xref: /freebsd/sys/kern/vfs_subr.c (revision 6af83ee0d2941d18880b6aaa2b4facd1d30c6106)
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 <machine/stdarg.h>
72 
73 #include <vm/vm.h>
74 #include <vm/vm_object.h>
75 #include <vm/vm_extern.h>
76 #include <vm/pmap.h>
77 #include <vm/vm_map.h>
78 #include <vm/vm_page.h>
79 #include <vm/vm_kern.h>
80 #include <vm/uma.h>
81 
82 static MALLOC_DEFINE(M_NETADDR, "Export Host", "Export host address structure");
83 
84 static void	delmntque(struct vnode *vp);
85 static void	insmntque(struct vnode *vp, struct mount *mp);
86 static void	vlruvp(struct vnode *vp);
87 static int	flushbuflist(struct bufv *bufv, int flags, struct vnode *vp,
88 		    int slpflag, int slptimeo);
89 static void	syncer_shutdown(void *arg, int howto);
90 static int	vtryrecycle(struct vnode *vp);
91 static void	vx_lock(struct vnode *vp);
92 static void	vx_unlock(struct vnode *vp);
93 static void	vbusy(struct vnode *vp);
94 static void	vdropl(struct vnode *vp);
95 static void	vholdl(struct vnode *);
96 
97 /*
98  * Enable Giant pushdown based on whether or not the vm is mpsafe in this
99  * build.  Without mpsafevm the buffer cache can not run Giant free.
100  */
101 int mpsafe_vfs = 0;
102 TUNABLE_INT("debug.mpsafevfs", &mpsafe_vfs);
103 SYSCTL_INT(_debug, OID_AUTO, mpsafevfs, CTLFLAG_RD, &mpsafe_vfs, 0,
104     "MPSAFE VFS");
105 
106 /*
107  * Number of vnodes in existence.  Increased whenever getnewvnode()
108  * allocates a new vnode, never decreased.
109  */
110 static unsigned long	numvnodes;
111 
112 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, "");
113 
114 /*
115  * Conversion tables for conversion from vnode types to inode formats
116  * and back.
117  */
118 enum vtype iftovt_tab[16] = {
119 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
120 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
121 };
122 int vttoif_tab[9] = {
123 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
124 	S_IFSOCK, S_IFIFO, S_IFMT,
125 };
126 
127 /*
128  * List of vnodes that are ready for recycling.
129  */
130 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
131 
132 /*
133  * Minimum number of free vnodes.  If there are fewer than this free vnodes,
134  * getnewvnode() will return a newly allocated vnode.
135  */
136 static u_long wantfreevnodes = 25;
137 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, "");
138 /* Number of vnodes in the free list. */
139 static u_long freevnodes;
140 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, "");
141 
142 /*
143  * Various variables used for debugging the new implementation of
144  * reassignbuf().
145  * XXX these are probably of (very) limited utility now.
146  */
147 static int reassignbufcalls;
148 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, "");
149 static int nameileafonly;
150 SYSCTL_INT(_vfs, OID_AUTO, nameileafonly, CTLFLAG_RW, &nameileafonly, 0, "");
151 
152 /*
153  * Cache for the mount type id assigned to NFS.  This is used for
154  * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c.
155  */
156 int	nfs_mount_type = -1;
157 
158 /* To keep more than one thread at a time from running vfs_getnewfsid */
159 static struct mtx mntid_mtx;
160 
161 /*
162  * Lock for any access to the following:
163  *	vnode_free_list
164  *	numvnodes
165  *	freevnodes
166  */
167 static struct mtx vnode_free_list_mtx;
168 
169 /* Publicly exported FS */
170 struct nfs_public nfs_pub;
171 
172 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
173 static uma_zone_t vnode_zone;
174 static uma_zone_t vnodepoll_zone;
175 
176 /* Set to 1 to print out reclaim of active vnodes */
177 int	prtactive;
178 
179 /*
180  * The workitem queue.
181  *
182  * It is useful to delay writes of file data and filesystem metadata
183  * for tens of seconds so that quickly created and deleted files need
184  * not waste disk bandwidth being created and removed. To realize this,
185  * we append vnodes to a "workitem" queue. When running with a soft
186  * updates implementation, most pending metadata dependencies should
187  * not wait for more than a few seconds. Thus, mounted on block devices
188  * are delayed only about a half the time that file data is delayed.
189  * Similarly, directory updates are more critical, so are only delayed
190  * about a third the time that file data is delayed. Thus, there are
191  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
192  * one each second (driven off the filesystem syncer process). The
193  * syncer_delayno variable indicates the next queue that is to be processed.
194  * Items that need to be processed soon are placed in this queue:
195  *
196  *	syncer_workitem_pending[syncer_delayno]
197  *
198  * A delay of fifteen seconds is done by placing the request fifteen
199  * entries later in the queue:
200  *
201  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
202  *
203  */
204 static int syncer_delayno;
205 static long syncer_mask;
206 LIST_HEAD(synclist, bufobj);
207 static struct synclist *syncer_workitem_pending;
208 /*
209  * The sync_mtx protects:
210  *	bo->bo_synclist
211  *	sync_vnode_count
212  *	syncer_delayno
213  *	syncer_state
214  *	syncer_workitem_pending
215  *	syncer_worklist_len
216  *	rushjob
217  */
218 static struct mtx sync_mtx;
219 
220 #define SYNCER_MAXDELAY		32
221 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
222 static int syncdelay = 30;		/* max time to delay syncing data */
223 static int filedelay = 30;		/* time to delay syncing files */
224 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, "");
225 static int dirdelay = 29;		/* time to delay syncing directories */
226 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, "");
227 static int metadelay = 28;		/* time to delay syncing metadata */
228 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, "");
229 static int rushjob;		/* number of slots to run ASAP */
230 static int stat_rush_requests;	/* number of times I/O speeded up */
231 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, "");
232 
233 /*
234  * When shutting down the syncer, run it at four times normal speed.
235  */
236 #define SYNCER_SHUTDOWN_SPEEDUP		4
237 static int sync_vnode_count;
238 static int syncer_worklist_len;
239 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
240     syncer_state;
241 
242 /*
243  * Number of vnodes we want to exist at any one time.  This is mostly used
244  * to size hash tables in vnode-related code.  It is normally not used in
245  * getnewvnode(), as wantfreevnodes is normally nonzero.)
246  *
247  * XXX desiredvnodes is historical cruft and should not exist.
248  */
249 int desiredvnodes;
250 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW,
251     &desiredvnodes, 0, "Maximum number of vnodes");
252 static int minvnodes;
253 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
254     &minvnodes, 0, "Minimum number of vnodes");
255 static int vnlru_nowhere;
256 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
257     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
258 
259 /* Hook for calling soft updates. */
260 int (*softdep_process_worklist_hook)(struct mount *);
261 
262 /*
263  * Initialize the vnode management data structures.
264  */
265 #ifndef	MAXVNODES_MAX
266 #define	MAXVNODES_MAX	100000
267 #endif
268 static void
269 vntblinit(void *dummy __unused)
270 {
271 
272 	/*
273 	 * Desiredvnodes is a function of the physical memory size and
274 	 * the kernel's heap size.  Specifically, desiredvnodes scales
275 	 * in proportion to the physical memory size until two fifths
276 	 * of the kernel's heap size is consumed by vnodes and vm
277 	 * objects.
278 	 */
279 	desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size /
280 	    (5 * (sizeof(struct vm_object) + sizeof(struct vnode))));
281 	if (desiredvnodes > MAXVNODES_MAX) {
282 		if (bootverbose)
283 			printf("Reducing kern.maxvnodes %d -> %d\n",
284 			    desiredvnodes, MAXVNODES_MAX);
285 		desiredvnodes = MAXVNODES_MAX;
286 	}
287 	minvnodes = desiredvnodes / 4;
288 	mtx_init(&mountlist_mtx, "mountlist", NULL, MTX_DEF);
289 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
290 	TAILQ_INIT(&vnode_free_list);
291 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
292 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
293 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
294 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
295 	      NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
296 	/*
297 	 * Initialize the filesystem syncer.
298 	 */
299 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
300 		&syncer_mask);
301 	syncer_maxdelay = syncer_mask + 1;
302 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
303 }
304 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL)
305 
306 
307 /*
308  * Mark a mount point as busy. Used to synchronize access and to delay
309  * unmounting. Interlock is not released on failure.
310  */
311 int
312 vfs_busy(mp, flags, interlkp, td)
313 	struct mount *mp;
314 	int flags;
315 	struct mtx *interlkp;
316 	struct thread *td;
317 {
318 	int lkflags;
319 
320 	MNT_ILOCK(mp);
321 	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
322 		if (flags & LK_NOWAIT) {
323 			MNT_IUNLOCK(mp);
324 			return (ENOENT);
325 		}
326 		if (interlkp)
327 			mtx_unlock(interlkp);
328 		mp->mnt_kern_flag |= MNTK_MWAIT;
329 		/*
330 		 * Since all busy locks are shared except the exclusive
331 		 * lock granted when unmounting, the only place that a
332 		 * wakeup needs to be done is at the release of the
333 		 * exclusive lock at the end of dounmount.
334 		 */
335 		msleep(mp, MNT_MTX(mp), PVFS|PDROP, "vfs_busy", 0);
336 		if (interlkp)
337 			mtx_lock(interlkp);
338 		return (ENOENT);
339 	}
340 	if (interlkp)
341 		mtx_unlock(interlkp);
342 	lkflags = LK_SHARED | LK_NOPAUSE | LK_INTERLOCK;
343 	if (lockmgr(&mp->mnt_lock, lkflags, MNT_MTX(mp), td))
344 		panic("vfs_busy: unexpected lock failure");
345 	return (0);
346 }
347 
348 /*
349  * Free a busy filesystem.
350  */
351 void
352 vfs_unbusy(mp, td)
353 	struct mount *mp;
354 	struct thread *td;
355 {
356 
357 	lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td);
358 }
359 
360 /*
361  * Lookup a mount point by filesystem identifier.
362  */
363 struct mount *
364 vfs_getvfs(fsid)
365 	fsid_t *fsid;
366 {
367 	struct mount *mp;
368 
369 	mtx_lock(&mountlist_mtx);
370 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
371 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
372 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
373 			mtx_unlock(&mountlist_mtx);
374 			return (mp);
375 		}
376 	}
377 	mtx_unlock(&mountlist_mtx);
378 	return ((struct mount *) 0);
379 }
380 
381 /*
382  * Check if a user can access priveledged mount options.
383  */
384 int
385 vfs_suser(struct mount *mp, struct thread *td)
386 {
387 	int error;
388 
389 	if ((mp->mnt_flag & MNT_USER) == 0 ||
390 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
391 		if ((error = suser(td)) != 0)
392 			return (error);
393 	}
394 	return (0);
395 }
396 
397 /*
398  * Get a new unique fsid.  Try to make its val[0] unique, since this value
399  * will be used to create fake device numbers for stat().  Also try (but
400  * not so hard) make its val[0] unique mod 2^16, since some emulators only
401  * support 16-bit device numbers.  We end up with unique val[0]'s for the
402  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
403  *
404  * Keep in mind that several mounts may be running in parallel.  Starting
405  * the search one past where the previous search terminated is both a
406  * micro-optimization and a defense against returning the same fsid to
407  * different mounts.
408  */
409 void
410 vfs_getnewfsid(mp)
411 	struct mount *mp;
412 {
413 	static u_int16_t mntid_base;
414 	fsid_t tfsid;
415 	int mtype;
416 
417 	mtx_lock(&mntid_mtx);
418 	mtype = mp->mnt_vfc->vfc_typenum;
419 	tfsid.val[1] = mtype;
420 	mtype = (mtype & 0xFF) << 24;
421 	for (;;) {
422 		tfsid.val[0] = makedev(255,
423 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
424 		mntid_base++;
425 		if (vfs_getvfs(&tfsid) == NULL)
426 			break;
427 	}
428 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
429 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
430 	mtx_unlock(&mntid_mtx);
431 }
432 
433 /*
434  * Knob to control the precision of file timestamps:
435  *
436  *   0 = seconds only; nanoseconds zeroed.
437  *   1 = seconds and nanoseconds, accurate within 1/HZ.
438  *   2 = seconds and nanoseconds, truncated to microseconds.
439  * >=3 = seconds and nanoseconds, maximum precision.
440  */
441 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
442 
443 static int timestamp_precision = TSP_SEC;
444 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
445     &timestamp_precision, 0, "");
446 
447 /*
448  * Get a current timestamp.
449  */
450 void
451 vfs_timestamp(tsp)
452 	struct timespec *tsp;
453 {
454 	struct timeval tv;
455 
456 	switch (timestamp_precision) {
457 	case TSP_SEC:
458 		tsp->tv_sec = time_second;
459 		tsp->tv_nsec = 0;
460 		break;
461 	case TSP_HZ:
462 		getnanotime(tsp);
463 		break;
464 	case TSP_USEC:
465 		microtime(&tv);
466 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
467 		break;
468 	case TSP_NSEC:
469 	default:
470 		nanotime(tsp);
471 		break;
472 	}
473 }
474 
475 /*
476  * Set vnode attributes to VNOVAL
477  */
478 void
479 vattr_null(vap)
480 	struct vattr *vap;
481 {
482 
483 	vap->va_type = VNON;
484 	vap->va_size = VNOVAL;
485 	vap->va_bytes = VNOVAL;
486 	vap->va_mode = VNOVAL;
487 	vap->va_nlink = VNOVAL;
488 	vap->va_uid = VNOVAL;
489 	vap->va_gid = VNOVAL;
490 	vap->va_fsid = VNOVAL;
491 	vap->va_fileid = VNOVAL;
492 	vap->va_blocksize = VNOVAL;
493 	vap->va_rdev = VNOVAL;
494 	vap->va_atime.tv_sec = VNOVAL;
495 	vap->va_atime.tv_nsec = VNOVAL;
496 	vap->va_mtime.tv_sec = VNOVAL;
497 	vap->va_mtime.tv_nsec = VNOVAL;
498 	vap->va_ctime.tv_sec = VNOVAL;
499 	vap->va_ctime.tv_nsec = VNOVAL;
500 	vap->va_birthtime.tv_sec = VNOVAL;
501 	vap->va_birthtime.tv_nsec = VNOVAL;
502 	vap->va_flags = VNOVAL;
503 	vap->va_gen = VNOVAL;
504 	vap->va_vaflags = 0;
505 }
506 
507 /*
508  * This routine is called when we have too many vnodes.  It attempts
509  * to free <count> vnodes and will potentially free vnodes that still
510  * have VM backing store (VM backing store is typically the cause
511  * of a vnode blowout so we want to do this).  Therefore, this operation
512  * is not considered cheap.
513  *
514  * A number of conditions may prevent a vnode from being reclaimed.
