xref: /freebsd/sys/kern/vfs_subr.c (revision 2be1a816b9ff69588e55be0a84cbe2a31efc0f2f)
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/dirent.h>
53 #include <sys/event.h>
54 #include <sys/eventhandler.h>
55 #include <sys/extattr.h>
56 #include <sys/file.h>
57 #include <sys/fcntl.h>
58 #include <sys/jail.h>
59 #include <sys/kdb.h>
60 #include <sys/kernel.h>
61 #include <sys/kthread.h>
62 #include <sys/lockf.h>
63 #include <sys/malloc.h>
64 #include <sys/mount.h>
65 #include <sys/namei.h>
66 #include <sys/priv.h>
67 #include <sys/reboot.h>
68 #include <sys/sleepqueue.h>
69 #include <sys/stat.h>
70 #include <sys/sysctl.h>
71 #include <sys/syslog.h>
72 #include <sys/vmmeter.h>
73 #include <sys/vnode.h>
74 
75 #include <machine/stdarg.h>
76 
77 #include <security/mac/mac_framework.h>
78 
79 #include <vm/vm.h>
80 #include <vm/vm_object.h>
81 #include <vm/vm_extern.h>
82 #include <vm/pmap.h>
83 #include <vm/vm_map.h>
84 #include <vm/vm_page.h>
85 #include <vm/vm_kern.h>
86 #include <vm/uma.h>
87 
88 #ifdef DDB
89 #include <ddb/ddb.h>
90 #endif
91 
92 static MALLOC_DEFINE(M_NETADDR, "subr_export_host", "Export host address structure");
93 
94 static void	delmntque(struct vnode *vp);
95 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
96 		    int slpflag, int slptimeo);
97 static void	syncer_shutdown(void *arg, int howto);
98 static int	vtryrecycle(struct vnode *vp);
99 static void	vbusy(struct vnode *vp);
100 static void	vinactive(struct vnode *, struct thread *);
101 static void	v_incr_usecount(struct vnode *);
102 static void	v_decr_usecount(struct vnode *);
103 static void	v_decr_useonly(struct vnode *);
104 static void	v_upgrade_usecount(struct vnode *);
105 static void	vfree(struct vnode *);
106 static void	vnlru_free(int);
107 static void	vdestroy(struct vnode *);
108 static void	vgonel(struct vnode *);
109 static void	vfs_knllock(void *arg);
110 static void	vfs_knlunlock(void *arg);
111 static int	vfs_knllocked(void *arg);
112 
113 
114 /*
115  * Enable Giant pushdown based on whether or not the vm is mpsafe in this
116  * build.  Without mpsafevm the buffer cache can not run Giant free.
117  */
118 int mpsafe_vfs = 1;
119 TUNABLE_INT("debug.mpsafevfs", &mpsafe_vfs);
120 SYSCTL_INT(_debug, OID_AUTO, mpsafevfs, CTLFLAG_RD, &mpsafe_vfs, 0,
121     "MPSAFE VFS");
122 
123 /*
124  * Number of vnodes in existence.  Increased whenever getnewvnode()
125  * allocates a new vnode, decreased on vdestroy() called on VI_DOOMed
126  * vnode.
127  */
128 static unsigned long	numvnodes;
129 
130 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, "");
131 
132 /*
133  * Conversion tables for conversion from vnode types to inode formats
134  * and back.
135  */
136 enum vtype iftovt_tab[16] = {
137 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
138 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
139 };
140 int vttoif_tab[10] = {
141 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
142 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
143 };
144 
145 /*
146  * List of vnodes that are ready for recycling.
147  */
148 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
149 
150 /*
151  * Free vnode target.  Free vnodes may simply be files which have been stat'd
152  * but not read.  This is somewhat common, and a small cache of such files
153  * should be kept to avoid recreation costs.
154  */
155 static u_long wantfreevnodes;
156 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, "");
157 /* Number of vnodes in the free list. */
158 static u_long freevnodes;
159 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, "");
160 
161 /*
162  * Various variables used for debugging the new implementation of
163  * reassignbuf().
164  * XXX these are probably of (very) limited utility now.
165  */
166 static int reassignbufcalls;
167 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, "");
168 
169 /*
170  * Cache for the mount type id assigned to NFS.  This is used for
171  * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c.
172  */
173 int	nfs_mount_type = -1;
174 
175 /* To keep more than one thread at a time from running vfs_getnewfsid */
176 static struct mtx mntid_mtx;
177 
178 /*
179  * Lock for any access to the following:
180  *	vnode_free_list
181  *	numvnodes
182  *	freevnodes
183  */
184 static struct mtx vnode_free_list_mtx;
185 
186 /* Publicly exported FS */
187 struct nfs_public nfs_pub;
188 
189 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
190 static uma_zone_t vnode_zone;
191 static uma_zone_t vnodepoll_zone;
192 
193 /* Set to 1 to print out reclaim of active vnodes */
194 int	prtactive;
195 
196 /*
197  * The workitem queue.
198  *
199  * It is useful to delay writes of file data and filesystem metadata
200  * for tens of seconds so that quickly created and deleted files need
201  * not waste disk bandwidth being created and removed. To realize this,
202  * we append vnodes to a "workitem" queue. When running with a soft
203  * updates implementation, most pending metadata dependencies should
204  * not wait for more than a few seconds. Thus, mounted on block devices
205  * are delayed only about a half the time that file data is delayed.
206  * Similarly, directory updates are more critical, so are only delayed
207  * about a third the time that file data is delayed. Thus, there are
208  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
209  * one each second (driven off the filesystem syncer process). The
210  * syncer_delayno variable indicates the next queue that is to be processed.
211  * Items that need to be processed soon are placed in this queue:
212  *
213  *	syncer_workitem_pending[syncer_delayno]
214  *
215  * A delay of fifteen seconds is done by placing the request fifteen
216  * entries later in the queue:
217  *
218  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
219  *
220  */
221 static int syncer_delayno;
222 static long syncer_mask;
223 LIST_HEAD(synclist, bufobj);
224 static struct synclist *syncer_workitem_pending;
225 /*
226  * The sync_mtx protects:
227  *	bo->bo_synclist
228  *	sync_vnode_count
229  *	syncer_delayno
230  *	syncer_state
231  *	syncer_workitem_pending
232  *	syncer_worklist_len
233  *	rushjob
234  */
235 static struct mtx sync_mtx;
236 
237 #define SYNCER_MAXDELAY		32
238 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
239 static int syncdelay = 30;		/* max time to delay syncing data */
240 static int filedelay = 30;		/* time to delay syncing files */
241 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, "");
242 static int dirdelay = 29;		/* time to delay syncing directories */
243 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, "");
244 static int metadelay = 28;		/* time to delay syncing metadata */
245 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, "");
246 static int rushjob;		/* number of slots to run ASAP */
247 static int stat_rush_requests;	/* number of times I/O speeded up */
248 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, "");
249 
250 /*
251  * When shutting down the syncer, run it at four times normal speed.
252  */
253 #define SYNCER_SHUTDOWN_SPEEDUP		4
254 static int sync_vnode_count;
255 static int syncer_worklist_len;
256 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
257     syncer_state;
258 
259 /*
260  * Number of vnodes we want to exist at any one time.  This is mostly used
261  * to size hash tables in vnode-related code.  It is normally not used in
262  * getnewvnode(), as wantfreevnodes is normally nonzero.)
263  *
264  * XXX desiredvnodes is historical cruft and should not exist.
265  */
266 int desiredvnodes;
267 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW,
268     &desiredvnodes, 0, "Maximum number of vnodes");
269 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
270     &wantfreevnodes, 0, "Minimum number of vnodes (legacy)");
271 static int vnlru_nowhere;
272 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
273     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
274 
275 /*
276  * Macros to control when a vnode is freed and recycled.  All require
277  * the vnode interlock.
278  */
279 #define VCANRECYCLE(vp) (((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt)
280 #define VSHOULDFREE(vp) (!((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt)
281 #define VSHOULDBUSY(vp) (((vp)->v_iflag & VI_FREE) && (vp)->v_holdcnt)
282 
283 
284 /*
285  * Initialize the vnode management data structures.
286  */
287 #ifndef	MAXVNODES_MAX
288 #define	MAXVNODES_MAX	100000
289 #endif
290 static void
291 vntblinit(void *dummy __unused)
292 {
293 
294 	/*
295 	 * Desiredvnodes is a function of the physical memory size and
296 	 * the kernel's heap size.  Specifically, desiredvnodes scales
297 	 * in proportion to the physical memory size until two fifths
298 	 * of the kernel's heap size is consumed by vnodes and vm
299 	 * objects.
300 	 */
301 	desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size /
302 	    (5 * (sizeof(struct vm_object) + sizeof(struct vnode))));
303 	if (desiredvnodes > MAXVNODES_MAX) {
304 		if (bootverbose)
305 			printf("Reducing kern.maxvnodes %d -> %d\n",
306 			    desiredvnodes, MAXVNODES_MAX);
307 		desiredvnodes = MAXVNODES_MAX;
308 	}
309 	wantfreevnodes = desiredvnodes / 4;
310 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
311 	TAILQ_INIT(&vnode_free_list);
312 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
313 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
314 	    NULL, NULL, UMA_ALIGN_PTR, 0);
315 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
316 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
317 	/*
318 	 * Initialize the filesystem syncer.
319 	 */
320 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
321 		&syncer_mask);
322 	syncer_maxdelay = syncer_mask + 1;
323 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
324 }
325 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
326 
327 
328 /*
329  * Mark a mount point as busy. Used to synchronize access and to delay
330  * unmounting. Interlock is not released on failure.
331  */
332 int
333 vfs_busy(struct mount *mp, int flags, struct mtx *interlkp,
334     struct thread *td)
335 {
336 	int lkflags;
337 
338 	MNT_ILOCK(mp);
339 	MNT_REF(mp);
340 	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
341 		if (flags & LK_NOWAIT) {
342 			MNT_REL(mp);
343 			MNT_IUNLOCK(mp);
344 			return (ENOENT);
345 		}
346 		if (interlkp)
347 			mtx_unlock(interlkp);
348 		mp->mnt_kern_flag |= MNTK_MWAIT;
349 		/*
350 		 * Since all busy locks are shared except the exclusive
351 		 * lock granted when unmounting, the only place that a
352 		 * wakeup needs to be done is at the release of the
353 		 * exclusive lock at the end of dounmount.
354 		 */
355 		msleep(mp, MNT_MTX(mp), PVFS, "vfs_busy", 0);
356 		MNT_REL(mp);
357 		MNT_IUNLOCK(mp);
358 		if (interlkp)
359 			mtx_lock(interlkp);
360 		return (ENOENT);
361 	}
362 	if (interlkp)
363 		mtx_unlock(interlkp);
364 	lkflags = LK_SHARED | LK_INTERLOCK;
365 	if (lockmgr(&mp->mnt_lock, lkflags, MNT_MTX(mp)))
366 		panic("vfs_busy: unexpected lock failure");
367 	return (0);
368 }
369 
370 /*
371  * Free a busy filesystem.
372  */
373 void
374 vfs_unbusy(struct mount *mp, struct thread *td)
375 {
376 
377 	lockmgr(&mp->mnt_lock, LK_RELEASE, NULL);
378 	vfs_rel(mp);
379 }
380 
381 /*
382  * Lookup a mount point by filesystem identifier.
383  */
384 struct mount *
385 vfs_getvfs(fsid_t *fsid)
386 {
387 	struct mount *mp;
388 
389 	mtx_lock(&mountlist_mtx);
390 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
391 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
392 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
393 			vfs_ref(mp);
394 			mtx_unlock(&mountlist_mtx);
395 			return (mp);
396 		}
397 	}
398 	mtx_unlock(&mountlist_mtx);
399 	return ((struct mount *) 0);
400 }
401 
402 /*
403  * Check if a user can access privileged mount options.
404  */
405 int
406 vfs_suser(struct mount *mp, struct thread *td)
407 {
408 	int error;
409 
410 	/*
411 	 * If the thread is jailed, but this is not a jail-friendly file
412 	 * system, deny immediately.
413 	 */
414 	if (jailed(td->td_ucred) && !(mp->mnt_vfc->vfc_flags & VFCF_JAIL))
415 		return (EPERM);
416 
417 	/*
418 	 * If the file system was mounted outside a jail and a jailed thread
419 	 * tries to access it, deny immediately.
420 	 */
421 	if (!jailed(mp->mnt_cred) && jailed(td->td_ucred))
422 		return (EPERM);
423 
424 	/*
425 	 * If the file system was mounted inside different jail that the jail of
426 	 * the calling thread, deny immediately.
427 	 */
428 	if (jailed(mp->mnt_cred) && jailed(td->td_ucred) &&
429 	    mp->mnt_cred->cr_prison != td->td_ucred->cr_prison) {
430 		return (EPERM);
431 	}
432 
433 	if ((mp->mnt_flag & MNT_USER) == 0 ||
434 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
435 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
436 			return (error);
437 	}
438 	return (0);
439 }
440 
441 /*
442  * Get a new unique fsid.  Try to make its val[0] unique, since this value
443  * will be used to create fake device numbers for stat().  Also try (but
444  * not so hard) make its val[0] unique mod 2^16, since some emulators only
445  * support 16-bit device numbers.  We end up with unique val[0]'s for the
446  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
447  *
448  * Keep in mind that several mounts may be running in parallel.  Starting
449  * the search one past where the previous search terminated is both a
450  * micro-optimization and a defense against returning the same fsid to
451  * different mounts.
452  */
453 void
454 vfs_getnewfsid(struct mount *mp)
455 {
456 	static u_int16_t mntid_base;
457 	struct mount *nmp;
458 	fsid_t tfsid;
459 	int mtype;
460 
461 	mtx_lock(&mntid_mtx);
462 	mtype = mp->mnt_vfc->vfc_typenum;
463 	tfsid.val[1] = mtype;
464 	mtype = (mtype & 0xFF) << 24;
465 	for (;;) {
466 		tfsid.val[0] = makedev(255,
467 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
468 		mntid_base++;
469 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
470 			break;
471 		vfs_rel(nmp);
472 	}
473 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
474 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
475 	mtx_unlock(&mntid_mtx);
476 }
477 
478 /*
479  * Knob to control the precision of file timestamps:
480  *
481  *   0 = seconds only; nanoseconds zeroed.
482  *   1 = seconds and nanoseconds, accurate within 1/HZ.
483  *   2 = seconds and nanoseconds, truncated to microseconds.
484  * >=3 = seconds and nanoseconds, maximum precision.
485  */
486 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
487 
488 static int timestamp_precision = TSP_SEC;
489 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
490     &timestamp_precision, 0, "");
491 
492 /*
493  * Get a current timestamp.
494  */
495 void
496 vfs_timestamp(struct timespec *tsp)
497 {
498 	struct timeval tv;
499 
500 	switch (timestamp_precision) {
501 	case TSP_SEC:
502 		tsp->tv_sec = time_second;
503 		tsp->tv_nsec = 0;
504 		break;
505 	case TSP_HZ:
506 		getnanotime(tsp);
507 		break;
508 	case TSP_USEC:
509 		microtime(&tv);
510 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
511 		break;
512 	case TSP_NSEC:
513 	default:
514 		nanotime(tsp);
515 		break;
516 	}
517 }
518 
519 /*
520  * Set vnode attributes to VNOVAL
521  */
522 void
523 vattr_null(struct vattr *vap)
524 {
525 
526 	vap->va_type = VNON;
527 	vap->va_size = VNOVAL;
528 	vap->va_bytes = VNOVAL;
529 	vap->va_mode = VNOVAL;
530 	vap->va_nlink = VNOVAL;
531 	vap->va_uid = VNOVAL;
532 	vap->va_gid = VNOVAL;
533 	vap->va_fsid = VNOVAL;
534 	vap->va_fileid = VNOVAL;
535 	vap->va_blocksize = VNOVAL;
536 	vap->va_rdev = VNOVAL;
537 	vap->va_atime.tv_sec = VNOVAL;
538 	vap->va_atime.tv_nsec = VNOVAL;
539 	vap->va_mtime.tv_sec = VNOVAL;
540 	vap->va_mtime.tv_nsec = VNOVAL;
541 	vap->va_ctime.tv_sec = VNOVAL;
542 	vap->va_ctime.tv_nsec = VNOVAL;
543 	vap->va_birthtime.tv_sec = VNOVAL;
544 	vap->va_birthtime.tv_nsec = VNOVAL;
545 	vap->va_flags = VNOVAL;
546 	vap->va_gen = VNOVAL;
547 	vap->va_vaflags = 0;
548 }
549 
550 /*
551  * This routine is called when we have too many vnodes.  It attempts
552  * to free <count> vnodes and will potentially free vnodes that still
553  * have VM backing store (VM backing store is typically the cause
554  * of a vnode blowout so we want to do this).  Therefore, this operation
555  * is not considered cheap.
