xref: /freebsd/sys/kern/vfs_subr.c (revision c6a33c8e88c5684876e670c8189d03ad25108d8a)
1 /*-
2  * Copyright (c) 1989, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  * (c) UNIX System Laboratories, Inc.
5  * All or some portions of this file are derived from material licensed
6  * to the University of California by American Telephone and Telegraph
7  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8  * the permission of UNIX System Laboratories, Inc.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 4. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	@(#)vfs_subr.c	8.31 (Berkeley) 5/26/95
35  */
36 
37 /*
38  * External virtual filesystem routines
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include "opt_compat.h"
45 #include "opt_ddb.h"
46 #include "opt_watchdog.h"
47 
48 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/bio.h>
51 #include <sys/buf.h>
52 #include <sys/condvar.h>
53 #include <sys/conf.h>
54 #include <sys/dirent.h>
55 #include <sys/event.h>
56 #include <sys/eventhandler.h>
57 #include <sys/extattr.h>
58 #include <sys/file.h>
59 #include <sys/fcntl.h>
60 #include <sys/jail.h>
61 #include <sys/kdb.h>
62 #include <sys/kernel.h>
63 #include <sys/kthread.h>
64 #include <sys/lockf.h>
65 #include <sys/malloc.h>
66 #include <sys/mount.h>
67 #include <sys/namei.h>
68 #include <sys/pctrie.h>
69 #include <sys/priv.h>
70 #include <sys/reboot.h>
71 #include <sys/refcount.h>
72 #include <sys/rwlock.h>
73 #include <sys/sched.h>
74 #include <sys/sleepqueue.h>
75 #include <sys/smp.h>
76 #include <sys/stat.h>
77 #include <sys/sysctl.h>
78 #include <sys/syslog.h>
79 #include <sys/vmmeter.h>
80 #include <sys/vnode.h>
81 #include <sys/watchdog.h>
82 
83 #include <machine/stdarg.h>
84 
85 #include <security/mac/mac_framework.h>
86 
87 #include <vm/vm.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_extern.h>
90 #include <vm/pmap.h>
91 #include <vm/vm_map.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_kern.h>
94 #include <vm/uma.h>
95 
96 #ifdef DDB
97 #include <ddb/ddb.h>
98 #endif
99 
100 static void	delmntque(struct vnode *vp);
101 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
102 		    int slpflag, int slptimeo);
103 static void	syncer_shutdown(void *arg, int howto);
104 static int	vtryrecycle(struct vnode *vp);
105 static void	v_init_counters(struct vnode *);
106 static void	v_incr_usecount(struct vnode *);
107 static void	v_incr_devcount(struct vnode *);
108 static void	v_decr_devcount(struct vnode *);
109 static void	vnlru_free(int);
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 static void
410 vntblinit(void *dummy __unused)
411 {
412 	u_int i;
413 	int physvnodes, virtvnodes;
414 
415 	/*
416 	 * Desiredvnodes is a function of the physical memory size and the
417 	 * kernel's heap size.  Generally speaking, it scales with the
418 	 * physical memory size.  The ratio of desiredvnodes to the physical
419 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
420 	 * Thereafter, the
421 	 * marginal ratio of desiredvnodes to the physical memory size is
422 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
423 	 * size.  The memory required by desiredvnodes vnodes and vm objects
424 	 * must not exceed 1/7th of the kernel's heap size.
425 	 */
426 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
427 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
428 	virtvnodes = vm_kmem_size / (7 * (sizeof(struct vm_object) +
429 	    sizeof(struct vnode)));
430 	desiredvnodes = min(physvnodes, virtvnodes);
431 	if (desiredvnodes > MAXVNODES_MAX) {
432 		if (bootverbose)
433 			printf("Reducing kern.maxvnodes %d -> %d\n",
434 			    desiredvnodes, MAXVNODES_MAX);
435 		desiredvnodes = MAXVNODES_MAX;
436 	}
437 	wantfreevnodes = desiredvnodes / 4;
438 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
439 	TAILQ_INIT(&vnode_free_list);
440 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
441 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
442 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
443 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
444 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
445 	/*
446 	 * Preallocate enough nodes to support one-per buf so that
447 	 * we can not fail an insert.  reassignbuf() callers can not
448 	 * tolerate the insertion failure.
449 	 */
450 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
451 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
452 	    UMA_ZONE_NOFREE | UMA_ZONE_VM);
453 	uma_prealloc(buf_trie_zone, nbuf);
454 	/*
455 	 * Initialize the filesystem syncer.
456 	 */
457 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
458 	    &syncer_mask);
459 	syncer_maxdelay = syncer_mask + 1;
460 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
461 	cv_init(&sync_wakeup, "syncer");
462 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
463 		vnsz2log++;
464 	vnsz2log--;
465 }
466 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
467 
468 
469 /*
470  * Mark a mount point as busy. Used to synchronize access and to delay
471  * unmounting. Eventually, mountlist_mtx is not released on failure.
472  *
473  * vfs_busy() is a custom lock, it can block the caller.
474  * vfs_busy() only sleeps if the unmount is active on the mount point.
475  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
476  * vnode belonging to mp.
477  *
478  * Lookup uses vfs_busy() to traverse mount points.
479  * root fs			var fs
480  * / vnode lock		A	/ vnode lock (/var)		D
481  * /var vnode lock	B	/log vnode lock(/var/log)	E
482  * vfs_busy lock	C	vfs_busy lock			F
483  *
484  * Within each file system, the lock order is C->A->B and F->D->E.
485  *
486  * When traversing across mounts, the system follows that lock order:
487  *
488  *        C->A->B
489  *              |
490  *              +->F->D->E
491  *
492  * The lookup() process for namei("/var") illustrates the process:
493  *  VOP_LOOKUP() obtains B while A is held
494  *  vfs_busy() obtains a shared lock on F while A and B are held
495  *  vput() releases lock on B
496  *  vput() releases lock on A
497  *  VFS_ROOT() obtains lock on D while shared lock on F is held
498  *  vfs_unbusy() releases shared lock on F
499  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
500  *    Attempt to lock A (instead of vp_crossmp) while D is held would
501  *    violate the global order, causing deadlocks.
502  *
503  * dounmount() locks B while F is drained.
504  */
505 int
506 vfs_busy(struct mount *mp, int flags)
507 {
508 
509 	MPASS((flags & ~MBF_MASK) == 0);
510 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
511 
512 	MNT_ILOCK(mp);
513 	MNT_REF(mp);
514 	/*
515 	 * If mount point is currenly being unmounted, sleep until the
516 	 * mount point fate is decided.  If thread doing the unmounting fails,
517 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
518 	 * that this mount point has survived the unmount attempt and vfs_busy
519 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
520 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
521 	 * about to be really destroyed.  vfs_busy needs to release its
522 	 * reference on the mount point in this case and return with ENOENT,
523 	 * telling the caller that mount mount it tried to busy is no longer
524 	 * valid.
525 	 */
526 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
527 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
528 			MNT_REL(mp);
529 			MNT_IUNLOCK(mp);
530 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
531 			    __func__);
532 			return (ENOENT);
533 		}
534 		if (flags & MBF_MNTLSTLOCK)
535 			mtx_unlock(&mountlist_mtx);
536 		mp->mnt_kern_flag |= MNTK_MWAIT;
537 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
538 		if (flags & MBF_MNTLSTLOCK)
539 			mtx_lock(&mountlist_mtx);
540 		MNT_ILOCK(mp);
541 	}
542 	if (flags & MBF_MNTLSTLOCK)
543 		mtx_unlock(&mountlist_mtx);
544 	mp->mnt_lockref++;
545 	MNT_IUNLOCK(mp);
546 	return (0);
547 }
548 
549 /*
550  * Free a busy filesystem.
551  */
552 void
553 vfs_unbusy(struct mount *mp)
554 {
555 
556 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
557 	MNT_ILOCK(mp);
558 	MNT_REL(mp);
559 	KASSERT(mp->mnt_lockref > 0, ("negative mnt_lockref"));
560 	mp->mnt_lockref--;
561 	if (mp->mnt_lockref == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
562 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
563 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
564 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
565 		wakeup(&mp->mnt_lockref);
566 	}
567 	MNT_IUNLOCK(mp);
568 }
569 
570 /*
571  * Lookup a mount point by filesystem identifier.
572  */
573 struct mount *
574 vfs_getvfs(fsid_t *fsid)
575 {
576 	struct mount *mp;
577 
578 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
579 	mtx_lock(&mountlist_mtx);
580 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
581 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
582 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
583 			vfs_ref(mp);
584 			mtx_unlock(&mountlist_mtx);
585 			return (mp);
586 		}
587 	}
588 	mtx_unlock(&mountlist_mtx);
589 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
590 	return ((struct mount *) 0);
591 }
592 
593 /*
594  * Lookup a mount point by filesystem identifier, busying it before
595  * returning.
596  *
597  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
598  * cache for popular filesystem identifiers.  The cache is lockess, using
599  * the fact that struct mount's are never freed.  In worst case we may
600  * get pointer to unmounted or even different filesystem, so we have to
601  * check what we got, and go slow way if so.
602  */
603 struct mount *
604 vfs_busyfs(fsid_t *fsid)
605 {
606 #define	FSID_CACHE_SIZE	256
607 	typedef struct mount * volatile vmp_t;
608 	static vmp_t cache[FSID_CACHE_SIZE];
609 	struct mount *mp;
610 	int error;
611 	uint32_t hash;
612 
613 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
614 	hash = fsid->val[0] ^ fsid->val[1];
615 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
616 	mp = cache[hash];
617 	if (mp == NULL ||
618 	    mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
619 	    mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
620 		goto slow;
621 	if (vfs_busy(mp, 0) != 0) {
622 		cache[hash] = NULL;
623 		goto slow;
624 	}
625 	if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
626 	    mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
627 		return (mp);
628 	else
629 	    vfs_unbusy(mp);
630 
631 slow:
632 	mtx_lock(&mountlist_mtx);
633 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
634 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
635 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
636 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
637 			if (error) {
638 				cache[hash] = NULL;
639 				mtx_unlock(&mountlist_mtx);
640 				return (NULL);
641 			}
642 			cache[hash] = mp;
643 			return (mp);
644 		}
645 	}
646 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
647 	mtx_unlock(&mountlist_mtx);
648 	return ((struct mount *) 0);
649 }
650 
651 /*
652  * Check if a user can access privileged mount options.
653  */
654 int
655 vfs_suser(struct mount *mp, struct thread *td)
656 {
657 	int error;
658 
659 	/*
660 	 * If the thread is jailed, but this is not a jail-friendly file
661 	 * system, deny immediately.
662 	 */
663 	if (!(mp->mnt_vfc->vfc_flags & VFCF_JAIL) && jailed(td->td_ucred))
664 		return (EPERM);
665 
666 	/*
667 	 * If the file system was mounted outside the jail of the calling
668 	 * thread, deny immediately.
669 	 */
670 	if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
671 		return (EPERM);
672 
673 	/*
674 	 * If file system supports delegated administration, we don't check
675 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
676 	 * by the file system itself.
677 	 * If this is not the user that did original mount, we check for
678 	 * the PRIV_VFS_MOUNT_OWNER privilege.
679 	 */
680 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
681 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
682 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
683 			return (error);
684 	}
685 	return (0);
686 }
687 
688 /*
689  * Get a new unique fsid.  Try to make its val[0] unique, since this value
690  * will be used to create fake device numbers for stat().  Also try (but
691  * not so hard) make its val[0] unique mod 2^16, since some emulators only
692  * support 16-bit device numbers.  We end up with unique val[0]'s for the
693  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
694  *
695  * Keep in mind that several mounts may be running in parallel.  Starting
696  * the search one past where the previous search terminated is both a
697  * micro-optimization and a defense against returning the same fsid to
698  * different mounts.
699  */
700 void
701 vfs_getnewfsid(struct mount *mp)
702 {
703 	static uint16_t mntid_base;
704 	struct mount *nmp;
705 	fsid_t tfsid;
706 	int mtype;
707 
708 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
709 	mtx_lock(&mntid_mtx);
710 	mtype = mp->mnt_vfc->vfc_typenum;
711 	tfsid.val[1] = mtype;
712 	mtype = (mtype & 0xFF) << 24;
713 	for (;;) {
714 		tfsid.val[0] = makedev(255,
715 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
716 		mntid_base++;
717 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
718 			break;
719 		vfs_rel(nmp);
720 	}
721 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
722 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
723 	mtx_unlock(&mntid_mtx);
724 }
725 
726 /*
727  * Knob to control the precision of file timestamps:
728  *
729  *   0 = seconds only; nanoseconds zeroed.
730  *   1 = seconds and nanoseconds, accurate within 1/HZ.
731  *   2 = seconds and nanoseconds, truncated to microseconds.
732  * >=3 = seconds and nanoseconds, maximum precision.
733  */
734 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
735 
736 static int timestamp_precision = TSP_USEC;
737 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
738     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
739     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to ms, "
740     "3+: sec + ns (max. precision))");
741 
742 /*
743  * Get a current timestamp.
744  */
745 void
746 vfs_timestamp(struct timespec *tsp)
747 {
748 	struct timeval tv;
749 
750 	switch (timestamp_precision) {
751 	case TSP_SEC:
752 		tsp->tv_sec = time_second;
753 		tsp->tv_nsec = 0;
754 		break;
755 	case TSP_HZ:
756 		getnanotime(tsp);
757 		break;
758 	case TSP_USEC:
759 		microtime(&tv);
760 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
761 		break;
762 	case TSP_NSEC:
763 	default:
764 		nanotime(tsp);
765 		break;
766 	}
767 }
768 
769 /*
770  * Set vnode attributes to VNOVAL
771  */
772 void
773 vattr_null(struct vattr *vap)
774 {
775 
776 	vap->va_type = VNON;
777 	vap->va_size = VNOVAL;
778 	vap->va_bytes = VNOVAL;
779 	vap->va_mode = VNOVAL;
780 	vap->va_nlink = VNOVAL;
781 	vap->va_uid = VNOVAL;
782 	vap->va_gid = VNOVAL;
783 	vap->va_fsid = VNOVAL;
784 	vap->va_fileid = VNOVAL;
785 	vap->va_blocksize = VNOVAL;
786 	vap->va_rdev = VNOVAL;
787 	vap->va_atime.tv_sec = VNOVAL;
788 	vap->va_atime.tv_nsec = VNOVAL;
789 	vap->va_mtime.tv_sec = VNOVAL;
790 	vap->va_mtime.tv_nsec = VNOVAL;
791 	vap->va_ctime.tv_sec = VNOVAL;
792 	vap->va_ctime.tv_nsec = VNOVAL;
793 	vap->va_birthtime.tv_sec = VNOVAL;
794 	vap->va_birthtime.tv_nsec = VNOVAL;
795 	vap->va_flags = VNOVAL;
796 	vap->va_gen = VNOVAL;
797 	vap->va_vaflags = 0;
798 }
799 
800 /*
801  * This routine is called when we have too many vnodes.  It attempts
802  * to free <count> vnodes and will potentially free vnodes that still
803  * have VM backing store (VM backing store is typically the cause
804  * of a vnode blowout so we want to do this).  Therefore, this operation
805  * is not considered cheap.
806  *
807  * A number of conditions may prevent a vnode from being reclaimed.
808  * the buffer cache may have references on the vnode, a directory
809  * vnode may still have references due to the namei cache representing
810  * underlying files, or the vnode may be in active use.   It is not
811  * desireable to reuse such vnodes.  These conditions may cause the
812  * number of vnodes to reach some minimum value regardless of what
813  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
814  */
815 static int
816 vlrureclaim(struct mount *mp, int reclaim_nc_src, int trigger)
817 {
818 	struct vnode *vp;
819 	int count, done, target;
820 
821 	done = 0;
822 	vn_start_write(NULL, &mp, V_WAIT);
823 	MNT_ILOCK(mp);
824 	count = mp->mnt_nvnodelistsize;
825 	target = count * (int64_t)gapvnodes / imax(desiredvnodes, 1);
826 	target = target / 10 + 1;
827 	while (count != 0 && done < target) {
828 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
829 		while (vp != NULL && vp->v_type == VMARKER)
830 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
831 		if (vp == NULL)
832 			break;
833 		/*
834 		 * XXX LRU is completely broken for non-free vnodes.  First
835 		 * by calling here in mountpoint order, then by moving
836 		 * unselected vnodes to the end here, and most grossly by
837 		 * removing the vlruvp() function that was supposed to
838 		 * maintain the order.  (This function was born broken
839 		 * since syncer problems prevented it doing anything.)  The
840 		 * order is closer to LRC (C = Created).
841 		 *
842 		 * LRU reclaiming of vnodes seems to have last worked in
843 		 * FreeBSD-3 where LRU wasn't mentioned under any spelling.
844 		 * Then there was no hold count, and inactive vnodes were
845 		 * simply put on the free list in LRU order.  The separate
846 		 * lists also break LRU.  We prefer to reclaim from the
847 		 * free list for technical reasons.  This tends to thrash
848 		 * the free list to keep very unrecently used held vnodes.
849 		 * The problem is mitigated by keeping the free list large.
850 		 */
851 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
852 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
853 		--count;
854 		if (!VI_TRYLOCK(vp))
855 			goto next_iter;
856 		/*
857 		 * If it's been deconstructed already, it's still
858 		 * referenced, or it exceeds the trigger, skip it.
859 		 * Also skip free vnodes.  We are trying to make space
860 		 * to expand the free list, not reduce it.
861 		 */
862 		if (vp->v_usecount ||
863 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
864 		    ((vp->v_iflag & VI_FREE) != 0) ||
865 		    (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
866 		    vp->v_object->resident_page_count > trigger)) {
867 			VI_UNLOCK(vp);
868 			goto next_iter;
869 		}
870 		MNT_IUNLOCK(mp);
871 		vholdl(vp);
872 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) {
873 			vdrop(vp);
874 			goto next_iter_mntunlocked;
875 		}
876 		VI_LOCK(vp);
877 		/*
878 		 * v_usecount may have been bumped after VOP_LOCK() dropped
879 		 * the vnode interlock and before it was locked again.
880 		 *
881 		 * It is not necessary to recheck VI_DOOMED because it can
882 		 * only be set by another thread that holds both the vnode
883 		 * lock and vnode interlock.  If another thread has the
884 		 * vnode lock before we get to VOP_LOCK() and obtains the
885 		 * vnode interlock after VOP_LOCK() drops the vnode
886 		 * interlock, the other thread will be unable to drop the
887 		 * vnode lock before our VOP_LOCK() call fails.
888 		 */
889 		if (vp->v_usecount ||
890 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
891 		    (vp->v_iflag & VI_FREE) != 0 ||
892 		    (vp->v_object != NULL &&
893 		    vp->v_object->resident_page_count > trigger)) {
894 			VOP_UNLOCK(vp, LK_INTERLOCK);
895 			vdrop(vp);
896 			goto next_iter_mntunlocked;
897 		}
898 		KASSERT((vp->v_iflag & VI_DOOMED) == 0,
899 		    ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
900 		atomic_add_long(&recycles_count, 1);
901 		vgonel(vp);
902 		VOP_UNLOCK(vp, 0);
903 		vdropl(vp);
904 		done++;
905 next_iter_mntunlocked:
906 		if (!should_yield())
907 			goto relock_mnt;
908 		goto yield;
909 next_iter:
910 		if (!should_yield())
911 			continue;
912 		MNT_IUNLOCK(mp);
913 yield:
914 		kern_yield(PRI_USER);
915 relock_mnt:
916 		MNT_ILOCK(mp);
917 	}
918 	MNT_IUNLOCK(mp);
919 	vn_finished_write(mp);
920 	return done;
921 }
922 
923 /*
924  * Attempt to reduce the free list by the requested amount.
