xref: /freebsd/sys/kern/vfs_subr.c (revision f0a75d274af375d15b97b830966b99a02b7db911)
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	vinactive(struct vnode *, struct thread *);
99 static void	v_incr_usecount(struct vnode *);
100 static void	v_decr_usecount(struct vnode *);
101 static void	v_decr_useonly(struct vnode *);
102 static void	v_upgrade_usecount(struct vnode *);
103 static void	vfree(struct vnode *);
104 static void	vnlru_free(int);
105 static void	vdestroy(struct vnode *);
106 static void	vgonel(struct vnode *);
107 static void	vfs_knllock(void *arg);
108 static void	vfs_knlunlock(void *arg);
109 static int	vfs_knllocked(void *arg);
110 
111 
112 /*
113  * Enable Giant pushdown based on whether or not the vm is mpsafe in this
114  * build.  Without mpsafevm the buffer cache can not run Giant free.
115  */
116 int mpsafe_vfs = 1;
117 TUNABLE_INT("debug.mpsafevfs", &mpsafe_vfs);
118 SYSCTL_INT(_debug, OID_AUTO, mpsafevfs, CTLFLAG_RD, &mpsafe_vfs, 0,
119     "MPSAFE VFS");
120 
121 /*
122  * Number of vnodes in existence.  Increased whenever getnewvnode()
123  * allocates a new vnode, decreased on vdestroy() called on VI_DOOMed
124  * vnode.
125  */
126 static unsigned long	numvnodes;
127 
128 SYSCTL_LONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, "");
129 
130 /*
131  * Conversion tables for conversion from vnode types to inode formats
132  * and back.
133  */
134 enum vtype iftovt_tab[16] = {
135 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
136 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
137 };
138 int vttoif_tab[10] = {
139 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
140 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
141 };
142 
143 /*
144  * List of vnodes that are ready for recycling.
145  */
146 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
147 
148 /*
149  * Free vnode target.  Free vnodes may simply be files which have been stat'd
150  * but not read.  This is somewhat common, and a small cache of such files
151  * should be kept to avoid recreation costs.
152  */
153 static u_long wantfreevnodes;
154 SYSCTL_LONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW, &wantfreevnodes, 0, "");
155 /* Number of vnodes in the free list. */
156 static u_long freevnodes;
157 SYSCTL_LONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, "");
158 
159 /*
160  * Various variables used for debugging the new implementation of
161  * reassignbuf().
162  * XXX these are probably of (very) limited utility now.
163  */
164 static int reassignbufcalls;
165 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0, "");
166 
167 /*
168  * Cache for the mount type id assigned to NFS.  This is used for
169  * special checks in nfs/nfs_nqlease.c and vm/vnode_pager.c.
170  */
171 int	nfs_mount_type = -1;
172 
173 /* To keep more than one thread at a time from running vfs_getnewfsid */
174 static struct mtx mntid_mtx;
175 
176 /*
177  * Lock for any access to the following:
178  *	vnode_free_list
179  *	numvnodes
180  *	freevnodes
181  */
182 static struct mtx vnode_free_list_mtx;
183 
184 /* Publicly exported FS */
185 struct nfs_public nfs_pub;
186 
187 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
188 static uma_zone_t vnode_zone;
189 static uma_zone_t vnodepoll_zone;
190 
191 /* Set to 1 to print out reclaim of active vnodes */
192 int	prtactive;
193 
194 /*
195  * The workitem queue.
196  *
197  * It is useful to delay writes of file data and filesystem metadata
198  * for tens of seconds so that quickly created and deleted files need
199  * not waste disk bandwidth being created and removed. To realize this,
200  * we append vnodes to a "workitem" queue. When running with a soft
201  * updates implementation, most pending metadata dependencies should
202  * not wait for more than a few seconds. Thus, mounted on block devices
203  * are delayed only about a half the time that file data is delayed.
204  * Similarly, directory updates are more critical, so are only delayed
205  * about a third the time that file data is delayed. Thus, there are
206  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
207  * one each second (driven off the filesystem syncer process). The
208  * syncer_delayno variable indicates the next queue that is to be processed.
209  * Items that need to be processed soon are placed in this queue:
210  *
211  *	syncer_workitem_pending[syncer_delayno]
212  *
213  * A delay of fifteen seconds is done by placing the request fifteen
214  * entries later in the queue:
215  *
216  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
217  *
218  */
219 static int syncer_delayno;
220 static long syncer_mask;
221 LIST_HEAD(synclist, bufobj);
222 static struct synclist *syncer_workitem_pending;
223 /*
224  * The sync_mtx protects:
225  *	bo->bo_synclist
226  *	sync_vnode_count
227  *	syncer_delayno
228  *	syncer_state
229  *	syncer_workitem_pending
230  *	syncer_worklist_len
231  *	rushjob
232  */
233 static struct mtx sync_mtx;
234 
235 #define SYNCER_MAXDELAY		32
236 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
237 static int syncdelay = 30;		/* max time to delay syncing data */
238 static int filedelay = 30;		/* time to delay syncing files */
239 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, "");
240 static int dirdelay = 29;		/* time to delay syncing directories */
241 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, "");
242 static int metadelay = 28;		/* time to delay syncing metadata */
243 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, "");
244 static int rushjob;		/* number of slots to run ASAP */
245 static int stat_rush_requests;	/* number of times I/O speeded up */
246 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, "");
247 
248 /*
249  * When shutting down the syncer, run it at four times normal speed.
250  */
251 #define SYNCER_SHUTDOWN_SPEEDUP		4
252 static int sync_vnode_count;
253 static int syncer_worklist_len;
254 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
255     syncer_state;
256 
257 /*
258  * Number of vnodes we want to exist at any one time.  This is mostly used
259  * to size hash tables in vnode-related code.  It is normally not used in
260  * getnewvnode(), as wantfreevnodes is normally nonzero.)
261  *
262  * XXX desiredvnodes is historical cruft and should not exist.
263  */
264 int desiredvnodes;
265 SYSCTL_INT(_kern, KERN_MAXVNODES, maxvnodes, CTLFLAG_RW,
266     &desiredvnodes, 0, "Maximum number of vnodes");
267 SYSCTL_INT(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
268     &wantfreevnodes, 0, "Minimum number of vnodes (legacy)");
269 static int vnlru_nowhere;
270 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
271     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
272 
273 /*
274  * Macros to control when a vnode is freed and recycled.  All require
275  * the vnode interlock.
276  */
277 #define VCANRECYCLE(vp) (((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt)
278 #define VSHOULDFREE(vp) (!((vp)->v_iflag & VI_FREE) && !(vp)->v_holdcnt)
279 #define VSHOULDBUSY(vp) (((vp)->v_iflag & VI_FREE) && (vp)->v_holdcnt)
280 
281 
282 /*
283  * Initialize the vnode management data structures.
284  */
285 #ifndef	MAXVNODES_MAX
286 #define	MAXVNODES_MAX	100000
287 #endif
288 static void
289 vntblinit(void *dummy __unused)
290 {
291 
292 	/*
293 	 * Desiredvnodes is a function of the physical memory size and
294 	 * the kernel's heap size.  Specifically, desiredvnodes scales
295 	 * in proportion to the physical memory size until two fifths
296 	 * of the kernel's heap size is consumed by vnodes and vm
297 	 * objects.
298 	 */
299 	desiredvnodes = min(maxproc + cnt.v_page_count / 4, 2 * vm_kmem_size /
300 	    (5 * (sizeof(struct vm_object) + sizeof(struct vnode))));
301 	if (desiredvnodes > MAXVNODES_MAX) {
302 		if (bootverbose)
303 			printf("Reducing kern.maxvnodes %d -> %d\n",
304 			    desiredvnodes, MAXVNODES_MAX);
305 		desiredvnodes = MAXVNODES_MAX;
306 	}
307 	wantfreevnodes = desiredvnodes / 4;
308 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
309 	TAILQ_INIT(&vnode_free_list);
310 	mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
311 	vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
312 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
313 	vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
314 	      NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
315 	/*
316 	 * Initialize the filesystem syncer.
317 	 */
318 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
319 		&syncer_mask);
320 	syncer_maxdelay = syncer_mask + 1;
321 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
322 }
323 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL)
324 
325 
326 /*
327  * Mark a mount point as busy. Used to synchronize access and to delay
328  * unmounting. Interlock is not released on failure.
329  */
330 int
331 vfs_busy(struct mount *mp, int flags, struct mtx *interlkp,
332     struct thread *td)
333 {
334 	int lkflags;
335 
336 	MNT_ILOCK(mp);
337 	MNT_REF(mp);
338 	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
339 		if (flags & LK_NOWAIT) {
340 			MNT_REL(mp);
341 			MNT_IUNLOCK(mp);
342 			return (ENOENT);
343 		}
344 		if (interlkp)
345 			mtx_unlock(interlkp);
346 		mp->mnt_kern_flag |= MNTK_MWAIT;
347 		/*
348 		 * Since all busy locks are shared except the exclusive
349 		 * lock granted when unmounting, the only place that a
350 		 * wakeup needs to be done is at the release of the
351 		 * exclusive lock at the end of dounmount.
352 		 */
353 		msleep(mp, MNT_MTX(mp), PVFS, "vfs_busy", 0);
354 		MNT_REL(mp);
355 		MNT_IUNLOCK(mp);
356 		if (interlkp)
357 			mtx_lock(interlkp);
358 		return (ENOENT);
359 	}
360 	if (interlkp)
361 		mtx_unlock(interlkp);
362 	lkflags = LK_SHARED | LK_INTERLOCK;
363 	if (lockmgr(&mp->mnt_lock, lkflags, MNT_MTX(mp), td))
364 		panic("vfs_busy: unexpected lock failure");
365 	return (0);
366 }
367 
368 /*
369  * Free a busy filesystem.
370  */
371 void
372 vfs_unbusy(struct mount *mp, struct thread *td)
373 {
374 
375 	lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td);
376 	vfs_rel(mp);
377 }
378 
379 /*
380  * Lookup a mount point by filesystem identifier.
381  */
382 struct mount *
383 vfs_getvfs(fsid_t *fsid)
384 {
385 	struct mount *mp;
386 
387 	mtx_lock(&mountlist_mtx);
388 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
389 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
390 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
391 			vfs_ref(mp);
392 			mtx_unlock(&mountlist_mtx);
393 			return (mp);
394 		}
395 	}
396 	mtx_unlock(&mountlist_mtx);
397 	return ((struct mount *) 0);
398 }
399 
400 /*
401  * Check if a user can access priveledged mount options.
402  */
403 int
404 vfs_suser(struct mount *mp, struct thread *td)
405 {
406 	int error;
407 
408 	if ((mp->mnt_flag & MNT_USER) == 0 ||
409 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
410 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
411 			return (error);
412 	}
413 	return (0);
414 }
415 
416 /*
417  * Get a new unique fsid.  Try to make its val[0] unique, since this value
418  * will be used to create fake device numbers for stat().  Also try (but
419  * not so hard) make its val[0] unique mod 2^16, since some emulators only
420  * support 16-bit device numbers.  We end up with unique val[0]'s for the
421  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
422  *
423  * Keep in mind that several mounts may be running in parallel.  Starting
424  * the search one past where the previous search terminated is both a
425  * micro-optimization and a defense against returning the same fsid to
426  * different mounts.
427  */
428 void
429 vfs_getnewfsid(struct mount *mp)
430 {
431 	static u_int16_t mntid_base;
432 	struct mount *nmp;
433 	fsid_t tfsid;
434 	int mtype;
435 
436 	mtx_lock(&mntid_mtx);
437 	mtype = mp->mnt_vfc->vfc_typenum;
438 	tfsid.val[1] = mtype;
439 	mtype = (mtype & 0xFF) << 24;
440 	for (;;) {
441 		tfsid.val[0] = makedev(255,
442 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
443 		mntid_base++;
444 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
445 			break;
446 		vfs_rel(nmp);
447 	}
448 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
449 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
450 	mtx_unlock(&mntid_mtx);
451 }
452 
453 /*
454  * Knob to control the precision of file timestamps:
455  *
456  *   0 = seconds only; nanoseconds zeroed.
457  *   1 = seconds and nanoseconds, accurate within 1/HZ.
458  *   2 = seconds and nanoseconds, truncated to microseconds.
459  * >=3 = seconds and nanoseconds, maximum precision.
460  */
461 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
462 
463 static int timestamp_precision = TSP_SEC;
464 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
465     &timestamp_precision, 0, "");
466 
467 /*
468  * Get a current timestamp.
