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize));
1537 	active = vp->v_iflag & VI_ACTIVE;
1538 	vp->v_iflag &= ~VI_ACTIVE;
1539 	if (active) {
1540 		mtx_lock(&mp->mnt_listmtx);
1541 		TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist);
1542 		mp->mnt_activevnodelistsize--;
1543 		mtx_unlock(&mp->mnt_listmtx);
1544 	}
1545 	vp->v_mount = NULL;
1546 	VI_UNLOCK(vp);
1547 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1548 		("bad mount point vnode list size"));
1549 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1550 	mp->mnt_nvnodelistsize--;
1551 	MNT_REL(mp);
1552 	MNT_IUNLOCK(mp);
1553 }
1554 
1555 static void
1556 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1557 {
1558 
1559 	vp->v_data = NULL;
1560 	vp->v_op = &dead_vnodeops;
1561 	vgone(vp);
1562 	vput(vp);
1563 }
1564 
1565 /*
1566  * Insert into list of vnodes for the new mount point, if available.
1567  */
1568 int
1569 insmntque1(struct vnode *vp, struct mount *mp,
1570 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1571 {
1572 
1573 	KASSERT(vp->v_mount == NULL,
1574 		("insmntque: vnode already on per mount vnode list"));
1575 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1576 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1577 
1578 	/*
1579 	 * We acquire the vnode interlock early to ensure that the
1580 	 * vnode cannot be recycled by another process releasing a
1581 	 * holdcnt on it before we get it on both the vnode list
1582 	 * and the active vnode list. The mount mutex protects only
1583 	 * manipulation of the vnode list and the vnode freelist
1584 	 * mutex protects only manipulation of the active vnode list.
1585 	 * Hence the need to hold the vnode interlock throughout.
1586 	 */
1587 	MNT_ILOCK(mp);
1588 	VI_LOCK(vp);
1589 	if (((mp->mnt_kern_flag & MNTK_NOINSMNTQ) != 0 &&
1590 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1591 	    mp->mnt_nvnodelistsize == 0)) &&
1592 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1593 		VI_UNLOCK(vp);
1594 		MNT_IUNLOCK(mp);
1595 		if (dtr != NULL)
1596 			dtr(vp, dtr_arg);
1597 		return (EBUSY);
1598 	}
1599 	vp->v_mount = mp;
1600 	MNT_REF(mp);
1601 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1602 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1603 		("neg mount point vnode list size"));
1604 	mp->mnt_nvnodelistsize++;
1605 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1606 	    ("Activating already active vnode"));
1607 	vp->v_iflag |= VI_ACTIVE;
1608 	mtx_lock(&mp->mnt_listmtx);
1609 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
1610 	mp->mnt_activevnodelistsize++;
1611 	mtx_unlock(&mp->mnt_listmtx);
1612 	VI_UNLOCK(vp);
1613 	MNT_IUNLOCK(mp);
1614 	return (0);
1615 }
1616 
1617 int
1618 insmntque(struct vnode *vp, struct mount *mp)
1619 {
1620 
1621 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1622 }
1623 
1624 /*
1625  * Flush out and invalidate all buffers associated with a bufobj
1626  * Called with the underlying object locked.
1627  */
1628 int
1629 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1630 {
1631 	int error;
1632 
1633 	BO_LOCK(bo);
1634 	if (flags & V_SAVE) {
1635 		error = bufobj_wwait(bo, slpflag, slptimeo);
1636 		if (error) {
1637 			BO_UNLOCK(bo);
1638 			return (error);
1639 		}
1640 		if (bo->bo_dirty.bv_cnt > 0) {
1641 			BO_UNLOCK(bo);
1642 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1643 				return (error);
1644 			/*
1645 			 * XXX We could save a lock/unlock if this was only
1646 			 * enabled under INVARIANTS
1647 			 */
1648 			BO_LOCK(bo);
1649 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1650 				panic("vinvalbuf: dirty bufs");
1651 		}
1652 	}
1653 	/*
1654 	 * If you alter this loop please notice that interlock is dropped and
1655 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1656 	 * no race conditions occur from this.
1657 	 */
1658 	do {
1659 		error = flushbuflist(&bo->bo_clean,
1660 		    flags, bo, slpflag, slptimeo);
1661 		if (error == 0 && !(flags & V_CLEANONLY))
1662 			error = flushbuflist(&bo->bo_dirty,
1663 			    flags, bo, slpflag, slptimeo);
1664 		if (error != 0 && error != EAGAIN) {
1665 			BO_UNLOCK(bo);
1666 			return (error);
1667 		}
1668 	} while (error != 0);
1669 
1670 	/*
1671 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1672 	 * have write I/O in-progress but if there is a VM object then the
1673 	 * VM object can also have read-I/O in-progress.
1674 	 */
1675 	do {
1676 		bufobj_wwait(bo, 0, 0);
1677 		if ((flags & V_VMIO) == 0) {
1678 			BO_UNLOCK(bo);
1679 			if (bo->bo_object != NULL) {
1680 				VM_OBJECT_WLOCK(bo->bo_object);
1681 				vm_object_pip_wait(bo->bo_object, "bovlbx");
1682 				VM_OBJECT_WUNLOCK(bo->bo_object);
1683 			}
1684 			BO_LOCK(bo);
1685 		}
1686 	} while (bo->bo_numoutput > 0);
1687 	BO_UNLOCK(bo);
1688 
1689 	/*
1690 	 * Destroy the copy in the VM cache, too.
1691 	 */
1692 	if (bo->bo_object != NULL &&
1693 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1694 		VM_OBJECT_WLOCK(bo->bo_object);
1695 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1696 		    OBJPR_CLEANONLY : 0);
1697 		VM_OBJECT_WUNLOCK(bo->bo_object);
1698 	}
1699 
1700 #ifdef INVARIANTS
1701 	BO_LOCK(bo);
1702 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1703 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1704 	    bo->bo_clean.bv_cnt > 0))
1705 		panic("vinvalbuf: flush failed");
1706 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1707 	    bo->bo_dirty.bv_cnt > 0)
1708 		panic("vinvalbuf: flush dirty failed");
1709 	BO_UNLOCK(bo);
1710 #endif
1711 	return (0);
1712 }
1713 
1714 /*
1715  * Flush out and invalidate all buffers associated with a vnode.
1716  * Called with the underlying object locked.
1717  */
1718 int
1719 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1720 {
1721 
1722 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1723 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1724 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1725 		return (0);
1726 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1727 }
1728 
1729 /*
1730  * Flush out buffers on the specified list.
1731  *
1732  */
1733 static int
1734 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1735     int slptimeo)
1736 {
1737 	struct buf *bp, *nbp;
1738 	int retval, error;
1739 	daddr_t lblkno;
1740 	b_xflags_t xflags;
1741 
1742 	ASSERT_BO_WLOCKED(bo);
1743 
1744 	retval = 0;
1745 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1746 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1747 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1748 			continue;
1749 		}
1750 		if (nbp != NULL) {
1751 			lblkno = nbp->b_lblkno;
1752 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
1753 		}
1754 		retval = EAGAIN;
1755 		error = BUF_TIMELOCK(bp,
1756 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
1757 		    "flushbuf", slpflag, slptimeo);
1758 		if (error) {
1759 			BO_LOCK(bo);
1760 			return (error != ENOLCK ? error : EAGAIN);
1761 		}
1762 		KASSERT(bp->b_bufobj == bo,
1763 		    ("bp %p wrong b_bufobj %p should be %p",
1764 		    bp, bp->b_bufobj, bo));
1765 		/*
1766 		 * XXX Since there are no node locks for NFS, I
1767 		 * believe there is a slight chance that a delayed
1768 		 * write will occur while sleeping just above, so
1769 		 * check for it.
1770 		 */
1771 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1772 		    (flags & V_SAVE)) {
1773 			bremfree(bp);
1774 			bp->b_flags |= B_ASYNC;
1775 			bwrite(bp);
1776 			BO_LOCK(bo);
1777 			return (EAGAIN);	/* XXX: why not loop ? */
1778 		}
1779 		bremfree(bp);
1780 		bp->b_flags |= (B_INVAL | B_RELBUF);
1781 		bp->b_flags &= ~B_ASYNC;
1782 		brelse(bp);
1783 		BO_LOCK(bo);
1784 		if (nbp == NULL)
1785 			break;
1786 		nbp = gbincore(bo, lblkno);
1787 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1788 		    != xflags)
1789 			break;			/* nbp invalid */
1790 	}
1791 	return (retval);
1792 }
1793 
1794 int
1795 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
1796 {
1797 	struct buf *bp;
1798 	int error;
1799 	daddr_t lblkno;
1800 
1801 	ASSERT_BO_LOCKED(bo);
1802 
1803 	for (lblkno = startn;;) {
1804 again:
1805 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
1806 		if (bp == NULL || bp->b_lblkno >= endn ||
1807 		    bp->b_lblkno < startn)
1808 			break;
1809 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
1810 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
1811 		if (error != 0) {
1812 			BO_RLOCK(bo);
1813 			if (error == ENOLCK)
1814 				goto again;
1815 			return (error);
1816 		}
1817 		KASSERT(bp->b_bufobj == bo,
1818 		    ("bp %p wrong b_bufobj %p should be %p",
1819 		    bp, bp->b_bufobj, bo));
1820 		lblkno = bp->b_lblkno + 1;
1821 		if ((bp->b_flags & B_MANAGED) == 0)
1822 			bremfree(bp);
1823 		bp->b_flags |= B_RELBUF;
1824 		/*
1825 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
1826 		 * pages backing each buffer in the range are unlikely to be
1827 		 * reused.  Dirty buffers will have the hint applied once
1828 		 * they've been written.
1829 		 */
1830 		if (bp->b_vp->v_object != NULL)
1831 			bp->b_flags |= B_NOREUSE;
1832 		brelse(bp);
1833 		BO_RLOCK(bo);
1834 	}
1835 	return (0);
1836 }
1837 
1838 /*
1839  * Truncate a file's buffer and pages to a specified length.  This
1840  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1841  * sync activity.
1842  */
1843 int
1844 vtruncbuf(struct vnode *vp, struct ucred *cred, off_t length, int blksize)
1845 {
1846 	struct buf *bp, *nbp;
1847 	int anyfreed;
1848 	int trunclbn;
1849 	struct bufobj *bo;
1850 
1851 	CTR5(KTR_VFS, "%s: vp %p with cred %p and block %d:%ju", __func__,
1852 	    vp, cred, blksize, (uintmax_t)length);
1853 
1854 	/*
1855 	 * Round up to the *next* lbn.
1856 	 */
1857 	trunclbn = howmany(length, blksize);
1858 
1859 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1860 restart:
1861 	bo = &vp->v_bufobj;
1862 	BO_LOCK(bo);
1863 	anyfreed = 1;
1864 	for (;anyfreed;) {
1865 		anyfreed = 0;
1866 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1867 			if (bp->b_lblkno < trunclbn)
1868 				continue;
1869 			if (BUF_LOCK(bp,
1870 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1871 			    BO_LOCKPTR(bo)) == ENOLCK)
1872 				goto restart;
1873 
1874 			bremfree(bp);
1875 			bp->b_flags |= (B_INVAL | B_RELBUF);
1876 			bp->b_flags &= ~B_ASYNC;
1877 			brelse(bp);
1878 			anyfreed = 1;
1879 
1880 			BO_LOCK(bo);
1881 			if (nbp != NULL &&
1882 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1883 			    (nbp->b_vp != vp) ||
1884 			    (nbp->b_flags & B_DELWRI))) {
1885 				BO_UNLOCK(bo);
1886 				goto restart;
1887 			}
1888 		}
1889 
1890 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1891 			if (bp->b_lblkno < trunclbn)
1892 				continue;
1893 			if (BUF_LOCK(bp,
1894 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1895 			    BO_LOCKPTR(bo)) == ENOLCK)
1896 				goto restart;
1897 			bremfree(bp);
1898 			bp->b_flags |= (B_INVAL | B_RELBUF);
1899 			bp->b_flags &= ~B_ASYNC;
1900 			brelse(bp);
1901 			anyfreed = 1;
1902 
1903 			BO_LOCK(bo);
1904 			if (nbp != NULL &&
1905 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1906 			    (nbp->b_vp != vp) ||
1907 			    (nbp->b_flags & B_DELWRI) == 0)) {
1908 				BO_UNLOCK(bo);
1909 				goto restart;
1910 			}
1911 		}
1912 	}
1913 
1914 	if (length > 0) {
1915 restartsync:
1916 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1917 			if (bp->b_lblkno > 0)
1918 				continue;
1919 			/*
1920 			 * Since we hold the vnode lock this should only
1921 			 * fail if we're racing with the buf daemon.
1922 			 */
1923 			if (BUF_LOCK(bp,
1924 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1925 			    BO_LOCKPTR(bo)) == ENOLCK) {
1926 				goto restart;
1927 			}
1928 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1929 			    ("buf(%p) on dirty queue without DELWRI", bp));
1930 
1931 			bremfree(bp);
1932 			bawrite(bp);
1933 			BO_LOCK(bo);
1934 			goto restartsync;
1935 		}
1936 	}
1937 
1938 	bufobj_wwait(bo, 0, 0);
1939 	BO_UNLOCK(bo);
1940 	vnode_pager_setsize(vp, length);
1941 
1942 	return (0);
1943 }
1944 
1945 static void
1946 buf_vlist_remove(struct buf *bp)
1947 {
1948 	struct bufv *bv;
1949 
1950 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1951 	ASSERT_BO_WLOCKED(bp->b_bufobj);
1952 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
1953 	    (BX_VNDIRTY|BX_VNCLEAN),
1954 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
1955 	if (bp->b_xflags & BX_VNDIRTY)
1956 		bv = &bp->b_bufobj->bo_dirty;
1957 	else
1958 		bv = &bp->b_bufobj->bo_clean;
1959 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
1960 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1961 	bv->bv_cnt--;
1962 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1963 }
1964 
1965 /*
1966  * Add the buffer to the sorted clean or dirty block list.
1967  *
1968  * NOTE: xflags is passed as a constant, optimizing this inline function!
1969  */
1970 static void
1971 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1972 {
1973 	struct bufv *bv;
1974 	struct buf *n;
1975 	int error;
1976 
1977 	ASSERT_BO_WLOCKED(bo);
1978 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
1979 	    ("dead bo %p", bo));
1980 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1981 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
1982 	bp->b_xflags |= xflags;
1983 	if (xflags & BX_VNDIRTY)
1984 		bv = &bo->bo_dirty;
1985 	else
1986 		bv = &bo->bo_clean;
1987 
1988 	/*
1989 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
1990 	 * we tend to grow at the tail so lookup_le should usually be cheaper
1991 	 * than _ge.
1992 	 */
1993 	if (bv->bv_cnt == 0 ||
1994 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
1995 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1996 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
1997 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
1998 	else
1999 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2000 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2001 	if (error)
2002 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
2003 	bv->bv_cnt++;
2004 }
2005 
2006 /*
2007  * Look up a buffer using the buffer tries.
2008  */
2009 struct buf *
2010 gbincore(struct bufobj *bo, daddr_t lblkno)
2011 {
2012 	struct buf *bp;
2013 
2014 	ASSERT_BO_LOCKED(bo);
2015 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2016 	if (bp != NULL)
2017 		return (bp);
2018 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
2019 }
2020 
2021 /*
2022  * Associate a buffer with a vnode.
2023  */
2024 void
2025 bgetvp(struct vnode *vp, struct buf *bp)
2026 {
2027 	struct bufobj *bo;
2028 
2029 	bo = &vp->v_bufobj;
2030 	ASSERT_BO_WLOCKED(bo);
2031 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2032 
2033 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2034 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2035 	    ("bgetvp: bp already attached! %p", bp));
2036 
2037 	vhold(vp);
2038 	bp->b_vp = vp;
2039 	bp->b_bufobj = bo;
2040 	/*
2041 	 * Insert onto list for new vnode.
