xref: /freebsd/sys/kern/vfs_subr.c (revision eaa797943eeac5614edfdc8f6309f332343c3dd2)
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  * 3. 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	vgonel(struct vnode *);
111 static void	vfs_knllock(void *arg);
112 static void	vfs_knlunlock(void *arg);
113 static void	vfs_knl_assert_locked(void *arg);
114 static void	vfs_knl_assert_unlocked(void *arg);
115 static void	destroy_vpollinfo(struct vpollinfo *vi);
116 
117 /*
118  * Number of vnodes in existence.  Increased whenever getnewvnode()
119  * allocates a new vnode, decreased in vdropl() for VI_DOOMED vnode.
120  */
121 static unsigned long	numvnodes;
122 
123 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
124     "Number of vnodes in existence");
125 
126 static u_long vnodes_created;
127 SYSCTL_ULONG(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
128     0, "Number of vnodes created by getnewvnode");
129 
130 /*
131  * Conversion tables for conversion from vnode types to inode formats
132  * and back.
133  */
134 enum vtype iftovt_tab[16] = {
135 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
136 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
137 };
138 int vttoif_tab[10] = {
139 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
140 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
141 };
142 
143 /*
144  * List of vnodes that are ready for recycling.
145  */
146 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
147 
148 /*
149  * "Free" vnode target.  Free vnodes are rarely completely free, but are
150  * just ones that are cheap to recycle.  Usually they are for files which
151  * have been stat'd but not read; these usually have inode and namecache
152  * data attached to them.  This target is the preferred minimum size of a
153  * sub-cache consisting mostly of such files. The system balances the size
154  * of this sub-cache with its complement to try to prevent either from
155  * thrashing while the other is relatively inactive.  The targets express
156  * a preference for the best balance.
157  *
158  * "Above" this target there are 2 further targets (watermarks) related
159  * to recyling of free vnodes.  In the best-operating case, the cache is
160  * exactly full, the free list has size between vlowat and vhiwat above the
161  * free target, and recycling from it and normal use maintains this state.
162  * Sometimes the free list is below vlowat or even empty, but this state
163  * is even better for immediate use provided the cache is not full.
164  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
165  * ones) to reach one of these states.  The watermarks are currently hard-
166  * coded as 4% and 9% of the available space higher.  These and the default
167  * of 25% for wantfreevnodes are too large if the memory size is large.
168  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
169  * whenever vnlru_proc() becomes active.
170  */
171 static u_long wantfreevnodes;
172 SYSCTL_ULONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW,
173     &wantfreevnodes, 0, "Target for minimum number of \"free\" vnodes");
174 static u_long freevnodes;
175 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
176     &freevnodes, 0, "Number of \"free\" vnodes");
177 
178 static u_long recycles_count;
179 SYSCTL_ULONG(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count, 0,
180     "Number of vnodes recycled to meet vnode cache targets");
181 
182 /*
183  * Various variables used for debugging the new implementation of
184  * reassignbuf().
185  * XXX these are probably of (very) limited utility now.
186  */
187 static int reassignbufcalls;
188 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0,
189     "Number of calls to reassignbuf");
190 
191 static u_long free_owe_inact;
192 SYSCTL_ULONG(_vfs, OID_AUTO, free_owe_inact, CTLFLAG_RD, &free_owe_inact, 0,
193     "Number of times free vnodes kept on active list due to VFS "
194     "owing inactivation");
195 
196 /* To keep more than one thread at a time from running vfs_getnewfsid */
197 static struct mtx mntid_mtx;
198 
199 /*
200  * Lock for any access to the following:
201  *	vnode_free_list
202  *	numvnodes
203  *	freevnodes
204  */
205 static struct mtx vnode_free_list_mtx;
206 
207 /* Publicly exported FS */
208 struct nfs_public nfs_pub;
209 
210 static uma_zone_t buf_trie_zone;
211 
212 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
213 static uma_zone_t vnode_zone;
214 static uma_zone_t vnodepoll_zone;
215 
216 /*
217  * The workitem queue.
218  *
219  * It is useful to delay writes of file data and filesystem metadata
220  * for tens of seconds so that quickly created and deleted files need
221  * not waste disk bandwidth being created and removed. To realize this,
222  * we append vnodes to a "workitem" queue. When running with a soft
223  * updates implementation, most pending metadata dependencies should
224  * not wait for more than a few seconds. Thus, mounted on block devices
225  * are delayed only about a half the time that file data is delayed.
226  * Similarly, directory updates are more critical, so are only delayed
227  * about a third the time that file data is delayed. Thus, there are
228  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
229  * one each second (driven off the filesystem syncer process). The
230  * syncer_delayno variable indicates the next queue that is to be processed.
231  * Items that need to be processed soon are placed in this queue:
232  *
233  *	syncer_workitem_pending[syncer_delayno]
234  *
235  * A delay of fifteen seconds is done by placing the request fifteen
236  * entries later in the queue:
237  *
238  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
239  *
240  */
241 static int syncer_delayno;
242 static long syncer_mask;
243 LIST_HEAD(synclist, bufobj);
244 static struct synclist *syncer_workitem_pending;
245 /*
246  * The sync_mtx protects:
247  *	bo->bo_synclist
248  *	sync_vnode_count
249  *	syncer_delayno
250  *	syncer_state
251  *	syncer_workitem_pending
252  *	syncer_worklist_len
253  *	rushjob
254  */
255 static struct mtx sync_mtx;
256 static struct cv sync_wakeup;
257 
258 #define SYNCER_MAXDELAY		32
259 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
260 static int syncdelay = 30;		/* max time to delay syncing data */
261 static int filedelay = 30;		/* time to delay syncing files */
262 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
263     "Time to delay syncing files (in seconds)");
264 static int dirdelay = 29;		/* time to delay syncing directories */
265 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
266     "Time to delay syncing directories (in seconds)");
267 static int metadelay = 28;		/* time to delay syncing metadata */
268 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
269     "Time to delay syncing metadata (in seconds)");
270 static int rushjob;		/* number of slots to run ASAP */
271 static int stat_rush_requests;	/* number of times I/O speeded up */
272 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
273     "Number of times I/O speeded up (rush requests)");
274 
275 /*
276  * When shutting down the syncer, run it at four times normal speed.
277  */
278 #define SYNCER_SHUTDOWN_SPEEDUP		4
279 static int sync_vnode_count;
280 static int syncer_worklist_len;
281 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
282     syncer_state;
283 
284 /* Target for maximum number of vnodes. */
285 int desiredvnodes;
286 static int gapvnodes;		/* gap between wanted and desired */
287 static int vhiwat;		/* enough extras after expansion */
288 static int vlowat;		/* minimal extras before expansion */
289 static int vstir;		/* nonzero to stir non-free vnodes */
290 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
291 
292 static int
293 sysctl_update_desiredvnodes(SYSCTL_HANDLER_ARGS)
294 {
295 	int error, old_desiredvnodes;
296 
297 	old_desiredvnodes = desiredvnodes;
298 	if ((error = sysctl_handle_int(oidp, arg1, arg2, req)) != 0)
299 		return (error);
300 	if (old_desiredvnodes != desiredvnodes) {
301 		wantfreevnodes = desiredvnodes / 4;
302 		/* XXX locking seems to be incomplete. */
303 		vfs_hash_changesize(desiredvnodes);
304 		cache_changesize(desiredvnodes);
305 	}
306 	return (0);
307 }
308 
309 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
310     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, &desiredvnodes, 0,
311     sysctl_update_desiredvnodes, "I", "Target for maximum number of vnodes");
312 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
313     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
314 static int vnlru_nowhere;
315 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
316     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
317 
318 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
319 static int vnsz2log;
320 
321 /*
322  * Support for the bufobj clean & dirty pctrie.
323  */
324 static void *
325 buf_trie_alloc(struct pctrie *ptree)
326 {
327 
328 	return uma_zalloc(buf_trie_zone, M_NOWAIT);
329 }
330 
331 static void
332 buf_trie_free(struct pctrie *ptree, void *node)
333 {
334 
335 	uma_zfree(buf_trie_zone, node);
336 }
337 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
338 
339 /*
340  * Initialize the vnode management data structures.
341  *
342  * Reevaluate the following cap on the number of vnodes after the physical
343  * memory size exceeds 512GB.  In the limit, as the physical memory size
344  * grows, the ratio of the memory size in KB to to vnodes approaches 64:1.
345  */
346 #ifndef	MAXVNODES_MAX
347 #define	MAXVNODES_MAX	(512 * 1024 * 1024 / 64)	/* 8M */
348 #endif
349 
350 /*
351  * Initialize a vnode as it first enters the zone.
352  */
353 static int
354 vnode_init(void *mem, int size, int flags)
355 {
356 	struct vnode *vp;
357 	struct bufobj *bo;
358 
359 	vp = mem;
360 	bzero(vp, size);
361 	/*
362 	 * Setup locks.
363 	 */
364 	vp->v_vnlock = &vp->v_lock;
365 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
366 	/*
367 	 * By default, don't allow shared locks unless filesystems opt-in.
368 	 */
369 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
370 	    LK_NOSHARE | LK_IS_VNODE);
371 	/*
372 	 * Initialize bufobj.
373 	 */
374 	bo = &vp->v_bufobj;
375 	bo->__bo_vnode = vp;
376 	rw_init(BO_LOCKPTR(bo), "bufobj interlock");
377 	bo->bo_private = vp;
378 	TAILQ_INIT(&bo->bo_clean.bv_hd);
379 	TAILQ_INIT(&bo->bo_dirty.bv_hd);
380 	/*
381 	 * Initialize namecache.
382 	 */
383 	LIST_INIT(&vp->v_cache_src);
384 	TAILQ_INIT(&vp->v_cache_dst);
385 	/*
386 	 * Initialize rangelocks.
387 	 */
388 	rangelock_init(&vp->v_rl);
389 	return (0);
390 }
391 
392 /*
393  * Free a vnode when it is cleared from the zone.
394  */
395 static void
396 vnode_fini(void *mem, int size)
397 {
398 	struct vnode *vp;
399 	struct bufobj *bo;
400 
401 	vp = mem;
402 	rangelock_destroy(&vp->v_rl);
403 	lockdestroy(vp->v_vnlock);
404 	mtx_destroy(&vp->v_interlock);
405 	bo = &vp->v_bufobj;
406 	rw_destroy(BO_LOCKPTR(bo));
407 }
408 
409 /*
410  * Provide the size of NFS nclnode and NFS fh for calculation of the
411  * vnode memory consumption.  The size is specified directly to
412  * eliminate dependency on NFS-private header.
413  *
414  * Other filesystems may use bigger or smaller (like UFS and ZFS)
415  * private inode data, but the NFS-based estimation is ample enough.
416  * Still, we care about differences in the size between 64- and 32-bit
417  * platforms.
418  *
419  * Namecache structure size is heuristically
420  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
421  */
422 #ifdef _LP64
423 #define	NFS_NCLNODE_SZ	(528 + 64)
424 #define	NC_SZ		148
425 #else
426 #define	NFS_NCLNODE_SZ	(360 + 32)
427 #define	NC_SZ		92
428 #endif
429 
430 static void
431 vntblinit(void *dummy __unused)
432 {
433 	u_int i;
434 	int physvnodes, virtvnodes;
435 
436 	/*
437 	 * Desiredvnodes is a function of the physical memory size and the
438 	 * kernel's heap size.  Generally speaking, it scales with the
439 	 * physical memory size.  The ratio of desiredvnodes to the physical
440 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
441 	 * Thereafter, the
442 	 * marginal ratio of desiredvnodes to the physical memory size is
443 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
444 	 * size.  The memory required by desiredvnodes vnodes and vm objects
445 	 * must not exceed 1/10th of the kernel's heap size.
446 	 */
447 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
448 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
449 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
450 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
451 	desiredvnodes = min(physvnodes, virtvnodes);
452 	if (desiredvnodes > MAXVNODES_MAX) {
453 		if (bootverbose)
454 			printf("Reducing kern.maxvnodes %d -> %d\n",
455 			    desiredvnodes, MAXVNODES_MAX);
456 		desiredvnodes = MAXVNODES_MAX;
457 	}
458 	wantfreevnodes = desiredvnodes / 4;
459 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
460 	TAILQ_INIT(&vnode_free_list);
461 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
462 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
463 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
464 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
465 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
466 	/*
467 	 * Preallocate enough nodes to support one-per buf so that
468 	 * we can not fail an insert.  reassignbuf() callers can not
469 	 * tolerate the insertion failure.
470 	 */
471 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
472 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
473 	    UMA_ZONE_NOFREE | UMA_ZONE_VM);
474 	uma_prealloc(buf_trie_zone, nbuf);
475 	/*
476 	 * Initialize the filesystem syncer.
477 	 */
478 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
479 	    &syncer_mask);
480 	syncer_maxdelay = syncer_mask + 1;
481 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
482 	cv_init(&sync_wakeup, "syncer");
483 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
484 		vnsz2log++;
485 	vnsz2log--;
486 }
487 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
488 
489 
490 /*
491  * Mark a mount point as busy. Used to synchronize access and to delay
492  * unmounting. Eventually, mountlist_mtx is not released on failure.
493  *
494  * vfs_busy() is a custom lock, it can block the caller.
495  * vfs_busy() only sleeps if the unmount is active on the mount point.
496  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
497  * vnode belonging to mp.
498  *
499  * Lookup uses vfs_busy() to traverse mount points.
500  * root fs			var fs
501  * / vnode lock		A	/ vnode lock (/var)		D
502  * /var vnode lock	B	/log vnode lock(/var/log)	E
503  * vfs_busy lock	C	vfs_busy lock			F
504  *
505  * Within each file system, the lock order is C->A->B and F->D->E.
506  *
507  * When traversing across mounts, the system follows that lock order:
508  *
509  *        C->A->B
510  *              |
511  *              +->F->D->E
512  *
513  * The lookup() process for namei("/var") illustrates the process:
514  *  VOP_LOOKUP() obtains B while A is held
515  *  vfs_busy() obtains a shared lock on F while A and B are held
516  *  vput() releases lock on B
517  *  vput() releases lock on A
518  *  VFS_ROOT() obtains lock on D while shared lock on F is held
519  *  vfs_unbusy() releases shared lock on F
520  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
521  *    Attempt to lock A (instead of vp_crossmp) while D is held would
522  *    violate the global order, causing deadlocks.
523  *
524  * dounmount() locks B while F is drained.
525  */
526 int
527 vfs_busy(struct mount *mp, int flags)
528 {
529 
530 	MPASS((flags & ~MBF_MASK) == 0);
531 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
532 
533 	MNT_ILOCK(mp);
534 	MNT_REF(mp);
535 	/*
536 	 * If mount point is currently being unmounted, sleep until the
537 	 * mount point fate is decided.  If thread doing the unmounting fails,
538 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
539 	 * that this mount point has survived the unmount attempt and vfs_busy
540 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
541 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
542 	 * about to be really destroyed.  vfs_busy needs to release its
543 	 * reference on the mount point in this case and return with ENOENT,
544 	 * telling the caller that mount mount it tried to busy is no longer
545 	 * valid.
546 	 */
547 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
548 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
549 			MNT_REL(mp);
550 			MNT_IUNLOCK(mp);
551 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
552 			    __func__);
553 			return (ENOENT);
554 		}
555 		if (flags & MBF_MNTLSTLOCK)
556 			mtx_unlock(&mountlist_mtx);
557 		mp->mnt_kern_flag |= MNTK_MWAIT;
558 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
559 		if (flags & MBF_MNTLSTLOCK)
560 			mtx_lock(&mountlist_mtx);
561 		MNT_ILOCK(mp);
562 	}
563 	if (flags & MBF_MNTLSTLOCK)
564 		mtx_unlock(&mountlist_mtx);
565 	mp->mnt_lockref++;
566 	MNT_IUNLOCK(mp);
567 	return (0);
568 }
569 
570 /*
571  * Free a busy filesystem.
572  */
573 void
574 vfs_unbusy(struct mount *mp)
575 {
576 
577 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
578 	MNT_ILOCK(mp);
579 	MNT_REL(mp);
580 	KASSERT(mp->mnt_lockref > 0, ("negative mnt_lockref"));
581 	mp->mnt_lockref--;
582 	if (mp->mnt_lockref == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
583 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
584 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
585 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
586 		wakeup(&mp->mnt_lockref);
587 	}
588 	MNT_IUNLOCK(mp);
589 }
590 
591 /*
592  * Lookup a mount point by filesystem identifier.
593  */
594 struct mount *
595 vfs_getvfs(fsid_t *fsid)
596 {
597 	struct mount *mp;
598 
599 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
600 	mtx_lock(&mountlist_mtx);
601 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
602 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
603 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
604 			vfs_ref(mp);
605 			mtx_unlock(&mountlist_mtx);
606 			return (mp);
607 		}
608 	}
609 	mtx_unlock(&mountlist_mtx);
610 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
611 	return ((struct mount *) 0);
612 }
613 
614 /*
615  * Lookup a mount point by filesystem identifier, busying it before
616  * returning.
617  *
618  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
619  * cache for popular filesystem identifiers.  The cache is lockess, using
620  * the fact that struct mount's are never freed.  In worst case we may
621  * get pointer to unmounted or even different filesystem, so we have to
622  * check what we got, and go slow way if so.
623  */
624 struct mount *
625 vfs_busyfs(fsid_t *fsid)
626 {
627 #define	FSID_CACHE_SIZE	256
628 	typedef struct mount * volatile vmp_t;
629 	static vmp_t cache[FSID_CACHE_SIZE];
630 	struct mount *mp;
631 	int error;
632 	uint32_t hash;
633 
634 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
635 	hash = fsid->val[0] ^ fsid->val[1];
636 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
637 	mp = cache[hash];
638 	if (mp == NULL ||
639 	    mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
640 	    mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
641 		goto slow;
642 	if (vfs_busy(mp, 0) != 0) {
643 		cache[hash] = NULL;
644 		goto slow;
645 	}
646 	if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
647 	    mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
648 		return (mp);
649 	else
650 	    vfs_unbusy(mp);
651 
652 slow:
653 	mtx_lock(&mountlist_mtx);
654 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
655 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
656 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
657 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
658 			if (error) {
659 				cache[hash] = NULL;
660 				mtx_unlock(&mountlist_mtx);
661 				return (NULL);
662 			}
663 			cache[hash] = mp;
664 			return (mp);
665 		}
666 	}
667 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
668 	mtx_unlock(&mountlist_mtx);
669 	return ((struct mount *) 0);
670 }
671 
672 /*
673  * Check if a user can access privileged mount options.
674  */
675 int
676 vfs_suser(struct mount *mp, struct thread *td)
677 {
678 	int error;
679 
680 	/*
681 	 * If the thread is jailed, but this is not a jail-friendly file
682 	 * system, deny immediately.
683 	 */
684 	if (!(mp->mnt_vfc->vfc_flags & VFCF_JAIL) && jailed(td->td_ucred))
685 		return (EPERM);
686 
687 	/*
688 	 * If the file system was mounted outside the jail of the calling
689 	 * thread, deny immediately.
690 	 */
691 	if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
692 		return (EPERM);
693 
694 	/*
695 	 * If file system supports delegated administration, we don't check
696 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
697 	 * by the file system itself.
698 	 * If this is not the user that did original mount, we check for
699 	 * the PRIV_VFS_MOUNT_OWNER privilege.
700 	 */
701 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
702 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
703 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
704 			return (error);
705 	}
706 	return (0);
707 }
708 
709 /*
710  * Get a new unique fsid.  Try to make its val[0] unique, since this value
711  * will be used to create fake device numbers for stat().  Also try (but
712  * not so hard) make its val[0] unique mod 2^16, since some emulators only
713  * support 16-bit device numbers.  We end up with unique val[0]'s for the
714  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
715  *
716  * Keep in mind that several mounts may be running in parallel.  Starting
717  * the search one past where the previous search terminated is both a
718  * micro-optimization and a defense against returning the same fsid to
719  * different mounts.
720  */
721 void
722 vfs_getnewfsid(struct mount *mp)
723 {
724 	static uint16_t mntid_base;
725 	struct mount *nmp;
726 	fsid_t tfsid;
727 	int mtype;
728 
729 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
730 	mtx_lock(&mntid_mtx);
731 	mtype = mp->mnt_vfc->vfc_typenum;
732 	tfsid.val[1] = mtype;
733 	mtype = (mtype & 0xFF) << 24;
734 	for (;;) {
735 		tfsid.val[0] = makedev(255,
736 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
737 		mntid_base++;
738 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
739 			break;
740 		vfs_rel(nmp);
741 	}
742 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
743 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
744 	mtx_unlock(&mntid_mtx);
745 }
746 
747 /*
748  * Knob to control the precision of file timestamps:
749  *
750  *   0 = seconds only; nanoseconds zeroed.
751  *   1 = seconds and nanoseconds, accurate within 1/HZ.