556  *
557  * A number of conditions may prevent a vnode from being reclaimed.
558  * the buffer cache may have references on the vnode, a directory
559  * vnode may still have references due to the namei cache representing
560  * underlying files, or the vnode may be in active use.   It is not
561  * desireable to reuse such vnodes.  These conditions may cause the
562  * number of vnodes to reach some minimum value regardless of what
563  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
564  */
565 static int
566 vlrureclaim(struct mount *mp)
567 {
568 	struct thread *td;
569 	struct vnode *vp;
570 	int done;
571 	int trigger;
572 	int usevnodes;
573 	int count;
574 
575 	/*
576 	 * Calculate the trigger point, don't allow user
577 	 * screwups to blow us up.   This prevents us from
578 	 * recycling vnodes with lots of resident pages.  We
579 	 * aren't trying to free memory, we are trying to
580 	 * free vnodes.
581 	 */
582 	usevnodes = desiredvnodes;
583 	if (usevnodes <= 0)
584 		usevnodes = 1;
585 	trigger = cnt.v_page_count * 2 / usevnodes;
586 	done = 0;
587 	td = curthread;
588 	vn_start_write(NULL, &mp, V_WAIT);
589 	MNT_ILOCK(mp);
590 	count = mp->mnt_nvnodelistsize / 10 + 1;
591 	while (count != 0) {
592 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
593 		while (vp != NULL && vp->v_type == VMARKER)
594 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
595 		if (vp == NULL)
596 			break;
597 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
598 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
599 		--count;
600 		if (!VI_TRYLOCK(vp))
601 			goto next_iter;
602 		/*
603 		 * If it's been deconstructed already, it's still
604 		 * referenced, or it exceeds the trigger, skip it.
605 		 */
606 		if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) ||
607 		    (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
608 		    vp->v_object->resident_page_count > trigger)) {
609 			VI_UNLOCK(vp);
610 			goto next_iter;
611 		}
612 		MNT_IUNLOCK(mp);
613 		vholdl(vp);
614 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) {
615 			vdrop(vp);
616 			goto next_iter_mntunlocked;
617 		}
618 		VI_LOCK(vp);
619 		/*
620 		 * v_usecount may have been bumped after VOP_LOCK() dropped
621 		 * the vnode interlock and before it was locked again.
622 		 *
623 		 * It is not necessary to recheck VI_DOOMED because it can
624 		 * only be set by another thread that holds both the vnode
625 		 * lock and vnode interlock.  If another thread has the
626 		 * vnode lock before we get to VOP_LOCK() and obtains the
627 		 * vnode interlock after VOP_LOCK() drops the vnode
628 		 * interlock, the other thread will be unable to drop the
629 		 * vnode lock before our VOP_LOCK() call fails.
630 		 */
631 		if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) ||
632 		    (vp->v_object != NULL &&
633 		    vp->v_object->resident_page_count > trigger)) {
634 			VOP_UNLOCK(vp, LK_INTERLOCK);
635 			goto next_iter_mntunlocked;
636 		}
637 		KASSERT((vp->v_iflag & VI_DOOMED) == 0,
638 		    ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
639 		vgonel(vp);
640 		VOP_UNLOCK(vp, 0);
641 		vdropl(vp);
642 		done++;
643 next_iter_mntunlocked:
644 		if ((count % 256) != 0)
645 			goto relock_mnt;
646 		goto yield;
647 next_iter:
648 		if ((count % 256) != 0)
649 			continue;
650 		MNT_IUNLOCK(mp);
651 yield:
652 		uio_yield();
653 relock_mnt:
654 		MNT_ILOCK(mp);
655 	}
656 	MNT_IUNLOCK(mp);
657 	vn_finished_write(mp);
658 	return done;
659 }
660 
661 /*
662  * Attempt to keep the free list at wantfreevnodes length.
663  */
664 static void
665 vnlru_free(int count)
666 {
667 	struct vnode *vp;
668 	int vfslocked;
669 
670 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
671 	for (; count > 0; count--) {
672 		vp = TAILQ_FIRST(&vnode_free_list);
673 		/*
674 		 * The list can be modified while the free_list_mtx
675 		 * has been dropped and vp could be NULL here.
676 		 */
677 		if (!vp)
678 			break;
679 		VNASSERT(vp->v_op != NULL, vp,
680 		    ("vnlru_free: vnode already reclaimed."));
681 		TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
682 		/*
683 		 * Don't recycle if we can't get the interlock.
684 		 */
685 		if (!VI_TRYLOCK(vp)) {
686 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
687 			continue;
688 		}
689 		VNASSERT(VCANRECYCLE(vp), vp,
690 		    ("vp inconsistent on freelist"));
691 		freevnodes--;
692 		vp->v_iflag &= ~VI_FREE;
693 		vholdl(vp);
694 		mtx_unlock(&vnode_free_list_mtx);
695 		VI_UNLOCK(vp);
696 		vfslocked = VFS_LOCK_GIANT(vp->v_mount);
697 		vtryrecycle(vp);
698 		VFS_UNLOCK_GIANT(vfslocked);
699 		/*
700 		 * If the recycled succeeded this vdrop will actually free
701 		 * the vnode.  If not it will simply place it back on
702 		 * the free list.
703 		 */
704 		vdrop(vp);
705 		mtx_lock(&vnode_free_list_mtx);
706 	}
707 }
708 /*
709  * Attempt to recycle vnodes in a context that is always safe to block.
710  * Calling vlrurecycle() from the bowels of filesystem code has some
711  * interesting deadlock problems.
712  */
713 static struct proc *vnlruproc;
714 static int vnlruproc_sig;
715 
716 static void
717 vnlru_proc(void)
718 {
719 	struct mount *mp, *nmp;
720 	int done;
721 	struct proc *p = vnlruproc;
722 	struct thread *td = curthread;
723 
724 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p,
725 	    SHUTDOWN_PRI_FIRST);
726 
727 	mtx_lock(&Giant);
728 
729 	for (;;) {
730 		kproc_suspend_check(p);
731 		mtx_lock(&vnode_free_list_mtx);
732 		if (freevnodes > wantfreevnodes)
733 			vnlru_free(freevnodes - wantfreevnodes);
734 		if (numvnodes <= desiredvnodes * 9 / 10) {
735 			vnlruproc_sig = 0;
736 			wakeup(&vnlruproc_sig);
737 			msleep(vnlruproc, &vnode_free_list_mtx,
738 			    PVFS|PDROP, "vlruwt", hz);
739 			continue;
740 		}
741 		mtx_unlock(&vnode_free_list_mtx);
742 		done = 0;
743 		mtx_lock(&mountlist_mtx);
744 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
745 			int vfsunlocked;
746 			if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) {
747 				nmp = TAILQ_NEXT(mp, mnt_list);
748 				continue;
749 			}
750 			if (!VFS_NEEDSGIANT(mp)) {
751 				mtx_unlock(&Giant);
752 				vfsunlocked = 1;
753 			} else
754 				vfsunlocked = 0;
755 			done += vlrureclaim(mp);
756 			if (vfsunlocked)
757 				mtx_lock(&Giant);
758 			mtx_lock(&mountlist_mtx);
759 			nmp = TAILQ_NEXT(mp, mnt_list);
760 			vfs_unbusy(mp, td);
761 		}
762 		mtx_unlock(&mountlist_mtx);
763 		if (done == 0) {
764 			EVENTHANDLER_INVOKE(vfs_lowvnodes, desiredvnodes / 10);
765 #if 0
766 			/* These messages are temporary debugging aids */
767 			if (vnlru_nowhere < 5)
768 				printf("vnlru process getting nowhere..\n");
769 			else if (vnlru_nowhere == 5)
770 				printf("vnlru process messages stopped.\n");
771 #endif
772 			vnlru_nowhere++;
773 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
774 		} else
775 			uio_yield();
776 	}
777 }
778 
779 static struct kproc_desc vnlru_kp = {
780 	"vnlru",
781 	vnlru_proc,
782 	&vnlruproc
783 };
784 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
785     &vnlru_kp);
786 
787 /*
788  * Routines having to do with the management of the vnode table.
789  */
790 
791 static void
792 vdestroy(struct vnode *vp)
793 {
794 	struct bufobj *bo;
795 
796 	CTR1(KTR_VFS, "vdestroy vp %p", vp);
797 	mtx_lock(&vnode_free_list_mtx);
798 	numvnodes--;
799 	mtx_unlock(&vnode_free_list_mtx);
800 	bo = &vp->v_bufobj;
801 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
802 	    ("cleaned vnode still on the free list."));
803 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
804 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
805 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
806 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
807 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
808 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
809 	VNASSERT(bo->bo_clean.bv_root == NULL, vp, ("cleanblkroot not NULL"));
810 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
811 	VNASSERT(bo->bo_dirty.bv_root == NULL, vp, ("dirtyblkroot not NULL"));
812 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
813 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
814 	VI_UNLOCK(vp);
815 #ifdef MAC
816 	mac_vnode_destroy(vp);
817 #endif
818 	if (vp->v_pollinfo != NULL) {
819 		knlist_destroy(&vp->v_pollinfo->vpi_selinfo.si_note);
820 		mtx_destroy(&vp->v_pollinfo->vpi_lock);
821 		uma_zfree(vnodepoll_zone, vp->v_pollinfo);
822 	}
823 #ifdef INVARIANTS
824 	/* XXX Elsewhere we can detect an already freed vnode via NULL v_op. */
825 	vp->v_op = NULL;
826 #endif
827 	lockdestroy(vp->v_vnlock);
828 	mtx_destroy(&vp->v_interlock);
829 	mtx_destroy(BO_MTX(bo));
830 	uma_zfree(vnode_zone, vp);
831 }
832 
833 /*
834  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
835  * before we actually vgone().  This function must be called with the vnode
836  * held to prevent the vnode from being returned to the free list midway
837  * through vgone().
838  */
839 static int
840 vtryrecycle(struct vnode *vp)
841 {
842 	struct mount *vnmp;
843 
844 	CTR1(KTR_VFS, "vtryrecycle: trying vp %p", vp);
845 	VNASSERT(vp->v_holdcnt, vp,
846 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
847 	/*
848 	 * This vnode may found and locked via some other list, if so we
849 	 * can't recycle it yet.
850 	 */
851 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0)
852 		return (EWOULDBLOCK);
853 	/*
854 	 * Don't recycle if its filesystem is being suspended.
855 	 */
856 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
857 		VOP_UNLOCK(vp, 0);
858 		return (EBUSY);
859 	}
860 	/*
861 	 * If we got this far, we need to acquire the interlock and see if
862 	 * anyone picked up this vnode from another list.  If not, we will
863 	 * mark it with DOOMED via vgonel() so that anyone who does find it
864 	 * will skip over it.
865 	 */
866 	VI_LOCK(vp);
867 	if (vp->v_usecount) {
868 		VOP_UNLOCK(vp, LK_INTERLOCK);
869 		vn_finished_write(vnmp);
870 		return (EBUSY);
871 	}
872 	if ((vp->v_iflag & VI_DOOMED) == 0)
873 		vgonel(vp);
874 	VOP_UNLOCK(vp, LK_INTERLOCK);
875 	vn_finished_write(vnmp);
876 	CTR1(KTR_VFS, "vtryrecycle: recycled vp %p", vp);
877 	return (0);
878 }
879 
880 /*
881  * Return the next vnode from the free list.
882  */
883 int
884 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
885     struct vnode **vpp)
886 {
887 	struct vnode *vp = NULL;
888 	struct bufobj *bo;
889 
890 	mtx_lock(&vnode_free_list_mtx);
891 	/*
892 	 * Lend our context to reclaim vnodes if they've exceeded the max.
893 	 */
894 	if (freevnodes > wantfreevnodes)
895 		vnlru_free(1);
896 	/*
897 	 * Wait for available vnodes.
898 	 */
899 	if (numvnodes > desiredvnodes) {
900 		if (mp != NULL && (mp->mnt_kern_flag & MNTK_SUSPEND)) {
901 			/*
902 			 * File system is beeing suspended, we cannot risk a
903 			 * deadlock here, so allocate new vnode anyway.
904 			 */
905 			if (freevnodes > wantfreevnodes)
906 				vnlru_free(freevnodes - wantfreevnodes);
907 			goto alloc;
908 		}
909 		if (vnlruproc_sig == 0) {
910 			vnlruproc_sig = 1;	/* avoid unnecessary wakeups */
911 			wakeup(vnlruproc);
912 		}
913 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
914 		    "vlruwk", hz);
915 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
916 		if (numvnodes > desiredvnodes) {
917 			mtx_unlock(&vnode_free_list_mtx);
918 			return (ENFILE);
919 		}
920 #endif
921 	}
922 alloc:
923 	numvnodes++;
924 	mtx_unlock(&vnode_free_list_mtx);
925 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO);
926 	/*
927 	 * Setup locks.
928 	 */
929 	vp->v_vnlock = &vp->v_lock;
930 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
931 	/*
932 	 * By default, don't allow shared locks unless filesystems
933 	 * opt-in.
934 	 */
935 	lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOSHARE);
936 	/*
937 	 * Initialize bufobj.
938 	 */
939 	bo = &vp->v_bufobj;
940 	bo->__bo_vnode = vp;
941 	mtx_init(BO_MTX(bo), "bufobj interlock", NULL, MTX_DEF);
942 	bo->bo_ops = &buf_ops_bio;
943 	bo->bo_private = vp;
944 	TAILQ_INIT(&bo->bo_clean.bv_hd);
945 	TAILQ_INIT(&bo->bo_dirty.bv_hd);
946 	/*
947 	 * Initialize namecache.
948 	 */
949 	LIST_INIT(&vp->v_cache_src);
950 	TAILQ_INIT(&vp->v_cache_dst);
951 	/*
952 	 * Finalize various vnode identity bits.
953 	 */
954 	vp->v_type = VNON;
955 	vp->v_tag = tag;
956 	vp->v_op = vops;
957 	v_incr_usecount(vp);
958 	vp->v_data = 0;
959 #ifdef MAC
960 	mac_vnode_init(vp);
961 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
962 		mac_vnode_associate_singlelabel(mp, vp);
963 	else if (mp == NULL)
964 		printf("NULL mp in getnewvnode()\n");
965 #endif
966 	if (mp != NULL) {
967 		bo->bo_bsize = mp->mnt_stat.f_iosize;
968 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
969 			vp->v_vflag |= VV_NOKNOTE;
970 	}
971 
972 	CTR2(KTR_VFS, "getnewvnode: mp %p vp %p", mp, vp);
973 	*vpp = vp;
974 	return (0);
975 }
976 
977 /*
978  * Delete from old mount point vnode list, if on one.
979  */
980 static void
981 delmntque(struct vnode *vp)
982 {
983 	struct mount *mp;
984 
985 	mp = vp->v_mount;
986 	if (mp == NULL)
987 		return;
988 	MNT_ILOCK(mp);
989 	vp->v_mount = NULL;
990 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
991 		("bad mount point vnode list size"));
992 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
993 	mp->mnt_nvnodelistsize--;
994 	MNT_REL(mp);
995 	MNT_IUNLOCK(mp);
996 }
997 
998 static void
999 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1000 {
1001 
1002 	vp->v_data = NULL;
1003 	vp->v_op = &dead_vnodeops;
1004 	/* XXX non mp-safe fs may still call insmntque with vnode
1005 	   unlocked */
1006 	if (!VOP_ISLOCKED(vp))
1007 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1008 	vgone(vp);
1009 	vput(vp);
1010 }
1011 
1012 /*
1013  * Insert into list of vnodes for the new mount point, if available.
1014  */
1015 int
1016 insmntque1(struct vnode *vp, struct mount *mp,
1017 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
1018 {
1019 
1020 	KASSERT(vp->v_mount == NULL,
1021 		("insmntque: vnode already on per mount vnode list"));
1022 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1023 	MNT_ILOCK(mp);
1024 	if ((mp->mnt_kern_flag & MNTK_NOINSMNTQ) != 0 &&
1025 	    mp->mnt_nvnodelistsize == 0) {
1026 		MNT_IUNLOCK(mp);
1027 		if (dtr != NULL)
1028 			dtr(vp, dtr_arg);
1029 		return (EBUSY);
1030 	}
1031 	vp->v_mount = mp;
1032 	MNT_REF(mp);
1033 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1034 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1035 		("neg mount point vnode list size"));
1036 	mp->mnt_nvnodelistsize++;
1037 	MNT_IUNLOCK(mp);
1038 	return (0);
1039 }
1040 
1041 int
1042 insmntque(struct vnode *vp, struct mount *mp)
1043 {
1044 
1045 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1046 }
1047 
1048 /*
1049  * Flush out and invalidate all buffers associated with a bufobj
1050  * Called with the underlying object locked.