2042 	 */
2043 	buf_vlist_add(bp, bo, BX_VNCLEAN);
2044 }
2045 
2046 /*
2047  * Disassociate a buffer from a vnode.
2048  */
2049 void
2050 brelvp(struct buf *bp)
2051 {
2052 	struct bufobj *bo;
2053 	struct vnode *vp;
2054 
2055 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2056 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2057 
2058 	/*
2059 	 * Delete from old vnode list, if on one.
2060 	 */
2061 	vp = bp->b_vp;		/* XXX */
2062 	bo = bp->b_bufobj;
2063 	BO_LOCK(bo);
2064 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2065 		buf_vlist_remove(bp);
2066 	else
2067 		panic("brelvp: Buffer %p not on queue.", bp);
2068 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2069 		bo->bo_flag &= ~BO_ONWORKLST;
2070 		mtx_lock(&sync_mtx);
2071 		LIST_REMOVE(bo, bo_synclist);
2072 		syncer_worklist_len--;
2073 		mtx_unlock(&sync_mtx);
2074 	}
2075 	bp->b_vp = NULL;
2076 	bp->b_bufobj = NULL;
2077 	BO_UNLOCK(bo);
2078 	vdrop(vp);
2079 }
2080 
2081 /*
2082  * Add an item to the syncer work queue.
2083  */
2084 static void
2085 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2086 {
2087 	int slot;
2088 
2089 	ASSERT_BO_WLOCKED(bo);
2090 
2091 	mtx_lock(&sync_mtx);
2092 	if (bo->bo_flag & BO_ONWORKLST)
2093 		LIST_REMOVE(bo, bo_synclist);
2094 	else {
2095 		bo->bo_flag |= BO_ONWORKLST;
2096 		syncer_worklist_len++;
2097 	}
2098 
2099 	if (delay > syncer_maxdelay - 2)
2100 		delay = syncer_maxdelay - 2;
2101 	slot = (syncer_delayno + delay) & syncer_mask;
2102 
2103 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2104 	mtx_unlock(&sync_mtx);
2105 }
2106 
2107 static int
2108 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2109 {
2110 	int error, len;
2111 
2112 	mtx_lock(&sync_mtx);
2113 	len = syncer_worklist_len - sync_vnode_count;
2114 	mtx_unlock(&sync_mtx);
2115 	error = SYSCTL_OUT(req, &len, sizeof(len));
2116 	return (error);
2117 }
2118 
2119 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
2120     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2121 
2122 static struct proc *updateproc;
2123 static void sched_sync(void);
2124 static struct kproc_desc up_kp = {
2125 	"syncer",
2126 	sched_sync,
2127 	&updateproc
2128 };
2129 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2130 
2131 static int
2132 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2133 {
2134 	struct vnode *vp;
2135 	struct mount *mp;
2136 
2137 	*bo = LIST_FIRST(slp);
2138 	if (*bo == NULL)
2139 		return (0);
2140 	vp = bo2vnode(*bo);
2141 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2142 		return (1);
2143 	/*
2144 	 * We use vhold in case the vnode does not
2145 	 * successfully sync.  vhold prevents the vnode from
2146 	 * going away when we unlock the sync_mtx so that
2147 	 * we can acquire the vnode interlock.
2148 	 */
2149 	vholdl(vp);
2150 	mtx_unlock(&sync_mtx);
2151 	VI_UNLOCK(vp);
2152 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2153 		vdrop(vp);
2154 		mtx_lock(&sync_mtx);
2155 		return (*bo == LIST_FIRST(slp));
2156 	}
2157 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2158 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2159 	VOP_UNLOCK(vp, 0);
2160 	vn_finished_write(mp);
2161 	BO_LOCK(*bo);
2162 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2163 		/*
2164 		 * Put us back on the worklist.  The worklist
2165 		 * routine will remove us from our current
2166 		 * position and then add us back in at a later
2167 		 * position.
2168 		 */
2169 		vn_syncer_add_to_worklist(*bo, syncdelay);
2170 	}
2171 	BO_UNLOCK(*bo);
2172 	vdrop(vp);
2173 	mtx_lock(&sync_mtx);
2174 	return (0);
2175 }
2176 
2177 static int first_printf = 1;
2178 
2179 /*
2180  * System filesystem synchronizer daemon.
2181  */
2182 static void
2183 sched_sync(void)
2184 {
2185 	struct synclist *next, *slp;
2186 	struct bufobj *bo;
2187 	long starttime;
2188 	struct thread *td = curthread;
2189 	int last_work_seen;
2190 	int net_worklist_len;
2191 	int syncer_final_iter;
2192 	int error;
2193 
2194 	last_work_seen = 0;
2195 	syncer_final_iter = 0;
2196 	syncer_state = SYNCER_RUNNING;
2197 	starttime = time_uptime;
2198 	td->td_pflags |= TDP_NORUNNINGBUF;
2199 
2200 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2201 	    SHUTDOWN_PRI_LAST);
2202 
2203 	mtx_lock(&sync_mtx);
2204 	for (;;) {
2205 		if (syncer_state == SYNCER_FINAL_DELAY &&
2206 		    syncer_final_iter == 0) {
2207 			mtx_unlock(&sync_mtx);
2208 			kproc_suspend_check(td->td_proc);
2209 			mtx_lock(&sync_mtx);
2210 		}
2211 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2212 		if (syncer_state != SYNCER_RUNNING &&
2213 		    starttime != time_uptime) {
2214 			if (first_printf) {
2215 				printf("\nSyncing disks, vnodes remaining... ");
2216 				first_printf = 0;
2217 			}
2218 			printf("%d ", net_worklist_len);
2219 		}
2220 		starttime = time_uptime;
2221 
2222 		/*
2223 		 * Push files whose dirty time has expired.  Be careful
2224 		 * of interrupt race on slp queue.
2225 		 *
2226 		 * Skip over empty worklist slots when shutting down.
2227 		 */
2228 		do {
2229 			slp = &syncer_workitem_pending[syncer_delayno];
2230 			syncer_delayno += 1;
2231 			if (syncer_delayno == syncer_maxdelay)
2232 				syncer_delayno = 0;
2233 			next = &syncer_workitem_pending[syncer_delayno];
2234 			/*
2235 			 * If the worklist has wrapped since the
2236 			 * it was emptied of all but syncer vnodes,
2237 			 * switch to the FINAL_DELAY state and run
2238 			 * for one more second.
2239 			 */
2240 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2241 			    net_worklist_len == 0 &&
2242 			    last_work_seen == syncer_delayno) {
2243 				syncer_state = SYNCER_FINAL_DELAY;
2244 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2245 			}
2246 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2247 		    syncer_worklist_len > 0);
2248 
2249 		/*
2250 		 * Keep track of the last time there was anything
2251 		 * on the worklist other than syncer vnodes.
2252 		 * Return to the SHUTTING_DOWN state if any
2253 		 * new work appears.
2254 		 */
2255 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2256 			last_work_seen = syncer_delayno;
2257 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2258 			syncer_state = SYNCER_SHUTTING_DOWN;
2259 		while (!LIST_EMPTY(slp)) {
2260 			error = sync_vnode(slp, &bo, td);
2261 			if (error == 1) {
2262 				LIST_REMOVE(bo, bo_synclist);
2263 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2264 				continue;
2265 			}
2266 
2267 			if (first_printf == 0) {
2268 				/*
2269 				 * Drop the sync mutex, because some watchdog
2270 				 * drivers need to sleep while patting
2271 				 */
2272 				mtx_unlock(&sync_mtx);
2273 				wdog_kern_pat(WD_LASTVAL);
2274 				mtx_lock(&sync_mtx);
2275 			}
2276 
2277 		}
2278 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2279 			syncer_final_iter--;
2280 		/*
2281 		 * The variable rushjob allows the kernel to speed up the
2282 		 * processing of the filesystem syncer process. A rushjob
2283 		 * value of N tells the filesystem syncer to process the next
2284 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2285 		 * is used by the soft update code to speed up the filesystem
2286 		 * syncer process when the incore state is getting so far
2287 		 * ahead of the disk that the kernel memory pool is being
2288 		 * threatened with exhaustion.
2289 		 */
2290 		if (rushjob > 0) {
2291 			rushjob -= 1;
2292 			continue;
2293 		}
2294 		/*
2295 		 * Just sleep for a short period of time between
2296 		 * iterations when shutting down to allow some I/O
2297 		 * to happen.
2298 		 *
2299 		 * If it has taken us less than a second to process the
2300 		 * current work, then wait. Otherwise start right over
2301 		 * again. We can still lose time if any single round
2302 		 * takes more than two seconds, but it does not really
2303 		 * matter as we are just trying to generally pace the
2304 		 * filesystem activity.
2305 		 */
2306 		if (syncer_state != SYNCER_RUNNING ||
2307 		    time_uptime == starttime) {
2308 			thread_lock(td);
2309 			sched_prio(td, PPAUSE);
2310 			thread_unlock(td);
2311 		}
2312 		if (syncer_state != SYNCER_RUNNING)
2313 			cv_timedwait(&sync_wakeup, &sync_mtx,
2314 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2315 		else if (time_uptime == starttime)
2316 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2317 	}
2318 }
2319 
2320 /*
2321  * Request the syncer daemon to speed up its work.
2322  * We never push it to speed up more than half of its
2323  * normal turn time, otherwise it could take over the cpu.
2324  */
2325 int
2326 speedup_syncer(void)
2327 {
2328 	int ret = 0;
2329 
2330 	mtx_lock(&sync_mtx);
2331 	if (rushjob < syncdelay / 2) {
2332 		rushjob += 1;
2333 		stat_rush_requests += 1;
2334 		ret = 1;
2335 	}
2336 	mtx_unlock(&sync_mtx);
2337 	cv_broadcast(&sync_wakeup);
2338 	return (ret);
2339 }
2340 
2341 /*
2342  * Tell the syncer to speed up its work and run though its work
2343  * list several times, then tell it to shut down.
2344  */
2345 static void
2346 syncer_shutdown(void *arg, int howto)
2347 {
2348 
2349 	if (howto & RB_NOSYNC)
2350 		return;
2351 	mtx_lock(&sync_mtx);
2352 	syncer_state = SYNCER_SHUTTING_DOWN;
2353 	rushjob = 0;
2354 	mtx_unlock(&sync_mtx);
2355 	cv_broadcast(&sync_wakeup);
2356 	kproc_shutdown(arg, howto);
2357 }
2358 
2359 void
2360 syncer_suspend(void)
2361 {
2362 
2363 	syncer_shutdown(updateproc, 0);
2364 }
2365 
2366 void
2367 syncer_resume(void)
2368 {
2369 
2370 	mtx_lock(&sync_mtx);
2371 	first_printf = 1;
2372 	syncer_state = SYNCER_RUNNING;
2373 	mtx_unlock(&sync_mtx);
2374 	cv_broadcast(&sync_wakeup);
2375 	kproc_resume(updateproc);
2376 }
2377 
2378 /*
2379  * Reassign a buffer from one vnode to another.
2380  * Used to assign file specific control information
2381  * (indirect blocks) to the vnode to which they belong.
2382  */
2383 void
2384 reassignbuf(struct buf *bp)
2385 {
2386 	struct vnode *vp;
2387 	struct bufobj *bo;
2388 	int delay;
2389 #ifdef INVARIANTS
2390 	struct bufv *bv;
2391 #endif
2392 
2393 	vp = bp->b_vp;
2394 	bo = bp->b_bufobj;
2395 	++reassignbufcalls;
2396 
2397 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2398 	    bp, bp->b_vp, bp->b_flags);
2399 	/*
2400 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2401 	 * is not fully linked in.
2402 	 */
2403 	if (bp->b_flags & B_PAGING)
2404 		panic("cannot reassign paging buffer");
2405 
2406 	/*
2407 	 * Delete from old vnode list, if on one.
2408 	 */
2409 	BO_LOCK(bo);
2410 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2411 		buf_vlist_remove(bp);
2412 	else
2413 		panic("reassignbuf: Buffer %p not on queue.", bp);
2414 	/*
2415 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2416 	 * of clean buffers.
2417 	 */
2418 	if (bp->b_flags & B_DELWRI) {
2419 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2420 			switch (vp->v_type) {
2421 			case VDIR:
2422 				delay = dirdelay;
2423 				break;
2424 			case VCHR:
2425 				delay = metadelay;
2426 				break;
2427 			default:
2428 				delay = filedelay;
2429 			}
2430 			vn_syncer_add_to_worklist(bo, delay);
2431 		}
2432 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2433 	} else {
2434 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2435 
2436 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2437 			mtx_lock(&sync_mtx);
2438 			LIST_REMOVE(bo, bo_synclist);
2439 			syncer_worklist_len--;
2440 			mtx_unlock(&sync_mtx);
2441 			bo->bo_flag &= ~BO_ONWORKLST;
2442 		}
2443 	}
2444 #ifdef INVARIANTS
2445 	bv = &bo->bo_clean;
2446 	bp = TAILQ_FIRST(&bv->bv_hd);
2447 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2448 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2449 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2450 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2451 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2452 	bv = &bo->bo_dirty;
2453 	bp = TAILQ_FIRST(&bv->bv_hd);
2454 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2455 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2456 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2457 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2458 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2459 #endif
2460 	BO_UNLOCK(bo);
2461 }
2462 
2463 /*
2464  * A temporary hack until refcount_* APIs are sorted out.
2465  */
2466 static __inline int
2467 vfs_refcount_acquire_if_not_zero(volatile u_int *count)
2468 {
2469 	u_int old;
2470 
2471 	old = *count;
2472 	for (;;) {
2473 		if (old == 0)
2474 			return (0);
2475 		if (atomic_fcmpset_int(count, &old, old + 1))
2476 			return (1);
2477 	}
2478 }
2479 
2480 static __inline int
2481 vfs_refcount_release_if_not_last(volatile u_int *count)
2482 {
2483 	u_int old;
2484 
2485 	old = *count;
2486 	for (;;) {
2487 		if (old == 1)
2488 			return (0);
2489 		if (atomic_fcmpset_int(count, &old, old - 1))
2490 			return (1);
2491 	}
2492 }
2493 
2494 static void
2495 v_init_counters(struct vnode *vp)
2496 {
2497 
2498 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2499 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2500 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2501 
2502 	refcount_init(&vp->v_holdcnt, 1);
2503 	refcount_init(&vp->v_usecount, 1);
2504 }
2505 
2506 static void
2507 v_incr_usecount_locked(struct vnode *vp)
2508 {
2509 
2510 	ASSERT_VI_LOCKED(vp, __func__);
2511 	if ((vp->v_iflag & VI_OWEINACT) != 0) {
2512 		VNASSERT(vp->v_usecount == 0, vp,
2513 		    ("vnode with usecount and VI_OWEINACT set"));
2514 		vp->v_iflag &= ~VI_OWEINACT;
2515 	}
2516 	refcount_acquire(&vp->v_usecount);
2517 	v_incr_devcount(vp);
2518 }
2519 
2520 /*
2521  * Increment the use count on the vnode, taking care to reference
2522  * the driver's usecount if this is a chardev.
2523  */
2524 static void
2525 v_incr_usecount(struct vnode *vp)
2526 {
2527 
2528 	ASSERT_VI_UNLOCKED(vp, __func__);
2529 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2530 
2531 	if (vp->v_type != VCHR &&
2532 	    vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) {
2533 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2534 		    ("vnode with usecount and VI_OWEINACT set"));
2535 	} else {
2536 		VI_LOCK(vp);
2537 		v_incr_usecount_locked(vp);
2538 		VI_UNLOCK(vp);
2539 	}
2540 }
2541 
2542 /*
2543  * Increment si_usecount of the associated device, if any.
2544  */
2545 static void
2546 v_incr_devcount(struct vnode *vp)
2547 {
2548 
2549 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2550 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2551 		dev_lock();
2552 		vp->v_rdev->si_usecount++;
2553 		dev_unlock();
2554 	}
2555 }
2556 
2557 /*
2558  * Decrement si_usecount of the associated device, if any.
