xref: /freebsd/sys/kern/vfs_subr.c (revision 71651a2743acfa3756ab1e7c3983e6861c6fada1)
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 bufobj *bo,
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 #if defined(__alpha__) || defined(__amd64__) || defined(__i386__)
102 int mpsafe_vfs = 1;
103 #else
104 int mpsafe_vfs;
105 #endif
106 TUNABLE_INT("debug.mpsafevfs", &mpsafe_vfs);
107 SYSCTL_INT(_debug, OID_AUTO, mpsafevfs, CTLFLAG_RD, &mpsafe_vfs, 0,
108     "MPSAFE VFS");
109 
110 /*
111  * Number of vnodes in existence.  Increased whenever getnewvnode()
112  * allocates a new vnode, never decreased.
113  */
114 static unsigned long	numvnodes;
115 
116 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, "");
117 
118 /*
119  * Conversion tables for conversion from vnode types to inode formats
120  * and back.
121  */
122 enum vtype iftovt_tab[16] = {
123 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
124 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
125 };
126 int vttoif_tab[9] = {
127 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
128 	S_IFSOCK, S_IFIFO, S_IFMT,
129 };
130 
131 /*
132  * List of vnodes that are ready for recycling.
133  */
134 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
135 
136 /*
137  * Minimum number of free vnodes.  If there are fewer than this free vnodes,
138  * getnewvnode() will return a newly allocated vnode.
139  */
140 static u_long wantfreevnodes = 25;
141 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, "");
142 /* Number of vnodes in the free list. */
143 static u_long freevnodes;
144 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, "");
145 
146 /*
147  * Various variables used for debugging the new implementation of
148  * reassignbuf().
149  * XXX these are probably of (very) limited utility now.
150  */
151 static int reassignbufcalls;
152 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, "");
153 static int nameileafonly;
154 SYSCTL_INT(_vfs, OID_AUTO, nameileafonly, CTLFLAG_RW, &nameileafonly, 0, "");
155 
156 /*
157  * Cache for the mount type id assigned to NFS.  This is used for
158  * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c.
159  */
160 int	nfs_mount_type = -1;
161 
162 /* To keep more than one thread at a time from running vfs_getnewfsid */
163 static struct mtx mntid_mtx;
164 
165 /*
166  * Lock for any access to the following:
167  *	vnode_free_list
168  *	numvnodes
169  *	freevnodes
170  */
171 static struct mtx vnode_free_list_mtx;
172 
173 /* Publicly exported FS */
174 struct nfs_public nfs_pub;
175 
176 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
177 static uma_zone_t vnode_zone;
178 static uma_zone_t vnodepoll_zone;
179 
180 /* Set to 1 to print out reclaim of active vnodes */
181 int	prtactive;
182 
183 /*
184  * The workitem queue.
185  *
186  * It is useful to delay writes of file data and filesystem metadata
187  * for tens of seconds so that quickly created and deleted files need
188  * not waste disk bandwidth being created and removed. To realize this,
189  * we append vnodes to a "workitem" queue. When running with a soft
190  * updates implementation, most pending metadata dependencies should
191  * not wait for more than a few seconds. Thus, mounted on block devices
192  * are delayed only about a half the time that file data is delayed.
193  * Similarly, directory updates are more critical, so are only delayed
194  * about a third the time that file data is delayed. Thus, there are
195  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
196  * one each second (driven off the filesystem syncer process). The
197  * syncer_delayno variable indicates the next queue that is to be processed.
198  * Items that need to be processed soon are placed in this queue:
199  *
200  *	syncer_workitem_pending[syncer_delayno]
201  *
202  * A delay of fifteen seconds is done by placing the request fifteen
203  * entries later in the queue:
204  *
205  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
206  *
207  */
208 static int syncer_delayno;
209 static long syncer_mask;
210 LIST_HEAD(synclist, bufobj);
211 static struct synclist *syncer_workitem_pending;
212 /*
213  * The sync_mtx protects:
214  *	bo->bo_synclist
215  *	sync_vnode_count
216  *	syncer_delayno
217  *	syncer_state
218  *	syncer_workitem_pending
219  *	syncer_worklist_len
220  *	rushjob
221  */
222 static struct mtx sync_mtx;
223 
224 #define SYNCER_MAXDELAY		32
225 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
226 static int syncdelay = 30;		/* max time to delay syncing data */
227 static int filedelay = 30;		/* time to delay syncing files */
228 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, "");
229 static int dirdelay = 29;		/* time to delay syncing directories */
230 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, "");
231 static int metadelay = 28;		/* time to delay syncing metadata */
232 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, "");
233 static int rushjob;		/* number of slots to run ASAP */
234 static int stat_rush_requests;	/* number of times I/O speeded up */
235 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, "");
236 
237 /*
238  * When shutting down the syncer, run it at four times normal speed.
239  */
240 #define SYNCER_SHUTDOWN_SPEEDUP		4
241 static int sync_vnode_count;
242 static int syncer_worklist_len;
243 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
244     syncer_state;
245 
246 /*
247  * Number of vnodes we want to exist at any one time.  This is mostly used
248  * to size hash tables in vnode-related code.  It is normally not used in
249  * getnewvnode(), as wantfreevnodes is normally nonzero.)
250  *
251  * XXX desiredvnodes is historical cruft and should not exist.
252  */
253 int desiredvnodes;
254 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW,
255     &desiredvnodes, 0, "Maximum number of vnodes");
256 static int minvnodes;
257 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
258     &minvnodes, 0, "Minimum number of vnodes");
259 static int vnlru_nowhere;
260 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
261     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
262 
263 /* Hook for calling soft updates. */
264 int (*softdep_process_worklist_hook)(struct mount *);
265 
266 /*
267  * Initialize the vnode management data structures.
268  */
269 #ifndef	MAXVNODES_MAX
270 #define	MAXVNODES_MAX	100000
271 #endif
272 static void
273 vntblinit(void *dummy __unused)
274 {
275 
276 	/*
277 	 * Desiredvnodes is a function of the physical memory size and
278 	 * the kernel's heap size.  Specifically, desiredvnodes scales
279 	 * in proportion to the physical memory size until two fifths
280 	 * of the kernel's heap size is consumed by vnodes and vm
281 	 * objects.
282 	 */
283 	desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size /
284 	    (5 * (sizeof(struct vm_object) + sizeof(struct vnode))));
285 	if (desiredvnodes > MAXVNODES_MAX) {
286 		if (bootverbose)
287 			printf("Reducing kern.maxvnodes %d -> %d\n",
288 			    desiredvnodes, MAXVNODES_MAX);
289 		desiredvnodes = MAXVNODES_MAX;
290 	}
291 	minvnodes = desiredvnodes / 4;
292 	mtx_init(&mountlist_mtx, "mountlist", NULL, MTX_DEF);
293 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
294 	TAILQ_INIT(&vnode_free_list);
295 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
296 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
297 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
298 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
299 	      NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
300 	/*
301 	 * Initialize the filesystem syncer.
302 	 */
303 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
304 		&syncer_mask);
305 	syncer_maxdelay = syncer_mask + 1;
306 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
307 }
308 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL)
309 
310 
311 /*
312  * Mark a mount point as busy. Used to synchronize access and to delay
313  * unmounting. Interlock is not released on failure.
314  */
315 int
316 vfs_busy(mp, flags, interlkp, td)
317 	struct mount *mp;
318 	int flags;
319 	struct mtx *interlkp;
320 	struct thread *td;
321 {
322 	int lkflags;
323 
324 	MNT_ILOCK(mp);
325 	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
326 		if (flags & LK_NOWAIT) {
327 			MNT_IUNLOCK(mp);
328 			return (ENOENT);
329 		}
330 		if (interlkp)
331 			mtx_unlock(interlkp);
332 		mp->mnt_kern_flag |= MNTK_MWAIT;
333 		/*
334 		 * Since all busy locks are shared except the exclusive
335 		 * lock granted when unmounting, the only place that a
336 		 * wakeup needs to be done is at the release of the
337 		 * exclusive lock at the end of dounmount.
338 		 */
339 		msleep(mp, MNT_MTX(mp), PVFS|PDROP, "vfs_busy", 0);
340 		if (interlkp)
341 			mtx_lock(interlkp);
342 		return (ENOENT);
343 	}
344 	if (interlkp)
345 		mtx_unlock(interlkp);
346 	lkflags = LK_SHARED | LK_NOPAUSE | LK_INTERLOCK;
347 	if (lockmgr(&mp->mnt_lock, lkflags, MNT_MTX(mp), td))
348 		panic("vfs_busy: unexpected lock failure");
349 	return (0);
350 }
351 
352 /*
353  * Free a busy filesystem.
354  */
355 void
356 vfs_unbusy(mp, td)
357 	struct mount *mp;
358 	struct thread *td;
359 {
360 
361 	lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td);
362 }
363 
364 /*
365  * Lookup a mount point by filesystem identifier.
366  */
367 struct mount *
368 vfs_getvfs(fsid)
369 	fsid_t *fsid;
370 {
371 	struct mount *mp;
372 
373 	mtx_lock(&mountlist_mtx);
374 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
375 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
376 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
377 			mtx_unlock(&mountlist_mtx);
378 			return (mp);
379 		}
380 	}
381 	mtx_unlock(&mountlist_mtx);
382 	return ((struct mount *) 0);
383 }
384 
385 /*
386  * Check if a user can access priveledged mount options.
387  */
388 int
389 vfs_suser(struct mount *mp, struct thread *td)
390 {
391 	int error;
392 
393 	if ((mp->mnt_flag & MNT_USER) == 0 ||
394 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
395 		if ((error = suser(td)) != 0)
396 			return (error);
397 	}
398 	return (0);
399 }
400 
401 /*
402  * Get a new unique fsid.  Try to make its val[0] unique, since this value
403  * will be used to create fake device numbers for stat().  Also try (but
404  * not so hard) make its val[0] unique mod 2^16, since some emulators only
405  * support 16-bit device numbers.  We end up with unique val[0]'s for the
406  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
407  *
408  * Keep in mind that several mounts may be running in parallel.  Starting
409  * the search one past where the previous search terminated is both a
410  * micro-optimization and a defense against returning the same fsid to
411  * different mounts.
412  */
413 void
414 vfs_getnewfsid(mp)
415 	struct mount *mp;
416 {
417 	static u_int16_t mntid_base;
418 	fsid_t tfsid;
419 	int mtype;
420 
421 	mtx_lock(&mntid_mtx);
422 	mtype = mp->mnt_vfc->vfc_typenum;
423 	tfsid.val[1] = mtype;
424 	mtype = (mtype & 0xFF) << 24;
425 	for (;;) {
426 		tfsid.val[0] = makedev(255,
427 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
428 		mntid_base++;
429 		if (vfs_getvfs(&tfsid) == NULL)
430 			break;
431 	}
432 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
433 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
434 	mtx_unlock(&mntid_mtx);
435 }
436 
437 /*
438  * Knob to control the precision of file timestamps:
439  *
440  *   0 = seconds only; nanoseconds zeroed.
441  *   1 = seconds and nanoseconds, accurate within 1/HZ.
442  *   2 = seconds and nanoseconds, truncated to microseconds.
443  * >=3 = seconds and nanoseconds, maximum precision.
444  */
445 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
446 
447 static int timestamp_precision = TSP_SEC;
448 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
449     &timestamp_precision, 0, "");
450 
451 /*
452  * Get a current timestamp.
453  */
454 void
455 vfs_timestamp(tsp)
456 	struct timespec *tsp;
457 {
458 	struct timeval tv;
459 
460 	switch (timestamp_precision) {
461 	case TSP_SEC:
462 		tsp->tv_sec = time_second;
463 		tsp->tv_nsec = 0;
464 		break;
465 	case TSP_HZ:
466 		getnanotime(tsp);
467 		break;
468 	case TSP_USEC:
469 		microtime(&tv);
470 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
471 		break;
472 	case TSP_NSEC:
473 	default:
474 		nanotime(tsp);
475 		break;
476 	}
477 }
478 
479 /*
480  * Set vnode attributes to VNOVAL
481  */
482 void
483 vattr_null(vap)
484 	struct vattr *vap;
485 {
486 
487 	vap->va_type = VNON;
488 	vap->va_size = VNOVAL;
489 	vap->va_bytes = VNOVAL;
490 	vap->va_mode = VNOVAL;
491 	vap->va_nlink = VNOVAL;
492 	vap->va_uid = VNOVAL;
493 	vap->va_gid = VNOVAL;
494 	vap->va_fsid = VNOVAL;
495 	vap->va_fileid = VNOVAL;
496 	vap->va_blocksize = VNOVAL;
497 	vap->va_rdev = VNOVAL;
498 	vap->va_atime.tv_sec = VNOVAL;
499 	vap->va_atime.tv_nsec = VNOVAL;
500 	vap->va_mtime.tv_sec = VNOVAL;
501 	vap->va_mtime.tv_nsec = VNOVAL;
502 	vap->va_ctime.tv_sec = VNOVAL;
503 	vap->va_ctime.tv_nsec = VNOVAL;
504 	vap->va_birthtime.tv_sec = VNOVAL;
505 	vap->va_birthtime.tv_nsec = VNOVAL;
506 	vap->va_flags = VNOVAL;
507 	vap->va_gen = VNOVAL;
508 	vap->va_vaflags = 0;
509 }
510 
511 /*
512  * This routine is called when we have too many vnodes.  It attempts
513  * to free <count> vnodes and will potentially free vnodes that still
514  * have VM backing store (VM backing store is typically the cause
515  * of a vnode blowout so we want to do this).  Therefore, this operation
516  * is not considered cheap.
517  *
518  * A number of conditions may prevent a vnode from being reclaimed.
519  * the buffer cache may have references on the vnode, a directory
520  * vnode may still have references due to the namei cache representing
521  * underlying files, or the vnode may be in active use.   It is not
522  * desireable to reuse such vnodes.  These conditions may cause the
523  * number of vnodes to reach some minimum value regardless of what
524  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
525  */
526 static int
527 vlrureclaim(struct mount *mp)
528 {
529 	struct vnode *vp;
530 	int done;
531 	int trigger;
532 	int usevnodes;
533 	int count;
534 
535 	/*
536 	 * Calculate the trigger point, don't allow user
537 	 * screwups to blow us up.   This prevents us from
538 	 * recycling vnodes with lots of resident pages.  We
539 	 * aren't trying to free memory, we are trying to
540 	 * free vnodes.
