xref: /freebsd/sys/kern/vfs_subr.c (revision 6e0da4f753ed6b5d26395001a6194b4fdea70177)
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_ddb.h"
45 #include "opt_mac.h"
46 
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/bio.h>
50 #include <sys/buf.h>
51 #include <sys/conf.h>
52 #include <sys/event.h>
53 #include <sys/eventhandler.h>
54 #include <sys/extattr.h>
55 #include <sys/fcntl.h>
56 #include <sys/kdb.h>
57 #include <sys/kernel.h>
58 #include <sys/kthread.h>
59 #include <sys/mac.h>
60 #include <sys/malloc.h>
61 #include <sys/mount.h>
62 #include <sys/namei.h>
63 #include <sys/reboot.h>
64 #include <sys/sleepqueue.h>
65 #include <sys/stat.h>
66 #include <sys/sysctl.h>
67 #include <sys/syslog.h>
68 #include <sys/vmmeter.h>
69 #include <sys/vnode.h>
70 
71 #include <vm/vm.h>
72 #include <vm/vm_object.h>
73 #include <vm/vm_extern.h>
74 #include <vm/pmap.h>
75 #include <vm/vm_map.h>
76 #include <vm/vm_page.h>
77 #include <vm/vm_kern.h>
78 #include <vm/uma.h>
79 
80 static MALLOC_DEFINE(M_NETADDR, "Export Host", "Export host address structure");
81 
82 static void	delmntque(struct vnode *vp);
83 static void	insmntque(struct vnode *vp, struct mount *mp);
84 static void	vclean(struct vnode *vp, int flags, struct thread *td);
85 static void	vlruvp(struct vnode *vp);
86 static int	flushbuflist(struct bufv *bufv, int flags, struct vnode *vp,
87 		    int slpflag, int slptimeo);
88 static void	syncer_shutdown(void *arg, int howto);
89 static int	vtryrecycle(struct vnode *vp);
90 static void	vx_lock(struct vnode *vp);
91 static void	vx_unlock(struct vnode *vp);
92 
93 
94 /*
95  * Number of vnodes in existence.  Increased whenever getnewvnode()
96  * allocates a new vnode, never decreased.
97  */
98 static unsigned long	numvnodes;
99 
100 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, "");
101 
102 /*
103  * Conversion tables for conversion from vnode types to inode formats
104  * and back.
105  */
106 enum vtype iftovt_tab[16] = {
107 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
108 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
109 };
110 int vttoif_tab[9] = {
111 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
112 	S_IFSOCK, S_IFIFO, S_IFMT,
113 };
114 
115 /*
116  * List of vnodes that are ready for recycling.
117  */
118 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
119 
120 /*
121  * Minimum number of free vnodes.  If there are fewer than this free vnodes,
122  * getnewvnode() will return a newly allocated vnode.
123  */
124 static u_long wantfreevnodes = 25;
125 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, "");
126 /* Number of vnodes in the free list. */
127 static u_long freevnodes;
128 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, "");
129 
130 /*
131  * Various variables used for debugging the new implementation of
132  * reassignbuf().
133  * XXX these are probably of (very) limited utility now.
134  */
135 static int reassignbufcalls;
136 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, "");
137 static int nameileafonly;
138 SYSCTL_INT(_vfs, OID_AUTO, nameileafonly, CTLFLAG_RW, &nameileafonly, 0, "");
139 
140 /*
141  * Cache for the mount type id assigned to NFS.  This is used for
142  * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c.
143  */
144 int	nfs_mount_type = -1;
145 
146 /* To keep more than one thread at a time from running vfs_getnewfsid */
147 static struct mtx mntid_mtx;
148 
149 /*
150  * Lock for any access to the following:
151  *	vnode_free_list
152  *	numvnodes
153  *	freevnodes
154  */
155 static struct mtx vnode_free_list_mtx;
156 
157 /* Publicly exported FS */
158 struct nfs_public nfs_pub;
159 
160 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
161 static uma_zone_t vnode_zone;
162 static uma_zone_t vnodepoll_zone;
163 
164 /* Set to 1 to print out reclaim of active vnodes */
165 int	prtactive;
166 
167 /*
168  * The workitem queue.
169  *
170  * It is useful to delay writes of file data and filesystem metadata
171  * for tens of seconds so that quickly created and deleted files need
172  * not waste disk bandwidth being created and removed. To realize this,
173  * we append vnodes to a "workitem" queue. When running with a soft
174  * updates implementation, most pending metadata dependencies should
175  * not wait for more than a few seconds. Thus, mounted on block devices
176  * are delayed only about a half the time that file data is delayed.
177  * Similarly, directory updates are more critical, so are only delayed
178  * about a third the time that file data is delayed. Thus, there are
179  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
180  * one each second (driven off the filesystem syncer process). The
181  * syncer_delayno variable indicates the next queue that is to be processed.
182  * Items that need to be processed soon are placed in this queue:
183  *
184  *	syncer_workitem_pending[syncer_delayno]
185  *
186  * A delay of fifteen seconds is done by placing the request fifteen
187  * entries later in the queue:
188  *
189  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
190  *
191  */
192 static int syncer_delayno;
193 static long syncer_mask;
194 LIST_HEAD(synclist, bufobj);
195 static struct synclist *syncer_workitem_pending;
196 /*
197  * The sync_mtx protects:
198  *	bo->bo_synclist
199  *	sync_vnode_count
200  *	syncer_delayno
201  *	syncer_state
202  *	syncer_workitem_pending
203  *	syncer_worklist_len
204  *	rushjob
205  */
206 static struct mtx sync_mtx;
207 
208 #define SYNCER_MAXDELAY		32
209 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
210 static int syncdelay = 30;		/* max time to delay syncing data */
211 static int filedelay = 30;		/* time to delay syncing files */
212 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, "");
213 static int dirdelay = 29;		/* time to delay syncing directories */
214 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, "");
215 static int metadelay = 28;		/* time to delay syncing metadata */
216 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, "");
217 static int rushjob;		/* number of slots to run ASAP */
218 static int stat_rush_requests;	/* number of times I/O speeded up */
219 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, "");
220 
221 /*
222  * When shutting down the syncer, run it at four times normal speed.
223  */
224 #define SYNCER_SHUTDOWN_SPEEDUP		4
225 static int sync_vnode_count;
226 static int syncer_worklist_len;
227 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
228     syncer_state;
229 
230 /*
231  * Number of vnodes we want to exist at any one time.  This is mostly used
232  * to size hash tables in vnode-related code.  It is normally not used in
233  * getnewvnode(), as wantfreevnodes is normally nonzero.)
234  *
235  * XXX desiredvnodes is historical cruft and should not exist.
236  */
237 int desiredvnodes;
238 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW,
239     &desiredvnodes, 0, "Maximum number of vnodes");
240 static int minvnodes;
241 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
242     &minvnodes, 0, "Minimum number of vnodes");
243 static int vnlru_nowhere;
244 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
245     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
246 
247 /* Hook for calling soft updates. */
248 int (*softdep_process_worklist_hook)(struct mount *);
249 
250 /*
251  * Initialize the vnode management data structures.
252  */
253 #ifndef	MAXVNODES_MAX
254 #define	MAXVNODES_MAX	100000
255 #endif
256 static void
257 vntblinit(void *dummy __unused)
258 {
259 
260 	/*
261 	 * Desiredvnodes is a function of the physical memory size and
262 	 * the kernel's heap size.  Specifically, desiredvnodes scales
263 	 * in proportion to the physical memory size until two fifths
264 	 * of the kernel's heap size is consumed by vnodes and vm
265 	 * objects.
266 	 */
267 	desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size /
268 	    (5 * (sizeof(struct vm_object) + sizeof(struct vnode))));
269 	if (desiredvnodes > MAXVNODES_MAX) {
270 		if (bootverbose)
271 			printf("Reducing kern.maxvnodes %d -> %d\n",
272 			    desiredvnodes, MAXVNODES_MAX);
273 		desiredvnodes = MAXVNODES_MAX;
274 	}
275 	minvnodes = desiredvnodes / 4;
276 	mtx_init(&mountlist_mtx, "mountlist", NULL, MTX_DEF);
277 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
278 	TAILQ_INIT(&vnode_free_list);
279 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
280 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
281 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
282 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
283 	      NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
284 	/*
285 	 * Initialize the filesystem syncer.
286 	 */
287 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
288 		&syncer_mask);
289 	syncer_maxdelay = syncer_mask + 1;
290 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
291 }
292 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL)
293 
294 
295 /*
296  * Mark a mount point as busy. Used to synchronize access and to delay
297  * unmounting. Interlock is not released on failure.
298  */
299 int
300 vfs_busy(mp, flags, interlkp, td)
301 	struct mount *mp;
302 	int flags;
303 	struct mtx *interlkp;
304 	struct thread *td;
305 {
306 	int lkflags;
307 
308 	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
309 		if (flags & LK_NOWAIT)
310 			return (ENOENT);
311 		mp->mnt_kern_flag |= MNTK_MWAIT;
312 		/*
313 		 * Since all busy locks are shared except the exclusive
314 		 * lock granted when unmounting, the only place that a
315 		 * wakeup needs to be done is at the release of the
316 		 * exclusive lock at the end of dounmount.
317 		 */
318 		msleep(mp, interlkp, PVFS, "vfs_busy", 0);
319 		return (ENOENT);
320 	}
321 	lkflags = LK_SHARED | LK_NOPAUSE;
322 	if (interlkp)
323 		lkflags |= LK_INTERLOCK;
324 	if (lockmgr(&mp->mnt_lock, lkflags, interlkp, td))
325 		panic("vfs_busy: unexpected lock failure");
326 	return (0);
327 }
328 
329 /*
330  * Free a busy filesystem.
331  */
332 void
333 vfs_unbusy(mp, td)
334 	struct mount *mp;
335 	struct thread *td;
336 {
337 
338 	lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td);
339 }
340 
341 /*
342  * Lookup a mount point by filesystem identifier.
343  */
344 struct mount *
345 vfs_getvfs(fsid)
346 	fsid_t *fsid;
347 {
348 	register struct mount *mp;
349 
350 	mtx_lock(&mountlist_mtx);
351 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
352 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
353 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
354 			mtx_unlock(&mountlist_mtx);
355 			return (mp);
356 		}
357 	}
358 	mtx_unlock(&mountlist_mtx);
359 	return ((struct mount *) 0);
360 }
361 
362 /*
363  * Check if a user can access priveledged mount options.
364  */
365 int
366 vfs_suser(struct mount *mp, struct thread *td)
367 {
368 	int error;
369 
370 	if ((mp->mnt_flag & MNT_USER) == 0 ||
371 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
372 		if ((error = suser(td)) != 0)
373 			return (error);
374 	}
375 	return (0);
376 }
377 
378 /*
379  * Get a new unique fsid.  Try to make its val[0] unique, since this value
380  * will be used to create fake device numbers for stat().  Also try (but
381  * not so hard) make its val[0] unique mod 2^16, since some emulators only
382  * support 16-bit device numbers.  We end up with unique val[0]'s for the
383  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
384  *
385  * Keep in mind that several mounts may be running in parallel.  Starting
386  * the search one past where the previous search terminated is both a
387  * micro-optimization and a defense against returning the same fsid to
388  * different mounts.
389  */
390 void
391 vfs_getnewfsid(mp)
392 	struct mount *mp;
393 {
394 	static u_int16_t mntid_base;
395 	fsid_t tfsid;
396 	int mtype;
397 
398 	mtx_lock(&mntid_mtx);
399 	mtype = mp->mnt_vfc->vfc_typenum;
400 	tfsid.val[1] = mtype;
401 	mtype = (mtype & 0xFF) << 24;
402 	for (;;) {
403 		tfsid.val[0] = makedev(255,
404 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
405 		mntid_base++;
406 		if (vfs_getvfs(&tfsid) == NULL)
407 			break;
408 	}
409 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
410 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
411 	mtx_unlock(&mntid_mtx);
412 }
413 
414 /*
415  * Knob to control the precision of file timestamps:
416  *
417  *   0 = seconds only; nanoseconds zeroed.
418  *   1 = seconds and nanoseconds, accurate within 1/HZ.
419  *   2 = seconds and nanoseconds, truncated to microseconds.
420  * >=3 = seconds and nanoseconds, maximum precision.
421  */
422 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
423 
424 static int timestamp_precision = TSP_SEC;
425 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
426     &timestamp_precision, 0, "");
427 
428 /*
429  * Get a current timestamp.
430  */
431 void
432 vfs_timestamp(tsp)
433 	struct timespec *tsp;
434 {
435 	struct timeval tv;
436 
437 	switch (timestamp_precision) {
438 	case TSP_SEC:
439 		tsp->tv_sec = time_second;
440 		tsp->tv_nsec = 0;
441 		break;
442 	case TSP_HZ:
443 		getnanotime(tsp);
444 		break;
445 	case TSP_USEC:
446 		microtime(&tv);
447 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
448 		break;
449 	case TSP_NSEC:
450 	default:
451 		nanotime(tsp);
452 		break;
453 	}
454 }
455 
456 /*
457  * Set vnode attributes to VNOVAL
458  */
459 void
460 vattr_null(vap)
461 	register struct vattr *vap;
462 {
463 
464 	vap->va_type = VNON;
465 	vap->va_size = VNOVAL;
466 	vap->va_bytes = VNOVAL;
467 	vap->va_mode = VNOVAL;
468 	vap->va_nlink = VNOVAL;
469 	vap->va_uid = VNOVAL;
470 	vap->va_gid = VNOVAL;
471 	vap->va_fsid = VNOVAL;
472 	vap->va_fileid = VNOVAL;
473 	vap->va_blocksize = VNOVAL;
474 	vap->va_rdev = VNOVAL;
475 	vap->va_atime.tv_sec = VNOVAL;
476 	vap->va_atime.tv_nsec = VNOVAL;
477 	vap->va_mtime.tv_sec = VNOVAL;
478 	vap->va_mtime.tv_nsec = VNOVAL;
479 	vap->va_ctime.tv_sec = VNOVAL;
480 	vap->va_ctime.tv_nsec = VNOVAL;
481 	vap->va_birthtime.tv_sec = VNOVAL;
482 	vap->va_birthtime.tv_nsec = VNOVAL;
483 	vap->va_flags = VNOVAL;
484 	vap->va_gen = VNOVAL;
485 	vap->va_vaflags = 0;
486 }
487 
488 /*
489  * This routine is called when we have too many vnodes.  It attempts
490  * to free <count> vnodes and will potentially free vnodes that still
491  * have VM backing store (VM backing store is typically the cause
492  * of a vnode blowout so we want to do this).  Therefore, this operation
493  * is not considered cheap.
494  *
495  * A number of conditions may prevent a vnode from being reclaimed.
496  * the buffer cache may have references on the vnode, a directory
497  * vnode may still have references due to the namei cache representing
498  * underlying files, or the vnode may be in active use.   It is not
499  * desireable to reuse such vnodes.  These conditions may cause the
500  * number of vnodes to reach some minimum value regardless of what
501  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
502  */
503 static int
504 vlrureclaim(struct mount *mp)
505 {
506 	struct vnode *vp;
507 	int done;
508 	int trigger;
509 	int usevnodes;
510 	int count;
511 
512 	/*
513 	 * Calculate the trigger point, don't allow user
514 	 * screwups to blow us up.   This prevents us from
515 	 * recycling vnodes with lots of resident pages.  We
516 	 * aren't trying to free memory, we are trying to
517 	 * free vnodes.