2559  */
2560 static void
2561 v_decr_devcount(struct vnode *vp)
2562 {
2563 
2564 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2565 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2566 		dev_lock();
2567 		vp->v_rdev->si_usecount--;
2568 		dev_unlock();
2569 	}
2570 }
2571 
2572 /*
2573  * Grab a particular vnode from the free list, increment its
2574  * reference count and lock it.  VI_DOOMED is set if the vnode
2575  * is being destroyed.  Only callers who specify LK_RETRY will
2576  * see doomed vnodes.  If inactive processing was delayed in
2577  * vput try to do it here.
2578  *
2579  * Notes on lockless counter manipulation:
2580  * _vhold, vputx and other routines make various decisions based
2581  * on either holdcnt or usecount being 0. As long as either counter
2582  * is not transitioning 0->1 nor 1->0, the manipulation can be done
2583  * with atomic operations. Otherwise the interlock is taken covering
2584  * both the atomic and additional actions.
2585  */
2586 int
2587 vget(struct vnode *vp, int flags, struct thread *td)
2588 {
2589 	int error, oweinact;
2590 
2591 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2592 	    ("vget: invalid lock operation"));
2593 
2594 	if ((flags & LK_INTERLOCK) != 0)
2595 		ASSERT_VI_LOCKED(vp, __func__);
2596 	else
2597 		ASSERT_VI_UNLOCKED(vp, __func__);
2598 	if ((flags & LK_VNHELD) != 0)
2599 		VNASSERT((vp->v_holdcnt > 0), vp,
2600 		    ("vget: LK_VNHELD passed but vnode not held"));
2601 
2602 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2603 
2604 	if ((flags & LK_VNHELD) == 0)
2605 		_vhold(vp, (flags & LK_INTERLOCK) != 0);
2606 
2607 	if ((error = vn_lock(vp, flags)) != 0) {
2608 		vdrop(vp);
2609 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2610 		    vp);
2611 		return (error);
2612 	}
2613 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
2614 		panic("vget: vn_lock failed to return ENOENT\n");
2615 	/*
2616 	 * We don't guarantee that any particular close will
2617 	 * trigger inactive processing so just make a best effort
2618 	 * here at preventing a reference to a removed file.  If
2619 	 * we don't succeed no harm is done.
2620 	 *
2621 	 * Upgrade our holdcnt to a usecount.
2622 	 */
2623 	if (vp->v_type == VCHR ||
2624 	    !vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) {
2625 		VI_LOCK(vp);
2626 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2627 			oweinact = 0;
2628 		} else {
2629 			oweinact = 1;
2630 			vp->v_iflag &= ~VI_OWEINACT;
2631 		}
2632 		refcount_acquire(&vp->v_usecount);
2633 		v_incr_devcount(vp);
2634 		if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2635 		    (flags & LK_NOWAIT) == 0)
2636 			vinactive(vp, td);
2637 		VI_UNLOCK(vp);
2638 	}
2639 	return (0);
2640 }
2641 
2642 /*
2643  * Increase the reference (use) and hold count of a vnode.
2644  * This will also remove the vnode from the free list if it is presently free.
2645  */
2646 void
2647 vref(struct vnode *vp)
2648 {
2649 
2650 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2651 	_vhold(vp, false);
2652 	v_incr_usecount(vp);
2653 }
2654 
2655 void
2656 vrefl(struct vnode *vp)
2657 {
2658 
2659 	ASSERT_VI_LOCKED(vp, __func__);
2660 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2661 	_vhold(vp, true);
2662 	v_incr_usecount_locked(vp);
2663 }
2664 
2665 void
2666 vrefact(struct vnode *vp)
2667 {
2668 
2669 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2670 	if (__predict_false(vp->v_type == VCHR)) {
2671 		VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp,
2672 		    ("%s: wrong ref counts", __func__));
2673 		vref(vp);
2674 		return;
2675 	}
2676 #ifdef INVARIANTS
2677 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
2678 	VNASSERT(old > 0, vp, ("%s: wrong hold count", __func__));
2679 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
2680 	VNASSERT(old > 0, vp, ("%s: wrong use count", __func__));
2681 #else
2682 	refcount_acquire(&vp->v_holdcnt);
2683 	refcount_acquire(&vp->v_usecount);
2684 #endif
2685 }
2686 
2687 /*
2688  * Return reference count of a vnode.
2689  *
2690  * The results of this call are only guaranteed when some mechanism is used to
2691  * stop other processes from gaining references to the vnode.  This may be the
2692  * case if the caller holds the only reference.  This is also useful when stale
2693  * data is acceptable as race conditions may be accounted for by some other
2694  * means.
2695  */
2696 int
2697 vrefcnt(struct vnode *vp)
2698 {
2699 
2700 	return (vp->v_usecount);
2701 }
2702 
2703 #define	VPUTX_VRELE	1
2704 #define	VPUTX_VPUT	2
2705 #define	VPUTX_VUNREF	3
2706 
2707 /*
2708  * Decrement the use and hold counts for a vnode.
2709  *
2710  * See an explanation near vget() as to why atomic operation is safe.
2711  */
2712 static void
2713 vputx(struct vnode *vp, int func)
2714 {
2715 	int error;
2716 
2717 	KASSERT(vp != NULL, ("vputx: null vp"));
2718 	if (func == VPUTX_VUNREF)
2719 		ASSERT_VOP_LOCKED(vp, "vunref");
2720 	else if (func == VPUTX_VPUT)
2721 		ASSERT_VOP_LOCKED(vp, "vput");
2722 	else
2723 		KASSERT(func == VPUTX_VRELE, ("vputx: wrong func"));
2724 	ASSERT_VI_UNLOCKED(vp, __func__);
2725 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2726 
2727 	if (vp->v_type != VCHR &&
2728 	    vfs_refcount_release_if_not_last(&vp->v_usecount)) {
2729 		if (func == VPUTX_VPUT)
2730 			VOP_UNLOCK(vp, 0);
2731 		vdrop(vp);
2732 		return;
2733 	}
2734 
2735 	VI_LOCK(vp);
2736 
2737 	/*
2738 	 * We want to hold the vnode until the inactive finishes to
2739 	 * prevent vgone() races.  We drop the use count here and the
2740 	 * hold count below when we're done.
2741 	 */
2742 	if (!refcount_release(&vp->v_usecount) ||
2743 	    (vp->v_iflag & VI_DOINGINACT)) {
2744 		if (func == VPUTX_VPUT)
2745 			VOP_UNLOCK(vp, 0);
2746 		v_decr_devcount(vp);
2747 		vdropl(vp);
2748 		return;
2749 	}
2750 
2751 	v_decr_devcount(vp);
2752 
2753 	error = 0;
2754 
2755 	if (vp->v_usecount != 0) {
2756 		vn_printf(vp, "vputx: usecount not zero for vnode ");
2757 		panic("vputx: usecount not zero");
2758 	}
2759 
2760 	CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp);
2761 
2762 	/*
2763 	 * We must call VOP_INACTIVE with the node locked. Mark
2764 	 * as VI_DOINGINACT to avoid recursion.
2765 	 */
2766 	vp->v_iflag |= VI_OWEINACT;
2767 	switch (func) {
2768 	case VPUTX_VRELE:
2769 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
2770 		VI_LOCK(vp);
2771 		break;
2772 	case VPUTX_VPUT:
2773 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2774 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
2775 			    LK_NOWAIT);
2776 			VI_LOCK(vp);
2777 		}
2778 		break;
2779 	case VPUTX_VUNREF:
2780 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2781 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
2782 			VI_LOCK(vp);
2783 		}
2784 		break;
2785 	}
2786 	VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
2787 	    ("vnode with usecount and VI_OWEINACT set"));
2788 	if (error == 0) {
2789 		if (vp->v_iflag & VI_OWEINACT)
2790 			vinactive(vp, curthread);
2791 		if (func != VPUTX_VUNREF)
2792 			VOP_UNLOCK(vp, 0);
2793 	}
2794 	vdropl(vp);
2795 }
2796 
2797 /*
2798  * Vnode put/release.
2799  * If count drops to zero, call inactive routine and return to freelist.
2800  */
2801 void
2802 vrele(struct vnode *vp)
2803 {
2804 
2805 	vputx(vp, VPUTX_VRELE);
2806 }
2807 
2808 /*
2809  * Release an already locked vnode.  This give the same effects as
2810  * unlock+vrele(), but takes less time and avoids releasing and
2811  * re-aquiring the lock (as vrele() acquires the lock internally.)
2812  */
2813 void
2814 vput(struct vnode *vp)
2815 {
2816 
2817 	vputx(vp, VPUTX_VPUT);
2818 }
2819 
2820 /*
2821  * Release an exclusively locked vnode. Do not unlock the vnode lock.
2822  */
2823 void
2824 vunref(struct vnode *vp)
2825 {
2826 
2827 	vputx(vp, VPUTX_VUNREF);
2828 }
2829 
2830 /*
2831  * Increase the hold count and activate if this is the first reference.
2832  */
2833 void
2834 _vhold(struct vnode *vp, bool locked)
2835 {
2836 	struct mount *mp;
2837 
2838 	if (locked)
2839 		ASSERT_VI_LOCKED(vp, __func__);
2840 	else
2841 		ASSERT_VI_UNLOCKED(vp, __func__);
2842 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2843 	if (!locked) {
2844 		if (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 		VI_LOCK(vp);
2850 	}
2851 	if ((vp->v_iflag & VI_FREE) == 0) {
2852 		refcount_acquire(&vp->v_holdcnt);
2853 		if (!locked)
2854 			VI_UNLOCK(vp);
2855 		return;
2856 	}
2857 	VNASSERT(vp->v_holdcnt == 0, vp,
2858 	    ("%s: wrong hold count", __func__));
2859 	VNASSERT(vp->v_op != NULL, vp,
2860 	    ("%s: vnode already reclaimed.", __func__));
2861 	/*
2862 	 * Remove a vnode from the free list, mark it as in use,
2863 	 * and put it on the active list.
2864 	 */
2865 	VNASSERT(vp->v_mount != NULL, vp,
2866 	    ("_vhold: vnode not on per mount vnode list"));
2867 	mp = vp->v_mount;
2868 	mtx_lock(&mp->mnt_listmtx);
2869 	if ((vp->v_mflag & VMP_TMPMNTFREELIST) != 0) {
2870 		TAILQ_REMOVE(&mp->mnt_tmpfreevnodelist, vp, v_actfreelist);
2871 		mp->mnt_tmpfreevnodelistsize--;
2872 		vp->v_mflag &= ~VMP_TMPMNTFREELIST;
2873 	} else {
2874 		mtx_lock(&vnode_free_list_mtx);
2875 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
2876 		freevnodes--;
2877 		mtx_unlock(&vnode_free_list_mtx);
2878 	}
2879 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
2880 	    ("Activating already active vnode"));
2881 	vp->v_iflag &= ~VI_FREE;
2882 	vp->v_iflag |= VI_ACTIVE;
2883 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
2884 	mp->mnt_activevnodelistsize++;
2885 	mtx_unlock(&mp->mnt_listmtx);
2886 	refcount_acquire(&vp->v_holdcnt);
2887 	if (!locked)
2888 		VI_UNLOCK(vp);
2889 }
2890 
2891 /*
2892  * Drop the hold count of the vnode.  If this is the last reference to
2893  * the vnode we place it on the free list unless it has been vgone'd
2894  * (marked VI_DOOMED) in which case we will free it.
2895  *
2896  * Because the vnode vm object keeps a hold reference on the vnode if
2897  * there is at least one resident non-cached page, the vnode cannot
2898  * leave the active list without the page cleanup done.
2899  */
2900 void
2901 _vdrop(struct vnode *vp, bool locked)
2902 {
2903 	struct bufobj *bo;
2904 	struct mount *mp;
2905 	int active;
2906 
2907 	if (locked)
2908 		ASSERT_VI_LOCKED(vp, __func__);
2909 	else
2910 		ASSERT_VI_UNLOCKED(vp, __func__);
2911 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2912 	if ((int)vp->v_holdcnt <= 0)
2913 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
2914 	if (!locked) {
2915 		if (vfs_refcount_release_if_not_last(&vp->v_holdcnt))
2916 			return;
2917 		VI_LOCK(vp);
2918 	}
2919 	if (refcount_release(&vp->v_holdcnt) == 0) {
2920 		VI_UNLOCK(vp);
2921 		return;
2922 	}
2923 	if ((vp->v_iflag & VI_DOOMED) == 0) {
2924 		/*
2925 		 * Mark a vnode as free: remove it from its active list
2926 		 * and put it up for recycling on the freelist.
2927 		 */
2928 		VNASSERT(vp->v_op != NULL, vp,
2929 		    ("vdropl: vnode already reclaimed."));
2930 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2931 		    ("vnode already free"));
2932 		VNASSERT(vp->v_holdcnt == 0, vp,
2933 		    ("vdropl: freeing when we shouldn't"));
2934 		active = vp->v_iflag & VI_ACTIVE;
2935 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2936 			vp->v_iflag &= ~VI_ACTIVE;
2937 			mp = vp->v_mount;
2938 			if (mp != NULL) {
2939 				mtx_lock(&mp->mnt_listmtx);
2940 				if (active) {
2941 					TAILQ_REMOVE(&mp->mnt_activevnodelist,
2942 					    vp, v_actfreelist);
2943 					mp->mnt_activevnodelistsize--;
2944 				}
2945 				TAILQ_INSERT_TAIL(&mp->mnt_tmpfreevnodelist,
2946 				    vp, v_actfreelist);
2947 				mp->mnt_tmpfreevnodelistsize++;
2948 				vp->v_iflag |= VI_FREE;
2949 				vp->v_mflag |= VMP_TMPMNTFREELIST;
2950 				VI_UNLOCK(vp);
2951 				if (mp->mnt_tmpfreevnodelistsize >=
2952 				    mnt_free_list_batch)
2953 					vnlru_return_batch_locked(mp);
2954 				mtx_unlock(&mp->mnt_listmtx);
2955 			} else {
2956 				VNASSERT(active == 0, vp,
2957 				    ("vdropl: active vnode not on per mount "
2958 				    "vnode list"));
2959 				mtx_lock(&vnode_free_list_mtx);
2960 				TAILQ_INSERT_TAIL(&vnode_free_list, vp,
2961 				    v_actfreelist);
2962 				freevnodes++;
2963 				vp->v_iflag |= VI_FREE;
2964 				VI_UNLOCK(vp);
2965 				mtx_unlock(&vnode_free_list_mtx);
2966 			}
2967 		} else {
2968 			VI_UNLOCK(vp);
2969 			counter_u64_add(free_owe_inact, 1);
2970 		}
2971 		return;
2972 	}
2973 	/*
2974 	 * The vnode has been marked for destruction, so free it.
2975 	 *
2976 	 * The vnode will be returned to the zone where it will
2977 	 * normally remain until it is needed for another vnode. We
2978 	 * need to cleanup (or verify that the cleanup has already
2979 	 * been done) any residual data left from its current use
2980 	 * so as not to contaminate the freshly allocated vnode.
2981 	 */
2982 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
2983 	atomic_subtract_long(&numvnodes, 1);
2984 	bo = &vp->v_bufobj;
2985 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2986 	    ("cleaned vnode still on the free list."));
2987 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
2988 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
2989 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
2990 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
2991 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
2992 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
2993 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
2994 	    ("clean blk trie not empty"));
2995 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
2996 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
2997 	    ("dirty blk trie not empty"));
2998 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
2999 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
3000 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
3001 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
3002 	    ("Dangling rangelock waiters"));
3003 	VI_UNLOCK(vp);
3004 #ifdef MAC
3005 	mac_vnode_destroy(vp);
3006 #endif
3007 	if (vp->v_pollinfo != NULL) {
3008 		destroy_vpollinfo(vp->v_pollinfo);
3009 		vp->v_pollinfo = NULL;
3010 	}
3011 #ifdef INVARIANTS
3012 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
3013 	vp->v_op = NULL;
3014 #endif
3015 	vp->v_mountedhere = NULL;
3016 	vp->v_unpcb = NULL;
3017 	vp->v_rdev = NULL;
3018 	vp->v_fifoinfo = NULL;
3019 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
3020 	vp->v_iflag = 0;
3021 	vp->v_vflag = 0;
3022 	bo->bo_flag = 0;
3023 	uma_zfree(vnode_zone, vp);
3024 }
3025 
3026 /*
3027  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3028  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
3029  * OWEINACT tracks whether a vnode missed a call to inactive due to a
3030  * failed lock upgrade.
