xref: /freebsd/sys/kern/vfs_subr.c (revision 7afc53b8dfcc7d5897920ce6cc7e842fbb4ab813)
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	vbusy(struct vnode *vp);
92 static void	vdropl(struct vnode *vp);
93 static void	vinactive(struct vnode *, struct thread *);
94 static void	v_incr_usecount(struct vnode *, int);
95 static void	vfree(struct vnode *);
96 static void	vfreehead(struct vnode *);
97 static void	vnlru_free(int);
98 static void	vdestroy(struct vnode *);
99 
100 /*
101  * Enable Giant pushdown based on whether or not the vm is mpsafe in this
102  * build.  Without mpsafevm the buffer cache can not run Giant free.
103  */
104 #if defined(__alpha__) || defined(__amd64__) || defined(__i386__)
105 int mpsafe_vfs = 1;
106 #else
107 int mpsafe_vfs;
108 #endif
109 TUNABLE_INT("debug.mpsafevfs", &mpsafe_vfs);
110 SYSCTL_INT(_debug, OID_AUTO, mpsafevfs, CTLFLAG_RD, &mpsafe_vfs, 0,
111     "MPSAFE VFS");
112 
113 /*
114  * Number of vnodes in existence.  Increased whenever getnewvnode()
115  * allocates a new vnode, never decreased.
116  */
117 static unsigned long	numvnodes;
118 
119 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, "");
120 
121 /*
122  * Conversion tables for conversion from vnode types to inode formats
123  * and back.
124  */
125 enum vtype iftovt_tab[16] = {
126 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
127 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
128 };
129 int vttoif_tab[9] = {
130 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
131 	S_IFSOCK, S_IFIFO, S_IFMT,
132 };
133 
134 /*
135  * List of vnodes that are ready for recycling.
136  */
137 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
138 
139 /*
140  * Free vnode target.  Free vnodes may simply be files which have been stat'd
141  * but not read.  This is somewhat common, and a small cache of such files
142  * should be kept to avoid recreation costs.
143  */
144 static u_long wantfreevnodes;
145 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, "");
146 /* Number of vnodes in the free list. */
147 static u_long freevnodes;
148 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, "");
149 
150 /*
151  * Various variables used for debugging the new implementation of
152  * reassignbuf().
153  * XXX these are probably of (very) limited utility now.
154  */
155 static int reassignbufcalls;
156 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, "");
157 
158 /*
159  * Cache for the mount type id assigned to NFS.  This is used for
160  * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c.
161  */
162 int	nfs_mount_type = -1;
163 
164 /* To keep more than one thread at a time from running vfs_getnewfsid */
165 static struct mtx mntid_mtx;
166 
167 /*
168  * Lock for any access to the following:
169  *	vnode_free_list
170  *	numvnodes
171  *	freevnodes
172  */
173 static struct mtx vnode_free_list_mtx;
174 
175 /* Publicly exported FS */
176 struct nfs_public nfs_pub;
177 
178 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
179 static uma_zone_t vnode_zone;
180 static uma_zone_t vnodepoll_zone;
181 
182 /* Set to 1 to print out reclaim of active vnodes */
183 int	prtactive;
184 
185 /*
186  * The workitem queue.
187  *
188  * It is useful to delay writes of file data and filesystem metadata
189  * for tens of seconds so that quickly created and deleted files need
190  * not waste disk bandwidth being created and removed. To realize this,
191  * we append vnodes to a "workitem" queue. When running with a soft
192  * updates implementation, most pending metadata dependencies should
193  * not wait for more than a few seconds. Thus, mounted on block devices
194  * are delayed only about a half the time that file data is delayed.
195  * Similarly, directory updates are more critical, so are only delayed
196  * about a third the time that file data is delayed. Thus, there are
197  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
198  * one each second (driven off the filesystem syncer process). The
199  * syncer_delayno variable indicates the next queue that is to be processed.
200  * Items that need to be processed soon are placed in this queue:
201  *
202  *	syncer_workitem_pending[syncer_delayno]
203  *
204  * A delay of fifteen seconds is done by placing the request fifteen
205  * entries later in the queue:
206  *
207  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
208  *
209  */
210 static int syncer_delayno;
211 static long syncer_mask;
212 LIST_HEAD(synclist, bufobj);
213 static struct synclist *syncer_workitem_pending;
214 /*
215  * The sync_mtx protects:
216  *	bo->bo_synclist
217  *	sync_vnode_count
218  *	syncer_delayno
219  *	syncer_state
220  *	syncer_workitem_pending
221  *	syncer_worklist_len
222  *	rushjob
223  */
224 static struct mtx sync_mtx;
225 
226 #define SYNCER_MAXDELAY		32
227 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
228 static int syncdelay = 30;		/* max time to delay syncing data */
229 static int filedelay = 30;		/* time to delay syncing files */
230 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, "");
231 static int dirdelay = 29;		/* time to delay syncing directories */
232 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, "");
233 static int metadelay = 28;		/* time to delay syncing metadata */
234 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, "");
235 static int rushjob;		/* number of slots to run ASAP */
236 static int stat_rush_requests;	/* number of times I/O speeded up */
237 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, "");
238 
239 /*
240  * When shutting down the syncer, run it at four times normal speed.
241  */
242 #define SYNCER_SHUTDOWN_SPEEDUP		4
243 static int sync_vnode_count;
244 static int syncer_worklist_len;
245 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
246     syncer_state;
247 
248 /*
249  * Number of vnodes we want to exist at any one time.  This is mostly used
250  * to size hash tables in vnode-related code.  It is normally not used in
251  * getnewvnode(), as wantfreevnodes is normally nonzero.)
252  *
253  * XXX desiredvnodes is historical cruft and should not exist.
254  */
255 int desiredvnodes;
256 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW,
257     &desiredvnodes, 0, "Maximum number of vnodes");
258 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
259     &wantfreevnodes, 0, "Minimum number of vnodes (legacy)");
260 static int vnlru_nowhere;
261 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
262     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
263 
264 /* Hook for calling soft updates. */
265 int (*softdep_process_worklist_hook)(struct mount *);
266 
267 /*
268  * Macros to control when a vnode is freed and recycled.  All require
269  * the vnode interlock.
270  */
271 #define VCANRECYCLE(vp) (((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt)
272 #define VSHOULDFREE(vp) (!((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt)
273 #define VSHOULDBUSY(vp) (((vp)->v_iflag & VI_FREE) && (vp)->v_holdcnt)
274 
275 
276 /*
277  * Initialize the vnode management data structures.
278  */
279 #ifndef	MAXVNODES_MAX
280 #define	MAXVNODES_MAX	100000
281 #endif
282 static void
283 vntblinit(void *dummy __unused)
284 {
285 
286 	/*
287 	 * Desiredvnodes is a function of the physical memory size and
288 	 * the kernel's heap size.  Specifically, desiredvnodes scales
289 	 * in proportion to the physical memory size until two fifths
290 	 * of the kernel's heap size is consumed by vnodes and vm
291 	 * objects.
292 	 */
293 	desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size /
294 	    (5 * (sizeof(struct vm_object) + sizeof(struct vnode))));
295 	if (desiredvnodes > MAXVNODES_MAX) {
296 		if (bootverbose)
297 			printf("Reducing kern.maxvnodes %d -> %d\n",
298 			    desiredvnodes, MAXVNODES_MAX);
299 		desiredvnodes = MAXVNODES_MAX;
300 	}
301 	wantfreevnodes = desiredvnodes / 4;
302 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
303 	TAILQ_INIT(&vnode_free_list);
304 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
305 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
306 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
307 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
308 	      NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
309 	/*
310 	 * Initialize the filesystem syncer.
311 	 */
312 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
313 		&syncer_mask);
314 	syncer_maxdelay = syncer_mask + 1;
315 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
316 }
317 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL)
318 
319 
320 /*
321  * Mark a mount point as busy. Used to synchronize access and to delay
322  * unmounting. Interlock is not released on failure.
323  */
324 int
325 vfs_busy(mp, flags, interlkp, td)
326 	struct mount *mp;
327 	int flags;
328 	struct mtx *interlkp;
329 	struct thread *td;
330 {
331 	int lkflags;
332 
333 	MNT_ILOCK(mp);
334 	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
335 		if (flags & LK_NOWAIT) {
336 			MNT_IUNLOCK(mp);
337 			return (ENOENT);
338 		}
339 		if (interlkp)
340 			mtx_unlock(interlkp);
341 		mp->mnt_kern_flag |= MNTK_MWAIT;
342 		/*
343 		 * Since all busy locks are shared except the exclusive
344 		 * lock granted when unmounting, the only place that a
345 		 * wakeup needs to be done is at the release of the
346 		 * exclusive lock at the end of dounmount.
347 		 */
348 		msleep(mp, MNT_MTX(mp), PVFS|PDROP, "vfs_busy", 0);
349 		if (interlkp)
350 			mtx_lock(interlkp);
351 		return (ENOENT);
352 	}
353 	if (interlkp)
354 		mtx_unlock(interlkp);
355 	lkflags = LK_SHARED | LK_INTERLOCK;
356 	if (lockmgr(&mp->mnt_lock, lkflags, MNT_MTX(mp), td))
357 		panic("vfs_busy: unexpected lock failure");
358 	return (0);
359 }
360 
361 /*
362  * Free a busy filesystem.
363  */
364 void
365 vfs_unbusy(mp, td)
366 	struct mount *mp;
367 	struct thread *td;
368 {
369 
370 	lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td);
371 }
372 
373 /*
374  * Lookup a mount point by filesystem identifier.
375  */
376 struct mount *
377 vfs_getvfs(fsid)
378 	fsid_t *fsid;
379 {
380 	struct mount *mp;
381 
382 	mtx_lock(&mountlist_mtx);
383 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
384 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
385 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
386 			mtx_unlock(&mountlist_mtx);
387 			return (mp);
388 		}
389 	}
390 	mtx_unlock(&mountlist_mtx);
391 	return ((struct mount *) 0);
392 }
393 
394 /*
395  * Check if a user can access priveledged mount options.
396  */
397 int
398 vfs_suser(struct mount *mp, struct thread *td)
399 {
400 	int error;
401 
402 	if ((mp->mnt_flag & MNT_USER) == 0 ||
403 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
404 		if ((error = suser(td)) != 0)
405 			return (error);
406 	}
407 	return (0);
408 }
409 
410 /*
411  * Get a new unique fsid.  Try to make its val[0] unique, since this value
412  * will be used to create fake device numbers for stat().  Also try (but
413  * not so hard) make its val[0] unique mod 2^16, since some emulators only
414  * support 16-bit device numbers.  We end up with unique val[0]'s for the
415  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
416  *
417  * Keep in mind that several mounts may be running in parallel.  Starting
418  * the search one past where the previous search terminated is both a
419  * micro-optimization and a defense against returning the same fsid to
420  * different mounts.
421  */
422 void
423 vfs_getnewfsid(mp)
424 	struct mount *mp;
425 {
426 	static u_int16_t mntid_base;
427 	fsid_t tfsid;
428 	int mtype;
429 
430 	mtx_lock(&mntid_mtx);
431 	mtype = mp->mnt_vfc->vfc_typenum;
432 	tfsid.val[1] = mtype;
433 	mtype = (mtype & 0xFF) << 24;
434 	for (;;) {
435 		tfsid.val[0] = makedev(255,
436 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
437 		mntid_base++;
438 		if (vfs_getvfs(&tfsid) == NULL)
439 			break;
440 	}
441 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
442 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
443 	mtx_unlock(&mntid_mtx);
444 }
445 
446 /*
447  * Knob to control the precision of file timestamps:
448  *
449  *   0 = seconds only; nanoseconds zeroed.
450  *   1 = seconds and nanoseconds, accurate within 1/HZ.
451  *   2 = seconds and nanoseconds, truncated to microseconds.
452  * >=3 = seconds and nanoseconds, maximum precision.
453  */
454 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
455 
456 static int timestamp_precision = TSP_SEC;
457 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
458     &timestamp_precision, 0, "");
459 
460 /*
461  * Get a current timestamp.
462  */
463 void
464 vfs_timestamp(tsp)
465 	struct timespec *tsp;
466 {
467 	struct timeval tv;
468 
469 	switch (timestamp_precision) {
470 	case TSP_SEC:
471 		tsp->tv_sec = time_second;
472 		tsp->tv_nsec = 0;
473 		break;
474 	case TSP_HZ:
475 		getnanotime(tsp);
476 		break;
477 	case TSP_USEC:
478 		microtime(&tv);
479 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
480 		break;
481 	case TSP_NSEC:
482 	default:
483 		nanotime(tsp);
484 		break;
485 	}
486 }
487 
488 /*
489  * Set vnode attributes to VNOVAL
490  */
491 void
492 vattr_null(vap)
493 	struct vattr *vap;
494 {
495 
496 	vap->va_type = VNON;
497 	vap->va_size = VNOVAL;
498 	vap->va_bytes = VNOVAL;
499 	vap->va_mode = VNOVAL;
500 	vap->va_nlink = VNOVAL;
501 	vap->va_uid = VNOVAL;
502 	vap->va_gid = VNOVAL;
503 	vap->va_fsid = VNOVAL;
504 	vap->va_fileid = VNOVAL;
505 	vap->va_blocksize = VNOVAL;
506 	vap->va_rdev = VNOVAL;
507 	vap->va_atime.tv_sec = VNOVAL;
508 	vap->va_atime.tv_nsec = VNOVAL;
509 	vap->va_mtime.tv_sec = VNOVAL;
510 	vap->va_mtime.tv_nsec = VNOVAL;
511 	vap->va_ctime.tv_sec = VNOVAL;
512 	vap->va_ctime.tv_nsec = VNOVAL;
513 	vap->va_birthtime.tv_sec = VNOVAL;
514 	vap->va_birthtime.tv_nsec = VNOVAL;
515 	vap->va_flags = VNOVAL;
516 	vap->va_gen = VNOVAL;
517 	vap->va_vaflags = 0;
518 }
519 
520 /*
521  * This routine is called when we have too many vnodes.  It attempts
522  * to free <count> vnodes and will potentially free vnodes that still
523  * have VM backing store (VM backing store is typically the cause
524  * of a vnode blowout so we want to do this).  Therefore, this operation
525  * is not considered cheap.