469  */
470 void
471 vfs_timestamp(struct timespec *tsp)
472 {
473 	struct timeval tv;
474 
475 	switch (timestamp_precision) {
476 	case TSP_SEC:
477 		tsp->tv_sec = time_second;
478 		tsp->tv_nsec = 0;
479 		break;
480 	case TSP_HZ:
481 		getnanotime(tsp);
482 		break;
483 	case TSP_USEC:
484 		microtime(&tv);
485 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
486 		break;
487 	case TSP_NSEC:
488 	default:
489 		nanotime(tsp);
490 		break;
491 	}
492 }
493 
494 /*
495  * Set vnode attributes to VNOVAL
496  */
497 void
498 vattr_null(struct vattr *vap)
499 {
500 
501 	vap->va_type = VNON;
502 	vap->va_size = VNOVAL;
503 	vap->va_bytes = VNOVAL;
504 	vap->va_mode = VNOVAL;
505 	vap->va_nlink = VNOVAL;
506 	vap->va_uid = VNOVAL;
507 	vap->va_gid = VNOVAL;
508 	vap->va_fsid = VNOVAL;
509 	vap->va_fileid = VNOVAL;
510 	vap->va_blocksize = VNOVAL;
511 	vap->va_rdev = VNOVAL;
512 	vap->va_atime.tv_sec = VNOVAL;
513 	vap->va_atime.tv_nsec = VNOVAL;
514 	vap->va_mtime.tv_sec = VNOVAL;
515 	vap->va_mtime.tv_nsec = VNOVAL;
516 	vap->va_ctime.tv_sec = VNOVAL;
517 	vap->va_ctime.tv_nsec = VNOVAL;
518 	vap->va_birthtime.tv_sec = VNOVAL;
519 	vap->va_birthtime.tv_nsec = VNOVAL;
520 	vap->va_flags = VNOVAL;
521 	vap->va_gen = VNOVAL;
522 	vap->va_vaflags = 0;
523 }
524 
525 /*
526  * This routine is called when we have too many vnodes.  It attempts
527  * to free <count> vnodes and will potentially free vnodes that still
528  * have VM backing store (VM backing store is typically the cause
529  * of a vnode blowout so we want to do this).  Therefore, this operation
530  * is not considered cheap.
531  *
532  * A number of conditions may prevent a vnode from being reclaimed.
533  * the buffer cache may have references on the vnode, a directory
534  * vnode may still have references due to the namei cache representing
535  * underlying files, or the vnode may be in active use.   It is not
536  * desireable to reuse such vnodes.  These conditions may cause the
537  * number of vnodes to reach some minimum value regardless of what
538  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
539  */
540 static int
541 vlrureclaim(struct mount *mp)
542 {
543 	struct thread *td;
544 	struct vnode *vp;
545 	int done;
546 	int trigger;
547 	int usevnodes;
548 	int count;
549 
550 	/*
551 	 * Calculate the trigger point, don't allow user
552 	 * screwups to blow us up.   This prevents us from
553 	 * recycling vnodes with lots of resident pages.  We
554 	 * aren't trying to free memory, we are trying to
555 	 * free vnodes.
556 	 */
557 	usevnodes = desiredvnodes;
558 	if (usevnodes <= 0)
559 		usevnodes = 1;
560 	trigger = cnt.v_page_count * 2 / usevnodes;
561 	done = 0;
562 	td = curthread;
563 	vn_start_write(NULL, &mp, V_WAIT);
564 	MNT_ILOCK(mp);
565 	count = mp->mnt_nvnodelistsize / 10 + 1;
566 	while (count != 0) {
567 		vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
568 		while (vp != NULL && vp->v_type == VMARKER)
569 			vp = TAILQ_NEXT(vp, v_nmntvnodes);
570 		if (vp == NULL)
571 			break;
572 		TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
573 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
574 		--count;
575 		if (!VI_TRYLOCK(vp))
576 			goto next_iter;
577 		/*
578 		 * If it's been deconstructed already, it's still
579 		 * referenced, or it exceeds the trigger, skip it.
580 		 */
581 		if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) ||
582 		    (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
583 		    vp->v_object->resident_page_count > trigger)) {
584 			VI_UNLOCK(vp);
585 			goto next_iter;
586 		}
587 		MNT_IUNLOCK(mp);
588 		vholdl(vp);
589 		if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT, td)) {
590 			vdrop(vp);
591 			goto next_iter_mntunlocked;
592 		}
593 		VI_LOCK(vp);
594 		/*
595 		 * v_usecount may have been bumped after VOP_LOCK() dropped
596 		 * the vnode interlock and before it was locked again.
597 		 *
598 		 * It is not necessary to recheck VI_DOOMED because it can
599 		 * only be set by another thread that holds both the vnode
600 		 * lock and vnode interlock.  If another thread has the
601 		 * vnode lock before we get to VOP_LOCK() and obtains the
602 		 * vnode interlock after VOP_LOCK() drops the vnode
603 		 * interlock, the other thread will be unable to drop the
604 		 * vnode lock before our VOP_LOCK() call fails.
605 		 */
606 		if (vp->v_usecount || !LIST_EMPTY(&(vp)->v_cache_src) ||
607 		    (vp->v_object != NULL &&
608 		    vp->v_object->resident_page_count > trigger)) {
609 			VOP_UNLOCK(vp, LK_INTERLOCK, td);
610 			goto next_iter_mntunlocked;
611 		}
612 		KASSERT((vp->v_iflag & VI_DOOMED) == 0,
613 		    ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
614 		vgonel(vp);
615 		VOP_UNLOCK(vp, 0, td);
616 		vdropl(vp);
617 		done++;
618 next_iter_mntunlocked:
619 		if ((count % 256) != 0)
620 			goto relock_mnt;
621 		goto yield;
622 next_iter:
623 		if ((count % 256) != 0)
624 			continue;
625 		MNT_IUNLOCK(mp);
626 yield:
627 		uio_yield();
628 relock_mnt:
629 		MNT_ILOCK(mp);
630 	}
631 	MNT_IUNLOCK(mp);
632 	vn_finished_write(mp);
633 	return done;
634 }
635 
636 /*
637  * Attempt to keep the free list at wantfreevnodes length.
638  */
639 static void
640 vnlru_free(int count)
641 {
642 	struct vnode *vp;
643 	int vfslocked;
644 
645 	mtx_assert(&vnode_free_list_mtx, MA_OWNED);
646 	for (; count > 0; count--) {
647 		vp = TAILQ_FIRST(&vnode_free_list);
648 		/*
649 		 * The list can be modified while the free_list_mtx
650 		 * has been dropped and vp could be NULL here.
651 		 */
652 		if (!vp)
653 			break;
654 		VNASSERT(vp->v_op != NULL, vp,
655 		    ("vnlru_free: vnode already reclaimed."));
656 		TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
657 		/*
658 		 * Don't recycle if we can't get the interlock.
659 		 */
660 		if (!VI_TRYLOCK(vp)) {
661 			TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
662 			continue;
663 		}
664 		VNASSERT(VCANRECYCLE(vp), vp,
665 		    ("vp inconsistent on freelist"));
666 		freevnodes--;
667 		vp->v_iflag &= ~VI_FREE;
668 		vholdl(vp);
669 		mtx_unlock(&vnode_free_list_mtx);
670 		VI_UNLOCK(vp);
671 		vfslocked = VFS_LOCK_GIANT(vp->v_mount);
672 		vtryrecycle(vp);
673 		VFS_UNLOCK_GIANT(vfslocked);
674 		/*
675 		 * If the recycled succeeded this vdrop will actually free
676 		 * the vnode.  If not it will simply place it back on
677 		 * the free list.
678 		 */
679 		vdrop(vp);
680 		mtx_lock(&vnode_free_list_mtx);
681 	}
682 }
683 /*
684  * Attempt to recycle vnodes in a context that is always safe to block.
685  * Calling vlrurecycle() from the bowels of filesystem code has some
686  * interesting deadlock problems.
687  */
688 static struct proc *vnlruproc;
689 static int vnlruproc_sig;
690 
691 static void
692 vnlru_proc(void)
693 {
694 	struct mount *mp, *nmp;
695 	int done;
696 	struct proc *p = vnlruproc;
697 	struct thread *td = FIRST_THREAD_IN_PROC(p);
698 
699 	mtx_lock(&Giant);
700 
701 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, p,
702 	    SHUTDOWN_PRI_FIRST);
703 
704 	for (;;) {
705 		kthread_suspend_check(p);
706 		mtx_lock(&vnode_free_list_mtx);
707 		if (freevnodes > wantfreevnodes)
708 			vnlru_free(freevnodes - wantfreevnodes);
709 		if (numvnodes <= desiredvnodes * 9 / 10) {
710 			vnlruproc_sig = 0;
711 			wakeup(&vnlruproc_sig);
712 			msleep(vnlruproc, &vnode_free_list_mtx,
713 			    PVFS|PDROP, "vlruwt", hz);
714 			continue;
715 		}
716 		mtx_unlock(&vnode_free_list_mtx);
717 		done = 0;
718 		mtx_lock(&mountlist_mtx);
719 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
720 			int vfsunlocked;
721 			if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td)) {
722 				nmp = TAILQ_NEXT(mp, mnt_list);
723 				continue;
724 			}
725 			if (!VFS_NEEDSGIANT(mp)) {
726 				mtx_unlock(&Giant);
727 				vfsunlocked = 1;
728 			} else
729 				vfsunlocked = 0;
730 			done += vlrureclaim(mp);
731 			if (vfsunlocked)
732 				mtx_lock(&Giant);
733 			mtx_lock(&mountlist_mtx);
734 			nmp = TAILQ_NEXT(mp, mnt_list);
735 			vfs_unbusy(mp, td);
736 		}
737 		mtx_unlock(&mountlist_mtx);
738 		if (done == 0) {
739 #if 0
740 			/* These messages are temporary debugging aids */
741 			if (vnlru_nowhere < 5)
742 				printf("vnlru process getting nowhere..\n");
743 			else if (vnlru_nowhere == 5)
744 				printf("vnlru process messages stopped.\n");
745 #endif
746 			vnlru_nowhere++;
747 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
748 		} else
749 			uio_yield();
750 	}
751 }
752 
753 static struct kproc_desc vnlru_kp = {
754 	"vnlru",
755 	vnlru_proc,
756 	&vnlruproc
757 };
758 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
759 
760 /*
761  * Routines having to do with the management of the vnode table.
762  */
763 
764 static void
765 vdestroy(struct vnode *vp)
766 {
767 	struct bufobj *bo;
768 
769 	CTR1(KTR_VFS, "vdestroy vp %p", vp);
770 	mtx_lock(&vnode_free_list_mtx);
771 	numvnodes--;
772 	mtx_unlock(&vnode_free_list_mtx);
773 	bo = &vp->v_bufobj;
774 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
775 	    ("cleaned vnode still on the free list."));
776 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
777 	VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
778 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
779 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
780 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
781 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
782 	VNASSERT(bo->bo_clean.bv_root == NULL, vp, ("cleanblkroot not NULL"));
783 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
784 	VNASSERT(bo->bo_dirty.bv_root == NULL, vp, ("dirtyblkroot not NULL"));
785 	VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
786 	VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
787 	VI_UNLOCK(vp);
788 #ifdef MAC
789 	mac_destroy_vnode(vp);
790 #endif
791 	if (vp->v_pollinfo != NULL) {
792 		knlist_destroy(&vp->v_pollinfo->vpi_selinfo.si_note);
793 		mtx_destroy(&vp->v_pollinfo->vpi_lock);
794 		uma_zfree(vnodepoll_zone, vp->v_pollinfo);
795 	}
796 #ifdef INVARIANTS
797 	/* XXX Elsewhere we can detect an already freed vnode via NULL v_op. */
798 	vp->v_op = NULL;
799 #endif
800 	lockdestroy(vp->v_vnlock);
801 	mtx_destroy(&vp->v_interlock);
802 	uma_zfree(vnode_zone, vp);
803 }
804 
805 /*
806  * Try to recycle a freed vnode.  We abort if anyone picks up a reference
807  * before we actually vgone().  This function must be called with the vnode
808  * held to prevent the vnode from being returned to the free list midway
809  * through vgone().
810  */
811 static int
812 vtryrecycle(struct vnode *vp)
813 {
814 	struct thread *td = curthread;
815 	struct mount *vnmp;
816 
817 	CTR1(KTR_VFS, "vtryrecycle: trying vp %p", vp);
818 	VNASSERT(vp->v_holdcnt, vp,
819 	    ("vtryrecycle: Recycling vp %p without a reference.", vp));
820 	/*
821 	 * This vnode may found and locked via some other list, if so we
822 	 * can't recycle it yet.
823 	 */
824 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT, td) != 0)
825 		return (EWOULDBLOCK);
826 	/*
827 	 * Don't recycle if its filesystem is being suspended.
828 	 */
829 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
830 		VOP_UNLOCK(vp, 0, td);
831 		return (EBUSY);
832 	}
833 	/*
834 	 * If we got this far, we need to acquire the interlock and see if
835 	 * anyone picked up this vnode from another list.  If not, we will
836 	 * mark it with DOOMED via vgonel() so that anyone who does find it
837 	 * will skip over it.
