xref: /freebsd/sys/kern/vfs_subr.c (revision 63518eccca27064285cf2e680510ba9a4c3e2231)
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/kdb.h>
59 #include <sys/kernel.h>
60 #include <sys/kthread.h>
61 #include <sys/malloc.h>
62 #include <sys/mount.h>
63 #include <sys/namei.h>
64 #include <sys/priv.h>
65 #include <sys/reboot.h>
66 #include <sys/sleepqueue.h>
67 #include <sys/stat.h>
68 #include <sys/sysctl.h>
69 #include <sys/syslog.h>
70 #include <sys/vmmeter.h>
71 #include <sys/vnode.h>
72 
73 #include <machine/stdarg.h>
74 
75 #include <security/mac/mac_framework.h>
76 
77 #include <vm/vm.h>
78 #include <vm/vm_object.h>
79 #include <vm/vm_extern.h>
80 #include <vm/pmap.h>
81 #include <vm/vm_map.h>
82 #include <vm/vm_page.h>
83 #include <vm/vm_kern.h>
84 #include <vm/uma.h>
85 
86 #ifdef DDB
87 #include <ddb/ddb.h>
88 #endif
89 
90 static MALLOC_DEFINE(M_NETADDR, "subr_export_host", "Export host address structure");
91 
92 static void	delmntque(struct vnode *vp);
93 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
94 		    int slpflag, int slptimeo);
95 static void	syncer_shutdown(void *arg, int howto);
96 static int	vtryrecycle(struct vnode *vp);
97 static void	vbusy(struct vnode *vp);
98 static void	vdropl(struct vnode *vp);
99 static void	vinactive(struct vnode *, struct thread *);
100 static void	v_incr_usecount(struct vnode *);
101 static void	v_decr_usecount(struct vnode *);
102 static void	v_decr_useonly(struct vnode *);
103 static void	v_upgrade_usecount(struct vnode *);
104 static void	vfree(struct vnode *);
105 static void	vnlru_free(int);
106 static void	vdestroy(struct vnode *);
107 static void	vgonel(struct vnode *);
108 static void	vfs_knllock(void *arg);
109 static void	vfs_knlunlock(void *arg);
110 static int	vfs_knllocked(void *arg);
111 
112 
113 /*
114  * Enable Giant pushdown based on whether or not the vm is mpsafe in this
115  * build.  Without mpsafevm the buffer cache can not run Giant free.
116  */
117 #if !defined(__powerpc__)
118 int mpsafe_vfs = 1;
119 #else
120 int mpsafe_vfs;
121 #endif
122 TUNABLE_INT("debug.mpsafevfs", &mpsafe_vfs);
123 SYSCTL_INT(_debug, OID_AUTO, mpsafevfs, CTLFLAG_RD, &mpsafe_vfs, 0,
124     "MPSAFE VFS");
125 
126 /*
127  * Number of vnodes in existence.  Increased whenever getnewvnode()
128  * allocates a new vnode, decreased on vdestroy() called on VI_DOOMed
129  * vnode.
130  */
131 static unsigned long	numvnodes;
132 
133 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, "");
134 
135 /*
136  * Conversion tables for conversion from vnode types to inode formats
137  * and back.
138  */
139 enum vtype iftovt_tab[16] = {
140 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
141 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
142 };
143 int vttoif_tab[10] = {
144 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
145 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
146 };
147 
148 /*
149  * List of vnodes that are ready for recycling.
150  */
151 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
152 
153 /*
154  * Free vnode target.  Free vnodes may simply be files which have been stat'd
155  * but not read.  This is somewhat common, and a small cache of such files
156  * should be kept to avoid recreation costs.
157  */
158 static u_long wantfreevnodes;
159 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, "");
160 /* Number of vnodes in the free list. */
161 static u_long freevnodes;
162 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, "");
163 
164 /*
165  * Various variables used for debugging the new implementation of
166  * reassignbuf().
167  * XXX these are probably of (very) limited utility now.
168  */
169 static int reassignbufcalls;
170 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, "");
171 
172 /*
173  * Cache for the mount type id assigned to NFS.  This is used for
174  * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c.
175  */
176 int	nfs_mount_type = -1;
177 
178 /* To keep more than one thread at a time from running vfs_getnewfsid */
179 static struct mtx mntid_mtx;
180 
181 /*
182  * Lock for any access to the following:
183  *	vnode_free_list
184  *	numvnodes
185  *	freevnodes
186  */
187 static struct mtx vnode_free_list_mtx;
188 
189 /* Publicly exported FS */
190 struct nfs_public nfs_pub;
191 
192 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
193 static uma_zone_t vnode_zone;
194 static uma_zone_t vnodepoll_zone;
195 
196 /* Set to 1 to print out reclaim of active vnodes */
197 int	prtactive;
198 
199 /*
200  * The workitem queue.
201  *
202  * It is useful to delay writes of file data and filesystem metadata
203  * for tens of seconds so that quickly created and deleted files need
204  * not waste disk bandwidth being created and removed. To realize this,
205  * we append vnodes to a "workitem" queue. When running with a soft
206  * updates implementation, most pending metadata dependencies should
207  * not wait for more than a few seconds. Thus, mounted on block devices
208  * are delayed only about a half the time that file data is delayed.
209  * Similarly, directory updates are more critical, so are only delayed
210  * about a third the time that file data is delayed. Thus, there are
211  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
212  * one each second (driven off the filesystem syncer process). The
213  * syncer_delayno variable indicates the next queue that is to be processed.
214  * Items that need to be processed soon are placed in this queue:
215  *
216  *	syncer_workitem_pending[syncer_delayno]
217  *
218  * A delay of fifteen seconds is done by placing the request fifteen
219  * entries later in the queue:
220  *
221  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
222  *
223  */
224 static int syncer_delayno;
225 static long syncer_mask;
226 LIST_HEAD(synclist, bufobj);
227 static struct synclist *syncer_workitem_pending;
228 /*
229  * The sync_mtx protects:
230  *	bo->bo_synclist
231  *	sync_vnode_count
232  *	syncer_delayno
233  *	syncer_state
234  *	syncer_workitem_pending
235  *	syncer_worklist_len
236  *	rushjob
237  */
238 static struct mtx sync_mtx;
239 
240 #define SYNCER_MAXDELAY		32
241 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
242 static int syncdelay = 30;		/* max time to delay syncing data */
243 static int filedelay = 30;		/* time to delay syncing files */
244 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, "");
245 static int dirdelay = 29;		/* time to delay syncing directories */
246 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, "");
247 static int metadelay = 28;		/* time to delay syncing metadata */
248 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, "");
249 static int rushjob;		/* number of slots to run ASAP */
250 static int stat_rush_requests;	/* number of times I/O speeded up */
251 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, "");
252 
253 /*
254  * When shutting down the syncer, run it at four times normal speed.
255  */
256 #define SYNCER_SHUTDOWN_SPEEDUP		4
257 static int sync_vnode_count;
258 static int syncer_worklist_len;
259 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
260     syncer_state;
261 
262 /*
263  * Number of vnodes we want to exist at any one time.  This is mostly used
264  * to size hash tables in vnode-related code.  It is normally not used in
265  * getnewvnode(), as wantfreevnodes is normally nonzero.)
266  *
267  * XXX desiredvnodes is historical cruft and should not exist.
268  */
269 int desiredvnodes;
270 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW,
271     &desiredvnodes, 0, "Maximum number of vnodes");
272 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
273     &wantfreevnodes, 0, "Minimum number of vnodes (legacy)");
274 static int vnlru_nowhere;
275 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
276     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
277 
278 /*
279  * Macros to control when a vnode is freed and recycled.  All require
280  * the vnode interlock.
281  */
282 #define VCANRECYCLE(vp) (((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt)
283 #define VSHOULDFREE(vp) (!((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt)
284 #define VSHOULDBUSY(vp) (((vp)->v_iflag & VI_FREE) && (vp)->v_holdcnt)
285 
286 
287 /*
288  * Initialize the vnode management data structures.
289  */
290 #ifndef	MAXVNODES_MAX
291 #define	MAXVNODES_MAX	100000
292 #endif
293 static void
294 vntblinit(void *dummy __unused)
295 {
296 
297 	/*
298 	 * Desiredvnodes is a function of the physical memory size and
299 	 * the kernel's heap size.  Specifically, desiredvnodes scales
300 	 * in proportion to the physical memory size until two fifths
301 	 * of the kernel's heap size is consumed by vnodes and vm
302 	 * objects.
303 	 */
304 	desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size /
305 	    (5 * (sizeof(struct vm_object) + sizeof(struct vnode))));
306 	if (desiredvnodes > MAXVNODES_MAX) {
307 		if (bootverbose)
308 			printf("Reducing kern.maxvnodes %d -> %d\n",
309 			    desiredvnodes, MAXVNODES_MAX);
310 		desiredvnodes = MAXVNODES_MAX;
311 	}
312 	wantfreevnodes = desiredvnodes / 4;
313 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
314 	TAILQ_INIT(&vnode_free_list);
315 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
316 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
317 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
318 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
319 	      NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
320 	/*
321 	 * Initialize the filesystem syncer.
322 	 */
323 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
324 		&syncer_mask);
325 	syncer_maxdelay = syncer_mask + 1;
326 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
327 }
328 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL)
329 
330 
331 /*
332  * Mark a mount point as busy. Used to synchronize access and to delay
333  * unmounting. Interlock is not released on failure.
334  */
335 int
336 vfs_busy(struct mount *mp, int flags, struct mtx *interlkp,
337     struct thread *td)
338 {
339 	int lkflags;
340 
341 	MNT_ILOCK(mp);
342 	MNT_REF(mp);
343 	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
344 		if (flags & LK_NOWAIT) {
345 			MNT_REL(mp);
346 			MNT_IUNLOCK(mp);
347 			return (ENOENT);
348 		}
349 		if (interlkp)
350 			mtx_unlock(interlkp);
351 		mp->mnt_kern_flag |= MNTK_MWAIT;
352 		/*
353 		 * Since all busy locks are shared except the exclusive
354 		 * lock granted when unmounting, the only place that a
355 		 * wakeup needs to be done is at the release of the
356 		 * exclusive lock at the end of dounmount.
357 		 */
358 		msleep(mp, MNT_MTX(mp), PVFS, "vfs_busy", 0);
359 		MNT_REL(mp);
360 		MNT_IUNLOCK(mp);
361 		if (interlkp)
362 			mtx_lock(interlkp);
363 		return (ENOENT);
364 	}
365 	if (interlkp)
366 		mtx_unlock(interlkp);
367 	lkflags = LK_SHARED | LK_INTERLOCK;
368 	if (lockmgr(&mp->mnt_lock, lkflags, MNT_MTX(mp), td))
369 		panic("vfs_busy: unexpected lock failure");
370 	return (0);
371 }
372 
373 /*
374  * Free a busy filesystem.
375  */
376 void
377 vfs_unbusy(struct mount *mp, struct thread *td)
378 {
379 
380 	lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td);
381 	vfs_rel(mp);
382 }
383 
384 /*
385  * Lookup a mount point by filesystem identifier.
386  */
387 struct mount *
388 vfs_getvfs(fsid_t *fsid)
389 {
390 	struct mount *mp;
391 
392 	mtx_lock(&mountlist_mtx);
393 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
394 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
395 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
396 			vfs_ref(mp);
397 			mtx_unlock(&mountlist_mtx);
398 			return (mp);
399 		}
400 	}
401 	mtx_unlock(&mountlist_mtx);
402 	return ((struct mount *) 0);
403 }
404 
405 /*
406  * Check if a user can access priveledged mount options.
407  */
408 int
409 vfs_suser(struct mount *mp, struct thread *td)
410 {
411 	int error;
412 
413 	if ((mp->mnt_flag & MNT_USER) == 0 ||
414 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
415 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
416 			return (error);
417 	}
418 	return (0);
419 }
420 
421 /*
422  * Get a new unique fsid.  Try to make its val[0] unique, since this value
423  * will be used to create fake device numbers for stat().  Also try (but
424  * not so hard) make its val[0] unique mod 2^16, since some emulators only
425  * support 16-bit device numbers.  We end up with unique val[0]'s for the
426  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
427  *
428  * Keep in mind that several mounts may be running in parallel.  Starting
429  * the search one past where the previous search terminated is both a
430  * micro-optimization and a defense against returning the same fsid to
431  * different mounts.
432  */
433 void
434 vfs_getnewfsid(struct mount *mp)
435 {
436 	static u_int16_t mntid_base;
437 	struct mount *nmp;
438 	fsid_t tfsid;
439 	int mtype;
440 
441 	mtx_lock(&mntid_mtx);
442 	mtype = mp->mnt_vfc->vfc_typenum;
443 	tfsid.val[1] = mtype;
444 	mtype = (mtype & 0xFF) << 24;
445 	for (;;) {
446 		tfsid.val[0] = makedev(255,
447 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
448 		mntid_base++;
449 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
450 			break;
451 		vfs_rel(nmp);
452 	}
453 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
454 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
455 	mtx_unlock(&mntid_mtx);
456 }
457 
458 /*
459  * Knob to control the precision of file timestamps:
460  *
461  *   0 = seconds only; nanoseconds zeroed.
462  *   1 = seconds and nanoseconds, accurate within 1/HZ.
463  *   2 = seconds and nanoseconds, truncated to microseconds.
464  * >=3 = seconds and nanoseconds, maximum precision.
465  */
466 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
467 
468 static int timestamp_precision = TSP_SEC;
469 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
470     &timestamp_precision, 0, "");
471 
472 /*
473  * Get a current timestamp.
474  */
475 void
476 vfs_timestamp(struct timespec *tsp)
477 {
478 	struct timeval tv;
479 
480 	switch (timestamp_precision) {
481 	case TSP_SEC:
482 		tsp->tv_sec = time_second;
483 		tsp->tv_nsec = 0;
484 		break;
485 	case TSP_HZ:
486 		getnanotime(tsp);
487 		break;
488 	case TSP_USEC:
489 		microtime(&tv);
490 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
491 		break;
492 	case TSP_NSEC:
493 	default:
494 		nanotime(tsp);
495 		break;
496 	}
497 }
498 
499 /*
500  * Set vnode attributes to VNOVAL
501  */
502 void
503 vattr_null(struct vattr *vap)
504 {
505 
506 	vap->va_type = VNON;
507 	vap->va_size = VNOVAL;
508 	vap->va_bytes = VNOVAL;
509 	vap->va_mode = VNOVAL;
510 	vap->va_nlink = VNOVAL;
511 	vap->va_uid = VNOVAL;
512 	vap->va_gid = VNOVAL;
513 	vap->va_fsid = VNOVAL;
514 	vap->va_fileid = VNOVAL;
515 	vap->va_blocksize = VNOVAL;
516 	vap->va_rdev = VNOVAL;
517 	vap->va_atime.tv_sec = VNOVAL;
518 	vap->va_atime.tv_nsec = VNOVAL;
519 	vap->va_mtime.tv_sec = VNOVAL;
520 	vap->va_mtime.tv_nsec = VNOVAL;
521 	vap->va_ctime.tv_sec = VNOVAL;
522 	vap->va_ctime.tv_nsec = VNOVAL;
523 	vap->va_birthtime.tv_sec = VNOVAL;
524 	vap->va_birthtime.tv_nsec = VNOVAL;
525 	vap->va_flags = VNOVAL;
526 	vap->va_gen = VNOVAL;
527 	vap->va_vaflags = 0;
528 }
529 
530 /*
531  * This routine is called when we have too many vnodes.  It attempts
532  * to free <count> vnodes and will potentially free vnodes that still
533  * have VM backing store (VM backing store is typically the cause
534  * of a vnode blowout so we want to do this).  Therefore, this operation
535  * is not considered cheap.
536  *
537  * A number of conditions may prevent a vnode from being reclaimed.
538  * the buffer cache may have references on the vnode, a directory
539  * vnode may still have references due to the namei cache representing
540  * underlying files, or the vnode may be in active use.   It is not
541  * desireable to reuse such vnodes.  These conditions may cause the
542  * number of vnodes to reach some minimum value regardless of what
543  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
544  */
545 static int
546 vlrureclaim(struct mount *mp)
547 {
548 	struct thread *td;
549 	struct vnode *vp;
550 	int done;
551 	int trigger;
552 	int usevnodes;
553 	int count;
554 
555 	/*
556 	 * Calculate the trigger point, don't allow user
557 	 * screwups to blow us up.   This prevents us from
558 	 * recycling vnodes with lots of resident pages.  We
559 	 * aren't trying to free memory, we are trying to
560 	 * free vnodes.