518 	 */
519 	usevnodes = desiredvnodes;
520 	if (usevnodes <= 0)
521 		usevnodes = 1;
522 	trigger = cnt.v_page_count * 2 / usevnodes;
523 
524 	done = 0;
525 	MNT_ILOCK(mp);
526 	count = mp->mnt_nvnodelistsize / 10 + 1;
527 	while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) {
528 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
529 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
530 
531 		if (vp->v_type != VNON &&
532 		    vp->v_type != VBAD &&
533 		    VI_TRYLOCK(vp)) {
534 			if (VMIGHTFREE(vp) &&           /* critical path opt */
535 			    (vp->v_object == NULL ||
536 			    vp->v_object->resident_page_count < trigger)) {
537 				MNT_IUNLOCK(mp);
538 				vgonel(vp, curthread);
539 				done++;
540 				MNT_ILOCK(mp);
541 			} else
542 				VI_UNLOCK(vp);
543 		}
544 		--count;
545 	}
546 	MNT_IUNLOCK(mp);
547 	return done;
548 }
549 
550 /*
551  * Attempt to recycle vnodes in a context that is always safe to block.
552  * Calling vlrurecycle() from the bowels of filesystem code has some
553  * interesting deadlock problems.
554  */
555 static struct proc *vnlruproc;
556 static int vnlruproc_sig;
557 
558 static void
559 vnlru_proc(void)
560 {
561 	struct mount *mp, *nmp;
562 	int done;
563 	struct proc *p = vnlruproc;
564 	struct thread *td = FIRST_THREAD_IN_PROC(p);
565 
566 	mtx_lock(&Giant);
567 
568 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p,
569 	    SHUTDOWN_PRI_FIRST);
570 
571 	for (;;) {
572 		kthread_suspend_check(p);
573 		mtx_lock(&vnode_free_list_mtx);
574 		if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) {
575 			mtx_unlock(&vnode_free_list_mtx);
576 			vnlruproc_sig = 0;
577 			wakeup(&vnlruproc_sig);
578 			tsleep(vnlruproc, PVFS, "vlruwt", hz);
579 			continue;
580 		}
581 		mtx_unlock(&vnode_free_list_mtx);
582 		done = 0;
583 		mtx_lock(&mountlist_mtx);
584 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
585 			if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) {
586 				nmp = TAILQ_NEXT(mp, mnt_list);
587 				continue;
588 			}
589 			done += vlrureclaim(mp);
590 			mtx_lock(&mountlist_mtx);
591 			nmp = TAILQ_NEXT(mp, mnt_list);
592 			vfs_unbusy(mp, td);
593 		}
594 		mtx_unlock(&mountlist_mtx);
595 		if (done == 0) {
596 #if 0
597 			/* These messages are temporary debugging aids */
598 			if (vnlru_nowhere < 5)
599 				printf("vnlru process getting nowhere..\n");
600 			else if (vnlru_nowhere == 5)
601 				printf("vnlru process messages stopped.\n");
602 #endif
603 			vnlru_nowhere++;
604 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
605 		}
606 	}
607 }
608 
609 static struct kproc_desc vnlru_kp = {
610 	"vnlru",
611 	vnlru_proc,
612 	&vnlruproc
613 };
614 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
615 
616 
617 /*
618  * Routines having to do with the management of the vnode table.
619  */
620 
621 /*
622  * Check to see if a free vnode can be recycled. If it can,
623  * recycle it and return it with the vnode interlock held.
624  */
625 static int
626 vtryrecycle(struct vnode *vp)
627 {
628 	struct thread *td = curthread;
629 	vm_object_t object;
630 	struct mount *vnmp;
631 	int error;
632 
633 	/* Don't recycle if we can't get the interlock */
634 	if (!VI_TRYLOCK(vp))
635 		return (EWOULDBLOCK);
636 	/*
637 	 * This vnode may found and locked via some other list, if so we
638 	 * can't recycle it yet.
639 	 */
640 	if (vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
641 		return (EWOULDBLOCK);
642 	/*
643 	 * Don't recycle if its filesystem is being suspended.
644 	 */
645 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
646 		VOP_UNLOCK(vp, 0, td);
647 		return (EBUSY);
648 	}
649 
650 	/*
651 	 * Don't recycle if we still have cached pages.
652 	 */
653 	if (VOP_GETVOBJECT(vp, &object) == 0) {
654 		VM_OBJECT_LOCK(object);
655 		if (object->resident_page_count ||
656 		    object->ref_count) {
657 			VM_OBJECT_UNLOCK(object);
658 			error = EBUSY;
659 			goto done;
660 		}
661 		VM_OBJECT_UNLOCK(object);
662 	}
663 	if (LIST_FIRST(&vp->v_cache_src)) {
664 		/*
665 		 * note: nameileafonly sysctl is temporary,
666 		 * for debugging only, and will eventually be
667 		 * removed.
668 		 */
669 		if (nameileafonly > 0) {
670 			/*
671 			 * Do not reuse namei-cached directory
672 			 * vnodes that have cached
673 			 * subdirectories.
674 			 */
675 			if (cache_leaf_test(vp) < 0) {
676 				error = EISDIR;
677 				goto done;
678 			}
679 		} else if (nameileafonly < 0 ||
680 			    vmiodirenable == 0) {
681 			/*
682 			 * Do not reuse namei-cached directory
683 			 * vnodes if nameileafonly is -1 or
684 			 * if VMIO backing for directories is
685 			 * turned off (otherwise we reuse them
686 			 * too quickly).
687 			 */
688 			error = EBUSY;
689 			goto done;
690 		}
691 	}
692 	/*
693 	 * If we got this far, we need to acquire the interlock and see if
694 	 * anyone picked up this vnode from another list.  If not, we will
695 	 * mark it with XLOCK via vgonel() so that anyone who does find it
696 	 * will skip over it.
697 	 */
698 	VI_LOCK(vp);
699 	if (VSHOULDBUSY(vp) && (vp->v_iflag & VI_XLOCK) == 0) {
700 		VI_UNLOCK(vp);
701 		error = EBUSY;
702 		goto done;
703 	}
704 	mtx_lock(&vnode_free_list_mtx);
705 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
706 	vp->v_iflag &= ~VI_FREE;
707 	mtx_unlock(&vnode_free_list_mtx);
708 	vp->v_iflag |= VI_DOOMED;
709 	if ((vp->v_type != VBAD) || (vp->v_data != NULL)) {
710 		VOP_UNLOCK(vp, 0, td);
711 		vgonel(vp, td);
712 		VI_LOCK(vp);
713 	} else
714 		VOP_UNLOCK(vp, 0, td);
715 	vn_finished_write(vnmp);
716 	return (0);
717 done:
718 	VOP_UNLOCK(vp, 0, td);
719 	vn_finished_write(vnmp);
720 	return (error);
721 }
722 
723 /*
724  * Return the next vnode from the free list.
725  */
726 int
727 getnewvnode(tag, mp, vops, vpp)
728 	const char *tag;
729 	struct mount *mp;
730 	struct vop_vector *vops;
731 	struct vnode **vpp;
732 {
733 	struct vnode *vp = NULL;
734 	struct vpollinfo *pollinfo = NULL;
735 	struct bufobj *bo;
736 
737 	mtx_lock(&vnode_free_list_mtx);
738 
739 	/*
740 	 * Try to reuse vnodes if we hit the max.  This situation only
741 	 * occurs in certain large-memory (2G+) situations.  We cannot
742 	 * attempt to directly reclaim vnodes due to nasty recursion
743 	 * problems.
744 	 */
745 	while (numvnodes - freevnodes > desiredvnodes) {
746 		if (vnlruproc_sig == 0) {
747 			vnlruproc_sig = 1;      /* avoid unnecessary wakeups */
748 			wakeup(vnlruproc);
749 		}
750 		mtx_unlock(&vnode_free_list_mtx);
751 		tsleep(&vnlruproc_sig, PVFS, "vlruwk", hz);
752 		mtx_lock(&vnode_free_list_mtx);
753 	}
754 
755 	/*
756 	 * Attempt to reuse a vnode already on the free list, allocating
757 	 * a new vnode if we can't find one or if we have not reached a
758 	 * good minimum for good LRU performance.
759 	 */
760 
761 	if (freevnodes >= wantfreevnodes && numvnodes >= minvnodes) {
762 		int error;
763 		int count;
764 
765 		for (count = 0; count < freevnodes; count++) {
766 			vp = TAILQ_FIRST(&vnode_free_list);
767 
768 			KASSERT(vp->v_usecount == 0 &&
769 			    (vp->v_iflag & VI_DOINGINACT) == 0,
770 			    ("getnewvnode: free vnode isn't"));
771 
772 			TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
773 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
774 			mtx_unlock(&vnode_free_list_mtx);
775 			error = vtryrecycle(vp);
776 			mtx_lock(&vnode_free_list_mtx);
777 			if (error == 0)
778 				break;
779 			vp = NULL;
780 		}
781 	}
782 	if (vp) {
783 		freevnodes--;
784 		bo = &vp->v_bufobj;
785 		mtx_unlock(&vnode_free_list_mtx);
786 
787 #ifdef INVARIANTS
788 		{
789 			if (vp->v_data)
790 				printf("cleaned vnode isn't, "
791 				       "address %p, inode %p\n",
792 				       vp, vp->v_data);
793 			if (bo->bo_numoutput)
794 				panic("Clean vnode has pending I/O's");
795 			if (vp->v_writecount != 0)
796 				panic("Non-zero write count");
797 		}
798 #endif
799 		if ((pollinfo = vp->v_pollinfo) != NULL) {
800 			/*
801 			 * To avoid lock order reversals, the call to
802 			 * uma_zfree() must be delayed until the vnode
803 			 * interlock is released.
804 			 */
805 			vp->v_pollinfo = NULL;
806 		}
807 #ifdef MAC
808 		mac_destroy_vnode(vp);
809 #endif
810 		vp->v_iflag = 0;
811 		vp->v_vflag = 0;
812 		vp->v_lastw = 0;
813 		vp->v_lasta = 0;
814 		vp->v_cstart = 0;
815 		vp->v_clen = 0;
816 		vp->v_socket = 0;
817 		lockdestroy(vp->v_vnlock);
818 		lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE);
819 		KASSERT(bo->bo_clean.bv_cnt == 0, ("cleanbufcnt not 0"));
820 		KASSERT(bo->bo_clean.bv_root == NULL, ("cleanblkroot not NULL"));
821 		KASSERT(bo->bo_dirty.bv_cnt == 0, ("dirtybufcnt not 0"));
822 		KASSERT(bo->bo_dirty.bv_root == NULL, ("dirtyblkroot not NULL"));
823 	} else {
824 		numvnodes++;
825 		mtx_unlock(&vnode_free_list_mtx);
826 
827 		vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO);
828 		mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
829 		VI_LOCK(vp);
830 		vp->v_dd = vp;
831 		bo = &vp->v_bufobj;
832 		bo->__bo_vnode = vp;
833 		bo->bo_mtx = &vp->v_interlock;
834 		vp->v_vnlock = &vp->v_lock;
835 		lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE);
836 		cache_purge(vp);		/* Sets up v_id. */
837 		LIST_INIT(&vp->v_cache_src);
838 		TAILQ_INIT(&vp->v_cache_dst);
839 	}
840 
841 	TAILQ_INIT(&bo->bo_clean.bv_hd);
842 	TAILQ_INIT(&bo->bo_dirty.bv_hd);
843 	bo->bo_ops = &buf_ops_bio;
844 	bo->bo_private = vp;
845 	vp->v_type = VNON;
846 	vp->v_tag = tag;
847 	vp->v_op = vops;
848 	*vpp = vp;
849 	vp->v_usecount = 1;
850 	vp->v_data = 0;
851 	VI_UNLOCK(vp);
852 	if (pollinfo != NULL) {
853 		knlist_destroy(&pollinfo->vpi_selinfo.si_note);
854 		mtx_destroy(&pollinfo->vpi_lock);
855 		uma_zfree(vnodepoll_zone, pollinfo);
856 	}
857 #ifdef MAC
858 	mac_init_vnode(vp);
859 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
860 		mac_associate_vnode_singlelabel(mp, vp);
861 	else if (mp == NULL)
862 		printf("NULL mp in getnewvnode()\n");
863 #endif
864 	delmntque(vp);
865 	if (mp != NULL) {
866 		insmntque(vp, mp);
867 		bo->bo_bsize = mp->mnt_stat.f_iosize;
868 	}
869 
870 	return (0);
871 }
872 
873 /*
874  * Delete from old mount point vnode list, if on one.
875  */
876 static void
877 delmntque(struct vnode *vp)
878 {
879 	struct mount *mp;
880 
881 	if (vp->v_mount == NULL)
882 		return;
883 	mp = vp->v_mount;
884 	MNT_ILOCK(mp);
885 	vp->v_mount = NULL;
886 	KASSERT(mp->mnt_nvnodelistsize > 0,
887 		("bad mount point vnode list size"));
888 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
889 	mp->mnt_nvnodelistsize--;
890 	MNT_IUNLOCK(mp);
891 }
892 
893 /*
894  * Insert into list of vnodes for the new mount point, if available.
895  */
896 static void
897 insmntque(struct vnode *vp, struct mount *mp)
898 {
899 
900 	vp->v_mount = mp;
901 	KASSERT(mp != NULL, ("Don't call insmntque(foo, NULL)"));
902 	MNT_ILOCK(vp->v_mount);
903 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
904 	mp->mnt_nvnodelistsize++;
905 	MNT_IUNLOCK(vp->v_mount);
906 }
907 
908 /*
909  * Flush out and invalidate all buffers associated with a vnode.
910  * Called with the underlying object locked.
911  */
912 int
913 vinvalbuf(vp, flags, td, slpflag, slptimeo)
914 	struct vnode *vp;
915 	int flags;
916 	struct thread *td;
917 	int slpflag, slptimeo;
918 {
919 	int error;
920 	vm_object_t object;
921 	struct bufobj *bo;
922 
923 	GIANT_REQUIRED;
924 
925 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
926 
927 	bo = &vp->v_bufobj;
928 	BO_LOCK(bo);
929 	if (flags & V_SAVE) {
930 		error = bufobj_wwait(bo, slpflag, slptimeo);
931 		if (error) {
932 			BO_UNLOCK(bo);
933 			return (error);
934 		}
935 		if (bo->bo_dirty.bv_cnt > 0) {
936 			BO_UNLOCK(bo);
937 			if ((error = BO_SYNC(bo, MNT_WAIT, td)) != 0)
938 				return (error);
939 			/*
940 			 * XXX We could save a lock/unlock if this was only
941 			 * enabled under INVARIANTS
942 			 */
943 			BO_LOCK(bo);
944 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
945 				panic("vinvalbuf: dirty bufs");
946 		}
947 	}
948 	/*
949 	 * If you alter this loop please notice that interlock is dropped and
950 	 * reacquired in flushbuflist.  Special care is needed to ensure that
951 	 * no race conditions occur from this.