838 	 */
839 	VI_LOCK(vp);
840 	if (vp->v_usecount) {
841 		VOP_UNLOCK(vp, LK_INTERLOCK, td);
842 		vn_finished_write(vnmp);
843 		return (EBUSY);
844 	}
845 	if ((vp->v_iflag & VI_DOOMED) == 0)
846 		vgonel(vp);
847 	VOP_UNLOCK(vp, LK_INTERLOCK, td);
848 	vn_finished_write(vnmp);
849 	CTR1(KTR_VFS, "vtryrecycle: recycled vp %p", vp);
850 	return (0);
851 }
852 
853 /*
854  * Return the next vnode from the free list.
855  */
856 int
857 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
858     struct vnode **vpp)
859 {
860 	struct vnode *vp = NULL;
861 	struct bufobj *bo;
862 
863 	mtx_lock(&vnode_free_list_mtx);
864 	/*
865 	 * Lend our context to reclaim vnodes if they've exceeded the max.
866 	 */
867 	if (freevnodes > wantfreevnodes)
868 		vnlru_free(1);
869 	/*
870 	 * Wait for available vnodes.
871 	 */
872 	if (numvnodes > desiredvnodes) {
873 		if (mp != NULL && (mp->mnt_kern_flag & MNTK_SUSPEND)) {
874 			/*
875 			 * File system is beeing suspended, we cannot risk a
876 			 * deadlock here, so allocate new vnode anyway.
877 			 */
878 			if (freevnodes > wantfreevnodes)
879 				vnlru_free(freevnodes - wantfreevnodes);
880 			goto alloc;
881 		}
882 		if (vnlruproc_sig == 0) {
883 			vnlruproc_sig = 1;      /* avoid unnecessary wakeups */
884 			wakeup(vnlruproc);
885 		}
886 		msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
887 		    "vlruwk", hz);
888 #if 0	/* XXX Not all VFS_VGET/ffs_vget callers check returns. */
889 		if (numvnodes > desiredvnodes) {
890 			mtx_unlock(&vnode_free_list_mtx);
891 			return (ENFILE);
892 		}
893 #endif
894 	}
895 alloc:
896 	numvnodes++;
897 	mtx_unlock(&vnode_free_list_mtx);
898 	vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK|M_ZERO);
899 	/*
900 	 * Setup locks.
901 	 */
902 	vp->v_vnlock = &vp->v_lock;
903 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
904 	/*
905 	 * By default, don't allow shared locks unless filesystems
906 	 * opt-in.
907 	 */
908 	lockinit(vp->v_vnlock, PVFS, tag, VLKTIMEOUT, LK_NOSHARE);
909 	/*
910 	 * Initialize bufobj.
911 	 */
912 	bo = &vp->v_bufobj;
913 	bo->__bo_vnode = vp;
914 	bo->bo_mtx = &vp->v_interlock;
915 	bo->bo_ops = &buf_ops_bio;
916 	bo->bo_private = vp;
917 	TAILQ_INIT(&bo->bo_clean.bv_hd);
918 	TAILQ_INIT(&bo->bo_dirty.bv_hd);
919 	/*
920 	 * Initialize namecache.
921 	 */
922 	LIST_INIT(&vp->v_cache_src);
923 	TAILQ_INIT(&vp->v_cache_dst);
924 	/*
925 	 * Finalize various vnode identity bits.
926 	 */
927 	vp->v_type = VNON;
928 	vp->v_tag = tag;
929 	vp->v_op = vops;
930 	v_incr_usecount(vp);
931 	vp->v_data = 0;
932 #ifdef MAC
933 	mac_init_vnode(vp);
934 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
935 		mac_associate_vnode_singlelabel(mp, vp);
936 	else if (mp == NULL)
937 		printf("NULL mp in getnewvnode()\n");
938 #endif
939 	if (mp != NULL) {
940 		bo->bo_bsize = mp->mnt_stat.f_iosize;
941 		if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
942 			vp->v_vflag |= VV_NOKNOTE;
943 	}
944 
945 	CTR2(KTR_VFS, "getnewvnode: mp %p vp %p", mp, vp);
946 	*vpp = vp;
947 	return (0);
948 }
949 
950 /*
951  * Delete from old mount point vnode list, if on one.
952  */
953 static void
954 delmntque(struct vnode *vp)
955 {
956 	struct mount *mp;
957 
958 	mp = vp->v_mount;
959 	if (mp == NULL)
960 		return;
961 	MNT_ILOCK(mp);
962 	vp->v_mount = NULL;
963 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
964 		("bad mount point vnode list size"));
965 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
966 	mp->mnt_nvnodelistsize--;
967 	MNT_REL(mp);
968 	MNT_IUNLOCK(mp);
969 }
970 
971 static void
972 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
973 {
974 	struct thread *td;
975 
976 	td = curthread; /* XXX ? */
977 	vp->v_data = NULL;
978 	vp->v_op = &dead_vnodeops;
979 	/* XXX non mp-safe fs may still call insmntque with vnode
980 	   unlocked */
981 	if (!VOP_ISLOCKED(vp, td))
982 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
983 	vgone(vp);
984 	vput(vp);
985 }
986 
987 /*
988  * Insert into list of vnodes for the new mount point, if available.
989  */
990 int
991 insmntque1(struct vnode *vp, struct mount *mp,
992 	void (*dtr)(struct vnode *, void *), void *dtr_arg)
993 {
994 
995 	KASSERT(vp->v_mount == NULL,
996 		("insmntque: vnode already on per mount vnode list"));
997 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
998 	MNT_ILOCK(mp);
999 	if ((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
1000 	    mp->mnt_nvnodelistsize == 0) {
1001 		MNT_IUNLOCK(mp);
1002 		if (dtr != NULL)
1003 			dtr(vp, dtr_arg);
1004 		return (EBUSY);
1005 	}
1006 	vp->v_mount = mp;
1007 	MNT_REF(mp);
1008 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1009 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1010 		("neg mount point vnode list size"));
1011 	mp->mnt_nvnodelistsize++;
1012 	MNT_IUNLOCK(mp);
1013 	return (0);
1014 }
1015 
1016 int
1017 insmntque(struct vnode *vp, struct mount *mp)
1018 {
1019 
1020 	return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1021 }
1022 
1023 /*
1024  * Flush out and invalidate all buffers associated with a bufobj
1025  * Called with the underlying object locked.
1026  */
1027 int
1028 bufobj_invalbuf(struct bufobj *bo, int flags, struct thread *td, int slpflag,
1029     int slptimeo)
1030 {
1031 	int error;
1032 
1033 	BO_LOCK(bo);
1034 	if (flags & V_SAVE) {
1035 		error = bufobj_wwait(bo, slpflag, slptimeo);
1036 		if (error) {
1037 			BO_UNLOCK(bo);
1038 			return (error);
1039 		}
1040 		if (bo->bo_dirty.bv_cnt > 0) {
1041 			BO_UNLOCK(bo);
1042 			if ((error = BO_SYNC(bo, MNT_WAIT, td)) != 0)
1043 				return (error);
1044 			/*
1045 			 * XXX We could save a lock/unlock if this was only
1046 			 * enabled under INVARIANTS
1047 			 */
1048 			BO_LOCK(bo);
1049 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1050 				panic("vinvalbuf: dirty bufs");
1051 		}
1052 	}
1053 	/*
1054 	 * If you alter this loop please notice that interlock is dropped and
1055 	 * reacquired in flushbuflist.  Special care is needed to ensure that
1056 	 * no race conditions occur from this.
1057 	 */
1058 	do {
1059 		error = flushbuflist(&bo->bo_clean,
1060 		    flags, bo, slpflag, slptimeo);
1061 		if (error == 0)
1062 			error = flushbuflist(&bo->bo_dirty,
1063 			    flags, bo, slpflag, slptimeo);
1064 		if (error != 0 && error != EAGAIN) {
1065 			BO_UNLOCK(bo);
1066 			return (error);
1067 		}
1068 	} while (error != 0);
1069 
1070 	/*
1071 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
1072 	 * have write I/O in-progress but if there is a VM object then the
1073 	 * VM object can also have read-I/O in-progress.
1074 	 */
1075 	do {
1076 		bufobj_wwait(bo, 0, 0);
1077 		BO_UNLOCK(bo);
1078 		if (bo->bo_object != NULL) {
1079 			VM_OBJECT_LOCK(bo->bo_object);
1080 			vm_object_pip_wait(bo->bo_object, "bovlbx");
1081 			VM_OBJECT_UNLOCK(bo->bo_object);
1082 		}
1083 		BO_LOCK(bo);
1084 	} while (bo->bo_numoutput > 0);
1085 	BO_UNLOCK(bo);
1086 
1087 	/*
1088 	 * Destroy the copy in the VM cache, too.
1089 	 */
1090 	if (bo->bo_object != NULL) {
1091 		VM_OBJECT_LOCK(bo->bo_object);
1092 		vm_object_page_remove(bo->bo_object, 0, 0,
1093 			(flags & V_SAVE) ? TRUE : FALSE);
1094 		VM_OBJECT_UNLOCK(bo->bo_object);
1095 	}
1096 
1097 #ifdef INVARIANTS
1098 	BO_LOCK(bo);
1099 	if ((flags & (V_ALT | V_NORMAL)) == 0 &&
1100 	    (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0))
1101 		panic("vinvalbuf: flush failed");
1102 	BO_UNLOCK(bo);
1103 #endif
1104 	return (0);
1105 }
1106 
1107 /*
1108  * Flush out and invalidate all buffers associated with a vnode.
1109  * Called with the underlying object locked.
1110  */
1111 int
1112 vinvalbuf(struct vnode *vp, int flags, struct thread *td, int slpflag,
1113     int slptimeo)
1114 {
1115 
1116 	CTR2(KTR_VFS, "vinvalbuf vp %p flags %d", vp, flags);
1117 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1118 	return (bufobj_invalbuf(&vp->v_bufobj, flags, td, slpflag, slptimeo));
1119 }
1120 
1121 /*
1122  * Flush out buffers on the specified list.
1123  *
1124  */
1125 static int
1126 flushbuflist( struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1127     int slptimeo)
1128 {
1129 	struct buf *bp, *nbp;
1130 	int retval, error;
1131 	daddr_t lblkno;
1132 	b_xflags_t xflags;
1133 
1134 	ASSERT_BO_LOCKED(bo);
1135 
1136 	retval = 0;
1137 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1138 		if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1139 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1140 			continue;
1141 		}
1142 		lblkno = 0;
1143 		xflags = 0;
1144 		if (nbp != NULL) {
1145 			lblkno = nbp->b_lblkno;
1146 			xflags = nbp->b_xflags &
1147 				(BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN);
1148 		}
1149 		retval = EAGAIN;
1150 		error = BUF_TIMELOCK(bp,
1151 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_MTX(bo),
1152 		    "flushbuf", slpflag, slptimeo);
1153 		if (error) {
1154 			BO_LOCK(bo);
1155 			return (error != ENOLCK ? error : EAGAIN);
1156 		}
1157 		KASSERT(bp->b_bufobj == bo,
1158 	            ("bp %p wrong b_bufobj %p should be %p",
1159 		    bp, bp->b_bufobj, bo));
1160 		if (bp->b_bufobj != bo) {	/* XXX: necessary ? */
1161 			BUF_UNLOCK(bp);
1162 			BO_LOCK(bo);
1163 			return (EAGAIN);
1164 		}
1165 		/*
1166 		 * XXX Since there are no node locks for NFS, I
1167 		 * believe there is a slight chance that a delayed
1168 		 * write will occur while sleeping just above, so
1169 		 * check for it.
1170 		 */
1171 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1172 		    (flags & V_SAVE)) {
1173 			bremfree(bp);
1174 			bp->b_flags |= B_ASYNC;
1175 			bwrite(bp);
1176 			BO_LOCK(bo);
1177 			return (EAGAIN);	/* XXX: why not loop ? */
1178 		}
1179 		bremfree(bp);
1180 		bp->b_flags |= (B_INVAL | B_RELBUF);
1181 		bp->b_flags &= ~B_ASYNC;
1182 		brelse(bp);
1183 		BO_LOCK(bo);
1184 		if (nbp != NULL &&
1185 		    (nbp->b_bufobj != bo ||
1186 		     nbp->b_lblkno != lblkno ||
1187 		     (nbp->b_xflags &
1188 		      (BX_BKGRDMARKER | BX_VNDIRTY | BX_VNCLEAN)) != xflags))
1189 			break;			/* nbp invalid */
1190 	}
1191 	return (retval);
1192 }
1193 
1194 /*
1195  * Truncate a file's buffer and pages to a specified length.  This
1196  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1197  * sync activity.
1198  */
1199 int
1200 vtruncbuf(struct vnode *vp, struct ucred *cred, struct thread *td,
1201     off_t length, int blksize)
1202 {
1203 	struct buf *bp, *nbp;
1204 	int anyfreed;
1205 	int trunclbn;
1206 	struct bufobj *bo;
1207 
1208 	CTR2(KTR_VFS, "vtruncbuf vp %p length %jd", vp, length);
1209 	/*
1210 	 * Round up to the *next* lbn.