561 	 */
562 	usevnodes = desiredvnodes;
563 	if (usevnodes <= 0)
564 		usevnodes = 1;
565 	trigger = cnt.v_page_count * 2 / usevnodes;
566 	done = 0;
567 	td = curthread;
568 	vn_start_write(NULL, &mp, V_WAIT);
569 	MNT_ILOCK(mp);
570 	count = mp->mnt_nvnodelistsize / 10 + 1;
571 	while (count != 0) {
572 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
573 		while (vp != NULL && vp->v_type == VMARKER)
574 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
575 		if (vp == NULL)
576 			break;
577 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
578 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
579 		--count;
580 		if (!VI_TRYLOCK(vp))
581 			goto next_iter;
582 		/*
583 		 * If it's been deconstructed already, it's still
584 		 * referenced, or it exceeds the trigger, skip it.
585 		 */
586 		if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) ||
587 		    (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
588 		    vp->v_object->resident_page_count > trigger)) {
589 			VI_UNLOCK(vp);
590 			goto next_iter;
591 		}
592 		MNT_IUNLOCK(mp);
593 		vholdl(vp);
594 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT, td)) {
595 			vdrop(vp);
596 			goto next_iter_mntunlocked;
597 		}
598 		VI_LOCK(vp);
599 		/*
600 		 * v_usecount may have been bumped after VOP_LOCK() dropped
601 		 * the vnode interlock and before it was locked again.
602 		 *
603 		 * It is not necessary to recheck VI_DOOMED because it can
604 		 * only be set by another thread that holds both the vnode
605 		 * lock and vnode interlock.  If another thread has the
606 		 * vnode lock before we get to VOP_LOCK() and obtains the
607 		 * vnode interlock after VOP_LOCK() drops the vnode
608 		 * interlock, the other thread will be unable to drop the
609 		 * vnode lock before our VOP_LOCK() call fails.
610 		 */
611 		if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) ||
612 		    (vp->v_object != NULL &&
613 		    vp->v_object->resident_page_count > trigger)) {
614 			VOP_UNLOCK(vp, LK_INTERLOCK, td);
615 			goto next_iter_mntunlocked;
616 		}
617 		KASSERT((vp->v_iflag & VI_DOOMED) == 0,
618 		    ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
619 		vgonel(vp);
620 		VOP_UNLOCK(vp, 0, td);
621 		vdropl(vp);
622 		done++;
623 next_iter_mntunlocked:
624 		if ((count % 256) != 0)
625 			goto relock_mnt;
626 		goto yield;
627 next_iter:
628 		if ((count % 256) != 0)
629 			continue;
630 		MNT_IUNLOCK(mp);
631 yield:
632 		uio_yield();
633 relock_mnt:
634 		MNT_ILOCK(mp);
635 	}
636 	MNT_IUNLOCK(mp);
637 	vn_finished_write(mp);
638 	return done;
639 }
640 
641 /*
642  * Attempt to keep the free list at wantfreevnodes length.
643  */
644 static void
645 vnlru_free(int count)
646 {
647 	struct vnode *vp;
648 	int vfslocked;
649 
650 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
651 	for (; count > 0; count--) {
652 		vp = TAILQ_FIRST(&vnode_free_list);
653 		/*
654 		 * The list can be modified while the free_list_mtx
655 		 * has been dropped and vp could be NULL here.
656 		 */
657 		if (!vp)
658 			break;
659 		VNASSERT(vp->v_op != NULL, vp,
660 		    ("vnlru_free: vnode already reclaimed."));
661 		TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
662 		/*
663 		 * Don't recycle if we can't get the interlock.
664 		 */
665 		if (!VI_TRYLOCK(vp)) {
666 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
667 			continue;
668 		}
669 		VNASSERT(VCANRECYCLE(vp), vp,
670 		    ("vp inconsistent on freelist"));
671 		freevnodes--;
672 		vp->v_iflag &= ~VI_FREE;
673 		vholdl(vp);
674 		mtx_unlock(&vnode_free_list_mtx);
675 		VI_UNLOCK(vp);
676 		vfslocked = VFS_LOCK_GIANT(vp->v_mount);
677 		vtryrecycle(vp);
678 		VFS_UNLOCK_GIANT(vfslocked);
679 		/*
680 		 * If the recycled succeeded this vdrop will actually free
681 		 * the vnode.  If not it will simply place it back on
682 		 * the free list.
683 		 */
684 		vdrop(vp);
685 		mtx_lock(&vnode_free_list_mtx);
686 	}
687 }
688 /*
689  * Attempt to recycle vnodes in a context that is always safe to block.
690  * Calling vlrurecycle() from the bowels of filesystem code has some
691  * interesting deadlock problems.
692  */
693 static struct proc *vnlruproc;
694 static int vnlruproc_sig;
695 
696 static void
697 vnlru_proc(void)
698 {
699 	struct mount *mp, *nmp;
700 	int done;
701 	struct proc *p = vnlruproc;
702 	struct thread *td = FIRST_THREAD_IN_PROC(p);
703 
704 	mtx_lock(&Giant);
705 
706 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p,
707 	    SHUTDOWN_PRI_FIRST);
708 
709 	for (;;) {
710 		kthread_suspend_check(p);
711 		mtx_lock(&vnode_free_list_mtx);
712 		if (freevnodes > wantfreevnodes)
713 			vnlru_free(freevnodes - wantfreevnodes);
714 		if (numvnodes <= desiredvnodes * 9 / 10) {
715 			vnlruproc_sig = 0;
716 			wakeup(&vnlruproc_sig);
717 			msleep(vnlruproc, &vnode_free_list_mtx,
718 			    PVFS|PDROP, "vlruwt", hz);
719 			continue;
720 		}
721 		mtx_unlock(&vnode_free_list_mtx);
722 		done = 0;
723 		mtx_lock(&mountlist_mtx);
724 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
725 			int vfsunlocked;
726 			if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) {
727 				nmp = TAILQ_NEXT(mp, mnt_list);
728 				continue;
729 			}
730 			if (!VFS_NEEDSGIANT(mp)) {
731 				mtx_unlock(&Giant);
732 				vfsunlocked = 1;
733 			} else
734 				vfsunlocked = 0;
735 			done += vlrureclaim(mp);
736 			if (vfsunlocked)
737 				mtx_lock(&Giant);
738 			mtx_lock(&mountlist_mtx);
739 			nmp = TAILQ_NEXT(mp, mnt_list);
740 			vfs_unbusy(mp, td);
741 		}
742 		mtx_unlock(&mountlist_mtx);
743 		if (done == 0) {
744 #if 0
745 			/* These messages are temporary debugging aids */
746 			if (vnlru_nowhere < 5)
747 				printf("vnlru process getting nowhere..\n");
748 			else if (vnlru_nowhere == 5)
749 				printf("vnlru process messages stopped.\n");
750 #endif
751 			vnlru_nowhere++;
752 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
753 		} else
754 			uio_yield();
755 	}
756 }
757 
758 static struct kproc_desc vnlru_kp = {
759 	"vnlru",
760 	vnlru_proc,
761 	&vnlruproc
762 };
763 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
764 
765 /*
766  * Routines having to do with the management of the vnode table.
767  */
768 
769 static void
770 vdestroy(struct vnode *vp)
771 {
772 	struct bufobj *bo;
773 
774 	CTR1(KTR_VFS, "vdestroy vp %p", vp);
775 	mtx_lock(&vnode_free_list_mtx);
776 	numvnodes--;
777 	mtx_unlock(&vnode_free_list_mtx);
778 	bo = &vp->v_bufobj;
779 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
780 	    ("cleaned vnode still on the free list."));
781 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
782 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
783 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
784 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
785 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
786 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
787 	VNASSERT(bo->bo_clean.bv_root == NULL, vp, ("cleanblkroot not NULL"));
788 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
789 	VNASSERT(bo->bo_dirty.bv_root == NULL, vp, ("dirtyblkroot not NULL"));
790 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
791 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
792 	VI_UNLOCK(vp);
793 #ifdef MAC
794 	mac_destroy_vnode(vp);
795 #endif
796 	if (vp->v_pollinfo != NULL) {
797 		knlist_destroy(&vp->v_pollinfo->vpi_selinfo.si_note);
798 		mtx_destroy(&vp->v_pollinfo->vpi_lock);
799 		uma_zfree(vnodepoll_zone, vp->v_pollinfo);
800 	}
801 #ifdef INVARIANTS
802 	/* XXX Elsewhere we can detect an already freed vnode via NULL v_op. */
803 	vp->v_op = NULL;
804 #endif
805 	lockdestroy(vp->v_vnlock);
806 	mtx_destroy(&vp->v_interlock);
807 	uma_zfree(vnode_zone, vp);
808 }
809 
810 /*
811  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
812  * before we actually vgone().  This function must be called with the vnode
813  * held to prevent the vnode from being returned to the free list midway
814  * through vgone().
815  */
816 static int
817 vtryrecycle(struct vnode *vp)
818 {
819 	struct thread *td = curthread;
820 	struct mount *vnmp;
821 
822 	CTR1(KTR_VFS, "vtryrecycle: trying vp %p", vp);
823 	VNASSERT(vp->v_holdcnt, vp,
824 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
825 	/*
826 	 * This vnode may found and locked via some other list, if so we
827 	 * can't recycle it yet.
828 	 */
829 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
830 		return (EWOULDBLOCK);
831 	/*
832 	 * Don't recycle if its filesystem is being suspended.
833 	 */
834 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
835 		VOP_UNLOCK(vp, 0, td);
836 		return (EBUSY);
837 	}
838 	/*
839 	 * If we got this far, we need to acquire the interlock and see if
840 	 * anyone picked up this vnode from another list.  If not, we will
841 	 * mark it with DOOMED via vgonel() so that anyone who does find it
842 	 * will skip over it.
843 	 */
844 	VI_LOCK(vp);
845 	if (vp->v_usecount) {
846 		VOP_UNLOCK(vp, LK_INTERLOCK, td);
847 		vn_finished_write(vnmp);
848 		return (EBUSY);
849 	}
850 	if ((vp->v_iflag & VI_DOOMED) == 0)
851 		vgonel(vp);
852 	VOP_UNLOCK(vp, LK_INTERLOCK, td);
853 	vn_finished_write(vnmp);
854 	CTR1(KTR_VFS, "vtryrecycle: recycled vp %p", vp);
855 	return (0);
856 }
857 
858 /*
859  * Return the next vnode from the free list.
860  */
861 int
862 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
863     struct vnode **vpp)
864 {
865 	struct vnode *vp = NULL;
866 	struct bufobj *bo;
867 
868 	mtx_lock(&vnode_free_list_mtx);
869 	/*
870 	 * Lend our context to reclaim vnodes if they've exceeded the max.
871 	 */
872 	if (freevnodes > wantfreevnodes)
873 		vnlru_free(1);
874 	/*
875 	 * Wait for available vnodes.
876 	 */
877 	if (numvnodes > desiredvnodes) {
878 		if (mp != NULL && (mp->mnt_kern_flag & MNTK_SUSPEND)) {
879 			/*
880 			 * File system is beeing suspended, we cannot risk a
881 			 * deadlock here, so allocate new vnode anyway.
882 			 */
883 			if (freevnodes > wantfreevnodes)
884 				vnlru_free(freevnodes - wantfreevnodes);
885 			goto alloc;
886 		}
887 		if (vnlruproc_sig == 0) {
888 			vnlruproc_sig = 1;      /* avoid unnecessary wakeups */
889 			wakeup(vnlruproc);
890 		}
891 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
892 		    "vlruwk", hz);
893 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
894 		if (numvnodes > desiredvnodes) {
895 			mtx_unlock(&vnode_free_list_mtx);
896 			return (ENFILE);
897 		}
898 #endif
899 	}
900 alloc:
901 	numvnodes++;
902 	mtx_unlock(&vnode_free_list_mtx);
903 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO);
904 	/*
905 	 * Setup locks.
906 	 */
907 	vp->v_vnlock = &vp->v_lock;
908 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
909 	/*
910 	 * By default, don't allow shared locks unless filesystems
911 	 * opt-in.
912 	 */
913 	lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOSHARE);
914 	/*
915 	 * Initialize bufobj.
916 	 */
917 	bo = &vp->v_bufobj;
918 	bo->__bo_vnode = vp;
919 	bo->bo_mtx = &vp->v_interlock;
920 	bo->bo_ops = &buf_ops_bio;
921 	bo->bo_private = vp;
922 	TAILQ_INIT(&bo->bo_clean.bv_hd);
923 	TAILQ_INIT(&bo->bo_dirty.bv_hd);
924 	/*
925 	 * Initialize namecache.
926 	 */
927 	LIST_INIT(&vp->v_cache_src);
928 	TAILQ_INIT(&vp->v_cache_dst);
929 	/*
930 	 * Finalize various vnode identity bits.
931 	 */
932 	vp->v_type = VNON;
933 	vp->v_tag = tag;
934 	vp->v_op = vops;
935 	v_incr_usecount(vp);
936 	vp->v_data = 0;
937 #ifdef MAC
938 	mac_init_vnode(vp);
939 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
940 		mac_associate_vnode_singlelabel(mp, vp);
941 	else if (mp == NULL)
942 		printf("NULL mp in getnewvnode()\n");
943 #endif
944 	if (mp != NULL) {
945 		bo->bo_bsize = mp->mnt_stat.f_iosize;
946 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
947 			vp->v_vflag |= VV_NOKNOTE;
948 	}
949 
950 	CTR2(KTR_VFS, "getnewvnode: mp %p vp %p", mp, vp);
951 	*vpp = vp;
952 	return (0);
953 }
954 
955 /*
956  * Delete from old mount point vnode list, if on one.
957  */
958 static void
959 delmntque(struct vnode *vp)
960 {
961 	struct mount *mp;
962 
963 	mp = vp->v_mount;
964 	if (mp == NULL)
965 		return;
966 	MNT_ILOCK(mp);
967 	vp->v_mount = NULL;
968 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
969 		("bad mount point vnode list size"));
970 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
971 	mp->mnt_nvnodelistsize--;
972 	MNT_REL(mp);
973 	MNT_IUNLOCK(mp);
974 }
975 
976 static void
977 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
978 {
979 	struct thread *td;
980 
981 	td = curthread; /* XXX ? */
982 	vp->v_data = NULL;
983 	vp->v_op = &dead_vnodeops;
984 	/* XXX non mp-safe fs may still call insmntque with vnode
985 	   unlocked */
986 	if (!VOP_ISLOCKED(vp, td))
987 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
988 	vgone(vp);
989 	vput(vp);
990 }
991 
992 /*
993  * Insert into list of vnodes for the new mount point, if available.
994  */
995 int
996 insmntque1(struct vnode *vp, struct mount *mp,
997 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
998 {
999 
1000 	KASSERT(vp->v_mount == NULL,
1001 		("insmntque: vnode already on per mount vnode list"));
1002 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1003 	MNT_ILOCK(mp);
1004 	if ((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1005 	    mp->mnt_nvnodelistsize == 0) {
1006 		MNT_IUNLOCK(mp);
1007 		if (dtr != NULL)
1008 			dtr(vp, dtr_arg);
1009 		return (EBUSY);
1010 	}
1011 	vp->v_mount = mp;
1012 	MNT_REF(mp);
1013 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1014 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1015 		("neg mount point vnode list size"));
1016 	mp->mnt_nvnodelistsize++;
1017 	MNT_IUNLOCK(mp);
1018 	return (0);
1019 }
1020 
1021 int
1022 insmntque(struct vnode *vp, struct mount *mp)
1023 {
1024 
1025 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1026 }
1027 
1028 /*
1029  * Flush out and invalidate all buffers associated with a bufobj
1030  * Called with the underlying object locked.