752  *   2 = seconds and nanoseconds, truncated to microseconds.
753  * >=3 = seconds and nanoseconds, maximum precision.
754  */
755 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
756 
757 static int timestamp_precision = TSP_USEC;
758 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
759     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
760     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to ms, "
761     "3+: sec + ns (max. precision))");
762 
763 /*
764  * Get a current timestamp.
765  */
766 void
767 vfs_timestamp(struct timespec *tsp)
768 {
769 	struct timeval tv;
770 
771 	switch (timestamp_precision) {
772 	case TSP_SEC:
773 		tsp->tv_sec = time_second;
774 		tsp->tv_nsec = 0;
775 		break;
776 	case TSP_HZ:
777 		getnanotime(tsp);
778 		break;
779 	case TSP_USEC:
780 		microtime(&tv);
781 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
782 		break;
783 	case TSP_NSEC:
784 	default:
785 		nanotime(tsp);
786 		break;
787 	}
788 }
789 
790 /*
791  * Set vnode attributes to VNOVAL
792  */
793 void
794 vattr_null(struct vattr *vap)
795 {
796 
797 	vap->va_type = VNON;
798 	vap->va_size = VNOVAL;
799 	vap->va_bytes = VNOVAL;
800 	vap->va_mode = VNOVAL;
801 	vap->va_nlink = VNOVAL;
802 	vap->va_uid = VNOVAL;
803 	vap->va_gid = VNOVAL;
804 	vap->va_fsid = VNOVAL;
805 	vap->va_fileid = VNOVAL;
806 	vap->va_blocksize = VNOVAL;
807 	vap->va_rdev = VNOVAL;
808 	vap->va_atime.tv_sec = VNOVAL;
809 	vap->va_atime.tv_nsec = VNOVAL;
810 	vap->va_mtime.tv_sec = VNOVAL;
811 	vap->va_mtime.tv_nsec = VNOVAL;
812 	vap->va_ctime.tv_sec = VNOVAL;
813 	vap->va_ctime.tv_nsec = VNOVAL;
814 	vap->va_birthtime.tv_sec = VNOVAL;
815 	vap->va_birthtime.tv_nsec = VNOVAL;
816 	vap->va_flags = VNOVAL;
817 	vap->va_gen = VNOVAL;
818 	vap->va_vaflags = 0;
819 }
820 
821 /*
822  * This routine is called when we have too many vnodes.  It attempts
823  * to free <count> vnodes and will potentially free vnodes that still
824  * have VM backing store (VM backing store is typically the cause
825  * of a vnode blowout so we want to do this).  Therefore, this operation
826  * is not considered cheap.
827  *
828  * A number of conditions may prevent a vnode from being reclaimed.
829  * the buffer cache may have references on the vnode, a directory
830  * vnode may still have references due to the namei cache representing
831  * underlying files, or the vnode may be in active use.   It is not
832  * desirable to reuse such vnodes.  These conditions may cause the
833  * number of vnodes to reach some minimum value regardless of what
834  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
835  */
836 static int
837 vlrureclaim(struct mount *mp, int reclaim_nc_src, int trigger)
838 {
839 	struct vnode *vp;
840 	int count, done, target;
841 
842 	done = 0;
843 	vn_start_write(NULL, &mp, V_WAIT);
844 	MNT_ILOCK(mp);
845 	count = mp->mnt_nvnodelistsize;
846 	target = count * (int64_t)gapvnodes / imax(desiredvnodes, 1);
847 	target = target / 10 + 1;
848 	while (count != 0 && done < target) {
849 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
850 		while (vp != NULL && vp->v_type == VMARKER)
851 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
852 		if (vp == NULL)
853 			break;
854 		/*
855 		 * XXX LRU is completely broken for non-free vnodes.  First
856 		 * by calling here in mountpoint order, then by moving
857 		 * unselected vnodes to the end here, and most grossly by
858 		 * removing the vlruvp() function that was supposed to
859 		 * maintain the order.  (This function was born broken
860 		 * since syncer problems prevented it doing anything.)  The
861 		 * order is closer to LRC (C = Created).
862 		 *
863 		 * LRU reclaiming of vnodes seems to have last worked in
864 		 * FreeBSD-3 where LRU wasn't mentioned under any spelling.
865 		 * Then there was no hold count, and inactive vnodes were
866 		 * simply put on the free list in LRU order.  The separate
867 		 * lists also break LRU.  We prefer to reclaim from the
868 		 * free list for technical reasons.  This tends to thrash
869 		 * the free list to keep very unrecently used held vnodes.
870 		 * The problem is mitigated by keeping the free list large.
871 		 */
872 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
873 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
874 		--count;
875 		if (!VI_TRYLOCK(vp))
876 			goto next_iter;
877 		/*
878 		 * If it's been deconstructed already, it's still
879 		 * referenced, or it exceeds the trigger, skip it.
880 		 * Also skip free vnodes.  We are trying to make space
881 		 * to expand the free list, not reduce it.
882 		 */
883 		if (vp->v_usecount ||
884 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
885 		    ((vp->v_iflag & VI_FREE) != 0) ||
886 		    (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
887 		    vp->v_object->resident_page_count > trigger)) {
888 			VI_UNLOCK(vp);
889 			goto next_iter;
890 		}
891 		MNT_IUNLOCK(mp);
892 		vholdl(vp);
893 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) {
894 			vdrop(vp);
895 			goto next_iter_mntunlocked;
896 		}
897 		VI_LOCK(vp);
898 		/*
899 		 * v_usecount may have been bumped after VOP_LOCK() dropped
900 		 * the vnode interlock and before it was locked again.
901 		 *
902 		 * It is not necessary to recheck VI_DOOMED because it can
903 		 * only be set by another thread that holds both the vnode
904 		 * lock and vnode interlock.  If another thread has the
905 		 * vnode lock before we get to VOP_LOCK() and obtains the
906 		 * vnode interlock after VOP_LOCK() drops the vnode
907 		 * interlock, the other thread will be unable to drop the
908 		 * vnode lock before our VOP_LOCK() call fails.
909 		 */
910 		if (vp->v_usecount ||
911 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
912 		    (vp->v_iflag & VI_FREE) != 0 ||
913 		    (vp->v_object != NULL &&
914 		    vp->v_object->resident_page_count > trigger)) {
915 			VOP_UNLOCK(vp, LK_INTERLOCK);
916 			vdrop(vp);
917 			goto next_iter_mntunlocked;
918 		}
919 		KASSERT((vp->v_iflag & VI_DOOMED) == 0,
920 		    ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
921 		atomic_add_long(&recycles_count, 1);
922 		vgonel(vp);
923 		VOP_UNLOCK(vp, 0);
924 		vdropl(vp);
925 		done++;
926 next_iter_mntunlocked:
927 		if (!should_yield())
928 			goto relock_mnt;
929 		goto yield;
930 next_iter:
931 		if (!should_yield())
932 			continue;
933 		MNT_IUNLOCK(mp);
934 yield:
935 		kern_yield(PRI_USER);
936 relock_mnt:
937 		MNT_ILOCK(mp);
938 	}
939 	MNT_IUNLOCK(mp);
940 	vn_finished_write(mp);
941 	return done;
942 }
943 
944 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
945 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
946     0,
947     "limit on vnode free requests per call to the vnlru_free routine");
948 
949 /*
950  * Attempt to reduce the free list by the requested amount.
951  */
952 static void
953 vnlru_free_locked(int count, struct vfsops *mnt_op)
954 {
955 	struct vnode *vp;
956 	struct mount *mp;
957 
958 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
959 	if (count > max_vnlru_free)
960 		count = max_vnlru_free;
961 	for (; count > 0; count--) {
962 		vp = TAILQ_FIRST(&vnode_free_list);
963 		/*
964 		 * The list can be modified while the free_list_mtx
965 		 * has been dropped and vp could be NULL here.
966 		 */
967 		if (!vp)
968 			break;
969 		VNASSERT(vp->v_op != NULL, vp,
970 		    ("vnlru_free: vnode already reclaimed."));
971 		KASSERT((vp->v_iflag & VI_FREE) != 0,
972 		    ("Removing vnode not on freelist"));
973 		KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
974 		    ("Mangling active vnode"));
975 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
976 
977 		/*
978 		 * Don't recycle if our vnode is from different type
979 		 * of mount point.  Note that mp is type-safe, the
980 		 * check does not reach unmapped address even if
981 		 * vnode is reclaimed.
982 		 * Don't recycle if we can't get the interlock without
983 		 * blocking.
984 		 */
985 		if ((mnt_op != NULL && (mp = vp->v_mount) != NULL &&
986 		    mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) {
987 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist);
988 			continue;
989 		}
990 		VNASSERT((vp->v_iflag & VI_FREE) != 0 && vp->v_holdcnt == 0,
991 		    vp, ("vp inconsistent on freelist"));
992 
993 		/*
994 		 * The clear of VI_FREE prevents activation of the
995 		 * vnode.  There is no sense in putting the vnode on
996 		 * the mount point active list, only to remove it
997 		 * later during recycling.  Inline the relevant part
998 		 * of vholdl(), to avoid triggering assertions or
999 		 * activating.
1000 		 */
1001 		freevnodes--;
1002 		vp->v_iflag &= ~VI_FREE;
1003 		refcount_acquire(&vp->v_holdcnt);
1004 
1005 		mtx_unlock(&vnode_free_list_mtx);
1006 		VI_UNLOCK(vp);
1007 		vtryrecycle(vp);
1008 		/*
1009 		 * If the recycled succeeded this vdrop will actually free
1010 		 * the vnode.  If not it will simply place it back on
1011 		 * the free list.
1012 		 */
1013 		vdrop(vp);
1014 		mtx_lock(&vnode_free_list_mtx);
1015 	}
1016 }
1017 
1018 void
1019 vnlru_free(int count, struct vfsops *mnt_op)
1020 {
1021 
1022 	mtx_lock(&vnode_free_list_mtx);
1023 	vnlru_free_locked(count, mnt_op);
1024 	mtx_unlock(&vnode_free_list_mtx);
1025 }
1026 
1027 
1028 /* XXX some names and initialization are bad for limits and watermarks. */
1029 static int
1030 vspace(void)
1031 {
1032 	int space;
1033 
1034 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1035 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1036 	vlowat = vhiwat / 2;
1037 	if (numvnodes > desiredvnodes)
1038 		return (0);
1039 	space = desiredvnodes - numvnodes;
1040 	if (freevnodes > wantfreevnodes)
1041 		space += freevnodes - wantfreevnodes;
1042 	return (space);
1043 }
1044 
1045 /*
1046  * Attempt to recycle vnodes in a context that is always safe to block.
1047  * Calling vlrurecycle() from the bowels of filesystem code has some
1048  * interesting deadlock problems.
1049  */
1050 static struct proc *vnlruproc;
1051 static int vnlruproc_sig;
1052 
1053 static void
1054 vnlru_proc(void)
1055 {
1056 	struct mount *mp, *nmp;
1057 	unsigned long ofreevnodes, onumvnodes;
1058 	int done, force, reclaim_nc_src, trigger, usevnodes;
1059 
1060 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1061 	    SHUTDOWN_PRI_FIRST);
1062 
1063 	force = 0;
1064 	for (;;) {
1065 		kproc_suspend_check(vnlruproc);
1066 		mtx_lock(&vnode_free_list_mtx);
1067 		/*
1068 		 * If numvnodes is too large (due to desiredvnodes being
1069 		 * adjusted using its sysctl, or emergency growth), first
1070 		 * try to reduce it by discarding from the free list.
1071 		 */
1072 		if (numvnodes > desiredvnodes && freevnodes > 0)
1073 			vnlru_free_locked(ulmin(numvnodes - desiredvnodes,
1074 			    freevnodes), NULL);
1075 		/*
1076 		 * Sleep if the vnode cache is in a good state.  This is
1077 		 * when it is not over-full and has space for about a 4%
1078 		 * or 9% expansion (by growing its size or inexcessively
1079 		 * reducing its free list).  Otherwise, try to reclaim
1080 		 * space for a 10% expansion.
1081 		 */
1082 		if (vstir && force == 0) {
1083 			force = 1;
1084 			vstir = 0;
1085 		}
1086 		if (vspace() >= vlowat && force == 0) {
1087 			vnlruproc_sig = 0;
1088 			wakeup(&vnlruproc_sig);
1089 			msleep(vnlruproc, &vnode_free_list_mtx,
1090 			    PVFS|PDROP, "vlruwt", hz);
1091 			continue;
1092 		}
1093 		mtx_unlock(&vnode_free_list_mtx);
1094 		done = 0;
1095 		ofreevnodes = freevnodes;
1096 		onumvnodes = numvnodes;
1097 		/*
1098 		 * Calculate parameters for recycling.  These are the same
1099 		 * throughout the loop to give some semblance of fairness.
1100 		 * The trigger point is to avoid recycling vnodes with lots
1101 		 * of resident pages.  We aren't trying to free memory; we
1102 		 * are trying to recycle or at least free vnodes.
1103 		 */
1104 		if (numvnodes <= desiredvnodes)
1105 			usevnodes = numvnodes - freevnodes;
1106 		else
1107 			usevnodes = numvnodes;
1108 		if (usevnodes <= 0)
1109 			usevnodes = 1;
1110 		/*
1111 		 * The trigger value is is chosen to give a conservatively
1112 		 * large value to ensure that it alone doesn't prevent
1113 		 * making progress.  The value can easily be so large that
1114 		 * it is effectively infinite in some congested and
1115 		 * misconfigured cases, and this is necessary.  Normally
1116 		 * it is about 8 to 100 (pages), which is quite large.
1117 		 */
1118 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1119 		if (force < 2)
1120 			trigger = vsmalltrigger;
1121 		reclaim_nc_src = force >= 3;
1122 		mtx_lock(&mountlist_mtx);
1123 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1124 			if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) {
1125 				nmp = TAILQ_NEXT(mp, mnt_list);
1126 				continue;
1127 			}
1128 			done += vlrureclaim(mp, reclaim_nc_src, trigger);
1129 			mtx_lock(&mountlist_mtx);
1130 			nmp = TAILQ_NEXT(mp, mnt_list);
1131 			vfs_unbusy(mp);
1132 		}
1133 		mtx_unlock(&mountlist_mtx);
1134 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1135 			uma_reclaim();
1136 		if (done == 0) {
1137 			if (force == 0 || force == 1) {
1138 				force = 2;
1139 				continue;
1140 			}
1141 			if (force == 2) {
1142 				force = 3;
1143 				continue;
1144 			}
1145 			force = 0;
1146 			vnlru_nowhere++;
1147 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1148 		} else
1149 			kern_yield(PRI_USER);
1150 		/*
1151 		 * After becoming active to expand above low water, keep
1152 		 * active until above high water.
1153 		 */
1154 		force = vspace() < vhiwat;
1155 	}
1156 }
1157 
1158 static struct kproc_desc vnlru_kp = {
1159 	"vnlru",
1160 	vnlru_proc,
1161 	&vnlruproc
1162 };
1163 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1164     &vnlru_kp);
1165 
1166 /*
1167  * Routines having to do with the management of the vnode table.
1168  */
1169 
1170 /*
1171  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1172  * before we actually vgone().  This function must be called with the vnode
1173  * held to prevent the vnode from being returned to the free list midway
1174  * through vgone().
1175  */
1176 static int
1177 vtryrecycle(struct vnode *vp)
1178 {
1179 	struct mount *vnmp;
1180 
1181 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1182 	VNASSERT(vp->v_holdcnt, vp,
1183 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1184 	/*
1185 	 * This vnode may found and locked via some other list, if so we
1186 	 * can't recycle it yet.
1187 	 */
1188 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1189 		CTR2(KTR_VFS,
1190 		    "%s: impossible to recycle, vp %p lock is already held",
1191 		    __func__, vp);
1192 		return (EWOULDBLOCK);
1193 	}
1194 	/*
1195 	 * Don't recycle if its filesystem is being suspended.
1196 	 */
1197 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1198 		VOP_UNLOCK(vp, 0);
1199 		CTR2(KTR_VFS,
1200 		    "%s: impossible to recycle, cannot start the write for %p",
1201 		    __func__, vp);
1202 		return (EBUSY);
1203 	}
1204 	/*
1205 	 * If we got this far, we need to acquire the interlock and see if
1206 	 * anyone picked up this vnode from another list.  If not, we will
1207 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1208 	 * will skip over it.
1209 	 */
1210 	VI_LOCK(vp);
1211 	if (vp->v_usecount) {
1212 		VOP_UNLOCK(vp, LK_INTERLOCK);
1213 		vn_finished_write(vnmp);
1214 		CTR2(KTR_VFS,
1215 		    "%s: impossible to recycle, %p is already referenced",
1216 		    __func__, vp);
1217 		return (EBUSY);
1218 	}
1219 	if ((vp->v_iflag & VI_DOOMED) == 0) {
1220 		atomic_add_long(&recycles_count, 1);
1221 		vgonel(vp);
1222 	}
1223 	VOP_UNLOCK(vp, LK_INTERLOCK);
1224 	vn_finished_write(vnmp);
1225 	return (0);
1226 }
1227 
1228 static void
1229 vcheckspace(void)
1230 {
1231 
1232 	if (vspace() < vlowat && vnlruproc_sig == 0) {
1233 		vnlruproc_sig = 1;
1234 		wakeup(vnlruproc);
1235 	}
1236 }
1237 
1238 /*
1239  * Wait if necessary for space for a new vnode.
1240  */
1241 static int
1242 getnewvnode_wait(int suspended)
1243 {
1244 
1245 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1246 	if (numvnodes >= desiredvnodes) {
1247 		if (suspended) {
1248 			/*
1249 			 * The file system is being suspended.  We cannot
1250 			 * risk a deadlock here, so allow allocation of
1251 			 * another vnode even if this would give too many.
1252 			 */
1253 			return (0);
1254 		}
1255 		if (vnlruproc_sig == 0) {
1256 			vnlruproc_sig = 1;	/* avoid unnecessary wakeups */
1257 			wakeup(vnlruproc);
1258 		}
1259 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
1260 		    "vlruwk", hz);
1261 	}
1262 	/* Post-adjust like the pre-adjust in getnewvnode(). */
1263 	if (numvnodes + 1 > desiredvnodes && freevnodes > 1)
1264 		vnlru_free_locked(1, NULL);
1265 	return (numvnodes >= desiredvnodes ? ENFILE : 0);
1266 }
1267 
1268 /*
1269  * This hack is fragile, and probably not needed any more now that the
1270  * watermark handling works.
1271  */
1272 void
1273 getnewvnode_reserve(u_int count)
1274 {
1275 	struct thread *td;
1276 
1277 	/* Pre-adjust like the pre-adjust in getnewvnode(), with any count. */
1278 	/* XXX no longer so quick, but this part is not racy. */
1279 	mtx_lock(&vnode_free_list_mtx);
1280 	if (numvnodes + count > desiredvnodes && freevnodes > wantfreevnodes)
1281 		vnlru_free_locked(ulmin(numvnodes + count - desiredvnodes,
1282 		    freevnodes - wantfreevnodes), NULL);
1283 	mtx_unlock(&vnode_free_list_mtx);
1284 
1285 	td = curthread;
1286 	/* First try to be quick and racy. */
1287 	if (atomic_fetchadd_long(&numvnodes, count) + count <= desiredvnodes) {
1288 		td->td_vp_reserv += count;
1289 		vcheckspace();	/* XXX no longer so quick, but more racy */
1290 		return;
1291 	} else
1292 		atomic_subtract_long(&numvnodes, count);
1293 
1294 	mtx_lock(&vnode_free_list_mtx);
1295 	while (count > 0) {
1296 		if (getnewvnode_wait(0) == 0) {
1297 			count--;
1298 			td->td_vp_reserv++;
1299 			atomic_add_long(&numvnodes, 1);
1300 		}
1301 	}
1302 	vcheckspace();
1303 	mtx_unlock(&vnode_free_list_mtx);
1304 }
1305 
1306 /*
1307  * This hack is fragile, especially if desiredvnodes or wantvnodes are
1308  * misconfgured or changed significantly.  Reducing desiredvnodes below
1309  * the reserved amount should cause bizarre behaviour like reducing it
1310  * below the number of active vnodes -- the system will try to reduce
1311  * numvnodes to match, but should fail, so the subtraction below should
1312  * not overflow.
1313  */
1314 void
1315 getnewvnode_drop_reserve(void)
1316 {
1317 	struct thread *td;
1318 
1319 	td = curthread;
1320 	atomic_subtract_long(&numvnodes, td->td_vp_reserv);
1321 	td->td_vp_reserv = 0;
1322 }
1323 
1324 /*
1325  * Return the next vnode from the free list.
