xref: /freebsd/sys/kern/vfs_subr.c (revision 430f7286a566b1407c7b32ce13585caf5aa59b92)
1 /*-
2  * Copyright (c) 1989, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  * (c) UNIX System Laboratories, Inc.
5  * All or some portions of this file are derived from material licensed
6  * to the University of California by American Telephone and Telegraph
7  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8  * the permission of UNIX System Laboratories, Inc.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 4. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	@(#)vfs_subr.c	8.31 (Berkeley) 5/26/95
35  */
36 
37 /*
38  * External virtual filesystem routines
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include "opt_compat.h"
45 #include "opt_ddb.h"
46 #include "opt_watchdog.h"
47 
48 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/bio.h>
51 #include <sys/buf.h>
52 #include <sys/condvar.h>
53 #include <sys/conf.h>
54 #include <sys/dirent.h>
55 #include <sys/event.h>
56 #include <sys/eventhandler.h>
57 #include <sys/extattr.h>
58 #include <sys/file.h>
59 #include <sys/fcntl.h>
60 #include <sys/jail.h>
61 #include <sys/kdb.h>
62 #include <sys/kernel.h>
63 #include <sys/kthread.h>
64 #include <sys/lockf.h>
65 #include <sys/malloc.h>
66 #include <sys/mount.h>
67 #include <sys/namei.h>
68 #include <sys/pctrie.h>
69 #include <sys/priv.h>
70 #include <sys/reboot.h>
71 #include <sys/refcount.h>
72 #include <sys/rwlock.h>
73 #include <sys/sched.h>
74 #include <sys/sleepqueue.h>
75 #include <sys/smp.h>
76 #include <sys/stat.h>
77 #include <sys/sysctl.h>
78 #include <sys/syslog.h>
79 #include <sys/vmmeter.h>
80 #include <sys/vnode.h>
81 #include <sys/watchdog.h>
82 
83 #include <machine/stdarg.h>
84 
85 #include <security/mac/mac_framework.h>
86 
87 #include <vm/vm.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_extern.h>
90 #include <vm/pmap.h>
91 #include <vm/vm_map.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_kern.h>
94 #include <vm/uma.h>
95 
96 #ifdef DDB
97 #include <ddb/ddb.h>
98 #endif
99 
100 static void	delmntque(struct vnode *vp);
101 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
102 		    int slpflag, int slptimeo);
103 static void	syncer_shutdown(void *arg, int howto);
104 static int	vtryrecycle(struct vnode *vp);
105 static void	v_init_counters(struct vnode *);
106 static void	v_incr_usecount(struct vnode *);
107 static void	v_incr_usecount_locked(struct vnode *);
108 static void	v_incr_devcount(struct vnode *);
109 static void	v_decr_devcount(struct vnode *);
110 static void	vnlru_free(int);
111 static void	vgonel(struct vnode *);
112 static void	vfs_knllock(void *arg);
113 static void	vfs_knlunlock(void *arg);
114 static void	vfs_knl_assert_locked(void *arg);
115 static void	vfs_knl_assert_unlocked(void *arg);
116 static void	destroy_vpollinfo(struct vpollinfo *vi);
117 
118 /*
119  * Number of vnodes in existence.  Increased whenever getnewvnode()
120  * allocates a new vnode, decreased in vdropl() for VI_DOOMED vnode.
121  */
122 static unsigned long	numvnodes;
123 
124 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
125     "Number of vnodes in existence");
126 
127 static u_long vnodes_created;
128 SYSCTL_ULONG(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
129     0, "Number of vnodes created by getnewvnode");
130 
131 /*
132  * Conversion tables for conversion from vnode types to inode formats
133  * and back.
134  */
135 enum vtype iftovt_tab[16] = {
136 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
137 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
138 };
139 int vttoif_tab[10] = {
140 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
141 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
142 };
143 
144 /*
145  * List of vnodes that are ready for recycling.
146  */
147 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
148 
149 /*
150  * "Free" vnode target.  Free vnodes are rarely completely free, but are
151  * just ones that are cheap to recycle.  Usually they are for files which
152  * have been stat'd but not read; these usually have inode and namecache
153  * data attached to them.  This target is the preferred minimum size of a
154  * sub-cache consisting mostly of such files. The system balances the size
155  * of this sub-cache with its complement to try to prevent either from
156  * thrashing while the other is relatively inactive.  The targets express
157  * a preference for the best balance.
158  *
159  * "Above" this target there are 2 further targets (watermarks) related
160  * to recyling of free vnodes.  In the best-operating case, the cache is
161  * exactly full, the free list has size between vlowat and vhiwat above the
162  * free target, and recycling from it and normal use maintains this state.
163  * Sometimes the free list is below vlowat or even empty, but this state
164  * is even better for immediate use provided the cache is not full.
165  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
166  * ones) to reach one of these states.  The watermarks are currently hard-
167  * coded as 4% and 9% of the available space higher.  These and the default
168  * of 25% for wantfreevnodes are too large if the memory size is large.
169  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
170  * whenever vnlru_proc() becomes active.
171  */
172 static u_long wantfreevnodes;
173 SYSCTL_ULONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW,
174     &wantfreevnodes, 0, "Target for minimum number of \"free\" vnodes");
175 static u_long freevnodes;
176 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
177     &freevnodes, 0, "Number of \"free\" vnodes");
178 
179 static u_long recycles_count;
180 SYSCTL_ULONG(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count, 0,
181     "Number of vnodes recycled to meet vnode cache targets");
182 
183 /*
184  * Various variables used for debugging the new implementation of
185  * reassignbuf().
186  * XXX these are probably of (very) limited utility now.
187  */
188 static int reassignbufcalls;
189 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0,
190     "Number of calls to reassignbuf");
191 
192 static u_long free_owe_inact;
193 SYSCTL_ULONG(_vfs, OID_AUTO, free_owe_inact, CTLFLAG_RD, &free_owe_inact, 0,
194     "Number of times free vnodes kept on active list due to VFS "
195     "owing inactivation");
196 
197 /* To keep more than one thread at a time from running vfs_getnewfsid */
198 static struct mtx mntid_mtx;
199 
200 /*
201  * Lock for any access to the following:
202  *	vnode_free_list
203  *	numvnodes
204  *	freevnodes
205  */
206 static struct mtx vnode_free_list_mtx;
207 
208 /* Publicly exported FS */
209 struct nfs_public nfs_pub;
210 
211 static uma_zone_t buf_trie_zone;
212 
213 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
214 static uma_zone_t vnode_zone;
215 static uma_zone_t vnodepoll_zone;
216 
217 /*
218  * The workitem queue.
219  *
220  * It is useful to delay writes of file data and filesystem metadata
221  * for tens of seconds so that quickly created and deleted files need
222  * not waste disk bandwidth being created and removed. To realize this,
223  * we append vnodes to a "workitem" queue. When running with a soft
224  * updates implementation, most pending metadata dependencies should
225  * not wait for more than a few seconds. Thus, mounted on block devices
226  * are delayed only about a half the time that file data is delayed.
227  * Similarly, directory updates are more critical, so are only delayed
228  * about a third the time that file data is delayed. Thus, there are
229  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
230  * one each second (driven off the filesystem syncer process). The
231  * syncer_delayno variable indicates the next queue that is to be processed.
232  * Items that need to be processed soon are placed in this queue:
233  *
234  *	syncer_workitem_pending[syncer_delayno]
235  *
236  * A delay of fifteen seconds is done by placing the request fifteen
237  * entries later in the queue:
238  *
239  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
240  *
241  */
242 static int syncer_delayno;
243 static long syncer_mask;
244 LIST_HEAD(synclist, bufobj);
245 static struct synclist *syncer_workitem_pending;
246 /*
247  * The sync_mtx protects:
248  *	bo->bo_synclist
249  *	sync_vnode_count
250  *	syncer_delayno
251  *	syncer_state
252  *	syncer_workitem_pending
253  *	syncer_worklist_len
254  *	rushjob
255  */
256 static struct mtx sync_mtx;
257 static struct cv sync_wakeup;
258 
259 #define SYNCER_MAXDELAY		32
260 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
261 static int syncdelay = 30;		/* max time to delay syncing data */
262 static int filedelay = 30;		/* time to delay syncing files */
263 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
264     "Time to delay syncing files (in seconds)");
265 static int dirdelay = 29;		/* time to delay syncing directories */
266 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
267     "Time to delay syncing directories (in seconds)");
268 static int metadelay = 28;		/* time to delay syncing metadata */
269 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
270     "Time to delay syncing metadata (in seconds)");
271 static int rushjob;		/* number of slots to run ASAP */
272 static int stat_rush_requests;	/* number of times I/O speeded up */
273 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
274     "Number of times I/O speeded up (rush requests)");
275 
276 /*
277  * When shutting down the syncer, run it at four times normal speed.
278  */
279 #define SYNCER_SHUTDOWN_SPEEDUP		4
280 static int sync_vnode_count;
281 static int syncer_worklist_len;
282 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
283     syncer_state;
284 
285 /* Target for maximum number of vnodes. */
286 int desiredvnodes;
287 static int gapvnodes;		/* gap between wanted and desired */
288 static int vhiwat;		/* enough extras after expansion */
289 static int vlowat;		/* minimal extras before expansion */
290 static int vstir;		/* nonzero to stir non-free vnodes */
291 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
292 
293 static int
294 sysctl_update_desiredvnodes(SYSCTL_HANDLER_ARGS)
295 {
296 	int error, old_desiredvnodes;
297 
298 	old_desiredvnodes = desiredvnodes;
299 	if ((error = sysctl_handle_int(oidp, arg1, arg2, req)) != 0)
300 		return (error);
301 	if (old_desiredvnodes != desiredvnodes) {
302 		wantfreevnodes = desiredvnodes / 4;
303 		/* XXX locking seems to be incomplete. */
304 		vfs_hash_changesize(desiredvnodes);
305 		cache_changesize(desiredvnodes);
306 	}
307 	return (0);
308 }
309 
310 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
311     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, &desiredvnodes, 0,
312     sysctl_update_desiredvnodes, "I", "Target for maximum number of vnodes");
313 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
314     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
315 static int vnlru_nowhere;
316 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
317     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
318 
319 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
320 static int vnsz2log;
321 
322 /*
323  * Support for the bufobj clean & dirty pctrie.
324  */
325 static void *
326 buf_trie_alloc(struct pctrie *ptree)
327 {
328 
329 	return uma_zalloc(buf_trie_zone, M_NOWAIT);
330 }
331 
332 static void
333 buf_trie_free(struct pctrie *ptree, void *node)
334 {
335 
336 	uma_zfree(buf_trie_zone, node);
337 }
338 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
339 
340 /*
341  * Initialize the vnode management data structures.
342  *
343  * Reevaluate the following cap on the number of vnodes after the physical
344  * memory size exceeds 512GB.  In the limit, as the physical memory size
345  * grows, the ratio of the memory size in KB to to vnodes approaches 64:1.
346  */
347 #ifndef	MAXVNODES_MAX
348 #define	MAXVNODES_MAX	(512 * 1024 * 1024 / 64)	/* 8M */
349 #endif
350 
351 /*
352  * Initialize a vnode as it first enters the zone.
353  */
354 static int
355 vnode_init(void *mem, int size, int flags)
356 {
357 	struct vnode *vp;
358 	struct bufobj *bo;
359 
360 	vp = mem;
361 	bzero(vp, size);
362 	/*
363 	 * Setup locks.
364 	 */
365 	vp->v_vnlock = &vp->v_lock;
366 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
367 	/*
368 	 * By default, don't allow shared locks unless filesystems opt-in.
369 	 */
370 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
371 	    LK_NOSHARE | LK_IS_VNODE);
372 	/*
373 	 * Initialize bufobj.
374 	 */
375 	bo = &vp->v_bufobj;
376 	bo->__bo_vnode = vp;
377 	rw_init(BO_LOCKPTR(bo), "bufobj interlock");
378 	bo->bo_private = vp;
379 	TAILQ_INIT(&bo->bo_clean.bv_hd);
380 	TAILQ_INIT(&bo->bo_dirty.bv_hd);
381 	/*
382 	 * Initialize namecache.
383 	 */
384 	LIST_INIT(&vp->v_cache_src);
385 	TAILQ_INIT(&vp->v_cache_dst);
386 	/*
387 	 * Initialize rangelocks.
388 	 */
389 	rangelock_init(&vp->v_rl);
390 	return (0);
391 }
392 
393 /*
394  * Free a vnode when it is cleared from the zone.
395  */
396 static void
397 vnode_fini(void *mem, int size)
398 {
399 	struct vnode *vp;
400 	struct bufobj *bo;
401 
402 	vp = mem;
403 	rangelock_destroy(&vp->v_rl);
404 	lockdestroy(vp->v_vnlock);
405 	mtx_destroy(&vp->v_interlock);
406 	bo = &vp->v_bufobj;
407 	rw_destroy(BO_LOCKPTR(bo));
408 }
409 
410 /*
411  * Provide the size of NFS nclnode and NFS fh for calculation of the
412  * vnode memory consumption.  The size is specified directly to
413  * eliminate dependency on NFS-private header.
414  *
415  * Other filesystems may use bigger or smaller (like UFS and ZFS)
416  * private inode data, but the NFS-based estimation is ample enough.
417  * Still, we care about differences in the size between 64- and 32-bit
418  * platforms.
419  *
420  * Namecache structure size is heuristically
421  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
422  */
423 #ifdef _LP64
424 #define	NFS_NCLNODE_SZ	(528 + 64)
425 #define	NC_SZ		148
426 #else
427 #define	NFS_NCLNODE_SZ	(360 + 32)
428 #define	NC_SZ		92
429 #endif
430 
431 static void
432 vntblinit(void *dummy __unused)
433 {
434 	u_int i;
435 	int physvnodes, virtvnodes;
436 
437 	/*
438 	 * Desiredvnodes is a function of the physical memory size and the
439 	 * kernel's heap size.  Generally speaking, it scales with the
440 	 * physical memory size.  The ratio of desiredvnodes to the physical
441 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
442 	 * Thereafter, the
443 	 * marginal ratio of desiredvnodes to the physical memory size is
444 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
445 	 * size.  The memory required by desiredvnodes vnodes and vm objects
446 	 * must not exceed 1/10th of the kernel's heap size.
447 	 */
448 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
449 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
450 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
451 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
452 	desiredvnodes = min(physvnodes, virtvnodes);
453 	if (desiredvnodes > MAXVNODES_MAX) {
454 		if (bootverbose)
455 			printf("Reducing kern.maxvnodes %d -> %d\n",
456 			    desiredvnodes, MAXVNODES_MAX);
457 		desiredvnodes = MAXVNODES_MAX;
458 	}
459 	wantfreevnodes = desiredvnodes / 4;
460 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
461 	TAILQ_INIT(&vnode_free_list);
462 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
463 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
464 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
465 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
466 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
467 	/*
468 	 * Preallocate enough nodes to support one-per buf so that
469 	 * we can not fail an insert.  reassignbuf() callers can not
470 	 * tolerate the insertion failure.
471 	 */
472 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
473 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
474 	    UMA_ZONE_NOFREE | UMA_ZONE_VM);
475 	uma_prealloc(buf_trie_zone, nbuf);
476 	/*
477 	 * Initialize the filesystem syncer.
478 	 */
479 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
480 	    &syncer_mask);
481 	syncer_maxdelay = syncer_mask + 1;
482 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
483 	cv_init(&sync_wakeup, "syncer");
484 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
485 		vnsz2log++;
486 	vnsz2log--;
487 }
488 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
489 
490 
491 /*
492  * Mark a mount point as busy. Used to synchronize access and to delay
493  * unmounting. Eventually, mountlist_mtx is not released on failure.
494  *
495  * vfs_busy() is a custom lock, it can block the caller.
496  * vfs_busy() only sleeps if the unmount is active on the mount point.
497  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
498  * vnode belonging to mp.
499  *
500  * Lookup uses vfs_busy() to traverse mount points.
501  * root fs			var fs
502  * / vnode lock		A	/ vnode lock (/var)		D
503  * /var vnode lock	B	/log vnode lock(/var/log)	E
504  * vfs_busy lock	C	vfs_busy lock			F
505  *
506  * Within each file system, the lock order is C->A->B and F->D->E.
507  *
508  * When traversing across mounts, the system follows that lock order:
509  *
510  *        C->A->B
511  *              |
512  *              +->F->D->E
513  *
514  * The lookup() process for namei("/var") illustrates the process:
515  *  VOP_LOOKUP() obtains B while A is held
516  *  vfs_busy() obtains a shared lock on F while A and B are held
517  *  vput() releases lock on B
518  *  vput() releases lock on A
519  *  VFS_ROOT() obtains lock on D while shared lock on F is held
520  *  vfs_unbusy() releases shared lock on F
521  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
522  *    Attempt to lock A (instead of vp_crossmp) while D is held would
523  *    violate the global order, causing deadlocks.
524  *
525  * dounmount() locks B while F is drained.
526  */
527 int
528 vfs_busy(struct mount *mp, int flags)
529 {
530 
531 	MPASS((flags & ~MBF_MASK) == 0);
532 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
533 
534 	MNT_ILOCK(mp);
535 	MNT_REF(mp);
536 	/*
537 	 * If mount point is currently being unmounted, sleep until the
538 	 * mount point fate is decided.  If thread doing the unmounting fails,
539 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
540 	 * that this mount point has survived the unmount attempt and vfs_busy
541 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
542 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
543 	 * about to be really destroyed.  vfs_busy needs to release its
544 	 * reference on the mount point in this case and return with ENOENT,
545 	 * telling the caller that mount mount it tried to busy is no longer
546 	 * valid.
547 	 */
548 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
549 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
550 			MNT_REL(mp);
551 			MNT_IUNLOCK(mp);
552 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
553 			    __func__);
554 			return (ENOENT);
555 		}
556 		if (flags & MBF_MNTLSTLOCK)
557 			mtx_unlock(&mountlist_mtx);
558 		mp->mnt_kern_flag |= MNTK_MWAIT;
559 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
560 		if (flags & MBF_MNTLSTLOCK)
561 			mtx_lock(&mountlist_mtx);
562 		MNT_ILOCK(mp);
563 	}
564 	if (flags & MBF_MNTLSTLOCK)
565 		mtx_unlock(&mountlist_mtx);
566 	mp->mnt_lockref++;
567 	MNT_IUNLOCK(mp);
568 	return (0);
569 }
570 
571 /*
572  * Free a busy filesystem.
573  */
574 void
575 vfs_unbusy(struct mount *mp)
576 {
577 
578 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
579 	MNT_ILOCK(mp);
580 	MNT_REL(mp);
581 	KASSERT(mp->mnt_lockref > 0, ("negative mnt_lockref"));
582 	mp->mnt_lockref--;
583 	if (mp->mnt_lockref == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
584 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
585 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
586 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
587 		wakeup(&mp->mnt_lockref);
588 	}
589 	MNT_IUNLOCK(mp);
590 }
591 
592 /*
593  * Lookup a mount point by filesystem identifier.
594  */
595 struct mount *
596 vfs_getvfs(fsid_t *fsid)
597 {
598 	struct mount *mp;
599 
600 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
601 	mtx_lock(&mountlist_mtx);
602 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
603 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
604 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
605 			vfs_ref(mp);
606 			mtx_unlock(&mountlist_mtx);
607 			return (mp);
608 		}
609 	}
610 	mtx_unlock(&mountlist_mtx);
611 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
612 	return ((struct mount *) 0);
613 }
614 
615 /*
616  * Lookup a mount point by filesystem identifier, busying it before
617  * returning.
618  *
619  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
620  * cache for popular filesystem identifiers.  The cache is lockess, using
621  * the fact that struct mount's are never freed.  In worst case we may
622  * get pointer to unmounted or even different filesystem, so we have to
623  * check what we got, and go slow way if so.
624  */
625 struct mount *
626 vfs_busyfs(fsid_t *fsid)
627 {
628 #define	FSID_CACHE_SIZE	256
629 	typedef struct mount * volatile vmp_t;
630 	static vmp_t cache[FSID_CACHE_SIZE];
631 	struct mount *mp;
632 	int error;
633 	uint32_t hash;
634 
635 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
636 	hash = fsid->val[0] ^ fsid->val[1];
637 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
638 	mp = cache[hash];
639 	if (mp == NULL ||
640 	    mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
641 	    mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
642 		goto slow;
643 	if (vfs_busy(mp, 0) != 0) {
644 		cache[hash] = NULL;
645 		goto slow;
646 	}
647 	if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
648 	    mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
649 		return (mp);
650 	else
651 	    vfs_unbusy(mp);
652 
653 slow:
654 	mtx_lock(&mountlist_mtx);
655 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
656 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
657 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
658 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
659 			if (error) {
660 				cache[hash] = NULL;
661 				mtx_unlock(&mountlist_mtx);
662 				return (NULL);
663 			}
664 			cache[hash] = mp;
665 			return (mp);
666 		}
667 	}
668 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
669 	mtx_unlock(&mountlist_mtx);
670 	return ((struct mount *) 0);
671 }
672 
673 /*
674  * Check if a user can access privileged mount options.