1031  */
1032 int
1033 bufobj_invalbuf(struct bufobj *bo, int flags, struct thread *td, int slpflag,
1034     int slptimeo)
1035 {
1036 	int error;
1037 
1038 	BO_LOCK(bo);
1039 	if (flags & V_SAVE) {
1040 		error = bufobj_wwait(bo, slpflag, slptimeo);
1041 		if (error) {
1042 			BO_UNLOCK(bo);
1043 			return (error);
1044 		}
1045 		if (bo->bo_dirty.bv_cnt > 0) {
1046 			BO_UNLOCK(bo);
1047 			if ((error = BO_SYNC(bo, MNT_WAIT, td)) != 0)
1048 				return (error);
1049 			/*
1050 			 * XXX We could save a lock/unlock if this was only
1051 			 * enabled under INVARIANTS
1052 			 */
1053 			BO_LOCK(bo);
1054 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1055 				panic("vinvalbuf: dirty bufs");
1056 		}
1057 	}
1058 	/*
1059 	 * If you alter this loop please notice that interlock is dropped and
1060 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1061 	 * no race conditions occur from this.
1062 	 */
1063 	do {
1064 		error = flushbuflist(&bo->bo_clean,
1065 		    flags, bo, slpflag, slptimeo);
1066 		if (error == 0)
1067 			error = flushbuflist(&bo->bo_dirty,
1068 			    flags, bo, slpflag, slptimeo);
1069 		if (error != 0 && error != EAGAIN) {
1070 			BO_UNLOCK(bo);
1071 			return (error);
1072 		}
1073 	} while (error != 0);
1074 
1075 	/*
1076 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1077 	 * have write I/O in-progress but if there is a VM object then the
1078 	 * VM object can also have read-I/O in-progress.
1079 	 */
1080 	do {
1081 		bufobj_wwait(bo, 0, 0);
1082 		BO_UNLOCK(bo);
1083 		if (bo->bo_object != NULL) {
1084 			VM_OBJECT_LOCK(bo->bo_object);
1085 			vm_object_pip_wait(bo->bo_object, "bovlbx");
1086 			VM_OBJECT_UNLOCK(bo->bo_object);
1087 		}
1088 		BO_LOCK(bo);
1089 	} while (bo->bo_numoutput > 0);
1090 	BO_UNLOCK(bo);
1091 
1092 	/*
1093 	 * Destroy the copy in the VM cache, too.
1094 	 */
1095 	if (bo->bo_object != NULL) {
1096 		VM_OBJECT_LOCK(bo->bo_object);
1097 		vm_object_page_remove(bo->bo_object, 0, 0,
1098 			(flags & V_SAVE) ? TRUE : FALSE);
1099 		VM_OBJECT_UNLOCK(bo->bo_object);
1100 	}
1101 
1102 #ifdef INVARIANTS
1103 	BO_LOCK(bo);
1104 	if ((flags & (V_ALT | V_NORMAL)) == 0 &&
1105 	    (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0))
1106 		panic("vinvalbuf: flush failed");
1107 	BO_UNLOCK(bo);
1108 #endif
1109 	return (0);
1110 }
1111 
1112 /*
1113  * Flush out and invalidate all buffers associated with a vnode.
1114  * Called with the underlying object locked.
1115  */
1116 int
1117 vinvalbuf(struct vnode *vp, int flags, struct thread *td, int slpflag,
1118     int slptimeo)
1119 {
1120 
1121 	CTR2(KTR_VFS, "vinvalbuf vp %p flags %d", vp, flags);
1122 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1123 	return (bufobj_invalbuf(&vp->v_bufobj, flags, td, slpflag, slptimeo));
1124 }
1125 
1126 /*
1127  * Flush out buffers on the specified list.
1128  *
1129  */
1130 static int
1131 flushbuflist( struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1132     int slptimeo)
1133 {
1134 	struct buf *bp, *nbp;
1135 	int retval, error;
1136 	daddr_t lblkno;
1137 	b_xflags_t xflags;
1138 
1139 	ASSERT_BO_LOCKED(bo);
1140 
1141 	retval = 0;
1142 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1143 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1144 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1145 			continue;
1146 		}
1147 		lblkno = 0;
1148 		xflags = 0;
1149 		if (nbp != NULL) {
1150 			lblkno = nbp->b_lblkno;
1151 			xflags = nbp->b_xflags &
1152 				(BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN);
1153 		}
1154 		retval = EAGAIN;
1155 		error = BUF_TIMELOCK(bp,
1156 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo),
1157 		    "flushbuf", slpflag, slptimeo);
1158 		if (error) {
1159 			BO_LOCK(bo);
1160 			return (error != ENOLCK ? error : EAGAIN);
1161 		}
1162 		KASSERT(bp->b_bufobj == bo,
1163 	            ("bp %p wrong b_bufobj %p should be %p",
1164 		    bp, bp->b_bufobj, bo));
1165 		if (bp->b_bufobj != bo) {	/* XXX: necessary ? */
1166 			BUF_UNLOCK(bp);
1167 			BO_LOCK(bo);
1168 			return (EAGAIN);
1169 		}
1170 		/*
1171 		 * XXX Since there are no node locks for NFS, I
1172 		 * believe there is a slight chance that a delayed
1173 		 * write will occur while sleeping just above, so
1174 		 * check for it.
1175 		 */
1176 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1177 		    (flags & V_SAVE)) {
1178 			bremfree(bp);
1179 			bp->b_flags |= B_ASYNC;
1180 			bwrite(bp);
1181 			BO_LOCK(bo);
1182 			return (EAGAIN);	/* XXX: why not loop ? */
1183 		}
1184 		bremfree(bp);
1185 		bp->b_flags |= (B_INVAL | B_RELBUF);
1186 		bp->b_flags &= ~B_ASYNC;
1187 		brelse(bp);
1188 		BO_LOCK(bo);
1189 		if (nbp != NULL &&
1190 		    (nbp->b_bufobj != bo ||
1191 		     nbp->b_lblkno != lblkno ||
1192 		     (nbp->b_xflags &
1193 		      (BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN)) != xflags))
1194 			break;			/* nbp invalid */
1195 	}
1196 	return (retval);
1197 }
1198 
1199 /*
1200  * Truncate a file's buffer and pages to a specified length.  This
1201  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1202  * sync activity.
1203  */
1204 int
1205 vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td,
1206     off_t length, int blksize)
1207 {
1208 	struct buf *bp, *nbp;
1209 	int anyfreed;
1210 	int trunclbn;
1211 	struct bufobj *bo;
1212 
1213 	CTR2(KTR_VFS, "vtruncbuf vp %p length %jd", vp, length);
1214 	/*
1215 	 * Round up to the *next* lbn.
1216 	 */
1217 	trunclbn = (length + blksize - 1) / blksize;
1218 
1219 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1220 restart:
1221 	VI_LOCK(vp);
1222 	bo = &vp->v_bufobj;
1223 	anyfreed = 1;
1224 	for (;anyfreed;) {
1225 		anyfreed = 0;
1226 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1227 			if (bp->b_lblkno < trunclbn)
1228 				continue;
1229 			if (BUF_LOCK(bp,
1230 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1231 			    VI_MTX(vp)) == ENOLCK)
1232 				goto restart;
1233 
1234 			bremfree(bp);
1235 			bp->b_flags |= (B_INVAL | B_RELBUF);
1236 			bp->b_flags &= ~B_ASYNC;
1237 			brelse(bp);
1238 			anyfreed = 1;
1239 
1240 			if (nbp != NULL &&
1241 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1242 			    (nbp->b_vp != vp) ||
1243 			    (nbp->b_flags & B_DELWRI))) {
1244 				goto restart;
1245 			}
1246 			VI_LOCK(vp);
1247 		}
1248 
1249 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1250 			if (bp->b_lblkno < trunclbn)
1251 				continue;
1252 			if (BUF_LOCK(bp,
1253 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1254 			    VI_MTX(vp)) == ENOLCK)
1255 				goto restart;
1256 			bremfree(bp);
1257 			bp->b_flags |= (B_INVAL | B_RELBUF);
1258 			bp->b_flags &= ~B_ASYNC;
1259 			brelse(bp);
1260 			anyfreed = 1;
1261 			if (nbp != NULL &&
1262 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1263 			    (nbp->b_vp != vp) ||
1264 			    (nbp->b_flags & B_DELWRI) == 0)) {
1265 				goto restart;
1266 			}
1267 			VI_LOCK(vp);
1268 		}
1269 	}
1270 
1271 	if (length > 0) {
1272 restartsync:
1273 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1274 			if (bp->b_lblkno > 0)
1275 				continue;
1276 			/*
1277 			 * Since we hold the vnode lock this should only
1278 			 * fail if we're racing with the buf daemon.
1279 			 */
1280 			if (BUF_LOCK(bp,
1281 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1282 			    VI_MTX(vp)) == ENOLCK) {
1283 				goto restart;
1284 			}
1285 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1286 			    ("buf(%p) on dirty queue without DELWRI", bp));
1287 
1288 			bremfree(bp);
1289 			bawrite(bp);
1290 			VI_LOCK(vp);
1291 			goto restartsync;
1292 		}
1293 	}
1294 
1295 	bufobj_wwait(bo, 0, 0);
1296 	VI_UNLOCK(vp);
1297 	vnode_pager_setsize(vp, length);
1298 
1299 	return (0);
1300 }
1301 
1302 /*
1303  * buf_splay() - splay tree core for the clean/dirty list of buffers in
1304  * 		 a vnode.
1305  *
1306  *	NOTE: We have to deal with the special case of a background bitmap
1307  *	buffer, a situation where two buffers will have the same logical
1308  *	block offset.  We want (1) only the foreground buffer to be accessed
1309  *	in a lookup and (2) must differentiate between the foreground and
1310  *	background buffer in the splay tree algorithm because the splay
1311  *	tree cannot normally handle multiple entities with the same 'index'.
1312  *	We accomplish this by adding differentiating flags to the splay tree's
1313  *	numerical domain.
1314  */
1315 static
1316 struct buf *
1317 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root)
1318 {
1319 	struct buf dummy;
1320 	struct buf *lefttreemax, *righttreemin, *y;
1321 
1322 	if (root == NULL)
1323 		return (NULL);
1324 	lefttreemax = righttreemin = &dummy;
1325 	for (;;) {
1326 		if (lblkno < root->b_lblkno ||
1327 		    (lblkno == root->b_lblkno &&
1328 		    (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1329 			if ((y = root->b_left) == NULL)
1330 				break;
1331 			if (lblkno < y->b_lblkno) {
1332 				/* Rotate right. */
1333 				root->b_left = y->b_right;
1334 				y->b_right = root;
1335 				root = y;
1336 				if ((y = root->b_left) == NULL)
1337 					break;
1338 			}
1339 			/* Link into the new root's right tree. */
1340 			righttreemin->b_left = root;
1341 			righttreemin = root;
1342 		} else if (lblkno > root->b_lblkno ||
1343 		    (lblkno == root->b_lblkno &&
1344 		    (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) {
1345 			if ((y = root->b_right) == NULL)
1346 				break;
1347 			if (lblkno > y->b_lblkno) {
1348 				/* Rotate left. */
1349 				root->b_right = y->b_left;
1350 				y->b_left = root;
1351 				root = y;
1352 				if ((y = root->b_right) == NULL)
1353 					break;
1354 			}
1355 			/* Link into the new root's left tree. */
1356 			lefttreemax->b_right = root;
1357 			lefttreemax = root;
1358 		} else {
1359 			break;
1360 		}
1361 		root = y;
1362 	}
1363 	/* Assemble the new root. */
1364 	lefttreemax->b_right = root->b_left;
1365 	righttreemin->b_left = root->b_right;
1366 	root->b_left = dummy.b_right;
1367 	root->b_right = dummy.b_left;
1368 	return (root);
1369 }
1370 
1371 static void
1372 buf_vlist_remove(struct buf *bp)
1373 {
1374 	struct buf *root;
1375 	struct bufv *bv;
1376 
1377 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1378 	ASSERT_BO_LOCKED(bp->b_bufobj);
1379 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
1380 	    (BX_VNDIRTY|BX_VNCLEAN),
1381 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
1382 	if (bp->b_xflags & BX_VNDIRTY)
1383 		bv = &bp->b_bufobj->bo_dirty;
1384 	else
1385 		bv = &bp->b_bufobj->bo_clean;
1386 	if (bp != bv->bv_root) {
1387 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1388 		KASSERT(root == bp, ("splay lookup failed in remove"));
1389 	}
1390 	if (bp->b_left == NULL) {
1391 		root = bp->b_right;
1392 	} else {
1393 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left);
1394 		root->b_right = bp->b_right;
1395 	}
1396 	bv->bv_root = root;
1397 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1398 	bv->bv_cnt--;
1399 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1400 }
1401 
1402 /*
1403  * Add the buffer to the sorted clean or dirty block list using a
1404  * splay tree algorithm.
1405  *
1406  * NOTE: xflags is passed as a constant, optimizing this inline function!
1407  */
1408 static void
1409 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1410 {
1411 	struct buf *root;
1412 	struct bufv *bv;
1413 
1414 	ASSERT_BO_LOCKED(bo);
1415 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1416 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
1417 	bp->b_xflags |= xflags;
1418 	if (xflags & BX_VNDIRTY)
1419 		bv = &bo->bo_dirty;
1420 	else
1421 		bv = &bo->bo_clean;
1422 
1423 	root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1424 	if (root == NULL) {
1425 		bp->b_left = NULL;
1426 		bp->b_right = NULL;
1427 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1428 	} else if (bp->b_lblkno < root->b_lblkno ||
1429 	    (bp->b_lblkno == root->b_lblkno &&
1430 	    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1431 		bp->b_left = root->b_left;
1432 		bp->b_right = root;
1433 		root->b_left = NULL;
1434 		TAILQ_INSERT_BEFORE(root, bp, b_bobufs);
1435 	} else {
1436 		bp->b_right = root->b_right;
1437 		bp->b_left = root;
1438 		root->b_right = NULL;
1439 		TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs);
1440 	}
1441 	bv->bv_cnt++;
1442 	bv->bv_root = bp;
1443 }
1444 
1445 /*
1446  * Lookup a buffer using the splay tree.  Note that we specifically avoid
1447  * shadow buffers used in background bitmap writes.
1448  *
1449  * This code isn't quite efficient as it could be because we are maintaining
1450  * two sorted lists and do not know which list the block resides in.
1451  *
1452  * During a "make buildworld" the desired buffer is found at one of
1453  * the roots more than 60% of the time.  Thus, checking both roots
1454  * before performing either splay eliminates unnecessary splays on the
1455  * first tree splayed.
1456  */
1457 struct buf *
1458 gbincore(struct bufobj *bo, daddr_t lblkno)
1459 {
1460 	struct buf *bp;
1461 
1462 	ASSERT_BO_LOCKED(bo);
1463 	if ((bp = bo->bo_clean.bv_root) != NULL &&
1464 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1465 		return (bp);
1466 	if ((bp = bo->bo_dirty.bv_root) != NULL &&
1467 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1468 		return (bp);
1469 	if ((bp = bo->bo_clean.bv_root) != NULL) {
1470 		bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp);
1471 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1472 			return (bp);
1473 	}
1474 	if ((bp = bo->bo_dirty.bv_root) != NULL) {
1475 		bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp);
1476 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1477 			return (bp);
1478 	}
1479 	return (NULL);
1480 }
1481 
1482 /*
1483  * Associate a buffer with a vnode.
1484  */
1485 void
1486 bgetvp(struct vnode *vp, struct buf *bp)
1487 {
1488 
1489 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
1490 
1491 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
1492 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
1493 	    ("bgetvp: bp already attached! %p", bp));
1494 
1495 	ASSERT_VI_LOCKED(vp, "bgetvp");
1496 	vholdl(vp);
1497 	if (VFS_NEEDSGIANT(vp->v_mount) ||
1498 	    vp->v_bufobj.bo_flag & BO_NEEDSGIANT)
1499 		bp->b_flags |= B_NEEDSGIANT;
1500 	bp->b_vp = vp;
1501 	bp->b_bufobj = &vp->v_bufobj;
1502 	/*
1503 	 * Insert onto list for new vnode.
1504 	 */
1505 	buf_vlist_add(bp, &vp->v_bufobj, BX_VNCLEAN);
1506 }
1507 
1508 /*
1509  * Disassociate a buffer from a vnode.