1326  */
1327 int
1328 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1329     struct vnode **vpp)
1330 {
1331 	struct vnode *vp;
1332 	struct thread *td;
1333 	struct lock_object *lo;
1334 	static int cyclecount;
1335 	int error;
1336 
1337 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1338 	vp = NULL;
1339 	td = curthread;
1340 	if (td->td_vp_reserv > 0) {
1341 		td->td_vp_reserv -= 1;
1342 		goto alloc;
1343 	}
1344 	mtx_lock(&vnode_free_list_mtx);
1345 	if (numvnodes < desiredvnodes)
1346 		cyclecount = 0;
1347 	else if (cyclecount++ >= freevnodes) {
1348 		cyclecount = 0;
1349 		vstir = 1;
1350 	}
1351 	/*
1352 	 * Grow the vnode cache if it will not be above its target max
1353 	 * after growing.  Otherwise, if the free list is nonempty, try
1354 	 * to reclaim 1 item from it before growing the cache (possibly
1355 	 * above its target max if the reclamation failed or is delayed).
1356 	 * Otherwise, wait for some space.  In all cases, schedule
1357 	 * vnlru_proc() if we are getting short of space.  The watermarks
1358 	 * should be chosen so that we never wait or even reclaim from
1359 	 * the free list to below its target minimum.
1360 	 */
1361 	if (numvnodes + 1 <= desiredvnodes)
1362 		;
1363 	else if (freevnodes > 0)
1364 		vnlru_free_locked(1, NULL);
1365 	else {
1366 		error = getnewvnode_wait(mp != NULL && (mp->mnt_kern_flag &
1367 		    MNTK_SUSPEND));
1368 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
1369 		if (error != 0) {
1370 			mtx_unlock(&vnode_free_list_mtx);
1371 			return (error);
1372 		}
1373 #endif
1374 	}
1375 	vcheckspace();
1376 	atomic_add_long(&numvnodes, 1);
1377 	mtx_unlock(&vnode_free_list_mtx);
1378 alloc:
1379 	atomic_add_long(&vnodes_created, 1);
1380 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK);
1381 	/*
1382 	 * Locks are given the generic name "vnode" when created.
1383 	 * Follow the historic practice of using the filesystem
1384 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1385 	 *
1386 	 * Locks live in a witness group keyed on their name. Thus,
1387 	 * when a lock is renamed, it must also move from the witness
1388 	 * group of its old name to the witness group of its new name.
1389 	 *
1390 	 * The change only needs to be made when the vnode moves
1391 	 * from one filesystem type to another. We ensure that each
1392 	 * filesystem use a single static name pointer for its tag so
1393 	 * that we can compare pointers rather than doing a strcmp().
1394 	 */
1395 	lo = &vp->v_vnlock->lock_object;
1396 	if (lo->lo_name != tag) {
1397 		lo->lo_name = tag;
1398 		WITNESS_DESTROY(lo);
1399 		WITNESS_INIT(lo, tag);
1400 	}
1401 	/*
1402 	 * By default, don't allow shared locks unless filesystems opt-in.
1403 	 */
1404 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1405 	/*
1406 	 * Finalize various vnode identity bits.
1407 	 */
1408 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1409 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1410 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1411 	vp->v_type = VNON;
1412 	vp->v_tag = tag;
1413 	vp->v_op = vops;
1414 	v_init_counters(vp);
1415 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1416 #ifdef MAC
1417 	mac_vnode_init(vp);
1418 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1419 		mac_vnode_associate_singlelabel(mp, vp);
1420 	else if (mp == NULL && vops != &dead_vnodeops)
1421 		printf("NULL mp in getnewvnode()\n");
1422 #endif
1423 	if (mp != NULL) {
1424 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1425 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1426 			vp->v_vflag |= VV_NOKNOTE;
1427 	}
1428 
1429 	/*
1430 	 * For the filesystems which do not use vfs_hash_insert(),
1431 	 * still initialize v_hash to have vfs_hash_index() useful.
1432 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1433 	 * its own hashing.
1434 	 */
1435 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1436 
1437 	*vpp = vp;
1438 	return (0);
1439 }
1440 
1441 /*
1442  * Delete from old mount point vnode list, if on one.
1443  */
1444 static void
1445 delmntque(struct vnode *vp)
1446 {
1447 	struct mount *mp;
1448 	int active;
1449 
1450 	mp = vp->v_mount;
1451 	if (mp == NULL)
1452 		return;
1453 	MNT_ILOCK(mp);
1454 	VI_LOCK(vp);
1455 	KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize,
1456 	    ("Active vnode list size %d > Vnode list size %d",
1457 	     mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize));
1458 	active = vp->v_iflag & VI_ACTIVE;
1459 	vp->v_iflag &= ~VI_ACTIVE;
1460 	if (active) {
1461 		mtx_lock(&vnode_free_list_mtx);
1462 		TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist);
1463 		mp->mnt_activevnodelistsize--;
1464 		mtx_unlock(&vnode_free_list_mtx);
1465 	}
1466 	vp->v_mount = NULL;
1467 	VI_UNLOCK(vp);
1468 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1469 		("bad mount point vnode list size"));
1470 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1471 	mp->mnt_nvnodelistsize--;
1472 	MNT_REL(mp);
1473 	MNT_IUNLOCK(mp);
1474 }
1475 
1476 static void
1477 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1478 {
1479 
1480 	vp->v_data = NULL;
1481 	vp->v_op = &dead_vnodeops;
1482 	vgone(vp);
1483 	vput(vp);
1484 }
1485 
1486 /*
1487  * Insert into list of vnodes for the new mount point, if available.
1488  */
1489 int
1490 insmntque1(struct vnode *vp, struct mount *mp,
1491 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1492 {
1493 
1494 	KASSERT(vp->v_mount == NULL,
1495 		("insmntque: vnode already on per mount vnode list"));
1496 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1497 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1498 
1499 	/*
1500 	 * We acquire the vnode interlock early to ensure that the
1501 	 * vnode cannot be recycled by another process releasing a
1502 	 * holdcnt on it before we get it on both the vnode list
1503 	 * and the active vnode list. The mount mutex protects only
1504 	 * manipulation of the vnode list and the vnode freelist
1505 	 * mutex protects only manipulation of the active vnode list.
1506 	 * Hence the need to hold the vnode interlock throughout.
1507 	 */
1508 	MNT_ILOCK(mp);
1509 	VI_LOCK(vp);
1510 	if (((mp->mnt_kern_flag & MNTK_NOINSMNTQ) != 0 &&
1511 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1512 	    mp->mnt_nvnodelistsize == 0)) &&
1513 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1514 		VI_UNLOCK(vp);
1515 		MNT_IUNLOCK(mp);
1516 		if (dtr != NULL)
1517 			dtr(vp, dtr_arg);
1518 		return (EBUSY);
1519 	}
1520 	vp->v_mount = mp;
1521 	MNT_REF(mp);
1522 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1523 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1524 		("neg mount point vnode list size"));
1525 	mp->mnt_nvnodelistsize++;
1526 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1527 	    ("Activating already active vnode"));
1528 	vp->v_iflag |= VI_ACTIVE;
1529 	mtx_lock(&vnode_free_list_mtx);
1530 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
1531 	mp->mnt_activevnodelistsize++;
1532 	mtx_unlock(&vnode_free_list_mtx);
1533 	VI_UNLOCK(vp);
1534 	MNT_IUNLOCK(mp);
1535 	return (0);
1536 }
1537 
1538 int
1539 insmntque(struct vnode *vp, struct mount *mp)
1540 {
1541 
1542 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1543 }
1544 
1545 /*
1546  * Flush out and invalidate all buffers associated with a bufobj
1547  * Called with the underlying object locked.
1548  */
1549 int
1550 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1551 {
1552 	int error;
1553 
1554 	BO_LOCK(bo);
1555 	if (flags & V_SAVE) {
1556 		error = bufobj_wwait(bo, slpflag, slptimeo);
1557 		if (error) {
1558 			BO_UNLOCK(bo);
1559 			return (error);
1560 		}
1561 		if (bo->bo_dirty.bv_cnt > 0) {
1562 			BO_UNLOCK(bo);
1563 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1564 				return (error);
1565 			/*
1566 			 * XXX We could save a lock/unlock if this was only
1567 			 * enabled under INVARIANTS
1568 			 */
1569 			BO_LOCK(bo);
1570 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1571 				panic("vinvalbuf: dirty bufs");
1572 		}
1573 	}
1574 	/*
1575 	 * If you alter this loop please notice that interlock is dropped and
1576 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1577 	 * no race conditions occur from this.
1578 	 */
1579 	do {
1580 		error = flushbuflist(&bo->bo_clean,
1581 		    flags, bo, slpflag, slptimeo);
1582 		if (error == 0 && !(flags & V_CLEANONLY))
1583 			error = flushbuflist(&bo->bo_dirty,
1584 			    flags, bo, slpflag, slptimeo);
1585 		if (error != 0 && error != EAGAIN) {
1586 			BO_UNLOCK(bo);
1587 			return (error);
1588 		}
1589 	} while (error != 0);
1590 
1591 	/*
1592 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1593 	 * have write I/O in-progress but if there is a VM object then the
1594 	 * VM object can also have read-I/O in-progress.
1595 	 */
1596 	do {
1597 		bufobj_wwait(bo, 0, 0);
1598 		BO_UNLOCK(bo);
1599 		if (bo->bo_object != NULL) {
1600 			VM_OBJECT_WLOCK(bo->bo_object);
1601 			vm_object_pip_wait(bo->bo_object, "bovlbx");
1602 			VM_OBJECT_WUNLOCK(bo->bo_object);
1603 		}
1604 		BO_LOCK(bo);
1605 	} while (bo->bo_numoutput > 0);
1606 	BO_UNLOCK(bo);
1607 
1608 	/*
1609 	 * Destroy the copy in the VM cache, too.
1610 	 */
1611 	if (bo->bo_object != NULL &&
1612 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY)) == 0) {
1613 		VM_OBJECT_WLOCK(bo->bo_object);
1614 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1615 		    OBJPR_CLEANONLY : 0);
1616 		VM_OBJECT_WUNLOCK(bo->bo_object);
1617 	}
1618 
1619 #ifdef INVARIANTS
1620 	BO_LOCK(bo);
1621 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY)) == 0 &&
1622 	    (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0))
1623 		panic("vinvalbuf: flush failed");
1624 	BO_UNLOCK(bo);
1625 #endif
1626 	return (0);
1627 }
1628 
1629 /*
1630  * Flush out and invalidate all buffers associated with a vnode.
1631  * Called with the underlying object locked.
1632  */
1633 int
1634 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1635 {
1636 
1637 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1638 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1639 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1640 		return (0);
1641 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1642 }
1643 
1644 /*
1645  * Flush out buffers on the specified list.
1646  *
1647  */
1648 static int
1649 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1650     int slptimeo)
1651 {
1652 	struct buf *bp, *nbp;
1653 	int retval, error;
1654 	daddr_t lblkno;
1655 	b_xflags_t xflags;
1656 
1657 	ASSERT_BO_WLOCKED(bo);
1658 
1659 	retval = 0;
1660 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1661 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1662 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1663 			continue;
1664 		}
1665 		lblkno = 0;
1666 		xflags = 0;
1667 		if (nbp != NULL) {
1668 			lblkno = nbp->b_lblkno;
1669 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
1670 		}
1671 		retval = EAGAIN;
1672 		error = BUF_TIMELOCK(bp,
1673 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
1674 		    "flushbuf", slpflag, slptimeo);
1675 		if (error) {
1676 			BO_LOCK(bo);
1677 			return (error != ENOLCK ? error : EAGAIN);
1678 		}
1679 		KASSERT(bp->b_bufobj == bo,
1680 		    ("bp %p wrong b_bufobj %p should be %p",
1681 		    bp, bp->b_bufobj, bo));
1682 		/*
1683 		 * XXX Since there are no node locks for NFS, I
1684 		 * believe there is a slight chance that a delayed
1685 		 * write will occur while sleeping just above, so
1686 		 * check for it.
1687 		 */
1688 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1689 		    (flags & V_SAVE)) {
1690 			bremfree(bp);
1691 			bp->b_flags |= B_ASYNC;
1692 			bwrite(bp);
1693 			BO_LOCK(bo);
1694 			return (EAGAIN);	/* XXX: why not loop ? */
1695 		}
1696 		bremfree(bp);
1697 		bp->b_flags |= (B_INVAL | B_RELBUF);
1698 		bp->b_flags &= ~B_ASYNC;
1699 		brelse(bp);
1700 		BO_LOCK(bo);
1701 		nbp = gbincore(bo, lblkno);
1702 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1703 		    != xflags)
1704 			break;			/* nbp invalid */
1705 	}
1706 	return (retval);
1707 }
1708 
1709 int
1710 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
1711 {
1712 	struct buf *bp;
1713 	int error;
1714 	daddr_t lblkno;
1715 
1716 	ASSERT_BO_LOCKED(bo);
1717 
1718 	for (lblkno = startn;;) {
1719 again:
1720 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
1721 		if (bp == NULL || bp->b_lblkno >= endn ||
1722 		    bp->b_lblkno < startn)
1723 			break;
1724 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
1725 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
1726 		if (error != 0) {
1727 			BO_RLOCK(bo);
1728 			if (error == ENOLCK)
1729 				goto again;
1730 			return (error);
1731 		}
1732 		KASSERT(bp->b_bufobj == bo,
1733 		    ("bp %p wrong b_bufobj %p should be %p",
1734 		    bp, bp->b_bufobj, bo));
1735 		lblkno = bp->b_lblkno + 1;
1736 		if ((bp->b_flags & B_MANAGED) == 0)
1737 			bremfree(bp);
1738 		bp->b_flags |= B_RELBUF;
1739 		/*
1740 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
1741 		 * pages backing each buffer in the range are unlikely to be
1742 		 * reused.  Dirty buffers will have the hint applied once
1743 		 * they've been written.
1744 		 */
1745 		if (bp->b_vp->v_object != NULL)
1746 			bp->b_flags |= B_NOREUSE;
1747 		brelse(bp);
1748 		BO_RLOCK(bo);
1749 	}
1750 	return (0);
1751 }
1752 
1753 /*
1754  * Truncate a file's buffer and pages to a specified length.  This
1755  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1756  * sync activity.
1757  */
1758 int
1759 vtruncbuf(struct vnode *vp, struct ucred *cred, off_t length, int blksize)
1760 {
1761 	struct buf *bp, *nbp;
1762 	int anyfreed;
1763 	int trunclbn;
1764 	struct bufobj *bo;
1765 
1766 	CTR5(KTR_VFS, "%s: vp %p with cred %p and block %d:%ju", __func__,
1767 	    vp, cred, blksize, (uintmax_t)length);
1768 
1769 	/*
1770 	 * Round up to the *next* lbn.
1771 	 */
1772 	trunclbn = howmany(length, blksize);
1773 
1774 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1775 restart:
1776 	bo = &vp->v_bufobj;
1777 	BO_LOCK(bo);
1778 	anyfreed = 1;
1779 	for (;anyfreed;) {
1780 		anyfreed = 0;
1781 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.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 
1789 			bremfree(bp);
1790 			bp->b_flags |= (B_INVAL | B_RELBUF);
1791 			bp->b_flags &= ~B_ASYNC;
1792 			brelse(bp);
1793 			anyfreed = 1;
1794 
1795 			BO_LOCK(bo);
1796 			if (nbp != NULL &&
1797 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1798 			    (nbp->b_vp != vp) ||
1799 			    (nbp->b_flags & B_DELWRI))) {
1800 				BO_UNLOCK(bo);
1801 				goto restart;
1802 			}
1803 		}
1804 
1805 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1806 			if (bp->b_lblkno < trunclbn)
1807 				continue;
1808 			if (BUF_LOCK(bp,
1809 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1810 			    BO_LOCKPTR(bo)) == ENOLCK)
1811 				goto restart;
1812 			bremfree(bp);
1813 			bp->b_flags |= (B_INVAL | B_RELBUF);
1814 			bp->b_flags &= ~B_ASYNC;
1815 			brelse(bp);
1816 			anyfreed = 1;
1817 
1818 			BO_LOCK(bo);
1819 			if (nbp != NULL &&
1820 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1821 			    (nbp->b_vp != vp) ||
1822 			    (nbp->b_flags & B_DELWRI) == 0)) {
1823 				BO_UNLOCK(bo);
1824 				goto restart;
1825 			}
1826 		}
1827 	}
1828 
1829 	if (length > 0) {
1830 restartsync:
1831 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1832 			if (bp->b_lblkno > 0)
1833 				continue;
1834 			/*
1835 			 * Since we hold the vnode lock this should only
1836 			 * fail if we're racing with the buf daemon.
1837 			 */
1838 			if (BUF_LOCK(bp,
1839 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1840 			    BO_LOCKPTR(bo)) == ENOLCK) {
1841 				goto restart;
1842 			}
1843 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1844 			    ("buf(%p) on dirty queue without DELWRI", bp));
1845 
1846 			bremfree(bp);
1847 			bawrite(bp);
1848 			BO_LOCK(bo);
1849 			goto restartsync;
1850 		}
1851 	}
1852 
1853 	bufobj_wwait(bo, 0, 0);
1854 	BO_UNLOCK(bo);
1855 	vnode_pager_setsize(vp, length);
1856 
1857 	return (0);
1858 }
1859 
1860 static void
1861 buf_vlist_remove(struct buf *bp)
1862 {
1863 	struct bufv *bv;
1864 
1865 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1866 	ASSERT_BO_WLOCKED(bp->b_bufobj);
1867 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
1868 	    (BX_VNDIRTY|BX_VNCLEAN),
1869 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
1870 	if (bp->b_xflags & BX_VNDIRTY)
1871 		bv = &bp->b_bufobj->bo_dirty;
1872 	else
1873 		bv = &bp->b_bufobj->bo_clean;
1874 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
1875 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1876 	bv->bv_cnt--;
1877 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1878 }
1879 
1880 /*
1881  * Add the buffer to the sorted clean or dirty block list.
1882  *
1883  * NOTE: xflags is passed as a constant, optimizing this inline function!
1884  */
1885 static void
1886 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1887 {
1888 	struct bufv *bv;
1889 	struct buf *n;
1890 	int error;
1891 
1892 	ASSERT_BO_WLOCKED(bo);
1893 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
1894 	    ("dead bo %p", bo));
1895 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1896 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
1897 	bp->b_xflags |= xflags;
1898 	if (xflags & BX_VNDIRTY)
1899 		bv = &bo->bo_dirty;
1900 	else
1901 		bv = &bo->bo_clean;
1902 
1903 	/*
1904 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
1905 	 * we tend to grow at the tail so lookup_le should usually be cheaper
1906 	 * than _ge.
1907 	 */
1908 	if (bv->bv_cnt == 0 ||
1909 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
1910 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1911 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
1912 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
1913 	else
1914 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
1915 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
1916 	if (error)
1917 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
1918 	bv->bv_cnt++;
1919 }
1920 
1921 /*
1922  * Look up a buffer using the buffer tries.
1923  */
1924 struct buf *
1925 gbincore(struct bufobj *bo, daddr_t lblkno)
1926 {
1927 	struct buf *bp;
1928 
1929 	ASSERT_BO_LOCKED(bo);
1930 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
1931 	if (bp != NULL)
1932 		return (bp);
1933 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
1934 }
1935 
1936 /*
1937  * Associate a buffer with a vnode.
1938  */
1939 void
1940 bgetvp(struct vnode *vp, struct buf *bp)
1941 {
1942 	struct bufobj *bo;
1943 
1944 	bo = &vp->v_bufobj;
1945 	ASSERT_BO_WLOCKED(bo);
1946 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
1947 
1948 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
1949 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
1950 	    ("bgetvp: bp already attached! %p", bp));
1951 
1952 	vhold(vp);
1953 	bp->b_vp = vp;
1954 	bp->b_bufobj = bo;
1955 	/*
1956 	 * Insert onto list for new vnode.
1957 	 */
1958 	buf_vlist_add(bp, bo, BX_VNCLEAN);
1959 }
1960 
1961 /*
1962  * Disassociate a buffer from a vnode.
1963  */
1964 void
1965 brelvp(struct buf *bp)
1966 {
1967 	struct bufobj *bo;
1968 	struct vnode *vp;
1969 
1970 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1971 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1972 
1973 	/*
1974 	 * Delete from old vnode list, if on one.
1975 	 */
1976 	vp = bp->b_vp;		/* XXX */
1977 	bo = bp->b_bufobj;
1978 	BO_LOCK(bo);
1979 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1980 		buf_vlist_remove(bp);
1981 	else
1982 		panic("brelvp: Buffer %p not on queue.", bp);
1983 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1984 		bo->bo_flag &= ~BO_ONWORKLST;
1985 		mtx_lock(&sync_mtx);
1986 		LIST_REMOVE(bo, bo_synclist);
1987 		syncer_worklist_len--;
1988 		mtx_unlock(&sync_mtx);
1989 	}
1990 	bp->b_vp = NULL;
1991 	bp->b_bufobj = NULL;
1992 	BO_UNLOCK(bo);
1993 	vdrop(vp);
1994 }
1995 
1996 /*
1997  * Add an item to the syncer work queue.