3031  */
3032 void
3033 vinactive(struct vnode *vp, struct thread *td)
3034 {
3035 	struct vm_object *obj;
3036 
3037 	ASSERT_VOP_ELOCKED(vp, "vinactive");
3038 	ASSERT_VI_LOCKED(vp, "vinactive");
3039 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3040 	    ("vinactive: recursed on VI_DOINGINACT"));
3041 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3042 	vp->v_iflag |= VI_DOINGINACT;
3043 	vp->v_iflag &= ~VI_OWEINACT;
3044 	VI_UNLOCK(vp);
3045 	/*
3046 	 * Before moving off the active list, we must be sure that any
3047 	 * modified pages are converted into the vnode's dirty
3048 	 * buffers, since these will no longer be checked once the
3049 	 * vnode is on the inactive list.
3050 	 *
3051 	 * The write-out of the dirty pages is asynchronous.  At the
3052 	 * point that VOP_INACTIVE() is called, there could still be
3053 	 * pending I/O and dirty pages in the object.
3054 	 */
3055 	if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3056 	    (obj->flags & OBJ_MIGHTBEDIRTY) != 0) {
3057 		VM_OBJECT_WLOCK(obj);
3058 		vm_object_page_clean(obj, 0, 0, 0);
3059 		VM_OBJECT_WUNLOCK(obj);
3060 	}
3061 	VOP_INACTIVE(vp, td);
3062 	VI_LOCK(vp);
3063 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3064 	    ("vinactive: lost VI_DOINGINACT"));
3065 	vp->v_iflag &= ~VI_DOINGINACT;
3066 }
3067 
3068 /*
3069  * Remove any vnodes in the vnode table belonging to mount point mp.
3070  *
3071  * If FORCECLOSE is not specified, there should not be any active ones,
3072  * return error if any are found (nb: this is a user error, not a
3073  * system error). If FORCECLOSE is specified, detach any active vnodes
3074  * that are found.
3075  *
3076  * If WRITECLOSE is set, only flush out regular file vnodes open for
3077  * writing.
3078  *
3079  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3080  *
3081  * `rootrefs' specifies the base reference count for the root vnode
3082  * of this filesystem. The root vnode is considered busy if its
3083  * v_usecount exceeds this value. On a successful return, vflush(, td)
3084  * will call vrele() on the root vnode exactly rootrefs times.
3085  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3086  * be zero.
3087  */
3088 #ifdef DIAGNOSTIC
3089 static int busyprt = 0;		/* print out busy vnodes */
3090 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3091 #endif
3092 
3093 int
3094 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3095 {
3096 	struct vnode *vp, *mvp, *rootvp = NULL;
3097 	struct vattr vattr;
3098 	int busy = 0, error;
3099 
3100 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3101 	    rootrefs, flags);
3102 	if (rootrefs > 0) {
3103 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3104 		    ("vflush: bad args"));
3105 		/*
3106 		 * Get the filesystem root vnode. We can vput() it
3107 		 * immediately, since with rootrefs > 0, it won't go away.
3108 		 */
3109 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3110 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3111 			    __func__, error);
3112 			return (error);
3113 		}
3114 		vput(rootvp);
3115 	}
3116 loop:
3117 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3118 		vholdl(vp);
3119 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3120 		if (error) {
3121 			vdrop(vp);
3122 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3123 			goto loop;
3124 		}
3125 		/*
3126 		 * Skip over a vnodes marked VV_SYSTEM.
3127 		 */
3128 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3129 			VOP_UNLOCK(vp, 0);
3130 			vdrop(vp);
3131 			continue;
3132 		}
3133 		/*
3134 		 * If WRITECLOSE is set, flush out unlinked but still open
3135 		 * files (even if open only for reading) and regular file
3136 		 * vnodes open for writing.
3137 		 */
3138 		if (flags & WRITECLOSE) {
3139 			if (vp->v_object != NULL) {
3140 				VM_OBJECT_WLOCK(vp->v_object);
3141 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3142 				VM_OBJECT_WUNLOCK(vp->v_object);
3143 			}
3144 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3145 			if (error != 0) {
3146 				VOP_UNLOCK(vp, 0);
3147 				vdrop(vp);
3148 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3149 				return (error);
3150 			}
3151 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3152 			VI_LOCK(vp);
3153 
3154 			if ((vp->v_type == VNON ||
3155 			    (error == 0 && vattr.va_nlink > 0)) &&
3156 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
3157 				VOP_UNLOCK(vp, 0);
3158 				vdropl(vp);
3159 				continue;
3160 			}
3161 		} else
3162 			VI_LOCK(vp);
3163 		/*
3164 		 * With v_usecount == 0, all we need to do is clear out the
3165 		 * vnode data structures and we are done.
3166 		 *
3167 		 * If FORCECLOSE is set, forcibly close the vnode.
3168 		 */
3169 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3170 			vgonel(vp);
3171 		} else {
3172 			busy++;
3173 #ifdef DIAGNOSTIC
3174 			if (busyprt)
3175 				vn_printf(vp, "vflush: busy vnode ");
3176 #endif
3177 		}
3178 		VOP_UNLOCK(vp, 0);
3179 		vdropl(vp);
3180 	}
3181 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3182 		/*
3183 		 * If just the root vnode is busy, and if its refcount
3184 		 * is equal to `rootrefs', then go ahead and kill it.
3185 		 */
3186 		VI_LOCK(rootvp);
3187 		KASSERT(busy > 0, ("vflush: not busy"));
3188 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3189 		    ("vflush: usecount %d < rootrefs %d",
3190 		     rootvp->v_usecount, rootrefs));
3191 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3192 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3193 			vgone(rootvp);
3194 			VOP_UNLOCK(rootvp, 0);
3195 			busy = 0;
3196 		} else
3197 			VI_UNLOCK(rootvp);
3198 	}
3199 	if (busy) {
3200 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3201 		    busy);
3202 		return (EBUSY);
3203 	}
3204 	for (; rootrefs > 0; rootrefs--)
3205 		vrele(rootvp);
3206 	return (0);
3207 }
3208 
3209 /*
3210  * Recycle an unused vnode to the front of the free list.
3211  */
3212 int
3213 vrecycle(struct vnode *vp)
3214 {
3215 	int recycled;
3216 
3217 	VI_LOCK(vp);
3218 	recycled = vrecyclel(vp);
3219 	VI_UNLOCK(vp);
3220 	return (recycled);
3221 }
3222 
3223 /*
3224  * vrecycle, with the vp interlock held.
3225  */
3226 int
3227 vrecyclel(struct vnode *vp)
3228 {
3229 	int recycled;
3230 
3231 	ASSERT_VOP_ELOCKED(vp, __func__);
3232 	ASSERT_VI_LOCKED(vp, __func__);
3233 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3234 	recycled = 0;
3235 	if (vp->v_usecount == 0) {
3236 		recycled = 1;
3237 		vgonel(vp);
3238 	}
3239 	return (recycled);
3240 }
3241 
3242 /*
3243  * Eliminate all activity associated with a vnode
3244  * in preparation for reuse.
3245  */
3246 void
3247 vgone(struct vnode *vp)
3248 {
3249 	VI_LOCK(vp);
3250 	vgonel(vp);
3251 	VI_UNLOCK(vp);
3252 }
3253 
3254 static void
3255 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3256     struct vnode *lowervp __unused)
3257 {
3258 }
3259 
3260 /*
3261  * Notify upper mounts about reclaimed or unlinked vnode.
3262  */
3263 void
3264 vfs_notify_upper(struct vnode *vp, int event)
3265 {
3266 	static struct vfsops vgonel_vfsops = {
3267 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3268 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3269 	};
3270 	struct mount *mp, *ump, *mmp;
3271 
3272 	mp = vp->v_mount;
3273 	if (mp == NULL)
3274 		return;
3275 
3276 	MNT_ILOCK(mp);
3277 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3278 		goto unlock;
3279 	MNT_IUNLOCK(mp);
3280 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3281 	mmp->mnt_op = &vgonel_vfsops;
3282 	mmp->mnt_kern_flag |= MNTK_MARKER;
3283 	MNT_ILOCK(mp);
3284 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3285 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3286 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3287 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3288 			continue;
3289 		}
3290 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3291 		MNT_IUNLOCK(mp);
3292 		switch (event) {
3293 		case VFS_NOTIFY_UPPER_RECLAIM:
3294 			VFS_RECLAIM_LOWERVP(ump, vp);
3295 			break;
3296 		case VFS_NOTIFY_UPPER_UNLINK:
3297 			VFS_UNLINK_LOWERVP(ump, vp);
3298 			break;
3299 		default:
3300 			KASSERT(0, ("invalid event %d", event));
3301 			break;
3302 		}
3303 		MNT_ILOCK(mp);
3304 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3305 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3306 	}
3307 	free(mmp, M_TEMP);
3308 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3309 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3310 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3311 		wakeup(&mp->mnt_uppers);
3312 	}
3313 unlock:
3314 	MNT_IUNLOCK(mp);
3315 }
3316 
3317 /*
3318  * vgone, with the vp interlock held.
3319  */
3320 static void
3321 vgonel(struct vnode *vp)
3322 {
3323 	struct thread *td;
3324 	int oweinact;
3325 	int active;
3326 	struct mount *mp;
3327 
3328 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3329 	ASSERT_VI_LOCKED(vp, "vgonel");
3330 	VNASSERT(vp->v_holdcnt, vp,
3331 	    ("vgonel: vp %p has no reference.", vp));
3332 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3333 	td = curthread;
3334 
3335 	/*
3336 	 * Don't vgonel if we're already doomed.
3337 	 */
3338 	if (vp->v_iflag & VI_DOOMED)
3339 		return;
3340 	vp->v_iflag |= VI_DOOMED;
3341 
3342 	/*
3343 	 * Check to see if the vnode is in use.  If so, we have to call
3344 	 * VOP_CLOSE() and VOP_INACTIVE().
3345 	 */
3346 	active = vp->v_usecount;
3347 	oweinact = (vp->v_iflag & VI_OWEINACT);
3348 	VI_UNLOCK(vp);
3349 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3350 
3351 	/*
3352 	 * If purging an active vnode, it must be closed and
3353 	 * deactivated before being reclaimed.
3354 	 */
3355 	if (active)
3356 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3357 	if (oweinact || active) {
3358 		VI_LOCK(vp);
3359 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
3360 			vinactive(vp, td);
3361 		VI_UNLOCK(vp);
3362 	}
3363 	if (vp->v_type == VSOCK)
3364 		vfs_unp_reclaim(vp);
3365 
3366 	/*
3367 	 * Clean out any buffers associated with the vnode.
3368 	 * If the flush fails, just toss the buffers.
3369 	 */
3370 	mp = NULL;
3371 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3372 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3373 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3374 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3375 			;
3376 	}
3377 
3378 	BO_LOCK(&vp->v_bufobj);
3379 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3380 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3381 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3382 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3383 	    ("vp %p bufobj not invalidated", vp));
3384 
3385 	/*
3386 	 * For VMIO bufobj, BO_DEAD is set in vm_object_terminate()
3387 	 * after the object's page queue is flushed.
3388 	 */
3389 	if (vp->v_bufobj.bo_object == NULL)
3390 		vp->v_bufobj.bo_flag |= BO_DEAD;
3391 	BO_UNLOCK(&vp->v_bufobj);
3392 
3393 	/*
3394 	 * Reclaim the vnode.
3395 	 */
3396 	if (VOP_RECLAIM(vp, td))
3397 		panic("vgone: cannot reclaim");
3398 	if (mp != NULL)
3399 		vn_finished_secondary_write(mp);
3400 	VNASSERT(vp->v_object == NULL, vp,
3401 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
3402 	/*
3403 	 * Clear the advisory locks and wake up waiting threads.
3404 	 */
3405 	(void)VOP_ADVLOCKPURGE(vp);
3406 	vp->v_lockf = NULL;
3407 	/*
3408 	 * Delete from old mount point vnode list.
3409 	 */
3410 	delmntque(vp);
3411 	cache_purge(vp);
3412 	/*
3413 	 * Done with purge, reset to the standard lock and invalidate
3414 	 * the vnode.
3415 	 */
3416 	VI_LOCK(vp);
3417 	vp->v_vnlock = &vp->v_lock;
3418 	vp->v_op = &dead_vnodeops;
3419 	vp->v_tag = "none";
3420 	vp->v_type = VBAD;
3421 }
3422 
3423 /*
3424  * Calculate the total number of references to a special device.
3425  */
3426 int
3427 vcount(struct vnode *vp)
3428 {
3429 	int count;
3430 
3431 	dev_lock();
3432 	count = vp->v_rdev->si_usecount;
3433 	dev_unlock();
3434 	return (count);
3435 }
3436 
3437 /*
3438  * Same as above, but using the struct cdev *as argument
3439  */
3440 int
3441 count_dev(struct cdev *dev)
3442 {
3443 	int count;
3444 
3445 	dev_lock();
3446 	count = dev->si_usecount;
3447 	dev_unlock();
3448 	return(count);
3449 }
3450 
3451 /*
3452  * Print out a description of a vnode.
3453  */
3454 static char *typename[] =
3455 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3456  "VMARKER"};
3457 
3458 void
3459 vn_printf(struct vnode *vp, const char *fmt, ...)
3460 {
3461 	va_list ap;
3462 	char buf[256], buf2[16];
3463 	u_long flags;
3464 
3465 	va_start(ap, fmt);
3466 	vprintf(fmt, ap);
3467 	va_end(ap);
3468 	printf("%p: ", (void *)vp);
3469 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
3470 	printf("    usecount %d, writecount %d, refcount %d",
3471 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
3472 	switch (vp->v_type) {
3473 	case VDIR:
3474 		printf(" mountedhere %p\n", vp->v_mountedhere);
3475 		break;
3476 	case VCHR:
3477 		printf(" rdev %p\n", vp->v_rdev);
3478 		break;
3479 	case VSOCK:
3480 		printf(" socket %p\n", vp->v_unpcb);
3481 		break;
3482 	case VFIFO:
3483 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
3484 		break;
3485 	default:
3486 		printf("\n");
3487 		break;
3488 	}
3489 	buf[0] = '\0';
3490 	buf[1] = '\0';
3491 	if (vp->v_vflag & VV_ROOT)
3492 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3493 	if (vp->v_vflag & VV_ISTTY)
3494 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3495 	if (vp->v_vflag & VV_NOSYNC)
3496 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3497 	if (vp->v_vflag & VV_ETERNALDEV)
3498 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3499 	if (vp->v_vflag & VV_CACHEDLABEL)
3500 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3501 	if (vp->v_vflag & VV_TEXT)
3502 		strlcat(buf, "|VV_TEXT", sizeof(buf));
3503 	if (vp->v_vflag & VV_COPYONWRITE)
3504 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3505 	if (vp->v_vflag & VV_SYSTEM)
3506 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3507 	if (vp->v_vflag & VV_PROCDEP)
3508 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3509 	if (vp->v_vflag & VV_NOKNOTE)
3510 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3511 	if (vp->v_vflag & VV_DELETED)
3512 		strlcat(buf, "|VV_DELETED", sizeof(buf));
3513 	if (vp->v_vflag & VV_MD)
3514 		strlcat(buf, "|VV_MD", sizeof(buf));
3515 	if (vp->v_vflag & VV_FORCEINSMQ)
3516 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
3517 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
3518 	    VV_CACHEDLABEL | VV_TEXT | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
3519 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
3520 	if (flags != 0) {
3521 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
3522 		strlcat(buf, buf2, sizeof(buf));
3523 	}
3524 	if (vp->v_iflag & VI_MOUNT)
3525 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
3526 	if (vp->v_iflag & VI_DOOMED)
3527 		strlcat(buf, "|VI_DOOMED", sizeof(buf));
3528 	if (vp->v_iflag & VI_FREE)
3529 		strlcat(buf, "|VI_FREE", sizeof(buf));
3530 	if (vp->v_iflag & VI_ACTIVE)
3531 		strlcat(buf, "|VI_ACTIVE", sizeof(buf));
3532 	if (vp->v_iflag & VI_DOINGINACT)
3533 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
3534 	if (vp->v_iflag & VI_OWEINACT)
3535 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
3536 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOOMED | VI_FREE |
3537 	    VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT);
3538 	if (flags != 0) {
3539 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
3540 		strlcat(buf, buf2, sizeof(buf));
3541 	}
3542 	printf("    flags (%s)\n", buf + 1);
3543 	if (mtx_owned(VI_MTX(vp)))
3544 		printf(" VI_LOCKed");
3545 	if (vp->v_object != NULL)
3546 		printf("    v_object %p ref %d pages %d "
3547 		    "cleanbuf %d dirtybuf %d\n",
3548 		    vp->v_object, vp->v_object->ref_count,
3549 		    vp->v_object->resident_page_count,
3550 		    vp->v_bufobj.bo_clean.bv_cnt,
3551 		    vp->v_bufobj.bo_dirty.bv_cnt);
3552 	printf("    ");
3553 	lockmgr_printinfo(vp->v_vnlock);
3554 	if (vp->v_data != NULL)
3555 		VOP_PRINT(vp);
3556 }
3557 
3558 #ifdef DDB
3559 /*
3560  * List all of the locked vnodes in the system.