1211 	 */
1212 	trunclbn = (length + blksize - 1) / blksize;
1213 
1214 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1215 restart:
1216 	VI_LOCK(vp);
1217 	bo = &vp->v_bufobj;
1218 	anyfreed = 1;
1219 	for (;anyfreed;) {
1220 		anyfreed = 0;
1221 		TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1222 			if (bp->b_lblkno < trunclbn)
1223 				continue;
1224 			if (BUF_LOCK(bp,
1225 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1226 			    VI_MTX(vp)) == ENOLCK)
1227 				goto restart;
1228 
1229 			bremfree(bp);
1230 			bp->b_flags |= (B_INVAL | B_RELBUF);
1231 			bp->b_flags &= ~B_ASYNC;
1232 			brelse(bp);
1233 			anyfreed = 1;
1234 
1235 			if (nbp != NULL &&
1236 			    (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1237 			    (nbp->b_vp != vp) ||
1238 			    (nbp->b_flags & B_DELWRI))) {
1239 				goto restart;
1240 			}
1241 			VI_LOCK(vp);
1242 		}
1243 
1244 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1245 			if (bp->b_lblkno < trunclbn)
1246 				continue;
1247 			if (BUF_LOCK(bp,
1248 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1249 			    VI_MTX(vp)) == ENOLCK)
1250 				goto restart;
1251 			bremfree(bp);
1252 			bp->b_flags |= (B_INVAL | B_RELBUF);
1253 			bp->b_flags &= ~B_ASYNC;
1254 			brelse(bp);
1255 			anyfreed = 1;
1256 			if (nbp != NULL &&
1257 			    (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1258 			    (nbp->b_vp != vp) ||
1259 			    (nbp->b_flags & B_DELWRI) == 0)) {
1260 				goto restart;
1261 			}
1262 			VI_LOCK(vp);
1263 		}
1264 	}
1265 
1266 	if (length > 0) {
1267 restartsync:
1268 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1269 			if (bp->b_lblkno > 0)
1270 				continue;
1271 			/*
1272 			 * Since we hold the vnode lock this should only
1273 			 * fail if we're racing with the buf daemon.
1274 			 */
1275 			if (BUF_LOCK(bp,
1276 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1277 			    VI_MTX(vp)) == ENOLCK) {
1278 				goto restart;
1279 			}
1280 			VNASSERT((bp->b_flags & B_DELWRI), vp,
1281 			    ("buf(%p) on dirty queue without DELWRI", bp));
1282 
1283 			bremfree(bp);
1284 			bawrite(bp);
1285 			VI_LOCK(vp);
1286 			goto restartsync;
1287 		}
1288 	}
1289 
1290 	bufobj_wwait(bo, 0, 0);
1291 	VI_UNLOCK(vp);
1292 	vnode_pager_setsize(vp, length);
1293 
1294 	return (0);
1295 }
1296 
1297 /*
1298  * buf_splay() - splay tree core for the clean/dirty list of buffers in
1299  * 		 a vnode.
1300  *
1301  *	NOTE: We have to deal with the special case of a background bitmap
1302  *	buffer, a situation where two buffers will have the same logical
1303  *	block offset.  We want (1) only the foreground buffer to be accessed
1304  *	in a lookup and (2) must differentiate between the foreground and
1305  *	background buffer in the splay tree algorithm because the splay
1306  *	tree cannot normally handle multiple entities with the same 'index'.
1307  *	We accomplish this by adding differentiating flags to the splay tree's
1308  *	numerical domain.
1309  */
1310 static
1311 struct buf *
1312 buf_splay(daddr_t lblkno, b_xflags_t xflags, struct buf *root)
1313 {
1314 	struct buf dummy;
1315 	struct buf *lefttreemax, *righttreemin, *y;
1316 
1317 	if (root == NULL)
1318 		return (NULL);
1319 	lefttreemax = righttreemin = &dummy;
1320 	for (;;) {
1321 		if (lblkno < root->b_lblkno ||
1322 		    (lblkno == root->b_lblkno &&
1323 		    (xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1324 			if ((y = root->b_left) == NULL)
1325 				break;
1326 			if (lblkno < y->b_lblkno) {
1327 				/* Rotate right. */
1328 				root->b_left = y->b_right;
1329 				y->b_right = root;
1330 				root = y;
1331 				if ((y = root->b_left) == NULL)
1332 					break;
1333 			}
1334 			/* Link into the new root's right tree. */
1335 			righttreemin->b_left = root;
1336 			righttreemin = root;
1337 		} else if (lblkno > root->b_lblkno ||
1338 		    (lblkno == root->b_lblkno &&
1339 		    (xflags & BX_BKGRDMARKER) > (root->b_xflags & BX_BKGRDMARKER))) {
1340 			if ((y = root->b_right) == NULL)
1341 				break;
1342 			if (lblkno > y->b_lblkno) {
1343 				/* Rotate left. */
1344 				root->b_right = y->b_left;
1345 				y->b_left = root;
1346 				root = y;
1347 				if ((y = root->b_right) == NULL)
1348 					break;
1349 			}
1350 			/* Link into the new root's left tree. */
1351 			lefttreemax->b_right = root;
1352 			lefttreemax = root;
1353 		} else {
1354 			break;
1355 		}
1356 		root = y;
1357 	}
1358 	/* Assemble the new root. */
1359 	lefttreemax->b_right = root->b_left;
1360 	righttreemin->b_left = root->b_right;
1361 	root->b_left = dummy.b_right;
1362 	root->b_right = dummy.b_left;
1363 	return (root);
1364 }
1365 
1366 static void
1367 buf_vlist_remove(struct buf *bp)
1368 {
1369 	struct buf *root;
1370 	struct bufv *bv;
1371 
1372 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1373 	ASSERT_BO_LOCKED(bp->b_bufobj);
1374 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
1375 	    (BX_VNDIRTY|BX_VNCLEAN),
1376 	    ("buf_vlist_remove: Buf %p is on two lists", bp));
1377 	if (bp->b_xflags & BX_VNDIRTY)
1378 		bv = &bp->b_bufobj->bo_dirty;
1379 	else
1380 		bv = &bp->b_bufobj->bo_clean;
1381 	if (bp != bv->bv_root) {
1382 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1383 		KASSERT(root == bp, ("splay lookup failed in remove"));
1384 	}
1385 	if (bp->b_left == NULL) {
1386 		root = bp->b_right;
1387 	} else {
1388 		root = buf_splay(bp->b_lblkno, bp->b_xflags, bp->b_left);
1389 		root->b_right = bp->b_right;
1390 	}
1391 	bv->bv_root = root;
1392 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1393 	bv->bv_cnt--;
1394 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1395 }
1396 
1397 /*
1398  * Add the buffer to the sorted clean or dirty block list using a
1399  * splay tree algorithm.
1400  *
1401  * NOTE: xflags is passed as a constant, optimizing this inline function!
1402  */
1403 static void
1404 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1405 {
1406 	struct buf *root;
1407 	struct bufv *bv;
1408 
1409 	ASSERT_BO_LOCKED(bo);
1410 	KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1411 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
1412 	bp->b_xflags |= xflags;
1413 	if (xflags & BX_VNDIRTY)
1414 		bv = &bo->bo_dirty;
1415 	else
1416 		bv = &bo->bo_clean;
1417 
1418 	root = buf_splay(bp->b_lblkno, bp->b_xflags, bv->bv_root);
1419 	if (root == NULL) {
1420 		bp->b_left = NULL;
1421 		bp->b_right = NULL;
1422 		TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1423 	} else if (bp->b_lblkno < root->b_lblkno ||
1424 	    (bp->b_lblkno == root->b_lblkno &&
1425 	    (bp->b_xflags & BX_BKGRDMARKER) < (root->b_xflags & BX_BKGRDMARKER))) {
1426 		bp->b_left = root->b_left;
1427 		bp->b_right = root;
1428 		root->b_left = NULL;
1429 		TAILQ_INSERT_BEFORE(root, bp, b_bobufs);
1430 	} else {
1431 		bp->b_right = root->b_right;
1432 		bp->b_left = root;
1433 		root->b_right = NULL;
1434 		TAILQ_INSERT_AFTER(&bv->bv_hd, root, bp, b_bobufs);
1435 	}
1436 	bv->bv_cnt++;
1437 	bv->bv_root = bp;
1438 }
1439 
1440 /*
1441  * Lookup a buffer using the splay tree.  Note that we specifically avoid
1442  * shadow buffers used in background bitmap writes.
1443  *
1444  * This code isn't quite efficient as it could be because we are maintaining
1445  * two sorted lists and do not know which list the block resides in.
1446  *
1447  * During a "make buildworld" the desired buffer is found at one of
1448  * the roots more than 60% of the time.  Thus, checking both roots
1449  * before performing either splay eliminates unnecessary splays on the
1450  * first tree splayed.
1451  */
1452 struct buf *
1453 gbincore(struct bufobj *bo, daddr_t lblkno)
1454 {
1455 	struct buf *bp;
1456 
1457 	ASSERT_BO_LOCKED(bo);
1458 	if ((bp = bo->bo_clean.bv_root) != NULL &&
1459 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1460 		return (bp);
1461 	if ((bp = bo->bo_dirty.bv_root) != NULL &&
1462 	    bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1463 		return (bp);
1464 	if ((bp = bo->bo_clean.bv_root) != NULL) {
1465 		bo->bo_clean.bv_root = bp = buf_splay(lblkno, 0, bp);
1466 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1467 			return (bp);
1468 	}
1469 	if ((bp = bo->bo_dirty.bv_root) != NULL) {
1470 		bo->bo_dirty.bv_root = bp = buf_splay(lblkno, 0, bp);
1471 		if (bp->b_lblkno == lblkno && !(bp->b_xflags & BX_BKGRDMARKER))
1472 			return (bp);
1473 	}
1474 	return (NULL);
1475 }
1476 
1477 /*
1478  * Associate a buffer with a vnode.
1479  */
1480 void
1481 bgetvp(struct vnode *vp, struct buf *bp)
1482 {
1483 
1484 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
1485 
1486 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
1487 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
1488 	    ("bgetvp: bp already attached! %p", bp));
1489 
1490 	ASSERT_VI_LOCKED(vp, "bgetvp");
1491 	vholdl(vp);
1492 	if (VFS_NEEDSGIANT(vp->v_mount) ||
1493 	    vp->v_bufobj.bo_flag & BO_NEEDSGIANT)
1494 		bp->b_flags |= B_NEEDSGIANT;
1495 	bp->b_vp = vp;
1496 	bp->b_bufobj = &vp->v_bufobj;
1497 	/*
1498 	 * Insert onto list for new vnode.
1499 	 */
1500 	buf_vlist_add(bp, &vp->v_bufobj, BX_VNCLEAN);
1501 }
1502 
1503 /*
1504  * Disassociate a buffer from a vnode.
1505  */
1506 void
1507 brelvp(struct buf *bp)
1508 {
1509 	struct bufobj *bo;
1510 	struct vnode *vp;
1511 
1512 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1513 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1514 
1515 	/*
1516 	 * Delete from old vnode list, if on one.
1517 	 */
1518 	vp = bp->b_vp;		/* XXX */
1519 	bo = bp->b_bufobj;
1520 	BO_LOCK(bo);
1521 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1522 		buf_vlist_remove(bp);
1523 	else
1524 		panic("brelvp: Buffer %p not on queue.", bp);
1525 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1526 		bo->bo_flag &= ~BO_ONWORKLST;
1527 		mtx_lock(&sync_mtx);
1528 		LIST_REMOVE(bo, bo_synclist);
1529  		syncer_worklist_len--;
1530 		mtx_unlock(&sync_mtx);
1531 	}
1532 	bp->b_flags &= ~B_NEEDSGIANT;
1533 	bp->b_vp = NULL;
1534 	bp->b_bufobj = NULL;
1535 	vdropl(vp);
1536 }
1537 
1538 /*
1539  * Add an item to the syncer work queue.
1540  */
1541 static void
1542 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
1543 {
1544 	int slot;
1545 
1546 	ASSERT_BO_LOCKED(bo);
1547 
1548 	mtx_lock(&sync_mtx);
1549 	if (bo->bo_flag & BO_ONWORKLST)
1550 		LIST_REMOVE(bo, bo_synclist);
1551 	else {
1552 		bo->bo_flag |= BO_ONWORKLST;
1553  		syncer_worklist_len++;
1554 	}
1555 
1556 	if (delay > syncer_maxdelay - 2)
1557 		delay = syncer_maxdelay - 2;
1558 	slot = (syncer_delayno + delay) & syncer_mask;
1559 
1560 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
1561 	mtx_unlock(&sync_mtx);
1562 }
1563 
1564 static int
1565 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
1566 {
1567 	int error, len;
1568 
1569 	mtx_lock(&sync_mtx);
1570 	len = syncer_worklist_len - sync_vnode_count;
1571 	mtx_unlock(&sync_mtx);
1572 	error = SYSCTL_OUT(req, &len, sizeof(len));
1573 	return (error);
1574 }
1575 
1576 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
1577     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
1578 
1579 static struct proc *updateproc;
1580 static void sched_sync(void);
1581 static struct kproc_desc up_kp = {
1582 	"syncer",
1583 	sched_sync,
1584 	&updateproc
1585 };
1586 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp)
1587 
1588 static int
1589 sync_vnode(struct bufobj *bo, struct thread *td)
1590 {
1591 	struct vnode *vp;
1592 	struct mount *mp;
1593 
1594 	vp = bo->__bo_vnode; 	/* XXX */
1595 	if (VOP_ISLOCKED(vp, NULL) != 0)
1596 		return (1);
1597 	if (VI_TRYLOCK(vp) == 0)
1598 		return (1);
1599 	/*
1600 	 * We use vhold in case the vnode does not
1601 	 * successfully sync.  vhold prevents the vnode from
1602 	 * going away when we unlock the sync_mtx so that
1603 	 * we can acquire the vnode interlock.