925  */
926 static void
927 vnlru_free(int count)
928 {
929 	struct vnode *vp;
930 
931 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
932 	for (; count > 0; count--) {
933 		vp = TAILQ_FIRST(&vnode_free_list);
934 		/*
935 		 * The list can be modified while the free_list_mtx
936 		 * has been dropped and vp could be NULL here.
937 		 */
938 		if (!vp)
939 			break;
940 		VNASSERT(vp->v_op != NULL, vp,
941 		    ("vnlru_free: vnode already reclaimed."));
942 		KASSERT((vp->v_iflag & VI_FREE) != 0,
943 		    ("Removing vnode not on freelist"));
944 		KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
945 		    ("Mangling active vnode"));
946 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
947 		/*
948 		 * Don't recycle if we can't get the interlock.
949 		 */
950 		if (!VI_TRYLOCK(vp)) {
951 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist);
952 			continue;
953 		}
954 		VNASSERT((vp->v_iflag & VI_FREE) != 0 && vp->v_holdcnt == 0,
955 		    vp, ("vp inconsistent on freelist"));
956 
957 		/*
958 		 * The clear of VI_FREE prevents activation of the
959 		 * vnode.  There is no sense in putting the vnode on
960 		 * the mount point active list, only to remove it
961 		 * later during recycling.  Inline the relevant part
962 		 * of vholdl(), to avoid triggering assertions or
963 		 * activating.
964 		 */
965 		freevnodes--;
966 		vp->v_iflag &= ~VI_FREE;
967 		refcount_acquire(&vp->v_holdcnt);
968 
969 		mtx_unlock(&vnode_free_list_mtx);
970 		VI_UNLOCK(vp);
971 		vtryrecycle(vp);
972 		/*
973 		 * If the recycled succeeded this vdrop will actually free
974 		 * the vnode.  If not it will simply place it back on
975 		 * the free list.
976 		 */
977 		vdrop(vp);
978 		mtx_lock(&vnode_free_list_mtx);
979 	}
980 }
981 
982 /* XXX some names and initialization are bad for limits and watermarks. */
983 static int
984 vspace(void)
985 {
986 	int space;
987 
988 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
989 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
990 	vlowat = vhiwat / 2;
991 	if (numvnodes > desiredvnodes)
992 		return (0);
993 	space = desiredvnodes - numvnodes;
994 	if (freevnodes > wantfreevnodes)
995 		space += freevnodes - wantfreevnodes;
996 	return (space);
997 }
998 
999 /*
1000  * Attempt to recycle vnodes in a context that is always safe to block.
1001  * Calling vlrurecycle() from the bowels of filesystem code has some
1002  * interesting deadlock problems.
1003  */
1004 static struct proc *vnlruproc;
1005 static int vnlruproc_sig;
1006 
1007 static void
1008 vnlru_proc(void)
1009 {
1010 	struct mount *mp, *nmp;
1011 	unsigned long ofreevnodes, onumvnodes;
1012 	int done, force, reclaim_nc_src, trigger, usevnodes;
1013 
1014 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1015 	    SHUTDOWN_PRI_FIRST);
1016 
1017 	force = 0;
1018 	for (;;) {
1019 		kproc_suspend_check(vnlruproc);
1020 		mtx_lock(&vnode_free_list_mtx);
1021 		/*
1022 		 * If numvnodes is too large (due to desiredvnodes being
1023 		 * adjusted using its sysctl, or emergency growth), first
1024 		 * try to reduce it by discarding from the free list.
1025 		 */
1026 		if (numvnodes > desiredvnodes && freevnodes > 0)
1027 			vnlru_free(ulmin(numvnodes - desiredvnodes,
1028 			    freevnodes));
1029 		/*
1030 		 * Sleep if the vnode cache is in a good state.  This is
1031 		 * when it is not over-full and has space for about a 4%
1032 		 * or 9% expansion (by growing its size or inexcessively
1033 		 * reducing its free list).  Otherwise, try to reclaim
1034 		 * space for a 10% expansion.
1035 		 */
1036 		if (vstir && force == 0) {
1037 			force = 1;
1038 			vstir = 0;
1039 		}
1040 		if (vspace() >= vlowat && force == 0) {
1041 			vnlruproc_sig = 0;
1042 			wakeup(&vnlruproc_sig);
1043 			msleep(vnlruproc, &vnode_free_list_mtx,
1044 			    PVFS|PDROP, "vlruwt", hz);
1045 			continue;
1046 		}
1047 		mtx_unlock(&vnode_free_list_mtx);
1048 		done = 0;
1049 		ofreevnodes = freevnodes;
1050 		onumvnodes = numvnodes;
1051 		/*
1052 		 * Calculate parameters for recycling.  These are the same
1053 		 * throughout the loop to give some semblance of fairness.
1054 		 * The trigger point is to avoid recycling vnodes with lots
1055 		 * of resident pages.  We aren't trying to free memory; we
1056 		 * are trying to recycle or at least free vnodes.
1057 		 */
1058 		if (numvnodes <= desiredvnodes)
1059 			usevnodes = numvnodes - freevnodes;
1060 		else
1061 			usevnodes = numvnodes;
1062 		if (usevnodes <= 0)
1063 			usevnodes = 1;
1064 		/*
1065 		 * The trigger value is is chosen to give a conservatively
1066 		 * large value to ensure that it alone doesn't prevent
1067 		 * making progress.  The value can easily be so large that
1068 		 * it is effectively infinite in some congested and
1069 		 * misconfigured cases, and this is necessary.  Normally
1070 		 * it is about 8 to 100 (pages), which is quite large.
1071 		 */
1072 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1073 		if (force < 2)
1074 			trigger = vsmalltrigger;
1075 		reclaim_nc_src = force >= 3;
1076 		mtx_lock(&mountlist_mtx);
1077 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1078 			if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) {
1079 				nmp = TAILQ_NEXT(mp, mnt_list);
1080 				continue;
1081 			}
1082 			done += vlrureclaim(mp, reclaim_nc_src, trigger);
1083 			mtx_lock(&mountlist_mtx);
1084 			nmp = TAILQ_NEXT(mp, mnt_list);
1085 			vfs_unbusy(mp);
1086 		}
1087 		mtx_unlock(&mountlist_mtx);
1088 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1089 			uma_reclaim();
1090 		if (done == 0) {
1091 			if (force == 0 || force == 1) {
1092 				force = 2;
1093 				continue;
1094 			}
1095 			if (force == 2) {
1096 				force = 3;
1097 				continue;
1098 			}
1099 			force = 0;
1100 			vnlru_nowhere++;
1101 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1102 		} else
1103 			kern_yield(PRI_USER);
1104 		/*
1105 		 * After becoming active to expand above low water, keep
1106 		 * active until above high water.
1107 		 */
1108 		force = vspace() < vhiwat;
1109 	}
1110 }
1111 
1112 static struct kproc_desc vnlru_kp = {
1113 	"vnlru",
1114 	vnlru_proc,
1115 	&vnlruproc
1116 };
1117 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1118     &vnlru_kp);
1119 
1120 /*
1121  * Routines having to do with the management of the vnode table.
1122  */
1123 
1124 /*
1125  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1126  * before we actually vgone().  This function must be called with the vnode
1127  * held to prevent the vnode from being returned to the free list midway
1128  * through vgone().
1129  */
1130 static int
1131 vtryrecycle(struct vnode *vp)
1132 {
1133 	struct mount *vnmp;
1134 
1135 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1136 	VNASSERT(vp->v_holdcnt, vp,
1137 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1138 	/*
1139 	 * This vnode may found and locked via some other list, if so we
1140 	 * can't recycle it yet.
1141 	 */
1142 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1143 		CTR2(KTR_VFS,
1144 		    "%s: impossible to recycle, vp %p lock is already held",
1145 		    __func__, vp);
1146 		return (EWOULDBLOCK);
1147 	}
1148 	/*
1149 	 * Don't recycle if its filesystem is being suspended.
1150 	 */
1151 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1152 		VOP_UNLOCK(vp, 0);
1153 		CTR2(KTR_VFS,
1154 		    "%s: impossible to recycle, cannot start the write for %p",
1155 		    __func__, vp);
1156 		return (EBUSY);
1157 	}
1158 	/*
1159 	 * If we got this far, we need to acquire the interlock and see if
1160 	 * anyone picked up this vnode from another list.  If not, we will
1161 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1162 	 * will skip over it.
1163 	 */
1164 	VI_LOCK(vp);
1165 	if (vp->v_usecount) {
1166 		VOP_UNLOCK(vp, LK_INTERLOCK);
1167 		vn_finished_write(vnmp);
1168 		CTR2(KTR_VFS,
1169 		    "%s: impossible to recycle, %p is already referenced",
1170 		    __func__, vp);
1171 		return (EBUSY);
1172 	}
1173 	if ((vp->v_iflag & VI_DOOMED) == 0) {
1174 		atomic_add_long(&recycles_count, 1);
1175 		vgonel(vp);
1176 	}
1177 	VOP_UNLOCK(vp, LK_INTERLOCK);
1178 	vn_finished_write(vnmp);
1179 	return (0);
1180 }
1181 
1182 static void
1183 vcheckspace(void)
1184 {
1185 
1186 	if (vspace() < vlowat && vnlruproc_sig == 0) {
1187 		vnlruproc_sig = 1;
1188 		wakeup(vnlruproc);
1189 	}
1190 }
1191 
1192 /*
1193  * Wait if necessary for space for a new vnode.
1194  */
1195 static int
1196 getnewvnode_wait(int suspended)
1197 {
1198 
1199 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1200 	if (numvnodes >= desiredvnodes) {
1201 		if (suspended) {
1202 			/*
1203 			 * The file system is being suspended.  We cannot
1204 			 * risk a deadlock here, so allow allocation of
1205 			 * another vnode even if this would give too many.
1206 			 */
1207 			return (0);
1208 		}
1209 		if (vnlruproc_sig == 0) {
1210 			vnlruproc_sig = 1;	/* avoid unnecessary wakeups */
1211 			wakeup(vnlruproc);
1212 		}
1213 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
1214 		    "vlruwk", hz);
1215 	}
1216 	/* Post-adjust like the pre-adjust in getnewvnode(). */
1217 	if (numvnodes + 1 > desiredvnodes && freevnodes > 1)
1218 		vnlru_free(1);
1219 	return (numvnodes >= desiredvnodes ? ENFILE : 0);
1220 }
1221 
1222 /*
1223  * This hack is fragile, and probably not needed any more now that the
1224  * watermark handling works.
1225  */
1226 void
1227 getnewvnode_reserve(u_int count)
1228 {
1229 	struct thread *td;
1230 
1231 	/* Pre-adjust like the pre-adjust in getnewvnode(), with any count. */
1232 	/* XXX no longer so quick, but this part is not racy. */
1233 	mtx_lock(&vnode_free_list_mtx);
1234 	if (numvnodes + count > desiredvnodes && freevnodes > wantfreevnodes)
1235 		vnlru_free(ulmin(numvnodes + count - desiredvnodes,
1236 		    freevnodes - wantfreevnodes));
1237 	mtx_unlock(&vnode_free_list_mtx);
1238 
1239 	td = curthread;
1240 	/* First try to be quick and racy. */
1241 	if (atomic_fetchadd_long(&numvnodes, count) + count <= desiredvnodes) {
1242 		td->td_vp_reserv += count;
1243 		vcheckspace();	/* XXX no longer so quick, but more racy */
1244 		return;
1245 	} else
1246 		atomic_subtract_long(&numvnodes, count);
1247 
1248 	mtx_lock(&vnode_free_list_mtx);
1249 	while (count > 0) {
1250 		if (getnewvnode_wait(0) == 0) {
1251 			count--;
1252 			td->td_vp_reserv++;
1253 			atomic_add_long(&numvnodes, 1);
1254 		}
1255 	}
1256 	vcheckspace();
1257 	mtx_unlock(&vnode_free_list_mtx);
1258 }
1259 
1260 /*
1261  * This hack is fragile, especially if desiredvnodes or wantvnodes are
1262  * misconfgured or changed significantly.  Reducing desiredvnodes below
1263  * the reserved amount should cause bizarre behaviour like reducing it
1264  * below the number of active vnodes -- the system will try to reduce
1265  * numvnodes to match, but should fail, so the subtraction below should
1266  * not overflow.
1267  */
1268 void
1269 getnewvnode_drop_reserve(void)
1270 {
1271 	struct thread *td;
1272 
1273 	td = curthread;
1274 	atomic_subtract_long(&numvnodes, td->td_vp_reserv);
1275 	td->td_vp_reserv = 0;
1276 }
1277 
1278 /*
1279  * Return the next vnode from the free list.
1280  */
1281 int
1282 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1283     struct vnode **vpp)
1284 {
1285 	struct vnode *vp;
1286 	struct thread *td;
1287 	struct lock_object *lo;
1288 	static int cyclecount;
1289 	int error;
1290 
1291 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1292 	vp = NULL;
1293 	td = curthread;
1294 	if (td->td_vp_reserv > 0) {
1295 		td->td_vp_reserv -= 1;
1296 		goto alloc;
1297 	}
1298 	mtx_lock(&vnode_free_list_mtx);
1299 	if (numvnodes < desiredvnodes)
1300 		cyclecount = 0;
1301 	else if (cyclecount++ >= freevnodes) {
1302 		cyclecount = 0;
1303 		vstir = 1;
1304 	}
1305 	/*
1306 	 * Grow the vnode cache if it will not be above its target max
1307 	 * after growing.  Otherwise, if the free list is nonempty, try
1308 	 * to reclaim 1 item from it before growing the cache (possibly
1309 	 * above its target max if the reclamation failed or is delayed).
1310 	 * Otherwise, wait for some space.  In all cases, schedule
1311 	 * vnlru_proc() if we are getting short of space.  The watermarks
1312 	 * should be chosen so that we never wait or even reclaim from
1313 	 * the free list to below its target minimum.
1314 	 */
1315 	if (numvnodes + 1 <= desiredvnodes)
1316 		;
1317 	else if (freevnodes > 0)
1318 		vnlru_free(1);
1319 	else {
1320 		error = getnewvnode_wait(mp != NULL && (mp->mnt_kern_flag &
1321 		    MNTK_SUSPEND));
1322 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
1323 		if (error != 0) {
1324 			mtx_unlock(&vnode_free_list_mtx);
1325 			return (error);
1326 		}
1327 #endif
1328 	}
1329 	vcheckspace();
1330 	atomic_add_long(&numvnodes, 1);
1331 	mtx_unlock(&vnode_free_list_mtx);
1332 alloc:
1333 	atomic_add_long(&vnodes_created, 1);
1334 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK);
1335 	/*
1336 	 * Locks are given the generic name "vnode" when created.
1337 	 * Follow the historic practice of using the filesystem
1338 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1339 	 *
1340 	 * Locks live in a witness group keyed on their name. Thus,
1341 	 * when a lock is renamed, it must also move from the witness
1342 	 * group of its old name to the witness group of its new name.
1343 	 *
1344 	 * The change only needs to be made when the vnode moves
1345 	 * from one filesystem type to another. We ensure that each
1346 	 * filesystem use a single static name pointer for its tag so
1347 	 * that we can compare pointers rather than doing a strcmp().
1348 	 */
1349 	lo = &vp->v_vnlock->lock_object;
1350 	if (lo->lo_name != tag) {
1351 		lo->lo_name = tag;
1352 		WITNESS_DESTROY(lo);
1353 		WITNESS_INIT(lo, tag);
1354 	}
1355 	/*
1356 	 * By default, don't allow shared locks unless filesystems opt-in.
1357 	 */
1358 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1359 	/*
1360 	 * Finalize various vnode identity bits.
1361 	 */
1362 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1363 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1364 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1365 	vp->v_type = VNON;
1366 	vp->v_tag = tag;
1367 	vp->v_op = vops;
1368 	v_init_counters(vp);
1369 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1370 #ifdef MAC
1371 	mac_vnode_init(vp);
1372 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1373 		mac_vnode_associate_singlelabel(mp, vp);
1374 	else if (mp == NULL && vops != &dead_vnodeops)
1375 		printf("NULL mp in getnewvnode()\n");
1376 #endif
1377 	if (mp != NULL) {
1378 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1379 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1380 			vp->v_vflag |= VV_NOKNOTE;
1381 	}
1382 
1383 	/*
1384 	 * For the filesystems which do not use vfs_hash_insert(),
1385 	 * still initialize v_hash to have vfs_hash_index() useful.
1386 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1387 	 * its own hashing.
1388 	 */
1389 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1390 
1391 	*vpp = vp;
1392 	return (0);
1393 }
1394 
1395 /*
1396  * Delete from old mount point vnode list, if on one.
1397  */
1398 static void
1399 delmntque(struct vnode *vp)
1400 {
1401 	struct mount *mp;
1402 	int active;
1403 
1404 	mp = vp->v_mount;
1405 	if (mp == NULL)
1406 		return;
1407 	MNT_ILOCK(mp);
1408 	VI_LOCK(vp);
1409 	KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize,
1410 	    ("Active vnode list size %d > Vnode list size %d",
1411 	     mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize));
1412 	active = vp->v_iflag & VI_ACTIVE;
1413 	vp->v_iflag &= ~VI_ACTIVE;
1414 	if (active) {
1415 		mtx_lock(&vnode_free_list_mtx);
1416 		TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist);
1417 		mp->mnt_activevnodelistsize--;
1418 		mtx_unlock(&vnode_free_list_mtx);
1419 	}
1420 	vp->v_mount = NULL;
1421 	VI_UNLOCK(vp);
1422 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1423 		("bad mount point vnode list size"));
1424 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1425 	mp->mnt_nvnodelistsize--;
1426 	MNT_REL(mp);
1427 	MNT_IUNLOCK(mp);
1428 }
1429 
1430 static void
1431 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1432 {
1433 
1434 	vp->v_data = NULL;
1435 	vp->v_op = &dead_vnodeops;
1436 	vgone(vp);
1437 	vput(vp);
1438 }
1439 
1440 /*
1441  * Insert into list of vnodes for the new mount point, if available.
1442  */
1443 int
1444 insmntque1(struct vnode *vp, struct mount *mp,
1445 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1446 {
1447 
1448 	KASSERT(vp->v_mount == NULL,
1449 		("insmntque: vnode already on per mount vnode list"));
1450 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1451 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1452 
1453 	/*
1454 	 * We acquire the vnode interlock early to ensure that the
1455 	 * vnode cannot be recycled by another process releasing a
1456 	 * holdcnt on it before we get it on both the vnode list
1457 	 * and the active vnode list. The mount mutex protects only
1458 	 * manipulation of the vnode list and the vnode freelist
1459 	 * mutex protects only manipulation of the active vnode list.
1460 	 * Hence the need to hold the vnode interlock throughout.