1051  */
1052 int
1053 bufobj_invalbuf(struct bufobj *bo, int flags, struct thread *td, int slpflag,
1054     int slptimeo)
1055 {
1056 	int error;
1057 
1058 	BO_LOCK(bo);
1059 	if (flags & V_SAVE) {
1060 		error = bufobj_wwait(bo, slpflag, slptimeo);
1061 		if (error) {
1062 			BO_UNLOCK(bo);
1063 			return (error);
1064 		}
1065 		if (bo->bo_dirty.bv_cnt > 0) {
1066 			BO_UNLOCK(bo);
1067 			if ((error = BO_SYNC(bo, MNT_WAIT, td)) != 0)
1068 				return (error);
1069 			/*
1070 			 * XXX We could save a lock/unlock if this was only
1071 			 * enabled under INVARIANTS
1072 			 */
1073 			BO_LOCK(bo);
1074 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1075 				panic("vinvalbuf: dirty bufs");
1076 		}
1077 	}
1078 	/*
1079 	 * If you alter this loop please notice that interlock is dropped and
1080 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1081 	 * no race conditions occur from this.
1082 	 */
1083 	do {
1084 		error = flushbuflist(&bo->bo_clean,
1085 		    flags, bo, slpflag, slptimeo);
1086 		if (error == 0)
1087 			error = flushbuflist(&bo->bo_dirty,
1088 			    flags, bo, slpflag, slptimeo);
1089 		if (error != 0 && error != EAGAIN) {
1090 			BO_UNLOCK(bo);
1091 			return (error);
1092 		}
1093 	} while (error != 0);
1094 
1095 	/*
1096 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1097 	 * have write I/O in-progress but if there is a VM object then the
1098 	 * VM object can also have read-I/O in-progress.
1099 	 */
1100 	do {
1101 		bufobj_wwait(bo, 0, 0);
1102 		BO_UNLOCK(bo);
1103 		if (bo->bo_object != NULL) {
1104 			VM_OBJECT_LOCK(bo->bo_object);
1105 			vm_object_pip_wait(bo->bo_object, "bovlbx");
1106 			VM_OBJECT_UNLOCK(bo->bo_object);
1107 		}
1108 		BO_LOCK(bo);
1109 	} while (bo->bo_numoutput > 0);
1110 	BO_UNLOCK(bo);
1111 
1112 	/*
1113 	 * Destroy the copy in the VM cache, too.
1114 	 */
1115 	if (bo->bo_object != NULL) {
1116 		VM_OBJECT_LOCK(bo->bo_object);
1117 		vm_object_page_remove(bo->bo_object, 0, 0,
1118 			(flags & V_SAVE) ? TRUE : FALSE);
1119 		VM_OBJECT_UNLOCK(bo->bo_object);
1120 	}
1121 
1122 #ifdef INVARIANTS
1123 	BO_LOCK(bo);
1124 	if ((flags & (V_ALT | V_NORMAL)) == 0 &&
1125 	    (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0))
1126 		panic("vinvalbuf: flush failed");
1127 	BO_UNLOCK(bo);
1128 #endif
1129 	return (0);
1130 }
1131 
1132 /*
1133  * Flush out and invalidate all buffers associated with a vnode.
1134  * Called with the underlying object locked.
1135  */
1136 int
1137 vinvalbuf(struct vnode *vp, int flags, struct thread *td, int slpflag,
1138     int slptimeo)
1139 {
1140 
1141 	CTR2(KTR_VFS, "vinvalbuf vp %p flags %d", vp, flags);
1142 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1143 	return (bufobj_invalbuf(&vp->v_bufobj, flags, td, slpflag, slptimeo));
1144 }
1145 
1146 /*
1147  * Flush out buffers on the specified list.
1148  *
1149  */
1150 static int
1151 flushbuflist( struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1152     int slptimeo)
1153 {
1154 	struct buf *bp, *nbp;
1155 	int retval, error;
1156 	daddr_t lblkno;
1157 	b_xflags_t xflags;
1158 
1159 	ASSERT_BO_LOCKED(bo);
1160 
1161 	retval = 0;
1162 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1163 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1164 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1165 			continue;
1166 		}
1167 		lblkno = 0;
1168 		xflags = 0;
1169 		if (nbp != NULL) {
1170 			lblkno = nbp->b_lblkno;
1171 			xflags = nbp->b_xflags &
1172 				(BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN);
1173 		}
1174 		retval = EAGAIN;
1175 		error = BUF_TIMELOCK(bp,
1176 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo),
1177 		    "flushbuf", slpflag, slptimeo);
1178 		if (error) {
1179 			BO_LOCK(bo);
1180 			return (error != ENOLCK ? error : EAGAIN);
1181 		}
1182 		KASSERT(bp->b_bufobj == bo,
1183 		    ("bp %p wrong b_bufobj %p should be %p",
1184 		    bp, bp->b_bufobj, bo));
1185 		if (bp->b_bufobj != bo) {	/* XXX: necessary ? */
1186 			BUF_UNLOCK(bp);
1187 			BO_LOCK(bo);
1188 			return (EAGAIN);
1189 		}
1190 		/*
1191 		 * XXX Since there are no node locks for NFS, I
1192 		 * believe there is a slight chance that a delayed
1193 		 * write will occur while sleeping just above, so
1194 		 * check for it.
1195 		 */
1196 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1197 		    (flags & V_SAVE)) {
1198 			bremfree(bp);
1199 			bp->b_flags |= B_ASYNC;
1200 			bwrite(bp);
1201 			BO_LOCK(bo);
1202 			return (EAGAIN);	/* XXX: why not loop ? */
1203 		}
1204 		bremfree(bp);
1205 		bp->b_flags |= (B_INVAL | B_RELBUF);
1206 		bp->b_flags &= ~B_ASYNC;
1207 		brelse(bp);
1208 		BO_LOCK(bo);
1209 		if (nbp != NULL &&
1210 		    (nbp->b_bufobj != bo ||
1211 		     nbp->b_lblkno != lblkno ||
1212 		     (nbp->b_xflags &
1213 		      (BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN)) != xflags))
1214 			break;			/* nbp invalid */
1215 	}
1216 	return (retval);
1217 }
1218 
1219 /*
1220  * Truncate a file's buffer and pages to a specified length.  This
1221  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1222  * sync activity.
1223  */
1224 int
1225 vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td,
1226     off_t length, int blksize)
1227 {
1228 	struct buf *bp, *nbp;
1229 	int anyfreed;
1230 	int trunclbn;
1231 	struct bufobj *bo;
1232 
1233 	CTR2(KTR_VFS, "vtruncbuf vp %p length %jd", vp, length);
1234 	/*
1235 	 * Round up to the *next* lbn.
1236 	 */
1237 	trunclbn = (length + blksize - 1) / blksize;
1238 
1239 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1240 restart:
1241 	bo = &vp->v_bufobj;
1242 	BO_LOCK(bo);
1243 	anyfreed = 1;
1244 	for (;anyfreed;) {
1245 		anyfreed = 0;
1246 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1247 			if (bp->b_lblkno < trunclbn)
1248 				continue;
1249 			if (BUF_LOCK(bp,
1250 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1251 			    BO_MTX(bo)) == ENOLCK)
1252 				goto restart;
1253 
1254 			bremfree(bp);
1255 			bp->b_flags |= (B_INVAL | B_RELBUF);
1256 			bp->b_flags &= ~B_ASYNC;
1257 			brelse(bp);
1258 			anyfreed = 1;
1259 
1260 			if (nbp != NULL &&
1261 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1262 			    (nbp->b_vp != vp) ||
1263 			    (nbp->b_flags & B_DELWRI))) {
1264 				goto restart;
1265 			}
1266 			BO_LOCK(bo);
1267 		}
1268 
1269 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1270 			if (bp->b_lblkno < trunclbn)
1271 				continue;
1272 			if (BUF_LOCK(bp,
1273 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1274 			    BO_MTX(bo)) == ENOLCK)
1275 				goto restart;
1276 			bremfree(bp);
1277 			bp->b_flags |= (B_INVAL | B_RELBUF);
1278 			bp->b_flags &= ~B_ASYNC;
1279 			brelse(bp);
1280 			anyfreed = 1;
1281 			if (nbp != NULL &&
1282 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1283 			    (nbp->b_vp != vp) ||
1284 			    (nbp->b_flags & B_DELWRI) == 0)) {
1285 				goto restart;
1286 			}
1287 			BO_LOCK(bo);
1288 		}
1289 	}
1290 
1291 	if (length > 0) {
1292 restartsync:
1293 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1294 			if (bp->b_lblkno > 0)
1295 				continue;
1296 			/*
1297 			 * Since we hold the vnode lock this should only
1298 			 * fail if we're racing with the buf daemon.
1299 			 */
1300 			if (BUF_LOCK(bp,
1301 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1302 			    BO_MTX(bo)) == ENOLCK) {
1303 				goto restart;
1304 			}
1305 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1306 			    ("buf(%p) on dirty queue without DELWRI", bp));
1307 
1308 			bremfree(bp);
1309 			bawrite(bp);
1310 			BO_LOCK(bo);
1311 			goto restartsync;
1312 		}
1313 	}
1314 
1315 	bufobj_wwait(bo, 0, 0);
1316 	BO_UNLOCK(bo);
1317 	vnode_pager_setsize(vp, length);
1318 
1319 	return (0);
1320 }
1321 
1322 /*
1323  * buf_splay() - splay tree core for the clean/dirty list of buffers in
1324  * 		 a vnode.
1325  *
1326  *	NOTE: We have to deal with the special case of a background bitmap
1327  *	buffer, a situation where two buffers will have the same logical
1328  *	block offset.  We want (1) only the foreground buffer to be accessed
1329  *	in a lookup and (2) must differentiate between the foreground and
1330  *	background buffer in the splay tree algorithm because the splay
1331  *	tree cannot normally handle multiple entities with the same 'index'.
1332  *	We accomplish this by adding differentiating flags to the splay tree's
1333  *	numerical domain.
1334  */
1335 static
1336 struct buf *
1337 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root)
1338 {
1339 	struct buf dummy;
1340 	struct buf *lefttreemax, *righttreemin, *y;
1341 
1342 	if (root == NULL)
1343 		return (NULL);
1344 	lefttreemax = righttreemin = &dummy;
1345 	for (;;) {
1346 		if (lblkno < root->b_lblkno ||
1347 		    (lblkno == root->b_lblkno &&
1348 		    (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1349 			if ((y = root->b_left) == NULL)
1350 				break;
1351 			if (lblkno < y->b_lblkno) {
1352 				/* Rotate right. */
1353 				root->b_left = y->b_right;
1354 				y->b_right = root;
1355 				root = y;
1356 				if ((y = root->b_left) == NULL)
1357 					break;
1358 			}
1359 			/* Link into the new root's right tree. */
1360 			righttreemin->b_left = root;
1361 			righttreemin = root;
1362 		} else if (lblkno > root->b_lblkno ||
1363 		    (lblkno == root->b_lblkno &&
1364 		    (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) {
1365 			if ((y = root->b_right) == NULL)
1366 				break;
1367 			if (lblkno > y->b_lblkno) {
1368 				/* Rotate left. */
1369 				root->b_right = y->b_left;
1370 				y->b_left = root;
1371 				root = y;
1372 				if ((y = root->b_right) == NULL)
1373 					break;
1374 			}
1375 			/* Link into the new root's left tree. */
1376 			lefttreemax->b_right = root;
1377 			lefttreemax = root;
1378 		} else {
1379 			break;
1380 		}
1381 		root = y;
1382 	}
1383 	/* Assemble the new root. */
1384 	lefttreemax->b_right = root->b_left;
1385 	righttreemin->b_left = root->b_right;
1386 	root->b_left = dummy.b_right;
1387 	root->b_right = dummy.b_left;
1388 	return (root);
1389 }
1390 
1391 static void
1392 buf_vlist_remove(struct buf *bp)
1393 {
1394 	struct buf *root;
1395 	struct bufv *bv;
1396 
1397 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1398 	ASSERT_BO_LOCKED(bp->b_bufobj);
1399 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
1400 	    (BX_VNDIRTY|BX_VNCLEAN),
1401 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
1402 	if (bp->b_xflags & BX_VNDIRTY)
1403 		bv = &bp->b_bufobj->bo_dirty;
1404 	else
1405 		bv = &bp->b_bufobj->bo_clean;
1406 	if (bp != bv->bv_root) {
1407 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1408 		KASSERT(root == bp, ("splay lookup failed in remove"));
1409 	}
1410 	if (bp->b_left == NULL) {
1411 		root = bp->b_right;
1412 	} else {
1413 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left);
1414 		root->b_right = bp->b_right;
1415 	}
1416 	bv->bv_root = root;
1417 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1418 	bv->bv_cnt--;
1419 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1420 }
1421 
1422 /*
1423  * Add the buffer to the sorted clean or dirty block list using a
1424  * splay tree algorithm.
1425  *
1426  * NOTE: xflags is passed as a constant, optimizing this inline function!
1427  */
1428 static void
1429 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1430 {
1431 	struct buf *root;
1432 	struct bufv *bv;
1433 
1434 	ASSERT_BO_LOCKED(bo);
1435 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1436 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
1437 	bp->b_xflags |= xflags;
1438 	if (xflags & BX_VNDIRTY)
1439 		bv = &bo->bo_dirty;
1440 	else
1441 		bv = &bo->bo_clean;
1442 
1443 	root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1444 	if (root == NULL) {
1445 		bp->b_left = NULL;
1446 		bp->b_right = NULL;
1447 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1448 	} else if (bp->b_lblkno < root->b_lblkno ||
1449 	    (bp->b_lblkno == root->b_lblkno &&
1450 	    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1451 		bp->b_left = root->b_left;
1452 		bp->b_right = root;
1453 		root->b_left = NULL;
1454 		TAILQ_INSERT_BEFORE(root, bp, b_bobufs);
1455 	} else {
1456 		bp->b_right = root->b_right;
1457 		bp->b_left = root;
1458 		root->b_right = NULL;
1459 		TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs);
1460 	}
1461 	bv->bv_cnt++;
1462 	bv->bv_root = bp;
1463 }
1464 
1465 /*
1466  * Lookup a buffer using the splay tree.  Note that we specifically avoid
1467  * shadow buffers used in background bitmap writes.
1468  *
1469  * This code isn't quite efficient as it could be because we are maintaining
1470  * two sorted lists and do not know which list the block resides in.
1471  *
1472  * During a "make buildworld" the desired buffer is found at one of
1473  * the roots more than 60% of the time.  Thus, checking both roots
1474  * before performing either splay eliminates unnecessary splays on the
1475  * first tree splayed.
1476  */
1477 struct buf *
1478 gbincore(struct bufobj *bo, daddr_t lblkno)
1479 {
1480 	struct buf *bp;
1481 
1482 	ASSERT_BO_LOCKED(bo);
1483 	if ((bp = bo->bo_clean.bv_root) != NULL &&
1484 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1485 		return (bp);
1486 	if ((bp = bo->bo_dirty.bv_root) != NULL &&
1487 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1488 		return (bp);
1489 	if ((bp = bo->bo_clean.bv_root) != NULL) {
1490 		bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp);
1491 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1492 			return (bp);
1493 	}
1494 	if ((bp = bo->bo_dirty.bv_root) != NULL) {
1495 		bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp);
1496 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1497 			return (bp);
1498 	}
1499 	return (NULL);
1500 }
1501 
1502 /*
1503  * Associate a buffer with a vnode.
1504  */
1505 void
1506 bgetvp(struct vnode *vp, struct buf *bp)
1507 {
1508 	struct bufobj *bo;
1509 
1510 	bo = &vp->v_bufobj;
1511 	ASSERT_BO_LOCKED(bo);
1512 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
1513 
1514 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
1515 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
1516 	    ("bgetvp: bp already attached! %p", bp));
1517 
1518 	vhold(vp);
1519 	if (VFS_NEEDSGIANT(vp->v_mount) || bo->bo_flag & BO_NEEDSGIANT)
1520 		bp->b_flags |= B_NEEDSGIANT;
1521 	bp->b_vp = vp;
1522 	bp->b_bufobj = bo;
1523 	/*
1524 	 * Insert onto list for new vnode.
1525 	 */
1526 	buf_vlist_add(bp, bo, BX_VNCLEAN);
1527 }
1528 
1529 /*
1530  * Disassociate a buffer from a vnode.