3561  * Called when debugging the kernel.
3562  */
3563 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
3564 {
3565 	struct mount *mp;
3566 	struct vnode *vp;
3567 
3568 	/*
3569 	 * Note: because this is DDB, we can't obey the locking semantics
3570 	 * for these structures, which means we could catch an inconsistent
3571 	 * state and dereference a nasty pointer.  Not much to be done
3572 	 * about that.
3573 	 */
3574 	db_printf("Locked vnodes\n");
3575 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3576 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3577 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
3578 				vn_printf(vp, "vnode ");
3579 		}
3580 	}
3581 }
3582 
3583 /*
3584  * Show details about the given vnode.
3585  */
3586 DB_SHOW_COMMAND(vnode, db_show_vnode)
3587 {
3588 	struct vnode *vp;
3589 
3590 	if (!have_addr)
3591 		return;
3592 	vp = (struct vnode *)addr;
3593 	vn_printf(vp, "vnode ");
3594 }
3595 
3596 /*
3597  * Show details about the given mount point.
3598  */
3599 DB_SHOW_COMMAND(mount, db_show_mount)
3600 {
3601 	struct mount *mp;
3602 	struct vfsopt *opt;
3603 	struct statfs *sp;
3604 	struct vnode *vp;
3605 	char buf[512];
3606 	uint64_t mflags;
3607 	u_int flags;
3608 
3609 	if (!have_addr) {
3610 		/* No address given, print short info about all mount points. */
3611 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3612 			db_printf("%p %s on %s (%s)\n", mp,
3613 			    mp->mnt_stat.f_mntfromname,
3614 			    mp->mnt_stat.f_mntonname,
3615 			    mp->mnt_stat.f_fstypename);
3616 			if (db_pager_quit)
3617 				break;
3618 		}
3619 		db_printf("\nMore info: show mount <addr>\n");
3620 		return;
3621 	}
3622 
3623 	mp = (struct mount *)addr;
3624 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
3625 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
3626 
3627 	buf[0] = '\0';
3628 	mflags = mp->mnt_flag;
3629 #define	MNT_FLAG(flag)	do {						\
3630 	if (mflags & (flag)) {						\
3631 		if (buf[0] != '\0')					\
3632 			strlcat(buf, ", ", sizeof(buf));		\
3633 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
3634 		mflags &= ~(flag);					\
3635 	}								\
3636 } while (0)
3637 	MNT_FLAG(MNT_RDONLY);
3638 	MNT_FLAG(MNT_SYNCHRONOUS);
3639 	MNT_FLAG(MNT_NOEXEC);
3640 	MNT_FLAG(MNT_NOSUID);
3641 	MNT_FLAG(MNT_NFS4ACLS);
3642 	MNT_FLAG(MNT_UNION);
3643 	MNT_FLAG(MNT_ASYNC);
3644 	MNT_FLAG(MNT_SUIDDIR);
3645 	MNT_FLAG(MNT_SOFTDEP);
3646 	MNT_FLAG(MNT_NOSYMFOLLOW);
3647 	MNT_FLAG(MNT_GJOURNAL);
3648 	MNT_FLAG(MNT_MULTILABEL);
3649 	MNT_FLAG(MNT_ACLS);
3650 	MNT_FLAG(MNT_NOATIME);
3651 	MNT_FLAG(MNT_NOCLUSTERR);
3652 	MNT_FLAG(MNT_NOCLUSTERW);
3653 	MNT_FLAG(MNT_SUJ);
3654 	MNT_FLAG(MNT_EXRDONLY);
3655 	MNT_FLAG(MNT_EXPORTED);
3656 	MNT_FLAG(MNT_DEFEXPORTED);
3657 	MNT_FLAG(MNT_EXPORTANON);
3658 	MNT_FLAG(MNT_EXKERB);
3659 	MNT_FLAG(MNT_EXPUBLIC);
3660 	MNT_FLAG(MNT_LOCAL);
3661 	MNT_FLAG(MNT_QUOTA);
3662 	MNT_FLAG(MNT_ROOTFS);
3663 	MNT_FLAG(MNT_USER);
3664 	MNT_FLAG(MNT_IGNORE);
3665 	MNT_FLAG(MNT_UPDATE);
3666 	MNT_FLAG(MNT_DELEXPORT);
3667 	MNT_FLAG(MNT_RELOAD);
3668 	MNT_FLAG(MNT_FORCE);
3669 	MNT_FLAG(MNT_SNAPSHOT);
3670 	MNT_FLAG(MNT_BYFSID);
3671 #undef MNT_FLAG
3672 	if (mflags != 0) {
3673 		if (buf[0] != '\0')
3674 			strlcat(buf, ", ", sizeof(buf));
3675 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3676 		    "0x%016jx", mflags);
3677 	}
3678 	db_printf("    mnt_flag = %s\n", buf);
3679 
3680 	buf[0] = '\0';
3681 	flags = mp->mnt_kern_flag;
3682 #define	MNT_KERN_FLAG(flag)	do {					\
3683 	if (flags & (flag)) {						\
3684 		if (buf[0] != '\0')					\
3685 			strlcat(buf, ", ", sizeof(buf));		\
3686 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
3687 		flags &= ~(flag);					\
3688 	}								\
3689 } while (0)
3690 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
3691 	MNT_KERN_FLAG(MNTK_ASYNC);
3692 	MNT_KERN_FLAG(MNTK_SOFTDEP);
3693 	MNT_KERN_FLAG(MNTK_NOINSMNTQ);
3694 	MNT_KERN_FLAG(MNTK_DRAINING);
3695 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
3696 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
3697 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
3698 	MNT_KERN_FLAG(MNTK_NO_IOPF);
3699 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
3700 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
3701 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
3702 	MNT_KERN_FLAG(MNTK_MARKER);
3703 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
3704 	MNT_KERN_FLAG(MNTK_NOASYNC);
3705 	MNT_KERN_FLAG(MNTK_UNMOUNT);
3706 	MNT_KERN_FLAG(MNTK_MWAIT);
3707 	MNT_KERN_FLAG(MNTK_SUSPEND);
3708 	MNT_KERN_FLAG(MNTK_SUSPEND2);
3709 	MNT_KERN_FLAG(MNTK_SUSPENDED);
3710 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
3711 	MNT_KERN_FLAG(MNTK_NOKNOTE);
3712 #undef MNT_KERN_FLAG
3713 	if (flags != 0) {
3714 		if (buf[0] != '\0')
3715 			strlcat(buf, ", ", sizeof(buf));
3716 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3717 		    "0x%08x", flags);
3718 	}
3719 	db_printf("    mnt_kern_flag = %s\n", buf);
3720 
3721 	db_printf("    mnt_opt = ");
3722 	opt = TAILQ_FIRST(mp->mnt_opt);
3723 	if (opt != NULL) {
3724 		db_printf("%s", opt->name);
3725 		opt = TAILQ_NEXT(opt, link);
3726 		while (opt != NULL) {
3727 			db_printf(", %s", opt->name);
3728 			opt = TAILQ_NEXT(opt, link);
3729 		}
3730 	}
3731 	db_printf("\n");
3732 
3733 	sp = &mp->mnt_stat;
3734 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
3735 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
3736 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
3737 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
3738 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
3739 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
3740 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
3741 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
3742 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
3743 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
3744 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
3745 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
3746 
3747 	db_printf("    mnt_cred = { uid=%u ruid=%u",
3748 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
3749 	if (jailed(mp->mnt_cred))
3750 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
3751 	db_printf(" }\n");
3752 	db_printf("    mnt_ref = %d\n", mp->mnt_ref);
3753 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
3754 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
3755 	db_printf("    mnt_activevnodelistsize = %d\n",
3756 	    mp->mnt_activevnodelistsize);
3757 	db_printf("    mnt_writeopcount = %d\n", mp->mnt_writeopcount);
3758 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
3759 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
3760 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
3761 	db_printf("    mnt_lockref = %d\n", mp->mnt_lockref);
3762 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
3763 	db_printf("    mnt_secondary_accwrites = %d\n",
3764 	    mp->mnt_secondary_accwrites);
3765 	db_printf("    mnt_gjprovider = %s\n",
3766 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
3767 
3768 	db_printf("\n\nList of active vnodes\n");
3769 	TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) {
3770 		if (vp->v_type != VMARKER) {
3771 			vn_printf(vp, "vnode ");
3772 			if (db_pager_quit)
3773 				break;
3774 		}
3775 	}
3776 	db_printf("\n\nList of inactive vnodes\n");
3777 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3778 		if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) {
3779 			vn_printf(vp, "vnode ");
3780 			if (db_pager_quit)
3781 				break;
3782 		}
3783 	}
3784 }
3785 #endif	/* DDB */
3786 
3787 /*
3788  * Fill in a struct xvfsconf based on a struct vfsconf.
3789  */
3790 static int
3791 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
3792 {
3793 	struct xvfsconf xvfsp;
3794 
3795 	bzero(&xvfsp, sizeof(xvfsp));
3796 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3797 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3798 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3799 	xvfsp.vfc_flags = vfsp->vfc_flags;
3800 	/*
3801 	 * These are unused in userland, we keep them
3802 	 * to not break binary compatibility.
3803 	 */
3804 	xvfsp.vfc_vfsops = NULL;
3805 	xvfsp.vfc_next = NULL;
3806 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3807 }
3808 
3809 #ifdef COMPAT_FREEBSD32
3810 struct xvfsconf32 {
3811 	uint32_t	vfc_vfsops;
3812 	char		vfc_name[MFSNAMELEN];
3813 	int32_t		vfc_typenum;
3814 	int32_t		vfc_refcount;
3815 	int32_t		vfc_flags;
3816 	uint32_t	vfc_next;
3817 };
3818 
3819 static int
3820 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
3821 {
3822 	struct xvfsconf32 xvfsp;
3823 
3824 	bzero(&xvfsp, sizeof(xvfsp));
3825 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3826 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3827 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3828 	xvfsp.vfc_flags = vfsp->vfc_flags;
3829 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3830 }
3831 #endif
3832 
3833 /*
3834  * Top level filesystem related information gathering.
3835  */
3836 static int
3837 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
3838 {
3839 	struct vfsconf *vfsp;
3840 	int error;
3841 
3842 	error = 0;
3843 	vfsconf_slock();
3844 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3845 #ifdef COMPAT_FREEBSD32
3846 		if (req->flags & SCTL_MASK32)
3847 			error = vfsconf2x32(req, vfsp);
3848 		else
3849 #endif
3850 			error = vfsconf2x(req, vfsp);
3851 		if (error)
3852 			break;
3853 	}
3854 	vfsconf_sunlock();
3855 	return (error);
3856 }
3857 
3858 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
3859     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
3860     "S,xvfsconf", "List of all configured filesystems");
3861 
3862 #ifndef BURN_BRIDGES
3863 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
3864 
3865 static int
3866 vfs_sysctl(SYSCTL_HANDLER_ARGS)
3867 {
3868 	int *name = (int *)arg1 - 1;	/* XXX */
3869 	u_int namelen = arg2 + 1;	/* XXX */
3870 	struct vfsconf *vfsp;
3871 
3872 	log(LOG_WARNING, "userland calling deprecated sysctl, "
3873 	    "please rebuild world\n");
3874 
3875 #if 1 || defined(COMPAT_PRELITE2)
3876 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
3877 	if (namelen == 1)
3878 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
3879 #endif
3880 
3881 	switch (name[1]) {
3882 	case VFS_MAXTYPENUM:
3883 		if (namelen != 2)
3884 			return (ENOTDIR);
3885 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
3886 	case VFS_CONF:
3887 		if (namelen != 3)
3888 			return (ENOTDIR);	/* overloaded */
3889 		vfsconf_slock();
3890 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3891 			if (vfsp->vfc_typenum == name[2])
3892 				break;
3893 		}
3894 		vfsconf_sunlock();
3895 		if (vfsp == NULL)
3896 			return (EOPNOTSUPP);
3897 #ifdef COMPAT_FREEBSD32
3898 		if (req->flags & SCTL_MASK32)
3899 			return (vfsconf2x32(req, vfsp));
3900 		else
3901 #endif
3902 			return (vfsconf2x(req, vfsp));
3903 	}
3904 	return (EOPNOTSUPP);
3905 }
3906 
3907 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
3908     CTLFLAG_MPSAFE, vfs_sysctl,
3909     "Generic filesystem");
3910 
3911 #if 1 || defined(COMPAT_PRELITE2)
3912 
3913 static int
3914 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
3915 {
3916 	int error;
3917 	struct vfsconf *vfsp;
3918 	struct ovfsconf ovfs;
3919 
3920 	vfsconf_slock();
3921 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3922 		bzero(&ovfs, sizeof(ovfs));
3923 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
3924 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
3925 		ovfs.vfc_index = vfsp->vfc_typenum;
3926 		ovfs.vfc_refcount = vfsp->vfc_refcount;
3927 		ovfs.vfc_flags = vfsp->vfc_flags;
3928 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
3929 		if (error != 0) {
3930 			vfsconf_sunlock();
3931 			return (error);
3932 		}
3933 	}
3934 	vfsconf_sunlock();
3935 	return (0);
3936 }
3937 
3938 #endif /* 1 || COMPAT_PRELITE2 */
3939 #endif /* !BURN_BRIDGES */
3940 
3941 #define KINFO_VNODESLOP		10
3942 #ifdef notyet
3943 /*
3944  * Dump vnode list (via sysctl).
3945  */
3946 /* ARGSUSED */
3947 static int
3948 sysctl_vnode(SYSCTL_HANDLER_ARGS)
3949 {
3950 	struct xvnode *xvn;
3951 	struct mount *mp;
3952 	struct vnode *vp;
3953 	int error, len, n;
3954 
3955 	/*
3956 	 * Stale numvnodes access is not fatal here.