675  */
676 int
677 vfs_suser(struct mount *mp, struct thread *td)
678 {
679 	int error;
680 
681 	/*
682 	 * If the thread is jailed, but this is not a jail-friendly file
683 	 * system, deny immediately.
684 	 */
685 	if (!(mp->mnt_vfc->vfc_flags & VFCF_JAIL) && jailed(td->td_ucred))
686 		return (EPERM);
687 
688 	/*
689 	 * If the file system was mounted outside the jail of the calling
690 	 * thread, deny immediately.
691 	 */
692 	if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
693 		return (EPERM);
694 
695 	/*
696 	 * If file system supports delegated administration, we don't check
697 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
698 	 * by the file system itself.
699 	 * If this is not the user that did original mount, we check for
700 	 * the PRIV_VFS_MOUNT_OWNER privilege.
701 	 */
702 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
703 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
704 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
705 			return (error);
706 	}
707 	return (0);
708 }
709 
710 /*
711  * Get a new unique fsid.  Try to make its val[0] unique, since this value
712  * will be used to create fake device numbers for stat().  Also try (but
713  * not so hard) make its val[0] unique mod 2^16, since some emulators only
714  * support 16-bit device numbers.  We end up with unique val[0]'s for the
715  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
716  *
717  * Keep in mind that several mounts may be running in parallel.  Starting
718  * the search one past where the previous search terminated is both a
719  * micro-optimization and a defense against returning the same fsid to
720  * different mounts.
721  */
722 void
723 vfs_getnewfsid(struct mount *mp)
724 {
725 	static uint16_t mntid_base;
726 	struct mount *nmp;
727 	fsid_t tfsid;
728 	int mtype;
729 
730 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
731 	mtx_lock(&mntid_mtx);
732 	mtype = mp->mnt_vfc->vfc_typenum;
733 	tfsid.val[1] = mtype;
734 	mtype = (mtype & 0xFF) << 24;
735 	for (;;) {
736 		tfsid.val[0] = makedev(255,
737 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
738 		mntid_base++;
739 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
740 			break;
741 		vfs_rel(nmp);
742 	}
743 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
744 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
745 	mtx_unlock(&mntid_mtx);
746 }
747 
748 /*
749  * Knob to control the precision of file timestamps:
750  *
751  *   0 = seconds only; nanoseconds zeroed.
752  *   1 = seconds and nanoseconds, accurate within 1/HZ.
753  *   2 = seconds and nanoseconds, truncated to microseconds.
754  * >=3 = seconds and nanoseconds, maximum precision.
755  */
756 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
757 
758 static int timestamp_precision = TSP_USEC;
759 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
760     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
761     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to ms, "
762     "3+: sec + ns (max. precision))");
763 
764 /*
765  * Get a current timestamp.
766  */
767 void
768 vfs_timestamp(struct timespec *tsp)
769 {
770 	struct timeval tv;
771 
772 	switch (timestamp_precision) {
773 	case TSP_SEC:
774 		tsp->tv_sec = time_second;
775 		tsp->tv_nsec = 0;
776 		break;
777 	case TSP_HZ:
778 		getnanotime(tsp);
779 		break;
780 	case TSP_USEC:
781 		microtime(&tv);
782 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
783 		break;
784 	case TSP_NSEC:
785 	default:
786 		nanotime(tsp);
787 		break;
788 	}
789 }
790 
791 /*
792  * Set vnode attributes to VNOVAL
793  */
794 void
795 vattr_null(struct vattr *vap)
796 {
797 
798 	vap->va_type = VNON;
799 	vap->va_size = VNOVAL;
800 	vap->va_bytes = VNOVAL;
801 	vap->va_mode = VNOVAL;
802 	vap->va_nlink = VNOVAL;
803 	vap->va_uid = VNOVAL;
804 	vap->va_gid = VNOVAL;
805 	vap->va_fsid = VNOVAL;
806 	vap->va_fileid = VNOVAL;
807 	vap->va_blocksize = VNOVAL;
808 	vap->va_rdev = VNOVAL;
809 	vap->va_atime.tv_sec = VNOVAL;
810 	vap->va_atime.tv_nsec = VNOVAL;
811 	vap->va_mtime.tv_sec = VNOVAL;
812 	vap->va_mtime.tv_nsec = VNOVAL;
813 	vap->va_ctime.tv_sec = VNOVAL;
814 	vap->va_ctime.tv_nsec = VNOVAL;
815 	vap->va_birthtime.tv_sec = VNOVAL;
816 	vap->va_birthtime.tv_nsec = VNOVAL;
817 	vap->va_flags = VNOVAL;
818 	vap->va_gen = VNOVAL;
819 	vap->va_vaflags = 0;
820 }
821 
822 /*
823  * This routine is called when we have too many vnodes.  It attempts
824  * to free <count> vnodes and will potentially free vnodes that still
825  * have VM backing store (VM backing store is typically the cause
826  * of a vnode blowout so we want to do this).  Therefore, this operation
827  * is not considered cheap.
828  *
829  * A number of conditions may prevent a vnode from being reclaimed.
830  * the buffer cache may have references on the vnode, a directory
831  * vnode may still have references due to the namei cache representing
832  * underlying files, or the vnode may be in active use.   It is not
833  * desirable to reuse such vnodes.  These conditions may cause the
834  * number of vnodes to reach some minimum value regardless of what
835  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
836  */
837 static int
838 vlrureclaim(struct mount *mp, int reclaim_nc_src, int trigger)
839 {
840 	struct vnode *vp;
841 	int count, done, target;
842 
843 	done = 0;
844 	vn_start_write(NULL, &mp, V_WAIT);
845 	MNT_ILOCK(mp);
846 	count = mp->mnt_nvnodelistsize;
847 	target = count * (int64_t)gapvnodes / imax(desiredvnodes, 1);
848 	target = target / 10 + 1;
849 	while (count != 0 && done < target) {
850 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
851 		while (vp != NULL && vp->v_type == VMARKER)
852 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
853 		if (vp == NULL)
854 			break;
855 		/*
856 		 * XXX LRU is completely broken for non-free vnodes.  First
857 		 * by calling here in mountpoint order, then by moving
858 		 * unselected vnodes to the end here, and most grossly by
859 		 * removing the vlruvp() function that was supposed to
860 		 * maintain the order.  (This function was born broken
861 		 * since syncer problems prevented it doing anything.)  The
862 		 * order is closer to LRC (C = Created).
863 		 *
864 		 * LRU reclaiming of vnodes seems to have last worked in
865 		 * FreeBSD-3 where LRU wasn't mentioned under any spelling.
866 		 * Then there was no hold count, and inactive vnodes were
867 		 * simply put on the free list in LRU order.  The separate
868 		 * lists also break LRU.  We prefer to reclaim from the
869 		 * free list for technical reasons.  This tends to thrash
870 		 * the free list to keep very unrecently used held vnodes.
871 		 * The problem is mitigated by keeping the free list large.
872 		 */
873 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
874 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
875 		--count;
876 		if (!VI_TRYLOCK(vp))
877 			goto next_iter;
878 		/*
879 		 * If it's been deconstructed already, it's still
880 		 * referenced, or it exceeds the trigger, skip it.
881 		 * Also skip free vnodes.  We are trying to make space
882 		 * to expand the free list, not reduce it.
883 		 */
884 		if (vp->v_usecount ||
885 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
886 		    ((vp->v_iflag & VI_FREE) != 0) ||
887 		    (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
888 		    vp->v_object->resident_page_count > trigger)) {
889 			VI_UNLOCK(vp);
890 			goto next_iter;
891 		}
892 		MNT_IUNLOCK(mp);
893 		vholdl(vp);
894 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) {
895 			vdrop(vp);
896 			goto next_iter_mntunlocked;
897 		}
898 		VI_LOCK(vp);
899 		/*
900 		 * v_usecount may have been bumped after VOP_LOCK() dropped
901 		 * the vnode interlock and before it was locked again.
902 		 *
903 		 * It is not necessary to recheck VI_DOOMED because it can
904 		 * only be set by another thread that holds both the vnode
905 		 * lock and vnode interlock.  If another thread has the
906 		 * vnode lock before we get to VOP_LOCK() and obtains the
907 		 * vnode interlock after VOP_LOCK() drops the vnode
908 		 * interlock, the other thread will be unable to drop the
909 		 * vnode lock before our VOP_LOCK() call fails.
910 		 */
911 		if (vp->v_usecount ||
912 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
913 		    (vp->v_iflag & VI_FREE) != 0 ||
914 		    (vp->v_object != NULL &&
915 		    vp->v_object->resident_page_count > trigger)) {
916 			VOP_UNLOCK(vp, LK_INTERLOCK);
917 			vdrop(vp);
918 			goto next_iter_mntunlocked;
919 		}
920 		KASSERT((vp->v_iflag & VI_DOOMED) == 0,
921 		    ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
922 		atomic_add_long(&recycles_count, 1);
923 		vgonel(vp);
924 		VOP_UNLOCK(vp, 0);
925 		vdropl(vp);
926 		done++;
927 next_iter_mntunlocked:
928 		if (!should_yield())
929 			goto relock_mnt;
930 		goto yield;
931 next_iter:
932 		if (!should_yield())
933 			continue;
934 		MNT_IUNLOCK(mp);
935 yield:
936 		kern_yield(PRI_USER);
937 relock_mnt:
938 		MNT_ILOCK(mp);
939 	}
940 	MNT_IUNLOCK(mp);
941 	vn_finished_write(mp);
942 	return done;
943 }
944 
945 /*
946  * Attempt to reduce the free list by the requested amount.
947  */
948 static void
949 vnlru_free(int count)
950 {
951 	struct vnode *vp;
952 
953 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
954 	for (; count > 0; count--) {
955 		vp = TAILQ_FIRST(&vnode_free_list);
956 		/*
957 		 * The list can be modified while the free_list_mtx
958 		 * has been dropped and vp could be NULL here.
959 		 */
960 		if (!vp)
961 			break;
962 		VNASSERT(vp->v_op != NULL, vp,
963 		    ("vnlru_free: vnode already reclaimed."));
964 		KASSERT((vp->v_iflag & VI_FREE) != 0,
965 		    ("Removing vnode not on freelist"));
966 		KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
967 		    ("Mangling active vnode"));
968 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
969 		/*
970 		 * Don't recycle if we can't get the interlock.
971 		 */
972 		if (!VI_TRYLOCK(vp)) {
973 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist);
974 			continue;
975 		}
976 		VNASSERT((vp->v_iflag & VI_FREE) != 0 && vp->v_holdcnt == 0,
977 		    vp, ("vp inconsistent on freelist"));
978 
979 		/*
980 		 * The clear of VI_FREE prevents activation of the
981 		 * vnode.  There is no sense in putting the vnode on
982 		 * the mount point active list, only to remove it
983 		 * later during recycling.  Inline the relevant part
984 		 * of vholdl(), to avoid triggering assertions or
985 		 * activating.
986 		 */
987 		freevnodes--;
988 		vp->v_iflag &= ~VI_FREE;
989 		refcount_acquire(&vp->v_holdcnt);
990 
991 		mtx_unlock(&vnode_free_list_mtx);
992 		VI_UNLOCK(vp);
993 		vtryrecycle(vp);
994 		/*
995 		 * If the recycled succeeded this vdrop will actually free
996 		 * the vnode.  If not it will simply place it back on
997 		 * the free list.
998 		 */
999 		vdrop(vp);
1000 		mtx_lock(&vnode_free_list_mtx);
1001 	}
1002 }
1003 
1004 /* XXX some names and initialization are bad for limits and watermarks. */
1005 static int
1006 vspace(void)
1007 {
1008 	int space;
1009 
1010 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1011 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1012 	vlowat = vhiwat / 2;
1013 	if (numvnodes > desiredvnodes)
1014 		return (0);
1015 	space = desiredvnodes - numvnodes;
1016 	if (freevnodes > wantfreevnodes)
1017 		space += freevnodes - wantfreevnodes;
1018 	return (space);
1019 }
1020 
1021 /*
1022  * Attempt to recycle vnodes in a context that is always safe to block.
1023  * Calling vlrurecycle() from the bowels of filesystem code has some
1024  * interesting deadlock problems.
1025  */
1026 static struct proc *vnlruproc;
1027 static int vnlruproc_sig;
1028 
1029 static void
1030 vnlru_proc(void)
1031 {
1032 	struct mount *mp, *nmp;
1033 	unsigned long ofreevnodes, onumvnodes;
1034 	int done, force, reclaim_nc_src, trigger, usevnodes;
1035 
1036 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1037 	    SHUTDOWN_PRI_FIRST);
1038 
1039 	force = 0;
1040 	for (;;) {
1041 		kproc_suspend_check(vnlruproc);
1042 		mtx_lock(&vnode_free_list_mtx);
1043 		/*
1044 		 * If numvnodes is too large (due to desiredvnodes being
1045 		 * adjusted using its sysctl, or emergency growth), first
1046 		 * try to reduce it by discarding from the free list.
1047 		 */
1048 		if (numvnodes > desiredvnodes && freevnodes > 0)
1049 			vnlru_free(ulmin(numvnodes - desiredvnodes,
1050 			    freevnodes));
1051 		/*
1052 		 * Sleep if the vnode cache is in a good state.  This is
1053 		 * when it is not over-full and has space for about a 4%
1054 		 * or 9% expansion (by growing its size or inexcessively
1055 		 * reducing its free list).  Otherwise, try to reclaim
1056 		 * space for a 10% expansion.
1057 		 */
1058 		if (vstir && force == 0) {
1059 			force = 1;
1060 			vstir = 0;
1061 		}
1062 		if (vspace() >= vlowat && force == 0) {
1063 			vnlruproc_sig = 0;
1064 			wakeup(&vnlruproc_sig);
1065 			msleep(vnlruproc, &vnode_free_list_mtx,
1066 			    PVFS|PDROP, "vlruwt", hz);
1067 			continue;
1068 		}
1069 		mtx_unlock(&vnode_free_list_mtx);
1070 		done = 0;
1071 		ofreevnodes = freevnodes;
1072 		onumvnodes = numvnodes;
1073 		/*
1074 		 * Calculate parameters for recycling.  These are the same
1075 		 * throughout the loop to give some semblance of fairness.
1076 		 * The trigger point is to avoid recycling vnodes with lots
1077 		 * of resident pages.  We aren't trying to free memory; we
1078 		 * are trying to recycle or at least free vnodes.
1079 		 */
1080 		if (numvnodes <= desiredvnodes)
1081 			usevnodes = numvnodes - freevnodes;
1082 		else
1083 			usevnodes = numvnodes;
1084 		if (usevnodes <= 0)
1085 			usevnodes = 1;
1086 		/*
1087 		 * The trigger value is is chosen to give a conservatively
1088 		 * large value to ensure that it alone doesn't prevent
1089 		 * making progress.  The value can easily be so large that
1090 		 * it is effectively infinite in some congested and
1091 		 * misconfigured cases, and this is necessary.  Normally
1092 		 * it is about 8 to 100 (pages), which is quite large.
1093 		 */
1094 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1095 		if (force < 2)
1096 			trigger = vsmalltrigger;
1097 		reclaim_nc_src = force >= 3;
1098 		mtx_lock(&mountlist_mtx);
1099 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1100 			if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) {
1101 				nmp = TAILQ_NEXT(mp, mnt_list);
1102 				continue;
1103 			}
1104 			done += vlrureclaim(mp, reclaim_nc_src, trigger);
1105 			mtx_lock(&mountlist_mtx);
1106 			nmp = TAILQ_NEXT(mp, mnt_list);
1107 			vfs_unbusy(mp);
1108 		}
1109 		mtx_unlock(&mountlist_mtx);
1110 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1111 			uma_reclaim();
1112 		if (done == 0) {
1113 			if (force == 0 || force == 1) {
1114 				force = 2;
1115 				continue;
1116 			}
1117 			if (force == 2) {
1118 				force = 3;
1119 				continue;
1120 			}
1121 			force = 0;
1122 			vnlru_nowhere++;
1123 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1124 		} else
1125 			kern_yield(PRI_USER);
1126 		/*
1127 		 * After becoming active to expand above low water, keep
1128 		 * active until above high water.
1129 		 */
1130 		force = vspace() < vhiwat;
1131 	}
1132 }
1133 
1134 static struct kproc_desc vnlru_kp = {
1135 	"vnlru",
1136 	vnlru_proc,
1137 	&vnlruproc
1138 };
1139 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1140     &vnlru_kp);
1141 
1142 /*
1143  * Routines having to do with the management of the vnode table.
1144  */
1145 
1146 /*
1147  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1148  * before we actually vgone().  This function must be called with the vnode
1149  * held to prevent the vnode from being returned to the free list midway
1150  * through vgone().
1151  */
1152 static int
1153 vtryrecycle(struct vnode *vp)
1154 {
1155 	struct mount *vnmp;
1156 
1157 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1158 	VNASSERT(vp->v_holdcnt, vp,
1159 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1160 	/*
1161 	 * This vnode may found and locked via some other list, if so we
1162 	 * can't recycle it yet.
1163 	 */
1164 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1165 		CTR2(KTR_VFS,
1166 		    "%s: impossible to recycle, vp %p lock is already held",
1167 		    __func__, vp);
1168 		return (EWOULDBLOCK);
1169 	}
1170 	/*
1171 	 * Don't recycle if its filesystem is being suspended.
1172 	 */
1173 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1174 		VOP_UNLOCK(vp, 0);
1175 		CTR2(KTR_VFS,
1176 		    "%s: impossible to recycle, cannot start the write for %p",
1177 		    __func__, vp);
1178 		return (EBUSY);
1179 	}
1180 	/*
1181 	 * If we got this far, we need to acquire the interlock and see if
1182 	 * anyone picked up this vnode from another list.  If not, we will
1183 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1184 	 * will skip over it.
1185 	 */
1186 	VI_LOCK(vp);
1187 	if (vp->v_usecount) {
1188 		VOP_UNLOCK(vp, LK_INTERLOCK);
1189 		vn_finished_write(vnmp);
1190 		CTR2(KTR_VFS,
1191 		    "%s: impossible to recycle, %p is already referenced",
1192 		    __func__, vp);
1193 		return (EBUSY);
1194 	}
1195 	if ((vp->v_iflag & VI_DOOMED) == 0) {
1196 		atomic_add_long(&recycles_count, 1);
1197 		vgonel(vp);
1198 	}
1199 	VOP_UNLOCK(vp, LK_INTERLOCK);
1200 	vn_finished_write(vnmp);
1201 	return (0);
1202 }
1203 
1204 static void
1205 vcheckspace(void)
1206 {
1207 
1208 	if (vspace() < vlowat && vnlruproc_sig == 0) {
1209 		vnlruproc_sig = 1;
1210 		wakeup(vnlruproc);
1211 	}
1212 }
1213 
1214 /*
1215  * Wait if necessary for space for a new vnode.
1216  */
1217 static int
1218 getnewvnode_wait(int suspended)
1219 {
1220 
1221 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1222 	if (numvnodes >= desiredvnodes) {
1223 		if (suspended) {
1224 			/*
1225 			 * The file system is being suspended.  We cannot
1226 			 * risk a deadlock here, so allow allocation of
1227 			 * another vnode even if this would give too many.
1228 			 */
1229 			return (0);
1230 		}
1231 		if (vnlruproc_sig == 0) {
1232 			vnlruproc_sig = 1;	/* avoid unnecessary wakeups */
1233 			wakeup(vnlruproc);
1234 		}
1235 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
1236 		    "vlruwk", hz);
1237 	}
1238 	/* Post-adjust like the pre-adjust in getnewvnode(). */
1239 	if (numvnodes + 1 > desiredvnodes && freevnodes > 1)
1240 		vnlru_free(1);
1241 	return (numvnodes >= desiredvnodes ? ENFILE : 0);
1242 }
1243 
1244 /*
1245  * This hack is fragile, and probably not needed any more now that the
1246  * watermark handling works.
1247  */
1248 void
1249 getnewvnode_reserve(u_int count)
1250 {
1251 	struct thread *td;
1252 
1253 	/* Pre-adjust like the pre-adjust in getnewvnode(), with any count. */
1254 	/* XXX no longer so quick, but this part is not racy. */
1255 	mtx_lock(&vnode_free_list_mtx);
1256 	if (numvnodes + count > desiredvnodes && freevnodes > wantfreevnodes)
1257 		vnlru_free(ulmin(numvnodes + count - desiredvnodes,
1258 		    freevnodes - wantfreevnodes));
1259 	mtx_unlock(&vnode_free_list_mtx);
1260 
1261 	td = curthread;
1262 	/* First try to be quick and racy. */
1263 	if (atomic_fetchadd_long(&numvnodes, count) + count <= desiredvnodes) {
1264 		td->td_vp_reserv += count;
1265 		vcheckspace();	/* XXX no longer so quick, but more racy */
1266 		return;
1267 	} else
1268 		atomic_subtract_long(&numvnodes, count);
1269 
1270 	mtx_lock(&vnode_free_list_mtx);
1271 	while (count > 0) {
1272 		if (getnewvnode_wait(0) == 0) {
1273 			count--;
1274 			td->td_vp_reserv++;
1275 			atomic_add_long(&numvnodes, 1);
1276 		}
1277 	}
1278 	vcheckspace();
1279 	mtx_unlock(&vnode_free_list_mtx);
1280 }
1281 
1282 /*
1283  * This hack is fragile, especially if desiredvnodes or wantvnodes are
1284  * misconfgured or changed significantly.  Reducing desiredvnodes below
1285  * the reserved amount should cause bizarre behaviour like reducing it
1286  * below the number of active vnodes -- the system will try to reduce
1287  * numvnodes to match, but should fail, so the subtraction below should
1288  * not overflow.