515  * the buffer cache may have references on the vnode, a directory
516  * vnode may still have references due to the namei cache representing
517  * underlying files, or the vnode may be in active use.   It is not
518  * desireable to reuse such vnodes.  These conditions may cause the
519  * number of vnodes to reach some minimum value regardless of what
520  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
521  */
522 static int
523 vlrureclaim(struct mount *mp)
524 {
525 	struct vnode *vp;
526 	int done;
527 	int trigger;
528 	int usevnodes;
529 	int count;
530 
531 	/*
532 	 * Calculate the trigger point, don't allow user
533 	 * screwups to blow us up.   This prevents us from
534 	 * recycling vnodes with lots of resident pages.  We
535 	 * aren't trying to free memory, we are trying to
536 	 * free vnodes.
537 	 */
538 	usevnodes = desiredvnodes;
539 	if (usevnodes <= 0)
540 		usevnodes = 1;
541 	trigger = cnt.v_page_count * 2 / usevnodes;
542 
543 	done = 0;
544 	MNT_ILOCK(mp);
545 	count = mp->mnt_nvnodelistsize / 10 + 1;
546 	while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) {
547 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
548 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
549 
550 		if (vp->v_type != VNON &&
551 		    vp->v_type != VBAD &&
552 		    VI_TRYLOCK(vp)) {
553 			if (VMIGHTFREE(vp) &&           /* critical path opt */
554 			    (vp->v_object == NULL ||
555 			    vp->v_object->resident_page_count < trigger)) {
556 				MNT_IUNLOCK(mp);
557 				vgonel(vp, curthread);
558 				done++;
559 				MNT_ILOCK(mp);
560 			} else
561 				VI_UNLOCK(vp);
562 		}
563 		--count;
564 	}
565 	MNT_IUNLOCK(mp);
566 	return done;
567 }
568 
569 /*
570  * Attempt to recycle vnodes in a context that is always safe to block.
571  * Calling vlrurecycle() from the bowels of filesystem code has some
572  * interesting deadlock problems.
573  */
574 static struct proc *vnlruproc;
575 static int vnlruproc_sig;
576 
577 static void
578 vnlru_proc(void)
579 {
580 	struct mount *mp, *nmp;
581 	int done;
582 	struct proc *p = vnlruproc;
583 	struct thread *td = FIRST_THREAD_IN_PROC(p);
584 
585 	mtx_lock(&Giant);
586 
587 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p,
588 	    SHUTDOWN_PRI_FIRST);
589 
590 	for (;;) {
591 		kthread_suspend_check(p);
592 		mtx_lock(&vnode_free_list_mtx);
593 		if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) {
594 			vnlruproc_sig = 0;
595 			wakeup(&vnlruproc_sig);
596 			msleep(vnlruproc, &vnode_free_list_mtx,
597 			    PVFS|PDROP, "vlruwt", hz);
598 			continue;
599 		}
600 		mtx_unlock(&vnode_free_list_mtx);
601 		done = 0;
602 		mtx_lock(&mountlist_mtx);
603 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
604 			if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) {
605 				nmp = TAILQ_NEXT(mp, mnt_list);
606 				continue;
607 			}
608 			done += vlrureclaim(mp);
609 			mtx_lock(&mountlist_mtx);
610 			nmp = TAILQ_NEXT(mp, mnt_list);
611 			vfs_unbusy(mp, td);
612 		}
613 		mtx_unlock(&mountlist_mtx);
614 		if (done == 0) {
615 #if 0
616 			/* These messages are temporary debugging aids */
617 			if (vnlru_nowhere < 5)
618 				printf("vnlru process getting nowhere..\n");
619 			else if (vnlru_nowhere == 5)
620 				printf("vnlru process messages stopped.\n");
621 #endif
622 			vnlru_nowhere++;
623 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
624 		}
625 	}
626 }
627 
628 static struct kproc_desc vnlru_kp = {
629 	"vnlru",
630 	vnlru_proc,
631 	&vnlruproc
632 };
633 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
634 
635 
636 /*
637  * Routines having to do with the management of the vnode table.
638  */
639 
640 /*
641  * Check to see if a free vnode can be recycled. If it can,
642  * recycle it and return it with the vnode interlock held.
643  */
644 static int
645 vtryrecycle(struct vnode *vp)
646 {
647 	struct thread *td = curthread;
648 	vm_object_t object;
649 	struct mount *vnmp;
650 	int error;
651 
652 	/* Don't recycle if we can't get the interlock */
653 	if (!VI_TRYLOCK(vp))
654 		return (EWOULDBLOCK);
655 	if (!VCANRECYCLE(vp)) {
656 		VI_UNLOCK(vp);
657 		return (EBUSY);
658 	}
659 	/*
660 	 * This vnode may found and locked via some other list, if so we
661 	 * can't recycle it yet.
662 	 */
663 	if (vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
664 		return (EWOULDBLOCK);
665 	/*
666 	 * Don't recycle if its filesystem is being suspended.
667 	 */
668 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
669 		VOP_UNLOCK(vp, 0, td);
670 		return (EBUSY);
671 	}
672 
673 	/*
674 	 * Don't recycle if we still have cached pages.
675 	 */
676 	object = vp->v_object;
677 	if (object != NULL) {
678 		VM_OBJECT_LOCK(object);
679 		if (object->resident_page_count ||
680 		    object->ref_count) {
681 			VM_OBJECT_UNLOCK(object);
682 			error = EBUSY;
683 			goto done;
684 		}
685 		VM_OBJECT_UNLOCK(object);
686 	}
687 	if (LIST_FIRST(&vp->v_cache_src)) {
688 		/*
689 		 * note: nameileafonly sysctl is temporary,
690 		 * for debugging only, and will eventually be
691 		 * removed.
692 		 */
693 		if (nameileafonly > 0) {
694 			/*
695 			 * Do not reuse namei-cached directory
696 			 * vnodes that have cached
697 			 * subdirectories.
698 			 */
699 			if (cache_leaf_test(vp) < 0) {
700 				error = EISDIR;
701 				goto done;
702 			}
703 		} else if (nameileafonly < 0 ||
704 			    vmiodirenable == 0) {
705 			/*
706 			 * Do not reuse namei-cached directory
707 			 * vnodes if nameileafonly is -1 or
708 			 * if VMIO backing for directories is
709 			 * turned off (otherwise we reuse them
710 			 * too quickly).
711 			 */
712 			error = EBUSY;
713 			goto done;
714 		}
715 	}
716 	/*
717 	 * If we got this far, we need to acquire the interlock and see if
718 	 * anyone picked up this vnode from another list.  If not, we will
719 	 * mark it with XLOCK via vgonel() so that anyone who does find it
720 	 * will skip over it.
721 	 */
722 	VI_LOCK(vp);
723 	if (!VCANRECYCLE(vp)) {
724 		VI_UNLOCK(vp);
725 		error = EBUSY;
726 		goto done;
727 	}
728 	mtx_lock(&vnode_free_list_mtx);
729 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
730 	vp->v_iflag &= ~VI_FREE;
731 	mtx_unlock(&vnode_free_list_mtx);
732 	vp->v_iflag |= VI_DOOMED;
733 	if ((vp->v_type != VBAD) || (vp->v_data != NULL)) {
734 		VOP_UNLOCK(vp, 0, td);
735 		vgonel(vp, td);
736 	} else
737 		VOP_UNLOCK(vp, LK_INTERLOCK, td);
738 	vn_finished_write(vnmp);
739 	return (0);
740 done:
741 	VOP_UNLOCK(vp, 0, td);
742 	vn_finished_write(vnmp);
743 	return (error);
744 }
745 
746 /*
747  * Return the next vnode from the free list.
748  */
749 int
750 getnewvnode(tag, mp, vops, vpp)
751 	const char *tag;
752 	struct mount *mp;
753 	struct vop_vector *vops;
754 	struct vnode **vpp;
755 {
756 	struct vnode *vp = NULL;
757 	struct vpollinfo *pollinfo = NULL;
758 	struct bufobj *bo;
759 
760 	mtx_lock(&vnode_free_list_mtx);
761 
762 	/*
763 	 * Try to reuse vnodes if we hit the max.  This situation only
764 	 * occurs in certain large-memory (2G+) situations.  We cannot
765 	 * attempt to directly reclaim vnodes due to nasty recursion
766 	 * problems.
767 	 */
768 	while (numvnodes - freevnodes > desiredvnodes) {
769 		if (vnlruproc_sig == 0) {
770 			vnlruproc_sig = 1;      /* avoid unnecessary wakeups */
771 			wakeup(vnlruproc);
772 		}
773 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
774 		    "vlruwk", hz);
775 	}
776 
777 	/*
778 	 * Attempt to reuse a vnode already on the free list, allocating
779 	 * a new vnode if we can't find one or if we have not reached a
780 	 * good minimum for good LRU performance.
781 	 */
782 
783 	if (freevnodes >= wantfreevnodes && numvnodes >= minvnodes) {
784 		int error;
785 		int count;
786 
787 		for (count = 0; count < freevnodes; count++) {
788 			vp = TAILQ_FIRST(&vnode_free_list);
789 			TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
790 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
791 			mtx_unlock(&vnode_free_list_mtx);
792 			error = vtryrecycle(vp);
793 			mtx_lock(&vnode_free_list_mtx);
794 			if (error == 0)
795 				break;
796 			vp = NULL;
797 		}
798 	}
799 	if (vp) {
800 		freevnodes--;
801 		bo = &vp->v_bufobj;
802 		mtx_unlock(&vnode_free_list_mtx);
803 
804 #ifdef INVARIANTS
805 		{
806 			if (vp->v_data)
807 				printf("cleaned vnode isn't, "
808 				       "address %p, inode %p\n",
809 				       vp, vp->v_data);
810 			if (bo->bo_numoutput)
811 				panic("%p: Clean vnode has pending I/O's", vp);
812 			if (vp->v_usecount != 0)
813 				panic("%p: Non-zero use count", vp);
814 			if (vp->v_writecount != 0)
815 				panic("%p: Non-zero write count", vp);
816 		}
817 #endif
818 		if ((pollinfo = vp->v_pollinfo) != NULL) {
819 			/*
820 			 * To avoid lock order reversals, the call to
821 			 * uma_zfree() must be delayed until the vnode
822 			 * interlock is released.
823 			 */
824 			vp->v_pollinfo = NULL;
825 		}
826 #ifdef MAC
827 		mac_destroy_vnode(vp);
828 #endif
829 		vp->v_iflag = 0;
830 		vp->v_vflag = 0;
831 		vp->v_lastw = 0;
832 		vp->v_lasta = 0;
833 		vp->v_cstart = 0;
834 		vp->v_clen = 0;
835 		vp->v_socket = 0;
836 		lockdestroy(vp->v_vnlock);
837 		lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE);
838 		KASSERT(bo->bo_clean.bv_cnt == 0, ("cleanbufcnt not 0"));
839 		KASSERT(bo->bo_clean.bv_root == NULL, ("cleanblkroot not NULL"));
840 		KASSERT(bo->bo_dirty.bv_cnt == 0, ("dirtybufcnt not 0"));
841 		KASSERT(bo->bo_dirty.bv_root == NULL, ("dirtyblkroot not NULL"));
842 	} else {
843 		numvnodes++;
844 		mtx_unlock(&vnode_free_list_mtx);
845 
846 		vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO);
847 		mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
848 		vp->v_dd = vp;
849 		bo = &vp->v_bufobj;
850 		bo->__bo_vnode = vp;
851 		bo->bo_mtx = &vp->v_interlock;
852 		vp->v_vnlock = &vp->v_lock;
853 		lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE);
854 		cache_purge(vp);		/* Sets up v_id. */
855 		LIST_INIT(&vp->v_cache_src);
856 		TAILQ_INIT(&vp->v_cache_dst);
857 	}
858 
859 	TAILQ_INIT(&bo->bo_clean.bv_hd);
860 	TAILQ_INIT(&bo->bo_dirty.bv_hd);
861 	bo->bo_ops = &buf_ops_bio;
862 	bo->bo_private = vp;
863 	vp->v_type = VNON;
864 	vp->v_tag = tag;
865 	vp->v_op = vops;
866 	*vpp = vp;
867 	vp->v_usecount = 1;
868 	vp->v_data = 0;
869 	if (pollinfo != NULL) {
870 		knlist_destroy(&pollinfo->vpi_selinfo.si_note);
871 		mtx_destroy(&pollinfo->vpi_lock);
872 		uma_zfree(vnodepoll_zone, pollinfo);
873 	}
874 #ifdef MAC
875 	mac_init_vnode(vp);
876 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
877 		mac_associate_vnode_singlelabel(mp, vp);
878 	else if (mp == NULL)
879 		printf("NULL mp in getnewvnode()\n");
880 #endif
881 	delmntque(vp);
882 	if (mp != NULL) {
883 		insmntque(vp, mp);
884 		bo->bo_bsize = mp->mnt_stat.f_iosize;
885 	}
886 
887 	return (0);
888 }
889 
890 /*
891  * Delete from old mount point vnode list, if on one.
892  */
893 static void
894 delmntque(struct vnode *vp)
895 {
896 	struct mount *mp;
897 
898 	if (vp->v_mount == NULL)
899 		return;
900 	mp = vp->v_mount;
901 	MNT_ILOCK(mp);
902 	vp->v_mount = NULL;
903 	KASSERT(mp->mnt_nvnodelistsize > 0,
904 		("bad mount point vnode list size"));
905 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
906 	mp->mnt_nvnodelistsize--;
907 	MNT_IUNLOCK(mp);
908 }
909 
910 /*
911  * Insert into list of vnodes for the new mount point, if available.
912  */
913 static void
914 insmntque(struct vnode *vp, struct mount *mp)
915 {
916 
917 	vp->v_mount = mp;
918 	KASSERT(mp != NULL, ("Don't call insmntque(foo, NULL)"));
919 	MNT_ILOCK(vp->v_mount);
920 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
921 	mp->mnt_nvnodelistsize++;
922 	MNT_IUNLOCK(vp->v_mount);
923 }
924 
925 /*
926  * Flush out and invalidate all buffers associated with a vnode.
927  * Called with the underlying object locked.
928  */
929 int
930 vinvalbuf(vp, flags, td, slpflag, slptimeo)
931 	struct vnode *vp;
932 	int flags;
933 	struct thread *td;
934 	int slpflag, slptimeo;
935 {
936 	int error;
937 	struct bufobj *bo;
938 
939 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
940 
941 	bo = &vp->v_bufobj;
942 	BO_LOCK(bo);
943 	if (flags & V_SAVE) {
944 		error = bufobj_wwait(bo, slpflag, slptimeo);
945 		if (error) {
946 			BO_UNLOCK(bo);
947 			return (error);
948 		}
949 		if (bo->bo_dirty.bv_cnt > 0) {
950 			BO_UNLOCK(bo);
951 			if ((error = BO_SYNC(bo, MNT_WAIT, td)) != 0)
952 				return (error);
953 			/*
954 			 * XXX We could save a lock/unlock if this was only
955 			 * enabled under INVARIANTS
956 			 */
957 			BO_LOCK(bo);
958 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
959 				panic("vinvalbuf: dirty bufs");
960 		}
961 	}
962 	/*
963 	 * If you alter this loop please notice that interlock is dropped and
964 	 * reacquired in flushbuflist.  Special care is needed to ensure that
965 	 * no race conditions occur from this.