952 	 */
953 	do {
954 		error = flushbuflist(&bo->bo_clean,
955 		    flags, vp, slpflag, slptimeo);
956 		if (error == 0)
957 			error = flushbuflist(&bo->bo_dirty,
958 			    flags, vp, slpflag, slptimeo);
959 		if (error != 0 && error != EAGAIN) {
960 			BO_UNLOCK(bo);
961 			return (error);
962 		}
963 	} while (error != 0);
964 
965 	/*
966 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
967 	 * have write I/O in-progress but if there is a VM object then the
968 	 * VM object can also have read-I/O in-progress.
969 	 */
970 	do {
971 		bufobj_wwait(bo, 0, 0);
972 		VI_UNLOCK(vp);
973 		if (VOP_GETVOBJECT(vp, &object) == 0) {
974 			VM_OBJECT_LOCK(object);
975 			vm_object_pip_wait(object, "vnvlbx");
976 			VM_OBJECT_UNLOCK(object);
977 		}
978 		VI_LOCK(vp);
979 	} while (bo->bo_numoutput > 0);
980 	VI_UNLOCK(vp);
981 
982 	/*
983 	 * Destroy the copy in the VM cache, too.
984 	 */
985 	if (VOP_GETVOBJECT(vp, &object) == 0) {
986 		VM_OBJECT_LOCK(object);
987 		vm_object_page_remove(object, 0, 0,
988 			(flags & V_SAVE) ? TRUE : FALSE);
989 		VM_OBJECT_UNLOCK(object);
990 	}
991 
992 #ifdef INVARIANTS
993 	BO_LOCK(bo);
994 	if ((flags & (V_ALT | V_NORMAL)) == 0 &&
995 	    (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0))
996 		panic("vinvalbuf: flush failed");
997 	BO_UNLOCK(bo);
998 #endif
999 	return (0);
1000 }
1001 
1002 /*
1003  * Flush out buffers on the specified list.
1004  *
1005  */
1006 static int
1007 flushbuflist(bufv, flags, vp, slpflag, slptimeo)
1008 	struct bufv *bufv;
1009 	int flags;
1010 	struct vnode *vp;
1011 	int slpflag, slptimeo;
1012 {
1013 	struct buf *bp, *nbp;
1014 	int retval, error;
1015 	struct bufobj *bo;
1016 
1017 	bo = &vp->v_bufobj;
1018 	ASSERT_BO_LOCKED(bo);
1019 
1020 	retval = 0;
1021 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1022 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1023 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1024 			continue;
1025 		}
1026 		retval = EAGAIN;
1027 		error = BUF_TIMELOCK(bp,
1028 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo),
1029 		    "flushbuf", slpflag, slptimeo);
1030 		if (error) {
1031 			BO_LOCK(bo);
1032 			return (error != ENOLCK ? error : EAGAIN);
1033 		}
1034 		/*
1035 		 * XXX Since there are no node locks for NFS, I
1036 		 * believe there is a slight chance that a delayed
1037 		 * write will occur while sleeping just above, so
1038 		 * check for it.  Note that vfs_bio_awrite expects
1039 		 * buffers to reside on a queue, while bwrite and
1040 		 * brelse do not.
1041 		 */
1042 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1043 			(flags & V_SAVE)) {
1044 
1045 			if (bp->b_vp == vp) {
1046 				if (bp->b_flags & B_CLUSTEROK) {
1047 					vfs_bio_awrite(bp);
1048 				} else {
1049 					bremfree(bp);
1050 					bp->b_flags |= B_ASYNC;
1051 					bwrite(bp);
1052 				}
1053 			} else {
1054 				bremfree(bp);
1055 				(void) bwrite(bp);
1056 			}
1057 			BO_LOCK(bo);
1058 			return (EAGAIN);
1059 		}
1060 		bremfree(bp);
1061 		bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF);
1062 		bp->b_flags &= ~B_ASYNC;
1063 		brelse(bp);
1064 		BO_LOCK(bo);
1065 	}
1066 	return (retval);
1067 }
1068 
1069 /*
1070  * Truncate a file's buffer and pages to a specified length.  This
1071  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1072  * sync activity.
1073  */
1074 int
1075 vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td, off_t length, int blksize)
1076 {
1077 	struct buf *bp, *nbp;
1078 	int anyfreed;
1079 	int trunclbn;
1080 	struct bufobj *bo;
1081 
1082 	/*
1083 	 * Round up to the *next* lbn.
1084 	 */
1085 	trunclbn = (length + blksize - 1) / blksize;
1086 
1087 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1088 restart:
1089 	VI_LOCK(vp);
1090 	bo = &vp->v_bufobj;
1091 	anyfreed = 1;
1092 	for (;anyfreed;) {
1093 		anyfreed = 0;
1094 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1095 			if (bp->b_lblkno < trunclbn)
1096 				continue;
1097 			if (BUF_LOCK(bp,
1098 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1099 			    VI_MTX(vp)) == ENOLCK)
1100 				goto restart;
1101 
1102 			bremfree(bp);
1103 			bp->b_flags |= (B_INVAL | B_RELBUF);
1104 			bp->b_flags &= ~B_ASYNC;
1105 			brelse(bp);
1106 			anyfreed = 1;
1107 
1108 			if (nbp != NULL &&
1109 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1110 			    (nbp->b_vp != vp) ||
1111 			    (nbp->b_flags & B_DELWRI))) {
1112 				goto restart;
1113 			}
1114 			VI_LOCK(vp);
1115 		}
1116 
1117 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1118 			if (bp->b_lblkno < trunclbn)
1119 				continue;
1120 			if (BUF_LOCK(bp,
1121 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1122 			    VI_MTX(vp)) == ENOLCK)
1123 				goto restart;
1124 			bremfree(bp);
1125 			bp->b_flags |= (B_INVAL | B_RELBUF);
1126 			bp->b_flags &= ~B_ASYNC;
1127 			brelse(bp);
1128 			anyfreed = 1;
1129 			if (nbp != NULL &&
1130 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1131 			    (nbp->b_vp != vp) ||
1132 			    (nbp->b_flags & B_DELWRI) == 0)) {
1133 				goto restart;
1134 			}
1135 			VI_LOCK(vp);
1136 		}
1137 	}
1138 
1139 	if (length > 0) {
1140 restartsync:
1141 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1142 			if (bp->b_lblkno > 0)
1143 				continue;
1144 			/*
1145 			 * Since we hold the vnode lock this should only
1146 			 * fail if we're racing with the buf daemon.
1147 			 */
1148 			if (BUF_LOCK(bp,
1149 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1150 			    VI_MTX(vp)) == ENOLCK) {
1151 				goto restart;
1152 			}
1153 			KASSERT((bp->b_flags & B_DELWRI),
1154 			    ("buf(%p) on dirty queue without DELWRI", bp));
1155 
1156 			bremfree(bp);
1157 			bawrite(bp);
1158 			VI_LOCK(vp);
1159 			goto restartsync;
1160 		}
1161 	}
1162 
1163 	bufobj_wwait(bo, 0, 0);
1164 	VI_UNLOCK(vp);
1165 	vnode_pager_setsize(vp, length);
1166 
1167 	return (0);
1168 }
1169 
1170 /*
1171  * buf_splay() - splay tree core for the clean/dirty list of buffers in
1172  * 		 a vnode.
1173  *
1174  *	NOTE: We have to deal with the special case of a background bitmap
1175  *	buffer, a situation where two buffers will have the same logical
1176  *	block offset.  We want (1) only the foreground buffer to be accessed
1177  *	in a lookup and (2) must differentiate between the foreground and
1178  *	background buffer in the splay tree algorithm because the splay
1179  *	tree cannot normally handle multiple entities with the same 'index'.
1180  *	We accomplish this by adding differentiating flags to the splay tree's
1181  *	numerical domain.
1182  */
1183 static
1184 struct buf *
1185 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root)
1186 {
1187 	struct buf dummy;
1188 	struct buf *lefttreemax, *righttreemin, *y;
1189 
1190 	if (root == NULL)
1191 		return (NULL);
1192 	lefttreemax = righttreemin = &dummy;
1193 	for (;;) {
1194 		if (lblkno < root->b_lblkno ||
1195 		    (lblkno == root->b_lblkno &&
1196 		    (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1197 			if ((y = root->b_left) == NULL)
1198 				break;
1199 			if (lblkno < y->b_lblkno) {
1200 				/* Rotate right. */
1201 				root->b_left = y->b_right;
1202 				y->b_right = root;
1203 				root = y;
1204 				if ((y = root->b_left) == NULL)
1205 					break;
1206 			}
1207 			/* Link into the new root's right tree. */
1208 			righttreemin->b_left = root;
1209 			righttreemin = root;
1210 		} else if (lblkno > root->b_lblkno ||
1211 		    (lblkno == root->b_lblkno &&
1212 		    (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) {
1213 			if ((y = root->b_right) == NULL)
1214 				break;
1215 			if (lblkno > y->b_lblkno) {
1216 				/* Rotate left. */
1217 				root->b_right = y->b_left;
1218 				y->b_left = root;
1219 				root = y;
1220 				if ((y = root->b_right) == NULL)
1221 					break;
1222 			}
1223 			/* Link into the new root's left tree. */
1224 			lefttreemax->b_right = root;
1225 			lefttreemax = root;
1226 		} else {
1227 			break;
1228 		}
1229 		root = y;
1230 	}
1231 	/* Assemble the new root. */
1232 	lefttreemax->b_right = root->b_left;
1233 	righttreemin->b_left = root->b_right;
1234 	root->b_left = dummy.b_right;
1235 	root->b_right = dummy.b_left;
1236 	return (root);
1237 }
1238 
1239 static void
1240 buf_vlist_remove(struct buf *bp)
1241 {
1242 	struct buf *root;
1243 	struct bufv *bv;
1244 
1245 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1246 	ASSERT_BO_LOCKED(bp->b_bufobj);
1247 	if (bp->b_xflags & BX_VNDIRTY)
1248 		bv = &bp->b_bufobj->bo_dirty;
1249 	else
1250 		bv = &bp->b_bufobj->bo_clean;
1251 	if (bp != bv->bv_root) {
1252 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1253 		KASSERT(root == bp, ("splay lookup failed in remove"));
1254 	}
1255 	if (bp->b_left == NULL) {
1256 		root = bp->b_right;
1257 	} else {
1258 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left);
1259 		root->b_right = bp->b_right;
1260 	}
1261 	bv->bv_root = root;
1262 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1263 	bv->bv_cnt--;
1264 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1265 }
1266 
1267 /*
1268  * Add the buffer to the sorted clean or dirty block list using a
1269  * splay tree algorithm.
1270  *
1271  * NOTE: xflags is passed as a constant, optimizing this inline function!
1272  */
1273 static void
1274 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1275 {
1276 	struct buf *root;
1277 	struct bufv *bv;
1278 
1279 	ASSERT_BO_LOCKED(bo);
1280 	bp->b_xflags |= xflags;
1281 	if (xflags & BX_VNDIRTY)
1282 		bv = &bo->bo_dirty;
1283 	else
1284 		bv = &bo->bo_clean;
1285 
1286 	root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1287 	if (root == NULL) {
1288 		bp->b_left = NULL;
1289 		bp->b_right = NULL;
1290 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1291 	} else if (bp->b_lblkno < root->b_lblkno ||
1292 	    (bp->b_lblkno == root->b_lblkno &&
1293 	    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1294 		bp->b_left = root->b_left;
1295 		bp->b_right = root;
1296 		root->b_left = NULL;
1297 		TAILQ_INSERT_BEFORE(root, bp, b_bobufs);
1298 	} else {
1299 		bp->b_right = root->b_right;
1300 		bp->b_left = root;
1301 		root->b_right = NULL;
1302 		TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs);
1303 	}
1304 	bv->bv_cnt++;
1305 	bv->bv_root = bp;
1306 }
1307 
1308 /*
1309  * Lookup a buffer using the splay tree.  Note that we specifically avoid
1310  * shadow buffers used in background bitmap writes.
1311  *
1312  * This code isn't quite efficient as it could be because we are maintaining
1313  * two sorted lists and do not know which list the block resides in.
1314  *
1315  * During a "make buildworld" the desired buffer is found at one of
1316  * the roots more than 60% of the time.  Thus, checking both roots
1317  * before performing either splay eliminates unnecessary splays on the
1318  * first tree splayed.
1319  */
1320 struct buf *
1321 gbincore(struct bufobj *bo, daddr_t lblkno)
1322 {
1323 	struct buf *bp;
1324 
1325 	GIANT_REQUIRED;
1326 
1327 	ASSERT_BO_LOCKED(bo);
1328 	if ((bp = bo->bo_clean.bv_root) != NULL &&
1329 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1330 		return (bp);
1331 	if ((bp = bo->bo_dirty.bv_root) != NULL &&
1332 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1333 		return (bp);
1334 	if ((bp = bo->bo_clean.bv_root) != NULL) {
1335 		bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp);
1336 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1337 			return (bp);
1338 	}
1339 	if ((bp = bo->bo_dirty.bv_root) != NULL) {
1340 		bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp);
1341 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1342 			return (bp);
1343 	}
1344 	return (NULL);
1345 }
1346 
1347 /*
1348  * Associate a buffer with a vnode.
1349  */
1350 void
1351 bgetvp(struct vnode *vp, struct buf *bp)
1352 {
1353 
1354 	KASSERT(bp->b_vp == NULL, ("bgetvp: not free"));
1355 
1356 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1357 	    ("bgetvp: bp already attached! %p", bp));
1358 
1359 	ASSERT_VI_LOCKED(vp, "bgetvp");
1360 	vholdl(vp);
1361 	bp->b_vp = vp;
1362 	bp->b_bufobj = &vp->v_bufobj;
1363 	/*
1364 	 * Insert onto list for new vnode.
1365 	 */
1366 	buf_vlist_add(bp, &vp->v_bufobj, BX_VNCLEAN);
1367 }
1368 
1369 /*
1370  * Disassociate a buffer from a vnode.
1371  */
1372 void
1373 brelvp(struct buf *bp)
1374 {
1375 	struct bufobj *bo;
1376 	struct vnode *vp;
1377 
1378 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1379 
1380 	/*
1381 	 * Delete from old vnode list, if on one.
1382 	 */
1383 	vp = bp->b_vp;		/* XXX */
1384 	bo = bp->b_bufobj;
1385 	BO_LOCK(bo);
1386 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1387 		buf_vlist_remove(bp);
1388 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1389 		bo->bo_flag &= ~BO_ONWORKLST;
1390 		mtx_lock(&sync_mtx);
1391 		LIST_REMOVE(bo, bo_synclist);
1392  		syncer_worklist_len--;
1393 		mtx_unlock(&sync_mtx);
1394 	}
1395 	vdropl(vp);
1396 	bp->b_vp = NULL;
1397 	bp->b_bufobj = NULL;
1398 	BO_UNLOCK(bo);
1399 }
1400 
1401 /*
1402  * Add an item to the syncer work queue.