1531  */
1532 void
1533 brelvp(struct buf *bp)
1534 {
1535 	struct bufobj *bo;
1536 	struct vnode *vp;
1537 
1538 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1539 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1540 
1541 	/*
1542 	 * Delete from old vnode list, if on one.
1543 	 */
1544 	vp = bp->b_vp;		/* XXX */
1545 	bo = bp->b_bufobj;
1546 	BO_LOCK(bo);
1547 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1548 		buf_vlist_remove(bp);
1549 	else
1550 		panic("brelvp: Buffer %p not on queue.", bp);
1551 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1552 		bo->bo_flag &= ~BO_ONWORKLST;
1553 		mtx_lock(&sync_mtx);
1554 		LIST_REMOVE(bo, bo_synclist);
1555 		syncer_worklist_len--;
1556 		mtx_unlock(&sync_mtx);
1557 	}
1558 	bp->b_flags &= ~B_NEEDSGIANT;
1559 	bp->b_vp = NULL;
1560 	bp->b_bufobj = NULL;
1561 	BO_UNLOCK(bo);
1562 	vdrop(vp);
1563 }
1564 
1565 /*
1566  * Add an item to the syncer work queue.
1567  */
1568 static void
1569 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
1570 {
1571 	int slot;
1572 
1573 	ASSERT_BO_LOCKED(bo);
1574 
1575 	mtx_lock(&sync_mtx);
1576 	if (bo->bo_flag & BO_ONWORKLST)
1577 		LIST_REMOVE(bo, bo_synclist);
1578 	else {
1579 		bo->bo_flag |= BO_ONWORKLST;
1580 		syncer_worklist_len++;
1581 	}
1582 
1583 	if (delay > syncer_maxdelay - 2)
1584 		delay = syncer_maxdelay - 2;
1585 	slot = (syncer_delayno + delay) & syncer_mask;
1586 
1587 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
1588 	mtx_unlock(&sync_mtx);
1589 }
1590 
1591 static int
1592 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
1593 {
1594 	int error, len;
1595 
1596 	mtx_lock(&sync_mtx);
1597 	len = syncer_worklist_len - sync_vnode_count;
1598 	mtx_unlock(&sync_mtx);
1599 	error = SYSCTL_OUT(req, &len, sizeof(len));
1600 	return (error);
1601 }
1602 
1603 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
1604     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
1605 
1606 static struct proc *updateproc;
1607 static void sched_sync(void);
1608 static struct kproc_desc up_kp = {
1609 	"syncer",
1610 	sched_sync,
1611 	&updateproc
1612 };
1613 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
1614 
1615 static int
1616 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
1617 {
1618 	struct vnode *vp;
1619 	struct mount *mp;
1620 	int vfslocked;
1621 
1622 	vfslocked = 0;
1623 restart:
1624 	*bo = LIST_FIRST(slp);
1625 	if (*bo == NULL) {
1626 		VFS_UNLOCK_GIANT(vfslocked);
1627 		return (0);
1628 	}
1629 	vp = (*bo)->__bo_vnode;	/* XXX */
1630 	if (VFS_NEEDSGIANT(vp->v_mount)) {
1631 		if (!vfslocked) {
1632 			vfslocked = 1;
1633 			if (mtx_trylock(&Giant) == 0) {
1634 				mtx_unlock(&sync_mtx);
1635 				mtx_lock(&Giant);
1636 				mtx_lock(&sync_mtx);
1637 				goto restart;
1638 			}
1639 		}
1640 	} else {
1641 		VFS_UNLOCK_GIANT(vfslocked);
1642 		vfslocked = 0;
1643 	}
1644 	if (VOP_ISLOCKED(vp) != 0) {
1645 		VFS_UNLOCK_GIANT(vfslocked);
1646 		return (1);
1647 	}
1648 	if (VI_TRYLOCK(vp) == 0) {
1649 		VFS_UNLOCK_GIANT(vfslocked);
1650 		return (1);
1651 	}
1652 	/*
1653 	 * We use vhold in case the vnode does not
1654 	 * successfully sync.  vhold prevents the vnode from
1655 	 * going away when we unlock the sync_mtx so that
1656 	 * we can acquire the vnode interlock.
1657 	 */
1658 	vholdl(vp);
1659 	mtx_unlock(&sync_mtx);
1660 	VI_UNLOCK(vp);
1661 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1662 		vdrop(vp);
1663 		VFS_UNLOCK_GIANT(vfslocked);
1664 		mtx_lock(&sync_mtx);
1665 		return (*bo == LIST_FIRST(slp));
1666 	}
1667 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1668 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
1669 	VOP_UNLOCK(vp, 0);
1670 	vn_finished_write(mp);
1671 	BO_LOCK(*bo);
1672 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
1673 		/*
1674 		 * Put us back on the worklist.  The worklist
1675 		 * routine will remove us from our current
1676 		 * position and then add us back in at a later
1677 		 * position.
1678 		 */
1679 		vn_syncer_add_to_worklist(*bo, syncdelay);
1680 	}
1681 	BO_UNLOCK(*bo);
1682 	vdrop(vp);
1683 	VFS_UNLOCK_GIANT(vfslocked);
1684 	mtx_lock(&sync_mtx);
1685 	return (0);
1686 }
1687 
1688 /*
1689  * System filesystem synchronizer daemon.
1690  */
1691 static void
1692 sched_sync(void)
1693 {
1694 	struct synclist *next;
1695 	struct synclist *slp;
1696 	struct bufobj *bo;
1697 	long starttime;
1698 	struct thread *td = curthread;
1699 	static int dummychan;
1700 	int last_work_seen;
1701 	int net_worklist_len;
1702 	int syncer_final_iter;
1703 	int first_printf;
1704 	int error;
1705 
1706 	last_work_seen = 0;
1707 	syncer_final_iter = 0;
1708 	first_printf = 1;
1709 	syncer_state = SYNCER_RUNNING;
1710 	starttime = time_uptime;
1711 	td->td_pflags |= TDP_NORUNNINGBUF;
1712 
1713 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
1714 	    SHUTDOWN_PRI_LAST);
1715 
1716 	mtx_lock(&sync_mtx);
1717 	for (;;) {
1718 		if (syncer_state == SYNCER_FINAL_DELAY &&
1719 		    syncer_final_iter == 0) {
1720 			mtx_unlock(&sync_mtx);
1721 			kproc_suspend_check(td->td_proc);
1722 			mtx_lock(&sync_mtx);
1723 		}
1724 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
1725 		if (syncer_state != SYNCER_RUNNING &&
1726 		    starttime != time_uptime) {
1727 			if (first_printf) {
1728 				printf("\nSyncing disks, vnodes remaining...");
1729 				first_printf = 0;
1730 			}
1731 			printf("%d ", net_worklist_len);
1732 		}
1733 		starttime = time_uptime;
1734 
1735 		/*
1736 		 * Push files whose dirty time has expired.  Be careful
1737 		 * of interrupt race on slp queue.
1738 		 *
1739 		 * Skip over empty worklist slots when shutting down.
1740 		 */
1741 		do {
1742 			slp = &syncer_workitem_pending[syncer_delayno];
1743 			syncer_delayno += 1;
1744 			if (syncer_delayno == syncer_maxdelay)
1745 				syncer_delayno = 0;
1746 			next = &syncer_workitem_pending[syncer_delayno];
1747 			/*
1748 			 * If the worklist has wrapped since the
1749 			 * it was emptied of all but syncer vnodes,
1750 			 * switch to the FINAL_DELAY state and run
1751 			 * for one more second.
1752 			 */
1753 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
1754 			    net_worklist_len == 0 &&
1755 			    last_work_seen == syncer_delayno) {
1756 				syncer_state = SYNCER_FINAL_DELAY;
1757 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
1758 			}
1759 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
1760 		    syncer_worklist_len > 0);
1761 
1762 		/*
1763 		 * Keep track of the last time there was anything
1764 		 * on the worklist other than syncer vnodes.
1765 		 * Return to the SHUTTING_DOWN state if any
1766 		 * new work appears.
1767 		 */
1768 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
1769 			last_work_seen = syncer_delayno;
1770 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
1771 			syncer_state = SYNCER_SHUTTING_DOWN;
1772 		while (!LIST_EMPTY(slp)) {
1773 			error = sync_vnode(slp, &bo, td);
1774 			if (error == 1) {
1775 				LIST_REMOVE(bo, bo_synclist);
1776 				LIST_INSERT_HEAD(next, bo, bo_synclist);
1777 				continue;
1778 			}
1779 		}
1780 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
1781 			syncer_final_iter--;
1782 		/*
1783 		 * The variable rushjob allows the kernel to speed up the
1784 		 * processing of the filesystem syncer process. A rushjob
1785 		 * value of N tells the filesystem syncer to process the next
1786 		 * N seconds worth of work on its queue ASAP. Currently rushjob
1787 		 * is used by the soft update code to speed up the filesystem
1788 		 * syncer process when the incore state is getting so far
1789 		 * ahead of the disk that the kernel memory pool is being
1790 		 * threatened with exhaustion.
1791 		 */
1792 		if (rushjob > 0) {
1793 			rushjob -= 1;
1794 			continue;
1795 		}
1796 		/*
1797 		 * Just sleep for a short period of time between
1798 		 * iterations when shutting down to allow some I/O
1799 		 * to happen.
1800 		 *
1801 		 * If it has taken us less than a second to process the
1802 		 * current work, then wait. Otherwise start right over
1803 		 * again. We can still lose time if any single round
1804 		 * takes more than two seconds, but it does not really
1805 		 * matter as we are just trying to generally pace the
1806 		 * filesystem activity.
1807 		 */
1808 		if (syncer_state != SYNCER_RUNNING)
1809 			msleep(&dummychan, &sync_mtx, PPAUSE, "syncfnl",
1810 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
1811 		else if (time_uptime == starttime)
1812 			msleep(&lbolt, &sync_mtx, PPAUSE, "syncer", 0);
1813 	}
1814 }
1815 
1816 /*
1817  * Request the syncer daemon to speed up its work.
1818  * We never push it to speed up more than half of its
1819  * normal turn time, otherwise it could take over the cpu.
1820  */
1821 int
1822 speedup_syncer(void)
1823 {
1824 	struct thread *td;
1825 	int ret = 0;
1826 
1827 	td = FIRST_THREAD_IN_PROC(updateproc);
1828 	mtx_lock(&sync_mtx);
1829 	if (rushjob < syncdelay / 2) {
1830 		rushjob += 1;
1831 		stat_rush_requests += 1;
1832 		ret = 1;
1833 	}
1834 	mtx_unlock(&sync_mtx);
1835 	sleepq_remove(td, &lbolt);
1836 	return (ret);
1837 }
1838 
1839 /*
1840  * Tell the syncer to speed up its work and run though its work
1841  * list several times, then tell it to shut down.
1842  */
1843 static void
1844 syncer_shutdown(void *arg, int howto)
1845 {
1846 	struct thread *td;
1847 
1848 	if (howto & RB_NOSYNC)
1849 		return;
1850 	td = FIRST_THREAD_IN_PROC(updateproc);
1851 	mtx_lock(&sync_mtx);
1852 	syncer_state = SYNCER_SHUTTING_DOWN;
1853 	rushjob = 0;
1854 	mtx_unlock(&sync_mtx);
1855 	sleepq_remove(td, &lbolt);
1856 	kproc_shutdown(arg, howto);
1857 }
1858 
1859 /*
1860  * Reassign a buffer from one vnode to another.
1861  * Used to assign file specific control information
1862  * (indirect blocks) to the vnode to which they belong.
1863  */
1864 void
1865 reassignbuf(struct buf *bp)
1866 {
1867 	struct vnode *vp;
1868 	struct bufobj *bo;
1869 	int delay;
1870 #ifdef INVARIANTS
1871 	struct bufv *bv;
1872 #endif
1873 
1874 	vp = bp->b_vp;
1875 	bo = bp->b_bufobj;
1876 	++reassignbufcalls;
1877 
1878 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
1879 	    bp, bp->b_vp, bp->b_flags);
1880 	/*
1881 	 * B_PAGING flagged buffers cannot be reassigned because their vp
1882 	 * is not fully linked in.
1883 	 */
1884 	if (bp->b_flags & B_PAGING)
1885 		panic("cannot reassign paging buffer");
1886 
1887 	/*
1888 	 * Delete from old vnode list, if on one.
1889 	 */
1890 	BO_LOCK(bo);
1891 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1892 		buf_vlist_remove(bp);
1893 	else
1894 		panic("reassignbuf: Buffer %p not on queue.", bp);
1895 	/*
1896 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
1897 	 * of clean buffers.
1898 	 */
1899 	if (bp->b_flags & B_DELWRI) {
1900 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
1901 			switch (vp->v_type) {
1902 			case VDIR:
1903 				delay = dirdelay;
1904 				break;
1905 			case VCHR:
1906 				delay = metadelay;
1907 				break;
1908 			default:
1909 				delay = filedelay;
1910 			}
1911 			vn_syncer_add_to_worklist(bo, delay);
1912 		}
1913 		buf_vlist_add(bp, bo, BX_VNDIRTY);
1914 	} else {
1915 		buf_vlist_add(bp, bo, BX_VNCLEAN);
1916 
1917 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1918 			mtx_lock(&sync_mtx);
1919 			LIST_REMOVE(bo, bo_synclist);
1920 			syncer_worklist_len--;
1921 			mtx_unlock(&sync_mtx);
1922 			bo->bo_flag &= ~BO_ONWORKLST;
1923 		}
1924 	}
1925 #ifdef INVARIANTS
1926 	bv = &bo->bo_clean;
1927 	bp = TAILQ_FIRST(&bv->bv_hd);
1928 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1929 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1930 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
1931 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1932 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1933 	bv = &bo->bo_dirty;
1934 	bp = TAILQ_FIRST(&bv->bv_hd);
1935 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1936 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1937 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
1938 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1939 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1940 #endif
1941 	BO_UNLOCK(bo);
1942 }
1943 
1944 /*
1945  * Increment the use and hold counts on the vnode, taking care to reference
1946  * the driver's usecount if this is a chardev.  The vholdl() will remove
1947  * the vnode from the free list if it is presently free.  Requires the
1948  * vnode interlock and returns with it held.
1949  */
1950 static void
1951 v_incr_usecount(struct vnode *vp)
1952 {
1953 
1954 	CTR3(KTR_VFS, "v_incr_usecount: vp %p holdcnt %d usecount %d\n",
1955 	    vp, vp->v_holdcnt, vp->v_usecount);
1956 	vp->v_usecount++;
1957 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1958 		dev_lock();
1959 		vp->v_rdev->si_usecount++;
1960 		dev_unlock();
1961 	}
1962 	vholdl(vp);
1963 }
1964 
1965 /*
1966  * Turn a holdcnt into a use+holdcnt such that only one call to
1967  * v_decr_usecount is needed.
1968  */
1969 static void
1970 v_upgrade_usecount(struct vnode *vp)
1971 {
1972 
1973 	CTR3(KTR_VFS, "v_upgrade_usecount: vp %p holdcnt %d usecount %d\n",
1974 	    vp, vp->v_holdcnt, vp->v_usecount);
1975 	vp->v_usecount++;
1976 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1977 		dev_lock();
1978 		vp->v_rdev->si_usecount++;
1979 		dev_unlock();
1980 	}
1981 }
1982 
1983 /*
1984  * Decrement the vnode use and hold count along with the driver's usecount
1985  * if this is a chardev.  The vdropl() below releases the vnode interlock
1986  * as it may free the vnode.
1987  */
1988 static void
1989 v_decr_usecount(struct vnode *vp)
1990 {
1991 
1992 	CTR3(KTR_VFS, "v_decr_usecount: vp %p holdcnt %d usecount %d\n",
1993 	    vp, vp->v_holdcnt, vp->v_usecount);
1994 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
1995 	VNASSERT(vp->v_usecount > 0, vp,
1996 	    ("v_decr_usecount: negative usecount"));
1997 	vp->v_usecount--;
1998 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1999 		dev_lock();
2000 		vp->v_rdev->si_usecount--;
2001 		dev_unlock();
2002 	}
2003 	vdropl(vp);
2004 }
2005 
2006 /*
2007  * Decrement only the use count and driver use count.  This is intended to
2008  * be paired with a follow on vdropl() to release the remaining hold count.
2009  * In this way we may vgone() a vnode with a 0 usecount without risk of
2010  * having it end up on a free list because the hold count is kept above 0.