966 	 */
967 	do {
968 		error = flushbuflist(&bo->bo_clean,
969 		    flags, vp, slpflag, slptimeo);
970 		if (error == 0)
971 			error = flushbuflist(&bo->bo_dirty,
972 			    flags, vp, slpflag, slptimeo);
973 		if (error != 0 && error != EAGAIN) {
974 			BO_UNLOCK(bo);
975 			return (error);
976 		}
977 	} while (error != 0);
978 
979 	/*
980 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
981 	 * have write I/O in-progress but if there is a VM object then the
982 	 * VM object can also have read-I/O in-progress.
983 	 */
984 	do {
985 		bufobj_wwait(bo, 0, 0);
986 		BO_UNLOCK(bo);
987 		if (bo->bo_object != NULL) {
988 			VM_OBJECT_LOCK(bo->bo_object);
989 			vm_object_pip_wait(bo->bo_object, "vnvlbx");
990 			VM_OBJECT_UNLOCK(bo->bo_object);
991 		}
992 		BO_LOCK(bo);
993 	} while (bo->bo_numoutput > 0);
994 	BO_UNLOCK(bo);
995 
996 	/*
997 	 * Destroy the copy in the VM cache, too.
998 	 */
999 	if (bo->bo_object != NULL) {
1000 		VM_OBJECT_LOCK(bo->bo_object);
1001 		vm_object_page_remove(bo->bo_object, 0, 0,
1002 			(flags & V_SAVE) ? TRUE : FALSE);
1003 		VM_OBJECT_UNLOCK(bo->bo_object);
1004 	}
1005 
1006 #ifdef INVARIANTS
1007 	BO_LOCK(bo);
1008 	if ((flags & (V_ALT | V_NORMAL)) == 0 &&
1009 	    (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0))
1010 		panic("vinvalbuf: flush failed");
1011 	BO_UNLOCK(bo);
1012 #endif
1013 	return (0);
1014 }
1015 
1016 /*
1017  * Flush out buffers on the specified list.
1018  *
1019  */
1020 static int
1021 flushbuflist(bufv, flags, vp, slpflag, slptimeo)
1022 	struct bufv *bufv;
1023 	int flags;
1024 	struct vnode *vp;
1025 	int slpflag, slptimeo;
1026 {
1027 	struct buf *bp, *nbp;
1028 	int retval, error;
1029 	struct bufobj *bo;
1030 
1031 	bo = &vp->v_bufobj;
1032 	ASSERT_BO_LOCKED(bo);
1033 
1034 	retval = 0;
1035 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1036 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1037 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1038 			continue;
1039 		}
1040 		retval = EAGAIN;
1041 		error = BUF_TIMELOCK(bp,
1042 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo),
1043 		    "flushbuf", slpflag, slptimeo);
1044 		if (error) {
1045 			BO_LOCK(bo);
1046 			return (error != ENOLCK ? error : EAGAIN);
1047 		}
1048 		/*
1049 		 * XXX Since there are no node locks for NFS, I
1050 		 * believe there is a slight chance that a delayed
1051 		 * write will occur while sleeping just above, so
1052 		 * check for it.  Note that vfs_bio_awrite expects
1053 		 * buffers to reside on a queue, while bwrite and
1054 		 * brelse do not.
1055 		 */
1056 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1057 			(flags & V_SAVE)) {
1058 
1059 			if (bp->b_vp == vp) {
1060 				if (bp->b_flags & B_CLUSTEROK) {
1061 					vfs_bio_awrite(bp);
1062 				} else {
1063 					bremfree(bp);
1064 					bp->b_flags |= B_ASYNC;
1065 					bwrite(bp);
1066 				}
1067 			} else {
1068 				bremfree(bp);
1069 				(void) bwrite(bp);
1070 			}
1071 			BO_LOCK(bo);
1072 			return (EAGAIN);
1073 		}
1074 		bremfree(bp);
1075 		bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF);
1076 		bp->b_flags &= ~B_ASYNC;
1077 		brelse(bp);
1078 		BO_LOCK(bo);
1079 	}
1080 	return (retval);
1081 }
1082 
1083 /*
1084  * Truncate a file's buffer and pages to a specified length.  This
1085  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1086  * sync activity.
1087  */
1088 int
1089 vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td, off_t length, int blksize)
1090 {
1091 	struct buf *bp, *nbp;
1092 	int anyfreed;
1093 	int trunclbn;
1094 	struct bufobj *bo;
1095 
1096 	/*
1097 	 * Round up to the *next* lbn.
1098 	 */
1099 	trunclbn = (length + blksize - 1) / blksize;
1100 
1101 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1102 restart:
1103 	VI_LOCK(vp);
1104 	bo = &vp->v_bufobj;
1105 	anyfreed = 1;
1106 	for (;anyfreed;) {
1107 		anyfreed = 0;
1108 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1109 			if (bp->b_lblkno < trunclbn)
1110 				continue;
1111 			if (BUF_LOCK(bp,
1112 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1113 			    VI_MTX(vp)) == ENOLCK)
1114 				goto restart;
1115 
1116 			bremfree(bp);
1117 			bp->b_flags |= (B_INVAL | B_RELBUF);
1118 			bp->b_flags &= ~B_ASYNC;
1119 			brelse(bp);
1120 			anyfreed = 1;
1121 
1122 			if (nbp != NULL &&
1123 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1124 			    (nbp->b_vp != vp) ||
1125 			    (nbp->b_flags & B_DELWRI))) {
1126 				goto restart;
1127 			}
1128 			VI_LOCK(vp);
1129 		}
1130 
1131 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1132 			if (bp->b_lblkno < trunclbn)
1133 				continue;
1134 			if (BUF_LOCK(bp,
1135 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1136 			    VI_MTX(vp)) == ENOLCK)
1137 				goto restart;
1138 			bremfree(bp);
1139 			bp->b_flags |= (B_INVAL | B_RELBUF);
1140 			bp->b_flags &= ~B_ASYNC;
1141 			brelse(bp);
1142 			anyfreed = 1;
1143 			if (nbp != NULL &&
1144 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1145 			    (nbp->b_vp != vp) ||
1146 			    (nbp->b_flags & B_DELWRI) == 0)) {
1147 				goto restart;
1148 			}
1149 			VI_LOCK(vp);
1150 		}
1151 	}
1152 
1153 	if (length > 0) {
1154 restartsync:
1155 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1156 			if (bp->b_lblkno > 0)
1157 				continue;
1158 			/*
1159 			 * Since we hold the vnode lock this should only
1160 			 * fail if we're racing with the buf daemon.
1161 			 */
1162 			if (BUF_LOCK(bp,
1163 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1164 			    VI_MTX(vp)) == ENOLCK) {
1165 				goto restart;
1166 			}
1167 			KASSERT((bp->b_flags & B_DELWRI),
1168 			    ("buf(%p) on dirty queue without DELWRI", bp));
1169 
1170 			bremfree(bp);
1171 			bawrite(bp);
1172 			VI_LOCK(vp);
1173 			goto restartsync;
1174 		}
1175 	}
1176 
1177 	bufobj_wwait(bo, 0, 0);
1178 	VI_UNLOCK(vp);
1179 	vnode_pager_setsize(vp, length);
1180 
1181 	return (0);
1182 }
1183 
1184 /*
1185  * buf_splay() - splay tree core for the clean/dirty list of buffers in
1186  * 		 a vnode.
1187  *
1188  *	NOTE: We have to deal with the special case of a background bitmap
1189  *	buffer, a situation where two buffers will have the same logical
1190  *	block offset.  We want (1) only the foreground buffer to be accessed
1191  *	in a lookup and (2) must differentiate between the foreground and
1192  *	background buffer in the splay tree algorithm because the splay
1193  *	tree cannot normally handle multiple entities with the same 'index'.
1194  *	We accomplish this by adding differentiating flags to the splay tree's
1195  *	numerical domain.
1196  */
1197 static
1198 struct buf *
1199 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root)
1200 {
1201 	struct buf dummy;
1202 	struct buf *lefttreemax, *righttreemin, *y;
1203 
1204 	if (root == NULL)
1205 		return (NULL);
1206 	lefttreemax = righttreemin = &dummy;
1207 	for (;;) {
1208 		if (lblkno < root->b_lblkno ||
1209 		    (lblkno == root->b_lblkno &&
1210 		    (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1211 			if ((y = root->b_left) == NULL)
1212 				break;
1213 			if (lblkno < y->b_lblkno) {
1214 				/* Rotate right. */
1215 				root->b_left = y->b_right;
1216 				y->b_right = root;
1217 				root = y;
1218 				if ((y = root->b_left) == NULL)
1219 					break;
1220 			}
1221 			/* Link into the new root's right tree. */
1222 			righttreemin->b_left = root;
1223 			righttreemin = root;
1224 		} else if (lblkno > root->b_lblkno ||
1225 		    (lblkno == root->b_lblkno &&
1226 		    (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) {
1227 			if ((y = root->b_right) == NULL)
1228 				break;
1229 			if (lblkno > y->b_lblkno) {
1230 				/* Rotate left. */
1231 				root->b_right = y->b_left;
1232 				y->b_left = root;
1233 				root = y;
1234 				if ((y = root->b_right) == NULL)
1235 					break;
1236 			}
1237 			/* Link into the new root's left tree. */
1238 			lefttreemax->b_right = root;
1239 			lefttreemax = root;
1240 		} else {
1241 			break;
1242 		}
1243 		root = y;
1244 	}
1245 	/* Assemble the new root. */
1246 	lefttreemax->b_right = root->b_left;
1247 	righttreemin->b_left = root->b_right;
1248 	root->b_left = dummy.b_right;
1249 	root->b_right = dummy.b_left;
1250 	return (root);
1251 }
1252 
1253 static void
1254 buf_vlist_remove(struct buf *bp)
1255 {
1256 	struct buf *root;
1257 	struct bufv *bv;
1258 
1259 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1260 	ASSERT_BO_LOCKED(bp->b_bufobj);
1261 	if (bp->b_xflags & BX_VNDIRTY)
1262 		bv = &bp->b_bufobj->bo_dirty;
1263 	else
1264 		bv = &bp->b_bufobj->bo_clean;
1265 	if (bp != bv->bv_root) {
1266 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1267 		KASSERT(root == bp, ("splay lookup failed in remove"));
1268 	}
1269 	if (bp->b_left == NULL) {
1270 		root = bp->b_right;
1271 	} else {
1272 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left);
1273 		root->b_right = bp->b_right;
1274 	}
1275 	bv->bv_root = root;
1276 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1277 	bv->bv_cnt--;
1278 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1279 }
1280 
1281 /*
1282  * Add the buffer to the sorted clean or dirty block list using a
1283  * splay tree algorithm.
1284  *
1285  * NOTE: xflags is passed as a constant, optimizing this inline function!
1286  */
1287 static void
1288 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1289 {
1290 	struct buf *root;
1291 	struct bufv *bv;
1292 
1293 	ASSERT_BO_LOCKED(bo);
1294 	bp->b_xflags |= xflags;
1295 	if (xflags & BX_VNDIRTY)
1296 		bv = &bo->bo_dirty;
1297 	else
1298 		bv = &bo->bo_clean;
1299 
1300 	root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1301 	if (root == NULL) {
1302 		bp->b_left = NULL;
1303 		bp->b_right = NULL;
1304 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1305 	} else if (bp->b_lblkno < root->b_lblkno ||
1306 	    (bp->b_lblkno == root->b_lblkno &&
1307 	    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1308 		bp->b_left = root->b_left;
1309 		bp->b_right = root;
1310 		root->b_left = NULL;
1311 		TAILQ_INSERT_BEFORE(root, bp, b_bobufs);
1312 	} else {
1313 		bp->b_right = root->b_right;
1314 		bp->b_left = root;
1315 		root->b_right = NULL;
1316 		TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs);
1317 	}
1318 	bv->bv_cnt++;
1319 	bv->bv_root = bp;
1320 }
1321 
1322 /*
1323  * Lookup a buffer using the splay tree.  Note that we specifically avoid
1324  * shadow buffers used in background bitmap writes.
1325  *
1326  * This code isn't quite efficient as it could be because we are maintaining
1327  * two sorted lists and do not know which list the block resides in.
1328  *
1329  * During a "make buildworld" the desired buffer is found at one of
1330  * the roots more than 60% of the time.  Thus, checking both roots
1331  * before performing either splay eliminates unnecessary splays on the
1332  * first tree splayed.
1333  */
1334 struct buf *
1335 gbincore(struct bufobj *bo, daddr_t lblkno)
1336 {
1337 	struct buf *bp;
1338 
1339 	ASSERT_BO_LOCKED(bo);
1340 	if ((bp = bo->bo_clean.bv_root) != NULL &&
1341 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1342 		return (bp);
1343 	if ((bp = bo->bo_dirty.bv_root) != NULL &&
1344 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1345 		return (bp);
1346 	if ((bp = bo->bo_clean.bv_root) != NULL) {
1347 		bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp);
1348 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1349 			return (bp);
1350 	}
1351 	if ((bp = bo->bo_dirty.bv_root) != NULL) {
1352 		bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp);
1353 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1354 			return (bp);
1355 	}
1356 	return (NULL);
1357 }
1358 
1359 /*
1360  * Associate a buffer with a vnode.
1361  */
1362 void
1363 bgetvp(struct vnode *vp, struct buf *bp)
1364 {
1365 	KASSERT(bp->b_vp == NULL, ("bgetvp: not free"));
1366 
1367 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
1368 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1369 	    ("bgetvp: bp already attached! %p", bp));
1370 
1371 	ASSERT_VI_LOCKED(vp, "bgetvp");
1372 	vholdl(vp);
1373 	bp->b_vp = vp;
1374 	bp->b_bufobj = &vp->v_bufobj;
1375 	/*
1376 	 * Insert onto list for new vnode.
1377 	 */
1378 	buf_vlist_add(bp, &vp->v_bufobj, BX_VNCLEAN);
1379 }
1380 
1381 /*
1382  * Disassociate a buffer from a vnode.
1383  */
1384 void
1385 brelvp(struct buf *bp)
1386 {
1387 	struct bufobj *bo;
1388 	struct vnode *vp;
1389 
1390 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1391 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1392 
1393 	/*
1394 	 * Delete from old vnode list, if on one.