1403  */
1404 static void
1405 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
1406 {
1407 	int slot;
1408 
1409 	ASSERT_BO_LOCKED(bo);
1410 
1411 	mtx_lock(&sync_mtx);
1412 	if (bo->bo_flag & BO_ONWORKLST)
1413 		LIST_REMOVE(bo, bo_synclist);
1414 	else {
1415 		bo->bo_flag |= BO_ONWORKLST;
1416  		syncer_worklist_len++;
1417 	}
1418 
1419 	if (delay > syncer_maxdelay - 2)
1420 		delay = syncer_maxdelay - 2;
1421 	slot = (syncer_delayno + delay) & syncer_mask;
1422 
1423 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
1424 	mtx_unlock(&sync_mtx);
1425 }
1426 
1427 static int
1428 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
1429 {
1430 	int error, len;
1431 
1432 	mtx_lock(&sync_mtx);
1433 	len = syncer_worklist_len - sync_vnode_count;
1434 	mtx_unlock(&sync_mtx);
1435 	error = SYSCTL_OUT(req, &len, sizeof(len));
1436 	return (error);
1437 }
1438 
1439 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
1440     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
1441 
1442 struct  proc *updateproc;
1443 static void sched_sync(void);
1444 static struct kproc_desc up_kp = {
1445 	"syncer",
1446 	sched_sync,
1447 	&updateproc
1448 };
1449 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp)
1450 
1451 static int
1452 sync_vnode(struct bufobj *bo, struct thread *td)
1453 {
1454 	struct vnode *vp;
1455 	struct mount *mp;
1456 
1457 	vp = bo->__bo_vnode; 	/* XXX */
1458 	if (VOP_ISLOCKED(vp, NULL) != 0)
1459 		return (1);
1460 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0)
1461 		return (1);
1462 	if (VI_TRYLOCK(vp) == 0) {
1463 		vn_finished_write(mp);
1464 		return (1);
1465 	}
1466 	/*
1467 	 * We use vhold in case the vnode does not
1468 	 * successfully sync.  vhold prevents the vnode from
1469 	 * going away when we unlock the sync_mtx so that
1470 	 * we can acquire the vnode interlock.
1471 	 */
1472 	vholdl(vp);
1473 	mtx_unlock(&sync_mtx);
1474 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK, td);
1475 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
1476 	VOP_UNLOCK(vp, 0, td);
1477 	vn_finished_write(mp);
1478 	VI_LOCK(vp);
1479 	if ((bo->bo_flag & BO_ONWORKLST) != 0) {
1480 		/*
1481 		 * Put us back on the worklist.  The worklist
1482 		 * routine will remove us from our current
1483 		 * position and then add us back in at a later
1484 		 * position.
1485 		 */
1486 		vn_syncer_add_to_worklist(bo, syncdelay);
1487 	}
1488 	vdropl(vp);
1489 	VI_UNLOCK(vp);
1490 	mtx_lock(&sync_mtx);
1491 	return (0);
1492 }
1493 
1494 /*
1495  * System filesystem synchronizer daemon.
1496  */
1497 static void
1498 sched_sync(void)
1499 {
1500 	struct synclist *next;
1501 	struct synclist *slp;
1502 	struct bufobj *bo;
1503 	long starttime;
1504 	struct thread *td = FIRST_THREAD_IN_PROC(updateproc);
1505 	static int dummychan;
1506 	int last_work_seen;
1507 	int net_worklist_len;
1508 	int syncer_final_iter;
1509 	int first_printf;
1510 	int error;
1511 
1512 	mtx_lock(&Giant);
1513 	last_work_seen = 0;
1514 	syncer_final_iter = 0;
1515 	first_printf = 1;
1516 	syncer_state = SYNCER_RUNNING;
1517 	starttime = time_second;
1518 
1519 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
1520 	    SHUTDOWN_PRI_LAST);
1521 
1522 	for (;;) {
1523 		mtx_lock(&sync_mtx);
1524 		if (syncer_state == SYNCER_FINAL_DELAY &&
1525 		    syncer_final_iter == 0) {
1526 			mtx_unlock(&sync_mtx);
1527 			kthread_suspend_check(td->td_proc);
1528 			mtx_lock(&sync_mtx);
1529 		}
1530 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
1531 		if (syncer_state != SYNCER_RUNNING &&
1532 		    starttime != time_second) {
1533 			if (first_printf) {
1534 				printf("\nSyncing disks, vnodes remaining...");
1535 				first_printf = 0;
1536 			}
1537 			printf("%d ", net_worklist_len);
1538 		}
1539 		starttime = time_second;
1540 
1541 		/*
1542 		 * Push files whose dirty time has expired.  Be careful
1543 		 * of interrupt race on slp queue.
1544 		 *
1545 		 * Skip over empty worklist slots when shutting down.
1546 		 */
1547 		do {
1548 			slp = &syncer_workitem_pending[syncer_delayno];
1549 			syncer_delayno += 1;
1550 			if (syncer_delayno == syncer_maxdelay)
1551 				syncer_delayno = 0;
1552 			next = &syncer_workitem_pending[syncer_delayno];
1553 			/*
1554 			 * If the worklist has wrapped since the
1555 			 * it was emptied of all but syncer vnodes,
1556 			 * switch to the FINAL_DELAY state and run
1557 			 * for one more second.
1558 			 */
1559 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
1560 			    net_worklist_len == 0 &&
1561 			    last_work_seen == syncer_delayno) {
1562 				syncer_state = SYNCER_FINAL_DELAY;
1563 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
1564 			}
1565 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
1566 		    syncer_worklist_len > 0);
1567 
1568 		/*
1569 		 * Keep track of the last time there was anything
1570 		 * on the worklist other than syncer vnodes.
1571 		 * Return to the SHUTTING_DOWN state if any
1572 		 * new work appears.
1573 		 */
1574 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
1575 			last_work_seen = syncer_delayno;
1576 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
1577 			syncer_state = SYNCER_SHUTTING_DOWN;
1578 		while ((bo = LIST_FIRST(slp)) != NULL) {
1579 			error = sync_vnode(bo, td);
1580 			if (error == 1) {
1581 				LIST_REMOVE(bo, bo_synclist);
1582 				LIST_INSERT_HEAD(next, bo, bo_synclist);
1583 				continue;
1584 			}
1585 		}
1586 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
1587 			syncer_final_iter--;
1588 		mtx_unlock(&sync_mtx);
1589 
1590 		/*
1591 		 * Do soft update processing.
1592 		 */
1593 		if (softdep_process_worklist_hook != NULL)
1594 			(*softdep_process_worklist_hook)(NULL);
1595 
1596 		/*
1597 		 * The variable rushjob allows the kernel to speed up the
1598 		 * processing of the filesystem syncer process. A rushjob
1599 		 * value of N tells the filesystem syncer to process the next
1600 		 * N seconds worth of work on its queue ASAP. Currently rushjob
1601 		 * is used by the soft update code to speed up the filesystem
1602 		 * syncer process when the incore state is getting so far
1603 		 * ahead of the disk that the kernel memory pool is being
1604 		 * threatened with exhaustion.
1605 		 */
1606 		mtx_lock(&sync_mtx);
1607 		if (rushjob > 0) {
1608 			rushjob -= 1;
1609 			mtx_unlock(&sync_mtx);
1610 			continue;
1611 		}
1612 		mtx_unlock(&sync_mtx);
1613 		/*
1614 		 * Just sleep for a short period if time between
1615 		 * iterations when shutting down to allow some I/O
1616 		 * to happen.
1617 		 *
1618 		 * If it has taken us less than a second to process the
1619 		 * current work, then wait. Otherwise start right over
1620 		 * again. We can still lose time if any single round
1621 		 * takes more than two seconds, but it does not really
1622 		 * matter as we are just trying to generally pace the
1623 		 * filesystem activity.
1624 		 */
1625 		if (syncer_state != SYNCER_RUNNING)
1626 			tsleep(&dummychan, PPAUSE, "syncfnl",
1627 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
1628 		else if (time_second == starttime)
1629 			tsleep(&lbolt, PPAUSE, "syncer", 0);
1630 	}
1631 }
1632 
1633 /*
1634  * Request the syncer daemon to speed up its work.
1635  * We never push it to speed up more than half of its
1636  * normal turn time, otherwise it could take over the cpu.
1637  */
1638 int
1639 speedup_syncer()
1640 {
1641 	struct thread *td;
1642 	int ret = 0;
1643 
1644 	td = FIRST_THREAD_IN_PROC(updateproc);
1645 	sleepq_remove(td, &lbolt);
1646 	mtx_lock(&sync_mtx);
1647 	if (rushjob < syncdelay / 2) {
1648 		rushjob += 1;
1649 		stat_rush_requests += 1;
1650 		ret = 1;
1651 	}
1652 	mtx_unlock(&sync_mtx);
1653 	return (ret);
1654 }
1655 
1656 /*
1657  * Tell the syncer to speed up its work and run though its work
1658  * list several times, then tell it to shut down.
1659  */
1660 static void
1661 syncer_shutdown(void *arg, int howto)
1662 {
1663 	struct thread *td;
1664 
1665 	if (howto & RB_NOSYNC)
1666 		return;
1667 	td = FIRST_THREAD_IN_PROC(updateproc);
1668 	sleepq_remove(td, &lbolt);
1669 	mtx_lock(&sync_mtx);
1670 	syncer_state = SYNCER_SHUTTING_DOWN;
1671 	rushjob = 0;
1672 	mtx_unlock(&sync_mtx);
1673 	kproc_shutdown(arg, howto);
1674 }
1675 
1676 /*
1677  * Reassign a buffer from one vnode to another.
1678  * Used to assign file specific control information
1679  * (indirect blocks) to the vnode to which they belong.
1680  */
1681 void
1682 reassignbuf(struct buf *bp)
1683 {
1684 	struct vnode *vp;
1685 	struct bufobj *bo;
1686 	int delay;
1687 
1688 	vp = bp->b_vp;
1689 	bo = bp->b_bufobj;
1690 	++reassignbufcalls;
1691 
1692 	/*
1693 	 * B_PAGING flagged buffers cannot be reassigned because their vp
1694 	 * is not fully linked in.
1695 	 */
1696 	if (bp->b_flags & B_PAGING)
1697 		panic("cannot reassign paging buffer");
1698 
1699 	/*
1700 	 * Delete from old vnode list, if on one.
1701 	 */
1702 	VI_LOCK(vp);
1703 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1704 		buf_vlist_remove(bp);
1705 	/*
1706 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
1707 	 * of clean buffers.
1708 	 */
1709 	if (bp->b_flags & B_DELWRI) {
1710 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
1711 			switch (vp->v_type) {
1712 			case VDIR:
1713 				delay = dirdelay;
1714 				break;
1715 			case VCHR:
1716 				delay = metadelay;
1717 				break;
1718 			default:
1719 				delay = filedelay;
1720 			}
1721 			vn_syncer_add_to_worklist(bo, delay);
1722 		}
1723 		buf_vlist_add(bp, bo, BX_VNDIRTY);
1724 	} else {
1725 		buf_vlist_add(bp, bo, BX_VNCLEAN);
1726 
1727 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1728 			mtx_lock(&sync_mtx);
1729 			LIST_REMOVE(bo, bo_synclist);
1730  			syncer_worklist_len--;
1731 			mtx_unlock(&sync_mtx);
1732 			bo->bo_flag &= ~BO_ONWORKLST;
1733 		}
1734 	}
1735 	VI_UNLOCK(vp);
1736 }
1737 
1738 static void
1739 v_incr_usecount(struct vnode *vp, int delta)
1740 {
1741 
1742 	vp->v_usecount += delta;
1743 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1744 		dev_lock();
1745 		vp->v_rdev->si_usecount += delta;
1746 		dev_unlock();
1747 	}
1748 }
1749 
1750 /*
1751  * Grab a particular vnode from the free list, increment its
1752  * reference count and lock it. The vnode lock bit is set if the
1753  * vnode is being eliminated in vgone. The process is awakened
1754  * when the transition is completed, and an error returned to
1755  * indicate that the vnode is no longer usable (possibly having
1756  * been changed to a new filesystem type).
1757  */
1758 int
1759 vget(vp, flags, td)
1760 	register struct vnode *vp;
1761 	int flags;
1762 	struct thread *td;
1763 {
1764 	int error;
1765 
1766 	/*
1767 	 * If the vnode is in the process of being cleaned out for
1768 	 * another use, we wait for the cleaning to finish and then
1769 	 * return failure. Cleaning is determined by checking that
1770 	 * the VI_XLOCK flag is set.
1771 	 */
1772 	if ((flags & LK_INTERLOCK) == 0)
1773 		VI_LOCK(vp);
1774 	if (vp->v_iflag & VI_XLOCK && vp->v_vxthread != curthread) {
1775 		if ((flags & LK_NOWAIT) == 0) {
1776 			vp->v_iflag |= VI_XWANT;
1777 			msleep(vp, VI_MTX(vp), PINOD | PDROP, "vget", 0);
1778 			return (ENOENT);
1779 		}
1780 		VI_UNLOCK(vp);
1781 		return (EBUSY);
1782 	}
1783 
1784 	v_incr_usecount(vp, 1);
1785 
1786 	if (VSHOULDBUSY(vp))
1787 		vbusy(vp);
1788 	if (flags & LK_TYPE_MASK) {
1789 		if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) {
1790 			/*
1791 			 * must expand vrele here because we do not want
1792 			 * to call VOP_INACTIVE if the reference count
1793 			 * drops back to zero since it was never really
1794 			 * active. We must remove it from the free list
1795 			 * before sleeping so that multiple processes do
1796 			 * not try to recycle it.
1797 			 */
1798 			VI_LOCK(vp);
1799 			v_incr_usecount(vp, -1);
1800 			if (VSHOULDFREE(vp))
1801 				vfree(vp);
1802 			else
1803 				vlruvp(vp);
1804 			VI_UNLOCK(vp);
1805 		}
1806 		return (error);
1807 	}
1808 	VI_UNLOCK(vp);
1809 	return (0);
1810 }
1811 
1812 /*
1813  * Increase the reference count of a vnode.
1814  */
1815 void
1816 vref(struct vnode *vp)
1817 {
1818 
1819 	VI_LOCK(vp);
1820 	v_incr_usecount(vp, 1);
1821 	VI_UNLOCK(vp);
1822 }
1823 
1824 /*
1825  * Return reference count of a vnode.
1826  *
1827  * The results of this call are only guaranteed when some mechanism other
1828  * than the VI lock is used to stop other processes from gaining references
1829  * to the vnode.  This may be the case if the caller holds the only reference.
1830  * This is also useful when stale data is acceptable as race conditions may
1831  * be accounted for by some other means.
1832  */
1833 int
1834 vrefcnt(struct vnode *vp)
1835 {
1836 	int usecnt;
1837 
1838 	VI_LOCK(vp);
1839 	usecnt = vp->v_usecount;
1840 	VI_UNLOCK(vp);
1841 
1842 	return (usecnt);
1843 }
1844 
1845 
1846 /*
1847  * Vnode put/release.
1848  * If count drops to zero, call inactive routine and return to freelist.
1849  */
1850 void
1851 vrele(vp)
1852 	struct vnode *vp;
1853 {
1854 	struct thread *td = curthread;	/* XXX */
1855 
1856 	GIANT_REQUIRED;
1857 
1858 	KASSERT(vp != NULL, ("vrele: null vp"));
1859 
1860 	VI_LOCK(vp);
1861 
1862 	/* Skip this v_writecount check if we're going to panic below. */
1863 	KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1,
1864 	    ("vrele: missed vn_close"));
1865 
1866 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
1867 	    vp->v_usecount == 1)) {
1868 		v_incr_usecount(vp, -1);
1869 		VI_UNLOCK(vp);
1870 
1871 		return;
1872 	}
1873 
1874 	if (vp->v_usecount == 1) {
1875 		v_incr_usecount(vp, -1);
1876 		/*
1877 		 * We must call VOP_INACTIVE with the node locked. Mark
1878 		 * as VI_DOINGINACT to avoid recursion.