1461 	 */
1462 	MNT_ILOCK(mp);
1463 	VI_LOCK(vp);
1464 	if (((mp->mnt_kern_flag & MNTK_NOINSMNTQ) != 0 &&
1465 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1466 	    mp->mnt_nvnodelistsize == 0)) &&
1467 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1468 		VI_UNLOCK(vp);
1469 		MNT_IUNLOCK(mp);
1470 		if (dtr != NULL)
1471 			dtr(vp, dtr_arg);
1472 		return (EBUSY);
1473 	}
1474 	vp->v_mount = mp;
1475 	MNT_REF(mp);
1476 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1477 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1478 		("neg mount point vnode list size"));
1479 	mp->mnt_nvnodelistsize++;
1480 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1481 	    ("Activating already active vnode"));
1482 	vp->v_iflag |= VI_ACTIVE;
1483 	mtx_lock(&vnode_free_list_mtx);
1484 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
1485 	mp->mnt_activevnodelistsize++;
1486 	mtx_unlock(&vnode_free_list_mtx);
1487 	VI_UNLOCK(vp);
1488 	MNT_IUNLOCK(mp);
1489 	return (0);
1490 }
1491 
1492 int
1493 insmntque(struct vnode *vp, struct mount *mp)
1494 {
1495 
1496 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1497 }
1498 
1499 /*
1500  * Flush out and invalidate all buffers associated with a bufobj
1501  * Called with the underlying object locked.
1502  */
1503 int
1504 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1505 {
1506 	int error;
1507 
1508 	BO_LOCK(bo);
1509 	if (flags & V_SAVE) {
1510 		error = bufobj_wwait(bo, slpflag, slptimeo);
1511 		if (error) {
1512 			BO_UNLOCK(bo);
1513 			return (error);
1514 		}
1515 		if (bo->bo_dirty.bv_cnt > 0) {
1516 			BO_UNLOCK(bo);
1517 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1518 				return (error);
1519 			/*
1520 			 * XXX We could save a lock/unlock if this was only
1521 			 * enabled under INVARIANTS
1522 			 */
1523 			BO_LOCK(bo);
1524 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1525 				panic("vinvalbuf: dirty bufs");
1526 		}
1527 	}
1528 	/*
1529 	 * If you alter this loop please notice that interlock is dropped and
1530 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1531 	 * no race conditions occur from this.
1532 	 */
1533 	do {
1534 		error = flushbuflist(&bo->bo_clean,
1535 		    flags, bo, slpflag, slptimeo);
1536 		if (error == 0 && !(flags & V_CLEANONLY))
1537 			error = flushbuflist(&bo->bo_dirty,
1538 			    flags, bo, slpflag, slptimeo);
1539 		if (error != 0 && error != EAGAIN) {
1540 			BO_UNLOCK(bo);
1541 			return (error);
1542 		}
1543 	} while (error != 0);
1544 
1545 	/*
1546 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1547 	 * have write I/O in-progress but if there is a VM object then the
1548 	 * VM object can also have read-I/O in-progress.
1549 	 */
1550 	do {
1551 		bufobj_wwait(bo, 0, 0);
1552 		BO_UNLOCK(bo);
1553 		if (bo->bo_object != NULL) {
1554 			VM_OBJECT_WLOCK(bo->bo_object);
1555 			vm_object_pip_wait(bo->bo_object, "bovlbx");
1556 			VM_OBJECT_WUNLOCK(bo->bo_object);
1557 		}
1558 		BO_LOCK(bo);
1559 	} while (bo->bo_numoutput > 0);
1560 	BO_UNLOCK(bo);
1561 
1562 	/*
1563 	 * Destroy the copy in the VM cache, too.
1564 	 */
1565 	if (bo->bo_object != NULL &&
1566 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY)) == 0) {
1567 		VM_OBJECT_WLOCK(bo->bo_object);
1568 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1569 		    OBJPR_CLEANONLY : 0);
1570 		VM_OBJECT_WUNLOCK(bo->bo_object);
1571 	}
1572 
1573 #ifdef INVARIANTS
1574 	BO_LOCK(bo);
1575 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY)) == 0 &&
1576 	    (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0))
1577 		panic("vinvalbuf: flush failed");
1578 	BO_UNLOCK(bo);
1579 #endif
1580 	return (0);
1581 }
1582 
1583 /*
1584  * Flush out and invalidate all buffers associated with a vnode.
1585  * Called with the underlying object locked.
1586  */
1587 int
1588 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1589 {
1590 
1591 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1592 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1593 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1594 		return (0);
1595 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1596 }
1597 
1598 /*
1599  * Flush out buffers on the specified list.
1600  *
1601  */
1602 static int
1603 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1604     int slptimeo)
1605 {
1606 	struct buf *bp, *nbp;
1607 	int retval, error;
1608 	daddr_t lblkno;
1609 	b_xflags_t xflags;
1610 
1611 	ASSERT_BO_WLOCKED(bo);
1612 
1613 	retval = 0;
1614 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1615 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1616 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1617 			continue;
1618 		}
1619 		lblkno = 0;
1620 		xflags = 0;
1621 		if (nbp != NULL) {
1622 			lblkno = nbp->b_lblkno;
1623 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
1624 		}
1625 		retval = EAGAIN;
1626 		error = BUF_TIMELOCK(bp,
1627 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
1628 		    "flushbuf", slpflag, slptimeo);
1629 		if (error) {
1630 			BO_LOCK(bo);
1631 			return (error != ENOLCK ? error : EAGAIN);
1632 		}
1633 		KASSERT(bp->b_bufobj == bo,
1634 		    ("bp %p wrong b_bufobj %p should be %p",
1635 		    bp, bp->b_bufobj, bo));
1636 		/*
1637 		 * XXX Since there are no node locks for NFS, I
1638 		 * believe there is a slight chance that a delayed
1639 		 * write will occur while sleeping just above, so
1640 		 * check for it.
1641 		 */
1642 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1643 		    (flags & V_SAVE)) {
1644 			bremfree(bp);
1645 			bp->b_flags |= B_ASYNC;
1646 			bwrite(bp);
1647 			BO_LOCK(bo);
1648 			return (EAGAIN);	/* XXX: why not loop ? */
1649 		}
1650 		bremfree(bp);
1651 		bp->b_flags |= (B_INVAL | B_RELBUF);
1652 		bp->b_flags &= ~B_ASYNC;
1653 		brelse(bp);
1654 		BO_LOCK(bo);
1655 		if (nbp != NULL &&
1656 		    (nbp->b_bufobj != bo ||
1657 		     nbp->b_lblkno != lblkno ||
1658 		     (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) != xflags))
1659 			break;			/* nbp invalid */
1660 	}
1661 	return (retval);
1662 }
1663 
1664 /*
1665  * Truncate a file's buffer and pages to a specified length.  This
1666  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1667  * sync activity.
1668  */
1669 int
1670 vtruncbuf(struct vnode *vp, struct ucred *cred, off_t length, int blksize)
1671 {
1672 	struct buf *bp, *nbp;
1673 	int anyfreed;
1674 	int trunclbn;
1675 	struct bufobj *bo;
1676 
1677 	CTR5(KTR_VFS, "%s: vp %p with cred %p and block %d:%ju", __func__,
1678 	    vp, cred, blksize, (uintmax_t)length);
1679 
1680 	/*
1681 	 * Round up to the *next* lbn.
1682 	 */
1683 	trunclbn = (length + blksize - 1) / blksize;
1684 
1685 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1686 restart:
1687 	bo = &vp->v_bufobj;
1688 	BO_LOCK(bo);
1689 	anyfreed = 1;
1690 	for (;anyfreed;) {
1691 		anyfreed = 0;
1692 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1693 			if (bp->b_lblkno < trunclbn)
1694 				continue;
1695 			if (BUF_LOCK(bp,
1696 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1697 			    BO_LOCKPTR(bo)) == ENOLCK)
1698 				goto restart;
1699 
1700 			bremfree(bp);
1701 			bp->b_flags |= (B_INVAL | B_RELBUF);
1702 			bp->b_flags &= ~B_ASYNC;
1703 			brelse(bp);
1704 			anyfreed = 1;
1705 
1706 			BO_LOCK(bo);
1707 			if (nbp != NULL &&
1708 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1709 			    (nbp->b_vp != vp) ||
1710 			    (nbp->b_flags & B_DELWRI))) {
1711 				BO_UNLOCK(bo);
1712 				goto restart;
1713 			}
1714 		}
1715 
1716 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1717 			if (bp->b_lblkno < trunclbn)
1718 				continue;
1719 			if (BUF_LOCK(bp,
1720 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1721 			    BO_LOCKPTR(bo)) == ENOLCK)
1722 				goto restart;
1723 			bremfree(bp);
1724 			bp->b_flags |= (B_INVAL | B_RELBUF);
1725 			bp->b_flags &= ~B_ASYNC;
1726 			brelse(bp);
1727 			anyfreed = 1;
1728 
1729 			BO_LOCK(bo);
1730 			if (nbp != NULL &&
1731 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1732 			    (nbp->b_vp != vp) ||
1733 			    (nbp->b_flags & B_DELWRI) == 0)) {
1734 				BO_UNLOCK(bo);
1735 				goto restart;
1736 			}
1737 		}
1738 	}
1739 
1740 	if (length > 0) {
1741 restartsync:
1742 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1743 			if (bp->b_lblkno > 0)
1744 				continue;
1745 			/*
1746 			 * Since we hold the vnode lock this should only
1747 			 * fail if we're racing with the buf daemon.
1748 			 */
1749 			if (BUF_LOCK(bp,
1750 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1751 			    BO_LOCKPTR(bo)) == ENOLCK) {
1752 				goto restart;
1753 			}
1754 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1755 			    ("buf(%p) on dirty queue without DELWRI", bp));
1756 
1757 			bremfree(bp);
1758 			bawrite(bp);
1759 			BO_LOCK(bo);
1760 			goto restartsync;
1761 		}
1762 	}
1763 
1764 	bufobj_wwait(bo, 0, 0);
1765 	BO_UNLOCK(bo);
1766 	vnode_pager_setsize(vp, length);
1767 
1768 	return (0);
1769 }
1770 
1771 static void
1772 buf_vlist_remove(struct buf *bp)
1773 {
1774 	struct bufv *bv;
1775 
1776 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1777 	ASSERT_BO_WLOCKED(bp->b_bufobj);
1778 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
1779 	    (BX_VNDIRTY|BX_VNCLEAN),
1780 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
1781 	if (bp->b_xflags & BX_VNDIRTY)
1782 		bv = &bp->b_bufobj->bo_dirty;
1783 	else
1784 		bv = &bp->b_bufobj->bo_clean;
1785 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
1786 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1787 	bv->bv_cnt--;
1788 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1789 }
1790 
1791 /*
1792  * Add the buffer to the sorted clean or dirty block list.
1793  *
1794  * NOTE: xflags is passed as a constant, optimizing this inline function!
1795  */
1796 static void
1797 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1798 {
1799 	struct bufv *bv;
1800 	struct buf *n;
1801 	int error;
1802 
1803 	ASSERT_BO_WLOCKED(bo);
1804 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
1805 	    ("dead bo %p", bo));
1806 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1807 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
1808 	bp->b_xflags |= xflags;
1809 	if (xflags & BX_VNDIRTY)
1810 		bv = &bo->bo_dirty;
1811 	else
1812 		bv = &bo->bo_clean;
1813 
1814 	/*
1815 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
1816 	 * we tend to grow at the tail so lookup_le should usually be cheaper
1817 	 * than _ge.
1818 	 */
1819 	if (bv->bv_cnt == 0 ||
1820 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
1821 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1822 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
1823 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
1824 	else
1825 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
1826 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
1827 	if (error)
1828 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
1829 	bv->bv_cnt++;
1830 }
1831 
1832 /*
1833  * Look up a buffer using the buffer tries.
1834  */
1835 struct buf *
1836 gbincore(struct bufobj *bo, daddr_t lblkno)
1837 {
1838 	struct buf *bp;
1839 
1840 	ASSERT_BO_LOCKED(bo);
1841 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
1842 	if (bp != NULL)
1843 		return (bp);
1844 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
1845 }
1846 
1847 /*
1848  * Associate a buffer with a vnode.
1849  */
1850 void
1851 bgetvp(struct vnode *vp, struct buf *bp)
1852 {
1853 	struct bufobj *bo;
1854 
1855 	bo = &vp->v_bufobj;
1856 	ASSERT_BO_WLOCKED(bo);
1857 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
1858 
1859 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
1860 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
1861 	    ("bgetvp: bp already attached! %p", bp));
1862 
1863 	vhold(vp);
1864 	bp->b_vp = vp;
1865 	bp->b_bufobj = bo;
1866 	/*
1867 	 * Insert onto list for new vnode.
1868 	 */
1869 	buf_vlist_add(bp, bo, BX_VNCLEAN);
1870 }
1871 
1872 /*
1873  * Disassociate a buffer from a vnode.
1874  */
1875 void
1876 brelvp(struct buf *bp)
1877 {
1878 	struct bufobj *bo;
1879 	struct vnode *vp;
1880 
1881 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1882 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1883 
1884 	/*
1885 	 * Delete from old vnode list, if on one.
1886 	 */
1887 	vp = bp->b_vp;		/* XXX */
1888 	bo = bp->b_bufobj;
1889 	BO_LOCK(bo);
1890 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1891 		buf_vlist_remove(bp);
1892 	else
1893 		panic("brelvp: Buffer %p not on queue.", bp);
1894 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1895 		bo->bo_flag &= ~BO_ONWORKLST;
1896 		mtx_lock(&sync_mtx);
1897 		LIST_REMOVE(bo, bo_synclist);
1898 		syncer_worklist_len--;
1899 		mtx_unlock(&sync_mtx);
1900 	}
1901 	bp->b_vp = NULL;
1902 	bp->b_bufobj = NULL;
1903 	BO_UNLOCK(bo);
1904 	vdrop(vp);
1905 }
1906 
1907 /*
1908  * Add an item to the syncer work queue.
1909  */
1910 static void
1911 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
1912 {
1913 	int slot;
1914 
1915 	ASSERT_BO_WLOCKED(bo);
1916 
1917 	mtx_lock(&sync_mtx);
1918 	if (bo->bo_flag & BO_ONWORKLST)
1919 		LIST_REMOVE(bo, bo_synclist);
1920 	else {
1921 		bo->bo_flag |= BO_ONWORKLST;
1922 		syncer_worklist_len++;
1923 	}
1924 
1925 	if (delay > syncer_maxdelay - 2)
1926 		delay = syncer_maxdelay - 2;
1927 	slot = (syncer_delayno + delay) & syncer_mask;
1928 
1929 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
1930 	mtx_unlock(&sync_mtx);
1931 }
1932 
1933 static int
1934 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
1935 {
1936 	int error, len;
1937 
1938 	mtx_lock(&sync_mtx);
1939 	len = syncer_worklist_len - sync_vnode_count;
1940 	mtx_unlock(&sync_mtx);
1941 	error = SYSCTL_OUT(req, &len, sizeof(len));
1942 	return (error);
1943 }
1944 
1945 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
1946     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
1947 
1948 static struct proc *updateproc;
1949 static void sched_sync(void);
1950 static struct kproc_desc up_kp = {
1951 	"syncer",
1952 	sched_sync,
1953 	&updateproc
1954 };
1955 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
1956 
1957 static int
1958 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
1959 {
1960 	struct vnode *vp;
1961 	struct mount *mp;
1962 
1963 	*bo = LIST_FIRST(slp);
1964 	if (*bo == NULL)
1965 		return (0);
1966 	vp = (*bo)->__bo_vnode;	/* XXX */
1967 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
1968 		return (1);
1969 	/*
1970 	 * We use vhold in case the vnode does not
1971 	 * successfully sync.  vhold prevents the vnode from
1972 	 * going away when we unlock the sync_mtx so that
1973 	 * we can acquire the vnode interlock.
1974 	 */
1975 	vholdl(vp);
1976 	mtx_unlock(&sync_mtx);
1977 	VI_UNLOCK(vp);
1978 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1979 		vdrop(vp);
1980 		mtx_lock(&sync_mtx);
1981 		return (*bo == LIST_FIRST(slp));
1982 	}
1983 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1984 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
1985 	VOP_UNLOCK(vp, 0);
1986 	vn_finished_write(mp);
1987 	BO_LOCK(*bo);
1988 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
1989 		/*
1990 		 * Put us back on the worklist.  The worklist
1991 		 * routine will remove us from our current
1992 		 * position and then add us back in at a later
1993 		 * position.
1994 		 */
1995 		vn_syncer_add_to_worklist(*bo, syncdelay);
1996 	}
1997 	BO_UNLOCK(*bo);
1998 	vdrop(vp);
1999 	mtx_lock(&sync_mtx);
2000 	return (0);
2001 }
2002 
2003 static int first_printf = 1;
2004 
2005 /*
2006  * System filesystem synchronizer daemon.
2007  */
2008 static void
2009 sched_sync(void)
2010 {
2011 	struct synclist *next, *slp;
2012 	struct bufobj *bo;
2013 	long starttime;
2014 	struct thread *td = curthread;
2015 	int last_work_seen;
2016 	int net_worklist_len;
2017 	int syncer_final_iter;
2018 	int error;
2019 
2020 	last_work_seen = 0;
2021 	syncer_final_iter = 0;
2022 	syncer_state = SYNCER_RUNNING;
2023 	starttime = time_uptime;
2024 	td->td_pflags |= TDP_NORUNNINGBUF;
2025 
2026 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2027 	    SHUTDOWN_PRI_LAST);
2028 
2029 	mtx_lock(&sync_mtx);
2030 	for (;;) {
2031 		if (syncer_state == SYNCER_FINAL_DELAY &&
2032 		    syncer_final_iter == 0) {
2033 			mtx_unlock(&sync_mtx);
2034 			kproc_suspend_check(td->td_proc);
2035 			mtx_lock(&sync_mtx);
2036 		}
2037 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2038 		if (syncer_state != SYNCER_RUNNING &&
2039 		    starttime != time_uptime) {
2040 			if (first_printf) {
2041 				printf("\nSyncing disks, vnodes remaining...");
2042 				first_printf = 0;
2043 			}
2044 			printf("%d ", net_worklist_len);
2045 		}
2046 		starttime = time_uptime;
2047 
2048 		/*
2049 		 * Push files whose dirty time has expired.  Be careful
2050 		 * of interrupt race on slp queue.
2051 		 *
2052 		 * Skip over empty worklist slots when shutting down.
2053 		 */
2054 		do {
2055 			slp = &syncer_workitem_pending[syncer_delayno];
2056 			syncer_delayno += 1;
2057 			if (syncer_delayno == syncer_maxdelay)
2058 				syncer_delayno = 0;
2059 			next = &syncer_workitem_pending[syncer_delayno];
2060 			/*
2061 			 * If the worklist has wrapped since the
2062 			 * it was emptied of all but syncer vnodes,
2063 			 * switch to the FINAL_DELAY state and run
2064 			 * for one more second.
2065 			 */
2066 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2067 			    net_worklist_len == 0 &&
2068 			    last_work_seen == syncer_delayno) {
2069 				syncer_state = SYNCER_FINAL_DELAY;
2070 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2071 			}
2072 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2073 		    syncer_worklist_len > 0);
2074 
2075 		/*
2076 		 * Keep track of the last time there was anything
2077 		 * on the worklist other than syncer vnodes.
2078 		 * Return to the SHUTTING_DOWN state if any
2079 		 * new work appears.