2011  */
2012 static void
2013 v_decr_useonly(struct vnode *vp)
2014 {
2015 
2016 	CTR3(KTR_VFS, "v_decr_useonly: vp %p holdcnt %d usecount %d\n",
2017 	    vp, vp->v_holdcnt, vp->v_usecount);
2018 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
2019 	VNASSERT(vp->v_usecount > 0, vp,
2020 	    ("v_decr_useonly: negative usecount"));
2021 	vp->v_usecount--;
2022 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2023 		dev_lock();
2024 		vp->v_rdev->si_usecount--;
2025 		dev_unlock();
2026 	}
2027 }
2028 
2029 /*
2030  * Grab a particular vnode from the free list, increment its
2031  * reference count and lock it.  VI_DOOMED is set if the vnode
2032  * is being destroyed.  Only callers who specify LK_RETRY will
2033  * see doomed vnodes.  If inactive processing was delayed in
2034  * vput try to do it here.
2035  */
2036 int
2037 vget(struct vnode *vp, int flags, struct thread *td)
2038 {
2039 	int error;
2040 
2041 	error = 0;
2042 	VFS_ASSERT_GIANT(vp->v_mount);
2043 	if ((flags & LK_INTERLOCK) == 0)
2044 		VI_LOCK(vp);
2045 	vholdl(vp);
2046 	if ((error = vn_lock(vp, flags | LK_INTERLOCK)) != 0) {
2047 		vdrop(vp);
2048 		return (error);
2049 	}
2050 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
2051 		panic("vget: vn_lock failed to return ENOENT\n");
2052 	VI_LOCK(vp);
2053 	/* Upgrade our holdcnt to a usecount. */
2054 	v_upgrade_usecount(vp);
2055 	/*
2056  	 * We don't guarantee that any particular close will
2057 	 * trigger inactive processing so just make a best effort
2058 	 * here at preventing a reference to a removed file.  If
2059 	 * we don't succeed no harm is done.
2060 	 */
2061 	if (vp->v_iflag & VI_OWEINACT) {
2062 		if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2063 		    (flags & LK_NOWAIT) == 0)
2064 			vinactive(vp, td);
2065 		vp->v_iflag &= ~VI_OWEINACT;
2066 	}
2067 	VI_UNLOCK(vp);
2068 	return (0);
2069 }
2070 
2071 /*
2072  * Increase the reference count of a vnode.
2073  */
2074 void
2075 vref(struct vnode *vp)
2076 {
2077 
2078 	VI_LOCK(vp);
2079 	v_incr_usecount(vp);
2080 	VI_UNLOCK(vp);
2081 }
2082 
2083 /*
2084  * Return reference count of a vnode.
2085  *
2086  * The results of this call are only guaranteed when some mechanism other
2087  * than the VI lock is used to stop other processes from gaining references
2088  * to the vnode.  This may be the case if the caller holds the only reference.
2089  * This is also useful when stale data is acceptable as race conditions may
2090  * be accounted for by some other means.
2091  */
2092 int
2093 vrefcnt(struct vnode *vp)
2094 {
2095 	int usecnt;
2096 
2097 	VI_LOCK(vp);
2098 	usecnt = vp->v_usecount;
2099 	VI_UNLOCK(vp);
2100 
2101 	return (usecnt);
2102 }
2103 
2104 
2105 /*
2106  * Vnode put/release.
2107  * If count drops to zero, call inactive routine and return to freelist.
2108  */
2109 void
2110 vrele(struct vnode *vp)
2111 {
2112 	struct thread *td = curthread;	/* XXX */
2113 
2114 	KASSERT(vp != NULL, ("vrele: null vp"));
2115 	VFS_ASSERT_GIANT(vp->v_mount);
2116 
2117 	VI_LOCK(vp);
2118 
2119 	/* Skip this v_writecount check if we're going to panic below. */
2120 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
2121 	    ("vrele: missed vn_close"));
2122 
2123 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2124 	    vp->v_usecount == 1)) {
2125 		v_decr_usecount(vp);
2126 		return;
2127 	}
2128 	if (vp->v_usecount != 1) {
2129 #ifdef DIAGNOSTIC
2130 		vprint("vrele: negative ref count", vp);
2131 #endif
2132 		VI_UNLOCK(vp);
2133 		panic("vrele: negative ref cnt");
2134 	}
2135 	/*
2136 	 * We want to hold the vnode until the inactive finishes to
2137 	 * prevent vgone() races.  We drop the use count here and the
2138 	 * hold count below when we're done.
2139 	 */
2140 	v_decr_useonly(vp);
2141 	/*
2142 	 * We must call VOP_INACTIVE with the node locked. Mark
2143 	 * as VI_DOINGINACT to avoid recursion.
2144 	 */
2145 	vp->v_iflag |= VI_OWEINACT;
2146 	if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK) == 0) {
2147 		VI_LOCK(vp);
2148 		if (vp->v_usecount > 0)
2149 			vp->v_iflag &= ~VI_OWEINACT;
2150 		if (vp->v_iflag & VI_OWEINACT)
2151 			vinactive(vp, td);
2152 		VOP_UNLOCK(vp, 0);
2153 	} else {
2154 		VI_LOCK(vp);
2155 		if (vp->v_usecount > 0)
2156 			vp->v_iflag &= ~VI_OWEINACT;
2157 	}
2158 	vdropl(vp);
2159 }
2160 
2161 /*
2162  * Release an already locked vnode.  This give the same effects as
2163  * unlock+vrele(), but takes less time and avoids releasing and
2164  * re-aquiring the lock (as vrele() acquires the lock internally.)
2165  */
2166 void
2167 vput(struct vnode *vp)
2168 {
2169 	struct thread *td = curthread;	/* XXX */
2170 	int error;
2171 
2172 	KASSERT(vp != NULL, ("vput: null vp"));
2173 	ASSERT_VOP_LOCKED(vp, "vput");
2174 	VFS_ASSERT_GIANT(vp->v_mount);
2175 	VI_LOCK(vp);
2176 	/* Skip this v_writecount check if we're going to panic below. */
2177 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
2178 	    ("vput: missed vn_close"));
2179 	error = 0;
2180 
2181 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2182 	    vp->v_usecount == 1)) {
2183 		VOP_UNLOCK(vp, 0);
2184 		v_decr_usecount(vp);
2185 		return;
2186 	}
2187 
2188 	if (vp->v_usecount != 1) {
2189 #ifdef DIAGNOSTIC
2190 		vprint("vput: negative ref count", vp);
2191 #endif
2192 		panic("vput: negative ref cnt");
2193 	}
2194 	/*
2195 	 * We want to hold the vnode until the inactive finishes to
2196 	 * prevent vgone() races.  We drop the use count here and the
2197 	 * hold count below when we're done.
2198 	 */
2199 	v_decr_useonly(vp);
2200 	vp->v_iflag |= VI_OWEINACT;
2201 	if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2202 		error = VOP_LOCK(vp, LK_UPGRADE|LK_INTERLOCK|LK_NOWAIT);
2203 		VI_LOCK(vp);
2204 		if (error) {
2205 			if (vp->v_usecount > 0)
2206 				vp->v_iflag &= ~VI_OWEINACT;
2207 			goto done;
2208 		}
2209 	}
2210 	if (vp->v_usecount > 0)
2211 		vp->v_iflag &= ~VI_OWEINACT;
2212 	if (vp->v_iflag & VI_OWEINACT)
2213 		vinactive(vp, td);
2214 	VOP_UNLOCK(vp, 0);
2215 done:
2216 	vdropl(vp);
2217 }
2218 
2219 /*
2220  * Somebody doesn't want the vnode recycled.
2221  */
2222 void
2223 vhold(struct vnode *vp)
2224 {
2225 
2226 	VI_LOCK(vp);
2227 	vholdl(vp);
2228 	VI_UNLOCK(vp);
2229 }
2230 
2231 void
2232 vholdl(struct vnode *vp)
2233 {
2234 
2235 	vp->v_holdcnt++;
2236 	if (VSHOULDBUSY(vp))
2237 		vbusy(vp);
2238 }
2239 
2240 /*
2241  * Note that there is one less who cares about this vnode.  vdrop() is the
2242  * opposite of vhold().
2243  */
2244 void
2245 vdrop(struct vnode *vp)
2246 {
2247 
2248 	VI_LOCK(vp);
2249 	vdropl(vp);
2250 }
2251 
2252 /*
2253  * Drop the hold count of the vnode.  If this is the last reference to
2254  * the vnode we will free it if it has been vgone'd otherwise it is
2255  * placed on the free list.
2256  */
2257 void
2258 vdropl(struct vnode *vp)
2259 {
2260 
2261 	ASSERT_VI_LOCKED(vp, "vdropl");
2262 	if (vp->v_holdcnt <= 0)
2263 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
2264 	vp->v_holdcnt--;
2265 	if (vp->v_holdcnt == 0) {
2266 		if (vp->v_iflag & VI_DOOMED) {
2267 			vdestroy(vp);
2268 			return;
2269 		} else
2270 			vfree(vp);
2271 	}
2272 	VI_UNLOCK(vp);
2273 }
2274 
2275 /*
2276  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
2277  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
2278  * OWEINACT tracks whether a vnode missed a call to inactive due to a
2279  * failed lock upgrade.
2280  */
2281 static void
2282 vinactive(struct vnode *vp, struct thread *td)
2283 {
2284 
2285 	ASSERT_VOP_LOCKED(vp, "vinactive");
2286 	ASSERT_VI_LOCKED(vp, "vinactive");
2287 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2288 	    ("vinactive: recursed on VI_DOINGINACT"));
2289 	vp->v_iflag |= VI_DOINGINACT;
2290 	vp->v_iflag &= ~VI_OWEINACT;
2291 	VI_UNLOCK(vp);
2292 	VOP_INACTIVE(vp, td);
2293 	VI_LOCK(vp);
2294 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2295 	    ("vinactive: lost VI_DOINGINACT"));
2296 	vp->v_iflag &= ~VI_DOINGINACT;
2297 }
2298 
2299 /*
2300  * Remove any vnodes in the vnode table belonging to mount point mp.
2301  *
2302  * If FORCECLOSE is not specified, there should not be any active ones,
2303  * return error if any are found (nb: this is a user error, not a
2304  * system error). If FORCECLOSE is specified, detach any active vnodes
2305  * that are found.
2306  *
2307  * If WRITECLOSE is set, only flush out regular file vnodes open for
2308  * writing.
2309  *
2310  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2311  *
2312  * `rootrefs' specifies the base reference count for the root vnode
2313  * of this filesystem. The root vnode is considered busy if its
2314  * v_usecount exceeds this value. On a successful return, vflush(, td)
2315  * will call vrele() on the root vnode exactly rootrefs times.
2316  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2317  * be zero.
2318  */
2319 #ifdef DIAGNOSTIC
2320 static int busyprt = 0;		/* print out busy vnodes */
2321 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
2322 #endif
2323 
2324 int
2325 vflush( struct mount *mp, int rootrefs, int flags, struct thread *td)
2326 {
2327 	struct vnode *vp, *mvp, *rootvp = NULL;
2328 	struct vattr vattr;
2329 	int busy = 0, error;
2330 
2331 	CTR1(KTR_VFS, "vflush: mp %p", mp);
2332 	if (rootrefs > 0) {
2333 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2334 		    ("vflush: bad args"));
2335 		/*
2336 		 * Get the filesystem root vnode. We can vput() it
2337 		 * immediately, since with rootrefs > 0, it won't go away.
2338 		 */
2339 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp, td)) != 0)
2340 			return (error);
2341 		vput(rootvp);
2342 
2343 	}
2344 	MNT_ILOCK(mp);
2345 loop:
2346 	MNT_VNODE_FOREACH(vp, mp, mvp) {
2347 
2348 		VI_LOCK(vp);
2349 		vholdl(vp);
2350 		MNT_IUNLOCK(mp);
2351 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
2352 		if (error) {
2353 			vdrop(vp);
2354 			MNT_ILOCK(mp);
2355 			MNT_VNODE_FOREACH_ABORT_ILOCKED(mp, mvp);
2356 			goto loop;
2357 		}
2358 		/*
2359 		 * Skip over a vnodes marked VV_SYSTEM.
2360 		 */
2361 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2362 			VOP_UNLOCK(vp, 0);
2363 			vdrop(vp);
2364 			MNT_ILOCK(mp);
2365 			continue;
2366 		}
2367 		/*
2368 		 * If WRITECLOSE is set, flush out unlinked but still open
2369 		 * files (even if open only for reading) and regular file
2370 		 * vnodes open for writing.
2371 		 */
2372 		if (flags & WRITECLOSE) {
2373 			error = VOP_GETATTR(vp, &vattr, td->td_ucred, td);
2374 			VI_LOCK(vp);
2375 
2376 			if ((vp->v_type == VNON ||
2377 			    (error == 0 && vattr.va_nlink > 0)) &&
2378 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2379 				VOP_UNLOCK(vp, 0);
2380 				vdropl(vp);
2381 				MNT_ILOCK(mp);
2382 				continue;
2383 			}
2384 		} else
2385 			VI_LOCK(vp);
2386 		/*
2387 		 * With v_usecount == 0, all we need to do is clear out the
2388 		 * vnode data structures and we are done.
2389 		 *
2390 		 * If FORCECLOSE is set, forcibly close the vnode.
2391 		 */
2392 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
2393 			VNASSERT(vp->v_usecount == 0 ||
2394 			    (vp->v_type != VCHR && vp->v_type != VBLK), vp,
2395 			    ("device VNODE %p is FORCECLOSED", vp));
2396 			vgonel(vp);
2397 		} else {
2398 			busy++;
2399 #ifdef DIAGNOSTIC
2400 			if (busyprt)
2401 				vprint("vflush: busy vnode", vp);
2402 #endif
2403 		}
2404 		VOP_UNLOCK(vp, 0);
2405 		vdropl(vp);
2406 		MNT_ILOCK(mp);
2407 	}
2408 	MNT_IUNLOCK(mp);
2409 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2410 		/*
2411 		 * If just the root vnode is busy, and if its refcount
2412 		 * is equal to `rootrefs', then go ahead and kill it.
2413 		 */
2414 		VI_LOCK(rootvp);
2415 		KASSERT(busy > 0, ("vflush: not busy"));
2416 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
2417 		    ("vflush: usecount %d < rootrefs %d",
2418 		     rootvp->v_usecount, rootrefs));
2419 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2420 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
2421 			vgone(rootvp);
2422 			VOP_UNLOCK(rootvp, 0);
2423 			busy = 0;
2424 		} else
2425 			VI_UNLOCK(rootvp);
2426 	}
2427 	if (busy)
2428 		return (EBUSY);
2429 	for (; rootrefs > 0; rootrefs--)
2430 		vrele(rootvp);
2431 	return (0);
2432 }
2433 
2434 /*
2435  * Recycle an unused vnode to the front of the free list.
2436  */
2437 int
2438 vrecycle(struct vnode *vp, struct thread *td)
2439 {
2440 	int recycled;
2441 
2442 	ASSERT_VOP_LOCKED(vp, "vrecycle");
2443 	recycled = 0;
2444 	VI_LOCK(vp);
2445 	if (vp->v_usecount == 0) {
2446 		recycled = 1;
2447 		vgonel(vp);
2448 	}
2449 	VI_UNLOCK(vp);
2450 	return (recycled);
2451 }
2452 
2453 /*
2454  * Eliminate all activity associated with a vnode
2455  * in preparation for reuse.
2456  */
2457 void
2458 vgone(struct vnode *vp)
2459 {
2460 	VI_LOCK(vp);
2461 	vgonel(vp);
2462 	VI_UNLOCK(vp);
2463 }
2464 
2465 /*
2466  * vgone, with the vp interlock held.
2467  */
2468 void
2469 vgonel(struct vnode *vp)
2470 {
2471 	struct thread *td;
2472 	int oweinact;
2473 	int active;
2474 	struct mount *mp;
2475 
2476 	CTR1(KTR_VFS, "vgonel: vp %p", vp);
2477 	ASSERT_VOP_LOCKED(vp, "vgonel");
2478 	ASSERT_VI_LOCKED(vp, "vgonel");
2479 	VNASSERT(vp->v_holdcnt, vp,
2480 	    ("vgonel: vp %p has no reference.", vp));
2481 	td = curthread;
2482 
2483 	/*
2484 	 * Don't vgonel if we're already doomed.
2485 	 */
2486 	if (vp->v_iflag & VI_DOOMED)
2487 		return;
2488 	vp->v_iflag |= VI_DOOMED;
2489 	/*
2490 	 * Check to see if the vnode is in use.  If so, we have to call
2491 	 * VOP_CLOSE() and VOP_INACTIVE().
2492 	 */
2493 	active = vp->v_usecount;
2494 	oweinact = (vp->v_iflag & VI_OWEINACT);
2495 	VI_UNLOCK(vp);
2496 	/*
2497 	 * Clean out any buffers associated with the vnode.