541 	 */
542 	usevnodes = desiredvnodes;
543 	if (usevnodes <= 0)
544 		usevnodes = 1;
545 	trigger = cnt.v_page_count * 2 / usevnodes;
546 
547 	done = 0;
548 	MNT_ILOCK(mp);
549 	count = mp->mnt_nvnodelistsize / 10 + 1;
550 	while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) {
551 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
552 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
553 
554 		if (vp->v_type != VNON &&
555 		    vp->v_type != VBAD &&
556 		    VI_TRYLOCK(vp)) {
557 			if (VMIGHTFREE(vp) &&           /* critical path opt */
558 			    (vp->v_object == NULL ||
559 			    vp->v_object->resident_page_count < trigger)) {
560 				MNT_IUNLOCK(mp);
561 				vgonel(vp, curthread);
562 				done++;
563 				MNT_ILOCK(mp);
564 			} else
565 				VI_UNLOCK(vp);
566 		}
567 		--count;
568 	}
569 	MNT_IUNLOCK(mp);
570 	return done;
571 }
572 
573 /*
574  * Attempt to recycle vnodes in a context that is always safe to block.
575  * Calling vlrurecycle() from the bowels of filesystem code has some
576  * interesting deadlock problems.
577  */
578 static struct proc *vnlruproc;
579 static int vnlruproc_sig;
580 
581 static void
582 vnlru_proc(void)
583 {
584 	struct mount *mp, *nmp;
585 	int done;
586 	struct proc *p = vnlruproc;
587 	struct thread *td = FIRST_THREAD_IN_PROC(p);
588 
589 	mtx_lock(&Giant);
590 
591 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p,
592 	    SHUTDOWN_PRI_FIRST);
593 
594 	for (;;) {
595 		kthread_suspend_check(p);
596 		mtx_lock(&vnode_free_list_mtx);
597 		if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) {
598 			vnlruproc_sig = 0;
599 			wakeup(&vnlruproc_sig);
600 			msleep(vnlruproc, &vnode_free_list_mtx,
601 			    PVFS|PDROP, "vlruwt", hz);
602 			continue;
603 		}
604 		mtx_unlock(&vnode_free_list_mtx);
605 		done = 0;
606 		mtx_lock(&mountlist_mtx);
607 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
608 			if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) {
609 				nmp = TAILQ_NEXT(mp, mnt_list);
610 				continue;
611 			}
612 			done += vlrureclaim(mp);
613 			mtx_lock(&mountlist_mtx);
614 			nmp = TAILQ_NEXT(mp, mnt_list);
615 			vfs_unbusy(mp, td);
616 		}
617 		mtx_unlock(&mountlist_mtx);
618 		if (done == 0) {
619 #if 0
620 			/* These messages are temporary debugging aids */
621 			if (vnlru_nowhere < 5)
622 				printf("vnlru process getting nowhere..\n");
623 			else if (vnlru_nowhere == 5)
624 				printf("vnlru process messages stopped.\n");
625 #endif
626 			vnlru_nowhere++;
627 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
628 		}
629 	}
630 }
631 
632 static struct kproc_desc vnlru_kp = {
633 	"vnlru",
634 	vnlru_proc,
635 	&vnlruproc
636 };
637 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
638 
639 
640 /*
641  * Routines having to do with the management of the vnode table.
642  */
643 
644 /*
645  * Check to see if a free vnode can be recycled. If it can,
646  * recycle it and return it with the vnode interlock held.
647  */
648 static int
649 vtryrecycle(struct vnode *vp)
650 {
651 	struct thread *td = curthread;
652 	vm_object_t object;
653 	struct mount *vnmp;
654 	int error;
655 
656 	/* Don't recycle if we can't get the interlock */
657 	if (!VI_TRYLOCK(vp))
658 		return (EWOULDBLOCK);
659 	if (!VCANRECYCLE(vp)) {
660 		VI_UNLOCK(vp);
661 		return (EBUSY);
662 	}
663 	/*
664 	 * This vnode may found and locked via some other list, if so we
665 	 * can't recycle it yet.
666 	 */
667 	if (vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
668 		return (EWOULDBLOCK);
669 	/*
670 	 * Don't recycle if its filesystem is being suspended.
671 	 */
672 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
673 		VOP_UNLOCK(vp, 0, td);
674 		return (EBUSY);
675 	}
676 
677 	/*
678 	 * Don't recycle if we still have cached pages.
679 	 */
680 	object = vp->v_object;
681 	if (object != NULL) {
682 		VM_OBJECT_LOCK(object);
683 		if (object->resident_page_count ||
684 		    object->ref_count) {
685 			VM_OBJECT_UNLOCK(object);
686 			error = EBUSY;
687 			goto done;
688 		}
689 		VM_OBJECT_UNLOCK(object);
690 	}
691 	if (LIST_FIRST(&vp->v_cache_src)) {
692 		/*
693 		 * note: nameileafonly sysctl is temporary,
694 		 * for debugging only, and will eventually be
695 		 * removed.
696 		 */
697 		if (nameileafonly > 0) {
698 			/*
699 			 * Do not reuse namei-cached directory
700 			 * vnodes that have cached
701 			 * subdirectories.
702 			 */
703 			if (cache_leaf_test(vp) < 0) {
704 				error = EISDIR;
705 				goto done;
706 			}
707 		} else if (nameileafonly < 0 ||
708 			    vmiodirenable == 0) {
709 			/*
710 			 * Do not reuse namei-cached directory
711 			 * vnodes if nameileafonly is -1 or
712 			 * if VMIO backing for directories is
713 			 * turned off (otherwise we reuse them
714 			 * too quickly).
715 			 */
716 			error = EBUSY;
717 			goto done;
718 		}
719 	}
720 	/*
721 	 * If we got this far, we need to acquire the interlock and see if
722 	 * anyone picked up this vnode from another list.  If not, we will
723 	 * mark it with XLOCK via vgonel() so that anyone who does find it
724 	 * will skip over it.
725 	 */
726 	VI_LOCK(vp);
727 	if (!VCANRECYCLE(vp)) {
728 		VI_UNLOCK(vp);
729 		error = EBUSY;
730 		goto done;
731 	}
732 	mtx_lock(&vnode_free_list_mtx);
733 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
734 	vp->v_iflag &= ~VI_FREE;
735 	mtx_unlock(&vnode_free_list_mtx);
736 	vp->v_iflag |= VI_DOOMED;
737 	if ((vp->v_type != VBAD) || (vp->v_data != NULL)) {
738 		VOP_UNLOCK(vp, 0, td);
739 		vgonel(vp, td);
740 	} else
741 		VOP_UNLOCK(vp, LK_INTERLOCK, td);
742 	vn_finished_write(vnmp);
743 	return (0);
744 done:
745 	VOP_UNLOCK(vp, 0, td);
746 	vn_finished_write(vnmp);
747 	return (error);
748 }
749 
750 /*
751  * Return the next vnode from the free list.
752  */
753 int
754 getnewvnode(tag, mp, vops, vpp)
755 	const char *tag;
756 	struct mount *mp;
757 	struct vop_vector *vops;
758 	struct vnode **vpp;
759 {
760 	struct vnode *vp = NULL;
761 	struct vpollinfo *pollinfo = NULL;
762 	struct bufobj *bo;
763 
764 	mtx_lock(&vnode_free_list_mtx);
765 
766 	/*
767 	 * Try to reuse vnodes if we hit the max.  This situation only
768 	 * occurs in certain large-memory (2G+) situations.  We cannot
769 	 * attempt to directly reclaim vnodes due to nasty recursion
770 	 * problems.
771 	 */
772 	while (numvnodes - freevnodes > desiredvnodes) {
773 		if (vnlruproc_sig == 0) {
774 			vnlruproc_sig = 1;      /* avoid unnecessary wakeups */
775 			wakeup(vnlruproc);
776 		}
777 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
778 		    "vlruwk", hz);
779 	}
780 
781 	/*
782 	 * Attempt to reuse a vnode already on the free list, allocating
783 	 * a new vnode if we can't find one or if we have not reached a
784 	 * good minimum for good LRU performance.
785 	 */
786 
787 	if (freevnodes >= wantfreevnodes && numvnodes >= minvnodes) {
788 		int error;
789 		int count;
790 
791 		for (count = 0; count < freevnodes; count++) {
792 			vp = TAILQ_FIRST(&vnode_free_list);
793 			TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
794 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
795 			mtx_unlock(&vnode_free_list_mtx);
796 			error = vtryrecycle(vp);
797 			mtx_lock(&vnode_free_list_mtx);
798 			if (error == 0)
799 				break;
800 			vp = NULL;
801 		}
802 	}
803 	if (vp) {
804 		freevnodes--;
805 		bo = &vp->v_bufobj;
806 		mtx_unlock(&vnode_free_list_mtx);
807 
808 #ifdef INVARIANTS
809 		{
810 			if (vp->v_data)
811 				printf("cleaned vnode isn't, "
812 				       "address %p, inode %p\n",
813 				       vp, vp->v_data);
814 			if (bo->bo_numoutput)
815 				panic("%p: Clean vnode has pending I/O's", vp);
816 			if (vp->v_usecount != 0)
817 				panic("%p: Non-zero use count", vp);
818 			if (vp->v_writecount != 0)
819 				panic("%p: Non-zero write count", vp);
820 		}
821 #endif
822 		if ((pollinfo = vp->v_pollinfo) != NULL) {
823 			/*
824 			 * To avoid lock order reversals, the call to
825 			 * uma_zfree() must be delayed until the vnode
826 			 * interlock is released.
827 			 */
828 			vp->v_pollinfo = NULL;
829 		}
830 #ifdef MAC
831 		mac_destroy_vnode(vp);
832 #endif
833 		vp->v_iflag = 0;
834 		vp->v_vflag = 0;
835 		vp->v_lastw = 0;
836 		vp->v_lasta = 0;
837 		vp->v_cstart = 0;
838 		vp->v_clen = 0;
839 		bzero(&vp->v_un, sizeof vp->v_un);
840 		lockdestroy(vp->v_vnlock);
841 		lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE);
842 		VNASSERT(bo->bo_clean.bv_cnt == 0, vp,
843 		    ("cleanbufcnt not 0"));
844 		VNASSERT(bo->bo_clean.bv_root == NULL, vp,
845 		    ("cleanblkroot not NULL"));
846 		VNASSERT(bo->bo_dirty.bv_cnt == 0, vp,
847 		    ("dirtybufcnt not 0"));
848 		VNASSERT(bo->bo_dirty.bv_root == NULL, vp,
849 		    ("dirtyblkroot not NULL"));
850 	} else {
851 		numvnodes++;
852 		mtx_unlock(&vnode_free_list_mtx);
853 
854 		vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO);
855 		mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
856 		vp->v_dd = vp;
857 		bo = &vp->v_bufobj;
858 		bo->__bo_vnode = vp;
859 		bo->bo_mtx = &vp->v_interlock;
860 		vp->v_vnlock = &vp->v_lock;
861 		lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOPAUSE);
862 		cache_purge(vp);		/* Sets up v_id. */
863 		LIST_INIT(&vp->v_cache_src);
864 		TAILQ_INIT(&vp->v_cache_dst);
865 	}
866 
867 	TAILQ_INIT(&bo->bo_clean.bv_hd);
868 	TAILQ_INIT(&bo->bo_dirty.bv_hd);
869 	bo->bo_ops = &buf_ops_bio;
870 	bo->bo_private = vp;
871 	vp->v_type = VNON;
872 	vp->v_tag = tag;
873 	vp->v_op = vops;
874 	*vpp = vp;
875 	vp->v_usecount = 1;
876 	vp->v_data = 0;
877 	if (pollinfo != NULL) {
878 		knlist_destroy(&pollinfo->vpi_selinfo.si_note);
879 		mtx_destroy(&pollinfo->vpi_lock);
880 		uma_zfree(vnodepoll_zone, pollinfo);
881 	}
882 #ifdef MAC
883 	mac_init_vnode(vp);
884 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
885 		mac_associate_vnode_singlelabel(mp, vp);
886 	else if (mp == NULL)
887 		printf("NULL mp in getnewvnode()\n");
888 #endif
889 	delmntque(vp);
890 	if (mp != NULL) {
891 		insmntque(vp, mp);
892 		bo->bo_bsize = mp->mnt_stat.f_iosize;
893 	}
894 
895 	return (0);
896 }
897 
898 /*
899  * Delete from old mount point vnode list, if on one.
900  */
901 static void
902 delmntque(struct vnode *vp)
903 {
904 	struct mount *mp;
905 
906 	if (vp->v_mount == NULL)
907 		return;
908 	mp = vp->v_mount;
909 	MNT_ILOCK(mp);
910 	vp->v_mount = NULL;
911 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
912 		("bad mount point vnode list size"));
913 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
914 	mp->mnt_nvnodelistsize--;
915 	MNT_IUNLOCK(mp);
916 }
917 
918 /*
919  * Insert into list of vnodes for the new mount point, if available.
920  */
921 static void
922 insmntque(struct vnode *vp, struct mount *mp)
923 {
924 
925 	vp->v_mount = mp;
926 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
927 	MNT_ILOCK(vp->v_mount);
928 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
929 	mp->mnt_nvnodelistsize++;
930 	MNT_IUNLOCK(vp->v_mount);
931 }
932 
933 /*
934  * Flush out and invalidate all buffers associated with a bufobj
935  * Called with the underlying object locked.
936  */
937 int
938 bufobj_invalbuf(struct bufobj *bo, int flags, struct thread *td, int slpflag, int slptimeo)
939 {
940 	int error;
941 
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, bo, slpflag, slptimeo);
970 		if (error == 0)
971 			error = flushbuflist(&bo->bo_dirty,
972 			    flags, bo, 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, "bovlbx");
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 and invalidate all buffers associated with a vnode.
1018  * Called with the underlying object locked.
1019  */
1020 int
1021 vinvalbuf(struct vnode *vp, int flags, struct thread *td, int slpflag, int slptimeo)
1022 {
1023 
1024 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1025 	return (bufobj_invalbuf(&vp->v_bufobj, flags, td, slpflag, slptimeo));
1026 }
1027 
1028 /*
1029  * Flush out buffers on the specified list.