2080 		 */
2081 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2082 			last_work_seen = syncer_delayno;
2083 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2084 			syncer_state = SYNCER_SHUTTING_DOWN;
2085 		while (!LIST_EMPTY(slp)) {
2086 			error = sync_vnode(slp, &bo, td);
2087 			if (error == 1) {
2088 				LIST_REMOVE(bo, bo_synclist);
2089 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2090 				continue;
2091 			}
2092 
2093 			if (first_printf == 0) {
2094 				/*
2095 				 * Drop the sync mutex, because some watchdog
2096 				 * drivers need to sleep while patting
2097 				 */
2098 				mtx_unlock(&sync_mtx);
2099 				wdog_kern_pat(WD_LASTVAL);
2100 				mtx_lock(&sync_mtx);
2101 			}
2102 
2103 		}
2104 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2105 			syncer_final_iter--;
2106 		/*
2107 		 * The variable rushjob allows the kernel to speed up the
2108 		 * processing of the filesystem syncer process. A rushjob
2109 		 * value of N tells the filesystem syncer to process the next
2110 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2111 		 * is used by the soft update code to speed up the filesystem
2112 		 * syncer process when the incore state is getting so far
2113 		 * ahead of the disk that the kernel memory pool is being
2114 		 * threatened with exhaustion.
2115 		 */
2116 		if (rushjob > 0) {
2117 			rushjob -= 1;
2118 			continue;
2119 		}
2120 		/*
2121 		 * Just sleep for a short period of time between
2122 		 * iterations when shutting down to allow some I/O
2123 		 * to happen.
2124 		 *
2125 		 * If it has taken us less than a second to process the
2126 		 * current work, then wait. Otherwise start right over
2127 		 * again. We can still lose time if any single round
2128 		 * takes more than two seconds, but it does not really
2129 		 * matter as we are just trying to generally pace the
2130 		 * filesystem activity.
2131 		 */
2132 		if (syncer_state != SYNCER_RUNNING ||
2133 		    time_uptime == starttime) {
2134 			thread_lock(td);
2135 			sched_prio(td, PPAUSE);
2136 			thread_unlock(td);
2137 		}
2138 		if (syncer_state != SYNCER_RUNNING)
2139 			cv_timedwait(&sync_wakeup, &sync_mtx,
2140 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2141 		else if (time_uptime == starttime)
2142 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2143 	}
2144 }
2145 
2146 /*
2147  * Request the syncer daemon to speed up its work.
2148  * We never push it to speed up more than half of its
2149  * normal turn time, otherwise it could take over the cpu.
2150  */
2151 int
2152 speedup_syncer(void)
2153 {
2154 	int ret = 0;
2155 
2156 	mtx_lock(&sync_mtx);
2157 	if (rushjob < syncdelay / 2) {
2158 		rushjob += 1;
2159 		stat_rush_requests += 1;
2160 		ret = 1;
2161 	}
2162 	mtx_unlock(&sync_mtx);
2163 	cv_broadcast(&sync_wakeup);
2164 	return (ret);
2165 }
2166 
2167 /*
2168  * Tell the syncer to speed up its work and run though its work
2169  * list several times, then tell it to shut down.
2170  */
2171 static void
2172 syncer_shutdown(void *arg, int howto)
2173 {
2174 
2175 	if (howto & RB_NOSYNC)
2176 		return;
2177 	mtx_lock(&sync_mtx);
2178 	syncer_state = SYNCER_SHUTTING_DOWN;
2179 	rushjob = 0;
2180 	mtx_unlock(&sync_mtx);
2181 	cv_broadcast(&sync_wakeup);
2182 	kproc_shutdown(arg, howto);
2183 }
2184 
2185 void
2186 syncer_suspend(void)
2187 {
2188 
2189 	syncer_shutdown(updateproc, 0);
2190 }
2191 
2192 void
2193 syncer_resume(void)
2194 {
2195 
2196 	mtx_lock(&sync_mtx);
2197 	first_printf = 1;
2198 	syncer_state = SYNCER_RUNNING;
2199 	mtx_unlock(&sync_mtx);
2200 	cv_broadcast(&sync_wakeup);
2201 	kproc_resume(updateproc);
2202 }
2203 
2204 /*
2205  * Reassign a buffer from one vnode to another.
2206  * Used to assign file specific control information
2207  * (indirect blocks) to the vnode to which they belong.
2208  */
2209 void
2210 reassignbuf(struct buf *bp)
2211 {
2212 	struct vnode *vp;
2213 	struct bufobj *bo;
2214 	int delay;
2215 #ifdef INVARIANTS
2216 	struct bufv *bv;
2217 #endif
2218 
2219 	vp = bp->b_vp;
2220 	bo = bp->b_bufobj;
2221 	++reassignbufcalls;
2222 
2223 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2224 	    bp, bp->b_vp, bp->b_flags);
2225 	/*
2226 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2227 	 * is not fully linked in.
2228 	 */
2229 	if (bp->b_flags & B_PAGING)
2230 		panic("cannot reassign paging buffer");
2231 
2232 	/*
2233 	 * Delete from old vnode list, if on one.
2234 	 */
2235 	BO_LOCK(bo);
2236 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2237 		buf_vlist_remove(bp);
2238 	else
2239 		panic("reassignbuf: Buffer %p not on queue.", bp);
2240 	/*
2241 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2242 	 * of clean buffers.
2243 	 */
2244 	if (bp->b_flags & B_DELWRI) {
2245 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2246 			switch (vp->v_type) {
2247 			case VDIR:
2248 				delay = dirdelay;
2249 				break;
2250 			case VCHR:
2251 				delay = metadelay;
2252 				break;
2253 			default:
2254 				delay = filedelay;
2255 			}
2256 			vn_syncer_add_to_worklist(bo, delay);
2257 		}
2258 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2259 	} else {
2260 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2261 
2262 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2263 			mtx_lock(&sync_mtx);
2264 			LIST_REMOVE(bo, bo_synclist);
2265 			syncer_worklist_len--;
2266 			mtx_unlock(&sync_mtx);
2267 			bo->bo_flag &= ~BO_ONWORKLST;
2268 		}
2269 	}
2270 #ifdef INVARIANTS
2271 	bv = &bo->bo_clean;
2272 	bp = TAILQ_FIRST(&bv->bv_hd);
2273 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2274 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2275 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2276 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2277 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2278 	bv = &bo->bo_dirty;
2279 	bp = TAILQ_FIRST(&bv->bv_hd);
2280 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2281 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2282 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2283 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2284 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2285 #endif
2286 	BO_UNLOCK(bo);
2287 }
2288 
2289 /*
2290  * A temporary hack until refcount_* APIs are sorted out.
2291  */
2292 static __inline int
2293 vfs_refcount_acquire_if_not_zero(volatile u_int *count)
2294 {
2295 	u_int old;
2296 
2297 	for (;;) {
2298 		old = *count;
2299 		if (old == 0)
2300 			return (0);
2301 		if (atomic_cmpset_int(count, old, old + 1))
2302 			return (1);
2303 	}
2304 }
2305 
2306 static __inline int
2307 vfs_refcount_release_if_not_last(volatile u_int *count)
2308 {
2309 	u_int old;
2310 
2311 	for (;;) {
2312 		old = *count;
2313 		if (old == 1)
2314 			return (0);
2315 		if (atomic_cmpset_int(count, old, old - 1))
2316 			return (1);
2317 	}
2318 }
2319 
2320 static void
2321 v_init_counters(struct vnode *vp)
2322 {
2323 
2324 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2325 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2326 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2327 
2328 	refcount_init(&vp->v_holdcnt, 1);
2329 	refcount_init(&vp->v_usecount, 1);
2330 }
2331 
2332 /*
2333  * Increment the use and hold counts on the vnode, taking care to reference
2334  * the driver's usecount if this is a chardev.  The _vhold() will remove
2335  * the vnode from the free list if it is presently free.
2336  */
2337 static void
2338 v_incr_usecount(struct vnode *vp)
2339 {
2340 
2341 	ASSERT_VI_UNLOCKED(vp, __func__);
2342 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2343 
2344 	if (vp->v_type == VCHR) {
2345 		VI_LOCK(vp);
2346 		_vhold(vp, true);
2347 		if (vp->v_iflag & VI_OWEINACT) {
2348 			VNASSERT(vp->v_usecount == 0, vp,
2349 			    ("vnode with usecount and VI_OWEINACT set"));
2350 			vp->v_iflag &= ~VI_OWEINACT;
2351 		}
2352 		refcount_acquire(&vp->v_usecount);
2353 		v_incr_devcount(vp);
2354 		VI_UNLOCK(vp);
2355 		return;
2356 	}
2357 
2358 	_vhold(vp, false);
2359 	if (vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) {
2360 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2361 		    ("vnode with usecount and VI_OWEINACT set"));
2362 	} else {
2363 		VI_LOCK(vp);
2364 		if (vp->v_iflag & VI_OWEINACT)
2365 			vp->v_iflag &= ~VI_OWEINACT;
2366 		refcount_acquire(&vp->v_usecount);
2367 		VI_UNLOCK(vp);
2368 	}
2369 }
2370 
2371 /*
2372  * Increment si_usecount of the associated device, if any.
2373  */
2374 static void
2375 v_incr_devcount(struct vnode *vp)
2376 {
2377 
2378 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2379 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2380 		dev_lock();
2381 		vp->v_rdev->si_usecount++;
2382 		dev_unlock();
2383 	}
2384 }
2385 
2386 /*
2387  * Decrement si_usecount of the associated device, if any.
2388  */
2389 static void
2390 v_decr_devcount(struct vnode *vp)
2391 {
2392 
2393 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2394 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2395 		dev_lock();
2396 		vp->v_rdev->si_usecount--;
2397 		dev_unlock();
2398 	}
2399 }
2400 
2401 /*
2402  * Grab a particular vnode from the free list, increment its
2403  * reference count and lock it.  VI_DOOMED is set if the vnode
2404  * is being destroyed.  Only callers who specify LK_RETRY will
2405  * see doomed vnodes.  If inactive processing was delayed in
2406  * vput try to do it here.
2407  *
2408  * Notes on lockless counter manipulation:
2409  * _vhold, vputx and other routines make various decisions based
2410  * on either holdcnt or usecount being 0. As long as either contuner
2411  * is not transitioning 0->1 nor 1->0, the manipulation can be done
2412  * with atomic operations. Otherwise the interlock is taken.
2413  */
2414 int
2415 vget(struct vnode *vp, int flags, struct thread *td)
2416 {
2417 	int error, oweinact;
2418 
2419 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2420 	    ("vget: invalid lock operation"));
2421 
2422 	if ((flags & LK_INTERLOCK) != 0)
2423 		ASSERT_VI_LOCKED(vp, __func__);
2424 	else
2425 		ASSERT_VI_UNLOCKED(vp, __func__);
2426 	if ((flags & LK_VNHELD) != 0)
2427 		VNASSERT((vp->v_holdcnt > 0), vp,
2428 		    ("vget: LK_VNHELD passed but vnode not held"));
2429 
2430 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2431 
2432 	if ((flags & LK_VNHELD) == 0)
2433 		_vhold(vp, (flags & LK_INTERLOCK) != 0);
2434 
2435 	if ((error = vn_lock(vp, flags)) != 0) {
2436 		vdrop(vp);
2437 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2438 		    vp);
2439 		return (error);
2440 	}
2441 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
2442 		panic("vget: vn_lock failed to return ENOENT\n");
2443 	/*
2444 	 * We don't guarantee that any particular close will
2445 	 * trigger inactive processing so just make a best effort
2446 	 * here at preventing a reference to a removed file.  If
2447 	 * we don't succeed no harm is done.
2448 	 *
2449 	 * Upgrade our holdcnt to a usecount.
2450 	 */
2451 	if (vp->v_type != VCHR &&
2452 	    vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) {
2453 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2454 		    ("vnode with usecount and VI_OWEINACT set"));
2455 	} else {
2456 		VI_LOCK(vp);
2457 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2458 			oweinact = 0;
2459 		} else {
2460 			oweinact = 1;
2461 			vp->v_iflag &= ~VI_OWEINACT;
2462 		}
2463 		refcount_acquire(&vp->v_usecount);
2464 		v_incr_devcount(vp);
2465 		if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2466 		    (flags & LK_NOWAIT) == 0)
2467 			vinactive(vp, td);
2468 		VI_UNLOCK(vp);
2469 	}
2470 	return (0);
2471 }
2472 
2473 /*
2474  * Increase the reference count of a vnode.
2475  */
2476 void
2477 vref(struct vnode *vp)
2478 {
2479 
2480 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2481 	v_incr_usecount(vp);
2482 }
2483 
2484 /*
2485  * Return reference count of a vnode.
2486  *
2487  * The results of this call are only guaranteed when some mechanism is used to
2488  * stop other processes from gaining references to the vnode.  This may be the
2489  * case if the caller holds the only reference.  This is also useful when stale
2490  * data is acceptable as race conditions may be accounted for by some other
2491  * means.
2492  */
2493 int
2494 vrefcnt(struct vnode *vp)
2495 {
2496 
2497 	return (vp->v_usecount);
2498 }
2499 
2500 #define	VPUTX_VRELE	1
2501 #define	VPUTX_VPUT	2
2502 #define	VPUTX_VUNREF	3
2503 
2504 /*
2505  * Decrement the use and hold counts for a vnode.
2506  *
2507  * See an explanation near vget() as to why atomic operation is safe.
2508  */
2509 static void
2510 vputx(struct vnode *vp, int func)
2511 {
2512 	int error;
2513 
2514 	KASSERT(vp != NULL, ("vputx: null vp"));
2515 	if (func == VPUTX_VUNREF)
2516 		ASSERT_VOP_LOCKED(vp, "vunref");
2517 	else if (func == VPUTX_VPUT)
2518 		ASSERT_VOP_LOCKED(vp, "vput");
2519 	else
2520 		KASSERT(func == VPUTX_VRELE, ("vputx: wrong func"));
2521 	ASSERT_VI_UNLOCKED(vp, __func__);
2522 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2523 
2524 	if (vp->v_type != VCHR &&
2525 	    vfs_refcount_release_if_not_last(&vp->v_usecount)) {
2526 		if (func == VPUTX_VPUT)
2527 			VOP_UNLOCK(vp, 0);
2528 		vdrop(vp);
2529 		return;
2530 	}
2531 
2532 	VI_LOCK(vp);
2533 
2534 	/*
2535 	 * We want to hold the vnode until the inactive finishes to
2536 	 * prevent vgone() races.  We drop the use count here and the
2537 	 * hold count below when we're done.
2538 	 */
2539 	if (!refcount_release(&vp->v_usecount) ||
2540 	    (vp->v_iflag & VI_DOINGINACT)) {
2541 		if (func == VPUTX_VPUT)
2542 			VOP_UNLOCK(vp, 0);
2543 		v_decr_devcount(vp);
2544 		vdropl(vp);
2545 		return;
2546 	}
2547 
2548 	v_decr_devcount(vp);
2549 
2550 	error = 0;
2551 
2552 	if (vp->v_usecount != 0) {
2553 		vprint("vputx: usecount not zero", vp);
2554 		panic("vputx: usecount not zero");
2555 	}
2556 
2557 	CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp);
2558 
2559 	/*
2560 	 * We must call VOP_INACTIVE with the node locked. Mark
2561 	 * as VI_DOINGINACT to avoid recursion.
2562 	 */
2563 	vp->v_iflag |= VI_OWEINACT;
2564 	switch (func) {
2565 	case VPUTX_VRELE:
2566 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
2567 		VI_LOCK(vp);
2568 		break;
2569 	case VPUTX_VPUT:
2570 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2571 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
2572 			    LK_NOWAIT);
2573 			VI_LOCK(vp);
2574 		}
2575 		break;
2576 	case VPUTX_VUNREF:
2577 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2578 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
2579 			VI_LOCK(vp);
2580 		}
2581 		break;
2582 	}
2583 	VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
2584 	    ("vnode with usecount and VI_OWEINACT set"));
2585 	if (error == 0) {
2586 		if (vp->v_iflag & VI_OWEINACT)
2587 			vinactive(vp, curthread);
2588 		if (func != VPUTX_VUNREF)
2589 			VOP_UNLOCK(vp, 0);
2590 	}
2591 	vdropl(vp);
2592 }
2593 
2594 /*
2595  * Vnode put/release.
2596  * If count drops to zero, call inactive routine and return to freelist.
2597  */
2598 void
2599 vrele(struct vnode *vp)
2600 {
2601 
2602 	vputx(vp, VPUTX_VRELE);
2603 }
2604 
2605 /*
2606  * Release an already locked vnode.  This give the same effects as
2607  * unlock+vrele(), but takes less time and avoids releasing and
2608  * re-aquiring the lock (as vrele() acquires the lock internally.)
2609  */
2610 void
2611 vput(struct vnode *vp)
2612 {
2613 
2614 	vputx(vp, VPUTX_VPUT);
2615 }
2616 
2617 /*
2618  * Release an exclusively locked vnode. Do not unlock the vnode lock.
2619  */
2620 void
2621 vunref(struct vnode *vp)
2622 {
2623 
2624 	vputx(vp, VPUTX_VUNREF);
2625 }
2626 
2627 /*
2628  * Increase the hold count and activate if this is the first reference.
2629  */
2630 void
2631 _vhold(struct vnode *vp, bool locked)
2632 {
2633 	struct mount *mp;
2634 
2635 	if (locked)
2636 		ASSERT_VI_LOCKED(vp, __func__);
2637 	else
2638 		ASSERT_VI_UNLOCKED(vp, __func__);
2639 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2640 	if (!locked && vfs_refcount_acquire_if_not_zero(&vp->v_holdcnt)) {
2641 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2642 		    ("_vhold: vnode with holdcnt is free"));
2643 		return;
2644 	}
2645 
2646 	if (!locked)
2647 		VI_LOCK(vp);
2648 	if ((vp->v_iflag & VI_FREE) == 0) {
2649 		refcount_acquire(&vp->v_holdcnt);
2650 		if (!locked)
2651 			VI_UNLOCK(vp);
2652 		return;
2653 	}
2654 	VNASSERT(vp->v_holdcnt == 0, vp,
2655 	    ("%s: wrong hold count", __func__));
2656 	VNASSERT(vp->v_op != NULL, vp,
2657 	    ("%s: vnode already reclaimed.", __func__));
2658 	/*
2659 	 * Remove a vnode from the free list, mark it as in use,
2660 	 * and put it on the active list.
2661 	 */
2662 	mtx_lock(&vnode_free_list_mtx);
2663 	TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
2664 	freevnodes--;
2665 	vp->v_iflag &= ~VI_FREE;
2666 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
2667 	    ("Activating already active vnode"));
2668 	vp->v_iflag |= VI_ACTIVE;
2669 	mp = vp->v_mount;
2670 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
2671 	mp->mnt_activevnodelistsize++;
2672 	mtx_unlock(&vnode_free_list_mtx);
2673 	refcount_acquire(&vp->v_holdcnt);
2674 	if (!locked)
2675 		VI_UNLOCK(vp);
2676 }
2677 
2678 /*
2679  * Drop the hold count of the vnode.  If this is the last reference to
2680  * the vnode we place it on the free list unless it has been vgone'd
2681  * (marked VI_DOOMED) in which case we will free it.
2682  *
2683  * Because the vnode vm object keeps a hold reference on the vnode if
2684  * there is at least one resident non-cached page, the vnode cannot
2685  * leave the active list without the page cleanup done.