1510  */
1511 void
1512 brelvp(struct buf *bp)
1513 {
1514 	struct bufobj *bo;
1515 	struct vnode *vp;
1516 
1517 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1518 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1519 
1520 	/*
1521 	 * Delete from old vnode list, if on one.
1522 	 */
1523 	vp = bp->b_vp;		/* XXX */
1524 	bo = bp->b_bufobj;
1525 	BO_LOCK(bo);
1526 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1527 		buf_vlist_remove(bp);
1528 	else
1529 		panic("brelvp: Buffer %p not on queue.", bp);
1530 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1531 		bo->bo_flag &= ~BO_ONWORKLST;
1532 		mtx_lock(&sync_mtx);
1533 		LIST_REMOVE(bo, bo_synclist);
1534  		syncer_worklist_len--;
1535 		mtx_unlock(&sync_mtx);
1536 	}
1537 	bp->b_flags &= ~B_NEEDSGIANT;
1538 	bp->b_vp = NULL;
1539 	bp->b_bufobj = NULL;
1540 	vdropl(vp);
1541 }
1542 
1543 /*
1544  * Add an item to the syncer work queue.
1545  */
1546 static void
1547 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
1548 {
1549 	int slot;
1550 
1551 	ASSERT_BO_LOCKED(bo);
1552 
1553 	mtx_lock(&sync_mtx);
1554 	if (bo->bo_flag & BO_ONWORKLST)
1555 		LIST_REMOVE(bo, bo_synclist);
1556 	else {
1557 		bo->bo_flag |= BO_ONWORKLST;
1558  		syncer_worklist_len++;
1559 	}
1560 
1561 	if (delay > syncer_maxdelay - 2)
1562 		delay = syncer_maxdelay - 2;
1563 	slot = (syncer_delayno + delay) & syncer_mask;
1564 
1565 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
1566 	mtx_unlock(&sync_mtx);
1567 }
1568 
1569 static int
1570 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
1571 {
1572 	int error, len;
1573 
1574 	mtx_lock(&sync_mtx);
1575 	len = syncer_worklist_len - sync_vnode_count;
1576 	mtx_unlock(&sync_mtx);
1577 	error = SYSCTL_OUT(req, &len, sizeof(len));
1578 	return (error);
1579 }
1580 
1581 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
1582     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
1583 
1584 static struct proc *updateproc;
1585 static void sched_sync(void);
1586 static struct kproc_desc up_kp = {
1587 	"syncer",
1588 	sched_sync,
1589 	&updateproc
1590 };
1591 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp)
1592 
1593 static int
1594 sync_vnode(struct bufobj *bo, struct thread *td)
1595 {
1596 	struct vnode *vp;
1597 	struct mount *mp;
1598 
1599 	vp = bo->__bo_vnode; 	/* XXX */
1600 	if (VOP_ISLOCKED(vp, NULL) != 0)
1601 		return (1);
1602 	if (VI_TRYLOCK(vp) == 0)
1603 		return (1);
1604 	/*
1605 	 * We use vhold in case the vnode does not
1606 	 * successfully sync.  vhold prevents the vnode from
1607 	 * going away when we unlock the sync_mtx so that
1608 	 * we can acquire the vnode interlock.
1609 	 */
1610 	vholdl(vp);
1611 	mtx_unlock(&sync_mtx);
1612 	VI_UNLOCK(vp);
1613 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1614 		vdrop(vp);
1615 		mtx_lock(&sync_mtx);
1616 		return (1);
1617 	}
1618 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
1619 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
1620 	VOP_UNLOCK(vp, 0, td);
1621 	vn_finished_write(mp);
1622 	VI_LOCK(vp);
1623 	if ((bo->bo_flag & BO_ONWORKLST) != 0) {
1624 		/*
1625 		 * Put us back on the worklist.  The worklist
1626 		 * routine will remove us from our current
1627 		 * position and then add us back in at a later
1628 		 * position.
1629 		 */
1630 		vn_syncer_add_to_worklist(bo, syncdelay);
1631 	}
1632 	vdropl(vp);
1633 	mtx_lock(&sync_mtx);
1634 	return (0);
1635 }
1636 
1637 /*
1638  * System filesystem synchronizer daemon.
1639  */
1640 static void
1641 sched_sync(void)
1642 {
1643 	struct synclist *next;
1644 	struct synclist *slp;
1645 	struct bufobj *bo;
1646 	long starttime;
1647 	struct thread *td = FIRST_THREAD_IN_PROC(updateproc);
1648 	static int dummychan;
1649 	int last_work_seen;
1650 	int net_worklist_len;
1651 	int syncer_final_iter;
1652 	int first_printf;
1653 	int error;
1654 
1655 	mtx_lock(&Giant);
1656 	last_work_seen = 0;
1657 	syncer_final_iter = 0;
1658 	first_printf = 1;
1659 	syncer_state = SYNCER_RUNNING;
1660 	starttime = time_uptime;
1661 	td->td_pflags |= TDP_NORUNNINGBUF;
1662 
1663 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
1664 	    SHUTDOWN_PRI_LAST);
1665 
1666 	mtx_lock(&sync_mtx);
1667 	for (;;) {
1668 		if (syncer_state == SYNCER_FINAL_DELAY &&
1669 		    syncer_final_iter == 0) {
1670 			mtx_unlock(&sync_mtx);
1671 			kthread_suspend_check(td->td_proc);
1672 			mtx_lock(&sync_mtx);
1673 		}
1674 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
1675 		if (syncer_state != SYNCER_RUNNING &&
1676 		    starttime != time_uptime) {
1677 			if (first_printf) {
1678 				printf("\nSyncing disks, vnodes remaining...");
1679 				first_printf = 0;
1680 			}
1681 			printf("%d ", net_worklist_len);
1682 		}
1683 		starttime = time_uptime;
1684 
1685 		/*
1686 		 * Push files whose dirty time has expired.  Be careful
1687 		 * of interrupt race on slp queue.
1688 		 *
1689 		 * Skip over empty worklist slots when shutting down.
1690 		 */
1691 		do {
1692 			slp = &syncer_workitem_pending[syncer_delayno];
1693 			syncer_delayno += 1;
1694 			if (syncer_delayno == syncer_maxdelay)
1695 				syncer_delayno = 0;
1696 			next = &syncer_workitem_pending[syncer_delayno];
1697 			/*
1698 			 * If the worklist has wrapped since the
1699 			 * it was emptied of all but syncer vnodes,
1700 			 * switch to the FINAL_DELAY state and run
1701 			 * for one more second.
1702 			 */
1703 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
1704 			    net_worklist_len == 0 &&
1705 			    last_work_seen == syncer_delayno) {
1706 				syncer_state = SYNCER_FINAL_DELAY;
1707 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
1708 			}
1709 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
1710 		    syncer_worklist_len > 0);
1711 
1712 		/*
1713 		 * Keep track of the last time there was anything
1714 		 * on the worklist other than syncer vnodes.
1715 		 * Return to the SHUTTING_DOWN state if any
1716 		 * new work appears.
1717 		 */
1718 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
1719 			last_work_seen = syncer_delayno;
1720 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
1721 			syncer_state = SYNCER_SHUTTING_DOWN;
1722 		while ((bo = LIST_FIRST(slp)) != NULL) {
1723 			error = sync_vnode(bo, td);
1724 			if (error == 1) {
1725 				LIST_REMOVE(bo, bo_synclist);
1726 				LIST_INSERT_HEAD(next, bo, bo_synclist);
1727 				continue;
1728 			}
1729 		}
1730 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
1731 			syncer_final_iter--;
1732 		/*
1733 		 * The variable rushjob allows the kernel to speed up the
1734 		 * processing of the filesystem syncer process. A rushjob
1735 		 * value of N tells the filesystem syncer to process the next
1736 		 * N seconds worth of work on its queue ASAP. Currently rushjob
1737 		 * is used by the soft update code to speed up the filesystem
1738 		 * syncer process when the incore state is getting so far
1739 		 * ahead of the disk that the kernel memory pool is being
1740 		 * threatened with exhaustion.
1741 		 */
1742 		if (rushjob > 0) {
1743 			rushjob -= 1;
1744 			continue;
1745 		}
1746 		/*
1747 		 * Just sleep for a short period of time between
1748 		 * iterations when shutting down to allow some I/O
1749 		 * to happen.
1750 		 *
1751 		 * If it has taken us less than a second to process the
1752 		 * current work, then wait. Otherwise start right over
1753 		 * again. We can still lose time if any single round
1754 		 * takes more than two seconds, but it does not really
1755 		 * matter as we are just trying to generally pace the
1756 		 * filesystem activity.
1757 		 */
1758 		if (syncer_state != SYNCER_RUNNING)
1759 			msleep(&dummychan, &sync_mtx, PPAUSE, "syncfnl",
1760 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
1761 		else if (time_uptime == starttime)
1762 			msleep(&lbolt, &sync_mtx, PPAUSE, "syncer", 0);
1763 	}
1764 }
1765 
1766 /*
1767  * Request the syncer daemon to speed up its work.
1768  * We never push it to speed up more than half of its
1769  * normal turn time, otherwise it could take over the cpu.
1770  */
1771 int
1772 speedup_syncer(void)
1773 {
1774 	struct thread *td;
1775 	int ret = 0;
1776 
1777 	td = FIRST_THREAD_IN_PROC(updateproc);
1778 	sleepq_remove(td, &lbolt);
1779 	mtx_lock(&sync_mtx);
1780 	if (rushjob < syncdelay / 2) {
1781 		rushjob += 1;
1782 		stat_rush_requests += 1;
1783 		ret = 1;
1784 	}
1785 	mtx_unlock(&sync_mtx);
1786 	return (ret);
1787 }
1788 
1789 /*
1790  * Tell the syncer to speed up its work and run though its work
1791  * list several times, then tell it to shut down.
1792  */
1793 static void
1794 syncer_shutdown(void *arg, int howto)
1795 {
1796 	struct thread *td;
1797 
1798 	if (howto & RB_NOSYNC)
1799 		return;
1800 	td = FIRST_THREAD_IN_PROC(updateproc);
1801 	sleepq_remove(td, &lbolt);
1802 	mtx_lock(&sync_mtx);
1803 	syncer_state = SYNCER_SHUTTING_DOWN;
1804 	rushjob = 0;
1805 	mtx_unlock(&sync_mtx);
1806 	kproc_shutdown(arg, howto);
1807 }
1808 
1809 /*
1810  * Reassign a buffer from one vnode to another.
1811  * Used to assign file specific control information
1812  * (indirect blocks) to the vnode to which they belong.
1813  */
1814 void
1815 reassignbuf(struct buf *bp)
1816 {
1817 	struct vnode *vp;
1818 	struct bufobj *bo;
1819 	int delay;
1820 #ifdef INVARIANTS
1821 	struct bufv *bv;
1822 #endif
1823 
1824 	vp = bp->b_vp;
1825 	bo = bp->b_bufobj;
1826 	++reassignbufcalls;
1827 
1828 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
1829 	    bp, bp->b_vp, bp->b_flags);
1830 	/*
1831 	 * B_PAGING flagged buffers cannot be reassigned because their vp
1832 	 * is not fully linked in.
1833 	 */
1834 	if (bp->b_flags & B_PAGING)
1835 		panic("cannot reassign paging buffer");
1836 
1837 	/*
1838 	 * Delete from old vnode list, if on one.
1839 	 */
1840 	VI_LOCK(vp);
1841 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1842 		buf_vlist_remove(bp);
1843 	else
1844 		panic("reassignbuf: Buffer %p not on queue.", bp);
1845 	/*
1846 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
1847 	 * of clean buffers.
1848 	 */
1849 	if (bp->b_flags & B_DELWRI) {
1850 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
1851 			switch (vp->v_type) {
1852 			case VDIR:
1853 				delay = dirdelay;
1854 				break;
1855 			case VCHR:
1856 				delay = metadelay;
1857 				break;
1858 			default:
1859 				delay = filedelay;
1860 			}
1861 			vn_syncer_add_to_worklist(bo, delay);
1862 		}
1863 		buf_vlist_add(bp, bo, BX_VNDIRTY);
1864 	} else {
1865 		buf_vlist_add(bp, bo, BX_VNCLEAN);
1866 
1867 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1868 			mtx_lock(&sync_mtx);
1869 			LIST_REMOVE(bo, bo_synclist);
1870  			syncer_worklist_len--;
1871 			mtx_unlock(&sync_mtx);
1872 			bo->bo_flag &= ~BO_ONWORKLST;
1873 		}
1874 	}
1875 #ifdef INVARIANTS
1876 	bv = &bo->bo_clean;
1877 	bp = TAILQ_FIRST(&bv->bv_hd);
1878 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1879 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1880 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
1881 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1882 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1883 	bv = &bo->bo_dirty;
1884 	bp = TAILQ_FIRST(&bv->bv_hd);
1885 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1886 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1887 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
1888 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1889 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1890 #endif
1891 	VI_UNLOCK(vp);
1892 }
1893 
1894 /*
1895  * Increment the use and hold counts on the vnode, taking care to reference
1896  * the driver's usecount if this is a chardev.  The vholdl() will remove
1897  * the vnode from the free list if it is presently free.  Requires the
1898  * vnode interlock and returns with it held.
1899  */
1900 static void
1901 v_incr_usecount(struct vnode *vp)
1902 {
1903 
1904 	CTR3(KTR_VFS, "v_incr_usecount: vp %p holdcnt %d usecount %d\n",
1905 	    vp, vp->v_holdcnt, vp->v_usecount);
1906 	vp->v_usecount++;
1907 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1908 		dev_lock();
1909 		vp->v_rdev->si_usecount++;
1910 		dev_unlock();
1911 	}
1912 	vholdl(vp);
1913 }
1914 
1915 /*
1916  * Turn a holdcnt into a use+holdcnt such that only one call to
1917  * v_decr_usecount is needed.
1918  */
1919 static void
1920 v_upgrade_usecount(struct vnode *vp)
1921 {
1922 
1923 	CTR3(KTR_VFS, "v_upgrade_usecount: vp %p holdcnt %d usecount %d\n",
1924 	    vp, vp->v_holdcnt, vp->v_usecount);
1925 	vp->v_usecount++;
1926 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1927 		dev_lock();
1928 		vp->v_rdev->si_usecount++;
1929 		dev_unlock();
1930 	}
1931 }
1932 
1933 /*
1934  * Decrement the vnode use and hold count along with the driver's usecount
1935  * if this is a chardev.  The vdropl() below releases the vnode interlock
1936  * as it may free the vnode.
1937  */
1938 static void
1939 v_decr_usecount(struct vnode *vp)
1940 {
1941 
1942 	CTR3(KTR_VFS, "v_decr_usecount: vp %p holdcnt %d usecount %d\n",
1943 	    vp, vp->v_holdcnt, vp->v_usecount);
1944 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
1945 	VNASSERT(vp->v_usecount > 0, vp,
1946 	    ("v_decr_usecount: negative usecount"));
1947 	vp->v_usecount--;
1948 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1949 		dev_lock();
1950 		vp->v_rdev->si_usecount--;
1951 		dev_unlock();
1952 	}
1953 	vdropl(vp);
1954 }
1955 
1956 /*
1957  * Decrement only the use count and driver use count.  This is intended to
1958  * be paired with a follow on vdropl() to release the remaining hold count.
1959  * In this way we may vgone() a vnode with a 0 usecount without risk of
1960  * having it end up on a free list because the hold count is kept above 0.
1961  */
1962 static void
1963 v_decr_useonly(struct vnode *vp)
1964 {
1965 
1966 	CTR3(KTR_VFS, "v_decr_useonly: vp %p holdcnt %d usecount %d\n",
1967 	    vp, vp->v_holdcnt, vp->v_usecount);
1968 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
1969 	VNASSERT(vp->v_usecount > 0, vp,
1970 	    ("v_decr_useonly: negative usecount"));
1971 	vp->v_usecount--;
1972 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1973 		dev_lock();
1974 		vp->v_rdev->si_usecount--;
1975 		dev_unlock();
1976 	}
1977 }
1978 
1979 /*
1980  * Grab a particular vnode from the free list, increment its
1981  * reference count and lock it. The vnode lock bit is set if the
1982  * vnode is being eliminated in vgone. The process is awakened
1983  * when the transition is completed, and an error returned to
1984  * indicate that the vnode is no longer usable (possibly having
1985  * been changed to a new filesystem type).