1879 		 */
1880 		if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) {
1881 			VI_LOCK(vp);
1882 			vp->v_iflag |= VI_DOINGINACT;
1883 			VI_UNLOCK(vp);
1884 			VOP_INACTIVE(vp, td);
1885 			VI_LOCK(vp);
1886 			KASSERT(vp->v_iflag & VI_DOINGINACT,
1887 			    ("vrele: lost VI_DOINGINACT"));
1888 			vp->v_iflag &= ~VI_DOINGINACT;
1889 		} else
1890 			VI_LOCK(vp);
1891 		if (VSHOULDFREE(vp))
1892 			vfree(vp);
1893 		else
1894 			vlruvp(vp);
1895 		VI_UNLOCK(vp);
1896 
1897 	} else {
1898 #ifdef DIAGNOSTIC
1899 		vprint("vrele: negative ref count", vp);
1900 #endif
1901 		VI_UNLOCK(vp);
1902 		panic("vrele: negative ref cnt");
1903 	}
1904 }
1905 
1906 /*
1907  * Release an already locked vnode.  This give the same effects as
1908  * unlock+vrele(), but takes less time and avoids releasing and
1909  * re-aquiring the lock (as vrele() aquires the lock internally.)
1910  */
1911 void
1912 vput(vp)
1913 	struct vnode *vp;
1914 {
1915 	struct thread *td = curthread;	/* XXX */
1916 
1917 	GIANT_REQUIRED;
1918 
1919 	KASSERT(vp != NULL, ("vput: null vp"));
1920 	VI_LOCK(vp);
1921 	/* Skip this v_writecount check if we're going to panic below. */
1922 	KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1,
1923 	    ("vput: missed vn_close"));
1924 
1925 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
1926 	    vp->v_usecount == 1)) {
1927 		v_incr_usecount(vp, -1);
1928 		VOP_UNLOCK(vp, LK_INTERLOCK, td);
1929 		return;
1930 	}
1931 
1932 	if (vp->v_usecount == 1) {
1933 		v_incr_usecount(vp, -1);
1934 		/*
1935 		 * We must call VOP_INACTIVE with the node locked, so
1936 		 * we just need to release the vnode mutex. Mark as
1937 		 * as VI_DOINGINACT to avoid recursion.
1938 		 */
1939 		vp->v_iflag |= VI_DOINGINACT;
1940 		VI_UNLOCK(vp);
1941 		VOP_INACTIVE(vp, td);
1942 		VI_LOCK(vp);
1943 		KASSERT(vp->v_iflag & VI_DOINGINACT,
1944 		    ("vput: lost VI_DOINGINACT"));
1945 		vp->v_iflag &= ~VI_DOINGINACT;
1946 		if (VSHOULDFREE(vp))
1947 			vfree(vp);
1948 		else
1949 			vlruvp(vp);
1950 		VI_UNLOCK(vp);
1951 
1952 	} else {
1953 #ifdef DIAGNOSTIC
1954 		vprint("vput: negative ref count", vp);
1955 #endif
1956 		panic("vput: negative ref cnt");
1957 	}
1958 }
1959 
1960 /*
1961  * Somebody doesn't want the vnode recycled.
1962  */
1963 void
1964 vhold(struct vnode *vp)
1965 {
1966 
1967 	VI_LOCK(vp);
1968 	vholdl(vp);
1969 	VI_UNLOCK(vp);
1970 }
1971 
1972 void
1973 vholdl(struct vnode *vp)
1974 {
1975 
1976 	vp->v_holdcnt++;
1977 	if (VSHOULDBUSY(vp))
1978 		vbusy(vp);
1979 }
1980 
1981 /*
1982  * Note that there is one less who cares about this vnode.  vdrop() is the
1983  * opposite of vhold().
1984  */
1985 void
1986 vdrop(struct vnode *vp)
1987 {
1988 
1989 	VI_LOCK(vp);
1990 	vdropl(vp);
1991 	VI_UNLOCK(vp);
1992 }
1993 
1994 void
1995 vdropl(vp)
1996 	register struct vnode *vp;
1997 {
1998 
1999 	if (vp->v_holdcnt <= 0)
2000 		panic("vdrop: holdcnt");
2001 	vp->v_holdcnt--;
2002 	if (VSHOULDFREE(vp))
2003 		vfree(vp);
2004 	else
2005 		vlruvp(vp);
2006 }
2007 
2008 /*
2009  * Remove any vnodes in the vnode table belonging to mount point mp.
2010  *
2011  * If FORCECLOSE is not specified, there should not be any active ones,
2012  * return error if any are found (nb: this is a user error, not a
2013  * system error). If FORCECLOSE is specified, detach any active vnodes
2014  * that are found.
2015  *
2016  * If WRITECLOSE is set, only flush out regular file vnodes open for
2017  * writing.
2018  *
2019  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2020  *
2021  * `rootrefs' specifies the base reference count for the root vnode
2022  * of this filesystem. The root vnode is considered busy if its
2023  * v_usecount exceeds this value. On a successful return, vflush(, td)
2024  * will call vrele() on the root vnode exactly rootrefs times.
2025  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2026  * be zero.
2027  */
2028 #ifdef DIAGNOSTIC
2029 static int busyprt = 0;		/* print out busy vnodes */
2030 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
2031 #endif
2032 
2033 int
2034 vflush(mp, rootrefs, flags, td)
2035 	struct mount *mp;
2036 	int rootrefs;
2037 	int flags;
2038 	struct thread *td;
2039 {
2040 	struct vnode *vp, *nvp, *rootvp = NULL;
2041 	struct vattr vattr;
2042 	int busy = 0, error;
2043 
2044 	if (rootrefs > 0) {
2045 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2046 		    ("vflush: bad args"));
2047 		/*
2048 		 * Get the filesystem root vnode. We can vput() it
2049 		 * immediately, since with rootrefs > 0, it won't go away.
2050 		 */
2051 		if ((error = VFS_ROOT(mp, &rootvp, td)) != 0)
2052 			return (error);
2053 		vput(rootvp);
2054 
2055 	}
2056 	MNT_ILOCK(mp);
2057 loop:
2058 	MNT_VNODE_FOREACH(vp, mp, nvp) {
2059 
2060 		VI_LOCK(vp);
2061 		MNT_IUNLOCK(mp);
2062 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td);
2063 		if (error) {
2064 			MNT_ILOCK(mp);
2065 			goto loop;
2066 		}
2067 		/*
2068 		 * Skip over a vnodes marked VV_SYSTEM.
2069 		 */
2070 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2071 			VOP_UNLOCK(vp, 0, td);
2072 			MNT_ILOCK(mp);
2073 			continue;
2074 		}
2075 		/*
2076 		 * If WRITECLOSE is set, flush out unlinked but still open
2077 		 * files (even if open only for reading) and regular file
2078 		 * vnodes open for writing.
2079 		 */
2080 		if (flags & WRITECLOSE) {
2081 			error = VOP_GETATTR(vp, &vattr, td->td_ucred, td);
2082 			VI_LOCK(vp);
2083 
2084 			if ((vp->v_type == VNON ||
2085 			    (error == 0 && vattr.va_nlink > 0)) &&
2086 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2087 				VOP_UNLOCK(vp, LK_INTERLOCK, td);
2088 				MNT_ILOCK(mp);
2089 				continue;
2090 			}
2091 		} else
2092 			VI_LOCK(vp);
2093 
2094 		VOP_UNLOCK(vp, 0, td);
2095 
2096 		/*
2097 		 * With v_usecount == 0, all we need to do is clear out the
2098 		 * vnode data structures and we are done.
2099 		 */
2100 		if (vp->v_usecount == 0) {
2101 			vgonel(vp, td);
2102 			MNT_ILOCK(mp);
2103 			continue;
2104 		}
2105 
2106 		/*
2107 		 * If FORCECLOSE is set, forcibly close the vnode. For block
2108 		 * or character devices, revert to an anonymous device. For
2109 		 * all other files, just kill them.
2110 		 */
2111 		if (flags & FORCECLOSE) {
2112 			KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
2113 			    ("device VNODE %p is FORCECLOSED", vp));
2114 			vgonel(vp, td);
2115 			MNT_ILOCK(mp);
2116 			continue;
2117 		}
2118 #ifdef DIAGNOSTIC
2119 		if (busyprt)
2120 			vprint("vflush: busy vnode", vp);
2121 #endif
2122 		VI_UNLOCK(vp);
2123 		MNT_ILOCK(mp);
2124 		busy++;
2125 	}
2126 	MNT_IUNLOCK(mp);
2127 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2128 		/*
2129 		 * If just the root vnode is busy, and if its refcount
2130 		 * is equal to `rootrefs', then go ahead and kill it.
2131 		 */
2132 		VI_LOCK(rootvp);
2133 		KASSERT(busy > 0, ("vflush: not busy"));
2134 		KASSERT(rootvp->v_usecount >= rootrefs,
2135 		    ("vflush: usecount %d < rootrefs %d",
2136 		     rootvp->v_usecount, rootrefs));
2137 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2138 			vgonel(rootvp, td);
2139 			busy = 0;
2140 		} else
2141 			VI_UNLOCK(rootvp);
2142 	}
2143 	if (busy)
2144 		return (EBUSY);
2145 	for (; rootrefs > 0; rootrefs--)
2146 		vrele(rootvp);
2147 	return (0);
2148 }
2149 
2150 /*
2151  * This moves a now (likely recyclable) vnode to the end of the
2152  * mountlist.  XXX However, it is temporarily disabled until we
2153  * can clean up ffs_sync() and friends, which have loop restart
2154  * conditions which this code causes to operate O(N^2).
2155  */
2156 static void
2157 vlruvp(struct vnode *vp)
2158 {
2159 #if 0
2160 	struct mount *mp;
2161 
2162 	if ((mp = vp->v_mount) != NULL) {
2163 		MNT_ILOCK(mp);
2164 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2165 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2166 		MNT_IUNLOCK(mp);
2167 	}
2168 #endif
2169 }
2170 
2171 static void
2172 vx_lock(struct vnode *vp)
2173 {
2174 
2175 	ASSERT_VI_LOCKED(vp, "vx_lock");
2176 
2177 	/*
2178 	 * Prevent the vnode from being recycled or brought into use while we
2179 	 * clean it out.
2180 	 */
2181 	if (vp->v_iflag & VI_XLOCK)
2182 		panic("vclean: deadlock");
2183 	vp->v_iflag |= VI_XLOCK;
2184 	vp->v_vxthread = curthread;
2185 }
2186 
2187 static void
2188 vx_unlock(struct vnode *vp)
2189 {
2190 	ASSERT_VI_LOCKED(vp, "vx_unlock");
2191 	vp->v_iflag &= ~VI_XLOCK;
2192 	vp->v_vxthread = NULL;
2193 	if (vp->v_iflag & VI_XWANT) {
2194 		vp->v_iflag &= ~VI_XWANT;
2195 		wakeup(vp);
2196 	}
2197 }
2198 
2199 /*
2200  * Disassociate the underlying filesystem from a vnode.
2201  */
2202 static void
2203 vclean(vp, flags, td)
2204 	struct vnode *vp;
2205 	int flags;
2206 	struct thread *td;
2207 {
2208 	int active;
2209 
2210 	ASSERT_VI_LOCKED(vp, "vclean");
2211 	/*
2212 	 * Check to see if the vnode is in use. If so we have to reference it
2213 	 * before we clean it out so that its count cannot fall to zero and
2214 	 * generate a race against ourselves to recycle it.
2215 	 */
2216 	if ((active = vp->v_usecount))
2217 		v_incr_usecount(vp, 1);
2218 
2219 	/*
2220 	 * Even if the count is zero, the VOP_INACTIVE routine may still
2221 	 * have the object locked while it cleans it out. The VOP_LOCK
2222 	 * ensures that the VOP_INACTIVE routine is done with its work.
2223 	 * For active vnodes, it ensures that no other activity can
2224 	 * occur while the underlying object is being cleaned out.
2225 	 */
2226 	VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td);
2227 
2228 	/*
2229 	 * Clean out any buffers associated with the vnode.
2230 	 * If the flush fails, just toss the buffers.
2231 	 */
2232 	if (flags & DOCLOSE) {
2233 		struct buf *bp;
2234 		bp = TAILQ_FIRST(&vp->v_bufobj.bo_dirty.bv_hd);
2235 		if (bp != NULL)
2236 			(void) vn_write_suspend_wait(vp, NULL, V_WAIT);
2237 		if (vinvalbuf(vp, V_SAVE, td, 0, 0) != 0)
2238 			vinvalbuf(vp, 0, td, 0, 0);
2239 	}
2240 
2241 	VOP_DESTROYVOBJECT(vp);
2242 
2243 	/*
2244 	 * Any other processes trying to obtain this lock must first
2245 	 * wait for VXLOCK to clear, then call the new lock operation.
2246 	 */
2247 	VOP_UNLOCK(vp, 0, td);
2248 
2249 	/*
2250 	 * If purging an active vnode, it must be closed and
2251 	 * deactivated before being reclaimed. Note that the
2252 	 * VOP_INACTIVE will unlock the vnode.
2253 	 */
2254 	if (active) {
2255 		if (flags & DOCLOSE)
2256 			VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
2257 		VI_LOCK(vp);
2258 		if ((vp->v_iflag & VI_DOINGINACT) == 0) {
2259 			vp->v_iflag |= VI_DOINGINACT;
2260 			VI_UNLOCK(vp);
2261 			if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
2262 				panic("vclean: cannot relock.");
2263 			VOP_INACTIVE(vp, td);
2264 			VI_LOCK(vp);
2265 			KASSERT(vp->v_iflag & VI_DOINGINACT,
2266 			    ("vclean: lost VI_DOINGINACT"));
2267 			vp->v_iflag &= ~VI_DOINGINACT;
2268 		}
2269 		VI_UNLOCK(vp);
2270 	}
2271 	/*
2272 	 * Reclaim the vnode.
2273 	 */
2274 	if (VOP_RECLAIM(vp, td))
2275 		panic("vclean: cannot reclaim");
2276 
2277 	if (active) {
2278 		/*
2279 		 * Inline copy of vrele() since VOP_INACTIVE
2280 		 * has already been called.
2281 		 */
2282 		VI_LOCK(vp);
2283 		v_incr_usecount(vp, -1);
2284 		if (vp->v_usecount <= 0) {
2285 #ifdef INVARIANTS
2286 			if (vp->v_usecount < 0 || vp->v_writecount != 0) {
2287 				vprint("vclean: bad ref count", vp);
2288 				panic("vclean: ref cnt");
2289 			}
2290 #endif
2291 			if (VSHOULDFREE(vp))
2292 				vfree(vp);
2293 		}
2294 		VI_UNLOCK(vp);
2295 	}
2296 	/*
2297 	 * Delete from old mount point vnode list.
2298 	 */
2299 	delmntque(vp);
2300 	cache_purge(vp);
2301 	VI_LOCK(vp);
2302 	if (VSHOULDFREE(vp))
2303 		vfree(vp);
2304 
2305 	/*
2306 	 * Done with purge, reset to the standard lock and
2307 	 * notify sleepers of the grim news.