2498 	 * If the flush fails, just toss the buffers.
2499 	 */
2500 	mp = NULL;
2501 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
2502 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
2503 	if (vinvalbuf(vp, V_SAVE, td, 0, 0) != 0)
2504 		vinvalbuf(vp, 0, td, 0, 0);
2505 
2506 	/*
2507 	 * If purging an active vnode, it must be closed and
2508 	 * deactivated before being reclaimed.
2509 	 */
2510 	if (active)
2511 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
2512 	if (oweinact || active) {
2513 		VI_LOCK(vp);
2514 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
2515 			vinactive(vp, td);
2516 		VI_UNLOCK(vp);
2517 	}
2518 	/*
2519 	 * Reclaim the vnode.
2520 	 */
2521 	if (VOP_RECLAIM(vp, td))
2522 		panic("vgone: cannot reclaim");
2523 	if (mp != NULL)
2524 		vn_finished_secondary_write(mp);
2525 	VNASSERT(vp->v_object == NULL, vp,
2526 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
2527 	/*
2528 	 * Clear the advisory locks and wake up waiting threads.
2529 	 */
2530 	lf_purgelocks(vp, &(vp->v_lockf));
2531 	/*
2532 	 * Delete from old mount point vnode list.
2533 	 */
2534 	delmntque(vp);
2535 	cache_purge(vp);
2536 	/*
2537 	 * Done with purge, reset to the standard lock and invalidate
2538 	 * the vnode.
2539 	 */
2540 	VI_LOCK(vp);
2541 	vp->v_vnlock = &vp->v_lock;
2542 	vp->v_op = &dead_vnodeops;
2543 	vp->v_tag = "none";
2544 	vp->v_type = VBAD;
2545 }
2546 
2547 /*
2548  * Calculate the total number of references to a special device.
2549  */
2550 int
2551 vcount(struct vnode *vp)
2552 {
2553 	int count;
2554 
2555 	dev_lock();
2556 	count = vp->v_rdev->si_usecount;
2557 	dev_unlock();
2558 	return (count);
2559 }
2560 
2561 /*
2562  * Same as above, but using the struct cdev *as argument
2563  */
2564 int
2565 count_dev(struct cdev *dev)
2566 {
2567 	int count;
2568 
2569 	dev_lock();
2570 	count = dev->si_usecount;
2571 	dev_unlock();
2572 	return(count);
2573 }
2574 
2575 /*
2576  * Print out a description of a vnode.
2577  */
2578 static char *typename[] =
2579 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
2580  "VMARKER"};
2581 
2582 void
2583 vn_printf(struct vnode *vp, const char *fmt, ...)
2584 {
2585 	va_list ap;
2586 	char buf[256], buf2[16];
2587 	u_long flags;
2588 
2589 	va_start(ap, fmt);
2590 	vprintf(fmt, ap);
2591 	va_end(ap);
2592 	printf("%p: ", (void *)vp);
2593 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
2594 	printf("    usecount %d, writecount %d, refcount %d mountedhere %p\n",
2595 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere);
2596 	buf[0] = '\0';
2597 	buf[1] = '\0';
2598 	if (vp->v_vflag & VV_ROOT)
2599 		strlcat(buf, "|VV_ROOT", sizeof(buf));
2600 	if (vp->v_vflag & VV_ISTTY)
2601 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
2602 	if (vp->v_vflag & VV_NOSYNC)
2603 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
2604 	if (vp->v_vflag & VV_CACHEDLABEL)
2605 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
2606 	if (vp->v_vflag & VV_TEXT)
2607 		strlcat(buf, "|VV_TEXT", sizeof(buf));
2608 	if (vp->v_vflag & VV_COPYONWRITE)
2609 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
2610 	if (vp->v_vflag & VV_SYSTEM)
2611 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
2612 	if (vp->v_vflag & VV_PROCDEP)
2613 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
2614 	if (vp->v_vflag & VV_NOKNOTE)
2615 		strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
2616 	if (vp->v_vflag & VV_DELETED)
2617 		strlcat(buf, "|VV_DELETED", sizeof(buf));
2618 	if (vp->v_vflag & VV_MD)
2619 		strlcat(buf, "|VV_MD", sizeof(buf));
2620 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC |
2621 	    VV_CACHEDLABEL | VV_TEXT | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
2622 	    VV_NOKNOTE | VV_DELETED | VV_MD);
2623 	if (flags != 0) {
2624 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
2625 		strlcat(buf, buf2, sizeof(buf));
2626 	}
2627 	if (vp->v_iflag & VI_MOUNT)
2628 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
2629 	if (vp->v_iflag & VI_AGE)
2630 		strlcat(buf, "|VI_AGE", sizeof(buf));
2631 	if (vp->v_iflag & VI_DOOMED)
2632 		strlcat(buf, "|VI_DOOMED", sizeof(buf));
2633 	if (vp->v_iflag & VI_FREE)
2634 		strlcat(buf, "|VI_FREE", sizeof(buf));
2635 	if (vp->v_iflag & VI_OBJDIRTY)
2636 		strlcat(buf, "|VI_OBJDIRTY", sizeof(buf));
2637 	if (vp->v_iflag & VI_DOINGINACT)
2638 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
2639 	if (vp->v_iflag & VI_OWEINACT)
2640 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
2641 	flags = vp->v_iflag & ~(VI_MOUNT | VI_AGE | VI_DOOMED | VI_FREE |
2642 	    VI_OBJDIRTY | VI_DOINGINACT | VI_OWEINACT);
2643 	if (flags != 0) {
2644 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
2645 		strlcat(buf, buf2, sizeof(buf));
2646 	}
2647 	printf("    flags (%s)\n", buf + 1);
2648 	if (mtx_owned(VI_MTX(vp)))
2649 		printf(" VI_LOCKed");
2650 	if (vp->v_object != NULL)
2651 		printf("    v_object %p ref %d pages %d\n",
2652 		    vp->v_object, vp->v_object->ref_count,
2653 		    vp->v_object->resident_page_count);
2654 	printf("    ");
2655 	lockmgr_printinfo(vp->v_vnlock);
2656 	printf("\n");
2657 	if (vp->v_data != NULL)
2658 		VOP_PRINT(vp);
2659 }
2660 
2661 #ifdef DDB
2662 /*
2663  * List all of the locked vnodes in the system.
2664  * Called when debugging the kernel.
2665  */
2666 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
2667 {
2668 	struct mount *mp, *nmp;
2669 	struct vnode *vp;
2670 
2671 	/*
2672 	 * Note: because this is DDB, we can't obey the locking semantics
2673 	 * for these structures, which means we could catch an inconsistent
2674 	 * state and dereference a nasty pointer.  Not much to be done
2675 	 * about that.
2676 	 */
2677 	db_printf("Locked vnodes\n");
2678 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
2679 		nmp = TAILQ_NEXT(mp, mnt_list);
2680 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2681 			if (vp->v_type != VMARKER &&
2682 			    VOP_ISLOCKED(vp))
2683 				vprint("", vp);
2684 		}
2685 		nmp = TAILQ_NEXT(mp, mnt_list);
2686 	}
2687 }
2688 
2689 /*
2690  * Show details about the given vnode.
2691  */
2692 DB_SHOW_COMMAND(vnode, db_show_vnode)
2693 {
2694 	struct vnode *vp;
2695 
2696 	if (!have_addr)
2697 		return;
2698 	vp = (struct vnode *)addr;
2699 	vn_printf(vp, "vnode ");
2700 }
2701 #endif	/* DDB */
2702 
2703 /*
2704  * Fill in a struct xvfsconf based on a struct vfsconf.
2705  */
2706 static void
2707 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp)
2708 {
2709 
2710 	strcpy(xvfsp->vfc_name, vfsp->vfc_name);
2711 	xvfsp->vfc_typenum = vfsp->vfc_typenum;
2712 	xvfsp->vfc_refcount = vfsp->vfc_refcount;
2713 	xvfsp->vfc_flags = vfsp->vfc_flags;
2714 	/*
2715 	 * These are unused in userland, we keep them
2716 	 * to not break binary compatibility.
2717 	 */
2718 	xvfsp->vfc_vfsops = NULL;
2719 	xvfsp->vfc_next = NULL;
2720 }
2721 
2722 /*
2723  * Top level filesystem related information gathering.
2724  */
2725 static int
2726 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
2727 {
2728 	struct vfsconf *vfsp;
2729 	struct xvfsconf xvfsp;
2730 	int error;
2731 
2732 	error = 0;
2733 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2734 		bzero(&xvfsp, sizeof(xvfsp));
2735 		vfsconf2x(vfsp, &xvfsp);
2736 		error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp);
2737 		if (error)
2738 			break;
2739 	}
2740 	return (error);
2741 }
2742 
2743 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist,
2744     "S,xvfsconf", "List of all configured filesystems");
2745 
2746 #ifndef BURN_BRIDGES
2747 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
2748 
2749 static int
2750 vfs_sysctl(SYSCTL_HANDLER_ARGS)
2751 {
2752 	int *name = (int *)arg1 - 1;	/* XXX */
2753 	u_int namelen = arg2 + 1;	/* XXX */
2754 	struct vfsconf *vfsp;
2755 	struct xvfsconf xvfsp;
2756 
2757 	printf("WARNING: userland calling deprecated sysctl, "
2758 	    "please rebuild world\n");
2759 
2760 #if 1 || defined(COMPAT_PRELITE2)
2761 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
2762 	if (namelen == 1)
2763 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
2764 #endif
2765 
2766 	switch (name[1]) {
2767 	case VFS_MAXTYPENUM:
2768 		if (namelen != 2)
2769 			return (ENOTDIR);
2770 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
2771 	case VFS_CONF:
2772 		if (namelen != 3)
2773 			return (ENOTDIR);	/* overloaded */
2774 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list)
2775 			if (vfsp->vfc_typenum == name[2])
2776 				break;
2777 		if (vfsp == NULL)
2778 			return (EOPNOTSUPP);
2779 		bzero(&xvfsp, sizeof(xvfsp));
2780 		vfsconf2x(vfsp, &xvfsp);
2781 		return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
2782 	}
2783 	return (EOPNOTSUPP);
2784 }
2785 
2786 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP,
2787 	vfs_sysctl, "Generic filesystem");
2788 
2789 #if 1 || defined(COMPAT_PRELITE2)
2790 
2791 static int
2792 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
2793 {
2794 	int error;
2795 	struct vfsconf *vfsp;
2796 	struct ovfsconf ovfs;
2797 
2798 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2799 		bzero(&ovfs, sizeof(ovfs));
2800 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
2801 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
2802 		ovfs.vfc_index = vfsp->vfc_typenum;
2803 		ovfs.vfc_refcount = vfsp->vfc_refcount;
2804 		ovfs.vfc_flags = vfsp->vfc_flags;
2805 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
2806 		if (error)
2807 			return error;
2808 	}
2809 	return 0;
2810 }
2811 
2812 #endif /* 1 || COMPAT_PRELITE2 */
2813 #endif /* !BURN_BRIDGES */
2814 
2815 #define KINFO_VNODESLOP		10
2816 #ifdef notyet
2817 /*
2818  * Dump vnode list (via sysctl).
2819  */
2820 /* ARGSUSED */
2821 static int
2822 sysctl_vnode(SYSCTL_HANDLER_ARGS)
2823 {
2824 	struct xvnode *xvn;
2825 	struct thread *td = req->td;
2826 	struct mount *mp;
2827 	struct vnode *vp;
2828 	int error, len, n;
2829 
2830 	/*
2831 	 * Stale numvnodes access is not fatal here.
2832 	 */
2833 	req->lock = 0;
2834 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
2835 	if (!req->oldptr)
2836 		/* Make an estimate */
2837 		return (SYSCTL_OUT(req, 0, len));
2838 
2839 	error = sysctl_wire_old_buffer(req, 0);
2840 	if (error != 0)
2841 		return (error);
2842 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
2843 	n = 0;
2844 	mtx_lock(&mountlist_mtx);
2845 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
2846 		if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td))
2847 			continue;
2848 		MNT_ILOCK(mp);
2849 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2850 			if (n == len)
2851 				break;
2852 			vref(vp);
2853 			xvn[n].xv_size = sizeof *xvn;
2854 			xvn[n].xv_vnode = vp;
2855 			xvn[n].xv_id = 0;	/* XXX compat */
2856 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
2857 			XV_COPY(usecount);
2858 			XV_COPY(writecount);
2859 			XV_COPY(holdcnt);
2860 			XV_COPY(mount);
2861 			XV_COPY(numoutput);
2862 			XV_COPY(type);
2863 #undef XV_COPY
2864 			xvn[n].xv_flag = vp->v_vflag;
2865 
2866 			switch (vp->v_type) {
2867 			case VREG:
2868 			case VDIR:
2869 			case VLNK:
2870 				break;
2871 			case VBLK:
2872 			case VCHR:
2873 				if (vp->v_rdev == NULL) {
2874 					vrele(vp);
2875 					continue;
2876 				}
2877 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
2878 				break;
2879 			case VSOCK:
2880 				xvn[n].xv_socket = vp->v_socket;
2881 				break;
2882 			case VFIFO:
2883 				xvn[n].xv_fifo = vp->v_fifoinfo;
2884 				break;
2885 			case VNON:
2886 			case VBAD:
2887 			default:
2888 				/* shouldn't happen? */
2889 				vrele(vp);
2890 				continue;
2891 			}
2892 			vrele(vp);
2893 			++n;
2894 		}
2895 		MNT_IUNLOCK(mp);
2896 		mtx_lock(&mountlist_mtx);
2897 		vfs_unbusy(mp, td);
2898 		if (n == len)
2899 			break;
2900 	}
2901 	mtx_unlock(&mountlist_mtx);
2902 
2903 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
2904 	free(xvn, M_TEMP);
2905 	return (error);
2906 }
2907 
2908 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD,
2909 	0, 0, sysctl_vnode, "S,xvnode", "");
2910 #endif
2911 
2912 /*
2913  * Unmount all filesystems. The list is traversed in reverse order
2914  * of mounting to avoid dependencies.
2915  */
2916 void
2917 vfs_unmountall(void)
2918 {
2919 	struct mount *mp;
2920 	struct thread *td;
2921 	int error;
2922 
2923 	KASSERT(curthread != NULL, ("vfs_unmountall: NULL curthread"));
2924 	td = curthread;
2925 	/*
2926 	 * Since this only runs when rebooting, it is not interlocked.
2927 	 */
2928 	while(!TAILQ_EMPTY(&mountlist)) {
2929 		mp = TAILQ_LAST(&mountlist, mntlist);
2930 		error = dounmount(mp, MNT_FORCE, td);
2931 		if (error) {
2932 			TAILQ_REMOVE(&mountlist, mp, mnt_list);
2933 			/*
2934 			 * XXX: Due to the way in which we mount the root
2935 			 * file system off of devfs, devfs will generate a
2936 			 * "busy" warning when we try to unmount it before
2937 			 * the root.  Don't print a warning as a result in
2938 			 * order to avoid false positive errors that may
2939 			 * cause needless upset.
2940 			 */
2941 			if (strcmp(mp->mnt_vfc->vfc_name, "devfs") != 0) {
2942 				printf("unmount of %s failed (",
2943 				    mp->mnt_stat.f_mntonname);
2944 				if (error == EBUSY)
2945 					printf("BUSY)\n");
2946 				else
2947 					printf("%d)\n", error);
2948 			}
2949 		} else {
2950 			/* The unmount has removed mp from the mountlist */
2951 		}
2952 	}
2953 }
2954 
2955 /*
2956  * perform msync on all vnodes under a mount point
2957  * the mount point must be locked.
2958  */
2959 void
2960 vfs_msync(struct mount *mp, int flags)
2961 {
2962 	struct vnode *vp, *mvp;
2963 	struct vm_object *obj;
2964 
2965 	MNT_ILOCK(mp);
2966 	MNT_VNODE_FOREACH(vp, mp, mvp) {
2967 		VI_LOCK(vp);
2968 		if ((vp->v_iflag & VI_OBJDIRTY) &&
2969 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
2970 			MNT_IUNLOCK(mp);
2971 			if (!vget(vp,
2972 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
2973 			    curthread)) {
2974 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
2975 					vput(vp);
2976 					MNT_ILOCK(mp);
2977 					continue;
2978 				}
2979 
2980 				obj = vp->v_object;
2981 				if (obj != NULL) {
2982 					VM_OBJECT_LOCK(obj);
2983 					vm_object_page_clean(obj, 0, 0,
2984 					    flags == MNT_WAIT ?
