xref: /freebsd/sys/kern/vfs_subr.c (revision 0bc2abddc8d4abb89a210f2bb113e9e7c2d4ce18)
1 /*-
2  * Copyright (c) 1989, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  * (c) UNIX System Laboratories, Inc.
5  * All or some portions of this file are derived from material licensed
6  * to the University of California by American Telephone and Telegraph
7  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8  * the permission of UNIX System Laboratories, Inc.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 4. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	@(#)vfs_subr.c	8.31 (Berkeley) 5/26/95
35  */
36 
37 /*
38  * External virtual filesystem routines
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include "opt_compat.h"
45 #include "opt_ddb.h"
46 #include "opt_watchdog.h"
47 
48 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/bio.h>
51 #include <sys/buf.h>
52 #include <sys/condvar.h>
53 #include <sys/conf.h>
54 #include <sys/dirent.h>
55 #include <sys/event.h>
56 #include <sys/eventhandler.h>
57 #include <sys/extattr.h>
58 #include <sys/file.h>
59 #include <sys/fcntl.h>
60 #include <sys/jail.h>
61 #include <sys/kdb.h>
62 #include <sys/kernel.h>
63 #include <sys/kthread.h>
64 #include <sys/lockf.h>
65 #include <sys/malloc.h>
66 #include <sys/mount.h>
67 #include <sys/namei.h>
68 #include <sys/pctrie.h>
69 #include <sys/priv.h>
70 #include <sys/reboot.h>
71 #include <sys/refcount.h>
72 #include <sys/rwlock.h>
73 #include <sys/sched.h>
74 #include <sys/sleepqueue.h>
75 #include <sys/smp.h>
76 #include <sys/stat.h>
77 #include <sys/sysctl.h>
78 #include <sys/syslog.h>
79 #include <sys/vmmeter.h>
80 #include <sys/vnode.h>
81 #include <sys/watchdog.h>
82 
83 #include <machine/stdarg.h>
84 
85 #include <security/mac/mac_framework.h>
86 
87 #include <vm/vm.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_extern.h>
90 #include <vm/pmap.h>
91 #include <vm/vm_map.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_kern.h>
94 #include <vm/uma.h>
95 
96 #ifdef DDB
97 #include <ddb/ddb.h>
98 #endif
99 
100 static void	delmntque(struct vnode *vp);
101 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
102 		    int slpflag, int slptimeo);
103 static void	syncer_shutdown(void *arg, int howto);
104 static int	vtryrecycle(struct vnode *vp);
105 static void	v_init_counters(struct vnode *);
106 static void	v_incr_usecount(struct vnode *);
107 static void	v_incr_usecount_locked(struct vnode *);
108 static void	v_incr_devcount(struct vnode *);
109 static void	v_decr_devcount(struct vnode *);
110 static void	vnlru_free(int);
111 static void	vgonel(struct vnode *);
112 static void	vfs_knllock(void *arg);
113 static void	vfs_knlunlock(void *arg);
114 static void	vfs_knl_assert_locked(void *arg);
115 static void	vfs_knl_assert_unlocked(void *arg);
116 static void	destroy_vpollinfo(struct vpollinfo *vi);
117 
118 /*
119  * Number of vnodes in existence.  Increased whenever getnewvnode()
120  * allocates a new vnode, decreased in vdropl() for VI_DOOMED vnode.
121  */
122 static unsigned long	numvnodes;
123 
124 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
125     "Number of vnodes in existence");
126 
127 static u_long vnodes_created;
128 SYSCTL_ULONG(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
129     0, "Number of vnodes created by getnewvnode");
130 
131 /*
132  * Conversion tables for conversion from vnode types to inode formats
133  * and back.
134  */
135 enum vtype iftovt_tab[16] = {
136 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
137 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
138 };
139 int vttoif_tab[10] = {
140 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
141 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
142 };
143 
144 /*
145  * List of vnodes that are ready for recycling.
146  */
147 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
148 
149 /*
150  * "Free" vnode target.  Free vnodes are rarely completely free, but are
151  * just ones that are cheap to recycle.  Usually they are for files which
152  * have been stat'd but not read; these usually have inode and namecache
153  * data attached to them.  This target is the preferred minimum size of a
154  * sub-cache consisting mostly of such files. The system balances the size
155  * of this sub-cache with its complement to try to prevent either from
156  * thrashing while the other is relatively inactive.  The targets express
157  * a preference for the best balance.
158  *
159  * "Above" this target there are 2 further targets (watermarks) related
160  * to recyling of free vnodes.  In the best-operating case, the cache is
161  * exactly full, the free list has size between vlowat and vhiwat above the
162  * free target, and recycling from it and normal use maintains this state.
163  * Sometimes the free list is below vlowat or even empty, but this state
164  * is even better for immediate use provided the cache is not full.
165  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
166  * ones) to reach one of these states.  The watermarks are currently hard-
167  * coded as 4% and 9% of the available space higher.  These and the default
168  * of 25% for wantfreevnodes are too large if the memory size is large.
169  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
170  * whenever vnlru_proc() becomes active.
171  */
172 static u_long wantfreevnodes;
173 SYSCTL_ULONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW,
174     &wantfreevnodes, 0, "Target for minimum number of \"free\" vnodes");
175 static u_long freevnodes;
176 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
177     &freevnodes, 0, "Number of \"free\" vnodes");
178 
179 static u_long recycles_count;
180 SYSCTL_ULONG(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count, 0,
181     "Number of vnodes recycled to meet vnode cache targets");
182 
183 /*
184  * Various variables used for debugging the new implementation of
185  * reassignbuf().
186  * XXX these are probably of (very) limited utility now.
187  */
188 static int reassignbufcalls;
189 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0,
190     "Number of calls to reassignbuf");
191 
192 static u_long free_owe_inact;
193 SYSCTL_ULONG(_vfs, OID_AUTO, free_owe_inact, CTLFLAG_RD, &free_owe_inact, 0,
194     "Number of times free vnodes kept on active list due to VFS "
195     "owing inactivation");
196 
197 /* To keep more than one thread at a time from running vfs_getnewfsid */
198 static struct mtx mntid_mtx;
199 
200 /*
201  * Lock for any access to the following:
202  *	vnode_free_list
203  *	numvnodes
204  *	freevnodes
205  */
206 static struct mtx vnode_free_list_mtx;
207 
208 /* Publicly exported FS */
209 struct nfs_public nfs_pub;
210 
211 static uma_zone_t buf_trie_zone;
212 
213 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
214 static uma_zone_t vnode_zone;
215 static uma_zone_t vnodepoll_zone;
216 
217 /*
218  * The workitem queue.
219  *
220  * It is useful to delay writes of file data and filesystem metadata
221  * for tens of seconds so that quickly created and deleted files need
222  * not waste disk bandwidth being created and removed. To realize this,
223  * we append vnodes to a "workitem" queue. When running with a soft
224  * updates implementation, most pending metadata dependencies should
225  * not wait for more than a few seconds. Thus, mounted on block devices
226  * are delayed only about a half the time that file data is delayed.
227  * Similarly, directory updates are more critical, so are only delayed
228  * about a third the time that file data is delayed. Thus, there are
229  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
230  * one each second (driven off the filesystem syncer process). The
231  * syncer_delayno variable indicates the next queue that is to be processed.
232  * Items that need to be processed soon are placed in this queue:
233  *
234  *	syncer_workitem_pending[syncer_delayno]
235  *
236  * A delay of fifteen seconds is done by placing the request fifteen
237  * entries later in the queue:
238  *
239  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
240  *
241  */
242 static int syncer_delayno;
243 static long syncer_mask;
244 LIST_HEAD(synclist, bufobj);
245 static struct synclist *syncer_workitem_pending;
246 /*
247  * The sync_mtx protects:
248  *	bo->bo_synclist
249  *	sync_vnode_count
250  *	syncer_delayno
251  *	syncer_state
252  *	syncer_workitem_pending
253  *	syncer_worklist_len
254  *	rushjob
255  */
256 static struct mtx sync_mtx;
257 static struct cv sync_wakeup;
258 
259 #define SYNCER_MAXDELAY		32
260 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
261 static int syncdelay = 30;		/* max time to delay syncing data */
262 static int filedelay = 30;		/* time to delay syncing files */
263 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
264     "Time to delay syncing files (in seconds)");
265 static int dirdelay = 29;		/* time to delay syncing directories */
266 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
267     "Time to delay syncing directories (in seconds)");
268 static int metadelay = 28;		/* time to delay syncing metadata */
269 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
270     "Time to delay syncing metadata (in seconds)");
271 static int rushjob;		/* number of slots to run ASAP */
272 static int stat_rush_requests;	/* number of times I/O speeded up */
273 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
274     "Number of times I/O speeded up (rush requests)");
275 
276 /*
277  * When shutting down the syncer, run it at four times normal speed.
278  */
279 #define SYNCER_SHUTDOWN_SPEEDUP		4
280 static int sync_vnode_count;
281 static int syncer_worklist_len;
282 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
283     syncer_state;
284 
285 /* Target for maximum number of vnodes. */
286 int desiredvnodes;
287 static int gapvnodes;		/* gap between wanted and desired */
288 static int vhiwat;		/* enough extras after expansion */
289 static int vlowat;		/* minimal extras before expansion */
290 static int vstir;		/* nonzero to stir non-free vnodes */
291 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
292 
293 static int
294 sysctl_update_desiredvnodes(SYSCTL_HANDLER_ARGS)
295 {
296 	int error, old_desiredvnodes;
297 
298 	old_desiredvnodes = desiredvnodes;
299 	if ((error = sysctl_handle_int(oidp, arg1, arg2, req)) != 0)
300 		return (error);
301 	if (old_desiredvnodes != desiredvnodes) {
302 		wantfreevnodes = desiredvnodes / 4;
303 		/* XXX locking seems to be incomplete. */
304 		vfs_hash_changesize(desiredvnodes);
305 		cache_changesize(desiredvnodes);
306 	}
307 	return (0);
308 }
309 
310 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
311     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, &desiredvnodes, 0,
312     sysctl_update_desiredvnodes, "I", "Target for maximum number of vnodes");
313 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
314     &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
315 static int vnlru_nowhere;
316 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
317     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
318 
319 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
320 static int vnsz2log;
321 
322 /*
323  * Support for the bufobj clean & dirty pctrie.
324  */
325 static void *
326 buf_trie_alloc(struct pctrie *ptree)
327 {
328 
329 	return uma_zalloc(buf_trie_zone, M_NOWAIT);
330 }
331 
332 static void
333 buf_trie_free(struct pctrie *ptree, void *node)
334 {
335 
336 	uma_zfree(buf_trie_zone, node);
337 }
338 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
339 
340 /*
341  * Initialize the vnode management data structures.
342  *
343  * Reevaluate the following cap on the number of vnodes after the physical
344  * memory size exceeds 512GB.  In the limit, as the physical memory size
345  * grows, the ratio of the memory size in KB to to vnodes approaches 64:1.
346  */
347 #ifndef	MAXVNODES_MAX
348 #define	MAXVNODES_MAX	(512 * 1024 * 1024 / 64)	/* 8M */
349 #endif
350 
351 /*
352  * Initialize a vnode as it first enters the zone.
353  */
354 static int
355 vnode_init(void *mem, int size, int flags)
356 {
357 	struct vnode *vp;
358 	struct bufobj *bo;
359 
360 	vp = mem;
361 	bzero(vp, size);
362 	/*
363 	 * Setup locks.
364 	 */
365 	vp->v_vnlock = &vp->v_lock;
366 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
367 	/*
368 	 * By default, don't allow shared locks unless filesystems opt-in.
369 	 */
370 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
371 	    LK_NOSHARE | LK_IS_VNODE);
372 	/*
373 	 * Initialize bufobj.
374 	 */
375 	bo = &vp->v_bufobj;
376 	bo->__bo_vnode = vp;
377 	rw_init(BO_LOCKPTR(bo), "bufobj interlock");
378 	bo->bo_private = vp;
379 	TAILQ_INIT(&bo->bo_clean.bv_hd);
380 	TAILQ_INIT(&bo->bo_dirty.bv_hd);
381 	/*
382 	 * Initialize namecache.
383 	 */
384 	LIST_INIT(&vp->v_cache_src);
385 	TAILQ_INIT(&vp->v_cache_dst);
386 	/*
387 	 * Initialize rangelocks.
388 	 */
389 	rangelock_init(&vp->v_rl);
390 	return (0);
391 }
392 
393 /*
394  * Free a vnode when it is cleared from the zone.
395  */
396 static void
397 vnode_fini(void *mem, int size)
398 {
399 	struct vnode *vp;
400 	struct bufobj *bo;
401 
402 	vp = mem;
403 	rangelock_destroy(&vp->v_rl);
404 	lockdestroy(vp->v_vnlock);
405 	mtx_destroy(&vp->v_interlock);
406 	bo = &vp->v_bufobj;
407 	rw_destroy(BO_LOCKPTR(bo));
408 }
409 
410 static void
411 vntblinit(void *dummy __unused)
412 {
413 	u_int i;
414 	int physvnodes, virtvnodes;
415 
416 	/*
417 	 * Desiredvnodes is a function of the physical memory size and the
418 	 * kernel's heap size.  Generally speaking, it scales with the
419 	 * physical memory size.  The ratio of desiredvnodes to the physical
420 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
421 	 * Thereafter, the
422 	 * marginal ratio of desiredvnodes to the physical memory size is
423 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
424 	 * size.  The memory required by desiredvnodes vnodes and vm objects
425 	 * must not exceed 1/7th of the kernel's heap size.
426 	 */
427 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
428 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
429 	virtvnodes = vm_kmem_size / (7 * (sizeof(struct vm_object) +
430 	    sizeof(struct vnode)));
431 	desiredvnodes = min(physvnodes, virtvnodes);
432 	if (desiredvnodes > MAXVNODES_MAX) {
433 		if (bootverbose)
434 			printf("Reducing kern.maxvnodes %d -> %d\n",
435 			    desiredvnodes, MAXVNODES_MAX);
436 		desiredvnodes = MAXVNODES_MAX;
437 	}
438 	wantfreevnodes = desiredvnodes / 4;
439 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
440 	TAILQ_INIT(&vnode_free_list);
441 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
442 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
443 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
444 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
445 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
446 	/*
447 	 * Preallocate enough nodes to support one-per buf so that
448 	 * we can not fail an insert.  reassignbuf() callers can not
449 	 * tolerate the insertion failure.
450 	 */
451 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
452 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
453 	    UMA_ZONE_NOFREE | UMA_ZONE_VM);
454 	uma_prealloc(buf_trie_zone, nbuf);
455 	/*
456 	 * Initialize the filesystem syncer.
457 	 */
458 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
459 	    &syncer_mask);
460 	syncer_maxdelay = syncer_mask + 1;
461 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
462 	cv_init(&sync_wakeup, "syncer");
463 	for (i = 1; i <= sizeof(struct vnode); i <<= 1)
464 		vnsz2log++;
465 	vnsz2log--;
466 }
467 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
468 
469 
470 /*
471  * Mark a mount point as busy. Used to synchronize access and to delay
472  * unmounting. Eventually, mountlist_mtx is not released on failure.
473  *
474  * vfs_busy() is a custom lock, it can block the caller.
475  * vfs_busy() only sleeps if the unmount is active on the mount point.
476  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
477  * vnode belonging to mp.
478  *
479  * Lookup uses vfs_busy() to traverse mount points.
480  * root fs			var fs
481  * / vnode lock		A	/ vnode lock (/var)		D
482  * /var vnode lock	B	/log vnode lock(/var/log)	E
483  * vfs_busy lock	C	vfs_busy lock			F
484  *
485  * Within each file system, the lock order is C->A->B and F->D->E.
486  *
487  * When traversing across mounts, the system follows that lock order:
488  *
489  *        C->A->B
490  *              |
491  *              +->F->D->E
492  *
493  * The lookup() process for namei("/var") illustrates the process:
494  *  VOP_LOOKUP() obtains B while A is held
495  *  vfs_busy() obtains a shared lock on F while A and B are held
496  *  vput() releases lock on B
497  *  vput() releases lock on A
498  *  VFS_ROOT() obtains lock on D while shared lock on F is held
499  *  vfs_unbusy() releases shared lock on F
500  *  vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
501  *    Attempt to lock A (instead of vp_crossmp) while D is held would
502  *    violate the global order, causing deadlocks.
503  *
504  * dounmount() locks B while F is drained.
505  */
506 int
507 vfs_busy(struct mount *mp, int flags)
508 {
509 
510 	MPASS((flags & ~MBF_MASK) == 0);
511 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
512 
513 	MNT_ILOCK(mp);
514 	MNT_REF(mp);
515 	/*
516 	 * If mount point is currenly being unmounted, sleep until the
517 	 * mount point fate is decided.  If thread doing the unmounting fails,
518 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
519 	 * that this mount point has survived the unmount attempt and vfs_busy
520 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
521 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
522 	 * about to be really destroyed.  vfs_busy needs to release its
523 	 * reference on the mount point in this case and return with ENOENT,
524 	 * telling the caller that mount mount it tried to busy is no longer
525 	 * valid.
526 	 */
527 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
528 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
529 			MNT_REL(mp);
530 			MNT_IUNLOCK(mp);
531 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
532 			    __func__);
533 			return (ENOENT);
534 		}
535 		if (flags & MBF_MNTLSTLOCK)
536 			mtx_unlock(&mountlist_mtx);
537 		mp->mnt_kern_flag |= MNTK_MWAIT;
538 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
539 		if (flags & MBF_MNTLSTLOCK)
540 			mtx_lock(&mountlist_mtx);
541 		MNT_ILOCK(mp);
542 	}
543 	if (flags & MBF_MNTLSTLOCK)
544 		mtx_unlock(&mountlist_mtx);
545 	mp->mnt_lockref++;
546 	MNT_IUNLOCK(mp);
547 	return (0);
548 }
549 
550 /*
551  * Free a busy filesystem.
552  */
553 void
554 vfs_unbusy(struct mount *mp)
555 {
556 
557 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
558 	MNT_ILOCK(mp);
559 	MNT_REL(mp);
560 	KASSERT(mp->mnt_lockref > 0, ("negative mnt_lockref"));
561 	mp->mnt_lockref--;
562 	if (mp->mnt_lockref == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
563 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
564 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
565 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
566 		wakeup(&mp->mnt_lockref);
567 	}
568 	MNT_IUNLOCK(mp);
569 }
570 
571 /*
572  * Lookup a mount point by filesystem identifier.
573  */
574 struct mount *
575 vfs_getvfs(fsid_t *fsid)
576 {
577 	struct mount *mp;
578 
579 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
580 	mtx_lock(&mountlist_mtx);
581 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
582 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
583 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
584 			vfs_ref(mp);
585 			mtx_unlock(&mountlist_mtx);
586 			return (mp);
587 		}
588 	}
589 	mtx_unlock(&mountlist_mtx);
590 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
591 	return ((struct mount *) 0);
592 }
593 
594 /*
595  * Lookup a mount point by filesystem identifier, busying it before
596  * returning.
597  *
598  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
599  * cache for popular filesystem identifiers.  The cache is lockess, using
600  * the fact that struct mount's are never freed.  In worst case we may
601  * get pointer to unmounted or even different filesystem, so we have to
602  * check what we got, and go slow way if so.
603  */
604 struct mount *
605 vfs_busyfs(fsid_t *fsid)
606 {
607 #define	FSID_CACHE_SIZE	256
608 	typedef struct mount * volatile vmp_t;
609 	static vmp_t cache[FSID_CACHE_SIZE];
610 	struct mount *mp;
611 	int error;
612 	uint32_t hash;
613 
614 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
615 	hash = fsid->val[0] ^ fsid->val[1];
616 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
617 	mp = cache[hash];
618 	if (mp == NULL ||
619 	    mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
620 	    mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
621 		goto slow;
622 	if (vfs_busy(mp, 0) != 0) {
623 		cache[hash] = NULL;
624 		goto slow;
625 	}
626 	if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
627 	    mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
628 		return (mp);
629 	else
630 	    vfs_unbusy(mp);
631 
632 slow:
633 	mtx_lock(&mountlist_mtx);
634 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
635 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
636 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
637 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
638 			if (error) {
639 				cache[hash] = NULL;
640 				mtx_unlock(&mountlist_mtx);
641 				return (NULL);
642 			}
643 			cache[hash] = mp;
644 			return (mp);
645 		}
646 	}
647 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
648 	mtx_unlock(&mountlist_mtx);
649 	return ((struct mount *) 0);
650 }
651 
652 /*
653  * Check if a user can access privileged mount options.
654  */
655 int
656 vfs_suser(struct mount *mp, struct thread *td)
657 {
658 	int error;
659 
660 	/*
661 	 * If the thread is jailed, but this is not a jail-friendly file
662 	 * system, deny immediately.
663 	 */
664 	if (!(mp->mnt_vfc->vfc_flags & VFCF_JAIL) && jailed(td->td_ucred))
665 		return (EPERM);
666 
667 	/*
668 	 * If the file system was mounted outside the jail of the calling
669 	 * thread, deny immediately.
670 	 */
671 	if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
672 		return (EPERM);
673 
674 	/*
675 	 * If file system supports delegated administration, we don't check
676 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
677 	 * by the file system itself.