2308 	 */
2309 	vp->v_vnlock = &vp->v_lock;
2310 	vp->v_op = &dead_vnodeops;
2311 	if (vp->v_pollinfo != NULL)
2312 		vn_pollgone(vp);
2313 	vp->v_tag = "none";
2314 }
2315 
2316 /*
2317  * Recycle an unused vnode to the front of the free list.
2318  * Release the passed interlock if the vnode will be recycled.
2319  */
2320 int
2321 vrecycle(vp, inter_lkp, td)
2322 	struct vnode *vp;
2323 	struct mtx *inter_lkp;
2324 	struct thread *td;
2325 {
2326 
2327 	VI_LOCK(vp);
2328 	if (vp->v_usecount == 0) {
2329 		if (inter_lkp) {
2330 			mtx_unlock(inter_lkp);
2331 		}
2332 		vgonel(vp, td);
2333 		return (1);
2334 	}
2335 	VI_UNLOCK(vp);
2336 	return (0);
2337 }
2338 
2339 /*
2340  * Eliminate all activity associated with a vnode
2341  * in preparation for reuse.
2342  */
2343 void
2344 vgone(vp)
2345 	register struct vnode *vp;
2346 {
2347 	struct thread *td = curthread;	/* XXX */
2348 
2349 	VI_LOCK(vp);
2350 	vgonel(vp, td);
2351 }
2352 
2353 /*
2354  * vgone, with the vp interlock held.
2355  */
2356 void
2357 vgonel(vp, td)
2358 	struct vnode *vp;
2359 	struct thread *td;
2360 {
2361 	/*
2362 	 * If a vgone (or vclean) is already in progress,
2363 	 * wait until it is done and return.
2364 	 */
2365 	ASSERT_VI_LOCKED(vp, "vgonel");
2366 	if (vp->v_iflag & VI_XLOCK) {
2367 		vp->v_iflag |= VI_XWANT;
2368 		msleep(vp, VI_MTX(vp), PINOD | PDROP, "vgone", 0);
2369 		return;
2370 	}
2371 	vx_lock(vp);
2372 
2373 	/*
2374 	 * Clean out the filesystem specific data.
2375 	 */
2376 	vclean(vp, DOCLOSE, td);
2377 	VI_UNLOCK(vp);
2378 
2379 	/*
2380 	 * If special device, remove it from special device alias list
2381 	 * if it is on one.
2382 	 */
2383 	VI_LOCK(vp);
2384 	if (vp->v_type == VCHR && vp->v_rdev != NULL)
2385 		dev_rel(vp);
2386 
2387 	/*
2388 	 * If it is on the freelist and not already at the head,
2389 	 * move it to the head of the list. The test of the
2390 	 * VDOOMED flag and the reference count of zero is because
2391 	 * it will be removed from the free list by getnewvnode,
2392 	 * but will not have its reference count incremented until
2393 	 * after calling vgone. If the reference count were
2394 	 * incremented first, vgone would (incorrectly) try to
2395 	 * close the previous instance of the underlying object.
2396 	 */
2397 	if (vp->v_usecount == 0 && !(vp->v_iflag & VI_DOOMED)) {
2398 		mtx_lock(&vnode_free_list_mtx);
2399 		if (vp->v_iflag & VI_FREE) {
2400 			TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2401 		} else {
2402 			vp->v_iflag |= VI_FREE;
2403 			freevnodes++;
2404 		}
2405 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2406 		mtx_unlock(&vnode_free_list_mtx);
2407 	}
2408 
2409 	vp->v_type = VBAD;
2410 	vx_unlock(vp);
2411 	VI_UNLOCK(vp);
2412 }
2413 
2414 /*
2415  * Lookup a vnode by device number.
2416  */
2417 int
2418 vfinddev(dev, vpp)
2419 	struct cdev *dev;
2420 	struct vnode **vpp;
2421 {
2422 	struct vnode *vp;
2423 
2424 	dev_lock();
2425 	SLIST_FOREACH(vp, &dev->si_hlist, v_specnext) {
2426 		*vpp = vp;
2427 		dev_unlock();
2428 		return (1);
2429 	}
2430 	dev_unlock();
2431 	return (0);
2432 }
2433 
2434 /*
2435  * Calculate the total number of references to a special device.
2436  */
2437 int
2438 vcount(vp)
2439 	struct vnode *vp;
2440 {
2441 	int count;
2442 
2443 	dev_lock();
2444 	count = vp->v_rdev->si_usecount;
2445 	dev_unlock();
2446 	return (count);
2447 }
2448 
2449 /*
2450  * Same as above, but using the struct cdev *as argument
2451  */
2452 int
2453 count_dev(dev)
2454 	struct cdev *dev;
2455 {
2456 	int count;
2457 
2458 	dev_lock();
2459 	count = dev->si_usecount;
2460 	dev_unlock();
2461 	return(count);
2462 }
2463 
2464 /*
2465  * Print out a description of a vnode.
2466  */
2467 static char *typename[] =
2468 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"};
2469 
2470 void
2471 vprint(label, vp)
2472 	char *label;
2473 	struct vnode *vp;
2474 {
2475 	char buf[96];
2476 
2477 	if (label != NULL)
2478 		printf("%s: %p: ", label, (void *)vp);
2479 	else
2480 		printf("%p: ", (void *)vp);
2481 	printf("tag %s, type %s\n    ", vp->v_tag, typename[vp->v_type]);
2482 	printf("usecount %d, writecount %d, refcount %d mountedhere %p\n",
2483 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere);
2484 	buf[0] = '\0';
2485 	if (vp->v_vflag & VV_ROOT)
2486 		strcat(buf, "|VV_ROOT");
2487 	if (vp->v_vflag & VV_TEXT)
2488 		strcat(buf, "|VV_TEXT");
2489 	if (vp->v_vflag & VV_SYSTEM)
2490 		strcat(buf, "|VV_SYSTEM");
2491 	if (vp->v_iflag & VI_XLOCK)
2492 		strcat(buf, "|VI_XLOCK");
2493 	if (vp->v_iflag & VI_XWANT)
2494 		strcat(buf, "|VI_XWANT");
2495 	if (vp->v_iflag & VI_DOOMED)
2496 		strcat(buf, "|VI_DOOMED");
2497 	if (vp->v_iflag & VI_FREE)
2498 		strcat(buf, "|VI_FREE");
2499 	if (vp->v_vflag & VV_OBJBUF)
2500 		strcat(buf, "|VV_OBJBUF");
2501 	if (buf[0] != '\0')
2502 		printf("    flags (%s)", &buf[1]);
2503 	if (mtx_owned(VI_MTX(vp)))
2504 		printf(" VI_LOCKed");
2505 	printf("\n    ");
2506 	lockmgr_printinfo(vp->v_vnlock);
2507 	printf("\n");
2508 	if (vp->v_data != NULL)
2509 		VOP_PRINT(vp);
2510 }
2511 
2512 #ifdef DDB
2513 #include <ddb/ddb.h>
2514 /*
2515  * List all of the locked vnodes in the system.
2516  * Called when debugging the kernel.
2517  */
2518 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
2519 {
2520 	struct mount *mp, *nmp;
2521 	struct vnode *vp;
2522 
2523 	/*
2524 	 * Note: because this is DDB, we can't obey the locking semantics
2525 	 * for these structures, which means we could catch an inconsistent
2526 	 * state and dereference a nasty pointer.  Not much to be done
2527 	 * about that.
2528 	 */
2529 	printf("Locked vnodes\n");
2530 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
2531 		nmp = TAILQ_NEXT(mp, mnt_list);
2532 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2533 			if (VOP_ISLOCKED(vp, NULL))
2534 				vprint(NULL, vp);
2535 		}
2536 		nmp = TAILQ_NEXT(mp, mnt_list);
2537 	}
2538 }
2539 #endif
2540 
2541 /*
2542  * Fill in a struct xvfsconf based on a struct vfsconf.
2543  */
2544 static void
2545 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp)
2546 {
2547 
2548 	strcpy(xvfsp->vfc_name, vfsp->vfc_name);
2549 	xvfsp->vfc_typenum = vfsp->vfc_typenum;
2550 	xvfsp->vfc_refcount = vfsp->vfc_refcount;
2551 	xvfsp->vfc_flags = vfsp->vfc_flags;
2552 	/*
2553 	 * These are unused in userland, we keep them
2554 	 * to not break binary compatibility.
2555 	 */
2556 	xvfsp->vfc_vfsops = NULL;
2557 	xvfsp->vfc_next = NULL;
2558 }
2559 
2560 /*
2561  * Top level filesystem related information gathering.
2562  */
2563 static int
2564 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
2565 {
2566 	struct vfsconf *vfsp;
2567 	struct xvfsconf xvfsp;
2568 	int error;
2569 
2570 	error = 0;
2571 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2572 		vfsconf2x(vfsp, &xvfsp);
2573 		error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp);
2574 		if (error)
2575 			break;
2576 	}
2577 	return (error);
2578 }
2579 
2580 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist,
2581     "S,xvfsconf", "List of all configured filesystems");
2582 
2583 #ifndef BURN_BRIDGES
2584 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
2585 
2586 static int
2587 vfs_sysctl(SYSCTL_HANDLER_ARGS)
2588 {
2589 	int *name = (int *)arg1 - 1;	/* XXX */
2590 	u_int namelen = arg2 + 1;	/* XXX */
2591 	struct vfsconf *vfsp;
2592 	struct xvfsconf xvfsp;
2593 
2594 	printf("WARNING: userland calling deprecated sysctl, "
2595 	    "please rebuild world\n");
2596 
2597 #if 1 || defined(COMPAT_PRELITE2)
2598 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
2599 	if (namelen == 1)
2600 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
2601 #endif
2602 
2603 	switch (name[1]) {
2604 	case VFS_MAXTYPENUM:
2605 		if (namelen != 2)
2606 			return (ENOTDIR);
2607 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
2608 	case VFS_CONF:
2609 		if (namelen != 3)
2610 			return (ENOTDIR);	/* overloaded */
2611 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list)
2612 			if (vfsp->vfc_typenum == name[2])
2613 				break;
2614 		if (vfsp == NULL)
2615 			return (EOPNOTSUPP);
2616 		vfsconf2x(vfsp, &xvfsp);
2617 		return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
2618 	}
2619 	return (EOPNOTSUPP);
2620 }
2621 
2622 SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP, vfs_sysctl,
2623 	"Generic filesystem");
2624 
2625 #if 1 || defined(COMPAT_PRELITE2)
2626 
2627 static int
2628 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
2629 {
2630 	int error;
2631 	struct vfsconf *vfsp;
2632 	struct ovfsconf ovfs;
2633 
2634 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2635 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
2636 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
2637 		ovfs.vfc_index = vfsp->vfc_typenum;
2638 		ovfs.vfc_refcount = vfsp->vfc_refcount;
2639 		ovfs.vfc_flags = vfsp->vfc_flags;
2640 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
2641 		if (error)
2642 			return error;
2643 	}
2644 	return 0;
2645 }
2646 
2647 #endif /* 1 || COMPAT_PRELITE2 */
2648 #endif /* !BURN_BRIDGES */
2649 
2650 #define KINFO_VNODESLOP		10
2651 #ifdef notyet
2652 /*
2653  * Dump vnode list (via sysctl).
2654  */
2655 /* ARGSUSED */
2656 static int
2657 sysctl_vnode(SYSCTL_HANDLER_ARGS)
2658 {
2659 	struct xvnode *xvn;
2660 	struct thread *td = req->td;
2661 	struct mount *mp;
2662 	struct vnode *vp;
2663 	int error, len, n;
2664 
2665 	/*
2666 	 * Stale numvnodes access is not fatal here.
2667 	 */
2668 	req->lock = 0;
2669 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
2670 	if (!req->oldptr)
2671 		/* Make an estimate */
2672 		return (SYSCTL_OUT(req, 0, len));
2673 
2674 	error = sysctl_wire_old_buffer(req, 0);
2675 	if (error != 0)
2676 		return (error);
2677 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
2678 	n = 0;
2679 	mtx_lock(&mountlist_mtx);
2680 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
2681 		if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td))
2682 			continue;
2683 		MNT_ILOCK(mp);
2684 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2685 			if (n == len)
2686 				break;
2687 			vref(vp);
2688 			xvn[n].xv_size = sizeof *xvn;
2689 			xvn[n].xv_vnode = vp;
2690 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
2691 			XV_COPY(usecount);
2692 			XV_COPY(writecount);
2693 			XV_COPY(holdcnt);
2694 			XV_COPY(id);
2695 			XV_COPY(mount);
2696 			XV_COPY(numoutput);
2697 			XV_COPY(type);
2698 #undef XV_COPY
2699 			xvn[n].xv_flag = vp->v_vflag;
2700 
2701 			switch (vp->v_type) {
2702 			case VREG:
2703 			case VDIR:
2704 			case VLNK:
2705 				break;
2706 			case VBLK:
2707 			case VCHR:
2708 				if (vp->v_rdev == NULL) {
2709 					vrele(vp);
2710 					continue;
2711 				}
2712 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
2713 				break;
2714 			case VSOCK:
2715 				xvn[n].xv_socket = vp->v_socket;
2716 				break;
2717 			case VFIFO:
2718 				xvn[n].xv_fifo = vp->v_fifoinfo;
2719 				break;
2720 			case VNON:
2721 			case VBAD:
2722 			default:
2723 				/* shouldn't happen? */
2724 				vrele(vp);
2725 				continue;
2726 			}
2727 			vrele(vp);
2728 			++n;
2729 		}
2730 		MNT_IUNLOCK(mp);
2731 		mtx_lock(&mountlist_mtx);
2732 		vfs_unbusy(mp, td);
2733 		if (n == len)
2734 			break;
2735 	}
2736 	mtx_unlock(&mountlist_mtx);
2737 
2738 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
2739 	free(xvn, M_TEMP);
2740 	return (error);
2741 }
2742 
2743 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD,
2744 	0, 0, sysctl_vnode, "S,xvnode", "");
2745 #endif
2746 
2747 /*
2748  * Unmount all filesystems. The list is traversed in reverse order
2749  * of mounting to avoid dependencies.
2750  */
2751 void
2752 vfs_unmountall()
2753 {
2754 	struct mount *mp;
2755 	struct thread *td;
2756 	int error;
2757 
2758 	if (curthread != NULL)
2759 		td = curthread;
2760 	else
2761 		td = FIRST_THREAD_IN_PROC(initproc); /* XXX XXX proc0? */
2762 	/*
2763 	 * Since this only runs when rebooting, it is not interlocked.
2764 	 */
2765 	while(!TAILQ_EMPTY(&mountlist)) {
2766 		mp = TAILQ_LAST(&mountlist, mntlist);
2767 		error = dounmount(mp, MNT_FORCE, td);
2768 		if (error) {
2769 			TAILQ_REMOVE(&mountlist, mp, mnt_list);
2770 			printf("unmount of %s failed (",
2771 			    mp->mnt_stat.f_mntonname);
2772 			if (error == EBUSY)
2773 				printf("BUSY)\n");
2774 			else
2775 				printf("%d)\n", error);
2776 		} else {
2777 			/* The unmount has removed mp from the mountlist */
2778 		}
2779 	}
2780 }
2781 
2782 /*
2783  * perform msync on all vnodes under a mount point
2784  * the mount point must be locked.