1030  *
1031  */
1032 static int
1033 flushbuflist(bufv, flags, bo, slpflag, slptimeo)
1034 	struct bufv *bufv;
1035 	int flags;
1036 	struct bufobj *bo;
1037 	int slpflag, slptimeo;
1038 {
1039 	struct buf *bp, *nbp;
1040 	int retval, error;
1041 
1042 	ASSERT_BO_LOCKED(bo);
1043 
1044 	retval = 0;
1045 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1046 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1047 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1048 			continue;
1049 		}
1050 		retval = EAGAIN;
1051 		error = BUF_TIMELOCK(bp,
1052 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo),
1053 		    "flushbuf", slpflag, slptimeo);
1054 		if (error) {
1055 			BO_LOCK(bo);
1056 			return (error != ENOLCK ? error : EAGAIN);
1057 		}
1058 		if (bp->b_bufobj != bo) {	/* XXX: necessary ? */
1059 			BO_LOCK(bo);
1060 			return (EAGAIN);
1061 		}
1062 		/*
1063 		 * XXX Since there are no node locks for NFS, I
1064 		 * believe there is a slight chance that a delayed
1065 		 * write will occur while sleeping just above, so
1066 		 * check for it.
1067 		 */
1068 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1069 		    (flags & V_SAVE)) {
1070 			bremfree(bp);
1071 			bp->b_flags |= B_ASYNC;
1072 			bwrite(bp);
1073 			BO_LOCK(bo);
1074 			return (EAGAIN);	/* XXX: why not loop ? */
1075 		}
1076 		bremfree(bp);
1077 		bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF);
1078 		bp->b_flags &= ~B_ASYNC;
1079 		brelse(bp);
1080 		BO_LOCK(bo);
1081 	}
1082 	return (retval);
1083 }
1084 
1085 /*
1086  * Truncate a file's buffer and pages to a specified length.  This
1087  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1088  * sync activity.
1089  */
1090 int
1091 vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td, off_t length, int blksize)
1092 {
1093 	struct buf *bp, *nbp;
1094 	int anyfreed;
1095 	int trunclbn;
1096 	struct bufobj *bo;
1097 
1098 	/*
1099 	 * Round up to the *next* lbn.
1100 	 */
1101 	trunclbn = (length + blksize - 1) / blksize;
1102 
1103 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1104 restart:
1105 	VI_LOCK(vp);
1106 	bo = &vp->v_bufobj;
1107 	anyfreed = 1;
1108 	for (;anyfreed;) {
1109 		anyfreed = 0;
1110 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1111 			if (bp->b_lblkno < trunclbn)
1112 				continue;
1113 			if (BUF_LOCK(bp,
1114 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1115 			    VI_MTX(vp)) == ENOLCK)
1116 				goto restart;
1117 
1118 			bremfree(bp);
1119 			bp->b_flags |= (B_INVAL | B_RELBUF);
1120 			bp->b_flags &= ~B_ASYNC;
1121 			brelse(bp);
1122 			anyfreed = 1;
1123 
1124 			if (nbp != NULL &&
1125 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1126 			    (nbp->b_vp != vp) ||
1127 			    (nbp->b_flags & B_DELWRI))) {
1128 				goto restart;
1129 			}
1130 			VI_LOCK(vp);
1131 		}
1132 
1133 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1134 			if (bp->b_lblkno < trunclbn)
1135 				continue;
1136 			if (BUF_LOCK(bp,
1137 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1138 			    VI_MTX(vp)) == ENOLCK)
1139 				goto restart;
1140 			bremfree(bp);
1141 			bp->b_flags |= (B_INVAL | B_RELBUF);
1142 			bp->b_flags &= ~B_ASYNC;
1143 			brelse(bp);
1144 			anyfreed = 1;
1145 			if (nbp != NULL &&
1146 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1147 			    (nbp->b_vp != vp) ||
1148 			    (nbp->b_flags & B_DELWRI) == 0)) {
1149 				goto restart;
1150 			}
1151 			VI_LOCK(vp);
1152 		}
1153 	}
1154 
1155 	if (length > 0) {
1156 restartsync:
1157 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1158 			if (bp->b_lblkno > 0)
1159 				continue;
1160 			/*
1161 			 * Since we hold the vnode lock this should only
1162 			 * fail if we're racing with the buf daemon.
1163 			 */
1164 			if (BUF_LOCK(bp,
1165 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1166 			    VI_MTX(vp)) == ENOLCK) {
1167 				goto restart;
1168 			}
1169 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1170 			    ("buf(%p) on dirty queue without DELWRI", bp));
1171 
1172 			bremfree(bp);
1173 			bawrite(bp);
1174 			VI_LOCK(vp);
1175 			goto restartsync;
1176 		}
1177 	}
1178 
1179 	bufobj_wwait(bo, 0, 0);
1180 	VI_UNLOCK(vp);
1181 	vnode_pager_setsize(vp, length);
1182 
1183 	return (0);
1184 }
1185 
1186 /*
1187  * buf_splay() - splay tree core for the clean/dirty list of buffers in
1188  * 		 a vnode.
1189  *
1190  *	NOTE: We have to deal with the special case of a background bitmap
1191  *	buffer, a situation where two buffers will have the same logical
1192  *	block offset.  We want (1) only the foreground buffer to be accessed
1193  *	in a lookup and (2) must differentiate between the foreground and
1194  *	background buffer in the splay tree algorithm because the splay
1195  *	tree cannot normally handle multiple entities with the same 'index'.
1196  *	We accomplish this by adding differentiating flags to the splay tree's
1197  *	numerical domain.
1198  */
1199 static
1200 struct buf *
1201 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root)
1202 {
1203 	struct buf dummy;
1204 	struct buf *lefttreemax, *righttreemin, *y;
1205 
1206 	if (root == NULL)
1207 		return (NULL);
1208 	lefttreemax = righttreemin = &dummy;
1209 	for (;;) {
1210 		if (lblkno < root->b_lblkno ||
1211 		    (lblkno == root->b_lblkno &&
1212 		    (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1213 			if ((y = root->b_left) == NULL)
1214 				break;
1215 			if (lblkno < y->b_lblkno) {
1216 				/* Rotate right. */
1217 				root->b_left = y->b_right;
1218 				y->b_right = root;
1219 				root = y;
1220 				if ((y = root->b_left) == NULL)
1221 					break;
1222 			}
1223 			/* Link into the new root's right tree. */
1224 			righttreemin->b_left = root;
1225 			righttreemin = root;
1226 		} else if (lblkno > root->b_lblkno ||
1227 		    (lblkno == root->b_lblkno &&
1228 		    (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) {
1229 			if ((y = root->b_right) == NULL)
1230 				break;
1231 			if (lblkno > y->b_lblkno) {
1232 				/* Rotate left. */
1233 				root->b_right = y->b_left;
1234 				y->b_left = root;
1235 				root = y;
1236 				if ((y = root->b_right) == NULL)
1237 					break;
1238 			}
1239 			/* Link into the new root's left tree. */
1240 			lefttreemax->b_right = root;
1241 			lefttreemax = root;
1242 		} else {
1243 			break;
1244 		}
1245 		root = y;
1246 	}
1247 	/* Assemble the new root. */
1248 	lefttreemax->b_right = root->b_left;
1249 	righttreemin->b_left = root->b_right;
1250 	root->b_left = dummy.b_right;
1251 	root->b_right = dummy.b_left;
1252 	return (root);
1253 }
1254 
1255 static void
1256 buf_vlist_remove(struct buf *bp)
1257 {
1258 	struct buf *root;
1259 	struct bufv *bv;
1260 
1261 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1262 	ASSERT_BO_LOCKED(bp->b_bufobj);
1263 	if (bp->b_xflags & BX_VNDIRTY)
1264 		bv = &bp->b_bufobj->bo_dirty;
1265 	else
1266 		bv = &bp->b_bufobj->bo_clean;
1267 	if (bp != bv->bv_root) {
1268 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1269 		KASSERT(root == bp, ("splay lookup failed in remove"));
1270 	}
1271 	if (bp->b_left == NULL) {
1272 		root = bp->b_right;
1273 	} else {
1274 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left);
1275 		root->b_right = bp->b_right;
1276 	}
1277 	bv->bv_root = root;
1278 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1279 	bv->bv_cnt--;
1280 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1281 }
1282 
1283 /*
1284  * Add the buffer to the sorted clean or dirty block list using a
1285  * splay tree algorithm.
1286  *
1287  * NOTE: xflags is passed as a constant, optimizing this inline function!
1288  */
1289 static void
1290 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1291 {
1292 	struct buf *root;
1293 	struct bufv *bv;
1294 
1295 	ASSERT_BO_LOCKED(bo);
1296 	bp->b_xflags |= xflags;
1297 	if (xflags & BX_VNDIRTY)
1298 		bv = &bo->bo_dirty;
1299 	else
1300 		bv = &bo->bo_clean;
1301 
1302 	root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1303 	if (root == NULL) {
1304 		bp->b_left = NULL;
1305 		bp->b_right = NULL;
1306 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1307 	} else if (bp->b_lblkno < root->b_lblkno ||
1308 	    (bp->b_lblkno == root->b_lblkno &&
1309 	    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1310 		bp->b_left = root->b_left;
1311 		bp->b_right = root;
1312 		root->b_left = NULL;
1313 		TAILQ_INSERT_BEFORE(root, bp, b_bobufs);
1314 	} else {
1315 		bp->b_right = root->b_right;
1316 		bp->b_left = root;
1317 		root->b_right = NULL;
1318 		TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs);
1319 	}
1320 	bv->bv_cnt++;
1321 	bv->bv_root = bp;
1322 }
1323 
1324 /*
1325  * Lookup a buffer using the splay tree.  Note that we specifically avoid
1326  * shadow buffers used in background bitmap writes.
1327  *
1328  * This code isn't quite efficient as it could be because we are maintaining
1329  * two sorted lists and do not know which list the block resides in.
1330  *
1331  * During a "make buildworld" the desired buffer is found at one of
1332  * the roots more than 60% of the time.  Thus, checking both roots
1333  * before performing either splay eliminates unnecessary splays on the
1334  * first tree splayed.
1335  */
1336 struct buf *
1337 gbincore(struct bufobj *bo, daddr_t lblkno)
1338 {
1339 	struct buf *bp;
1340 
1341 	ASSERT_BO_LOCKED(bo);
1342 	if ((bp = bo->bo_clean.bv_root) != NULL &&
1343 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1344 		return (bp);
1345 	if ((bp = bo->bo_dirty.bv_root) != NULL &&
1346 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1347 		return (bp);
1348 	if ((bp = bo->bo_clean.bv_root) != NULL) {
1349 		bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp);
1350 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1351 			return (bp);
1352 	}
1353 	if ((bp = bo->bo_dirty.bv_root) != NULL) {
1354 		bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp);
1355 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1356 			return (bp);
1357 	}
1358 	return (NULL);
1359 }
1360 
1361 /*
1362  * Associate a buffer with a vnode.
1363  */
1364 void
1365 bgetvp(struct vnode *vp, struct buf *bp)
1366 {
1367 
1368 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
1369 
1370 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
1371 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
1372 	    ("bgetvp: bp already attached! %p", bp));
1373 
1374 	ASSERT_VI_LOCKED(vp, "bgetvp");
1375 	vholdl(vp);
1376 	bp->b_vp = vp;
1377 	bp->b_bufobj = &vp->v_bufobj;
1378 	/*
1379 	 * Insert onto list for new vnode.
1380 	 */
1381 	buf_vlist_add(bp, &vp->v_bufobj, BX_VNCLEAN);
1382 }
1383 
1384 /*
1385  * Disassociate a buffer from a vnode.
1386  */
1387 void
1388 brelvp(struct buf *bp)
1389 {
1390 	struct bufobj *bo;
1391 	struct vnode *vp;
1392 
1393 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1394 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1395 
1396 	/*
1397 	 * Delete from old vnode list, if on one.
1398 	 */
1399 	vp = bp->b_vp;		/* XXX */
1400 	bo = bp->b_bufobj;
1401 	BO_LOCK(bo);
1402 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1403 		buf_vlist_remove(bp);
1404 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1405 		bo->bo_flag &= ~BO_ONWORKLST;
1406 		mtx_lock(&sync_mtx);
1407 		LIST_REMOVE(bo, bo_synclist);
1408  		syncer_worklist_len--;
1409 		mtx_unlock(&sync_mtx);
1410 	}
1411 	vdropl(vp);
1412 	bp->b_vp = NULL;
1413 	bp->b_bufobj = NULL;
1414 	BO_UNLOCK(bo);
1415 }
1416 
1417 /*
1418  * Add an item to the syncer work queue.
1419  */
1420 static void
1421 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
1422 {
1423 	int slot;
1424 
1425 	ASSERT_BO_LOCKED(bo);
1426 
1427 	mtx_lock(&sync_mtx);
1428 	if (bo->bo_flag & BO_ONWORKLST)
1429 		LIST_REMOVE(bo, bo_synclist);
1430 	else {
1431 		bo->bo_flag |= BO_ONWORKLST;
1432  		syncer_worklist_len++;
1433 	}
1434 
1435 	if (delay > syncer_maxdelay - 2)
1436 		delay = syncer_maxdelay - 2;
1437 	slot = (syncer_delayno + delay) & syncer_mask;
1438 
1439 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
1440 	mtx_unlock(&sync_mtx);
1441 }
1442 
1443 static int
1444 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
1445 {
1446 	int error, len;
1447 
1448 	mtx_lock(&sync_mtx);
1449 	len = syncer_worklist_len - sync_vnode_count;
1450 	mtx_unlock(&sync_mtx);
1451 	error = SYSCTL_OUT(req, &len, sizeof(len));
1452 	return (error);
1453 }
1454 
1455 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
1456     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
1457 
1458 struct  proc *updateproc;
1459 static void sched_sync(void);
1460 static struct kproc_desc up_kp = {
1461 	"syncer",
1462 	sched_sync,
1463 	&updateproc
1464 };
1465 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp)
1466 
1467 static int
1468 sync_vnode(struct bufobj *bo, struct thread *td)
1469 {
1470 	struct vnode *vp;
1471 	struct mount *mp;
1472 
1473 	vp = bo->__bo_vnode; 	/* XXX */
1474 	if (VOP_ISLOCKED(vp, NULL) != 0)
1475 		return (1);
1476 	if (VI_TRYLOCK(vp) == 0)
1477 		return (1);
1478 	/*
1479 	 * We use vhold in case the vnode does not
1480 	 * successfully sync.  vhold prevents the vnode from
1481 	 * going away when we unlock the sync_mtx so that
1482 	 * we can acquire the vnode interlock.