1395 	 */
1396 	vp = bp->b_vp;		/* XXX */
1397 	bo = bp->b_bufobj;
1398 	BO_LOCK(bo);
1399 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1400 		buf_vlist_remove(bp);
1401 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1402 		bo->bo_flag &= ~BO_ONWORKLST;
1403 		mtx_lock(&sync_mtx);
1404 		LIST_REMOVE(bo, bo_synclist);
1405  		syncer_worklist_len--;
1406 		mtx_unlock(&sync_mtx);
1407 	}
1408 	vdropl(vp);
1409 	bp->b_vp = NULL;
1410 	bp->b_bufobj = NULL;
1411 	BO_UNLOCK(bo);
1412 }
1413 
1414 /*
1415  * Add an item to the syncer work queue.
1416  */
1417 static void
1418 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
1419 {
1420 	int slot;
1421 
1422 	ASSERT_BO_LOCKED(bo);
1423 
1424 	mtx_lock(&sync_mtx);
1425 	if (bo->bo_flag & BO_ONWORKLST)
1426 		LIST_REMOVE(bo, bo_synclist);
1427 	else {
1428 		bo->bo_flag |= BO_ONWORKLST;
1429  		syncer_worklist_len++;
1430 	}
1431 
1432 	if (delay > syncer_maxdelay - 2)
1433 		delay = syncer_maxdelay - 2;
1434 	slot = (syncer_delayno + delay) & syncer_mask;
1435 
1436 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
1437 	mtx_unlock(&sync_mtx);
1438 }
1439 
1440 static int
1441 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
1442 {
1443 	int error, len;
1444 
1445 	mtx_lock(&sync_mtx);
1446 	len = syncer_worklist_len - sync_vnode_count;
1447 	mtx_unlock(&sync_mtx);
1448 	error = SYSCTL_OUT(req, &len, sizeof(len));
1449 	return (error);
1450 }
1451 
1452 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
1453     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
1454 
1455 struct  proc *updateproc;
1456 static void sched_sync(void);
1457 static struct kproc_desc up_kp = {
1458 	"syncer",
1459 	sched_sync,
1460 	&updateproc
1461 };
1462 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp)
1463 
1464 static int
1465 sync_vnode(struct bufobj *bo, struct thread *td)
1466 {
1467 	struct vnode *vp;
1468 	struct mount *mp;
1469 
1470 	vp = bo->__bo_vnode; 	/* XXX */
1471 	if (VOP_ISLOCKED(vp, NULL) != 0)
1472 		return (1);
1473 	if (VI_TRYLOCK(vp) == 0)
1474 		return (1);
1475 	/*
1476 	 * We use vhold in case the vnode does not
1477 	 * successfully sync.  vhold prevents the vnode from
1478 	 * going away when we unlock the sync_mtx so that
1479 	 * we can acquire the vnode interlock.
1480 	 */
1481 	vholdl(vp);
1482 	mtx_unlock(&sync_mtx);
1483 	VI_UNLOCK(vp);
1484 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1485 		vdrop(vp);
1486 		mtx_lock(&sync_mtx);
1487 		return (1);
1488 	}
1489 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
1490 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
1491 	VOP_UNLOCK(vp, 0, td);
1492 	vn_finished_write(mp);
1493 	VI_LOCK(vp);
1494 	if ((bo->bo_flag & BO_ONWORKLST) != 0) {
1495 		/*
1496 		 * Put us back on the worklist.  The worklist
1497 		 * routine will remove us from our current
1498 		 * position and then add us back in at a later
1499 		 * position.
1500 		 */
1501 		vn_syncer_add_to_worklist(bo, syncdelay);
1502 	}
1503 	vdropl(vp);
1504 	VI_UNLOCK(vp);
1505 	mtx_lock(&sync_mtx);
1506 	return (0);
1507 }
1508 
1509 /*
1510  * System filesystem synchronizer daemon.
1511  */
1512 static void
1513 sched_sync(void)
1514 {
1515 	struct synclist *next;
1516 	struct synclist *slp;
1517 	struct bufobj *bo;
1518 	long starttime;
1519 	struct thread *td = FIRST_THREAD_IN_PROC(updateproc);
1520 	static int dummychan;
1521 	int last_work_seen;
1522 	int net_worklist_len;
1523 	int syncer_final_iter;
1524 	int first_printf;
1525 	int error;
1526 
1527 	mtx_lock(&Giant);
1528 	last_work_seen = 0;
1529 	syncer_final_iter = 0;
1530 	first_printf = 1;
1531 	syncer_state = SYNCER_RUNNING;
1532 	starttime = time_second;
1533 
1534 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
1535 	    SHUTDOWN_PRI_LAST);
1536 
1537 	for (;;) {
1538 		mtx_lock(&sync_mtx);
1539 		if (syncer_state == SYNCER_FINAL_DELAY &&
1540 		    syncer_final_iter == 0) {
1541 			mtx_unlock(&sync_mtx);
1542 			kthread_suspend_check(td->td_proc);
1543 			mtx_lock(&sync_mtx);
1544 		}
1545 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
1546 		if (syncer_state != SYNCER_RUNNING &&
1547 		    starttime != time_second) {
1548 			if (first_printf) {
1549 				printf("\nSyncing disks, vnodes remaining...");
1550 				first_printf = 0;
1551 			}
1552 			printf("%d ", net_worklist_len);
1553 		}
1554 		starttime = time_second;
1555 
1556 		/*
1557 		 * Push files whose dirty time has expired.  Be careful
1558 		 * of interrupt race on slp queue.
1559 		 *
1560 		 * Skip over empty worklist slots when shutting down.
1561 		 */
1562 		do {
1563 			slp = &syncer_workitem_pending[syncer_delayno];
1564 			syncer_delayno += 1;
1565 			if (syncer_delayno == syncer_maxdelay)
1566 				syncer_delayno = 0;
1567 			next = &syncer_workitem_pending[syncer_delayno];
1568 			/*
1569 			 * If the worklist has wrapped since the
1570 			 * it was emptied of all but syncer vnodes,
1571 			 * switch to the FINAL_DELAY state and run
1572 			 * for one more second.
1573 			 */
1574 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
1575 			    net_worklist_len == 0 &&
1576 			    last_work_seen == syncer_delayno) {
1577 				syncer_state = SYNCER_FINAL_DELAY;
1578 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
1579 			}
1580 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
1581 		    syncer_worklist_len > 0);
1582 
1583 		/*
1584 		 * Keep track of the last time there was anything
1585 		 * on the worklist other than syncer vnodes.
1586 		 * Return to the SHUTTING_DOWN state if any
1587 		 * new work appears.
1588 		 */
1589 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
1590 			last_work_seen = syncer_delayno;
1591 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
1592 			syncer_state = SYNCER_SHUTTING_DOWN;
1593 		while ((bo = LIST_FIRST(slp)) != NULL) {
1594 			error = sync_vnode(bo, td);
1595 			if (error == 1) {
1596 				LIST_REMOVE(bo, bo_synclist);
1597 				LIST_INSERT_HEAD(next, bo, bo_synclist);
1598 				continue;
1599 			}
1600 		}
1601 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
1602 			syncer_final_iter--;
1603 		mtx_unlock(&sync_mtx);
1604 
1605 		/*
1606 		 * Do soft update processing.
1607 		 */
1608 		if (softdep_process_worklist_hook != NULL)
1609 			(*softdep_process_worklist_hook)(NULL);
1610 
1611 		/*
1612 		 * The variable rushjob allows the kernel to speed up the
1613 		 * processing of the filesystem syncer process. A rushjob
1614 		 * value of N tells the filesystem syncer to process the next
1615 		 * N seconds worth of work on its queue ASAP. Currently rushjob
1616 		 * is used by the soft update code to speed up the filesystem
1617 		 * syncer process when the incore state is getting so far
1618 		 * ahead of the disk that the kernel memory pool is being
1619 		 * threatened with exhaustion.
1620 		 */
1621 		mtx_lock(&sync_mtx);
1622 		if (rushjob > 0) {
1623 			rushjob -= 1;
1624 			mtx_unlock(&sync_mtx);
1625 			continue;
1626 		}
1627 		mtx_unlock(&sync_mtx);
1628 		/*
1629 		 * Just sleep for a short period if time between
1630 		 * iterations when shutting down to allow some I/O
1631 		 * to happen.
1632 		 *
1633 		 * If it has taken us less than a second to process the
1634 		 * current work, then wait. Otherwise start right over
1635 		 * again. We can still lose time if any single round
1636 		 * takes more than two seconds, but it does not really
1637 		 * matter as we are just trying to generally pace the
1638 		 * filesystem activity.
1639 		 */
1640 		if (syncer_state != SYNCER_RUNNING)
1641 			tsleep(&dummychan, PPAUSE, "syncfnl",
1642 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
1643 		else if (time_second == starttime)
1644 			tsleep(&lbolt, PPAUSE, "syncer", 0);
1645 	}
1646 }
1647 
1648 /*
1649  * Request the syncer daemon to speed up its work.
1650  * We never push it to speed up more than half of its
1651  * normal turn time, otherwise it could take over the cpu.
1652  */
1653 int
1654 speedup_syncer()
1655 {
1656 	struct thread *td;
1657 	int ret = 0;
1658 
1659 	td = FIRST_THREAD_IN_PROC(updateproc);
1660 	sleepq_remove(td, &lbolt);
1661 	mtx_lock(&sync_mtx);
1662 	if (rushjob < syncdelay / 2) {
1663 		rushjob += 1;
1664 		stat_rush_requests += 1;
1665 		ret = 1;
1666 	}
1667 	mtx_unlock(&sync_mtx);
1668 	return (ret);
1669 }
1670 
1671 /*
1672  * Tell the syncer to speed up its work and run though its work
1673  * list several times, then tell it to shut down.
1674  */
1675 static void
1676 syncer_shutdown(void *arg, int howto)
1677 {
1678 	struct thread *td;
1679 
1680 	if (howto & RB_NOSYNC)
1681 		return;
1682 	td = FIRST_THREAD_IN_PROC(updateproc);
1683 	sleepq_remove(td, &lbolt);
1684 	mtx_lock(&sync_mtx);
1685 	syncer_state = SYNCER_SHUTTING_DOWN;
1686 	rushjob = 0;
1687 	mtx_unlock(&sync_mtx);
1688 	kproc_shutdown(arg, howto);
1689 }
1690 
1691 /*
1692  * Reassign a buffer from one vnode to another.
1693  * Used to assign file specific control information
1694  * (indirect blocks) to the vnode to which they belong.
1695  */
1696 void
1697 reassignbuf(struct buf *bp)
1698 {
1699 	struct vnode *vp;
1700 	struct bufobj *bo;
1701 	int delay;
1702 
1703 	vp = bp->b_vp;
1704 	bo = bp->b_bufobj;
1705 	++reassignbufcalls;
1706 
1707 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
1708 	    bp, bp->b_vp, bp->b_flags);
1709 	/*
1710 	 * B_PAGING flagged buffers cannot be reassigned because their vp
1711 	 * is not fully linked in.
1712 	 */
1713 	if (bp->b_flags & B_PAGING)
1714 		panic("cannot reassign paging buffer");
1715 
1716 	/*
1717 	 * Delete from old vnode list, if on one.
1718 	 */
1719 	VI_LOCK(vp);
1720 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1721 		buf_vlist_remove(bp);
1722 	/*
1723 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
1724 	 * of clean buffers.
1725 	 */
1726 	if (bp->b_flags & B_DELWRI) {
1727 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
1728 			switch (vp->v_type) {
1729 			case VDIR:
1730 				delay = dirdelay;
1731 				break;
1732 			case VCHR:
1733 				delay = metadelay;
1734 				break;
1735 			default:
1736 				delay = filedelay;
1737 			}
1738 			vn_syncer_add_to_worklist(bo, delay);
1739 		}
1740 		buf_vlist_add(bp, bo, BX_VNDIRTY);
1741 	} else {
1742 		buf_vlist_add(bp, bo, BX_VNCLEAN);
1743 
1744 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1745 			mtx_lock(&sync_mtx);
1746 			LIST_REMOVE(bo, bo_synclist);
1747  			syncer_worklist_len--;
1748 			mtx_unlock(&sync_mtx);
1749 			bo->bo_flag &= ~BO_ONWORKLST;
1750 		}
1751 	}
1752 	VI_UNLOCK(vp);
1753 }
1754 
1755 static void
1756 v_incr_usecount(struct vnode *vp, int delta)
1757 {
1758 
1759 	vp->v_usecount += delta;
1760 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1761 		dev_lock();
1762 		vp->v_rdev->si_usecount += delta;
1763 		dev_unlock();
1764 	}
1765 }
1766 
1767 /*
1768  * Grab a particular vnode from the free list, increment its
1769  * reference count and lock it. The vnode lock bit is set if the
1770  * vnode is being eliminated in vgone. The process is awakened
1771  * when the transition is completed, and an error returned to
1772  * indicate that the vnode is no longer usable (possibly having
1773  * been changed to a new filesystem type).
1774  */
1775 int
1776 vget(vp, flags, td)
1777 	struct vnode *vp;
1778 	int flags;
1779 	struct thread *td;
1780 {
1781 	int error;
1782 
1783 	/*
1784 	 * If the vnode is in the process of being cleaned out for
1785 	 * another use, we wait for the cleaning to finish and then
1786 	 * return failure. Cleaning is determined by checking that
1787 	 * the VI_XLOCK flag is set.
1788 	 */
1789 	if ((flags & LK_INTERLOCK) == 0)
1790 		VI_LOCK(vp);
1791 	if (vp->v_iflag & VI_XLOCK && vp->v_vxthread != curthread) {
1792 		if ((flags & LK_NOWAIT) == 0) {
1793 			vp->v_iflag |= VI_XWANT;
1794 			msleep(vp, VI_MTX(vp), PINOD | PDROP, "vget", 0);
1795 			return (ENOENT);
1796 		}
1797 		VI_UNLOCK(vp);
1798 		return (EBUSY);
1799 	}
1800 
1801 	v_incr_usecount(vp, 1);
1802 
1803 	if (VSHOULDBUSY(vp))
1804 		vbusy(vp);
1805 	if (flags & LK_TYPE_MASK) {
1806 		if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) {
1807 			/*
1808 			 * must expand vrele here because we do not want
1809 			 * to call VOP_INACTIVE if the reference count
1810 			 * drops back to zero since it was never really
1811 			 * active. We must remove it from the free list
1812 			 * before sleeping so that multiple processes do
1813 			 * not try to recycle it.
1814 			 */
1815 			VI_LOCK(vp);
1816 			v_incr_usecount(vp, -1);
1817 			if (VSHOULDFREE(vp))
1818 				vfree(vp);
1819 			else
1820 				vlruvp(vp);
1821 			VI_UNLOCK(vp);
1822 		}
1823 		return (error);
1824 	}
1825 	VI_UNLOCK(vp);
1826 	return (0);
1827 }
1828 
1829 /*
1830  * Increase the reference count of a vnode.
1831  */
1832 void
1833 vref(struct vnode *vp)
1834 {
1835 
1836 	VI_LOCK(vp);
1837 	v_incr_usecount(vp, 1);
1838 	VI_UNLOCK(vp);
1839 }
1840 
1841 /*
1842  * Return reference count of a vnode.