678 	 * If this is not the user that did original mount, we check for
679 	 * the PRIV_VFS_MOUNT_OWNER privilege.
680 	 */
681 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
682 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
683 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
684 			return (error);
685 	}
686 	return (0);
687 }
688 
689 /*
690  * Get a new unique fsid.  Try to make its val[0] unique, since this value
691  * will be used to create fake device numbers for stat().  Also try (but
692  * not so hard) make its val[0] unique mod 2^16, since some emulators only
693  * support 16-bit device numbers.  We end up with unique val[0]'s for the
694  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
695  *
696  * Keep in mind that several mounts may be running in parallel.  Starting
697  * the search one past where the previous search terminated is both a
698  * micro-optimization and a defense against returning the same fsid to
699  * different mounts.
700  */
701 void
702 vfs_getnewfsid(struct mount *mp)
703 {
704 	static uint16_t mntid_base;
705 	struct mount *nmp;
706 	fsid_t tfsid;
707 	int mtype;
708 
709 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
710 	mtx_lock(&mntid_mtx);
711 	mtype = mp->mnt_vfc->vfc_typenum;
712 	tfsid.val[1] = mtype;
713 	mtype = (mtype & 0xFF) << 24;
714 	for (;;) {
715 		tfsid.val[0] = makedev(255,
716 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
717 		mntid_base++;
718 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
719 			break;
720 		vfs_rel(nmp);
721 	}
722 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
723 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
724 	mtx_unlock(&mntid_mtx);
725 }
726 
727 /*
728  * Knob to control the precision of file timestamps:
729  *
730  *   0 = seconds only; nanoseconds zeroed.
731  *   1 = seconds and nanoseconds, accurate within 1/HZ.
732  *   2 = seconds and nanoseconds, truncated to microseconds.
733  * >=3 = seconds and nanoseconds, maximum precision.
734  */
735 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
736 
737 static int timestamp_precision = TSP_USEC;
738 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
739     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
740     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to ms, "
741     "3+: sec + ns (max. precision))");
742 
743 /*
744  * Get a current timestamp.
745  */
746 void
747 vfs_timestamp(struct timespec *tsp)
748 {
749 	struct timeval tv;
750 
751 	switch (timestamp_precision) {
752 	case TSP_SEC:
753 		tsp->tv_sec = time_second;
754 		tsp->tv_nsec = 0;
755 		break;
756 	case TSP_HZ:
757 		getnanotime(tsp);
758 		break;
759 	case TSP_USEC:
760 		microtime(&tv);
761 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
762 		break;
763 	case TSP_NSEC:
764 	default:
765 		nanotime(tsp);
766 		break;
767 	}
768 }
769 
770 /*
771  * Set vnode attributes to VNOVAL
772  */
773 void
774 vattr_null(struct vattr *vap)
775 {
776 
777 	vap->va_type = VNON;
778 	vap->va_size = VNOVAL;
779 	vap->va_bytes = VNOVAL;
780 	vap->va_mode = VNOVAL;
781 	vap->va_nlink = VNOVAL;
782 	vap->va_uid = VNOVAL;
783 	vap->va_gid = VNOVAL;
784 	vap->va_fsid = VNOVAL;
785 	vap->va_fileid = VNOVAL;
786 	vap->va_blocksize = VNOVAL;
787 	vap->va_rdev = VNOVAL;
788 	vap->va_atime.tv_sec = VNOVAL;
789 	vap->va_atime.tv_nsec = VNOVAL;
790 	vap->va_mtime.tv_sec = VNOVAL;
791 	vap->va_mtime.tv_nsec = VNOVAL;
792 	vap->va_ctime.tv_sec = VNOVAL;
793 	vap->va_ctime.tv_nsec = VNOVAL;
794 	vap->va_birthtime.tv_sec = VNOVAL;
795 	vap->va_birthtime.tv_nsec = VNOVAL;
796 	vap->va_flags = VNOVAL;
797 	vap->va_gen = VNOVAL;
798 	vap->va_vaflags = 0;
799 }
800 
801 /*
802  * This routine is called when we have too many vnodes.  It attempts
803  * to free <count> vnodes and will potentially free vnodes that still
804  * have VM backing store (VM backing store is typically the cause
805  * of a vnode blowout so we want to do this).  Therefore, this operation
806  * is not considered cheap.
807  *
808  * A number of conditions may prevent a vnode from being reclaimed.
809  * the buffer cache may have references on the vnode, a directory
810  * vnode may still have references due to the namei cache representing
811  * underlying files, or the vnode may be in active use.   It is not
812  * desireable to reuse such vnodes.  These conditions may cause the
813  * number of vnodes to reach some minimum value regardless of what
814  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
815  */
816 static int
817 vlrureclaim(struct mount *mp, int reclaim_nc_src, int trigger)
818 {
819 	struct vnode *vp;
820 	int count, done, target;
821 
822 	done = 0;
823 	vn_start_write(NULL, &mp, V_WAIT);
824 	MNT_ILOCK(mp);
825 	count = mp->mnt_nvnodelistsize;
826 	target = count * (int64_t)gapvnodes / imax(desiredvnodes, 1);
827 	target = target / 10 + 1;
828 	while (count != 0 && done < target) {
829 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
830 		while (vp != NULL && vp->v_type == VMARKER)
831 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
832 		if (vp == NULL)
833 			break;
834 		/*
835 		 * XXX LRU is completely broken for non-free vnodes.  First
836 		 * by calling here in mountpoint order, then by moving
837 		 * unselected vnodes to the end here, and most grossly by
838 		 * removing the vlruvp() function that was supposed to
839 		 * maintain the order.  (This function was born broken
840 		 * since syncer problems prevented it doing anything.)  The
841 		 * order is closer to LRC (C = Created).
842 		 *
843 		 * LRU reclaiming of vnodes seems to have last worked in
844 		 * FreeBSD-3 where LRU wasn't mentioned under any spelling.
845 		 * Then there was no hold count, and inactive vnodes were
846 		 * simply put on the free list in LRU order.  The separate
847 		 * lists also break LRU.  We prefer to reclaim from the
848 		 * free list for technical reasons.  This tends to thrash
849 		 * the free list to keep very unrecently used held vnodes.
850 		 * The problem is mitigated by keeping the free list large.
851 		 */
852 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
853 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
854 		--count;
855 		if (!VI_TRYLOCK(vp))
856 			goto next_iter;
857 		/*
858 		 * If it's been deconstructed already, it's still
859 		 * referenced, or it exceeds the trigger, skip it.
860 		 * Also skip free vnodes.  We are trying to make space
861 		 * to expand the free list, not reduce it.
862 		 */
863 		if (vp->v_usecount ||
864 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
865 		    ((vp->v_iflag & VI_FREE) != 0) ||
866 		    (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
867 		    vp->v_object->resident_page_count > trigger)) {
868 			VI_UNLOCK(vp);
869 			goto next_iter;
870 		}
871 		MNT_IUNLOCK(mp);
872 		vholdl(vp);
873 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) {
874 			vdrop(vp);
875 			goto next_iter_mntunlocked;
876 		}
877 		VI_LOCK(vp);
878 		/*
879 		 * v_usecount may have been bumped after VOP_LOCK() dropped
880 		 * the vnode interlock and before it was locked again.
881 		 *
882 		 * It is not necessary to recheck VI_DOOMED because it can
883 		 * only be set by another thread that holds both the vnode
884 		 * lock and vnode interlock.  If another thread has the
885 		 * vnode lock before we get to VOP_LOCK() and obtains the
886 		 * vnode interlock after VOP_LOCK() drops the vnode
887 		 * interlock, the other thread will be unable to drop the
888 		 * vnode lock before our VOP_LOCK() call fails.
889 		 */
890 		if (vp->v_usecount ||
891 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
892 		    (vp->v_iflag & VI_FREE) != 0 ||
893 		    (vp->v_object != NULL &&
894 		    vp->v_object->resident_page_count > trigger)) {
895 			VOP_UNLOCK(vp, LK_INTERLOCK);
896 			vdrop(vp);
897 			goto next_iter_mntunlocked;
898 		}
899 		KASSERT((vp->v_iflag & VI_DOOMED) == 0,
900 		    ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
901 		atomic_add_long(&recycles_count, 1);
902 		vgonel(vp);
903 		VOP_UNLOCK(vp, 0);
904 		vdropl(vp);
905 		done++;
906 next_iter_mntunlocked:
907 		if (!should_yield())
908 			goto relock_mnt;
909 		goto yield;
910 next_iter:
911 		if (!should_yield())
912 			continue;
913 		MNT_IUNLOCK(mp);
914 yield:
915 		kern_yield(PRI_USER);
916 relock_mnt:
917 		MNT_ILOCK(mp);
918 	}
919 	MNT_IUNLOCK(mp);
920 	vn_finished_write(mp);
921 	return done;
922 }
923 
924 /*
925  * Attempt to reduce the free list by the requested amount.
926  */
927 static void
928 vnlru_free(int count)
929 {
930 	struct vnode *vp;
931 
932 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
933 	for (; count > 0; count--) {
934 		vp = TAILQ_FIRST(&vnode_free_list);
935 		/*
936 		 * The list can be modified while the free_list_mtx
937 		 * has been dropped and vp could be NULL here.
938 		 */
939 		if (!vp)
940 			break;
941 		VNASSERT(vp->v_op != NULL, vp,
942 		    ("vnlru_free: vnode already reclaimed."));
943 		KASSERT((vp->v_iflag & VI_FREE) != 0,
944 		    ("Removing vnode not on freelist"));
945 		KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
946 		    ("Mangling active vnode"));
947 		TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
948 		/*
949 		 * Don't recycle if we can't get the interlock.
950 		 */
951 		if (!VI_TRYLOCK(vp)) {
952 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist);
953 			continue;
954 		}
955 		VNASSERT((vp->v_iflag & VI_FREE) != 0 && vp->v_holdcnt == 0,
956 		    vp, ("vp inconsistent on freelist"));
957 
958 		/*
959 		 * The clear of VI_FREE prevents activation of the
960 		 * vnode.  There is no sense in putting the vnode on
961 		 * the mount point active list, only to remove it
962 		 * later during recycling.  Inline the relevant part
963 		 * of vholdl(), to avoid triggering assertions or
964 		 * activating.
965 		 */
966 		freevnodes--;
967 		vp->v_iflag &= ~VI_FREE;
968 		refcount_acquire(&vp->v_holdcnt);
969 
970 		mtx_unlock(&vnode_free_list_mtx);
971 		VI_UNLOCK(vp);
972 		vtryrecycle(vp);
973 		/*
974 		 * If the recycled succeeded this vdrop will actually free
975 		 * the vnode.  If not it will simply place it back on
976 		 * the free list.
977 		 */
978 		vdrop(vp);
979 		mtx_lock(&vnode_free_list_mtx);
980 	}
981 }
982 
983 /* XXX some names and initialization are bad for limits and watermarks. */
984 static int
985 vspace(void)
986 {
987 	int space;
988 
989 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
990 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
991 	vlowat = vhiwat / 2;
992 	if (numvnodes > desiredvnodes)
993 		return (0);
994 	space = desiredvnodes - numvnodes;
995 	if (freevnodes > wantfreevnodes)
996 		space += freevnodes - wantfreevnodes;
997 	return (space);
998 }
999 
1000 /*
1001  * Attempt to recycle vnodes in a context that is always safe to block.
1002  * Calling vlrurecycle() from the bowels of filesystem code has some
1003  * interesting deadlock problems.
1004  */
1005 static struct proc *vnlruproc;
1006 static int vnlruproc_sig;
1007 
1008 static void
1009 vnlru_proc(void)
1010 {
1011 	struct mount *mp, *nmp;
1012 	unsigned long ofreevnodes, onumvnodes;
1013 	int done, force, reclaim_nc_src, trigger, usevnodes;
1014 
1015 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1016 	    SHUTDOWN_PRI_FIRST);
1017 
1018 	force = 0;
1019 	for (;;) {
1020 		kproc_suspend_check(vnlruproc);
1021 		mtx_lock(&vnode_free_list_mtx);
1022 		/*
1023 		 * If numvnodes is too large (due to desiredvnodes being
1024 		 * adjusted using its sysctl, or emergency growth), first
1025 		 * try to reduce it by discarding from the free list.
1026 		 */
1027 		if (numvnodes > desiredvnodes && freevnodes > 0)
1028 			vnlru_free(ulmin(numvnodes - desiredvnodes,
1029 			    freevnodes));
1030 		/*
1031 		 * Sleep if the vnode cache is in a good state.  This is
1032 		 * when it is not over-full and has space for about a 4%
1033 		 * or 9% expansion (by growing its size or inexcessively
1034 		 * reducing its free list).  Otherwise, try to reclaim
1035 		 * space for a 10% expansion.
1036 		 */
1037 		if (vstir && force == 0) {
1038 			force = 1;
1039 			vstir = 0;
1040 		}
1041 		if (vspace() >= vlowat && force == 0) {
1042 			vnlruproc_sig = 0;
1043 			wakeup(&vnlruproc_sig);
1044 			msleep(vnlruproc, &vnode_free_list_mtx,
1045 			    PVFS|PDROP, "vlruwt", hz);
1046 			continue;
1047 		}
1048 		mtx_unlock(&vnode_free_list_mtx);
1049 		done = 0;
1050 		ofreevnodes = freevnodes;
1051 		onumvnodes = numvnodes;
1052 		/*
1053 		 * Calculate parameters for recycling.  These are the same
1054 		 * throughout the loop to give some semblance of fairness.
1055 		 * The trigger point is to avoid recycling vnodes with lots
1056 		 * of resident pages.  We aren't trying to free memory; we
1057 		 * are trying to recycle or at least free vnodes.
1058 		 */
1059 		if (numvnodes <= desiredvnodes)
1060 			usevnodes = numvnodes - freevnodes;
1061 		else
1062 			usevnodes = numvnodes;
1063 		if (usevnodes <= 0)
1064 			usevnodes = 1;
1065 		/*
1066 		 * The trigger value is is chosen to give a conservatively
1067 		 * large value to ensure that it alone doesn't prevent
1068 		 * making progress.  The value can easily be so large that
1069 		 * it is effectively infinite in some congested and
1070 		 * misconfigured cases, and this is necessary.  Normally
1071 		 * it is about 8 to 100 (pages), which is quite large.
1072 		 */
1073 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1074 		if (force < 2)
1075 			trigger = vsmalltrigger;
1076 		reclaim_nc_src = force >= 3;
1077 		mtx_lock(&mountlist_mtx);
1078 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1079 			if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) {
1080 				nmp = TAILQ_NEXT(mp, mnt_list);
1081 				continue;
1082 			}
1083 			done += vlrureclaim(mp, reclaim_nc_src, trigger);
1084 			mtx_lock(&mountlist_mtx);
1085 			nmp = TAILQ_NEXT(mp, mnt_list);
1086 			vfs_unbusy(mp);
1087 		}
1088 		mtx_unlock(&mountlist_mtx);
1089 		if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1090 			uma_reclaim();
1091 		if (done == 0) {
1092 			if (force == 0 || force == 1) {
1093 				force = 2;
1094 				continue;
1095 			}
1096 			if (force == 2) {
1097 				force = 3;
1098 				continue;
1099 			}
1100 			force = 0;
1101 			vnlru_nowhere++;
1102 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1103 		} else
1104 			kern_yield(PRI_USER);
1105 		/*
1106 		 * After becoming active to expand above low water, keep
1107 		 * active until above high water.
1108 		 */
1109 		force = vspace() < vhiwat;
1110 	}
1111 }
1112 
1113 static struct kproc_desc vnlru_kp = {
1114 	"vnlru",
1115 	vnlru_proc,
1116 	&vnlruproc
1117 };
1118 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1119     &vnlru_kp);
1120 
1121 /*
1122  * Routines having to do with the management of the vnode table.
1123  */
1124 
1125 /*
1126  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
1127  * before we actually vgone().  This function must be called with the vnode
1128  * held to prevent the vnode from being returned to the free list midway
1129  * through vgone().
1130  */
1131 static int
1132 vtryrecycle(struct vnode *vp)
1133 {
1134 	struct mount *vnmp;
1135 
1136 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1137 	VNASSERT(vp->v_holdcnt, vp,
1138 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
1139 	/*
1140 	 * This vnode may found and locked via some other list, if so we
1141 	 * can't recycle it yet.
1142 	 */
1143 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1144 		CTR2(KTR_VFS,
1145 		    "%s: impossible to recycle, vp %p lock is already held",
1146 		    __func__, vp);
1147 		return (EWOULDBLOCK);
1148 	}
1149 	/*
1150 	 * Don't recycle if its filesystem is being suspended.
1151 	 */
1152 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1153 		VOP_UNLOCK(vp, 0);
1154 		CTR2(KTR_VFS,
1155 		    "%s: impossible to recycle, cannot start the write for %p",
1156 		    __func__, vp);
1157 		return (EBUSY);
1158 	}
1159 	/*
1160 	 * If we got this far, we need to acquire the interlock and see if
1161 	 * anyone picked up this vnode from another list.  If not, we will
1162 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1163 	 * will skip over it.
1164 	 */
1165 	VI_LOCK(vp);
1166 	if (vp->v_usecount) {
1167 		VOP_UNLOCK(vp, LK_INTERLOCK);
1168 		vn_finished_write(vnmp);
1169 		CTR2(KTR_VFS,
1170 		    "%s: impossible to recycle, %p is already referenced",
1171 		    __func__, vp);
1172 		return (EBUSY);
1173 	}
1174 	if ((vp->v_iflag & VI_DOOMED) == 0) {
1175 		atomic_add_long(&recycles_count, 1);
1176 		vgonel(vp);
1177 	}
1178 	VOP_UNLOCK(vp, LK_INTERLOCK);
1179 	vn_finished_write(vnmp);
1180 	return (0);
1181 }
1182 
1183 static void
1184 vcheckspace(void)
1185 {
1186 
1187 	if (vspace() < vlowat && vnlruproc_sig == 0) {
1188 		vnlruproc_sig = 1;
1189 		wakeup(vnlruproc);
1190 	}
1191 }
1192 
1193 /*
1194  * Wait if necessary for space for a new vnode.
1195  */
1196 static int
1197 getnewvnode_wait(int suspended)
1198 {
1199 
1200 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1201 	if (numvnodes >= desiredvnodes) {
1202 		if (suspended) {
1203 			/*
1204 			 * The file system is being suspended.  We cannot
1205 			 * risk a deadlock here, so allow allocation of
1206 			 * another vnode even if this would give too many.
1207 			 */
1208 			return (0);
1209 		}
1210 		if (vnlruproc_sig == 0) {
1211 			vnlruproc_sig = 1;	/* avoid unnecessary wakeups */
1212 			wakeup(vnlruproc);
1213 		}
1214 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
1215 		    "vlruwk", hz);
1216 	}
1217 	/* Post-adjust like the pre-adjust in getnewvnode(). */
1218 	if (numvnodes + 1 > desiredvnodes && freevnodes > 1)
1219 		vnlru_free(1);
1220 	return (numvnodes >= desiredvnodes ? ENFILE : 0);
1221 }
1222 
1223 /*
1224  * This hack is fragile, and probably not needed any more now that the
1225  * watermark handling works.
1226  */
1227 void
1228 getnewvnode_reserve(u_int count)
1229 {
1230 	struct thread *td;
1231 
1232 	/* Pre-adjust like the pre-adjust in getnewvnode(), with any count. */
1233 	/* XXX no longer so quick, but this part is not racy. */
1234 	mtx_lock(&vnode_free_list_mtx);
1235 	if (numvnodes + count > desiredvnodes && freevnodes > wantfreevnodes)
1236 		vnlru_free(ulmin(numvnodes + count - desiredvnodes,
1237 		    freevnodes - wantfreevnodes));
1238 	mtx_unlock(&vnode_free_list_mtx);
1239 
1240 	td = curthread;
1241 	/* First try to be quick and racy. */
1242 	if (atomic_fetchadd_long(&numvnodes, count) + count <= desiredvnodes) {
1243 		td->td_vp_reserv += count;
1244 		vcheckspace();	/* XXX no longer so quick, but more racy */
1245 		return;
1246 	} else
1247 		atomic_subtract_long(&numvnodes, count);
1248 
1249 	mtx_lock(&vnode_free_list_mtx);
1250 	while (count > 0) {
1251 		if (getnewvnode_wait(0) == 0) {
1252 			count--;
1253 			td->td_vp_reserv++;
1254 			atomic_add_long(&numvnodes, 1);
1255 		}
1256 	}
1257 	vcheckspace();
1258 	mtx_unlock(&vnode_free_list_mtx);
1259 }
1260 
1261 /*
1262  * This hack is fragile, especially if desiredvnodes or wantvnodes are
1263  * misconfgured or changed significantly.  Reducing desiredvnodes below
1264  * the reserved amount should cause bizarre behaviour like reducing it
1265  * below the number of active vnodes -- the system will try to reduce
1266  * numvnodes to match, but should fail, so the subtraction below should
1267  * not overflow.