526  *
527  * A number of conditions may prevent a vnode from being reclaimed.
528  * the buffer cache may have references on the vnode, a directory
529  * vnode may still have references due to the namei cache representing
530  * underlying files, or the vnode may be in active use.   It is not
531  * desireable to reuse such vnodes.  These conditions may cause the
532  * number of vnodes to reach some minimum value regardless of what
533  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
534  */
535 static int
536 vlrureclaim(struct mount *mp)
537 {
538 	struct vnode *vp;
539 	int done;
540 	int trigger;
541 	int usevnodes;
542 	int count;
543 
544 	/*
545 	 * Calculate the trigger point, don't allow user
546 	 * screwups to blow us up.   This prevents us from
547 	 * recycling vnodes with lots of resident pages.  We
548 	 * aren't trying to free memory, we are trying to
549 	 * free vnodes.
550 	 */
551 	usevnodes = desiredvnodes;
552 	if (usevnodes <= 0)
553 		usevnodes = 1;
554 	trigger = cnt.v_page_count * 2 / usevnodes;
555 
556 	done = 0;
557 	vn_start_write(NULL, &mp, V_WAIT);
558 	MNT_ILOCK(mp);
559 	count = mp->mnt_nvnodelistsize / 10 + 1;
560 	while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) {
561 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
562 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
563 
564 		if (vp->v_type != VNON &&
565 		    vp->v_type != VBAD &&
566 		    VI_TRYLOCK(vp)) {
567 			/* critical path opt */
568 			if (LIST_EMPTY(&(vp)->v_cache_src) &&
569 			    !(vp)->v_usecount &&
570 			    (vp->v_object == NULL ||
571 			    vp->v_object->resident_page_count < trigger)) {
572 				struct thread *td;
573 
574 				td = curthread;
575 				MNT_IUNLOCK(mp);
576 				VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE, td);
577 				if ((vp->v_iflag & VI_DOOMED) == 0)
578 					vgone(vp);
579 				VOP_UNLOCK(vp, 0, td);
580 				done++;
581 				MNT_ILOCK(mp);
582 			} else
583 				VI_UNLOCK(vp);
584 		}
585 		--count;
586 	}
587 	MNT_IUNLOCK(mp);
588 	vn_finished_write(mp);
589 	return done;
590 }
591 
592 /*
593  * Attempt to keep the free list at wantfreevnodes length.
594  */
595 static void
596 vnlru_free(int count)
597 {
598 	struct vnode *vp;
599 
600 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
601 	for (; count > 0; count--) {
602 		vp = TAILQ_FIRST(&vnode_free_list);
603 		/*
604 		 * The list can be modified while the free_list_mtx
605 		 * has been dropped and vp could be NULL here.
606 		 */
607 		if (!vp)
608 			break;
609 		TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
610 		/*
611 		 * Don't recycle if we can't get the interlock.
612 		 */
613 		if (!VI_TRYLOCK(vp)) {
614 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
615 			continue;
616 		}
617 		if (!VCANRECYCLE(vp)) {
618 			VI_UNLOCK(vp);
619 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
620 			continue;
621 		}
622 		freevnodes--;
623 		vp->v_iflag &= ~VI_FREE;
624 		mtx_unlock(&vnode_free_list_mtx);
625 		if (vtryrecycle(vp) != 0) {
626 			mtx_lock(&vnode_free_list_mtx);
627 			continue;
628 		}
629 		vdestroy(vp);
630 		mtx_lock(&vnode_free_list_mtx);
631 		numvnodes--;
632 	}
633 }
634 /*
635  * Attempt to recycle vnodes in a context that is always safe to block.
636  * Calling vlrurecycle() from the bowels of filesystem code has some
637  * interesting deadlock problems.
638  */
639 static struct proc *vnlruproc;
640 static int vnlruproc_sig;
641 
642 static void
643 vnlru_proc(void)
644 {
645 	struct mount *mp, *nmp;
646 	int done;
647 	struct proc *p = vnlruproc;
648 	struct thread *td = FIRST_THREAD_IN_PROC(p);
649 
650 	mtx_lock(&Giant);
651 
652 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p,
653 	    SHUTDOWN_PRI_FIRST);
654 
655 	for (;;) {
656 		kthread_suspend_check(p);
657 		mtx_lock(&vnode_free_list_mtx);
658 		if (freevnodes > wantfreevnodes)
659 			vnlru_free(freevnodes - wantfreevnodes);
660 		if (numvnodes <= desiredvnodes * 9 / 10) {
661 			vnlruproc_sig = 0;
662 			wakeup(&vnlruproc_sig);
663 			msleep(vnlruproc, &vnode_free_list_mtx,
664 			    PVFS|PDROP, "vlruwt", hz);
665 			continue;
666 		}
667 		mtx_unlock(&vnode_free_list_mtx);
668 		done = 0;
669 		mtx_lock(&mountlist_mtx);
670 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
671 			if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) {
672 				nmp = TAILQ_NEXT(mp, mnt_list);
673 				continue;
674 			}
675 			done += vlrureclaim(mp);
676 			mtx_lock(&mountlist_mtx);
677 			nmp = TAILQ_NEXT(mp, mnt_list);
678 			vfs_unbusy(mp, td);
679 		}
680 		mtx_unlock(&mountlist_mtx);
681 		if (done == 0) {
682 #if 0
683 			/* These messages are temporary debugging aids */
684 			if (vnlru_nowhere < 5)
685 				printf("vnlru process getting nowhere..\n");
686 			else if (vnlru_nowhere == 5)
687 				printf("vnlru process messages stopped.\n");
688 #endif
689 			vnlru_nowhere++;
690 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
691 		}
692 	}
693 }
694 
695 static struct kproc_desc vnlru_kp = {
696 	"vnlru",
697 	vnlru_proc,
698 	&vnlruproc
699 };
700 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
701 
702 /*
703  * Routines having to do with the management of the vnode table.
704  */
705 
706 static void
707 vdestroy(struct vnode *vp)
708 {
709 	struct bufobj *bo;
710 
711 	bo = &vp->v_bufobj;
712 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
713 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
714 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
715 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
716 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
717 	VNASSERT(bo->bo_clean.bv_root == NULL, vp, ("cleanblkroot not NULL"));
718 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
719 	VNASSERT(bo->bo_dirty.bv_root == NULL, vp, ("dirtyblkroot not NULL"));
720 #ifdef MAC
721 	mac_destroy_vnode(vp);
722 #endif
723 	if (vp->v_pollinfo != NULL) {
724 		knlist_destroy(&vp->v_pollinfo->vpi_selinfo.si_note);
725 		mtx_destroy(&vp->v_pollinfo->vpi_lock);
726 		uma_zfree(vnodepoll_zone, vp->v_pollinfo);
727 	}
728 	lockdestroy(vp->v_vnlock);
729 	mtx_destroy(&vp->v_interlock);
730 	uma_zfree(vnode_zone, vp);
731 }
732 
733 /*
734  * Check to see if a free vnode can be recycled. If it can,
735  * recycle it and return it with the vnode interlock held.
736  */
737 static int
738 vtryrecycle(struct vnode *vp)
739 {
740 	struct thread *td = curthread;
741 	struct mount *vnmp;
742 	int error;
743 
744 	ASSERT_VI_LOCKED(vp, "vtryrecycle");
745 	error = 0;
746 	vnmp = NULL;
747 	/*
748 	 * This vnode may found and locked via some other list, if so we
749 	 * can't recycle it yet.
750 	 */
751 	if (VOP_LOCK(vp, LK_INTERLOCK | LK_EXCLUSIVE | LK_NOWAIT, td) != 0) {
752 		VI_LOCK(vp);
753 		if (VSHOULDFREE(vp))
754 			vfree(vp);
755 		VI_UNLOCK(vp);
756 		return (EWOULDBLOCK);
757 	}
758 	/*
759 	 * Don't recycle if its filesystem is being suspended.
760 	 */
761 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
762 		vnmp = NULL;
763 		error = EBUSY;
764 		VI_LOCK(vp);
765 		goto err;
766 	}
767 	/*
768 	 * If we got this far, we need to acquire the interlock and see if
769 	 * anyone picked up this vnode from another list.  If not, we will
770 	 * mark it with DOOMED via vgonel() so that anyone who does find it
771 	 * will skip over it.
772 	 */
773 	VI_LOCK(vp);
774 	if (vp->v_holdcnt) {
775 		error = EBUSY;
776 		goto err;
777 	}
778 	if ((vp->v_iflag & VI_DOOMED) == 0) {
779 		vp->v_iflag |= VI_DOOMED;
780 		vgonel(vp, td);
781 		VI_LOCK(vp);
782 	}
783 	/*
784 	 * If someone ref'd the vnode while we were cleaning, we have to
785 	 * free it once the last ref is dropped.
786 	 */
787 	if (vp->v_holdcnt) {
788 		error = EBUSY;
789 		goto err;
790 	}
791 	if (vp->v_iflag & VI_FREE)
792 		vbusy(vp);
793 	VI_UNLOCK(vp);
794 	VOP_UNLOCK(vp, 0, td);
795 	vn_finished_write(vnmp);
796 	return (0);
797 err:
798 	if (VSHOULDFREE(vp))
799 		vfree(vp);
800 	VI_UNLOCK(vp);
801 	VOP_UNLOCK(vp, 0, td);
802 	if (vnmp != NULL)
803 		vn_finished_write(vnmp);
804 	return (error);
805 }
806 
807 /*
808  * Return the next vnode from the free list.
809  */
810 int
811 getnewvnode(tag, mp, vops, vpp)
812 	const char *tag;
813 	struct mount *mp;
814 	struct vop_vector *vops;
815 	struct vnode **vpp;
816 {
817 	struct vnode *vp = NULL;
818 	struct bufobj *bo;
819 
820 	mtx_lock(&vnode_free_list_mtx);
821 	/*
822 	 * Lend our context to reclaim vnodes if they've exceeded the max.
823 	 */
824 	if (freevnodes > wantfreevnodes)
825 		vnlru_free(1);
826 	/*
827 	 * Wait for available vnodes.
828 	 */
829 	if (numvnodes > desiredvnodes) {
830 		if (vnlruproc_sig == 0) {
831 			vnlruproc_sig = 1;      /* avoid unnecessary wakeups */
832 			wakeup(vnlruproc);
833 		}
834 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
835 		    "vlruwk", hz);
836 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
837 		if (numvnodes > desiredvnodes) {
838 			mtx_unlock(&vnode_free_list_mtx);
839 			return (ENFILE);
840 		}
841 #endif
842 	}
843 	numvnodes++;
844 	mtx_unlock(&vnode_free_list_mtx);
845 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO);
846 	/*
847 	 * Setup locks.
848 	 */
849 	vp->v_vnlock = &vp->v_lock;
850 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
851 	/*
852 	 * By default, don't allow shared locks unless filesystems
853 	 * opt-in.
854 	 */
855 	lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOSHARE);
856 	/*
857 	 * Initialize bufobj.
858 	 */
859 	bo = &vp->v_bufobj;
860 	bo->__bo_vnode = vp;
861 	bo->bo_mtx = &vp->v_interlock;
862 	bo->bo_ops = &buf_ops_bio;
863 	bo->bo_private = vp;
864 	TAILQ_INIT(&bo->bo_clean.bv_hd);
865 	TAILQ_INIT(&bo->bo_dirty.bv_hd);
866 	/*
867 	 * Initialize namecache.
868 	 */
869 	LIST_INIT(&vp->v_cache_src);
870 	TAILQ_INIT(&vp->v_cache_dst);
871 	/*
872 	 * Finalize various vnode identity bits.
873 	 */
874 	vp->v_type = VNON;
875 	vp->v_tag = tag;
876 	vp->v_op = vops;
877 	v_incr_usecount(vp, 1);
878 	vp->v_data = 0;
879 #ifdef MAC
880 	mac_init_vnode(vp);
881 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
882 		mac_associate_vnode_singlelabel(mp, vp);
883 	else if (mp == NULL)
884 		printf("NULL mp in getnewvnode()\n");
885 #endif
886 	delmntque(vp);
887 	if (mp != NULL) {
888 		insmntque(vp, mp);
889 		bo->bo_bsize = mp->mnt_stat.f_iosize;
890 	}
891 
892 	*vpp = vp;
893 	return (0);
894 }
895 
896 /*
897  * Delete from old mount point vnode list, if on one.
898  */
899 static void
900 delmntque(struct vnode *vp)
901 {
902 	struct mount *mp;
903 
904 	if (vp->v_mount == NULL)
905 		return;
906 	mp = vp->v_mount;
907 	MNT_ILOCK(mp);
908 	vp->v_mount = NULL;
909 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
910 		("bad mount point vnode list size"));
911 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
912 	mp->mnt_nvnodelistsize--;
913 	MNT_IUNLOCK(mp);
914 }
915 
916 /*
917  * Insert into list of vnodes for the new mount point, if available.
918  */
919 static void
920 insmntque(struct vnode *vp, struct mount *mp)
921 {
922 
923 	vp->v_mount = mp;
924 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
925 	MNT_ILOCK(vp->v_mount);
926 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
927 	mp->mnt_nvnodelistsize++;
928 	MNT_IUNLOCK(vp->v_mount);
929 }
930 
931 /*
932  * Flush out and invalidate all buffers associated with a bufobj
933  * Called with the underlying object locked.
934  */
935 int
936 bufobj_invalbuf(struct bufobj *bo, int flags, struct thread *td, int slpflag, int slptimeo)
937 {
938 	int error;
939 
940 	BO_LOCK(bo);
941 	if (flags & V_SAVE) {
942 		error = bufobj_wwait(bo, slpflag, slptimeo);
943 		if (error) {
944 			BO_UNLOCK(bo);
945 			return (error);
946 		}
947 		if (bo->bo_dirty.bv_cnt > 0) {
948 			BO_UNLOCK(bo);
949 			if ((error = BO_SYNC(bo, MNT_WAIT, td)) != 0)
950 				return (error);
951 			/*
952 			 * XXX We could save a lock/unlock if this was only
953 			 * enabled under INVARIANTS
954 			 */
955 			BO_LOCK(bo);
956 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
957 				panic("vinvalbuf: dirty bufs");
958 		}
959 	}
960 	/*
961 	 * If you alter this loop please notice that interlock is dropped and
962 	 * reacquired in flushbuflist.  Special care is needed to ensure that
963 	 * no race conditions occur from this.