1843  *
1844  * The results of this call are only guaranteed when some mechanism other
1845  * than the VI lock is used to stop other processes from gaining references
1846  * to the vnode.  This may be the case if the caller holds the only reference.
1847  * This is also useful when stale data is acceptable as race conditions may
1848  * be accounted for by some other means.
1849  */
1850 int
1851 vrefcnt(struct vnode *vp)
1852 {
1853 	int usecnt;
1854 
1855 	VI_LOCK(vp);
1856 	usecnt = vp->v_usecount;
1857 	VI_UNLOCK(vp);
1858 
1859 	return (usecnt);
1860 }
1861 
1862 
1863 /*
1864  * Vnode put/release.
1865  * If count drops to zero, call inactive routine and return to freelist.
1866  */
1867 void
1868 vrele(vp)
1869 	struct vnode *vp;
1870 {
1871 	struct thread *td = curthread;	/* XXX */
1872 
1873 	KASSERT(vp != NULL, ("vrele: null vp"));
1874 
1875 	VI_LOCK(vp);
1876 
1877 	/* Skip this v_writecount check if we're going to panic below. */
1878 	KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1,
1879 	    ("vrele: missed vn_close"));
1880 
1881 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
1882 	    vp->v_usecount == 1)) {
1883 		v_incr_usecount(vp, -1);
1884 		VI_UNLOCK(vp);
1885 
1886 		return;
1887 	}
1888 
1889 	if (vp->v_usecount == 1) {
1890 		v_incr_usecount(vp, -1);
1891 		/*
1892 		 * We must call VOP_INACTIVE with the node locked. Mark
1893 		 * as VI_DOINGINACT to avoid recursion.
1894 		 */
1895 		if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) {
1896 			VI_LOCK(vp);
1897 			KASSERT((vp->v_iflag & VI_DOINGINACT) == 0,
1898 			    ("vrele: recursed on VI_DOINGINACT"));
1899 			vp->v_iflag |= VI_DOINGINACT;
1900 			VI_UNLOCK(vp);
1901 			VOP_INACTIVE(vp, td);
1902 			VI_LOCK(vp);
1903 			KASSERT(vp->v_iflag & VI_DOINGINACT,
1904 			    ("vrele: lost VI_DOINGINACT"));
1905 			vp->v_iflag &= ~VI_DOINGINACT;
1906 		} else
1907 			VI_LOCK(vp);
1908 		if (VSHOULDFREE(vp))
1909 			vfree(vp);
1910 		else
1911 			vlruvp(vp);
1912 		VI_UNLOCK(vp);
1913 
1914 	} else {
1915 #ifdef DIAGNOSTIC
1916 		vprint("vrele: negative ref count", vp);
1917 #endif
1918 		VI_UNLOCK(vp);
1919 		panic("vrele: negative ref cnt");
1920 	}
1921 }
1922 
1923 /*
1924  * Release an already locked vnode.  This give the same effects as
1925  * unlock+vrele(), but takes less time and avoids releasing and
1926  * re-aquiring the lock (as vrele() aquires the lock internally.)
1927  */
1928 void
1929 vput(vp)
1930 	struct vnode *vp;
1931 {
1932 	struct thread *td = curthread;	/* XXX */
1933 
1934 	KASSERT(vp != NULL, ("vput: null vp"));
1935 	VI_LOCK(vp);
1936 	/* Skip this v_writecount check if we're going to panic below. */
1937 	KASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1,
1938 	    ("vput: missed vn_close"));
1939 
1940 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
1941 	    vp->v_usecount == 1)) {
1942 		v_incr_usecount(vp, -1);
1943 		VOP_UNLOCK(vp, LK_INTERLOCK, td);
1944 		return;
1945 	}
1946 
1947 	if (vp->v_usecount == 1) {
1948 		v_incr_usecount(vp, -1);
1949 		/*
1950 		 * We must call VOP_INACTIVE with the node locked, so
1951 		 * we just need to release the vnode mutex. Mark as
1952 		 * as VI_DOINGINACT to avoid recursion.
1953 		 */
1954 		KASSERT((vp->v_iflag & VI_DOINGINACT) == 0,
1955 		    ("vput: recursed on VI_DOINGINACT"));
1956 		vp->v_iflag |= VI_DOINGINACT;
1957 		VI_UNLOCK(vp);
1958 		VOP_INACTIVE(vp, td);
1959 		VI_LOCK(vp);
1960 		KASSERT(vp->v_iflag & VI_DOINGINACT,
1961 		    ("vput: lost VI_DOINGINACT"));
1962 		vp->v_iflag &= ~VI_DOINGINACT;
1963 		if (VSHOULDFREE(vp))
1964 			vfree(vp);
1965 		else
1966 			vlruvp(vp);
1967 		VI_UNLOCK(vp);
1968 
1969 	} else {
1970 #ifdef DIAGNOSTIC
1971 		vprint("vput: negative ref count", vp);
1972 #endif
1973 		panic("vput: negative ref cnt");
1974 	}
1975 }
1976 
1977 /*
1978  * Somebody doesn't want the vnode recycled.
1979  */
1980 void
1981 vhold(struct vnode *vp)
1982 {
1983 
1984 	VI_LOCK(vp);
1985 	vholdl(vp);
1986 	VI_UNLOCK(vp);
1987 }
1988 
1989 static void
1990 vholdl(struct vnode *vp)
1991 {
1992 
1993 	vp->v_holdcnt++;
1994 	if (VSHOULDBUSY(vp))
1995 		vbusy(vp);
1996 }
1997 
1998 /*
1999  * Note that there is one less who cares about this vnode.  vdrop() is the
2000  * opposite of vhold().
2001  */
2002 void
2003 vdrop(struct vnode *vp)
2004 {
2005 
2006 	VI_LOCK(vp);
2007 	vdropl(vp);
2008 	VI_UNLOCK(vp);
2009 }
2010 
2011 static void
2012 vdropl(struct vnode *vp)
2013 {
2014 
2015 	if (vp->v_holdcnt <= 0)
2016 		panic("vdrop: holdcnt");
2017 	vp->v_holdcnt--;
2018 	if (VSHOULDFREE(vp))
2019 		vfree(vp);
2020 	else
2021 		vlruvp(vp);
2022 }
2023 
2024 /*
2025  * Remove any vnodes in the vnode table belonging to mount point mp.
2026  *
2027  * If FORCECLOSE is not specified, there should not be any active ones,
2028  * return error if any are found (nb: this is a user error, not a
2029  * system error). If FORCECLOSE is specified, detach any active vnodes
2030  * that are found.
2031  *
2032  * If WRITECLOSE is set, only flush out regular file vnodes open for
2033  * writing.
2034  *
2035  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2036  *
2037  * `rootrefs' specifies the base reference count for the root vnode
2038  * of this filesystem. The root vnode is considered busy if its
2039  * v_usecount exceeds this value. On a successful return, vflush(, td)
2040  * will call vrele() on the root vnode exactly rootrefs times.
2041  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2042  * be zero.
2043  */
2044 #ifdef DIAGNOSTIC
2045 static int busyprt = 0;		/* print out busy vnodes */
2046 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
2047 #endif
2048 
2049 int
2050 vflush(mp, rootrefs, flags, td)
2051 	struct mount *mp;
2052 	int rootrefs;
2053 	int flags;
2054 	struct thread *td;
2055 {
2056 	struct vnode *vp, *nvp, *rootvp = NULL;
2057 	struct vattr vattr;
2058 	int busy = 0, error;
2059 
2060 	if (rootrefs > 0) {
2061 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2062 		    ("vflush: bad args"));
2063 		/*
2064 		 * Get the filesystem root vnode. We can vput() it
2065 		 * immediately, since with rootrefs > 0, it won't go away.
2066 		 */
2067 		if ((error = VFS_ROOT(mp, &rootvp, td)) != 0)
2068 			return (error);
2069 		vput(rootvp);
2070 
2071 	}
2072 	MNT_ILOCK(mp);
2073 loop:
2074 	MNT_VNODE_FOREACH(vp, mp, nvp) {
2075 
2076 		VI_LOCK(vp);
2077 		MNT_IUNLOCK(mp);
2078 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td);
2079 		if (error) {
2080 			MNT_ILOCK(mp);
2081 			goto loop;
2082 		}
2083 		/*
2084 		 * Skip over a vnodes marked VV_SYSTEM.
2085 		 */
2086 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2087 			VOP_UNLOCK(vp, 0, td);
2088 			MNT_ILOCK(mp);
2089 			continue;
2090 		}
2091 		/*
2092 		 * If WRITECLOSE is set, flush out unlinked but still open
2093 		 * files (even if open only for reading) and regular file
2094 		 * vnodes open for writing.
2095 		 */
2096 		if (flags & WRITECLOSE) {
2097 			error = VOP_GETATTR(vp, &vattr, td->td_ucred, td);
2098 			VI_LOCK(vp);
2099 
2100 			if ((vp->v_type == VNON ||
2101 			    (error == 0 && vattr.va_nlink > 0)) &&
2102 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2103 				VOP_UNLOCK(vp, LK_INTERLOCK, td);
2104 				MNT_ILOCK(mp);
2105 				continue;
2106 			}
2107 		} else
2108 			VI_LOCK(vp);
2109 
2110 		VOP_UNLOCK(vp, 0, td);
2111 
2112 		/*
2113 		 * With v_usecount == 0, all we need to do is clear out the
2114 		 * vnode data structures and we are done.
2115 		 */
2116 		if (vp->v_usecount == 0) {
2117 			vgonel(vp, td);
2118 			MNT_ILOCK(mp);
2119 			continue;
2120 		}
2121 
2122 		/*
2123 		 * If FORCECLOSE is set, forcibly close the vnode. For block
2124 		 * or character devices, revert to an anonymous device. For
2125 		 * all other files, just kill them.
2126 		 */
2127 		if (flags & FORCECLOSE) {
2128 			KASSERT(vp->v_type != VCHR && vp->v_type != VBLK,
2129 			    ("device VNODE %p is FORCECLOSED", vp));
2130 			vgonel(vp, td);
2131 			MNT_ILOCK(mp);
2132 			continue;
2133 		}
2134 #ifdef DIAGNOSTIC
2135 		if (busyprt)
2136 			vprint("vflush: busy vnode", vp);
2137 #endif
2138 		VI_UNLOCK(vp);
2139 		MNT_ILOCK(mp);
2140 		busy++;
2141 	}
2142 	MNT_IUNLOCK(mp);
2143 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2144 		/*
2145 		 * If just the root vnode is busy, and if its refcount
2146 		 * is equal to `rootrefs', then go ahead and kill it.
2147 		 */
2148 		VI_LOCK(rootvp);
2149 		KASSERT(busy > 0, ("vflush: not busy"));
2150 		KASSERT(rootvp->v_usecount >= rootrefs,
2151 		    ("vflush: usecount %d < rootrefs %d",
2152 		     rootvp->v_usecount, rootrefs));
2153 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2154 			vgonel(rootvp, td);
2155 			busy = 0;
2156 		} else
2157 			VI_UNLOCK(rootvp);
2158 	}
2159 	if (busy)
2160 		return (EBUSY);
2161 	for (; rootrefs > 0; rootrefs--)
2162 		vrele(rootvp);
2163 	return (0);
2164 }
2165 
2166 /*
2167  * This moves a now (likely recyclable) vnode to the end of the
2168  * mountlist.  XXX However, it is temporarily disabled until we
2169  * can clean up ffs_sync() and friends, which have loop restart
2170  * conditions which this code causes to operate O(N^2).
2171  */
2172 static void
2173 vlruvp(struct vnode *vp)
2174 {
2175 #if 0
2176 	struct mount *mp;
2177 
2178 	if ((mp = vp->v_mount) != NULL) {
2179 		MNT_ILOCK(mp);
2180 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2181 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2182 		MNT_IUNLOCK(mp);
2183 	}
2184 #endif
2185 }
2186 
2187 static void
2188 vx_lock(struct vnode *vp)
2189 {
2190 
2191 	ASSERT_VI_LOCKED(vp, "vx_lock");
2192 
2193 	/*
2194 	 * Prevent the vnode from being recycled or brought into use while we
2195 	 * clean it out.
2196 	 */
2197 	if (vp->v_iflag & VI_XLOCK)
2198 		panic("vx_lock: deadlock");
2199 	vp->v_iflag |= VI_XLOCK;
2200 	vp->v_vxthread = curthread;
2201 }
2202 
2203 static void
2204 vx_unlock(struct vnode *vp)
2205 {
2206 	ASSERT_VI_LOCKED(vp, "vx_unlock");
2207 	vp->v_iflag &= ~VI_XLOCK;
2208 	vp->v_vxthread = NULL;
2209 	if (vp->v_iflag & VI_XWANT) {
2210 		vp->v_iflag &= ~VI_XWANT;
2211 		wakeup(vp);
2212 	}
2213 }
2214 /*
2215  * Recycle an unused vnode to the front of the free list.
2216  * Release the passed interlock if the vnode will be recycled.
2217  */
2218 int
2219 vrecycle(struct vnode *vp, struct thread *td)
2220 {
2221 
2222 	VI_LOCK(vp);
2223 	if (vp->v_usecount == 0) {
2224 		vgonel(vp, td);
2225 		return (1);
2226 	}
2227 	VI_UNLOCK(vp);
2228 	return (0);
2229 }
2230 
2231 /*
2232  * Eliminate all activity associated with a vnode
2233  * in preparation for reuse.
2234  */
2235 void
2236 vgone(struct vnode *vp)
2237 {
2238 	struct thread *td = curthread;	/* XXX */
2239 
2240 	VI_LOCK(vp);
2241 	vgonel(vp, td);
2242 }
2243 
2244 /*
2245  * vgone, with the vp interlock held.
2246  */
2247 void
2248 vgonel(struct vnode *vp, struct thread *td)
2249 {
2250 	int active;
2251 
2252 	/*
2253 	 * If a vgone (or vclean) is already in progress,
2254 	 * wait until it is done and return.
2255 	 */
2256 	ASSERT_VI_LOCKED(vp, "vgonel");
2257 	if (vp->v_iflag & VI_XLOCK) {
2258 		vp->v_iflag |= VI_XWANT;
2259 		msleep(vp, VI_MTX(vp), PINOD | PDROP, "vgone", 0);
2260 		return;
2261 	}
2262 	vx_lock(vp);
2263 
2264 	/*
2265 	 * Check to see if the vnode is in use. If so we have to reference it
2266 	 * before we clean it out so that its count cannot fall to zero and
2267 	 * generate a race against ourselves to recycle it.
2268 	 */
2269 	if ((active = vp->v_usecount))
2270 		v_incr_usecount(vp, 1);
2271 
2272 	/*
2273 	 * Even if the count is zero, the VOP_INACTIVE routine may still
2274 	 * have the object locked while it cleans it out. The VOP_LOCK
2275 	 * ensures that the VOP_INACTIVE routine is done with its work.