2985 					    OBJPC_SYNC : OBJPC_NOSYNC);
2986 					VM_OBJECT_UNLOCK(obj);
2987 				}
2988 				vput(vp);
2989 			}
2990 			MNT_ILOCK(mp);
2991 		} else
2992 			VI_UNLOCK(vp);
2993 	}
2994 	MNT_IUNLOCK(mp);
2995 }
2996 
2997 /*
2998  * Mark a vnode as free, putting it up for recycling.
2999  */
3000 static void
3001 vfree(struct vnode *vp)
3002 {
3003 
3004 	CTR1(KTR_VFS, "vfree vp %p", vp);
3005 	ASSERT_VI_LOCKED(vp, "vfree");
3006 	mtx_lock(&vnode_free_list_mtx);
3007 	VNASSERT(vp->v_op != NULL, vp, ("vfree: vnode already reclaimed."));
3008 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, ("vnode already free"));
3009 	VNASSERT(VSHOULDFREE(vp), vp, ("vfree: freeing when we shouldn't"));
3010 	VNASSERT((vp->v_iflag & VI_DOOMED) == 0, vp,
3011 	    ("vfree: Freeing doomed vnode"));
3012 	if (vp->v_iflag & VI_AGE) {
3013 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
3014 	} else {
3015 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
3016 	}
3017 	freevnodes++;
3018 	vp->v_iflag &= ~VI_AGE;
3019 	vp->v_iflag |= VI_FREE;
3020 	mtx_unlock(&vnode_free_list_mtx);
3021 }
3022 
3023 /*
3024  * Opposite of vfree() - mark a vnode as in use.
3025  */
3026 static void
3027 vbusy(struct vnode *vp)
3028 {
3029 	CTR1(KTR_VFS, "vbusy vp %p", vp);
3030 	ASSERT_VI_LOCKED(vp, "vbusy");
3031 	VNASSERT((vp->v_iflag & VI_FREE) != 0, vp, ("vnode not free"));
3032 	VNASSERT(vp->v_op != NULL, vp, ("vbusy: vnode already reclaimed."));
3033 
3034 	mtx_lock(&vnode_free_list_mtx);
3035 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
3036 	freevnodes--;
3037 	vp->v_iflag &= ~(VI_FREE|VI_AGE);
3038 	mtx_unlock(&vnode_free_list_mtx);
3039 }
3040 
3041 /*
3042  * Initalize per-vnode helper structure to hold poll-related state.
3043  */
3044 void
3045 v_addpollinfo(struct vnode *vp)
3046 {
3047 	struct vpollinfo *vi;
3048 
3049 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK);
3050 	if (vp->v_pollinfo != NULL) {
3051 		uma_zfree(vnodepoll_zone, vi);
3052 		return;
3053 	}
3054 	vp->v_pollinfo = vi;
3055 	mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
3056 	knlist_init(&vp->v_pollinfo->vpi_selinfo.si_note, vp, vfs_knllock,
3057 	    vfs_knlunlock, vfs_knllocked);
3058 }
3059 
3060 /*
3061  * Record a process's interest in events which might happen to
3062  * a vnode.  Because poll uses the historic select-style interface
3063  * internally, this routine serves as both the ``check for any
3064  * pending events'' and the ``record my interest in future events''
3065  * functions.  (These are done together, while the lock is held,
3066  * to avoid race conditions.)
3067  */
3068 int
3069 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
3070 {
3071 
3072 	if (vp->v_pollinfo == NULL)
3073 		v_addpollinfo(vp);
3074 	mtx_lock(&vp->v_pollinfo->vpi_lock);
3075 	if (vp->v_pollinfo->vpi_revents & events) {
3076 		/*
3077 		 * This leaves events we are not interested
3078 		 * in available for the other process which
3079 		 * which presumably had requested them
3080 		 * (otherwise they would never have been
3081 		 * recorded).
3082 		 */
3083 		events &= vp->v_pollinfo->vpi_revents;
3084 		vp->v_pollinfo->vpi_revents &= ~events;
3085 
3086 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
3087 		return events;
3088 	}
3089 	vp->v_pollinfo->vpi_events |= events;
3090 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
3091 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3092 	return 0;
3093 }
3094 
3095 /*
3096  * Routine to create and manage a filesystem syncer vnode.
3097  */
3098 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
3099 static int	sync_fsync(struct  vop_fsync_args *);
3100 static int	sync_inactive(struct  vop_inactive_args *);
3101 static int	sync_reclaim(struct  vop_reclaim_args *);
3102 
3103 static struct vop_vector sync_vnodeops = {
3104 	.vop_bypass =	VOP_EOPNOTSUPP,
3105 	.vop_close =	sync_close,		/* close */
3106 	.vop_fsync =	sync_fsync,		/* fsync */
3107 	.vop_inactive =	sync_inactive,	/* inactive */
3108 	.vop_reclaim =	sync_reclaim,	/* reclaim */
3109 	.vop_lock1 =	vop_stdlock,	/* lock */
3110 	.vop_unlock =	vop_stdunlock,	/* unlock */
3111 	.vop_islocked =	vop_stdislocked,	/* islocked */
3112 };
3113 
3114 /*
3115  * Create a new filesystem syncer vnode for the specified mount point.
3116  */
3117 int
3118 vfs_allocate_syncvnode(struct mount *mp)
3119 {
3120 	struct vnode *vp;
3121 	struct bufobj *bo;
3122 	static long start, incr, next;
3123 	int error;
3124 
3125 	/* Allocate a new vnode */
3126 	if ((error = getnewvnode("syncer", mp, &sync_vnodeops, &vp)) != 0) {
3127 		mp->mnt_syncer = NULL;
3128 		return (error);
3129 	}
3130 	vp->v_type = VNON;
3131 	error = insmntque(vp, mp);
3132 	if (error != 0)
3133 		panic("vfs_allocate_syncvnode: insmntque failed");
3134 	/*
3135 	 * Place the vnode onto the syncer worklist. We attempt to
3136 	 * scatter them about on the list so that they will go off
3137 	 * at evenly distributed times even if all the filesystems
3138 	 * are mounted at once.
3139 	 */
3140 	next += incr;
3141 	if (next == 0 || next > syncer_maxdelay) {
3142 		start /= 2;
3143 		incr /= 2;
3144 		if (start == 0) {
3145 			start = syncer_maxdelay / 2;
3146 			incr = syncer_maxdelay;
3147 		}
3148 		next = start;
3149 	}
3150 	bo = &vp->v_bufobj;
3151 	BO_LOCK(bo);
3152 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
3153 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
3154 	mtx_lock(&sync_mtx);
3155 	sync_vnode_count++;
3156 	mtx_unlock(&sync_mtx);
3157 	BO_UNLOCK(bo);
3158 	mp->mnt_syncer = vp;
3159 	return (0);
3160 }
3161 
3162 /*
3163  * Do a lazy sync of the filesystem.
3164  */
3165 static int
3166 sync_fsync(struct vop_fsync_args *ap)
3167 {
3168 	struct vnode *syncvp = ap->a_vp;
3169 	struct mount *mp = syncvp->v_mount;
3170 	struct thread *td = ap->a_td;
3171 	int error;
3172 	struct bufobj *bo;
3173 
3174 	/*
3175 	 * We only need to do something if this is a lazy evaluation.
3176 	 */
3177 	if (ap->a_waitfor != MNT_LAZY)
3178 		return (0);
3179 
3180 	/*
3181 	 * Move ourselves to the back of the sync list.
3182 	 */
3183 	bo = &syncvp->v_bufobj;
3184 	BO_LOCK(bo);
3185 	vn_syncer_add_to_worklist(bo, syncdelay);
3186 	BO_UNLOCK(bo);
3187 
3188 	/*
3189 	 * Walk the list of vnodes pushing all that are dirty and
3190 	 * not already on the sync list.
3191 	 */
3192 	mtx_lock(&mountlist_mtx);
3193 	if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) {
3194 		mtx_unlock(&mountlist_mtx);
3195 		return (0);
3196 	}
3197 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
3198 		vfs_unbusy(mp, td);
3199 		return (0);
3200 	}
3201 	MNT_ILOCK(mp);
3202 	mp->mnt_noasync++;
3203 	mp->mnt_kern_flag &= ~MNTK_ASYNC;
3204 	MNT_IUNLOCK(mp);
3205 	vfs_msync(mp, MNT_NOWAIT);
3206 	error = VFS_SYNC(mp, MNT_LAZY, td);
3207 	MNT_ILOCK(mp);
3208 	mp->mnt_noasync--;
3209 	if ((mp->mnt_flag & MNT_ASYNC) != 0 && mp->mnt_noasync == 0)
3210 		mp->mnt_kern_flag |= MNTK_ASYNC;
3211 	MNT_IUNLOCK(mp);
3212 	vn_finished_write(mp);
3213 	vfs_unbusy(mp, td);
3214 	return (error);
3215 }
3216 
3217 /*
3218  * The syncer vnode is no referenced.
3219  */
3220 static int
3221 sync_inactive(struct vop_inactive_args *ap)
3222 {
3223 
3224 	vgone(ap->a_vp);
3225 	return (0);
3226 }
3227 
3228 /*
3229  * The syncer vnode is no longer needed and is being decommissioned.
3230  *
3231  * Modifications to the worklist must be protected by sync_mtx.
3232  */
3233 static int
3234 sync_reclaim(struct vop_reclaim_args *ap)
3235 {
3236 	struct vnode *vp = ap->a_vp;
3237 	struct bufobj *bo;
3238 
3239 	bo = &vp->v_bufobj;
3240 	BO_LOCK(bo);
3241 	vp->v_mount->mnt_syncer = NULL;
3242 	if (bo->bo_flag & BO_ONWORKLST) {
3243 		mtx_lock(&sync_mtx);
3244 		LIST_REMOVE(bo, bo_synclist);
3245 		syncer_worklist_len--;
3246 		sync_vnode_count--;
3247 		mtx_unlock(&sync_mtx);
3248 		bo->bo_flag &= ~BO_ONWORKLST;
3249 	}
3250 	BO_UNLOCK(bo);
3251 
3252 	return (0);
3253 }
3254 
3255 /*
3256  * Check if vnode represents a disk device
3257  */
3258 int
3259 vn_isdisk(struct vnode *vp, int *errp)
3260 {
3261 	int error;
3262 
3263 	error = 0;
3264 	dev_lock();
3265 	if (vp->v_type != VCHR)
3266 		error = ENOTBLK;
3267 	else if (vp->v_rdev == NULL)
3268 		error = ENXIO;
3269 	else if (vp->v_rdev->si_devsw == NULL)
3270 		error = ENXIO;
3271 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
3272 		error = ENOTBLK;
3273 	dev_unlock();
3274 	if (errp != NULL)
3275 		*errp = error;
3276 	return (error == 0);
3277 }
3278 
3279 /*
3280  * Common filesystem object access control check routine.  Accepts a
3281  * vnode's type, "mode", uid and gid, requested access mode, credentials,
3282  * and optional call-by-reference privused argument allowing vaccess()
3283  * to indicate to the caller whether privilege was used to satisfy the
3284  * request (obsoleted).  Returns 0 on success, or an errno on failure.
3285  *
3286  * The ifdef'd CAPABILITIES version is here for reference, but is not
3287  * actually used.
3288  */
3289 int
3290 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
3291     mode_t acc_mode, struct ucred *cred, int *privused)
3292 {
3293 	mode_t dac_granted;
3294 	mode_t priv_granted;
3295 
3296 	/*
3297 	 * Look for a normal, non-privileged way to access the file/directory
3298 	 * as requested.  If it exists, go with that.
3299 	 */
3300 
3301 	if (privused != NULL)
3302 		*privused = 0;
3303 
3304 	dac_granted = 0;
3305 
3306 	/* Check the owner. */
3307 	if (cred->cr_uid == file_uid) {
3308 		dac_granted |= VADMIN;
3309 		if (file_mode & S_IXUSR)
3310 			dac_granted |= VEXEC;
3311 		if (file_mode & S_IRUSR)
3312 			dac_granted |= VREAD;
3313 		if (file_mode & S_IWUSR)
3314 			dac_granted |= (VWRITE | VAPPEND);
3315 
3316 		if ((acc_mode & dac_granted) == acc_mode)
3317 			return (0);
3318 
3319 		goto privcheck;
3320 	}
3321 
3322 	/* Otherwise, check the groups (first match) */
3323 	if (groupmember(file_gid, cred)) {
3324 		if (file_mode & S_IXGRP)
3325 			dac_granted |= VEXEC;
3326 		if (file_mode & S_IRGRP)
3327 			dac_granted |= VREAD;
3328 		if (file_mode & S_IWGRP)
3329 			dac_granted |= (VWRITE | VAPPEND);
3330 
3331 		if ((acc_mode & dac_granted) == acc_mode)
3332 			return (0);
3333 
3334 		goto privcheck;
3335 	}
3336 
3337 	/* Otherwise, check everyone else. */
3338 	if (file_mode & S_IXOTH)
3339 		dac_granted |= VEXEC;
3340 	if (file_mode & S_IROTH)
3341 		dac_granted |= VREAD;
3342 	if (file_mode & S_IWOTH)
3343 		dac_granted |= (VWRITE | VAPPEND);
3344 	if ((acc_mode & dac_granted) == acc_mode)
3345 		return (0);
3346 
3347 privcheck:
3348 	/*
3349 	 * Build a privilege mask to determine if the set of privileges
3350 	 * satisfies the requirements when combined with the granted mask
3351 	 * from above.  For each privilege, if the privilege is required,
3352 	 * bitwise or the request type onto the priv_granted mask.
3353 	 */
3354 	priv_granted = 0;
3355 
3356 	if (type == VDIR) {
3357 		/*
3358 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
3359 		 * requests, instead of PRIV_VFS_EXEC.
3360 		 */
3361 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3362 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP, 0))
3363 			priv_granted |= VEXEC;
3364 	} else {
3365 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3366 		    !priv_check_cred(cred, PRIV_VFS_EXEC, 0))
3367 			priv_granted |= VEXEC;
3368 	}
3369 
3370 	if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) &&
3371 	    !priv_check_cred(cred, PRIV_VFS_READ, 0))
3372 		priv_granted |= VREAD;
3373 
3374 	if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
3375 	    !priv_check_cred(cred, PRIV_VFS_WRITE, 0))
3376 		priv_granted |= (VWRITE | VAPPEND);
3377 
3378 	if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
3379 	    !priv_check_cred(cred, PRIV_VFS_ADMIN, 0))
3380 		priv_granted |= VADMIN;
3381 
3382 	if ((acc_mode & (priv_granted | dac_granted)) == acc_mode) {
3383 		/* XXX audit: privilege used */
3384 		if (privused != NULL)
3385 			*privused = 1;
3386 		return (0);
3387 	}
3388 
3389 	return ((acc_mode & VADMIN) ? EPERM : EACCES);
3390 }
3391 
3392 /*
3393  * Credential check based on process requesting service, and per-attribute
3394  * permissions.
3395  */
3396 int
3397 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
3398     struct thread *td, int access)
3399 {
3400 
3401 	/*
3402 	 * Kernel-invoked always succeeds.
3403 	 */
3404 	if (cred == NOCRED)
3405 		return (0);
3406 
3407 	/*
3408 	 * Do not allow privileged processes in jail to directly manipulate
3409 	 * system attributes.
3410 	 */
3411 	switch (attrnamespace) {
3412 	case EXTATTR_NAMESPACE_SYSTEM:
3413 		/* Potentially should be: return (EPERM); */
3414 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0));
3415 	case EXTATTR_NAMESPACE_USER:
3416 		return (VOP_ACCESS(vp, access, cred, td));
3417 	default:
3418 		return (EPERM);
3419 	}
3420 }
3421 
3422 #ifdef DEBUG_VFS_LOCKS
3423 /*
3424  * This only exists to supress warnings from unlocked specfs accesses.  It is
3425  * no longer ok to have an unlocked VFS.