1998  */
1999 static void
2000 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2001 {
2002 	int slot;
2003 
2004 	ASSERT_BO_WLOCKED(bo);
2005 
2006 	mtx_lock(&sync_mtx);
2007 	if (bo->bo_flag & BO_ONWORKLST)
2008 		LIST_REMOVE(bo, bo_synclist);
2009 	else {
2010 		bo->bo_flag |= BO_ONWORKLST;
2011 		syncer_worklist_len++;
2012 	}
2013 
2014 	if (delay > syncer_maxdelay - 2)
2015 		delay = syncer_maxdelay - 2;
2016 	slot = (syncer_delayno + delay) & syncer_mask;
2017 
2018 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2019 	mtx_unlock(&sync_mtx);
2020 }
2021 
2022 static int
2023 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2024 {
2025 	int error, len;
2026 
2027 	mtx_lock(&sync_mtx);
2028 	len = syncer_worklist_len - sync_vnode_count;
2029 	mtx_unlock(&sync_mtx);
2030 	error = SYSCTL_OUT(req, &len, sizeof(len));
2031 	return (error);
2032 }
2033 
2034 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
2035     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2036 
2037 static struct proc *updateproc;
2038 static void sched_sync(void);
2039 static struct kproc_desc up_kp = {
2040 	"syncer",
2041 	sched_sync,
2042 	&updateproc
2043 };
2044 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2045 
2046 static int
2047 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2048 {
2049 	struct vnode *vp;
2050 	struct mount *mp;
2051 
2052 	*bo = LIST_FIRST(slp);
2053 	if (*bo == NULL)
2054 		return (0);
2055 	vp = (*bo)->__bo_vnode;	/* XXX */
2056 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2057 		return (1);
2058 	/*
2059 	 * We use vhold in case the vnode does not
2060 	 * successfully sync.  vhold prevents the vnode from
2061 	 * going away when we unlock the sync_mtx so that
2062 	 * we can acquire the vnode interlock.
2063 	 */
2064 	vholdl(vp);
2065 	mtx_unlock(&sync_mtx);
2066 	VI_UNLOCK(vp);
2067 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2068 		vdrop(vp);
2069 		mtx_lock(&sync_mtx);
2070 		return (*bo == LIST_FIRST(slp));
2071 	}
2072 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2073 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2074 	VOP_UNLOCK(vp, 0);
2075 	vn_finished_write(mp);
2076 	BO_LOCK(*bo);
2077 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2078 		/*
2079 		 * Put us back on the worklist.  The worklist
2080 		 * routine will remove us from our current
2081 		 * position and then add us back in at a later
2082 		 * position.
2083 		 */
2084 		vn_syncer_add_to_worklist(*bo, syncdelay);
2085 	}
2086 	BO_UNLOCK(*bo);
2087 	vdrop(vp);
2088 	mtx_lock(&sync_mtx);
2089 	return (0);
2090 }
2091 
2092 static int first_printf = 1;
2093 
2094 /*
2095  * System filesystem synchronizer daemon.
2096  */
2097 static void
2098 sched_sync(void)
2099 {
2100 	struct synclist *next, *slp;
2101 	struct bufobj *bo;
2102 	long starttime;
2103 	struct thread *td = curthread;
2104 	int last_work_seen;
2105 	int net_worklist_len;
2106 	int syncer_final_iter;
2107 	int error;
2108 
2109 	last_work_seen = 0;
2110 	syncer_final_iter = 0;
2111 	syncer_state = SYNCER_RUNNING;
2112 	starttime = time_uptime;
2113 	td->td_pflags |= TDP_NORUNNINGBUF;
2114 
2115 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2116 	    SHUTDOWN_PRI_LAST);
2117 
2118 	mtx_lock(&sync_mtx);
2119 	for (;;) {
2120 		if (syncer_state == SYNCER_FINAL_DELAY &&
2121 		    syncer_final_iter == 0) {
2122 			mtx_unlock(&sync_mtx);
2123 			kproc_suspend_check(td->td_proc);
2124 			mtx_lock(&sync_mtx);
2125 		}
2126 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2127 		if (syncer_state != SYNCER_RUNNING &&
2128 		    starttime != time_uptime) {
2129 			if (first_printf) {
2130 				printf("\nSyncing disks, vnodes remaining... ");
2131 				first_printf = 0;
2132 			}
2133 			printf("%d ", net_worklist_len);
2134 		}
2135 		starttime = time_uptime;
2136 
2137 		/*
2138 		 * Push files whose dirty time has expired.  Be careful
2139 		 * of interrupt race on slp queue.
2140 		 *
2141 		 * Skip over empty worklist slots when shutting down.
2142 		 */
2143 		do {
2144 			slp = &syncer_workitem_pending[syncer_delayno];
2145 			syncer_delayno += 1;
2146 			if (syncer_delayno == syncer_maxdelay)
2147 				syncer_delayno = 0;
2148 			next = &syncer_workitem_pending[syncer_delayno];
2149 			/*
2150 			 * If the worklist has wrapped since the
2151 			 * it was emptied of all but syncer vnodes,
2152 			 * switch to the FINAL_DELAY state and run
2153 			 * for one more second.
2154 			 */
2155 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2156 			    net_worklist_len == 0 &&
2157 			    last_work_seen == syncer_delayno) {
2158 				syncer_state = SYNCER_FINAL_DELAY;
2159 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2160 			}
2161 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2162 		    syncer_worklist_len > 0);
2163 
2164 		/*
2165 		 * Keep track of the last time there was anything
2166 		 * on the worklist other than syncer vnodes.
2167 		 * Return to the SHUTTING_DOWN state if any
2168 		 * new work appears.
2169 		 */
2170 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2171 			last_work_seen = syncer_delayno;
2172 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2173 			syncer_state = SYNCER_SHUTTING_DOWN;
2174 		while (!LIST_EMPTY(slp)) {
2175 			error = sync_vnode(slp, &bo, td);
2176 			if (error == 1) {
2177 				LIST_REMOVE(bo, bo_synclist);
2178 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2179 				continue;
2180 			}
2181 
2182 			if (first_printf == 0) {
2183 				/*
2184 				 * Drop the sync mutex, because some watchdog
2185 				 * drivers need to sleep while patting
2186 				 */
2187 				mtx_unlock(&sync_mtx);
2188 				wdog_kern_pat(WD_LASTVAL);
2189 				mtx_lock(&sync_mtx);
2190 			}
2191 
2192 		}
2193 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2194 			syncer_final_iter--;
2195 		/*
2196 		 * The variable rushjob allows the kernel to speed up the
2197 		 * processing of the filesystem syncer process. A rushjob
2198 		 * value of N tells the filesystem syncer to process the next
2199 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2200 		 * is used by the soft update code to speed up the filesystem
2201 		 * syncer process when the incore state is getting so far
2202 		 * ahead of the disk that the kernel memory pool is being
2203 		 * threatened with exhaustion.
2204 		 */
2205 		if (rushjob > 0) {
2206 			rushjob -= 1;
2207 			continue;
2208 		}
2209 		/*
2210 		 * Just sleep for a short period of time between
2211 		 * iterations when shutting down to allow some I/O
2212 		 * to happen.
2213 		 *
2214 		 * If it has taken us less than a second to process the
2215 		 * current work, then wait. Otherwise start right over
2216 		 * again. We can still lose time if any single round
2217 		 * takes more than two seconds, but it does not really
2218 		 * matter as we are just trying to generally pace the
2219 		 * filesystem activity.
2220 		 */
2221 		if (syncer_state != SYNCER_RUNNING ||
2222 		    time_uptime == starttime) {
2223 			thread_lock(td);
2224 			sched_prio(td, PPAUSE);
2225 			thread_unlock(td);
2226 		}
2227 		if (syncer_state != SYNCER_RUNNING)
2228 			cv_timedwait(&sync_wakeup, &sync_mtx,
2229 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2230 		else if (time_uptime == starttime)
2231 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2232 	}
2233 }
2234 
2235 /*
2236  * Request the syncer daemon to speed up its work.
2237  * We never push it to speed up more than half of its
2238  * normal turn time, otherwise it could take over the cpu.
2239  */
2240 int
2241 speedup_syncer(void)
2242 {
2243 	int ret = 0;
2244 
2245 	mtx_lock(&sync_mtx);
2246 	if (rushjob < syncdelay / 2) {
2247 		rushjob += 1;
2248 		stat_rush_requests += 1;
2249 		ret = 1;
2250 	}
2251 	mtx_unlock(&sync_mtx);
2252 	cv_broadcast(&sync_wakeup);
2253 	return (ret);
2254 }
2255 
2256 /*
2257  * Tell the syncer to speed up its work and run though its work
2258  * list several times, then tell it to shut down.
2259  */
2260 static void
2261 syncer_shutdown(void *arg, int howto)
2262 {
2263 
2264 	if (howto & RB_NOSYNC)
2265 		return;
2266 	mtx_lock(&sync_mtx);
2267 	syncer_state = SYNCER_SHUTTING_DOWN;
2268 	rushjob = 0;
2269 	mtx_unlock(&sync_mtx);
2270 	cv_broadcast(&sync_wakeup);
2271 	kproc_shutdown(arg, howto);
2272 }
2273 
2274 void
2275 syncer_suspend(void)
2276 {
2277 
2278 	syncer_shutdown(updateproc, 0);
2279 }
2280 
2281 void
2282 syncer_resume(void)
2283 {
2284 
2285 	mtx_lock(&sync_mtx);
2286 	first_printf = 1;
2287 	syncer_state = SYNCER_RUNNING;
2288 	mtx_unlock(&sync_mtx);
2289 	cv_broadcast(&sync_wakeup);
2290 	kproc_resume(updateproc);
2291 }
2292 
2293 /*
2294  * Reassign a buffer from one vnode to another.
2295  * Used to assign file specific control information
2296  * (indirect blocks) to the vnode to which they belong.
2297  */
2298 void
2299 reassignbuf(struct buf *bp)
2300 {
2301 	struct vnode *vp;
2302 	struct bufobj *bo;
2303 	int delay;
2304 #ifdef INVARIANTS
2305 	struct bufv *bv;
2306 #endif
2307 
2308 	vp = bp->b_vp;
2309 	bo = bp->b_bufobj;
2310 	++reassignbufcalls;
2311 
2312 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2313 	    bp, bp->b_vp, bp->b_flags);
2314 	/*
2315 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2316 	 * is not fully linked in.
2317 	 */
2318 	if (bp->b_flags & B_PAGING)
2319 		panic("cannot reassign paging buffer");
2320 
2321 	/*
2322 	 * Delete from old vnode list, if on one.
2323 	 */
2324 	BO_LOCK(bo);
2325 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2326 		buf_vlist_remove(bp);
2327 	else
2328 		panic("reassignbuf: Buffer %p not on queue.", bp);
2329 	/*
2330 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2331 	 * of clean buffers.
2332 	 */
2333 	if (bp->b_flags & B_DELWRI) {
2334 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2335 			switch (vp->v_type) {
2336 			case VDIR:
2337 				delay = dirdelay;
2338 				break;
2339 			case VCHR:
2340 				delay = metadelay;
2341 				break;
2342 			default:
2343 				delay = filedelay;
2344 			}
2345 			vn_syncer_add_to_worklist(bo, delay);
2346 		}
2347 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2348 	} else {
2349 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2350 
2351 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2352 			mtx_lock(&sync_mtx);
2353 			LIST_REMOVE(bo, bo_synclist);
2354 			syncer_worklist_len--;
2355 			mtx_unlock(&sync_mtx);
2356 			bo->bo_flag &= ~BO_ONWORKLST;
2357 		}
2358 	}
2359 #ifdef INVARIANTS
2360 	bv = &bo->bo_clean;
2361 	bp = TAILQ_FIRST(&bv->bv_hd);
2362 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2363 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2364 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2365 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2366 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2367 	bv = &bo->bo_dirty;
2368 	bp = TAILQ_FIRST(&bv->bv_hd);
2369 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2370 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2371 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2372 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2373 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2374 #endif
2375 	BO_UNLOCK(bo);
2376 }
2377 
2378 /*
2379  * A temporary hack until refcount_* APIs are sorted out.
2380  */
2381 static __inline int
2382 vfs_refcount_acquire_if_not_zero(volatile u_int *count)
2383 {
2384 	u_int old;
2385 
2386 	for (;;) {
2387 		old = *count;
2388 		if (old == 0)
2389 			return (0);
2390 		if (atomic_cmpset_int(count, old, old + 1))
2391 			return (1);
2392 	}
2393 }
2394 
2395 static __inline int
2396 vfs_refcount_release_if_not_last(volatile u_int *count)
2397 {
2398 	u_int old;
2399 
2400 	for (;;) {
2401 		old = *count;
2402 		if (old == 1)
2403 			return (0);
2404 		if (atomic_cmpset_int(count, old, old - 1))
2405 			return (1);
2406 	}
2407 }
2408 
2409 static void
2410 v_init_counters(struct vnode *vp)
2411 {
2412 
2413 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2414 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2415 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2416 
2417 	refcount_init(&vp->v_holdcnt, 1);
2418 	refcount_init(&vp->v_usecount, 1);
2419 }
2420 
2421 static void
2422 v_incr_usecount_locked(struct vnode *vp)
2423 {
2424 
2425 	ASSERT_VI_LOCKED(vp, __func__);
2426 	if ((vp->v_iflag & VI_OWEINACT) != 0) {
2427 		VNASSERT(vp->v_usecount == 0, vp,
2428 		    ("vnode with usecount and VI_OWEINACT set"));
2429 		vp->v_iflag &= ~VI_OWEINACT;
2430 	}
2431 	refcount_acquire(&vp->v_usecount);
2432 	v_incr_devcount(vp);
2433 }
2434 
2435 /*
2436  * Increment the use and hold counts on the vnode, taking care to reference
2437  * the driver's usecount if this is a chardev.  The _vhold() will remove
2438  * the vnode from the free list if it is presently free.
2439  */
2440 static void
2441 v_incr_usecount(struct vnode *vp)
2442 {
2443 
2444 	ASSERT_VI_UNLOCKED(vp, __func__);
2445 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2446 
2447 	if (vp->v_type != VCHR &&
2448 	    vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) {
2449 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2450 		    ("vnode with usecount and VI_OWEINACT set"));
2451 	} else {
2452 		VI_LOCK(vp);
2453 		v_incr_usecount_locked(vp);
2454 		VI_UNLOCK(vp);
2455 	}
2456 }
2457 
2458 /*
2459  * Increment si_usecount of the associated device, if any.
2460  */
2461 static void
2462 v_incr_devcount(struct vnode *vp)
2463 {
2464 
2465 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2466 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2467 		dev_lock();
2468 		vp->v_rdev->si_usecount++;
2469 		dev_unlock();
2470 	}
2471 }
2472 
2473 /*
2474  * Decrement si_usecount of the associated device, if any.
2475  */
2476 static void
2477 v_decr_devcount(struct vnode *vp)
2478 {
2479 
2480 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2481 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2482 		dev_lock();
2483 		vp->v_rdev->si_usecount--;
2484 		dev_unlock();
2485 	}
2486 }
2487 
2488 /*
2489  * Grab a particular vnode from the free list, increment its
2490  * reference count and lock it.  VI_DOOMED is set if the vnode
2491  * is being destroyed.  Only callers who specify LK_RETRY will
2492  * see doomed vnodes.  If inactive processing was delayed in
2493  * vput try to do it here.
2494  *
2495  * Notes on lockless counter manipulation:
2496  * _vhold, vputx and other routines make various decisions based
2497  * on either holdcnt or usecount being 0. As long as either counter
2498  * is not transitioning 0->1 nor 1->0, the manipulation can be done
2499  * with atomic operations. Otherwise the interlock is taken covering
2500  * both the atomic and additional actions.
2501  */
2502 int
2503 vget(struct vnode *vp, int flags, struct thread *td)
2504 {
2505 	int error, oweinact;
2506 
2507 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2508 	    ("vget: invalid lock operation"));
2509 
2510 	if ((flags & LK_INTERLOCK) != 0)
2511 		ASSERT_VI_LOCKED(vp, __func__);
2512 	else
2513 		ASSERT_VI_UNLOCKED(vp, __func__);
2514 	if ((flags & LK_VNHELD) != 0)
2515 		VNASSERT((vp->v_holdcnt > 0), vp,
2516 		    ("vget: LK_VNHELD passed but vnode not held"));
2517 
2518 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2519 
2520 	if ((flags & LK_VNHELD) == 0)
2521 		_vhold(vp, (flags & LK_INTERLOCK) != 0);
2522 
2523 	if ((error = vn_lock(vp, flags)) != 0) {
2524 		vdrop(vp);
2525 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2526 		    vp);
2527 		return (error);
2528 	}
2529 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
2530 		panic("vget: vn_lock failed to return ENOENT\n");
2531 	/*
2532 	 * We don't guarantee that any particular close will
2533 	 * trigger inactive processing so just make a best effort
2534 	 * here at preventing a reference to a removed file.  If
2535 	 * we don't succeed no harm is done.
2536 	 *
2537 	 * Upgrade our holdcnt to a usecount.
2538 	 */
2539 	if (vp->v_type == VCHR ||
2540 	    !vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) {
2541 		VI_LOCK(vp);
2542 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2543 			oweinact = 0;
2544 		} else {
2545 			oweinact = 1;
2546 			vp->v_iflag &= ~VI_OWEINACT;
2547 		}
2548 		refcount_acquire(&vp->v_usecount);
2549 		v_incr_devcount(vp);
2550 		if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2551 		    (flags & LK_NOWAIT) == 0)
2552 			vinactive(vp, td);
2553 		VI_UNLOCK(vp);
2554 	}
2555 	return (0);
2556 }
2557 
2558 /*
2559  * Increase the reference count of a vnode.
2560  */
2561 void
2562 vref(struct vnode *vp)
2563 {
2564 
2565 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2566 	_vhold(vp, false);
2567 	v_incr_usecount(vp);
2568 }
2569 
2570 void
2571 vrefl(struct vnode *vp)
2572 {
2573 
2574 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2575 	_vhold(vp, true);
2576 	v_incr_usecount_locked(vp);
2577 }
2578 
2579 /*
2580  * Return reference count of a vnode.
2581  *
2582  * The results of this call are only guaranteed when some mechanism is used to
2583  * stop other processes from gaining references to the vnode.  This may be the
2584  * case if the caller holds the only reference.  This is also useful when stale
2585  * data is acceptable as race conditions may be accounted for by some other
2586  * means.
2587  */
2588 int
2589 vrefcnt(struct vnode *vp)
2590 {
2591 
2592 	return (vp->v_usecount);
2593 }
2594 
2595 #define	VPUTX_VRELE	1
2596 #define	VPUTX_VPUT	2
2597 #define	VPUTX_VUNREF	3
2598 
2599 /*
2600  * Decrement the use and hold counts for a vnode.
2601  *
2602  * See an explanation near vget() as to why atomic operation is safe.
2603  */
2604 static void
2605 vputx(struct vnode *vp, int func)
2606 {
2607 	int error;
2608 
2609 	KASSERT(vp != NULL, ("vputx: null vp"));
2610 	if (func == VPUTX_VUNREF)
2611 		ASSERT_VOP_LOCKED(vp, "vunref");
2612 	else if (func == VPUTX_VPUT)
2613 		ASSERT_VOP_LOCKED(vp, "vput");
2614 	else
2615 		KASSERT(func == VPUTX_VRELE, ("vputx: wrong func"));
2616 	ASSERT_VI_UNLOCKED(vp, __func__);
2617 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2618 
2619 	if (vp->v_type != VCHR &&
2620 	    vfs_refcount_release_if_not_last(&vp->v_usecount)) {
2621 		if (func == VPUTX_VPUT)
2622 			VOP_UNLOCK(vp, 0);
2623 		vdrop(vp);
2624 		return;
2625 	}
2626 
2627 	VI_LOCK(vp);
2628 
2629 	/*
2630 	 * We want to hold the vnode until the inactive finishes to
2631 	 * prevent vgone() races.  We drop the use count here and the
2632 	 * hold count below when we're done.
2633 	 */
2634 	if (!refcount_release(&vp->v_usecount) ||
2635 	    (vp->v_iflag & VI_DOINGINACT)) {
2636 		if (func == VPUTX_VPUT)
2637 			VOP_UNLOCK(vp, 0);
2638 		v_decr_devcount(vp);
2639 		vdropl(vp);
2640 		return;
2641 	}
2642 
2643 	v_decr_devcount(vp);
2644 
2645 	error = 0;
2646 
2647 	if (vp->v_usecount != 0) {
2648 		vn_printf(vp, "vputx: usecount not zero for vnode ");
2649 		panic("vputx: usecount not zero");
2650 	}
2651 
2652 	CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp);
2653 
2654 	/*
2655 	 * We must call VOP_INACTIVE with the node locked. Mark
2656 	 * as VI_DOINGINACT to avoid recursion.