1268  */
1269 void
1270 getnewvnode_drop_reserve(void)
1271 {
1272 	struct thread *td;
1273 
1274 	td = curthread;
1275 	atomic_subtract_long(&numvnodes, td->td_vp_reserv);
1276 	td->td_vp_reserv = 0;
1277 }
1278 
1279 /*
1280  * Return the next vnode from the free list.
1281  */
1282 int
1283 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1284     struct vnode **vpp)
1285 {
1286 	struct vnode *vp;
1287 	struct thread *td;
1288 	struct lock_object *lo;
1289 	static int cyclecount;
1290 	int error;
1291 
1292 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1293 	vp = NULL;
1294 	td = curthread;
1295 	if (td->td_vp_reserv > 0) {
1296 		td->td_vp_reserv -= 1;
1297 		goto alloc;
1298 	}
1299 	mtx_lock(&vnode_free_list_mtx);
1300 	if (numvnodes < desiredvnodes)
1301 		cyclecount = 0;
1302 	else if (cyclecount++ >= freevnodes) {
1303 		cyclecount = 0;
1304 		vstir = 1;
1305 	}
1306 	/*
1307 	 * Grow the vnode cache if it will not be above its target max
1308 	 * after growing.  Otherwise, if the free list is nonempty, try
1309 	 * to reclaim 1 item from it before growing the cache (possibly
1310 	 * above its target max if the reclamation failed or is delayed).
1311 	 * Otherwise, wait for some space.  In all cases, schedule
1312 	 * vnlru_proc() if we are getting short of space.  The watermarks
1313 	 * should be chosen so that we never wait or even reclaim from
1314 	 * the free list to below its target minimum.
1315 	 */
1316 	if (numvnodes + 1 <= desiredvnodes)
1317 		;
1318 	else if (freevnodes > 0)
1319 		vnlru_free(1);
1320 	else {
1321 		error = getnewvnode_wait(mp != NULL && (mp->mnt_kern_flag &
1322 		    MNTK_SUSPEND));
1323 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
1324 		if (error != 0) {
1325 			mtx_unlock(&vnode_free_list_mtx);
1326 			return (error);
1327 		}
1328 #endif
1329 	}
1330 	vcheckspace();
1331 	atomic_add_long(&numvnodes, 1);
1332 	mtx_unlock(&vnode_free_list_mtx);
1333 alloc:
1334 	atomic_add_long(&vnodes_created, 1);
1335 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK);
1336 	/*
1337 	 * Locks are given the generic name "vnode" when created.
1338 	 * Follow the historic practice of using the filesystem
1339 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1340 	 *
1341 	 * Locks live in a witness group keyed on their name. Thus,
1342 	 * when a lock is renamed, it must also move from the witness
1343 	 * group of its old name to the witness group of its new name.
1344 	 *
1345 	 * The change only needs to be made when the vnode moves
1346 	 * from one filesystem type to another. We ensure that each
1347 	 * filesystem use a single static name pointer for its tag so
1348 	 * that we can compare pointers rather than doing a strcmp().
1349 	 */
1350 	lo = &vp->v_vnlock->lock_object;
1351 	if (lo->lo_name != tag) {
1352 		lo->lo_name = tag;
1353 		WITNESS_DESTROY(lo);
1354 		WITNESS_INIT(lo, tag);
1355 	}
1356 	/*
1357 	 * By default, don't allow shared locks unless filesystems opt-in.
1358 	 */
1359 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1360 	/*
1361 	 * Finalize various vnode identity bits.
1362 	 */
1363 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1364 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1365 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1366 	vp->v_type = VNON;
1367 	vp->v_tag = tag;
1368 	vp->v_op = vops;
1369 	v_init_counters(vp);
1370 	vp->v_bufobj.bo_ops = &buf_ops_bio;
1371 #ifdef MAC
1372 	mac_vnode_init(vp);
1373 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1374 		mac_vnode_associate_singlelabel(mp, vp);
1375 	else if (mp == NULL && vops != &dead_vnodeops)
1376 		printf("NULL mp in getnewvnode()\n");
1377 #endif
1378 	if (mp != NULL) {
1379 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1380 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1381 			vp->v_vflag |= VV_NOKNOTE;
1382 	}
1383 
1384 	/*
1385 	 * For the filesystems which do not use vfs_hash_insert(),
1386 	 * still initialize v_hash to have vfs_hash_index() useful.
1387 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1388 	 * its own hashing.
1389 	 */
1390 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
1391 
1392 	*vpp = vp;
1393 	return (0);
1394 }
1395 
1396 /*
1397  * Delete from old mount point vnode list, if on one.
1398  */
1399 static void
1400 delmntque(struct vnode *vp)
1401 {
1402 	struct mount *mp;
1403 	int active;
1404 
1405 	mp = vp->v_mount;
1406 	if (mp == NULL)
1407 		return;
1408 	MNT_ILOCK(mp);
1409 	VI_LOCK(vp);
1410 	KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize,
1411 	    ("Active vnode list size %d > Vnode list size %d",
1412 	     mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize));
1413 	active = vp->v_iflag & VI_ACTIVE;
1414 	vp->v_iflag &= ~VI_ACTIVE;
1415 	if (active) {
1416 		mtx_lock(&vnode_free_list_mtx);
1417 		TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist);
1418 		mp->mnt_activevnodelistsize--;
1419 		mtx_unlock(&vnode_free_list_mtx);
1420 	}
1421 	vp->v_mount = NULL;
1422 	VI_UNLOCK(vp);
1423 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1424 		("bad mount point vnode list size"));
1425 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1426 	mp->mnt_nvnodelistsize--;
1427 	MNT_REL(mp);
1428 	MNT_IUNLOCK(mp);
1429 }
1430 
1431 static void
1432 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1433 {
1434 
1435 	vp->v_data = NULL;
1436 	vp->v_op = &dead_vnodeops;
1437 	vgone(vp);
1438 	vput(vp);
1439 }
1440 
1441 /*
1442  * Insert into list of vnodes for the new mount point, if available.
1443  */
1444 int
1445 insmntque1(struct vnode *vp, struct mount *mp,
1446 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1447 {
1448 
1449 	KASSERT(vp->v_mount == NULL,
1450 		("insmntque: vnode already on per mount vnode list"));
1451 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1452 	ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1453 
1454 	/*
1455 	 * We acquire the vnode interlock early to ensure that the
1456 	 * vnode cannot be recycled by another process releasing a
1457 	 * holdcnt on it before we get it on both the vnode list
1458 	 * and the active vnode list. The mount mutex protects only
1459 	 * manipulation of the vnode list and the vnode freelist
1460 	 * mutex protects only manipulation of the active vnode list.
1461 	 * Hence the need to hold the vnode interlock throughout.
1462 	 */
1463 	MNT_ILOCK(mp);
1464 	VI_LOCK(vp);
1465 	if (((mp->mnt_kern_flag & MNTK_NOINSMNTQ) != 0 &&
1466 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1467 	    mp->mnt_nvnodelistsize == 0)) &&
1468 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1469 		VI_UNLOCK(vp);
1470 		MNT_IUNLOCK(mp);
1471 		if (dtr != NULL)
1472 			dtr(vp, dtr_arg);
1473 		return (EBUSY);
1474 	}
1475 	vp->v_mount = mp;
1476 	MNT_REF(mp);
1477 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1478 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1479 		("neg mount point vnode list size"));
1480 	mp->mnt_nvnodelistsize++;
1481 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1482 	    ("Activating already active vnode"));
1483 	vp->v_iflag |= VI_ACTIVE;
1484 	mtx_lock(&vnode_free_list_mtx);
1485 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
1486 	mp->mnt_activevnodelistsize++;
1487 	mtx_unlock(&vnode_free_list_mtx);
1488 	VI_UNLOCK(vp);
1489 	MNT_IUNLOCK(mp);
1490 	return (0);
1491 }
1492 
1493 int
1494 insmntque(struct vnode *vp, struct mount *mp)
1495 {
1496 
1497 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1498 }
1499 
1500 /*
1501  * Flush out and invalidate all buffers associated with a bufobj
1502  * Called with the underlying object locked.
1503  */
1504 int
1505 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1506 {
1507 	int error;
1508 
1509 	BO_LOCK(bo);
1510 	if (flags & V_SAVE) {
1511 		error = bufobj_wwait(bo, slpflag, slptimeo);
1512 		if (error) {
1513 			BO_UNLOCK(bo);
1514 			return (error);
1515 		}
1516 		if (bo->bo_dirty.bv_cnt > 0) {
1517 			BO_UNLOCK(bo);
1518 			if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1519 				return (error);
1520 			/*
1521 			 * XXX We could save a lock/unlock if this was only
1522 			 * enabled under INVARIANTS
1523 			 */
1524 			BO_LOCK(bo);
1525 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1526 				panic("vinvalbuf: dirty bufs");
1527 		}
1528 	}
1529 	/*
1530 	 * If you alter this loop please notice that interlock is dropped and
1531 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1532 	 * no race conditions occur from this.
1533 	 */
1534 	do {
1535 		error = flushbuflist(&bo->bo_clean,
1536 		    flags, bo, slpflag, slptimeo);
1537 		if (error == 0 && !(flags & V_CLEANONLY))
1538 			error = flushbuflist(&bo->bo_dirty,
1539 			    flags, bo, slpflag, slptimeo);
1540 		if (error != 0 && error != EAGAIN) {
1541 			BO_UNLOCK(bo);
1542 			return (error);
1543 		}
1544 	} while (error != 0);
1545 
1546 	/*
1547 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1548 	 * have write I/O in-progress but if there is a VM object then the
1549 	 * VM object can also have read-I/O in-progress.
1550 	 */
1551 	do {
1552 		bufobj_wwait(bo, 0, 0);
1553 		BO_UNLOCK(bo);
1554 		if (bo->bo_object != NULL) {
1555 			VM_OBJECT_WLOCK(bo->bo_object);
1556 			vm_object_pip_wait(bo->bo_object, "bovlbx");
1557 			VM_OBJECT_WUNLOCK(bo->bo_object);
1558 		}
1559 		BO_LOCK(bo);
1560 	} while (bo->bo_numoutput > 0);
1561 	BO_UNLOCK(bo);
1562 
1563 	/*
1564 	 * Destroy the copy in the VM cache, too.
1565 	 */
1566 	if (bo->bo_object != NULL &&
1567 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY)) == 0) {
1568 		VM_OBJECT_WLOCK(bo->bo_object);
1569 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1570 		    OBJPR_CLEANONLY : 0);
1571 		VM_OBJECT_WUNLOCK(bo->bo_object);
1572 	}
1573 
1574 #ifdef INVARIANTS
1575 	BO_LOCK(bo);
1576 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY)) == 0 &&
1577 	    (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0))
1578 		panic("vinvalbuf: flush failed");
1579 	BO_UNLOCK(bo);
1580 #endif
1581 	return (0);
1582 }
1583 
1584 /*
1585  * Flush out and invalidate all buffers associated with a vnode.
1586  * Called with the underlying object locked.
1587  */
1588 int
1589 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1590 {
1591 
1592 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1593 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1594 	if (vp->v_object != NULL && vp->v_object->handle != vp)
1595 		return (0);
1596 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1597 }
1598 
1599 /*
1600  * Flush out buffers on the specified list.
1601  *
1602  */
1603 static int
1604 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1605     int slptimeo)
1606 {
1607 	struct buf *bp, *nbp;
1608 	int retval, error;
1609 	daddr_t lblkno;
1610 	b_xflags_t xflags;
1611 
1612 	ASSERT_BO_WLOCKED(bo);
1613 
1614 	retval = 0;
1615 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1616 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1617 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1618 			continue;
1619 		}
1620 		lblkno = 0;
1621 		xflags = 0;
1622 		if (nbp != NULL) {
1623 			lblkno = nbp->b_lblkno;
1624 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
1625 		}
1626 		retval = EAGAIN;
1627 		error = BUF_TIMELOCK(bp,
1628 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
1629 		    "flushbuf", slpflag, slptimeo);
1630 		if (error) {
1631 			BO_LOCK(bo);
1632 			return (error != ENOLCK ? error : EAGAIN);
1633 		}
1634 		KASSERT(bp->b_bufobj == bo,
1635 		    ("bp %p wrong b_bufobj %p should be %p",
1636 		    bp, bp->b_bufobj, bo));
1637 		/*
1638 		 * XXX Since there are no node locks for NFS, I
1639 		 * believe there is a slight chance that a delayed
1640 		 * write will occur while sleeping just above, so
1641 		 * check for it.
1642 		 */
1643 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1644 		    (flags & V_SAVE)) {
1645 			bremfree(bp);
1646 			bp->b_flags |= B_ASYNC;
1647 			bwrite(bp);
1648 			BO_LOCK(bo);
1649 			return (EAGAIN);	/* XXX: why not loop ? */
1650 		}
1651 		bremfree(bp);
1652 		bp->b_flags |= (B_INVAL | B_RELBUF);
1653 		bp->b_flags &= ~B_ASYNC;
1654 		brelse(bp);
1655 		BO_LOCK(bo);
1656 		nbp = gbincore(bo, lblkno);
1657 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1658 		    != xflags)
1659 			break;			/* nbp invalid */
1660 	}
1661 	return (retval);
1662 }
1663 
1664 int
1665 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
1666 {
1667 	struct buf *bp;
1668 	int error;
1669 	daddr_t lblkno;
1670 
1671 	ASSERT_BO_LOCKED(bo);
1672 
1673 	for (lblkno = startn;;) {
1674 again:
1675 		bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
1676 		if (bp == NULL || bp->b_lblkno >= endn ||
1677 		    bp->b_lblkno < startn)
1678 			break;
1679 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
1680 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
1681 		if (error != 0) {
1682 			BO_RLOCK(bo);
1683 			if (error == ENOLCK)
1684 				goto again;
1685 			return (error);
1686 		}
1687 		KASSERT(bp->b_bufobj == bo,
1688 		    ("bp %p wrong b_bufobj %p should be %p",
1689 		    bp, bp->b_bufobj, bo));
1690 		lblkno = bp->b_lblkno + 1;
1691 		if ((bp->b_flags & B_MANAGED) == 0)
1692 			bremfree(bp);
1693 		bp->b_flags |= B_RELBUF;
1694 		/*
1695 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
1696 		 * pages backing each buffer in the range are unlikely to be
1697 		 * reused.  Dirty buffers will have the hint applied once
1698 		 * they've been written.
1699 		 */
1700 		if (bp->b_vp->v_object != NULL)
1701 			bp->b_flags |= B_NOREUSE;
1702 		brelse(bp);
1703 		BO_RLOCK(bo);
1704 	}
1705 	return (0);
1706 }
1707 
1708 /*
1709  * Truncate a file's buffer and pages to a specified length.  This
1710  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1711  * sync activity.
1712  */
1713 int
1714 vtruncbuf(struct vnode *vp, struct ucred *cred, off_t length, int blksize)
1715 {
1716 	struct buf *bp, *nbp;
1717 	int anyfreed;
1718 	int trunclbn;
1719 	struct bufobj *bo;
1720 
1721 	CTR5(KTR_VFS, "%s: vp %p with cred %p and block %d:%ju", __func__,
1722 	    vp, cred, blksize, (uintmax_t)length);
1723 
1724 	/*
1725 	 * Round up to the *next* lbn.
1726 	 */
1727 	trunclbn = (length + blksize - 1) / blksize;
1728 
1729 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1730 restart:
1731 	bo = &vp->v_bufobj;
1732 	BO_LOCK(bo);
1733 	anyfreed = 1;
1734 	for (;anyfreed;) {
1735 		anyfreed = 0;
1736 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1737 			if (bp->b_lblkno < trunclbn)
1738 				continue;
1739 			if (BUF_LOCK(bp,
1740 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1741 			    BO_LOCKPTR(bo)) == ENOLCK)
1742 				goto restart;
1743 
1744 			bremfree(bp);
1745 			bp->b_flags |= (B_INVAL | B_RELBUF);
1746 			bp->b_flags &= ~B_ASYNC;
1747 			brelse(bp);
1748 			anyfreed = 1;
1749 
1750 			BO_LOCK(bo);
1751 			if (nbp != NULL &&
1752 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1753 			    (nbp->b_vp != vp) ||
1754 			    (nbp->b_flags & B_DELWRI))) {
1755 				BO_UNLOCK(bo);
1756 				goto restart;
1757 			}
1758 		}
1759 
1760 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1761 			if (bp->b_lblkno < trunclbn)
1762 				continue;
1763 			if (BUF_LOCK(bp,
1764 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1765 			    BO_LOCKPTR(bo)) == ENOLCK)
1766 				goto restart;
1767 			bremfree(bp);
1768 			bp->b_flags |= (B_INVAL | B_RELBUF);
1769 			bp->b_flags &= ~B_ASYNC;
1770 			brelse(bp);
1771 			anyfreed = 1;
1772 
1773 			BO_LOCK(bo);
1774 			if (nbp != NULL &&
1775 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1776 			    (nbp->b_vp != vp) ||
1777 			    (nbp->b_flags & B_DELWRI) == 0)) {
1778 				BO_UNLOCK(bo);
1779 				goto restart;
1780 			}
1781 		}
1782 	}
1783 
1784 	if (length > 0) {
1785 restartsync:
1786 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1787 			if (bp->b_lblkno > 0)
1788 				continue;
1789 			/*
1790 			 * Since we hold the vnode lock this should only
1791 			 * fail if we're racing with the buf daemon.
1792 			 */
1793 			if (BUF_LOCK(bp,
1794 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1795 			    BO_LOCKPTR(bo)) == ENOLCK) {
1796 				goto restart;
1797 			}
1798 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1799 			    ("buf(%p) on dirty queue without DELWRI", bp));
1800 
1801 			bremfree(bp);
1802 			bawrite(bp);
1803 			BO_LOCK(bo);
1804 			goto restartsync;
1805 		}
1806 	}
1807 
1808 	bufobj_wwait(bo, 0, 0);
1809 	BO_UNLOCK(bo);
1810 	vnode_pager_setsize(vp, length);
1811 
1812 	return (0);
1813 }
1814 
1815 static void
1816 buf_vlist_remove(struct buf *bp)
1817 {
1818 	struct bufv *bv;
1819 
1820 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1821 	ASSERT_BO_WLOCKED(bp->b_bufobj);
1822 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
1823 	    (BX_VNDIRTY|BX_VNCLEAN),
1824 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
1825 	if (bp->b_xflags & BX_VNDIRTY)
1826 		bv = &bp->b_bufobj->bo_dirty;
1827 	else
1828 		bv = &bp->b_bufobj->bo_clean;
1829 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
1830 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1831 	bv->bv_cnt--;
1832 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1833 }
1834 
1835 /*
1836  * Add the buffer to the sorted clean or dirty block list.
1837  *
1838  * NOTE: xflags is passed as a constant, optimizing this inline function!
1839  */
1840 static void
1841 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1842 {
1843 	struct bufv *bv;
1844 	struct buf *n;
1845 	int error;
1846 
1847 	ASSERT_BO_WLOCKED(bo);
1848 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
1849 	    ("dead bo %p", bo));
1850 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1851 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
1852 	bp->b_xflags |= xflags;
1853 	if (xflags & BX_VNDIRTY)
1854 		bv = &bo->bo_dirty;
1855 	else
1856 		bv = &bo->bo_clean;
1857 
1858 	/*
1859 	 * Keep the list ordered.  Optimize empty list insertion.  Assume
1860 	 * we tend to grow at the tail so lookup_le should usually be cheaper
1861 	 * than _ge.
1862 	 */
1863 	if (bv->bv_cnt == 0 ||
1864 	    bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
1865 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1866 	else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
1867 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
1868 	else
1869 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
1870 	error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
1871 	if (error)
1872 		panic("buf_vlist_add:  Preallocated nodes insufficient.");
1873 	bv->bv_cnt++;
1874 }
1875 
1876 /*
1877  * Look up a buffer using the buffer tries.
1878  */
1879 struct buf *
1880 gbincore(struct bufobj *bo, daddr_t lblkno)
1881 {
1882 	struct buf *bp;
1883 
1884 	ASSERT_BO_LOCKED(bo);
1885 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
1886 	if (bp != NULL)
1887 		return (bp);
1888 	return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
1889 }
1890 
1891 /*
1892  * Associate a buffer with a vnode.
1893  */
1894 void
1895 bgetvp(struct vnode *vp, struct buf *bp)
1896 {
1897 	struct bufobj *bo;
1898 
1899 	bo = &vp->v_bufobj;
1900 	ASSERT_BO_WLOCKED(bo);
1901 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
1902 
1903 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
1904 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
1905 	    ("bgetvp: bp already attached! %p", bp));
1906 
1907 	vhold(vp);
1908 	bp->b_vp = vp;
1909 	bp->b_bufobj = bo;
1910 	/*
1911 	 * Insert onto list for new vnode.
1912 	 */
1913 	buf_vlist_add(bp, bo, BX_VNCLEAN);
1914 }
1915 
1916 /*
1917  * Disassociate a buffer from a vnode.
1918  */
1919 void
1920 brelvp(struct buf *bp)
1921 {
1922 	struct bufobj *bo;
1923 	struct vnode *vp;
1924 
1925 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1926 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1927 
1928 	/*
1929 	 * Delete from old vnode list, if on one.