1986  */
1987 int
1988 vget(struct vnode *vp, int flags, struct thread *td)
1989 {
1990 	int oweinact;
1991 	int oldflags;
1992 	int error;
1993 
1994 	error = 0;
1995 	oldflags = flags;
1996 	oweinact = 0;
1997 	VFS_ASSERT_GIANT(vp->v_mount);
1998 	if ((flags & LK_INTERLOCK) == 0)
1999 		VI_LOCK(vp);
2000 	/*
2001 	 * If the inactive call was deferred because vput() was called
2002 	 * with a shared lock, we have to do it here before another thread
2003 	 * gets a reference to data that should be dead.
2004 	 */
2005 	if (vp->v_iflag & VI_OWEINACT) {
2006 		if (flags & LK_NOWAIT) {
2007 			VI_UNLOCK(vp);
2008 			return (EBUSY);
2009 		}
2010 		flags &= ~LK_TYPE_MASK;
2011 		flags |= LK_EXCLUSIVE;
2012 		oweinact = 1;
2013 	}
2014 	vholdl(vp);
2015 	if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) {
2016 		vdrop(vp);
2017 		return (error);
2018 	}
2019 	VI_LOCK(vp);
2020 	/* Upgrade our holdcnt to a usecount. */
2021 	v_upgrade_usecount(vp);
2022 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
2023 		panic("vget: vn_lock failed to return ENOENT\n");
2024 	if (oweinact) {
2025 		if (vp->v_iflag & VI_OWEINACT)
2026 			vinactive(vp, td);
2027 		VI_UNLOCK(vp);
2028 		if ((oldflags & LK_TYPE_MASK) == 0)
2029 			VOP_UNLOCK(vp, 0, td);
2030 	} else
2031 		VI_UNLOCK(vp);
2032 	return (0);
2033 }
2034 
2035 /*
2036  * Increase the reference count of a vnode.
2037  */
2038 void
2039 vref(struct vnode *vp)
2040 {
2041 
2042 	VI_LOCK(vp);
2043 	v_incr_usecount(vp);
2044 	VI_UNLOCK(vp);
2045 }
2046 
2047 /*
2048  * Return reference count of a vnode.
2049  *
2050  * The results of this call are only guaranteed when some mechanism other
2051  * than the VI lock is used to stop other processes from gaining references
2052  * to the vnode.  This may be the case if the caller holds the only reference.
2053  * This is also useful when stale data is acceptable as race conditions may
2054  * be accounted for by some other means.
2055  */
2056 int
2057 vrefcnt(struct vnode *vp)
2058 {
2059 	int usecnt;
2060 
2061 	VI_LOCK(vp);
2062 	usecnt = vp->v_usecount;
2063 	VI_UNLOCK(vp);
2064 
2065 	return (usecnt);
2066 }
2067 
2068 
2069 /*
2070  * Vnode put/release.
2071  * If count drops to zero, call inactive routine and return to freelist.
2072  */
2073 void
2074 vrele(struct vnode *vp)
2075 {
2076 	struct thread *td = curthread;	/* XXX */
2077 
2078 	KASSERT(vp != NULL, ("vrele: null vp"));
2079 	VFS_ASSERT_GIANT(vp->v_mount);
2080 
2081 	VI_LOCK(vp);
2082 
2083 	/* Skip this v_writecount check if we're going to panic below. */
2084 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
2085 	    ("vrele: missed vn_close"));
2086 
2087 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2088 	    vp->v_usecount == 1)) {
2089 		v_decr_usecount(vp);
2090 		return;
2091 	}
2092 	if (vp->v_usecount != 1) {
2093 #ifdef DIAGNOSTIC
2094 		vprint("vrele: negative ref count", vp);
2095 #endif
2096 		VI_UNLOCK(vp);
2097 		panic("vrele: negative ref cnt");
2098 	}
2099 	/*
2100 	 * We want to hold the vnode until the inactive finishes to
2101 	 * prevent vgone() races.  We drop the use count here and the
2102 	 * hold count below when we're done.
2103 	 */
2104 	v_decr_useonly(vp);
2105 	/*
2106 	 * We must call VOP_INACTIVE with the node locked. Mark
2107 	 * as VI_DOINGINACT to avoid recursion.
2108 	 */
2109 	vp->v_iflag |= VI_OWEINACT;
2110 	if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) {
2111 		VI_LOCK(vp);
2112 		if (vp->v_usecount > 0)
2113 			vp->v_iflag &= ~VI_OWEINACT;
2114 		if (vp->v_iflag & VI_OWEINACT)
2115 			vinactive(vp, td);
2116 		VOP_UNLOCK(vp, 0, td);
2117 	} else {
2118 		VI_LOCK(vp);
2119 		if (vp->v_usecount > 0)
2120 			vp->v_iflag &= ~VI_OWEINACT;
2121 	}
2122 	vdropl(vp);
2123 }
2124 
2125 /*
2126  * Release an already locked vnode.  This give the same effects as
2127  * unlock+vrele(), but takes less time and avoids releasing and
2128  * re-aquiring the lock (as vrele() aquires the lock internally.)
2129  */
2130 void
2131 vput(struct vnode *vp)
2132 {
2133 	struct thread *td = curthread;	/* XXX */
2134 	int error;
2135 
2136 	KASSERT(vp != NULL, ("vput: null vp"));
2137 	ASSERT_VOP_LOCKED(vp, "vput");
2138 	VFS_ASSERT_GIANT(vp->v_mount);
2139 	VI_LOCK(vp);
2140 	/* Skip this v_writecount check if we're going to panic below. */
2141 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
2142 	    ("vput: missed vn_close"));
2143 	error = 0;
2144 
2145 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2146 	    vp->v_usecount == 1)) {
2147 		VOP_UNLOCK(vp, 0, td);
2148 		v_decr_usecount(vp);
2149 		return;
2150 	}
2151 
2152 	if (vp->v_usecount != 1) {
2153 #ifdef DIAGNOSTIC
2154 		vprint("vput: negative ref count", vp);
2155 #endif
2156 		panic("vput: negative ref cnt");
2157 	}
2158 	/*
2159 	 * We want to hold the vnode until the inactive finishes to
2160 	 * prevent vgone() races.  We drop the use count here and the
2161 	 * hold count below when we're done.
2162 	 */
2163 	v_decr_useonly(vp);
2164 	vp->v_iflag |= VI_OWEINACT;
2165 	if (VOP_ISLOCKED(vp, NULL) != LK_EXCLUSIVE) {
2166 		error = VOP_LOCK(vp, LK_EXCLUPGRADE|LK_INTERLOCK|LK_NOWAIT, td);
2167 		VI_LOCK(vp);
2168 		if (error) {
2169 			if (vp->v_usecount > 0)
2170 				vp->v_iflag &= ~VI_OWEINACT;
2171 			goto done;
2172 		}
2173 	}
2174 	if (vp->v_usecount > 0)
2175 		vp->v_iflag &= ~VI_OWEINACT;
2176 	if (vp->v_iflag & VI_OWEINACT)
2177 		vinactive(vp, td);
2178 	VOP_UNLOCK(vp, 0, td);
2179 done:
2180 	vdropl(vp);
2181 }
2182 
2183 /*
2184  * Somebody doesn't want the vnode recycled.
2185  */
2186 void
2187 vhold(struct vnode *vp)
2188 {
2189 
2190 	VI_LOCK(vp);
2191 	vholdl(vp);
2192 	VI_UNLOCK(vp);
2193 }
2194 
2195 void
2196 vholdl(struct vnode *vp)
2197 {
2198 
2199 	vp->v_holdcnt++;
2200 	if (VSHOULDBUSY(vp))
2201 		vbusy(vp);
2202 }
2203 
2204 /*
2205  * Note that there is one less who cares about this vnode.  vdrop() is the
2206  * opposite of vhold().
2207  */
2208 void
2209 vdrop(struct vnode *vp)
2210 {
2211 
2212 	VI_LOCK(vp);
2213 	vdropl(vp);
2214 }
2215 
2216 /*
2217  * Drop the hold count of the vnode.  If this is the last reference to
2218  * the vnode we will free it if it has been vgone'd otherwise it is
2219  * placed on the free list.
2220  */
2221 static void
2222 vdropl(struct vnode *vp)
2223 {
2224 
2225 	if (vp->v_holdcnt <= 0)
2226 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
2227 	vp->v_holdcnt--;
2228 	if (vp->v_holdcnt == 0) {
2229 		if (vp->v_iflag & VI_DOOMED) {
2230 			vdestroy(vp);
2231 			return;
2232 		} else
2233 			vfree(vp);
2234 	}
2235 	VI_UNLOCK(vp);
2236 }
2237 
2238 /*
2239  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
2240  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
2241  * OWEINACT tracks whether a vnode missed a call to inactive due to a
2242  * failed lock upgrade.
2243  */
2244 static void
2245 vinactive(struct vnode *vp, struct thread *td)
2246 {
2247 
2248 	ASSERT_VOP_LOCKED(vp, "vinactive");
2249 	ASSERT_VI_LOCKED(vp, "vinactive");
2250 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2251 	    ("vinactive: recursed on VI_DOINGINACT"));
2252 	vp->v_iflag |= VI_DOINGINACT;
2253 	vp->v_iflag &= ~VI_OWEINACT;
2254 	VI_UNLOCK(vp);
2255 	VOP_INACTIVE(vp, td);
2256 	VI_LOCK(vp);
2257 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2258 	    ("vinactive: lost VI_DOINGINACT"));
2259 	vp->v_iflag &= ~VI_DOINGINACT;
2260 }
2261 
2262 /*
2263  * Remove any vnodes in the vnode table belonging to mount point mp.
2264  *
2265  * If FORCECLOSE is not specified, there should not be any active ones,
2266  * return error if any are found (nb: this is a user error, not a
2267  * system error). If FORCECLOSE is specified, detach any active vnodes
2268  * that are found.
2269  *
2270  * If WRITECLOSE is set, only flush out regular file vnodes open for
2271  * writing.
2272  *
2273  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2274  *
2275  * `rootrefs' specifies the base reference count for the root vnode
2276  * of this filesystem. The root vnode is considered busy if its
2277  * v_usecount exceeds this value. On a successful return, vflush(, td)
2278  * will call vrele() on the root vnode exactly rootrefs times.
2279  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2280  * be zero.
2281  */
2282 #ifdef DIAGNOSTIC
2283 static int busyprt = 0;		/* print out busy vnodes */
2284 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
2285 #endif
2286 
2287 int
2288 vflush( struct mount *mp, int rootrefs, int flags, struct thread *td)
2289 {
2290 	struct vnode *vp, *mvp, *rootvp = NULL;
2291 	struct vattr vattr;
2292 	int busy = 0, error;
2293 
2294 	CTR1(KTR_VFS, "vflush: mp %p", mp);
2295 	if (rootrefs > 0) {
2296 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2297 		    ("vflush: bad args"));
2298 		/*
2299 		 * Get the filesystem root vnode. We can vput() it
2300 		 * immediately, since with rootrefs > 0, it won't go away.
2301 		 */
2302 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp, td)) != 0)
2303 			return (error);
2304 		vput(rootvp);
2305 
2306 	}
2307 	MNT_ILOCK(mp);
2308 loop:
2309 	MNT_VNODE_FOREACH(vp, mp, mvp) {
2310 
2311 		VI_LOCK(vp);
2312 		vholdl(vp);
2313 		MNT_IUNLOCK(mp);
2314 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td);
2315 		if (error) {
2316 			vdrop(vp);
2317 			MNT_ILOCK(mp);
2318 			MNT_VNODE_FOREACH_ABORT_ILOCKED(mp, mvp);
2319 			goto loop;
2320 		}
2321 		/*
2322 		 * Skip over a vnodes marked VV_SYSTEM.
2323 		 */
2324 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2325 			VOP_UNLOCK(vp, 0, td);
2326 			vdrop(vp);
2327 			MNT_ILOCK(mp);
2328 			continue;
2329 		}
2330 		/*
2331 		 * If WRITECLOSE is set, flush out unlinked but still open
2332 		 * files (even if open only for reading) and regular file
2333 		 * vnodes open for writing.
2334 		 */
2335 		if (flags & WRITECLOSE) {
2336 			error = VOP_GETATTR(vp, &vattr, td->td_ucred, td);
2337 			VI_LOCK(vp);
2338 
2339 			if ((vp->v_type == VNON ||
2340 			    (error == 0 && vattr.va_nlink > 0)) &&
2341 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2342 				VOP_UNLOCK(vp, 0, td);
2343 				vdropl(vp);
2344 				MNT_ILOCK(mp);
2345 				continue;
2346 			}
2347 		} else
2348 			VI_LOCK(vp);
2349 		/*
2350 		 * With v_usecount == 0, all we need to do is clear out the
2351 		 * vnode data structures and we are done.
2352 		 *
2353 		 * If FORCECLOSE is set, forcibly close the vnode.
2354 		 */
2355 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
2356 			VNASSERT(vp->v_usecount == 0 ||
2357 			    (vp->v_type != VCHR && vp->v_type != VBLK), vp,
2358 			    ("device VNODE %p is FORCECLOSED", vp));
2359 			vgonel(vp);
2360 		} else {
2361 			busy++;
2362 #ifdef DIAGNOSTIC
2363 			if (busyprt)
2364 				vprint("vflush: busy vnode", vp);
2365 #endif
2366 		}
2367 		VOP_UNLOCK(vp, 0, td);
2368 		vdropl(vp);
2369 		MNT_ILOCK(mp);
2370 	}
2371 	MNT_IUNLOCK(mp);
2372 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2373 		/*
2374 		 * If just the root vnode is busy, and if its refcount
2375 		 * is equal to `rootrefs', then go ahead and kill it.
2376 		 */
2377 		VI_LOCK(rootvp);
2378 		KASSERT(busy > 0, ("vflush: not busy"));
2379 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
2380 		    ("vflush: usecount %d < rootrefs %d",
2381 		     rootvp->v_usecount, rootrefs));
2382 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2383 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK, td);
2384 			vgone(rootvp);
2385 			VOP_UNLOCK(rootvp, 0, td);
2386 			busy = 0;
2387 		} else
2388 			VI_UNLOCK(rootvp);
2389 	}
2390 	if (busy)
2391 		return (EBUSY);
2392 	for (; rootrefs > 0; rootrefs--)
2393 		vrele(rootvp);
2394 	return (0);
2395 }
2396 
2397 /*
2398  * Recycle an unused vnode to the front of the free list.
2399  */
2400 int
2401 vrecycle(struct vnode *vp, struct thread *td)
2402 {
2403 	int recycled;
2404 
2405 	ASSERT_VOP_LOCKED(vp, "vrecycle");
2406 	recycled = 0;
2407 	VI_LOCK(vp);
2408 	if (vp->v_usecount == 0) {
2409 		recycled = 1;
2410 		vgonel(vp);
2411 	}
2412 	VI_UNLOCK(vp);
2413 	return (recycled);
2414 }
2415 
2416 /*
2417  * Eliminate all activity associated with a vnode
2418  * in preparation for reuse.
2419  */
2420 void
2421 vgone(struct vnode *vp)
2422 {
2423 	VI_LOCK(vp);
2424 	vgonel(vp);
2425 	VI_UNLOCK(vp);
2426 }
2427 
2428 /*
2429  * vgone, with the vp interlock held.
2430  */
2431 void
2432 vgonel(struct vnode *vp)
2433 {
2434 	struct thread *td;
2435 	int oweinact;
2436 	int active;
2437 	struct mount *mp;
2438 
2439 	CTR1(KTR_VFS, "vgonel: vp %p", vp);
2440 	ASSERT_VOP_LOCKED(vp, "vgonel");
2441 	ASSERT_VI_LOCKED(vp, "vgonel");
2442 	VNASSERT(vp->v_holdcnt, vp,
2443 	    ("vgonel: vp %p has no reference.", vp));
2444 	td = curthread;
2445 
2446 	/*
2447 	 * Don't vgonel if we're already doomed.