2657 	 */
2658 	vp->v_iflag |= VI_OWEINACT;
2659 	switch (func) {
2660 	case VPUTX_VRELE:
2661 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
2662 		VI_LOCK(vp);
2663 		break;
2664 	case VPUTX_VPUT:
2665 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2666 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
2667 			    LK_NOWAIT);
2668 			VI_LOCK(vp);
2669 		}
2670 		break;
2671 	case VPUTX_VUNREF:
2672 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2673 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
2674 			VI_LOCK(vp);
2675 		}
2676 		break;
2677 	}
2678 	VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
2679 	    ("vnode with usecount and VI_OWEINACT set"));
2680 	if (error == 0) {
2681 		if (vp->v_iflag & VI_OWEINACT)
2682 			vinactive(vp, curthread);
2683 		if (func != VPUTX_VUNREF)
2684 			VOP_UNLOCK(vp, 0);
2685 	}
2686 	vdropl(vp);
2687 }
2688 
2689 /*
2690  * Vnode put/release.
2691  * If count drops to zero, call inactive routine and return to freelist.
2692  */
2693 void
2694 vrele(struct vnode *vp)
2695 {
2696 
2697 	vputx(vp, VPUTX_VRELE);
2698 }
2699 
2700 /*
2701  * Release an already locked vnode.  This give the same effects as
2702  * unlock+vrele(), but takes less time and avoids releasing and
2703  * re-aquiring the lock (as vrele() acquires the lock internally.)
2704  */
2705 void
2706 vput(struct vnode *vp)
2707 {
2708 
2709 	vputx(vp, VPUTX_VPUT);
2710 }
2711 
2712 /*
2713  * Release an exclusively locked vnode. Do not unlock the vnode lock.
2714  */
2715 void
2716 vunref(struct vnode *vp)
2717 {
2718 
2719 	vputx(vp, VPUTX_VUNREF);
2720 }
2721 
2722 /*
2723  * Increase the hold count and activate if this is the first reference.
2724  */
2725 void
2726 _vhold(struct vnode *vp, bool locked)
2727 {
2728 	struct mount *mp;
2729 
2730 	if (locked)
2731 		ASSERT_VI_LOCKED(vp, __func__);
2732 	else
2733 		ASSERT_VI_UNLOCKED(vp, __func__);
2734 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2735 	if (!locked && vfs_refcount_acquire_if_not_zero(&vp->v_holdcnt)) {
2736 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2737 		    ("_vhold: vnode with holdcnt is free"));
2738 		return;
2739 	}
2740 
2741 	if (!locked)
2742 		VI_LOCK(vp);
2743 	if ((vp->v_iflag & VI_FREE) == 0) {
2744 		refcount_acquire(&vp->v_holdcnt);
2745 		if (!locked)
2746 			VI_UNLOCK(vp);
2747 		return;
2748 	}
2749 	VNASSERT(vp->v_holdcnt == 0, vp,
2750 	    ("%s: wrong hold count", __func__));
2751 	VNASSERT(vp->v_op != NULL, vp,
2752 	    ("%s: vnode already reclaimed.", __func__));
2753 	/*
2754 	 * Remove a vnode from the free list, mark it as in use,
2755 	 * and put it on the active list.
2756 	 */
2757 	mtx_lock(&vnode_free_list_mtx);
2758 	TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
2759 	freevnodes--;
2760 	vp->v_iflag &= ~VI_FREE;
2761 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
2762 	    ("Activating already active vnode"));
2763 	vp->v_iflag |= VI_ACTIVE;
2764 	mp = vp->v_mount;
2765 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
2766 	mp->mnt_activevnodelistsize++;
2767 	mtx_unlock(&vnode_free_list_mtx);
2768 	refcount_acquire(&vp->v_holdcnt);
2769 	if (!locked)
2770 		VI_UNLOCK(vp);
2771 }
2772 
2773 /*
2774  * Drop the hold count of the vnode.  If this is the last reference to
2775  * the vnode we place it on the free list unless it has been vgone'd
2776  * (marked VI_DOOMED) in which case we will free it.
2777  *
2778  * Because the vnode vm object keeps a hold reference on the vnode if
2779  * there is at least one resident non-cached page, the vnode cannot
2780  * leave the active list without the page cleanup done.
2781  */
2782 void
2783 _vdrop(struct vnode *vp, bool locked)
2784 {
2785 	struct bufobj *bo;
2786 	struct mount *mp;
2787 	int active;
2788 
2789 	if (locked)
2790 		ASSERT_VI_LOCKED(vp, __func__);
2791 	else
2792 		ASSERT_VI_UNLOCKED(vp, __func__);
2793 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2794 	if ((int)vp->v_holdcnt <= 0)
2795 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
2796 	if (vfs_refcount_release_if_not_last(&vp->v_holdcnt)) {
2797 		if (locked)
2798 			VI_UNLOCK(vp);
2799 		return;
2800 	}
2801 
2802 	if (!locked)
2803 		VI_LOCK(vp);
2804 	if (refcount_release(&vp->v_holdcnt) == 0) {
2805 		VI_UNLOCK(vp);
2806 		return;
2807 	}
2808 	if ((vp->v_iflag & VI_DOOMED) == 0) {
2809 		/*
2810 		 * Mark a vnode as free: remove it from its active list
2811 		 * and put it up for recycling on the freelist.
2812 		 */
2813 		VNASSERT(vp->v_op != NULL, vp,
2814 		    ("vdropl: vnode already reclaimed."));
2815 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2816 		    ("vnode already free"));
2817 		VNASSERT(vp->v_holdcnt == 0, vp,
2818 		    ("vdropl: freeing when we shouldn't"));
2819 		active = vp->v_iflag & VI_ACTIVE;
2820 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2821 			vp->v_iflag &= ~VI_ACTIVE;
2822 			mp = vp->v_mount;
2823 			mtx_lock(&vnode_free_list_mtx);
2824 			if (active) {
2825 				TAILQ_REMOVE(&mp->mnt_activevnodelist, vp,
2826 				    v_actfreelist);
2827 				mp->mnt_activevnodelistsize--;
2828 			}
2829 			TAILQ_INSERT_TAIL(&vnode_free_list, vp,
2830 			    v_actfreelist);
2831 			freevnodes++;
2832 			vp->v_iflag |= VI_FREE;
2833 			mtx_unlock(&vnode_free_list_mtx);
2834 		} else {
2835 			atomic_add_long(&free_owe_inact, 1);
2836 		}
2837 		VI_UNLOCK(vp);
2838 		return;
2839 	}
2840 	/*
2841 	 * The vnode has been marked for destruction, so free it.
2842 	 *
2843 	 * The vnode will be returned to the zone where it will
2844 	 * normally remain until it is needed for another vnode. We
2845 	 * need to cleanup (or verify that the cleanup has already
2846 	 * been done) any residual data left from its current use
2847 	 * so as not to contaminate the freshly allocated vnode.
2848 	 */
2849 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
2850 	atomic_subtract_long(&numvnodes, 1);
2851 	bo = &vp->v_bufobj;
2852 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2853 	    ("cleaned vnode still on the free list."));
2854 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
2855 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
2856 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
2857 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
2858 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
2859 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
2860 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
2861 	    ("clean blk trie not empty"));
2862 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
2863 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
2864 	    ("dirty blk trie not empty"));
2865 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
2866 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
2867 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
2868 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
2869 	    ("Dangling rangelock waiters"));
2870 	VI_UNLOCK(vp);
2871 #ifdef MAC
2872 	mac_vnode_destroy(vp);
2873 #endif
2874 	if (vp->v_pollinfo != NULL) {
2875 		destroy_vpollinfo(vp->v_pollinfo);
2876 		vp->v_pollinfo = NULL;
2877 	}
2878 #ifdef INVARIANTS
2879 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
2880 	vp->v_op = NULL;
2881 #endif
2882 	bzero(&vp->v_un, sizeof(vp->v_un));
2883 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
2884 	vp->v_iflag = 0;
2885 	vp->v_vflag = 0;
2886 	bo->bo_flag = 0;
2887 	uma_zfree(vnode_zone, vp);
2888 }
2889 
2890 /*
2891  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
2892  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
2893  * OWEINACT tracks whether a vnode missed a call to inactive due to a
2894  * failed lock upgrade.
2895  */
2896 void
2897 vinactive(struct vnode *vp, struct thread *td)
2898 {
2899 	struct vm_object *obj;
2900 
2901 	ASSERT_VOP_ELOCKED(vp, "vinactive");
2902 	ASSERT_VI_LOCKED(vp, "vinactive");
2903 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2904 	    ("vinactive: recursed on VI_DOINGINACT"));
2905 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2906 	vp->v_iflag |= VI_DOINGINACT;
2907 	vp->v_iflag &= ~VI_OWEINACT;
2908 	VI_UNLOCK(vp);
2909 	/*
2910 	 * Before moving off the active list, we must be sure that any
2911 	 * modified pages are converted into the vnode's dirty
2912 	 * buffers, since these will no longer be checked once the
2913 	 * vnode is on the inactive list.
2914 	 *
2915 	 * The write-out of the dirty pages is asynchronous.  At the
2916 	 * point that VOP_INACTIVE() is called, there could still be
2917 	 * pending I/O and dirty pages in the object.
2918 	 */
2919 	obj = vp->v_object;
2920 	if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0) {
2921 		VM_OBJECT_WLOCK(obj);
2922 		vm_object_page_clean(obj, 0, 0, OBJPC_NOSYNC);
2923 		VM_OBJECT_WUNLOCK(obj);
2924 	}
2925 	VOP_INACTIVE(vp, td);
2926 	VI_LOCK(vp);
2927 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2928 	    ("vinactive: lost VI_DOINGINACT"));
2929 	vp->v_iflag &= ~VI_DOINGINACT;
2930 }
2931 
2932 /*
2933  * Remove any vnodes in the vnode table belonging to mount point mp.
2934  *
2935  * If FORCECLOSE is not specified, there should not be any active ones,
2936  * return error if any are found (nb: this is a user error, not a
2937  * system error). If FORCECLOSE is specified, detach any active vnodes
2938  * that are found.
2939  *
2940  * If WRITECLOSE is set, only flush out regular file vnodes open for
2941  * writing.
2942  *
2943  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2944  *
2945  * `rootrefs' specifies the base reference count for the root vnode
2946  * of this filesystem. The root vnode is considered busy if its
2947  * v_usecount exceeds this value. On a successful return, vflush(, td)
2948  * will call vrele() on the root vnode exactly rootrefs times.
2949  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2950  * be zero.
2951  */
2952 #ifdef DIAGNOSTIC
2953 static int busyprt = 0;		/* print out busy vnodes */
2954 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
2955 #endif
2956 
2957 int
2958 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
2959 {
2960 	struct vnode *vp, *mvp, *rootvp = NULL;
2961 	struct vattr vattr;
2962 	int busy = 0, error;
2963 
2964 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
2965 	    rootrefs, flags);
2966 	if (rootrefs > 0) {
2967 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2968 		    ("vflush: bad args"));
2969 		/*
2970 		 * Get the filesystem root vnode. We can vput() it
2971 		 * immediately, since with rootrefs > 0, it won't go away.
2972 		 */
2973 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
2974 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
2975 			    __func__, error);
2976 			return (error);
2977 		}
2978 		vput(rootvp);
2979 	}
2980 loop:
2981 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
2982 		vholdl(vp);
2983 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
2984 		if (error) {
2985 			vdrop(vp);
2986 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
2987 			goto loop;
2988 		}
2989 		/*
2990 		 * Skip over a vnodes marked VV_SYSTEM.
2991 		 */
2992 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2993 			VOP_UNLOCK(vp, 0);
2994 			vdrop(vp);
2995 			continue;
2996 		}
2997 		/*
2998 		 * If WRITECLOSE is set, flush out unlinked but still open
2999 		 * files (even if open only for reading) and regular file
3000 		 * vnodes open for writing.
3001 		 */
3002 		if (flags & WRITECLOSE) {
3003 			if (vp->v_object != NULL) {
3004 				VM_OBJECT_WLOCK(vp->v_object);
3005 				vm_object_page_clean(vp->v_object, 0, 0, 0);
3006 				VM_OBJECT_WUNLOCK(vp->v_object);
3007 			}
3008 			error = VOP_FSYNC(vp, MNT_WAIT, td);
3009 			if (error != 0) {
3010 				VOP_UNLOCK(vp, 0);
3011 				vdrop(vp);
3012 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3013 				return (error);
3014 			}
3015 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3016 			VI_LOCK(vp);
3017 
3018 			if ((vp->v_type == VNON ||
3019 			    (error == 0 && vattr.va_nlink > 0)) &&
3020 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
3021 				VOP_UNLOCK(vp, 0);
3022 				vdropl(vp);
3023 				continue;
3024 			}
3025 		} else
3026 			VI_LOCK(vp);
3027 		/*
3028 		 * With v_usecount == 0, all we need to do is clear out the
3029 		 * vnode data structures and we are done.
3030 		 *
3031 		 * If FORCECLOSE is set, forcibly close the vnode.
3032 		 */
3033 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3034 			vgonel(vp);
3035 		} else {
3036 			busy++;
3037 #ifdef DIAGNOSTIC
3038 			if (busyprt)
3039 				vn_printf(vp, "vflush: busy vnode ");
3040 #endif
3041 		}
3042 		VOP_UNLOCK(vp, 0);
3043 		vdropl(vp);
3044 	}
3045 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3046 		/*
3047 		 * If just the root vnode is busy, and if its refcount
3048 		 * is equal to `rootrefs', then go ahead and kill it.
3049 		 */
3050 		VI_LOCK(rootvp);
3051 		KASSERT(busy > 0, ("vflush: not busy"));
3052 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3053 		    ("vflush: usecount %d < rootrefs %d",
3054 		     rootvp->v_usecount, rootrefs));
3055 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3056 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3057 			vgone(rootvp);
3058 			VOP_UNLOCK(rootvp, 0);
3059 			busy = 0;
3060 		} else
3061 			VI_UNLOCK(rootvp);
3062 	}
3063 	if (busy) {
3064 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3065 		    busy);
3066 		return (EBUSY);
3067 	}
3068 	for (; rootrefs > 0; rootrefs--)
3069 		vrele(rootvp);
3070 	return (0);
3071 }
3072 
3073 /*
3074  * Recycle an unused vnode to the front of the free list.
3075  */
3076 int
3077 vrecycle(struct vnode *vp)
3078 {
3079 	int recycled;
3080 
3081 	ASSERT_VOP_ELOCKED(vp, "vrecycle");
3082 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3083 	recycled = 0;
3084 	VI_LOCK(vp);
3085 	if (vp->v_usecount == 0) {
3086 		recycled = 1;
3087 		vgonel(vp);
3088 	}
3089 	VI_UNLOCK(vp);
3090 	return (recycled);
3091 }
3092 
3093 /*
3094  * Eliminate all activity associated with a vnode
3095  * in preparation for reuse.
3096  */
3097 void
3098 vgone(struct vnode *vp)
3099 {
3100 	VI_LOCK(vp);
3101 	vgonel(vp);
3102 	VI_UNLOCK(vp);
3103 }
3104 
3105 static void
3106 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3107     struct vnode *lowervp __unused)
3108 {
3109 }
3110 
3111 /*
3112  * Notify upper mounts about reclaimed or unlinked vnode.
3113  */
3114 void
3115 vfs_notify_upper(struct vnode *vp, int event)
3116 {
3117 	static struct vfsops vgonel_vfsops = {
3118 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3119 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3120 	};
3121 	struct mount *mp, *ump, *mmp;
3122 
3123 	mp = vp->v_mount;
3124 	if (mp == NULL)
3125 		return;
3126 
3127 	MNT_ILOCK(mp);
3128 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3129 		goto unlock;
3130 	MNT_IUNLOCK(mp);
3131 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3132 	mmp->mnt_op = &vgonel_vfsops;
3133 	mmp->mnt_kern_flag |= MNTK_MARKER;
3134 	MNT_ILOCK(mp);
3135 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3136 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3137 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3138 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3139 			continue;
3140 		}
3141 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3142 		MNT_IUNLOCK(mp);
3143 		switch (event) {
3144 		case VFS_NOTIFY_UPPER_RECLAIM:
3145 			VFS_RECLAIM_LOWERVP(ump, vp);
3146 			break;
3147 		case VFS_NOTIFY_UPPER_UNLINK:
3148 			VFS_UNLINK_LOWERVP(ump, vp);
3149 			break;
3150 		default:
3151 			KASSERT(0, ("invalid event %d", event));
3152 			break;
3153 		}
3154 		MNT_ILOCK(mp);
3155 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3156 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3157 	}
3158 	free(mmp, M_TEMP);
3159 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3160 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3161 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3162 		wakeup(&mp->mnt_uppers);
3163 	}
3164 unlock:
3165 	MNT_IUNLOCK(mp);
3166 }
3167 
3168 /*
3169  * vgone, with the vp interlock held.
3170  */
3171 static void
3172 vgonel(struct vnode *vp)
3173 {
3174 	struct thread *td;
3175 	int oweinact;
3176 	int active;
3177 	struct mount *mp;
3178 
3179 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3180 	ASSERT_VI_LOCKED(vp, "vgonel");
3181 	VNASSERT(vp->v_holdcnt, vp,
3182 	    ("vgonel: vp %p has no reference.", vp));
3183 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3184 	td = curthread;
3185 
3186 	/*
3187 	 * Don't vgonel if we're already doomed.
3188 	 */
3189 	if (vp->v_iflag & VI_DOOMED)
3190 		return;
3191 	vp->v_iflag |= VI_DOOMED;
3192 
3193 	/*
3194 	 * Check to see if the vnode is in use.  If so, we have to call
3195 	 * VOP_CLOSE() and VOP_INACTIVE().
3196 	 */
3197 	active = vp->v_usecount;
3198 	oweinact = (vp->v_iflag & VI_OWEINACT);
3199 	VI_UNLOCK(vp);
3200 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3201 
3202 	/*
3203 	 * If purging an active vnode, it must be closed and
3204 	 * deactivated before being reclaimed.
3205 	 */
3206 	if (active)
3207 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3208 	if (oweinact || active) {
3209 		VI_LOCK(vp);
3210 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
3211 			vinactive(vp, td);
3212 		VI_UNLOCK(vp);
3213 	}
3214 	if (vp->v_type == VSOCK)
3215 		vfs_unp_reclaim(vp);
3216 
3217 	/*
3218 	 * Clean out any buffers associated with the vnode.
3219 	 * If the flush fails, just toss the buffers.
3220 	 */
3221 	mp = NULL;
3222 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3223 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3224 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3225 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3226 			;
3227 	}
3228 
3229 	BO_LOCK(&vp->v_bufobj);
3230 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3231 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3232 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3233 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3234 	    ("vp %p bufobj not invalidated", vp));
3235 
3236 	/*
3237 	 * For VMIO bufobj, BO_DEAD is set in vm_object_terminate()
3238 	 * after the object's page queue is flushed.
3239 	 */
3240 	if (vp->v_bufobj.bo_object == NULL)
3241 		vp->v_bufobj.bo_flag |= BO_DEAD;
3242 	BO_UNLOCK(&vp->v_bufobj);
3243 
3244 	/*
3245 	 * Reclaim the vnode.
3246 	 */
3247 	if (VOP_RECLAIM(vp, td))
3248 		panic("vgone: cannot reclaim");
3249 	if (mp != NULL)
3250 		vn_finished_secondary_write(mp);
3251 	VNASSERT(vp->v_object == NULL, vp,
3252 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
3253 	/*
3254 	 * Clear the advisory locks and wake up waiting threads.
3255 	 */
3256 	(void)VOP_ADVLOCKPURGE(vp);
3257 	vp->v_lockf = NULL;
3258 	/*
3259 	 * Delete from old mount point vnode list.
3260 	 */
3261 	delmntque(vp);
3262 	cache_purge(vp);
3263 	/*
3264 	 * Done with purge, reset to the standard lock and invalidate
3265 	 * the vnode.
3266 	 */
3267 	VI_LOCK(vp);
3268 	vp->v_vnlock = &vp->v_lock;
3269 	vp->v_op = &dead_vnodeops;
3270 	vp->v_tag = "none";
3271 	vp->v_type = VBAD;
3272 }
3273 
3274 /*
3275  * Calculate the total number of references to a special device.
3276  */
3277 int
3278 vcount(struct vnode *vp)
3279 {
3280 	int count;
3281 
3282 	dev_lock();
3283 	count = vp->v_rdev->si_usecount;
3284 	dev_unlock();
3285 	return (count);
3286 }
3287 
3288 /*
3289  * Same as above, but using the struct cdev *as argument
3290  */
3291 int
3292 count_dev(struct cdev *dev)
3293 {
3294 	int count;
3295 
3296 	dev_lock();
3297 	count = dev->si_usecount;
3298 	dev_unlock();
3299 	return(count);
3300 }
3301 
3302 /*
3303  * Print out a description of a vnode.