1930 	 */
1931 	vp = bp->b_vp;		/* XXX */
1932 	bo = bp->b_bufobj;
1933 	BO_LOCK(bo);
1934 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1935 		buf_vlist_remove(bp);
1936 	else
1937 		panic("brelvp: Buffer %p not on queue.", bp);
1938 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1939 		bo->bo_flag &= ~BO_ONWORKLST;
1940 		mtx_lock(&sync_mtx);
1941 		LIST_REMOVE(bo, bo_synclist);
1942 		syncer_worklist_len--;
1943 		mtx_unlock(&sync_mtx);
1944 	}
1945 	bp->b_vp = NULL;
1946 	bp->b_bufobj = NULL;
1947 	BO_UNLOCK(bo);
1948 	vdrop(vp);
1949 }
1950 
1951 /*
1952  * Add an item to the syncer work queue.
1953  */
1954 static void
1955 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
1956 {
1957 	int slot;
1958 
1959 	ASSERT_BO_WLOCKED(bo);
1960 
1961 	mtx_lock(&sync_mtx);
1962 	if (bo->bo_flag & BO_ONWORKLST)
1963 		LIST_REMOVE(bo, bo_synclist);
1964 	else {
1965 		bo->bo_flag |= BO_ONWORKLST;
1966 		syncer_worklist_len++;
1967 	}
1968 
1969 	if (delay > syncer_maxdelay - 2)
1970 		delay = syncer_maxdelay - 2;
1971 	slot = (syncer_delayno + delay) & syncer_mask;
1972 
1973 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
1974 	mtx_unlock(&sync_mtx);
1975 }
1976 
1977 static int
1978 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
1979 {
1980 	int error, len;
1981 
1982 	mtx_lock(&sync_mtx);
1983 	len = syncer_worklist_len - sync_vnode_count;
1984 	mtx_unlock(&sync_mtx);
1985 	error = SYSCTL_OUT(req, &len, sizeof(len));
1986 	return (error);
1987 }
1988 
1989 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
1990     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
1991 
1992 static struct proc *updateproc;
1993 static void sched_sync(void);
1994 static struct kproc_desc up_kp = {
1995 	"syncer",
1996 	sched_sync,
1997 	&updateproc
1998 };
1999 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2000 
2001 static int
2002 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2003 {
2004 	struct vnode *vp;
2005 	struct mount *mp;
2006 
2007 	*bo = LIST_FIRST(slp);
2008 	if (*bo == NULL)
2009 		return (0);
2010 	vp = (*bo)->__bo_vnode;	/* XXX */
2011 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2012 		return (1);
2013 	/*
2014 	 * We use vhold in case the vnode does not
2015 	 * successfully sync.  vhold prevents the vnode from
2016 	 * going away when we unlock the sync_mtx so that
2017 	 * we can acquire the vnode interlock.
2018 	 */
2019 	vholdl(vp);
2020 	mtx_unlock(&sync_mtx);
2021 	VI_UNLOCK(vp);
2022 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2023 		vdrop(vp);
2024 		mtx_lock(&sync_mtx);
2025 		return (*bo == LIST_FIRST(slp));
2026 	}
2027 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2028 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2029 	VOP_UNLOCK(vp, 0);
2030 	vn_finished_write(mp);
2031 	BO_LOCK(*bo);
2032 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2033 		/*
2034 		 * Put us back on the worklist.  The worklist
2035 		 * routine will remove us from our current
2036 		 * position and then add us back in at a later
2037 		 * position.
2038 		 */
2039 		vn_syncer_add_to_worklist(*bo, syncdelay);
2040 	}
2041 	BO_UNLOCK(*bo);
2042 	vdrop(vp);
2043 	mtx_lock(&sync_mtx);
2044 	return (0);
2045 }
2046 
2047 static int first_printf = 1;
2048 
2049 /*
2050  * System filesystem synchronizer daemon.
2051  */
2052 static void
2053 sched_sync(void)
2054 {
2055 	struct synclist *next, *slp;
2056 	struct bufobj *bo;
2057 	long starttime;
2058 	struct thread *td = curthread;
2059 	int last_work_seen;
2060 	int net_worklist_len;
2061 	int syncer_final_iter;
2062 	int error;
2063 
2064 	last_work_seen = 0;
2065 	syncer_final_iter = 0;
2066 	syncer_state = SYNCER_RUNNING;
2067 	starttime = time_uptime;
2068 	td->td_pflags |= TDP_NORUNNINGBUF;
2069 
2070 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2071 	    SHUTDOWN_PRI_LAST);
2072 
2073 	mtx_lock(&sync_mtx);
2074 	for (;;) {
2075 		if (syncer_state == SYNCER_FINAL_DELAY &&
2076 		    syncer_final_iter == 0) {
2077 			mtx_unlock(&sync_mtx);
2078 			kproc_suspend_check(td->td_proc);
2079 			mtx_lock(&sync_mtx);
2080 		}
2081 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
2082 		if (syncer_state != SYNCER_RUNNING &&
2083 		    starttime != time_uptime) {
2084 			if (first_printf) {
2085 				printf("\nSyncing disks, vnodes remaining...");
2086 				first_printf = 0;
2087 			}
2088 			printf("%d ", net_worklist_len);
2089 		}
2090 		starttime = time_uptime;
2091 
2092 		/*
2093 		 * Push files whose dirty time has expired.  Be careful
2094 		 * of interrupt race on slp queue.
2095 		 *
2096 		 * Skip over empty worklist slots when shutting down.
2097 		 */
2098 		do {
2099 			slp = &syncer_workitem_pending[syncer_delayno];
2100 			syncer_delayno += 1;
2101 			if (syncer_delayno == syncer_maxdelay)
2102 				syncer_delayno = 0;
2103 			next = &syncer_workitem_pending[syncer_delayno];
2104 			/*
2105 			 * If the worklist has wrapped since the
2106 			 * it was emptied of all but syncer vnodes,
2107 			 * switch to the FINAL_DELAY state and run
2108 			 * for one more second.
2109 			 */
2110 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
2111 			    net_worklist_len == 0 &&
2112 			    last_work_seen == syncer_delayno) {
2113 				syncer_state = SYNCER_FINAL_DELAY;
2114 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2115 			}
2116 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2117 		    syncer_worklist_len > 0);
2118 
2119 		/*
2120 		 * Keep track of the last time there was anything
2121 		 * on the worklist other than syncer vnodes.
2122 		 * Return to the SHUTTING_DOWN state if any
2123 		 * new work appears.
2124 		 */
2125 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2126 			last_work_seen = syncer_delayno;
2127 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2128 			syncer_state = SYNCER_SHUTTING_DOWN;
2129 		while (!LIST_EMPTY(slp)) {
2130 			error = sync_vnode(slp, &bo, td);
2131 			if (error == 1) {
2132 				LIST_REMOVE(bo, bo_synclist);
2133 				LIST_INSERT_HEAD(next, bo, bo_synclist);
2134 				continue;
2135 			}
2136 
2137 			if (first_printf == 0) {
2138 				/*
2139 				 * Drop the sync mutex, because some watchdog
2140 				 * drivers need to sleep while patting
2141 				 */
2142 				mtx_unlock(&sync_mtx);
2143 				wdog_kern_pat(WD_LASTVAL);
2144 				mtx_lock(&sync_mtx);
2145 			}
2146 
2147 		}
2148 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2149 			syncer_final_iter--;
2150 		/*
2151 		 * The variable rushjob allows the kernel to speed up the
2152 		 * processing of the filesystem syncer process. A rushjob
2153 		 * value of N tells the filesystem syncer to process the next
2154 		 * N seconds worth of work on its queue ASAP. Currently rushjob
2155 		 * is used by the soft update code to speed up the filesystem
2156 		 * syncer process when the incore state is getting so far
2157 		 * ahead of the disk that the kernel memory pool is being
2158 		 * threatened with exhaustion.
2159 		 */
2160 		if (rushjob > 0) {
2161 			rushjob -= 1;
2162 			continue;
2163 		}
2164 		/*
2165 		 * Just sleep for a short period of time between
2166 		 * iterations when shutting down to allow some I/O
2167 		 * to happen.
2168 		 *
2169 		 * If it has taken us less than a second to process the
2170 		 * current work, then wait. Otherwise start right over
2171 		 * again. We can still lose time if any single round
2172 		 * takes more than two seconds, but it does not really
2173 		 * matter as we are just trying to generally pace the
2174 		 * filesystem activity.
2175 		 */
2176 		if (syncer_state != SYNCER_RUNNING ||
2177 		    time_uptime == starttime) {
2178 			thread_lock(td);
2179 			sched_prio(td, PPAUSE);
2180 			thread_unlock(td);
2181 		}
2182 		if (syncer_state != SYNCER_RUNNING)
2183 			cv_timedwait(&sync_wakeup, &sync_mtx,
2184 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
2185 		else if (time_uptime == starttime)
2186 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2187 	}
2188 }
2189 
2190 /*
2191  * Request the syncer daemon to speed up its work.
2192  * We never push it to speed up more than half of its
2193  * normal turn time, otherwise it could take over the cpu.
2194  */
2195 int
2196 speedup_syncer(void)
2197 {
2198 	int ret = 0;
2199 
2200 	mtx_lock(&sync_mtx);
2201 	if (rushjob < syncdelay / 2) {
2202 		rushjob += 1;
2203 		stat_rush_requests += 1;
2204 		ret = 1;
2205 	}
2206 	mtx_unlock(&sync_mtx);
2207 	cv_broadcast(&sync_wakeup);
2208 	return (ret);
2209 }
2210 
2211 /*
2212  * Tell the syncer to speed up its work and run though its work
2213  * list several times, then tell it to shut down.
2214  */
2215 static void
2216 syncer_shutdown(void *arg, int howto)
2217 {
2218 
2219 	if (howto & RB_NOSYNC)
2220 		return;
2221 	mtx_lock(&sync_mtx);
2222 	syncer_state = SYNCER_SHUTTING_DOWN;
2223 	rushjob = 0;
2224 	mtx_unlock(&sync_mtx);
2225 	cv_broadcast(&sync_wakeup);
2226 	kproc_shutdown(arg, howto);
2227 }
2228 
2229 void
2230 syncer_suspend(void)
2231 {
2232 
2233 	syncer_shutdown(updateproc, 0);
2234 }
2235 
2236 void
2237 syncer_resume(void)
2238 {
2239 
2240 	mtx_lock(&sync_mtx);
2241 	first_printf = 1;
2242 	syncer_state = SYNCER_RUNNING;
2243 	mtx_unlock(&sync_mtx);
2244 	cv_broadcast(&sync_wakeup);
2245 	kproc_resume(updateproc);
2246 }
2247 
2248 /*
2249  * Reassign a buffer from one vnode to another.
2250  * Used to assign file specific control information
2251  * (indirect blocks) to the vnode to which they belong.
2252  */
2253 void
2254 reassignbuf(struct buf *bp)
2255 {
2256 	struct vnode *vp;
2257 	struct bufobj *bo;
2258 	int delay;
2259 #ifdef INVARIANTS
2260 	struct bufv *bv;
2261 #endif
2262 
2263 	vp = bp->b_vp;
2264 	bo = bp->b_bufobj;
2265 	++reassignbufcalls;
2266 
2267 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2268 	    bp, bp->b_vp, bp->b_flags);
2269 	/*
2270 	 * B_PAGING flagged buffers cannot be reassigned because their vp
2271 	 * is not fully linked in.
2272 	 */
2273 	if (bp->b_flags & B_PAGING)
2274 		panic("cannot reassign paging buffer");
2275 
2276 	/*
2277 	 * Delete from old vnode list, if on one.
2278 	 */
2279 	BO_LOCK(bo);
2280 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2281 		buf_vlist_remove(bp);
2282 	else
2283 		panic("reassignbuf: Buffer %p not on queue.", bp);
2284 	/*
2285 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
2286 	 * of clean buffers.
2287 	 */
2288 	if (bp->b_flags & B_DELWRI) {
2289 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2290 			switch (vp->v_type) {
2291 			case VDIR:
2292 				delay = dirdelay;
2293 				break;
2294 			case VCHR:
2295 				delay = metadelay;
2296 				break;
2297 			default:
2298 				delay = filedelay;
2299 			}
2300 			vn_syncer_add_to_worklist(bo, delay);
2301 		}
2302 		buf_vlist_add(bp, bo, BX_VNDIRTY);
2303 	} else {
2304 		buf_vlist_add(bp, bo, BX_VNCLEAN);
2305 
2306 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2307 			mtx_lock(&sync_mtx);
2308 			LIST_REMOVE(bo, bo_synclist);
2309 			syncer_worklist_len--;
2310 			mtx_unlock(&sync_mtx);
2311 			bo->bo_flag &= ~BO_ONWORKLST;
2312 		}
2313 	}
2314 #ifdef INVARIANTS
2315 	bv = &bo->bo_clean;
2316 	bp = TAILQ_FIRST(&bv->bv_hd);
2317 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2318 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2319 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2320 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2321 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2322 	bv = &bo->bo_dirty;
2323 	bp = TAILQ_FIRST(&bv->bv_hd);
2324 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2325 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2326 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
2327 	KASSERT(bp == NULL || bp->b_bufobj == bo,
2328 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2329 #endif
2330 	BO_UNLOCK(bo);
2331 }
2332 
2333 /*
2334  * A temporary hack until refcount_* APIs are sorted out.
2335  */
2336 static __inline int
2337 vfs_refcount_acquire_if_not_zero(volatile u_int *count)
2338 {
2339 	u_int old;
2340 
2341 	for (;;) {
2342 		old = *count;
2343 		if (old == 0)
2344 			return (0);
2345 		if (atomic_cmpset_int(count, old, old + 1))
2346 			return (1);
2347 	}
2348 }
2349 
2350 static __inline int
2351 vfs_refcount_release_if_not_last(volatile u_int *count)
2352 {
2353 	u_int old;
2354 
2355 	for (;;) {
2356 		old = *count;
2357 		if (old == 1)
2358 			return (0);
2359 		if (atomic_cmpset_int(count, old, old - 1))
2360 			return (1);
2361 	}
2362 }
2363 
2364 static void
2365 v_init_counters(struct vnode *vp)
2366 {
2367 
2368 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2369 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
2370 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2371 
2372 	refcount_init(&vp->v_holdcnt, 1);
2373 	refcount_init(&vp->v_usecount, 1);
2374 }
2375 
2376 static void
2377 v_incr_usecount_locked(struct vnode *vp)
2378 {
2379 
2380 	ASSERT_VI_LOCKED(vp, __func__);
2381 	if ((vp->v_iflag & VI_OWEINACT) != 0) {
2382 		VNASSERT(vp->v_usecount == 0, vp,
2383 		    ("vnode with usecount and VI_OWEINACT set"));
2384 		vp->v_iflag &= ~VI_OWEINACT;
2385 	}
2386 	refcount_acquire(&vp->v_usecount);
2387 	v_incr_devcount(vp);
2388 }
2389 
2390 /*
2391  * Increment the use and hold counts on the vnode, taking care to reference
2392  * the driver's usecount if this is a chardev.  The _vhold() will remove
2393  * the vnode from the free list if it is presently free.
2394  */
2395 static void
2396 v_incr_usecount(struct vnode *vp)
2397 {
2398 
2399 	ASSERT_VI_UNLOCKED(vp, __func__);
2400 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2401 
2402 	if (vp->v_type != VCHR &&
2403 	    vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) {
2404 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2405 		    ("vnode with usecount and VI_OWEINACT set"));
2406 	} else {
2407 		VI_LOCK(vp);
2408 		v_incr_usecount_locked(vp);
2409 		VI_UNLOCK(vp);
2410 	}
2411 }
2412 
2413 /*
2414  * Increment si_usecount of the associated device, if any.
2415  */
2416 static void
2417 v_incr_devcount(struct vnode *vp)
2418 {
2419 
2420 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2421 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2422 		dev_lock();
2423 		vp->v_rdev->si_usecount++;
2424 		dev_unlock();
2425 	}
2426 }
2427 
2428 /*
2429  * Decrement si_usecount of the associated device, if any.
2430  */
2431 static void
2432 v_decr_devcount(struct vnode *vp)
2433 {
2434 
2435 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2436 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2437 		dev_lock();
2438 		vp->v_rdev->si_usecount--;
2439 		dev_unlock();
2440 	}
2441 }
2442 
2443 /*
2444  * Grab a particular vnode from the free list, increment its
2445  * reference count and lock it.  VI_DOOMED is set if the vnode
2446  * is being destroyed.  Only callers who specify LK_RETRY will
2447  * see doomed vnodes.  If inactive processing was delayed in
2448  * vput try to do it here.
2449  *
2450  * Notes on lockless counter manipulation:
2451  * _vhold, vputx and other routines make various decisions based
2452  * on either holdcnt or usecount being 0. As long as either contuner
2453  * is not transitioning 0->1 nor 1->0, the manipulation can be done
2454  * with atomic operations. Otherwise the interlock is taken.
2455  */
2456 int
2457 vget(struct vnode *vp, int flags, struct thread *td)
2458 {
2459 	int error, oweinact;
2460 
2461 	VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2462 	    ("vget: invalid lock operation"));
2463 
2464 	if ((flags & LK_INTERLOCK) != 0)
2465 		ASSERT_VI_LOCKED(vp, __func__);
2466 	else
2467 		ASSERT_VI_UNLOCKED(vp, __func__);
2468 	if ((flags & LK_VNHELD) != 0)
2469 		VNASSERT((vp->v_holdcnt > 0), vp,
2470 		    ("vget: LK_VNHELD passed but vnode not held"));
2471 
2472 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2473 
2474 	if ((flags & LK_VNHELD) == 0)
2475 		_vhold(vp, (flags & LK_INTERLOCK) != 0);
2476 
2477 	if ((error = vn_lock(vp, flags)) != 0) {
2478 		vdrop(vp);
2479 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2480 		    vp);
2481 		return (error);
2482 	}
2483 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
2484 		panic("vget: vn_lock failed to return ENOENT\n");
2485 	/*
2486 	 * We don't guarantee that any particular close will
2487 	 * trigger inactive processing so just make a best effort
2488 	 * here at preventing a reference to a removed file.  If
2489 	 * we don't succeed no harm is done.
2490 	 *
2491 	 * Upgrade our holdcnt to a usecount.
2492 	 */
2493 	if (vp->v_type != VCHR &&
2494 	    vfs_refcount_acquire_if_not_zero(&vp->v_usecount)) {
2495 		VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2496 		    ("vnode with usecount and VI_OWEINACT set"));
2497 	} else {
2498 		VI_LOCK(vp);
2499 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2500 			oweinact = 0;
2501 		} else {
2502 			oweinact = 1;
2503 			vp->v_iflag &= ~VI_OWEINACT;
2504 		}
2505 		refcount_acquire(&vp->v_usecount);
2506 		v_incr_devcount(vp);
2507 		if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2508 		    (flags & LK_NOWAIT) == 0)
2509 			vinactive(vp, td);
2510 		VI_UNLOCK(vp);
2511 	}
2512 	return (0);
2513 }
2514 
2515 /*
2516  * Increase the reference count of a vnode.
2517  */
2518 void
2519 vref(struct vnode *vp)
2520 {
2521 
2522 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2523 	_vhold(vp, false);
2524 	v_incr_usecount(vp);
2525 }
2526 
2527 void
2528 vrefl(struct vnode *vp)
2529 {
2530 
2531 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2532 	_vhold(vp, true);
2533 	v_incr_usecount_locked(vp);
2534 }
2535 
2536 /*
2537  * Return reference count of a vnode.
2538  *
2539  * The results of this call are only guaranteed when some mechanism is used to
2540  * stop other processes from gaining references to the vnode.  This may be the
2541  * case if the caller holds the only reference.  This is also useful when stale
2542  * data is acceptable as race conditions may be accounted for by some other
2543  * means.
2544  */
2545 int
2546 vrefcnt(struct vnode *vp)
2547 {
2548 
2549 	return (vp->v_usecount);
2550 }
2551 
2552 #define	VPUTX_VRELE	1
2553 #define	VPUTX_VPUT	2
2554 #define	VPUTX_VUNREF	3
2555 
2556 /*
2557  * Decrement the use and hold counts for a vnode.
2558  *
2559  * See an explanation near vget() as to why atomic operation is safe.
2560  */
2561 static void
2562 vputx(struct vnode *vp, int func)
2563 {
2564 	int error;
2565 
2566 	KASSERT(vp != NULL, ("vputx: null vp"));
2567 	if (func == VPUTX_VUNREF)
2568 		ASSERT_VOP_LOCKED(vp, "vunref");
2569 	else if (func == VPUTX_VPUT)
2570 		ASSERT_VOP_LOCKED(vp, "vput");
2571 	else
2572 		KASSERT(func == VPUTX_VRELE, ("vputx: wrong func"));
2573 	ASSERT_VI_UNLOCKED(vp, __func__);
2574 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2575 
2576 	if (vp->v_type != VCHR &&
2577 	    vfs_refcount_release_if_not_last(&vp->v_usecount)) {
2578 		if (func == VPUTX_VPUT)
2579 			VOP_UNLOCK(vp, 0);
2580 		vdrop(vp);
2581 		return;
2582 	}
2583 
2584 	VI_LOCK(vp);
2585 
2586 	/*
2587 	 * We want to hold the vnode until the inactive finishes to
2588 	 * prevent vgone() races.  We drop the use count here and the
2589 	 * hold count below when we're done.