2686  */
2687 void
2688 _vdrop(struct vnode *vp, bool locked)
2689 {
2690 	struct bufobj *bo;
2691 	struct mount *mp;
2692 	int active;
2693 
2694 	if (locked)
2695 		ASSERT_VI_LOCKED(vp, __func__);
2696 	else
2697 		ASSERT_VI_UNLOCKED(vp, __func__);
2698 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2699 	if ((int)vp->v_holdcnt <= 0)
2700 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
2701 	if (vfs_refcount_release_if_not_last(&vp->v_holdcnt)) {
2702 		if (locked)
2703 			VI_UNLOCK(vp);
2704 		return;
2705 	}
2706 
2707 	if (!locked)
2708 		VI_LOCK(vp);
2709 	if (refcount_release(&vp->v_holdcnt) == 0) {
2710 		VI_UNLOCK(vp);
2711 		return;
2712 	}
2713 	if ((vp->v_iflag & VI_DOOMED) == 0) {
2714 		/*
2715 		 * Mark a vnode as free: remove it from its active list
2716 		 * and put it up for recycling on the freelist.
2717 		 */
2718 		VNASSERT(vp->v_op != NULL, vp,
2719 		    ("vdropl: vnode already reclaimed."));
2720 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2721 		    ("vnode already free"));
2722 		VNASSERT(vp->v_holdcnt == 0, vp,
2723 		    ("vdropl: freeing when we shouldn't"));
2724 		active = vp->v_iflag & VI_ACTIVE;
2725 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2726 			vp->v_iflag &= ~VI_ACTIVE;
2727 			mp = vp->v_mount;
2728 			mtx_lock(&vnode_free_list_mtx);
2729 			if (active) {
2730 				TAILQ_REMOVE(&mp->mnt_activevnodelist, vp,
2731 				    v_actfreelist);
2732 				mp->mnt_activevnodelistsize--;
2733 			}
2734 			TAILQ_INSERT_TAIL(&vnode_free_list, vp,
2735 			    v_actfreelist);
2736 			freevnodes++;
2737 			vp->v_iflag |= VI_FREE;
2738 			mtx_unlock(&vnode_free_list_mtx);
2739 		} else {
2740 			atomic_add_long(&free_owe_inact, 1);
2741 		}
2742 		VI_UNLOCK(vp);
2743 		return;
2744 	}
2745 	/*
2746 	 * The vnode has been marked for destruction, so free it.
2747 	 *
2748 	 * The vnode will be returned to the zone where it will
2749 	 * normally remain until it is needed for another vnode. We
2750 	 * need to cleanup (or verify that the cleanup has already
2751 	 * been done) any residual data left from its current use
2752 	 * so as not to contaminate the freshly allocated vnode.
2753 	 */
2754 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
2755 	atomic_subtract_long(&numvnodes, 1);
2756 	bo = &vp->v_bufobj;
2757 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2758 	    ("cleaned vnode still on the free list."));
2759 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
2760 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
2761 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
2762 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
2763 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
2764 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
2765 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
2766 	    ("clean blk trie not empty"));
2767 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
2768 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
2769 	    ("dirty blk trie not empty"));
2770 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
2771 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
2772 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
2773 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
2774 	    ("Dangling rangelock waiters"));
2775 	VI_UNLOCK(vp);
2776 #ifdef MAC
2777 	mac_vnode_destroy(vp);
2778 #endif
2779 	if (vp->v_pollinfo != NULL) {
2780 		destroy_vpollinfo(vp->v_pollinfo);
2781 		vp->v_pollinfo = NULL;
2782 	}
2783 #ifdef INVARIANTS
2784 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
2785 	vp->v_op = NULL;
2786 #endif
2787 	bzero(&vp->v_un, sizeof(vp->v_un));
2788 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
2789 	vp->v_iflag = 0;
2790 	vp->v_vflag = 0;
2791 	bo->bo_flag = 0;
2792 	uma_zfree(vnode_zone, vp);
2793 }
2794 
2795 /*
2796  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
2797  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
2798  * OWEINACT tracks whether a vnode missed a call to inactive due to a
2799  * failed lock upgrade.
2800  */
2801 void
2802 vinactive(struct vnode *vp, struct thread *td)
2803 {
2804 	struct vm_object *obj;
2805 
2806 	ASSERT_VOP_ELOCKED(vp, "vinactive");
2807 	ASSERT_VI_LOCKED(vp, "vinactive");
2808 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2809 	    ("vinactive: recursed on VI_DOINGINACT"));
2810 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2811 	vp->v_iflag |= VI_DOINGINACT;
2812 	vp->v_iflag &= ~VI_OWEINACT;
2813 	VI_UNLOCK(vp);
2814 	/*
2815 	 * Before moving off the active list, we must be sure that any
2816 	 * modified pages are converted into the vnode's dirty
2817 	 * buffers, since these will no longer be checked once the
2818 	 * vnode is on the inactive list.
2819 	 *
2820 	 * The write-out of the dirty pages is asynchronous.  At the
2821 	 * point that VOP_INACTIVE() is called, there could still be
2822 	 * pending I/O and dirty pages in the object.
2823 	 */
2824 	obj = vp->v_object;
2825 	if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0) {
2826 		VM_OBJECT_WLOCK(obj);
2827 		vm_object_page_clean(obj, 0, 0, OBJPC_NOSYNC);
2828 		VM_OBJECT_WUNLOCK(obj);
2829 	}
2830 	VOP_INACTIVE(vp, td);
2831 	VI_LOCK(vp);
2832 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2833 	    ("vinactive: lost VI_DOINGINACT"));
2834 	vp->v_iflag &= ~VI_DOINGINACT;
2835 }
2836 
2837 /*
2838  * Remove any vnodes in the vnode table belonging to mount point mp.
2839  *
2840  * If FORCECLOSE is not specified, there should not be any active ones,
2841  * return error if any are found (nb: this is a user error, not a
2842  * system error). If FORCECLOSE is specified, detach any active vnodes
2843  * that are found.
2844  *
2845  * If WRITECLOSE is set, only flush out regular file vnodes open for
2846  * writing.
2847  *
2848  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2849  *
2850  * `rootrefs' specifies the base reference count for the root vnode
2851  * of this filesystem. The root vnode is considered busy if its
2852  * v_usecount exceeds this value. On a successful return, vflush(, td)
2853  * will call vrele() on the root vnode exactly rootrefs times.
2854  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2855  * be zero.
2856  */
2857 #ifdef DIAGNOSTIC
2858 static int busyprt = 0;		/* print out busy vnodes */
2859 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
2860 #endif
2861 
2862 int
2863 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
2864 {
2865 	struct vnode *vp, *mvp, *rootvp = NULL;
2866 	struct vattr vattr;
2867 	int busy = 0, error;
2868 
2869 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
2870 	    rootrefs, flags);
2871 	if (rootrefs > 0) {
2872 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2873 		    ("vflush: bad args"));
2874 		/*
2875 		 * Get the filesystem root vnode. We can vput() it
2876 		 * immediately, since with rootrefs > 0, it won't go away.
2877 		 */
2878 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
2879 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
2880 			    __func__, error);
2881 			return (error);
2882 		}
2883 		vput(rootvp);
2884 	}
2885 loop:
2886 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
2887 		vholdl(vp);
2888 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
2889 		if (error) {
2890 			vdrop(vp);
2891 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
2892 			goto loop;
2893 		}
2894 		/*
2895 		 * Skip over a vnodes marked VV_SYSTEM.
2896 		 */
2897 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2898 			VOP_UNLOCK(vp, 0);
2899 			vdrop(vp);
2900 			continue;
2901 		}
2902 		/*
2903 		 * If WRITECLOSE is set, flush out unlinked but still open
2904 		 * files (even if open only for reading) and regular file
2905 		 * vnodes open for writing.
2906 		 */
2907 		if (flags & WRITECLOSE) {
2908 			if (vp->v_object != NULL) {
2909 				VM_OBJECT_WLOCK(vp->v_object);
2910 				vm_object_page_clean(vp->v_object, 0, 0, 0);
2911 				VM_OBJECT_WUNLOCK(vp->v_object);
2912 			}
2913 			error = VOP_FSYNC(vp, MNT_WAIT, td);
2914 			if (error != 0) {
2915 				VOP_UNLOCK(vp, 0);
2916 				vdrop(vp);
2917 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
2918 				return (error);
2919 			}
2920 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
2921 			VI_LOCK(vp);
2922 
2923 			if ((vp->v_type == VNON ||
2924 			    (error == 0 && vattr.va_nlink > 0)) &&
2925 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2926 				VOP_UNLOCK(vp, 0);
2927 				vdropl(vp);
2928 				continue;
2929 			}
2930 		} else
2931 			VI_LOCK(vp);
2932 		/*
2933 		 * With v_usecount == 0, all we need to do is clear out the
2934 		 * vnode data structures and we are done.
2935 		 *
2936 		 * If FORCECLOSE is set, forcibly close the vnode.
2937 		 */
2938 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
2939 			vgonel(vp);
2940 		} else {
2941 			busy++;
2942 #ifdef DIAGNOSTIC
2943 			if (busyprt)
2944 				vprint("vflush: busy vnode", vp);
2945 #endif
2946 		}
2947 		VOP_UNLOCK(vp, 0);
2948 		vdropl(vp);
2949 	}
2950 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2951 		/*
2952 		 * If just the root vnode is busy, and if its refcount
2953 		 * is equal to `rootrefs', then go ahead and kill it.
2954 		 */
2955 		VI_LOCK(rootvp);
2956 		KASSERT(busy > 0, ("vflush: not busy"));
2957 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
2958 		    ("vflush: usecount %d < rootrefs %d",
2959 		     rootvp->v_usecount, rootrefs));
2960 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2961 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
2962 			vgone(rootvp);
2963 			VOP_UNLOCK(rootvp, 0);
2964 			busy = 0;
2965 		} else
2966 			VI_UNLOCK(rootvp);
2967 	}
2968 	if (busy) {
2969 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
2970 		    busy);
2971 		return (EBUSY);
2972 	}
2973 	for (; rootrefs > 0; rootrefs--)
2974 		vrele(rootvp);
2975 	return (0);
2976 }
2977 
2978 /*
2979  * Recycle an unused vnode to the front of the free list.
2980  */
2981 int
2982 vrecycle(struct vnode *vp)
2983 {
2984 	int recycled;
2985 
2986 	ASSERT_VOP_ELOCKED(vp, "vrecycle");
2987 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2988 	recycled = 0;
2989 	VI_LOCK(vp);
2990 	if (vp->v_usecount == 0) {
2991 		recycled = 1;
2992 		vgonel(vp);
2993 	}
2994 	VI_UNLOCK(vp);
2995 	return (recycled);
2996 }
2997 
2998 /*
2999  * Eliminate all activity associated with a vnode
3000  * in preparation for reuse.
3001  */
3002 void
3003 vgone(struct vnode *vp)
3004 {
3005 	VI_LOCK(vp);
3006 	vgonel(vp);
3007 	VI_UNLOCK(vp);
3008 }
3009 
3010 static void
3011 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3012     struct vnode *lowervp __unused)
3013 {
3014 }
3015 
3016 /*
3017  * Notify upper mounts about reclaimed or unlinked vnode.
3018  */
3019 void
3020 vfs_notify_upper(struct vnode *vp, int event)
3021 {
3022 	static struct vfsops vgonel_vfsops = {
3023 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3024 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3025 	};
3026 	struct mount *mp, *ump, *mmp;
3027 
3028 	mp = vp->v_mount;
3029 	if (mp == NULL)
3030 		return;
3031 
3032 	MNT_ILOCK(mp);
3033 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3034 		goto unlock;
3035 	MNT_IUNLOCK(mp);
3036 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3037 	mmp->mnt_op = &vgonel_vfsops;
3038 	mmp->mnt_kern_flag |= MNTK_MARKER;
3039 	MNT_ILOCK(mp);
3040 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3041 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3042 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3043 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3044 			continue;
3045 		}
3046 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3047 		MNT_IUNLOCK(mp);
3048 		switch (event) {
3049 		case VFS_NOTIFY_UPPER_RECLAIM:
3050 			VFS_RECLAIM_LOWERVP(ump, vp);
3051 			break;
3052 		case VFS_NOTIFY_UPPER_UNLINK:
3053 			VFS_UNLINK_LOWERVP(ump, vp);
3054 			break;
3055 		default:
3056 			KASSERT(0, ("invalid event %d", event));
3057 			break;
3058 		}
3059 		MNT_ILOCK(mp);
3060 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3061 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3062 	}
3063 	free(mmp, M_TEMP);
3064 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3065 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3066 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3067 		wakeup(&mp->mnt_uppers);
3068 	}
3069 unlock:
3070 	MNT_IUNLOCK(mp);
3071 }
3072 
3073 /*
3074  * vgone, with the vp interlock held.
3075  */
3076 static void
3077 vgonel(struct vnode *vp)
3078 {
3079 	struct thread *td;
3080 	int oweinact;
3081 	int active;
3082 	struct mount *mp;
3083 
3084 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3085 	ASSERT_VI_LOCKED(vp, "vgonel");
3086 	VNASSERT(vp->v_holdcnt, vp,
3087 	    ("vgonel: vp %p has no reference.", vp));
3088 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3089 	td = curthread;
3090 
3091 	/*
3092 	 * Don't vgonel if we're already doomed.
3093 	 */
3094 	if (vp->v_iflag & VI_DOOMED)
3095 		return;
3096 	vp->v_iflag |= VI_DOOMED;
3097 
3098 	/*
3099 	 * Check to see if the vnode is in use.  If so, we have to call
3100 	 * VOP_CLOSE() and VOP_INACTIVE().
3101 	 */
3102 	active = vp->v_usecount;
3103 	oweinact = (vp->v_iflag & VI_OWEINACT);
3104 	VI_UNLOCK(vp);
3105 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3106 
3107 	/*
3108 	 * If purging an active vnode, it must be closed and
3109 	 * deactivated before being reclaimed.
3110 	 */
3111 	if (active)
3112 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3113 	if (oweinact || active) {
3114 		VI_LOCK(vp);
3115 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
3116 			vinactive(vp, td);
3117 		VI_UNLOCK(vp);
3118 	}
3119 	if (vp->v_type == VSOCK)
3120 		vfs_unp_reclaim(vp);
3121 
3122 	/*
3123 	 * Clean out any buffers associated with the vnode.
3124 	 * If the flush fails, just toss the buffers.
3125 	 */
3126 	mp = NULL;
3127 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3128 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3129 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3130 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3131 			;
3132 	}
3133 
3134 	BO_LOCK(&vp->v_bufobj);
3135 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3136 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3137 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3138 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3139 	    ("vp %p bufobj not invalidated", vp));
3140 	vp->v_bufobj.bo_flag |= BO_DEAD;
3141 	BO_UNLOCK(&vp->v_bufobj);
3142 
3143 	/*
3144 	 * Reclaim the vnode.
3145 	 */
3146 	if (VOP_RECLAIM(vp, td))
3147 		panic("vgone: cannot reclaim");
3148 	if (mp != NULL)
3149 		vn_finished_secondary_write(mp);
3150 	VNASSERT(vp->v_object == NULL, vp,
3151 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
3152 	/*
3153 	 * Clear the advisory locks and wake up waiting threads.
3154 	 */
3155 	(void)VOP_ADVLOCKPURGE(vp);
3156 	vp->v_lockf = NULL;
3157 	/*
3158 	 * Delete from old mount point vnode list.
3159 	 */
3160 	delmntque(vp);
3161 	cache_purge(vp);
3162 	/*
3163 	 * Done with purge, reset to the standard lock and invalidate
3164 	 * the vnode.
3165 	 */
3166 	VI_LOCK(vp);
3167 	vp->v_vnlock = &vp->v_lock;
3168 	vp->v_op = &dead_vnodeops;
3169 	vp->v_tag = "none";
3170 	vp->v_type = VBAD;
3171 }
3172 
3173 /*
3174  * Calculate the total number of references to a special device.
3175  */
3176 int
3177 vcount(struct vnode *vp)
3178 {
3179 	int count;
3180 
3181 	dev_lock();
3182 	count = vp->v_rdev->si_usecount;
3183 	dev_unlock();
3184 	return (count);
3185 }
3186 
3187 /*
3188  * Same as above, but using the struct cdev *as argument
3189  */
3190 int
3191 count_dev(struct cdev *dev)
3192 {
3193 	int count;
3194 
3195 	dev_lock();
3196 	count = dev->si_usecount;
3197 	dev_unlock();
3198 	return(count);
3199 }
3200 
3201 /*
3202  * Print out a description of a vnode.
3203  */
3204 static char *typename[] =
3205 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3206  "VMARKER"};
3207 
3208 void
3209 vn_printf(struct vnode *vp, const char *fmt, ...)
3210 {
3211 	va_list ap;
3212 	char buf[256], buf2[16];
3213 	u_long flags;
3214 
3215 	va_start(ap, fmt);
3216 	vprintf(fmt, ap);
3217 	va_end(ap);
3218 	printf("%p: ", (void *)vp);
3219 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
3220 	printf("    usecount %d, writecount %d, refcount %d mountedhere %p\n",
3221 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere);
3222 	buf[0] = '\0';
3223 	buf[1] = '\0';
3224 	if (vp->v_vflag & VV_ROOT)
3225 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3226 	if (vp->v_vflag & VV_ISTTY)
3227 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3228 	if (vp->v_vflag & VV_NOSYNC)
3229 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3230 	if (vp->v_vflag & VV_ETERNALDEV)
3231 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3232 	if (vp->v_vflag & VV_CACHEDLABEL)
3233 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3234 	if (vp->v_vflag & VV_TEXT)
3235 		strlcat(buf, "|VV_TEXT", sizeof(buf));
3236 	if (vp->v_vflag & VV_COPYONWRITE)
3237 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3238 	if (vp->v_vflag & VV_SYSTEM)
3239 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3240 	if (vp->v_vflag & VV_PROCDEP)
3241 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3242 	if (vp->v_vflag & VV_NOKNOTE)
3243 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3244 	if (vp->v_vflag & VV_DELETED)
3245 		strlcat(buf, "|VV_DELETED", sizeof(buf));
3246 	if (vp->v_vflag & VV_MD)
3247 		strlcat(buf, "|VV_MD", sizeof(buf));
3248 	if (vp->v_vflag & VV_FORCEINSMQ)
3249 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
3250 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
3251 	    VV_CACHEDLABEL | VV_TEXT | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
3252 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
3253 	if (flags != 0) {
3254 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
3255 		strlcat(buf, buf2, sizeof(buf));
3256 	}
3257 	if (vp->v_iflag & VI_MOUNT)
3258 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
3259 	if (vp->v_iflag & VI_DOOMED)
3260 		strlcat(buf, "|VI_DOOMED", sizeof(buf));
3261 	if (vp->v_iflag & VI_FREE)
3262 		strlcat(buf, "|VI_FREE", sizeof(buf));
3263 	if (vp->v_iflag & VI_ACTIVE)
3264 		strlcat(buf, "|VI_ACTIVE", sizeof(buf));
3265 	if (vp->v_iflag & VI_DOINGINACT)
3266 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
3267 	if (vp->v_iflag & VI_OWEINACT)
3268 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
3269 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOOMED | VI_FREE |
3270 	    VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT);
3271 	if (flags != 0) {
3272 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
3273 		strlcat(buf, buf2, sizeof(buf));
3274 	}
3275 	printf("    flags (%s)\n", buf + 1);
3276 	if (mtx_owned(VI_MTX(vp)))
3277 		printf(" VI_LOCKed");
3278 	if (vp->v_object != NULL)
3279 		printf("    v_object %p ref %d pages %d "
3280 		    "cleanbuf %d dirtybuf %d\n",
3281 		    vp->v_object, vp->v_object->ref_count,
3282 		    vp->v_object->resident_page_count,
3283 		    vp->v_bufobj.bo_clean.bv_cnt,
3284 		    vp->v_bufobj.bo_dirty.bv_cnt);
3285 	printf("    ");
3286 	lockmgr_printinfo(vp->v_vnlock);
3287 	if (vp->v_data != NULL)
3288 		VOP_PRINT(vp);
3289 }
3290 
3291 #ifdef DDB
3292 /*
3293  * List all of the locked vnodes in the system.