2785  */
2786 void
2787 vfs_msync(struct mount *mp, int flags)
2788 {
2789 	struct vnode *vp, *nvp;
2790 	struct vm_object *obj;
2791 	int tries;
2792 
2793 	GIANT_REQUIRED;
2794 
2795 	tries = 5;
2796 	MNT_ILOCK(mp);
2797 loop:
2798 	TAILQ_FOREACH_SAFE(vp, &mp->mnt_nvnodelist, v_nmntvnodes, nvp) {
2799 		if (vp->v_mount != mp) {
2800 			if (--tries > 0)
2801 				goto loop;
2802 			break;
2803 		}
2804 
2805 		VI_LOCK(vp);
2806 		if (vp->v_iflag & VI_XLOCK) {
2807 			VI_UNLOCK(vp);
2808 			continue;
2809 		}
2810 
2811 		if ((vp->v_iflag & VI_OBJDIRTY) &&
2812 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) {
2813 			MNT_IUNLOCK(mp);
2814 			if (!vget(vp,
2815 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
2816 			    curthread)) {
2817 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
2818 					vput(vp);
2819 					MNT_ILOCK(mp);
2820 					continue;
2821 				}
2822 
2823 				if (VOP_GETVOBJECT(vp, &obj) == 0) {
2824 					VM_OBJECT_LOCK(obj);
2825 					vm_object_page_clean(obj, 0, 0,
2826 					    flags == MNT_WAIT ?
2827 					    OBJPC_SYNC : OBJPC_NOSYNC);
2828 					VM_OBJECT_UNLOCK(obj);
2829 				}
2830 				vput(vp);
2831 			}
2832 			MNT_ILOCK(mp);
2833 			if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) {
2834 				if (--tries > 0)
2835 					goto loop;
2836 				break;
2837 			}
2838 		} else
2839 			VI_UNLOCK(vp);
2840 	}
2841 	MNT_IUNLOCK(mp);
2842 }
2843 
2844 /*
2845  * Mark a vnode as free, putting it up for recycling.
2846  */
2847 void
2848 vfree(struct vnode *vp)
2849 {
2850 
2851 	ASSERT_VI_LOCKED(vp, "vfree");
2852 	mtx_lock(&vnode_free_list_mtx);
2853 	KASSERT((vp->v_iflag & VI_FREE) == 0, ("vnode already free"));
2854 	if (vp->v_iflag & VI_AGE) {
2855 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2856 	} else {
2857 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
2858 	}
2859 	freevnodes++;
2860 	mtx_unlock(&vnode_free_list_mtx);
2861 	vp->v_iflag &= ~VI_AGE;
2862 	vp->v_iflag |= VI_FREE;
2863 }
2864 
2865 /*
2866  * Opposite of vfree() - mark a vnode as in use.
2867  */
2868 void
2869 vbusy(struct vnode *vp)
2870 {
2871 
2872 	ASSERT_VI_LOCKED(vp, "vbusy");
2873 	KASSERT((vp->v_iflag & VI_FREE) != 0, ("vnode not free"));
2874 
2875 	mtx_lock(&vnode_free_list_mtx);
2876 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2877 	freevnodes--;
2878 	mtx_unlock(&vnode_free_list_mtx);
2879 
2880 	vp->v_iflag &= ~(VI_FREE|VI_AGE);
2881 }
2882 
2883 /*
2884  * Initalize per-vnode helper structure to hold poll-related state.
2885  */
2886 void
2887 v_addpollinfo(struct vnode *vp)
2888 {
2889 	struct vpollinfo *vi;
2890 
2891 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK);
2892 	if (vp->v_pollinfo != NULL) {
2893 		uma_zfree(vnodepoll_zone, vi);
2894 		return;
2895 	}
2896 	vp->v_pollinfo = vi;
2897 	mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
2898 	knlist_init(&vp->v_pollinfo->vpi_selinfo.si_note,
2899 	    &vp->v_pollinfo->vpi_lock);
2900 }
2901 
2902 /*
2903  * Record a process's interest in events which might happen to
2904  * a vnode.  Because poll uses the historic select-style interface
2905  * internally, this routine serves as both the ``check for any
2906  * pending events'' and the ``record my interest in future events''
2907  * functions.  (These are done together, while the lock is held,
2908  * to avoid race conditions.)
2909  */
2910 int
2911 vn_pollrecord(vp, td, events)
2912 	struct vnode *vp;
2913 	struct thread *td;
2914 	short events;
2915 {
2916 
2917 	if (vp->v_pollinfo == NULL)
2918 		v_addpollinfo(vp);
2919 	mtx_lock(&vp->v_pollinfo->vpi_lock);
2920 	if (vp->v_pollinfo->vpi_revents & events) {
2921 		/*
2922 		 * This leaves events we are not interested
2923 		 * in available for the other process which
2924 		 * which presumably had requested them
2925 		 * (otherwise they would never have been
2926 		 * recorded).
2927 		 */
2928 		events &= vp->v_pollinfo->vpi_revents;
2929 		vp->v_pollinfo->vpi_revents &= ~events;
2930 
2931 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
2932 		return events;
2933 	}
2934 	vp->v_pollinfo->vpi_events |= events;
2935 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
2936 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
2937 	return 0;
2938 }
2939 
2940 /*
2941  * Note the occurrence of an event.  If the VN_POLLEVENT macro is used,
2942  * it is possible for us to miss an event due to race conditions, but
2943  * that condition is expected to be rare, so for the moment it is the
2944  * preferred interface.
2945  */
2946 void
2947 vn_pollevent(vp, events)
2948 	struct vnode *vp;
2949 	short events;
2950 {
2951 
2952 	if (vp->v_pollinfo == NULL)
2953 		v_addpollinfo(vp);
2954 	mtx_lock(&vp->v_pollinfo->vpi_lock);
2955 	if (vp->v_pollinfo->vpi_events & events) {
2956 		/*
2957 		 * We clear vpi_events so that we don't
2958 		 * call selwakeup() twice if two events are
2959 		 * posted before the polling process(es) is
2960 		 * awakened.  This also ensures that we take at
2961 		 * most one selwakeup() if the polling process
2962 		 * is no longer interested.  However, it does
2963 		 * mean that only one event can be noticed at
2964 		 * a time.  (Perhaps we should only clear those
2965 		 * event bits which we note?) XXX
2966 		 */
2967 		vp->v_pollinfo->vpi_events = 0;	/* &= ~events ??? */
2968 		vp->v_pollinfo->vpi_revents |= events;
2969 		selwakeuppri(&vp->v_pollinfo->vpi_selinfo, PRIBIO);
2970 	}
2971 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
2972 }
2973 
2974 /*
2975  * Wake up anyone polling on vp because it is being revoked.
2976  * This depends on dead_poll() returning POLLHUP for correct
2977  * behavior.
2978  */
2979 void
2980 vn_pollgone(vp)
2981 	struct vnode *vp;
2982 {
2983 
2984 	mtx_lock(&vp->v_pollinfo->vpi_lock);
2985 	VN_KNOTE_LOCKED(vp, NOTE_REVOKE);
2986 	if (vp->v_pollinfo->vpi_events) {
2987 		vp->v_pollinfo->vpi_events = 0;
2988 		selwakeuppri(&vp->v_pollinfo->vpi_selinfo, PRIBIO);
2989 	}
2990 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
2991 }
2992 
2993 
2994 
2995 /*
2996  * Routine to create and manage a filesystem syncer vnode.
2997  */
2998 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
2999 static int	sync_fsync(struct  vop_fsync_args *);
3000 static int	sync_inactive(struct  vop_inactive_args *);
3001 static int	sync_reclaim(struct  vop_reclaim_args *);
3002 
3003 static struct vop_vector sync_vnodeops = {
3004 	.vop_bypass =	VOP_EOPNOTSUPP,
3005 	.vop_close =	sync_close,		/* close */
3006 	.vop_fsync =	sync_fsync,		/* fsync */
3007 	.vop_inactive =	sync_inactive,	/* inactive */
3008 	.vop_reclaim =	sync_reclaim,	/* reclaim */
3009 	.vop_lock =	vop_stdlock,	/* lock */
3010 	.vop_unlock =	vop_stdunlock,	/* unlock */
3011 	.vop_islocked =	vop_stdislocked,	/* islocked */
3012 };
3013 
3014 /*
3015  * Create a new filesystem syncer vnode for the specified mount point.
3016  */
3017 int
3018 vfs_allocate_syncvnode(mp)
3019 	struct mount *mp;
3020 {
3021 	struct vnode *vp;
3022 	static long start, incr, next;
3023 	int error;
3024 
3025 	/* Allocate a new vnode */
3026 	if ((error = getnewvnode("syncer", mp, &sync_vnodeops, &vp)) != 0) {
3027 		mp->mnt_syncer = NULL;
3028 		return (error);
3029 	}
3030 	vp->v_type = VNON;
3031 	/*
3032 	 * Place the vnode onto the syncer worklist. We attempt to
3033 	 * scatter them about on the list so that they will go off
3034 	 * at evenly distributed times even if all the filesystems
3035 	 * are mounted at once.
3036 	 */
3037 	next += incr;
3038 	if (next == 0 || next > syncer_maxdelay) {
3039 		start /= 2;
3040 		incr /= 2;
3041 		if (start == 0) {
3042 			start = syncer_maxdelay / 2;
3043 			incr = syncer_maxdelay;
3044 		}
3045 		next = start;
3046 	}
3047 	VI_LOCK(vp);
3048 	vn_syncer_add_to_worklist(&vp->v_bufobj,
3049 	    syncdelay > 0 ? next % syncdelay : 0);
3050 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
3051 	mtx_lock(&sync_mtx);
3052 	sync_vnode_count++;
3053 	mtx_unlock(&sync_mtx);
3054 	VI_UNLOCK(vp);
3055 	mp->mnt_syncer = vp;
3056 	return (0);
3057 }
3058 
3059 /*
3060  * Do a lazy sync of the filesystem.
3061  */
3062 static int
3063 sync_fsync(ap)
3064 	struct vop_fsync_args /* {
3065 		struct vnode *a_vp;
3066 		struct ucred *a_cred;
3067 		int a_waitfor;
3068 		struct thread *a_td;
3069 	} */ *ap;
3070 {
3071 	struct vnode *syncvp = ap->a_vp;
3072 	struct mount *mp = syncvp->v_mount;
3073 	struct thread *td = ap->a_td;
3074 	int error, asyncflag;
3075 	struct bufobj *bo;
3076 
3077 	/*
3078 	 * We only need to do something if this is a lazy evaluation.
3079 	 */
3080 	if (ap->a_waitfor != MNT_LAZY)
3081 		return (0);
3082 
3083 	/*
3084 	 * Move ourselves to the back of the sync list.
3085 	 */
3086 	bo = &syncvp->v_bufobj;
3087 	BO_LOCK(bo);
3088 	vn_syncer_add_to_worklist(bo, syncdelay);
3089 	BO_UNLOCK(bo);
3090 
3091 	/*
3092 	 * Walk the list of vnodes pushing all that are dirty and
3093 	 * not already on the sync list.
3094 	 */
3095 	mtx_lock(&mountlist_mtx);
3096 	if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) {
3097 		mtx_unlock(&mountlist_mtx);
3098 		return (0);
3099 	}
3100 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
3101 		vfs_unbusy(mp, td);
3102 		return (0);
3103 	}
3104 	asyncflag = mp->mnt_flag & MNT_ASYNC;
3105 	mp->mnt_flag &= ~MNT_ASYNC;
3106 	vfs_msync(mp, MNT_NOWAIT);
3107 	error = VFS_SYNC(mp, MNT_LAZY, td);
3108 	if (asyncflag)
3109 		mp->mnt_flag |= MNT_ASYNC;
3110 	vn_finished_write(mp);
3111 	vfs_unbusy(mp, td);
3112 	return (error);
3113 }
3114 
3115 /*
3116  * The syncer vnode is no referenced.
3117  */
3118 static int
3119 sync_inactive(ap)
3120 	struct vop_inactive_args /* {
3121 		struct vnode *a_vp;
3122 		struct thread *a_td;
3123 	} */ *ap;
3124 {
3125 
3126 	VOP_UNLOCK(ap->a_vp, 0, ap->a_td);
3127 	vgone(ap->a_vp);
3128 	return (0);
3129 }
3130 
3131 /*
3132  * The syncer vnode is no longer needed and is being decommissioned.
3133  *
3134  * Modifications to the worklist must be protected by sync_mtx.
3135  */
3136 static int
3137 sync_reclaim(ap)
3138 	struct vop_reclaim_args /* {
3139 		struct vnode *a_vp;
3140 	} */ *ap;
3141 {
3142 	struct vnode *vp = ap->a_vp;
3143 	struct bufobj *bo;
3144 
3145 	VI_LOCK(vp);
3146 	bo = &vp->v_bufobj;
3147 	vp->v_mount->mnt_syncer = NULL;
3148 	if (bo->bo_flag & BO_ONWORKLST) {
3149 		mtx_lock(&sync_mtx);
3150 		LIST_REMOVE(bo, bo_synclist);
3151  		syncer_worklist_len--;
3152 		sync_vnode_count--;
3153 		mtx_unlock(&sync_mtx);
3154 		bo->bo_flag &= ~BO_ONWORKLST;
3155 	}
3156 	VI_UNLOCK(vp);
3157 
3158 	return (0);
3159 }
3160 
3161 /*
3162  * Check if vnode represents a disk device
3163  */
3164 int
3165 vn_isdisk(vp, errp)
3166 	struct vnode *vp;
3167 	int *errp;
3168 {
3169 	int error;
3170 
3171 	error = 0;
3172 	dev_lock();
3173 	if (vp->v_type != VCHR)
3174 		error = ENOTBLK;
3175 	else if (vp->v_rdev == NULL)
3176 		error = ENXIO;
3177 	else if (vp->v_rdev->si_devsw == NULL)
3178 		error = ENXIO;
3179 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
3180 		error = ENOTBLK;
3181 	dev_unlock();
3182 	if (errp != NULL)
3183 		*errp = error;
3184 	return (error == 0);
3185 }
3186 
3187 /*
3188  * Free data allocated by namei(); see namei(9) for details.