964 	 */
965 	do {
966 		error = flushbuflist(&bo->bo_clean,
967 		    flags, bo, slpflag, slptimeo);
968 		if (error == 0)
969 			error = flushbuflist(&bo->bo_dirty,
970 			    flags, bo, slpflag, slptimeo);
971 		if (error != 0 && error != EAGAIN) {
972 			BO_UNLOCK(bo);
973 			return (error);
974 		}
975 	} while (error != 0);
976 
977 	/*
978 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
979 	 * have write I/O in-progress but if there is a VM object then the
980 	 * VM object can also have read-I/O in-progress.
981 	 */
982 	do {
983 		bufobj_wwait(bo, 0, 0);
984 		BO_UNLOCK(bo);
985 		if (bo->bo_object != NULL) {
986 			VM_OBJECT_LOCK(bo->bo_object);
987 			vm_object_pip_wait(bo->bo_object, "bovlbx");
988 			VM_OBJECT_UNLOCK(bo->bo_object);
989 		}
990 		BO_LOCK(bo);
991 	} while (bo->bo_numoutput > 0);
992 	BO_UNLOCK(bo);
993 
994 	/*
995 	 * Destroy the copy in the VM cache, too.
996 	 */
997 	if (bo->bo_object != NULL) {
998 		VM_OBJECT_LOCK(bo->bo_object);
999 		vm_object_page_remove(bo->bo_object, 0, 0,
1000 			(flags & V_SAVE) ? TRUE : FALSE);
1001 		VM_OBJECT_UNLOCK(bo->bo_object);
1002 	}
1003 
1004 #ifdef INVARIANTS
1005 	BO_LOCK(bo);
1006 	if ((flags & (V_ALT | V_NORMAL)) == 0 &&
1007 	    (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0))
1008 		panic("vinvalbuf: flush failed");
1009 	BO_UNLOCK(bo);
1010 #endif
1011 	return (0);
1012 }
1013 
1014 /*
1015  * Flush out and invalidate all buffers associated with a vnode.
1016  * Called with the underlying object locked.
1017  */
1018 int
1019 vinvalbuf(struct vnode *vp, int flags, struct thread *td, int slpflag, int slptimeo)
1020 {
1021 
1022 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1023 	return (bufobj_invalbuf(&vp->v_bufobj, flags, td, slpflag, slptimeo));
1024 }
1025 
1026 /*
1027  * Flush out buffers on the specified list.
1028  *
1029  */
1030 static int
1031 flushbuflist(bufv, flags, bo, slpflag, slptimeo)
1032 	struct bufv *bufv;
1033 	int flags;
1034 	struct bufobj *bo;
1035 	int slpflag, slptimeo;
1036 {
1037 	struct buf *bp, *nbp;
1038 	int retval, error;
1039 
1040 	ASSERT_BO_LOCKED(bo);
1041 
1042 	retval = 0;
1043 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1044 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1045 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1046 			continue;
1047 		}
1048 		retval = EAGAIN;
1049 		error = BUF_TIMELOCK(bp,
1050 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo),
1051 		    "flushbuf", slpflag, slptimeo);
1052 		if (error) {
1053 			BO_LOCK(bo);
1054 			return (error != ENOLCK ? error : EAGAIN);
1055 		}
1056 		KASSERT(bp->b_bufobj == bo,
1057 	            ("wrong b_bufobj %p should be %p", bp->b_bufobj, bo));
1058 		if (bp->b_bufobj != bo) {	/* XXX: necessary ? */
1059 			BUF_UNLOCK(bp);
1060 			BO_LOCK(bo);
1061 			return (EAGAIN);
1062 		}
1063 		/*
1064 		 * XXX Since there are no node locks for NFS, I
1065 		 * believe there is a slight chance that a delayed
1066 		 * write will occur while sleeping just above, so
1067 		 * check for it.
1068 		 */
1069 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1070 		    (flags & V_SAVE)) {
1071 			bremfree(bp);
1072 			bp->b_flags |= B_ASYNC;
1073 			bwrite(bp);
1074 			BO_LOCK(bo);
1075 			return (EAGAIN);	/* XXX: why not loop ? */
1076 		}
1077 		bremfree(bp);
1078 		bp->b_flags |= (B_INVAL | B_NOCACHE | B_RELBUF);
1079 		bp->b_flags &= ~B_ASYNC;
1080 		brelse(bp);
1081 		BO_LOCK(bo);
1082 	}
1083 	return (retval);
1084 }
1085 
1086 /*
1087  * Truncate a file's buffer and pages to a specified length.  This
1088  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1089  * sync activity.
1090  */
1091 int
1092 vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td, off_t length, int blksize)
1093 {
1094 	struct buf *bp, *nbp;
1095 	int anyfreed;
1096 	int trunclbn;
1097 	struct bufobj *bo;
1098 
1099 	/*
1100 	 * Round up to the *next* lbn.
1101 	 */
1102 	trunclbn = (length + blksize - 1) / blksize;
1103 
1104 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1105 restart:
1106 	VI_LOCK(vp);
1107 	bo = &vp->v_bufobj;
1108 	anyfreed = 1;
1109 	for (;anyfreed;) {
1110 		anyfreed = 0;
1111 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1112 			if (bp->b_lblkno < trunclbn)
1113 				continue;
1114 			if (BUF_LOCK(bp,
1115 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1116 			    VI_MTX(vp)) == ENOLCK)
1117 				goto restart;
1118 
1119 			bremfree(bp);
1120 			bp->b_flags |= (B_INVAL | B_RELBUF);
1121 			bp->b_flags &= ~B_ASYNC;
1122 			brelse(bp);
1123 			anyfreed = 1;
1124 
1125 			if (nbp != NULL &&
1126 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1127 			    (nbp->b_vp != vp) ||
1128 			    (nbp->b_flags & B_DELWRI))) {
1129 				goto restart;
1130 			}
1131 			VI_LOCK(vp);
1132 		}
1133 
1134 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1135 			if (bp->b_lblkno < trunclbn)
1136 				continue;
1137 			if (BUF_LOCK(bp,
1138 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1139 			    VI_MTX(vp)) == ENOLCK)
1140 				goto restart;
1141 			bremfree(bp);
1142 			bp->b_flags |= (B_INVAL | B_RELBUF);
1143 			bp->b_flags &= ~B_ASYNC;
1144 			brelse(bp);
1145 			anyfreed = 1;
1146 			if (nbp != NULL &&
1147 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1148 			    (nbp->b_vp != vp) ||
1149 			    (nbp->b_flags & B_DELWRI) == 0)) {
1150 				goto restart;
1151 			}
1152 			VI_LOCK(vp);
1153 		}
1154 	}
1155 
1156 	if (length > 0) {
1157 restartsync:
1158 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1159 			if (bp->b_lblkno > 0)
1160 				continue;
1161 			/*
1162 			 * Since we hold the vnode lock this should only
1163 			 * fail if we're racing with the buf daemon.
1164 			 */
1165 			if (BUF_LOCK(bp,
1166 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1167 			    VI_MTX(vp)) == ENOLCK) {
1168 				goto restart;
1169 			}
1170 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1171 			    ("buf(%p) on dirty queue without DELWRI", bp));
1172 
1173 			bremfree(bp);
1174 			bawrite(bp);
1175 			VI_LOCK(vp);
1176 			goto restartsync;
1177 		}
1178 	}
1179 
1180 	bufobj_wwait(bo, 0, 0);
1181 	VI_UNLOCK(vp);
1182 	vnode_pager_setsize(vp, length);
1183 
1184 	return (0);
1185 }
1186 
1187 /*
1188  * buf_splay() - splay tree core for the clean/dirty list of buffers in
1189  * 		 a vnode.
1190  *
1191  *	NOTE: We have to deal with the special case of a background bitmap
1192  *	buffer, a situation where two buffers will have the same logical
1193  *	block offset.  We want (1) only the foreground buffer to be accessed
1194  *	in a lookup and (2) must differentiate between the foreground and
1195  *	background buffer in the splay tree algorithm because the splay
1196  *	tree cannot normally handle multiple entities with the same 'index'.
1197  *	We accomplish this by adding differentiating flags to the splay tree's
1198  *	numerical domain.
1199  */
1200 static
1201 struct buf *
1202 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root)
1203 {
1204 	struct buf dummy;
1205 	struct buf *lefttreemax, *righttreemin, *y;
1206 
1207 	if (root == NULL)
1208 		return (NULL);
1209 	lefttreemax = righttreemin = &dummy;
1210 	for (;;) {
1211 		if (lblkno < root->b_lblkno ||
1212 		    (lblkno == root->b_lblkno &&
1213 		    (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1214 			if ((y = root->b_left) == NULL)
1215 				break;
1216 			if (lblkno < y->b_lblkno) {
1217 				/* Rotate right. */
1218 				root->b_left = y->b_right;
1219 				y->b_right = root;
1220 				root = y;
1221 				if ((y = root->b_left) == NULL)
1222 					break;
1223 			}
1224 			/* Link into the new root's right tree. */
1225 			righttreemin->b_left = root;
1226 			righttreemin = root;
1227 		} else if (lblkno > root->b_lblkno ||
1228 		    (lblkno == root->b_lblkno &&
1229 		    (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) {
1230 			if ((y = root->b_right) == NULL)
1231 				break;
1232 			if (lblkno > y->b_lblkno) {
1233 				/* Rotate left. */
1234 				root->b_right = y->b_left;
1235 				y->b_left = root;
1236 				root = y;
1237 				if ((y = root->b_right) == NULL)
1238 					break;
1239 			}
1240 			/* Link into the new root's left tree. */
1241 			lefttreemax->b_right = root;
1242 			lefttreemax = root;
1243 		} else {
1244 			break;
1245 		}
1246 		root = y;
1247 	}
1248 	/* Assemble the new root. */
1249 	lefttreemax->b_right = root->b_left;
1250 	righttreemin->b_left = root->b_right;
1251 	root->b_left = dummy.b_right;
1252 	root->b_right = dummy.b_left;
1253 	return (root);
1254 }
1255 
1256 static void
1257 buf_vlist_remove(struct buf *bp)
1258 {
1259 	struct buf *root;
1260 	struct bufv *bv;
1261 
1262 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1263 	ASSERT_BO_LOCKED(bp->b_bufobj);
1264 	if (bp->b_xflags & BX_VNDIRTY)
1265 		bv = &bp->b_bufobj->bo_dirty;
1266 	else
1267 		bv = &bp->b_bufobj->bo_clean;
1268 	if (bp != bv->bv_root) {
1269 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1270 		KASSERT(root == bp, ("splay lookup failed in remove"));
1271 	}
1272 	if (bp->b_left == NULL) {
1273 		root = bp->b_right;
1274 	} else {
1275 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left);
1276 		root->b_right = bp->b_right;
1277 	}
1278 	bv->bv_root = root;
1279 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1280 	bv->bv_cnt--;
1281 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1282 }
1283 
1284 /*
1285  * Add the buffer to the sorted clean or dirty block list using a
1286  * splay tree algorithm.
1287  *
1288  * NOTE: xflags is passed as a constant, optimizing this inline function!
1289  */
1290 static void
1291 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1292 {
1293 	struct buf *root;
1294 	struct bufv *bv;
1295 
1296 	ASSERT_BO_LOCKED(bo);
1297 	bp->b_xflags |= xflags;
1298 	if (xflags & BX_VNDIRTY)
1299 		bv = &bo->bo_dirty;
1300 	else
1301 		bv = &bo->bo_clean;
1302 
1303 	root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1304 	if (root == NULL) {
1305 		bp->b_left = NULL;
1306 		bp->b_right = NULL;
1307 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1308 	} else if (bp->b_lblkno < root->b_lblkno ||
1309 	    (bp->b_lblkno == root->b_lblkno &&
1310 	    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1311 		bp->b_left = root->b_left;
1312 		bp->b_right = root;
1313 		root->b_left = NULL;
1314 		TAILQ_INSERT_BEFORE(root, bp, b_bobufs);
1315 	} else {
1316 		bp->b_right = root->b_right;
1317 		bp->b_left = root;
1318 		root->b_right = NULL;
1319 		TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs);
1320 	}
1321 	bv->bv_cnt++;
1322 	bv->bv_root = bp;
1323 }
1324 
1325 /*
1326  * Lookup a buffer using the splay tree.  Note that we specifically avoid
1327  * shadow buffers used in background bitmap writes.
1328  *
1329  * This code isn't quite efficient as it could be because we are maintaining
1330  * two sorted lists and do not know which list the block resides in.
1331  *
1332  * During a "make buildworld" the desired buffer is found at one of
1333  * the roots more than 60% of the time.  Thus, checking both roots
1334  * before performing either splay eliminates unnecessary splays on the
1335  * first tree splayed.
1336  */
1337 struct buf *
1338 gbincore(struct bufobj *bo, daddr_t lblkno)
1339 {
1340 	struct buf *bp;
1341 
1342 	ASSERT_BO_LOCKED(bo);
1343 	if ((bp = bo->bo_clean.bv_root) != NULL &&
1344 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1345 		return (bp);
1346 	if ((bp = bo->bo_dirty.bv_root) != NULL &&
1347 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1348 		return (bp);
1349 	if ((bp = bo->bo_clean.bv_root) != NULL) {
1350 		bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp);
1351 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1352 			return (bp);
1353 	}
1354 	if ((bp = bo->bo_dirty.bv_root) != NULL) {
1355 		bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp);
1356 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1357 			return (bp);
1358 	}
1359 	return (NULL);
1360 }
1361 
1362 /*
1363  * Associate a buffer with a vnode.
1364  */
1365 void
1366 bgetvp(struct vnode *vp, struct buf *bp)
1367 {
1368 
1369 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
1370 
1371 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
1372 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
1373 	    ("bgetvp: bp already attached! %p", bp));
1374 
1375 	ASSERT_VI_LOCKED(vp, "bgetvp");
1376 	vholdl(vp);
1377 	bp->b_vp = vp;
1378 	bp->b_bufobj = &vp->v_bufobj;
1379 	/*
1380 	 * Insert onto list for new vnode.