1604 	 */
1605 	vholdl(vp);
1606 	mtx_unlock(&sync_mtx);
1607 	VI_UNLOCK(vp);
1608 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1609 		vdrop(vp);
1610 		mtx_lock(&sync_mtx);
1611 		return (1);
1612 	}
1613 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
1614 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
1615 	VOP_UNLOCK(vp, 0, td);
1616 	vn_finished_write(mp);
1617 	VI_LOCK(vp);
1618 	if ((bo->bo_flag & BO_ONWORKLST) != 0) {
1619 		/*
1620 		 * Put us back on the worklist.  The worklist
1621 		 * routine will remove us from our current
1622 		 * position and then add us back in at a later
1623 		 * position.
1624 		 */
1625 		vn_syncer_add_to_worklist(bo, syncdelay);
1626 	}
1627 	vdropl(vp);
1628 	mtx_lock(&sync_mtx);
1629 	return (0);
1630 }
1631 
1632 /*
1633  * System filesystem synchronizer daemon.
1634  */
1635 static void
1636 sched_sync(void)
1637 {
1638 	struct synclist *next;
1639 	struct synclist *slp;
1640 	struct bufobj *bo;
1641 	long starttime;
1642 	struct thread *td = FIRST_THREAD_IN_PROC(updateproc);
1643 	static int dummychan;
1644 	int last_work_seen;
1645 	int net_worklist_len;
1646 	int syncer_final_iter;
1647 	int first_printf;
1648 	int error;
1649 
1650 	mtx_lock(&Giant);
1651 	last_work_seen = 0;
1652 	syncer_final_iter = 0;
1653 	first_printf = 1;
1654 	syncer_state = SYNCER_RUNNING;
1655 	starttime = time_uptime;
1656 	td->td_pflags |= TDP_NORUNNINGBUF;
1657 
1658 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
1659 	    SHUTDOWN_PRI_LAST);
1660 
1661 	mtx_lock(&sync_mtx);
1662 	for (;;) {
1663 		if (syncer_state == SYNCER_FINAL_DELAY &&
1664 		    syncer_final_iter == 0) {
1665 			mtx_unlock(&sync_mtx);
1666 			kthread_suspend_check(td->td_proc);
1667 			mtx_lock(&sync_mtx);
1668 		}
1669 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
1670 		if (syncer_state != SYNCER_RUNNING &&
1671 		    starttime != time_uptime) {
1672 			if (first_printf) {
1673 				printf("\nSyncing disks, vnodes remaining...");
1674 				first_printf = 0;
1675 			}
1676 			printf("%d ", net_worklist_len);
1677 		}
1678 		starttime = time_uptime;
1679 
1680 		/*
1681 		 * Push files whose dirty time has expired.  Be careful
1682 		 * of interrupt race on slp queue.
1683 		 *
1684 		 * Skip over empty worklist slots when shutting down.
1685 		 */
1686 		do {
1687 			slp = &syncer_workitem_pending[syncer_delayno];
1688 			syncer_delayno += 1;
1689 			if (syncer_delayno == syncer_maxdelay)
1690 				syncer_delayno = 0;
1691 			next = &syncer_workitem_pending[syncer_delayno];
1692 			/*
1693 			 * If the worklist has wrapped since the
1694 			 * it was emptied of all but syncer vnodes,
1695 			 * switch to the FINAL_DELAY state and run
1696 			 * for one more second.
1697 			 */
1698 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
1699 			    net_worklist_len == 0 &&
1700 			    last_work_seen == syncer_delayno) {
1701 				syncer_state = SYNCER_FINAL_DELAY;
1702 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
1703 			}
1704 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
1705 		    syncer_worklist_len > 0);
1706 
1707 		/*
1708 		 * Keep track of the last time there was anything
1709 		 * on the worklist other than syncer vnodes.
1710 		 * Return to the SHUTTING_DOWN state if any
1711 		 * new work appears.
1712 		 */
1713 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
1714 			last_work_seen = syncer_delayno;
1715 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
1716 			syncer_state = SYNCER_SHUTTING_DOWN;
1717 		while ((bo = LIST_FIRST(slp)) != NULL) {
1718 			error = sync_vnode(bo, td);
1719 			if (error == 1) {
1720 				LIST_REMOVE(bo, bo_synclist);
1721 				LIST_INSERT_HEAD(next, bo, bo_synclist);
1722 				continue;
1723 			}
1724 		}
1725 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
1726 			syncer_final_iter--;
1727 		/*
1728 		 * The variable rushjob allows the kernel to speed up the
1729 		 * processing of the filesystem syncer process. A rushjob
1730 		 * value of N tells the filesystem syncer to process the next
1731 		 * N seconds worth of work on its queue ASAP. Currently rushjob
1732 		 * is used by the soft update code to speed up the filesystem
1733 		 * syncer process when the incore state is getting so far
1734 		 * ahead of the disk that the kernel memory pool is being
1735 		 * threatened with exhaustion.
1736 		 */
1737 		if (rushjob > 0) {
1738 			rushjob -= 1;
1739 			continue;
1740 		}
1741 		/*
1742 		 * Just sleep for a short period of time between
1743 		 * iterations when shutting down to allow some I/O
1744 		 * to happen.
1745 		 *
1746 		 * If it has taken us less than a second to process the
1747 		 * current work, then wait. Otherwise start right over
1748 		 * again. We can still lose time if any single round
1749 		 * takes more than two seconds, but it does not really
1750 		 * matter as we are just trying to generally pace the
1751 		 * filesystem activity.
1752 		 */
1753 		if (syncer_state != SYNCER_RUNNING)
1754 			msleep(&dummychan, &sync_mtx, PPAUSE, "syncfnl",
1755 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
1756 		else if (time_uptime == starttime)
1757 			msleep(&lbolt, &sync_mtx, PPAUSE, "syncer", 0);
1758 	}
1759 }
1760 
1761 /*
1762  * Request the syncer daemon to speed up its work.
1763  * We never push it to speed up more than half of its
1764  * normal turn time, otherwise it could take over the cpu.
1765  */
1766 int
1767 speedup_syncer(void)
1768 {
1769 	struct thread *td;
1770 	int ret = 0;
1771 
1772 	td = FIRST_THREAD_IN_PROC(updateproc);
1773 	sleepq_remove(td, &lbolt);
1774 	mtx_lock(&sync_mtx);
1775 	if (rushjob < syncdelay / 2) {
1776 		rushjob += 1;
1777 		stat_rush_requests += 1;
1778 		ret = 1;
1779 	}
1780 	mtx_unlock(&sync_mtx);
1781 	return (ret);
1782 }
1783 
1784 /*
1785  * Tell the syncer to speed up its work and run though its work
1786  * list several times, then tell it to shut down.
1787  */
1788 static void
1789 syncer_shutdown(void *arg, int howto)
1790 {
1791 	struct thread *td;
1792 
1793 	if (howto & RB_NOSYNC)
1794 		return;
1795 	td = FIRST_THREAD_IN_PROC(updateproc);
1796 	sleepq_remove(td, &lbolt);
1797 	mtx_lock(&sync_mtx);
1798 	syncer_state = SYNCER_SHUTTING_DOWN;
1799 	rushjob = 0;
1800 	mtx_unlock(&sync_mtx);
1801 	kproc_shutdown(arg, howto);
1802 }
1803 
1804 /*
1805  * Reassign a buffer from one vnode to another.
1806  * Used to assign file specific control information
1807  * (indirect blocks) to the vnode to which they belong.
1808  */
1809 void
1810 reassignbuf(struct buf *bp)
1811 {
1812 	struct vnode *vp;
1813 	struct bufobj *bo;
1814 	int delay;
1815 #ifdef INVARIANTS
1816 	struct bufv *bv;
1817 #endif
1818 
1819 	vp = bp->b_vp;
1820 	bo = bp->b_bufobj;
1821 	++reassignbufcalls;
1822 
1823 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
1824 	    bp, bp->b_vp, bp->b_flags);
1825 	/*
1826 	 * B_PAGING flagged buffers cannot be reassigned because their vp
1827 	 * is not fully linked in.
1828 	 */
1829 	if (bp->b_flags & B_PAGING)
1830 		panic("cannot reassign paging buffer");
1831 
1832 	/*
1833 	 * Delete from old vnode list, if on one.
1834 	 */
1835 	VI_LOCK(vp);
1836 	if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1837 		buf_vlist_remove(bp);
1838 	else
1839 		panic("reassignbuf: Buffer %p not on queue.", bp);
1840 	/*
1841 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
1842 	 * of clean buffers.
1843 	 */
1844 	if (bp->b_flags & B_DELWRI) {
1845 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
1846 			switch (vp->v_type) {
1847 			case VDIR:
1848 				delay = dirdelay;
1849 				break;
1850 			case VCHR:
1851 				delay = metadelay;
1852 				break;
1853 			default:
1854 				delay = filedelay;
1855 			}
1856 			vn_syncer_add_to_worklist(bo, delay);
1857 		}
1858 		buf_vlist_add(bp, bo, BX_VNDIRTY);
1859 	} else {
1860 		buf_vlist_add(bp, bo, BX_VNCLEAN);
1861 
1862 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1863 			mtx_lock(&sync_mtx);
1864 			LIST_REMOVE(bo, bo_synclist);
1865  			syncer_worklist_len--;
1866 			mtx_unlock(&sync_mtx);
1867 			bo->bo_flag &= ~BO_ONWORKLST;
1868 		}
1869 	}
1870 #ifdef INVARIANTS
1871 	bv = &bo->bo_clean;
1872 	bp = TAILQ_FIRST(&bv->bv_hd);
1873 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1874 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1875 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
1876 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1877 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1878 	bv = &bo->bo_dirty;
1879 	bp = TAILQ_FIRST(&bv->bv_hd);
1880 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1881 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1882 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
1883 	KASSERT(bp == NULL || bp->b_bufobj == bo,
1884 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
1885 #endif
1886 	VI_UNLOCK(vp);
1887 }
1888 
1889 /*
1890  * Increment the use and hold counts on the vnode, taking care to reference
1891  * the driver's usecount if this is a chardev.  The vholdl() will remove
1892  * the vnode from the free list if it is presently free.  Requires the
1893  * vnode interlock and returns with it held.
1894  */
1895 static void
1896 v_incr_usecount(struct vnode *vp)
1897 {
1898 
1899 	CTR3(KTR_VFS, "v_incr_usecount: vp %p holdcnt %d usecount %d\n",
1900 	    vp, vp->v_holdcnt, vp->v_usecount);
1901 	vp->v_usecount++;
1902 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1903 		dev_lock();
1904 		vp->v_rdev->si_usecount++;
1905 		dev_unlock();
1906 	}
1907 	vholdl(vp);
1908 }
1909 
1910 /*
1911  * Turn a holdcnt into a use+holdcnt such that only one call to
1912  * v_decr_usecount is needed.
1913  */
1914 static void
1915 v_upgrade_usecount(struct vnode *vp)
1916 {
1917 
1918 	CTR3(KTR_VFS, "v_upgrade_usecount: vp %p holdcnt %d usecount %d\n",
1919 	    vp, vp->v_holdcnt, vp->v_usecount);
1920 	vp->v_usecount++;
1921 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1922 		dev_lock();
1923 		vp->v_rdev->si_usecount++;
1924 		dev_unlock();
1925 	}
1926 }
1927 
1928 /*
1929  * Decrement the vnode use and hold count along with the driver's usecount
1930  * if this is a chardev.  The vdropl() below releases the vnode interlock
1931  * as it may free the vnode.
1932  */
1933 static void
1934 v_decr_usecount(struct vnode *vp)
1935 {
1936 
1937 	CTR3(KTR_VFS, "v_decr_usecount: vp %p holdcnt %d usecount %d\n",
1938 	    vp, vp->v_holdcnt, vp->v_usecount);
1939 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
1940 	VNASSERT(vp->v_usecount > 0, vp,
1941 	    ("v_decr_usecount: negative usecount"));
1942 	vp->v_usecount--;
1943 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1944 		dev_lock();
1945 		vp->v_rdev->si_usecount--;
1946 		dev_unlock();
1947 	}
1948 	vdropl(vp);
1949 }
1950 
1951 /*
1952  * Decrement only the use count and driver use count.  This is intended to
1953  * be paired with a follow on vdropl() to release the remaining hold count.