3189  */
3190 void
3191 NDFREE(ndp, flags)
3192      struct nameidata *ndp;
3193      const u_int flags;
3194 {
3195 
3196 	if (!(flags & NDF_NO_FREE_PNBUF) &&
3197 	    (ndp->ni_cnd.cn_flags & HASBUF)) {
3198 		uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf);
3199 		ndp->ni_cnd.cn_flags &= ~HASBUF;
3200 	}
3201 	if (!(flags & NDF_NO_DVP_UNLOCK) &&
3202 	    (ndp->ni_cnd.cn_flags & LOCKPARENT) &&
3203 	    ndp->ni_dvp != ndp->ni_vp)
3204 		VOP_UNLOCK(ndp->ni_dvp, 0, ndp->ni_cnd.cn_thread);
3205 	if (!(flags & NDF_NO_DVP_RELE) &&
3206 	    (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) {
3207 		vrele(ndp->ni_dvp);
3208 		ndp->ni_dvp = NULL;
3209 	}
3210 	if (!(flags & NDF_NO_VP_UNLOCK) &&
3211 	    (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp)
3212 		VOP_UNLOCK(ndp->ni_vp, 0, ndp->ni_cnd.cn_thread);
3213 	if (!(flags & NDF_NO_VP_RELE) &&
3214 	    ndp->ni_vp) {
3215 		vrele(ndp->ni_vp);
3216 		ndp->ni_vp = NULL;
3217 	}
3218 	if (!(flags & NDF_NO_STARTDIR_RELE) &&
3219 	    (ndp->ni_cnd.cn_flags & SAVESTART)) {
3220 		vrele(ndp->ni_startdir);
3221 		ndp->ni_startdir = NULL;
3222 	}
3223 }
3224 
3225 /*
3226  * Common filesystem object access control check routine.  Accepts a
3227  * vnode's type, "mode", uid and gid, requested access mode, credentials,
3228  * and optional call-by-reference privused argument allowing vaccess()
3229  * to indicate to the caller whether privilege was used to satisfy the
3230  * request (obsoleted).  Returns 0 on success, or an errno on failure.
3231  */
3232 int
3233 vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused)
3234 	enum vtype type;
3235 	mode_t file_mode;
3236 	uid_t file_uid;
3237 	gid_t file_gid;
3238 	mode_t acc_mode;
3239 	struct ucred *cred;
3240 	int *privused;
3241 {
3242 	mode_t dac_granted;
3243 #ifdef CAPABILITIES
3244 	mode_t cap_granted;
3245 #endif
3246 
3247 	/*
3248 	 * Look for a normal, non-privileged way to access the file/directory
3249 	 * as requested.  If it exists, go with that.
3250 	 */
3251 
3252 	if (privused != NULL)
3253 		*privused = 0;
3254 
3255 	dac_granted = 0;
3256 
3257 	/* Check the owner. */
3258 	if (cred->cr_uid == file_uid) {
3259 		dac_granted |= VADMIN;
3260 		if (file_mode & S_IXUSR)
3261 			dac_granted |= VEXEC;
3262 		if (file_mode & S_IRUSR)
3263 			dac_granted |= VREAD;
3264 		if (file_mode & S_IWUSR)
3265 			dac_granted |= (VWRITE | VAPPEND);
3266 
3267 		if ((acc_mode & dac_granted) == acc_mode)
3268 			return (0);
3269 
3270 		goto privcheck;
3271 	}
3272 
3273 	/* Otherwise, check the groups (first match) */
3274 	if (groupmember(file_gid, cred)) {
3275 		if (file_mode & S_IXGRP)
3276 			dac_granted |= VEXEC;
3277 		if (file_mode & S_IRGRP)
3278 			dac_granted |= VREAD;
3279 		if (file_mode & S_IWGRP)
3280 			dac_granted |= (VWRITE | VAPPEND);
3281 
3282 		if ((acc_mode & dac_granted) == acc_mode)
3283 			return (0);
3284 
3285 		goto privcheck;
3286 	}
3287 
3288 	/* Otherwise, check everyone else. */
3289 	if (file_mode & S_IXOTH)
3290 		dac_granted |= VEXEC;
3291 	if (file_mode & S_IROTH)
3292 		dac_granted |= VREAD;
3293 	if (file_mode & S_IWOTH)
3294 		dac_granted |= (VWRITE | VAPPEND);
3295 	if ((acc_mode & dac_granted) == acc_mode)
3296 		return (0);
3297 
3298 privcheck:
3299 	if (!suser_cred(cred, SUSER_ALLOWJAIL)) {
3300 		/* XXX audit: privilege used */
3301 		if (privused != NULL)
3302 			*privused = 1;
3303 		return (0);
3304 	}
3305 
3306 #ifdef CAPABILITIES
3307 	/*
3308 	 * Build a capability mask to determine if the set of capabilities
3309 	 * satisfies the requirements when combined with the granted mask
3310 	 * from above.
3311 	 * For each capability, if the capability is required, bitwise
3312 	 * or the request type onto the cap_granted mask.
3313 	 */
3314 	cap_granted = 0;
3315 
3316 	if (type == VDIR) {
3317 		/*
3318 		 * For directories, use CAP_DAC_READ_SEARCH to satisfy
3319 		 * VEXEC requests, instead of CAP_DAC_EXECUTE.
3320 		 */
3321 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3322 		    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL))
3323 			cap_granted |= VEXEC;
3324 	} else {
3325 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3326 		    !cap_check(cred, NULL, CAP_DAC_EXECUTE, SUSER_ALLOWJAIL))
3327 			cap_granted |= VEXEC;
3328 	}
3329 
3330 	if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) &&
3331 	    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL))
3332 		cap_granted |= VREAD;
3333 
3334 	if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
3335 	    !cap_check(cred, NULL, CAP_DAC_WRITE, SUSER_ALLOWJAIL))
3336 		cap_granted |= (VWRITE | VAPPEND);
3337 
3338 	if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
3339 	    !cap_check(cred, NULL, CAP_FOWNER, SUSER_ALLOWJAIL))
3340 		cap_granted |= VADMIN;
3341 
3342 	if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) {
3343 		/* XXX audit: privilege used */
3344 		if (privused != NULL)
3345 			*privused = 1;
3346 		return (0);
3347 	}
3348 #endif
3349 
3350 	return ((acc_mode & VADMIN) ? EPERM : EACCES);
3351 }
3352 
3353 /*
3354  * Credential check based on process requesting service, and per-attribute
3355  * permissions.
3356  */
3357 int
3358 extattr_check_cred(struct vnode *vp, int attrnamespace,
3359     struct ucred *cred, struct thread *td, int access)
3360 {
3361 
3362 	/*
3363 	 * Kernel-invoked always succeeds.
3364 	 */
3365 	if (cred == NOCRED)
3366 		return (0);
3367 
3368 	/*
3369 	 * Do not allow privileged processes in jail to directly
3370 	 * manipulate system attributes.
3371 	 *
3372 	 * XXX What capability should apply here?
3373 	 * Probably CAP_SYS_SETFFLAG.
3374 	 */
3375 	switch (attrnamespace) {
3376 	case EXTATTR_NAMESPACE_SYSTEM:
3377 		/* Potentially should be: return (EPERM); */
3378 		return (suser_cred(cred, 0));
3379 	case EXTATTR_NAMESPACE_USER:
3380 		return (VOP_ACCESS(vp, access, cred, td));
3381 	default:
3382 		return (EPERM);
3383 	}
3384 }
3385 
3386 #ifdef DEBUG_VFS_LOCKS
3387 /*
3388  * This only exists to supress warnings from unlocked specfs accesses.  It is
3389  * no longer ok to have an unlocked VFS.
3390  */
3391 #define	IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD)
3392 
3393 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
3394 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, "");
3395 
3396 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
3397 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, "");
3398 
3399 int vfs_badlock_print = 1;	/* Print lock violations. */
3400 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, "");
3401 
3402 #ifdef KDB
3403 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
3404 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, "");
3405 #endif
3406 
3407 static void
3408 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
3409 {
3410 
3411 #ifdef KDB
3412 	if (vfs_badlock_backtrace)
3413 		kdb_backtrace();
3414 #endif
3415 	if (vfs_badlock_print)
3416 		printf("%s: %p %s\n", str, (void *)vp, msg);
3417 	if (vfs_badlock_ddb)
3418 		kdb_enter("lock violation");
3419 }
3420 
3421 void
3422 assert_vi_locked(struct vnode *vp, const char *str)
3423 {
3424 
3425 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
3426 		vfs_badlock("interlock is not locked but should be", str, vp);
3427 }
3428 
3429 void
3430 assert_vi_unlocked(struct vnode *vp, const char *str)
3431 {
3432 
3433 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
3434 		vfs_badlock("interlock is locked but should not be", str, vp);
3435 }
3436 
3437 void
3438 assert_vop_locked(struct vnode *vp, const char *str)
3439 {
3440 
3441 	if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0)
3442 		vfs_badlock("is not locked but should be", str, vp);
3443 }
3444 
3445 void
3446 assert_vop_unlocked(struct vnode *vp, const char *str)
3447 {
3448 
3449 	if (vp && !IGNORE_LOCK(vp) &&
3450 	    VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE)
3451 		vfs_badlock("is locked but should not be", str, vp);
3452 }
3453 
3454 #if 0
3455 void
3456 assert_vop_elocked(struct vnode *vp, const char *str)
3457 {
3458 
3459 	if (vp && !IGNORE_LOCK(vp) &&
3460 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE)
3461 		vfs_badlock("is not exclusive locked but should be", str, vp);
3462 }
3463 
3464 void
3465 assert_vop_elocked_other(struct vnode *vp, const char *str)
3466 {
3467 
3468 	if (vp && !IGNORE_LOCK(vp) &&
3469 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER)
3470 		vfs_badlock("is not exclusive locked by another thread",
3471 		    str, vp);
3472 }
3473 
3474 void
3475 assert_vop_slocked(struct vnode *vp, const char *str)
3476 {
3477 
3478 	if (vp && !IGNORE_LOCK(vp) &&
3479 	    VOP_ISLOCKED(vp, curthread) != LK_SHARED)
3480 		vfs_badlock("is not locked shared but should be", str, vp);
3481 }
3482 #endif /* 0 */
3483 
3484 void
3485 vop_rename_pre(void *ap)
3486 {
3487 	struct vop_rename_args *a = ap;
3488 
3489 	if (a->a_tvp)
3490 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
3491 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
3492 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
3493 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
3494 
3495 	/* Check the source (from). */
3496 	if (a->a_tdvp != a->a_fdvp)
3497 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
3498 	if (a->a_tvp != a->a_fvp)
3499 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked");
3500 
3501 	/* Check the target. */
3502 	if (a->a_tvp)
3503 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
3504 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
3505 }
3506 
3507 void
3508 vop_strategy_pre(void *ap)
3509 {
3510 	struct vop_strategy_args *a;
3511 	struct buf *bp;
3512 
3513 	a = ap;
3514 	bp = a->a_bp;
3515 
3516 	/*
3517 	 * Cluster ops lock their component buffers but not the IO container.
3518 	 */
3519 	if ((bp->b_flags & B_CLUSTER) != 0)
3520 		return;
3521 
3522 	if (BUF_REFCNT(bp) < 1) {
3523 		if (vfs_badlock_print)
3524 			printf(
3525 			    "VOP_STRATEGY: bp is not locked but should be\n");
3526 		if (vfs_badlock_ddb)
3527 			kdb_enter("lock violation");
3528 	}
3529 }
3530 
3531 void
3532 vop_lookup_pre(void *ap)
3533 {
3534 	struct vop_lookup_args *a;
3535 	struct vnode *dvp;
3536 
3537 	a = ap;
3538 	dvp = a->a_dvp;
3539 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3540 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3541 }
3542 
3543 void
3544 vop_lookup_post(void *ap, int rc)
3545 {
3546 	struct vop_lookup_args *a;
3547 	struct componentname *cnp;
3548 	struct vnode *dvp;
3549 	struct vnode *vp;
3550 	int flags;
3551 
3552 	a = ap;
3553 	dvp = a->a_dvp;
3554 	cnp = a->a_cnp;
3555 	vp = *(a->a_vpp);
3556 	flags = cnp->cn_flags;
3557 
3558 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3559 
3560 	/*
3561 	 * If this is the last path component for this lookup and LOCKPARENT
3562 	 * is set, OR if there is an error the directory has to be locked.
3563 	 */
3564 	if ((flags & LOCKPARENT) && (flags & ISLASTCN))
3565 		ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (LOCKPARENT)");
3566 	else if (rc != 0)
3567 		ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (error)");
3568 	else if (dvp != vp)
3569 		ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (dvp)");
3570 	if (flags & PDIRUNLOCK)
3571 		ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (PDIRUNLOCK)");
3572 }
3573 
3574 void
3575 vop_lock_pre(void *ap)
3576 {
3577 	struct vop_lock_args *a = ap;
3578 
3579 	if ((a->a_flags & LK_INTERLOCK) == 0)
3580 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3581 	else
3582 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
3583 }
3584 
3585 void
3586 vop_lock_post(void *ap, int rc)
3587 {
3588 	struct vop_lock_args *a = ap;
3589 
3590 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3591 	if (rc == 0)
3592 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
3593 }
3594 
3595 void
3596 vop_unlock_pre(void *ap)
3597 {
3598 	struct vop_unlock_args *a = ap;
3599 
3600 	if (a->a_flags & LK_INTERLOCK)
3601 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
3602 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
3603 }
3604 
3605 void
3606 vop_unlock_post(void *ap, int rc)
3607 {
3608 	struct vop_unlock_args *a = ap;
3609 
3610 	if (a->a_flags & LK_INTERLOCK)
3611 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
3612 }
3613 #endif /* DEBUG_VFS_LOCKS */
3614 
3615 static struct knlist fs_knlist;
3616 
3617 static void
3618 vfs_event_init(void *arg)
3619 {
3620 	knlist_init(&fs_knlist, NULL);
3621 }
3622 /* XXX - correct order? */
3623 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
3624 
3625 void
3626 vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused)
3627 {
3628 
3629 	KNOTE_UNLOCKED(&fs_knlist, event);
3630 }
3631 
3632 static int	filt_fsattach(struct knote *kn);
3633 static void	filt_fsdetach(struct knote *kn);
3634 static int	filt_fsevent(struct knote *kn, long hint);
3635 
3636 struct filterops fs_filtops =
3637 	{ 0, filt_fsattach, filt_fsdetach, filt_fsevent };
3638 
3639 static int
3640 filt_fsattach(struct knote *kn)
3641 {
3642 
3643 	kn->kn_flags |= EV_CLEAR;
3644 	knlist_add(&fs_knlist, kn, 0);
3645 	return (0);
3646 }
3647 
3648 static void
3649 filt_fsdetach(struct knote *kn)
3650 {
3651 
3652 	knlist_remove(&fs_knlist, kn, 0);
3653 }
3654 
3655 static int
3656 filt_fsevent(struct knote *kn, long hint)
3657 {
3658 
3659 	kn->kn_fflags |= hint;
3660 	return (kn->kn_fflags != 0);
3661 }
3662 
3663 static int
3664 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
3665 {
3666 	struct vfsidctl vc;
3667 	int error;
3668 	struct mount *mp;
3669 
3670 	error = SYSCTL_IN(req, &vc, sizeof(vc));
3671 	if (error)
3672 		return (error);
3673 	if (vc.vc_vers != VFS_CTL_VERS1)
3674 		return (EINVAL);
3675 	mp = vfs_getvfs(&vc.vc_fsid);
3676 	if (mp == NULL)
3677 		return (ENOENT);
3678 	/* ensure that a specific sysctl goes to the right filesystem. */
3679 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
3680 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
3681 		return (EINVAL);
3682 	}
3683 	VCTLTOREQ(&vc, req);
3684 	return (VFS_SYSCTL(mp, vc.vc_op, req));
3685 }
3686 
3687 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR,
3688         NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid");
3689 
3690 /*
3691  * Function to initialize a va_filerev field sensibly.
3692  * XXX: Wouldn't a random number make a lot more sense ??
3693  */
3694 u_quad_t
3695 init_va_filerev(void)
3696 {
3697 	struct bintime bt;
3698 
3699 	getbinuptime(&bt);
3700 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
3701 }
3702