1381 	 */
1382 	buf_vlist_add(bp, &vp->v_bufobj, BX_VNCLEAN);
1383 }
1384 
1385 /*
1386  * Disassociate a buffer from a vnode.
1387  */
1388 void
1389 brelvp(struct buf *bp)
1390 {
1391 	struct bufobj *bo;
1392 	struct vnode *vp;
1393 
1394 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1395 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1396 
1397 	/*
1398 	 * Delete from old vnode list, if on one.
1399 	 */
1400 	vp = bp->b_vp;		/* XXX */
1401 	bo = bp->b_bufobj;
1402 	BO_LOCK(bo);
1403 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1404 		buf_vlist_remove(bp);
1405 	else
1406 		panic("brelvp: Buffer %p not on queue.", bp);
1407 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1408 		bo->bo_flag &= ~BO_ONWORKLST;
1409 		mtx_lock(&sync_mtx);
1410 		LIST_REMOVE(bo, bo_synclist);
1411  		syncer_worklist_len--;
1412 		mtx_unlock(&sync_mtx);
1413 	}
1414 	vdropl(vp);
1415 	bp->b_vp = NULL;
1416 	bp->b_bufobj = NULL;
1417 	BO_UNLOCK(bo);
1418 }
1419 
1420 /*
1421  * Add an item to the syncer work queue.
1422  */
1423 static void
1424 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
1425 {
1426 	int slot;
1427 
1428 	ASSERT_BO_LOCKED(bo);
1429 
1430 	mtx_lock(&sync_mtx);
1431 	if (bo->bo_flag & BO_ONWORKLST)
1432 		LIST_REMOVE(bo, bo_synclist);
1433 	else {
1434 		bo->bo_flag |= BO_ONWORKLST;
1435  		syncer_worklist_len++;
1436 	}
1437 
1438 	if (delay > syncer_maxdelay - 2)
1439 		delay = syncer_maxdelay - 2;
1440 	slot = (syncer_delayno + delay) & syncer_mask;
1441 
1442 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
1443 	mtx_unlock(&sync_mtx);
1444 }
1445 
1446 static int
1447 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
1448 {
1449 	int error, len;
1450 
1451 	mtx_lock(&sync_mtx);
1452 	len = syncer_worklist_len - sync_vnode_count;
1453 	mtx_unlock(&sync_mtx);
1454 	error = SYSCTL_OUT(req, &len, sizeof(len));
1455 	return (error);
1456 }
1457 
1458 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
1459     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
1460 
1461 struct  proc *updateproc;
1462 static void sched_sync(void);
1463 static struct kproc_desc up_kp = {
1464 	"syncer",
1465 	sched_sync,
1466 	&updateproc
1467 };
1468 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp)
1469 
1470 static int
1471 sync_vnode(struct bufobj *bo, struct thread *td)
1472 {
1473 	struct vnode *vp;
1474 	struct mount *mp;
1475 
1476 	vp = bo->__bo_vnode; 	/* XXX */
1477 	if (VOP_ISLOCKED(vp, NULL) != 0)
1478 		return (1);
1479 	if (VI_TRYLOCK(vp) == 0)
1480 		return (1);
1481 	/*
1482 	 * We use vhold in case the vnode does not
1483 	 * successfully sync.  vhold prevents the vnode from
1484 	 * going away when we unlock the sync_mtx so that
1485 	 * we can acquire the vnode interlock.
1486 	 */
1487 	vholdl(vp);
1488 	mtx_unlock(&sync_mtx);
1489 	VI_UNLOCK(vp);
1490 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1491 		vdrop(vp);
1492 		mtx_lock(&sync_mtx);
1493 		return (1);
1494 	}
1495 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
1496 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
1497 	VOP_UNLOCK(vp, 0, td);
1498 	vn_finished_write(mp);
1499 	VI_LOCK(vp);
1500 	if ((bo->bo_flag & BO_ONWORKLST) != 0) {
1501 		/*
1502 		 * Put us back on the worklist.  The worklist
1503 		 * routine will remove us from our current
1504 		 * position and then add us back in at a later
1505 		 * position.
1506 		 */
1507 		vn_syncer_add_to_worklist(bo, syncdelay);
1508 	}
1509 	vdropl(vp);
1510 	VI_UNLOCK(vp);
1511 	mtx_lock(&sync_mtx);
1512 	return (0);
1513 }
1514 
1515 /*
1516  * System filesystem synchronizer daemon.
1517  */
1518 static void
1519 sched_sync(void)
1520 {
1521 	struct synclist *next;
1522 	struct synclist *slp;
1523 	struct bufobj *bo;
1524 	long starttime;
1525 	struct thread *td = FIRST_THREAD_IN_PROC(updateproc);
1526 	static int dummychan;
1527 	int last_work_seen;
1528 	int net_worklist_len;
1529 	int syncer_final_iter;
1530 	int first_printf;
1531 	int error;
1532 
1533 	mtx_lock(&Giant);
1534 	last_work_seen = 0;
1535 	syncer_final_iter = 0;
1536 	first_printf = 1;
1537 	syncer_state = SYNCER_RUNNING;
1538 	starttime = time_second;
1539 
1540 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
1541 	    SHUTDOWN_PRI_LAST);
1542 
1543 	for (;;) {
1544 		mtx_lock(&sync_mtx);
1545 		if (syncer_state == SYNCER_FINAL_DELAY &&
1546 		    syncer_final_iter == 0) {
1547 			mtx_unlock(&sync_mtx);
1548 			kthread_suspend_check(td->td_proc);
1549 			mtx_lock(&sync_mtx);
1550 		}
1551 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
1552 		if (syncer_state != SYNCER_RUNNING &&
1553 		    starttime != time_second) {
1554 			if (first_printf) {
1555 				printf("\nSyncing disks, vnodes remaining...");
1556 				first_printf = 0;
1557 			}
1558 			printf("%d ", net_worklist_len);
1559 		}
1560 		starttime = time_second;
1561 
1562 		/*
1563 		 * Push files whose dirty time has expired.  Be careful
1564 		 * of interrupt race on slp queue.
1565 		 *
1566 		 * Skip over empty worklist slots when shutting down.
1567 		 */
1568 		do {
1569 			slp = &syncer_workitem_pending[syncer_delayno];
1570 			syncer_delayno += 1;
1571 			if (syncer_delayno == syncer_maxdelay)
1572 				syncer_delayno = 0;
1573 			next = &syncer_workitem_pending[syncer_delayno];
1574 			/*
1575 			 * If the worklist has wrapped since the
1576 			 * it was emptied of all but syncer vnodes,
1577 			 * switch to the FINAL_DELAY state and run
1578 			 * for one more second.
1579 			 */
1580 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
1581 			    net_worklist_len == 0 &&
1582 			    last_work_seen == syncer_delayno) {
1583 				syncer_state = SYNCER_FINAL_DELAY;
1584 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
1585 			}
1586 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
1587 		    syncer_worklist_len > 0);
1588 
1589 		/*
1590 		 * Keep track of the last time there was anything
1591 		 * on the worklist other than syncer vnodes.
1592 		 * Return to the SHUTTING_DOWN state if any
1593 		 * new work appears.
1594 		 */
1595 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
1596 			last_work_seen = syncer_delayno;
1597 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
1598 			syncer_state = SYNCER_SHUTTING_DOWN;
1599 		while ((bo = LIST_FIRST(slp)) != NULL) {
1600 			error = sync_vnode(bo, td);
1601 			if (error == 1) {
1602 				LIST_REMOVE(bo, bo_synclist);
1603 				LIST_INSERT_HEAD(next, bo, bo_synclist);
1604 				continue;
1605 			}
1606 		}
1607 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
1608 			syncer_final_iter--;
1609 		mtx_unlock(&sync_mtx);
1610 
1611 		/*
1612 		 * Do soft update processing.
1613 		 */
1614 		if (softdep_process_worklist_hook != NULL)
1615 			(*softdep_process_worklist_hook)(NULL);
1616 
1617 		/*
1618 		 * The variable rushjob allows the kernel to speed up the
1619 		 * processing of the filesystem syncer process. A rushjob
1620 		 * value of N tells the filesystem syncer to process the next
1621 		 * N seconds worth of work on its queue ASAP. Currently rushjob
1622 		 * is used by the soft update code to speed up the filesystem
1623 		 * syncer process when the incore state is getting so far
1624 		 * ahead of the disk that the kernel memory pool is being
1625 		 * threatened with exhaustion.
1626 		 */
1627 		mtx_lock(&sync_mtx);
1628 		if (rushjob > 0) {
1629 			rushjob -= 1;
1630 			mtx_unlock(&sync_mtx);
1631 			continue;
1632 		}
1633 		mtx_unlock(&sync_mtx);
1634 		/*
1635 		 * Just sleep for a short period if time between
1636 		 * iterations when shutting down to allow some I/O
1637 		 * to happen.
1638 		 *
1639 		 * If it has taken us less than a second to process the
1640 		 * current work, then wait. Otherwise start right over
1641 		 * again. We can still lose time if any single round
1642 		 * takes more than two seconds, but it does not really
1643 		 * matter as we are just trying to generally pace the
1644 		 * filesystem activity.
1645 		 */
1646 		if (syncer_state != SYNCER_RUNNING)
1647 			tsleep(&dummychan, PPAUSE, "syncfnl",
1648 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
1649 		else if (time_second == starttime)
1650 			tsleep(&lbolt, PPAUSE, "syncer", 0);
1651 	}
1652 }
1653 
1654 /*
1655  * Request the syncer daemon to speed up its work.
1656  * We never push it to speed up more than half of its
1657  * normal turn time, otherwise it could take over the cpu.
1658  */
1659 int
1660 speedup_syncer()
1661 {
1662 	struct thread *td;
1663 	int ret = 0;
1664 
1665 	td = FIRST_THREAD_IN_PROC(updateproc);
1666 	sleepq_remove(td, &lbolt);
1667 	mtx_lock(&sync_mtx);
1668 	if (rushjob < syncdelay / 2) {
1669 		rushjob += 1;
1670 		stat_rush_requests += 1;
1671 		ret = 1;
1672 	}
1673 	mtx_unlock(&sync_mtx);
1674 	return (ret);
1675 }
1676 
1677 /*
1678  * Tell the syncer to speed up its work and run though its work
1679  * list several times, then tell it to shut down.
1680  */
1681 static void
1682 syncer_shutdown(void *arg, int howto)
1683 {
1684 	struct thread *td;
1685 
1686 	if (howto & RB_NOSYNC)
1687 		return;
1688 	td = FIRST_THREAD_IN_PROC(updateproc);
1689 	sleepq_remove(td, &lbolt);
1690 	mtx_lock(&sync_mtx);
1691 	syncer_state = SYNCER_SHUTTING_DOWN;
1692 	rushjob = 0;
1693 	mtx_unlock(&sync_mtx);
1694 	kproc_shutdown(arg, howto);
1695 }
1696 
1697 /*
1698  * Reassign a buffer from one vnode to another.
1699  * Used to assign file specific control information
1700  * (indirect blocks) to the vnode to which they belong.
1701  */
1702 void
1703 reassignbuf(struct buf *bp)
1704 {
1705 	struct vnode *vp;
1706 	struct bufobj *bo;
1707 	int delay;
1708 
1709 	vp = bp->b_vp;
1710 	bo = bp->b_bufobj;
1711 	++reassignbufcalls;
1712 
1713 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
1714 	    bp, bp->b_vp, bp->b_flags);
1715 	/*
1716 	 * B_PAGING flagged buffers cannot be reassigned because their vp
1717 	 * is not fully linked in.
1718 	 */
1719 	if (bp->b_flags & B_PAGING)
1720 		panic("cannot reassign paging buffer");
1721 
1722 	/*
1723 	 * Delete from old vnode list, if on one.
1724 	 */
1725 	VI_LOCK(vp);
1726 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1727 		buf_vlist_remove(bp);
1728 	else
1729 		panic("reassignbuf: Buffer %p not on queue.", bp);
1730 	/*
1731 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
1732 	 * of clean buffers.
1733 	 */
1734 	if (bp->b_flags & B_DELWRI) {
1735 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
1736 			switch (vp->v_type) {
1737 			case VDIR:
1738 				delay = dirdelay;
1739 				break;
1740 			case VCHR:
1741 				delay = metadelay;
1742 				break;
1743 			default:
1744 				delay = filedelay;
1745 			}
1746 			vn_syncer_add_to_worklist(bo, delay);
1747 		}
1748 		buf_vlist_add(bp, bo, BX_VNDIRTY);
1749 	} else {
1750 		buf_vlist_add(bp, bo, BX_VNCLEAN);
1751 
1752 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1753 			mtx_lock(&sync_mtx);
1754 			LIST_REMOVE(bo, bo_synclist);
1755  			syncer_worklist_len--;
1756 			mtx_unlock(&sync_mtx);
1757 			bo->bo_flag &= ~BO_ONWORKLST;
1758 		}
1759 	}
1760 	VI_UNLOCK(vp);
1761 }
1762 
1763 static void
1764 v_incr_usecount(struct vnode *vp, int delta)
1765 {
1766 
1767 	vp->v_usecount += delta;
1768 	vp->v_holdcnt += delta;
1769 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1770 		dev_lock();
1771 		vp->v_rdev->si_usecount += delta;
1772 		dev_unlock();
1773 	}
1774 }
1775 
1776 /*
1777  * Grab a particular vnode from the free list, increment its
1778  * reference count and lock it. The vnode lock bit is set if the
1779  * vnode is being eliminated in vgone. The process is awakened
1780  * when the transition is completed, and an error returned to
1781  * indicate that the vnode is no longer usable (possibly having
1782  * been changed to a new filesystem type).
1783  */
1784 int
1785 vget(vp, flags, td)
1786 	struct vnode *vp;
1787 	int flags;
1788 	struct thread *td;
1789 {
1790 	int oweinact;
1791 	int oldflags;
1792 	int error;
1793 
1794 	error = 0;
1795 	oldflags = flags;
1796 	oweinact = 0;
1797 	if ((flags & LK_INTERLOCK) == 0)
1798 		VI_LOCK(vp);
1799 	/*
1800 	 * If the inactive call was deferred because vput() was called
1801 	 * with a shared lock, we have to do it here before another thread
1802 	 * gets a reference to data that should be dead.