1954  * In this way we may vgone() a vnode with a 0 usecount without risk of
1955  * having it end up on a free list because the hold count is kept above 0.
1956  */
1957 static void
1958 v_decr_useonly(struct vnode *vp)
1959 {
1960 
1961 	CTR3(KTR_VFS, "v_decr_useonly: vp %p holdcnt %d usecount %d\n",
1962 	    vp, vp->v_holdcnt, vp->v_usecount);
1963 	ASSERT_VI_LOCKED(vp, __FUNCTION__);
1964 	VNASSERT(vp->v_usecount > 0, vp,
1965 	    ("v_decr_useonly: negative usecount"));
1966 	vp->v_usecount--;
1967 	if (vp->v_type == VCHR && vp->v_rdev != NULL) {
1968 		dev_lock();
1969 		vp->v_rdev->si_usecount--;
1970 		dev_unlock();
1971 	}
1972 }
1973 
1974 /*
1975  * Grab a particular vnode from the free list, increment its
1976  * reference count and lock it. The vnode lock bit is set if the
1977  * vnode is being eliminated in vgone. The process is awakened
1978  * when the transition is completed, and an error returned to
1979  * indicate that the vnode is no longer usable (possibly having
1980  * been changed to a new filesystem type).
1981  */
1982 int
1983 vget(struct vnode *vp, int flags, struct thread *td)
1984 {
1985 	int oweinact;
1986 	int oldflags;
1987 	int error;
1988 
1989 	error = 0;
1990 	oldflags = flags;
1991 	oweinact = 0;
1992 	VFS_ASSERT_GIANT(vp->v_mount);
1993 	if ((flags & LK_INTERLOCK) == 0)
1994 		VI_LOCK(vp);
1995 	/*
1996 	 * If the inactive call was deferred because vput() was called
1997 	 * with a shared lock, we have to do it here before another thread
1998 	 * gets a reference to data that should be dead.
1999 	 */
2000 	if (vp->v_iflag & VI_OWEINACT) {
2001 		if (flags & LK_NOWAIT) {
2002 			VI_UNLOCK(vp);
2003 			return (EBUSY);
2004 		}
2005 		flags &= ~LK_TYPE_MASK;
2006 		flags |= LK_EXCLUSIVE;
2007 		oweinact = 1;
2008 	}
2009 	vholdl(vp);
2010 	if ((error = vn_lock(vp, flags | LK_INTERLOCK, td)) != 0) {
2011 		vdrop(vp);
2012 		return (error);
2013 	}
2014 	VI_LOCK(vp);
2015 	/* Upgrade our holdcnt to a usecount. */
2016 	v_upgrade_usecount(vp);
2017 	if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
2018 		panic("vget: vn_lock failed to return ENOENT\n");
2019 	if (oweinact) {
2020 		if (vp->v_iflag & VI_OWEINACT)
2021 			vinactive(vp, td);
2022 		VI_UNLOCK(vp);
2023 		if ((oldflags & LK_TYPE_MASK) == 0)
2024 			VOP_UNLOCK(vp, 0, td);
2025 	} else
2026 		VI_UNLOCK(vp);
2027 	return (0);
2028 }
2029 
2030 /*
2031  * Increase the reference count of a vnode.
2032  */
2033 void
2034 vref(struct vnode *vp)
2035 {
2036 
2037 	VI_LOCK(vp);
2038 	v_incr_usecount(vp);
2039 	VI_UNLOCK(vp);
2040 }
2041 
2042 /*
2043  * Return reference count of a vnode.
2044  *
2045  * The results of this call are only guaranteed when some mechanism other
2046  * than the VI lock is used to stop other processes from gaining references
2047  * to the vnode.  This may be the case if the caller holds the only reference.
2048  * This is also useful when stale data is acceptable as race conditions may
2049  * be accounted for by some other means.
2050  */
2051 int
2052 vrefcnt(struct vnode *vp)
2053 {
2054 	int usecnt;
2055 
2056 	VI_LOCK(vp);
2057 	usecnt = vp->v_usecount;
2058 	VI_UNLOCK(vp);
2059 
2060 	return (usecnt);
2061 }
2062 
2063 
2064 /*
2065  * Vnode put/release.
2066  * If count drops to zero, call inactive routine and return to freelist.
2067  */
2068 void
2069 vrele(struct vnode *vp)
2070 {
2071 	struct thread *td = curthread;	/* XXX */
2072 
2073 	KASSERT(vp != NULL, ("vrele: null vp"));
2074 	VFS_ASSERT_GIANT(vp->v_mount);
2075 
2076 	VI_LOCK(vp);
2077 
2078 	/* Skip this v_writecount check if we're going to panic below. */
2079 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
2080 	    ("vrele: missed vn_close"));
2081 
2082 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2083 	    vp->v_usecount == 1)) {
2084 		v_decr_usecount(vp);
2085 		return;
2086 	}
2087 	if (vp->v_usecount != 1) {
2088 #ifdef DIAGNOSTIC
2089 		vprint("vrele: negative ref count", vp);
2090 #endif
2091 		VI_UNLOCK(vp);
2092 		panic("vrele: negative ref cnt");
2093 	}
2094 	/*
2095 	 * We want to hold the vnode until the inactive finishes to
2096 	 * prevent vgone() races.  We drop the use count here and the
2097 	 * hold count below when we're done.
2098 	 */
2099 	v_decr_useonly(vp);
2100 	/*
2101 	 * We must call VOP_INACTIVE with the node locked. Mark
2102 	 * as VI_DOINGINACT to avoid recursion.
2103 	 */
2104 	vp->v_iflag |= VI_OWEINACT;
2105 	if (vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK, td) == 0) {
2106 		VI_LOCK(vp);
2107 		if (vp->v_usecount > 0)
2108 			vp->v_iflag &= ~VI_OWEINACT;
2109 		if (vp->v_iflag & VI_OWEINACT)
2110 			vinactive(vp, td);
2111 		VOP_UNLOCK(vp, 0, td);
2112 	} else {
2113 		VI_LOCK(vp);
2114 		if (vp->v_usecount > 0)
2115 			vp->v_iflag &= ~VI_OWEINACT;
2116 	}
2117 	vdropl(vp);
2118 }
2119 
2120 /*
2121  * Release an already locked vnode.  This give the same effects as
2122  * unlock+vrele(), but takes less time and avoids releasing and
2123  * re-aquiring the lock (as vrele() aquires the lock internally.)
2124  */
2125 void
2126 vput(struct vnode *vp)
2127 {
2128 	struct thread *td = curthread;	/* XXX */
2129 	int error;
2130 
2131 	KASSERT(vp != NULL, ("vput: null vp"));
2132 	ASSERT_VOP_LOCKED(vp, "vput");
2133 	VFS_ASSERT_GIANT(vp->v_mount);
2134 	VI_LOCK(vp);
2135 	/* Skip this v_writecount check if we're going to panic below. */
2136 	VNASSERT(vp->v_writecount < vp->v_usecount || vp->v_usecount < 1, vp,
2137 	    ("vput: missed vn_close"));
2138 	error = 0;
2139 
2140 	if (vp->v_usecount > 1 || ((vp->v_iflag & VI_DOINGINACT) &&
2141 	    vp->v_usecount == 1)) {
2142 		VOP_UNLOCK(vp, 0, td);
2143 		v_decr_usecount(vp);
2144 		return;
2145 	}
2146 
2147 	if (vp->v_usecount != 1) {
2148 #ifdef DIAGNOSTIC
2149 		vprint("vput: negative ref count", vp);
2150 #endif
2151 		panic("vput: negative ref cnt");
2152 	}
2153 	/*
2154 	 * We want to hold the vnode until the inactive finishes to
2155 	 * prevent vgone() races.  We drop the use count here and the
2156 	 * hold count below when we're done.
2157 	 */
2158 	v_decr_useonly(vp);
2159 	vp->v_iflag |= VI_OWEINACT;
2160 	if (VOP_ISLOCKED(vp, NULL) != LK_EXCLUSIVE) {
2161 		error = VOP_LOCK(vp, LK_EXCLUPGRADE|LK_INTERLOCK|LK_NOWAIT, td);
2162 		VI_LOCK(vp);
2163 		if (error) {
2164 			if (vp->v_usecount > 0)
2165 				vp->v_iflag &= ~VI_OWEINACT;
2166 			goto done;
2167 		}
2168 	}
2169 	if (vp->v_usecount > 0)
2170 		vp->v_iflag &= ~VI_OWEINACT;
2171 	if (vp->v_iflag & VI_OWEINACT)
2172 		vinactive(vp, td);
2173 	VOP_UNLOCK(vp, 0, td);
2174 done:
2175 	vdropl(vp);
2176 }
2177 
2178 /*
2179  * Somebody doesn't want the vnode recycled.
2180  */
2181 void
2182 vhold(struct vnode *vp)
2183 {
2184 
2185 	VI_LOCK(vp);
2186 	vholdl(vp);
2187 	VI_UNLOCK(vp);
2188 }
2189 
2190 void
2191 vholdl(struct vnode *vp)
2192 {
2193 
2194 	vp->v_holdcnt++;
2195 	if (VSHOULDBUSY(vp))
2196 		vbusy(vp);
2197 }
2198 
2199 /*
2200  * Note that there is one less who cares about this vnode.  vdrop() is the
2201  * opposite of vhold().
2202  */
2203 void
2204 vdrop(struct vnode *vp)
2205 {
2206 
2207 	VI_LOCK(vp);
2208 	vdropl(vp);
2209 }
2210 
2211 /*
2212  * Drop the hold count of the vnode.  If this is the last reference to
2213  * the vnode we will free it if it has been vgone'd otherwise it is
2214  * placed on the free list.
2215  */
2216 void
2217 vdropl(struct vnode *vp)
2218 {
2219 
2220 	ASSERT_VI_LOCKED(vp, "vdropl");
2221 	if (vp->v_holdcnt <= 0)
2222 		panic("vdrop: holdcnt %d", vp->v_holdcnt);
2223 	vp->v_holdcnt--;
2224 	if (vp->v_holdcnt == 0) {
2225 		if (vp->v_iflag & VI_DOOMED) {
2226 			vdestroy(vp);
2227 			return;
2228 		} else
2229 			vfree(vp);
2230 	}
2231 	VI_UNLOCK(vp);
2232 }
2233 
2234 /*
2235  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
2236  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
2237  * OWEINACT tracks whether a vnode missed a call to inactive due to a
2238  * failed lock upgrade.
2239  */
2240 static void
2241 vinactive(struct vnode *vp, struct thread *td)
2242 {
2243 
2244 	ASSERT_VOP_LOCKED(vp, "vinactive");
2245 	ASSERT_VI_LOCKED(vp, "vinactive");
2246 	VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2247 	    ("vinactive: recursed on VI_DOINGINACT"));
2248 	vp->v_iflag |= VI_DOINGINACT;
2249 	vp->v_iflag &= ~VI_OWEINACT;
2250 	VI_UNLOCK(vp);
2251 	VOP_INACTIVE(vp, td);
2252 	VI_LOCK(vp);
2253 	VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2254 	    ("vinactive: lost VI_DOINGINACT"));
2255 	vp->v_iflag &= ~VI_DOINGINACT;
2256 }
2257 
2258 /*
2259  * Remove any vnodes in the vnode table belonging to mount point mp.
2260  *
2261  * If FORCECLOSE is not specified, there should not be any active ones,
2262  * return error if any are found (nb: this is a user error, not a
2263  * system error). If FORCECLOSE is specified, detach any active vnodes
2264  * that are found.
2265  *
2266  * If WRITECLOSE is set, only flush out regular file vnodes open for
2267  * writing.
2268  *
2269  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2270  *
2271  * `rootrefs' specifies the base reference count for the root vnode
2272  * of this filesystem. The root vnode is considered busy if its
2273  * v_usecount exceeds this value. On a successful return, vflush(, td)
2274  * will call vrele() on the root vnode exactly rootrefs times.
2275  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2276  * be zero.
2277  */
2278 #ifdef DIAGNOSTIC
2279 static int busyprt = 0;		/* print out busy vnodes */
2280 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
2281 #endif
2282 
2283 int
2284 vflush( struct mount *mp, int rootrefs, int flags, struct thread *td)
2285 {
2286 	struct vnode *vp, *mvp, *rootvp = NULL;
2287 	struct vattr vattr;
2288 	int busy = 0, error;
2289 
2290 	CTR1(KTR_VFS, "vflush: mp %p", mp);
2291 	if (rootrefs > 0) {
2292 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2293 		    ("vflush: bad args"));
2294 		/*
2295 		 * Get the filesystem root vnode. We can vput() it
2296 		 * immediately, since with rootrefs > 0, it won't go away.