1289  */
1290 void
1291 getnewvnode_drop_reserve(void)
1292 {
1293 	struct thread *td;
1294 
1295 	td = curthread;
1296 	atomic_subtract_long(&numvnodes, td->td_vp_reserv);
1297 	td->td_vp_reserv = 0;
1298 }
1299 
1300 /*
1301  * Return the next vnode from the free list.
1302  */
1303 int
1304 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1305     struct vnode **vpp)
1306 {
1307 	struct vnode *vp;
1308 	struct thread *td;
1309 	struct lock_object *lo;
1310 	static int cyclecount;
1311 	int error;
1312 
1313 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1314 	vp = NULL;
1315 	td = curthread;
1316 	if (td->td_vp_reserv > 0) {
1317 		td->td_vp_reserv -= 1;
1318 		goto alloc;
1319 	}
1320 	mtx_lock(&vnode_free_list_mtx);
1321 	if (numvnodes < desiredvnodes)
1322 		cyclecount = 0;
1323 	else if (cyclecount++ >= freevnodes) {
1324 		cyclecount = 0;
1325 		vstir = 1;
1326 	}
1327 	/*
1328 	 * Grow the vnode cache if it will not be above its target max
1329 	 * after growing.  Otherwise, if the free list is nonempty, try
1330 	 * to reclaim 1 item from it before growing the cache (possibly
1331 	 * above its target max if the reclamation failed or is delayed).
1332 	 * Otherwise, wait for some space.  In all cases, schedule
1333 	 * vnlru_proc() if we are getting short of space.  The watermarks
1334 	 * should be chosen so that we never wait or even reclaim from
1335 	 * the free list to below its target minimum.
1336 	 */
1337 	if (numvnodes + 1 <= desiredvnodes)
1338 		;
1339 	else if (freevnodes > 0)
1340 		vnlru_free(1);
1341 	else {
1342 		error = getnewvnode_wait(mp != NULL && (mp->mnt_kern_flag &
1343 		    MNTK_SUSPEND));
1344 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
1345 		if (error != 0) {
1346 			mtx_unlock(&vnode_free_list_mtx);
1347 			return (error);
1348 		}
1349 #endif
1350 	}
1351 	vcheckspace();
1352 	atomic_add_long(&numvnodes, 1);
1353 	mtx_unlock(&vnode_free_list_mtx);
1354 alloc:
1355 	atomic_add_long(&vnodes_created, 1);
1356 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK);
1357 	/*
1358 	 * Locks are given the generic name "vnode" when created.
1359 	 * Follow the historic practice of using the filesystem
1360 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1361 	 *
1362 	 * Locks live in a witness group keyed on their name. Thus,
1363 	 * when a lock is renamed, it must also move from the witness
1364 	 * group of its old name to the witness group of its new name.
1365 	 *
1366 	 * The change only needs to be made when the vnode moves
1367 	 * from one filesystem type to another. We ensure that each
1368 	 * filesystem use a single static name pointer for its tag so
1369 	 * that we can compare pointers rather than doing a strcmp().
1370 	 */
1371 	lo = &vp->v_vnlock->lock_object;
1372 	if (lo->lo_name != tag) {
1373 		lo->lo_name = tag;
1374 		WITNESS_DESTROY(lo);
1375 		WITNESS_INIT(lo, tag);
1376 	}
1377 	/*
1378 	 * By default, don't allow shared locks unless filesystems opt-in.
1379 	 */
1380 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1381 	/*
1382 	 * Finalize various vnode identity bits.
1383 	 */
1384 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1385 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1386 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1387 	vp->v_type = VNON;
1388 	vp->v_tag = tag;
1389 	vp->v_op = vops;
1390 	v_init_counters(vp);
1391 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1392 #ifdef MAC
1393 	mac_vnode_init(vp);
1394 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1395 		mac_vnode_associate_singlelabel(mp, vp);
1396 	else if (mp == NULL && vops != &dead_vnodeops)
1397 		printf("NULL mp in getnewvnode()\n");
1398 #endif
1399 	if (mp != NULL) {
1400 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1401 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1402 			vp->v_vflag |= VV_NOKNOTE;
1403 	}
1404 
1405 	/*
1406 	 * For the filesystems which do not use vfs_hash_insert(),
1407 	 * still initialize v_hash to have vfs_hash_index() useful.
1408 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1409 	 * its own hashing.
1410 	 */
1411 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1412 
1413 	*vpp = vp;
1414 	return (0);
1415 }
1416 
1417 /*
1418  * Delete from old mount point vnode list, if on one.
1419  */
1420 static void
1421 delmntque(struct vnode *vp)
1422 {
1423 	struct mount *mp;
1424 	int active;
1425 
1426 	mp = vp->v_mount;
1427 	if (mp == NULL)
1428 		return;
1429 	MNT_ILOCK(mp);
1430 	VI_LOCK(vp);
1431 	KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize,
1432 	    ("Active vnode list size %d > Vnode list size %d",
1433 	     mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize));
1434 	active = vp->v_iflag & VI_ACTIVE;
1435 	vp->v_iflag &= ~VI_ACTIVE;
1436 	if (active) {
1437 		mtx_lock(&vnode_free_list_mtx);
1438 		TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist);
1439 		mp->mnt_activevnodelistsize--;
1440 		mtx_unlock(&vnode_free_list_mtx);
1441 	}
1442 	vp->v_mount = NULL;
1443 	VI_UNLOCK(vp);
1444 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1445 		("bad mount point vnode list size"));
1446 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1447 	mp->mnt_nvnodelistsize--;
1448 	MNT_REL(mp);
1449 	MNT_IUNLOCK(mp);
1450 }
1451 
1452 static void
1453 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1454 {
1455 
1456 	vp->v_data = NULL;
1457 	vp->v_op = &dead_vnodeops;
1458 	vgone(vp);
1459 	vput(vp);
1460 }
1461 
1462 /*
1463  * Insert into list of vnodes for the new mount point, if available.
1464  */
1465 int
1466 insmntque1(struct vnode *vp, struct mount *mp,
1467 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1468 {
1469 
1470 	KASSERT(vp->v_mount == NULL,
1471 		("insmntque: vnode already on per mount vnode list"));
1472 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1473 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1474 
1475 	/*
1476 	 * We acquire the vnode interlock early to ensure that the
1477 	 * vnode cannot be recycled by another process releasing a
1478 	 * holdcnt on it before we get it on both the vnode list
1479 	 * and the active vnode list. The mount mutex protects only
1480 	 * manipulation of the vnode list and the vnode freelist
1481 	 * mutex protects only manipulation of the active vnode list.
1482 	 * Hence the need to hold the vnode interlock throughout.
1483 	 */
1484 	MNT_ILOCK(mp);
1485 	VI_LOCK(vp);
1486 	if (((mp->mnt_kern_flag & MNTK_NOINSMNTQ) != 0 &&
1487 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1488 	    mp->mnt_nvnodelistsize == 0)) &&
1489 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1490 		VI_UNLOCK(vp);
1491 		MNT_IUNLOCK(mp);
1492 		if (dtr != NULL)
1493 			dtr(vp, dtr_arg);
1494 		return (EBUSY);
1495 	}
1496 	vp->v_mount = mp;
1497 	MNT_REF(mp);
1498 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1499 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1500 		("neg mount point vnode list size"));
1501 	mp->mnt_nvnodelistsize++;
1502 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1503 	    ("Activating already active vnode"));
1504 	vp->v_iflag |= VI_ACTIVE;
1505 	mtx_lock(&vnode_free_list_mtx);
1506 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
1507 	mp->mnt_activevnodelistsize++;
1508 	mtx_unlock(&vnode_free_list_mtx);
1509 	VI_UNLOCK(vp);
1510 	MNT_IUNLOCK(mp);
1511 	return (0);
1512 }
1513 
1514 int
1515 insmntque(struct vnode *vp, struct mount *mp)
1516 {
1517 
1518 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1519 }
1520 
1521 /*
1522  * Flush out and invalidate all buffers associated with a bufobj
1523  * Called with the underlying object locked.
1524  */
1525 int
1526 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1527 {
1528 	int error;
1529 
1530 	BO_LOCK(bo);
1531 	if (flags & V_SAVE) {
1532 		error = bufobj_wwait(bo, slpflag, slptimeo);
1533 		if (error) {
1534 			BO_UNLOCK(bo);
1535 			return (error);
1536 		}
1537 		if (bo->bo_dirty.bv_cnt > 0) {
1538 			BO_UNLOCK(bo);
1539 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1540 				return (error);
1541 			/*
1542 			 * XXX We could save a lock/unlock if this was only
1543 			 * enabled under INVARIANTS
1544 			 */
1545 			BO_LOCK(bo);
1546 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1547 				panic("vinvalbuf: dirty bufs");
1548 		}
1549 	}
1550 	/*
1551 	 * If you alter this loop please notice that interlock is dropped and
1552 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1553 	 * no race conditions occur from this.
1554 	 */
1555 	do {
1556 		error = flushbuflist(&bo->bo_clean,
1557 		    flags, bo, slpflag, slptimeo);
1558 		if (error == 0 && !(flags & V_CLEANONLY))
1559 			error = flushbuflist(&bo->bo_dirty,
1560 			    flags, bo, slpflag, slptimeo);
1561 		if (error != 0 && error != EAGAIN) {
1562 			BO_UNLOCK(bo);
1563 			return (error);
1564 		}
1565 	} while (error != 0);
1566 
1567 	/*
1568 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1569 	 * have write I/O in-progress but if there is a VM object then the
1570 	 * VM object can also have read-I/O in-progress.
1571 	 */
1572 	do {
1573 		bufobj_wwait(bo, 0, 0);
1574 		BO_UNLOCK(bo);
1575 		if (bo->bo_object != NULL) {
1576 			VM_OBJECT_WLOCK(bo->bo_object);
1577 			vm_object_pip_wait(bo->bo_object, "bovlbx");
1578 			VM_OBJECT_WUNLOCK(bo->bo_object);
1579 		}
1580 		BO_LOCK(bo);
1581 	} while (bo->bo_numoutput > 0);
1582 	BO_UNLOCK(bo);
1583 
1584 	/*
1585 	 * Destroy the copy in the VM cache, too.
1586 	 */
1587 	if (bo->bo_object != NULL &&
1588 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY)) == 0) {
1589 		VM_OBJECT_WLOCK(bo->bo_object);
1590 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1591 		    OBJPR_CLEANONLY : 0);
1592 		VM_OBJECT_WUNLOCK(bo->bo_object);
1593 	}
1594 
1595 #ifdef INVARIANTS
1596 	BO_LOCK(bo);
1597 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY)) == 0 &&
1598 	    (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0))
1599 		panic("vinvalbuf: flush failed");
1600 	BO_UNLOCK(bo);
1601 #endif
1602 	return (0);
1603 }
1604 
1605 /*
1606  * Flush out and invalidate all buffers associated with a vnode.
1607  * Called with the underlying object locked.
1608  */
1609 int
1610 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1611 {
1612 
1613 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1614 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1615 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1616 		return (0);
1617 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1618 }
1619 
1620 /*
1621  * Flush out buffers on the specified list.
1622  *
1623  */
1624 static int
1625 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1626     int slptimeo)
1627 {
1628 	struct buf *bp, *nbp;
1629 	int retval, error;
1630 	daddr_t lblkno;
1631 	b_xflags_t xflags;
1632 
1633 	ASSERT_BO_WLOCKED(bo);
1634 
1635 	retval = 0;
1636 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1637 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1638 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1639 			continue;
1640 		}
1641 		lblkno = 0;
1642 		xflags = 0;
1643 		if (nbp != NULL) {
1644 			lblkno = nbp->b_lblkno;
1645 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
1646 		}
1647 		retval = EAGAIN;
1648 		error = BUF_TIMELOCK(bp,
1649 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
1650 		    "flushbuf", slpflag, slptimeo);
1651 		if (error) {
1652 			BO_LOCK(bo);
1653 			return (error != ENOLCK ? error : EAGAIN);
1654 		}
1655 		KASSERT(bp->b_bufobj == bo,
1656 		    ("bp %p wrong b_bufobj %p should be %p",
1657 		    bp, bp->b_bufobj, bo));
1658 		/*
1659 		 * XXX Since there are no node locks for NFS, I
1660 		 * believe there is a slight chance that a delayed
1661 		 * write will occur while sleeping just above, so
1662 		 * check for it.
1663 		 */
1664 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1665 		    (flags & V_SAVE)) {
1666 			bremfree(bp);
1667 			bp->b_flags |= B_ASYNC;
1668 			bwrite(bp);
1669 			BO_LOCK(bo);
1670 			return (EAGAIN);	/* XXX: why not loop ? */
1671 		}
1672 		bremfree(bp);
1673 		bp->b_flags |= (B_INVAL | B_RELBUF);
1674 		bp->b_flags &= ~B_ASYNC;
1675 		brelse(bp);
1676 		BO_LOCK(bo);
1677 		nbp = gbincore(bo, lblkno);
1678 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1679 		    != xflags)
1680 			break;			/* nbp invalid */
1681 	}
1682 	return (retval);
1683 }
1684 
1685 int
1686 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
1687 {
1688 	struct buf *bp;
1689 	int error;
1690 	daddr_t lblkno;
1691 
1692 	ASSERT_BO_LOCKED(bo);
1693 
1694 	for (lblkno = startn;;) {
1695 again:
1696 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
1697 		if (bp == NULL || bp->b_lblkno >= endn ||
1698 		    bp->b_lblkno < startn)
1699 			break;
1700 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
1701 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
1702 		if (error != 0) {
1703 			BO_RLOCK(bo);
1704 			if (error == ENOLCK)
1705 				goto again;
1706 			return (error);
1707 		}
1708 		KASSERT(bp->b_bufobj == bo,
1709 		    ("bp %p wrong b_bufobj %p should be %p",
1710 		    bp, bp->b_bufobj, bo));
1711 		lblkno = bp->b_lblkno + 1;
1712 		if ((bp->b_flags & B_MANAGED) == 0)
1713 			bremfree(bp);
1714 		bp->b_flags |= B_RELBUF;
1715 		/*
1716 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
1717 		 * pages backing each buffer in the range are unlikely to be
1718 		 * reused.  Dirty buffers will have the hint applied once
1719 		 * they've been written.
1720 		 */
1721 		if (bp->b_vp->v_object != NULL)
1722 			bp->b_flags |= B_NOREUSE;
1723 		brelse(bp);
1724 		BO_RLOCK(bo);
1725 	}
1726 	return (0);
1727 }
1728 
1729 /*
1730  * Truncate a file's buffer and pages to a specified length.  This
1731  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1732  * sync activity.
1733  */
1734 int
1735 vtruncbuf(struct vnode *vp, struct ucred *cred, off_t length, int blksize)
1736 {
1737 	struct buf *bp, *nbp;
1738 	int anyfreed;
1739 	int trunclbn;
1740 	struct bufobj *bo;
1741 
1742 	CTR5(KTR_VFS, "%s: vp %p with cred %p and block %d:%ju", __func__,
1743 	    vp, cred, blksize, (uintmax_t)length);
1744 
1745 	/*
1746 	 * Round up to the *next* lbn.
1747 	 */
1748 	trunclbn = howmany(length, blksize);
1749 
1750 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1751 restart:
1752 	bo = &vp->v_bufobj;
1753 	BO_LOCK(bo);
1754 	anyfreed = 1;
1755 	for (;anyfreed;) {
1756 		anyfreed = 0;
1757 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1758 			if (bp->b_lblkno < trunclbn)
1759 				continue;
1760 			if (BUF_LOCK(bp,
1761 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1762 			    BO_LOCKPTR(bo)) == ENOLCK)
1763 				goto restart;
1764 
1765 			bremfree(bp);
1766 			bp->b_flags |= (B_INVAL | B_RELBUF);
1767 			bp->b_flags &= ~B_ASYNC;
1768 			brelse(bp);
1769 			anyfreed = 1;
1770 
1771 			BO_LOCK(bo);
1772 			if (nbp != NULL &&
1773 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1774 			    (nbp->b_vp != vp) ||
1775 			    (nbp->b_flags & B_DELWRI))) {
1776 				BO_UNLOCK(bo);
1777 				goto restart;
1778 			}
1779 		}
1780 
1781 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1782 			if (bp->b_lblkno < trunclbn)
1783 				continue;
1784 			if (BUF_LOCK(bp,
1785 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1786 			    BO_LOCKPTR(bo)) == ENOLCK)
1787 				goto restart;
1788 			bremfree(bp);
1789 			bp->b_flags |= (B_INVAL | B_RELBUF);
1790 			bp->b_flags &= ~B_ASYNC;
1791 			brelse(bp);
1792 			anyfreed = 1;
1793 
1794 			BO_LOCK(bo);
1795 			if (nbp != NULL &&
1796 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1797 			    (nbp->b_vp != vp) ||
1798 			    (nbp->b_flags & B_DELWRI) == 0)) {
1799 				BO_UNLOCK(bo);
1800 				goto restart;
1801 			}
1802 		}
1803 	}
1804 
1805 	if (length > 0) {
1806 restartsync:
1807 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1808 			if (bp->b_lblkno > 0)
1809 				continue;
1810 			/*
1811 			 * Since we hold the vnode lock this should only
1812 			 * fail if we're racing with the buf daemon.
1813 			 */
1814 			if (BUF_LOCK(bp,
1815 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1816 			    BO_LOCKPTR(bo)) == ENOLCK) {
1817 				goto restart;
1818 			}
1819 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1820 			    ("buf(%p) on dirty queue without DELWRI", bp));
1821 
1822 			bremfree(bp);
1823 			bawrite(bp);
1824 			BO_LOCK(bo);
1825 			goto restartsync;
1826 		}
1827 	}
1828 
1829 	bufobj_wwait(bo, 0, 0);
1830 	BO_UNLOCK(bo);
1831 	vnode_pager_setsize(vp, length);
1832 
1833 	return (0);
1834 }
1835 
1836 static void
1837 buf_vlist_remove(struct buf *bp)
1838 {
1839 	struct bufv *bv;
1840 
1841 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1842 	ASSERT_BO_WLOCKED(bp->b_bufobj);
1843 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
1844 	    (BX_VNDIRTY|BX_VNCLEAN),
1845 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
1846 	if (bp->b_xflags & BX_VNDIRTY)
1847 		bv = &bp->b_bufobj->bo_dirty;
1848 	else
1849 		bv = &bp->b_bufobj->bo_clean;
1850 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
1851 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1852 	bv->bv_cnt--;
1853 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1854 }
1855 
1856 /*
1857  * Add the buffer to the sorted clean or dirty block list.
1858  *
1859  * NOTE: xflags is passed as a constant, optimizing this inline function!
1860  */
1861 static void
1862 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1863 {
1864 	struct bufv *bv;
1865 	struct buf *n;
1866 	int error;
1867 
1868 	ASSERT_BO_WLOCKED(bo);
1869 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
1870 	    ("dead bo %p", bo));
1871 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1872 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
1873 	bp->b_xflags |= xflags;
1874 	if (xflags & BX_VNDIRTY)
1875 		bv = &bo->bo_dirty;
1876 	else
1877 		bv = &bo->bo_clean;
1878 
1879 	/*
1880 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
1881 	 * we tend to grow at the tail so lookup_le should usually be cheaper
1882 	 * than _ge.
1883 	 */
1884 	if (bv->bv_cnt == 0 ||
1885 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
1886 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1887 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
1888 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
1889 	else
1890 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
1891 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
1892 	if (error)
1893 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
1894 	bv->bv_cnt++;
1895 }
1896 
1897 /*
1898  * Look up a buffer using the buffer tries.
1899  */
1900 struct buf *
1901 gbincore(struct bufobj *bo, daddr_t lblkno)
1902 {
1903 	struct buf *bp;
1904 
1905 	ASSERT_BO_LOCKED(bo);
1906 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
1907 	if (bp != NULL)
1908 		return (bp);
1909 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
1910 }
1911 
1912 /*
1913  * Associate a buffer with a vnode.
1914  */
1915 void
1916 bgetvp(struct vnode *vp, struct buf *bp)
1917 {
1918 	struct bufobj *bo;
1919 
1920 	bo = &vp->v_bufobj;
1921 	ASSERT_BO_WLOCKED(bo);
1922 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
1923 
1924 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
1925 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
1926 	    ("bgetvp: bp already attached! %p", bp));
1927 
1928 	vhold(vp);
1929 	bp->b_vp = vp;
1930 	bp->b_bufobj = bo;
1931 	/*
1932 	 * Insert onto list for new vnode.
1933 	 */
1934 	buf_vlist_add(bp, bo, BX_VNCLEAN);
1935 }
1936 
1937 /*
1938  * Disassociate a buffer from a vnode.