1803 	 */
1804 	if (vp->v_iflag & VI_OWEINACT) {
1805 		if (flags & LK_NOWAIT) {
1806 			VI_UNLOCK(vp);
1807 			return (EBUSY);
1808 		}
1809 		flags &= ~LK_TYPE_MASK;
1810 		flags |= LK_EXCLUSIVE;
1811 		oweinact = 1;
1812 	}
1813 	v_incr_usecount(vp, 1);
1814 	if (VSHOULDBUSY(vp))
1815 		vbusy(vp);
1816 	if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) {
1817 		VI_LOCK(vp);
1818 		/*
1819 		 * must expand vrele here because we do not want
1820 		 * to call VOP_INACTIVE if the reference count
1821 		 * drops back to zero since it was never really
1822 		 * active.
1823 		 */
1824 		v_incr_usecount(vp, -1);
1825 		if (VSHOULDFREE(vp))
1826 			vfree(vp);
1827 		else
1828 			vlruvp(vp);
1829 		VI_UNLOCK(vp);
1830 		return (error);
1831 	}
1832 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
1833 		panic("vget: vn_lock failed to return ENOENT\n");
1834 	if (oweinact) {
1835 		VI_LOCK(vp);
1836 		if (vp->v_iflag & VI_OWEINACT)
1837 			vinactive(vp, td);
1838 		VI_UNLOCK(vp);
1839 		if ((oldflags & LK_TYPE_MASK) == 0)
1840 			VOP_UNLOCK(vp, 0, td);
1841 	}
1842 	return (0);
1843 }
1844 
1845 /*
1846  * Increase the reference count of a vnode.
1847  */
1848 void
1849 vref(struct vnode *vp)
1850 {
1851 
1852 	VI_LOCK(vp);
1853 	v_incr_usecount(vp, 1);
1854 	VI_UNLOCK(vp);
1855 }
1856 
1857 /*
1858  * Return reference count of a vnode.
1859  *
1860  * The results of this call are only guaranteed when some mechanism other
1861  * than the VI lock is used to stop other processes from gaining references
1862  * to the vnode.  This may be the case if the caller holds the only reference.
1863  * This is also useful when stale data is acceptable as race conditions may
1864  * be accounted for by some other means.
1865  */
1866 int
1867 vrefcnt(struct vnode *vp)
1868 {
1869 	int usecnt;
1870 
1871 	VI_LOCK(vp);
1872 	usecnt = vp->v_usecount;
1873 	VI_UNLOCK(vp);
1874 
1875 	return (usecnt);
1876 }
1877 
1878 
1879 /*
1880  * Vnode put/release.
1881  * If count drops to zero, call inactive routine and return to freelist.
1882  */
1883 void
1884 vrele(vp)
1885 	struct vnode *vp;
1886 {
1887 	struct thread *td = curthread;	/* XXX */
1888 
1889 	KASSERT(vp != NULL, ("vrele: null vp"));
1890 
1891 	VI_LOCK(vp);
1892 
1893 	/* Skip this v_writecount check if we're going to panic below. */
1894 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
1895 	    ("vrele: missed vn_close"));
1896 
1897 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
1898 	    vp->v_usecount == 1)) {
1899 		v_incr_usecount(vp, -1);
1900 		VI_UNLOCK(vp);
1901 
1902 		return;
1903 	}
1904 	if (vp->v_usecount != 1) {
1905 #ifdef DIAGNOSTIC
1906 		vprint("vrele: negative ref count", vp);
1907 #endif
1908 		VI_UNLOCK(vp);
1909 		panic("vrele: negative ref cnt");
1910 	}
1911 	v_incr_usecount(vp, -1);
1912 	/*
1913 	 * We must call VOP_INACTIVE with the node locked. Mark
1914 	 * as VI_DOINGINACT to avoid recursion.
1915 	 */
1916 	if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) {
1917 		VI_LOCK(vp);
1918 		vinactive(vp, td);
1919 		VOP_UNLOCK(vp, 0, td);
1920 	} else
1921 		VI_LOCK(vp);
1922 	if (VSHOULDFREE(vp))
1923 		vfree(vp);
1924 	else
1925 		vlruvp(vp);
1926 	VI_UNLOCK(vp);
1927 }
1928 
1929 /*
1930  * Release an already locked vnode.  This give the same effects as
1931  * unlock+vrele(), but takes less time and avoids releasing and
1932  * re-aquiring the lock (as vrele() aquires the lock internally.)
1933  */
1934 void
1935 vput(vp)
1936 	struct vnode *vp;
1937 {
1938 	struct thread *td = curthread;	/* XXX */
1939 	int error;
1940 
1941 	KASSERT(vp != NULL, ("vput: null vp"));
1942 	ASSERT_VOP_LOCKED(vp, "vput");
1943 	VI_LOCK(vp);
1944 	/* Skip this v_writecount check if we're going to panic below. */
1945 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
1946 	    ("vput: missed vn_close"));
1947 	error = 0;
1948 
1949 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
1950 	    vp->v_usecount == 1)) {
1951 		v_incr_usecount(vp, -1);
1952 		VOP_UNLOCK(vp, LK_INTERLOCK, td);
1953 		return;
1954 	}
1955 
1956 	if (vp->v_usecount != 1) {
1957 #ifdef DIAGNOSTIC
1958 		vprint("vput: negative ref count", vp);
1959 #endif
1960 		panic("vput: negative ref cnt");
1961 	}
1962 	v_incr_usecount(vp, -1);
1963 	vp->v_iflag |= VI_OWEINACT;
1964 	if (VOP_ISLOCKED(vp, NULL) != LK_EXCLUSIVE) {
1965 		error = VOP_LOCK(vp, LK_EXCLUPGRADE|LK_INTERLOCK|LK_NOWAIT, td);
1966 		VI_LOCK(vp);
1967 		if (error)
1968 			goto done;
1969 	}
1970 	if (vp->v_iflag & VI_OWEINACT)
1971 		vinactive(vp, td);
1972 	VOP_UNLOCK(vp, 0, td);
1973 done:
1974 	if (VSHOULDFREE(vp))
1975 		vfree(vp);
1976 	else
1977 		vlruvp(vp);
1978 	VI_UNLOCK(vp);
1979 }
1980 
1981 /*
1982  * Somebody doesn't want the vnode recycled.
1983  */
1984 void
1985 vhold(struct vnode *vp)
1986 {
1987 
1988 	VI_LOCK(vp);
1989 	vholdl(vp);
1990 	VI_UNLOCK(vp);
1991 }
1992 
1993 void
1994 vholdl(struct vnode *vp)
1995 {
1996 
1997 	vp->v_holdcnt++;
1998 	if (VSHOULDBUSY(vp))
1999 		vbusy(vp);
2000 }
2001 
2002 /*
2003  * Note that there is one less who cares about this vnode.  vdrop() is the
2004  * opposite of vhold().
2005  */
2006 void
2007 vdrop(struct vnode *vp)
2008 {
2009 
2010 	VI_LOCK(vp);
2011 	vdropl(vp);
2012 	VI_UNLOCK(vp);
2013 }
2014 
2015 static void
2016 vdropl(struct vnode *vp)
2017 {
2018 
2019 	if (vp->v_holdcnt <= 0)
2020 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
2021 	vp->v_holdcnt--;
2022 	if (VSHOULDFREE(vp))
2023 		vfree(vp);
2024 	else
2025 		vlruvp(vp);
2026 }
2027 
2028 static void
2029 vinactive(struct vnode *vp, struct thread *td)
2030 {
2031 	ASSERT_VOP_LOCKED(vp, "vinactive");
2032 	ASSERT_VI_LOCKED(vp, "vinactive");
2033 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2034 	    ("vinactive: recursed on VI_DOINGINACT"));
2035 	vp->v_iflag |= VI_DOINGINACT;
2036 	VI_UNLOCK(vp);
2037 	VOP_INACTIVE(vp, td);
2038 	VI_LOCK(vp);
2039 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2040 	    ("vinactive: lost VI_DOINGINACT"));
2041 	vp->v_iflag &= ~(VI_DOINGINACT|VI_OWEINACT);
2042 }
2043 
2044 /*
2045  * Remove any vnodes in the vnode table belonging to mount point mp.
2046  *
2047  * If FORCECLOSE is not specified, there should not be any active ones,
2048  * return error if any are found (nb: this is a user error, not a
2049  * system error). If FORCECLOSE is specified, detach any active vnodes
2050  * that are found.
2051  *
2052  * If WRITECLOSE is set, only flush out regular file vnodes open for
2053  * writing.
2054  *
2055  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2056  *
2057  * `rootrefs' specifies the base reference count for the root vnode
2058  * of this filesystem. The root vnode is considered busy if its
2059  * v_usecount exceeds this value. On a successful return, vflush(, td)
2060  * will call vrele() on the root vnode exactly rootrefs times.
2061  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2062  * be zero.
2063  */
2064 #ifdef DIAGNOSTIC
2065 static int busyprt = 0;		/* print out busy vnodes */
2066 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
2067 #endif
2068 
2069 int
2070 vflush(mp, rootrefs, flags, td)
2071 	struct mount *mp;
2072 	int rootrefs;
2073 	int flags;
2074 	struct thread *td;
2075 {
2076 	struct vnode *vp, *nvp, *rootvp = NULL;
2077 	struct vattr vattr;
2078 	int busy = 0, error;
2079 
2080 	if (rootrefs > 0) {
2081 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2082 		    ("vflush: bad args"));
2083 		/*
2084 		 * Get the filesystem root vnode. We can vput() it
2085 		 * immediately, since with rootrefs > 0, it won't go away.
2086 		 */
2087 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp, td)) != 0)
2088 			return (error);
2089 		vput(rootvp);
2090 
2091 	}
2092 	MNT_ILOCK(mp);
2093 loop:
2094 	MNT_VNODE_FOREACH(vp, mp, nvp) {
2095 
2096 		VI_LOCK(vp);
2097 		MNT_IUNLOCK(mp);
2098 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td);
2099 		if (error) {
2100 			MNT_ILOCK(mp);
2101 			goto loop;
2102 		}
2103 		/*
2104 		 * Skip over a vnodes marked VV_SYSTEM.
2105 		 */
2106 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2107 			VOP_UNLOCK(vp, 0, td);
2108 			MNT_ILOCK(mp);
2109 			continue;
2110 		}
2111 		/*
2112 		 * If WRITECLOSE is set, flush out unlinked but still open
2113 		 * files (even if open only for reading) and regular file
2114 		 * vnodes open for writing.
2115 		 */
2116 		if (flags & WRITECLOSE) {
2117 			error = VOP_GETATTR(vp, &vattr, td->td_ucred, td);
2118 			VI_LOCK(vp);
2119 
2120 			if ((vp->v_type == VNON ||
2121 			    (error == 0 && vattr.va_nlink > 0)) &&
2122 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2123 				VOP_UNLOCK(vp, LK_INTERLOCK, td);
2124 				MNT_ILOCK(mp);
2125 				continue;
2126 			}
2127 		} else
2128 			VI_LOCK(vp);
2129 		/*
2130 		 * With v_usecount == 0, all we need to do is clear out the
2131 		 * vnode data structures and we are done.
2132 		 */
2133 		if (vp->v_usecount == 0) {
2134 			vgonel(vp, td);
2135 			VOP_UNLOCK(vp, 0, td);
2136 			MNT_ILOCK(mp);
2137 			continue;
2138 		}
2139 		/*
2140 		 * If FORCECLOSE is set, forcibly close the vnode. For block
2141 		 * or character devices, revert to an anonymous device. For
2142 		 * all other files, just kill them.
2143 		 */
2144 		if (flags & FORCECLOSE) {
2145 			VNASSERT(vp->v_type != VCHR && vp->v_type != VBLK, vp,
2146 			    ("device VNODE %p is FORCECLOSED", vp));
2147 			vgonel(vp, td);
2148 			VOP_UNLOCK(vp, 0, td);
2149 			MNT_ILOCK(mp);
2150 			continue;
2151 		}
2152 		VOP_UNLOCK(vp, 0, td);
2153 #ifdef DIAGNOSTIC
2154 		if (busyprt)
2155 			vprint("vflush: busy vnode", vp);
2156 #endif
2157 		VI_UNLOCK(vp);
2158 		MNT_ILOCK(mp);
2159 		busy++;
2160 	}
2161 	MNT_IUNLOCK(mp);
2162 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2163 		/*
2164 		 * If just the root vnode is busy, and if its refcount
2165 		 * is equal to `rootrefs', then go ahead and kill it.
2166 		 */
2167 		VI_LOCK(rootvp);
2168 		KASSERT(busy > 0, ("vflush: not busy"));
2169 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
2170 		    ("vflush: usecount %d < rootrefs %d",
2171 		     rootvp->v_usecount, rootrefs));
2172 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2173 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK, td);
2174 			vgone(rootvp);
2175 			VOP_UNLOCK(rootvp, 0, td);
2176 			busy = 0;
2177 		} else
2178 			VI_UNLOCK(rootvp);
2179 	}
2180 	if (busy)
2181 		return (EBUSY);
2182 	for (; rootrefs > 0; rootrefs--)
2183 		vrele(rootvp);
2184 	return (0);
2185 }
2186 
2187 /*
2188  * This moves a now (likely recyclable) vnode to the end of the
2189  * mountlist.  XXX However, it is temporarily disabled until we
2190  * can clean up ffs_sync() and friends, which have loop restart
2191  * conditions which this code causes to operate O(N^2).
2192  */
2193 static void
2194 vlruvp(struct vnode *vp)
2195 {
2196 #if 0
2197 	struct mount *mp;
2198 
2199 	if ((mp = vp->v_mount) != NULL) {
2200 		MNT_ILOCK(mp);
2201 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2202 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2203 		MNT_IUNLOCK(mp);
2204 	}
2205 #endif
2206 }
2207 
2208 /*
2209  * Recycle an unused vnode to the front of the free list.
2210  * Release the passed interlock if the vnode will be recycled.