2297 		 */
2298 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp, td)) != 0)
2299 			return (error);
2300 		vput(rootvp);
2301 
2302 	}
2303 	MNT_ILOCK(mp);
2304 loop:
2305 	MNT_VNODE_FOREACH(vp, mp, mvp) {
2306 
2307 		VI_LOCK(vp);
2308 		vholdl(vp);
2309 		MNT_IUNLOCK(mp);
2310 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE, td);
2311 		if (error) {
2312 			vdrop(vp);
2313 			MNT_ILOCK(mp);
2314 			MNT_VNODE_FOREACH_ABORT_ILOCKED(mp, mvp);
2315 			goto loop;
2316 		}
2317 		/*
2318 		 * Skip over a vnodes marked VV_SYSTEM.
2319 		 */
2320 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2321 			VOP_UNLOCK(vp, 0, td);
2322 			vdrop(vp);
2323 			MNT_ILOCK(mp);
2324 			continue;
2325 		}
2326 		/*
2327 		 * If WRITECLOSE is set, flush out unlinked but still open
2328 		 * files (even if open only for reading) and regular file
2329 		 * vnodes open for writing.
2330 		 */
2331 		if (flags & WRITECLOSE) {
2332 			error = VOP_GETATTR(vp, &vattr, td->td_ucred, td);
2333 			VI_LOCK(vp);
2334 
2335 			if ((vp->v_type == VNON ||
2336 			    (error == 0 && vattr.va_nlink > 0)) &&
2337 			    (vp->v_writecount == 0 || vp->v_type != VREG)) {
2338 				VOP_UNLOCK(vp, 0, td);
2339 				vdropl(vp);
2340 				MNT_ILOCK(mp);
2341 				continue;
2342 			}
2343 		} else
2344 			VI_LOCK(vp);
2345 		/*
2346 		 * With v_usecount == 0, all we need to do is clear out the
2347 		 * vnode data structures and we are done.
2348 		 *
2349 		 * If FORCECLOSE is set, forcibly close the vnode.
2350 		 */
2351 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
2352 			VNASSERT(vp->v_usecount == 0 ||
2353 			    (vp->v_type != VCHR && vp->v_type != VBLK), vp,
2354 			    ("device VNODE %p is FORCECLOSED", vp));
2355 			vgonel(vp);
2356 		} else {
2357 			busy++;
2358 #ifdef DIAGNOSTIC
2359 			if (busyprt)
2360 				vprint("vflush: busy vnode", vp);
2361 #endif
2362 		}
2363 		VOP_UNLOCK(vp, 0, td);
2364 		vdropl(vp);
2365 		MNT_ILOCK(mp);
2366 	}
2367 	MNT_IUNLOCK(mp);
2368 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
2369 		/*
2370 		 * If just the root vnode is busy, and if its refcount
2371 		 * is equal to `rootrefs', then go ahead and kill it.
2372 		 */
2373 		VI_LOCK(rootvp);
2374 		KASSERT(busy > 0, ("vflush: not busy"));
2375 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
2376 		    ("vflush: usecount %d < rootrefs %d",
2377 		     rootvp->v_usecount, rootrefs));
2378 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
2379 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK, td);
2380 			vgone(rootvp);
2381 			VOP_UNLOCK(rootvp, 0, td);
2382 			busy = 0;
2383 		} else
2384 			VI_UNLOCK(rootvp);
2385 	}
2386 	if (busy)
2387 		return (EBUSY);
2388 	for (; rootrefs > 0; rootrefs--)
2389 		vrele(rootvp);
2390 	return (0);
2391 }
2392 
2393 /*
2394  * Recycle an unused vnode to the front of the free list.
2395  */
2396 int
2397 vrecycle(struct vnode *vp, struct thread *td)
2398 {
2399 	int recycled;
2400 
2401 	ASSERT_VOP_LOCKED(vp, "vrecycle");
2402 	recycled = 0;
2403 	VI_LOCK(vp);
2404 	if (vp->v_usecount == 0) {
2405 		recycled = 1;
2406 		vgonel(vp);
2407 	}
2408 	VI_UNLOCK(vp);
2409 	return (recycled);
2410 }
2411 
2412 /*
2413  * Eliminate all activity associated with a vnode
2414  * in preparation for reuse.
2415  */
2416 void
2417 vgone(struct vnode *vp)
2418 {
2419 	VI_LOCK(vp);
2420 	vgonel(vp);
2421 	VI_UNLOCK(vp);
2422 }
2423 
2424 /*
2425  * vgone, with the vp interlock held.
2426  */
2427 void
2428 vgonel(struct vnode *vp)
2429 {
2430 	struct thread *td;
2431 	int oweinact;
2432 	int active;
2433 	struct mount *mp;
2434 
2435 	CTR1(KTR_VFS, "vgonel: vp %p", vp);
2436 	ASSERT_VOP_LOCKED(vp, "vgonel");
2437 	ASSERT_VI_LOCKED(vp, "vgonel");
2438 	VNASSERT(vp->v_holdcnt, vp,
2439 	    ("vgonel: vp %p has no reference.", vp));
2440 	td = curthread;
2441 
2442 	/*
2443 	 * Don't vgonel if we're already doomed.
2444 	 */
2445 	if (vp->v_iflag & VI_DOOMED)
2446 		return;
2447 	vp->v_iflag |= VI_DOOMED;
2448 	/*
2449 	 * Check to see if the vnode is in use.  If so, we have to call
2450 	 * VOP_CLOSE() and VOP_INACTIVE().
2451 	 */
2452 	active = vp->v_usecount;
2453 	oweinact = (vp->v_iflag & VI_OWEINACT);
2454 	VI_UNLOCK(vp);
2455 	/*
2456 	 * Clean out any buffers associated with the vnode.
2457 	 * If the flush fails, just toss the buffers.
2458 	 */
2459 	mp = NULL;
2460 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
2461 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
2462 	if (vinvalbuf(vp, V_SAVE, td, 0, 0) != 0)
2463 		vinvalbuf(vp, 0, td, 0, 0);
2464 
2465 	/*
2466 	 * If purging an active vnode, it must be closed and
2467 	 * deactivated before being reclaimed.
2468 	 */
2469 	if (active)
2470 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
2471 	if (oweinact || active) {
2472 		VI_LOCK(vp);
2473 		if ((vp->v_iflag & VI_DOINGINACT) == 0)
2474 			vinactive(vp, td);
2475 		VI_UNLOCK(vp);
2476 	}
2477 	/*
2478 	 * Reclaim the vnode.
2479 	 */
2480 	if (VOP_RECLAIM(vp, td))
2481 		panic("vgone: cannot reclaim");
2482 	if (mp != NULL)
2483 		vn_finished_secondary_write(mp);
2484 	VNASSERT(vp->v_object == NULL, vp,
2485 	    ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
2486 	/*
2487 	 * Delete from old mount point vnode list.
2488 	 */
2489 	delmntque(vp);
2490 	cache_purge(vp);
2491 	/*
2492 	 * Done with purge, reset to the standard lock and invalidate
2493 	 * the vnode.
2494 	 */
2495 	VI_LOCK(vp);
2496 	vp->v_vnlock = &vp->v_lock;
2497 	vp->v_op = &dead_vnodeops;
2498 	vp->v_tag = "none";
2499 	vp->v_type = VBAD;
2500 }
2501 
2502 /*
2503  * Calculate the total number of references to a special device.
2504  */
2505 int
2506 vcount(struct vnode *vp)
2507 {
2508 	int count;
2509 
2510 	dev_lock();
2511 	count = vp->v_rdev->si_usecount;
2512 	dev_unlock();
2513 	return (count);
2514 }
2515 
2516 /*
2517  * Same as above, but using the struct cdev *as argument
2518  */
2519 int
2520 count_dev(struct cdev *dev)
2521 {
2522 	int count;
2523 
2524 	dev_lock();
2525 	count = dev->si_usecount;
2526 	dev_unlock();
2527 	return(count);
2528 }
2529 
2530 /*
2531  * Print out a description of a vnode.
2532  */
2533 static char *typename[] =
2534 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
2535  "VMARKER"};
2536 
2537 void
2538 vn_printf(struct vnode *vp, const char *fmt, ...)
2539 {
2540 	va_list ap;
2541 	char buf[96];
2542 
2543 	va_start(ap, fmt);
2544 	vprintf(fmt, ap);
2545 	va_end(ap);
2546 	printf("%p: ", (void *)vp);
2547 	printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
2548 	printf("    usecount %d, writecount %d, refcount %d mountedhere %p\n",
2549 	    vp->v_usecount, vp->v_writecount, vp->v_holdcnt, vp->v_mountedhere);
2550 	buf[0] = '\0';
2551 	buf[1] = '\0';
2552 	if (vp->v_vflag & VV_ROOT)
2553 		strcat(buf, "|VV_ROOT");
2554 	if (vp->v_vflag & VV_TEXT)
2555 		strcat(buf, "|VV_TEXT");
2556 	if (vp->v_vflag & VV_SYSTEM)
2557 		strcat(buf, "|VV_SYSTEM");
2558 	if (vp->v_vflag & VV_DELETED)
2559 		strcat(buf, "|VV_DELETED");
2560 	if (vp->v_iflag & VI_DOOMED)
2561 		strcat(buf, "|VI_DOOMED");
2562 	if (vp->v_iflag & VI_FREE)
2563 		strcat(buf, "|VI_FREE");
2564 	printf("    flags (%s)\n", buf + 1);
2565 	if (mtx_owned(VI_MTX(vp)))
2566 		printf(" VI_LOCKed");
2567 	if (vp->v_object != NULL)
2568 		printf("    v_object %p ref %d pages %d\n",
2569 		    vp->v_object, vp->v_object->ref_count,
2570 		    vp->v_object->resident_page_count);
2571 	printf("    ");
2572 	lockmgr_printinfo(vp->v_vnlock);
2573 	printf("\n");
2574 	if (vp->v_data != NULL)
2575 		VOP_PRINT(vp);
2576 }
2577 
2578 #ifdef DDB
2579 /*
2580  * List all of the locked vnodes in the system.
2581  * Called when debugging the kernel.
2582  */
2583 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
2584 {
2585 	struct mount *mp, *nmp;
2586 	struct vnode *vp;
2587 
2588 	/*
2589 	 * Note: because this is DDB, we can't obey the locking semantics
2590 	 * for these structures, which means we could catch an inconsistent
2591 	 * state and dereference a nasty pointer.  Not much to be done
2592 	 * about that.
2593 	 */
2594 	printf("Locked vnodes\n");
2595 	for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
2596 		nmp = TAILQ_NEXT(mp, mnt_list);
2597 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2598 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp, NULL))
2599 				vprint("", vp);
2600 		}
2601 		nmp = TAILQ_NEXT(mp, mnt_list);
2602 	}
2603 }
2604 
2605 /*
2606  * Show details about the given vnode.
2607  */
2608 DB_SHOW_COMMAND(vnode, db_show_vnode)
2609 {
2610 	struct vnode *vp;
2611 
2612 	if (!have_addr)
2613 		return;
2614 	vp = (struct vnode *)addr;
2615 	vn_printf(vp, "vnode ");
2616 }
2617 #endif	/* DDB */
2618 
2619 /*
2620  * Fill in a struct xvfsconf based on a struct vfsconf.
2621  */
2622 static void
2623 vfsconf2x(struct vfsconf *vfsp, struct xvfsconf *xvfsp)
2624 {
2625 
2626 	strcpy(xvfsp->vfc_name, vfsp->vfc_name);
2627 	xvfsp->vfc_typenum = vfsp->vfc_typenum;
2628 	xvfsp->vfc_refcount = vfsp->vfc_refcount;
2629 	xvfsp->vfc_flags = vfsp->vfc_flags;
2630 	/*
2631 	 * These are unused in userland, we keep them
2632 	 * to not break binary compatibility.
2633 	 */
2634 	xvfsp->vfc_vfsops = NULL;
2635 	xvfsp->vfc_next = NULL;
2636 }
2637 
2638 /*
2639  * Top level filesystem related information gathering.