3957 	 */
3958 	req->lock = 0;
3959 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
3960 	if (!req->oldptr)
3961 		/* Make an estimate */
3962 		return (SYSCTL_OUT(req, 0, len));
3963 
3964 	error = sysctl_wire_old_buffer(req, 0);
3965 	if (error != 0)
3966 		return (error);
3967 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
3968 	n = 0;
3969 	mtx_lock(&mountlist_mtx);
3970 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3971 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
3972 			continue;
3973 		MNT_ILOCK(mp);
3974 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3975 			if (n == len)
3976 				break;
3977 			vref(vp);
3978 			xvn[n].xv_size = sizeof *xvn;
3979 			xvn[n].xv_vnode = vp;
3980 			xvn[n].xv_id = 0;	/* XXX compat */
3981 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
3982 			XV_COPY(usecount);
3983 			XV_COPY(writecount);
3984 			XV_COPY(holdcnt);
3985 			XV_COPY(mount);
3986 			XV_COPY(numoutput);
3987 			XV_COPY(type);
3988 #undef XV_COPY
3989 			xvn[n].xv_flag = vp->v_vflag;
3990 
3991 			switch (vp->v_type) {
3992 			case VREG:
3993 			case VDIR:
3994 			case VLNK:
3995 				break;
3996 			case VBLK:
3997 			case VCHR:
3998 				if (vp->v_rdev == NULL) {
3999 					vrele(vp);
4000 					continue;
4001 				}
4002 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
4003 				break;
4004 			case VSOCK:
4005 				xvn[n].xv_socket = vp->v_socket;
4006 				break;
4007 			case VFIFO:
4008 				xvn[n].xv_fifo = vp->v_fifoinfo;
4009 				break;
4010 			case VNON:
4011 			case VBAD:
4012 			default:
4013 				/* shouldn't happen? */
4014 				vrele(vp);
4015 				continue;
4016 			}
4017 			vrele(vp);
4018 			++n;
4019 		}
4020 		MNT_IUNLOCK(mp);
4021 		mtx_lock(&mountlist_mtx);
4022 		vfs_unbusy(mp);
4023 		if (n == len)
4024 			break;
4025 	}
4026 	mtx_unlock(&mountlist_mtx);
4027 
4028 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4029 	free(xvn, M_TEMP);
4030 	return (error);
4031 }
4032 
4033 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4034     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4035     "");
4036 #endif
4037 
4038 static void
4039 unmount_or_warn(struct mount *mp)
4040 {
4041 	int error;
4042 
4043 	error = dounmount(mp, MNT_FORCE, curthread);
4044 	if (error != 0) {
4045 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4046 		if (error == EBUSY)
4047 			printf("BUSY)\n");
4048 		else
4049 			printf("%d)\n", error);
4050 	}
4051 }
4052 
4053 /*
4054  * Unmount all filesystems. The list is traversed in reverse order
4055  * of mounting to avoid dependencies.
4056  */
4057 void
4058 vfs_unmountall(void)
4059 {
4060 	struct mount *mp, *tmp;
4061 
4062 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4063 
4064 	/*
4065 	 * Since this only runs when rebooting, it is not interlocked.
4066 	 */
4067 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4068 		vfs_ref(mp);
4069 
4070 		/*
4071 		 * Forcibly unmounting "/dev" before "/" would prevent clean
4072 		 * unmount of the latter.
4073 		 */
4074 		if (mp == rootdevmp)
4075 			continue;
4076 
4077 		unmount_or_warn(mp);
4078 	}
4079 
4080 	if (rootdevmp != NULL)
4081 		unmount_or_warn(rootdevmp);
4082 }
4083 
4084 /*
4085  * perform msync on all vnodes under a mount point
4086  * the mount point must be locked.
4087  */
4088 void
4089 vfs_msync(struct mount *mp, int flags)
4090 {
4091 	struct vnode *vp, *mvp;
4092 	struct vm_object *obj;
4093 
4094 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4095 
4096 	vnlru_return_batch(mp);
4097 
4098 	MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
4099 		obj = vp->v_object;
4100 		if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 &&
4101 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
4102 			if (!vget(vp,
4103 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
4104 			    curthread)) {
4105 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
4106 					vput(vp);
4107 					continue;
4108 				}
4109 
4110 				obj = vp->v_object;
4111 				if (obj != NULL) {
4112 					VM_OBJECT_WLOCK(obj);
4113 					vm_object_page_clean(obj, 0, 0,
4114 					    flags == MNT_WAIT ?
4115 					    OBJPC_SYNC : OBJPC_NOSYNC);
4116 					VM_OBJECT_WUNLOCK(obj);
4117 				}
4118 				vput(vp);
4119 			}
4120 		} else
4121 			VI_UNLOCK(vp);
4122 	}
4123 }
4124 
4125 static void
4126 destroy_vpollinfo_free(struct vpollinfo *vi)
4127 {
4128 
4129 	knlist_destroy(&vi->vpi_selinfo.si_note);
4130 	mtx_destroy(&vi->vpi_lock);
4131 	uma_zfree(vnodepoll_zone, vi);
4132 }
4133 
4134 static void
4135 destroy_vpollinfo(struct vpollinfo *vi)
4136 {
4137 
4138 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
4139 	seldrain(&vi->vpi_selinfo);
4140 	destroy_vpollinfo_free(vi);
4141 }
4142 
4143 /*
4144  * Initialize per-vnode helper structure to hold poll-related state.
4145  */
4146 void
4147 v_addpollinfo(struct vnode *vp)
4148 {
4149 	struct vpollinfo *vi;
4150 
4151 	if (vp->v_pollinfo != NULL)
4152 		return;
4153 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4154 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4155 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4156 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4157 	VI_LOCK(vp);
4158 	if (vp->v_pollinfo != NULL) {
4159 		VI_UNLOCK(vp);
4160 		destroy_vpollinfo_free(vi);
4161 		return;
4162 	}
4163 	vp->v_pollinfo = vi;
4164 	VI_UNLOCK(vp);
4165 }
4166 
4167 /*
4168  * Record a process's interest in events which might happen to
4169  * a vnode.  Because poll uses the historic select-style interface
4170  * internally, this routine serves as both the ``check for any
4171  * pending events'' and the ``record my interest in future events''
4172  * functions.  (These are done together, while the lock is held,
4173  * to avoid race conditions.)
4174  */
4175 int
4176 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4177 {
4178 
4179 	v_addpollinfo(vp);
4180 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4181 	if (vp->v_pollinfo->vpi_revents & events) {
4182 		/*
4183 		 * This leaves events we are not interested
4184 		 * in available for the other process which
4185 		 * which presumably had requested them
4186 		 * (otherwise they would never have been
4187 		 * recorded).
4188 		 */
4189 		events &= vp->v_pollinfo->vpi_revents;
4190 		vp->v_pollinfo->vpi_revents &= ~events;
4191 
4192 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4193 		return (events);
4194 	}
4195 	vp->v_pollinfo->vpi_events |= events;
4196 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4197 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4198 	return (0);
4199 }
4200 
4201 /*
4202  * Routine to create and manage a filesystem syncer vnode.
4203  */
4204 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4205 static int	sync_fsync(struct  vop_fsync_args *);
4206 static int	sync_inactive(struct  vop_inactive_args *);
4207 static int	sync_reclaim(struct  vop_reclaim_args *);
4208 
4209 static struct vop_vector sync_vnodeops = {
4210 	.vop_bypass =	VOP_EOPNOTSUPP,
4211 	.vop_close =	sync_close,		/* close */
4212 	.vop_fsync =	sync_fsync,		/* fsync */
4213 	.vop_inactive =	sync_inactive,	/* inactive */
4214 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4215 	.vop_lock1 =	vop_stdlock,	/* lock */
4216 	.vop_unlock =	vop_stdunlock,	/* unlock */
4217 	.vop_islocked =	vop_stdislocked,	/* islocked */
4218 };
4219 
4220 /*
4221  * Create a new filesystem syncer vnode for the specified mount point.
4222  */
4223 void
4224 vfs_allocate_syncvnode(struct mount *mp)
4225 {
4226 	struct vnode *vp;
4227 	struct bufobj *bo;
4228 	static long start, incr, next;
4229 	int error;
4230 
4231 	/* Allocate a new vnode */
4232 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4233 	if (error != 0)
4234 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4235 	vp->v_type = VNON;
4236 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4237 	vp->v_vflag |= VV_FORCEINSMQ;
4238 	error = insmntque(vp, mp);
4239 	if (error != 0)
4240 		panic("vfs_allocate_syncvnode: insmntque() failed");
4241 	vp->v_vflag &= ~VV_FORCEINSMQ;
4242 	VOP_UNLOCK(vp, 0);
4243 	/*
4244 	 * Place the vnode onto the syncer worklist. We attempt to
4245 	 * scatter them about on the list so that they will go off
4246 	 * at evenly distributed times even if all the filesystems
4247 	 * are mounted at once.
4248 	 */
4249 	next += incr;
4250 	if (next == 0 || next > syncer_maxdelay) {
4251 		start /= 2;
4252 		incr /= 2;
4253 		if (start == 0) {
4254 			start = syncer_maxdelay / 2;
4255 			incr = syncer_maxdelay;
4256 		}
4257 		next = start;
4258 	}
4259 	bo = &vp->v_bufobj;
4260 	BO_LOCK(bo);
4261 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4262 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4263 	mtx_lock(&sync_mtx);
4264 	sync_vnode_count++;
4265 	if (mp->mnt_syncer == NULL) {
4266 		mp->mnt_syncer = vp;
4267 		vp = NULL;
4268 	}
4269 	mtx_unlock(&sync_mtx);
4270 	BO_UNLOCK(bo);
4271 	if (vp != NULL) {
4272 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4273 		vgone(vp);
4274 		vput(vp);
4275 	}
4276 }
4277 
4278 void
4279 vfs_deallocate_syncvnode(struct mount *mp)
4280 {
4281 	struct vnode *vp;
4282 
4283 	mtx_lock(&sync_mtx);
4284 	vp = mp->mnt_syncer;
4285 	if (vp != NULL)
4286 		mp->mnt_syncer = NULL;
4287 	mtx_unlock(&sync_mtx);
4288 	if (vp != NULL)
4289 		vrele(vp);
4290 }
4291 
4292 /*
4293  * Do a lazy sync of the filesystem.
4294  */
4295 static int
4296 sync_fsync(struct vop_fsync_args *ap)
4297 {
4298 	struct vnode *syncvp = ap->a_vp;
4299 	struct mount *mp = syncvp->v_mount;
4300 	int error, save;
4301 	struct bufobj *bo;
4302 
4303 	/*
4304 	 * We only need to do something if this is a lazy evaluation.
4305 	 */
4306 	if (ap->a_waitfor != MNT_LAZY)
4307 		return (0);
4308 
4309 	/*
4310 	 * Move ourselves to the back of the sync list.
4311 	 */
4312 	bo = &syncvp->v_bufobj;
4313 	BO_LOCK(bo);
4314 	vn_syncer_add_to_worklist(bo, syncdelay);
4315 	BO_UNLOCK(bo);
4316 
4317 	/*
4318 	 * Walk the list of vnodes pushing all that are dirty and
4319 	 * not already on the sync list.
4320 	 */
4321 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4322 		return (0);
4323 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4324 		vfs_unbusy(mp);
4325 		return (0);
4326 	}
4327 	save = curthread_pflags_set(TDP_SYNCIO);
4328 	vfs_msync(mp, MNT_NOWAIT);
4329 	error = VFS_SYNC(mp, MNT_LAZY);
4330 	curthread_pflags_restore(save);
4331 	vn_finished_write(mp);
4332 	vfs_unbusy(mp);
4333 	return (error);
4334 }
4335 
4336 /*
4337  * The syncer vnode is no referenced.
4338  */
4339 static int
4340 sync_inactive(struct vop_inactive_args *ap)
4341 {
4342 
4343 	vgone(ap->a_vp);
4344 	return (0);
4345 }
4346 
4347 /*
4348  * The syncer vnode is no longer needed and is being decommissioned.
4349  *
4350  * Modifications to the worklist must be protected by sync_mtx.
4351  */
4352 static int
4353 sync_reclaim(struct vop_reclaim_args *ap)
4354 {
4355 	struct vnode *vp = ap->a_vp;
4356 	struct bufobj *bo;
4357 
4358 	bo = &vp->v_bufobj;
4359 	BO_LOCK(bo);
4360 	mtx_lock(&sync_mtx);
4361 	if (vp->v_mount->mnt_syncer == vp)
4362 		vp->v_mount->mnt_syncer = NULL;
4363 	if (bo->bo_flag & BO_ONWORKLST) {
4364 		LIST_REMOVE(bo, bo_synclist);
4365 		syncer_worklist_len--;
4366 		sync_vnode_count--;
4367 		bo->bo_flag &= ~BO_ONWORKLST;
4368 	}
4369 	mtx_unlock(&sync_mtx);
4370 	BO_UNLOCK(bo);
4371 
4372 	return (0);
4373 }
4374 
4375 /*
4376  * Check if vnode represents a disk device
4377  */
4378 int
4379 vn_isdisk(struct vnode *vp, int *errp)
4380 {
4381 	int error;
4382 
4383 	if (vp->v_type != VCHR) {
4384 		error = ENOTBLK;
4385 		goto out;
4386 	}
4387 	error = 0;
4388 	dev_lock();
4389 	if (vp->v_rdev == NULL)
4390 		error = ENXIO;
4391 	else if (vp->v_rdev->si_devsw == NULL)
4392 		error = ENXIO;
4393 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
4394 		error = ENOTBLK;
4395 	dev_unlock();
4396 out:
4397 	if (errp != NULL)
4398 		*errp = error;
4399 	return (error == 0);
4400 }
4401 
4402 /*
4403  * Common filesystem object access control check routine.  Accepts a
4404  * vnode's type, "mode", uid and gid, requested access mode, credentials,
4405  * and optional call-by-reference privused argument allowing vaccess()
4406  * to indicate to the caller whether privilege was used to satisfy the
4407  * request (obsoleted).  Returns 0 on success, or an errno on failure.
4408  */
4409 int
4410 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
4411     accmode_t accmode, struct ucred *cred, int *privused)
4412 {
4413 	accmode_t dac_granted;
4414 	accmode_t priv_granted;
4415 
4416 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
4417 	    ("invalid bit in accmode"));
4418 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
4419 	    ("VAPPEND without VWRITE"));
4420 
4421 	/*
4422 	 * Look for a normal, non-privileged way to access the file/directory
4423 	 * as requested.  If it exists, go with that.
4424 	 */
4425 
4426 	if (privused != NULL)
4427 		*privused = 0;
4428 
4429 	dac_granted = 0;
4430 
4431 	/* Check the owner. */
4432 	if (cred->cr_uid == file_uid) {
4433 		dac_granted |= VADMIN;
4434 		if (file_mode & S_IXUSR)
4435 			dac_granted |= VEXEC;
4436 		if (file_mode & S_IRUSR)
4437 			dac_granted |= VREAD;
4438 		if (file_mode & S_IWUSR)
4439 			dac_granted |= (VWRITE | VAPPEND);
4440 
4441 		if ((accmode & dac_granted) == accmode)
4442 			return (0);
4443 
4444 		goto privcheck;
4445 	}
4446 
4447 	/* Otherwise, check the groups (first match) */
4448 	if (groupmember(file_gid, cred)) {
4449 		if (file_mode & S_IXGRP)
4450 			dac_granted |= VEXEC;
4451 		if (file_mode & S_IRGRP)
4452 			dac_granted |= VREAD;
4453 		if (file_mode & S_IWGRP)
4454 			dac_granted |= (VWRITE | VAPPEND);
4455 
4456 		if ((accmode & dac_granted) == accmode)
4457 			return (0);
4458 
4459 		goto privcheck;
4460 	}
4461 
4462 	/* Otherwise, check everyone else. */
4463 	if (file_mode & S_IXOTH)
4464 		dac_granted |= VEXEC;
4465 	if (file_mode & S_IROTH)
4466 		dac_granted |= VREAD;
4467 	if (file_mode & S_IWOTH)
4468 		dac_granted |= (VWRITE | VAPPEND);
4469 	if ((accmode & dac_granted) == accmode)
4470 		return (0);
4471 
4472 privcheck:
4473 	/*
4474 	 * Build a privilege mask to determine if the set of privileges
4475 	 * satisfies the requirements when combined with the granted mask
4476 	 * from above.  For each privilege, if the privilege is required,
4477 	 * bitwise or the request type onto the priv_granted mask.