2211  */
2212 int
2213 vrecycle(struct vnode *vp, struct thread *td)
2214 {
2215 
2216 	ASSERT_VOP_LOCKED(vp, "vrecycle");
2217 	VI_LOCK(vp);
2218 	if (vp->v_usecount == 0 && (vp->v_iflag & VI_DOOMED) == 0) {
2219 		vgonel(vp, td);
2220 		return (1);
2221 	}
2222 	VI_UNLOCK(vp);
2223 	return (0);
2224 }
2225 
2226 /*
2227  * Eliminate all activity associated with a vnode
2228  * in preparation for reuse.
2229  */
2230 void
2231 vgone(struct vnode *vp)
2232 {
2233 	struct thread *td = curthread;	/* XXX */
2234 	ASSERT_VOP_LOCKED(vp, "vgone");
2235 
2236 	/*
2237 	 * Don't vgonel if we're already doomed.
2238 	 */
2239 	VI_LOCK(vp);
2240 	if (vp->v_iflag & VI_DOOMED) {
2241 		VI_UNLOCK(vp);
2242 		return;
2243 	}
2244 	vgonel(vp, td);
2245 }
2246 
2247 /*
2248  * vgone, with the vp interlock held.
2249  */
2250 void
2251 vgonel(struct vnode *vp, struct thread *td)
2252 {
2253 	int oweinact;
2254 	int active;
2255 	int doomed;
2256 
2257 	ASSERT_VOP_LOCKED(vp, "vgonel");
2258 	ASSERT_VI_LOCKED(vp, "vgonel");
2259 
2260 	/*
2261 	 * Check to see if the vnode is in use. If so we have to reference it
2262 	 * before we clean it out so that its count cannot fall to zero and
2263 	 * generate a race against ourselves to recycle it.
2264 	 */
2265 	if ((active = vp->v_usecount))
2266 		v_incr_usecount(vp, 1);
2267 
2268 	/*
2269 	 * See if we're already doomed, if so, this is coming from a
2270 	 * successful vtryrecycle();
2271 	 */
2272 	doomed = (vp->v_iflag & VI_DOOMED);
2273 	vp->v_iflag |= VI_DOOMED;
2274 	oweinact = (vp->v_iflag & VI_OWEINACT);
2275 	VI_UNLOCK(vp);
2276 
2277 	/*
2278 	 * Clean out any buffers associated with the vnode.
2279 	 * If the flush fails, just toss the buffers.
2280 	 */
2281 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
2282 		(void) vn_write_suspend_wait(vp, NULL, V_WAIT);
2283 	if (vinvalbuf(vp, V_SAVE, td, 0, 0) != 0)
2284 		vinvalbuf(vp, 0, td, 0, 0);
2285 
2286 	/*
2287 	 * If purging an active vnode, it must be closed and
2288 	 * deactivated before being reclaimed.
2289 	 */
2290 	if (active)
2291 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
2292 	if (oweinact || active) {
2293 		VI_LOCK(vp);
2294 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
2295 			vinactive(vp, td);
2296 		VI_UNLOCK(vp);
2297 	}
2298 	/*
2299 	 * Reclaim the vnode.
2300 	 */
2301 	if (VOP_RECLAIM(vp, td))
2302 		panic("vgone: cannot reclaim");
2303 
2304 	VNASSERT(vp->v_object == NULL, vp,
2305 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
2306 
2307 	/*
2308 	 * Delete from old mount point vnode list.
2309 	 */
2310 	delmntque(vp);
2311 	cache_purge(vp);
2312 	VI_LOCK(vp);
2313 	if (active) {
2314 		v_incr_usecount(vp, -1);
2315 		VNASSERT(vp->v_usecount >= 0, vp, ("vgone: bad ref count"));
2316 	}
2317 	/*
2318 	 * Done with purge, reset to the standard lock and
2319 	 * notify sleepers of the grim news.
2320 	 */
2321 	vp->v_vnlock = &vp->v_lock;
2322 	vp->v_op = &dead_vnodeops;
2323 	vp->v_tag = "none";
2324 	vp->v_type = VBAD;
2325 
2326 	/*
2327 	 * If it is on the freelist and not already at the head,
2328 	 * move it to the head of the list. The test of the
2329 	 * VDOOMED flag and the reference count of zero is because
2330 	 * it will be removed from the free list by vnlru_free,
2331 	 * but will not have its reference count incremented until
2332 	 * after calling vgone. If the reference count were
2333 	 * incremented first, vgone would (incorrectly) try to
2334 	 * close the previous instance of the underlying object.
2335 	 */
2336 	if (vp->v_holdcnt == 0 && !doomed)
2337 		vfreehead(vp);
2338 	VI_UNLOCK(vp);
2339 }
2340 
2341 /*
2342  * Calculate the total number of references to a special device.
2343  */
2344 int
2345 vcount(vp)
2346 	struct vnode *vp;
2347 {
2348 	int count;
2349 
2350 	dev_lock();
2351 	count = vp->v_rdev->si_usecount;
2352 	dev_unlock();
2353 	return (count);
2354 }
2355 
2356 /*
2357  * Same as above, but using the struct cdev *as argument
2358  */
2359 int
2360 count_dev(dev)
2361 	struct cdev *dev;
2362 {
2363 	int count;
2364 
2365 	dev_lock();
2366 	count = dev->si_usecount;
2367 	dev_unlock();
2368 	return(count);
2369 }
2370 
2371 /*
2372  * Print out a description of a vnode.
2373  */
2374 static char *typename[] =
2375 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD"};
2376 
2377 void
2378 vn_printf(struct vnode *vp, const char *fmt, ...)
2379 {
2380 	va_list ap;
2381 	char buf[96];
2382 
2383 	va_start(ap, fmt);
2384 	vprintf(fmt, ap);
2385 	va_end(ap);
2386 	printf("%p: ", (void *)vp);
2387 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
2388 	printf("    usecount %d, writecount %d, refcount %d mountedhere %p\n",
2389 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere);
2390 	buf[0] = '\0';
2391 	buf[1] = '\0';
2392 	if (vp->v_vflag & VV_ROOT)
2393 		strcat(buf, "|VV_ROOT");
2394 	if (vp->v_vflag & VV_TEXT)
2395 		strcat(buf, "|VV_TEXT");
2396 	if (vp->v_vflag & VV_SYSTEM)
2397 		strcat(buf, "|VV_SYSTEM");
2398 	if (vp->v_iflag & VI_DOOMED)
2399 		strcat(buf, "|VI_DOOMED");
2400 	if (vp->v_iflag & VI_FREE)
2401 		strcat(buf, "|VI_FREE");
2402 	printf("    flags (%s)\n", buf + 1);
2403 	if (mtx_owned(VI_MTX(vp)))
2404 		printf(" VI_LOCKed");
2405 	if (vp->v_object != NULL)
2406 		printf("    v_object %p ref %d pages %d\n",
2407 		    vp->v_object, vp->v_object->ref_count,
2408 		    vp->v_object->resident_page_count);
2409 	printf("    ");
2410 	lockmgr_printinfo(vp->v_vnlock);
2411 	printf("\n");
2412 	if (vp->v_data != NULL)
2413 		VOP_PRINT(vp);
2414 }
2415 
2416 #ifdef DDB
2417 #include <ddb/ddb.h>
2418 /*
2419  * List all of the locked vnodes in the system.
2420  * Called when debugging the kernel.
2421  */
2422 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
2423 {
2424 	struct mount *mp, *nmp;
2425 	struct vnode *vp;
2426 
2427 	/*
2428 	 * Note: because this is DDB, we can't obey the locking semantics
2429 	 * for these structures, which means we could catch an inconsistent
2430 	 * state and dereference a nasty pointer.  Not much to be done
2431 	 * about that.
2432 	 */
2433 	printf("Locked vnodes\n");
2434 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
2435 		nmp = TAILQ_NEXT(mp, mnt_list);
2436 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2437 			if (VOP_ISLOCKED(vp, NULL))
2438 				vprint("", vp);
2439 		}
2440 		nmp = TAILQ_NEXT(mp, mnt_list);
2441 	}
2442 }
2443 #endif
2444 
2445 /*
2446  * Fill in a struct xvfsconf based on a struct vfsconf.
2447  */
2448 static void
2449 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp)
2450 {
2451 
2452 	strcpy(xvfsp->vfc_name, vfsp->vfc_name);
2453 	xvfsp->vfc_typenum = vfsp->vfc_typenum;
2454 	xvfsp->vfc_refcount = vfsp->vfc_refcount;
2455 	xvfsp->vfc_flags = vfsp->vfc_flags;
2456 	/*
2457 	 * These are unused in userland, we keep them
2458 	 * to not break binary compatibility.
2459 	 */
2460 	xvfsp->vfc_vfsops = NULL;
2461 	xvfsp->vfc_next = NULL;
2462 }
2463 
2464 /*
2465  * Top level filesystem related information gathering.
2466  */
2467 static int
2468 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
2469 {
2470 	struct vfsconf *vfsp;
2471 	struct xvfsconf xvfsp;
2472 	int error;
2473 
2474 	error = 0;
2475 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2476 		bzero(&xvfsp, sizeof(xvfsp));
2477 		vfsconf2x(vfsp, &xvfsp);
2478 		error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp);
2479 		if (error)
2480 			break;
2481 	}
2482 	return (error);
2483 }
2484 
2485 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist,
2486     "S,xvfsconf", "List of all configured filesystems");
2487 
2488 #ifndef BURN_BRIDGES
2489 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
2490 
2491 static int
2492 vfs_sysctl(SYSCTL_HANDLER_ARGS)
2493 {
2494 	int *name = (int *)arg1 - 1;	/* XXX */
2495 	u_int namelen = arg2 + 1;	/* XXX */
2496 	struct vfsconf *vfsp;
2497 	struct xvfsconf xvfsp;
2498 
2499 	printf("WARNING: userland calling deprecated sysctl, "
2500 	    "please rebuild world\n");
2501 
2502 #if 1 || defined(COMPAT_PRELITE2)
2503 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
2504 	if (namelen == 1)
2505 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
2506 #endif
2507 
2508 	switch (name[1]) {
2509 	case VFS_MAXTYPENUM:
2510 		if (namelen != 2)
2511 			return (ENOTDIR);
2512 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
2513 	case VFS_CONF:
2514 		if (namelen != 3)
2515 			return (ENOTDIR);	/* overloaded */
2516 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list)
2517 			if (vfsp->vfc_typenum == name[2])
2518 				break;
2519 		if (vfsp == NULL)
2520 			return (EOPNOTSUPP);
2521 		bzero(&xvfsp, sizeof(xvfsp));
2522 		vfsconf2x(vfsp, &xvfsp);
2523 		return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
2524 	}
2525 	return (EOPNOTSUPP);
2526 }
2527 
2528 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP,
2529 	vfs_sysctl, "Generic filesystem");
2530 
2531 #if 1 || defined(COMPAT_PRELITE2)
2532 
2533 static int
2534 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
2535 {
2536 	int error;
2537 	struct vfsconf *vfsp;
2538 	struct ovfsconf ovfs;
2539 
2540 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2541 		bzero(&ovfs, sizeof(ovfs));
2542 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
2543 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
2544 		ovfs.vfc_index = vfsp->vfc_typenum;
2545 		ovfs.vfc_refcount = vfsp->vfc_refcount;
2546 		ovfs.vfc_flags = vfsp->vfc_flags;
2547 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
2548 		if (error)
2549 			return error;
2550 	}
2551 	return 0;
2552 }
2553 
2554 #endif /* 1 || COMPAT_PRELITE2 */
2555 #endif /* !BURN_BRIDGES */
2556 
2557 #define KINFO_VNODESLOP		10
2558 #ifdef notyet
2559 /*
2560  * Dump vnode list (via sysctl).
2561  */
2562 /* ARGSUSED */
2563 static int
2564 sysctl_vnode(SYSCTL_HANDLER_ARGS)
2565 {
2566 	struct xvnode *xvn;
2567 	struct thread *td = req->td;
2568 	struct mount *mp;
2569 	struct vnode *vp;
2570 	int error, len, n;
2571 
2572 	/*
2573 	 * Stale numvnodes access is not fatal here.
2574 	 */
2575 	req->lock = 0;
2576 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
2577 	if (!req->oldptr)
2578 		/* Make an estimate */
2579 		return (SYSCTL_OUT(req, 0, len));
2580 
2581 	error = sysctl_wire_old_buffer(req, 0);
2582 	if (error != 0)
2583 		return (error);
2584 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
2585 	n = 0;
2586 	mtx_lock(&mountlist_mtx);
2587 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
2588 		if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td))
2589 			continue;
2590 		MNT_ILOCK(mp);
2591 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2592 			if (n == len)
2593 				break;
2594 			vref(vp);
2595 			xvn[n].xv_size = sizeof *xvn;
2596 			xvn[n].xv_vnode = vp;
2597 			xvn[n].xv_id = 0;	/* XXX compat */
2598 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
2599 			XV_COPY(usecount);
2600 			XV_COPY(writecount);
2601 			XV_COPY(holdcnt);
2602 			XV_COPY(mount);
2603 			XV_COPY(numoutput);
2604 			XV_COPY(type);
2605 #undef XV_COPY
2606 			xvn[n].xv_flag = vp->v_vflag;
2607 
2608 			switch (vp->v_type) {
2609 			case VREG:
2610 			case VDIR:
2611 			case VLNK:
2612 				break;
2613 			case VBLK:
2614 			case VCHR:
2615 				if (vp->v_rdev == NULL) {
2616 					vrele(vp);
2617 					continue;
2618 				}
2619 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
2620 				break;
2621 			case VSOCK:
2622 				xvn[n].xv_socket = vp->v_socket;
2623 				break;
2624 			case VFIFO:
2625 				xvn[n].xv_fifo = vp->v_fifoinfo;
2626 				break;
2627 			case VNON:
2628 			case VBAD:
2629 			default:
2630 				/* shouldn't happen? */
2631 				vrele(vp);
2632 				continue;
2633 			}
2634 			vrele(vp);
2635 			++n;
2636 		}
2637 		MNT_IUNLOCK(mp);
2638 		mtx_lock(&mountlist_mtx);
2639 		vfs_unbusy(mp, td);
2640 		if (n == len)
2641 			break;
2642 	}
2643 	mtx_unlock(&mountlist_mtx);
2644 
2645 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
2646 	free(xvn, M_TEMP);
2647 	return (error);
2648 }
2649 
2650 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD,
2651 	0, 0, sysctl_vnode, "S,xvnode", "");
2652 #endif
2653 
2654 /*
2655  * Unmount all filesystems. The list is traversed in reverse order
2656  * of mounting to avoid dependencies.