2590 	 */
2591 	if (!refcount_release(&vp->v_usecount) ||
2592 	    (vp->v_iflag & VI_DOINGINACT)) {
2593 		if (func == VPUTX_VPUT)
2594 			VOP_UNLOCK(vp, 0);
2595 		v_decr_devcount(vp);
2596 		vdropl(vp);
2597 		return;
2598 	}
2599 
2600 	v_decr_devcount(vp);
2601 
2602 	error = 0;
2603 
2604 	if (vp->v_usecount != 0) {
2605 		vprint("vputx: usecount not zero", vp);
2606 		panic("vputx: usecount not zero");
2607 	}
2608 
2609 	CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp);
2610 
2611 	/*
2612 	 * We must call VOP_INACTIVE with the node locked. Mark
2613 	 * as VI_DOINGINACT to avoid recursion.
2614 	 */
2615 	vp->v_iflag |= VI_OWEINACT;
2616 	switch (func) {
2617 	case VPUTX_VRELE:
2618 		error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
2619 		VI_LOCK(vp);
2620 		break;
2621 	case VPUTX_VPUT:
2622 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2623 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
2624 			    LK_NOWAIT);
2625 			VI_LOCK(vp);
2626 		}
2627 		break;
2628 	case VPUTX_VUNREF:
2629 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2630 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
2631 			VI_LOCK(vp);
2632 		}
2633 		break;
2634 	}
2635 	VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
2636 	    ("vnode with usecount and VI_OWEINACT set"));
2637 	if (error == 0) {
2638 		if (vp->v_iflag & VI_OWEINACT)
2639 			vinactive(vp, curthread);
2640 		if (func != VPUTX_VUNREF)
2641 			VOP_UNLOCK(vp, 0);
2642 	}
2643 	vdropl(vp);
2644 }
2645 
2646 /*
2647  * Vnode put/release.
2648  * If count drops to zero, call inactive routine and return to freelist.
2649  */
2650 void
2651 vrele(struct vnode *vp)
2652 {
2653 
2654 	vputx(vp, VPUTX_VRELE);
2655 }
2656 
2657 /*
2658  * Release an already locked vnode.  This give the same effects as
2659  * unlock+vrele(), but takes less time and avoids releasing and
2660  * re-aquiring the lock (as vrele() acquires the lock internally.)
2661  */
2662 void
2663 vput(struct vnode *vp)
2664 {
2665 
2666 	vputx(vp, VPUTX_VPUT);
2667 }
2668 
2669 /*
2670  * Release an exclusively locked vnode. Do not unlock the vnode lock.
2671  */
2672 void
2673 vunref(struct vnode *vp)
2674 {
2675 
2676 	vputx(vp, VPUTX_VUNREF);
2677 }
2678 
2679 /*
2680  * Increase the hold count and activate if this is the first reference.
2681  */
2682 void
2683 _vhold(struct vnode *vp, bool locked)
2684 {
2685 	struct mount *mp;
2686 
2687 	if (locked)
2688 		ASSERT_VI_LOCKED(vp, __func__);
2689 	else
2690 		ASSERT_VI_UNLOCKED(vp, __func__);
2691 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2692 	if (!locked && vfs_refcount_acquire_if_not_zero(&vp->v_holdcnt)) {
2693 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2694 		    ("_vhold: vnode with holdcnt is free"));
2695 		return;
2696 	}
2697 
2698 	if (!locked)
2699 		VI_LOCK(vp);
2700 	if ((vp->v_iflag & VI_FREE) == 0) {
2701 		refcount_acquire(&vp->v_holdcnt);
2702 		if (!locked)
2703 			VI_UNLOCK(vp);
2704 		return;
2705 	}
2706 	VNASSERT(vp->v_holdcnt == 0, vp,
2707 	    ("%s: wrong hold count", __func__));
2708 	VNASSERT(vp->v_op != NULL, vp,
2709 	    ("%s: vnode already reclaimed.", __func__));
2710 	/*
2711 	 * Remove a vnode from the free list, mark it as in use,
2712 	 * and put it on the active list.
2713 	 */
2714 	mtx_lock(&vnode_free_list_mtx);
2715 	TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
2716 	freevnodes--;
2717 	vp->v_iflag &= ~VI_FREE;
2718 	KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
2719 	    ("Activating already active vnode"));
2720 	vp->v_iflag |= VI_ACTIVE;
2721 	mp = vp->v_mount;
2722 	TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
2723 	mp->mnt_activevnodelistsize++;
2724 	mtx_unlock(&vnode_free_list_mtx);
2725 	refcount_acquire(&vp->v_holdcnt);
2726 	if (!locked)
2727 		VI_UNLOCK(vp);
2728 }
2729 
2730 /*
2731  * Drop the hold count of the vnode.  If this is the last reference to
2732  * the vnode we place it on the free list unless it has been vgone'd
2733  * (marked VI_DOOMED) in which case we will free it.
2734  *
2735  * Because the vnode vm object keeps a hold reference on the vnode if
2736  * there is at least one resident non-cached page, the vnode cannot
2737  * leave the active list without the page cleanup done.
2738  */
2739 void
2740 _vdrop(struct vnode *vp, bool locked)
2741 {
2742 	struct bufobj *bo;
2743 	struct mount *mp;
2744 	int active;
2745 
2746 	if (locked)
2747 		ASSERT_VI_LOCKED(vp, __func__);
2748 	else
2749 		ASSERT_VI_UNLOCKED(vp, __func__);
2750 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2751 	if ((int)vp->v_holdcnt <= 0)
2752 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
2753 	if (vfs_refcount_release_if_not_last(&vp->v_holdcnt)) {
2754 		if (locked)
2755 			VI_UNLOCK(vp);
2756 		return;
2757 	}
2758 
2759 	if (!locked)
2760 		VI_LOCK(vp);
2761 	if (refcount_release(&vp->v_holdcnt) == 0) {
2762 		VI_UNLOCK(vp);
2763 		return;
2764 	}
2765 	if ((vp->v_iflag & VI_DOOMED) == 0) {
2766 		/*
2767 		 * Mark a vnode as free: remove it from its active list
2768 		 * and put it up for recycling on the freelist.
2769 		 */
2770 		VNASSERT(vp->v_op != NULL, vp,
2771 		    ("vdropl: vnode already reclaimed."));
2772 		VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2773 		    ("vnode already free"));
2774 		VNASSERT(vp->v_holdcnt == 0, vp,
2775 		    ("vdropl: freeing when we shouldn't"));
2776 		active = vp->v_iflag & VI_ACTIVE;
2777 		if ((vp->v_iflag & VI_OWEINACT) == 0) {
2778 			vp->v_iflag &= ~VI_ACTIVE;
2779 			mp = vp->v_mount;
2780 			mtx_lock(&vnode_free_list_mtx);
2781 			if (active) {
2782 				TAILQ_REMOVE(&mp->mnt_activevnodelist, vp,
2783 				    v_actfreelist);
2784 				mp->mnt_activevnodelistsize--;
2785 			}
2786 			TAILQ_INSERT_TAIL(&vnode_free_list, vp,
2787 			    v_actfreelist);
2788 			freevnodes++;
2789 			vp->v_iflag |= VI_FREE;
2790 			mtx_unlock(&vnode_free_list_mtx);
2791 		} else {
2792 			atomic_add_long(&free_owe_inact, 1);
2793 		}
2794 		VI_UNLOCK(vp);
2795 		return;
2796 	}
2797 	/*
2798 	 * The vnode has been marked for destruction, so free it.
2799 	 *
2800 	 * The vnode will be returned to the zone where it will
2801 	 * normally remain until it is needed for another vnode. We
2802 	 * need to cleanup (or verify that the cleanup has already
2803 	 * been done) any residual data left from its current use
2804 	 * so as not to contaminate the freshly allocated vnode.
2805 	 */
2806 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
2807 	atomic_subtract_long(&numvnodes, 1);
2808 	bo = &vp->v_bufobj;
2809 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2810 	    ("cleaned vnode still on the free list."));
2811 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
2812 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
2813 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
2814 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
2815 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
2816 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
2817 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
2818 	    ("clean blk trie not empty"));
2819 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
2820 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
2821 	    ("dirty blk trie not empty"));
2822 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
2823 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
2824 	VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
2825 	VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
2826 	    ("Dangling rangelock waiters"));
2827 	VI_UNLOCK(vp);
2828 #ifdef MAC
2829 	mac_vnode_destroy(vp);
2830 #endif
2831 	if (vp->v_pollinfo != NULL) {
2832 		destroy_vpollinfo(vp->v_pollinfo);
2833 		vp->v_pollinfo = NULL;
2834 	}
2835 #ifdef INVARIANTS
2836 	/* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
2837 	vp->v_op = NULL;
2838 #endif
2839 	bzero(&vp->v_un, sizeof(vp->v_un));
2840 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
2841 	vp->v_iflag = 0;
2842 	vp->v_vflag = 0;
2843 	bo->bo_flag = 0;
2844 	uma_zfree(vnode_zone, vp);
2845 }
2846 
2847 /*
2848  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
2849  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
2850  * OWEINACT tracks whether a vnode missed a call to inactive due to a
2851  * failed lock upgrade.
2852  */
2853 void
2854 vinactive(struct vnode *vp, struct thread *td)
2855 {
2856 	struct vm_object *obj;
2857 
2858 	ASSERT_VOP_ELOCKED(vp, "vinactive");
2859 	ASSERT_VI_LOCKED(vp, "vinactive");
2860 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2861 	    ("vinactive: recursed on VI_DOINGINACT"));
2862 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2863 	vp->v_iflag |= VI_DOINGINACT;
2864 	vp->v_iflag &= ~VI_OWEINACT;
2865 	VI_UNLOCK(vp);
2866 	/*
2867 	 * Before moving off the active list, we must be sure that any
2868 	 * modified pages are converted into the vnode's dirty
2869 	 * buffers, since these will no longer be checked once the
2870 	 * vnode is on the inactive list.
2871 	 *
2872 	 * The write-out of the dirty pages is asynchronous.  At the
2873 	 * point that VOP_INACTIVE() is called, there could still be
2874 	 * pending I/O and dirty pages in the object.
2875 	 */
2876 	obj = vp->v_object;
2877 	if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0) {
2878 		VM_OBJECT_WLOCK(obj);
2879 		vm_object_page_clean(obj, 0, 0, OBJPC_NOSYNC);
2880 		VM_OBJECT_WUNLOCK(obj);
2881 	}
2882 	VOP_INACTIVE(vp, td);
2883 	VI_LOCK(vp);
2884 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2885 	    ("vinactive: lost VI_DOINGINACT"));
2886 	vp->v_iflag &= ~VI_DOINGINACT;
2887 }
2888 
2889 /*
2890  * Remove any vnodes in the vnode table belonging to mount point mp.
2891  *
2892  * If FORCECLOSE is not specified, there should not be any active ones,
2893  * return error if any are found (nb: this is a user error, not a
2894  * system error). If FORCECLOSE is specified, detach any active vnodes
2895  * that are found.
2896  *
2897  * If WRITECLOSE is set, only flush out regular file vnodes open for
2898  * writing.
2899  *
2900  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2901  *
2902  * `rootrefs' specifies the base reference count for the root vnode
2903  * of this filesystem. The root vnode is considered busy if its
2904  * v_usecount exceeds this value. On a successful return, vflush(, td)
2905  * will call vrele() on the root vnode exactly rootrefs times.
2906  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2907  * be zero.
2908  */
2909 #ifdef DIAGNOSTIC
2910 static int busyprt = 0;		/* print out busy vnodes */
2911 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
2912 #endif
2913 
2914 int
2915 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
2916 {
2917 	struct vnode *vp, *mvp, *rootvp = NULL;
2918 	struct vattr vattr;
2919 	int busy = 0, error;
2920 
2921 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
2922 	    rootrefs, flags);
2923 	if (rootrefs > 0) {
2924 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2925 		    ("vflush: bad args"));
2926 		/*
2927 		 * Get the filesystem root vnode. We can vput() it
2928 		 * immediately, since with rootrefs > 0, it won't go away.
2929 		 */
2930 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
2931 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
2932 			    __func__, error);
2933 			return (error);
2934 		}
2935 		vput(rootvp);
2936 	}
2937 loop:
2938 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
2939 		vholdl(vp);
2940 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
2941 		if (error) {
2942 			vdrop(vp);
2943 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
2944 			goto loop;
2945 		}
2946 		/*
2947 		 * Skip over a vnodes marked VV_SYSTEM.
2948 		 */
2949 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2950 			VOP_UNLOCK(vp, 0);
2951 			vdrop(vp);
2952 			continue;
2953 		}
2954 		/*
2955 		 * If WRITECLOSE is set, flush out unlinked but still open
2956 		 * files (even if open only for reading) and regular file
2957 		 * vnodes open for writing.
2958 		 */
2959 		if (flags & WRITECLOSE) {
2960 			if (vp->v_object != NULL) {
2961 				VM_OBJECT_WLOCK(vp->v_object);
2962 				vm_object_page_clean(vp->v_object, 0, 0, 0);
2963 				VM_OBJECT_WUNLOCK(vp->v_object);
2964 			}
2965 			error = VOP_FSYNC(vp, MNT_WAIT, td);
2966 			if (error != 0) {
2967 				VOP_UNLOCK(vp, 0);
2968 				vdrop(vp);
2969 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
2970 				return (error);
2971 			}
2972 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
2973 			VI_LOCK(vp);
2974 
2975 			if ((vp->v_type == VNON ||
2976 			    (error == 0 && vattr.va_nlink > 0)) &&
2977 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2978 				VOP_UNLOCK(vp, 0);
2979 				vdropl(vp);
2980 				continue;
2981 			}
2982 		} else
2983 			VI_LOCK(vp);
2984 		/*
2985 		 * With v_usecount == 0, all we need to do is clear out the
2986 		 * vnode data structures and we are done.
2987 		 *
2988 		 * If FORCECLOSE is set, forcibly close the vnode.
2989 		 */
2990 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
2991 			vgonel(vp);
2992 		} else {
2993 			busy++;
2994 #ifdef DIAGNOSTIC
2995 			if (busyprt)
2996 				vprint("vflush: busy vnode", vp);
2997 #endif
2998 		}
2999 		VOP_UNLOCK(vp, 0);
3000 		vdropl(vp);
3001 	}
3002 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3003 		/*
3004 		 * If just the root vnode is busy, and if its refcount
3005 		 * is equal to `rootrefs', then go ahead and kill it.
3006 		 */
3007 		VI_LOCK(rootvp);
3008 		KASSERT(busy > 0, ("vflush: not busy"));
3009 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3010 		    ("vflush: usecount %d < rootrefs %d",
3011 		     rootvp->v_usecount, rootrefs));
3012 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
3013 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3014 			vgone(rootvp);
3015 			VOP_UNLOCK(rootvp, 0);
3016 			busy = 0;
3017 		} else
3018 			VI_UNLOCK(rootvp);
3019 	}
3020 	if (busy) {
3021 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3022 		    busy);
3023 		return (EBUSY);
3024 	}
3025 	for (; rootrefs > 0; rootrefs--)
3026 		vrele(rootvp);
3027 	return (0);
3028 }
3029 
3030 /*
3031  * Recycle an unused vnode to the front of the free list.
3032  */
3033 int
3034 vrecycle(struct vnode *vp)
3035 {
3036 	int recycled;
3037 
3038 	ASSERT_VOP_ELOCKED(vp, "vrecycle");
3039 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3040 	recycled = 0;
3041 	VI_LOCK(vp);
3042 	if (vp->v_usecount == 0) {
3043 		recycled = 1;
3044 		vgonel(vp);
3045 	}
3046 	VI_UNLOCK(vp);
3047 	return (recycled);
3048 }
3049 
3050 /*
3051  * Eliminate all activity associated with a vnode
3052  * in preparation for reuse.
3053  */
3054 void
3055 vgone(struct vnode *vp)
3056 {
3057 	VI_LOCK(vp);
3058 	vgonel(vp);
3059 	VI_UNLOCK(vp);
3060 }
3061 
3062 static void
3063 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3064     struct vnode *lowervp __unused)
3065 {
3066 }
3067 
3068 /*
3069  * Notify upper mounts about reclaimed or unlinked vnode.
3070  */
3071 void
3072 vfs_notify_upper(struct vnode *vp, int event)
3073 {
3074 	static struct vfsops vgonel_vfsops = {
3075 		.vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3076 		.vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3077 	};
3078 	struct mount *mp, *ump, *mmp;
3079 
3080 	mp = vp->v_mount;
3081 	if (mp == NULL)
3082 		return;
3083 
3084 	MNT_ILOCK(mp);
3085 	if (TAILQ_EMPTY(&mp->mnt_uppers))
3086 		goto unlock;
3087 	MNT_IUNLOCK(mp);
3088 	mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3089 	mmp->mnt_op = &vgonel_vfsops;
3090 	mmp->mnt_kern_flag |= MNTK_MARKER;
3091 	MNT_ILOCK(mp);
3092 	mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3093 	for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3094 		if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3095 			ump = TAILQ_NEXT(ump, mnt_upper_link);
3096 			continue;
3097 		}
3098 		TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3099 		MNT_IUNLOCK(mp);
3100 		switch (event) {
3101 		case VFS_NOTIFY_UPPER_RECLAIM:
3102 			VFS_RECLAIM_LOWERVP(ump, vp);
3103 			break;
3104 		case VFS_NOTIFY_UPPER_UNLINK:
3105 			VFS_UNLINK_LOWERVP(ump, vp);
3106 			break;
3107 		default:
3108 			KASSERT(0, ("invalid event %d", event));
3109 			break;
3110 		}
3111 		MNT_ILOCK(mp);
3112 		ump = TAILQ_NEXT(mmp, mnt_upper_link);
3113 		TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3114 	}
3115 	free(mmp, M_TEMP);
3116 	mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3117 	if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3118 		mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3119 		wakeup(&mp->mnt_uppers);
3120 	}
3121 unlock:
3122 	MNT_IUNLOCK(mp);
3123 }
3124 
3125 /*
3126  * vgone, with the vp interlock held.
3127  */
3128 static void
3129 vgonel(struct vnode *vp)
3130 {
3131 	struct thread *td;
3132 	int oweinact;
3133 	int active;
3134 	struct mount *mp;
3135 
3136 	ASSERT_VOP_ELOCKED(vp, "vgonel");
3137 	ASSERT_VI_LOCKED(vp, "vgonel");
3138 	VNASSERT(vp->v_holdcnt, vp,
3139 	    ("vgonel: vp %p has no reference.", vp));
3140 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3141 	td = curthread;
3142 
3143 	/*
3144 	 * Don't vgonel if we're already doomed.
3145 	 */
3146 	if (vp->v_iflag & VI_DOOMED)
3147 		return;
3148 	vp->v_iflag |= VI_DOOMED;
3149 
3150 	/*
3151 	 * Check to see if the vnode is in use.  If so, we have to call
3152 	 * VOP_CLOSE() and VOP_INACTIVE().
3153 	 */
3154 	active = vp->v_usecount;
3155 	oweinact = (vp->v_iflag & VI_OWEINACT);
3156 	VI_UNLOCK(vp);
3157 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3158 
3159 	/*
3160 	 * If purging an active vnode, it must be closed and
3161 	 * deactivated before being reclaimed.
3162 	 */
3163 	if (active)
3164 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3165 	if (oweinact || active) {
3166 		VI_LOCK(vp);
3167 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
3168 			vinactive(vp, td);
3169 		VI_UNLOCK(vp);
3170 	}
3171 	if (vp->v_type == VSOCK)
3172 		vfs_unp_reclaim(vp);
3173 
3174 	/*
3175 	 * Clean out any buffers associated with the vnode.
3176 	 * If the flush fails, just toss the buffers.
3177 	 */
3178 	mp = NULL;
3179 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3180 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
3181 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3182 		while (vinvalbuf(vp, 0, 0, 0) != 0)
3183 			;
3184 	}
3185 
3186 	BO_LOCK(&vp->v_bufobj);
3187 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3188 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3189 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3190 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
3191 	    ("vp %p bufobj not invalidated", vp));
3192 	vp->v_bufobj.bo_flag |= BO_DEAD;
3193 	BO_UNLOCK(&vp->v_bufobj);
3194 
3195 	/*
3196 	 * Reclaim the vnode.
3197 	 */
3198 	if (VOP_RECLAIM(vp, td))
3199 		panic("vgone: cannot reclaim");
3200 	if (mp != NULL)
3201 		vn_finished_secondary_write(mp);
3202 	VNASSERT(vp->v_object == NULL, vp,
3203 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
3204 	/*
3205 	 * Clear the advisory locks and wake up waiting threads.
3206 	 */
3207 	(void)VOP_ADVLOCKPURGE(vp);
3208 	vp->v_lockf = NULL;
3209 	/*
3210 	 * Delete from old mount point vnode list.