2448 	 */
2449 	if (vp->v_iflag & VI_DOOMED)
2450 		return;
2451 	vp->v_iflag |= VI_DOOMED;
2452 	/*
2453 	 * Check to see if the vnode is in use.  If so, we have to call
2454 	 * VOP_CLOSE() and VOP_INACTIVE().
2455 	 */
2456 	active = vp->v_usecount;
2457 	oweinact = (vp->v_iflag & VI_OWEINACT);
2458 	VI_UNLOCK(vp);
2459 	/*
2460 	 * Clean out any buffers associated with the vnode.
2461 	 * If the flush fails, just toss the buffers.
2462 	 */
2463 	mp = NULL;
2464 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
2465 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
2466 	if (vinvalbuf(vp, V_SAVE, td, 0, 0) != 0)
2467 		vinvalbuf(vp, 0, td, 0, 0);
2468 
2469 	/*
2470 	 * If purging an active vnode, it must be closed and
2471 	 * deactivated before being reclaimed.
2472 	 */
2473 	if (active)
2474 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
2475 	if (oweinact || active) {
2476 		VI_LOCK(vp);
2477 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
2478 			vinactive(vp, td);
2479 		VI_UNLOCK(vp);
2480 	}
2481 	/*
2482 	 * Reclaim the vnode.
2483 	 */
2484 	if (VOP_RECLAIM(vp, td))
2485 		panic("vgone: cannot reclaim");
2486 	if (mp != NULL)
2487 		vn_finished_secondary_write(mp);
2488 	VNASSERT(vp->v_object == NULL, vp,
2489 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
2490 	/*
2491 	 * Delete from old mount point vnode list.
2492 	 */
2493 	delmntque(vp);
2494 	cache_purge(vp);
2495 	/*
2496 	 * Done with purge, reset to the standard lock and invalidate
2497 	 * the vnode.
2498 	 */
2499 	VI_LOCK(vp);
2500 	vp->v_vnlock = &vp->v_lock;
2501 	vp->v_op = &dead_vnodeops;
2502 	vp->v_tag = "none";
2503 	vp->v_type = VBAD;
2504 }
2505 
2506 /*
2507  * Calculate the total number of references to a special device.
2508  */
2509 int
2510 vcount(struct vnode *vp)
2511 {
2512 	int count;
2513 
2514 	dev_lock();
2515 	count = vp->v_rdev->si_usecount;
2516 	dev_unlock();
2517 	return (count);
2518 }
2519 
2520 /*
2521  * Same as above, but using the struct cdev *as argument
2522  */
2523 int
2524 count_dev(struct cdev *dev)
2525 {
2526 	int count;
2527 
2528 	dev_lock();
2529 	count = dev->si_usecount;
2530 	dev_unlock();
2531 	return(count);
2532 }
2533 
2534 /*
2535  * Print out a description of a vnode.
2536  */
2537 static char *typename[] =
2538 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
2539  "VMARKER"};
2540 
2541 void
2542 vn_printf(struct vnode *vp, const char *fmt, ...)
2543 {
2544 	va_list ap;
2545 	char buf[96];
2546 
2547 	va_start(ap, fmt);
2548 	vprintf(fmt, ap);
2549 	va_end(ap);
2550 	printf("%p: ", (void *)vp);
2551 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
2552 	printf("    usecount %d, writecount %d, refcount %d mountedhere %p\n",
2553 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere);
2554 	buf[0] = '\0';
2555 	buf[1] = '\0';
2556 	if (vp->v_vflag & VV_ROOT)
2557 		strcat(buf, "|VV_ROOT");
2558 	if (vp->v_vflag & VV_TEXT)
2559 		strcat(buf, "|VV_TEXT");
2560 	if (vp->v_vflag & VV_SYSTEM)
2561 		strcat(buf, "|VV_SYSTEM");
2562 	if (vp->v_vflag & VV_DELETED)
2563 		strcat(buf, "|VV_DELETED");
2564 	if (vp->v_iflag & VI_DOOMED)
2565 		strcat(buf, "|VI_DOOMED");
2566 	if (vp->v_iflag & VI_FREE)
2567 		strcat(buf, "|VI_FREE");
2568 	printf("    flags (%s)\n", buf + 1);
2569 	if (mtx_owned(VI_MTX(vp)))
2570 		printf(" VI_LOCKed");
2571 	if (vp->v_object != NULL)
2572 		printf("    v_object %p ref %d pages %d\n",
2573 		    vp->v_object, vp->v_object->ref_count,
2574 		    vp->v_object->resident_page_count);
2575 	printf("    ");
2576 	lockmgr_printinfo(vp->v_vnlock);
2577 	printf("\n");
2578 	if (vp->v_data != NULL)
2579 		VOP_PRINT(vp);
2580 }
2581 
2582 #ifdef DDB
2583 /*
2584  * List all of the locked vnodes in the system.
2585  * Called when debugging the kernel.
2586  */
2587 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
2588 {
2589 	struct mount *mp, *nmp;
2590 	struct vnode *vp;
2591 
2592 	/*
2593 	 * Note: because this is DDB, we can't obey the locking semantics
2594 	 * for these structures, which means we could catch an inconsistent
2595 	 * state and dereference a nasty pointer.  Not much to be done
2596 	 * about that.
2597 	 */
2598 	printf("Locked vnodes\n");
2599 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
2600 		nmp = TAILQ_NEXT(mp, mnt_list);
2601 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2602 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp, NULL))
2603 				vprint("", vp);
2604 		}
2605 		nmp = TAILQ_NEXT(mp, mnt_list);
2606 	}
2607 }
2608 
2609 /*
2610  * Show details about the given vnode.
2611  */
2612 DB_SHOW_COMMAND(vnode, db_show_vnode)
2613 {
2614 	struct vnode *vp;
2615 
2616 	if (!have_addr)
2617 		return;
2618 	vp = (struct vnode *)addr;
2619 	vn_printf(vp, "vnode ");
2620 }
2621 #endif	/* DDB */
2622 
2623 /*
2624  * Fill in a struct xvfsconf based on a struct vfsconf.
2625  */
2626 static void
2627 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp)
2628 {
2629 
2630 	strcpy(xvfsp->vfc_name, vfsp->vfc_name);
2631 	xvfsp->vfc_typenum = vfsp->vfc_typenum;
2632 	xvfsp->vfc_refcount = vfsp->vfc_refcount;
2633 	xvfsp->vfc_flags = vfsp->vfc_flags;
2634 	/*
2635 	 * These are unused in userland, we keep them
2636 	 * to not break binary compatibility.
2637 	 */
2638 	xvfsp->vfc_vfsops = NULL;
2639 	xvfsp->vfc_next = NULL;
2640 }
2641 
2642 /*
2643  * Top level filesystem related information gathering.
2644  */
2645 static int
2646 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
2647 {
2648 	struct vfsconf *vfsp;
2649 	struct xvfsconf xvfsp;
2650 	int error;
2651 
2652 	error = 0;
2653 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2654 		bzero(&xvfsp, sizeof(xvfsp));
2655 		vfsconf2x(vfsp, &xvfsp);
2656 		error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp);
2657 		if (error)
2658 			break;
2659 	}
2660 	return (error);
2661 }
2662 
2663 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist,
2664     "S,xvfsconf", "List of all configured filesystems");
2665 
2666 #ifndef BURN_BRIDGES
2667 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
2668 
2669 static int
2670 vfs_sysctl(SYSCTL_HANDLER_ARGS)
2671 {
2672 	int *name = (int *)arg1 - 1;	/* XXX */
2673 	u_int namelen = arg2 + 1;	/* XXX */
2674 	struct vfsconf *vfsp;
2675 	struct xvfsconf xvfsp;
2676 
2677 	printf("WARNING: userland calling deprecated sysctl, "
2678 	    "please rebuild world\n");
2679 
2680 #if 1 || defined(COMPAT_PRELITE2)
2681 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
2682 	if (namelen == 1)
2683 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
2684 #endif
2685 
2686 	switch (name[1]) {
2687 	case VFS_MAXTYPENUM:
2688 		if (namelen != 2)
2689 			return (ENOTDIR);
2690 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
2691 	case VFS_CONF:
2692 		if (namelen != 3)
2693 			return (ENOTDIR);	/* overloaded */
2694 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list)
2695 			if (vfsp->vfc_typenum == name[2])
2696 				break;
2697 		if (vfsp == NULL)
2698 			return (EOPNOTSUPP);
2699 		bzero(&xvfsp, sizeof(xvfsp));
2700 		vfsconf2x(vfsp, &xvfsp);
2701 		return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
2702 	}
2703 	return (EOPNOTSUPP);
2704 }
2705 
2706 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP,
2707 	vfs_sysctl, "Generic filesystem");
2708 
2709 #if 1 || defined(COMPAT_PRELITE2)
2710 
2711 static int
2712 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
2713 {
2714 	int error;
2715 	struct vfsconf *vfsp;
2716 	struct ovfsconf ovfs;
2717 
2718 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2719 		bzero(&ovfs, sizeof(ovfs));
2720 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
2721 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
2722 		ovfs.vfc_index = vfsp->vfc_typenum;
2723 		ovfs.vfc_refcount = vfsp->vfc_refcount;
2724 		ovfs.vfc_flags = vfsp->vfc_flags;
2725 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
2726 		if (error)
2727 			return error;
2728 	}
2729 	return 0;
2730 }
2731 
2732 #endif /* 1 || COMPAT_PRELITE2 */
2733 #endif /* !BURN_BRIDGES */
2734 
2735 #define KINFO_VNODESLOP		10
2736 #ifdef notyet
2737 /*
2738  * Dump vnode list (via sysctl).
2739  */
2740 /* ARGSUSED */
2741 static int
2742 sysctl_vnode(SYSCTL_HANDLER_ARGS)
2743 {
2744 	struct xvnode *xvn;
2745 	struct thread *td = req->td;
2746 	struct mount *mp;
2747 	struct vnode *vp;
2748 	int error, len, n;
2749 
2750 	/*
2751 	 * Stale numvnodes access is not fatal here.
2752 	 */
2753 	req->lock = 0;
2754 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
2755 	if (!req->oldptr)
2756 		/* Make an estimate */
2757 		return (SYSCTL_OUT(req, 0, len));
2758 
2759 	error = sysctl_wire_old_buffer(req, 0);
2760 	if (error != 0)
2761 		return (error);
2762 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
2763 	n = 0;
2764 	mtx_lock(&mountlist_mtx);
2765 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
2766 		if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td))
2767 			continue;
2768 		MNT_ILOCK(mp);
2769 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2770 			if (n == len)
2771 				break;
2772 			vref(vp);
2773 			xvn[n].xv_size = sizeof *xvn;
2774 			xvn[n].xv_vnode = vp;
2775 			xvn[n].xv_id = 0;	/* XXX compat */
2776 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
2777 			XV_COPY(usecount);
2778 			XV_COPY(writecount);
2779 			XV_COPY(holdcnt);
2780 			XV_COPY(mount);
2781 			XV_COPY(numoutput);
2782 			XV_COPY(type);
2783 #undef XV_COPY
2784 			xvn[n].xv_flag = vp->v_vflag;
2785 
2786 			switch (vp->v_type) {
2787 			case VREG:
2788 			case VDIR:
2789 			case VLNK:
2790 				break;
2791 			case VBLK:
2792 			case VCHR:
2793 				if (vp->v_rdev == NULL) {
2794 					vrele(vp);
2795 					continue;
2796 				}
2797 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
2798 				break;
2799 			case VSOCK:
2800 				xvn[n].xv_socket = vp->v_socket;
2801 				break;
2802 			case VFIFO:
2803 				xvn[n].xv_fifo = vp->v_fifoinfo;
2804 				break;
2805 			case VNON:
2806 			case VBAD:
2807 			default:
2808 				/* shouldn't happen? */
2809 				vrele(vp);
2810 				continue;
2811 			}
2812 			vrele(vp);
2813 			++n;
2814 		}
2815 		MNT_IUNLOCK(mp);
2816 		mtx_lock(&mountlist_mtx);
2817 		vfs_unbusy(mp, td);
2818 		if (n == len)
2819 			break;
2820 	}
2821 	mtx_unlock(&mountlist_mtx);
2822 
2823 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
2824 	free(xvn, M_TEMP);
2825 	return (error);
2826 }
2827 
2828 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD,
2829 	0, 0, sysctl_vnode, "S,xvnode", "");
2830 #endif
2831 
2832 /*
2833  * Unmount all filesystems. The list is traversed in reverse order
2834  * of mounting to avoid dependencies.
2835  */
2836 void
2837 vfs_unmountall(void)
2838 {
2839 	struct mount *mp;
2840 	struct thread *td;
2841 	int error;
2842 
2843 	KASSERT(curthread != NULL, ("vfs_unmountall: NULL curthread"));
2844 	td = curthread;
2845 	/*
2846 	 * Since this only runs when rebooting, it is not interlocked.
2847 	 */
2848 	while(!TAILQ_EMPTY(&mountlist)) {
2849 		mp = TAILQ_LAST(&mountlist, mntlist);
2850 		error = dounmount(mp, MNT_FORCE, td);
2851 		if (error) {
2852 			TAILQ_REMOVE(&mountlist, mp, mnt_list);
2853 			/*
2854 			 * XXX: Due to the way in which we mount the root
2855 			 * file system off of devfs, devfs will generate a
2856 			 * "busy" warning when we try to unmount it before
2857 			 * the root.  Don't print a warning as a result in
2858 			 * order to avoid false positive errors that may
2859 			 * cause needless upset.
2860 			 */
2861 			if (strcmp(mp->mnt_vfc->vfc_name, "devfs") != 0) {
2862 				printf("unmount of %s failed (",
2863 				    mp->mnt_stat.f_mntonname);
2864 				if (error == EBUSY)
2865 					printf("BUSY)\n");
2866 				else
2867 					printf("%d)\n", error);
2868 			}
2869 		} else {
2870 			/* The unmount has removed mp from the mountlist */
2871 		}
2872 	}
2873 }
2874 
2875 /*
2876  * perform msync on all vnodes under a mount point
2877  * the mount point must be locked.
2878  */
2879 void
2880 vfs_msync(struct mount *mp, int flags)
2881 {
2882 	struct vnode *vp, *mvp;
2883 	struct vm_object *obj;
2884 
2885 	MNT_ILOCK(mp);
2886 	MNT_VNODE_FOREACH(vp, mp, mvp) {
2887 		VI_LOCK(vp);
2888 		if ((vp->v_iflag & VI_OBJDIRTY) &&
2889 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) {
2890 			MNT_IUNLOCK(mp);
2891 			if (!vget(vp,
2892 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
2893 			    curthread)) {
2894 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
2895 					vput(vp);
2896 					MNT_ILOCK(mp);
2897 					continue;
2898 				}
2899 
2900 				obj = vp->v_object;
2901 				if (obj != NULL) {
2902 					VM_OBJECT_LOCK(obj);
2903 					vm_object_page_clean(obj, 0, 0,
2904 					    flags == MNT_WAIT ?
2905 					    OBJPC_SYNC : OBJPC_NOSYNC);
2906 					VM_OBJECT_UNLOCK(obj);
2907 				}
2908 				vput(vp);
2909 			}
2910 			MNT_ILOCK(mp);
2911 		} else
2912 			VI_UNLOCK(vp);
2913 	}
2914 	MNT_IUNLOCK(mp);
2915 }
2916 
2917 /*
2918  * Mark a vnode as free, putting it up for recycling.
2919  */
2920 static void
2921 vfree(struct vnode *vp)
2922 {
2923 
2924 	CTR1(KTR_VFS, "vfree vp %p", vp);
2925 	ASSERT_VI_LOCKED(vp, "vfree");
2926 	mtx_lock(&vnode_free_list_mtx);
2927 	VNASSERT(vp->v_op != NULL, vp, ("vfree: vnode already reclaimed."));
2928 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, ("vnode already free"));
2929 	VNASSERT(VSHOULDFREE(vp), vp, ("vfree: freeing when we shouldn't"));
2930 	VNASSERT((vp->v_iflag & VI_DOOMED) == 0, vp,
2931 	    ("vfree: Freeing doomed vnode"));
2932 	if (vp->v_iflag & VI_AGE) {
2933 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2934 	} else {
2935 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
2936 	}
2937 	freevnodes++;
2938 	vp->v_iflag &= ~VI_AGE;
2939 	vp->v_iflag |= VI_FREE;
2940 	mtx_unlock(&vnode_free_list_mtx);
2941 }
2942 
2943 /*
2944  * Opposite of vfree() - mark a vnode as in use.