1939  */
1940 void
1941 brelvp(struct buf *bp)
1942 {
1943 	struct bufobj *bo;
1944 	struct vnode *vp;
1945 
1946 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1947 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1948 
1949 	/*
1950 	 * Delete from old vnode list, if on one.
1951 	 */
1952 	vp = bp->b_vp;		/* XXX */
1953 	bo = bp->b_bufobj;
1954 	BO_LOCK(bo);
1955 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1956 		buf_vlist_remove(bp);
1957 	else
1958 		panic("brelvp: Buffer %p not on queue.", bp);
1959 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1960 		bo->bo_flag &= ~BO_ONWORKLST;
1961 		mtx_lock(&sync_mtx);
1962 		LIST_REMOVE(bo, bo_synclist);
1963 		syncer_worklist_len--;
1964 		mtx_unlock(&sync_mtx);
1965 	}
1966 	bp->b_vp = NULL;
1967 	bp->b_bufobj = NULL;
1968 	BO_UNLOCK(bo);
1969 	vdrop(vp);
1970 }
1971 
1972 /*
1973  * Add an item to the syncer work queue.
1974  */
1975 static void
1976 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
1977 {
1978 	int slot;
1979 
1980 	ASSERT_BO_WLOCKED(bo);
1981 
1982 	mtx_lock(&sync_mtx);
1983 	if (bo->bo_flag & BO_ONWORKLST)
1984 		LIST_REMOVE(bo, bo_synclist);
1985 	else {
1986 		bo->bo_flag |= BO_ONWORKLST;
1987 		syncer_worklist_len++;
1988 	}
1989 
1990 	if (delay > syncer_maxdelay - 2)
1991 		delay = syncer_maxdelay - 2;
1992 	slot = (syncer_delayno + delay) & syncer_mask;
1993 
1994 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
1995 	mtx_unlock(&sync_mtx);
1996 }
1997 
1998 static int
1999 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2000 {
2001 	int error, len;
2002 
2003 	mtx_lock(&sync_mtx);
2004 	len = syncer_worklist_len - sync_vnode_count;
2005 	mtx_unlock(&sync_mtx);
2006 	error = SYSCTL_OUT(req, &len, sizeof(len));
2007 	return (error);
2008 }
2009 
2010 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
2011     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2012 
2013 static struct proc *updateproc;
2014 static void sched_sync(void);
2015 static struct kproc_desc up_kp = {
2016 	"syncer",
2017 	sched_sync,
2018 	&updateproc
2019 };
2020 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2021 
2022 static int
2023 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2024 {
2025 	struct vnode *vp;
2026 	struct mount *mp;
2027 
2028 	*bo = LIST_FIRST(slp);
2029 	if (*bo == NULL)
2030 		return (0);
2031 	vp = (*bo)->__bo_vnode;	/* XXX */
2032 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2033 		return (1);
2034 	/*
2035 	 * We use vhold in case the vnode does not
2036 	 * successfully sync.  vhold prevents the vnode from
2037 	 * going away when we unlock the sync_mtx so that
2038 	 * we can acquire the vnode interlock.
2039 	 */
2040 	vholdl(vp);
2041 	mtx_unlock(&sync_mtx);
2042 	VI_UNLOCK(vp);
2043 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2044 		vdrop(vp);
2045 		mtx_lock(&sync_mtx);
2046 		return (*bo == LIST_FIRST(slp));
2047 	}
2048 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2049 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2050 	VOP_UNLOCK(vp, 0);
2051 	vn_finished_write(mp);
2052 	BO_LOCK(*bo);
2053 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2054 		/*
2055 		 * Put us back on the worklist.  The worklist
2056 		 * routine will remove us from our current
2057 		 * position and then add us back in at a later
2058 		 * position.
2059 		 */
2060 		vn_syncer_add_to_worklist(*bo, syncdelay);
2061 	}
2062 	BO_UNLOCK(*bo);
2063 	vdrop(vp);
2064 	mtx_lock(&sync_mtx);
2065 	return (0);
2066 }
2067 
2068 static int first_printf = 1;
2069 
2070 /*
2071  * System filesystem synchronizer daemon.
2072  */
2073 static void
2074 sched_sync(void)
2075 {
2076 	struct synclist *next, *slp;
2077 	struct bufobj *bo;
2078 	long starttime;
2079 	struct thread *td = curthread;
2080 	int last_work_seen;
2081 	int net_worklist_len;
2082 	int syncer_final_iter;
2083 	int error;
2084 
2085 	last_work_seen = 0;
2086 	syncer_final_iter = 0;
2087 	syncer_state = SYNCER_RUNNING;
2088 	starttime = time_uptime;
2089 	td->td_pflags |= TDP_NORUNNINGBUF;
2090 
2091 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2092 	    SHUTDOWN_PRI_LAST);
2093 
2094 	mtx_lock(&sync_mtx);
2095 	for (;;) {
2096 		if (syncer_state == SYNCER_FINAL_DELAY &&
2097 		    syncer_final_iter == 0) {
2098 			mtx_unlock(&sync_mtx);
2099 			kproc_suspend_check(td->td_proc);
2100 			mtx_lock(&sync_mtx);
2101 		}
2102 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2103 		if (syncer_state != SYNCER_RUNNING &&
2104 		    starttime != time_uptime) {
2105 			if (first_printf) {
2106 				printf("\nSyncing disks, vnodes remaining...");
2107 				first_printf = 0;
2108 			}
2109 			printf("%d ", net_worklist_len);
2110 		}
2111 		starttime = time_uptime;
2112 
2113 		/*
2114 		 * Push files whose dirty time has expired.  Be careful
2115 		 * of interrupt race on slp queue.
2116 		 *
2117 		 * Skip over empty worklist slots when shutting down.
2118 		 */
2119 		do {
2120 			slp = &syncer_workitem_pending[syncer_delayno];
2121 			syncer_delayno += 1;
2122 			if (syncer_delayno == syncer_maxdelay)
2123 				syncer_delayno = 0;
2124 			next = &syncer_workitem_pending[syncer_delayno];
2125 			/*
2126 			 * If the worklist has wrapped since the
2127 			 * it was emptied of all but syncer vnodes,
2128 			 * switch to the FINAL_DELAY state and run
2129 			 * for one more second.
2130 			 */
2131 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2132 			    net_worklist_len == 0 &&
2133 			    last_work_seen == syncer_delayno) {
2134 				syncer_state = SYNCER_FINAL_DELAY;
2135 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2136 			}
2137 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2138 		    syncer_worklist_len > 0);
2139 
2140 		/*
2141 		 * Keep track of the last time there was anything
2142 		 * on the worklist other than syncer vnodes.
2143 		 * Return to the SHUTTING_DOWN state if any
2144 		 * new work appears.
2145 		 */
2146 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2147 			last_work_seen = syncer_delayno;
2148 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2149 			syncer_state = SYNCER_SHUTTING_DOWN;
2150 		while (!LIST_EMPTY(slp)) {
2151 			error = sync_vnode(slp, &bo, td);
2152 			if (error == 1) {
2153 				LIST_REMOVE(bo, bo_synclist);
2154 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2155 				continue;
2156 			}
2157 
2158 			if (first_printf == 0) {
2159 				/*
2160 				 * Drop the sync mutex, because some watchdog
2161 				 * drivers need to sleep while patting
2162 				 */
2163 				mtx_unlock(&sync_mtx);
2164 				wdog_kern_pat(WD_LASTVAL);
2165 				mtx_lock(&sync_mtx);
2166 			}
2167 
2168 		}
2169 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2170 			syncer_final_iter--;
2171 		/*
2172 		 * The variable rushjob allows the kernel to speed up the
2173 		 * processing of the filesystem syncer process. A rushjob
2174 		 * value of N tells the filesystem syncer to process the next
2175 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2176 		 * is used by the soft update code to speed up the filesystem
2177 		 * syncer process when the incore state is getting so far
2178 		 * ahead of the disk that the kernel memory pool is being
2179 		 * threatened with exhaustion.
2180 		 */
2181 		if (rushjob > 0) {
2182 			rushjob -= 1;
2183 			continue;
2184 		}
2185 		/*
2186 		 * Just sleep for a short period of time between
2187 		 * iterations when shutting down to allow some I/O
2188 		 * to happen.
2189 		 *
2190 		 * If it has taken us less than a second to process the
2191 		 * current work, then wait. Otherwise start right over
2192 		 * again. We can still lose time if any single round
2193 		 * takes more than two seconds, but it does not really
2194 		 * matter as we are just trying to generally pace the
2195 		 * filesystem activity.
2196 		 */
2197 		if (syncer_state != SYNCER_RUNNING ||
2198 		    time_uptime == starttime) {
2199 			thread_lock(td);
2200 			sched_prio(td, PPAUSE);
2201 			thread_unlock(td);
2202 		}
2203 		if (syncer_state != SYNCER_RUNNING)
2204 			cv_timedwait(&sync_wakeup, &sync_mtx,
2205 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2206 		else if (time_uptime == starttime)
2207 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2208 	}
2209 }
2210 
2211 /*
2212  * Request the syncer daemon to speed up its work.
2213  * We never push it to speed up more than half of its
2214  * normal turn time, otherwise it could take over the cpu.
2215  */
2216 int
2217 speedup_syncer(void)
2218 {
2219 	int ret = 0;
2220 
2221 	mtx_lock(&sync_mtx);
2222 	if (rushjob < syncdelay / 2) {
2223 		rushjob += 1;
2224 		stat_rush_requests += 1;
2225 		ret = 1;
2226 	}
2227 	mtx_unlock(&sync_mtx);
2228 	cv_broadcast(&sync_wakeup);
2229 	return (ret);
2230 }
2231 
2232 /*
2233  * Tell the syncer to speed up its work and run though its work
2234  * list several times, then tell it to shut down.
2235  */
2236 static void
2237 syncer_shutdown(void *arg, int howto)
2238 {
2239 
2240 	if (howto & RB_NOSYNC)
2241 		return;
2242 	mtx_lock(&sync_mtx);
2243 	syncer_state = SYNCER_SHUTTING_DOWN;
2244 	rushjob = 0;
2245 	mtx_unlock(&sync_mtx);
2246 	cv_broadcast(&sync_wakeup);
2247 	kproc_shutdown(arg, howto);
2248 }
2249 
2250 void
2251 syncer_suspend(void)
2252 {
2253 
2254 	syncer_shutdown(updateproc, 0);
2255 }
2256 
2257 void
2258 syncer_resume(void)
2259 {
2260 
2261 	mtx_lock(&sync_mtx);
2262 	first_printf = 1;
2263 	syncer_state = SYNCER_RUNNING;
2264 	mtx_unlock(&sync_mtx);
2265 	cv_broadcast(&sync_wakeup);
2266 	kproc_resume(updateproc);
2267 }
2268 
2269 /*
2270  * Reassign a buffer from one vnode to another.
2271  * Used to assign file specific control information
2272  * (indirect blocks) to the vnode to which they belong.
2273  */
2274 void
2275 reassignbuf(struct buf *bp)
2276 {
2277 	struct vnode *vp;
2278 	struct bufobj *bo;
2279 	int delay;
2280 #ifdef INVARIANTS
2281 	struct bufv *bv;
2282 #endif
2283 
2284 	vp = bp->b_vp;
2285 	bo = bp->b_bufobj;
2286 	++reassignbufcalls;
2287 
2288 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2289 	    bp, bp->b_vp, bp->b_flags);
2290 	/*
2291 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2292 	 * is not fully linked in.
2293 	 */
2294 	if (bp->b_flags & B_PAGING)
2295 		panic("cannot reassign paging buffer");
2296 
2297 	/*
2298 	 * Delete from old vnode list, if on one.
2299 	 */
2300 	BO_LOCK(bo);
2301 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2302 		buf_vlist_remove(bp);
2303 	else
2304 		panic("reassignbuf: Buffer %p not on queue.", bp);
2305 	/*
2306 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2307 	 * of clean buffers.
2308 	 */
2309 	if (bp->b_flags & B_DELWRI) {
2310 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2311 			switch (vp->v_type) {
2312 			case VDIR:
2313 				delay = dirdelay;
2314 				break;
2315 			case VCHR:
2316 				delay = metadelay;
2317 				break;
2318 			default:
2319 				delay = filedelay;
2320 			}
2321 			vn_syncer_add_to_worklist(bo, delay);
2322 		}
2323 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2324 	} else {
2325 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2326 
2327 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2328 			mtx_lock(&sync_mtx);
2329 			LIST_REMOVE(bo, bo_synclist);
2330 			syncer_worklist_len--;
2331 			mtx_unlock(&sync_mtx);
2332 			bo->bo_flag &= ~BO_ONWORKLST;
2333 		}
2334 	}
2335 #ifdef INVARIANTS
2336 	bv = &bo->bo_clean;
2337 	bp = TAILQ_FIRST(&bv->bv_hd);
2338 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2339 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2340 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2341 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2342 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2343 	bv = &bo->bo_dirty;
2344 	bp = TAILQ_FIRST(&bv->bv_hd);
2345 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2346 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2347 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2348 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2349 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2350 #endif
2351 	BO_UNLOCK(bo);
2352 }
2353 
2354 /*
2355  * A temporary hack until refcount_* APIs are sorted out.
2356  */
2357 static __inline int
2358 vfs_refcount_acquire_if_not_zero(volatile u_int *count)
2359 {
2360 	u_int old;
2361 
2362 	for (;;) {
2363 		old = *count;
2364 		if (old == 0)
2365 			return (0);
2366 		if (atomic_cmpset_int(count, old, old + 1))
2367 			return (1);
2368 	}
2369 }
2370 
2371 static __inline int
2372 vfs_refcount_release_if_not_last(volatile u_int *count)
2373 {
2374 	u_int old;
2375 
2376 	for (;;) {
2377 		old = *count;
2378 		if (old == 1)
2379 			return (0);
2380 		if (atomic_cmpset_int(count, old, old - 1))
2381 			return (1);
2382 	}
2383 }
2384 
2385 static void
2386 v_init_counters(struct vnode *vp)
2387 {
2388 
2389 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2390 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2391 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2392 
2393 	refcount_init(&vp->v_holdcnt, 1);
2394 	refcount_init(&vp->v_usecount, 1);
2395 }
2396 
2397 static void
2398 v_incr_usecount_locked(struct vnode *vp)
2399 {
2400 
2401 	ASSERT_VI_LOCKED(vp, __func__);
2402 	if ((vp->v_iflag & VI_OWEINACT) != 0) {
2403 		VNASSERT(vp->v_usecount == 0, vp,
2404 		    ("vnode with usecount and VI_OWEINACT set"));
2405 		vp->v_iflag &= ~VI_OWEINACT;
2406 	}
2407 	refcount_acquire(&vp->v_usecount);
2408 	v_incr_devcount(vp);
2409 }
2410 
2411 /*
2412  * Increment the use and hold counts on the vnode, taking care to reference
2413  * the driver's usecount if this is a chardev.  The _vhold() will remove
2414  * the vnode from the free list if it is presently free.
2415  */
2416 static void
2417 v_incr_usecount(struct vnode *vp)
2418 {
2419 
2420 	ASSERT_VI_UNLOCKED(vp, __func__);
2421 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2422 
2423 	if (vp->v_type != VCHR &&
2424 	    vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) {
2425 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2426 		    ("vnode with usecount and VI_OWEINACT set"));
2427 	} else {
2428 		VI_LOCK(vp);
2429 		v_incr_usecount_locked(vp);
2430 		VI_UNLOCK(vp);
2431 	}
2432 }
2433 
2434 /*
2435  * Increment si_usecount of the associated device, if any.
2436  */
2437 static void
2438 v_incr_devcount(struct vnode *vp)
2439 {
2440 
2441 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2442 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2443 		dev_lock();
2444 		vp->v_rdev->si_usecount++;
2445 		dev_unlock();
2446 	}
2447 }
2448 
2449 /*
2450  * Decrement si_usecount of the associated device, if any.
2451  */
2452 static void
2453 v_decr_devcount(struct vnode *vp)
2454 {
2455 
2456 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2457 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2458 		dev_lock();
2459 		vp->v_rdev->si_usecount--;
2460 		dev_unlock();
2461 	}
2462 }
2463 
2464 /*
2465  * Grab a particular vnode from the free list, increment its
2466  * reference count and lock it.  VI_DOOMED is set if the vnode
2467  * is being destroyed.  Only callers who specify LK_RETRY will
2468  * see doomed vnodes.  If inactive processing was delayed in
2469  * vput try to do it here.
2470  *
2471  * Notes on lockless counter manipulation:
2472  * _vhold, vputx and other routines make various decisions based
2473  * on either holdcnt or usecount being 0. As long as either contuner
2474  * is not transitioning 0->1 nor 1->0, the manipulation can be done
2475  * with atomic operations. Otherwise the interlock is taken.
2476  */
2477 int
2478 vget(struct vnode *vp, int flags, struct thread *td)
2479 {
2480 	int error, oweinact;
2481 
2482 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2483 	    ("vget: invalid lock operation"));
2484 
2485 	if ((flags & LK_INTERLOCK) != 0)
2486 		ASSERT_VI_LOCKED(vp, __func__);
2487 	else
2488 		ASSERT_VI_UNLOCKED(vp, __func__);
2489 	if ((flags & LK_VNHELD) != 0)
2490 		VNASSERT((vp->v_holdcnt > 0), vp,
2491 		    ("vget: LK_VNHELD passed but vnode not held"));
2492 
2493 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2494 
2495 	if ((flags & LK_VNHELD) == 0)
2496 		_vhold(vp, (flags & LK_INTERLOCK) != 0);
2497 
2498 	if ((error = vn_lock(vp, flags)) != 0) {
2499 		vdrop(vp);
2500 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2501 		    vp);
2502 		return (error);
2503 	}
2504 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
2505 		panic("vget: vn_lock failed to return ENOENT\n");
2506 	/*
2507 	 * We don't guarantee that any particular close will
2508 	 * trigger inactive processing so just make a best effort
2509 	 * here at preventing a reference to a removed file.  If
2510 	 * we don't succeed no harm is done.
2511 	 *
2512 	 * Upgrade our holdcnt to a usecount.
2513 	 */
2514 	if (vp->v_type != VCHR &&
2515 	    vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) {
2516 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2517 		    ("vnode with usecount and VI_OWEINACT set"));
2518 	} else {
2519 		VI_LOCK(vp);
2520 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2521 			oweinact = 0;
2522 		} else {
2523 			oweinact = 1;
2524 			vp->v_iflag &= ~VI_OWEINACT;
2525 		}
2526 		refcount_acquire(&vp->v_usecount);
2527 		v_incr_devcount(vp);
2528 		if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2529 		    (flags & LK_NOWAIT) == 0)
2530 			vinactive(vp, td);
2531 		VI_UNLOCK(vp);
2532 	}
2533 	return (0);
2534 }
2535 
2536 /*
2537  * Increase the reference count of a vnode.
2538  */
2539 void
2540 vref(struct vnode *vp)
2541 {
2542 
2543 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2544 	_vhold(vp, false);
2545 	v_incr_usecount(vp);
2546 }
2547 
2548 void
2549 vrefl(struct vnode *vp)
2550 {
2551 
2552 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2553 	_vhold(vp, true);
2554 	v_incr_usecount_locked(vp);
2555 }
2556 
2557 /*
2558  * Return reference count of a vnode.
2559  *
2560  * The results of this call are only guaranteed when some mechanism is used to
2561  * stop other processes from gaining references to the vnode.  This may be the
2562  * case if the caller holds the only reference.  This is also useful when stale
2563  * data is acceptable as race conditions may be accounted for by some other
2564  * means.
2565  */
2566 int
2567 vrefcnt(struct vnode *vp)
2568 {
2569 
2570 	return (vp->v_usecount);
2571 }
2572 
2573 #define	VPUTX_VRELE	1
2574 #define	VPUTX_VPUT	2
2575 #define	VPUTX_VUNREF	3
2576 
2577 /*
2578  * Decrement the use and hold counts for a vnode.
2579  *
2580  * See an explanation near vget() as to why atomic operation is safe.
2581  */
2582 static void
2583 vputx(struct vnode *vp, int func)
2584 {
2585 	int error;
2586 
2587 	KASSERT(vp != NULL, ("vputx: null vp"));
2588 	if (func == VPUTX_VUNREF)
2589 		ASSERT_VOP_LOCKED(vp, "vunref");
2590 	else if (func == VPUTX_VPUT)
2591 		ASSERT_VOP_LOCKED(vp, "vput");
2592 	else
2593 		KASSERT(func == VPUTX_VRELE, ("vputx: wrong func"));
2594 	ASSERT_VI_UNLOCKED(vp, __func__);
2595 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2596 
2597 	if (vp->v_type != VCHR &&
2598 	    vfs_refcount_release_if_not_last(&vp->v_usecount)) {
2599 		if (func == VPUTX_VPUT)
2600 			VOP_UNLOCK(vp, 0);
2601 		vdrop(vp);
2602 		return;
2603 	}
2604 
2605 	VI_LOCK(vp);
2606 
2607 	/*
2608 	 * We want to hold the vnode until the inactive finishes to
2609 	 * prevent vgone() races.  We drop the use count here and the
2610 	 * hold count below when we're done.