3304  */
3305 static char *typename[] =
3306 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3307  "VMARKER"};
3308 
3309 void
3310 vn_printf(struct vnode *vp, const char *fmt, ...)
3311 {
3312 	va_list ap;
3313 	char buf[256], buf2[16];
3314 	u_long flags;
3315 
3316 	va_start(ap, fmt);
3317 	vprintf(fmt, ap);
3318 	va_end(ap);
3319 	printf("%p: ", (void *)vp);
3320 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
3321 	printf("    usecount %d, writecount %d, refcount %d mountedhere %p\n",
3322 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere);
3323 	buf[0] = '\0';
3324 	buf[1] = '\0';
3325 	if (vp->v_vflag & VV_ROOT)
3326 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3327 	if (vp->v_vflag & VV_ISTTY)
3328 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3329 	if (vp->v_vflag & VV_NOSYNC)
3330 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3331 	if (vp->v_vflag & VV_ETERNALDEV)
3332 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3333 	if (vp->v_vflag & VV_CACHEDLABEL)
3334 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3335 	if (vp->v_vflag & VV_TEXT)
3336 		strlcat(buf, "|VV_TEXT", sizeof(buf));
3337 	if (vp->v_vflag & VV_COPYONWRITE)
3338 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3339 	if (vp->v_vflag & VV_SYSTEM)
3340 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3341 	if (vp->v_vflag & VV_PROCDEP)
3342 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3343 	if (vp->v_vflag & VV_NOKNOTE)
3344 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3345 	if (vp->v_vflag & VV_DELETED)
3346 		strlcat(buf, "|VV_DELETED", sizeof(buf));
3347 	if (vp->v_vflag & VV_MD)
3348 		strlcat(buf, "|VV_MD", sizeof(buf));
3349 	if (vp->v_vflag & VV_FORCEINSMQ)
3350 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
3351 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
3352 	    VV_CACHEDLABEL | VV_TEXT | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
3353 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
3354 	if (flags != 0) {
3355 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
3356 		strlcat(buf, buf2, sizeof(buf));
3357 	}
3358 	if (vp->v_iflag & VI_MOUNT)
3359 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
3360 	if (vp->v_iflag & VI_DOOMED)
3361 		strlcat(buf, "|VI_DOOMED", sizeof(buf));
3362 	if (vp->v_iflag & VI_FREE)
3363 		strlcat(buf, "|VI_FREE", sizeof(buf));
3364 	if (vp->v_iflag & VI_ACTIVE)
3365 		strlcat(buf, "|VI_ACTIVE", sizeof(buf));
3366 	if (vp->v_iflag & VI_DOINGINACT)
3367 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
3368 	if (vp->v_iflag & VI_OWEINACT)
3369 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
3370 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOOMED | VI_FREE |
3371 	    VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT);
3372 	if (flags != 0) {
3373 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
3374 		strlcat(buf, buf2, sizeof(buf));
3375 	}
3376 	printf("    flags (%s)\n", buf + 1);
3377 	if (mtx_owned(VI_MTX(vp)))
3378 		printf(" VI_LOCKed");
3379 	if (vp->v_object != NULL)
3380 		printf("    v_object %p ref %d pages %d "
3381 		    "cleanbuf %d dirtybuf %d\n",
3382 		    vp->v_object, vp->v_object->ref_count,
3383 		    vp->v_object->resident_page_count,
3384 		    vp->v_bufobj.bo_clean.bv_cnt,
3385 		    vp->v_bufobj.bo_dirty.bv_cnt);
3386 	printf("    ");
3387 	lockmgr_printinfo(vp->v_vnlock);
3388 	if (vp->v_data != NULL)
3389 		VOP_PRINT(vp);
3390 }
3391 
3392 #ifdef DDB
3393 /*
3394  * List all of the locked vnodes in the system.
3395  * Called when debugging the kernel.
3396  */
3397 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
3398 {
3399 	struct mount *mp;
3400 	struct vnode *vp;
3401 
3402 	/*
3403 	 * Note: because this is DDB, we can't obey the locking semantics
3404 	 * for these structures, which means we could catch an inconsistent
3405 	 * state and dereference a nasty pointer.  Not much to be done
3406 	 * about that.
3407 	 */
3408 	db_printf("Locked vnodes\n");
3409 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3410 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3411 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
3412 				vn_printf(vp, "vnode ");
3413 		}
3414 	}
3415 }
3416 
3417 /*
3418  * Show details about the given vnode.
3419  */
3420 DB_SHOW_COMMAND(vnode, db_show_vnode)
3421 {
3422 	struct vnode *vp;
3423 
3424 	if (!have_addr)
3425 		return;
3426 	vp = (struct vnode *)addr;
3427 	vn_printf(vp, "vnode ");
3428 }
3429 
3430 /*
3431  * Show details about the given mount point.
3432  */
3433 DB_SHOW_COMMAND(mount, db_show_mount)
3434 {
3435 	struct mount *mp;
3436 	struct vfsopt *opt;
3437 	struct statfs *sp;
3438 	struct vnode *vp;
3439 	char buf[512];
3440 	uint64_t mflags;
3441 	u_int flags;
3442 
3443 	if (!have_addr) {
3444 		/* No address given, print short info about all mount points. */
3445 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3446 			db_printf("%p %s on %s (%s)\n", mp,
3447 			    mp->mnt_stat.f_mntfromname,
3448 			    mp->mnt_stat.f_mntonname,
3449 			    mp->mnt_stat.f_fstypename);
3450 			if (db_pager_quit)
3451 				break;
3452 		}
3453 		db_printf("\nMore info: show mount <addr>\n");
3454 		return;
3455 	}
3456 
3457 	mp = (struct mount *)addr;
3458 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
3459 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
3460 
3461 	buf[0] = '\0';
3462 	mflags = mp->mnt_flag;
3463 #define	MNT_FLAG(flag)	do {						\
3464 	if (mflags & (flag)) {						\
3465 		if (buf[0] != '\0')					\
3466 			strlcat(buf, ", ", sizeof(buf));		\
3467 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
3468 		mflags &= ~(flag);					\
3469 	}								\
3470 } while (0)
3471 	MNT_FLAG(MNT_RDONLY);
3472 	MNT_FLAG(MNT_SYNCHRONOUS);
3473 	MNT_FLAG(MNT_NOEXEC);
3474 	MNT_FLAG(MNT_NOSUID);
3475 	MNT_FLAG(MNT_NFS4ACLS);
3476 	MNT_FLAG(MNT_UNION);
3477 	MNT_FLAG(MNT_ASYNC);
3478 	MNT_FLAG(MNT_SUIDDIR);
3479 	MNT_FLAG(MNT_SOFTDEP);
3480 	MNT_FLAG(MNT_NOSYMFOLLOW);
3481 	MNT_FLAG(MNT_GJOURNAL);
3482 	MNT_FLAG(MNT_MULTILABEL);
3483 	MNT_FLAG(MNT_ACLS);
3484 	MNT_FLAG(MNT_NOATIME);
3485 	MNT_FLAG(MNT_NOCLUSTERR);
3486 	MNT_FLAG(MNT_NOCLUSTERW);
3487 	MNT_FLAG(MNT_SUJ);
3488 	MNT_FLAG(MNT_EXRDONLY);
3489 	MNT_FLAG(MNT_EXPORTED);
3490 	MNT_FLAG(MNT_DEFEXPORTED);
3491 	MNT_FLAG(MNT_EXPORTANON);
3492 	MNT_FLAG(MNT_EXKERB);
3493 	MNT_FLAG(MNT_EXPUBLIC);
3494 	MNT_FLAG(MNT_LOCAL);
3495 	MNT_FLAG(MNT_QUOTA);
3496 	MNT_FLAG(MNT_ROOTFS);
3497 	MNT_FLAG(MNT_USER);
3498 	MNT_FLAG(MNT_IGNORE);
3499 	MNT_FLAG(MNT_UPDATE);
3500 	MNT_FLAG(MNT_DELEXPORT);
3501 	MNT_FLAG(MNT_RELOAD);
3502 	MNT_FLAG(MNT_FORCE);
3503 	MNT_FLAG(MNT_SNAPSHOT);
3504 	MNT_FLAG(MNT_BYFSID);
3505 #undef MNT_FLAG
3506 	if (mflags != 0) {
3507 		if (buf[0] != '\0')
3508 			strlcat(buf, ", ", sizeof(buf));
3509 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3510 		    "0x%016jx", mflags);
3511 	}
3512 	db_printf("    mnt_flag = %s\n", buf);
3513 
3514 	buf[0] = '\0';
3515 	flags = mp->mnt_kern_flag;
3516 #define	MNT_KERN_FLAG(flag)	do {					\
3517 	if (flags & (flag)) {						\
3518 		if (buf[0] != '\0')					\
3519 			strlcat(buf, ", ", sizeof(buf));		\
3520 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
3521 		flags &= ~(flag);					\
3522 	}								\
3523 } while (0)
3524 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
3525 	MNT_KERN_FLAG(MNTK_ASYNC);
3526 	MNT_KERN_FLAG(MNTK_SOFTDEP);
3527 	MNT_KERN_FLAG(MNTK_NOINSMNTQ);
3528 	MNT_KERN_FLAG(MNTK_DRAINING);
3529 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
3530 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
3531 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
3532 	MNT_KERN_FLAG(MNTK_NO_IOPF);
3533 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
3534 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
3535 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
3536 	MNT_KERN_FLAG(MNTK_MARKER);
3537 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
3538 	MNT_KERN_FLAG(MNTK_NOASYNC);
3539 	MNT_KERN_FLAG(MNTK_UNMOUNT);
3540 	MNT_KERN_FLAG(MNTK_MWAIT);
3541 	MNT_KERN_FLAG(MNTK_SUSPEND);
3542 	MNT_KERN_FLAG(MNTK_SUSPEND2);
3543 	MNT_KERN_FLAG(MNTK_SUSPENDED);
3544 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
3545 	MNT_KERN_FLAG(MNTK_NOKNOTE);
3546 #undef MNT_KERN_FLAG
3547 	if (flags != 0) {
3548 		if (buf[0] != '\0')
3549 			strlcat(buf, ", ", sizeof(buf));
3550 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3551 		    "0x%08x", flags);
3552 	}
3553 	db_printf("    mnt_kern_flag = %s\n", buf);
3554 
3555 	db_printf("    mnt_opt = ");
3556 	opt = TAILQ_FIRST(mp->mnt_opt);
3557 	if (opt != NULL) {
3558 		db_printf("%s", opt->name);
3559 		opt = TAILQ_NEXT(opt, link);
3560 		while (opt != NULL) {
3561 			db_printf(", %s", opt->name);
3562 			opt = TAILQ_NEXT(opt, link);
3563 		}
3564 	}
3565 	db_printf("\n");
3566 
3567 	sp = &mp->mnt_stat;
3568 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
3569 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
3570 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
3571 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
3572 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
3573 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
3574 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
3575 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
3576 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
3577 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
3578 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
3579 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
3580 
3581 	db_printf("    mnt_cred = { uid=%u ruid=%u",
3582 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
3583 	if (jailed(mp->mnt_cred))
3584 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
3585 	db_printf(" }\n");
3586 	db_printf("    mnt_ref = %d\n", mp->mnt_ref);
3587 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
3588 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
3589 	db_printf("    mnt_activevnodelistsize = %d\n",
3590 	    mp->mnt_activevnodelistsize);
3591 	db_printf("    mnt_writeopcount = %d\n", mp->mnt_writeopcount);
3592 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
3593 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
3594 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
3595 	db_printf("    mnt_lockref = %d\n", mp->mnt_lockref);
3596 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
3597 	db_printf("    mnt_secondary_accwrites = %d\n",
3598 	    mp->mnt_secondary_accwrites);
3599 	db_printf("    mnt_gjprovider = %s\n",
3600 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
3601 
3602 	db_printf("\n\nList of active vnodes\n");
3603 	TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) {
3604 		if (vp->v_type != VMARKER) {
3605 			vn_printf(vp, "vnode ");
3606 			if (db_pager_quit)
3607 				break;
3608 		}
3609 	}
3610 	db_printf("\n\nList of inactive vnodes\n");
3611 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3612 		if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) {
3613 			vn_printf(vp, "vnode ");
3614 			if (db_pager_quit)
3615 				break;
3616 		}
3617 	}
3618 }
3619 #endif	/* DDB */
3620 
3621 /*
3622  * Fill in a struct xvfsconf based on a struct vfsconf.
3623  */
3624 static int
3625 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
3626 {
3627 	struct xvfsconf xvfsp;
3628 
3629 	bzero(&xvfsp, sizeof(xvfsp));
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 	/*
3635 	 * These are unused in userland, we keep them
3636 	 * to not break binary compatibility.
3637 	 */
3638 	xvfsp.vfc_vfsops = NULL;
3639 	xvfsp.vfc_next = NULL;
3640 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3641 }
3642 
3643 #ifdef COMPAT_FREEBSD32
3644 struct xvfsconf32 {
3645 	uint32_t	vfc_vfsops;
3646 	char		vfc_name[MFSNAMELEN];
3647 	int32_t		vfc_typenum;
3648 	int32_t		vfc_refcount;
3649 	int32_t		vfc_flags;
3650 	uint32_t	vfc_next;
3651 };
3652 
3653 static int
3654 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
3655 {
3656 	struct xvfsconf32 xvfsp;
3657 
3658 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3659 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3660 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3661 	xvfsp.vfc_flags = vfsp->vfc_flags;
3662 	xvfsp.vfc_vfsops = 0;
3663 	xvfsp.vfc_next = 0;
3664 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3665 }
3666 #endif
3667 
3668 /*
3669  * Top level filesystem related information gathering.
3670  */
3671 static int
3672 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
3673 {
3674 	struct vfsconf *vfsp;
3675 	int error;
3676 
3677 	error = 0;
3678 	vfsconf_slock();
3679 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3680 #ifdef COMPAT_FREEBSD32
3681 		if (req->flags & SCTL_MASK32)
3682 			error = vfsconf2x32(req, vfsp);
3683 		else
3684 #endif
3685 			error = vfsconf2x(req, vfsp);
3686 		if (error)
3687 			break;
3688 	}
3689 	vfsconf_sunlock();
3690 	return (error);
3691 }
3692 
3693 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
3694     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
3695     "S,xvfsconf", "List of all configured filesystems");
3696 
3697 #ifndef BURN_BRIDGES
3698 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
3699 
3700 static int
3701 vfs_sysctl(SYSCTL_HANDLER_ARGS)
3702 {
3703 	int *name = (int *)arg1 - 1;	/* XXX */
3704 	u_int namelen = arg2 + 1;	/* XXX */
3705 	struct vfsconf *vfsp;
3706 
3707 	log(LOG_WARNING, "userland calling deprecated sysctl, "
3708 	    "please rebuild world\n");
3709 
3710 #if 1 || defined(COMPAT_PRELITE2)
3711 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
3712 	if (namelen == 1)
3713 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
3714 #endif
3715 
3716 	switch (name[1]) {
3717 	case VFS_MAXTYPENUM:
3718 		if (namelen != 2)
3719 			return (ENOTDIR);
3720 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
3721 	case VFS_CONF:
3722 		if (namelen != 3)
3723 			return (ENOTDIR);	/* overloaded */
3724 		vfsconf_slock();
3725 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3726 			if (vfsp->vfc_typenum == name[2])
3727 				break;
3728 		}
3729 		vfsconf_sunlock();
3730 		if (vfsp == NULL)
3731 			return (EOPNOTSUPP);
3732 #ifdef COMPAT_FREEBSD32
3733 		if (req->flags & SCTL_MASK32)
3734 			return (vfsconf2x32(req, vfsp));
3735 		else
3736 #endif
3737 			return (vfsconf2x(req, vfsp));
3738 	}
3739 	return (EOPNOTSUPP);
3740 }
3741 
3742 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
3743     CTLFLAG_MPSAFE, vfs_sysctl,
3744     "Generic filesystem");
3745 
3746 #if 1 || defined(COMPAT_PRELITE2)
3747 
3748 static int
3749 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
3750 {
3751 	int error;
3752 	struct vfsconf *vfsp;
3753 	struct ovfsconf ovfs;
3754 
3755 	vfsconf_slock();
3756 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3757 		bzero(&ovfs, sizeof(ovfs));
3758 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
3759 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
3760 		ovfs.vfc_index = vfsp->vfc_typenum;
3761 		ovfs.vfc_refcount = vfsp->vfc_refcount;
3762 		ovfs.vfc_flags = vfsp->vfc_flags;
3763 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
3764 		if (error != 0) {
3765 			vfsconf_sunlock();
3766 			return (error);
3767 		}
3768 	}
3769 	vfsconf_sunlock();
3770 	return (0);
3771 }
3772 
3773 #endif /* 1 || COMPAT_PRELITE2 */
3774 #endif /* !BURN_BRIDGES */
3775 
3776 #define KINFO_VNODESLOP		10
3777 #ifdef notyet
3778 /*
3779  * Dump vnode list (via sysctl).
3780  */
3781 /* ARGSUSED */
3782 static int
3783 sysctl_vnode(SYSCTL_HANDLER_ARGS)
3784 {
3785 	struct xvnode *xvn;
3786 	struct mount *mp;
3787 	struct vnode *vp;
3788 	int error, len, n;
3789 
3790 	/*
3791 	 * Stale numvnodes access is not fatal here.
3792 	 */
3793 	req->lock = 0;
3794 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
3795 	if (!req->oldptr)
3796 		/* Make an estimate */
3797 		return (SYSCTL_OUT(req, 0, len));
3798 
3799 	error = sysctl_wire_old_buffer(req, 0);
3800 	if (error != 0)
3801 		return (error);
3802 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
3803 	n = 0;
3804 	mtx_lock(&mountlist_mtx);
3805 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3806 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
3807 			continue;
3808 		MNT_ILOCK(mp);
3809 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3810 			if (n == len)
3811 				break;
3812 			vref(vp);
3813 			xvn[n].xv_size = sizeof *xvn;
3814 			xvn[n].xv_vnode = vp;
3815 			xvn[n].xv_id = 0;	/* XXX compat */
3816 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
3817 			XV_COPY(usecount);
3818 			XV_COPY(writecount);
3819 			XV_COPY(holdcnt);
3820 			XV_COPY(mount);
3821 			XV_COPY(numoutput);
3822 			XV_COPY(type);
3823 #undef XV_COPY
3824 			xvn[n].xv_flag = vp->v_vflag;
3825 
3826 			switch (vp->v_type) {
3827 			case VREG:
3828 			case VDIR:
3829 			case VLNK:
3830 				break;
3831 			case VBLK:
3832 			case VCHR:
3833 				if (vp->v_rdev == NULL) {
3834 					vrele(vp);
3835 					continue;
3836 				}
3837 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
3838 				break;
3839 			case VSOCK:
3840 				xvn[n].xv_socket = vp->v_socket;
3841 				break;
3842 			case VFIFO:
3843 				xvn[n].xv_fifo = vp->v_fifoinfo;
3844 				break;
3845 			case VNON:
3846 			case VBAD:
3847 			default:
3848 				/* shouldn't happen? */
3849 				vrele(vp);
3850 				continue;
3851 			}
3852 			vrele(vp);
3853 			++n;
3854 		}
3855 		MNT_IUNLOCK(mp);
3856 		mtx_lock(&mountlist_mtx);
3857 		vfs_unbusy(mp);
3858 		if (n == len)
3859 			break;
3860 	}
3861 	mtx_unlock(&mountlist_mtx);
3862 
3863 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
3864 	free(xvn, M_TEMP);
3865 	return (error);
3866 }
3867 
3868 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
3869     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
3870     "");
3871 #endif
3872 
3873 static void
3874 unmount_or_warn(struct mount *mp)
3875 {
3876 	int error;
3877 
3878 	error = dounmount(mp, MNT_FORCE, curthread);
3879 	if (error != 0) {
3880 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
3881 		if (error == EBUSY)
3882 			printf("BUSY)\n");
3883 		else
3884 			printf("%d)\n", error);
3885 	}
3886 }
3887 
3888 /*
3889  * Unmount all filesystems. The list is traversed in reverse order
3890  * of mounting to avoid dependencies.
3891  */
3892 void
3893 vfs_unmountall(void)
3894 {
3895 	struct mount *mp, *tmp;
3896 
3897 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
3898 
3899 	/*
3900 	 * Since this only runs when rebooting, it is not interlocked.
3901 	 */
3902 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
3903 		vfs_ref(mp);
3904 
3905 		/*
3906 		 * Forcibly unmounting "/dev" before "/" would prevent clean
3907 		 * unmount of the latter.