3294  * Called when debugging the kernel.
3295  */
3296 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
3297 {
3298 	struct mount *mp;
3299 	struct vnode *vp;
3300 
3301 	/*
3302 	 * Note: because this is DDB, we can't obey the locking semantics
3303 	 * for these structures, which means we could catch an inconsistent
3304 	 * state and dereference a nasty pointer.  Not much to be done
3305 	 * about that.
3306 	 */
3307 	db_printf("Locked vnodes\n");
3308 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3309 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3310 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
3311 				vprint("", vp);
3312 		}
3313 	}
3314 }
3315 
3316 /*
3317  * Show details about the given vnode.
3318  */
3319 DB_SHOW_COMMAND(vnode, db_show_vnode)
3320 {
3321 	struct vnode *vp;
3322 
3323 	if (!have_addr)
3324 		return;
3325 	vp = (struct vnode *)addr;
3326 	vn_printf(vp, "vnode ");
3327 }
3328 
3329 /*
3330  * Show details about the given mount point.
3331  */
3332 DB_SHOW_COMMAND(mount, db_show_mount)
3333 {
3334 	struct mount *mp;
3335 	struct vfsopt *opt;
3336 	struct statfs *sp;
3337 	struct vnode *vp;
3338 	char buf[512];
3339 	uint64_t mflags;
3340 	u_int flags;
3341 
3342 	if (!have_addr) {
3343 		/* No address given, print short info about all mount points. */
3344 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3345 			db_printf("%p %s on %s (%s)\n", mp,
3346 			    mp->mnt_stat.f_mntfromname,
3347 			    mp->mnt_stat.f_mntonname,
3348 			    mp->mnt_stat.f_fstypename);
3349 			if (db_pager_quit)
3350 				break;
3351 		}
3352 		db_printf("\nMore info: show mount <addr>\n");
3353 		return;
3354 	}
3355 
3356 	mp = (struct mount *)addr;
3357 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
3358 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
3359 
3360 	buf[0] = '\0';
3361 	mflags = mp->mnt_flag;
3362 #define	MNT_FLAG(flag)	do {						\
3363 	if (mflags & (flag)) {						\
3364 		if (buf[0] != '\0')					\
3365 			strlcat(buf, ", ", sizeof(buf));		\
3366 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
3367 		mflags &= ~(flag);					\
3368 	}								\
3369 } while (0)
3370 	MNT_FLAG(MNT_RDONLY);
3371 	MNT_FLAG(MNT_SYNCHRONOUS);
3372 	MNT_FLAG(MNT_NOEXEC);
3373 	MNT_FLAG(MNT_NOSUID);
3374 	MNT_FLAG(MNT_NFS4ACLS);
3375 	MNT_FLAG(MNT_UNION);
3376 	MNT_FLAG(MNT_ASYNC);
3377 	MNT_FLAG(MNT_SUIDDIR);
3378 	MNT_FLAG(MNT_SOFTDEP);
3379 	MNT_FLAG(MNT_NOSYMFOLLOW);
3380 	MNT_FLAG(MNT_GJOURNAL);
3381 	MNT_FLAG(MNT_MULTILABEL);
3382 	MNT_FLAG(MNT_ACLS);
3383 	MNT_FLAG(MNT_NOATIME);
3384 	MNT_FLAG(MNT_NOCLUSTERR);
3385 	MNT_FLAG(MNT_NOCLUSTERW);
3386 	MNT_FLAG(MNT_SUJ);
3387 	MNT_FLAG(MNT_EXRDONLY);
3388 	MNT_FLAG(MNT_EXPORTED);
3389 	MNT_FLAG(MNT_DEFEXPORTED);
3390 	MNT_FLAG(MNT_EXPORTANON);
3391 	MNT_FLAG(MNT_EXKERB);
3392 	MNT_FLAG(MNT_EXPUBLIC);
3393 	MNT_FLAG(MNT_LOCAL);
3394 	MNT_FLAG(MNT_QUOTA);
3395 	MNT_FLAG(MNT_ROOTFS);
3396 	MNT_FLAG(MNT_USER);
3397 	MNT_FLAG(MNT_IGNORE);
3398 	MNT_FLAG(MNT_UPDATE);
3399 	MNT_FLAG(MNT_DELEXPORT);
3400 	MNT_FLAG(MNT_RELOAD);
3401 	MNT_FLAG(MNT_FORCE);
3402 	MNT_FLAG(MNT_SNAPSHOT);
3403 	MNT_FLAG(MNT_BYFSID);
3404 #undef MNT_FLAG
3405 	if (mflags != 0) {
3406 		if (buf[0] != '\0')
3407 			strlcat(buf, ", ", sizeof(buf));
3408 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3409 		    "0x%016jx", mflags);
3410 	}
3411 	db_printf("    mnt_flag = %s\n", buf);
3412 
3413 	buf[0] = '\0';
3414 	flags = mp->mnt_kern_flag;
3415 #define	MNT_KERN_FLAG(flag)	do {					\
3416 	if (flags & (flag)) {						\
3417 		if (buf[0] != '\0')					\
3418 			strlcat(buf, ", ", sizeof(buf));		\
3419 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
3420 		flags &= ~(flag);					\
3421 	}								\
3422 } while (0)
3423 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
3424 	MNT_KERN_FLAG(MNTK_ASYNC);
3425 	MNT_KERN_FLAG(MNTK_SOFTDEP);
3426 	MNT_KERN_FLAG(MNTK_NOINSMNTQ);
3427 	MNT_KERN_FLAG(MNTK_DRAINING);
3428 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
3429 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
3430 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
3431 	MNT_KERN_FLAG(MNTK_NO_IOPF);
3432 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
3433 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
3434 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
3435 	MNT_KERN_FLAG(MNTK_MARKER);
3436 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
3437 	MNT_KERN_FLAG(MNTK_NOASYNC);
3438 	MNT_KERN_FLAG(MNTK_UNMOUNT);
3439 	MNT_KERN_FLAG(MNTK_MWAIT);
3440 	MNT_KERN_FLAG(MNTK_SUSPEND);
3441 	MNT_KERN_FLAG(MNTK_SUSPEND2);
3442 	MNT_KERN_FLAG(MNTK_SUSPENDED);
3443 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
3444 	MNT_KERN_FLAG(MNTK_NOKNOTE);
3445 #undef MNT_KERN_FLAG
3446 	if (flags != 0) {
3447 		if (buf[0] != '\0')
3448 			strlcat(buf, ", ", sizeof(buf));
3449 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3450 		    "0x%08x", flags);
3451 	}
3452 	db_printf("    mnt_kern_flag = %s\n", buf);
3453 
3454 	db_printf("    mnt_opt = ");
3455 	opt = TAILQ_FIRST(mp->mnt_opt);
3456 	if (opt != NULL) {
3457 		db_printf("%s", opt->name);
3458 		opt = TAILQ_NEXT(opt, link);
3459 		while (opt != NULL) {
3460 			db_printf(", %s", opt->name);
3461 			opt = TAILQ_NEXT(opt, link);
3462 		}
3463 	}
3464 	db_printf("\n");
3465 
3466 	sp = &mp->mnt_stat;
3467 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
3468 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
3469 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
3470 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
3471 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
3472 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
3473 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
3474 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
3475 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
3476 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
3477 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
3478 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
3479 
3480 	db_printf("    mnt_cred = { uid=%u ruid=%u",
3481 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
3482 	if (jailed(mp->mnt_cred))
3483 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
3484 	db_printf(" }\n");
3485 	db_printf("    mnt_ref = %d\n", mp->mnt_ref);
3486 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
3487 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
3488 	db_printf("    mnt_activevnodelistsize = %d\n",
3489 	    mp->mnt_activevnodelistsize);
3490 	db_printf("    mnt_writeopcount = %d\n", mp->mnt_writeopcount);
3491 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
3492 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
3493 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
3494 	db_printf("    mnt_lockref = %d\n", mp->mnt_lockref);
3495 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
3496 	db_printf("    mnt_secondary_accwrites = %d\n",
3497 	    mp->mnt_secondary_accwrites);
3498 	db_printf("    mnt_gjprovider = %s\n",
3499 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
3500 
3501 	db_printf("\n\nList of active vnodes\n");
3502 	TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) {
3503 		if (vp->v_type != VMARKER) {
3504 			vn_printf(vp, "vnode ");
3505 			if (db_pager_quit)
3506 				break;
3507 		}
3508 	}
3509 	db_printf("\n\nList of inactive vnodes\n");
3510 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3511 		if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) {
3512 			vn_printf(vp, "vnode ");
3513 			if (db_pager_quit)
3514 				break;
3515 		}
3516 	}
3517 }
3518 #endif	/* DDB */
3519 
3520 /*
3521  * Fill in a struct xvfsconf based on a struct vfsconf.
3522  */
3523 static int
3524 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
3525 {
3526 	struct xvfsconf xvfsp;
3527 
3528 	bzero(&xvfsp, sizeof(xvfsp));
3529 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3530 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3531 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3532 	xvfsp.vfc_flags = vfsp->vfc_flags;
3533 	/*
3534 	 * These are unused in userland, we keep them
3535 	 * to not break binary compatibility.
3536 	 */
3537 	xvfsp.vfc_vfsops = NULL;
3538 	xvfsp.vfc_next = NULL;
3539 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3540 }
3541 
3542 #ifdef COMPAT_FREEBSD32
3543 struct xvfsconf32 {
3544 	uint32_t	vfc_vfsops;
3545 	char		vfc_name[MFSNAMELEN];
3546 	int32_t		vfc_typenum;
3547 	int32_t		vfc_refcount;
3548 	int32_t		vfc_flags;
3549 	uint32_t	vfc_next;
3550 };
3551 
3552 static int
3553 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
3554 {
3555 	struct xvfsconf32 xvfsp;
3556 
3557 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3558 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3559 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3560 	xvfsp.vfc_flags = vfsp->vfc_flags;
3561 	xvfsp.vfc_vfsops = 0;
3562 	xvfsp.vfc_next = 0;
3563 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3564 }
3565 #endif
3566 
3567 /*
3568  * Top level filesystem related information gathering.
3569  */
3570 static int
3571 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
3572 {
3573 	struct vfsconf *vfsp;
3574 	int error;
3575 
3576 	error = 0;
3577 	vfsconf_slock();
3578 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3579 #ifdef COMPAT_FREEBSD32
3580 		if (req->flags & SCTL_MASK32)
3581 			error = vfsconf2x32(req, vfsp);
3582 		else
3583 #endif
3584 			error = vfsconf2x(req, vfsp);
3585 		if (error)
3586 			break;
3587 	}
3588 	vfsconf_sunlock();
3589 	return (error);
3590 }
3591 
3592 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
3593     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
3594     "S,xvfsconf", "List of all configured filesystems");
3595 
3596 #ifndef BURN_BRIDGES
3597 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
3598 
3599 static int
3600 vfs_sysctl(SYSCTL_HANDLER_ARGS)
3601 {
3602 	int *name = (int *)arg1 - 1;	/* XXX */
3603 	u_int namelen = arg2 + 1;	/* XXX */
3604 	struct vfsconf *vfsp;
3605 
3606 	log(LOG_WARNING, "userland calling deprecated sysctl, "
3607 	    "please rebuild world\n");
3608 
3609 #if 1 || defined(COMPAT_PRELITE2)
3610 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
3611 	if (namelen == 1)
3612 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
3613 #endif
3614 
3615 	switch (name[1]) {
3616 	case VFS_MAXTYPENUM:
3617 		if (namelen != 2)
3618 			return (ENOTDIR);
3619 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
3620 	case VFS_CONF:
3621 		if (namelen != 3)
3622 			return (ENOTDIR);	/* overloaded */
3623 		vfsconf_slock();
3624 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3625 			if (vfsp->vfc_typenum == name[2])
3626 				break;
3627 		}
3628 		vfsconf_sunlock();
3629 		if (vfsp == NULL)
3630 			return (EOPNOTSUPP);
3631 #ifdef COMPAT_FREEBSD32
3632 		if (req->flags & SCTL_MASK32)
3633 			return (vfsconf2x32(req, vfsp));
3634 		else
3635 #endif
3636 			return (vfsconf2x(req, vfsp));
3637 	}
3638 	return (EOPNOTSUPP);
3639 }
3640 
3641 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
3642     CTLFLAG_MPSAFE, vfs_sysctl,
3643     "Generic filesystem");
3644 
3645 #if 1 || defined(COMPAT_PRELITE2)
3646 
3647 static int
3648 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
3649 {
3650 	int error;
3651 	struct vfsconf *vfsp;
3652 	struct ovfsconf ovfs;
3653 
3654 	vfsconf_slock();
3655 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3656 		bzero(&ovfs, sizeof(ovfs));
3657 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
3658 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
3659 		ovfs.vfc_index = vfsp->vfc_typenum;
3660 		ovfs.vfc_refcount = vfsp->vfc_refcount;
3661 		ovfs.vfc_flags = vfsp->vfc_flags;
3662 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
3663 		if (error != 0) {
3664 			vfsconf_sunlock();
3665 			return (error);
3666 		}
3667 	}
3668 	vfsconf_sunlock();
3669 	return (0);
3670 }
3671 
3672 #endif /* 1 || COMPAT_PRELITE2 */
3673 #endif /* !BURN_BRIDGES */
3674 
3675 #define KINFO_VNODESLOP		10
3676 #ifdef notyet
3677 /*
3678  * Dump vnode list (via sysctl).
3679  */
3680 /* ARGSUSED */
3681 static int
3682 sysctl_vnode(SYSCTL_HANDLER_ARGS)
3683 {
3684 	struct xvnode *xvn;
3685 	struct mount *mp;
3686 	struct vnode *vp;
3687 	int error, len, n;
3688 
3689 	/*
3690 	 * Stale numvnodes access is not fatal here.
3691 	 */
3692 	req->lock = 0;
3693 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
3694 	if (!req->oldptr)
3695 		/* Make an estimate */
3696 		return (SYSCTL_OUT(req, 0, len));
3697 
3698 	error = sysctl_wire_old_buffer(req, 0);
3699 	if (error != 0)
3700 		return (error);
3701 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
3702 	n = 0;
3703 	mtx_lock(&mountlist_mtx);
3704 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3705 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
3706 			continue;
3707 		MNT_ILOCK(mp);
3708 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3709 			if (n == len)
3710 				break;
3711 			vref(vp);
3712 			xvn[n].xv_size = sizeof *xvn;
3713 			xvn[n].xv_vnode = vp;
3714 			xvn[n].xv_id = 0;	/* XXX compat */
3715 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
3716 			XV_COPY(usecount);
3717 			XV_COPY(writecount);
3718 			XV_COPY(holdcnt);
3719 			XV_COPY(mount);
3720 			XV_COPY(numoutput);
3721 			XV_COPY(type);
3722 #undef XV_COPY
3723 			xvn[n].xv_flag = vp->v_vflag;
3724 
3725 			switch (vp->v_type) {
3726 			case VREG:
3727 			case VDIR:
3728 			case VLNK:
3729 				break;
3730 			case VBLK:
3731 			case VCHR:
3732 				if (vp->v_rdev == NULL) {
3733 					vrele(vp);
3734 					continue;
3735 				}
3736 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
3737 				break;
3738 			case VSOCK:
3739 				xvn[n].xv_socket = vp->v_socket;
3740 				break;
3741 			case VFIFO:
3742 				xvn[n].xv_fifo = vp->v_fifoinfo;
3743 				break;
3744 			case VNON:
3745 			case VBAD:
3746 			default:
3747 				/* shouldn't happen? */
3748 				vrele(vp);
3749 				continue;
3750 			}
3751 			vrele(vp);
3752 			++n;
3753 		}
3754 		MNT_IUNLOCK(mp);
3755 		mtx_lock(&mountlist_mtx);
3756 		vfs_unbusy(mp);
3757 		if (n == len)
3758 			break;
3759 	}
3760 	mtx_unlock(&mountlist_mtx);
3761 
3762 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
3763 	free(xvn, M_TEMP);
3764 	return (error);
3765 }
3766 
3767 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
3768     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
3769     "");
3770 #endif
3771 
3772 static void
3773 unmount_or_warn(struct mount *mp)
3774 {
3775 	int error;
3776 
3777 	error = dounmount(mp, MNT_FORCE, curthread);
3778 	if (error != 0) {
3779 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
3780 		if (error == EBUSY)
3781 			printf("BUSY)\n");
3782 		else
3783 			printf("%d)\n", error);
3784 	}
3785 }
3786 
3787 /*
3788  * Unmount all filesystems. The list is traversed in reverse order
3789  * of mounting to avoid dependencies.
3790  */
3791 void
3792 vfs_unmountall(void)
3793 {
3794 	struct mount *mp, *tmp;
3795 
3796 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
3797 
3798 	/*
3799 	 * Since this only runs when rebooting, it is not interlocked.
3800 	 */
3801 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
3802 		vfs_ref(mp);
3803 
3804 		/*
3805 		 * Forcibly unmounting "/dev" before "/" would prevent clean
3806 		 * unmount of the latter.
3807 		 */
3808 		if (mp == rootdevmp)
3809 			continue;
3810 
3811 		unmount_or_warn(mp);
3812 	}
3813 
3814 	if (rootdevmp != NULL)
3815 		unmount_or_warn(rootdevmp);
3816 }
3817 
3818 /*
3819  * perform msync on all vnodes under a mount point
3820  * the mount point must be locked.
3821  */
3822 void
3823 vfs_msync(struct mount *mp, int flags)
3824 {
3825 	struct vnode *vp, *mvp;
3826 	struct vm_object *obj;
3827 
3828 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
3829 	MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
3830 		obj = vp->v_object;
3831 		if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 &&
3832 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
3833 			if (!vget(vp,
3834 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
3835 			    curthread)) {
3836 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
3837 					vput(vp);
3838 					continue;
3839 				}
3840 
3841 				obj = vp->v_object;
3842 				if (obj != NULL) {
3843 					VM_OBJECT_WLOCK(obj);
3844 					vm_object_page_clean(obj, 0, 0,
3845 					    flags == MNT_WAIT ?
3846 					    OBJPC_SYNC : OBJPC_NOSYNC);
3847 					VM_OBJECT_WUNLOCK(obj);
3848 				}
3849 				vput(vp);
3850 			}
3851 		} else
3852 			VI_UNLOCK(vp);
3853 	}
3854 }
3855 
3856 static void
3857 destroy_vpollinfo_free(struct vpollinfo *vi)
3858 {
3859 
3860 	knlist_destroy(&vi->vpi_selinfo.si_note);
3861 	mtx_destroy(&vi->vpi_lock);
3862 	uma_zfree(vnodepoll_zone, vi);
3863 }
3864 
3865 static void
3866 destroy_vpollinfo(struct vpollinfo *vi)
3867 {
3868 
3869 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
3870 	seldrain(&vi->vpi_selinfo);
3871 	destroy_vpollinfo_free(vi);
3872 }
3873 
3874 /*
3875  * Initalize per-vnode helper structure to hold poll-related state.