1483 	 */
1484 	vholdl(vp);
1485 	mtx_unlock(&sync_mtx);
1486 	VI_UNLOCK(vp);
1487 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1488 		vdrop(vp);
1489 		mtx_lock(&sync_mtx);
1490 		return (1);
1491 	}
1492 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
1493 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
1494 	VOP_UNLOCK(vp, 0, td);
1495 	vn_finished_write(mp);
1496 	VI_LOCK(vp);
1497 	if ((bo->bo_flag & BO_ONWORKLST) != 0) {
1498 		/*
1499 		 * Put us back on the worklist.  The worklist
1500 		 * routine will remove us from our current
1501 		 * position and then add us back in at a later
1502 		 * position.
1503 		 */
1504 		vn_syncer_add_to_worklist(bo, syncdelay);
1505 	}
1506 	vdropl(vp);
1507 	VI_UNLOCK(vp);
1508 	mtx_lock(&sync_mtx);
1509 	return (0);
1510 }
1511 
1512 /*
1513  * System filesystem synchronizer daemon.
1514  */
1515 static void
1516 sched_sync(void)
1517 {
1518 	struct synclist *next;
1519 	struct synclist *slp;
1520 	struct bufobj *bo;
1521 	long starttime;
1522 	struct thread *td = FIRST_THREAD_IN_PROC(updateproc);
1523 	static int dummychan;
1524 	int last_work_seen;
1525 	int net_worklist_len;
1526 	int syncer_final_iter;
1527 	int first_printf;
1528 	int error;
1529 
1530 	mtx_lock(&Giant);
1531 	last_work_seen = 0;
1532 	syncer_final_iter = 0;
1533 	first_printf = 1;
1534 	syncer_state = SYNCER_RUNNING;
1535 	starttime = time_second;
1536 
1537 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
1538 	    SHUTDOWN_PRI_LAST);
1539 
1540 	for (;;) {
1541 		mtx_lock(&sync_mtx);
1542 		if (syncer_state == SYNCER_FINAL_DELAY &&
1543 		    syncer_final_iter == 0) {
1544 			mtx_unlock(&sync_mtx);
1545 			kthread_suspend_check(td->td_proc);
1546 			mtx_lock(&sync_mtx);
1547 		}
1548 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
1549 		if (syncer_state != SYNCER_RUNNING &&
1550 		    starttime != time_second) {
1551 			if (first_printf) {
1552 				printf("\nSyncing disks, vnodes remaining...");
1553 				first_printf = 0;
1554 			}
1555 			printf("%d ", net_worklist_len);
1556 		}
1557 		starttime = time_second;
1558 
1559 		/*
1560 		 * Push files whose dirty time has expired.  Be careful
1561 		 * of interrupt race on slp queue.
1562 		 *
1563 		 * Skip over empty worklist slots when shutting down.
1564 		 */
1565 		do {
1566 			slp = &syncer_workitem_pending[syncer_delayno];
1567 			syncer_delayno += 1;
1568 			if (syncer_delayno == syncer_maxdelay)
1569 				syncer_delayno = 0;
1570 			next = &syncer_workitem_pending[syncer_delayno];
1571 			/*
1572 			 * If the worklist has wrapped since the
1573 			 * it was emptied of all but syncer vnodes,
1574 			 * switch to the FINAL_DELAY state and run
1575 			 * for one more second.
1576 			 */
1577 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
1578 			    net_worklist_len == 0 &&
1579 			    last_work_seen == syncer_delayno) {
1580 				syncer_state = SYNCER_FINAL_DELAY;
1581 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
1582 			}
1583 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
1584 		    syncer_worklist_len > 0);
1585 
1586 		/*
1587 		 * Keep track of the last time there was anything
1588 		 * on the worklist other than syncer vnodes.
1589 		 * Return to the SHUTTING_DOWN state if any
1590 		 * new work appears.
1591 		 */
1592 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
1593 			last_work_seen = syncer_delayno;
1594 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
1595 			syncer_state = SYNCER_SHUTTING_DOWN;
1596 		while ((bo = LIST_FIRST(slp)) != NULL) {
1597 			error = sync_vnode(bo, td);
1598 			if (error == 1) {
1599 				LIST_REMOVE(bo, bo_synclist);
1600 				LIST_INSERT_HEAD(next, bo, bo_synclist);
1601 				continue;
1602 			}
1603 		}
1604 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
1605 			syncer_final_iter--;
1606 		mtx_unlock(&sync_mtx);
1607 
1608 		/*
1609 		 * Do soft update processing.
1610 		 */
1611 		if (softdep_process_worklist_hook != NULL)
1612 			(*softdep_process_worklist_hook)(NULL);
1613 
1614 		/*
1615 		 * The variable rushjob allows the kernel to speed up the
1616 		 * processing of the filesystem syncer process. A rushjob
1617 		 * value of N tells the filesystem syncer to process the next
1618 		 * N seconds worth of work on its queue ASAP. Currently rushjob
1619 		 * is used by the soft update code to speed up the filesystem
1620 		 * syncer process when the incore state is getting so far
1621 		 * ahead of the disk that the kernel memory pool is being
1622 		 * threatened with exhaustion.
1623 		 */
1624 		mtx_lock(&sync_mtx);
1625 		if (rushjob > 0) {
1626 			rushjob -= 1;
1627 			mtx_unlock(&sync_mtx);
1628 			continue;
1629 		}
1630 		mtx_unlock(&sync_mtx);
1631 		/*
1632 		 * Just sleep for a short period if time between
1633 		 * iterations when shutting down to allow some I/O
1634 		 * to happen.
1635 		 *
1636 		 * If it has taken us less than a second to process the
1637 		 * current work, then wait. Otherwise start right over
1638 		 * again. We can still lose time if any single round
1639 		 * takes more than two seconds, but it does not really
1640 		 * matter as we are just trying to generally pace the
1641 		 * filesystem activity.
1642 		 */
1643 		if (syncer_state != SYNCER_RUNNING)
1644 			tsleep(&dummychan, PPAUSE, "syncfnl",
1645 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
1646 		else if (time_second == starttime)
1647 			tsleep(&lbolt, PPAUSE, "syncer", 0);
1648 	}
1649 }
1650 
1651 /*
1652  * Request the syncer daemon to speed up its work.
1653  * We never push it to speed up more than half of its
1654  * normal turn time, otherwise it could take over the cpu.
1655  */
1656 int
1657 speedup_syncer()
1658 {
1659 	struct thread *td;
1660 	int ret = 0;
1661 
1662 	td = FIRST_THREAD_IN_PROC(updateproc);
1663 	sleepq_remove(td, &lbolt);
1664 	mtx_lock(&sync_mtx);
1665 	if (rushjob < syncdelay / 2) {
1666 		rushjob += 1;
1667 		stat_rush_requests += 1;
1668 		ret = 1;
1669 	}
1670 	mtx_unlock(&sync_mtx);
1671 	return (ret);
1672 }
1673 
1674 /*
1675  * Tell the syncer to speed up its work and run though its work
1676  * list several times, then tell it to shut down.
1677  */
1678 static void
1679 syncer_shutdown(void *arg, int howto)
1680 {
1681 	struct thread *td;
1682 
1683 	if (howto & RB_NOSYNC)
1684 		return;
1685 	td = FIRST_THREAD_IN_PROC(updateproc);
1686 	sleepq_remove(td, &lbolt);
1687 	mtx_lock(&sync_mtx);
1688 	syncer_state = SYNCER_SHUTTING_DOWN;
1689 	rushjob = 0;
1690 	mtx_unlock(&sync_mtx);
1691 	kproc_shutdown(arg, howto);
1692 }
1693 
1694 /*
1695  * Reassign a buffer from one vnode to another.
1696  * Used to assign file specific control information
1697  * (indirect blocks) to the vnode to which they belong.
1698  */
1699 void
1700 reassignbuf(struct buf *bp)
1701 {
1702 	struct vnode *vp;
1703 	struct bufobj *bo;
1704 	int delay;
1705 
1706 	vp = bp->b_vp;
1707 	bo = bp->b_bufobj;
1708 	++reassignbufcalls;
1709 
1710 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
1711 	    bp, bp->b_vp, bp->b_flags);
1712 	/*
1713 	 * B_PAGING flagged buffers cannot be reassigned because their vp
1714 	 * is not fully linked in.
1715 	 */
1716 	if (bp->b_flags & B_PAGING)
1717 		panic("cannot reassign paging buffer");
1718 
1719 	/*
1720 	 * Delete from old vnode list, if on one.
1721 	 */
1722 	VI_LOCK(vp);
1723 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1724 		buf_vlist_remove(bp);
1725 	/*
1726 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
1727 	 * of clean buffers.
1728 	 */
1729 	if (bp->b_flags & B_DELWRI) {
1730 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
1731 			switch (vp->v_type) {
1732 			case VDIR:
1733 				delay = dirdelay;
1734 				break;
1735 			case VCHR:
1736 				delay = metadelay;
1737 				break;
1738 			default:
1739 				delay = filedelay;
1740 			}
1741 			vn_syncer_add_to_worklist(bo, delay);
1742 		}
1743 		buf_vlist_add(bp, bo, BX_VNDIRTY);
1744 	} else {
1745 		buf_vlist_add(bp, bo, BX_VNCLEAN);
1746 
1747 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1748 			mtx_lock(&sync_mtx);
1749 			LIST_REMOVE(bo, bo_synclist);
1750  			syncer_worklist_len--;
1751 			mtx_unlock(&sync_mtx);
1752 			bo->bo_flag &= ~BO_ONWORKLST;
1753 		}
1754 	}
1755 	VI_UNLOCK(vp);
1756 }
1757 
1758 static void
1759 v_incr_usecount(struct vnode *vp, int delta)
1760 {
1761 
1762 	vp->v_usecount += delta;
1763 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1764 		dev_lock();
1765 		vp->v_rdev->si_usecount += delta;
1766 		dev_unlock();
1767 	}
1768 }
1769 
1770 /*
1771  * Grab a particular vnode from the free list, increment its
1772  * reference count and lock it. The vnode lock bit is set if the
1773  * vnode is being eliminated in vgone. The process is awakened
1774  * when the transition is completed, and an error returned to
1775  * indicate that the vnode is no longer usable (possibly having
1776  * been changed to a new filesystem type).
1777  */
1778 int
1779 vget(vp, flags, td)
1780 	struct vnode *vp;
1781 	int flags;
1782 	struct thread *td;
1783 {
1784 	int error;
1785 
1786 	/*
1787 	 * If the vnode is in the process of being cleaned out for
1788 	 * another use, we wait for the cleaning to finish and then
1789 	 * return failure. Cleaning is determined by checking that
1790 	 * the VI_XLOCK flag is set.
1791 	 */
1792 	if ((flags & LK_INTERLOCK) == 0)
1793 		VI_LOCK(vp);
1794 	if (vp->v_iflag & VI_XLOCK && vp->v_vxthread != curthread) {
1795 		if ((flags & LK_NOWAIT) == 0) {
1796 			vx_waitl(vp);
1797 			VI_UNLOCK(vp);
1798 			return (ENOENT);
1799 		}
1800 		VI_UNLOCK(vp);
1801 		return (EBUSY);
1802 	}
1803 
1804 	v_incr_usecount(vp, 1);
1805 
1806 	if (VSHOULDBUSY(vp))
1807 		vbusy(vp);
1808 	if (flags & LK_TYPE_MASK) {
1809 		if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) {
1810 			/*
1811 			 * must expand vrele here because we do not want
1812 			 * to call VOP_INACTIVE if the reference count
1813 			 * drops back to zero since it was never really
1814 			 * active. We must remove it from the free list
1815 			 * before sleeping so that multiple processes do
1816 			 * not try to recycle it.
1817 			 */
1818 			VI_LOCK(vp);
1819 			v_incr_usecount(vp, -1);
1820 			if (VSHOULDFREE(vp))
1821 				vfree(vp);
1822 			else
1823 				vlruvp(vp);
1824 			VI_UNLOCK(vp);
1825 		}
1826 		return (error);
1827 	}
1828 	VI_UNLOCK(vp);
1829 	return (0);
1830 }
1831 
1832 /*
1833  * Increase the reference count of a vnode.
1834  */
1835 void
1836 vref(struct vnode *vp)
1837 {
1838 
1839 	VI_LOCK(vp);
1840 	v_incr_usecount(vp, 1);
1841 	VI_UNLOCK(vp);
1842 }
1843 
1844 /*
1845  * Return reference count of a vnode.
1846  *
1847  * The results of this call are only guaranteed when some mechanism other
1848  * than the VI lock is used to stop other processes from gaining references
1849  * to the vnode.  This may be the case if the caller holds the only reference.
1850  * This is also useful when stale data is acceptable as race conditions may
1851  * be accounted for by some other means.
1852  */
1853 int
1854 vrefcnt(struct vnode *vp)
1855 {
1856 	int usecnt;
1857 
1858 	VI_LOCK(vp);
1859 	usecnt = vp->v_usecount;
1860 	VI_UNLOCK(vp);
1861 
1862 	return (usecnt);
1863 }
1864 
1865 
1866 /*
1867  * Vnode put/release.
1868  * If count drops to zero, call inactive routine and return to freelist.
1869  */
1870 void
1871 vrele(vp)
1872 	struct vnode *vp;
1873 {
1874 	struct thread *td = curthread;	/* XXX */
1875 
1876 	KASSERT(vp != NULL, ("vrele: null vp"));
1877 
1878 	VI_LOCK(vp);
1879 
1880 	/* Skip this v_writecount check if we're going to panic below. */
1881 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
1882 	    ("vrele: missed vn_close"));
1883 
1884 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
1885 	    vp->v_usecount == 1)) {
1886 		v_incr_usecount(vp, -1);
1887 		VI_UNLOCK(vp);
1888 
1889 		return;
1890 	}
1891 
1892 	if (vp->v_usecount == 1) {
1893 		v_incr_usecount(vp, -1);
1894 		/*
1895 		 * We must call VOP_INACTIVE with the node locked. Mark
1896 		 * as VI_DOINGINACT to avoid recursion.