2945  */
2946 static void
2947 vbusy(struct vnode *vp)
2948 {
2949 	CTR1(KTR_VFS, "vbusy vp %p", vp);
2950 	ASSERT_VI_LOCKED(vp, "vbusy");
2951 	VNASSERT((vp->v_iflag & VI_FREE) != 0, vp, ("vnode not free"));
2952 	VNASSERT(vp->v_op != NULL, vp, ("vbusy: vnode already reclaimed."));
2953 
2954 	mtx_lock(&vnode_free_list_mtx);
2955 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2956 	freevnodes--;
2957 	vp->v_iflag &= ~(VI_FREE|VI_AGE);
2958 	mtx_unlock(&vnode_free_list_mtx);
2959 }
2960 
2961 /*
2962  * Initalize per-vnode helper structure to hold poll-related state.
2963  */
2964 void
2965 v_addpollinfo(struct vnode *vp)
2966 {
2967 	struct vpollinfo *vi;
2968 
2969 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK);
2970 	if (vp->v_pollinfo != NULL) {
2971 		uma_zfree(vnodepoll_zone, vi);
2972 		return;
2973 	}
2974 	vp->v_pollinfo = vi;
2975 	mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
2976 	knlist_init(&vp->v_pollinfo->vpi_selinfo.si_note, vp, vfs_knllock,
2977 	    vfs_knlunlock, vfs_knllocked);
2978 }
2979 
2980 /*
2981  * Record a process's interest in events which might happen to
2982  * a vnode.  Because poll uses the historic select-style interface
2983  * internally, this routine serves as both the ``check for any
2984  * pending events'' and the ``record my interest in future events''
2985  * functions.  (These are done together, while the lock is held,
2986  * to avoid race conditions.)
2987  */
2988 int
2989 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
2990 {
2991 
2992 	if (vp->v_pollinfo == NULL)
2993 		v_addpollinfo(vp);
2994 	mtx_lock(&vp->v_pollinfo->vpi_lock);
2995 	if (vp->v_pollinfo->vpi_revents & events) {
2996 		/*
2997 		 * This leaves events we are not interested
2998 		 * in available for the other process which
2999 		 * which presumably had requested them
3000 		 * (otherwise they would never have been
3001 		 * recorded).
3002 		 */
3003 		events &= vp->v_pollinfo->vpi_revents;
3004 		vp->v_pollinfo->vpi_revents &= ~events;
3005 
3006 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
3007 		return events;
3008 	}
3009 	vp->v_pollinfo->vpi_events |= events;
3010 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
3011 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3012 	return 0;
3013 }
3014 
3015 /*
3016  * Routine to create and manage a filesystem syncer vnode.
3017  */
3018 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
3019 static int	sync_fsync(struct  vop_fsync_args *);
3020 static int	sync_inactive(struct  vop_inactive_args *);
3021 static int	sync_reclaim(struct  vop_reclaim_args *);
3022 
3023 static struct vop_vector sync_vnodeops = {
3024 	.vop_bypass =	VOP_EOPNOTSUPP,
3025 	.vop_close =	sync_close,		/* close */
3026 	.vop_fsync =	sync_fsync,		/* fsync */
3027 	.vop_inactive =	sync_inactive,	/* inactive */
3028 	.vop_reclaim =	sync_reclaim,	/* reclaim */
3029 	._vop_lock =	vop_stdlock,	/* lock */
3030 	.vop_unlock =	vop_stdunlock,	/* unlock */
3031 	.vop_islocked =	vop_stdislocked,	/* islocked */
3032 };
3033 
3034 /*
3035  * Create a new filesystem syncer vnode for the specified mount point.
3036  */
3037 int
3038 vfs_allocate_syncvnode(struct mount *mp)
3039 {
3040 	struct vnode *vp;
3041 	static long start, incr, next;
3042 	int error;
3043 
3044 	/* Allocate a new vnode */
3045 	if ((error = getnewvnode("syncer", mp, &sync_vnodeops, &vp)) != 0) {
3046 		mp->mnt_syncer = NULL;
3047 		return (error);
3048 	}
3049 	vp->v_type = VNON;
3050 	error = insmntque(vp, mp);
3051 	if (error != 0)
3052 		panic("vfs_allocate_syncvnode: insmntque failed");
3053 	/*
3054 	 * Place the vnode onto the syncer worklist. We attempt to
3055 	 * scatter them about on the list so that they will go off
3056 	 * at evenly distributed times even if all the filesystems
3057 	 * are mounted at once.
3058 	 */
3059 	next += incr;
3060 	if (next == 0 || next > syncer_maxdelay) {
3061 		start /= 2;
3062 		incr /= 2;
3063 		if (start == 0) {
3064 			start = syncer_maxdelay / 2;
3065 			incr = syncer_maxdelay;
3066 		}
3067 		next = start;
3068 	}
3069 	VI_LOCK(vp);
3070 	vn_syncer_add_to_worklist(&vp->v_bufobj,
3071 	    syncdelay > 0 ? next % syncdelay : 0);
3072 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
3073 	mtx_lock(&sync_mtx);
3074 	sync_vnode_count++;
3075 	mtx_unlock(&sync_mtx);
3076 	VI_UNLOCK(vp);
3077 	mp->mnt_syncer = vp;
3078 	return (0);
3079 }
3080 
3081 /*
3082  * Do a lazy sync of the filesystem.
3083  */
3084 static int
3085 sync_fsync(struct vop_fsync_args *ap)
3086 {
3087 	struct vnode *syncvp = ap->a_vp;
3088 	struct mount *mp = syncvp->v_mount;
3089 	struct thread *td = ap->a_td;
3090 	int error;
3091 	struct bufobj *bo;
3092 
3093 	/*
3094 	 * We only need to do something if this is a lazy evaluation.
3095 	 */
3096 	if (ap->a_waitfor != MNT_LAZY)
3097 		return (0);
3098 
3099 	/*
3100 	 * Move ourselves to the back of the sync list.
3101 	 */
3102 	bo = &syncvp->v_bufobj;
3103 	BO_LOCK(bo);
3104 	vn_syncer_add_to_worklist(bo, syncdelay);
3105 	BO_UNLOCK(bo);
3106 
3107 	/*
3108 	 * Walk the list of vnodes pushing all that are dirty and
3109 	 * not already on the sync list.
3110 	 */
3111 	mtx_lock(&mountlist_mtx);
3112 	if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) {
3113 		mtx_unlock(&mountlist_mtx);
3114 		return (0);
3115 	}
3116 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
3117 		vfs_unbusy(mp, td);
3118 		return (0);
3119 	}
3120 	MNT_ILOCK(mp);
3121 	mp->mnt_noasync++;
3122 	mp->mnt_kern_flag &= ~MNTK_ASYNC;
3123 	MNT_IUNLOCK(mp);
3124 	vfs_msync(mp, MNT_NOWAIT);
3125 	error = VFS_SYNC(mp, MNT_LAZY, td);
3126 	MNT_ILOCK(mp);
3127 	mp->mnt_noasync--;
3128 	if ((mp->mnt_flag & MNT_ASYNC) != 0 && mp->mnt_noasync == 0)
3129 		mp->mnt_kern_flag |= MNTK_ASYNC;
3130 	MNT_IUNLOCK(mp);
3131 	vn_finished_write(mp);
3132 	vfs_unbusy(mp, td);
3133 	return (error);
3134 }
3135 
3136 /*
3137  * The syncer vnode is no referenced.
3138  */
3139 static int
3140 sync_inactive(struct vop_inactive_args *ap)
3141 {
3142 
3143 	vgone(ap->a_vp);
3144 	return (0);
3145 }
3146 
3147 /*
3148  * The syncer vnode is no longer needed and is being decommissioned.
3149  *
3150  * Modifications to the worklist must be protected by sync_mtx.
3151  */
3152 static int
3153 sync_reclaim(struct vop_reclaim_args *ap)
3154 {
3155 	struct vnode *vp = ap->a_vp;
3156 	struct bufobj *bo;
3157 
3158 	VI_LOCK(vp);
3159 	bo = &vp->v_bufobj;
3160 	vp->v_mount->mnt_syncer = NULL;
3161 	if (bo->bo_flag & BO_ONWORKLST) {
3162 		mtx_lock(&sync_mtx);
3163 		LIST_REMOVE(bo, bo_synclist);
3164  		syncer_worklist_len--;
3165 		sync_vnode_count--;
3166 		mtx_unlock(&sync_mtx);
3167 		bo->bo_flag &= ~BO_ONWORKLST;
3168 	}
3169 	VI_UNLOCK(vp);
3170 
3171 	return (0);
3172 }
3173 
3174 /*
3175  * Check if vnode represents a disk device
3176  */
3177 int
3178 vn_isdisk(struct vnode *vp, int *errp)
3179 {
3180 	int error;
3181 
3182 	error = 0;
3183 	dev_lock();
3184 	if (vp->v_type != VCHR)
3185 		error = ENOTBLK;
3186 	else if (vp->v_rdev == NULL)
3187 		error = ENXIO;
3188 	else if (vp->v_rdev->si_devsw == NULL)
3189 		error = ENXIO;
3190 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
3191 		error = ENOTBLK;
3192 	dev_unlock();
3193 	if (errp != NULL)
3194 		*errp = error;
3195 	return (error == 0);
3196 }
3197 
3198 /*
3199  * Common filesystem object access control check routine.  Accepts a
3200  * vnode's type, "mode", uid and gid, requested access mode, credentials,
3201  * and optional call-by-reference privused argument allowing vaccess()
3202  * to indicate to the caller whether privilege was used to satisfy the
3203  * request (obsoleted).  Returns 0 on success, or an errno on failure.
3204  *
3205  * The ifdef'd CAPABILITIES version is here for reference, but is not
3206  * actually used.
3207  */
3208 int
3209 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
3210     mode_t acc_mode, struct ucred *cred, int *privused)
3211 {
3212 	mode_t dac_granted;
3213 	mode_t priv_granted;
3214 
3215 	/*
3216 	 * Look for a normal, non-privileged way to access the file/directory
3217 	 * as requested.  If it exists, go with that.
3218 	 */
3219 
3220 	if (privused != NULL)
3221 		*privused = 0;
3222 
3223 	dac_granted = 0;
3224 
3225 	/* Check the owner. */
3226 	if (cred->cr_uid == file_uid) {
3227 		dac_granted |= VADMIN;
3228 		if (file_mode & S_IXUSR)
3229 			dac_granted |= VEXEC;
3230 		if (file_mode & S_IRUSR)
3231 			dac_granted |= VREAD;
3232 		if (file_mode & S_IWUSR)
3233 			dac_granted |= (VWRITE | VAPPEND);
3234 
3235 		if ((acc_mode & dac_granted) == acc_mode)
3236 			return (0);
3237 
3238 		goto privcheck;
3239 	}
3240 
3241 	/* Otherwise, check the groups (first match) */
3242 	if (groupmember(file_gid, cred)) {
3243 		if (file_mode & S_IXGRP)
3244 			dac_granted |= VEXEC;
3245 		if (file_mode & S_IRGRP)
3246 			dac_granted |= VREAD;
3247 		if (file_mode & S_IWGRP)
3248 			dac_granted |= (VWRITE | VAPPEND);
3249 
3250 		if ((acc_mode & dac_granted) == acc_mode)
3251 			return (0);
3252 
3253 		goto privcheck;
3254 	}
3255 
3256 	/* Otherwise, check everyone else. */
3257 	if (file_mode & S_IXOTH)
3258 		dac_granted |= VEXEC;
3259 	if (file_mode & S_IROTH)
3260 		dac_granted |= VREAD;
3261 	if (file_mode & S_IWOTH)
3262 		dac_granted |= (VWRITE | VAPPEND);
3263 	if ((acc_mode & dac_granted) == acc_mode)
3264 		return (0);
3265 
3266 privcheck:
3267 	/*
3268 	 * Build a privilege mask to determine if the set of privileges
3269 	 * satisfies the requirements when combined with the granted mask
3270 	 * from above.  For each privilege, if the privilege is required,
3271 	 * bitwise or the request type onto the priv_granted mask.
3272 	 */
3273 	priv_granted = 0;
3274 
3275 	if (type == VDIR) {
3276 		/*
3277 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
3278 		 * requests, instead of PRIV_VFS_EXEC.
3279 		 */
3280 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3281 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP, SUSER_ALLOWJAIL))
3282 			priv_granted |= VEXEC;
3283 	} else {
3284 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3285 		    !priv_check_cred(cred, PRIV_VFS_EXEC, SUSER_ALLOWJAIL))
3286 			priv_granted |= VEXEC;
3287 	}
3288 
3289 	if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) &&
3290 	    !priv_check_cred(cred, PRIV_VFS_READ, SUSER_ALLOWJAIL))
3291 		priv_granted |= VREAD;
3292 
3293 	if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
3294 	    !priv_check_cred(cred, PRIV_VFS_WRITE, SUSER_ALLOWJAIL))
3295 		priv_granted |= (VWRITE | VAPPEND);
3296 
3297 	if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
3298 	    !priv_check_cred(cred, PRIV_VFS_ADMIN, SUSER_ALLOWJAIL))
3299 		priv_granted |= VADMIN;
3300 
3301 	if ((acc_mode & (priv_granted | dac_granted)) == acc_mode) {
3302 		/* XXX audit: privilege used */
3303 		if (privused != NULL)
3304 			*privused = 1;
3305 		return (0);
3306 	}
3307 
3308 	return ((acc_mode & VADMIN) ? EPERM : EACCES);
3309 }
3310 
3311 /*
3312  * Credential check based on process requesting service, and per-attribute
3313  * permissions.
3314  */
3315 int
3316 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
3317     struct thread *td, int access)
3318 {
3319 
3320 	/*
3321 	 * Kernel-invoked always succeeds.
3322 	 */
3323 	if (cred == NOCRED)
3324 		return (0);
3325 
3326 	/*
3327 	 * Do not allow privileged processes in jail to directly manipulate
3328 	 * system attributes.
3329 	 */
3330 	switch (attrnamespace) {
3331 	case EXTATTR_NAMESPACE_SYSTEM:
3332 		/* Potentially should be: return (EPERM); */
3333 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0));
3334 	case EXTATTR_NAMESPACE_USER:
3335 		return (VOP_ACCESS(vp, access, cred, td));
3336 	default:
3337 		return (EPERM);
3338 	}
3339 }
3340 
3341 #ifdef DEBUG_VFS_LOCKS
3342 /*
3343  * This only exists to supress warnings from unlocked specfs accesses.  It is
3344  * no longer ok to have an unlocked VFS.