3876  */
3877 void
3878 v_addpollinfo(struct vnode *vp)
3879 {
3880 	struct vpollinfo *vi;
3881 
3882 	if (vp->v_pollinfo != NULL)
3883 		return;
3884 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
3885 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
3886 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
3887 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
3888 	VI_LOCK(vp);
3889 	if (vp->v_pollinfo != NULL) {
3890 		VI_UNLOCK(vp);
3891 		destroy_vpollinfo_free(vi);
3892 		return;
3893 	}
3894 	vp->v_pollinfo = vi;
3895 	VI_UNLOCK(vp);
3896 }
3897 
3898 /*
3899  * Record a process's interest in events which might happen to
3900  * a vnode.  Because poll uses the historic select-style interface
3901  * internally, this routine serves as both the ``check for any
3902  * pending events'' and the ``record my interest in future events''
3903  * functions.  (These are done together, while the lock is held,
3904  * to avoid race conditions.)
3905  */
3906 int
3907 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
3908 {
3909 
3910 	v_addpollinfo(vp);
3911 	mtx_lock(&vp->v_pollinfo->vpi_lock);
3912 	if (vp->v_pollinfo->vpi_revents & events) {
3913 		/*
3914 		 * This leaves events we are not interested
3915 		 * in available for the other process which
3916 		 * which presumably had requested them
3917 		 * (otherwise they would never have been
3918 		 * recorded).
3919 		 */
3920 		events &= vp->v_pollinfo->vpi_revents;
3921 		vp->v_pollinfo->vpi_revents &= ~events;
3922 
3923 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
3924 		return (events);
3925 	}
3926 	vp->v_pollinfo->vpi_events |= events;
3927 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
3928 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3929 	return (0);
3930 }
3931 
3932 /*
3933  * Routine to create and manage a filesystem syncer vnode.
3934  */
3935 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
3936 static int	sync_fsync(struct  vop_fsync_args *);
3937 static int	sync_inactive(struct  vop_inactive_args *);
3938 static int	sync_reclaim(struct  vop_reclaim_args *);
3939 
3940 static struct vop_vector sync_vnodeops = {
3941 	.vop_bypass =	VOP_EOPNOTSUPP,
3942 	.vop_close =	sync_close,		/* close */
3943 	.vop_fsync =	sync_fsync,		/* fsync */
3944 	.vop_inactive =	sync_inactive,	/* inactive */
3945 	.vop_reclaim =	sync_reclaim,	/* reclaim */
3946 	.vop_lock1 =	vop_stdlock,	/* lock */
3947 	.vop_unlock =	vop_stdunlock,	/* unlock */
3948 	.vop_islocked =	vop_stdislocked,	/* islocked */
3949 };
3950 
3951 /*
3952  * Create a new filesystem syncer vnode for the specified mount point.
3953  */
3954 void
3955 vfs_allocate_syncvnode(struct mount *mp)
3956 {
3957 	struct vnode *vp;
3958 	struct bufobj *bo;
3959 	static long start, incr, next;
3960 	int error;
3961 
3962 	/* Allocate a new vnode */
3963 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
3964 	if (error != 0)
3965 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
3966 	vp->v_type = VNON;
3967 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
3968 	vp->v_vflag |= VV_FORCEINSMQ;
3969 	error = insmntque(vp, mp);
3970 	if (error != 0)
3971 		panic("vfs_allocate_syncvnode: insmntque() failed");
3972 	vp->v_vflag &= ~VV_FORCEINSMQ;
3973 	VOP_UNLOCK(vp, 0);
3974 	/*
3975 	 * Place the vnode onto the syncer worklist. We attempt to
3976 	 * scatter them about on the list so that they will go off
3977 	 * at evenly distributed times even if all the filesystems
3978 	 * are mounted at once.
3979 	 */
3980 	next += incr;
3981 	if (next == 0 || next > syncer_maxdelay) {
3982 		start /= 2;
3983 		incr /= 2;
3984 		if (start == 0) {
3985 			start = syncer_maxdelay / 2;
3986 			incr = syncer_maxdelay;
3987 		}
3988 		next = start;
3989 	}
3990 	bo = &vp->v_bufobj;
3991 	BO_LOCK(bo);
3992 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
3993 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
3994 	mtx_lock(&sync_mtx);
3995 	sync_vnode_count++;
3996 	if (mp->mnt_syncer == NULL) {
3997 		mp->mnt_syncer = vp;
3998 		vp = NULL;
3999 	}
4000 	mtx_unlock(&sync_mtx);
4001 	BO_UNLOCK(bo);
4002 	if (vp != NULL) {
4003 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4004 		vgone(vp);
4005 		vput(vp);
4006 	}
4007 }
4008 
4009 void
4010 vfs_deallocate_syncvnode(struct mount *mp)
4011 {
4012 	struct vnode *vp;
4013 
4014 	mtx_lock(&sync_mtx);
4015 	vp = mp->mnt_syncer;
4016 	if (vp != NULL)
4017 		mp->mnt_syncer = NULL;
4018 	mtx_unlock(&sync_mtx);
4019 	if (vp != NULL)
4020 		vrele(vp);
4021 }
4022 
4023 /*
4024  * Do a lazy sync of the filesystem.
4025  */
4026 static int
4027 sync_fsync(struct vop_fsync_args *ap)
4028 {
4029 	struct vnode *syncvp = ap->a_vp;
4030 	struct mount *mp = syncvp->v_mount;
4031 	int error, save;
4032 	struct bufobj *bo;
4033 
4034 	/*
4035 	 * We only need to do something if this is a lazy evaluation.
4036 	 */
4037 	if (ap->a_waitfor != MNT_LAZY)
4038 		return (0);
4039 
4040 	/*
4041 	 * Move ourselves to the back of the sync list.
4042 	 */
4043 	bo = &syncvp->v_bufobj;
4044 	BO_LOCK(bo);
4045 	vn_syncer_add_to_worklist(bo, syncdelay);
4046 	BO_UNLOCK(bo);
4047 
4048 	/*
4049 	 * Walk the list of vnodes pushing all that are dirty and
4050 	 * not already on the sync list.
4051 	 */
4052 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4053 		return (0);
4054 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4055 		vfs_unbusy(mp);
4056 		return (0);
4057 	}
4058 	save = curthread_pflags_set(TDP_SYNCIO);
4059 	vfs_msync(mp, MNT_NOWAIT);
4060 	error = VFS_SYNC(mp, MNT_LAZY);
4061 	curthread_pflags_restore(save);
4062 	vn_finished_write(mp);
4063 	vfs_unbusy(mp);
4064 	return (error);
4065 }
4066 
4067 /*
4068  * The syncer vnode is no referenced.
4069  */
4070 static int
4071 sync_inactive(struct vop_inactive_args *ap)
4072 {
4073 
4074 	vgone(ap->a_vp);
4075 	return (0);
4076 }
4077 
4078 /*
4079  * The syncer vnode is no longer needed and is being decommissioned.
4080  *
4081  * Modifications to the worklist must be protected by sync_mtx.
4082  */
4083 static int
4084 sync_reclaim(struct vop_reclaim_args *ap)
4085 {
4086 	struct vnode *vp = ap->a_vp;
4087 	struct bufobj *bo;
4088 
4089 	bo = &vp->v_bufobj;
4090 	BO_LOCK(bo);
4091 	mtx_lock(&sync_mtx);
4092 	if (vp->v_mount->mnt_syncer == vp)
4093 		vp->v_mount->mnt_syncer = NULL;
4094 	if (bo->bo_flag & BO_ONWORKLST) {
4095 		LIST_REMOVE(bo, bo_synclist);
4096 		syncer_worklist_len--;
4097 		sync_vnode_count--;
4098 		bo->bo_flag &= ~BO_ONWORKLST;
4099 	}
4100 	mtx_unlock(&sync_mtx);
4101 	BO_UNLOCK(bo);
4102 
4103 	return (0);
4104 }
4105 
4106 /*
4107  * Check if vnode represents a disk device
4108  */
4109 int
4110 vn_isdisk(struct vnode *vp, int *errp)
4111 {
4112 	int error;
4113 
4114 	if (vp->v_type != VCHR) {
4115 		error = ENOTBLK;
4116 		goto out;
4117 	}
4118 	error = 0;
4119 	dev_lock();
4120 	if (vp->v_rdev == NULL)
4121 		error = ENXIO;
4122 	else if (vp->v_rdev->si_devsw == NULL)
4123 		error = ENXIO;
4124 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
4125 		error = ENOTBLK;
4126 	dev_unlock();
4127 out:
4128 	if (errp != NULL)
4129 		*errp = error;
4130 	return (error == 0);
4131 }
4132 
4133 /*
4134  * Common filesystem object access control check routine.  Accepts a
4135  * vnode's type, "mode", uid and gid, requested access mode, credentials,
4136  * and optional call-by-reference privused argument allowing vaccess()
4137  * to indicate to the caller whether privilege was used to satisfy the
4138  * request (obsoleted).  Returns 0 on success, or an errno on failure.
4139  */
4140 int
4141 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
4142     accmode_t accmode, struct ucred *cred, int *privused)
4143 {
4144 	accmode_t dac_granted;
4145 	accmode_t priv_granted;
4146 
4147 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
4148 	    ("invalid bit in accmode"));
4149 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
4150 	    ("VAPPEND without VWRITE"));
4151 
4152 	/*
4153 	 * Look for a normal, non-privileged way to access the file/directory
4154 	 * as requested.  If it exists, go with that.
4155 	 */
4156 
4157 	if (privused != NULL)
4158 		*privused = 0;
4159 
4160 	dac_granted = 0;
4161 
4162 	/* Check the owner. */
4163 	if (cred->cr_uid == file_uid) {
4164 		dac_granted |= VADMIN;
4165 		if (file_mode & S_IXUSR)
4166 			dac_granted |= VEXEC;
4167 		if (file_mode & S_IRUSR)
4168 			dac_granted |= VREAD;
4169 		if (file_mode & S_IWUSR)
4170 			dac_granted |= (VWRITE | VAPPEND);
4171 
4172 		if ((accmode & dac_granted) == accmode)
4173 			return (0);
4174 
4175 		goto privcheck;
4176 	}
4177 
4178 	/* Otherwise, check the groups (first match) */
4179 	if (groupmember(file_gid, cred)) {
4180 		if (file_mode & S_IXGRP)
4181 			dac_granted |= VEXEC;
4182 		if (file_mode & S_IRGRP)
4183 			dac_granted |= VREAD;
4184 		if (file_mode & S_IWGRP)
4185 			dac_granted |= (VWRITE | VAPPEND);
4186 
4187 		if ((accmode & dac_granted) == accmode)
4188 			return (0);
4189 
4190 		goto privcheck;
4191 	}
4192 
4193 	/* Otherwise, check everyone else. */
4194 	if (file_mode & S_IXOTH)
4195 		dac_granted |= VEXEC;
4196 	if (file_mode & S_IROTH)
4197 		dac_granted |= VREAD;
4198 	if (file_mode & S_IWOTH)
4199 		dac_granted |= (VWRITE | VAPPEND);
4200 	if ((accmode & dac_granted) == accmode)
4201 		return (0);
4202 
4203 privcheck:
4204 	/*
4205 	 * Build a privilege mask to determine if the set of privileges
4206 	 * satisfies the requirements when combined with the granted mask
4207 	 * from above.  For each privilege, if the privilege is required,
4208 	 * bitwise or the request type onto the priv_granted mask.
4209 	 */
4210 	priv_granted = 0;
4211 
4212 	if (type == VDIR) {
4213 		/*
4214 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
4215 		 * requests, instead of PRIV_VFS_EXEC.
4216 		 */
4217 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4218 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP, 0))
4219 			priv_granted |= VEXEC;
4220 	} else {
4221 		/*
4222 		 * Ensure that at least one execute bit is on. Otherwise,
4223 		 * a privileged user will always succeed, and we don't want
4224 		 * this to happen unless the file really is executable.
4225 		 */
4226 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4227 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
4228 		    !priv_check_cred(cred, PRIV_VFS_EXEC, 0))
4229 			priv_granted |= VEXEC;
4230 	}
4231 
4232 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
4233 	    !priv_check_cred(cred, PRIV_VFS_READ, 0))
4234 		priv_granted |= VREAD;
4235 
4236 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
4237 	    !priv_check_cred(cred, PRIV_VFS_WRITE, 0))
4238 		priv_granted |= (VWRITE | VAPPEND);
4239 
4240 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
4241 	    !priv_check_cred(cred, PRIV_VFS_ADMIN, 0))
4242 		priv_granted |= VADMIN;
4243 
4244 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
4245 		/* XXX audit: privilege used */
4246 		if (privused != NULL)
4247 			*privused = 1;
4248 		return (0);
4249 	}
4250 
4251 	return ((accmode & VADMIN) ? EPERM : EACCES);
4252 }
4253 
4254 /*
4255  * Credential check based on process requesting service, and per-attribute
4256  * permissions.
4257  */
4258 int
4259 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
4260     struct thread *td, accmode_t accmode)
4261 {
4262 
4263 	/*
4264 	 * Kernel-invoked always succeeds.
4265 	 */
4266 	if (cred == NOCRED)
4267 		return (0);
4268 
4269 	/*
4270 	 * Do not allow privileged processes in jail to directly manipulate
4271 	 * system attributes.
4272 	 */
4273 	switch (attrnamespace) {
4274 	case EXTATTR_NAMESPACE_SYSTEM:
4275 		/* Potentially should be: return (EPERM); */
4276 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0));
4277 	case EXTATTR_NAMESPACE_USER:
4278 		return (VOP_ACCESS(vp, accmode, cred, td));
4279 	default:
4280 		return (EPERM);
4281 	}
4282 }
4283 
4284 #ifdef DEBUG_VFS_LOCKS
4285 /*
4286  * This only exists to supress warnings from unlocked specfs accesses.  It is
4287  * no longer ok to have an unlocked VFS.
4288  */
4289 #define	IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL ||		\
4290 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
4291 
4292 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
4293 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
4294     "Drop into debugger on lock violation");
4295 
4296 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
4297 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
4298     0, "Check for interlock across VOPs");
4299 
4300 int vfs_badlock_print = 1;	/* Print lock violations. */
4301 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
4302     0, "Print lock violations");
4303 
4304 #ifdef KDB
4305 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
4306 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
4307     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
4308 #endif
4309 
4310 static void
4311 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
4312 {
4313 
4314 #ifdef KDB
4315 	if (vfs_badlock_backtrace)
4316 		kdb_backtrace();
4317 #endif
4318 	if (vfs_badlock_print)
4319 		printf("%s: %p %s\n", str, (void *)vp, msg);
4320 	if (vfs_badlock_ddb)
4321 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4322 }
4323 
4324 void
4325 assert_vi_locked(struct vnode *vp, const char *str)
4326 {
4327 
4328 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
4329 		vfs_badlock("interlock is not locked but should be", str, vp);
4330 }
4331 
4332 void
4333 assert_vi_unlocked(struct vnode *vp, const char *str)
4334 {
4335 
4336 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
4337 		vfs_badlock("interlock is locked but should not be", str, vp);
4338 }
4339 
4340 void
4341 assert_vop_locked(struct vnode *vp, const char *str)
4342 {
4343 	int locked;
4344 
4345 	if (!IGNORE_LOCK(vp)) {
4346 		locked = VOP_ISLOCKED(vp);
4347 		if (locked == 0 || locked == LK_EXCLOTHER)
4348 			vfs_badlock("is not locked but should be", str, vp);
4349 	}
4350 }
4351 
4352 void
4353 assert_vop_unlocked(struct vnode *vp, const char *str)
4354 {
4355 
4356 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
4357 		vfs_badlock("is locked but should not be", str, vp);
4358 }
4359 
4360 void
4361 assert_vop_elocked(struct vnode *vp, const char *str)
4362 {
4363 
4364 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
4365 		vfs_badlock("is not exclusive locked but should be", str, vp);
4366 }
4367 
4368 #if 0
4369 void
4370 assert_vop_elocked_other(struct vnode *vp, const char *str)
4371 {
4372 
4373 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLOTHER)
4374 		vfs_badlock("is not exclusive locked by another thread",
4375 		    str, vp);
4376 }
4377 
4378 void
4379 assert_vop_slocked(struct vnode *vp, const char *str)
4380 {
4381 
4382 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_SHARED)
4383 		vfs_badlock("is not locked shared but should be", str, vp);
4384 }
4385 #endif /* 0 */
4386 #endif /* DEBUG_VFS_LOCKS */
4387 
4388 void
4389 vop_rename_fail(struct vop_rename_args *ap)
4390 {
4391 
4392 	if (ap->a_tvp != NULL)
4393 		vput(ap->a_tvp);
4394 	if (ap->a_tdvp == ap->a_tvp)
4395 		vrele(ap->a_tdvp);
4396 	else
4397 		vput(ap->a_tdvp);
4398 	vrele(ap->a_fdvp);
4399 	vrele(ap->a_fvp);
4400 }
4401 
4402 void
4403 vop_rename_pre(void *ap)
4404 {
4405 	struct vop_rename_args *a = ap;
4406 
4407 #ifdef DEBUG_VFS_LOCKS
4408 	if (a->a_tvp)
4409 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
4410 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
4411 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
4412 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
4413 
4414 	/* Check the source (from). */
4415 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
4416 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
4417 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
4418 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
4419 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
4420 
4421 	/* Check the target. */
4422 	if (a->a_tvp)
4423 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
4424 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
4425 #endif
4426 	if (a->a_tdvp != a->a_fdvp)
4427 		vhold(a->a_fdvp);
4428 	if (a->a_tvp != a->a_fvp)
4429 		vhold(a->a_fvp);
4430 	vhold(a->a_tdvp);
4431 	if (a->a_tvp)
4432 		vhold(a->a_tvp);
4433 }
4434 
4435 void
4436 vop_strategy_pre(void *ap)
4437 {
4438 #ifdef DEBUG_VFS_LOCKS
4439 	struct vop_strategy_args *a;
4440 	struct buf *bp;
4441 
4442 	a = ap;
4443 	bp = a->a_bp;
4444 
4445 	/*
4446 	 * Cluster ops lock their component buffers but not the IO container.