1897 		 */
1898 		if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) {
1899 			VI_LOCK(vp);
1900 			VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
1901 			    ("vrele: recursed on VI_DOINGINACT"));
1902 			vp->v_iflag |= VI_DOINGINACT;
1903 			VI_UNLOCK(vp);
1904 			VOP_INACTIVE(vp, td);
1905 			VI_LOCK(vp);
1906 			VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
1907 			    ("vrele: lost VI_DOINGINACT"));
1908 			vp->v_iflag &= ~VI_DOINGINACT;
1909 		} else
1910 			VI_LOCK(vp);
1911 		if (VSHOULDFREE(vp))
1912 			vfree(vp);
1913 		else
1914 			vlruvp(vp);
1915 		VI_UNLOCK(vp);
1916 
1917 	} else {
1918 #ifdef DIAGNOSTIC
1919 		vprint("vrele: negative ref count", vp);
1920 #endif
1921 		VI_UNLOCK(vp);
1922 		panic("vrele: negative ref cnt");
1923 	}
1924 }
1925 
1926 /*
1927  * Release an already locked vnode.  This give the same effects as
1928  * unlock+vrele(), but takes less time and avoids releasing and
1929  * re-aquiring the lock (as vrele() aquires the lock internally.)
1930  */
1931 void
1932 vput(vp)
1933 	struct vnode *vp;
1934 {
1935 	struct thread *td = curthread;	/* XXX */
1936 
1937 	KASSERT(vp != NULL, ("vput: null vp"));
1938 	VI_LOCK(vp);
1939 	/* Skip this v_writecount check if we're going to panic below. */
1940 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
1941 	    ("vput: missed vn_close"));
1942 
1943 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
1944 	    vp->v_usecount == 1)) {
1945 		v_incr_usecount(vp, -1);
1946 		VOP_UNLOCK(vp, LK_INTERLOCK, td);
1947 		return;
1948 	}
1949 
1950 	if (vp->v_usecount == 1) {
1951 		v_incr_usecount(vp, -1);
1952 		/*
1953 		 * We must call VOP_INACTIVE with the node locked, so
1954 		 * we just need to release the vnode mutex. Mark as
1955 		 * as VI_DOINGINACT to avoid recursion.
1956 		 */
1957 		VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
1958 		    ("vput: recursed on VI_DOINGINACT"));
1959 		vp->v_iflag |= VI_DOINGINACT;
1960 		VI_UNLOCK(vp);
1961 		VOP_INACTIVE(vp, td);
1962 		VI_LOCK(vp);
1963 		VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
1964 		    ("vput: lost VI_DOINGINACT"));
1965 		vp->v_iflag &= ~VI_DOINGINACT;
1966 		if (VSHOULDFREE(vp))
1967 			vfree(vp);
1968 		else
1969 			vlruvp(vp);
1970 		VI_UNLOCK(vp);
1971 
1972 	} else {
1973 #ifdef DIAGNOSTIC
1974 		vprint("vput: negative ref count", vp);
1975 #endif
1976 		panic("vput: negative ref cnt");
1977 	}
1978 }
1979 
1980 /*
1981  * Somebody doesn't want the vnode recycled.
1982  */
1983 void
1984 vhold(struct vnode *vp)
1985 {
1986 
1987 	VI_LOCK(vp);
1988 	vholdl(vp);
1989 	VI_UNLOCK(vp);
1990 }
1991 
1992 static void
1993 vholdl(struct vnode *vp)
1994 {
1995 
1996 	vp->v_holdcnt++;
1997 	if (VSHOULDBUSY(vp))
1998 		vbusy(vp);
1999 }
2000 
2001 /*
2002  * Note that there is one less who cares about this vnode.  vdrop() is the
2003  * opposite of vhold().
2004  */
2005 void
2006 vdrop(struct vnode *vp)
2007 {
2008 
2009 	VI_LOCK(vp);
2010 	vdropl(vp);
2011 	VI_UNLOCK(vp);
2012 }
2013 
2014 static void
2015 vdropl(struct vnode *vp)
2016 {
2017 
2018 	if (vp->v_holdcnt <= 0)
2019 		panic("vdrop: holdcnt");
2020 	vp->v_holdcnt--;
2021 	if (VSHOULDFREE(vp))
2022 		vfree(vp);
2023 	else
2024 		vlruvp(vp);
2025 }
2026 
2027 /*
2028  * Remove any vnodes in the vnode table belonging to mount point mp.
2029  *
2030  * If FORCECLOSE is not specified, there should not be any active ones,
2031  * return error if any are found (nb: this is a user error, not a
2032  * system error). If FORCECLOSE is specified, detach any active vnodes
2033  * that are found.
2034  *
2035  * If WRITECLOSE is set, only flush out regular file vnodes open for
2036  * writing.
2037  *
2038  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2039  *
2040  * `rootrefs' specifies the base reference count for the root vnode
2041  * of this filesystem. The root vnode is considered busy if its
2042  * v_usecount exceeds this value. On a successful return, vflush(, td)
2043  * will call vrele() on the root vnode exactly rootrefs times.
2044  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2045  * be zero.
2046  */
2047 #ifdef DIAGNOSTIC
2048 static int busyprt = 0;		/* print out busy vnodes */
2049 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
2050 #endif
2051 
2052 int
2053 vflush(mp, rootrefs, flags, td)
2054 	struct mount *mp;
2055 	int rootrefs;
2056 	int flags;
2057 	struct thread *td;
2058 {
2059 	struct vnode *vp, *nvp, *rootvp = NULL;
2060 	struct vattr vattr;
2061 	int busy = 0, error;
2062 
2063 	if (rootrefs > 0) {
2064 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2065 		    ("vflush: bad args"));
2066 		/*
2067 		 * Get the filesystem root vnode. We can vput() it
2068 		 * immediately, since with rootrefs > 0, it won't go away.
2069 		 */
2070 		if ((error = VFS_ROOT(mp, &rootvp, td)) != 0)
2071 			return (error);
2072 		vput(rootvp);
2073 
2074 	}
2075 	MNT_ILOCK(mp);
2076 loop:
2077 	MNT_VNODE_FOREACH(vp, mp, nvp) {
2078 
2079 		VI_LOCK(vp);
2080 		MNT_IUNLOCK(mp);
2081 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td);
2082 		if (error) {
2083 			MNT_ILOCK(mp);
2084 			goto loop;
2085 		}
2086 		/*
2087 		 * Skip over a vnodes marked VV_SYSTEM.
2088 		 */
2089 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2090 			VOP_UNLOCK(vp, 0, td);
2091 			MNT_ILOCK(mp);
2092 			continue;
2093 		}
2094 		/*
2095 		 * If WRITECLOSE is set, flush out unlinked but still open
2096 		 * files (even if open only for reading) and regular file
2097 		 * vnodes open for writing.
2098 		 */
2099 		if (flags & WRITECLOSE) {
2100 			error = VOP_GETATTR(vp, &vattr, td->td_ucred, td);
2101 			VI_LOCK(vp);
2102 
2103 			if ((vp->v_type == VNON ||
2104 			    (error == 0 && vattr.va_nlink > 0)) &&
2105 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2106 				VOP_UNLOCK(vp, LK_INTERLOCK, td);
2107 				MNT_ILOCK(mp);
2108 				continue;
2109 			}
2110 		} else
2111 			VI_LOCK(vp);
2112 
2113 		VOP_UNLOCK(vp, 0, td);
2114 
2115 		/*
2116 		 * With v_usecount == 0, all we need to do is clear out the
2117 		 * vnode data structures and we are done.
2118 		 */
2119 		if (vp->v_usecount == 0) {
2120 			vgonel(vp, td);
2121 			MNT_ILOCK(mp);
2122 			continue;
2123 		}
2124 
2125 		/*
2126 		 * If FORCECLOSE is set, forcibly close the vnode. For block
2127 		 * or character devices, revert to an anonymous device. For
2128 		 * all other files, just kill them.
2129 		 */
2130 		if (flags & FORCECLOSE) {
2131 			VNASSERT(vp->v_type != VCHR && vp->v_type != VBLK, vp,
2132 			    ("device VNODE %p is FORCECLOSED", vp));
2133 			vgonel(vp, td);
2134 			MNT_ILOCK(mp);
2135 			continue;
2136 		}
2137 #ifdef DIAGNOSTIC
2138 		if (busyprt)
2139 			vprint("vflush: busy vnode", vp);
2140 #endif
2141 		VI_UNLOCK(vp);
2142 		MNT_ILOCK(mp);
2143 		busy++;
2144 	}
2145 	MNT_IUNLOCK(mp);
2146 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2147 		/*
2148 		 * If just the root vnode is busy, and if its refcount
2149 		 * is equal to `rootrefs', then go ahead and kill it.
2150 		 */
2151 		VI_LOCK(rootvp);
2152 		KASSERT(busy > 0, ("vflush: not busy"));
2153 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
2154 		    ("vflush: usecount %d < rootrefs %d",
2155 		     rootvp->v_usecount, rootrefs));
2156 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2157 			vgonel(rootvp, td);
2158 			busy = 0;
2159 		} else
2160 			VI_UNLOCK(rootvp);
2161 	}
2162 	if (busy)
2163 		return (EBUSY);
2164 	for (; rootrefs > 0; rootrefs--)
2165 		vrele(rootvp);
2166 	return (0);
2167 }
2168 
2169 /*
2170  * This moves a now (likely recyclable) vnode to the end of the
2171  * mountlist.  XXX However, it is temporarily disabled until we
2172  * can clean up ffs_sync() and friends, which have loop restart
2173  * conditions which this code causes to operate O(N^2).
2174  */
2175 static void
2176 vlruvp(struct vnode *vp)
2177 {
2178 #if 0
2179 	struct mount *mp;
2180 
2181 	if ((mp = vp->v_mount) != NULL) {
2182 		MNT_ILOCK(mp);
2183 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2184 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2185 		MNT_IUNLOCK(mp);
2186 	}
2187 #endif
2188 }
2189 
2190 static void
2191 vx_lock(struct vnode *vp)
2192 {
2193 
2194 	ASSERT_VI_LOCKED(vp, "vx_lock");
2195 
2196 	/*
2197 	 * Prevent the vnode from being recycled or brought into use while we
2198 	 * clean it out.
2199 	 */
2200 	if (vp->v_iflag & VI_XLOCK)
2201 		panic("vx_lock: deadlock");
2202 	vp->v_iflag |= VI_XLOCK;
2203 	vp->v_vxthread = curthread;
2204 }
2205 
2206 static void
2207 vx_unlock(struct vnode *vp)
2208 {
2209 	ASSERT_VI_LOCKED(vp, "vx_unlock");
2210 	vp->v_iflag &= ~VI_XLOCK;
2211 	vp->v_vxthread = NULL;
2212 	if (vp->v_iflag & VI_XWANT) {
2213 		vp->v_iflag &= ~VI_XWANT;
2214 		wakeup(vp);
2215 	}
2216 }
2217 
2218 int
2219 vx_wait(struct vnode *vp)
2220 {
2221 	int locked;
2222 
2223 	ASSERT_VI_UNLOCKED(vp, "vx_wait");
2224 	VI_LOCK(vp);
2225 	locked = vx_waitl(vp);
2226 	VI_UNLOCK(vp);
2227 	return (locked);
2228 }
2229 
2230 int
2231 vx_waitl(struct vnode *vp)
2232 {
2233 	int locked = 0;
2234 
2235 	ASSERT_VI_LOCKED(vp, "vx_wait");
2236 	while (vp->v_iflag & VI_XLOCK) {
2237 		locked = 1;
2238 		vp->v_iflag |= VI_XWANT;
2239 		msleep(vp, VI_MTX(vp), PINOD, "vxwait", 0);
2240 	}
2241 	return (locked);
2242 }
2243 
2244 /*
2245  * Recycle an unused vnode to the front of the free list.
2246  * Release the passed interlock if the vnode will be recycled.
2247  */
2248 int
2249 vrecycle(struct vnode *vp, struct thread *td)
2250 {
2251 
2252 	VI_LOCK(vp);
2253 	if (vp->v_usecount == 0) {
2254 		vgonel(vp, td);
2255 		return (1);
2256 	}
2257 	VI_UNLOCK(vp);
2258 	return (0);
2259 }
2260 
2261 /*
2262  * Eliminate all activity associated with a vnode
2263  * in preparation for reuse.
2264  */
2265 void
2266 vgone(struct vnode *vp)
2267 {
2268 	struct thread *td = curthread;	/* XXX */
2269 
2270 	VI_LOCK(vp);
2271 	vgonel(vp, td);
2272 }
2273 
2274 /*
2275  * vgone, with the vp interlock held.
2276  */
2277 void
2278 vgonel(struct vnode *vp, struct thread *td)
2279 {
2280 	int active;
2281 
2282 	/*
2283 	 * If a vgone (or vclean) is already in progress,
2284 	 * wait until it is done and return.
2285 	 */
2286 	ASSERT_VI_LOCKED(vp, "vgonel");
2287 	if (vx_waitl(vp)) {
2288 		VI_UNLOCK(vp);
2289 		return;
2290 	}
2291 
2292 	vx_lock(vp);
2293 
2294 	/*
2295 	 * Check to see if the vnode is in use. If so we have to reference it
2296 	 * before we clean it out so that its count cannot fall to zero and
2297 	 * generate a race against ourselves to recycle it.
2298 	 */
2299 	if ((active = vp->v_usecount))
2300 		v_incr_usecount(vp, 1);
2301 
2302 	/*
2303 	 * Even if the count is zero, the VOP_INACTIVE routine may still
2304 	 * have the object locked while it cleans it out. The VOP_LOCK
2305 	 * ensures that the VOP_INACTIVE routine is done with its work.
2306 	 * For active vnodes, it ensures that no other activity can
2307 	 * occur while the underlying object is being cleaned out.
2308 	 */
2309 	VOP_LOCK(vp, LK_DRAIN | LK_INTERLOCK, td);
2310 
2311 	/*
2312 	 * Clean out any buffers associated with the vnode.
2313 	 * If the flush fails, just toss the buffers.
2314 	 */
2315 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd));
2316 		(void) vn_write_suspend_wait(vp, NULL, V_WAIT);
2317 	if (vinvalbuf(vp, V_SAVE, td, 0, 0) != 0)
2318 		vinvalbuf(vp, 0, td, 0, 0);
2319 
2320 	/*
2321 	 * Any other processes trying to obtain this lock must first
2322 	 * wait for VXLOCK to clear, then call the new lock operation.
2323 	 */
2324 	VOP_UNLOCK(vp, 0, td);
2325 
2326 	/*
2327 	 * If purging an active vnode, it must be closed and
2328 	 * deactivated before being reclaimed. Note that the
2329 	 * VOP_INACTIVE will unlock the vnode.