2611 	 */
2612 	if (!refcount_release(&vp->v_usecount) ||
2613 	    (vp->v_iflag & VI_DOINGINACT)) {
2614 		if (func == VPUTX_VPUT)
2615 			VOP_UNLOCK(vp, 0);
2616 		v_decr_devcount(vp);
2617 		vdropl(vp);
2618 		return;
2619 	}
2620 
2621 	v_decr_devcount(vp);
2622 
2623 	error = 0;
2624 
2625 	if (vp->v_usecount != 0) {
2626 		vprint("vputx: usecount not zero", vp);
2627 		panic("vputx: usecount not zero");
2628 	}
2629 
2630 	CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp);
2631 
2632 	/*
2633 	 * We must call VOP_INACTIVE with the node locked. Mark
2634 	 * as VI_DOINGINACT to avoid recursion.
2635 	 */
2636 	vp->v_iflag |= VI_OWEINACT;
2637 	switch (func) {
2638 	case VPUTX_VRELE:
2639 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
2640 		VI_LOCK(vp);
2641 		break;
2642 	case VPUTX_VPUT:
2643 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2644 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
2645 			    LK_NOWAIT);
2646 			VI_LOCK(vp);
2647 		}
2648 		break;
2649 	case VPUTX_VUNREF:
2650 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2651 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
2652 			VI_LOCK(vp);
2653 		}
2654 		break;
2655 	}
2656 	VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
2657 	    ("vnode with usecount and VI_OWEINACT set"));
2658 	if (error == 0) {
2659 		if (vp->v_iflag & VI_OWEINACT)
2660 			vinactive(vp, curthread);
2661 		if (func != VPUTX_VUNREF)
2662 			VOP_UNLOCK(vp, 0);
2663 	}
2664 	vdropl(vp);
2665 }
2666 
2667 /*
2668  * Vnode put/release.
2669  * If count drops to zero, call inactive routine and return to freelist.
2670  */
2671 void
2672 vrele(struct vnode *vp)
2673 {
2674 
2675 	vputx(vp, VPUTX_VRELE);
2676 }
2677 
2678 /*
2679  * Release an already locked vnode.  This give the same effects as
2680  * unlock+vrele(), but takes less time and avoids releasing and
2681  * re-aquiring the lock (as vrele() acquires the lock internally.)
2682  */
2683 void
2684 vput(struct vnode *vp)
2685 {
2686 
2687 	vputx(vp, VPUTX_VPUT);
2688 }
2689 
2690 /*
2691  * Release an exclusively locked vnode. Do not unlock the vnode lock.
2692  */
2693 void
2694 vunref(struct vnode *vp)
2695 {
2696 
2697 	vputx(vp, VPUTX_VUNREF);
2698 }
2699 
2700 /*
2701  * Increase the hold count and activate if this is the first reference.
2702  */
2703 void
2704 _vhold(struct vnode *vp, bool locked)
2705 {
2706 	struct mount *mp;
2707 
2708 	if (locked)
2709 		ASSERT_VI_LOCKED(vp, __func__);
2710 	else
2711 		ASSERT_VI_UNLOCKED(vp, __func__);
2712 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2713 	if (!locked && vfs_refcount_acquire_if_not_zero(&vp->v_holdcnt)) {
2714 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2715 		    ("_vhold: vnode with holdcnt is free"));
2716 		return;
2717 	}
2718 
2719 	if (!locked)
2720 		VI_LOCK(vp);
2721 	if ((vp->v_iflag & VI_FREE) == 0) {
2722 		refcount_acquire(&vp->v_holdcnt);
2723 		if (!locked)
2724 			VI_UNLOCK(vp);
2725 		return;
2726 	}
2727 	VNASSERT(vp->v_holdcnt == 0, vp,
2728 	    ("%s: wrong hold count", __func__));
2729 	VNASSERT(vp->v_op != NULL, vp,
2730 	    ("%s: vnode already reclaimed.", __func__));
2731 	/*
2732 	 * Remove a vnode from the free list, mark it as in use,
2733 	 * and put it on the active list.
2734 	 */
2735 	mtx_lock(&vnode_free_list_mtx);
2736 	TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
2737 	freevnodes--;
2738 	vp->v_iflag &= ~VI_FREE;
2739 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
2740 	    ("Activating already active vnode"));
2741 	vp->v_iflag |= VI_ACTIVE;
2742 	mp = vp->v_mount;
2743 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
2744 	mp->mnt_activevnodelistsize++;
2745 	mtx_unlock(&vnode_free_list_mtx);
2746 	refcount_acquire(&vp->v_holdcnt);
2747 	if (!locked)
2748 		VI_UNLOCK(vp);
2749 }
2750 
2751 /*
2752  * Drop the hold count of the vnode.  If this is the last reference to
2753  * the vnode we place it on the free list unless it has been vgone'd
2754  * (marked VI_DOOMED) in which case we will free it.
2755  *
2756  * Because the vnode vm object keeps a hold reference on the vnode if
2757  * there is at least one resident non-cached page, the vnode cannot
2758  * leave the active list without the page cleanup done.
2759  */
2760 void
2761 _vdrop(struct vnode *vp, bool locked)
2762 {
2763 	struct bufobj *bo;
2764 	struct mount *mp;
2765 	int active;
2766 
2767 	if (locked)
2768 		ASSERT_VI_LOCKED(vp, __func__);
2769 	else
2770 		ASSERT_VI_UNLOCKED(vp, __func__);
2771 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2772 	if ((int)vp->v_holdcnt <= 0)
2773 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
2774 	if (vfs_refcount_release_if_not_last(&vp->v_holdcnt)) {
2775 		if (locked)
2776 			VI_UNLOCK(vp);
2777 		return;
2778 	}
2779 
2780 	if (!locked)
2781 		VI_LOCK(vp);
2782 	if (refcount_release(&vp->v_holdcnt) == 0) {
2783 		VI_UNLOCK(vp);
2784 		return;
2785 	}
2786 	if ((vp->v_iflag & VI_DOOMED) == 0) {
2787 		/*
2788 		 * Mark a vnode as free: remove it from its active list
2789 		 * and put it up for recycling on the freelist.
2790 		 */
2791 		VNASSERT(vp->v_op != NULL, vp,
2792 		    ("vdropl: vnode already reclaimed."));
2793 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2794 		    ("vnode already free"));
2795 		VNASSERT(vp->v_holdcnt == 0, vp,
2796 		    ("vdropl: freeing when we shouldn't"));
2797 		active = vp->v_iflag & VI_ACTIVE;
2798 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2799 			vp->v_iflag &= ~VI_ACTIVE;
2800 			mp = vp->v_mount;
2801 			mtx_lock(&vnode_free_list_mtx);
2802 			if (active) {
2803 				TAILQ_REMOVE(&mp->mnt_activevnodelist, vp,
2804 				    v_actfreelist);
2805 				mp->mnt_activevnodelistsize--;
2806 			}
2807 			TAILQ_INSERT_TAIL(&vnode_free_list, vp,
2808 			    v_actfreelist);
2809 			freevnodes++;
2810 			vp->v_iflag |= VI_FREE;
2811 			mtx_unlock(&vnode_free_list_mtx);
2812 		} else {
2813 			atomic_add_long(&free_owe_inact, 1);
2814 		}
2815 		VI_UNLOCK(vp);
2816 		return;
2817 	}
2818 	/*
2819 	 * The vnode has been marked for destruction, so free it.
2820 	 *
2821 	 * The vnode will be returned to the zone where it will
2822 	 * normally remain until it is needed for another vnode. We
2823 	 * need to cleanup (or verify that the cleanup has already
2824 	 * been done) any residual data left from its current use
2825 	 * so as not to contaminate the freshly allocated vnode.
2826 	 */
2827 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
2828 	atomic_subtract_long(&numvnodes, 1);
2829 	bo = &vp->v_bufobj;
2830 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2831 	    ("cleaned vnode still on the free list."));
2832 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
2833 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
2834 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
2835 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
2836 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
2837 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
2838 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
2839 	    ("clean blk trie not empty"));
2840 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
2841 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
2842 	    ("dirty blk trie not empty"));
2843 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
2844 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
2845 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
2846 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
2847 	    ("Dangling rangelock waiters"));
2848 	VI_UNLOCK(vp);
2849 #ifdef MAC
2850 	mac_vnode_destroy(vp);
2851 #endif
2852 	if (vp->v_pollinfo != NULL) {
2853 		destroy_vpollinfo(vp->v_pollinfo);
2854 		vp->v_pollinfo = NULL;
2855 	}
2856 #ifdef INVARIANTS
2857 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
2858 	vp->v_op = NULL;
2859 #endif
2860 	bzero(&vp->v_un, sizeof(vp->v_un));
2861 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
2862 	vp->v_iflag = 0;
2863 	vp->v_vflag = 0;
2864 	bo->bo_flag = 0;
2865 	uma_zfree(vnode_zone, vp);
2866 }
2867 
2868 /*
2869  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
2870  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
2871  * OWEINACT tracks whether a vnode missed a call to inactive due to a
2872  * failed lock upgrade.
2873  */
2874 void
2875 vinactive(struct vnode *vp, struct thread *td)
2876 {
2877 	struct vm_object *obj;
2878 
2879 	ASSERT_VOP_ELOCKED(vp, "vinactive");
2880 	ASSERT_VI_LOCKED(vp, "vinactive");
2881 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2882 	    ("vinactive: recursed on VI_DOINGINACT"));
2883 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2884 	vp->v_iflag |= VI_DOINGINACT;
2885 	vp->v_iflag &= ~VI_OWEINACT;
2886 	VI_UNLOCK(vp);
2887 	/*
2888 	 * Before moving off the active list, we must be sure that any
2889 	 * modified pages are converted into the vnode's dirty
2890 	 * buffers, since these will no longer be checked once the
2891 	 * vnode is on the inactive list.
2892 	 *
2893 	 * The write-out of the dirty pages is asynchronous.  At the
2894 	 * point that VOP_INACTIVE() is called, there could still be
2895 	 * pending I/O and dirty pages in the object.
2896 	 */
2897 	obj = vp->v_object;
2898 	if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0) {
2899 		VM_OBJECT_WLOCK(obj);
2900 		vm_object_page_clean(obj, 0, 0, OBJPC_NOSYNC);
2901 		VM_OBJECT_WUNLOCK(obj);
2902 	}
2903 	VOP_INACTIVE(vp, td);
2904 	VI_LOCK(vp);
2905 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2906 	    ("vinactive: lost VI_DOINGINACT"));
2907 	vp->v_iflag &= ~VI_DOINGINACT;
2908 }
2909 
2910 /*
2911  * Remove any vnodes in the vnode table belonging to mount point mp.
2912  *
2913  * If FORCECLOSE is not specified, there should not be any active ones,
2914  * return error if any are found (nb: this is a user error, not a
2915  * system error). If FORCECLOSE is specified, detach any active vnodes
2916  * that are found.
2917  *
2918  * If WRITECLOSE is set, only flush out regular file vnodes open for
2919  * writing.
2920  *
2921  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2922  *
2923  * `rootrefs' specifies the base reference count for the root vnode
2924  * of this filesystem. The root vnode is considered busy if its
2925  * v_usecount exceeds this value. On a successful return, vflush(, td)
2926  * will call vrele() on the root vnode exactly rootrefs times.
2927  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2928  * be zero.
2929  */
2930 #ifdef DIAGNOSTIC
2931 static int busyprt = 0;		/* print out busy vnodes */
2932 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
2933 #endif
2934 
2935 int
2936 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
2937 {
2938 	struct vnode *vp, *mvp, *rootvp = NULL;
2939 	struct vattr vattr;
2940 	int busy = 0, error;
2941 
2942 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
2943 	    rootrefs, flags);
2944 	if (rootrefs > 0) {
2945 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2946 		    ("vflush: bad args"));
2947 		/*
2948 		 * Get the filesystem root vnode. We can vput() it
2949 		 * immediately, since with rootrefs > 0, it won't go away.
2950 		 */
2951 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
2952 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
2953 			    __func__, error);
2954 			return (error);
2955 		}
2956 		vput(rootvp);
2957 	}
2958 loop:
2959 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
2960 		vholdl(vp);
2961 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
2962 		if (error) {
2963 			vdrop(vp);
2964 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
2965 			goto loop;
2966 		}
2967 		/*
2968 		 * Skip over a vnodes marked VV_SYSTEM.
2969 		 */
2970 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2971 			VOP_UNLOCK(vp, 0);
2972 			vdrop(vp);
2973 			continue;
2974 		}
2975 		/*
2976 		 * If WRITECLOSE is set, flush out unlinked but still open
2977 		 * files (even if open only for reading) and regular file
2978 		 * vnodes open for writing.
2979 		 */
2980 		if (flags & WRITECLOSE) {
2981 			if (vp->v_object != NULL) {
2982 				VM_OBJECT_WLOCK(vp->v_object);
2983 				vm_object_page_clean(vp->v_object, 0, 0, 0);
2984 				VM_OBJECT_WUNLOCK(vp->v_object);
2985 			}
2986 			error = VOP_FSYNC(vp, MNT_WAIT, td);
2987 			if (error != 0) {
2988 				VOP_UNLOCK(vp, 0);
2989 				vdrop(vp);
2990 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
2991 				return (error);
2992 			}
2993 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
2994 			VI_LOCK(vp);
2995 
2996 			if ((vp->v_type == VNON ||
2997 			    (error == 0 && vattr.va_nlink > 0)) &&
2998 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2999 				VOP_UNLOCK(vp, 0);
3000 				vdropl(vp);
3001 				continue;
3002 			}
3003 		} else
3004 			VI_LOCK(vp);
3005 		/*
3006 		 * With v_usecount == 0, all we need to do is clear out the
3007 		 * vnode data structures and we are done.
3008 		 *
3009 		 * If FORCECLOSE is set, forcibly close the vnode.
3010 		 */
3011 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3012 			vgonel(vp);
3013 		} else {
3014 			busy++;
3015 #ifdef DIAGNOSTIC
3016 			if (busyprt)
3017 				vprint("vflush: busy vnode", vp);
3018 #endif
3019 		}
3020 		VOP_UNLOCK(vp, 0);
3021 		vdropl(vp);
3022 	}
3023 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3024 		/*
3025 		 * If just the root vnode is busy, and if its refcount
3026 		 * is equal to `rootrefs', then go ahead and kill it.
3027 		 */
3028 		VI_LOCK(rootvp);
3029 		KASSERT(busy > 0, ("vflush: not busy"));
3030 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3031 		    ("vflush: usecount %d < rootrefs %d",
3032 		     rootvp->v_usecount, rootrefs));
3033 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3034 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3035 			vgone(rootvp);
3036 			VOP_UNLOCK(rootvp, 0);
3037 			busy = 0;
3038 		} else
3039 			VI_UNLOCK(rootvp);
3040 	}
3041 	if (busy) {
3042 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3043 		    busy);
3044 		return (EBUSY);
3045 	}
3046 	for (; rootrefs > 0; rootrefs--)
3047 		vrele(rootvp);
3048 	return (0);
3049 }
3050 
3051 /*
3052  * Recycle an unused vnode to the front of the free list.
3053  */
3054 int
3055 vrecycle(struct vnode *vp)
3056 {
3057 	int recycled;
3058 
3059 	ASSERT_VOP_ELOCKED(vp, "vrecycle");
3060 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3061 	recycled = 0;
3062 	VI_LOCK(vp);
3063 	if (vp->v_usecount == 0) {
3064 		recycled = 1;
3065 		vgonel(vp);
3066 	}
3067 	VI_UNLOCK(vp);
3068 	return (recycled);
3069 }
3070 
3071 /*
3072  * Eliminate all activity associated with a vnode
3073  * in preparation for reuse.
3074  */
3075 void
3076 vgone(struct vnode *vp)
3077 {
3078 	VI_LOCK(vp);
3079 	vgonel(vp);
3080 	VI_UNLOCK(vp);
3081 }
3082 
3083 static void
3084 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3085     struct vnode *lowervp __unused)
3086 {
3087 }
3088 
3089 /*
3090  * Notify upper mounts about reclaimed or unlinked vnode.
3091  */
3092 void
3093 vfs_notify_upper(struct vnode *vp, int event)
3094 {
3095 	static struct vfsops vgonel_vfsops = {
3096 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3097 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3098 	};
3099 	struct mount *mp, *ump, *mmp;
3100 
3101 	mp = vp->v_mount;
3102 	if (mp == NULL)
3103 		return;
3104 
3105 	MNT_ILOCK(mp);
3106 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3107 		goto unlock;
3108 	MNT_IUNLOCK(mp);
3109 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3110 	mmp->mnt_op = &vgonel_vfsops;
3111 	mmp->mnt_kern_flag |= MNTK_MARKER;
3112 	MNT_ILOCK(mp);
3113 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3114 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3115 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3116 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3117 			continue;
3118 		}
3119 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3120 		MNT_IUNLOCK(mp);
3121 		switch (event) {
3122 		case VFS_NOTIFY_UPPER_RECLAIM:
3123 			VFS_RECLAIM_LOWERVP(ump, vp);
3124 			break;
3125 		case VFS_NOTIFY_UPPER_UNLINK:
3126 			VFS_UNLINK_LOWERVP(ump, vp);
3127 			break;
3128 		default:
3129 			KASSERT(0, ("invalid event %d", event));
3130 			break;
3131 		}
3132 		MNT_ILOCK(mp);
3133 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3134 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3135 	}
3136 	free(mmp, M_TEMP);
3137 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3138 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3139 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3140 		wakeup(&mp->mnt_uppers);
3141 	}
3142 unlock:
3143 	MNT_IUNLOCK(mp);
3144 }
3145 
3146 /*
3147  * vgone, with the vp interlock held.
3148  */
3149 static void
3150 vgonel(struct vnode *vp)
3151 {
3152 	struct thread *td;
3153 	int oweinact;
3154 	int active;
3155 	struct mount *mp;
3156 
3157 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3158 	ASSERT_VI_LOCKED(vp, "vgonel");
3159 	VNASSERT(vp->v_holdcnt, vp,
3160 	    ("vgonel: vp %p has no reference.", vp));
3161 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3162 	td = curthread;
3163 
3164 	/*
3165 	 * Don't vgonel if we're already doomed.
3166 	 */
3167 	if (vp->v_iflag & VI_DOOMED)
3168 		return;
3169 	vp->v_iflag |= VI_DOOMED;
3170 
3171 	/*
3172 	 * Check to see if the vnode is in use.  If so, we have to call
3173 	 * VOP_CLOSE() and VOP_INACTIVE().
3174 	 */
3175 	active = vp->v_usecount;
3176 	oweinact = (vp->v_iflag & VI_OWEINACT);
3177 	VI_UNLOCK(vp);
3178 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3179 
3180 	/*
3181 	 * If purging an active vnode, it must be closed and
3182 	 * deactivated before being reclaimed.
3183 	 */
3184 	if (active)
3185 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3186 	if (oweinact || active) {
3187 		VI_LOCK(vp);
3188 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
3189 			vinactive(vp, td);
3190 		VI_UNLOCK(vp);
3191 	}
3192 	if (vp->v_type == VSOCK)
3193 		vfs_unp_reclaim(vp);
3194 
3195 	/*
3196 	 * Clean out any buffers associated with the vnode.
3197 	 * If the flush fails, just toss the buffers.
3198 	 */
3199 	mp = NULL;
3200 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3201 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3202 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3203 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3204 			;
3205 	}
3206 
3207 	BO_LOCK(&vp->v_bufobj);
3208 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3209 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3210 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3211 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3212 	    ("vp %p bufobj not invalidated", vp));
3213 	vp->v_bufobj.bo_flag |= BO_DEAD;
3214 	BO_UNLOCK(&vp->v_bufobj);
3215 
3216 	/*
3217 	 * Reclaim the vnode.
3218 	 */
3219 	if (VOP_RECLAIM(vp, td))
3220 		panic("vgone: cannot reclaim");
3221 	if (mp != NULL)
3222 		vn_finished_secondary_write(mp);
3223 	VNASSERT(vp->v_object == NULL, vp,
3224 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
3225 	/*
3226 	 * Clear the advisory locks and wake up waiting threads.
3227 	 */
3228 	(void)VOP_ADVLOCKPURGE(vp);
3229 	vp->v_lockf = NULL;
3230 	/*
3231 	 * Delete from old mount point vnode list.
3232 	 */
3233 	delmntque(vp);
3234 	cache_purge(vp);
3235 	/*
3236 	 * Done with purge, reset to the standard lock and invalidate
3237 	 * the vnode.
3238 	 */
3239 	VI_LOCK(vp);
3240 	vp->v_vnlock = &vp->v_lock;
3241 	vp->v_op = &dead_vnodeops;
3242 	vp->v_tag = "none";
3243 	vp->v_type = VBAD;
3244 }
3245 
3246 /*
3247  * Calculate the total number of references to a special device.