3211 	 */
3212 	delmntque(vp);
3213 	cache_purge(vp);
3214 	/*
3215 	 * Done with purge, reset to the standard lock and invalidate
3216 	 * the vnode.
3217 	 */
3218 	VI_LOCK(vp);
3219 	vp->v_vnlock = &vp->v_lock;
3220 	vp->v_op = &dead_vnodeops;
3221 	vp->v_tag = "none";
3222 	vp->v_type = VBAD;
3223 }
3224 
3225 /*
3226  * Calculate the total number of references to a special device.
3227  */
3228 int
3229 vcount(struct vnode *vp)
3230 {
3231 	int count;
3232 
3233 	dev_lock();
3234 	count = vp->v_rdev->si_usecount;
3235 	dev_unlock();
3236 	return (count);
3237 }
3238 
3239 /*
3240  * Same as above, but using the struct cdev *as argument
3241  */
3242 int
3243 count_dev(struct cdev *dev)
3244 {
3245 	int count;
3246 
3247 	dev_lock();
3248 	count = dev->si_usecount;
3249 	dev_unlock();
3250 	return(count);
3251 }
3252 
3253 /*
3254  * Print out a description of a vnode.
3255  */
3256 static char *typename[] =
3257 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3258  "VMARKER"};
3259 
3260 void
3261 vn_printf(struct vnode *vp, const char *fmt, ...)
3262 {
3263 	va_list ap;
3264 	char buf[256], buf2[16];
3265 	u_long flags;
3266 
3267 	va_start(ap, fmt);
3268 	vprintf(fmt, ap);
3269 	va_end(ap);
3270 	printf("%p: ", (void *)vp);
3271 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
3272 	printf("    usecount %d, writecount %d, refcount %d mountedhere %p\n",
3273 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere);
3274 	buf[0] = '\0';
3275 	buf[1] = '\0';
3276 	if (vp->v_vflag & VV_ROOT)
3277 		strlcat(buf, "|VV_ROOT", sizeof(buf));
3278 	if (vp->v_vflag & VV_ISTTY)
3279 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
3280 	if (vp->v_vflag & VV_NOSYNC)
3281 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3282 	if (vp->v_vflag & VV_ETERNALDEV)
3283 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3284 	if (vp->v_vflag & VV_CACHEDLABEL)
3285 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3286 	if (vp->v_vflag & VV_TEXT)
3287 		strlcat(buf, "|VV_TEXT", sizeof(buf));
3288 	if (vp->v_vflag & VV_COPYONWRITE)
3289 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3290 	if (vp->v_vflag & VV_SYSTEM)
3291 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3292 	if (vp->v_vflag & VV_PROCDEP)
3293 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3294 	if (vp->v_vflag & VV_NOKNOTE)
3295 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3296 	if (vp->v_vflag & VV_DELETED)
3297 		strlcat(buf, "|VV_DELETED", sizeof(buf));
3298 	if (vp->v_vflag & VV_MD)
3299 		strlcat(buf, "|VV_MD", sizeof(buf));
3300 	if (vp->v_vflag & VV_FORCEINSMQ)
3301 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
3302 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
3303 	    VV_CACHEDLABEL | VV_TEXT | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
3304 	    VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
3305 	if (flags != 0) {
3306 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
3307 		strlcat(buf, buf2, sizeof(buf));
3308 	}
3309 	if (vp->v_iflag & VI_MOUNT)
3310 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
3311 	if (vp->v_iflag & VI_DOOMED)
3312 		strlcat(buf, "|VI_DOOMED", sizeof(buf));
3313 	if (vp->v_iflag & VI_FREE)
3314 		strlcat(buf, "|VI_FREE", sizeof(buf));
3315 	if (vp->v_iflag & VI_ACTIVE)
3316 		strlcat(buf, "|VI_ACTIVE", sizeof(buf));
3317 	if (vp->v_iflag & VI_DOINGINACT)
3318 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
3319 	if (vp->v_iflag & VI_OWEINACT)
3320 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
3321 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOOMED | VI_FREE |
3322 	    VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT);
3323 	if (flags != 0) {
3324 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
3325 		strlcat(buf, buf2, sizeof(buf));
3326 	}
3327 	printf("    flags (%s)\n", buf + 1);
3328 	if (mtx_owned(VI_MTX(vp)))
3329 		printf(" VI_LOCKed");
3330 	if (vp->v_object != NULL)
3331 		printf("    v_object %p ref %d pages %d "
3332 		    "cleanbuf %d dirtybuf %d\n",
3333 		    vp->v_object, vp->v_object->ref_count,
3334 		    vp->v_object->resident_page_count,
3335 		    vp->v_bufobj.bo_clean.bv_cnt,
3336 		    vp->v_bufobj.bo_dirty.bv_cnt);
3337 	printf("    ");
3338 	lockmgr_printinfo(vp->v_vnlock);
3339 	if (vp->v_data != NULL)
3340 		VOP_PRINT(vp);
3341 }
3342 
3343 #ifdef DDB
3344 /*
3345  * List all of the locked vnodes in the system.
3346  * Called when debugging the kernel.
3347  */
3348 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
3349 {
3350 	struct mount *mp;
3351 	struct vnode *vp;
3352 
3353 	/*
3354 	 * Note: because this is DDB, we can't obey the locking semantics
3355 	 * for these structures, which means we could catch an inconsistent
3356 	 * state and dereference a nasty pointer.  Not much to be done
3357 	 * about that.
3358 	 */
3359 	db_printf("Locked vnodes\n");
3360 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3361 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3362 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
3363 				vprint("", vp);
3364 		}
3365 	}
3366 }
3367 
3368 /*
3369  * Show details about the given vnode.
3370  */
3371 DB_SHOW_COMMAND(vnode, db_show_vnode)
3372 {
3373 	struct vnode *vp;
3374 
3375 	if (!have_addr)
3376 		return;
3377 	vp = (struct vnode *)addr;
3378 	vn_printf(vp, "vnode ");
3379 }
3380 
3381 /*
3382  * Show details about the given mount point.
3383  */
3384 DB_SHOW_COMMAND(mount, db_show_mount)
3385 {
3386 	struct mount *mp;
3387 	struct vfsopt *opt;
3388 	struct statfs *sp;
3389 	struct vnode *vp;
3390 	char buf[512];
3391 	uint64_t mflags;
3392 	u_int flags;
3393 
3394 	if (!have_addr) {
3395 		/* No address given, print short info about all mount points. */
3396 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3397 			db_printf("%p %s on %s (%s)\n", mp,
3398 			    mp->mnt_stat.f_mntfromname,
3399 			    mp->mnt_stat.f_mntonname,
3400 			    mp->mnt_stat.f_fstypename);
3401 			if (db_pager_quit)
3402 				break;
3403 		}
3404 		db_printf("\nMore info: show mount <addr>\n");
3405 		return;
3406 	}
3407 
3408 	mp = (struct mount *)addr;
3409 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
3410 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
3411 
3412 	buf[0] = '\0';
3413 	mflags = mp->mnt_flag;
3414 #define	MNT_FLAG(flag)	do {						\
3415 	if (mflags & (flag)) {						\
3416 		if (buf[0] != '\0')					\
3417 			strlcat(buf, ", ", sizeof(buf));		\
3418 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
3419 		mflags &= ~(flag);					\
3420 	}								\
3421 } while (0)
3422 	MNT_FLAG(MNT_RDONLY);
3423 	MNT_FLAG(MNT_SYNCHRONOUS);
3424 	MNT_FLAG(MNT_NOEXEC);
3425 	MNT_FLAG(MNT_NOSUID);
3426 	MNT_FLAG(MNT_NFS4ACLS);
3427 	MNT_FLAG(MNT_UNION);
3428 	MNT_FLAG(MNT_ASYNC);
3429 	MNT_FLAG(MNT_SUIDDIR);
3430 	MNT_FLAG(MNT_SOFTDEP);
3431 	MNT_FLAG(MNT_NOSYMFOLLOW);
3432 	MNT_FLAG(MNT_GJOURNAL);
3433 	MNT_FLAG(MNT_MULTILABEL);
3434 	MNT_FLAG(MNT_ACLS);
3435 	MNT_FLAG(MNT_NOATIME);
3436 	MNT_FLAG(MNT_NOCLUSTERR);
3437 	MNT_FLAG(MNT_NOCLUSTERW);
3438 	MNT_FLAG(MNT_SUJ);
3439 	MNT_FLAG(MNT_EXRDONLY);
3440 	MNT_FLAG(MNT_EXPORTED);
3441 	MNT_FLAG(MNT_DEFEXPORTED);
3442 	MNT_FLAG(MNT_EXPORTANON);
3443 	MNT_FLAG(MNT_EXKERB);
3444 	MNT_FLAG(MNT_EXPUBLIC);
3445 	MNT_FLAG(MNT_LOCAL);
3446 	MNT_FLAG(MNT_QUOTA);
3447 	MNT_FLAG(MNT_ROOTFS);
3448 	MNT_FLAG(MNT_USER);
3449 	MNT_FLAG(MNT_IGNORE);
3450 	MNT_FLAG(MNT_UPDATE);
3451 	MNT_FLAG(MNT_DELEXPORT);
3452 	MNT_FLAG(MNT_RELOAD);
3453 	MNT_FLAG(MNT_FORCE);
3454 	MNT_FLAG(MNT_SNAPSHOT);
3455 	MNT_FLAG(MNT_BYFSID);
3456 #undef MNT_FLAG
3457 	if (mflags != 0) {
3458 		if (buf[0] != '\0')
3459 			strlcat(buf, ", ", sizeof(buf));
3460 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3461 		    "0x%016jx", mflags);
3462 	}
3463 	db_printf("    mnt_flag = %s\n", buf);
3464 
3465 	buf[0] = '\0';
3466 	flags = mp->mnt_kern_flag;
3467 #define	MNT_KERN_FLAG(flag)	do {					\
3468 	if (flags & (flag)) {						\
3469 		if (buf[0] != '\0')					\
3470 			strlcat(buf, ", ", sizeof(buf));		\
3471 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
3472 		flags &= ~(flag);					\
3473 	}								\
3474 } while (0)
3475 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
3476 	MNT_KERN_FLAG(MNTK_ASYNC);
3477 	MNT_KERN_FLAG(MNTK_SOFTDEP);
3478 	MNT_KERN_FLAG(MNTK_NOINSMNTQ);
3479 	MNT_KERN_FLAG(MNTK_DRAINING);
3480 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
3481 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
3482 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
3483 	MNT_KERN_FLAG(MNTK_NO_IOPF);
3484 	MNT_KERN_FLAG(MNTK_VGONE_UPPER);
3485 	MNT_KERN_FLAG(MNTK_VGONE_WAITER);
3486 	MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
3487 	MNT_KERN_FLAG(MNTK_MARKER);
3488 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
3489 	MNT_KERN_FLAG(MNTK_NOASYNC);
3490 	MNT_KERN_FLAG(MNTK_UNMOUNT);
3491 	MNT_KERN_FLAG(MNTK_MWAIT);
3492 	MNT_KERN_FLAG(MNTK_SUSPEND);
3493 	MNT_KERN_FLAG(MNTK_SUSPEND2);
3494 	MNT_KERN_FLAG(MNTK_SUSPENDED);
3495 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
3496 	MNT_KERN_FLAG(MNTK_NOKNOTE);
3497 #undef MNT_KERN_FLAG
3498 	if (flags != 0) {
3499 		if (buf[0] != '\0')
3500 			strlcat(buf, ", ", sizeof(buf));
3501 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3502 		    "0x%08x", flags);
3503 	}
3504 	db_printf("    mnt_kern_flag = %s\n", buf);
3505 
3506 	db_printf("    mnt_opt = ");
3507 	opt = TAILQ_FIRST(mp->mnt_opt);
3508 	if (opt != NULL) {
3509 		db_printf("%s", opt->name);
3510 		opt = TAILQ_NEXT(opt, link);
3511 		while (opt != NULL) {
3512 			db_printf(", %s", opt->name);
3513 			opt = TAILQ_NEXT(opt, link);
3514 		}
3515 	}
3516 	db_printf("\n");
3517 
3518 	sp = &mp->mnt_stat;
3519 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
3520 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
3521 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
3522 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
3523 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
3524 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
3525 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
3526 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
3527 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
3528 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
3529 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
3530 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
3531 
3532 	db_printf("    mnt_cred = { uid=%u ruid=%u",
3533 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
3534 	if (jailed(mp->mnt_cred))
3535 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
3536 	db_printf(" }\n");
3537 	db_printf("    mnt_ref = %d\n", mp->mnt_ref);
3538 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
3539 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
3540 	db_printf("    mnt_activevnodelistsize = %d\n",
3541 	    mp->mnt_activevnodelistsize);
3542 	db_printf("    mnt_writeopcount = %d\n", mp->mnt_writeopcount);
3543 	db_printf("    mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
3544 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
3545 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
3546 	db_printf("    mnt_lockref = %d\n", mp->mnt_lockref);
3547 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
3548 	db_printf("    mnt_secondary_accwrites = %d\n",
3549 	    mp->mnt_secondary_accwrites);
3550 	db_printf("    mnt_gjprovider = %s\n",
3551 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
3552 
3553 	db_printf("\n\nList of active vnodes\n");
3554 	TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) {
3555 		if (vp->v_type != VMARKER) {
3556 			vn_printf(vp, "vnode ");
3557 			if (db_pager_quit)
3558 				break;
3559 		}
3560 	}
3561 	db_printf("\n\nList of inactive vnodes\n");
3562 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3563 		if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) {
3564 			vn_printf(vp, "vnode ");
3565 			if (db_pager_quit)
3566 				break;
3567 		}
3568 	}
3569 }
3570 #endif	/* DDB */
3571 
3572 /*
3573  * Fill in a struct xvfsconf based on a struct vfsconf.
3574  */
3575 static int
3576 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
3577 {
3578 	struct xvfsconf xvfsp;
3579 
3580 	bzero(&xvfsp, sizeof(xvfsp));
3581 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3582 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3583 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3584 	xvfsp.vfc_flags = vfsp->vfc_flags;
3585 	/*
3586 	 * These are unused in userland, we keep them
3587 	 * to not break binary compatibility.
3588 	 */
3589 	xvfsp.vfc_vfsops = NULL;
3590 	xvfsp.vfc_next = NULL;
3591 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3592 }
3593 
3594 #ifdef COMPAT_FREEBSD32
3595 struct xvfsconf32 {
3596 	uint32_t	vfc_vfsops;
3597 	char		vfc_name[MFSNAMELEN];
3598 	int32_t		vfc_typenum;
3599 	int32_t		vfc_refcount;
3600 	int32_t		vfc_flags;
3601 	uint32_t	vfc_next;
3602 };
3603 
3604 static int
3605 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
3606 {
3607 	struct xvfsconf32 xvfsp;
3608 
3609 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3610 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
3611 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
3612 	xvfsp.vfc_flags = vfsp->vfc_flags;
3613 	xvfsp.vfc_vfsops = 0;
3614 	xvfsp.vfc_next = 0;
3615 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3616 }
3617 #endif
3618 
3619 /*
3620  * Top level filesystem related information gathering.
3621  */
3622 static int
3623 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
3624 {
3625 	struct vfsconf *vfsp;
3626 	int error;
3627 
3628 	error = 0;
3629 	vfsconf_slock();
3630 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3631 #ifdef COMPAT_FREEBSD32
3632 		if (req->flags & SCTL_MASK32)
3633 			error = vfsconf2x32(req, vfsp);
3634 		else
3635 #endif
3636 			error = vfsconf2x(req, vfsp);
3637 		if (error)
3638 			break;
3639 	}
3640 	vfsconf_sunlock();
3641 	return (error);
3642 }
3643 
3644 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
3645     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
3646     "S,xvfsconf", "List of all configured filesystems");
3647 
3648 #ifndef BURN_BRIDGES
3649 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
3650 
3651 static int
3652 vfs_sysctl(SYSCTL_HANDLER_ARGS)
3653 {
3654 	int *name = (int *)arg1 - 1;	/* XXX */
3655 	u_int namelen = arg2 + 1;	/* XXX */
3656 	struct vfsconf *vfsp;
3657 
3658 	log(LOG_WARNING, "userland calling deprecated sysctl, "
3659 	    "please rebuild world\n");
3660 
3661 #if 1 || defined(COMPAT_PRELITE2)
3662 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
3663 	if (namelen == 1)
3664 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
3665 #endif
3666 
3667 	switch (name[1]) {
3668 	case VFS_MAXTYPENUM:
3669 		if (namelen != 2)
3670 			return (ENOTDIR);
3671 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
3672 	case VFS_CONF:
3673 		if (namelen != 3)
3674 			return (ENOTDIR);	/* overloaded */
3675 		vfsconf_slock();
3676 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3677 			if (vfsp->vfc_typenum == name[2])
3678 				break;
3679 		}
3680 		vfsconf_sunlock();
3681 		if (vfsp == NULL)
3682 			return (EOPNOTSUPP);
3683 #ifdef COMPAT_FREEBSD32
3684 		if (req->flags & SCTL_MASK32)
3685 			return (vfsconf2x32(req, vfsp));
3686 		else
3687 #endif
3688 			return (vfsconf2x(req, vfsp));
3689 	}
3690 	return (EOPNOTSUPP);
3691 }
3692 
3693 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
3694     CTLFLAG_MPSAFE, vfs_sysctl,
3695     "Generic filesystem");
3696 
3697 #if 1 || defined(COMPAT_PRELITE2)
3698 
3699 static int
3700 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
3701 {
3702 	int error;
3703 	struct vfsconf *vfsp;
3704 	struct ovfsconf ovfs;
3705 
3706 	vfsconf_slock();
3707 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3708 		bzero(&ovfs, sizeof(ovfs));
3709 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
3710 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
3711 		ovfs.vfc_index = vfsp->vfc_typenum;
3712 		ovfs.vfc_refcount = vfsp->vfc_refcount;
3713 		ovfs.vfc_flags = vfsp->vfc_flags;
3714 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
3715 		if (error != 0) {
3716 			vfsconf_sunlock();
3717 			return (error);
3718 		}
3719 	}
3720 	vfsconf_sunlock();
3721 	return (0);
3722 }
3723 
3724 #endif /* 1 || COMPAT_PRELITE2 */
3725 #endif /* !BURN_BRIDGES */
3726 
3727 #define KINFO_VNODESLOP		10
3728 #ifdef notyet
3729 /*
3730  * Dump vnode list (via sysctl).
3731  */
3732 /* ARGSUSED */
3733 static int
3734 sysctl_vnode(SYSCTL_HANDLER_ARGS)
3735 {
3736 	struct xvnode *xvn;
3737 	struct mount *mp;
3738 	struct vnode *vp;
3739 	int error, len, n;
3740 
3741 	/*
3742 	 * Stale numvnodes access is not fatal here.
3743 	 */
3744 	req->lock = 0;
3745 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
3746 	if (!req->oldptr)
3747 		/* Make an estimate */
3748 		return (SYSCTL_OUT(req, 0, len));
3749 
3750 	error = sysctl_wire_old_buffer(req, 0);
3751 	if (error != 0)
3752 		return (error);
3753 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
3754 	n = 0;
3755 	mtx_lock(&mountlist_mtx);
3756 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3757 		if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
3758 			continue;
3759 		MNT_ILOCK(mp);
3760 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3761 			if (n == len)
3762 				break;
3763 			vref(vp);
3764 			xvn[n].xv_size = sizeof *xvn;
3765 			xvn[n].xv_vnode = vp;
3766 			xvn[n].xv_id = 0;	/* XXX compat */
3767 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
3768 			XV_COPY(usecount);
3769 			XV_COPY(writecount);
3770 			XV_COPY(holdcnt);
3771 			XV_COPY(mount);
3772 			XV_COPY(numoutput);
3773 			XV_COPY(type);
3774 #undef XV_COPY
3775 			xvn[n].xv_flag = vp->v_vflag;
3776 
3777 			switch (vp->v_type) {
3778 			case VREG:
3779 			case VDIR:
3780 			case VLNK:
3781 				break;
3782 			case VBLK:
3783 			case VCHR:
3784 				if (vp->v_rdev == NULL) {
3785 					vrele(vp);
3786 					continue;
3787 				}
3788 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
3789 				break;
3790 			case VSOCK:
3791 				xvn[n].xv_socket = vp->v_socket;
3792 				break;
3793 			case VFIFO:
3794 				xvn[n].xv_fifo = vp->v_fifoinfo;
3795 				break;
3796 			case VNON:
3797 			case VBAD:
3798 			default:
3799 				/* shouldn't happen? */
3800 				vrele(vp);
3801 				continue;
3802 			}
3803 			vrele(vp);
3804 			++n;
3805 		}
3806 		MNT_IUNLOCK(mp);
3807 		mtx_lock(&mountlist_mtx);
3808 		vfs_unbusy(mp);
3809 		if (n == len)
3810 			break;
3811 	}
3812 	mtx_unlock(&mountlist_mtx);
3813 
3814 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
3815 	free(xvn, M_TEMP);
3816 	return (error);
3817 }
3818 
3819 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
3820     CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
3821     "");
3822 #endif
3823 
3824 static void
3825 unmount_or_warn(struct mount *mp)
3826 {
3827 	int error;
3828 
3829 	error = dounmount(mp, MNT_FORCE, curthread);
3830 	if (error != 0) {
3831 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
3832 		if (error == EBUSY)
3833 			printf("BUSY)\n");
3834 		else
3835 			printf("%d)\n", error);
3836 	}
3837 }
3838 
3839 /*
3840  * Unmount all filesystems. The list is traversed in reverse order
3841  * of mounting to avoid dependencies.