2276 	 * For active vnodes, it ensures that no other activity can
2277 	 * occur while the underlying object is being cleaned out.
2278 	 */
2279 	VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td);
2280 
2281 	/*
2282 	 * Clean out any buffers associated with the vnode.
2283 	 * If the flush fails, just toss the buffers.
2284 	 */
2285 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd));
2286 		(void) vn_write_suspend_wait(vp, NULL, V_WAIT);
2287 	if (vinvalbuf(vp, V_SAVE, td, 0, 0) != 0)
2288 		vinvalbuf(vp, 0, td, 0, 0);
2289 
2290 	/*
2291 	 * Any other processes trying to obtain this lock must first
2292 	 * wait for VXLOCK to clear, then call the new lock operation.
2293 	 */
2294 	VOP_UNLOCK(vp, 0, td);
2295 
2296 	/*
2297 	 * If purging an active vnode, it must be closed and
2298 	 * deactivated before being reclaimed. Note that the
2299 	 * VOP_INACTIVE will unlock the vnode.
2300 	 */
2301 	if (active) {
2302 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
2303 		VI_LOCK(vp);
2304 		if ((vp->v_iflag & VI_DOINGINACT) == 0) {
2305 			KASSERT((vp->v_iflag & VI_DOINGINACT) == 0,
2306 			    ("vclean: recursed on VI_DOINGINACT"));
2307 			vp->v_iflag |= VI_DOINGINACT;
2308 			VI_UNLOCK(vp);
2309 			if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
2310 				panic("vclean: cannot relock.");
2311 			VOP_INACTIVE(vp, td);
2312 			VI_LOCK(vp);
2313 			KASSERT(vp->v_iflag & VI_DOINGINACT,
2314 			    ("vclean: lost VI_DOINGINACT"));
2315 			vp->v_iflag &= ~VI_DOINGINACT;
2316 		}
2317 		VI_UNLOCK(vp);
2318 	}
2319 	/*
2320 	 * Reclaim the vnode.
2321 	 */
2322 	if (VOP_RECLAIM(vp, td))
2323 		panic("vclean: cannot reclaim");
2324 
2325 	KASSERT(vp->v_object == NULL,
2326 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
2327 
2328 	if (active) {
2329 		/*
2330 		 * Inline copy of vrele() since VOP_INACTIVE
2331 		 * has already been called.
2332 		 */
2333 		VI_LOCK(vp);
2334 		v_incr_usecount(vp, -1);
2335 		if (vp->v_usecount <= 0) {
2336 #ifdef INVARIANTS
2337 			if (vp->v_usecount < 0 || vp->v_writecount != 0) {
2338 				vprint("vclean: bad ref count", vp);
2339 				panic("vclean: ref cnt");
2340 			}
2341 #endif
2342 			if (VSHOULDFREE(vp))
2343 				vfree(vp);
2344 		}
2345 		VI_UNLOCK(vp);
2346 	}
2347 	/*
2348 	 * Delete from old mount point vnode list.
2349 	 */
2350 	delmntque(vp);
2351 	cache_purge(vp);
2352 	VI_LOCK(vp);
2353 	if (VSHOULDFREE(vp))
2354 		vfree(vp);
2355 
2356 	/*
2357 	 * Done with purge, reset to the standard lock and
2358 	 * notify sleepers of the grim news.
2359 	 */
2360 	vp->v_vnlock = &vp->v_lock;
2361 	vp->v_op = &dead_vnodeops;
2362 	vp->v_tag = "none";
2363 
2364 	VI_UNLOCK(vp);
2365 
2366 	/*
2367 	 * If special device, remove it from special device alias list
2368 	 * if it is on one.
2369 	 */
2370 	VI_LOCK(vp);
2371 	if (vp->v_type == VCHR && vp->v_rdev != NULL)
2372 		dev_rel(vp);
2373 
2374 	/*
2375 	 * If it is on the freelist and not already at the head,
2376 	 * move it to the head of the list. The test of the
2377 	 * VDOOMED flag and the reference count of zero is because
2378 	 * it will be removed from the free list by getnewvnode,
2379 	 * but will not have its reference count incremented until
2380 	 * after calling vgone. If the reference count were
2381 	 * incremented first, vgone would (incorrectly) try to
2382 	 * close the previous instance of the underlying object.
2383 	 */
2384 	if (vp->v_usecount == 0 && !(vp->v_iflag & VI_DOOMED)) {
2385 		mtx_lock(&vnode_free_list_mtx);
2386 		if (vp->v_iflag & VI_FREE) {
2387 			TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2388 		} else {
2389 			vp->v_iflag |= VI_FREE;
2390 			freevnodes++;
2391 		}
2392 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2393 		mtx_unlock(&vnode_free_list_mtx);
2394 	}
2395 
2396 	vp->v_type = VBAD;
2397 	vx_unlock(vp);
2398 	VI_UNLOCK(vp);
2399 }
2400 
2401 /*
2402  * Lookup a vnode by device number.
2403  */
2404 int
2405 vfinddev(dev, vpp)
2406 	struct cdev *dev;
2407 	struct vnode **vpp;
2408 {
2409 	struct vnode *vp;
2410 
2411 	dev_lock();
2412 	SLIST_FOREACH(vp, &dev->si_hlist, v_specnext) {
2413 		*vpp = vp;
2414 		dev_unlock();
2415 		return (1);
2416 	}
2417 	dev_unlock();
2418 	return (0);
2419 }
2420 
2421 /*
2422  * Calculate the total number of references to a special device.
2423  */
2424 int
2425 vcount(vp)
2426 	struct vnode *vp;
2427 {
2428 	int count;
2429 
2430 	dev_lock();
2431 	count = vp->v_rdev->si_usecount;
2432 	dev_unlock();
2433 	return (count);
2434 }
2435 
2436 /*
2437  * Same as above, but using the struct cdev *as argument
2438  */
2439 int
2440 count_dev(dev)
2441 	struct cdev *dev;
2442 {
2443 	int count;
2444 
2445 	dev_lock();
2446 	count = dev->si_usecount;
2447 	dev_unlock();
2448 	return(count);
2449 }
2450 
2451 /*
2452  * Print out a description of a vnode.
2453  */
2454 static char *typename[] =
2455 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"};
2456 
2457 void
2458 vn_printf(struct vnode *vp, const char *fmt, ...)
2459 {
2460 	va_list ap;
2461 	char buf[96];
2462 
2463 	va_start(ap, fmt);
2464 	vprintf(fmt, ap);
2465 	va_end(ap);
2466 	printf("%p: ", (void *)vp);
2467 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
2468 	printf("    usecount %d, writecount %d, refcount %d mountedhere %p\n",
2469 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere);
2470 	buf[0] = '\0';
2471 	buf[1] = '\0';
2472 	if (vp->v_vflag & VV_ROOT)
2473 		strcat(buf, "|VV_ROOT");
2474 	if (vp->v_vflag & VV_TEXT)
2475 		strcat(buf, "|VV_TEXT");
2476 	if (vp->v_vflag & VV_SYSTEM)
2477 		strcat(buf, "|VV_SYSTEM");
2478 	if (vp->v_iflag & VI_XLOCK)
2479 		strcat(buf, "|VI_XLOCK");
2480 	if (vp->v_iflag & VI_XWANT)
2481 		strcat(buf, "|VI_XWANT");
2482 	if (vp->v_iflag & VI_DOOMED)
2483 		strcat(buf, "|VI_DOOMED");
2484 	if (vp->v_iflag & VI_FREE)
2485 		strcat(buf, "|VI_FREE");
2486 	printf("    flags (%s)\n", buf + 1);
2487 	if (mtx_owned(VI_MTX(vp)))
2488 		printf(" VI_LOCKed");
2489 	if (vp->v_object != NULL);
2490 		printf("    v_object %p\n", vp->v_object);
2491 	printf("    ");
2492 	lockmgr_printinfo(vp->v_vnlock);
2493 	printf("\n");
2494 	if (vp->v_data != NULL)
2495 		VOP_PRINT(vp);
2496 }
2497 
2498 #ifdef DDB
2499 #include <ddb/ddb.h>
2500 /*
2501  * List all of the locked vnodes in the system.
2502  * Called when debugging the kernel.
2503  */
2504 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
2505 {
2506 	struct mount *mp, *nmp;
2507 	struct vnode *vp;
2508 
2509 	/*
2510 	 * Note: because this is DDB, we can't obey the locking semantics
2511 	 * for these structures, which means we could catch an inconsistent
2512 	 * state and dereference a nasty pointer.  Not much to be done
2513 	 * about that.
2514 	 */
2515 	printf("Locked vnodes\n");
2516 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
2517 		nmp = TAILQ_NEXT(mp, mnt_list);
2518 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2519 			if (VOP_ISLOCKED(vp, NULL))
2520 				vprint("", vp);
2521 		}
2522 		nmp = TAILQ_NEXT(mp, mnt_list);
2523 	}
2524 }
2525 #endif
2526 
2527 /*
2528  * Fill in a struct xvfsconf based on a struct vfsconf.
2529  */
2530 static void
2531 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp)
2532 {
2533 
2534 	strcpy(xvfsp->vfc_name, vfsp->vfc_name);
2535 	xvfsp->vfc_typenum = vfsp->vfc_typenum;
2536 	xvfsp->vfc_refcount = vfsp->vfc_refcount;
2537 	xvfsp->vfc_flags = vfsp->vfc_flags;
2538 	/*
2539 	 * These are unused in userland, we keep them
2540 	 * to not break binary compatibility.
2541 	 */
2542 	xvfsp->vfc_vfsops = NULL;
2543 	xvfsp->vfc_next = NULL;
2544 }
2545 
2546 /*
2547  * Top level filesystem related information gathering.
2548  */
2549 static int
2550 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
2551 {
2552 	struct vfsconf *vfsp;
2553 	struct xvfsconf xvfsp;
2554 	int error;
2555 
2556 	error = 0;
2557 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2558 		vfsconf2x(vfsp, &xvfsp);
2559 		error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp);
2560 		if (error)
2561 			break;
2562 	}
2563 	return (error);
2564 }
2565 
2566 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist,
2567     "S,xvfsconf", "List of all configured filesystems");
2568 
2569 #ifndef BURN_BRIDGES
2570 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
2571 
2572 static int
2573 vfs_sysctl(SYSCTL_HANDLER_ARGS)
2574 {
2575 	int *name = (int *)arg1 - 1;	/* XXX */
2576 	u_int namelen = arg2 + 1;	/* XXX */
2577 	struct vfsconf *vfsp;
2578 	struct xvfsconf xvfsp;
2579 
2580 	printf("WARNING: userland calling deprecated sysctl, "
2581 	    "please rebuild world\n");
2582 
2583 #if 1 || defined(COMPAT_PRELITE2)
2584 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
2585 	if (namelen == 1)
2586 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
2587 #endif
2588 
2589 	switch (name[1]) {
2590 	case VFS_MAXTYPENUM:
2591 		if (namelen != 2)
2592 			return (ENOTDIR);
2593 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
2594 	case VFS_CONF:
2595 		if (namelen != 3)
2596 			return (ENOTDIR);	/* overloaded */
2597 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list)
2598 			if (vfsp->vfc_typenum == name[2])
2599 				break;
2600 		if (vfsp == NULL)
2601 			return (EOPNOTSUPP);
2602 		vfsconf2x(vfsp, &xvfsp);
2603 		return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
2604 	}
2605 	return (EOPNOTSUPP);
2606 }
2607 
2608 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP,
2609 	vfs_sysctl, "Generic filesystem");
2610 
2611 #if 1 || defined(COMPAT_PRELITE2)
2612 
2613 static int
2614 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
2615 {
2616 	int error;
2617 	struct vfsconf *vfsp;
2618 	struct ovfsconf ovfs;
2619 
2620 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2621 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
2622 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
2623 		ovfs.vfc_index = vfsp->vfc_typenum;
2624 		ovfs.vfc_refcount = vfsp->vfc_refcount;
2625 		ovfs.vfc_flags = vfsp->vfc_flags;
2626 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
2627 		if (error)
2628 			return error;
2629 	}
2630 	return 0;
2631 }
2632 
2633 #endif /* 1 || COMPAT_PRELITE2 */
2634 #endif /* !BURN_BRIDGES */
2635 
2636 #define KINFO_VNODESLOP		10
2637 #ifdef notyet
2638 /*
2639  * Dump vnode list (via sysctl).
2640  */
2641 /* ARGSUSED */
2642 static int
2643 sysctl_vnode(SYSCTL_HANDLER_ARGS)
2644 {
2645 	struct xvnode *xvn;
2646 	struct thread *td = req->td;
2647 	struct mount *mp;
2648 	struct vnode *vp;
2649 	int error, len, n;
2650 
2651 	/*
2652 	 * Stale numvnodes access is not fatal here.
2653 	 */
2654 	req->lock = 0;
2655 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
2656 	if (!req->oldptr)
2657 		/* Make an estimate */
2658 		return (SYSCTL_OUT(req, 0, len));
2659 
2660 	error = sysctl_wire_old_buffer(req, 0);
2661 	if (error != 0)
2662 		return (error);
2663 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
2664 	n = 0;
2665 	mtx_lock(&mountlist_mtx);
2666 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
2667 		if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td))
2668 			continue;
2669 		MNT_ILOCK(mp);
2670 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2671 			if (n == len)
2672 				break;
2673 			vref(vp);
2674 			xvn[n].xv_size = sizeof *xvn;
2675 			xvn[n].xv_vnode = vp;
2676 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
2677 			XV_COPY(usecount);
2678 			XV_COPY(writecount);
2679 			XV_COPY(holdcnt);
2680 			XV_COPY(id);
2681 			XV_COPY(mount);
2682 			XV_COPY(numoutput);
2683 			XV_COPY(type);
2684 #undef XV_COPY
2685 			xvn[n].xv_flag = vp->v_vflag;
2686 
2687 			switch (vp->v_type) {
2688 			case VREG:
2689 			case VDIR:
2690 			case VLNK:
2691 				break;
2692 			case VBLK:
2693 			case VCHR:
2694 				if (vp->v_rdev == NULL) {
2695 					vrele(vp);
2696 					continue;
2697 				}
2698 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
2699 				break;
2700 			case VSOCK:
2701 				xvn[n].xv_socket = vp->v_socket;
2702 				break;
2703 			case VFIFO:
2704 				xvn[n].xv_fifo = vp->v_fifoinfo;
2705 				break;
2706 			case VNON:
2707 			case VBAD:
2708 			default:
2709 				/* shouldn't happen? */
2710 				vrele(vp);
2711 				continue;
2712 			}
2713 			vrele(vp);
2714 			++n;
2715 		}
2716 		MNT_IUNLOCK(mp);
2717 		mtx_lock(&mountlist_mtx);
2718 		vfs_unbusy(mp, td);
2719 		if (n == len)
2720 			break;
2721 	}
2722 	mtx_unlock(&mountlist_mtx);
2723 
2724 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
2725 	free(xvn, M_TEMP);
2726 	return (error);
2727 }
2728 
2729 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD,
2730 	0, 0, sysctl_vnode, "S,xvnode", "");
2731 #endif
2732 
2733 /*
2734  * Unmount all filesystems. The list is traversed in reverse order
2735  * of mounting to avoid dependencies.