3248  */
3249 int
3250 vcount(struct vnode *vp)
3251 {
3252 	int count;
3253 
3254 	dev_lock();
3255 	count = vp->v_rdev->si_usecount;
3256 	dev_unlock();
3257 	return (count);
3258 }
3259 
3260 /*
3261  * Same as above, but using the struct cdev *as argument
3262  */
3263 int
3264 count_dev(struct cdev *dev)
3265 {
3266 	int count;
3267 
3268 	dev_lock();
3269 	count = dev->si_usecount;
3270 	dev_unlock();
3271 	return(count);
3272 }
3273 
3274 /*
3275  * Print out a description of a vnode.
3276  */
3277 static char *typename[] =
3278 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3279  "VMARKER"};
3280 
3281 void
3282 vn_printf(struct vnode *vp, const char *fmt, ...)
3283 {
3284 	va_list ap;
3285 	char buf[256], buf2[16];
3286 	u_long flags;
3287 
3288 	va_start(ap, fmt);
3289 	vprintf(fmt, ap);
3290 	va_end(ap);
3291 	printf("%p: ", (void *)vp);
3292 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
3293 	printf("    usecount %d, writecount %d, refcount %d mountedhere %p\n",
3294 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere);
3295 	buf[0] = '\0';
3296 	buf[1] = '\0';
3297 	if (vp->v_vflag & VV_ROOT)
3298 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3299 	if (vp->v_vflag & VV_ISTTY)
3300 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3301 	if (vp->v_vflag & VV_NOSYNC)
3302 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3303 	if (vp->v_vflag & VV_ETERNALDEV)
3304 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3305 	if (vp->v_vflag & VV_CACHEDLABEL)
3306 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3307 	if (vp->v_vflag & VV_TEXT)
3308 		strlcat(buf, "|VV_TEXT", sizeof(buf));
3309 	if (vp->v_vflag & VV_COPYONWRITE)
3310 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3311 	if (vp->v_vflag & VV_SYSTEM)
3312 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3313 	if (vp->v_vflag & VV_PROCDEP)
3314 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3315 	if (vp->v_vflag & VV_NOKNOTE)
3316 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3317 	if (vp->v_vflag & VV_DELETED)
3318 		strlcat(buf, "|VV_DELETED", sizeof(buf));
3319 	if (vp->v_vflag & VV_MD)
3320 		strlcat(buf, "|VV_MD", sizeof(buf));
3321 	if (vp->v_vflag & VV_FORCEINSMQ)
3322 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
3323 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
3324 	    VV_CACHEDLABEL | VV_TEXT | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
3325 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
3326 	if (flags != 0) {
3327 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
3328 		strlcat(buf, buf2, sizeof(buf));
3329 	}
3330 	if (vp->v_iflag & VI_MOUNT)
3331 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
3332 	if (vp->v_iflag & VI_DOOMED)
3333 		strlcat(buf, "|VI_DOOMED", sizeof(buf));
3334 	if (vp->v_iflag & VI_FREE)
3335 		strlcat(buf, "|VI_FREE", sizeof(buf));
3336 	if (vp->v_iflag & VI_ACTIVE)
3337 		strlcat(buf, "|VI_ACTIVE", sizeof(buf));
3338 	if (vp->v_iflag & VI_DOINGINACT)
3339 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
3340 	if (vp->v_iflag & VI_OWEINACT)
3341 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
3342 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOOMED | VI_FREE |
3343 	    VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT);
3344 	if (flags != 0) {
3345 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
3346 		strlcat(buf, buf2, sizeof(buf));
3347 	}
3348 	printf("    flags (%s)\n", buf + 1);
3349 	if (mtx_owned(VI_MTX(vp)))
3350 		printf(" VI_LOCKed");
3351 	if (vp->v_object != NULL)
3352 		printf("    v_object %p ref %d pages %d "
3353 		    "cleanbuf %d dirtybuf %d\n",
3354 		    vp->v_object, vp->v_object->ref_count,
3355 		    vp->v_object->resident_page_count,
3356 		    vp->v_bufobj.bo_clean.bv_cnt,
3357 		    vp->v_bufobj.bo_dirty.bv_cnt);
3358 	printf("    ");
3359 	lockmgr_printinfo(vp->v_vnlock);
3360 	if (vp->v_data != NULL)
3361 		VOP_PRINT(vp);
3362 }
3363 
3364 #ifdef DDB
3365 /*
3366  * List all of the locked vnodes in the system.
3367  * Called when debugging the kernel.
3368  */
3369 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
3370 {
3371 	struct mount *mp;
3372 	struct vnode *vp;
3373 
3374 	/*
3375 	 * Note: because this is DDB, we can't obey the locking semantics
3376 	 * for these structures, which means we could catch an inconsistent
3377 	 * state and dereference a nasty pointer.  Not much to be done
3378 	 * about that.
3379 	 */
3380 	db_printf("Locked vnodes\n");
3381 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3382 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3383 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
3384 				vprint("", vp);
3385 		}
3386 	}
3387 }
3388 
3389 /*
3390  * Show details about the given vnode.
3391  */
3392 DB_SHOW_COMMAND(vnode, db_show_vnode)
3393 {
3394 	struct vnode *vp;
3395 
3396 	if (!have_addr)
3397 		return;
3398 	vp = (struct vnode *)addr;
3399 	vn_printf(vp, "vnode ");
3400 }
3401 
3402 /*
3403  * Show details about the given mount point.
3404  */
3405 DB_SHOW_COMMAND(mount, db_show_mount)
3406 {
3407 	struct mount *mp;
3408 	struct vfsopt *opt;
3409 	struct statfs *sp;
3410 	struct vnode *vp;
3411 	char buf[512];
3412 	uint64_t mflags;
3413 	u_int flags;
3414 
3415 	if (!have_addr) {
3416 		/* No address given, print short info about all mount points. */
3417 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3418 			db_printf("%p %s on %s (%s)\n", mp,
3419 			    mp->mnt_stat.f_mntfromname,
3420 			    mp->mnt_stat.f_mntonname,
3421 			    mp->mnt_stat.f_fstypename);
3422 			if (db_pager_quit)
3423 				break;
3424 		}
3425 		db_printf("\nMore info: show mount <addr>\n");
3426 		return;
3427 	}
3428 
3429 	mp = (struct mount *)addr;
3430 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
3431 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
3432 
3433 	buf[0] = '\0';
3434 	mflags = mp->mnt_flag;
3435 #define	MNT_FLAG(flag)	do {						\
3436 	if (mflags & (flag)) {						\
3437 		if (buf[0] != '\0')					\
3438 			strlcat(buf, ", ", sizeof(buf));		\
3439 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
3440 		mflags &= ~(flag);					\
3441 	}								\
3442 } while (0)
3443 	MNT_FLAG(MNT_RDONLY);
3444 	MNT_FLAG(MNT_SYNCHRONOUS);
3445 	MNT_FLAG(MNT_NOEXEC);
3446 	MNT_FLAG(MNT_NOSUID);
3447 	MNT_FLAG(MNT_NFS4ACLS);
3448 	MNT_FLAG(MNT_UNION);
3449 	MNT_FLAG(MNT_ASYNC);
3450 	MNT_FLAG(MNT_SUIDDIR);
3451 	MNT_FLAG(MNT_SOFTDEP);
3452 	MNT_FLAG(MNT_NOSYMFOLLOW);
3453 	MNT_FLAG(MNT_GJOURNAL);
3454 	MNT_FLAG(MNT_MULTILABEL);
3455 	MNT_FLAG(MNT_ACLS);
3456 	MNT_FLAG(MNT_NOATIME);
3457 	MNT_FLAG(MNT_NOCLUSTERR);
3458 	MNT_FLAG(MNT_NOCLUSTERW);
3459 	MNT_FLAG(MNT_SUJ);
3460 	MNT_FLAG(MNT_EXRDONLY);
3461 	MNT_FLAG(MNT_EXPORTED);
3462 	MNT_FLAG(MNT_DEFEXPORTED);
3463 	MNT_FLAG(MNT_EXPORTANON);
3464 	MNT_FLAG(MNT_EXKERB);
3465 	MNT_FLAG(MNT_EXPUBLIC);
3466 	MNT_FLAG(MNT_LOCAL);
3467 	MNT_FLAG(MNT_QUOTA);
3468 	MNT_FLAG(MNT_ROOTFS);
3469 	MNT_FLAG(MNT_USER);
3470 	MNT_FLAG(MNT_IGNORE);
3471 	MNT_FLAG(MNT_UPDATE);
3472 	MNT_FLAG(MNT_DELEXPORT);
3473 	MNT_FLAG(MNT_RELOAD);
3474 	MNT_FLAG(MNT_FORCE);
3475 	MNT_FLAG(MNT_SNAPSHOT);
3476 	MNT_FLAG(MNT_BYFSID);
3477 #undef MNT_FLAG
3478 	if (mflags != 0) {
3479 		if (buf[0] != '\0')
3480 			strlcat(buf, ", ", sizeof(buf));
3481 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3482 		    "0x%016jx", mflags);
3483 	}
3484 	db_printf("    mnt_flag = %s\n", buf);
3485 
3486 	buf[0] = '\0';
3487 	flags = mp->mnt_kern_flag;
3488 #define	MNT_KERN_FLAG(flag)	do {					\
3489 	if (flags & (flag)) {						\
3490 		if (buf[0] != '\0')					\
3491 			strlcat(buf, ", ", sizeof(buf));		\
3492 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
3493 		flags &= ~(flag);					\
3494 	}								\
3495 } while (0)
3496 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
3497 	MNT_KERN_FLAG(MNTK_ASYNC);
3498 	MNT_KERN_FLAG(MNTK_SOFTDEP);
3499 	MNT_KERN_FLAG(MNTK_NOINSMNTQ);
3500 	MNT_KERN_FLAG(MNTK_DRAINING);
3501 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
3502 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
3503 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
3504 	MNT_KERN_FLAG(MNTK_NO_IOPF);
3505 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
3506 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
3507 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
3508 	MNT_KERN_FLAG(MNTK_MARKER);
3509 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
3510 	MNT_KERN_FLAG(MNTK_NOASYNC);
3511 	MNT_KERN_FLAG(MNTK_UNMOUNT);
3512 	MNT_KERN_FLAG(MNTK_MWAIT);
3513 	MNT_KERN_FLAG(MNTK_SUSPEND);
3514 	MNT_KERN_FLAG(MNTK_SUSPEND2);
3515 	MNT_KERN_FLAG(MNTK_SUSPENDED);
3516 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
3517 	MNT_KERN_FLAG(MNTK_NOKNOTE);
3518 #undef MNT_KERN_FLAG
3519 	if (flags != 0) {
3520 		if (buf[0] != '\0')
3521 			strlcat(buf, ", ", sizeof(buf));
3522 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3523 		    "0x%08x", flags);
3524 	}
3525 	db_printf("    mnt_kern_flag = %s\n", buf);
3526 
3527 	db_printf("    mnt_opt = ");
3528 	opt = TAILQ_FIRST(mp->mnt_opt);
3529 	if (opt != NULL) {
3530 		db_printf("%s", opt->name);
3531 		opt = TAILQ_NEXT(opt, link);
3532 		while (opt != NULL) {
3533 			db_printf(", %s", opt->name);
3534 			opt = TAILQ_NEXT(opt, link);
3535 		}
3536 	}
3537 	db_printf("\n");
3538 
3539 	sp = &mp->mnt_stat;
3540 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
3541 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
3542 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
3543 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
3544 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
3545 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
3546 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
3547 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
3548 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
3549 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
3550 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
3551 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
3552 
3553 	db_printf("    mnt_cred = { uid=%u ruid=%u",
3554 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
3555 	if (jailed(mp->mnt_cred))
3556 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
3557 	db_printf(" }\n");
3558 	db_printf("    mnt_ref = %d\n", mp->mnt_ref);
3559 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
3560 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
3561 	db_printf("    mnt_activevnodelistsize = %d\n",
3562 	    mp->mnt_activevnodelistsize);
3563 	db_printf("    mnt_writeopcount = %d\n", mp->mnt_writeopcount);
3564 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
3565 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
3566 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
3567 	db_printf("    mnt_lockref = %d\n", mp->mnt_lockref);
3568 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
3569 	db_printf("    mnt_secondary_accwrites = %d\n",
3570 	    mp->mnt_secondary_accwrites);
3571 	db_printf("    mnt_gjprovider = %s\n",
3572 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
3573 
3574 	db_printf("\n\nList of active vnodes\n");
3575 	TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) {
3576 		if (vp->v_type != VMARKER) {
3577 			vn_printf(vp, "vnode ");
3578 			if (db_pager_quit)
3579 				break;
3580 		}
3581 	}
3582 	db_printf("\n\nList of inactive vnodes\n");
3583 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3584 		if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) {
3585 			vn_printf(vp, "vnode ");
3586 			if (db_pager_quit)
3587 				break;
3588 		}
3589 	}
3590 }
3591 #endif	/* DDB */
3592 
3593 /*
3594  * Fill in a struct xvfsconf based on a struct vfsconf.
3595  */
3596 static int
3597 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
3598 {
3599 	struct xvfsconf xvfsp;
3600 
3601 	bzero(&xvfsp, sizeof(xvfsp));
3602 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3603 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3604 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3605 	xvfsp.vfc_flags = vfsp->vfc_flags;
3606 	/*
3607 	 * These are unused in userland, we keep them
3608 	 * to not break binary compatibility.
3609 	 */
3610 	xvfsp.vfc_vfsops = NULL;
3611 	xvfsp.vfc_next = NULL;
3612 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3613 }
3614 
3615 #ifdef COMPAT_FREEBSD32
3616 struct xvfsconf32 {
3617 	uint32_t	vfc_vfsops;
3618 	char		vfc_name[MFSNAMELEN];
3619 	int32_t		vfc_typenum;
3620 	int32_t		vfc_refcount;
3621 	int32_t		vfc_flags;
3622 	uint32_t	vfc_next;
3623 };
3624 
3625 static int
3626 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
3627 {
3628 	struct xvfsconf32 xvfsp;
3629 
3630 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3631 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3632 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3633 	xvfsp.vfc_flags = vfsp->vfc_flags;
3634 	xvfsp.vfc_vfsops = 0;
3635 	xvfsp.vfc_next = 0;
3636 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3637 }
3638 #endif
3639 
3640 /*
3641  * Top level filesystem related information gathering.
3642  */
3643 static int
3644 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
3645 {
3646 	struct vfsconf *vfsp;
3647 	int error;
3648 
3649 	error = 0;
3650 	vfsconf_slock();
3651 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3652 #ifdef COMPAT_FREEBSD32
3653 		if (req->flags & SCTL_MASK32)
3654 			error = vfsconf2x32(req, vfsp);
3655 		else
3656 #endif
3657 			error = vfsconf2x(req, vfsp);
3658 		if (error)
3659 			break;
3660 	}
3661 	vfsconf_sunlock();
3662 	return (error);
3663 }
3664 
3665 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
3666     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
3667     "S,xvfsconf", "List of all configured filesystems");
3668 
3669 #ifndef BURN_BRIDGES
3670 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
3671 
3672 static int
3673 vfs_sysctl(SYSCTL_HANDLER_ARGS)
3674 {
3675 	int *name = (int *)arg1 - 1;	/* XXX */
3676 	u_int namelen = arg2 + 1;	/* XXX */
3677 	struct vfsconf *vfsp;
3678 
3679 	log(LOG_WARNING, "userland calling deprecated sysctl, "
3680 	    "please rebuild world\n");
3681 
3682 #if 1 || defined(COMPAT_PRELITE2)
3683 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
3684 	if (namelen == 1)
3685 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
3686 #endif
3687 
3688 	switch (name[1]) {
3689 	case VFS_MAXTYPENUM:
3690 		if (namelen != 2)
3691 			return (ENOTDIR);
3692 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
3693 	case VFS_CONF:
3694 		if (namelen != 3)
3695 			return (ENOTDIR);	/* overloaded */
3696 		vfsconf_slock();
3697 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3698 			if (vfsp->vfc_typenum == name[2])
3699 				break;
3700 		}
3701 		vfsconf_sunlock();
3702 		if (vfsp == NULL)
3703 			return (EOPNOTSUPP);
3704 #ifdef COMPAT_FREEBSD32
3705 		if (req->flags & SCTL_MASK32)
3706 			return (vfsconf2x32(req, vfsp));
3707 		else
3708 #endif
3709 			return (vfsconf2x(req, vfsp));
3710 	}
3711 	return (EOPNOTSUPP);
3712 }
3713 
3714 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
3715     CTLFLAG_MPSAFE, vfs_sysctl,
3716     "Generic filesystem");
3717 
3718 #if 1 || defined(COMPAT_PRELITE2)
3719 
3720 static int
3721 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
3722 {
3723 	int error;
3724 	struct vfsconf *vfsp;
3725 	struct ovfsconf ovfs;
3726 
3727 	vfsconf_slock();
3728 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3729 		bzero(&ovfs, sizeof(ovfs));
3730 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
3731 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
3732 		ovfs.vfc_index = vfsp->vfc_typenum;
3733 		ovfs.vfc_refcount = vfsp->vfc_refcount;
3734 		ovfs.vfc_flags = vfsp->vfc_flags;
3735 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
3736 		if (error != 0) {
3737 			vfsconf_sunlock();
3738 			return (error);
3739 		}
3740 	}
3741 	vfsconf_sunlock();
3742 	return (0);
3743 }
3744 
3745 #endif /* 1 || COMPAT_PRELITE2 */
3746 #endif /* !BURN_BRIDGES */
3747 
3748 #define KINFO_VNODESLOP		10
3749 #ifdef notyet
3750 /*
3751  * Dump vnode list (via sysctl).
3752  */
3753 /* ARGSUSED */
3754 static int
3755 sysctl_vnode(SYSCTL_HANDLER_ARGS)
3756 {
3757 	struct xvnode *xvn;
3758 	struct mount *mp;
3759 	struct vnode *vp;
3760 	int error, len, n;
3761 
3762 	/*
3763 	 * Stale numvnodes access is not fatal here.
3764 	 */
3765 	req->lock = 0;
3766 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
3767 	if (!req->oldptr)
3768 		/* Make an estimate */
3769 		return (SYSCTL_OUT(req, 0, len));
3770 
3771 	error = sysctl_wire_old_buffer(req, 0);
3772 	if (error != 0)
3773 		return (error);
3774 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
3775 	n = 0;
3776 	mtx_lock(&mountlist_mtx);
3777 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3778 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
3779 			continue;
3780 		MNT_ILOCK(mp);
3781 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3782 			if (n == len)
3783 				break;
3784 			vref(vp);
3785 			xvn[n].xv_size = sizeof *xvn;
3786 			xvn[n].xv_vnode = vp;
3787 			xvn[n].xv_id = 0;	/* XXX compat */
3788 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
3789 			XV_COPY(usecount);
3790 			XV_COPY(writecount);
3791 			XV_COPY(holdcnt);
3792 			XV_COPY(mount);
3793 			XV_COPY(numoutput);
3794 			XV_COPY(type);
3795 #undef XV_COPY
3796 			xvn[n].xv_flag = vp->v_vflag;
3797 
3798 			switch (vp->v_type) {
3799 			case VREG:
3800 			case VDIR:
3801 			case VLNK:
3802 				break;
3803 			case VBLK:
3804 			case VCHR:
3805 				if (vp->v_rdev == NULL) {
3806 					vrele(vp);
3807 					continue;
3808 				}
3809 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
3810 				break;
3811 			case VSOCK:
3812 				xvn[n].xv_socket = vp->v_socket;
3813 				break;
3814 			case VFIFO:
3815 				xvn[n].xv_fifo = vp->v_fifoinfo;
3816 				break;
3817 			case VNON:
3818 			case VBAD:
3819 			default:
3820 				/* shouldn't happen? */
3821 				vrele(vp);
3822 				continue;
3823 			}
3824 			vrele(vp);
3825 			++n;
3826 		}
3827 		MNT_IUNLOCK(mp);
3828 		mtx_lock(&mountlist_mtx);
3829 		vfs_unbusy(mp);
3830 		if (n == len)
3831 			break;
3832 	}
3833 	mtx_unlock(&mountlist_mtx);
3834 
3835 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
3836 	free(xvn, M_TEMP);
3837 	return (error);
3838 }
3839 
3840 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
3841     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
3842     "");
3843 #endif
3844 
3845 static void
3846 unmount_or_warn(struct mount *mp)
3847 {
3848 	int error;
3849 
3850 	error = dounmount(mp, MNT_FORCE, curthread);
3851 	if (error != 0) {
3852 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
3853 		if (error == EBUSY)
3854 			printf("BUSY)\n");
3855 		else
3856 			printf("%d)\n", error);
3857 	}
3858 }
3859 
3860 /*
3861  * Unmount all filesystems. The list is traversed in reverse order
3862  * of mounting to avoid dependencies.
3863  */
3864 void
3865 vfs_unmountall(void)
3866 {
3867 	struct mount *mp, *tmp;
3868 
3869 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
3870 
3871 	/*
3872 	 * Since this only runs when rebooting, it is not interlocked.
3873 	 */
3874 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
3875 		vfs_ref(mp);
3876 
3877 		/*
3878 		 * Forcibly unmounting "/dev" before "/" would prevent clean
3879 		 * unmount of the latter.