2330 	 */
2331 	if (active) {
2332 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
2333 		VI_LOCK(vp);
2334 		if ((vp->v_iflag & VI_DOINGINACT) == 0) {
2335 			VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2336 			    ("vclean: recursed on VI_DOINGINACT"));
2337 			vp->v_iflag |= VI_DOINGINACT;
2338 			VI_UNLOCK(vp);
2339 			if (vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
2340 				panic("vclean: cannot relock.");
2341 			VOP_INACTIVE(vp, td);
2342 			VI_LOCK(vp);
2343 			VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2344 			    ("vclean: lost VI_DOINGINACT"));
2345 			vp->v_iflag &= ~VI_DOINGINACT;
2346 		}
2347 		VI_UNLOCK(vp);
2348 	}
2349 	/*
2350 	 * Reclaim the vnode.
2351 	 */
2352 	if (VOP_RECLAIM(vp, td))
2353 		panic("vclean: cannot reclaim");
2354 
2355 	VNASSERT(vp->v_object == NULL, vp,
2356 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
2357 
2358 	if (active) {
2359 		/*
2360 		 * Inline copy of vrele() since VOP_INACTIVE
2361 		 * has already been called.
2362 		 */
2363 		VI_LOCK(vp);
2364 		v_incr_usecount(vp, -1);
2365 		if (vp->v_usecount <= 0) {
2366 #ifdef INVARIANTS
2367 			if (vp->v_usecount < 0 || vp->v_writecount != 0) {
2368 				vprint("vclean: bad ref count", vp);
2369 				panic("vclean: ref cnt");
2370 			}
2371 #endif
2372 			if (VSHOULDFREE(vp))
2373 				vfree(vp);
2374 		}
2375 		VI_UNLOCK(vp);
2376 	}
2377 	/*
2378 	 * Delete from old mount point vnode list.
2379 	 */
2380 	delmntque(vp);
2381 	cache_purge(vp);
2382 	VI_LOCK(vp);
2383 	if (VSHOULDFREE(vp))
2384 		vfree(vp);
2385 
2386 	/*
2387 	 * Done with purge, reset to the standard lock and
2388 	 * notify sleepers of the grim news.
2389 	 */
2390 	vp->v_vnlock = &vp->v_lock;
2391 	vp->v_op = &dead_vnodeops;
2392 	vp->v_tag = "none";
2393 
2394 	VI_UNLOCK(vp);
2395 
2396 	/*
2397 	 * If special device, remove it from special device alias list
2398 	 * if it is on one.
2399 	 */
2400 	VI_LOCK(vp);
2401 
2402 	/*
2403 	 * If it is on the freelist and not already at the head,
2404 	 * move it to the head of the list. The test of the
2405 	 * VDOOMED flag and the reference count of zero is because
2406 	 * it will be removed from the free list by getnewvnode,
2407 	 * but will not have its reference count incremented until
2408 	 * after calling vgone. If the reference count were
2409 	 * incremented first, vgone would (incorrectly) try to
2410 	 * close the previous instance of the underlying object.
2411 	 */
2412 	if (vp->v_usecount == 0 && !(vp->v_iflag & VI_DOOMED)) {
2413 		mtx_lock(&vnode_free_list_mtx);
2414 		if (vp->v_iflag & VI_FREE) {
2415 			TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2416 		} else {
2417 			vp->v_iflag |= VI_FREE;
2418 			freevnodes++;
2419 		}
2420 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2421 		mtx_unlock(&vnode_free_list_mtx);
2422 	}
2423 
2424 	vp->v_type = VBAD;
2425 	vx_unlock(vp);
2426 	VI_UNLOCK(vp);
2427 }
2428 
2429 /*
2430  * Calculate the total number of references to a special device.
2431  */
2432 int
2433 vcount(vp)
2434 	struct vnode *vp;
2435 {
2436 	int count;
2437 
2438 	dev_lock();
2439 	count = vp->v_rdev->si_usecount;
2440 	dev_unlock();
2441 	return (count);
2442 }
2443 
2444 /*
2445  * Same as above, but using the struct cdev *as argument
2446  */
2447 int
2448 count_dev(dev)
2449 	struct cdev *dev;
2450 {
2451 	int count;
2452 
2453 	dev_lock();
2454 	count = dev->si_usecount;
2455 	dev_unlock();
2456 	return(count);
2457 }
2458 
2459 /*
2460  * Print out a description of a vnode.
2461  */
2462 static char *typename[] =
2463 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"};
2464 
2465 void
2466 vn_printf(struct vnode *vp, const char *fmt, ...)
2467 {
2468 	va_list ap;
2469 	char buf[96];
2470 
2471 	va_start(ap, fmt);
2472 	vprintf(fmt, ap);
2473 	va_end(ap);
2474 	printf("%p: ", (void *)vp);
2475 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
2476 	printf("    usecount %d, writecount %d, refcount %d mountedhere %p\n",
2477 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere);
2478 	buf[0] = '\0';
2479 	buf[1] = '\0';
2480 	if (vp->v_vflag & VV_ROOT)
2481 		strcat(buf, "|VV_ROOT");
2482 	if (vp->v_vflag & VV_TEXT)
2483 		strcat(buf, "|VV_TEXT");
2484 	if (vp->v_vflag & VV_SYSTEM)
2485 		strcat(buf, "|VV_SYSTEM");
2486 	if (vp->v_iflag & VI_XLOCK)
2487 		strcat(buf, "|VI_XLOCK");
2488 	if (vp->v_iflag & VI_XWANT)
2489 		strcat(buf, "|VI_XWANT");
2490 	if (vp->v_iflag & VI_DOOMED)
2491 		strcat(buf, "|VI_DOOMED");
2492 	if (vp->v_iflag & VI_FREE)
2493 		strcat(buf, "|VI_FREE");
2494 	printf("    flags (%s)\n", buf + 1);
2495 	if (mtx_owned(VI_MTX(vp)))
2496 		printf(" VI_LOCKed");
2497 	if (vp->v_object != NULL);
2498 		printf("    v_object %p\n", vp->v_object);
2499 	printf("    ");
2500 	lockmgr_printinfo(vp->v_vnlock);
2501 	printf("\n");
2502 	if (vp->v_data != NULL)
2503 		VOP_PRINT(vp);
2504 }
2505 
2506 #ifdef DDB
2507 #include <ddb/ddb.h>
2508 /*
2509  * List all of the locked vnodes in the system.
2510  * Called when debugging the kernel.
2511  */
2512 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
2513 {
2514 	struct mount *mp, *nmp;
2515 	struct vnode *vp;
2516 
2517 	/*
2518 	 * Note: because this is DDB, we can't obey the locking semantics
2519 	 * for these structures, which means we could catch an inconsistent
2520 	 * state and dereference a nasty pointer.  Not much to be done
2521 	 * about that.
2522 	 */
2523 	printf("Locked vnodes\n");
2524 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
2525 		nmp = TAILQ_NEXT(mp, mnt_list);
2526 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2527 			if (VOP_ISLOCKED(vp, NULL))
2528 				vprint("", vp);
2529 		}
2530 		nmp = TAILQ_NEXT(mp, mnt_list);
2531 	}
2532 }
2533 #endif
2534 
2535 /*
2536  * Fill in a struct xvfsconf based on a struct vfsconf.
2537  */
2538 static void
2539 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp)
2540 {
2541 
2542 	strcpy(xvfsp->vfc_name, vfsp->vfc_name);
2543 	xvfsp->vfc_typenum = vfsp->vfc_typenum;
2544 	xvfsp->vfc_refcount = vfsp->vfc_refcount;
2545 	xvfsp->vfc_flags = vfsp->vfc_flags;
2546 	/*
2547 	 * These are unused in userland, we keep them
2548 	 * to not break binary compatibility.
2549 	 */
2550 	xvfsp->vfc_vfsops = NULL;
2551 	xvfsp->vfc_next = NULL;
2552 }
2553 
2554 /*
2555  * Top level filesystem related information gathering.
2556  */
2557 static int
2558 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
2559 {
2560 	struct vfsconf *vfsp;
2561 	struct xvfsconf xvfsp;
2562 	int error;
2563 
2564 	error = 0;
2565 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2566 		vfsconf2x(vfsp, &xvfsp);
2567 		error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp);
2568 		if (error)
2569 			break;
2570 	}
2571 	return (error);
2572 }
2573 
2574 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist,
2575     "S,xvfsconf", "List of all configured filesystems");
2576 
2577 #ifndef BURN_BRIDGES
2578 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
2579 
2580 static int
2581 vfs_sysctl(SYSCTL_HANDLER_ARGS)
2582 {
2583 	int *name = (int *)arg1 - 1;	/* XXX */
2584 	u_int namelen = arg2 + 1;	/* XXX */
2585 	struct vfsconf *vfsp;
2586 	struct xvfsconf xvfsp;
2587 
2588 	printf("WARNING: userland calling deprecated sysctl, "
2589 	    "please rebuild world\n");
2590 
2591 #if 1 || defined(COMPAT_PRELITE2)
2592 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
2593 	if (namelen == 1)
2594 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
2595 #endif
2596 
2597 	switch (name[1]) {
2598 	case VFS_MAXTYPENUM:
2599 		if (namelen != 2)
2600 			return (ENOTDIR);
2601 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
2602 	case VFS_CONF:
2603 		if (namelen != 3)
2604 			return (ENOTDIR);	/* overloaded */
2605 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list)
2606 			if (vfsp->vfc_typenum == name[2])
2607 				break;
2608 		if (vfsp == NULL)
2609 			return (EOPNOTSUPP);
2610 		vfsconf2x(vfsp, &xvfsp);
2611 		return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
2612 	}
2613 	return (EOPNOTSUPP);
2614 }
2615 
2616 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP,
2617 	vfs_sysctl, "Generic filesystem");
2618 
2619 #if 1 || defined(COMPAT_PRELITE2)
2620 
2621 static int
2622 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
2623 {
2624 	int error;
2625 	struct vfsconf *vfsp;
2626 	struct ovfsconf ovfs;
2627 
2628 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2629 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
2630 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
2631 		ovfs.vfc_index = vfsp->vfc_typenum;
2632 		ovfs.vfc_refcount = vfsp->vfc_refcount;
2633 		ovfs.vfc_flags = vfsp->vfc_flags;
2634 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
2635 		if (error)
2636 			return error;
2637 	}
2638 	return 0;
2639 }
2640 
2641 #endif /* 1 || COMPAT_PRELITE2 */
2642 #endif /* !BURN_BRIDGES */
2643 
2644 #define KINFO_VNODESLOP		10
2645 #ifdef notyet
2646 /*
2647  * Dump vnode list (via sysctl).
2648  */
2649 /* ARGSUSED */
2650 static int
2651 sysctl_vnode(SYSCTL_HANDLER_ARGS)
2652 {
2653 	struct xvnode *xvn;
2654 	struct thread *td = req->td;
2655 	struct mount *mp;
2656 	struct vnode *vp;
2657 	int error, len, n;
2658 
2659 	/*
2660 	 * Stale numvnodes access is not fatal here.
2661 	 */
2662 	req->lock = 0;
2663 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
2664 	if (!req->oldptr)
2665 		/* Make an estimate */
2666 		return (SYSCTL_OUT(req, 0, len));
2667 
2668 	error = sysctl_wire_old_buffer(req, 0);
2669 	if (error != 0)
2670 		return (error);
2671 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
2672 	n = 0;
2673 	mtx_lock(&mountlist_mtx);
2674 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
2675 		if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td))
2676 			continue;
2677 		MNT_ILOCK(mp);
2678 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2679 			if (n == len)
2680 				break;
2681 			vref(vp);
2682 			xvn[n].xv_size = sizeof *xvn;
2683 			xvn[n].xv_vnode = vp;
2684 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
2685 			XV_COPY(usecount);
2686 			XV_COPY(writecount);
2687 			XV_COPY(holdcnt);
2688 			XV_COPY(id);
2689 			XV_COPY(mount);
2690 			XV_COPY(numoutput);
2691 			XV_COPY(type);
2692 #undef XV_COPY
2693 			xvn[n].xv_flag = vp->v_vflag;
2694 
2695 			switch (vp->v_type) {
2696 			case VREG:
2697 			case VDIR:
2698 			case VLNK:
2699 				break;
2700 			case VBLK:
2701 			case VCHR:
2702 				if (vp->v_rdev == NULL) {
2703 					vrele(vp);
2704 					continue;
2705 				}
2706 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
2707 				break;
2708 			case VSOCK:
2709 				xvn[n].xv_socket = vp->v_socket;
2710 				break;
2711 			case VFIFO:
2712 				xvn[n].xv_fifo = vp->v_fifoinfo;
2713 				break;
2714 			case VNON:
2715 			case VBAD:
2716 			default:
2717 				/* shouldn't happen? */
2718 				vrele(vp);
2719 				continue;
2720 			}
2721 			vrele(vp);
2722 			++n;
2723 		}
2724 		MNT_IUNLOCK(mp);
2725 		mtx_lock(&mountlist_mtx);
2726 		vfs_unbusy(mp, td);
2727 		if (n == len)
2728 			break;
2729 	}
2730 	mtx_unlock(&mountlist_mtx);
2731 
2732 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
2733 	free(xvn, M_TEMP);
2734 	return (error);
2735 }
2736 
2737 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD,
2738 	0, 0, sysctl_vnode, "S,xvnode", "");
2739 #endif
2740 
2741 /*
2742  * Unmount all filesystems. The list is traversed in reverse order
2743  * of mounting to avoid dependencies.
2744  */
2745 void
2746 vfs_unmountall()
2747 {
2748 	struct mount *mp;
2749 	struct thread *td;
2750 	int error;
2751 
2752 	if (curthread != NULL)
2753 		td = curthread;
2754 	else
2755 		td = FIRST_THREAD_IN_PROC(initproc); /* XXX XXX proc0? */
2756 	/*
2757 	 * Since this only runs when rebooting, it is not interlocked.
2758 	 */
2759 	while(!TAILQ_EMPTY(&mountlist)) {
2760 		mp = TAILQ_LAST(&mountlist, mntlist);
2761 		error = dounmount(mp, MNT_FORCE, td);
2762 		if (error) {
2763 			TAILQ_REMOVE(&mountlist, mp, mnt_list);
2764 			printf("unmount of %s failed (",
2765 			    mp->mnt_stat.f_mntonname);
2766 			if (error == EBUSY)
2767 				printf("BUSY)\n");
2768 			else
2769 				printf("%d)\n", error);
2770 		} else {
2771 			/* The unmount has removed mp from the mountlist */
2772 		}
2773 	}
2774 }
2775 
2776 /*
2777  * perform msync on all vnodes under a mount point
2778  * the mount point must be locked.