2736  */
2737 void
2738 vfs_unmountall()
2739 {
2740 	struct mount *mp;
2741 	struct thread *td;
2742 	int error;
2743 
2744 	if (curthread != NULL)
2745 		td = curthread;
2746 	else
2747 		td = FIRST_THREAD_IN_PROC(initproc); /* XXX XXX proc0? */
2748 	/*
2749 	 * Since this only runs when rebooting, it is not interlocked.
2750 	 */
2751 	while(!TAILQ_EMPTY(&mountlist)) {
2752 		mp = TAILQ_LAST(&mountlist, mntlist);
2753 		error = dounmount(mp, MNT_FORCE, td);
2754 		if (error) {
2755 			TAILQ_REMOVE(&mountlist, mp, mnt_list);
2756 			printf("unmount of %s failed (",
2757 			    mp->mnt_stat.f_mntonname);
2758 			if (error == EBUSY)
2759 				printf("BUSY)\n");
2760 			else
2761 				printf("%d)\n", error);
2762 		} else {
2763 			/* The unmount has removed mp from the mountlist */
2764 		}
2765 	}
2766 }
2767 
2768 /*
2769  * perform msync on all vnodes under a mount point
2770  * the mount point must be locked.
2771  */
2772 void
2773 vfs_msync(struct mount *mp, int flags)
2774 {
2775 	struct vnode *vp, *nvp;
2776 	struct vm_object *obj;
2777 	int tries;
2778 
2779 	tries = 5;
2780 	MNT_ILOCK(mp);
2781 loop:
2782 	TAILQ_FOREACH_SAFE(vp, &mp->mnt_nvnodelist, v_nmntvnodes, nvp) {
2783 		if (vp->v_mount != mp) {
2784 			if (--tries > 0)
2785 				goto loop;
2786 			break;
2787 		}
2788 
2789 		VI_LOCK(vp);
2790 		if (vp->v_iflag & VI_XLOCK) {
2791 			VI_UNLOCK(vp);
2792 			continue;
2793 		}
2794 
2795 		if ((vp->v_iflag & VI_OBJDIRTY) &&
2796 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) {
2797 			MNT_IUNLOCK(mp);
2798 			if (!vget(vp,
2799 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
2800 			    curthread)) {
2801 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
2802 					vput(vp);
2803 					MNT_ILOCK(mp);
2804 					continue;
2805 				}
2806 
2807 				obj = vp->v_object;
2808 				if (obj != NULL) {
2809 					VM_OBJECT_LOCK(obj);
2810 					vm_object_page_clean(obj, 0, 0,
2811 					    flags == MNT_WAIT ?
2812 					    OBJPC_SYNC : OBJPC_NOSYNC);
2813 					VM_OBJECT_UNLOCK(obj);
2814 				}
2815 				vput(vp);
2816 			}
2817 			MNT_ILOCK(mp);
2818 			if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) {
2819 				if (--tries > 0)
2820 					goto loop;
2821 				break;
2822 			}
2823 		} else
2824 			VI_UNLOCK(vp);
2825 	}
2826 	MNT_IUNLOCK(mp);
2827 }
2828 
2829 /*
2830  * Mark a vnode as free, putting it up for recycling.
2831  */
2832 void
2833 vfree(struct vnode *vp)
2834 {
2835 
2836 	ASSERT_VI_LOCKED(vp, "vfree");
2837 	mtx_lock(&vnode_free_list_mtx);
2838 	KASSERT((vp->v_iflag & VI_FREE) == 0, ("vnode already free"));
2839 	if (vp->v_iflag & VI_AGE) {
2840 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2841 	} else {
2842 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
2843 	}
2844 	freevnodes++;
2845 	mtx_unlock(&vnode_free_list_mtx);
2846 	vp->v_iflag &= ~VI_AGE;
2847 	vp->v_iflag |= VI_FREE;
2848 }
2849 
2850 /*
2851  * Opposite of vfree() - mark a vnode as in use.
2852  */
2853 static void
2854 vbusy(struct vnode *vp)
2855 {
2856 
2857 	ASSERT_VI_LOCKED(vp, "vbusy");
2858 	KASSERT((vp->v_iflag & VI_FREE) != 0, ("vnode not free"));
2859 
2860 	mtx_lock(&vnode_free_list_mtx);
2861 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2862 	freevnodes--;
2863 	mtx_unlock(&vnode_free_list_mtx);
2864 
2865 	vp->v_iflag &= ~(VI_FREE|VI_AGE);
2866 }
2867 
2868 /*
2869  * Initalize per-vnode helper structure to hold poll-related state.
2870  */
2871 void
2872 v_addpollinfo(struct vnode *vp)
2873 {
2874 	struct vpollinfo *vi;
2875 
2876 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK);
2877 	if (vp->v_pollinfo != NULL) {
2878 		uma_zfree(vnodepoll_zone, vi);
2879 		return;
2880 	}
2881 	vp->v_pollinfo = vi;
2882 	mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
2883 	knlist_init(&vp->v_pollinfo->vpi_selinfo.si_note,
2884 	    &vp->v_pollinfo->vpi_lock);
2885 }
2886 
2887 /*
2888  * Record a process's interest in events which might happen to
2889  * a vnode.  Because poll uses the historic select-style interface
2890  * internally, this routine serves as both the ``check for any
2891  * pending events'' and the ``record my interest in future events''
2892  * functions.  (These are done together, while the lock is held,
2893  * to avoid race conditions.)
2894  */
2895 int
2896 vn_pollrecord(vp, td, events)
2897 	struct vnode *vp;
2898 	struct thread *td;
2899 	short events;
2900 {
2901 
2902 	if (vp->v_pollinfo == NULL)
2903 		v_addpollinfo(vp);
2904 	mtx_lock(&vp->v_pollinfo->vpi_lock);
2905 	if (vp->v_pollinfo->vpi_revents & events) {
2906 		/*
2907 		 * This leaves events we are not interested
2908 		 * in available for the other process which
2909 		 * which presumably had requested them
2910 		 * (otherwise they would never have been
2911 		 * recorded).
2912 		 */
2913 		events &= vp->v_pollinfo->vpi_revents;
2914 		vp->v_pollinfo->vpi_revents &= ~events;
2915 
2916 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
2917 		return events;
2918 	}
2919 	vp->v_pollinfo->vpi_events |= events;
2920 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
2921 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
2922 	return 0;
2923 }
2924 
2925 /*
2926  * Routine to create and manage a filesystem syncer vnode.
2927  */
2928 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
2929 static int	sync_fsync(struct  vop_fsync_args *);
2930 static int	sync_inactive(struct  vop_inactive_args *);
2931 static int	sync_reclaim(struct  vop_reclaim_args *);
2932 
2933 static struct vop_vector sync_vnodeops = {
2934 	.vop_bypass =	VOP_EOPNOTSUPP,
2935 	.vop_close =	sync_close,		/* close */
2936 	.vop_fsync =	sync_fsync,		/* fsync */
2937 	.vop_inactive =	sync_inactive,	/* inactive */
2938 	.vop_reclaim =	sync_reclaim,	/* reclaim */
2939 	.vop_lock =	vop_stdlock,	/* lock */
2940 	.vop_unlock =	vop_stdunlock,	/* unlock */
2941 	.vop_islocked =	vop_stdislocked,	/* islocked */
2942 };
2943 
2944 /*
2945  * Create a new filesystem syncer vnode for the specified mount point.
2946  */
2947 int
2948 vfs_allocate_syncvnode(mp)
2949 	struct mount *mp;
2950 {
2951 	struct vnode *vp;
2952 	static long start, incr, next;
2953 	int error;
2954 
2955 	/* Allocate a new vnode */
2956 	if ((error = getnewvnode("syncer", mp, &sync_vnodeops, &vp)) != 0) {
2957 		mp->mnt_syncer = NULL;
2958 		return (error);
2959 	}
2960 	vp->v_type = VNON;
2961 	/*
2962 	 * Place the vnode onto the syncer worklist. We attempt to
2963 	 * scatter them about on the list so that they will go off
2964 	 * at evenly distributed times even if all the filesystems
2965 	 * are mounted at once.
2966 	 */
2967 	next += incr;
2968 	if (next == 0 || next > syncer_maxdelay) {
2969 		start /= 2;
2970 		incr /= 2;
2971 		if (start == 0) {
2972 			start = syncer_maxdelay / 2;
2973 			incr = syncer_maxdelay;
2974 		}
2975 		next = start;
2976 	}
2977 	VI_LOCK(vp);
2978 	vn_syncer_add_to_worklist(&vp->v_bufobj,
2979 	    syncdelay > 0 ? next % syncdelay : 0);
2980 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
2981 	mtx_lock(&sync_mtx);
2982 	sync_vnode_count++;
2983 	mtx_unlock(&sync_mtx);
2984 	VI_UNLOCK(vp);
2985 	mp->mnt_syncer = vp;
2986 	return (0);
2987 }
2988 
2989 /*
2990  * Do a lazy sync of the filesystem.
2991  */
2992 static int
2993 sync_fsync(ap)
2994 	struct vop_fsync_args /* {
2995 		struct vnode *a_vp;
2996 		struct ucred *a_cred;
2997 		int a_waitfor;
2998 		struct thread *a_td;
2999 	} */ *ap;
3000 {
3001 	struct vnode *syncvp = ap->a_vp;
3002 	struct mount *mp = syncvp->v_mount;
3003 	struct thread *td = ap->a_td;
3004 	int error, asyncflag;
3005 	struct bufobj *bo;
3006 
3007 	/*
3008 	 * We only need to do something if this is a lazy evaluation.
3009 	 */
3010 	if (ap->a_waitfor != MNT_LAZY)
3011 		return (0);
3012 
3013 	/*
3014 	 * Move ourselves to the back of the sync list.
3015 	 */
3016 	bo = &syncvp->v_bufobj;
3017 	BO_LOCK(bo);
3018 	vn_syncer_add_to_worklist(bo, syncdelay);
3019 	BO_UNLOCK(bo);
3020 
3021 	/*
3022 	 * Walk the list of vnodes pushing all that are dirty and
3023 	 * not already on the sync list.
3024 	 */
3025 	mtx_lock(&mountlist_mtx);
3026 	if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) {
3027 		mtx_unlock(&mountlist_mtx);
3028 		return (0);
3029 	}
3030 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
3031 		vfs_unbusy(mp, td);
3032 		return (0);
3033 	}
3034 	asyncflag = mp->mnt_flag & MNT_ASYNC;
3035 	mp->mnt_flag &= ~MNT_ASYNC;
3036 	vfs_msync(mp, MNT_NOWAIT);
3037 	error = VFS_SYNC(mp, MNT_LAZY, td);
3038 	if (asyncflag)
3039 		mp->mnt_flag |= MNT_ASYNC;
3040 	vn_finished_write(mp);
3041 	vfs_unbusy(mp, td);
3042 	return (error);
3043 }
3044 
3045 /*
3046  * The syncer vnode is no referenced.
3047  */
3048 static int
3049 sync_inactive(ap)
3050 	struct vop_inactive_args /* {
3051 		struct vnode *a_vp;
3052 		struct thread *a_td;
3053 	} */ *ap;
3054 {
3055 
3056 	VOP_UNLOCK(ap->a_vp, 0, ap->a_td);
3057 	vgone(ap->a_vp);
3058 	return (0);
3059 }
3060 
3061 /*
3062  * The syncer vnode is no longer needed and is being decommissioned.
3063  *
3064  * Modifications to the worklist must be protected by sync_mtx.
3065  */
3066 static int
3067 sync_reclaim(ap)
3068 	struct vop_reclaim_args /* {
3069 		struct vnode *a_vp;
3070 	} */ *ap;
3071 {
3072 	struct vnode *vp = ap->a_vp;
3073 	struct bufobj *bo;
3074 
3075 	VI_LOCK(vp);
3076 	bo = &vp->v_bufobj;
3077 	vp->v_mount->mnt_syncer = NULL;
3078 	if (bo->bo_flag & BO_ONWORKLST) {
3079 		mtx_lock(&sync_mtx);
3080 		LIST_REMOVE(bo, bo_synclist);
3081  		syncer_worklist_len--;
3082 		sync_vnode_count--;
3083 		mtx_unlock(&sync_mtx);
3084 		bo->bo_flag &= ~BO_ONWORKLST;
3085 	}
3086 	VI_UNLOCK(vp);
3087 
3088 	return (0);
3089 }
3090 
3091 /*
3092  * Check if vnode represents a disk device
3093  */
3094 int
3095 vn_isdisk(vp, errp)
3096 	struct vnode *vp;
3097 	int *errp;
3098 {
3099 	int error;
3100 
3101 	error = 0;
3102 	dev_lock();
3103 	if (vp->v_type != VCHR)
3104 		error = ENOTBLK;
3105 	else if (vp->v_rdev == NULL)
3106 		error = ENXIO;
3107 	else if (vp->v_rdev->si_devsw == NULL)
3108 		error = ENXIO;
3109 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
3110 		error = ENOTBLK;
3111 	dev_unlock();
3112 	if (errp != NULL)
3113 		*errp = error;
3114 	return (error == 0);
3115 }
3116 
3117 /*
3118  * Free data allocated by namei(); see namei(9) for details.
3119  */
3120 void
3121 NDFREE(ndp, flags)
3122      struct nameidata *ndp;
3123      const u_int flags;
3124 {
3125 
3126 	if (!(flags & NDF_NO_FREE_PNBUF) &&
3127 	    (ndp->ni_cnd.cn_flags & HASBUF)) {
3128 		uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf);
3129 		ndp->ni_cnd.cn_flags &= ~HASBUF;
3130 	}
3131 	if (!(flags & NDF_NO_DVP_UNLOCK) &&
3132 	    (ndp->ni_cnd.cn_flags & LOCKPARENT) &&
3133 	    ndp->ni_dvp != ndp->ni_vp)
3134 		VOP_UNLOCK(ndp->ni_dvp, 0, ndp->ni_cnd.cn_thread);
3135 	if (!(flags & NDF_NO_DVP_RELE) &&
3136 	    (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) {
3137 		vrele(ndp->ni_dvp);
3138 		ndp->ni_dvp = NULL;
3139 	}
3140 	if (!(flags & NDF_NO_VP_UNLOCK) &&
3141 	    (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp)
3142 		VOP_UNLOCK(ndp->ni_vp, 0, ndp->ni_cnd.cn_thread);
3143 	if (!(flags & NDF_NO_VP_RELE) &&
3144 	    ndp->ni_vp) {
3145 		vrele(ndp->ni_vp);
3146 		ndp->ni_vp = NULL;
3147 	}
3148 	if (!(flags & NDF_NO_STARTDIR_RELE) &&
3149 	    (ndp->ni_cnd.cn_flags & SAVESTART)) {
3150 		vrele(ndp->ni_startdir);
3151 		ndp->ni_startdir = NULL;
3152 	}
3153 }
3154 
3155 /*
3156  * Common filesystem object access control check routine.  Accepts a
3157  * vnode's type, "mode", uid and gid, requested access mode, credentials,
3158  * and optional call-by-reference privused argument allowing vaccess()
3159  * to indicate to the caller whether privilege was used to satisfy the
3160  * request (obsoleted).  Returns 0 on success, or an errno on failure.