4447 	 */
4448 	if ((bp->b_flags & B_CLUSTER) != 0)
4449 		return;
4450 
4451 	if (panicstr == NULL && !BUF_ISLOCKED(bp)) {
4452 		if (vfs_badlock_print)
4453 			printf(
4454 			    "VOP_STRATEGY: bp is not locked but should be\n");
4455 		if (vfs_badlock_ddb)
4456 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4457 	}
4458 #endif
4459 }
4460 
4461 void
4462 vop_lock_pre(void *ap)
4463 {
4464 #ifdef DEBUG_VFS_LOCKS
4465 	struct vop_lock1_args *a = ap;
4466 
4467 	if ((a->a_flags & LK_INTERLOCK) == 0)
4468 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4469 	else
4470 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
4471 #endif
4472 }
4473 
4474 void
4475 vop_lock_post(void *ap, int rc)
4476 {
4477 #ifdef DEBUG_VFS_LOCKS
4478 	struct vop_lock1_args *a = ap;
4479 
4480 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4481 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
4482 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
4483 #endif
4484 }
4485 
4486 void
4487 vop_unlock_pre(void *ap)
4488 {
4489 #ifdef DEBUG_VFS_LOCKS
4490 	struct vop_unlock_args *a = ap;
4491 
4492 	if (a->a_flags & LK_INTERLOCK)
4493 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
4494 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
4495 #endif
4496 }
4497 
4498 void
4499 vop_unlock_post(void *ap, int rc)
4500 {
4501 #ifdef DEBUG_VFS_LOCKS
4502 	struct vop_unlock_args *a = ap;
4503 
4504 	if (a->a_flags & LK_INTERLOCK)
4505 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
4506 #endif
4507 }
4508 
4509 void
4510 vop_create_post(void *ap, int rc)
4511 {
4512 	struct vop_create_args *a = ap;
4513 
4514 	if (!rc)
4515 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4516 }
4517 
4518 void
4519 vop_deleteextattr_post(void *ap, int rc)
4520 {
4521 	struct vop_deleteextattr_args *a = ap;
4522 
4523 	if (!rc)
4524 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4525 }
4526 
4527 void
4528 vop_link_post(void *ap, int rc)
4529 {
4530 	struct vop_link_args *a = ap;
4531 
4532 	if (!rc) {
4533 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
4534 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
4535 	}
4536 }
4537 
4538 void
4539 vop_mkdir_post(void *ap, int rc)
4540 {
4541 	struct vop_mkdir_args *a = ap;
4542 
4543 	if (!rc)
4544 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4545 }
4546 
4547 void
4548 vop_mknod_post(void *ap, int rc)
4549 {
4550 	struct vop_mknod_args *a = ap;
4551 
4552 	if (!rc)
4553 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4554 }
4555 
4556 void
4557 vop_reclaim_post(void *ap, int rc)
4558 {
4559 	struct vop_reclaim_args *a = ap;
4560 
4561 	if (!rc)
4562 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
4563 }
4564 
4565 void
4566 vop_remove_post(void *ap, int rc)
4567 {
4568 	struct vop_remove_args *a = ap;
4569 
4570 	if (!rc) {
4571 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4572 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4573 	}
4574 }
4575 
4576 void
4577 vop_rename_post(void *ap, int rc)
4578 {
4579 	struct vop_rename_args *a = ap;
4580 
4581 	if (!rc) {
4582 		VFS_KNOTE_UNLOCKED(a->a_fdvp, NOTE_WRITE);
4583 		VFS_KNOTE_UNLOCKED(a->a_tdvp, NOTE_WRITE);
4584 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
4585 		if (a->a_tvp)
4586 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
4587 	}
4588 	if (a->a_tdvp != a->a_fdvp)
4589 		vdrop(a->a_fdvp);
4590 	if (a->a_tvp != a->a_fvp)
4591 		vdrop(a->a_fvp);
4592 	vdrop(a->a_tdvp);
4593 	if (a->a_tvp)
4594 		vdrop(a->a_tvp);
4595 }
4596 
4597 void
4598 vop_rmdir_post(void *ap, int rc)
4599 {
4600 	struct vop_rmdir_args *a = ap;
4601 
4602 	if (!rc) {
4603 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4604 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4605 	}
4606 }
4607 
4608 void
4609 vop_setattr_post(void *ap, int rc)
4610 {
4611 	struct vop_setattr_args *a = ap;
4612 
4613 	if (!rc)
4614 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4615 }
4616 
4617 void
4618 vop_setextattr_post(void *ap, int rc)
4619 {
4620 	struct vop_setextattr_args *a = ap;
4621 
4622 	if (!rc)
4623 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4624 }
4625 
4626 void
4627 vop_symlink_post(void *ap, int rc)
4628 {
4629 	struct vop_symlink_args *a = ap;
4630 
4631 	if (!rc)
4632 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4633 }
4634 
4635 static struct knlist fs_knlist;
4636 
4637 static void
4638 vfs_event_init(void *arg)
4639 {
4640 	knlist_init_mtx(&fs_knlist, NULL);
4641 }
4642 /* XXX - correct order? */
4643 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
4644 
4645 void
4646 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
4647 {
4648 
4649 	KNOTE_UNLOCKED(&fs_knlist, event);
4650 }
4651 
4652 static int	filt_fsattach(struct knote *kn);
4653 static void	filt_fsdetach(struct knote *kn);
4654 static int	filt_fsevent(struct knote *kn, long hint);
4655 
4656 struct filterops fs_filtops = {
4657 	.f_isfd = 0,
4658 	.f_attach = filt_fsattach,
4659 	.f_detach = filt_fsdetach,
4660 	.f_event = filt_fsevent
4661 };
4662 
4663 static int
4664 filt_fsattach(struct knote *kn)
4665 {
4666 
4667 	kn->kn_flags |= EV_CLEAR;
4668 	knlist_add(&fs_knlist, kn, 0);
4669 	return (0);
4670 }
4671 
4672 static void
4673 filt_fsdetach(struct knote *kn)
4674 {
4675 
4676 	knlist_remove(&fs_knlist, kn, 0);
4677 }
4678 
4679 static int
4680 filt_fsevent(struct knote *kn, long hint)
4681 {
4682 
4683 	kn->kn_fflags |= hint;
4684 	return (kn->kn_fflags != 0);
4685 }
4686 
4687 static int
4688 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
4689 {
4690 	struct vfsidctl vc;
4691 	int error;
4692 	struct mount *mp;
4693 
4694 	error = SYSCTL_IN(req, &vc, sizeof(vc));
4695 	if (error)
4696 		return (error);
4697 	if (vc.vc_vers != VFS_CTL_VERS1)
4698 		return (EINVAL);
4699 	mp = vfs_getvfs(&vc.vc_fsid);
4700 	if (mp == NULL)
4701 		return (ENOENT);
4702 	/* ensure that a specific sysctl goes to the right filesystem. */
4703 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
4704 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
4705 		vfs_rel(mp);
4706 		return (EINVAL);
4707 	}
4708 	VCTLTOREQ(&vc, req);
4709 	error = VFS_SYSCTL(mp, vc.vc_op, req);
4710 	vfs_rel(mp);
4711 	return (error);
4712 }
4713 
4714 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR,
4715     NULL, 0, sysctl_vfs_ctl, "",
4716     "Sysctl by fsid");
4717 
4718 /*
4719  * Function to initialize a va_filerev field sensibly.
4720  * XXX: Wouldn't a random number make a lot more sense ??
4721  */
4722 u_quad_t
4723 init_va_filerev(void)
4724 {
4725 	struct bintime bt;
4726 
4727 	getbinuptime(&bt);
4728 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
4729 }
4730 
4731 static int	filt_vfsread(struct knote *kn, long hint);
4732 static int	filt_vfswrite(struct knote *kn, long hint);
4733 static int	filt_vfsvnode(struct knote *kn, long hint);
4734 static void	filt_vfsdetach(struct knote *kn);
4735 static struct filterops vfsread_filtops = {
4736 	.f_isfd = 1,
4737 	.f_detach = filt_vfsdetach,
4738 	.f_event = filt_vfsread
4739 };
4740 static struct filterops vfswrite_filtops = {
4741 	.f_isfd = 1,
4742 	.f_detach = filt_vfsdetach,
4743 	.f_event = filt_vfswrite
4744 };
4745 static struct filterops vfsvnode_filtops = {
4746 	.f_isfd = 1,
4747 	.f_detach = filt_vfsdetach,
4748 	.f_event = filt_vfsvnode
4749 };
4750 
4751 static void
4752 vfs_knllock(void *arg)
4753 {
4754 	struct vnode *vp = arg;
4755 
4756 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4757 }
4758 
4759 static void
4760 vfs_knlunlock(void *arg)
4761 {
4762 	struct vnode *vp = arg;
4763 
4764 	VOP_UNLOCK(vp, 0);
4765 }
4766 
4767 static void
4768 vfs_knl_assert_locked(void *arg)
4769 {
4770 #ifdef DEBUG_VFS_LOCKS
4771 	struct vnode *vp = arg;
4772 
4773 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
4774 #endif
4775 }
4776 
4777 static void
4778 vfs_knl_assert_unlocked(void *arg)
4779 {
4780 #ifdef DEBUG_VFS_LOCKS
4781 	struct vnode *vp = arg;
4782 
4783 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
4784 #endif
4785 }
4786 
4787 int
4788 vfs_kqfilter(struct vop_kqfilter_args *ap)
4789 {
4790 	struct vnode *vp = ap->a_vp;
4791 	struct knote *kn = ap->a_kn;
4792 	struct knlist *knl;
4793 
4794 	switch (kn->kn_filter) {
4795 	case EVFILT_READ:
4796 		kn->kn_fop = &vfsread_filtops;
4797 		break;
4798 	case EVFILT_WRITE:
4799 		kn->kn_fop = &vfswrite_filtops;
4800 		break;
4801 	case EVFILT_VNODE:
4802 		kn->kn_fop = &vfsvnode_filtops;
4803 		break;
4804 	default:
4805 		return (EINVAL);
4806 	}
4807 
4808 	kn->kn_hook = (caddr_t)vp;
4809 
4810 	v_addpollinfo(vp);
4811 	if (vp->v_pollinfo == NULL)
4812 		return (ENOMEM);
4813 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
4814 	vhold(vp);
4815 	knlist_add(knl, kn, 0);
4816 
4817 	return (0);
4818 }
4819 
4820 /*
4821  * Detach knote from vnode
4822  */
4823 static void
4824 filt_vfsdetach(struct knote *kn)
4825 {
4826 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4827 
4828 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
4829 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
4830 	vdrop(vp);
4831 }
4832 
4833 /*ARGSUSED*/
4834 static int
4835 filt_vfsread(struct knote *kn, long hint)
4836 {
4837 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4838 	struct vattr va;
4839 	int res;
4840 
4841 	/*
4842 	 * filesystem is gone, so set the EOF flag and schedule
4843 	 * the knote for deletion.
4844 	 */
4845 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
4846 		VI_LOCK(vp);
4847 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
4848 		VI_UNLOCK(vp);
4849 		return (1);
4850 	}
4851 
4852 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
4853 		return (0);
4854 
4855 	VI_LOCK(vp);
4856 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
4857 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
4858 	VI_UNLOCK(vp);
4859 	return (res);
4860 }
4861 
4862 /*ARGSUSED*/
4863 static int
4864 filt_vfswrite(struct knote *kn, long hint)
4865 {
4866 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4867 
4868 	VI_LOCK(vp);
4869 
4870 	/*
4871 	 * filesystem is gone, so set the EOF flag and schedule
4872 	 * the knote for deletion.
4873 	 */
4874 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
4875 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
4876 
4877 	kn->kn_data = 0;
4878 	VI_UNLOCK(vp);
4879 	return (1);
4880 }
4881 
4882 static int
4883 filt_vfsvnode(struct knote *kn, long hint)
4884 {
4885 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4886 	int res;
4887 
4888 	VI_LOCK(vp);
4889 	if (kn->kn_sfflags & hint)
4890 		kn->kn_fflags |= hint;
4891 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
4892 		kn->kn_flags |= EV_EOF;
4893 		VI_UNLOCK(vp);
4894 		return (1);
4895 	}
4896 	res = (kn->kn_fflags != 0);
4897 	VI_UNLOCK(vp);
4898 	return (res);
4899 }
4900 
4901 int
4902 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
4903 {
4904 	int error;
4905 
4906 	if (dp->d_reclen > ap->a_uio->uio_resid)
4907 		return (ENAMETOOLONG);
4908 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
4909 	if (error) {
4910 		if (ap->a_ncookies != NULL) {
4911 			if (ap->a_cookies != NULL)
4912 				free(ap->a_cookies, M_TEMP);
4913 			ap->a_cookies = NULL;
4914 			*ap->a_ncookies = 0;
4915 		}
4916 		return (error);
4917 	}
4918 	if (ap->a_ncookies == NULL)
4919 		return (0);
4920 
4921 	KASSERT(ap->a_cookies,
4922 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
4923 
4924 	*ap->a_cookies = realloc(*ap->a_cookies,
4925 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
4926 	(*ap->a_cookies)[*ap->a_ncookies] = off;
4927 	return (0);
4928 }
4929 
4930 /*
4931  * Mark for update the access time of the file if the filesystem
4932  * supports VOP_MARKATIME.  This functionality is used by execve and
4933  * mmap, so we want to avoid the I/O implied by directly setting
4934  * va_atime for the sake of efficiency.
4935  */
4936 void
4937 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
4938 {
4939 	struct mount *mp;
4940 
4941 	mp = vp->v_mount;
4942 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
4943 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
4944 		(void)VOP_MARKATIME(vp);
4945 }
4946 
4947 /*
4948  * The purpose of this routine is to remove granularity from accmode_t,
4949  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
4950  * VADMIN and VAPPEND.
4951  *
4952  * If it returns 0, the caller is supposed to continue with the usual
4953  * access checks using 'accmode' as modified by this routine.  If it
4954  * returns nonzero value, the caller is supposed to return that value
4955  * as errno.
4956  *
4957  * Note that after this routine runs, accmode may be zero.
4958  */
4959 int
4960 vfs_unixify_accmode(accmode_t *accmode)
4961 {
4962 	/*
4963 	 * There is no way to specify explicit "deny" rule using
4964 	 * file mode or POSIX.1e ACLs.
4965 	 */
4966 	if (*accmode & VEXPLICIT_DENY) {
4967 		*accmode = 0;
4968 		return (0);
4969 	}
4970 
4971 	/*
4972 	 * None of these can be translated into usual access bits.
4973 	 * Also, the common case for NFSv4 ACLs is to not contain
4974 	 * either of these bits. Caller should check for VWRITE
4975 	 * on the containing directory instead.
4976 	 */
4977 	if (*accmode & (VDELETE_CHILD | VDELETE))
4978 		return (EPERM);
4979 
4980 	if (*accmode & VADMIN_PERMS) {
4981 		*accmode &= ~VADMIN_PERMS;
4982 		*accmode |= VADMIN;
4983 	}
4984 
4985 	/*
4986 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
4987 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
4988 	 */
4989 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
4990 
4991 	return (0);
4992 }
4993 
4994 /*
4995  * These are helper functions for filesystems to traverse all
4996  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
4997  *
4998  * This interface replaces MNT_VNODE_FOREACH.
4999  */
5000 
5001 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
5002 
5003 struct vnode *
5004 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
5005 {
5006 	struct vnode *vp;
5007 
5008 	if (should_yield())
5009 		kern_yield(PRI_USER);
5010 	MNT_ILOCK(mp);
5011 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5012 	vp = TAILQ_NEXT(*mvp, v_nmntvnodes);
5013 	while (vp != NULL && (vp->v_type == VMARKER ||
5014 	    (vp->v_iflag & VI_DOOMED) != 0))
5015 		vp = TAILQ_NEXT(vp, v_nmntvnodes);
5016 
5017 	/* Check if we are done */
5018 	if (vp == NULL) {
5019 		__mnt_vnode_markerfree_all(mvp, mp);
5020 		/* MNT_IUNLOCK(mp); -- done in above function */
5021 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
5022 		return (NULL);
5023 	}
5024 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5025 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5026 	VI_LOCK(vp);
5027 	MNT_IUNLOCK(mp);
5028 	return (vp);
5029 }
5030 
5031 struct vnode *
5032 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
5033 {
5034 	struct vnode *vp;
5035 
5036 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5037 	MNT_ILOCK(mp);
5038 	MNT_REF(mp);
5039 	(*mvp)->v_type = VMARKER;
5040 
5041 	vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
5042 	while (vp != NULL && (vp->v_type == VMARKER ||
5043 	    (vp->v_iflag & VI_DOOMED) != 0))
5044 		vp = TAILQ_NEXT(vp, v_nmntvnodes);
5045 
5046 	/* Check if we are done */
5047 	if (vp == NULL) {
5048 		MNT_REL(mp);
5049 		MNT_IUNLOCK(mp);
5050 		free(*mvp, M_VNODE_MARKER);
5051 		*mvp = NULL;
5052 		return (NULL);
5053 	}
5054 	(*mvp)->v_mount = mp;
5055 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5056 	VI_LOCK(vp);
5057 	MNT_IUNLOCK(mp);
5058 	return (vp);
5059 }
5060 
5061 
5062 void
5063 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
5064 {
5065 
5066 	if (*mvp == NULL) {
5067 		MNT_IUNLOCK(mp);
5068 		return;
5069 	}
5070 
5071 	mtx_assert(MNT_MTX(mp), MA_OWNED);
5072 
5073 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5074 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5075 	MNT_REL(mp);
5076 	MNT_IUNLOCK(mp);
5077 	free(*mvp, M_VNODE_MARKER);
5078 	*mvp = NULL;
5079 }
5080 
5081 /*
5082  * These are helper functions for filesystems to traverse their
5083  * active vnodes.  See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h
5084  */
5085 static void
5086 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5087 {
5088 
5089 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5090 
5091 	MNT_ILOCK(mp);
5092 	MNT_REL(mp);
5093 	MNT_IUNLOCK(mp);
5094 	free(*mvp, M_VNODE_MARKER);
5095 	*mvp = NULL;
5096 }
5097 
5098 static struct vnode *
5099 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5100 {
5101 	struct vnode *vp, *nvp;
5102 
5103 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
5104 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5105 restart:
5106 	vp = TAILQ_NEXT(*mvp, v_actfreelist);
5107 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5108 	while (vp != NULL) {
5109 		if (vp->v_type == VMARKER) {
5110 			vp = TAILQ_NEXT(vp, v_actfreelist);
5111 			continue;
5112 		}
5113 		if (!VI_TRYLOCK(vp)) {
5114 			if (mp_ncpus == 1 || should_yield()) {
5115 				TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5116 				mtx_unlock(&vnode_free_list_mtx);
5117 				pause("vnacti", 1);
5118 				mtx_lock(&vnode_free_list_mtx);
5119 				goto restart;
5120 			}
5121 			continue;
5122 		}
5123 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
5124 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
5125 		    ("alien vnode on the active list %p %p", vp, mp));
5126 		if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0)
5127 			break;
5128 		nvp = TAILQ_NEXT(vp, v_actfreelist);
5129 		VI_UNLOCK(vp);
5130 		vp = nvp;
5131 	}
5132 
5133 	/* Check if we are done */
5134 	if (vp == NULL) {
5135 		mtx_unlock(&vnode_free_list_mtx);
5136 		mnt_vnode_markerfree_active(mvp, mp);
5137 		return (NULL);
5138 	}
5139 	TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist);
5140 	mtx_unlock(&vnode_free_list_mtx);
5141 	ASSERT_VI_LOCKED(vp, "active iter");
5142 	KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp));
5143 	return (vp);
5144 }
5145 
5146 struct vnode *
5147 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5148 {
5149 
5150 	if (should_yield())
5151 		kern_yield(PRI_USER);
5152 	mtx_lock(&vnode_free_list_mtx);
5153 	return (mnt_vnode_next_active(mvp, mp));
5154 }
5155 
5156 struct vnode *
5157 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp)
5158 {
5159 	struct vnode *vp;
5160 
5161 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5162 	MNT_ILOCK(mp);
5163 	MNT_REF(mp);
5164 	MNT_IUNLOCK(mp);
5165 	(*mvp)->v_type = VMARKER;
5166 	(*mvp)->v_mount = mp;
5167 
5168 	mtx_lock(&vnode_free_list_mtx);
5169 	vp = TAILQ_FIRST(&mp->mnt_activevnodelist);
5170 	if (vp == NULL) {
5171 		mtx_unlock(&vnode_free_list_mtx);
5172 		mnt_vnode_markerfree_active(mvp, mp);
5173 		return (NULL);
5174 	}
5175 	TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5176 	return (mnt_vnode_next_active(mvp, mp));
5177 }
5178 
5179 void
5180 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5181 {
5182 
5183 	if (*mvp == NULL)
5184 		return;
5185 
5186 	mtx_lock(&vnode_free_list_mtx);
5187 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5188 	mtx_unlock(&vnode_free_list_mtx);
5189 	mnt_vnode_markerfree_active(mvp, mp);
5190 }
5191