2657  */
2658 void
2659 vfs_unmountall()
2660 {
2661 	struct mount *mp;
2662 	struct thread *td;
2663 	int error;
2664 
2665 	KASSERT(curthread != NULL, ("vfs_unmountall: NULL curthread"));
2666 	td = curthread;
2667 	/*
2668 	 * Since this only runs when rebooting, it is not interlocked.
2669 	 */
2670 	while(!TAILQ_EMPTY(&mountlist)) {
2671 		mp = TAILQ_LAST(&mountlist, mntlist);
2672 		error = dounmount(mp, MNT_FORCE, td);
2673 		if (error) {
2674 			TAILQ_REMOVE(&mountlist, mp, mnt_list);
2675 			printf("unmount of %s failed (",
2676 			    mp->mnt_stat.f_mntonname);
2677 			if (error == EBUSY)
2678 				printf("BUSY)\n");
2679 			else
2680 				printf("%d)\n", error);
2681 		} else {
2682 			/* The unmount has removed mp from the mountlist */
2683 		}
2684 	}
2685 }
2686 
2687 /*
2688  * perform msync on all vnodes under a mount point
2689  * the mount point must be locked.
2690  */
2691 void
2692 vfs_msync(struct mount *mp, int flags)
2693 {
2694 	struct vnode *vp, *nvp;
2695 	struct vm_object *obj;
2696 	int tries;
2697 
2698 	tries = 5;
2699 	MNT_ILOCK(mp);
2700 loop:
2701 	TAILQ_FOREACH_SAFE(vp, &mp->mnt_nvnodelist, v_nmntvnodes, nvp) {
2702 		if (vp->v_mount != mp) {
2703 			if (--tries > 0)
2704 				goto loop;
2705 			break;
2706 		}
2707 
2708 		VI_LOCK(vp);
2709 		if ((vp->v_iflag & VI_OBJDIRTY) &&
2710 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) {
2711 			MNT_IUNLOCK(mp);
2712 			if (!vget(vp,
2713 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
2714 			    curthread)) {
2715 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
2716 					vput(vp);
2717 					MNT_ILOCK(mp);
2718 					continue;
2719 				}
2720 
2721 				obj = vp->v_object;
2722 				if (obj != NULL) {
2723 					VM_OBJECT_LOCK(obj);
2724 					vm_object_page_clean(obj, 0, 0,
2725 					    flags == MNT_WAIT ?
2726 					    OBJPC_SYNC : OBJPC_NOSYNC);
2727 					VM_OBJECT_UNLOCK(obj);
2728 				}
2729 				vput(vp);
2730 			}
2731 			MNT_ILOCK(mp);
2732 			if (TAILQ_NEXT(vp, v_nmntvnodes) != nvp) {
2733 				if (--tries > 0)
2734 					goto loop;
2735 				break;
2736 			}
2737 		} else
2738 			VI_UNLOCK(vp);
2739 	}
2740 	MNT_IUNLOCK(mp);
2741 }
2742 
2743 /*
2744  * Mark a vnode as free, putting it up for recycling.
2745  */
2746 static void
2747 vfree(struct vnode *vp)
2748 {
2749 
2750 	ASSERT_VI_LOCKED(vp, "vfree");
2751 	mtx_lock(&vnode_free_list_mtx);
2752 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, ("vnode already free"));
2753 	VNASSERT(VSHOULDFREE(vp), vp, ("vfree: freeing when we shouldn't"));
2754 	if (vp->v_iflag & (VI_AGE|VI_DOOMED)) {
2755 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2756 	} else {
2757 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
2758 	}
2759 	freevnodes++;
2760 	mtx_unlock(&vnode_free_list_mtx);
2761 	vp->v_iflag &= ~(VI_AGE|VI_DOOMED);
2762 	vp->v_iflag |= VI_FREE;
2763 }
2764 
2765 /*
2766  * Move a vnode to the head of the free list.
2767  */
2768 static void
2769 vfreehead(struct vnode *vp)
2770 {
2771 	mtx_lock(&vnode_free_list_mtx);
2772 	if (vp->v_iflag & VI_FREE) {
2773 		TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2774 	} else {
2775 		vp->v_iflag |= VI_FREE;
2776 		freevnodes++;
2777 	}
2778 	TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2779 	mtx_unlock(&vnode_free_list_mtx);
2780 }
2781 
2782 /*
2783  * Opposite of vfree() - mark a vnode as in use.
2784  */
2785 static void
2786 vbusy(struct vnode *vp)
2787 {
2788 
2789 	ASSERT_VI_LOCKED(vp, "vbusy");
2790 	VNASSERT((vp->v_iflag & VI_FREE) != 0, vp, ("vnode not free"));
2791 
2792 	mtx_lock(&vnode_free_list_mtx);
2793 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2794 	freevnodes--;
2795 	mtx_unlock(&vnode_free_list_mtx);
2796 
2797 	vp->v_iflag &= ~(VI_FREE|VI_AGE);
2798 }
2799 
2800 /*
2801  * Initalize per-vnode helper structure to hold poll-related state.
2802  */
2803 void
2804 v_addpollinfo(struct vnode *vp)
2805 {
2806 	struct vpollinfo *vi;
2807 
2808 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK);
2809 	if (vp->v_pollinfo != NULL) {
2810 		uma_zfree(vnodepoll_zone, vi);
2811 		return;
2812 	}
2813 	vp->v_pollinfo = vi;
2814 	mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
2815 	knlist_init(&vp->v_pollinfo->vpi_selinfo.si_note,
2816 	    &vp->v_pollinfo->vpi_lock);
2817 }
2818 
2819 /*
2820  * Record a process's interest in events which might happen to
2821  * a vnode.  Because poll uses the historic select-style interface
2822  * internally, this routine serves as both the ``check for any
2823  * pending events'' and the ``record my interest in future events''
2824  * functions.  (These are done together, while the lock is held,
2825  * to avoid race conditions.)
2826  */
2827 int
2828 vn_pollrecord(vp, td, events)
2829 	struct vnode *vp;
2830 	struct thread *td;
2831 	short events;
2832 {
2833 
2834 	if (vp->v_pollinfo == NULL)
2835 		v_addpollinfo(vp);
2836 	mtx_lock(&vp->v_pollinfo->vpi_lock);
2837 	if (vp->v_pollinfo->vpi_revents & events) {
2838 		/*
2839 		 * This leaves events we are not interested
2840 		 * in available for the other process which
2841 		 * which presumably had requested them
2842 		 * (otherwise they would never have been
2843 		 * recorded).
2844 		 */
2845 		events &= vp->v_pollinfo->vpi_revents;
2846 		vp->v_pollinfo->vpi_revents &= ~events;
2847 
2848 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
2849 		return events;
2850 	}
2851 	vp->v_pollinfo->vpi_events |= events;
2852 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
2853 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
2854 	return 0;
2855 }
2856 
2857 /*
2858  * Routine to create and manage a filesystem syncer vnode.
2859  */
2860 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
2861 static int	sync_fsync(struct  vop_fsync_args *);
2862 static int	sync_inactive(struct  vop_inactive_args *);
2863 static int	sync_reclaim(struct  vop_reclaim_args *);
2864 
2865 static struct vop_vector sync_vnodeops = {
2866 	.vop_bypass =	VOP_EOPNOTSUPP,
2867 	.vop_close =	sync_close,		/* close */
2868 	.vop_fsync =	sync_fsync,		/* fsync */
2869 	.vop_inactive =	sync_inactive,	/* inactive */
2870 	.vop_reclaim =	sync_reclaim,	/* reclaim */
2871 	.vop_lock =	vop_stdlock,	/* lock */
2872 	.vop_unlock =	vop_stdunlock,	/* unlock */
2873 	.vop_islocked =	vop_stdislocked,	/* islocked */
2874 };
2875 
2876 /*
2877  * Create a new filesystem syncer vnode for the specified mount point.
2878  */
2879 int
2880 vfs_allocate_syncvnode(mp)
2881 	struct mount *mp;
2882 {
2883 	struct vnode *vp;
2884 	static long start, incr, next;
2885 	int error;
2886 
2887 	/* Allocate a new vnode */
2888 	if ((error = getnewvnode("syncer", mp, &sync_vnodeops, &vp)) != 0) {
2889 		mp->mnt_syncer = NULL;
2890 		return (error);
2891 	}
2892 	vp->v_type = VNON;
2893 	/*
2894 	 * Place the vnode onto the syncer worklist. We attempt to
2895 	 * scatter them about on the list so that they will go off
2896 	 * at evenly distributed times even if all the filesystems
2897 	 * are mounted at once.
2898 	 */
2899 	next += incr;
2900 	if (next == 0 || next > syncer_maxdelay) {
2901 		start /= 2;
2902 		incr /= 2;
2903 		if (start == 0) {
2904 			start = syncer_maxdelay / 2;
2905 			incr = syncer_maxdelay;
2906 		}
2907 		next = start;
2908 	}
2909 	VI_LOCK(vp);
2910 	vn_syncer_add_to_worklist(&vp->v_bufobj,
2911 	    syncdelay > 0 ? next % syncdelay : 0);
2912 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
2913 	mtx_lock(&sync_mtx);
2914 	sync_vnode_count++;
2915 	mtx_unlock(&sync_mtx);
2916 	VI_UNLOCK(vp);
2917 	mp->mnt_syncer = vp;
2918 	return (0);
2919 }
2920 
2921 /*
2922  * Do a lazy sync of the filesystem.
2923  */
2924 static int
2925 sync_fsync(ap)
2926 	struct vop_fsync_args /* {
2927 		struct vnode *a_vp;
2928 		struct ucred *a_cred;
2929 		int a_waitfor;
2930 		struct thread *a_td;
2931 	} */ *ap;
2932 {
2933 	struct vnode *syncvp = ap->a_vp;
2934 	struct mount *mp = syncvp->v_mount;
2935 	struct thread *td = ap->a_td;
2936 	int error, asyncflag;
2937 	struct bufobj *bo;
2938 
2939 	/*
2940 	 * We only need to do something if this is a lazy evaluation.
2941 	 */
2942 	if (ap->a_waitfor != MNT_LAZY)
2943 		return (0);
2944 
2945 	/*
2946 	 * Move ourselves to the back of the sync list.
2947 	 */
2948 	bo = &syncvp->v_bufobj;
2949 	BO_LOCK(bo);
2950 	vn_syncer_add_to_worklist(bo, syncdelay);
2951 	BO_UNLOCK(bo);
2952 
2953 	/*
2954 	 * Walk the list of vnodes pushing all that are dirty and
2955 	 * not already on the sync list.
2956 	 */
2957 	mtx_lock(&mountlist_mtx);
2958 	if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) {
2959 		mtx_unlock(&mountlist_mtx);
2960 		return (0);
2961 	}
2962 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
2963 		vfs_unbusy(mp, td);
2964 		return (0);
2965 	}
2966 	asyncflag = mp->mnt_flag & MNT_ASYNC;
2967 	mp->mnt_flag &= ~MNT_ASYNC;
2968 	vfs_msync(mp, MNT_NOWAIT);
2969 	error = VFS_SYNC(mp, MNT_LAZY, td);
2970 	if (asyncflag)
2971 		mp->mnt_flag |= MNT_ASYNC;
2972 	vn_finished_write(mp);
2973 	vfs_unbusy(mp, td);
2974 	return (error);
2975 }
2976 
2977 /*
2978  * The syncer vnode is no referenced.
2979  */
2980 static int
2981 sync_inactive(ap)
2982 	struct vop_inactive_args /* {
2983 		struct vnode *a_vp;
2984 		struct thread *a_td;
2985 	} */ *ap;
2986 {
2987 
2988 	vgone(ap->a_vp);
2989 	return (0);
2990 }
2991 
2992 /*
2993  * The syncer vnode is no longer needed and is being decommissioned.
2994  *
2995  * Modifications to the worklist must be protected by sync_mtx.
2996  */
2997 static int
2998 sync_reclaim(ap)
2999 	struct vop_reclaim_args /* {
3000 		struct vnode *a_vp;
3001 	} */ *ap;
3002 {
3003 	struct vnode *vp = ap->a_vp;
3004 	struct bufobj *bo;
3005 
3006 	VI_LOCK(vp);
3007 	bo = &vp->v_bufobj;
3008 	vp->v_mount->mnt_syncer = NULL;
3009 	if (bo->bo_flag & BO_ONWORKLST) {
3010 		mtx_lock(&sync_mtx);
3011 		LIST_REMOVE(bo, bo_synclist);
3012  		syncer_worklist_len--;
3013 		sync_vnode_count--;
3014 		mtx_unlock(&sync_mtx);
3015 		bo->bo_flag &= ~BO_ONWORKLST;
3016 	}
3017 	VI_UNLOCK(vp);
3018 
3019 	return (0);
3020 }
3021 
3022 /*
3023  * Check if vnode represents a disk device
3024  */
3025 int
3026 vn_isdisk(vp, errp)
3027 	struct vnode *vp;
3028 	int *errp;
3029 {
3030 	int error;
3031 
3032 	error = 0;
3033 	dev_lock();
3034 	if (vp->v_type != VCHR)
3035 		error = ENOTBLK;
3036 	else if (vp->v_rdev == NULL)
3037 		error = ENXIO;
3038 	else if (vp->v_rdev->si_devsw == NULL)
3039 		error = ENXIO;
3040 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
3041 		error = ENOTBLK;
3042 	dev_unlock();
3043 	if (errp != NULL)
3044 		*errp = error;
3045 	return (error == 0);
3046 }
3047 
3048 /*
3049  * Common filesystem object access control check routine.  Accepts a
3050  * vnode's type, "mode", uid and gid, requested access mode, credentials,
3051  * and optional call-by-reference privused argument allowing vaccess()
3052  * to indicate to the caller whether privilege was used to satisfy the
3053  * request (obsoleted).  Returns 0 on success, or an errno on failure.