3161  */
3162 int
3163 vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused)
3164 	enum vtype type;
3165 	mode_t file_mode;
3166 	uid_t file_uid;
3167 	gid_t file_gid;
3168 	mode_t acc_mode;
3169 	struct ucred *cred;
3170 	int *privused;
3171 {
3172 	mode_t dac_granted;
3173 #ifdef CAPABILITIES
3174 	mode_t cap_granted;
3175 #endif
3176 
3177 	/*
3178 	 * Look for a normal, non-privileged way to access the file/directory
3179 	 * as requested.  If it exists, go with that.
3180 	 */
3181 
3182 	if (privused != NULL)
3183 		*privused = 0;
3184 
3185 	dac_granted = 0;
3186 
3187 	/* Check the owner. */
3188 	if (cred->cr_uid == file_uid) {
3189 		dac_granted |= VADMIN;
3190 		if (file_mode & S_IXUSR)
3191 			dac_granted |= VEXEC;
3192 		if (file_mode & S_IRUSR)
3193 			dac_granted |= VREAD;
3194 		if (file_mode & S_IWUSR)
3195 			dac_granted |= (VWRITE | VAPPEND);
3196 
3197 		if ((acc_mode & dac_granted) == acc_mode)
3198 			return (0);
3199 
3200 		goto privcheck;
3201 	}
3202 
3203 	/* Otherwise, check the groups (first match) */
3204 	if (groupmember(file_gid, cred)) {
3205 		if (file_mode & S_IXGRP)
3206 			dac_granted |= VEXEC;
3207 		if (file_mode & S_IRGRP)
3208 			dac_granted |= VREAD;
3209 		if (file_mode & S_IWGRP)
3210 			dac_granted |= (VWRITE | VAPPEND);
3211 
3212 		if ((acc_mode & dac_granted) == acc_mode)
3213 			return (0);
3214 
3215 		goto privcheck;
3216 	}
3217 
3218 	/* Otherwise, check everyone else. */
3219 	if (file_mode & S_IXOTH)
3220 		dac_granted |= VEXEC;
3221 	if (file_mode & S_IROTH)
3222 		dac_granted |= VREAD;
3223 	if (file_mode & S_IWOTH)
3224 		dac_granted |= (VWRITE | VAPPEND);
3225 	if ((acc_mode & dac_granted) == acc_mode)
3226 		return (0);
3227 
3228 privcheck:
3229 	if (!suser_cred(cred, SUSER_ALLOWJAIL)) {
3230 		/* XXX audit: privilege used */
3231 		if (privused != NULL)
3232 			*privused = 1;
3233 		return (0);
3234 	}
3235 
3236 #ifdef CAPABILITIES
3237 	/*
3238 	 * Build a capability mask to determine if the set of capabilities
3239 	 * satisfies the requirements when combined with the granted mask
3240 	 * from above.
3241 	 * For each capability, if the capability is required, bitwise
3242 	 * or the request type onto the cap_granted mask.
3243 	 */
3244 	cap_granted = 0;
3245 
3246 	if (type == VDIR) {
3247 		/*
3248 		 * For directories, use CAP_DAC_READ_SEARCH to satisfy
3249 		 * VEXEC requests, instead of CAP_DAC_EXECUTE.
3250 		 */
3251 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3252 		    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL))
3253 			cap_granted |= VEXEC;
3254 	} else {
3255 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3256 		    !cap_check(cred, NULL, CAP_DAC_EXECUTE, SUSER_ALLOWJAIL))
3257 			cap_granted |= VEXEC;
3258 	}
3259 
3260 	if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) &&
3261 	    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL))
3262 		cap_granted |= VREAD;
3263 
3264 	if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
3265 	    !cap_check(cred, NULL, CAP_DAC_WRITE, SUSER_ALLOWJAIL))
3266 		cap_granted |= (VWRITE | VAPPEND);
3267 
3268 	if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
3269 	    !cap_check(cred, NULL, CAP_FOWNER, SUSER_ALLOWJAIL))
3270 		cap_granted |= VADMIN;
3271 
3272 	if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) {
3273 		/* XXX audit: privilege used */
3274 		if (privused != NULL)
3275 			*privused = 1;
3276 		return (0);
3277 	}
3278 #endif
3279 
3280 	return ((acc_mode & VADMIN) ? EPERM : EACCES);
3281 }
3282 
3283 /*
3284  * Credential check based on process requesting service, and per-attribute
3285  * permissions.
3286  */
3287 int
3288 extattr_check_cred(struct vnode *vp, int attrnamespace,
3289     struct ucred *cred, struct thread *td, int access)
3290 {
3291 
3292 	/*
3293 	 * Kernel-invoked always succeeds.
3294 	 */
3295 	if (cred == NOCRED)
3296 		return (0);
3297 
3298 	/*
3299 	 * Do not allow privileged processes in jail to directly
3300 	 * manipulate system attributes.
3301 	 *
3302 	 * XXX What capability should apply here?
3303 	 * Probably CAP_SYS_SETFFLAG.
3304 	 */
3305 	switch (attrnamespace) {
3306 	case EXTATTR_NAMESPACE_SYSTEM:
3307 		/* Potentially should be: return (EPERM); */
3308 		return (suser_cred(cred, 0));
3309 	case EXTATTR_NAMESPACE_USER:
3310 		return (VOP_ACCESS(vp, access, cred, td));
3311 	default:
3312 		return (EPERM);
3313 	}
3314 }
3315 
3316 #ifdef DEBUG_VFS_LOCKS
3317 /*
3318  * This only exists to supress warnings from unlocked specfs accesses.  It is
3319  * no longer ok to have an unlocked VFS.
3320  */
3321 #define	IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD)
3322 
3323 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
3324 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, "");
3325 
3326 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
3327 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, "");
3328 
3329 int vfs_badlock_print = 1;	/* Print lock violations. */
3330 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, "");
3331 
3332 #ifdef KDB
3333 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
3334 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, "");
3335 #endif
3336 
3337 static void
3338 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
3339 {
3340 
3341 #ifdef KDB
3342 	if (vfs_badlock_backtrace)
3343 		kdb_backtrace();
3344 #endif
3345 	if (vfs_badlock_print)
3346 		printf("%s: %p %s\n", str, (void *)vp, msg);
3347 	if (vfs_badlock_ddb)
3348 		kdb_enter("lock violation");
3349 }
3350 
3351 void
3352 assert_vi_locked(struct vnode *vp, const char *str)
3353 {
3354 
3355 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
3356 		vfs_badlock("interlock is not locked but should be", str, vp);
3357 }
3358 
3359 void
3360 assert_vi_unlocked(struct vnode *vp, const char *str)
3361 {
3362 
3363 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
3364 		vfs_badlock("interlock is locked but should not be", str, vp);
3365 }
3366 
3367 void
3368 assert_vop_locked(struct vnode *vp, const char *str)
3369 {
3370 
3371 	if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0)
3372 		vfs_badlock("is not locked but should be", str, vp);
3373 }
3374 
3375 void
3376 assert_vop_unlocked(struct vnode *vp, const char *str)
3377 {
3378 
3379 	if (vp && !IGNORE_LOCK(vp) &&
3380 	    VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE)
3381 		vfs_badlock("is locked but should not be", str, vp);
3382 }
3383 
3384 #if 0
3385 void
3386 assert_vop_elocked(struct vnode *vp, const char *str)
3387 {
3388 
3389 	if (vp && !IGNORE_LOCK(vp) &&
3390 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE)
3391 		vfs_badlock("is not exclusive locked but should be", str, vp);
3392 }
3393 
3394 void
3395 assert_vop_elocked_other(struct vnode *vp, const char *str)
3396 {
3397 
3398 	if (vp && !IGNORE_LOCK(vp) &&
3399 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER)
3400 		vfs_badlock("is not exclusive locked by another thread",
3401 		    str, vp);
3402 }
3403 
3404 void
3405 assert_vop_slocked(struct vnode *vp, const char *str)
3406 {
3407 
3408 	if (vp && !IGNORE_LOCK(vp) &&
3409 	    VOP_ISLOCKED(vp, curthread) != LK_SHARED)
3410 		vfs_badlock("is not locked shared but should be", str, vp);
3411 }
3412 #endif /* 0 */
3413 
3414 void
3415 vop_rename_pre(void *ap)
3416 {
3417 	struct vop_rename_args *a = ap;
3418 
3419 	if (a->a_tvp)
3420 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
3421 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
3422 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
3423 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
3424 
3425 	/* Check the source (from). */
3426 	if (a->a_tdvp != a->a_fdvp)
3427 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
3428 	if (a->a_tvp != a->a_fvp)
3429 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked");
3430 
3431 	/* Check the target. */
3432 	if (a->a_tvp)
3433 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
3434 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
3435 }
3436 
3437 void
3438 vop_strategy_pre(void *ap)
3439 {
3440 	struct vop_strategy_args *a;
3441 	struct buf *bp;
3442 
3443 	a = ap;
3444 	bp = a->a_bp;
3445 
3446 	/*
3447 	 * Cluster ops lock their component buffers but not the IO container.
3448 	 */
3449 	if ((bp->b_flags & B_CLUSTER) != 0)
3450 		return;
3451 
3452 	if (BUF_REFCNT(bp) < 1) {
3453 		if (vfs_badlock_print)
3454 			printf(
3455 			    "VOP_STRATEGY: bp is not locked but should be\n");
3456 		if (vfs_badlock_ddb)
3457 			kdb_enter("lock violation");
3458 	}
3459 }
3460 
3461 void
3462 vop_lookup_pre(void *ap)
3463 {
3464 	struct vop_lookup_args *a;
3465 	struct vnode *dvp;
3466 
3467 	a = ap;
3468 	dvp = a->a_dvp;
3469 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3470 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3471 }
3472 
3473 void
3474 vop_lookup_post(void *ap, int rc)
3475 {
3476 	struct vop_lookup_args *a;
3477 	struct componentname *cnp;
3478 	struct vnode *dvp;
3479 	struct vnode *vp;
3480 	int flags;
3481 
3482 	a = ap;
3483 	dvp = a->a_dvp;
3484 	cnp = a->a_cnp;
3485 	vp = *(a->a_vpp);
3486 	flags = cnp->cn_flags;
3487 
3488 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3489 
3490 	/*
3491 	 * If this is the last path component for this lookup and LOCKPARENT
3492 	 * is set, OR if there is an error the directory has to be locked.
3493 	 */
3494 	if ((flags & LOCKPARENT) && (flags & ISLASTCN))
3495 		ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (LOCKPARENT)");
3496 	else if (rc != 0)
3497 		ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (error)");
3498 	else if (dvp != vp)
3499 		ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (dvp)");
3500 	if (flags & PDIRUNLOCK)
3501 		ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (PDIRUNLOCK)");
3502 }
3503 
3504 void
3505 vop_lock_pre(void *ap)
3506 {
3507 	struct vop_lock_args *a = ap;
3508 
3509 	if ((a->a_flags & LK_INTERLOCK) == 0)
3510 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3511 	else
3512 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
3513 }
3514 
3515 void
3516 vop_lock_post(void *ap, int rc)
3517 {
3518 	struct vop_lock_args *a = ap;
3519 
3520 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3521 	if (rc == 0)
3522 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
3523 }
3524 
3525 void
3526 vop_unlock_pre(void *ap)
3527 {
3528 	struct vop_unlock_args *a = ap;
3529 
3530 	if (a->a_flags & LK_INTERLOCK)
3531 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
3532 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
3533 }
3534 
3535 void
3536 vop_unlock_post(void *ap, int rc)
3537 {
3538 	struct vop_unlock_args *a = ap;
3539 
3540 	if (a->a_flags & LK_INTERLOCK)
3541 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
3542 }
3543 #endif /* DEBUG_VFS_LOCKS */
3544 
3545 static struct knlist fs_knlist;
3546 
3547 static void
3548 vfs_event_init(void *arg)
3549 {
3550 	knlist_init(&fs_knlist, NULL);
3551 }
3552 /* XXX - correct order? */
3553 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
3554 
3555 void
3556 vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused)
3557 {
3558 
3559 	KNOTE_UNLOCKED(&fs_knlist, event);
3560 }
3561 
3562 static int	filt_fsattach(struct knote *kn);
3563 static void	filt_fsdetach(struct knote *kn);
3564 static int	filt_fsevent(struct knote *kn, long hint);
3565 
3566 struct filterops fs_filtops =
3567 	{ 0, filt_fsattach, filt_fsdetach, filt_fsevent };
3568 
3569 static int
3570 filt_fsattach(struct knote *kn)
3571 {
3572 
3573 	kn->kn_flags |= EV_CLEAR;
3574 	knlist_add(&fs_knlist, kn, 0);
3575 	return (0);
3576 }
3577 
3578 static void
3579 filt_fsdetach(struct knote *kn)
3580 {
3581 
3582 	knlist_remove(&fs_knlist, kn, 0);
3583 }
3584 
3585 static int
3586 filt_fsevent(struct knote *kn, long hint)
3587 {
3588 
3589 	kn->kn_fflags |= hint;
3590 	return (kn->kn_fflags != 0);
3591 }
3592 
3593 static int
3594 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
3595 {
3596 	struct vfsidctl vc;
3597 	int error;
3598 	struct mount *mp;
3599 
3600 	error = SYSCTL_IN(req, &vc, sizeof(vc));
3601 	if (error)
3602 		return (error);
3603 	if (vc.vc_vers != VFS_CTL_VERS1)
3604 		return (EINVAL);
3605 	mp = vfs_getvfs(&vc.vc_fsid);
3606 	if (mp == NULL)
3607 		return (ENOENT);
3608 	/* ensure that a specific sysctl goes to the right filesystem. */
3609 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
3610 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
3611 		return (EINVAL);
3612 	}
3613 	VCTLTOREQ(&vc, req);
3614 	return (VFS_SYSCTL(mp, vc.vc_op, req));
3615 }
3616 
3617 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR,
3618         NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid");
3619 
3620 /*
3621  * Function to initialize a va_filerev field sensibly.
3622  * XXX: Wouldn't a random number make a lot more sense ??
3623  */
3624 u_quad_t
3625 init_va_filerev(void)
3626 {
3627 	struct bintime bt;
3628 
3629 	getbinuptime(&bt);
3630 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
3631 }
3632