3842  */
3843 void
3844 vfs_unmountall(void)
3845 {
3846 	struct mount *mp, *tmp;
3847 
3848 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
3849 
3850 	/*
3851 	 * Since this only runs when rebooting, it is not interlocked.
3852 	 */
3853 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
3854 		vfs_ref(mp);
3855 
3856 		/*
3857 		 * Forcibly unmounting "/dev" before "/" would prevent clean
3858 		 * unmount of the latter.
3859 		 */
3860 		if (mp == rootdevmp)
3861 			continue;
3862 
3863 		unmount_or_warn(mp);
3864 	}
3865 
3866 	if (rootdevmp != NULL)
3867 		unmount_or_warn(rootdevmp);
3868 }
3869 
3870 /*
3871  * perform msync on all vnodes under a mount point
3872  * the mount point must be locked.
3873  */
3874 void
3875 vfs_msync(struct mount *mp, int flags)
3876 {
3877 	struct vnode *vp, *mvp;
3878 	struct vm_object *obj;
3879 
3880 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
3881 	MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
3882 		obj = vp->v_object;
3883 		if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 &&
3884 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
3885 			if (!vget(vp,
3886 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
3887 			    curthread)) {
3888 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
3889 					vput(vp);
3890 					continue;
3891 				}
3892 
3893 				obj = vp->v_object;
3894 				if (obj != NULL) {
3895 					VM_OBJECT_WLOCK(obj);
3896 					vm_object_page_clean(obj, 0, 0,
3897 					    flags == MNT_WAIT ?
3898 					    OBJPC_SYNC : OBJPC_NOSYNC);
3899 					VM_OBJECT_WUNLOCK(obj);
3900 				}
3901 				vput(vp);
3902 			}
3903 		} else
3904 			VI_UNLOCK(vp);
3905 	}
3906 }
3907 
3908 static void
3909 destroy_vpollinfo_free(struct vpollinfo *vi)
3910 {
3911 
3912 	knlist_destroy(&vi->vpi_selinfo.si_note);
3913 	mtx_destroy(&vi->vpi_lock);
3914 	uma_zfree(vnodepoll_zone, vi);
3915 }
3916 
3917 static void
3918 destroy_vpollinfo(struct vpollinfo *vi)
3919 {
3920 
3921 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
3922 	seldrain(&vi->vpi_selinfo);
3923 	destroy_vpollinfo_free(vi);
3924 }
3925 
3926 /*
3927  * Initalize per-vnode helper structure to hold poll-related state.
3928  */
3929 void
3930 v_addpollinfo(struct vnode *vp)
3931 {
3932 	struct vpollinfo *vi;
3933 
3934 	if (vp->v_pollinfo != NULL)
3935 		return;
3936 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
3937 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
3938 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
3939 	    vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
3940 	VI_LOCK(vp);
3941 	if (vp->v_pollinfo != NULL) {
3942 		VI_UNLOCK(vp);
3943 		destroy_vpollinfo_free(vi);
3944 		return;
3945 	}
3946 	vp->v_pollinfo = vi;
3947 	VI_UNLOCK(vp);
3948 }
3949 
3950 /*
3951  * Record a process's interest in events which might happen to
3952  * a vnode.  Because poll uses the historic select-style interface
3953  * internally, this routine serves as both the ``check for any
3954  * pending events'' and the ``record my interest in future events''
3955  * functions.  (These are done together, while the lock is held,
3956  * to avoid race conditions.)
3957  */
3958 int
3959 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
3960 {
3961 
3962 	v_addpollinfo(vp);
3963 	mtx_lock(&vp->v_pollinfo->vpi_lock);
3964 	if (vp->v_pollinfo->vpi_revents & events) {
3965 		/*
3966 		 * This leaves events we are not interested
3967 		 * in available for the other process which
3968 		 * which presumably had requested them
3969 		 * (otherwise they would never have been
3970 		 * recorded).
3971 		 */
3972 		events &= vp->v_pollinfo->vpi_revents;
3973 		vp->v_pollinfo->vpi_revents &= ~events;
3974 
3975 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
3976 		return (events);
3977 	}
3978 	vp->v_pollinfo->vpi_events |= events;
3979 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
3980 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3981 	return (0);
3982 }
3983 
3984 /*
3985  * Routine to create and manage a filesystem syncer vnode.
3986  */
3987 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
3988 static int	sync_fsync(struct  vop_fsync_args *);
3989 static int	sync_inactive(struct  vop_inactive_args *);
3990 static int	sync_reclaim(struct  vop_reclaim_args *);
3991 
3992 static struct vop_vector sync_vnodeops = {
3993 	.vop_bypass =	VOP_EOPNOTSUPP,
3994 	.vop_close =	sync_close,		/* close */
3995 	.vop_fsync =	sync_fsync,		/* fsync */
3996 	.vop_inactive =	sync_inactive,	/* inactive */
3997 	.vop_reclaim =	sync_reclaim,	/* reclaim */
3998 	.vop_lock1 =	vop_stdlock,	/* lock */
3999 	.vop_unlock =	vop_stdunlock,	/* unlock */
4000 	.vop_islocked =	vop_stdislocked,	/* islocked */
4001 };
4002 
4003 /*
4004  * Create a new filesystem syncer vnode for the specified mount point.
4005  */
4006 void
4007 vfs_allocate_syncvnode(struct mount *mp)
4008 {
4009 	struct vnode *vp;
4010 	struct bufobj *bo;
4011 	static long start, incr, next;
4012 	int error;
4013 
4014 	/* Allocate a new vnode */
4015 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4016 	if (error != 0)
4017 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
4018 	vp->v_type = VNON;
4019 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4020 	vp->v_vflag |= VV_FORCEINSMQ;
4021 	error = insmntque(vp, mp);
4022 	if (error != 0)
4023 		panic("vfs_allocate_syncvnode: insmntque() failed");
4024 	vp->v_vflag &= ~VV_FORCEINSMQ;
4025 	VOP_UNLOCK(vp, 0);
4026 	/*
4027 	 * Place the vnode onto the syncer worklist. We attempt to
4028 	 * scatter them about on the list so that they will go off
4029 	 * at evenly distributed times even if all the filesystems
4030 	 * are mounted at once.
4031 	 */
4032 	next += incr;
4033 	if (next == 0 || next > syncer_maxdelay) {
4034 		start /= 2;
4035 		incr /= 2;
4036 		if (start == 0) {
4037 			start = syncer_maxdelay / 2;
4038 			incr = syncer_maxdelay;
4039 		}
4040 		next = start;
4041 	}
4042 	bo = &vp->v_bufobj;
4043 	BO_LOCK(bo);
4044 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4045 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4046 	mtx_lock(&sync_mtx);
4047 	sync_vnode_count++;
4048 	if (mp->mnt_syncer == NULL) {
4049 		mp->mnt_syncer = vp;
4050 		vp = NULL;
4051 	}
4052 	mtx_unlock(&sync_mtx);
4053 	BO_UNLOCK(bo);
4054 	if (vp != NULL) {
4055 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4056 		vgone(vp);
4057 		vput(vp);
4058 	}
4059 }
4060 
4061 void
4062 vfs_deallocate_syncvnode(struct mount *mp)
4063 {
4064 	struct vnode *vp;
4065 
4066 	mtx_lock(&sync_mtx);
4067 	vp = mp->mnt_syncer;
4068 	if (vp != NULL)
4069 		mp->mnt_syncer = NULL;
4070 	mtx_unlock(&sync_mtx);
4071 	if (vp != NULL)
4072 		vrele(vp);
4073 }
4074 
4075 /*
4076  * Do a lazy sync of the filesystem.
4077  */
4078 static int
4079 sync_fsync(struct vop_fsync_args *ap)
4080 {
4081 	struct vnode *syncvp = ap->a_vp;
4082 	struct mount *mp = syncvp->v_mount;
4083 	int error, save;
4084 	struct bufobj *bo;
4085 
4086 	/*
4087 	 * We only need to do something if this is a lazy evaluation.
4088 	 */
4089 	if (ap->a_waitfor != MNT_LAZY)
4090 		return (0);
4091 
4092 	/*
4093 	 * Move ourselves to the back of the sync list.
4094 	 */
4095 	bo = &syncvp->v_bufobj;
4096 	BO_LOCK(bo);
4097 	vn_syncer_add_to_worklist(bo, syncdelay);
4098 	BO_UNLOCK(bo);
4099 
4100 	/*
4101 	 * Walk the list of vnodes pushing all that are dirty and
4102 	 * not already on the sync list.
4103 	 */
4104 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
4105 		return (0);
4106 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4107 		vfs_unbusy(mp);
4108 		return (0);
4109 	}
4110 	save = curthread_pflags_set(TDP_SYNCIO);
4111 	vfs_msync(mp, MNT_NOWAIT);
4112 	error = VFS_SYNC(mp, MNT_LAZY);
4113 	curthread_pflags_restore(save);
4114 	vn_finished_write(mp);
4115 	vfs_unbusy(mp);
4116 	return (error);
4117 }
4118 
4119 /*
4120  * The syncer vnode is no referenced.
4121  */
4122 static int
4123 sync_inactive(struct vop_inactive_args *ap)
4124 {
4125 
4126 	vgone(ap->a_vp);
4127 	return (0);
4128 }
4129 
4130 /*
4131  * The syncer vnode is no longer needed and is being decommissioned.
4132  *
4133  * Modifications to the worklist must be protected by sync_mtx.
4134  */
4135 static int
4136 sync_reclaim(struct vop_reclaim_args *ap)
4137 {
4138 	struct vnode *vp = ap->a_vp;
4139 	struct bufobj *bo;
4140 
4141 	bo = &vp->v_bufobj;
4142 	BO_LOCK(bo);
4143 	mtx_lock(&sync_mtx);
4144 	if (vp->v_mount->mnt_syncer == vp)
4145 		vp->v_mount->mnt_syncer = NULL;
4146 	if (bo->bo_flag & BO_ONWORKLST) {
4147 		LIST_REMOVE(bo, bo_synclist);
4148 		syncer_worklist_len--;
4149 		sync_vnode_count--;
4150 		bo->bo_flag &= ~BO_ONWORKLST;
4151 	}
4152 	mtx_unlock(&sync_mtx);
4153 	BO_UNLOCK(bo);
4154 
4155 	return (0);
4156 }
4157 
4158 /*
4159  * Check if vnode represents a disk device
4160  */
4161 int
4162 vn_isdisk(struct vnode *vp, int *errp)
4163 {
4164 	int error;
4165 
4166 	if (vp->v_type != VCHR) {
4167 		error = ENOTBLK;
4168 		goto out;
4169 	}
4170 	error = 0;
4171 	dev_lock();
4172 	if (vp->v_rdev == NULL)
4173 		error = ENXIO;
4174 	else if (vp->v_rdev->si_devsw == NULL)
4175 		error = ENXIO;
4176 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
4177 		error = ENOTBLK;
4178 	dev_unlock();
4179 out:
4180 	if (errp != NULL)
4181 		*errp = error;
4182 	return (error == 0);
4183 }
4184 
4185 /*
4186  * Common filesystem object access control check routine.  Accepts a
4187  * vnode's type, "mode", uid and gid, requested access mode, credentials,
4188  * and optional call-by-reference privused argument allowing vaccess()
4189  * to indicate to the caller whether privilege was used to satisfy the
4190  * request (obsoleted).  Returns 0 on success, or an errno on failure.
4191  */
4192 int
4193 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
4194     accmode_t accmode, struct ucred *cred, int *privused)
4195 {
4196 	accmode_t dac_granted;
4197 	accmode_t priv_granted;
4198 
4199 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
4200 	    ("invalid bit in accmode"));
4201 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
4202 	    ("VAPPEND without VWRITE"));
4203 
4204 	/*
4205 	 * Look for a normal, non-privileged way to access the file/directory
4206 	 * as requested.  If it exists, go with that.
4207 	 */
4208 
4209 	if (privused != NULL)
4210 		*privused = 0;
4211 
4212 	dac_granted = 0;
4213 
4214 	/* Check the owner. */
4215 	if (cred->cr_uid == file_uid) {
4216 		dac_granted |= VADMIN;
4217 		if (file_mode & S_IXUSR)
4218 			dac_granted |= VEXEC;
4219 		if (file_mode & S_IRUSR)
4220 			dac_granted |= VREAD;
4221 		if (file_mode & S_IWUSR)
4222 			dac_granted |= (VWRITE | VAPPEND);
4223 
4224 		if ((accmode & dac_granted) == accmode)
4225 			return (0);
4226 
4227 		goto privcheck;
4228 	}
4229 
4230 	/* Otherwise, check the groups (first match) */
4231 	if (groupmember(file_gid, cred)) {
4232 		if (file_mode & S_IXGRP)
4233 			dac_granted |= VEXEC;
4234 		if (file_mode & S_IRGRP)
4235 			dac_granted |= VREAD;
4236 		if (file_mode & S_IWGRP)
4237 			dac_granted |= (VWRITE | VAPPEND);
4238 
4239 		if ((accmode & dac_granted) == accmode)
4240 			return (0);
4241 
4242 		goto privcheck;
4243 	}
4244 
4245 	/* Otherwise, check everyone else. */
4246 	if (file_mode & S_IXOTH)
4247 		dac_granted |= VEXEC;
4248 	if (file_mode & S_IROTH)
4249 		dac_granted |= VREAD;
4250 	if (file_mode & S_IWOTH)
4251 		dac_granted |= (VWRITE | VAPPEND);
4252 	if ((accmode & dac_granted) == accmode)
4253 		return (0);
4254 
4255 privcheck:
4256 	/*
4257 	 * Build a privilege mask to determine if the set of privileges
4258 	 * satisfies the requirements when combined with the granted mask
4259 	 * from above.  For each privilege, if the privilege is required,
4260 	 * bitwise or the request type onto the priv_granted mask.
4261 	 */
4262 	priv_granted = 0;
4263 
4264 	if (type == VDIR) {
4265 		/*
4266 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
4267 		 * requests, instead of PRIV_VFS_EXEC.
4268 		 */
4269 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4270 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP, 0))
4271 			priv_granted |= VEXEC;
4272 	} else {
4273 		/*
4274 		 * Ensure that at least one execute bit is on. Otherwise,
4275 		 * a privileged user will always succeed, and we don't want
4276 		 * this to happen unless the file really is executable.
4277 		 */
4278 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4279 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
4280 		    !priv_check_cred(cred, PRIV_VFS_EXEC, 0))
4281 			priv_granted |= VEXEC;
4282 	}
4283 
4284 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
4285 	    !priv_check_cred(cred, PRIV_VFS_READ, 0))
4286 		priv_granted |= VREAD;
4287 
4288 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
4289 	    !priv_check_cred(cred, PRIV_VFS_WRITE, 0))
4290 		priv_granted |= (VWRITE | VAPPEND);
4291 
4292 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
4293 	    !priv_check_cred(cred, PRIV_VFS_ADMIN, 0))
4294 		priv_granted |= VADMIN;
4295 
4296 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
4297 		/* XXX audit: privilege used */
4298 		if (privused != NULL)
4299 			*privused = 1;
4300 		return (0);
4301 	}
4302 
4303 	return ((accmode & VADMIN) ? EPERM : EACCES);
4304 }
4305 
4306 /*
4307  * Credential check based on process requesting service, and per-attribute
4308  * permissions.
4309  */
4310 int
4311 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
4312     struct thread *td, accmode_t accmode)
4313 {
4314 
4315 	/*
4316 	 * Kernel-invoked always succeeds.
4317 	 */
4318 	if (cred == NOCRED)
4319 		return (0);
4320 
4321 	/*
4322 	 * Do not allow privileged processes in jail to directly manipulate
4323 	 * system attributes.
4324 	 */
4325 	switch (attrnamespace) {
4326 	case EXTATTR_NAMESPACE_SYSTEM:
4327 		/* Potentially should be: return (EPERM); */
4328 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0));
4329 	case EXTATTR_NAMESPACE_USER:
4330 		return (VOP_ACCESS(vp, accmode, cred, td));
4331 	default:
4332 		return (EPERM);
4333 	}
4334 }
4335 
4336 #ifdef DEBUG_VFS_LOCKS
4337 /*
4338  * This only exists to supress warnings from unlocked specfs accesses.  It is
4339  * no longer ok to have an unlocked VFS.
4340  */
4341 #define	IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL ||		\
4342 	(vp)->v_type == VCHR ||	(vp)->v_type == VBAD)
4343 
4344 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
4345 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
4346     "Drop into debugger on lock violation");
4347 
4348 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
4349 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
4350     0, "Check for interlock across VOPs");
4351 
4352 int vfs_badlock_print = 1;	/* Print lock violations. */
4353 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
4354     0, "Print lock violations");
4355 
4356 #ifdef KDB
4357 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
4358 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
4359     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
4360 #endif
4361 
4362 static void
4363 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
4364 {
4365 
4366 #ifdef KDB
4367 	if (vfs_badlock_backtrace)
4368 		kdb_backtrace();
4369 #endif
4370 	if (vfs_badlock_print)
4371 		printf("%s: %p %s\n", str, (void *)vp, msg);
4372 	if (vfs_badlock_ddb)
4373 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4374 }
4375 
4376 void
4377 assert_vi_locked(struct vnode *vp, const char *str)
4378 {
4379 
4380 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
4381 		vfs_badlock("interlock is not locked but should be", str, vp);
4382 }
4383 
4384 void
4385 assert_vi_unlocked(struct vnode *vp, const char *str)
4386 {
4387 
4388 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
4389 		vfs_badlock("interlock is locked but should not be", str, vp);
4390 }
4391 
4392 void
4393 assert_vop_locked(struct vnode *vp, const char *str)
4394 {
4395 	int locked;
4396 
4397 	if (!IGNORE_LOCK(vp)) {
4398 		locked = VOP_ISLOCKED(vp);
4399 		if (locked == 0 || locked == LK_EXCLOTHER)
4400 			vfs_badlock("is not locked but should be", str, vp);
4401 	}
4402 }
4403 
4404 void
4405 assert_vop_unlocked(struct vnode *vp, const char *str)
4406 {
4407 
4408 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
4409 		vfs_badlock("is locked but should not be", str, vp);
4410 }
4411 
4412 void
4413 assert_vop_elocked(struct vnode *vp, const char *str)
4414 {
4415 
4416 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
4417 		vfs_badlock("is not exclusive locked but should be", str, vp);
4418 }
4419 
4420 #if 0
4421 void
4422 assert_vop_elocked_other(struct vnode *vp, const char *str)
4423 {
4424 
4425 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLOTHER)
4426 		vfs_badlock("is not exclusive locked by another thread",
4427 		    str, vp);
4428 }
4429 
4430 void
4431 assert_vop_slocked(struct vnode *vp, const char *str)
4432 {
4433 
4434 	if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_SHARED)
4435 		vfs_badlock("is not locked shared but should be", str, vp);
4436 }
4437 #endif /* 0 */
4438 #endif /* DEBUG_VFS_LOCKS */
4439 
4440 void
4441 vop_rename_fail(struct vop_rename_args *ap)
4442 {
4443 
4444 	if (ap->a_tvp != NULL)
4445 		vput(ap->a_tvp);
4446 	if (ap->a_tdvp == ap->a_tvp)
4447 		vrele(ap->a_tdvp);
4448 	else
4449 		vput(ap->a_tdvp);
4450 	vrele(ap->a_fdvp);
4451 	vrele(ap->a_fvp);
4452 }
4453 
4454 void
4455 vop_rename_pre(void *ap)
4456 {
4457 	struct vop_rename_args *a = ap;
4458 
4459 #ifdef DEBUG_VFS_LOCKS
4460 	if (a->a_tvp)
4461 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
4462 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
4463 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
4464 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
4465 
4466 	/* Check the source (from). */
4467 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
4468 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
4469 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
4470 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
4471 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
4472 
4473 	/* Check the target. */
4474 	if (a->a_tvp)
4475 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
4476 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
4477 #endif
4478 	if (a->a_tdvp != a->a_fdvp)
4479 		vhold(a->a_fdvp);
4480 	if (a->a_tvp != a->a_fvp)
4481 		vhold(a->a_fvp);
4482 	vhold(a->a_tdvp);
4483 	if (a->a_tvp)
4484 		vhold(a->a_tvp);
4485 }
4486 
4487 void
4488 vop_strategy_pre(void *ap)
4489 {
4490 #ifdef DEBUG_VFS_LOCKS
4491 	struct vop_strategy_args *a;
4492 	struct buf *bp;
4493 
4494 	a = ap;
4495 	bp = a->a_bp;
4496 
4497 	/*
4498 	 * Cluster ops lock their component buffers but not the IO container.