3345  */
3346 #define	IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD)
3347 
3348 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
3349 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, "");
3350 
3351 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
3352 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, "");
3353 
3354 int vfs_badlock_print = 1;	/* Print lock violations. */
3355 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, "");
3356 
3357 #ifdef KDB
3358 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
3359 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, "");
3360 #endif
3361 
3362 static void
3363 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
3364 {
3365 
3366 #ifdef KDB
3367 	if (vfs_badlock_backtrace)
3368 		kdb_backtrace();
3369 #endif
3370 	if (vfs_badlock_print)
3371 		printf("%s: %p %s\n", str, (void *)vp, msg);
3372 	if (vfs_badlock_ddb)
3373 		kdb_enter("lock violation");
3374 }
3375 
3376 void
3377 assert_vi_locked(struct vnode *vp, const char *str)
3378 {
3379 
3380 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
3381 		vfs_badlock("interlock is not locked but should be", str, vp);
3382 }
3383 
3384 void
3385 assert_vi_unlocked(struct vnode *vp, const char *str)
3386 {
3387 
3388 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
3389 		vfs_badlock("interlock is locked but should not be", str, vp);
3390 }
3391 
3392 void
3393 assert_vop_locked(struct vnode *vp, const char *str)
3394 {
3395 
3396 	if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0)
3397 		vfs_badlock("is not locked but should be", str, vp);
3398 }
3399 
3400 void
3401 assert_vop_unlocked(struct vnode *vp, const char *str)
3402 {
3403 
3404 	if (vp && !IGNORE_LOCK(vp) &&
3405 	    VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE)
3406 		vfs_badlock("is locked but should not be", str, vp);
3407 }
3408 
3409 void
3410 assert_vop_elocked(struct vnode *vp, const char *str)
3411 {
3412 
3413 	if (vp && !IGNORE_LOCK(vp) &&
3414 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE)
3415 		vfs_badlock("is not exclusive locked but should be", str, vp);
3416 }
3417 
3418 #if 0
3419 void
3420 assert_vop_elocked_other(struct vnode *vp, const char *str)
3421 {
3422 
3423 	if (vp && !IGNORE_LOCK(vp) &&
3424 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER)
3425 		vfs_badlock("is not exclusive locked by another thread",
3426 		    str, vp);
3427 }
3428 
3429 void
3430 assert_vop_slocked(struct vnode *vp, const char *str)
3431 {
3432 
3433 	if (vp && !IGNORE_LOCK(vp) &&
3434 	    VOP_ISLOCKED(vp, curthread) != LK_SHARED)
3435 		vfs_badlock("is not locked shared but should be", str, vp);
3436 }
3437 #endif /* 0 */
3438 #endif /* DEBUG_VFS_LOCKS */
3439 
3440 void
3441 vop_rename_pre(void *ap)
3442 {
3443 	struct vop_rename_args *a = ap;
3444 
3445 #ifdef DEBUG_VFS_LOCKS
3446 	if (a->a_tvp)
3447 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
3448 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
3449 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
3450 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
3451 
3452 	/* Check the source (from). */
3453 	if (a->a_tdvp != a->a_fdvp && a->a_tvp != a->a_fdvp)
3454 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
3455 	if (a->a_tvp != a->a_fvp)
3456 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
3457 
3458 	/* Check the target. */
3459 	if (a->a_tvp)
3460 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
3461 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
3462 #endif
3463 	if (a->a_tdvp != a->a_fdvp)
3464 		vhold(a->a_fdvp);
3465 	if (a->a_tvp != a->a_fvp)
3466 		vhold(a->a_fvp);
3467 	vhold(a->a_tdvp);
3468 	if (a->a_tvp)
3469 		vhold(a->a_tvp);
3470 }
3471 
3472 void
3473 vop_strategy_pre(void *ap)
3474 {
3475 #ifdef DEBUG_VFS_LOCKS
3476 	struct vop_strategy_args *a;
3477 	struct buf *bp;
3478 
3479 	a = ap;
3480 	bp = a->a_bp;
3481 
3482 	/*
3483 	 * Cluster ops lock their component buffers but not the IO container.
3484 	 */
3485 	if ((bp->b_flags & B_CLUSTER) != 0)
3486 		return;
3487 
3488 	if (BUF_REFCNT(bp) < 1) {
3489 		if (vfs_badlock_print)
3490 			printf(
3491 			    "VOP_STRATEGY: bp is not locked but should be\n");
3492 		if (vfs_badlock_ddb)
3493 			kdb_enter("lock violation");
3494 	}
3495 #endif
3496 }
3497 
3498 void
3499 vop_lookup_pre(void *ap)
3500 {
3501 #ifdef DEBUG_VFS_LOCKS
3502 	struct vop_lookup_args *a;
3503 	struct vnode *dvp;
3504 
3505 	a = ap;
3506 	dvp = a->a_dvp;
3507 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3508 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3509 #endif
3510 }
3511 
3512 void
3513 vop_lookup_post(void *ap, int rc)
3514 {
3515 #ifdef DEBUG_VFS_LOCKS
3516 	struct vop_lookup_args *a;
3517 	struct vnode *dvp;
3518 	struct vnode *vp;
3519 
3520 	a = ap;
3521 	dvp = a->a_dvp;
3522 	vp = *(a->a_vpp);
3523 
3524 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3525 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3526 
3527 	if (!rc)
3528 		ASSERT_VOP_LOCKED(vp, "VOP_LOOKUP (child)");
3529 #endif
3530 }
3531 
3532 void
3533 vop_lock_pre(void *ap)
3534 {
3535 #ifdef DEBUG_VFS_LOCKS
3536 	struct _vop_lock_args *a = ap;
3537 
3538 	if ((a->a_flags & LK_INTERLOCK) == 0)
3539 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3540 	else
3541 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
3542 #endif
3543 }
3544 
3545 void
3546 vop_lock_post(void *ap, int rc)
3547 {
3548 #ifdef DEBUG_VFS_LOCKS
3549 	struct _vop_lock_args *a = ap;
3550 
3551 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3552 	if (rc == 0)
3553 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
3554 #endif
3555 }
3556 
3557 void
3558 vop_unlock_pre(void *ap)
3559 {
3560 #ifdef DEBUG_VFS_LOCKS
3561 	struct vop_unlock_args *a = ap;
3562 
3563 	if (a->a_flags & LK_INTERLOCK)
3564 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
3565 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
3566 #endif
3567 }
3568 
3569 void
3570 vop_unlock_post(void *ap, int rc)
3571 {
3572 #ifdef DEBUG_VFS_LOCKS
3573 	struct vop_unlock_args *a = ap;
3574 
3575 	if (a->a_flags & LK_INTERLOCK)
3576 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
3577 #endif
3578 }
3579 
3580 void
3581 vop_create_post(void *ap, int rc)
3582 {
3583 	struct vop_create_args *a = ap;
3584 
3585 	if (!rc)
3586 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3587 }
3588 
3589 void
3590 vop_link_post(void *ap, int rc)
3591 {
3592 	struct vop_link_args *a = ap;
3593 
3594 	if (!rc) {
3595 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
3596 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
3597 	}
3598 }
3599 
3600 void
3601 vop_mkdir_post(void *ap, int rc)
3602 {
3603 	struct vop_mkdir_args *a = ap;
3604 
3605 	if (!rc)
3606 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
3607 }
3608 
3609 void
3610 vop_mknod_post(void *ap, int rc)
3611 {
3612 	struct vop_mknod_args *a = ap;
3613 
3614 	if (!rc)
3615 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3616 }
3617 
3618 void
3619 vop_remove_post(void *ap, int rc)
3620 {
3621 	struct vop_remove_args *a = ap;
3622 
3623 	if (!rc) {
3624 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3625 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
3626 	}
3627 }
3628 
3629 void
3630 vop_rename_post(void *ap, int rc)
3631 {
3632 	struct vop_rename_args *a = ap;
3633 
3634 	if (!rc) {
3635 		VFS_KNOTE_UNLOCKED(a->a_fdvp, NOTE_WRITE);
3636 		VFS_KNOTE_UNLOCKED(a->a_tdvp, NOTE_WRITE);
3637 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
3638 		if (a->a_tvp)
3639 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
3640 	}
3641 	if (a->a_tdvp != a->a_fdvp)
3642 		vdrop(a->a_fdvp);
3643 	if (a->a_tvp != a->a_fvp)
3644 		vdrop(a->a_fvp);
3645 	vdrop(a->a_tdvp);
3646 	if (a->a_tvp)
3647 		vdrop(a->a_tvp);
3648 }
3649 
3650 void
3651 vop_rmdir_post(void *ap, int rc)
3652 {
3653 	struct vop_rmdir_args *a = ap;
3654 
3655 	if (!rc) {
3656 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
3657 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
3658 	}
3659 }
3660 
3661 void
3662 vop_setattr_post(void *ap, int rc)
3663 {
3664 	struct vop_setattr_args *a = ap;
3665 
3666 	if (!rc)
3667 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
3668 }
3669 
3670 void
3671 vop_symlink_post(void *ap, int rc)
3672 {
3673 	struct vop_symlink_args *a = ap;
3674 
3675 	if (!rc)
3676 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3677 }
3678 
3679 static struct knlist fs_knlist;
3680 
3681 static void
3682 vfs_event_init(void *arg)
3683 {
3684 	knlist_init(&fs_knlist, NULL, NULL, NULL, NULL);
3685 }
3686 /* XXX - correct order? */
3687 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
3688 
3689 void
3690 vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused)
3691 {
3692 
3693 	KNOTE_UNLOCKED(&fs_knlist, event);
3694 }
3695 
3696 static int	filt_fsattach(struct knote *kn);
3697 static void	filt_fsdetach(struct knote *kn);
3698 static int	filt_fsevent(struct knote *kn, long hint);
3699 
3700 struct filterops fs_filtops =
3701 	{ 0, filt_fsattach, filt_fsdetach, filt_fsevent };
3702 
3703 static int
3704 filt_fsattach(struct knote *kn)
3705 {
3706 
3707 	kn->kn_flags |= EV_CLEAR;
3708 	knlist_add(&fs_knlist, kn, 0);
3709 	return (0);
3710 }
3711 
3712 static void
3713 filt_fsdetach(struct knote *kn)
3714 {
3715 
3716 	knlist_remove(&fs_knlist, kn, 0);
3717 }
3718 
3719 static int
3720 filt_fsevent(struct knote *kn, long hint)
3721 {
3722 
3723 	kn->kn_fflags |= hint;
3724 	return (kn->kn_fflags != 0);
3725 }
3726 
3727 static int
3728 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
3729 {
3730 	struct vfsidctl vc;
3731 	int error;
3732 	struct mount *mp;
3733 
3734 	error = SYSCTL_IN(req, &vc, sizeof(vc));
3735 	if (error)
3736 		return (error);
3737 	if (vc.vc_vers != VFS_CTL_VERS1)
3738 		return (EINVAL);
3739 	mp = vfs_getvfs(&vc.vc_fsid);
3740 	if (mp == NULL)
3741 		return (ENOENT);
3742 	/* ensure that a specific sysctl goes to the right filesystem. */
3743 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
3744 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
3745 		vfs_rel(mp);
3746 		return (EINVAL);
3747 	}
3748 	VCTLTOREQ(&vc, req);
3749 	error = VFS_SYSCTL(mp, vc.vc_op, req);
3750 	vfs_rel(mp);
3751 	return (error);
3752 }
3753 
3754 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR,
3755         NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid");
3756 
3757 /*
3758  * Function to initialize a va_filerev field sensibly.
3759  * XXX: Wouldn't a random number make a lot more sense ??
3760  */
3761 u_quad_t
3762 init_va_filerev(void)
3763 {
3764 	struct bintime bt;
3765 
3766 	getbinuptime(&bt);
3767 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
3768 }
3769 
3770 static int	filt_vfsread(struct knote *kn, long hint);
3771 static int	filt_vfswrite(struct knote *kn, long hint);
3772 static int	filt_vfsvnode(struct knote *kn, long hint);
3773 static void	filt_vfsdetach(struct knote *kn);
3774 static struct filterops vfsread_filtops =
3775 	{ 1, NULL, filt_vfsdetach, filt_vfsread };
3776 static struct filterops vfswrite_filtops =
3777 	{ 1, NULL, filt_vfsdetach, filt_vfswrite };
3778 static struct filterops vfsvnode_filtops =
3779 	{ 1, NULL, filt_vfsdetach, filt_vfsvnode };
3780 
3781 static void
3782 vfs_knllock(void *arg)
3783 {
3784 	struct vnode *vp = arg;
3785 
3786 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, curthread);
3787 }
3788 
3789 static void
3790 vfs_knlunlock(void *arg)
3791 {
3792 	struct vnode *vp = arg;
3793 
3794 	VOP_UNLOCK(vp, 0, curthread);
3795 }
3796 
3797 static int
3798 vfs_knllocked(void *arg)
3799 {
3800 	struct vnode *vp = arg;
3801 
3802 	return (VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE);
3803 }
3804 
3805 int
3806 vfs_kqfilter(struct vop_kqfilter_args *ap)
3807 {
3808 	struct vnode *vp = ap->a_vp;
3809 	struct knote *kn = ap->a_kn;
3810 	struct knlist *knl;
3811 
3812 	switch (kn->kn_filter) {
3813 	case EVFILT_READ:
3814 		kn->kn_fop = &vfsread_filtops;
3815 		break;
3816 	case EVFILT_WRITE:
3817 		kn->kn_fop = &vfswrite_filtops;
3818 		break;
3819 	case EVFILT_VNODE:
3820 		kn->kn_fop = &vfsvnode_filtops;
3821 		break;
3822 	default:
3823 		return (EINVAL);
3824 	}
3825 
3826 	kn->kn_hook = (caddr_t)vp;
3827 
3828 	if (vp->v_pollinfo == NULL)
3829 		v_addpollinfo(vp);
3830 	if (vp->v_pollinfo == NULL)
3831 		return (ENOMEM);
3832 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
3833 	knlist_add(knl, kn, 0);
3834 
3835 	return (0);
3836 }
3837 
3838 /*
3839  * Detach knote from vnode
3840  */
3841 static void
3842 filt_vfsdetach(struct knote *kn)
3843 {
3844 	struct vnode *vp = (struct vnode *)kn->kn_hook;
3845 
3846 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
3847 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
3848 }
3849 
3850 /*ARGSUSED*/
3851 static int
3852 filt_vfsread(struct knote *kn, long hint)
3853 {
3854 	struct vnode *vp = (struct vnode *)kn->kn_hook;
3855 	struct vattr va;
3856 
3857 	/*
3858 	 * filesystem is gone, so set the EOF flag and schedule
3859 	 * the knote for deletion.
3860 	 */
3861 	if (hint == NOTE_REVOKE) {
3862 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3863 		return (1);
3864 	}
3865 
3866 	if (VOP_GETATTR(vp, &va, curthread->td_ucred, curthread))
3867 		return (0);
3868 
3869 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
3870 	return (kn->kn_data != 0);
3871 }
3872 
3873 /*ARGSUSED*/
3874 static int
3875 filt_vfswrite(struct knote *kn, long hint)
3876 {
3877 	/*
3878 	 * filesystem is gone, so set the EOF flag and schedule
3879 	 * the knote for deletion.
3880 	 */
3881 	if (hint == NOTE_REVOKE)
3882 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3883 
3884 	kn->kn_data = 0;
3885 	return (1);
3886 }
3887 
3888 static int
3889 filt_vfsvnode(struct knote *kn, long hint)
3890 {
3891 	if (kn->kn_sfflags & hint)
3892 		kn->kn_fflags |= hint;
3893 	if (hint == NOTE_REVOKE) {
3894 		kn->kn_flags |= EV_EOF;
3895 		return (1);
3896 	}
3897 	return (kn->kn_fflags != 0);
3898 }
3899 
3900 int
3901 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
3902 {
3903 	int error;
3904 
3905 	if (dp->d_reclen > ap->a_uio->uio_resid)
3906 		return (ENAMETOOLONG);
3907 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
3908 	if (error) {
3909 		if (ap->a_ncookies != NULL) {
3910 			if (ap->a_cookies != NULL)
3911 				free(ap->a_cookies, M_TEMP);
3912 			ap->a_cookies = NULL;
3913 			*ap->a_ncookies = 0;
3914 		}
3915 		return (error);
3916 	}
3917 	if (ap->a_ncookies == NULL)
3918 		return (0);
3919 
3920 	KASSERT(ap->a_cookies,
3921 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
3922 
3923 	*ap->a_cookies = realloc(*ap->a_cookies,
3924 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
3925 	(*ap->a_cookies)[*ap->a_ncookies] = off;
3926 	return (0);
3927 }
3928 
3929 /*
3930  * Mark for update the access time of the file if the filesystem
3931  * supports VA_MARK_ATIME.  This functionality is used by execve
3932  * and mmap, so we want to avoid the synchronous I/O implied by
3933  * directly setting va_atime for the sake of efficiency.
3934  */
3935 void
3936 vfs_mark_atime(struct vnode *vp, struct thread *td)
3937 {
3938 	struct vattr atimeattr;
3939 
3940 	if ((vp->v_mount->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) {
3941 		VATTR_NULL(&atimeattr);
3942 		atimeattr.va_vaflags |= VA_MARK_ATIME;
3943 		(void)VOP_SETATTR(vp, &atimeattr, td->td_ucred, td);
3944 	}
3945 }
3946