2779  */
2780 void
2781 vfs_msync(struct mount *mp, int flags)
2782 {
2783 	struct vnode *vp, *nvp;
2784 	struct vm_object *obj;
2785 	int tries;
2786 
2787 	tries = 5;
2788 	MNT_ILOCK(mp);
2789 loop:
2790 	TAILQ_FOREACH_SAFE(vp, &mp->mnt_nvnodelist, v_nmntvnodes, nvp) {
2791 		if (vp->v_mount != mp) {
2792 			if (--tries > 0)
2793 				goto loop;
2794 			break;
2795 		}
2796 
2797 		VI_LOCK(vp);
2798 		if (vp->v_iflag & VI_XLOCK) {
2799 			VI_UNLOCK(vp);
2800 			continue;
2801 		}
2802 
2803 		if ((vp->v_iflag & VI_OBJDIRTY) &&
2804 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) {
2805 			MNT_IUNLOCK(mp);
2806 			if (!vget(vp,
2807 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
2808 			    curthread)) {
2809 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
2810 					vput(vp);
2811 					MNT_ILOCK(mp);
2812 					continue;
2813 				}
2814 
2815 				obj = vp->v_object;
2816 				if (obj != NULL) {
2817 					VM_OBJECT_LOCK(obj);
2818 					vm_object_page_clean(obj, 0, 0,
2819 					    flags == MNT_WAIT ?
2820 					    OBJPC_SYNC : OBJPC_NOSYNC);
2821 					VM_OBJECT_UNLOCK(obj);
2822 				}
2823 				vput(vp);
2824 			}
2825 			MNT_ILOCK(mp);
2826 			if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) {
2827 				if (--tries > 0)
2828 					goto loop;
2829 				break;
2830 			}
2831 		} else
2832 			VI_UNLOCK(vp);
2833 	}
2834 	MNT_IUNLOCK(mp);
2835 }
2836 
2837 /*
2838  * Mark a vnode as free, putting it up for recycling.
2839  */
2840 void
2841 vfree(struct vnode *vp)
2842 {
2843 
2844 	ASSERT_VI_LOCKED(vp, "vfree");
2845 	mtx_lock(&vnode_free_list_mtx);
2846 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, ("vnode already free"));
2847 	if (vp->v_iflag & VI_AGE) {
2848 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2849 	} else {
2850 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
2851 	}
2852 	freevnodes++;
2853 	mtx_unlock(&vnode_free_list_mtx);
2854 	vp->v_iflag &= ~VI_AGE;
2855 	vp->v_iflag |= VI_FREE;
2856 }
2857 
2858 /*
2859  * Opposite of vfree() - mark a vnode as in use.
2860  */
2861 static void
2862 vbusy(struct vnode *vp)
2863 {
2864 
2865 	ASSERT_VI_LOCKED(vp, "vbusy");
2866 	VNASSERT((vp->v_iflag & VI_FREE) != 0, vp, ("vnode not free"));
2867 
2868 	mtx_lock(&vnode_free_list_mtx);
2869 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2870 	freevnodes--;
2871 	mtx_unlock(&vnode_free_list_mtx);
2872 
2873 	vp->v_iflag &= ~(VI_FREE|VI_AGE);
2874 }
2875 
2876 /*
2877  * Initalize per-vnode helper structure to hold poll-related state.
2878  */
2879 void
2880 v_addpollinfo(struct vnode *vp)
2881 {
2882 	struct vpollinfo *vi;
2883 
2884 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK);
2885 	if (vp->v_pollinfo != NULL) {
2886 		uma_zfree(vnodepoll_zone, vi);
2887 		return;
2888 	}
2889 	vp->v_pollinfo = vi;
2890 	mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
2891 	knlist_init(&vp->v_pollinfo->vpi_selinfo.si_note,
2892 	    &vp->v_pollinfo->vpi_lock);
2893 }
2894 
2895 /*
2896  * Record a process's interest in events which might happen to
2897  * a vnode.  Because poll uses the historic select-style interface
2898  * internally, this routine serves as both the ``check for any
2899  * pending events'' and the ``record my interest in future events''
2900  * functions.  (These are done together, while the lock is held,
2901  * to avoid race conditions.)
2902  */
2903 int
2904 vn_pollrecord(vp, td, events)
2905 	struct vnode *vp;
2906 	struct thread *td;
2907 	short events;
2908 {
2909 
2910 	if (vp->v_pollinfo == NULL)
2911 		v_addpollinfo(vp);
2912 	mtx_lock(&vp->v_pollinfo->vpi_lock);
2913 	if (vp->v_pollinfo->vpi_revents & events) {
2914 		/*
2915 		 * This leaves events we are not interested
2916 		 * in available for the other process which
2917 		 * which presumably had requested them
2918 		 * (otherwise they would never have been
2919 		 * recorded).
2920 		 */
2921 		events &= vp->v_pollinfo->vpi_revents;
2922 		vp->v_pollinfo->vpi_revents &= ~events;
2923 
2924 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
2925 		return events;
2926 	}
2927 	vp->v_pollinfo->vpi_events |= events;
2928 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
2929 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
2930 	return 0;
2931 }
2932 
2933 /*
2934  * Routine to create and manage a filesystem syncer vnode.
2935  */
2936 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
2937 static int	sync_fsync(struct  vop_fsync_args *);
2938 static int	sync_inactive(struct  vop_inactive_args *);
2939 static int	sync_reclaim(struct  vop_reclaim_args *);
2940 
2941 static struct vop_vector sync_vnodeops = {
2942 	.vop_bypass =	VOP_EOPNOTSUPP,
2943 	.vop_close =	sync_close,		/* close */
2944 	.vop_fsync =	sync_fsync,		/* fsync */
2945 	.vop_inactive =	sync_inactive,	/* inactive */
2946 	.vop_reclaim =	sync_reclaim,	/* reclaim */
2947 	.vop_lock =	vop_stdlock,	/* lock */
2948 	.vop_unlock =	vop_stdunlock,	/* unlock */
2949 	.vop_islocked =	vop_stdislocked,	/* islocked */
2950 };
2951 
2952 /*
2953  * Create a new filesystem syncer vnode for the specified mount point.
2954  */
2955 int
2956 vfs_allocate_syncvnode(mp)
2957 	struct mount *mp;
2958 {
2959 	struct vnode *vp;
2960 	static long start, incr, next;
2961 	int error;
2962 
2963 	/* Allocate a new vnode */
2964 	if ((error = getnewvnode("syncer", mp, &sync_vnodeops, &vp)) != 0) {
2965 		mp->mnt_syncer = NULL;
2966 		return (error);
2967 	}
2968 	vp->v_type = VNON;
2969 	/*
2970 	 * Place the vnode onto the syncer worklist. We attempt to
2971 	 * scatter them about on the list so that they will go off
2972 	 * at evenly distributed times even if all the filesystems
2973 	 * are mounted at once.
2974 	 */
2975 	next += incr;
2976 	if (next == 0 || next > syncer_maxdelay) {
2977 		start /= 2;
2978 		incr /= 2;
2979 		if (start == 0) {
2980 			start = syncer_maxdelay / 2;
2981 			incr = syncer_maxdelay;
2982 		}
2983 		next = start;
2984 	}
2985 	VI_LOCK(vp);
2986 	vn_syncer_add_to_worklist(&vp->v_bufobj,
2987 	    syncdelay > 0 ? next % syncdelay : 0);
2988 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
2989 	mtx_lock(&sync_mtx);
2990 	sync_vnode_count++;
2991 	mtx_unlock(&sync_mtx);
2992 	VI_UNLOCK(vp);
2993 	mp->mnt_syncer = vp;
2994 	return (0);
2995 }
2996 
2997 /*
2998  * Do a lazy sync of the filesystem.
2999  */
3000 static int
3001 sync_fsync(ap)
3002 	struct vop_fsync_args /* {
3003 		struct vnode *a_vp;
3004 		struct ucred *a_cred;
3005 		int a_waitfor;
3006 		struct thread *a_td;
3007 	} */ *ap;
3008 {
3009 	struct vnode *syncvp = ap->a_vp;
3010 	struct mount *mp = syncvp->v_mount;
3011 	struct thread *td = ap->a_td;
3012 	int error, asyncflag;
3013 	struct bufobj *bo;
3014 
3015 	/*
3016 	 * We only need to do something if this is a lazy evaluation.
3017 	 */
3018 	if (ap->a_waitfor != MNT_LAZY)
3019 		return (0);
3020 
3021 	/*
3022 	 * Move ourselves to the back of the sync list.
3023 	 */
3024 	bo = &syncvp->v_bufobj;
3025 	BO_LOCK(bo);
3026 	vn_syncer_add_to_worklist(bo, syncdelay);
3027 	BO_UNLOCK(bo);
3028 
3029 	/*
3030 	 * Walk the list of vnodes pushing all that are dirty and
3031 	 * not already on the sync list.
3032 	 */
3033 	mtx_lock(&mountlist_mtx);
3034 	if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) {
3035 		mtx_unlock(&mountlist_mtx);
3036 		return (0);
3037 	}
3038 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
3039 		vfs_unbusy(mp, td);
3040 		return (0);
3041 	}
3042 	asyncflag = mp->mnt_flag & MNT_ASYNC;
3043 	mp->mnt_flag &= ~MNT_ASYNC;
3044 	vfs_msync(mp, MNT_NOWAIT);
3045 	error = VFS_SYNC(mp, MNT_LAZY, td);
3046 	if (asyncflag)
3047 		mp->mnt_flag |= MNT_ASYNC;
3048 	vn_finished_write(mp);
3049 	vfs_unbusy(mp, td);
3050 	return (error);
3051 }
3052 
3053 /*
3054  * The syncer vnode is no referenced.
3055  */
3056 static int
3057 sync_inactive(ap)
3058 	struct vop_inactive_args /* {
3059 		struct vnode *a_vp;
3060 		struct thread *a_td;
3061 	} */ *ap;
3062 {
3063 
3064 	VOP_UNLOCK(ap->a_vp, 0, ap->a_td);
3065 	vgone(ap->a_vp);
3066 	return (0);
3067 }
3068 
3069 /*
3070  * The syncer vnode is no longer needed and is being decommissioned.
3071  *
3072  * Modifications to the worklist must be protected by sync_mtx.
3073  */
3074 static int
3075 sync_reclaim(ap)
3076 	struct vop_reclaim_args /* {
3077 		struct vnode *a_vp;
3078 	} */ *ap;
3079 {
3080 	struct vnode *vp = ap->a_vp;
3081 	struct bufobj *bo;
3082 
3083 	VI_LOCK(vp);
3084 	bo = &vp->v_bufobj;
3085 	vp->v_mount->mnt_syncer = NULL;
3086 	if (bo->bo_flag & BO_ONWORKLST) {
3087 		mtx_lock(&sync_mtx);
3088 		LIST_REMOVE(bo, bo_synclist);
3089  		syncer_worklist_len--;
3090 		sync_vnode_count--;
3091 		mtx_unlock(&sync_mtx);
3092 		bo->bo_flag &= ~BO_ONWORKLST;
3093 	}
3094 	VI_UNLOCK(vp);
3095 
3096 	return (0);
3097 }
3098 
3099 /*
3100  * Check if vnode represents a disk device
3101  */
3102 int
3103 vn_isdisk(vp, errp)
3104 	struct vnode *vp;
3105 	int *errp;
3106 {
3107 	int error;
3108 
3109 	error = 0;
3110 	dev_lock();
3111 	if (vp->v_type != VCHR)
3112 		error = ENOTBLK;
3113 	else if (vp->v_rdev == NULL)
3114 		error = ENXIO;
3115 	else if (vp->v_rdev->si_devsw == NULL)
3116 		error = ENXIO;
3117 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
3118 		error = ENOTBLK;
3119 	dev_unlock();
3120 	if (errp != NULL)
3121 		*errp = error;
3122 	return (error == 0);
3123 }
3124 
3125 /*
3126  * Free data allocated by namei(); see namei(9) for details.
3127  */
3128 void
3129 NDFREE(ndp, flags)
3130      struct nameidata *ndp;
3131      const u_int flags;
3132 {
3133 
3134 	if (!(flags & NDF_NO_FREE_PNBUF) &&
3135 	    (ndp->ni_cnd.cn_flags & HASBUF)) {
3136 		uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf);
3137 		ndp->ni_cnd.cn_flags &= ~HASBUF;
3138 	}
3139 	if (!(flags & NDF_NO_DVP_UNLOCK) &&
3140 	    (ndp->ni_cnd.cn_flags & LOCKPARENT) &&
3141 	    ndp->ni_dvp != ndp->ni_vp)
3142 		VOP_UNLOCK(ndp->ni_dvp, 0, ndp->ni_cnd.cn_thread);
3143 	if (!(flags & NDF_NO_DVP_RELE) &&
3144 	    (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) {
3145 		vrele(ndp->ni_dvp);
3146 		ndp->ni_dvp = NULL;
3147 	}
3148 	if (!(flags & NDF_NO_VP_UNLOCK) &&
3149 	    (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp)
3150 		VOP_UNLOCK(ndp->ni_vp, 0, ndp->ni_cnd.cn_thread);
3151 	if (!(flags & NDF_NO_VP_RELE) &&
3152 	    ndp->ni_vp) {
3153 		vrele(ndp->ni_vp);
3154 		ndp->ni_vp = NULL;
3155 	}
3156 	if (!(flags & NDF_NO_STARTDIR_RELE) &&
3157 	    (ndp->ni_cnd.cn_flags & SAVESTART)) {
3158 		vrele(ndp->ni_startdir);
3159 		ndp->ni_startdir = NULL;
3160 	}
3161 }
3162 
3163 /*
3164  * Common filesystem object access control check routine.  Accepts a
3165  * vnode's type, "mode", uid and gid, requested access mode, credentials,
3166  * and optional call-by-reference privused argument allowing vaccess()
3167  * to indicate to the caller whether privilege was used to satisfy the
3168  * request (obsoleted).  Returns 0 on success, or an errno on failure.
3169  */
3170 int
3171 vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused)
3172 	enum vtype type;
3173 	mode_t file_mode;
3174 	uid_t file_uid;
3175 	gid_t file_gid;
3176 	mode_t acc_mode;
3177 	struct ucred *cred;
3178 	int *privused;
3179 {
3180 	mode_t dac_granted;
3181 #ifdef CAPABILITIES
3182 	mode_t cap_granted;
3183 #endif
3184 
3185 	/*
3186 	 * Look for a normal, non-privileged way to access the file/directory
3187 	 * as requested.  If it exists, go with that.