3908 		 */
3909 		if (mp == rootdevmp)
3910 			continue;
3911 
3912 		unmount_or_warn(mp);
3913 	}
3914 
3915 	if (rootdevmp != NULL)
3916 		unmount_or_warn(rootdevmp);
3917 }
3918 
3919 /*
3920  * perform msync on all vnodes under a mount point
3921  * the mount point must be locked.
3922  */
3923 void
3924 vfs_msync(struct mount *mp, int flags)
3925 {
3926 	struct vnode *vp, *mvp;
3927 	struct vm_object *obj;
3928 
3929 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
3930 	MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
3931 		obj = vp->v_object;
3932 		if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 &&
3933 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
3934 			if (!vget(vp,
3935 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
3936 			    curthread)) {
3937 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
3938 					vput(vp);
3939 					continue;
3940 				}
3941 
3942 				obj = vp->v_object;
3943 				if (obj != NULL) {
3944 					VM_OBJECT_WLOCK(obj);
3945 					vm_object_page_clean(obj, 0, 0,
3946 					    flags == MNT_WAIT ?
3947 					    OBJPC_SYNC : OBJPC_NOSYNC);
3948 					VM_OBJECT_WUNLOCK(obj);
3949 				}
3950 				vput(vp);
3951 			}
3952 		} else
3953 			VI_UNLOCK(vp);
3954 	}
3955 }
3956 
3957 static void
3958 destroy_vpollinfo_free(struct vpollinfo *vi)
3959 {
3960 
3961 	knlist_destroy(&vi->vpi_selinfo.si_note);
3962 	mtx_destroy(&vi->vpi_lock);
3963 	uma_zfree(vnodepoll_zone, vi);
3964 }
3965 
3966 static void
3967 destroy_vpollinfo(struct vpollinfo *vi)
3968 {
3969 
3970 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
3971 	seldrain(&vi->vpi_selinfo);
3972 	destroy_vpollinfo_free(vi);
3973 }
3974 
3975 /*
3976  * Initialize per-vnode helper structure to hold poll-related state.
3977  */
3978 void
3979 v_addpollinfo(struct vnode *vp)
3980 {
3981 	struct vpollinfo *vi;
3982 
3983 	if (vp->v_pollinfo != NULL)
3984 		return;
3985 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
3986 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
3987 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
3988 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
3989 	VI_LOCK(vp);
3990 	if (vp->v_pollinfo != NULL) {
3991 		VI_UNLOCK(vp);
3992 		destroy_vpollinfo_free(vi);
3993 		return;
3994 	}
3995 	vp->v_pollinfo = vi;
3996 	VI_UNLOCK(vp);
3997 }
3998 
3999 /*
4000  * Record a process's interest in events which might happen to
4001  * a vnode.  Because poll uses the historic select-style interface
4002  * internally, this routine serves as both the ``check for any
4003  * pending events'' and the ``record my interest in future events''
4004  * functions.  (These are done together, while the lock is held,
4005  * to avoid race conditions.)
4006  */
4007 int
4008 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4009 {
4010 
4011 	v_addpollinfo(vp);
4012 	mtx_lock(&vp->v_pollinfo->vpi_lock);
4013 	if (vp->v_pollinfo->vpi_revents & events) {
4014 		/*
4015 		 * This leaves events we are not interested
4016 		 * in available for the other process which
4017 		 * which presumably had requested them
4018 		 * (otherwise they would never have been
4019 		 * recorded).
4020 		 */
4021 		events &= vp->v_pollinfo->vpi_revents;
4022 		vp->v_pollinfo->vpi_revents &= ~events;
4023 
4024 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
4025 		return (events);
4026 	}
4027 	vp->v_pollinfo->vpi_events |= events;
4028 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4029 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
4030 	return (0);
4031 }
4032 
4033 /*
4034  * Routine to create and manage a filesystem syncer vnode.
4035  */
4036 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
4037 static int	sync_fsync(struct  vop_fsync_args *);
4038 static int	sync_inactive(struct  vop_inactive_args *);
4039 static int	sync_reclaim(struct  vop_reclaim_args *);
4040 
4041 static struct vop_vector sync_vnodeops = {
4042 	.vop_bypass =	VOP_EOPNOTSUPP,
4043 	.vop_close =	sync_close,		/* close */
4044 	.vop_fsync =	sync_fsync,		/* fsync */
4045 	.vop_inactive =	sync_inactive,	/* inactive */
4046 	.vop_reclaim =	sync_reclaim,	/* reclaim */
4047 	.vop_lock1 =	vop_stdlock,	/* lock */
4048 	.vop_unlock =	vop_stdunlock,	/* unlock */
4049 	.vop_islocked =	vop_stdislocked,	/* islocked */
4050 };
4051 
4052 /*
4053  * Create a new filesystem syncer vnode for the specified mount point.
4054  */
4055 void
4056 vfs_allocate_syncvnode(struct mount *mp)
4057 {
4058 	struct vnode *vp;
4059 	struct bufobj *bo;
4060 	static long start, incr, next;
4061 	int error;
4062 
4063 	/* Allocate a new vnode */
4064 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4065 	if (error != 0)
4066 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4067 	vp->v_type = VNON;
4068 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4069 	vp->v_vflag |= VV_FORCEINSMQ;
4070 	error = insmntque(vp, mp);
4071 	if (error != 0)
4072 		panic("vfs_allocate_syncvnode: insmntque() failed");
4073 	vp->v_vflag &= ~VV_FORCEINSMQ;
4074 	VOP_UNLOCK(vp, 0);
4075 	/*
4076 	 * Place the vnode onto the syncer worklist. We attempt to
4077 	 * scatter them about on the list so that they will go off
4078 	 * at evenly distributed times even if all the filesystems
4079 	 * are mounted at once.
4080 	 */
4081 	next += incr;
4082 	if (next == 0 || next > syncer_maxdelay) {
4083 		start /= 2;
4084 		incr /= 2;
4085 		if (start == 0) {
4086 			start = syncer_maxdelay / 2;
4087 			incr = syncer_maxdelay;
4088 		}
4089 		next = start;
4090 	}
4091 	bo = &vp->v_bufobj;
4092 	BO_LOCK(bo);
4093 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4094 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4095 	mtx_lock(&sync_mtx);
4096 	sync_vnode_count++;
4097 	if (mp->mnt_syncer == NULL) {
4098 		mp->mnt_syncer = vp;
4099 		vp = NULL;
4100 	}
4101 	mtx_unlock(&sync_mtx);
4102 	BO_UNLOCK(bo);
4103 	if (vp != NULL) {
4104 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4105 		vgone(vp);
4106 		vput(vp);
4107 	}
4108 }
4109 
4110 void
4111 vfs_deallocate_syncvnode(struct mount *mp)
4112 {
4113 	struct vnode *vp;
4114 
4115 	mtx_lock(&sync_mtx);
4116 	vp = mp->mnt_syncer;
4117 	if (vp != NULL)
4118 		mp->mnt_syncer = NULL;
4119 	mtx_unlock(&sync_mtx);
4120 	if (vp != NULL)
4121 		vrele(vp);
4122 }
4123 
4124 /*
4125  * Do a lazy sync of the filesystem.
4126  */
4127 static int
4128 sync_fsync(struct vop_fsync_args *ap)
4129 {
4130 	struct vnode *syncvp = ap->a_vp;
4131 	struct mount *mp = syncvp->v_mount;
4132 	int error, save;
4133 	struct bufobj *bo;
4134 
4135 	/*
4136 	 * We only need to do something if this is a lazy evaluation.
4137 	 */
4138 	if (ap->a_waitfor != MNT_LAZY)
4139 		return (0);
4140 
4141 	/*
4142 	 * Move ourselves to the back of the sync list.
4143 	 */
4144 	bo = &syncvp->v_bufobj;
4145 	BO_LOCK(bo);
4146 	vn_syncer_add_to_worklist(bo, syncdelay);
4147 	BO_UNLOCK(bo);
4148 
4149 	/*
4150 	 * Walk the list of vnodes pushing all that are dirty and
4151 	 * not already on the sync list.
4152 	 */
4153 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4154 		return (0);
4155 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4156 		vfs_unbusy(mp);
4157 		return (0);
4158 	}
4159 	save = curthread_pflags_set(TDP_SYNCIO);
4160 	vfs_msync(mp, MNT_NOWAIT);
4161 	error = VFS_SYNC(mp, MNT_LAZY);
4162 	curthread_pflags_restore(save);
4163 	vn_finished_write(mp);
4164 	vfs_unbusy(mp);
4165 	return (error);
4166 }
4167 
4168 /*
4169  * The syncer vnode is no referenced.
4170  */
4171 static int
4172 sync_inactive(struct vop_inactive_args *ap)
4173 {
4174 
4175 	vgone(ap->a_vp);
4176 	return (0);
4177 }
4178 
4179 /*
4180  * The syncer vnode is no longer needed and is being decommissioned.
4181  *
4182  * Modifications to the worklist must be protected by sync_mtx.
4183  */
4184 static int
4185 sync_reclaim(struct vop_reclaim_args *ap)
4186 {
4187 	struct vnode *vp = ap->a_vp;
4188 	struct bufobj *bo;
4189 
4190 	bo = &vp->v_bufobj;
4191 	BO_LOCK(bo);
4192 	mtx_lock(&sync_mtx);
4193 	if (vp->v_mount->mnt_syncer == vp)
4194 		vp->v_mount->mnt_syncer = NULL;
4195 	if (bo->bo_flag & BO_ONWORKLST) {
4196 		LIST_REMOVE(bo, bo_synclist);
4197 		syncer_worklist_len--;
4198 		sync_vnode_count--;
4199 		bo->bo_flag &= ~BO_ONWORKLST;
4200 	}
4201 	mtx_unlock(&sync_mtx);
4202 	BO_UNLOCK(bo);
4203 
4204 	return (0);
4205 }
4206 
4207 /*
4208  * Check if vnode represents a disk device
4209  */
4210 int
4211 vn_isdisk(struct vnode *vp, int *errp)
4212 {
4213 	int error;
4214 
4215 	if (vp->v_type != VCHR) {
4216 		error = ENOTBLK;
4217 		goto out;
4218 	}
4219 	error = 0;
4220 	dev_lock();
4221 	if (vp->v_rdev == NULL)
4222 		error = ENXIO;
4223 	else if (vp->v_rdev->si_devsw == NULL)
4224 		error = ENXIO;
4225 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
4226 		error = ENOTBLK;
4227 	dev_unlock();
4228 out:
4229 	if (errp != NULL)
4230 		*errp = error;
4231 	return (error == 0);
4232 }
4233 
4234 /*
4235  * Common filesystem object access control check routine.  Accepts a
4236  * vnode's type, "mode", uid and gid, requested access mode, credentials,
4237  * and optional call-by-reference privused argument allowing vaccess()
4238  * to indicate to the caller whether privilege was used to satisfy the
4239  * request (obsoleted).  Returns 0 on success, or an errno on failure.
4240  */
4241 int
4242 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
4243     accmode_t accmode, struct ucred *cred, int *privused)
4244 {
4245 	accmode_t dac_granted;
4246 	accmode_t priv_granted;
4247 
4248 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
4249 	    ("invalid bit in accmode"));
4250 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
4251 	    ("VAPPEND without VWRITE"));
4252 
4253 	/*
4254 	 * Look for a normal, non-privileged way to access the file/directory
4255 	 * as requested.  If it exists, go with that.
4256 	 */
4257 
4258 	if (privused != NULL)
4259 		*privused = 0;
4260 
4261 	dac_granted = 0;
4262 
4263 	/* Check the owner. */
4264 	if (cred->cr_uid == file_uid) {
4265 		dac_granted |= VADMIN;
4266 		if (file_mode & S_IXUSR)
4267 			dac_granted |= VEXEC;
4268 		if (file_mode & S_IRUSR)
4269 			dac_granted |= VREAD;
4270 		if (file_mode & S_IWUSR)
4271 			dac_granted |= (VWRITE | VAPPEND);
4272 
4273 		if ((accmode & dac_granted) == accmode)
4274 			return (0);
4275 
4276 		goto privcheck;
4277 	}
4278 
4279 	/* Otherwise, check the groups (first match) */
4280 	if (groupmember(file_gid, cred)) {
4281 		if (file_mode & S_IXGRP)
4282 			dac_granted |= VEXEC;
4283 		if (file_mode & S_IRGRP)
4284 			dac_granted |= VREAD;
4285 		if (file_mode & S_IWGRP)
4286 			dac_granted |= (VWRITE | VAPPEND);
4287 
4288 		if ((accmode & dac_granted) == accmode)
4289 			return (0);
4290 
4291 		goto privcheck;
4292 	}
4293 
4294 	/* Otherwise, check everyone else. */
4295 	if (file_mode & S_IXOTH)
4296 		dac_granted |= VEXEC;
4297 	if (file_mode & S_IROTH)
4298 		dac_granted |= VREAD;
4299 	if (file_mode & S_IWOTH)
4300 		dac_granted |= (VWRITE | VAPPEND);
4301 	if ((accmode & dac_granted) == accmode)
4302 		return (0);
4303 
4304 privcheck:
4305 	/*
4306 	 * Build a privilege mask to determine if the set of privileges
4307 	 * satisfies the requirements when combined with the granted mask
4308 	 * from above.  For each privilege, if the privilege is required,
4309 	 * bitwise or the request type onto the priv_granted mask.
4310 	 */
4311 	priv_granted = 0;
4312 
4313 	if (type == VDIR) {
4314 		/*
4315 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
4316 		 * requests, instead of PRIV_VFS_EXEC.
4317 		 */
4318 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4319 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP, 0))
4320 			priv_granted |= VEXEC;
4321 	} else {
4322 		/*
4323 		 * Ensure that at least one execute bit is on. Otherwise,
4324 		 * a privileged user will always succeed, and we don't want
4325 		 * this to happen unless the file really is executable.
4326 		 */
4327 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4328 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
4329 		    !priv_check_cred(cred, PRIV_VFS_EXEC, 0))
4330 			priv_granted |= VEXEC;
4331 	}
4332 
4333 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
4334 	    !priv_check_cred(cred, PRIV_VFS_READ, 0))
4335 		priv_granted |= VREAD;
4336 
4337 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
4338 	    !priv_check_cred(cred, PRIV_VFS_WRITE, 0))
4339 		priv_granted |= (VWRITE | VAPPEND);
4340 
4341 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
4342 	    !priv_check_cred(cred, PRIV_VFS_ADMIN, 0))
4343 		priv_granted |= VADMIN;
4344 
4345 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
4346 		/* XXX audit: privilege used */
4347 		if (privused != NULL)
4348 			*privused = 1;
4349 		return (0);
4350 	}
4351 
4352 	return ((accmode & VADMIN) ? EPERM : EACCES);
4353 }
4354 
4355 /*
4356  * Credential check based on process requesting service, and per-attribute
4357  * permissions.
4358  */
4359 int
4360 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
4361     struct thread *td, accmode_t accmode)
4362 {
4363 
4364 	/*
4365 	 * Kernel-invoked always succeeds.
4366 	 */
4367 	if (cred == NOCRED)
4368 		return (0);
4369 
4370 	/*
4371 	 * Do not allow privileged processes in jail to directly manipulate
4372 	 * system attributes.
4373 	 */
4374 	switch (attrnamespace) {
4375 	case EXTATTR_NAMESPACE_SYSTEM:
4376 		/* Potentially should be: return (EPERM); */
4377 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0));
4378 	case EXTATTR_NAMESPACE_USER:
4379 		return (VOP_ACCESS(vp, accmode, cred, td));
4380 	default:
4381 		return (EPERM);
4382 	}
4383 }
4384 
4385 #ifdef DEBUG_VFS_LOCKS
4386 /*
4387  * This only exists to suppress warnings from unlocked specfs accesses.  It is
4388  * no longer ok to have an unlocked VFS.
4389  */
4390 #define	IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL ||		\
4391 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
4392 
4393 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
4394 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
4395     "Drop into debugger on lock violation");
4396 
4397 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
4398 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
4399     0, "Check for interlock across VOPs");
4400 
4401 int vfs_badlock_print = 1;	/* Print lock violations. */
4402 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
4403     0, "Print lock violations");
4404 
4405 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
4406 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
4407     0, "Print vnode details on lock violations");
4408 
4409 #ifdef KDB
4410 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
4411 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
4412     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
4413 #endif
4414 
4415 static void
4416 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
4417 {
4418 
4419 #ifdef KDB
4420 	if (vfs_badlock_backtrace)
4421 		kdb_backtrace();
4422 #endif
4423 	if (vfs_badlock_vnode)
4424 		vn_printf(vp, "vnode ");
4425 	if (vfs_badlock_print)
4426 		printf("%s: %p %s\n", str, (void *)vp, msg);
4427 	if (vfs_badlock_ddb)
4428 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4429 }
4430 
4431 void
4432 assert_vi_locked(struct vnode *vp, const char *str)
4433 {
4434 
4435 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
4436 		vfs_badlock("interlock is not locked but should be", str, vp);
4437 }
4438 
4439 void
4440 assert_vi_unlocked(struct vnode *vp, const char *str)
4441 {
4442 
4443 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
4444 		vfs_badlock("interlock is locked but should not be", str, vp);
4445 }
4446 
4447 void
4448 assert_vop_locked(struct vnode *vp, const char *str)
4449 {
4450 	int locked;
4451 
4452 	if (!IGNORE_LOCK(vp)) {
4453 		locked = VOP_ISLOCKED(vp);
4454 		if (locked == 0 || locked == LK_EXCLOTHER)
4455 			vfs_badlock("is not locked but should be", str, vp);
4456 	}
4457 }
4458 
4459 void
4460 assert_vop_unlocked(struct vnode *vp, const char *str)
4461 {
4462 
4463 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
4464 		vfs_badlock("is locked but should not be", str, vp);
4465 }
4466 
4467 void
4468 assert_vop_elocked(struct vnode *vp, const char *str)
4469 {
4470 
4471 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
4472 		vfs_badlock("is not exclusive locked but should be", str, vp);
4473 }
4474 #endif /* DEBUG_VFS_LOCKS */
4475 
4476 void
4477 vop_rename_fail(struct vop_rename_args *ap)
4478 {
4479 
4480 	if (ap->a_tvp != NULL)
4481 		vput(ap->a_tvp);
4482 	if (ap->a_tdvp == ap->a_tvp)
4483 		vrele(ap->a_tdvp);
4484 	else
4485 		vput(ap->a_tdvp);
4486 	vrele(ap->a_fdvp);
4487 	vrele(ap->a_fvp);
4488 }
4489 
4490 void
4491 vop_rename_pre(void *ap)
4492 {
4493 	struct vop_rename_args *a = ap;
4494 
4495 #ifdef DEBUG_VFS_LOCKS
4496 	if (a->a_tvp)
4497 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
4498 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
4499 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
4500 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
4501 
4502 	/* Check the source (from). */
4503 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
4504 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
4505 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
4506 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
4507 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
4508 
4509 	/* Check the target. */
4510 	if (a->a_tvp)
4511 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
4512 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
4513 #endif
4514 	if (a->a_tdvp != a->a_fdvp)
4515 		vhold(a->a_fdvp);
4516 	if (a->a_tvp != a->a_fvp)
4517 		vhold(a->a_fvp);
4518 	vhold(a->a_tdvp);
4519 	if (a->a_tvp)
4520 		vhold(a->a_tvp);
4521 }
4522 
4523 #ifdef DEBUG_VFS_LOCKS
4524 void
4525 vop_strategy_pre(void *ap)
4526 {
4527 	struct vop_strategy_args *a;
4528 	struct buf *bp;
4529 
4530 	a = ap;
4531 	bp = a->a_bp;
4532 
4533 	/*
4534 	 * Cluster ops lock their component buffers but not the IO container.