3054  */
3055 int
3056 vaccess(type, file_mode, file_uid, file_gid, acc_mode, cred, privused)
3057 	enum vtype type;
3058 	mode_t file_mode;
3059 	uid_t file_uid;
3060 	gid_t file_gid;
3061 	mode_t acc_mode;
3062 	struct ucred *cred;
3063 	int *privused;
3064 {
3065 	mode_t dac_granted;
3066 #ifdef CAPABILITIES
3067 	mode_t cap_granted;
3068 #endif
3069 
3070 	/*
3071 	 * Look for a normal, non-privileged way to access the file/directory
3072 	 * as requested.  If it exists, go with that.
3073 	 */
3074 
3075 	if (privused != NULL)
3076 		*privused = 0;
3077 
3078 	dac_granted = 0;
3079 
3080 	/* Check the owner. */
3081 	if (cred->cr_uid == file_uid) {
3082 		dac_granted |= VADMIN;
3083 		if (file_mode & S_IXUSR)
3084 			dac_granted |= VEXEC;
3085 		if (file_mode & S_IRUSR)
3086 			dac_granted |= VREAD;
3087 		if (file_mode & S_IWUSR)
3088 			dac_granted |= (VWRITE | VAPPEND);
3089 
3090 		if ((acc_mode & dac_granted) == acc_mode)
3091 			return (0);
3092 
3093 		goto privcheck;
3094 	}
3095 
3096 	/* Otherwise, check the groups (first match) */
3097 	if (groupmember(file_gid, cred)) {
3098 		if (file_mode & S_IXGRP)
3099 			dac_granted |= VEXEC;
3100 		if (file_mode & S_IRGRP)
3101 			dac_granted |= VREAD;
3102 		if (file_mode & S_IWGRP)
3103 			dac_granted |= (VWRITE | VAPPEND);
3104 
3105 		if ((acc_mode & dac_granted) == acc_mode)
3106 			return (0);
3107 
3108 		goto privcheck;
3109 	}
3110 
3111 	/* Otherwise, check everyone else. */
3112 	if (file_mode & S_IXOTH)
3113 		dac_granted |= VEXEC;
3114 	if (file_mode & S_IROTH)
3115 		dac_granted |= VREAD;
3116 	if (file_mode & S_IWOTH)
3117 		dac_granted |= (VWRITE | VAPPEND);
3118 	if ((acc_mode & dac_granted) == acc_mode)
3119 		return (0);
3120 
3121 privcheck:
3122 	if (!suser_cred(cred, SUSER_ALLOWJAIL)) {
3123 		/* XXX audit: privilege used */
3124 		if (privused != NULL)
3125 			*privused = 1;
3126 		return (0);
3127 	}
3128 
3129 #ifdef CAPABILITIES
3130 	/*
3131 	 * Build a capability mask to determine if the set of capabilities
3132 	 * satisfies the requirements when combined with the granted mask
3133 	 * from above.
3134 	 * For each capability, if the capability is required, bitwise
3135 	 * or the request type onto the cap_granted mask.
3136 	 */
3137 	cap_granted = 0;
3138 
3139 	if (type == VDIR) {
3140 		/*
3141 		 * For directories, use CAP_DAC_READ_SEARCH to satisfy
3142 		 * VEXEC requests, instead of CAP_DAC_EXECUTE.
3143 		 */
3144 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3145 		    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL))
3146 			cap_granted |= VEXEC;
3147 	} else {
3148 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3149 		    !cap_check(cred, NULL, CAP_DAC_EXECUTE, SUSER_ALLOWJAIL))
3150 			cap_granted |= VEXEC;
3151 	}
3152 
3153 	if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) &&
3154 	    !cap_check(cred, NULL, CAP_DAC_READ_SEARCH, SUSER_ALLOWJAIL))
3155 		cap_granted |= VREAD;
3156 
3157 	if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
3158 	    !cap_check(cred, NULL, CAP_DAC_WRITE, SUSER_ALLOWJAIL))
3159 		cap_granted |= (VWRITE | VAPPEND);
3160 
3161 	if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
3162 	    !cap_check(cred, NULL, CAP_FOWNER, SUSER_ALLOWJAIL))
3163 		cap_granted |= VADMIN;
3164 
3165 	if ((acc_mode & (cap_granted | dac_granted)) == acc_mode) {
3166 		/* XXX audit: privilege used */
3167 		if (privused != NULL)
3168 			*privused = 1;
3169 		return (0);
3170 	}
3171 #endif
3172 
3173 	return ((acc_mode & VADMIN) ? EPERM : EACCES);
3174 }
3175 
3176 /*
3177  * Credential check based on process requesting service, and per-attribute
3178  * permissions.
3179  */
3180 int
3181 extattr_check_cred(struct vnode *vp, int attrnamespace,
3182     struct ucred *cred, struct thread *td, int access)
3183 {
3184 
3185 	/*
3186 	 * Kernel-invoked always succeeds.
3187 	 */
3188 	if (cred == NOCRED)
3189 		return (0);
3190 
3191 	/*
3192 	 * Do not allow privileged processes in jail to directly
3193 	 * manipulate system attributes.
3194 	 *
3195 	 * XXX What capability should apply here?
3196 	 * Probably CAP_SYS_SETFFLAG.
3197 	 */
3198 	switch (attrnamespace) {
3199 	case EXTATTR_NAMESPACE_SYSTEM:
3200 		/* Potentially should be: return (EPERM); */
3201 		return (suser_cred(cred, 0));
3202 	case EXTATTR_NAMESPACE_USER:
3203 		return (VOP_ACCESS(vp, access, cred, td));
3204 	default:
3205 		return (EPERM);
3206 	}
3207 }
3208 
3209 #ifdef DEBUG_VFS_LOCKS
3210 /*
3211  * This only exists to supress warnings from unlocked specfs accesses.  It is
3212  * no longer ok to have an unlocked VFS.
3213  */
3214 #define	IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD)
3215 
3216 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
3217 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, "");
3218 
3219 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
3220 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, "");
3221 
3222 int vfs_badlock_print = 1;	/* Print lock violations. */
3223 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, "");
3224 
3225 #ifdef KDB
3226 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
3227 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, "");
3228 #endif
3229 
3230 static void
3231 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
3232 {
3233 
3234 #ifdef KDB
3235 	if (vfs_badlock_backtrace)
3236 		kdb_backtrace();
3237 #endif
3238 	if (vfs_badlock_print)
3239 		printf("%s: %p %s\n", str, (void *)vp, msg);
3240 	if (vfs_badlock_ddb)
3241 		kdb_enter("lock violation");
3242 }
3243 
3244 void
3245 assert_vi_locked(struct vnode *vp, const char *str)
3246 {
3247 
3248 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
3249 		vfs_badlock("interlock is not locked but should be", str, vp);
3250 }
3251 
3252 void
3253 assert_vi_unlocked(struct vnode *vp, const char *str)
3254 {
3255 
3256 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
3257 		vfs_badlock("interlock is locked but should not be", str, vp);
3258 }
3259 
3260 void
3261 assert_vop_locked(struct vnode *vp, const char *str)
3262 {
3263 
3264 	if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0)
3265 		vfs_badlock("is not locked but should be", str, vp);
3266 }
3267 
3268 void
3269 assert_vop_unlocked(struct vnode *vp, const char *str)
3270 {
3271 
3272 	if (vp && !IGNORE_LOCK(vp) &&
3273 	    VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE)
3274 		vfs_badlock("is locked but should not be", str, vp);
3275 }
3276 
3277 void
3278 assert_vop_elocked(struct vnode *vp, const char *str)
3279 {
3280 
3281 	if (vp && !IGNORE_LOCK(vp) &&
3282 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE)
3283 		vfs_badlock("is not exclusive locked but should be", str, vp);
3284 }
3285 
3286 #if 0
3287 void
3288 assert_vop_elocked_other(struct vnode *vp, const char *str)
3289 {
3290 
3291 	if (vp && !IGNORE_LOCK(vp) &&
3292 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER)
3293 		vfs_badlock("is not exclusive locked by another thread",
3294 		    str, vp);
3295 }
3296 
3297 void
3298 assert_vop_slocked(struct vnode *vp, const char *str)
3299 {
3300 
3301 	if (vp && !IGNORE_LOCK(vp) &&
3302 	    VOP_ISLOCKED(vp, curthread) != LK_SHARED)
3303 		vfs_badlock("is not locked shared but should be", str, vp);
3304 }
3305 #endif /* 0 */
3306 
3307 void
3308 vop_rename_pre(void *ap)
3309 {
3310 	struct vop_rename_args *a = ap;
3311 
3312 	if (a->a_tvp)
3313 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
3314 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
3315 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
3316 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
3317 
3318 	/* Check the source (from). */
3319 	if (a->a_tdvp != a->a_fdvp)
3320 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
3321 	if (a->a_tvp != a->a_fvp)
3322 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: tvp locked");
3323 
3324 	/* Check the target. */
3325 	if (a->a_tvp)
3326 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
3327 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
3328 }
3329 
3330 void
3331 vop_strategy_pre(void *ap)
3332 {
3333 	struct vop_strategy_args *a;
3334 	struct buf *bp;
3335 
3336 	a = ap;
3337 	bp = a->a_bp;
3338 
3339 	/*
3340 	 * Cluster ops lock their component buffers but not the IO container.
3341 	 */
3342 	if ((bp->b_flags & B_CLUSTER) != 0)
3343 		return;
3344 
3345 	if (BUF_REFCNT(bp) < 1) {
3346 		if (vfs_badlock_print)
3347 			printf(
3348 			    "VOP_STRATEGY: bp is not locked but should be\n");
3349 		if (vfs_badlock_ddb)
3350 			kdb_enter("lock violation");
3351 	}
3352 }
3353 
3354 void
3355 vop_lookup_pre(void *ap)
3356 {
3357 	struct vop_lookup_args *a;
3358 	struct vnode *dvp;
3359 
3360 	a = ap;
3361 	dvp = a->a_dvp;
3362 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3363 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3364 }
3365 
3366 void
3367 vop_lookup_post(void *ap, int rc)
3368 {
3369 	struct vop_lookup_args *a;
3370 	struct vnode *dvp;
3371 	struct vnode *vp;
3372 
3373 	a = ap;
3374 	dvp = a->a_dvp;
3375 	vp = *(a->a_vpp);
3376 
3377 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3378 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3379 
3380 	if (!rc)
3381 		ASSERT_VOP_LOCKED(vp, "VOP_LOOKUP (child)");
3382 }
3383 
3384 void
3385 vop_lock_pre(void *ap)
3386 {
3387 	struct vop_lock_args *a = ap;
3388 
3389 	if ((a->a_flags & LK_INTERLOCK) == 0)
3390 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3391 	else
3392 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
3393 }
3394 
3395 void
3396 vop_lock_post(void *ap, int rc)
3397 {
3398 	struct vop_lock_args *a = ap;
3399 
3400 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3401 	if (rc == 0)
3402 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
3403 }
3404 
3405 void
3406 vop_unlock_pre(void *ap)
3407 {
3408 	struct vop_unlock_args *a = ap;
3409 
3410 	if (a->a_flags & LK_INTERLOCK)
3411 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
3412 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
3413 }
3414 
3415 void
3416 vop_unlock_post(void *ap, int rc)
3417 {
3418 	struct vop_unlock_args *a = ap;
3419 
3420 	if (a->a_flags & LK_INTERLOCK)
3421 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
3422 }
3423 #endif /* DEBUG_VFS_LOCKS */
3424 
3425 static struct knlist fs_knlist;
3426 
3427 static void
3428 vfs_event_init(void *arg)
3429 {
3430 	knlist_init(&fs_knlist, NULL);
3431 }
3432 /* XXX - correct order? */
3433 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
3434 
3435 void
3436 vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused)
3437 {
3438 
3439 	KNOTE_UNLOCKED(&fs_knlist, event);
3440 }
3441 
3442 static int	filt_fsattach(struct knote *kn);
3443 static void	filt_fsdetach(struct knote *kn);
3444 static int	filt_fsevent(struct knote *kn, long hint);
3445 
3446 struct filterops fs_filtops =
3447 	{ 0, filt_fsattach, filt_fsdetach, filt_fsevent };
3448 
3449 static int
3450 filt_fsattach(struct knote *kn)
3451 {
3452 
3453 	kn->kn_flags |= EV_CLEAR;
3454 	knlist_add(&fs_knlist, kn, 0);
3455 	return (0);
3456 }
3457 
3458 static void
3459 filt_fsdetach(struct knote *kn)
3460 {
3461 
3462 	knlist_remove(&fs_knlist, kn, 0);
3463 }
3464 
3465 static int
3466 filt_fsevent(struct knote *kn, long hint)
3467 {
3468 
3469 	kn->kn_fflags |= hint;
3470 	return (kn->kn_fflags != 0);
3471 }
3472 
3473 static int
3474 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
3475 {
3476 	struct vfsidctl vc;
3477 	int error;
3478 	struct mount *mp;
3479 
3480 	error = SYSCTL_IN(req, &vc, sizeof(vc));
3481 	if (error)
3482 		return (error);
3483 	if (vc.vc_vers != VFS_CTL_VERS1)
3484 		return (EINVAL);
3485 	mp = vfs_getvfs(&vc.vc_fsid);
3486 	if (mp == NULL)
3487 		return (ENOENT);
3488 	/* ensure that a specific sysctl goes to the right filesystem. */
3489 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
3490 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
3491 		return (EINVAL);
3492 	}
3493 	VCTLTOREQ(&vc, req);
3494 	return (VFS_SYSCTL(mp, vc.vc_op, req));
3495 }
3496 
3497 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR,
3498         NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid");
3499 
3500 /*
3501  * Function to initialize a va_filerev field sensibly.
3502  * XXX: Wouldn't a random number make a lot more sense ??
3503  */
3504 u_quad_t
3505 init_va_filerev(void)
3506 {
3507 	struct bintime bt;
3508 
3509 	getbinuptime(&bt);
3510 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
3511 }
3512