3880 		 */
3881 		if (mp == rootdevmp)
3882 			continue;
3883 
3884 		unmount_or_warn(mp);
3885 	}
3886 
3887 	if (rootdevmp != NULL)
3888 		unmount_or_warn(rootdevmp);
3889 }
3890 
3891 /*
3892  * perform msync on all vnodes under a mount point
3893  * the mount point must be locked.
3894  */
3895 void
3896 vfs_msync(struct mount *mp, int flags)
3897 {
3898 	struct vnode *vp, *mvp;
3899 	struct vm_object *obj;
3900 
3901 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
3902 	MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
3903 		obj = vp->v_object;
3904 		if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 &&
3905 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
3906 			if (!vget(vp,
3907 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
3908 			    curthread)) {
3909 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
3910 					vput(vp);
3911 					continue;
3912 				}
3913 
3914 				obj = vp->v_object;
3915 				if (obj != NULL) {
3916 					VM_OBJECT_WLOCK(obj);
3917 					vm_object_page_clean(obj, 0, 0,
3918 					    flags == MNT_WAIT ?
3919 					    OBJPC_SYNC : OBJPC_NOSYNC);
3920 					VM_OBJECT_WUNLOCK(obj);
3921 				}
3922 				vput(vp);
3923 			}
3924 		} else
3925 			VI_UNLOCK(vp);
3926 	}
3927 }
3928 
3929 static void
3930 destroy_vpollinfo_free(struct vpollinfo *vi)
3931 {
3932 
3933 	knlist_destroy(&vi->vpi_selinfo.si_note);
3934 	mtx_destroy(&vi->vpi_lock);
3935 	uma_zfree(vnodepoll_zone, vi);
3936 }
3937 
3938 static void
3939 destroy_vpollinfo(struct vpollinfo *vi)
3940 {
3941 
3942 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
3943 	seldrain(&vi->vpi_selinfo);
3944 	destroy_vpollinfo_free(vi);
3945 }
3946 
3947 /*
3948  * Initialize per-vnode helper structure to hold poll-related state.
3949  */
3950 void
3951 v_addpollinfo(struct vnode *vp)
3952 {
3953 	struct vpollinfo *vi;
3954 
3955 	if (vp->v_pollinfo != NULL)
3956 		return;
3957 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
3958 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
3959 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
3960 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
3961 	VI_LOCK(vp);
3962 	if (vp->v_pollinfo != NULL) {
3963 		VI_UNLOCK(vp);
3964 		destroy_vpollinfo_free(vi);
3965 		return;
3966 	}
3967 	vp->v_pollinfo = vi;
3968 	VI_UNLOCK(vp);
3969 }
3970 
3971 /*
3972  * Record a process's interest in events which might happen to
3973  * a vnode.  Because poll uses the historic select-style interface
3974  * internally, this routine serves as both the ``check for any
3975  * pending events'' and the ``record my interest in future events''
3976  * functions.  (These are done together, while the lock is held,
3977  * to avoid race conditions.)
3978  */
3979 int
3980 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
3981 {
3982 
3983 	v_addpollinfo(vp);
3984 	mtx_lock(&vp->v_pollinfo->vpi_lock);
3985 	if (vp->v_pollinfo->vpi_revents & events) {
3986 		/*
3987 		 * This leaves events we are not interested
3988 		 * in available for the other process which
3989 		 * which presumably had requested them
3990 		 * (otherwise they would never have been
3991 		 * recorded).
3992 		 */
3993 		events &= vp->v_pollinfo->vpi_revents;
3994 		vp->v_pollinfo->vpi_revents &= ~events;
3995 
3996 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
3997 		return (events);
3998 	}
3999 	vp->v_pollinfo->vpi_events |= events;
4000 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4001 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4002 	return (0);
4003 }
4004 
4005 /*
4006  * Routine to create and manage a filesystem syncer vnode.
4007  */
4008 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4009 static int	sync_fsync(struct  vop_fsync_args *);
4010 static int	sync_inactive(struct  vop_inactive_args *);
4011 static int	sync_reclaim(struct  vop_reclaim_args *);
4012 
4013 static struct vop_vector sync_vnodeops = {
4014 	.vop_bypass =	VOP_EOPNOTSUPP,
4015 	.vop_close =	sync_close,		/* close */
4016 	.vop_fsync =	sync_fsync,		/* fsync */
4017 	.vop_inactive =	sync_inactive,	/* inactive */
4018 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4019 	.vop_lock1 =	vop_stdlock,	/* lock */
4020 	.vop_unlock =	vop_stdunlock,	/* unlock */
4021 	.vop_islocked =	vop_stdislocked,	/* islocked */
4022 };
4023 
4024 /*
4025  * Create a new filesystem syncer vnode for the specified mount point.
4026  */
4027 void
4028 vfs_allocate_syncvnode(struct mount *mp)
4029 {
4030 	struct vnode *vp;
4031 	struct bufobj *bo;
4032 	static long start, incr, next;
4033 	int error;
4034 
4035 	/* Allocate a new vnode */
4036 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4037 	if (error != 0)
4038 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4039 	vp->v_type = VNON;
4040 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4041 	vp->v_vflag |= VV_FORCEINSMQ;
4042 	error = insmntque(vp, mp);
4043 	if (error != 0)
4044 		panic("vfs_allocate_syncvnode: insmntque() failed");
4045 	vp->v_vflag &= ~VV_FORCEINSMQ;
4046 	VOP_UNLOCK(vp, 0);
4047 	/*
4048 	 * Place the vnode onto the syncer worklist. We attempt to
4049 	 * scatter them about on the list so that they will go off
4050 	 * at evenly distributed times even if all the filesystems
4051 	 * are mounted at once.
4052 	 */
4053 	next += incr;
4054 	if (next == 0 || next > syncer_maxdelay) {
4055 		start /= 2;
4056 		incr /= 2;
4057 		if (start == 0) {
4058 			start = syncer_maxdelay / 2;
4059 			incr = syncer_maxdelay;
4060 		}
4061 		next = start;
4062 	}
4063 	bo = &vp->v_bufobj;
4064 	BO_LOCK(bo);
4065 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4066 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4067 	mtx_lock(&sync_mtx);
4068 	sync_vnode_count++;
4069 	if (mp->mnt_syncer == NULL) {
4070 		mp->mnt_syncer = vp;
4071 		vp = NULL;
4072 	}
4073 	mtx_unlock(&sync_mtx);
4074 	BO_UNLOCK(bo);
4075 	if (vp != NULL) {
4076 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4077 		vgone(vp);
4078 		vput(vp);
4079 	}
4080 }
4081 
4082 void
4083 vfs_deallocate_syncvnode(struct mount *mp)
4084 {
4085 	struct vnode *vp;
4086 
4087 	mtx_lock(&sync_mtx);
4088 	vp = mp->mnt_syncer;
4089 	if (vp != NULL)
4090 		mp->mnt_syncer = NULL;
4091 	mtx_unlock(&sync_mtx);
4092 	if (vp != NULL)
4093 		vrele(vp);
4094 }
4095 
4096 /*
4097  * Do a lazy sync of the filesystem.
4098  */
4099 static int
4100 sync_fsync(struct vop_fsync_args *ap)
4101 {
4102 	struct vnode *syncvp = ap->a_vp;
4103 	struct mount *mp = syncvp->v_mount;
4104 	int error, save;
4105 	struct bufobj *bo;
4106 
4107 	/*
4108 	 * We only need to do something if this is a lazy evaluation.
4109 	 */
4110 	if (ap->a_waitfor != MNT_LAZY)
4111 		return (0);
4112 
4113 	/*
4114 	 * Move ourselves to the back of the sync list.
4115 	 */
4116 	bo = &syncvp->v_bufobj;
4117 	BO_LOCK(bo);
4118 	vn_syncer_add_to_worklist(bo, syncdelay);
4119 	BO_UNLOCK(bo);
4120 
4121 	/*
4122 	 * Walk the list of vnodes pushing all that are dirty and
4123 	 * not already on the sync list.
4124 	 */
4125 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4126 		return (0);
4127 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4128 		vfs_unbusy(mp);
4129 		return (0);
4130 	}
4131 	save = curthread_pflags_set(TDP_SYNCIO);
4132 	vfs_msync(mp, MNT_NOWAIT);
4133 	error = VFS_SYNC(mp, MNT_LAZY);
4134 	curthread_pflags_restore(save);
4135 	vn_finished_write(mp);
4136 	vfs_unbusy(mp);
4137 	return (error);
4138 }
4139 
4140 /*
4141  * The syncer vnode is no referenced.
4142  */
4143 static int
4144 sync_inactive(struct vop_inactive_args *ap)
4145 {
4146 
4147 	vgone(ap->a_vp);
4148 	return (0);
4149 }
4150 
4151 /*
4152  * The syncer vnode is no longer needed and is being decommissioned.
4153  *
4154  * Modifications to the worklist must be protected by sync_mtx.
4155  */
4156 static int
4157 sync_reclaim(struct vop_reclaim_args *ap)
4158 {
4159 	struct vnode *vp = ap->a_vp;
4160 	struct bufobj *bo;
4161 
4162 	bo = &vp->v_bufobj;
4163 	BO_LOCK(bo);
4164 	mtx_lock(&sync_mtx);
4165 	if (vp->v_mount->mnt_syncer == vp)
4166 		vp->v_mount->mnt_syncer = NULL;
4167 	if (bo->bo_flag & BO_ONWORKLST) {
4168 		LIST_REMOVE(bo, bo_synclist);
4169 		syncer_worklist_len--;
4170 		sync_vnode_count--;
4171 		bo->bo_flag &= ~BO_ONWORKLST;
4172 	}
4173 	mtx_unlock(&sync_mtx);
4174 	BO_UNLOCK(bo);
4175 
4176 	return (0);
4177 }
4178 
4179 /*
4180  * Check if vnode represents a disk device
4181  */
4182 int
4183 vn_isdisk(struct vnode *vp, int *errp)
4184 {
4185 	int error;
4186 
4187 	if (vp->v_type != VCHR) {
4188 		error = ENOTBLK;
4189 		goto out;
4190 	}
4191 	error = 0;
4192 	dev_lock();
4193 	if (vp->v_rdev == NULL)
4194 		error = ENXIO;
4195 	else if (vp->v_rdev->si_devsw == NULL)
4196 		error = ENXIO;
4197 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
4198 		error = ENOTBLK;
4199 	dev_unlock();
4200 out:
4201 	if (errp != NULL)
4202 		*errp = error;
4203 	return (error == 0);
4204 }
4205 
4206 /*
4207  * Common filesystem object access control check routine.  Accepts a
4208  * vnode's type, "mode", uid and gid, requested access mode, credentials,
4209  * and optional call-by-reference privused argument allowing vaccess()
4210  * to indicate to the caller whether privilege was used to satisfy the
4211  * request (obsoleted).  Returns 0 on success, or an errno on failure.
4212  */
4213 int
4214 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
4215     accmode_t accmode, struct ucred *cred, int *privused)
4216 {
4217 	accmode_t dac_granted;
4218 	accmode_t priv_granted;
4219 
4220 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
4221 	    ("invalid bit in accmode"));
4222 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
4223 	    ("VAPPEND without VWRITE"));
4224 
4225 	/*
4226 	 * Look for a normal, non-privileged way to access the file/directory
4227 	 * as requested.  If it exists, go with that.
4228 	 */
4229 
4230 	if (privused != NULL)
4231 		*privused = 0;
4232 
4233 	dac_granted = 0;
4234 
4235 	/* Check the owner. */
4236 	if (cred->cr_uid == file_uid) {
4237 		dac_granted |= VADMIN;
4238 		if (file_mode & S_IXUSR)
4239 			dac_granted |= VEXEC;
4240 		if (file_mode & S_IRUSR)
4241 			dac_granted |= VREAD;
4242 		if (file_mode & S_IWUSR)
4243 			dac_granted |= (VWRITE | VAPPEND);
4244 
4245 		if ((accmode & dac_granted) == accmode)
4246 			return (0);
4247 
4248 		goto privcheck;
4249 	}
4250 
4251 	/* Otherwise, check the groups (first match) */
4252 	if (groupmember(file_gid, cred)) {
4253 		if (file_mode & S_IXGRP)
4254 			dac_granted |= VEXEC;
4255 		if (file_mode & S_IRGRP)
4256 			dac_granted |= VREAD;
4257 		if (file_mode & S_IWGRP)
4258 			dac_granted |= (VWRITE | VAPPEND);
4259 
4260 		if ((accmode & dac_granted) == accmode)
4261 			return (0);
4262 
4263 		goto privcheck;
4264 	}
4265 
4266 	/* Otherwise, check everyone else. */
4267 	if (file_mode & S_IXOTH)
4268 		dac_granted |= VEXEC;
4269 	if (file_mode & S_IROTH)
4270 		dac_granted |= VREAD;
4271 	if (file_mode & S_IWOTH)
4272 		dac_granted |= (VWRITE | VAPPEND);
4273 	if ((accmode & dac_granted) == accmode)
4274 		return (0);
4275 
4276 privcheck:
4277 	/*
4278 	 * Build a privilege mask to determine if the set of privileges
4279 	 * satisfies the requirements when combined with the granted mask
4280 	 * from above.  For each privilege, if the privilege is required,
4281 	 * bitwise or the request type onto the priv_granted mask.
4282 	 */
4283 	priv_granted = 0;
4284 
4285 	if (type == VDIR) {
4286 		/*
4287 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
4288 		 * requests, instead of PRIV_VFS_EXEC.
4289 		 */
4290 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4291 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP, 0))
4292 			priv_granted |= VEXEC;
4293 	} else {
4294 		/*
4295 		 * Ensure that at least one execute bit is on. Otherwise,
4296 		 * a privileged user will always succeed, and we don't want
4297 		 * this to happen unless the file really is executable.
4298 		 */
4299 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4300 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
4301 		    !priv_check_cred(cred, PRIV_VFS_EXEC, 0))
4302 			priv_granted |= VEXEC;
4303 	}
4304 
4305 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
4306 	    !priv_check_cred(cred, PRIV_VFS_READ, 0))
4307 		priv_granted |= VREAD;
4308 
4309 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
4310 	    !priv_check_cred(cred, PRIV_VFS_WRITE, 0))
4311 		priv_granted |= (VWRITE | VAPPEND);
4312 
4313 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
4314 	    !priv_check_cred(cred, PRIV_VFS_ADMIN, 0))
4315 		priv_granted |= VADMIN;
4316 
4317 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
4318 		/* XXX audit: privilege used */
4319 		if (privused != NULL)
4320 			*privused = 1;
4321 		return (0);
4322 	}
4323 
4324 	return ((accmode & VADMIN) ? EPERM : EACCES);
4325 }
4326 
4327 /*
4328  * Credential check based on process requesting service, and per-attribute
4329  * permissions.
4330  */
4331 int
4332 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
4333     struct thread *td, accmode_t accmode)
4334 {
4335 
4336 	/*
4337 	 * Kernel-invoked always succeeds.
4338 	 */
4339 	if (cred == NOCRED)
4340 		return (0);
4341 
4342 	/*
4343 	 * Do not allow privileged processes in jail to directly manipulate
4344 	 * system attributes.
4345 	 */
4346 	switch (attrnamespace) {
4347 	case EXTATTR_NAMESPACE_SYSTEM:
4348 		/* Potentially should be: return (EPERM); */
4349 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0));
4350 	case EXTATTR_NAMESPACE_USER:
4351 		return (VOP_ACCESS(vp, accmode, cred, td));
4352 	default:
4353 		return (EPERM);
4354 	}
4355 }
4356 
4357 #ifdef DEBUG_VFS_LOCKS
4358 /*
4359  * This only exists to suppress warnings from unlocked specfs accesses.  It is
4360  * no longer ok to have an unlocked VFS.
4361  */
4362 #define	IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL ||		\
4363 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
4364 
4365 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
4366 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
4367     "Drop into debugger on lock violation");
4368 
4369 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
4370 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
4371     0, "Check for interlock across VOPs");
4372 
4373 int vfs_badlock_print = 1;	/* Print lock violations. */
4374 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
4375     0, "Print lock violations");
4376 
4377 #ifdef KDB
4378 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
4379 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
4380     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
4381 #endif
4382 
4383 static void
4384 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
4385 {
4386 
4387 #ifdef KDB
4388 	if (vfs_badlock_backtrace)
4389 		kdb_backtrace();
4390 #endif
4391 	if (vfs_badlock_print)
4392 		printf("%s: %p %s\n", str, (void *)vp, msg);
4393 	if (vfs_badlock_ddb)
4394 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4395 }
4396 
4397 void
4398 assert_vi_locked(struct vnode *vp, const char *str)
4399 {
4400 
4401 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
4402 		vfs_badlock("interlock is not locked but should be", str, vp);
4403 }
4404 
4405 void
4406 assert_vi_unlocked(struct vnode *vp, const char *str)
4407 {
4408 
4409 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
4410 		vfs_badlock("interlock is locked but should not be", str, vp);
4411 }
4412 
4413 void
4414 assert_vop_locked(struct vnode *vp, const char *str)
4415 {
4416 	int locked;
4417 
4418 	if (!IGNORE_LOCK(vp)) {
4419 		locked = VOP_ISLOCKED(vp);
4420 		if (locked == 0 || locked == LK_EXCLOTHER)
4421 			vfs_badlock("is not locked but should be", str, vp);
4422 	}
4423 }
4424 
4425 void
4426 assert_vop_unlocked(struct vnode *vp, const char *str)
4427 {
4428 
4429 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
4430 		vfs_badlock("is locked but should not be", str, vp);
4431 }
4432 
4433 void
4434 assert_vop_elocked(struct vnode *vp, const char *str)
4435 {
4436 
4437 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
4438 		vfs_badlock("is not exclusive locked but should be", str, vp);
4439 }
4440 
4441 #if 0
4442 void
4443 assert_vop_elocked_other(struct vnode *vp, const char *str)
4444 {
4445 
4446 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLOTHER)
4447 		vfs_badlock("is not exclusive locked by another thread",
4448 		    str, vp);
4449 }
4450 
4451 void
4452 assert_vop_slocked(struct vnode *vp, const char *str)
4453 {
4454 
4455 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_SHARED)
4456 		vfs_badlock("is not locked shared but should be", str, vp);
4457 }
4458 #endif /* 0 */
4459 #endif /* DEBUG_VFS_LOCKS */
4460 
4461 void
4462 vop_rename_fail(struct vop_rename_args *ap)
4463 {
4464 
4465 	if (ap->a_tvp != NULL)
4466 		vput(ap->a_tvp);
4467 	if (ap->a_tdvp == ap->a_tvp)
4468 		vrele(ap->a_tdvp);
4469 	else
4470 		vput(ap->a_tdvp);
4471 	vrele(ap->a_fdvp);
4472 	vrele(ap->a_fvp);
4473 }
4474 
4475 void
4476 vop_rename_pre(void *ap)
4477 {
4478 	struct vop_rename_args *a = ap;
4479 
4480 #ifdef DEBUG_VFS_LOCKS
4481 	if (a->a_tvp)
4482 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
4483 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
4484 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
4485 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
4486 
4487 	/* Check the source (from). */
4488 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
4489 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
4490 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
4491 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
4492 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
4493 
4494 	/* Check the target. */
4495 	if (a->a_tvp)
4496 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
4497 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
4498 #endif
4499 	if (a->a_tdvp != a->a_fdvp)
4500 		vhold(a->a_fdvp);
4501 	if (a->a_tvp != a->a_fvp)
4502 		vhold(a->a_fvp);
4503 	vhold(a->a_tdvp);
4504 	if (a->a_tvp)
4505 		vhold(a->a_tvp);
4506 }
4507 
4508 void
4509 vop_strategy_pre(void *ap)
4510 {
4511 #ifdef DEBUG_VFS_LOCKS
4512 	struct vop_strategy_args *a;
4513 	struct buf *bp;
4514 
4515 	a = ap;
4516 	bp = a->a_bp;
4517 
4518 	/*
4519 	 * Cluster ops lock their component buffers but not the IO container.