4535 	 */
4536 	if ((bp->b_flags & B_CLUSTER) != 0)
4537 		return;
4538 
4539 	if (panicstr == NULL && !BUF_ISLOCKED(bp)) {
4540 		if (vfs_badlock_print)
4541 			printf(
4542 			    "VOP_STRATEGY: bp is not locked but should be\n");
4543 		if (vfs_badlock_ddb)
4544 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4545 	}
4546 }
4547 
4548 void
4549 vop_lock_pre(void *ap)
4550 {
4551 	struct vop_lock1_args *a = ap;
4552 
4553 	if ((a->a_flags & LK_INTERLOCK) == 0)
4554 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4555 	else
4556 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
4557 }
4558 
4559 void
4560 vop_lock_post(void *ap, int rc)
4561 {
4562 	struct vop_lock1_args *a = ap;
4563 
4564 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4565 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
4566 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
4567 }
4568 
4569 void
4570 vop_unlock_pre(void *ap)
4571 {
4572 	struct vop_unlock_args *a = ap;
4573 
4574 	if (a->a_flags & LK_INTERLOCK)
4575 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
4576 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
4577 }
4578 
4579 void
4580 vop_unlock_post(void *ap, int rc)
4581 {
4582 	struct vop_unlock_args *a = ap;
4583 
4584 	if (a->a_flags & LK_INTERLOCK)
4585 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
4586 }
4587 #endif
4588 
4589 void
4590 vop_create_post(void *ap, int rc)
4591 {
4592 	struct vop_create_args *a = ap;
4593 
4594 	if (!rc)
4595 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4596 }
4597 
4598 void
4599 vop_deleteextattr_post(void *ap, int rc)
4600 {
4601 	struct vop_deleteextattr_args *a = ap;
4602 
4603 	if (!rc)
4604 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4605 }
4606 
4607 void
4608 vop_link_post(void *ap, int rc)
4609 {
4610 	struct vop_link_args *a = ap;
4611 
4612 	if (!rc) {
4613 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
4614 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
4615 	}
4616 }
4617 
4618 void
4619 vop_mkdir_post(void *ap, int rc)
4620 {
4621 	struct vop_mkdir_args *a = ap;
4622 
4623 	if (!rc)
4624 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4625 }
4626 
4627 void
4628 vop_mknod_post(void *ap, int rc)
4629 {
4630 	struct vop_mknod_args *a = ap;
4631 
4632 	if (!rc)
4633 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4634 }
4635 
4636 void
4637 vop_reclaim_post(void *ap, int rc)
4638 {
4639 	struct vop_reclaim_args *a = ap;
4640 
4641 	if (!rc)
4642 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
4643 }
4644 
4645 void
4646 vop_remove_post(void *ap, int rc)
4647 {
4648 	struct vop_remove_args *a = ap;
4649 
4650 	if (!rc) {
4651 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4652 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4653 	}
4654 }
4655 
4656 void
4657 vop_rename_post(void *ap, int rc)
4658 {
4659 	struct vop_rename_args *a = ap;
4660 	long hint;
4661 
4662 	if (!rc) {
4663 		hint = NOTE_WRITE;
4664 		if (a->a_fdvp == a->a_tdvp) {
4665 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
4666 				hint |= NOTE_LINK;
4667 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
4668 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
4669 		} else {
4670 			hint |= NOTE_EXTEND;
4671 			if (a->a_fvp->v_type == VDIR)
4672 				hint |= NOTE_LINK;
4673 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
4674 
4675 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
4676 			    a->a_tvp->v_type == VDIR)
4677 				hint &= ~NOTE_LINK;
4678 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
4679 		}
4680 
4681 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
4682 		if (a->a_tvp)
4683 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
4684 	}
4685 	if (a->a_tdvp != a->a_fdvp)
4686 		vdrop(a->a_fdvp);
4687 	if (a->a_tvp != a->a_fvp)
4688 		vdrop(a->a_fvp);
4689 	vdrop(a->a_tdvp);
4690 	if (a->a_tvp)
4691 		vdrop(a->a_tvp);
4692 }
4693 
4694 void
4695 vop_rmdir_post(void *ap, int rc)
4696 {
4697 	struct vop_rmdir_args *a = ap;
4698 
4699 	if (!rc) {
4700 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4701 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4702 	}
4703 }
4704 
4705 void
4706 vop_setattr_post(void *ap, int rc)
4707 {
4708 	struct vop_setattr_args *a = ap;
4709 
4710 	if (!rc)
4711 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4712 }
4713 
4714 void
4715 vop_setextattr_post(void *ap, int rc)
4716 {
4717 	struct vop_setextattr_args *a = ap;
4718 
4719 	if (!rc)
4720 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4721 }
4722 
4723 void
4724 vop_symlink_post(void *ap, int rc)
4725 {
4726 	struct vop_symlink_args *a = ap;
4727 
4728 	if (!rc)
4729 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4730 }
4731 
4732 void
4733 vop_open_post(void *ap, int rc)
4734 {
4735 	struct vop_open_args *a = ap;
4736 
4737 	if (!rc)
4738 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
4739 }
4740 
4741 void
4742 vop_close_post(void *ap, int rc)
4743 {
4744 	struct vop_close_args *a = ap;
4745 
4746 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
4747 	    (a->a_vp->v_iflag & VI_DOOMED) == 0)) {
4748 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
4749 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
4750 	}
4751 }
4752 
4753 void
4754 vop_read_post(void *ap, int rc)
4755 {
4756 	struct vop_read_args *a = ap;
4757 
4758 	if (!rc)
4759 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
4760 }
4761 
4762 void
4763 vop_readdir_post(void *ap, int rc)
4764 {
4765 	struct vop_readdir_args *a = ap;
4766 
4767 	if (!rc)
4768 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
4769 }
4770 
4771 static struct knlist fs_knlist;
4772 
4773 static void
4774 vfs_event_init(void *arg)
4775 {
4776 	knlist_init_mtx(&fs_knlist, NULL);
4777 }
4778 /* XXX - correct order? */
4779 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
4780 
4781 void
4782 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
4783 {
4784 
4785 	KNOTE_UNLOCKED(&fs_knlist, event);
4786 }
4787 
4788 static int	filt_fsattach(struct knote *kn);
4789 static void	filt_fsdetach(struct knote *kn);
4790 static int	filt_fsevent(struct knote *kn, long hint);
4791 
4792 struct filterops fs_filtops = {
4793 	.f_isfd = 0,
4794 	.f_attach = filt_fsattach,
4795 	.f_detach = filt_fsdetach,
4796 	.f_event = filt_fsevent
4797 };
4798 
4799 static int
4800 filt_fsattach(struct knote *kn)
4801 {
4802 
4803 	kn->kn_flags |= EV_CLEAR;
4804 	knlist_add(&fs_knlist, kn, 0);
4805 	return (0);
4806 }
4807 
4808 static void
4809 filt_fsdetach(struct knote *kn)
4810 {
4811 
4812 	knlist_remove(&fs_knlist, kn, 0);
4813 }
4814 
4815 static int
4816 filt_fsevent(struct knote *kn, long hint)
4817 {
4818 
4819 	kn->kn_fflags |= hint;
4820 	return (kn->kn_fflags != 0);
4821 }
4822 
4823 static int
4824 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
4825 {
4826 	struct vfsidctl vc;
4827 	int error;
4828 	struct mount *mp;
4829 
4830 	error = SYSCTL_IN(req, &vc, sizeof(vc));
4831 	if (error)
4832 		return (error);
4833 	if (vc.vc_vers != VFS_CTL_VERS1)
4834 		return (EINVAL);
4835 	mp = vfs_getvfs(&vc.vc_fsid);
4836 	if (mp == NULL)
4837 		return (ENOENT);
4838 	/* ensure that a specific sysctl goes to the right filesystem. */
4839 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
4840 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
4841 		vfs_rel(mp);
4842 		return (EINVAL);
4843 	}
4844 	VCTLTOREQ(&vc, req);
4845 	error = VFS_SYSCTL(mp, vc.vc_op, req);
4846 	vfs_rel(mp);
4847 	return (error);
4848 }
4849 
4850 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR,
4851     NULL, 0, sysctl_vfs_ctl, "",
4852     "Sysctl by fsid");
4853 
4854 /*
4855  * Function to initialize a va_filerev field sensibly.
4856  * XXX: Wouldn't a random number make a lot more sense ??
4857  */
4858 u_quad_t
4859 init_va_filerev(void)
4860 {
4861 	struct bintime bt;
4862 
4863 	getbinuptime(&bt);
4864 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
4865 }
4866 
4867 static int	filt_vfsread(struct knote *kn, long hint);
4868 static int	filt_vfswrite(struct knote *kn, long hint);
4869 static int	filt_vfsvnode(struct knote *kn, long hint);
4870 static void	filt_vfsdetach(struct knote *kn);
4871 static struct filterops vfsread_filtops = {
4872 	.f_isfd = 1,
4873 	.f_detach = filt_vfsdetach,
4874 	.f_event = filt_vfsread
4875 };
4876 static struct filterops vfswrite_filtops = {
4877 	.f_isfd = 1,
4878 	.f_detach = filt_vfsdetach,
4879 	.f_event = filt_vfswrite
4880 };
4881 static struct filterops vfsvnode_filtops = {
4882 	.f_isfd = 1,
4883 	.f_detach = filt_vfsdetach,
4884 	.f_event = filt_vfsvnode
4885 };
4886 
4887 static void
4888 vfs_knllock(void *arg)
4889 {
4890 	struct vnode *vp = arg;
4891 
4892 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4893 }
4894 
4895 static void
4896 vfs_knlunlock(void *arg)
4897 {
4898 	struct vnode *vp = arg;
4899 
4900 	VOP_UNLOCK(vp, 0);
4901 }
4902 
4903 static void
4904 vfs_knl_assert_locked(void *arg)
4905 {
4906 #ifdef DEBUG_VFS_LOCKS
4907 	struct vnode *vp = arg;
4908 
4909 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
4910 #endif
4911 }
4912 
4913 static void
4914 vfs_knl_assert_unlocked(void *arg)
4915 {
4916 #ifdef DEBUG_VFS_LOCKS
4917 	struct vnode *vp = arg;
4918 
4919 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
4920 #endif
4921 }
4922 
4923 int
4924 vfs_kqfilter(struct vop_kqfilter_args *ap)
4925 {
4926 	struct vnode *vp = ap->a_vp;
4927 	struct knote *kn = ap->a_kn;
4928 	struct knlist *knl;
4929 
4930 	switch (kn->kn_filter) {
4931 	case EVFILT_READ:
4932 		kn->kn_fop = &vfsread_filtops;
4933 		break;
4934 	case EVFILT_WRITE:
4935 		kn->kn_fop = &vfswrite_filtops;
4936 		break;
4937 	case EVFILT_VNODE:
4938 		kn->kn_fop = &vfsvnode_filtops;
4939 		break;
4940 	default:
4941 		return (EINVAL);
4942 	}
4943 
4944 	kn->kn_hook = (caddr_t)vp;
4945 
4946 	v_addpollinfo(vp);
4947 	if (vp->v_pollinfo == NULL)
4948 		return (ENOMEM);
4949 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
4950 	vhold(vp);
4951 	knlist_add(knl, kn, 0);
4952 
4953 	return (0);
4954 }
4955 
4956 /*
4957  * Detach knote from vnode
4958  */
4959 static void
4960 filt_vfsdetach(struct knote *kn)
4961 {
4962 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4963 
4964 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
4965 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
4966 	vdrop(vp);
4967 }
4968 
4969 /*ARGSUSED*/
4970 static int
4971 filt_vfsread(struct knote *kn, long hint)
4972 {
4973 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4974 	struct vattr va;
4975 	int res;
4976 
4977 	/*
4978 	 * filesystem is gone, so set the EOF flag and schedule
4979 	 * the knote for deletion.
4980 	 */
4981 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
4982 		VI_LOCK(vp);
4983 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
4984 		VI_UNLOCK(vp);
4985 		return (1);
4986 	}
4987 
4988 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
4989 		return (0);
4990 
4991 	VI_LOCK(vp);
4992 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
4993 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
4994 	VI_UNLOCK(vp);
4995 	return (res);
4996 }
4997 
4998 /*ARGSUSED*/
4999 static int
5000 filt_vfswrite(struct knote *kn, long hint)
5001 {
5002 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5003 
5004 	VI_LOCK(vp);
5005 
5006 	/*
5007 	 * filesystem is gone, so set the EOF flag and schedule
5008 	 * the knote for deletion.
5009 	 */
5010 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5011 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5012 
5013 	kn->kn_data = 0;
5014 	VI_UNLOCK(vp);
5015 	return (1);
5016 }
5017 
5018 static int
5019 filt_vfsvnode(struct knote *kn, long hint)
5020 {
5021 	struct vnode *vp = (struct vnode *)kn->kn_hook;
5022 	int res;
5023 
5024 	VI_LOCK(vp);
5025 	if (kn->kn_sfflags & hint)
5026 		kn->kn_fflags |= hint;
5027 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5028 		kn->kn_flags |= EV_EOF;
5029 		VI_UNLOCK(vp);
5030 		return (1);
5031 	}
5032 	res = (kn->kn_fflags != 0);
5033 	VI_UNLOCK(vp);
5034 	return (res);
5035 }
5036 
5037 int
5038 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5039 {
5040 	int error;
5041 
5042 	if (dp->d_reclen > ap->a_uio->uio_resid)
5043 		return (ENAMETOOLONG);
5044 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
5045 	if (error) {
5046 		if (ap->a_ncookies != NULL) {
5047 			if (ap->a_cookies != NULL)
5048 				free(ap->a_cookies, M_TEMP);
5049 			ap->a_cookies = NULL;
5050 			*ap->a_ncookies = 0;
5051 		}
5052 		return (error);
5053 	}
5054 	if (ap->a_ncookies == NULL)
5055 		return (0);
5056 
5057 	KASSERT(ap->a_cookies,
5058 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
5059 
5060 	*ap->a_cookies = realloc(*ap->a_cookies,
5061 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
5062 	(*ap->a_cookies)[*ap->a_ncookies] = off;
5063 	*ap->a_ncookies += 1;
5064 	return (0);
5065 }
5066 
5067 /*
5068  * Mark for update the access time of the file if the filesystem
5069  * supports VOP_MARKATIME.  This functionality is used by execve and
5070  * mmap, so we want to avoid the I/O implied by directly setting
5071  * va_atime for the sake of efficiency.
5072  */
5073 void
5074 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
5075 {
5076 	struct mount *mp;
5077 
5078 	mp = vp->v_mount;
5079 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
5080 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
5081 		(void)VOP_MARKATIME(vp);
5082 }
5083 
5084 /*
5085  * The purpose of this routine is to remove granularity from accmode_t,
5086  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
5087  * VADMIN and VAPPEND.
5088  *
5089  * If it returns 0, the caller is supposed to continue with the usual
5090  * access checks using 'accmode' as modified by this routine.  If it
5091  * returns nonzero value, the caller is supposed to return that value
5092  * as errno.
5093  *
5094  * Note that after this routine runs, accmode may be zero.
5095  */
5096 int
5097 vfs_unixify_accmode(accmode_t *accmode)
5098 {
5099 	/*
5100 	 * There is no way to specify explicit "deny" rule using
5101 	 * file mode or POSIX.1e ACLs.
5102 	 */
5103 	if (*accmode & VEXPLICIT_DENY) {
5104 		*accmode = 0;
5105 		return (0);
5106 	}
5107 
5108 	/*
5109 	 * None of these can be translated into usual access bits.
5110 	 * Also, the common case for NFSv4 ACLs is to not contain
5111 	 * either of these bits. Caller should check for VWRITE
5112 	 * on the containing directory instead.
5113 	 */
5114 	if (*accmode & (VDELETE_CHILD | VDELETE))
5115 		return (EPERM);
5116 
5117 	if (*accmode & VADMIN_PERMS) {
5118 		*accmode &= ~VADMIN_PERMS;
5119 		*accmode |= VADMIN;
5120 	}
5121 
5122 	/*
5123 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5124 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5125 	 */
5126 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5127 
5128 	return (0);
5129 }
5130 
5131 /*
5132  * These are helper functions for filesystems to traverse all
5133  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
5134  *
5135  * This interface replaces MNT_VNODE_FOREACH.
5136  */
5137 
5138 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
5139 
5140 struct vnode *
5141 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
5142 {
5143 	struct vnode *vp;
5144 
5145 	if (should_yield())
5146 		kern_yield(PRI_USER);
5147 	MNT_ILOCK(mp);
5148 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5149 	vp = TAILQ_NEXT(*mvp, v_nmntvnodes);
5150 	while (vp != NULL && (vp->v_type == VMARKER ||
5151 	    (vp->v_iflag & VI_DOOMED) != 0))
5152 		vp = TAILQ_NEXT(vp, v_nmntvnodes);
5153 
5154 	/* Check if we are done */
5155 	if (vp == NULL) {
5156 		__mnt_vnode_markerfree_all(mvp, mp);
5157 		/* MNT_IUNLOCK(mp); -- done in above function */
5158 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
5159 		return (NULL);
5160 	}
5161 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5162 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5163 	VI_LOCK(vp);
5164 	MNT_IUNLOCK(mp);
5165 	return (vp);
5166 }
5167 
5168 struct vnode *
5169 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
5170 {
5171 	struct vnode *vp;
5172 
5173 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5174 	MNT_ILOCK(mp);
5175 	MNT_REF(mp);
5176 	(*mvp)->v_type = VMARKER;
5177 
5178 	vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
5179 	while (vp != NULL && (vp->v_type == VMARKER ||
5180 	    (vp->v_iflag & VI_DOOMED) != 0))
5181 		vp = TAILQ_NEXT(vp, v_nmntvnodes);
5182 
5183 	/* Check if we are done */
5184 	if (vp == NULL) {
5185 		MNT_REL(mp);
5186 		MNT_IUNLOCK(mp);
5187 		free(*mvp, M_VNODE_MARKER);
5188 		*mvp = NULL;
5189 		return (NULL);
5190 	}
5191 	(*mvp)->v_mount = mp;
5192 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5193 	VI_LOCK(vp);
5194 	MNT_IUNLOCK(mp);
5195 	return (vp);
5196 }
5197 
5198 
5199 void
5200 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
5201 {
5202 
5203 	if (*mvp == NULL) {
5204 		MNT_IUNLOCK(mp);
5205 		return;
5206 	}
5207 
5208 	mtx_assert(MNT_MTX(mp), MA_OWNED);
5209 
5210 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5211 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5212 	MNT_REL(mp);
5213 	MNT_IUNLOCK(mp);
5214 	free(*mvp, M_VNODE_MARKER);
5215 	*mvp = NULL;
5216 }
5217 
5218 /*
5219  * These are helper functions for filesystems to traverse their
5220  * active vnodes.  See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h
5221  */
5222 static void
5223 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5224 {
5225 
5226 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5227 
5228 	MNT_ILOCK(mp);
5229 	MNT_REL(mp);
5230 	MNT_IUNLOCK(mp);
5231 	free(*mvp, M_VNODE_MARKER);
5232 	*mvp = NULL;
5233 }
5234 
5235 static struct vnode *
5236 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5237 {
5238 	struct vnode *vp, *nvp;
5239 
5240 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
5241 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5242 restart:
5243 	vp = TAILQ_NEXT(*mvp, v_actfreelist);
5244 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5245 	while (vp != NULL) {
5246 		if (vp->v_type == VMARKER) {
5247 			vp = TAILQ_NEXT(vp, v_actfreelist);
5248 			continue;
5249 		}
5250 		if (!VI_TRYLOCK(vp)) {
5251 			if (mp_ncpus == 1 || should_yield()) {
5252 				TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5253 				mtx_unlock(&vnode_free_list_mtx);
5254 				pause("vnacti", 1);
5255 				mtx_lock(&vnode_free_list_mtx);
5256 				goto restart;
5257 			}
5258 			continue;
5259 		}
5260 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
5261 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
5262 		    ("alien vnode on the active list %p %p", vp, mp));
5263 		if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0)
5264 			break;
5265 		nvp = TAILQ_NEXT(vp, v_actfreelist);
5266 		VI_UNLOCK(vp);
5267 		vp = nvp;
5268 	}
5269 
5270 	/* Check if we are done */
5271 	if (vp == NULL) {
5272 		mtx_unlock(&vnode_free_list_mtx);
5273 		mnt_vnode_markerfree_active(mvp, mp);
5274 		return (NULL);
5275 	}
5276 	TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist);
5277 	mtx_unlock(&vnode_free_list_mtx);
5278 	ASSERT_VI_LOCKED(vp, "active iter");
5279 	KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp));
5280 	return (vp);
5281 }
5282 
5283 struct vnode *
5284 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5285 {
5286 
5287 	if (should_yield())
5288 		kern_yield(PRI_USER);
5289 	mtx_lock(&vnode_free_list_mtx);
5290 	return (mnt_vnode_next_active(mvp, mp));
5291 }
5292 
5293 struct vnode *
5294 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp)
5295 {
5296 	struct vnode *vp;
5297 
5298 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5299 	MNT_ILOCK(mp);
5300 	MNT_REF(mp);
5301 	MNT_IUNLOCK(mp);
5302 	(*mvp)->v_type = VMARKER;
5303 	(*mvp)->v_mount = mp;
5304 
5305 	mtx_lock(&vnode_free_list_mtx);
5306 	vp = TAILQ_FIRST(&mp->mnt_activevnodelist);
5307 	if (vp == NULL) {
5308 		mtx_unlock(&vnode_free_list_mtx);
5309 		mnt_vnode_markerfree_active(mvp, mp);
5310 		return (NULL);
5311 	}
5312 	TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5313 	return (mnt_vnode_next_active(mvp, mp));
5314 }
5315 
5316 void
5317 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5318 {
5319 
5320 	if (*mvp == NULL)
5321 		return;
5322 
5323 	mtx_lock(&vnode_free_list_mtx);
5324 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5325 	mtx_unlock(&vnode_free_list_mtx);
5326 	mnt_vnode_markerfree_active(mvp, mp);
5327 }
5328