3188 	 */
3189 
3190 	if (privused != NULL)
3191 		*privused = 0;
3192 
3193 	dac_granted = 0;
3194 
3195 	/* Check the owner. */
3196 	if (cred->cr_uid == file_uid) {
3197 		dac_granted |= VADMIN;
3198 		if (file_mode & S_IXUSR)
3199 			dac_granted |= VEXEC;
3200 		if (file_mode & S_IRUSR)
3201 			dac_granted |= VREAD;
3202 		if (file_mode & S_IWUSR)
3203 			dac_granted |= (VWRITE | VAPPEND);
3204 
3205 		if ((acc_mode & dac_granted) == acc_mode)
3206 			return (0);
3207 
3208 		goto privcheck;
3209 	}
3210 
3211 	/* Otherwise, check the groups (first match) */
3212 	if (groupmember(file_gid, cred)) {
3213 		if (file_mode & S_IXGRP)
3214 			dac_granted |= VEXEC;
3215 		if (file_mode & S_IRGRP)
3216 			dac_granted |= VREAD;
3217 		if (file_mode & S_IWGRP)
3218 			dac_granted |= (VWRITE | VAPPEND);
3219 
3220 		if ((acc_mode & dac_granted) == acc_mode)
3221 			return (0);
3222 
3223 		goto privcheck;
3224 	}
3225 
3226 	/* Otherwise, check everyone else. */
3227 	if (file_mode & S_IXOTH)
3228 		dac_granted |= VEXEC;
3229 	if (file_mode & S_IROTH)
3230 		dac_granted |= VREAD;
3231 	if (file_mode & S_IWOTH)
3232 		dac_granted |= (VWRITE | VAPPEND);
3233 	if ((acc_mode & dac_granted) == acc_mode)
3234 		return (0);
3235 
3236 privcheck:
3237 	if (!suser_cred(cred, SUSER_ALLOWJAIL)) {
3238 		/* XXX audit: privilege used */
3239 		if (privused != NULL)
3240 			*privused = 1;
3241 		return (0);
3242 	}
3243 
3244 #ifdef CAPABILITIES
3245 	/*
3246 	 * Build a capability mask to determine if the set of capabilities
3247 	 * satisfies the requirements when combined with the granted mask
3248 	 * from above.
3249 	 * For each capability, if the capability is required, bitwise
3250 	 * or the request type onto the cap_granted mask.
3251 	 */
3252 	cap_granted = 0;
3253 
3254 	if (type == VDIR) {
3255 		/*
3256 		 * For directories, use CAP_DAC_READ_SEARCH to satisfy
3257 		 * VEXEC requests, instead of CAP_DAC_EXECUTE.
3258 		 */
3259 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3260 		    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL))
3261 			cap_granted |= VEXEC;
3262 	} else {
3263 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3264 		    !cap_check(cred, NULL, CAP_DAC_EXECUTE, SUSER_ALLOWJAIL))
3265 			cap_granted |= VEXEC;
3266 	}
3267 
3268 	if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) &&
3269 	    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL))
3270 		cap_granted |= VREAD;
3271 
3272 	if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
3273 	    !cap_check(cred, NULL, CAP_DAC_WRITE, SUSER_ALLOWJAIL))
3274 		cap_granted |= (VWRITE | VAPPEND);
3275 
3276 	if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
3277 	    !cap_check(cred, NULL, CAP_FOWNER, SUSER_ALLOWJAIL))
3278 		cap_granted |= VADMIN;
3279 
3280 	if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) {
3281 		/* XXX audit: privilege used */
3282 		if (privused != NULL)
3283 			*privused = 1;
3284 		return (0);
3285 	}
3286 #endif
3287 
3288 	return ((acc_mode & VADMIN) ? EPERM : EACCES);
3289 }
3290 
3291 /*
3292  * Credential check based on process requesting service, and per-attribute
3293  * permissions.
3294  */
3295 int
3296 extattr_check_cred(struct vnode *vp, int attrnamespace,
3297     struct ucred *cred, struct thread *td, int access)
3298 {
3299 
3300 	/*
3301 	 * Kernel-invoked always succeeds.
3302 	 */
3303 	if (cred == NOCRED)
3304 		return (0);
3305 
3306 	/*
3307 	 * Do not allow privileged processes in jail to directly
3308 	 * manipulate system attributes.
3309 	 *
3310 	 * XXX What capability should apply here?
3311 	 * Probably CAP_SYS_SETFFLAG.
3312 	 */
3313 	switch (attrnamespace) {
3314 	case EXTATTR_NAMESPACE_SYSTEM:
3315 		/* Potentially should be: return (EPERM); */
3316 		return (suser_cred(cred, 0));
3317 	case EXTATTR_NAMESPACE_USER:
3318 		return (VOP_ACCESS(vp, access, cred, td));
3319 	default:
3320 		return (EPERM);
3321 	}
3322 }
3323 
3324 #ifdef DEBUG_VFS_LOCKS
3325 /*
3326  * This only exists to supress warnings from unlocked specfs accesses.  It is
3327  * no longer ok to have an unlocked VFS.
3328  */
3329 #define	IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD)
3330 
3331 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
3332 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, "");
3333 
3334 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
3335 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, "");
3336 
3337 int vfs_badlock_print = 1;	/* Print lock violations. */
3338 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, "");
3339 
3340 #ifdef KDB
3341 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
3342 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, "");
3343 #endif
3344 
3345 static void
3346 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
3347 {
3348 
3349 #ifdef KDB
3350 	if (vfs_badlock_backtrace)
3351 		kdb_backtrace();
3352 #endif
3353 	if (vfs_badlock_print)
3354 		printf("%s: %p %s\n", str, (void *)vp, msg);
3355 	if (vfs_badlock_ddb)
3356 		kdb_enter("lock violation");
3357 }
3358 
3359 void
3360 assert_vi_locked(struct vnode *vp, const char *str)
3361 {
3362 
3363 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
3364 		vfs_badlock("interlock is not locked but should be", str, vp);
3365 }
3366 
3367 void
3368 assert_vi_unlocked(struct vnode *vp, const char *str)
3369 {
3370 
3371 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
3372 		vfs_badlock("interlock is locked but should not be", str, vp);
3373 }
3374 
3375 void
3376 assert_vop_locked(struct vnode *vp, const char *str)
3377 {
3378 
3379 	if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0 &&
3380 	  !((vp->v_iflag & VI_XLOCK) && vp->v_vxthread == curthread))
3381 		vfs_badlock("is not locked but should be", str, vp);
3382 }
3383 
3384 void
3385 assert_vop_unlocked(struct vnode *vp, const char *str)
3386 {
3387 
3388 	if (vp && !IGNORE_LOCK(vp) &&
3389 	    VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE)
3390 		vfs_badlock("is locked but should not be", str, vp);
3391 }
3392 
3393 #if 0
3394 void
3395 assert_vop_elocked(struct vnode *vp, const char *str)
3396 {
3397 
3398 	if (vp && !IGNORE_LOCK(vp) &&
3399 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE)
3400 		vfs_badlock("is not exclusive locked but should be", str, vp);
3401 }
3402 
3403 void
3404 assert_vop_elocked_other(struct vnode *vp, const char *str)
3405 {
3406 
3407 	if (vp && !IGNORE_LOCK(vp) &&
3408 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER)
3409 		vfs_badlock("is not exclusive locked by another thread",
3410 		    str, vp);
3411 }
3412 
3413 void
3414 assert_vop_slocked(struct vnode *vp, const char *str)
3415 {
3416 
3417 	if (vp && !IGNORE_LOCK(vp) &&
3418 	    VOP_ISLOCKED(vp, curthread) != LK_SHARED)
3419 		vfs_badlock("is not locked shared but should be", str, vp);
3420 }
3421 #endif /* 0 */
3422 
3423 void
3424 vop_rename_pre(void *ap)
3425 {
3426 	struct vop_rename_args *a = ap;
3427 
3428 	if (a->a_tvp)
3429 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
3430 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
3431 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
3432 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
3433 
3434 	/* Check the source (from). */
3435 	if (a->a_tdvp != a->a_fdvp)
3436 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
3437 	if (a->a_tvp != a->a_fvp)
3438 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked");
3439 
3440 	/* Check the target. */
3441 	if (a->a_tvp)
3442 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
3443 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
3444 }
3445 
3446 void
3447 vop_strategy_pre(void *ap)
3448 {
3449 	struct vop_strategy_args *a;
3450 	struct buf *bp;
3451 
3452 	a = ap;
3453 	bp = a->a_bp;
3454 
3455 	/*
3456 	 * Cluster ops lock their component buffers but not the IO container.
3457 	 */
3458 	if ((bp->b_flags & B_CLUSTER) != 0)
3459 		return;
3460 
3461 	if (BUF_REFCNT(bp) < 1) {
3462 		if (vfs_badlock_print)
3463 			printf(
3464 			    "VOP_STRATEGY: bp is not locked but should be\n");
3465 		if (vfs_badlock_ddb)
3466 			kdb_enter("lock violation");
3467 	}
3468 }
3469 
3470 void
3471 vop_lookup_pre(void *ap)
3472 {
3473 	struct vop_lookup_args *a;
3474 	struct vnode *dvp;
3475 
3476 	a = ap;
3477 	dvp = a->a_dvp;
3478 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3479 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3480 }
3481 
3482 void
3483 vop_lookup_post(void *ap, int rc)
3484 {
3485 	struct vop_lookup_args *a;
3486 	struct componentname *cnp;
3487 	struct vnode *dvp;
3488 	struct vnode *vp;
3489 	int flags;
3490 
3491 	a = ap;
3492 	dvp = a->a_dvp;
3493 	cnp = a->a_cnp;
3494 	vp = *(a->a_vpp);
3495 	flags = cnp->cn_flags;
3496 
3497 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3498 
3499 	/*
3500 	 * If this is the last path component for this lookup and LOCKPARENT
3501 	 * is set, OR if there is an error the directory has to be locked.
3502 	 */
3503 	if ((flags & LOCKPARENT) && (flags & ISLASTCN))
3504 		ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (LOCKPARENT)");
3505 	else if (rc != 0)
3506 		ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP (error)");
3507 	else if (dvp != vp)
3508 		ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (dvp)");
3509 	if (flags & PDIRUNLOCK)
3510 		ASSERT_VOP_UNLOCKED(dvp, "VOP_LOOKUP (PDIRUNLOCK)");
3511 }
3512 
3513 void
3514 vop_lock_pre(void *ap)
3515 {
3516 	struct vop_lock_args *a = ap;
3517 
3518 	if (a->a_vp->v_iflag & VI_XLOCK &&
3519 	    a->a_vp->v_vxthread != curthread) {
3520 		vprint("vop_lock_pre:", a->a_vp);
3521 		panic("vop_lock_pre: locked while xlock held.\n");
3522 	}
3523 	if ((a->a_flags & LK_INTERLOCK) == 0)
3524 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3525 	else
3526 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
3527 }
3528 
3529 void
3530 vop_lock_post(void *ap, int rc)
3531 {
3532 	struct vop_lock_args *a = ap;
3533 
3534 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3535 	if (rc == 0)
3536 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
3537 }
3538 
3539 void
3540 vop_unlock_pre(void *ap)
3541 {
3542 	struct vop_unlock_args *a = ap;
3543 
3544 	if (a->a_flags & LK_INTERLOCK)
3545 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
3546 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
3547 }
3548 
3549 void
3550 vop_unlock_post(void *ap, int rc)
3551 {
3552 	struct vop_unlock_args *a = ap;
3553 
3554 	if (a->a_flags & LK_INTERLOCK)
3555 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
3556 }
3557 #endif /* DEBUG_VFS_LOCKS */
3558 
3559 static struct knlist fs_knlist;
3560 
3561 static void
3562 vfs_event_init(void *arg)
3563 {
3564 	knlist_init(&fs_knlist, NULL);
3565 }
3566 /* XXX - correct order? */
3567 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
3568 
3569 void
3570 vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused)
3571 {
3572 
3573 	KNOTE_UNLOCKED(&fs_knlist, event);
3574 }
3575 
3576 static int	filt_fsattach(struct knote *kn);
3577 static void	filt_fsdetach(struct knote *kn);
3578 static int	filt_fsevent(struct knote *kn, long hint);
3579 
3580 struct filterops fs_filtops =
3581 	{ 0, filt_fsattach, filt_fsdetach, filt_fsevent };
3582 
3583 static int
3584 filt_fsattach(struct knote *kn)
3585 {
3586 
3587 	kn->kn_flags |= EV_CLEAR;
3588 	knlist_add(&fs_knlist, kn, 0);
3589 	return (0);
3590 }
3591 
3592 static void
3593 filt_fsdetach(struct knote *kn)
3594 {
3595 
3596 	knlist_remove(&fs_knlist, kn, 0);
3597 }
3598 
3599 static int
3600 filt_fsevent(struct knote *kn, long hint)
3601 {
3602 
3603 	kn->kn_fflags |= hint;
3604 	return (kn->kn_fflags != 0);
3605 }
3606 
3607 static int
3608 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
3609 {
3610 	struct vfsidctl vc;
3611 	int error;
3612 	struct mount *mp;
3613 
3614 	error = SYSCTL_IN(req, &vc, sizeof(vc));
3615 	if (error)
3616 		return (error);
3617 	if (vc.vc_vers != VFS_CTL_VERS1)
3618 		return (EINVAL);
3619 	mp = vfs_getvfs(&vc.vc_fsid);
3620 	if (mp == NULL)
3621 		return (ENOENT);
3622 	/* ensure that a specific sysctl goes to the right filesystem. */
3623 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
3624 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
3625 		return (EINVAL);
3626 	}
3627 	VCTLTOREQ(&vc, req);
3628 	return (VFS_SYSCTL(mp, vc.vc_op, req));
3629 }
3630 
3631 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR,
3632         NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid");
3633 
3634 /*
3635  * Function to initialize a va_filerev field sensibly.
3636  * XXX: Wouldn't a random number make a lot more sense ??
3637  */
3638 u_quad_t
3639 init_va_filerev(void)
3640 {
3641 	struct bintime bt;
3642 
3643 	getbinuptime(&bt);
3644 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
3645 }
3646