3426  */
3427 #define	IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD)
3428 
3429 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
3430 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, "");
3431 
3432 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
3433 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, "");
3434 
3435 int vfs_badlock_print = 1;	/* Print lock violations. */
3436 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, "");
3437 
3438 #ifdef KDB
3439 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
3440 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, "");
3441 #endif
3442 
3443 static void
3444 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
3445 {
3446 
3447 #ifdef KDB
3448 	if (vfs_badlock_backtrace)
3449 		kdb_backtrace();
3450 #endif
3451 	if (vfs_badlock_print)
3452 		printf("%s: %p %s\n", str, (void *)vp, msg);
3453 	if (vfs_badlock_ddb)
3454 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
3455 }
3456 
3457 void
3458 assert_vi_locked(struct vnode *vp, const char *str)
3459 {
3460 
3461 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
3462 		vfs_badlock("interlock is not locked but should be", str, vp);
3463 }
3464 
3465 void
3466 assert_vi_unlocked(struct vnode *vp, const char *str)
3467 {
3468 
3469 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
3470 		vfs_badlock("interlock is locked but should not be", str, vp);
3471 }
3472 
3473 void
3474 assert_vop_locked(struct vnode *vp, const char *str)
3475 {
3476 
3477 	if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == 0)
3478 		vfs_badlock("is not locked but should be", str, vp);
3479 }
3480 
3481 void
3482 assert_vop_unlocked(struct vnode *vp, const char *str)
3483 {
3484 
3485 	if (vp && !IGNORE_LOCK(vp) &&
3486 	    VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
3487 		vfs_badlock("is locked but should not be", str, vp);
3488 }
3489 
3490 void
3491 assert_vop_elocked(struct vnode *vp, const char *str)
3492 {
3493 
3494 	if (vp && !IGNORE_LOCK(vp) &&
3495 	    VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
3496 		vfs_badlock("is not exclusive locked but should be", str, vp);
3497 }
3498 
3499 #if 0
3500 void
3501 assert_vop_elocked_other(struct vnode *vp, const char *str)
3502 {
3503 
3504 	if (vp && !IGNORE_LOCK(vp) &&
3505 	    VOP_ISLOCKED(vp) != LK_EXCLOTHER)
3506 		vfs_badlock("is not exclusive locked by another thread",
3507 		    str, vp);
3508 }
3509 
3510 void
3511 assert_vop_slocked(struct vnode *vp, const char *str)
3512 {
3513 
3514 	if (vp && !IGNORE_LOCK(vp) &&
3515 	    VOP_ISLOCKED(vp) != LK_SHARED)
3516 		vfs_badlock("is not locked shared but should be", str, vp);
3517 }
3518 #endif /* 0 */
3519 #endif /* DEBUG_VFS_LOCKS */
3520 
3521 void
3522 vop_rename_pre(void *ap)
3523 {
3524 	struct vop_rename_args *a = ap;
3525 
3526 #ifdef DEBUG_VFS_LOCKS
3527 	if (a->a_tvp)
3528 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
3529 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
3530 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
3531 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
3532 
3533 	/* Check the source (from). */
3534 	if (a->a_tdvp != a->a_fdvp && a->a_tvp != a->a_fdvp)
3535 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
3536 	if (a->a_tvp != a->a_fvp)
3537 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
3538 
3539 	/* Check the target. */
3540 	if (a->a_tvp)
3541 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
3542 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
3543 #endif
3544 	if (a->a_tdvp != a->a_fdvp)
3545 		vhold(a->a_fdvp);
3546 	if (a->a_tvp != a->a_fvp)
3547 		vhold(a->a_fvp);
3548 	vhold(a->a_tdvp);
3549 	if (a->a_tvp)
3550 		vhold(a->a_tvp);
3551 }
3552 
3553 void
3554 vop_strategy_pre(void *ap)
3555 {
3556 #ifdef DEBUG_VFS_LOCKS
3557 	struct vop_strategy_args *a;
3558 	struct buf *bp;
3559 
3560 	a = ap;
3561 	bp = a->a_bp;
3562 
3563 	/*
3564 	 * Cluster ops lock their component buffers but not the IO container.
3565 	 */
3566 	if ((bp->b_flags & B_CLUSTER) != 0)
3567 		return;
3568 
3569 	if (!BUF_ISLOCKED(bp)) {
3570 		if (vfs_badlock_print)
3571 			printf(
3572 			    "VOP_STRATEGY: bp is not locked but should be\n");
3573 		if (vfs_badlock_ddb)
3574 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
3575 	}
3576 #endif
3577 }
3578 
3579 void
3580 vop_lookup_pre(void *ap)
3581 {
3582 #ifdef DEBUG_VFS_LOCKS
3583 	struct vop_lookup_args *a;
3584 	struct vnode *dvp;
3585 
3586 	a = ap;
3587 	dvp = a->a_dvp;
3588 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3589 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3590 #endif
3591 }
3592 
3593 void
3594 vop_lookup_post(void *ap, int rc)
3595 {
3596 #ifdef DEBUG_VFS_LOCKS
3597 	struct vop_lookup_args *a;
3598 	struct vnode *dvp;
3599 	struct vnode *vp;
3600 
3601 	a = ap;
3602 	dvp = a->a_dvp;
3603 	vp = *(a->a_vpp);
3604 
3605 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3606 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3607 
3608 	if (!rc)
3609 		ASSERT_VOP_LOCKED(vp, "VOP_LOOKUP (child)");
3610 #endif
3611 }
3612 
3613 void
3614 vop_lock_pre(void *ap)
3615 {
3616 #ifdef DEBUG_VFS_LOCKS
3617 	struct vop_lock1_args *a = ap;
3618 
3619 	if ((a->a_flags & LK_INTERLOCK) == 0)
3620 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3621 	else
3622 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
3623 #endif
3624 }
3625 
3626 void
3627 vop_lock_post(void *ap, int rc)
3628 {
3629 #ifdef DEBUG_VFS_LOCKS
3630 	struct vop_lock1_args *a = ap;
3631 
3632 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3633 	if (rc == 0)
3634 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
3635 #endif
3636 }
3637 
3638 void
3639 vop_unlock_pre(void *ap)
3640 {
3641 #ifdef DEBUG_VFS_LOCKS
3642 	struct vop_unlock_args *a = ap;
3643 
3644 	if (a->a_flags & LK_INTERLOCK)
3645 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
3646 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
3647 #endif
3648 }
3649 
3650 void
3651 vop_unlock_post(void *ap, int rc)
3652 {
3653 #ifdef DEBUG_VFS_LOCKS
3654 	struct vop_unlock_args *a = ap;
3655 
3656 	if (a->a_flags & LK_INTERLOCK)
3657 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
3658 #endif
3659 }
3660 
3661 void
3662 vop_create_post(void *ap, int rc)
3663 {
3664 	struct vop_create_args *a = ap;
3665 
3666 	if (!rc)
3667 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3668 }
3669 
3670 void
3671 vop_link_post(void *ap, int rc)
3672 {
3673 	struct vop_link_args *a = ap;
3674 
3675 	if (!rc) {
3676 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
3677 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
3678 	}
3679 }
3680 
3681 void
3682 vop_mkdir_post(void *ap, int rc)
3683 {
3684 	struct vop_mkdir_args *a = ap;
3685 
3686 	if (!rc)
3687 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
3688 }
3689 
3690 void
3691 vop_mknod_post(void *ap, int rc)
3692 {
3693 	struct vop_mknod_args *a = ap;
3694 
3695 	if (!rc)
3696 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3697 }
3698 
3699 void
3700 vop_remove_post(void *ap, int rc)
3701 {
3702 	struct vop_remove_args *a = ap;
3703 
3704 	if (!rc) {
3705 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3706 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
3707 	}
3708 }
3709 
3710 void
3711 vop_rename_post(void *ap, int rc)
3712 {
3713 	struct vop_rename_args *a = ap;
3714 
3715 	if (!rc) {
3716 		VFS_KNOTE_UNLOCKED(a->a_fdvp, NOTE_WRITE);
3717 		VFS_KNOTE_UNLOCKED(a->a_tdvp, NOTE_WRITE);
3718 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
3719 		if (a->a_tvp)
3720 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
3721 	}
3722 	if (a->a_tdvp != a->a_fdvp)
3723 		vdrop(a->a_fdvp);
3724 	if (a->a_tvp != a->a_fvp)
3725 		vdrop(a->a_fvp);
3726 	vdrop(a->a_tdvp);
3727 	if (a->a_tvp)
3728 		vdrop(a->a_tvp);
3729 }
3730 
3731 void
3732 vop_rmdir_post(void *ap, int rc)
3733 {
3734 	struct vop_rmdir_args *a = ap;
3735 
3736 	if (!rc) {
3737 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
3738 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
3739 	}
3740 }
3741 
3742 void
3743 vop_setattr_post(void *ap, int rc)
3744 {
3745 	struct vop_setattr_args *a = ap;
3746 
3747 	if (!rc)
3748 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
3749 }
3750 
3751 void
3752 vop_symlink_post(void *ap, int rc)
3753 {
3754 	struct vop_symlink_args *a = ap;
3755 
3756 	if (!rc)
3757 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3758 }
3759 
3760 static struct knlist fs_knlist;
3761 
3762 static void
3763 vfs_event_init(void *arg)
3764 {
3765 	knlist_init(&fs_knlist, NULL, NULL, NULL, NULL);
3766 }
3767 /* XXX - correct order? */
3768 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
3769 
3770 void
3771 vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused)
3772 {
3773 
3774 	KNOTE_UNLOCKED(&fs_knlist, event);
3775 }
3776 
3777 static int	filt_fsattach(struct knote *kn);
3778 static void	filt_fsdetach(struct knote *kn);
3779 static int	filt_fsevent(struct knote *kn, long hint);
3780 
3781 struct filterops fs_filtops =
3782 	{ 0, filt_fsattach, filt_fsdetach, filt_fsevent };
3783 
3784 static int
3785 filt_fsattach(struct knote *kn)
3786 {
3787 
3788 	kn->kn_flags |= EV_CLEAR;
3789 	knlist_add(&fs_knlist, kn, 0);
3790 	return (0);
3791 }
3792 
3793 static void
3794 filt_fsdetach(struct knote *kn)
3795 {
3796 
3797 	knlist_remove(&fs_knlist, kn, 0);
3798 }
3799 
3800 static int
3801 filt_fsevent(struct knote *kn, long hint)
3802 {
3803 
3804 	kn->kn_fflags |= hint;
3805 	return (kn->kn_fflags != 0);
3806 }
3807 
3808 static int
3809 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
3810 {
3811 	struct vfsidctl vc;
3812 	int error;
3813 	struct mount *mp;
3814 
3815 	error = SYSCTL_IN(req, &vc, sizeof(vc));
3816 	if (error)
3817 		return (error);
3818 	if (vc.vc_vers != VFS_CTL_VERS1)
3819 		return (EINVAL);
3820 	mp = vfs_getvfs(&vc.vc_fsid);
3821 	if (mp == NULL)
3822 		return (ENOENT);
3823 	/* ensure that a specific sysctl goes to the right filesystem. */
3824 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
3825 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
3826 		vfs_rel(mp);
3827 		return (EINVAL);
3828 	}
3829 	VCTLTOREQ(&vc, req);
3830 	error = VFS_SYSCTL(mp, vc.vc_op, req);
3831 	vfs_rel(mp);
3832 	return (error);
3833 }
3834 
3835 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR, NULL, 0, sysctl_vfs_ctl, "",
3836     "Sysctl by fsid");
3837 
3838 /*
3839  * Function to initialize a va_filerev field sensibly.
3840  * XXX: Wouldn't a random number make a lot more sense ??
3841  */
3842 u_quad_t
3843 init_va_filerev(void)
3844 {
3845 	struct bintime bt;
3846 
3847 	getbinuptime(&bt);
3848 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
3849 }
3850 
3851 static int	filt_vfsread(struct knote *kn, long hint);
3852 static int	filt_vfswrite(struct knote *kn, long hint);
3853 static int	filt_vfsvnode(struct knote *kn, long hint);
3854 static void	filt_vfsdetach(struct knote *kn);
3855 static struct filterops vfsread_filtops =
3856 	{ 1, NULL, filt_vfsdetach, filt_vfsread };
3857 static struct filterops vfswrite_filtops =
3858 	{ 1, NULL, filt_vfsdetach, filt_vfswrite };
3859 static struct filterops vfsvnode_filtops =
3860 	{ 1, NULL, filt_vfsdetach, filt_vfsvnode };
3861 
3862 static void
3863 vfs_knllock(void *arg)
3864 {
3865 	struct vnode *vp = arg;
3866 
3867 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
3868 }
3869 
3870 static void
3871 vfs_knlunlock(void *arg)
3872 {
3873 	struct vnode *vp = arg;
3874 
3875 	VOP_UNLOCK(vp, 0);
3876 }
3877 
3878 static int
3879 vfs_knllocked(void *arg)
3880 {
3881 	struct vnode *vp = arg;
3882 
3883 	return (VOP_ISLOCKED(vp) == LK_EXCLUSIVE);
3884 }
3885 
3886 int
3887 vfs_kqfilter(struct vop_kqfilter_args *ap)
3888 {
3889 	struct vnode *vp = ap->a_vp;
3890 	struct knote *kn = ap->a_kn;
3891 	struct knlist *knl;
3892 
3893 	switch (kn->kn_filter) {
3894 	case EVFILT_READ:
3895 		kn->kn_fop = &vfsread_filtops;
3896 		break;
3897 	case EVFILT_WRITE:
3898 		kn->kn_fop = &vfswrite_filtops;
3899 		break;
3900 	case EVFILT_VNODE:
3901 		kn->kn_fop = &vfsvnode_filtops;
3902 		break;
3903 	default:
3904 		return (EINVAL);
3905 	}
3906 
3907 	kn->kn_hook = (caddr_t)vp;
3908 
3909 	if (vp->v_pollinfo == NULL)
3910 		v_addpollinfo(vp);
3911 	if (vp->v_pollinfo == NULL)
3912 		return (ENOMEM);
3913 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
3914 	knlist_add(knl, kn, 0);
3915 
3916 	return (0);
3917 }
3918 
3919 /*
3920  * Detach knote from vnode
3921  */
3922 static void
3923 filt_vfsdetach(struct knote *kn)
3924 {
3925 	struct vnode *vp = (struct vnode *)kn->kn_hook;
3926 
3927 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
3928 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
3929 }
3930 
3931 /*ARGSUSED*/
3932 static int
3933 filt_vfsread(struct knote *kn, long hint)
3934 {
3935 	struct vnode *vp = (struct vnode *)kn->kn_hook;
3936 	struct vattr va;
3937 
3938 	/*
3939 	 * filesystem is gone, so set the EOF flag and schedule
3940 	 * the knote for deletion.
3941 	 */
3942 	if (hint == NOTE_REVOKE) {
3943 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3944 		return (1);
3945 	}
3946 
3947 	if (VOP_GETATTR(vp, &va, curthread->td_ucred, curthread))
3948 		return (0);
3949 
3950 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
3951 	return (kn->kn_data != 0);
3952 }
3953 
3954 /*ARGSUSED*/
3955 static int
3956 filt_vfswrite(struct knote *kn, long hint)
3957 {
3958 	/*
3959 	 * filesystem is gone, so set the EOF flag and schedule
3960 	 * the knote for deletion.
3961 	 */
3962 	if (hint == NOTE_REVOKE)
3963 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3964 
3965 	kn->kn_data = 0;
3966 	return (1);
3967 }
3968 
3969 static int
3970 filt_vfsvnode(struct knote *kn, long hint)
3971 {
3972 	if (kn->kn_sfflags & hint)
3973 		kn->kn_fflags |= hint;
3974 	if (hint == NOTE_REVOKE) {
3975 		kn->kn_flags |= EV_EOF;
3976 		return (1);
3977 	}
3978 	return (kn->kn_fflags != 0);
3979 }
3980 
3981 int
3982 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
3983 {
3984 	int error;
3985 
3986 	if (dp->d_reclen > ap->a_uio->uio_resid)
3987 		return (ENAMETOOLONG);
3988 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
3989 	if (error) {
3990 		if (ap->a_ncookies != NULL) {
3991 			if (ap->a_cookies != NULL)
3992 				free(ap->a_cookies, M_TEMP);
3993 			ap->a_cookies = NULL;
3994 			*ap->a_ncookies = 0;
3995 		}
3996 		return (error);
3997 	}
3998 	if (ap->a_ncookies == NULL)
3999 		return (0);
4000 
4001 	KASSERT(ap->a_cookies,
4002 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
4003 
4004 	*ap->a_cookies = realloc(*ap->a_cookies,
4005 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
4006 	(*ap->a_cookies)[*ap->a_ncookies] = off;
4007 	return (0);
4008 }
4009 
4010 /*
4011  * Mark for update the access time of the file if the filesystem
4012  * supports VA_MARK_ATIME.  This functionality is used by execve
4013  * and mmap, so we want to avoid the synchronous I/O implied by
4014  * directly setting va_atime for the sake of efficiency.
4015  */
4016 void
4017 vfs_mark_atime(struct vnode *vp, struct thread *td)
4018 {
4019 	struct vattr atimeattr;
4020 
4021 	if ((vp->v_mount->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) {
4022 		VATTR_NULL(&atimeattr);
4023 		atimeattr.va_vaflags |= VA_MARK_ATIME;
4024 		(void)VOP_SETATTR(vp, &atimeattr, td->td_ucred, td);
4025 	}
4026 }
4027