4499 	 */
4500 	if ((bp->b_flags & B_CLUSTER) != 0)
4501 		return;
4502 
4503 	if (panicstr == NULL && !BUF_ISLOCKED(bp)) {
4504 		if (vfs_badlock_print)
4505 			printf(
4506 			    "VOP_STRATEGY: bp is not locked but should be\n");
4507 		if (vfs_badlock_ddb)
4508 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4509 	}
4510 #endif
4511 }
4512 
4513 void
4514 vop_lock_pre(void *ap)
4515 {
4516 #ifdef DEBUG_VFS_LOCKS
4517 	struct vop_lock1_args *a = ap;
4518 
4519 	if ((a->a_flags & LK_INTERLOCK) == 0)
4520 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4521 	else
4522 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
4523 #endif
4524 }
4525 
4526 void
4527 vop_lock_post(void *ap, int rc)
4528 {
4529 #ifdef DEBUG_VFS_LOCKS
4530 	struct vop_lock1_args *a = ap;
4531 
4532 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4533 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
4534 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
4535 #endif
4536 }
4537 
4538 void
4539 vop_unlock_pre(void *ap)
4540 {
4541 #ifdef DEBUG_VFS_LOCKS
4542 	struct vop_unlock_args *a = ap;
4543 
4544 	if (a->a_flags & LK_INTERLOCK)
4545 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
4546 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
4547 #endif
4548 }
4549 
4550 void
4551 vop_unlock_post(void *ap, int rc)
4552 {
4553 #ifdef DEBUG_VFS_LOCKS
4554 	struct vop_unlock_args *a = ap;
4555 
4556 	if (a->a_flags & LK_INTERLOCK)
4557 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
4558 #endif
4559 }
4560 
4561 void
4562 vop_create_post(void *ap, int rc)
4563 {
4564 	struct vop_create_args *a = ap;
4565 
4566 	if (!rc)
4567 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4568 }
4569 
4570 void
4571 vop_deleteextattr_post(void *ap, int rc)
4572 {
4573 	struct vop_deleteextattr_args *a = ap;
4574 
4575 	if (!rc)
4576 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4577 }
4578 
4579 void
4580 vop_link_post(void *ap, int rc)
4581 {
4582 	struct vop_link_args *a = ap;
4583 
4584 	if (!rc) {
4585 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
4586 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
4587 	}
4588 }
4589 
4590 void
4591 vop_mkdir_post(void *ap, int rc)
4592 {
4593 	struct vop_mkdir_args *a = ap;
4594 
4595 	if (!rc)
4596 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4597 }
4598 
4599 void
4600 vop_mknod_post(void *ap, int rc)
4601 {
4602 	struct vop_mknod_args *a = ap;
4603 
4604 	if (!rc)
4605 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4606 }
4607 
4608 void
4609 vop_reclaim_post(void *ap, int rc)
4610 {
4611 	struct vop_reclaim_args *a = ap;
4612 
4613 	if (!rc)
4614 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
4615 }
4616 
4617 void
4618 vop_remove_post(void *ap, int rc)
4619 {
4620 	struct vop_remove_args *a = ap;
4621 
4622 	if (!rc) {
4623 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4624 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4625 	}
4626 }
4627 
4628 void
4629 vop_rename_post(void *ap, int rc)
4630 {
4631 	struct vop_rename_args *a = ap;
4632 
4633 	if (!rc) {
4634 		VFS_KNOTE_UNLOCKED(a->a_fdvp, NOTE_WRITE);
4635 		VFS_KNOTE_UNLOCKED(a->a_tdvp, NOTE_WRITE);
4636 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
4637 		if (a->a_tvp)
4638 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
4639 	}
4640 	if (a->a_tdvp != a->a_fdvp)
4641 		vdrop(a->a_fdvp);
4642 	if (a->a_tvp != a->a_fvp)
4643 		vdrop(a->a_fvp);
4644 	vdrop(a->a_tdvp);
4645 	if (a->a_tvp)
4646 		vdrop(a->a_tvp);
4647 }
4648 
4649 void
4650 vop_rmdir_post(void *ap, int rc)
4651 {
4652 	struct vop_rmdir_args *a = ap;
4653 
4654 	if (!rc) {
4655 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4656 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4657 	}
4658 }
4659 
4660 void
4661 vop_setattr_post(void *ap, int rc)
4662 {
4663 	struct vop_setattr_args *a = ap;
4664 
4665 	if (!rc)
4666 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4667 }
4668 
4669 void
4670 vop_setextattr_post(void *ap, int rc)
4671 {
4672 	struct vop_setextattr_args *a = ap;
4673 
4674 	if (!rc)
4675 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4676 }
4677 
4678 void
4679 vop_symlink_post(void *ap, int rc)
4680 {
4681 	struct vop_symlink_args *a = ap;
4682 
4683 	if (!rc)
4684 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4685 }
4686 
4687 static struct knlist fs_knlist;
4688 
4689 static void
4690 vfs_event_init(void *arg)
4691 {
4692 	knlist_init_mtx(&fs_knlist, NULL);
4693 }
4694 /* XXX - correct order? */
4695 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
4696 
4697 void
4698 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
4699 {
4700 
4701 	KNOTE_UNLOCKED(&fs_knlist, event);
4702 }
4703 
4704 static int	filt_fsattach(struct knote *kn);
4705 static void	filt_fsdetach(struct knote *kn);
4706 static int	filt_fsevent(struct knote *kn, long hint);
4707 
4708 struct filterops fs_filtops = {
4709 	.f_isfd = 0,
4710 	.f_attach = filt_fsattach,
4711 	.f_detach = filt_fsdetach,
4712 	.f_event = filt_fsevent
4713 };
4714 
4715 static int
4716 filt_fsattach(struct knote *kn)
4717 {
4718 
4719 	kn->kn_flags |= EV_CLEAR;
4720 	knlist_add(&fs_knlist, kn, 0);
4721 	return (0);
4722 }
4723 
4724 static void
4725 filt_fsdetach(struct knote *kn)
4726 {
4727 
4728 	knlist_remove(&fs_knlist, kn, 0);
4729 }
4730 
4731 static int
4732 filt_fsevent(struct knote *kn, long hint)
4733 {
4734 
4735 	kn->kn_fflags |= hint;
4736 	return (kn->kn_fflags != 0);
4737 }
4738 
4739 static int
4740 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
4741 {
4742 	struct vfsidctl vc;
4743 	int error;
4744 	struct mount *mp;
4745 
4746 	error = SYSCTL_IN(req, &vc, sizeof(vc));
4747 	if (error)
4748 		return (error);
4749 	if (vc.vc_vers != VFS_CTL_VERS1)
4750 		return (EINVAL);
4751 	mp = vfs_getvfs(&vc.vc_fsid);
4752 	if (mp == NULL)
4753 		return (ENOENT);
4754 	/* ensure that a specific sysctl goes to the right filesystem. */
4755 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
4756 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
4757 		vfs_rel(mp);
4758 		return (EINVAL);
4759 	}
4760 	VCTLTOREQ(&vc, req);
4761 	error = VFS_SYSCTL(mp, vc.vc_op, req);
4762 	vfs_rel(mp);
4763 	return (error);
4764 }
4765 
4766 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR,
4767     NULL, 0, sysctl_vfs_ctl, "",
4768     "Sysctl by fsid");
4769 
4770 /*
4771  * Function to initialize a va_filerev field sensibly.
4772  * XXX: Wouldn't a random number make a lot more sense ??
4773  */
4774 u_quad_t
4775 init_va_filerev(void)
4776 {
4777 	struct bintime bt;
4778 
4779 	getbinuptime(&bt);
4780 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
4781 }
4782 
4783 static int	filt_vfsread(struct knote *kn, long hint);
4784 static int	filt_vfswrite(struct knote *kn, long hint);
4785 static int	filt_vfsvnode(struct knote *kn, long hint);
4786 static void	filt_vfsdetach(struct knote *kn);
4787 static struct filterops vfsread_filtops = {
4788 	.f_isfd = 1,
4789 	.f_detach = filt_vfsdetach,
4790 	.f_event = filt_vfsread
4791 };
4792 static struct filterops vfswrite_filtops = {
4793 	.f_isfd = 1,
4794 	.f_detach = filt_vfsdetach,
4795 	.f_event = filt_vfswrite
4796 };
4797 static struct filterops vfsvnode_filtops = {
4798 	.f_isfd = 1,
4799 	.f_detach = filt_vfsdetach,
4800 	.f_event = filt_vfsvnode
4801 };
4802 
4803 static void
4804 vfs_knllock(void *arg)
4805 {
4806 	struct vnode *vp = arg;
4807 
4808 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4809 }
4810 
4811 static void
4812 vfs_knlunlock(void *arg)
4813 {
4814 	struct vnode *vp = arg;
4815 
4816 	VOP_UNLOCK(vp, 0);
4817 }
4818 
4819 static void
4820 vfs_knl_assert_locked(void *arg)
4821 {
4822 #ifdef DEBUG_VFS_LOCKS
4823 	struct vnode *vp = arg;
4824 
4825 	ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
4826 #endif
4827 }
4828 
4829 static void
4830 vfs_knl_assert_unlocked(void *arg)
4831 {
4832 #ifdef DEBUG_VFS_LOCKS
4833 	struct vnode *vp = arg;
4834 
4835 	ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
4836 #endif
4837 }
4838 
4839 int
4840 vfs_kqfilter(struct vop_kqfilter_args *ap)
4841 {
4842 	struct vnode *vp = ap->a_vp;
4843 	struct knote *kn = ap->a_kn;
4844 	struct knlist *knl;
4845 
4846 	switch (kn->kn_filter) {
4847 	case EVFILT_READ:
4848 		kn->kn_fop = &vfsread_filtops;
4849 		break;
4850 	case EVFILT_WRITE:
4851 		kn->kn_fop = &vfswrite_filtops;
4852 		break;
4853 	case EVFILT_VNODE:
4854 		kn->kn_fop = &vfsvnode_filtops;
4855 		break;
4856 	default:
4857 		return (EINVAL);
4858 	}
4859 
4860 	kn->kn_hook = (caddr_t)vp;
4861 
4862 	v_addpollinfo(vp);
4863 	if (vp->v_pollinfo == NULL)
4864 		return (ENOMEM);
4865 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
4866 	vhold(vp);
4867 	knlist_add(knl, kn, 0);
4868 
4869 	return (0);
4870 }
4871 
4872 /*
4873  * Detach knote from vnode
4874  */
4875 static void
4876 filt_vfsdetach(struct knote *kn)
4877 {
4878 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4879 
4880 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
4881 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
4882 	vdrop(vp);
4883 }
4884 
4885 /*ARGSUSED*/
4886 static int
4887 filt_vfsread(struct knote *kn, long hint)
4888 {
4889 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4890 	struct vattr va;
4891 	int res;
4892 
4893 	/*
4894 	 * filesystem is gone, so set the EOF flag and schedule
4895 	 * the knote for deletion.
4896 	 */
4897 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
4898 		VI_LOCK(vp);
4899 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
4900 		VI_UNLOCK(vp);
4901 		return (1);
4902 	}
4903 
4904 	if (VOP_GETATTR(vp, &va, curthread->td_ucred))
4905 		return (0);
4906 
4907 	VI_LOCK(vp);
4908 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
4909 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
4910 	VI_UNLOCK(vp);
4911 	return (res);
4912 }
4913 
4914 /*ARGSUSED*/
4915 static int
4916 filt_vfswrite(struct knote *kn, long hint)
4917 {
4918 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4919 
4920 	VI_LOCK(vp);
4921 
4922 	/*
4923 	 * filesystem is gone, so set the EOF flag and schedule
4924 	 * the knote for deletion.
4925 	 */
4926 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
4927 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
4928 
4929 	kn->kn_data = 0;
4930 	VI_UNLOCK(vp);
4931 	return (1);
4932 }
4933 
4934 static int
4935 filt_vfsvnode(struct knote *kn, long hint)
4936 {
4937 	struct vnode *vp = (struct vnode *)kn->kn_hook;
4938 	int res;
4939 
4940 	VI_LOCK(vp);
4941 	if (kn->kn_sfflags & hint)
4942 		kn->kn_fflags |= hint;
4943 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
4944 		kn->kn_flags |= EV_EOF;
4945 		VI_UNLOCK(vp);
4946 		return (1);
4947 	}
4948 	res = (kn->kn_fflags != 0);
4949 	VI_UNLOCK(vp);
4950 	return (res);
4951 }
4952 
4953 int
4954 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
4955 {
4956 	int error;
4957 
4958 	if (dp->d_reclen > ap->a_uio->uio_resid)
4959 		return (ENAMETOOLONG);
4960 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
4961 	if (error) {
4962 		if (ap->a_ncookies != NULL) {
4963 			if (ap->a_cookies != NULL)
4964 				free(ap->a_cookies, M_TEMP);
4965 			ap->a_cookies = NULL;
4966 			*ap->a_ncookies = 0;
4967 		}
4968 		return (error);
4969 	}
4970 	if (ap->a_ncookies == NULL)
4971 		return (0);
4972 
4973 	KASSERT(ap->a_cookies,
4974 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
4975 
4976 	*ap->a_cookies = realloc(*ap->a_cookies,
4977 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
4978 	(*ap->a_cookies)[*ap->a_ncookies] = off;
4979 	return (0);
4980 }
4981 
4982 /*
4983  * Mark for update the access time of the file if the filesystem
4984  * supports VOP_MARKATIME.  This functionality is used by execve and
4985  * mmap, so we want to avoid the I/O implied by directly setting
4986  * va_atime for the sake of efficiency.
4987  */
4988 void
4989 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
4990 {
4991 	struct mount *mp;
4992 
4993 	mp = vp->v_mount;
4994 	ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
4995 	if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
4996 		(void)VOP_MARKATIME(vp);
4997 }
4998 
4999 /*
5000  * The purpose of this routine is to remove granularity from accmode_t,
5001  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
5002  * VADMIN and VAPPEND.
5003  *
5004  * If it returns 0, the caller is supposed to continue with the usual
5005  * access checks using 'accmode' as modified by this routine.  If it
5006  * returns nonzero value, the caller is supposed to return that value
5007  * as errno.
5008  *
5009  * Note that after this routine runs, accmode may be zero.
5010  */
5011 int
5012 vfs_unixify_accmode(accmode_t *accmode)
5013 {
5014 	/*
5015 	 * There is no way to specify explicit "deny" rule using
5016 	 * file mode or POSIX.1e ACLs.
5017 	 */
5018 	if (*accmode & VEXPLICIT_DENY) {
5019 		*accmode = 0;
5020 		return (0);
5021 	}
5022 
5023 	/*
5024 	 * None of these can be translated into usual access bits.
5025 	 * Also, the common case for NFSv4 ACLs is to not contain
5026 	 * either of these bits. Caller should check for VWRITE
5027 	 * on the containing directory instead.
5028 	 */
5029 	if (*accmode & (VDELETE_CHILD | VDELETE))
5030 		return (EPERM);
5031 
5032 	if (*accmode & VADMIN_PERMS) {
5033 		*accmode &= ~VADMIN_PERMS;
5034 		*accmode |= VADMIN;
5035 	}
5036 
5037 	/*
5038 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5039 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5040 	 */
5041 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5042 
5043 	return (0);
5044 }
5045 
5046 /*
5047  * These are helper functions for filesystems to traverse all
5048  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
5049  *
5050  * This interface replaces MNT_VNODE_FOREACH.
5051  */
5052 
5053 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
5054 
5055 struct vnode *
5056 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
5057 {
5058 	struct vnode *vp;
5059 
5060 	if (should_yield())
5061 		kern_yield(PRI_USER);
5062 	MNT_ILOCK(mp);
5063 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5064 	vp = TAILQ_NEXT(*mvp, v_nmntvnodes);
5065 	while (vp != NULL && (vp->v_type == VMARKER ||
5066 	    (vp->v_iflag & VI_DOOMED) != 0))
5067 		vp = TAILQ_NEXT(vp, v_nmntvnodes);
5068 
5069 	/* Check if we are done */
5070 	if (vp == NULL) {
5071 		__mnt_vnode_markerfree_all(mvp, mp);
5072 		/* MNT_IUNLOCK(mp); -- done in above function */
5073 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
5074 		return (NULL);
5075 	}
5076 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5077 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5078 	VI_LOCK(vp);
5079 	MNT_IUNLOCK(mp);
5080 	return (vp);
5081 }
5082 
5083 struct vnode *
5084 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
5085 {
5086 	struct vnode *vp;
5087 
5088 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5089 	MNT_ILOCK(mp);
5090 	MNT_REF(mp);
5091 	(*mvp)->v_type = VMARKER;
5092 
5093 	vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
5094 	while (vp != NULL && (vp->v_type == VMARKER ||
5095 	    (vp->v_iflag & VI_DOOMED) != 0))
5096 		vp = TAILQ_NEXT(vp, v_nmntvnodes);
5097 
5098 	/* Check if we are done */
5099 	if (vp == NULL) {
5100 		MNT_REL(mp);
5101 		MNT_IUNLOCK(mp);
5102 		free(*mvp, M_VNODE_MARKER);
5103 		*mvp = NULL;
5104 		return (NULL);
5105 	}
5106 	(*mvp)->v_mount = mp;
5107 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5108 	VI_LOCK(vp);
5109 	MNT_IUNLOCK(mp);
5110 	return (vp);
5111 }
5112 
5113 
5114 void
5115 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
5116 {
5117 
5118 	if (*mvp == NULL) {
5119 		MNT_IUNLOCK(mp);
5120 		return;
5121 	}
5122 
5123 	mtx_assert(MNT_MTX(mp), MA_OWNED);
5124 
5125 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5126 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5127 	MNT_REL(mp);
5128 	MNT_IUNLOCK(mp);
5129 	free(*mvp, M_VNODE_MARKER);
5130 	*mvp = NULL;
5131 }
5132 
5133 /*
5134  * These are helper functions for filesystems to traverse their
5135  * active vnodes.  See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h
5136  */
5137 static void
5138 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5139 {
5140 
5141 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5142 
5143 	MNT_ILOCK(mp);
5144 	MNT_REL(mp);
5145 	MNT_IUNLOCK(mp);
5146 	free(*mvp, M_VNODE_MARKER);
5147 	*mvp = NULL;
5148 }
5149 
5150 static struct vnode *
5151 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5152 {
5153 	struct vnode *vp, *nvp;
5154 
5155 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
5156 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5157 restart:
5158 	vp = TAILQ_NEXT(*mvp, v_actfreelist);
5159 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5160 	while (vp != NULL) {
5161 		if (vp->v_type == VMARKER) {
5162 			vp = TAILQ_NEXT(vp, v_actfreelist);
5163 			continue;
5164 		}
5165 		if (!VI_TRYLOCK(vp)) {
5166 			if (mp_ncpus == 1 || should_yield()) {
5167 				TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5168 				mtx_unlock(&vnode_free_list_mtx);
5169 				pause("vnacti", 1);
5170 				mtx_lock(&vnode_free_list_mtx);
5171 				goto restart;
5172 			}
5173 			continue;
5174 		}
5175 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
5176 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
5177 		    ("alien vnode on the active list %p %p", vp, mp));
5178 		if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0)
5179 			break;
5180 		nvp = TAILQ_NEXT(vp, v_actfreelist);
5181 		VI_UNLOCK(vp);
5182 		vp = nvp;
5183 	}
5184 
5185 	/* Check if we are done */
5186 	if (vp == NULL) {
5187 		mtx_unlock(&vnode_free_list_mtx);
5188 		mnt_vnode_markerfree_active(mvp, mp);
5189 		return (NULL);
5190 	}
5191 	TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist);
5192 	mtx_unlock(&vnode_free_list_mtx);
5193 	ASSERT_VI_LOCKED(vp, "active iter");
5194 	KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp));
5195 	return (vp);
5196 }
5197 
5198 struct vnode *
5199 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5200 {
5201 
5202 	if (should_yield())
5203 		kern_yield(PRI_USER);
5204 	mtx_lock(&vnode_free_list_mtx);
5205 	return (mnt_vnode_next_active(mvp, mp));
5206 }
5207 
5208 struct vnode *
5209 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp)
5210 {
5211 	struct vnode *vp;
5212 
5213 	*mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5214 	MNT_ILOCK(mp);
5215 	MNT_REF(mp);
5216 	MNT_IUNLOCK(mp);
5217 	(*mvp)->v_type = VMARKER;
5218 	(*mvp)->v_mount = mp;
5219 
5220 	mtx_lock(&vnode_free_list_mtx);
5221 	vp = TAILQ_FIRST(&mp->mnt_activevnodelist);
5222 	if (vp == NULL) {
5223 		mtx_unlock(&vnode_free_list_mtx);
5224 		mnt_vnode_markerfree_active(mvp, mp);
5225 		return (NULL);
5226 	}
5227 	TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5228 	return (mnt_vnode_next_active(mvp, mp));
5229 }
5230 
5231 void
5232 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5233 {
5234 
5235 	if (*mvp == NULL)
5236 		return;
5237 
5238 	mtx_lock(&vnode_free_list_mtx);
5239 	TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5240 	mtx_unlock(&vnode_free_list_mtx);
5241 	mnt_vnode_markerfree_active(mvp, mp);
5242 }
5243