4520 	 */
4521 	if ((bp->b_flags & B_CLUSTER) != 0)
4522 		return;
4523 
4524 	if (panicstr == NULL && !BUF_ISLOCKED(bp)) {
4525 		if (vfs_badlock_print)
4526 			printf(
4527 			    "VOP_STRATEGY: bp is not locked but should be\n");
4528 		if (vfs_badlock_ddb)
4529 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4530 	}
4531 #endif
4532 }
4533 
4534 void
4535 vop_lock_pre(void *ap)
4536 {
4537 #ifdef DEBUG_VFS_LOCKS
4538 	struct vop_lock1_args *a = ap;
4539 
4540 	if ((a->a_flags & LK_INTERLOCK) == 0)
4541 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4542 	else
4543 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
4544 #endif
4545 }
4546 
4547 void
4548 vop_lock_post(void *ap, int rc)
4549 {
4550 #ifdef DEBUG_VFS_LOCKS
4551 	struct vop_lock1_args *a = ap;
4552 
4553 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4554 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
4555 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
4556 #endif
4557 }
4558 
4559 void
4560 vop_unlock_pre(void *ap)
4561 {
4562 #ifdef DEBUG_VFS_LOCKS
4563 	struct vop_unlock_args *a = ap;
4564 
4565 	if (a->a_flags & LK_INTERLOCK)
4566 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
4567 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
4568 #endif
4569 }
4570 
4571 void
4572 vop_unlock_post(void *ap, int rc)
4573 {
4574 #ifdef DEBUG_VFS_LOCKS
4575 	struct vop_unlock_args *a = ap;
4576 
4577 	if (a->a_flags & LK_INTERLOCK)
4578 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
4579 #endif
4580 }
4581 
4582 void
4583 vop_create_post(void *ap, int rc)
4584 {
4585 	struct vop_create_args *a = ap;
4586 
4587 	if (!rc)
4588 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4589 }
4590 
4591 void
4592 vop_deleteextattr_post(void *ap, int rc)
4593 {
4594 	struct vop_deleteextattr_args *a = ap;
4595 
4596 	if (!rc)
4597 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4598 }
4599 
4600 void
4601 vop_link_post(void *ap, int rc)
4602 {
4603 	struct vop_link_args *a = ap;
4604 
4605 	if (!rc) {
4606 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
4607 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
4608 	}
4609 }
4610 
4611 void
4612 vop_mkdir_post(void *ap, int rc)
4613 {
4614 	struct vop_mkdir_args *a = ap;
4615 
4616 	if (!rc)
4617 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4618 }
4619 
4620 void
4621 vop_mknod_post(void *ap, int rc)
4622 {
4623 	struct vop_mknod_args *a = ap;
4624 
4625 	if (!rc)
4626 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4627 }
4628 
4629 void
4630 vop_reclaim_post(void *ap, int rc)
4631 {
4632 	struct vop_reclaim_args *a = ap;
4633 
4634 	if (!rc)
4635 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
4636 }
4637 
4638 void
4639 vop_remove_post(void *ap, int rc)
4640 {
4641 	struct vop_remove_args *a = ap;
4642 
4643 	if (!rc) {
4644 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4645 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4646 	}
4647 }
4648 
4649 void
4650 vop_rename_post(void *ap, int rc)
4651 {
4652 	struct vop_rename_args *a = ap;
4653 	long hint;
4654 
4655 	if (!rc) {
4656 		hint = NOTE_WRITE;
4657 		if (a->a_fdvp == a->a_tdvp) {
4658 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
4659 				hint |= NOTE_LINK;
4660 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
4661 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
4662 		} else {
4663 			hint |= NOTE_EXTEND;
4664 			if (a->a_fvp->v_type == VDIR)
4665 				hint |= NOTE_LINK;
4666 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
4667 
4668 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
4669 			    a->a_tvp->v_type == VDIR)
4670 				hint &= ~NOTE_LINK;
4671 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
4672 		}
4673 
4674 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
4675 		if (a->a_tvp)
4676 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
4677 	}
4678 	if (a->a_tdvp != a->a_fdvp)
4679 		vdrop(a->a_fdvp);
4680 	if (a->a_tvp != a->a_fvp)
4681 		vdrop(a->a_fvp);
4682 	vdrop(a->a_tdvp);
4683 	if (a->a_tvp)
4684 		vdrop(a->a_tvp);
4685 }
4686 
4687 void
4688 vop_rmdir_post(void *ap, int rc)
4689 {
4690 	struct vop_rmdir_args *a = ap;
4691 
4692 	if (!rc) {
4693 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4694 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4695 	}
4696 }
4697 
4698 void
4699 vop_setattr_post(void *ap, int rc)
4700 {
4701 	struct vop_setattr_args *a = ap;
4702 
4703 	if (!rc)
4704 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4705 }
4706 
4707 void
4708 vop_setextattr_post(void *ap, int rc)
4709 {
4710 	struct vop_setextattr_args *a = ap;
4711 
4712 	if (!rc)
4713 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4714 }
4715 
4716 void
4717 vop_symlink_post(void *ap, int rc)
4718 {
4719 	struct vop_symlink_args *a = ap;
4720 
4721 	if (!rc)
4722 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4723 }
4724 
4725 void
4726 vop_open_post(void *ap, int rc)
4727 {
4728 	struct vop_open_args *a = ap;
4729 
4730 	if (!rc)
4731 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
4732 }
4733 
4734 void
4735 vop_close_post(void *ap, int rc)
4736 {
4737 	struct vop_close_args *a = ap;
4738 
4739 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
4740 	    (a->a_vp->v_iflag & VI_DOOMED) == 0)) {
4741 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
4742 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
4743 	}
4744 }
4745 
4746 void
4747 vop_read_post(void *ap, int rc)
4748 {
4749 	struct vop_read_args *a = ap;
4750 
4751 	if (!rc)
4752 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
4753 }
4754 
4755 void
4756 vop_readdir_post(void *ap, int rc)
4757 {
4758 	struct vop_readdir_args *a = ap;
4759 
4760 	if (!rc)
4761 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
4762 }
4763 
4764 static struct knlist fs_knlist;
4765 
4766 static void
4767 vfs_event_init(void *arg)
4768 {
4769 	knlist_init_mtx(&fs_knlist, NULL);
4770 }
4771 /* XXX - correct order? */
4772 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
4773 
4774 void
4775 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
4776 {
4777 
4778 	KNOTE_UNLOCKED(&fs_knlist, event);
4779 }
4780 
4781 static int	filt_fsattach(struct knote *kn);
4782 static void	filt_fsdetach(struct knote *kn);
4783 static int	filt_fsevent(struct knote *kn, long hint);
4784 
4785 struct filterops fs_filtops = {
4786 	.f_isfd = 0,
4787 	.f_attach = filt_fsattach,
4788 	.f_detach = filt_fsdetach,
4789 	.f_event = filt_fsevent
4790 };
4791 
4792 static int
4793 filt_fsattach(struct knote *kn)
4794 {
4795 
4796 	kn->kn_flags |= EV_CLEAR;
4797 	knlist_add(&fs_knlist, kn, 0);
4798 	return (0);
4799 }
4800 
4801 static void
4802 filt_fsdetach(struct knote *kn)
4803 {
4804 
4805 	knlist_remove(&fs_knlist, kn, 0);
4806 }
4807 
4808 static int
4809 filt_fsevent(struct knote *kn, long hint)
4810 {
4811 
4812 	kn->kn_fflags |= hint;
4813 	return (kn->kn_fflags != 0);
4814 }
4815 
4816 static int
4817 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
4818 {
4819 	struct vfsidctl vc;
4820 	int error;
4821 	struct mount *mp;
4822 
4823 	error = SYSCTL_IN(req, &vc, sizeof(vc));
4824 	if (error)
4825 		return (error);
4826 	if (vc.vc_vers != VFS_CTL_VERS1)
4827 		return (EINVAL);
4828 	mp = vfs_getvfs(&vc.vc_fsid);
4829 	if (mp == NULL)
4830 		return (ENOENT);
4831 	/* ensure that a specific sysctl goes to the right filesystem. */
4832 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
4833 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
4834 		vfs_rel(mp);
4835 		return (EINVAL);
4836 	}
4837 	VCTLTOREQ(&vc, req);
4838 	error = VFS_SYSCTL(mp, vc.vc_op, req);
4839 	vfs_rel(mp);
4840 	return (error);
4841 }
4842 
4843 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR,
4844     NULL, 0, sysctl_vfs_ctl, "",
4845     "Sysctl by fsid");
4846 
4847 /*
4848  * Function to initialize a va_filerev field sensibly.
4849  * XXX: Wouldn't a random number make a lot more sense ??
4850  */
4851 u_quad_t
4852 init_va_filerev(void)
4853 {
4854 	struct bintime bt;
4855 
4856 	getbinuptime(&bt);
4857 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
4858 }
4859 
4860 static int	filt_vfsread(struct knote *kn, long hint);
4861 static int	filt_vfswrite(struct knote *kn, long hint);
4862 static int	filt_vfsvnode(struct knote *kn, long hint);
4863 static void	filt_vfsdetach(struct knote *kn);
4864 static struct filterops vfsread_filtops = {
4865 	.f_isfd = 1,
4866 	.f_detach = filt_vfsdetach,
4867 	.f_event = filt_vfsread
4868 };
4869 static struct filterops vfswrite_filtops = {
4870 	.f_isfd = 1,
4871 	.f_detach = filt_vfsdetach,
4872 	.f_event = filt_vfswrite
4873 };
4874 static struct filterops vfsvnode_filtops = {
4875 	.f_isfd = 1,
4876 	.f_detach = filt_vfsdetach,
4877 	.f_event = filt_vfsvnode
4878 };
4879 
4880 static void
4881 vfs_knllock(void *arg)
4882 {
4883 	struct vnode *vp = arg;
4884 
4885 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4886 }
4887 
4888 static void
4889 vfs_knlunlock(void *arg)
4890 {
4891 	struct vnode *vp = arg;
4892 
4893 	VOP_UNLOCK(vp, 0);
4894 }
4895 
4896 static void
4897 vfs_knl_assert_locked(void *arg)
4898 {
4899 #ifdef DEBUG_VFS_LOCKS
4900 	struct vnode *vp = arg;
4901 
4902 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
4903 #endif
4904 }
4905 
4906 static void
4907 vfs_knl_assert_unlocked(void *arg)
4908 {
4909 #ifdef DEBUG_VFS_LOCKS
4910 	struct vnode *vp = arg;
4911 
4912 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
4913 #endif
4914 }
4915 
4916 int
4917 vfs_kqfilter(struct vop_kqfilter_args *ap)
4918 {
4919 	struct vnode *vp = ap->a_vp;
4920 	struct knote *kn = ap->a_kn;
4921 	struct knlist *knl;
4922 
4923 	switch (kn->kn_filter) {
4924 	case EVFILT_READ:
4925 		kn->kn_fop = &vfsread_filtops;
4926 		break;
4927 	case EVFILT_WRITE:
4928 		kn->kn_fop = &vfswrite_filtops;
4929 		break;
4930 	case EVFILT_VNODE:
4931 		kn->kn_fop = &vfsvnode_filtops;
4932 		break;
4933 	default:
4934 		return (EINVAL);
4935 	}
4936 
4937 	kn->kn_hook = (caddr_t)vp;
4938 
4939 	v_addpollinfo(vp);
4940 	if (vp->v_pollinfo == NULL)
4941 		return (ENOMEM);
4942 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
4943 	vhold(vp);
4944 	knlist_add(knl, kn, 0);
4945 
4946 	return (0);
4947 }
4948 
4949 /*
4950  * Detach knote from vnode
4951  */
4952 static void
4953 filt_vfsdetach(struct knote *kn)
4954 {
4955 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4956 
4957 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
4958 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
4959 	vdrop(vp);
4960 }
4961 
4962 /*ARGSUSED*/
4963 static int
4964 filt_vfsread(struct knote *kn, long hint)
4965 {
4966 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4967 	struct vattr va;
4968 	int res;
4969 
4970 	/*
4971 	 * filesystem is gone, so set the EOF flag and schedule
4972 	 * the knote for deletion.
4973 	 */
4974 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
4975 		VI_LOCK(vp);
4976 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
4977 		VI_UNLOCK(vp);
4978 		return (1);
4979 	}
4980 
4981 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
4982 		return (0);
4983 
4984 	VI_LOCK(vp);
4985 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
4986 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
4987 	VI_UNLOCK(vp);
4988 	return (res);
4989 }
4990 
4991 /*ARGSUSED*/
4992 static int
4993 filt_vfswrite(struct knote *kn, long hint)
4994 {
4995 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4996 
4997 	VI_LOCK(vp);
4998 
4999 	/*
5000 	 * filesystem is gone, so set the EOF flag and schedule
5001 	 * the knote for deletion.
5002 	 */
5003 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5004 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5005 
5006 	kn->kn_data = 0;
5007 	VI_UNLOCK(vp);
5008 	return (1);
5009 }
5010 
5011 static int
5012 filt_vfsvnode(struct knote *kn, long hint)
5013 {
5014 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5015 	int res;
5016 
5017 	VI_LOCK(vp);
5018 	if (kn->kn_sfflags & hint)
5019 		kn->kn_fflags |= hint;
5020 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5021 		kn->kn_flags |= EV_EOF;
5022 		VI_UNLOCK(vp);
5023 		return (1);
5024 	}
5025 	res = (kn->kn_fflags != 0);
5026 	VI_UNLOCK(vp);
5027 	return (res);
5028 }
5029 
5030 int
5031 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5032 {
5033 	int error;
5034 
5035 	if (dp->d_reclen > ap->a_uio->uio_resid)
5036 		return (ENAMETOOLONG);
5037 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
5038 	if (error) {
5039 		if (ap->a_ncookies != NULL) {
5040 			if (ap->a_cookies != NULL)
5041 				free(ap->a_cookies, M_TEMP);
5042 			ap->a_cookies = NULL;
5043 			*ap->a_ncookies = 0;
5044 		}
5045 		return (error);
5046 	}
5047 	if (ap->a_ncookies == NULL)
5048 		return (0);
5049 
5050 	KASSERT(ap->a_cookies,
5051 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
5052 
5053 	*ap->a_cookies = realloc(*ap->a_cookies,
5054 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
5055 	(*ap->a_cookies)[*ap->a_ncookies] = off;
5056 	return (0);
5057 }
5058 
5059 /*
5060  * Mark for update the access time of the file if the filesystem
5061  * supports VOP_MARKATIME.  This functionality is used by execve and
5062  * mmap, so we want to avoid the I/O implied by directly setting
5063  * va_atime for the sake of efficiency.
5064  */
5065 void
5066 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
5067 {
5068 	struct mount *mp;
5069 
5070 	mp = vp->v_mount;
5071 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
5072 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
5073 		(void)VOP_MARKATIME(vp);
5074 }
5075 
5076 /*
5077  * The purpose of this routine is to remove granularity from accmode_t,
5078  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
5079  * VADMIN and VAPPEND.
5080  *
5081  * If it returns 0, the caller is supposed to continue with the usual
5082  * access checks using 'accmode' as modified by this routine.  If it
5083  * returns nonzero value, the caller is supposed to return that value
5084  * as errno.
5085  *
5086  * Note that after this routine runs, accmode may be zero.
5087  */
5088 int
5089 vfs_unixify_accmode(accmode_t *accmode)
5090 {
5091 	/*
5092 	 * There is no way to specify explicit "deny" rule using
5093 	 * file mode or POSIX.1e ACLs.
5094 	 */
5095 	if (*accmode & VEXPLICIT_DENY) {
5096 		*accmode = 0;
5097 		return (0);
5098 	}
5099 
5100 	/*
5101 	 * None of these can be translated into usual access bits.
5102 	 * Also, the common case for NFSv4 ACLs is to not contain
5103 	 * either of these bits. Caller should check for VWRITE
5104 	 * on the containing directory instead.
5105 	 */
5106 	if (*accmode & (VDELETE_CHILD | VDELETE))
5107 		return (EPERM);
5108 
5109 	if (*accmode & VADMIN_PERMS) {
5110 		*accmode &= ~VADMIN_PERMS;
5111 		*accmode |= VADMIN;
5112 	}
5113 
5114 	/*
5115 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5116 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5117 	 */
5118 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5119 
5120 	return (0);
5121 }
5122 
5123 /*
5124  * These are helper functions for filesystems to traverse all
5125  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
5126  *
5127  * This interface replaces MNT_VNODE_FOREACH.
5128  */
5129 
5130 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
5131 
5132 struct vnode *
5133 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
5134 {
5135 	struct vnode *vp;
5136 
5137 	if (should_yield())
5138 		kern_yield(PRI_USER);
5139 	MNT_ILOCK(mp);
5140 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5141 	vp = TAILQ_NEXT(*mvp, v_nmntvnodes);
5142 	while (vp != NULL && (vp->v_type == VMARKER ||
5143 	    (vp->v_iflag & VI_DOOMED) != 0))
5144 		vp = TAILQ_NEXT(vp, v_nmntvnodes);
5145 
5146 	/* Check if we are done */
5147 	if (vp == NULL) {
5148 		__mnt_vnode_markerfree_all(mvp, mp);
5149 		/* MNT_IUNLOCK(mp); -- done in above function */
5150 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
5151 		return (NULL);
5152 	}
5153 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5154 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5155 	VI_LOCK(vp);
5156 	MNT_IUNLOCK(mp);
5157 	return (vp);
5158 }
5159 
5160 struct vnode *
5161 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
5162 {
5163 	struct vnode *vp;
5164 
5165 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5166 	MNT_ILOCK(mp);
5167 	MNT_REF(mp);
5168 	(*mvp)->v_type = VMARKER;
5169 
5170 	vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
5171 	while (vp != NULL && (vp->v_type == VMARKER ||
5172 	    (vp->v_iflag & VI_DOOMED) != 0))
5173 		vp = TAILQ_NEXT(vp, v_nmntvnodes);
5174 
5175 	/* Check if we are done */
5176 	if (vp == NULL) {
5177 		MNT_REL(mp);
5178 		MNT_IUNLOCK(mp);
5179 		free(*mvp, M_VNODE_MARKER);
5180 		*mvp = NULL;
5181 		return (NULL);
5182 	}
5183 	(*mvp)->v_mount = mp;
5184 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5185 	VI_LOCK(vp);
5186 	MNT_IUNLOCK(mp);
5187 	return (vp);
5188 }
5189 
5190 
5191 void
5192 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
5193 {
5194 
5195 	if (*mvp == NULL) {
5196 		MNT_IUNLOCK(mp);
5197 		return;
5198 	}
5199 
5200 	mtx_assert(MNT_MTX(mp), MA_OWNED);
5201 
5202 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5203 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5204 	MNT_REL(mp);
5205 	MNT_IUNLOCK(mp);
5206 	free(*mvp, M_VNODE_MARKER);
5207 	*mvp = NULL;
5208 }
5209 
5210 /*
5211  * These are helper functions for filesystems to traverse their
5212  * active vnodes.  See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h
5213  */
5214 static void
5215 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5216 {
5217 
5218 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5219 
5220 	MNT_ILOCK(mp);
5221 	MNT_REL(mp);
5222 	MNT_IUNLOCK(mp);
5223 	free(*mvp, M_VNODE_MARKER);
5224 	*mvp = NULL;
5225 }
5226 
5227 static struct vnode *
5228 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5229 {
5230 	struct vnode *vp, *nvp;
5231 
5232 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
5233 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5234 restart:
5235 	vp = TAILQ_NEXT(*mvp, v_actfreelist);
5236 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5237 	while (vp != NULL) {
5238 		if (vp->v_type == VMARKER) {
5239 			vp = TAILQ_NEXT(vp, v_actfreelist);
5240 			continue;
5241 		}
5242 		if (!VI_TRYLOCK(vp)) {
5243 			if (mp_ncpus == 1 || should_yield()) {
5244 				TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5245 				mtx_unlock(&vnode_free_list_mtx);
5246 				pause("vnacti", 1);
5247 				mtx_lock(&vnode_free_list_mtx);
5248 				goto restart;
5249 			}
5250 			continue;
5251 		}
5252 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
5253 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
5254 		    ("alien vnode on the active list %p %p", vp, mp));
5255 		if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0)
5256 			break;
5257 		nvp = TAILQ_NEXT(vp, v_actfreelist);
5258 		VI_UNLOCK(vp);
5259 		vp = nvp;
5260 	}
5261 
5262 	/* Check if we are done */
5263 	if (vp == NULL) {
5264 		mtx_unlock(&vnode_free_list_mtx);
5265 		mnt_vnode_markerfree_active(mvp, mp);
5266 		return (NULL);
5267 	}
5268 	TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist);
5269 	mtx_unlock(&vnode_free_list_mtx);
5270 	ASSERT_VI_LOCKED(vp, "active iter");
5271 	KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp));
5272 	return (vp);
5273 }
5274 
5275 struct vnode *
5276 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5277 {
5278 
5279 	if (should_yield())
5280 		kern_yield(PRI_USER);
5281 	mtx_lock(&vnode_free_list_mtx);
5282 	return (mnt_vnode_next_active(mvp, mp));
5283 }
5284 
5285 struct vnode *
5286 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp)
5287 {
5288 	struct vnode *vp;
5289 
5290 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5291 	MNT_ILOCK(mp);
5292 	MNT_REF(mp);
5293 	MNT_IUNLOCK(mp);
5294 	(*mvp)->v_type = VMARKER;
5295 	(*mvp)->v_mount = mp;
5296 
5297 	mtx_lock(&vnode_free_list_mtx);
5298 	vp = TAILQ_FIRST(&mp->mnt_activevnodelist);
5299 	if (vp == NULL) {
5300 		mtx_unlock(&vnode_free_list_mtx);
5301 		mnt_vnode_markerfree_active(mvp, mp);
5302 		return (NULL);
5303 	}
5304 	TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5305 	return (mnt_vnode_next_active(mvp, mp));
5306 }
5307 
5308 void
5309 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5310 {
5311 
5312 	if (*mvp == NULL)
5313 		return;
5314 
5315 	mtx_lock(&vnode_free_list_mtx);
5316 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5317 	mtx_unlock(&vnode_free_list_mtx);
5318 	mnt_vnode_markerfree_active(mvp, mp);
5319 }
5320