4478 	 */
4479 	priv_granted = 0;
4480 
4481 	if (type == VDIR) {
4482 		/*
4483 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
4484 		 * requests, instead of PRIV_VFS_EXEC.
4485 		 */
4486 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4487 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP, 0))
4488 			priv_granted |= VEXEC;
4489 	} else {
4490 		/*
4491 		 * Ensure that at least one execute bit is on. Otherwise,
4492 		 * a privileged user will always succeed, and we don't want
4493 		 * this to happen unless the file really is executable.
4494 		 */
4495 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4496 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
4497 		    !priv_check_cred(cred, PRIV_VFS_EXEC, 0))
4498 			priv_granted |= VEXEC;
4499 	}
4500 
4501 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
4502 	    !priv_check_cred(cred, PRIV_VFS_READ, 0))
4503 		priv_granted |= VREAD;
4504 
4505 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
4506 	    !priv_check_cred(cred, PRIV_VFS_WRITE, 0))
4507 		priv_granted |= (VWRITE | VAPPEND);
4508 
4509 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
4510 	    !priv_check_cred(cred, PRIV_VFS_ADMIN, 0))
4511 		priv_granted |= VADMIN;
4512 
4513 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
4514 		/* XXX audit: privilege used */
4515 		if (privused != NULL)
4516 			*privused = 1;
4517 		return (0);
4518 	}
4519 
4520 	return ((accmode & VADMIN) ? EPERM : EACCES);
4521 }
4522 
4523 /*
4524  * Credential check based on process requesting service, and per-attribute
4525  * permissions.
4526  */
4527 int
4528 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
4529     struct thread *td, accmode_t accmode)
4530 {
4531 
4532 	/*
4533 	 * Kernel-invoked always succeeds.
4534 	 */
4535 	if (cred == NOCRED)
4536 		return (0);
4537 
4538 	/*
4539 	 * Do not allow privileged processes in jail to directly manipulate
4540 	 * system attributes.
4541 	 */
4542 	switch (attrnamespace) {
4543 	case EXTATTR_NAMESPACE_SYSTEM:
4544 		/* Potentially should be: return (EPERM); */
4545 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0));
4546 	case EXTATTR_NAMESPACE_USER:
4547 		return (VOP_ACCESS(vp, accmode, cred, td));
4548 	default:
4549 		return (EPERM);
4550 	}
4551 }
4552 
4553 #ifdef DEBUG_VFS_LOCKS
4554 /*
4555  * This only exists to suppress warnings from unlocked specfs accesses.  It is
4556  * no longer ok to have an unlocked VFS.
4557  */
4558 #define	IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL ||		\
4559 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
4560 
4561 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
4562 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
4563     "Drop into debugger on lock violation");
4564 
4565 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
4566 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
4567     0, "Check for interlock across VOPs");
4568 
4569 int vfs_badlock_print = 1;	/* Print lock violations. */
4570 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
4571     0, "Print lock violations");
4572 
4573 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
4574 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
4575     0, "Print vnode details on lock violations");
4576 
4577 #ifdef KDB
4578 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
4579 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
4580     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
4581 #endif
4582 
4583 static void
4584 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
4585 {
4586 
4587 #ifdef KDB
4588 	if (vfs_badlock_backtrace)
4589 		kdb_backtrace();
4590 #endif
4591 	if (vfs_badlock_vnode)
4592 		vn_printf(vp, "vnode ");
4593 	if (vfs_badlock_print)
4594 		printf("%s: %p %s\n", str, (void *)vp, msg);
4595 	if (vfs_badlock_ddb)
4596 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4597 }
4598 
4599 void
4600 assert_vi_locked(struct vnode *vp, const char *str)
4601 {
4602 
4603 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
4604 		vfs_badlock("interlock is not locked but should be", str, vp);
4605 }
4606 
4607 void
4608 assert_vi_unlocked(struct vnode *vp, const char *str)
4609 {
4610 
4611 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
4612 		vfs_badlock("interlock is locked but should not be", str, vp);
4613 }
4614 
4615 void
4616 assert_vop_locked(struct vnode *vp, const char *str)
4617 {
4618 	int locked;
4619 
4620 	if (!IGNORE_LOCK(vp)) {
4621 		locked = VOP_ISLOCKED(vp);
4622 		if (locked == 0 || locked == LK_EXCLOTHER)
4623 			vfs_badlock("is not locked but should be", str, vp);
4624 	}
4625 }
4626 
4627 void
4628 assert_vop_unlocked(struct vnode *vp, const char *str)
4629 {
4630 
4631 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
4632 		vfs_badlock("is locked but should not be", str, vp);
4633 }
4634 
4635 void
4636 assert_vop_elocked(struct vnode *vp, const char *str)
4637 {
4638 
4639 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
4640 		vfs_badlock("is not exclusive locked but should be", str, vp);
4641 }
4642 #endif /* DEBUG_VFS_LOCKS */
4643 
4644 void
4645 vop_rename_fail(struct vop_rename_args *ap)
4646 {
4647 
4648 	if (ap->a_tvp != NULL)
4649 		vput(ap->a_tvp);
4650 	if (ap->a_tdvp == ap->a_tvp)
4651 		vrele(ap->a_tdvp);
4652 	else
4653 		vput(ap->a_tdvp);
4654 	vrele(ap->a_fdvp);
4655 	vrele(ap->a_fvp);
4656 }
4657 
4658 void
4659 vop_rename_pre(void *ap)
4660 {
4661 	struct vop_rename_args *a = ap;
4662 
4663 #ifdef DEBUG_VFS_LOCKS
4664 	if (a->a_tvp)
4665 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
4666 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
4667 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
4668 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
4669 
4670 	/* Check the source (from). */
4671 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
4672 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
4673 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
4674 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
4675 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
4676 
4677 	/* Check the target. */
4678 	if (a->a_tvp)
4679 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
4680 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
4681 #endif
4682 	if (a->a_tdvp != a->a_fdvp)
4683 		vhold(a->a_fdvp);
4684 	if (a->a_tvp != a->a_fvp)
4685 		vhold(a->a_fvp);
4686 	vhold(a->a_tdvp);
4687 	if (a->a_tvp)
4688 		vhold(a->a_tvp);
4689 }
4690 
4691 #ifdef DEBUG_VFS_LOCKS
4692 void
4693 vop_strategy_pre(void *ap)
4694 {
4695 	struct vop_strategy_args *a;
4696 	struct buf *bp;
4697 
4698 	a = ap;
4699 	bp = a->a_bp;
4700 
4701 	/*
4702 	 * Cluster ops lock their component buffers but not the IO container.
4703 	 */
4704 	if ((bp->b_flags & B_CLUSTER) != 0)
4705 		return;
4706 
4707 	if (panicstr == NULL && !BUF_ISLOCKED(bp)) {
4708 		if (vfs_badlock_print)
4709 			printf(
4710 			    "VOP_STRATEGY: bp is not locked but should be\n");
4711 		if (vfs_badlock_ddb)
4712 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4713 	}
4714 }
4715 
4716 void
4717 vop_lock_pre(void *ap)
4718 {
4719 	struct vop_lock1_args *a = ap;
4720 
4721 	if ((a->a_flags & LK_INTERLOCK) == 0)
4722 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4723 	else
4724 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
4725 }
4726 
4727 void
4728 vop_lock_post(void *ap, int rc)
4729 {
4730 	struct vop_lock1_args *a = ap;
4731 
4732 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4733 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
4734 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
4735 }
4736 
4737 void
4738 vop_unlock_pre(void *ap)
4739 {
4740 	struct vop_unlock_args *a = ap;
4741 
4742 	if (a->a_flags & LK_INTERLOCK)
4743 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
4744 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
4745 }
4746 
4747 void
4748 vop_unlock_post(void *ap, int rc)
4749 {
4750 	struct vop_unlock_args *a = ap;
4751 
4752 	if (a->a_flags & LK_INTERLOCK)
4753 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
4754 }
4755 #endif
4756 
4757 void
4758 vop_create_post(void *ap, int rc)
4759 {
4760 	struct vop_create_args *a = ap;
4761 
4762 	if (!rc)
4763 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4764 }
4765 
4766 void
4767 vop_deleteextattr_post(void *ap, int rc)
4768 {
4769 	struct vop_deleteextattr_args *a = ap;
4770 
4771 	if (!rc)
4772 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4773 }
4774 
4775 void
4776 vop_link_post(void *ap, int rc)
4777 {
4778 	struct vop_link_args *a = ap;
4779 
4780 	if (!rc) {
4781 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
4782 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
4783 	}
4784 }
4785 
4786 void
4787 vop_mkdir_post(void *ap, int rc)
4788 {
4789 	struct vop_mkdir_args *a = ap;
4790 
4791 	if (!rc)
4792 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4793 }
4794 
4795 void
4796 vop_mknod_post(void *ap, int rc)
4797 {
4798 	struct vop_mknod_args *a = ap;
4799 
4800 	if (!rc)
4801 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4802 }
4803 
4804 void
4805 vop_reclaim_post(void *ap, int rc)
4806 {
4807 	struct vop_reclaim_args *a = ap;
4808 
4809 	if (!rc)
4810 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
4811 }
4812 
4813 void
4814 vop_remove_post(void *ap, int rc)
4815 {
4816 	struct vop_remove_args *a = ap;
4817 
4818 	if (!rc) {
4819 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4820 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4821 	}
4822 }
4823 
4824 void
4825 vop_rename_post(void *ap, int rc)
4826 {
4827 	struct vop_rename_args *a = ap;
4828 	long hint;
4829 
4830 	if (!rc) {
4831 		hint = NOTE_WRITE;
4832 		if (a->a_fdvp == a->a_tdvp) {
4833 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
4834 				hint |= NOTE_LINK;
4835 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
4836 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
4837 		} else {
4838 			hint |= NOTE_EXTEND;
4839 			if (a->a_fvp->v_type == VDIR)
4840 				hint |= NOTE_LINK;
4841 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
4842 
4843 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
4844 			    a->a_tvp->v_type == VDIR)
4845 				hint &= ~NOTE_LINK;
4846 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
4847 		}
4848 
4849 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
4850 		if (a->a_tvp)
4851 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
4852 	}
4853 	if (a->a_tdvp != a->a_fdvp)
4854 		vdrop(a->a_fdvp);
4855 	if (a->a_tvp != a->a_fvp)
4856 		vdrop(a->a_fvp);
4857 	vdrop(a->a_tdvp);
4858 	if (a->a_tvp)
4859 		vdrop(a->a_tvp);
4860 }
4861 
4862 void
4863 vop_rmdir_post(void *ap, int rc)
4864 {
4865 	struct vop_rmdir_args *a = ap;
4866 
4867 	if (!rc) {
4868 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4869 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4870 	}
4871 }
4872 
4873 void
4874 vop_setattr_post(void *ap, int rc)
4875 {
4876 	struct vop_setattr_args *a = ap;
4877 
4878 	if (!rc)
4879 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4880 }
4881 
4882 void
4883 vop_setextattr_post(void *ap, int rc)
4884 {
4885 	struct vop_setextattr_args *a = ap;
4886 
4887 	if (!rc)
4888 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4889 }
4890 
4891 void
4892 vop_symlink_post(void *ap, int rc)
4893 {
4894 	struct vop_symlink_args *a = ap;
4895 
4896 	if (!rc)
4897 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4898 }
4899 
4900 void
4901 vop_open_post(void *ap, int rc)
4902 {
4903 	struct vop_open_args *a = ap;
4904 
4905 	if (!rc)
4906 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
4907 }
4908 
4909 void
4910 vop_close_post(void *ap, int rc)
4911 {
4912 	struct vop_close_args *a = ap;
4913 
4914 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
4915 	    (a->a_vp->v_iflag & VI_DOOMED) == 0)) {
4916 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
4917 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
4918 	}
4919 }
4920 
4921 void
4922 vop_read_post(void *ap, int rc)
4923 {
4924 	struct vop_read_args *a = ap;
4925 
4926 	if (!rc)
4927 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
4928 }
4929 
4930 void
4931 vop_readdir_post(void *ap, int rc)
4932 {
4933 	struct vop_readdir_args *a = ap;
4934 
4935 	if (!rc)
4936 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
4937 }
4938 
4939 static struct knlist fs_knlist;
4940 
4941 static void
4942 vfs_event_init(void *arg)
4943 {
4944 	knlist_init_mtx(&fs_knlist, NULL);
4945 }
4946 /* XXX - correct order? */
4947 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
4948 
4949 void
4950 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
4951 {
4952 
4953 	KNOTE_UNLOCKED(&fs_knlist, event);
4954 }
4955 
4956 static int	filt_fsattach(struct knote *kn);
4957 static void	filt_fsdetach(struct knote *kn);
4958 static int	filt_fsevent(struct knote *kn, long hint);
4959 
4960 struct filterops fs_filtops = {
4961 	.f_isfd = 0,
4962 	.f_attach = filt_fsattach,
4963 	.f_detach = filt_fsdetach,
4964 	.f_event = filt_fsevent
4965 };
4966 
4967 static int
4968 filt_fsattach(struct knote *kn)
4969 {
4970 
4971 	kn->kn_flags |= EV_CLEAR;
4972 	knlist_add(&fs_knlist, kn, 0);
4973 	return (0);
4974 }
4975 
4976 static void
4977 filt_fsdetach(struct knote *kn)
4978 {
4979 
4980 	knlist_remove(&fs_knlist, kn, 0);
4981 }
4982 
4983 static int
4984 filt_fsevent(struct knote *kn, long hint)
4985 {
4986 
4987 	kn->kn_fflags |= hint;
4988 	return (kn->kn_fflags != 0);
4989 }
4990 
4991 static int
4992 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
4993 {
4994 	struct vfsidctl vc;
4995 	int error;
4996 	struct mount *mp;
4997 
4998 	error = SYSCTL_IN(req, &vc, sizeof(vc));
4999 	if (error)
5000 		return (error);
5001 	if (vc.vc_vers != VFS_CTL_VERS1)
5002 		return (EINVAL);
5003 	mp = vfs_getvfs(&vc.vc_fsid);
5004 	if (mp == NULL)
5005 		return (ENOENT);
5006 	/* ensure that a specific sysctl goes to the right filesystem. */
5007 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
5008 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
5009 		vfs_rel(mp);
5010 		return (EINVAL);
5011 	}
5012 	VCTLTOREQ(&vc, req);
5013 	error = VFS_SYSCTL(mp, vc.vc_op, req);
5014 	vfs_rel(mp);
5015 	return (error);
5016 }
5017 
5018 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR,
5019     NULL, 0, sysctl_vfs_ctl, "",
5020     "Sysctl by fsid");
5021 
5022 /*
5023  * Function to initialize a va_filerev field sensibly.
5024  * XXX: Wouldn't a random number make a lot more sense ??
5025  */
5026 u_quad_t
5027 init_va_filerev(void)
5028 {
5029 	struct bintime bt;
5030 
5031 	getbinuptime(&bt);
5032 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
5033 }
5034 
5035 static int	filt_vfsread(struct knote *kn, long hint);
5036 static int	filt_vfswrite(struct knote *kn, long hint);
5037 static int	filt_vfsvnode(struct knote *kn, long hint);
5038 static void	filt_vfsdetach(struct knote *kn);
5039 static struct filterops vfsread_filtops = {
5040 	.f_isfd = 1,
5041 	.f_detach = filt_vfsdetach,
5042 	.f_event = filt_vfsread
5043 };
5044 static struct filterops vfswrite_filtops = {
5045 	.f_isfd = 1,
5046 	.f_detach = filt_vfsdetach,
5047 	.f_event = filt_vfswrite
5048 };
5049 static struct filterops vfsvnode_filtops = {
5050 	.f_isfd = 1,
5051 	.f_detach = filt_vfsdetach,
5052 	.f_event = filt_vfsvnode
5053 };
5054 
5055 static void
5056 vfs_knllock(void *arg)
5057 {
5058 	struct vnode *vp = arg;
5059 
5060 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5061 }
5062 
5063 static void
5064 vfs_knlunlock(void *arg)
5065 {
5066 	struct vnode *vp = arg;
5067 
5068 	VOP_UNLOCK(vp, 0);
5069 }
5070 
5071 static void
5072 vfs_knl_assert_locked(void *arg)
5073 {
5074 #ifdef DEBUG_VFS_LOCKS
5075 	struct vnode *vp = arg;
5076 
5077 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
5078 #endif
5079 }
5080 
5081 static void
5082 vfs_knl_assert_unlocked(void *arg)
5083 {
5084 #ifdef DEBUG_VFS_LOCKS
5085 	struct vnode *vp = arg;
5086 
5087 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
5088 #endif
5089 }
5090 
5091 int
5092 vfs_kqfilter(struct vop_kqfilter_args *ap)
5093 {
5094 	struct vnode *vp = ap->a_vp;
5095 	struct knote *kn = ap->a_kn;
5096 	struct knlist *knl;
5097 
5098 	switch (kn->kn_filter) {
5099 	case EVFILT_READ:
5100 		kn->kn_fop = &vfsread_filtops;
5101 		break;
5102 	case EVFILT_WRITE:
5103 		kn->kn_fop = &vfswrite_filtops;
5104 		break;
5105 	case EVFILT_VNODE:
5106 		kn->kn_fop = &vfsvnode_filtops;
5107 		break;
5108 	default:
5109 		return (EINVAL);
5110 	}
5111 
5112 	kn->kn_hook = (caddr_t)vp;
5113 
5114 	v_addpollinfo(vp);
5115 	if (vp->v_pollinfo == NULL)
5116 		return (ENOMEM);
5117 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
5118 	vhold(vp);
5119 	knlist_add(knl, kn, 0);
5120 
5121 	return (0);
5122 }
5123 
5124 /*
5125  * Detach knote from vnode
5126  */
5127 static void
5128 filt_vfsdetach(struct knote *kn)
5129 {
5130 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5131 
5132 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5133 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5134 	vdrop(vp);
5135 }
5136 
5137 /*ARGSUSED*/
5138 static int
5139 filt_vfsread(struct knote *kn, long hint)
5140 {
5141 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5142 	struct vattr va;
5143 	int res;
5144 
5145 	/*
5146 	 * filesystem is gone, so set the EOF flag and schedule
5147 	 * the knote for deletion.
5148 	 */
5149 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5150 		VI_LOCK(vp);
5151 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5152 		VI_UNLOCK(vp);
5153 		return (1);
5154 	}
5155 
5156 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5157 		return (0);
5158 
5159 	VI_LOCK(vp);
5160 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5161 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5162 	VI_UNLOCK(vp);
5163 	return (res);
5164 }
5165 
5166 /*ARGSUSED*/
5167 static int
5168 filt_vfswrite(struct knote *kn, long hint)
5169 {
5170 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5171 
5172 	VI_LOCK(vp);
5173 
5174 	/*
5175 	 * filesystem is gone, so set the EOF flag and schedule
5176 	 * the knote for deletion.
5177 	 */
5178 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5179 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5180 
5181 	kn->kn_data = 0;
5182 	VI_UNLOCK(vp);
5183 	return (1);
5184 }
5185 
5186 static int
5187 filt_vfsvnode(struct knote *kn, long hint)
5188 {
5189 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5190 	int res;
5191 
5192 	VI_LOCK(vp);
5193 	if (kn->kn_sfflags & hint)
5194 		kn->kn_fflags |= hint;
5195 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5196 		kn->kn_flags |= EV_EOF;
5197 		VI_UNLOCK(vp);
5198 		return (1);
5199 	}
5200 	res = (kn->kn_fflags != 0);
5201 	VI_UNLOCK(vp);
5202 	return (res);
5203 }
5204 
5205 int
5206 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5207 {
5208 	int error;
5209 
5210 	if (dp->d_reclen > ap->a_uio->uio_resid)
5211 		return (ENAMETOOLONG);
5212 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
5213 	if (error) {
5214 		if (ap->a_ncookies != NULL) {
5215 			if (ap->a_cookies != NULL)
5216 				free(ap->a_cookies, M_TEMP);
5217 			ap->a_cookies = NULL;
5218 			*ap->a_ncookies = 0;
5219 		}
5220 		return (error);
5221 	}
5222 	if (ap->a_ncookies == NULL)
5223 		return (0);
5224 
5225 	KASSERT(ap->a_cookies,
5226 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
5227 
5228 	*ap->a_cookies = realloc(*ap->a_cookies,
5229 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
5230 	(*ap->a_cookies)[*ap->a_ncookies] = off;
5231 	*ap->a_ncookies += 1;
5232 	return (0);
5233 }
5234 
5235 /*
5236  * Mark for update the access time of the file if the filesystem
5237  * supports VOP_MARKATIME.  This functionality is used by execve and
5238  * mmap, so we want to avoid the I/O implied by directly setting
5239  * va_atime for the sake of efficiency.
5240  */
5241 void
5242 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
5243 {
5244 	struct mount *mp;
5245 
5246 	mp = vp->v_mount;
5247 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
5248 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
5249 		(void)VOP_MARKATIME(vp);
5250 }
5251 
5252 /*
5253  * The purpose of this routine is to remove granularity from accmode_t,
5254  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
5255  * VADMIN and VAPPEND.
5256  *
5257  * If it returns 0, the caller is supposed to continue with the usual
5258  * access checks using 'accmode' as modified by this routine.  If it
5259  * returns nonzero value, the caller is supposed to return that value
5260  * as errno.
5261  *
5262  * Note that after this routine runs, accmode may be zero.
5263  */
5264 int
5265 vfs_unixify_accmode(accmode_t *accmode)
5266 {
5267 	/*
5268 	 * There is no way to specify explicit "deny" rule using
5269 	 * file mode or POSIX.1e ACLs.
5270 	 */
5271 	if (*accmode & VEXPLICIT_DENY) {
5272 		*accmode = 0;
5273 		return (0);
5274 	}
5275 
5276 	/*
5277 	 * None of these can be translated into usual access bits.
5278 	 * Also, the common case for NFSv4 ACLs is to not contain
5279 	 * either of these bits. Caller should check for VWRITE
5280 	 * on the containing directory instead.
5281 	 */
5282 	if (*accmode & (VDELETE_CHILD | VDELETE))
5283 		return (EPERM);
5284 
5285 	if (*accmode & VADMIN_PERMS) {
5286 		*accmode &= ~VADMIN_PERMS;
5287 		*accmode |= VADMIN;
5288 	}
5289 
5290 	/*
5291 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5292 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5293 	 */
5294 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5295 
5296 	return (0);
5297 }
5298 
5299 /*
5300  * These are helper functions for filesystems to traverse all
5301  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
5302  *
5303  * This interface replaces MNT_VNODE_FOREACH.
5304  */
5305 
5306 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
5307 
5308 struct vnode *
5309 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
5310 {
5311 	struct vnode *vp;
5312 
5313 	if (should_yield())
5314 		kern_yield(PRI_USER);
5315 	MNT_ILOCK(mp);
5316 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5317 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
5318 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
5319 		/* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
5320 		if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
5321 			continue;
5322 		VI_LOCK(vp);
5323 		if ((vp->v_iflag & VI_DOOMED) != 0) {
5324 			VI_UNLOCK(vp);
5325 			continue;
5326 		}
5327 		break;
5328 	}
5329 	if (vp == NULL) {
5330 		__mnt_vnode_markerfree_all(mvp, mp);
5331 		/* MNT_IUNLOCK(mp); -- done in above function */
5332 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
5333 		return (NULL);
5334 	}
5335 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5336 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5337 	MNT_IUNLOCK(mp);
5338 	return (vp);
5339 }
5340 
5341 struct vnode *
5342 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
5343 {
5344 	struct vnode *vp;
5345 
5346 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5347 	MNT_ILOCK(mp);
5348 	MNT_REF(mp);
5349 	(*mvp)->v_mount = mp;
5350 	(*mvp)->v_type = VMARKER;
5351 
5352 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
5353 		/* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
5354 		if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
5355 			continue;
5356 		VI_LOCK(vp);
5357 		if ((vp->v_iflag & VI_DOOMED) != 0) {
5358 			VI_UNLOCK(vp);
5359 			continue;
5360 		}
5361 		break;
5362 	}
5363 	if (vp == NULL) {
5364 		MNT_REL(mp);
5365 		MNT_IUNLOCK(mp);
5366 		free(*mvp, M_VNODE_MARKER);
5367 		*mvp = NULL;
5368 		return (NULL);
5369 	}
5370 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5371 	MNT_IUNLOCK(mp);
5372 	return (vp);
5373 }
5374 
5375 void
5376 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
5377 {
5378 
5379 	if (*mvp == NULL) {
5380 		MNT_IUNLOCK(mp);
5381 		return;
5382 	}
5383 
5384 	mtx_assert(MNT_MTX(mp), MA_OWNED);
5385 
5386 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5387 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5388 	MNT_REL(mp);
5389 	MNT_IUNLOCK(mp);
5390 	free(*mvp, M_VNODE_MARKER);
5391 	*mvp = NULL;
5392 }
5393 
5394 /*
5395  * These are helper functions for filesystems to traverse their
5396  * active vnodes.  See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h
5397  */
5398 static void
5399 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5400 {
5401 
5402 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5403 
5404 	MNT_ILOCK(mp);
5405 	MNT_REL(mp);
5406 	MNT_IUNLOCK(mp);
5407 	free(*mvp, M_VNODE_MARKER);
5408 	*mvp = NULL;
5409 }
5410 
5411 /*
5412  * Relock the mp mount vnode list lock with the vp vnode interlock in the
5413  * conventional lock order during mnt_vnode_next_active iteration.
5414  *
5415  * On entry, the mount vnode list lock is held and the vnode interlock is not.
5416  * The list lock is dropped and reacquired.  On success, both locks are held.
5417  * On failure, the mount vnode list lock is held but the vnode interlock is
5418  * not, and the procedure may have yielded.
5419  */
5420 static bool
5421 mnt_vnode_next_active_relock(struct vnode *mvp, struct mount *mp,
5422     struct vnode *vp)
5423 {
5424 	const struct vnode *tmp;
5425 	bool held, ret;
5426 
5427 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
5428 	    TAILQ_NEXT(mvp, v_actfreelist) != NULL, mvp,
5429 	    ("%s: bad marker", __func__));
5430 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
5431 	    ("%s: inappropriate vnode", __func__));
5432 	ASSERT_VI_UNLOCKED(vp, __func__);
5433 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5434 
5435 	ret = false;
5436 
5437 	TAILQ_REMOVE(&mp->mnt_activevnodelist, mvp, v_actfreelist);
5438 	TAILQ_INSERT_BEFORE(vp, mvp, v_actfreelist);
5439 
5440 	/*
5441 	 * Use a hold to prevent vp from disappearing while the mount vnode
5442 	 * list lock is dropped and reacquired.  Normally a hold would be
5443 	 * acquired with vhold(), but that might try to acquire the vnode
5444 	 * interlock, which would be a LOR with the mount vnode list lock.
5445 	 */
5446 	held = vfs_refcount_acquire_if_not_zero(&vp->v_holdcnt);
5447 	mtx_unlock(&mp->mnt_listmtx);
5448 	if (!held)
5449 		goto abort;
5450 	VI_LOCK(vp);
5451 	if (!vfs_refcount_release_if_not_last(&vp->v_holdcnt)) {
5452 		vdropl(vp);
5453 		goto abort;
5454 	}
5455 	mtx_lock(&mp->mnt_listmtx);
5456 
5457 	/*
5458 	 * Determine whether the vnode is still the next one after the marker,
5459 	 * excepting any other markers.  If the vnode has not been doomed by
5460 	 * vgone() then the hold should have ensured that it remained on the
5461 	 * active list.  If it has been doomed but is still on the active list,
5462 	 * don't abort, but rather skip over it (avoid spinning on doomed
5463 	 * vnodes).
5464 	 */
5465 	tmp = mvp;
5466 	do {
5467 		tmp = TAILQ_NEXT(tmp, v_actfreelist);
5468 	} while (tmp != NULL && tmp->v_type == VMARKER);
5469 	if (tmp != vp) {
5470 		mtx_unlock(&mp->mnt_listmtx);
5471 		VI_UNLOCK(vp);
5472 		goto abort;
5473 	}
5474 
5475 	ret = true;
5476 	goto out;
5477 abort:
5478 	maybe_yield();
5479 	mtx_lock(&mp->mnt_listmtx);
5480 out:
5481 	if (ret)
5482 		ASSERT_VI_LOCKED(vp, __func__);
5483 	else
5484 		ASSERT_VI_UNLOCKED(vp, __func__);
5485 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5486 	return (ret);
5487 }
5488 
5489 static struct vnode *
5490 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5491 {
5492 	struct vnode *vp, *nvp;
5493 
5494 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
5495 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5496 restart:
5497 	vp = TAILQ_NEXT(*mvp, v_actfreelist);
5498 	while (vp != NULL) {
5499 		if (vp->v_type == VMARKER) {
5500 			vp = TAILQ_NEXT(vp, v_actfreelist);
5501 			continue;
5502 		}
5503 		/*
5504 		 * Try-lock because this is the wrong lock order.  If that does
5505 		 * not succeed, drop the mount vnode list lock and try to
5506 		 * reacquire it and the vnode interlock in the right order.
5507 		 */
5508 		if (!VI_TRYLOCK(vp) &&
5509 		    !mnt_vnode_next_active_relock(*mvp, mp, vp))
5510 			goto restart;
5511 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
5512 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
5513 		    ("alien vnode on the active list %p %p", vp, mp));
5514 		if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0)
5515 			break;
5516 		nvp = TAILQ_NEXT(vp, v_actfreelist);
5517 		VI_UNLOCK(vp);
5518 		vp = nvp;
5519 	}
5520 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5521 
5522 	/* Check if we are done */
5523 	if (vp == NULL) {
5524 		mtx_unlock(&mp->mnt_listmtx);
5525 		mnt_vnode_markerfree_active(mvp, mp);
5526 		return (NULL);
5527 	}
5528 	TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist);
5529 	mtx_unlock(&mp->mnt_listmtx);
5530 	ASSERT_VI_LOCKED(vp, "active iter");
5531 	KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp));
5532 	return (vp);
5533 }
5534 
5535 struct vnode *
5536 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5537 {
5538 
5539 	if (should_yield())
5540 		kern_yield(PRI_USER);
5541 	mtx_lock(&mp->mnt_listmtx);
5542 	return (mnt_vnode_next_active(mvp, mp));
5543 }
5544 
5545 struct vnode *
5546 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp)
5547 {
5548 	struct vnode *vp;
5549 
5550 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5551 	MNT_ILOCK(mp);
5552 	MNT_REF(mp);
5553 	MNT_IUNLOCK(mp);
5554 	(*mvp)->v_type = VMARKER;
5555 	(*mvp)->v_mount = mp;
5556 
5557 	mtx_lock(&mp->mnt_listmtx);
5558 	vp = TAILQ_FIRST(&mp->mnt_activevnodelist);
5559 	if (vp == NULL) {
5560 		mtx_unlock(&mp->mnt_listmtx);
5561 		mnt_vnode_markerfree_active(mvp, mp);
5562 		return (NULL);
5563 	}
5564 	TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5565 	return (mnt_vnode_next_active(mvp, mp));
5566 }
5567 
5568 void
5569 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5570 {
5571 
5572 	if (*mvp == NULL)
5573 		return;
5574 
5575 	mtx_lock(&mp->mnt_listmtx);
5576 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5577 	mtx_unlock(&mp->mnt_listmtx);
5578 	mnt_vnode_markerfree_active(mvp, mp);
5579 }
5580