2640  */
2641 static int
2642 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
2643 {
2644 	struct vfsconf *vfsp;
2645 	struct xvfsconf xvfsp;
2646 	int error;
2647 
2648 	error = 0;
2649 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2650 		bzero(&xvfsp, sizeof(xvfsp));
2651 		vfsconf2x(vfsp, &xvfsp);
2652 		error = SYSCTL_OUT(req, &xvfsp, sizeof xvfsp);
2653 		if (error)
2654 			break;
2655 	}
2656 	return (error);
2657 }
2658 
2659 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLFLAG_RD, NULL, 0, sysctl_vfs_conflist,
2660     "S,xvfsconf", "List of all configured filesystems");
2661 
2662 #ifndef BURN_BRIDGES
2663 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
2664 
2665 static int
2666 vfs_sysctl(SYSCTL_HANDLER_ARGS)
2667 {
2668 	int *name = (int *)arg1 - 1;	/* XXX */
2669 	u_int namelen = arg2 + 1;	/* XXX */
2670 	struct vfsconf *vfsp;
2671 	struct xvfsconf xvfsp;
2672 
2673 	printf("WARNING: userland calling deprecated sysctl, "
2674 	    "please rebuild world\n");
2675 
2676 #if 1 || defined(COMPAT_PRELITE2)
2677 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
2678 	if (namelen == 1)
2679 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
2680 #endif
2681 
2682 	switch (name[1]) {
2683 	case VFS_MAXTYPENUM:
2684 		if (namelen != 2)
2685 			return (ENOTDIR);
2686 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
2687 	case VFS_CONF:
2688 		if (namelen != 3)
2689 			return (ENOTDIR);	/* overloaded */
2690 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list)
2691 			if (vfsp->vfc_typenum == name[2])
2692 				break;
2693 		if (vfsp == NULL)
2694 			return (EOPNOTSUPP);
2695 		bzero(&xvfsp, sizeof(xvfsp));
2696 		vfsconf2x(vfsp, &xvfsp);
2697 		return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
2698 	}
2699 	return (EOPNOTSUPP);
2700 }
2701 
2702 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP,
2703 	vfs_sysctl, "Generic filesystem");
2704 
2705 #if 1 || defined(COMPAT_PRELITE2)
2706 
2707 static int
2708 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
2709 {
2710 	int error;
2711 	struct vfsconf *vfsp;
2712 	struct ovfsconf ovfs;
2713 
2714 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
2715 		bzero(&ovfs, sizeof(ovfs));
2716 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
2717 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
2718 		ovfs.vfc_index = vfsp->vfc_typenum;
2719 		ovfs.vfc_refcount = vfsp->vfc_refcount;
2720 		ovfs.vfc_flags = vfsp->vfc_flags;
2721 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
2722 		if (error)
2723 			return error;
2724 	}
2725 	return 0;
2726 }
2727 
2728 #endif /* 1 || COMPAT_PRELITE2 */
2729 #endif /* !BURN_BRIDGES */
2730 
2731 #define KINFO_VNODESLOP		10
2732 #ifdef notyet
2733 /*
2734  * Dump vnode list (via sysctl).
2735  */
2736 /* ARGSUSED */
2737 static int
2738 sysctl_vnode(SYSCTL_HANDLER_ARGS)
2739 {
2740 	struct xvnode *xvn;
2741 	struct thread *td = req->td;
2742 	struct mount *mp;
2743 	struct vnode *vp;
2744 	int error, len, n;
2745 
2746 	/*
2747 	 * Stale numvnodes access is not fatal here.
2748 	 */
2749 	req->lock = 0;
2750 	len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
2751 	if (!req->oldptr)
2752 		/* Make an estimate */
2753 		return (SYSCTL_OUT(req, 0, len));
2754 
2755 	error = sysctl_wire_old_buffer(req, 0);
2756 	if (error != 0)
2757 		return (error);
2758 	xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
2759 	n = 0;
2760 	mtx_lock(&mountlist_mtx);
2761 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
2762 		if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td))
2763 			continue;
2764 		MNT_ILOCK(mp);
2765 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
2766 			if (n == len)
2767 				break;
2768 			vref(vp);
2769 			xvn[n].xv_size = sizeof *xvn;
2770 			xvn[n].xv_vnode = vp;
2771 			xvn[n].xv_id = 0;	/* XXX compat */
2772 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
2773 			XV_COPY(usecount);
2774 			XV_COPY(writecount);
2775 			XV_COPY(holdcnt);
2776 			XV_COPY(mount);
2777 			XV_COPY(numoutput);
2778 			XV_COPY(type);
2779 #undef XV_COPY
2780 			xvn[n].xv_flag = vp->v_vflag;
2781 
2782 			switch (vp->v_type) {
2783 			case VREG:
2784 			case VDIR:
2785 			case VLNK:
2786 				break;
2787 			case VBLK:
2788 			case VCHR:
2789 				if (vp->v_rdev == NULL) {
2790 					vrele(vp);
2791 					continue;
2792 				}
2793 				xvn[n].xv_dev = dev2udev(vp->v_rdev);
2794 				break;
2795 			case VSOCK:
2796 				xvn[n].xv_socket = vp->v_socket;
2797 				break;
2798 			case VFIFO:
2799 				xvn[n].xv_fifo = vp->v_fifoinfo;
2800 				break;
2801 			case VNON:
2802 			case VBAD:
2803 			default:
2804 				/* shouldn't happen? */
2805 				vrele(vp);
2806 				continue;
2807 			}
2808 			vrele(vp);
2809 			++n;
2810 		}
2811 		MNT_IUNLOCK(mp);
2812 		mtx_lock(&mountlist_mtx);
2813 		vfs_unbusy(mp, td);
2814 		if (n == len)
2815 			break;
2816 	}
2817 	mtx_unlock(&mountlist_mtx);
2818 
2819 	error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
2820 	free(xvn, M_TEMP);
2821 	return (error);
2822 }
2823 
2824 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE|CTLFLAG_RD,
2825 	0, 0, sysctl_vnode, "S,xvnode", "");
2826 #endif
2827 
2828 /*
2829  * Unmount all filesystems. The list is traversed in reverse order
2830  * of mounting to avoid dependencies.
2831  */
2832 void
2833 vfs_unmountall(void)
2834 {
2835 	struct mount *mp;
2836 	struct thread *td;
2837 	int error;
2838 
2839 	KASSERT(curthread != NULL, ("vfs_unmountall: NULL curthread"));
2840 	td = curthread;
2841 	/*
2842 	 * Since this only runs when rebooting, it is not interlocked.
2843 	 */
2844 	while(!TAILQ_EMPTY(&mountlist)) {
2845 		mp = TAILQ_LAST(&mountlist, mntlist);
2846 		error = dounmount(mp, MNT_FORCE, td);
2847 		if (error) {
2848 			TAILQ_REMOVE(&mountlist, mp, mnt_list);
2849 			/*
2850 			 * XXX: Due to the way in which we mount the root
2851 			 * file system off of devfs, devfs will generate a
2852 			 * "busy" warning when we try to unmount it before
2853 			 * the root.  Don't print a warning as a result in
2854 			 * order to avoid false positive errors that may
2855 			 * cause needless upset.
2856 			 */
2857 			if (strcmp(mp->mnt_vfc->vfc_name, "devfs") != 0) {
2858 				printf("unmount of %s failed (",
2859 				    mp->mnt_stat.f_mntonname);
2860 				if (error == EBUSY)
2861 					printf("BUSY)\n");
2862 				else
2863 					printf("%d)\n", error);
2864 			}
2865 		} else {
2866 			/* The unmount has removed mp from the mountlist */
2867 		}
2868 	}
2869 }
2870 
2871 /*
2872  * perform msync on all vnodes under a mount point
2873  * the mount point must be locked.
2874  */
2875 void
2876 vfs_msync(struct mount *mp, int flags)
2877 {
2878 	struct vnode *vp, *mvp;
2879 	struct vm_object *obj;
2880 
2881 	MNT_ILOCK(mp);
2882 	MNT_VNODE_FOREACH(vp, mp, mvp) {
2883 		VI_LOCK(vp);
2884 		if ((vp->v_iflag & VI_OBJDIRTY) &&
2885 		    (flags == MNT_WAIT || VOP_ISLOCKED(vp, NULL) == 0)) {
2886 			MNT_IUNLOCK(mp);
2887 			if (!vget(vp,
2888 			    LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
2889 			    curthread)) {
2890 				if (vp->v_vflag & VV_NOSYNC) {	/* unlinked */
2891 					vput(vp);
2892 					MNT_ILOCK(mp);
2893 					continue;
2894 				}
2895 
2896 				obj = vp->v_object;
2897 				if (obj != NULL) {
2898 					VM_OBJECT_LOCK(obj);
2899 					vm_object_page_clean(obj, 0, 0,
2900 					    flags == MNT_WAIT ?
2901 					    OBJPC_SYNC : OBJPC_NOSYNC);
2902 					VM_OBJECT_UNLOCK(obj);
2903 				}
2904 				vput(vp);
2905 			}
2906 			MNT_ILOCK(mp);
2907 		} else
2908 			VI_UNLOCK(vp);
2909 	}
2910 	MNT_IUNLOCK(mp);
2911 }
2912 
2913 /*
2914  * Mark a vnode as free, putting it up for recycling.
2915  */
2916 static void
2917 vfree(struct vnode *vp)
2918 {
2919 
2920 	CTR1(KTR_VFS, "vfree vp %p", vp);
2921 	ASSERT_VI_LOCKED(vp, "vfree");
2922 	mtx_lock(&vnode_free_list_mtx);
2923 	VNASSERT(vp->v_op != NULL, vp, ("vfree: vnode already reclaimed."));
2924 	VNASSERT((vp->v_iflag & VI_FREE) == 0, vp, ("vnode already free"));
2925 	VNASSERT(VSHOULDFREE(vp), vp, ("vfree: freeing when we shouldn't"));
2926 	VNASSERT((vp->v_iflag & VI_DOOMED) == 0, vp,
2927 	    ("vfree: Freeing doomed vnode"));
2928 	if (vp->v_iflag & VI_AGE) {
2929 		TAILQ_INSERT_HEAD(&vnode_free_list, vp, v_freelist);
2930 	} else {
2931 		TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_freelist);
2932 	}
2933 	freevnodes++;
2934 	vp->v_iflag &= ~VI_AGE;
2935 	vp->v_iflag |= VI_FREE;
2936 	mtx_unlock(&vnode_free_list_mtx);
2937 }
2938 
2939 /*
2940  * Opposite of vfree() - mark a vnode as in use.
2941  */
2942 static void
2943 vbusy(struct vnode *vp)
2944 {
2945 	CTR1(KTR_VFS, "vbusy vp %p", vp);
2946 	ASSERT_VI_LOCKED(vp, "vbusy");
2947 	VNASSERT((vp->v_iflag & VI_FREE) != 0, vp, ("vnode not free"));
2948 	VNASSERT(vp->v_op != NULL, vp, ("vbusy: vnode already reclaimed."));
2949 
2950 	mtx_lock(&vnode_free_list_mtx);
2951 	TAILQ_REMOVE(&vnode_free_list, vp, v_freelist);
2952 	freevnodes--;
2953 	vp->v_iflag &= ~(VI_FREE|VI_AGE);
2954 	mtx_unlock(&vnode_free_list_mtx);
2955 }
2956 
2957 /*
2958  * Initalize per-vnode helper structure to hold poll-related state.
2959  */
2960 void
2961 v_addpollinfo(struct vnode *vp)
2962 {
2963 	struct vpollinfo *vi;
2964 
2965 	vi = uma_zalloc(vnodepoll_zone, M_WAITOK);
2966 	if (vp->v_pollinfo != NULL) {
2967 		uma_zfree(vnodepoll_zone, vi);
2968 		return;
2969 	}
2970 	vp->v_pollinfo = vi;
2971 	mtx_init(&vp->v_pollinfo->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
2972 	knlist_init(&vp->v_pollinfo->vpi_selinfo.si_note, vp, vfs_knllock,
2973 	    vfs_knlunlock, vfs_knllocked);
2974 }
2975 
2976 /*
2977  * Record a process's interest in events which might happen to
2978  * a vnode.  Because poll uses the historic select-style interface
2979  * internally, this routine serves as both the ``check for any
2980  * pending events'' and the ``record my interest in future events''
2981  * functions.  (These are done together, while the lock is held,
2982  * to avoid race conditions.)
2983  */
2984 int
2985 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
2986 {
2987 
2988 	if (vp->v_pollinfo == NULL)
2989 		v_addpollinfo(vp);
2990 	mtx_lock(&vp->v_pollinfo->vpi_lock);
2991 	if (vp->v_pollinfo->vpi_revents & events) {
2992 		/*
2993 		 * This leaves events we are not interested
2994 		 * in available for the other process which
2995 		 * which presumably had requested them
2996 		 * (otherwise they would never have been
2997 		 * recorded).
2998 		 */
2999 		events &= vp->v_pollinfo->vpi_revents;
3000 		vp->v_pollinfo->vpi_revents &= ~events;
3001 
3002 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
3003 		return events;
3004 	}
3005 	vp->v_pollinfo->vpi_events |= events;
3006 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
3007 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
3008 	return 0;
3009 }
3010 
3011 /*
3012  * Routine to create and manage a filesystem syncer vnode.
3013  */
3014 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
3015 static int	sync_fsync(struct  vop_fsync_args *);
3016 static int	sync_inactive(struct  vop_inactive_args *);
3017 static int	sync_reclaim(struct  vop_reclaim_args *);
3018 
3019 static struct vop_vector sync_vnodeops = {
3020 	.vop_bypass =	VOP_EOPNOTSUPP,
3021 	.vop_close =	sync_close,		/* close */
3022 	.vop_fsync =	sync_fsync,		/* fsync */
3023 	.vop_inactive =	sync_inactive,	/* inactive */
3024 	.vop_reclaim =	sync_reclaim,	/* reclaim */
3025 	._vop_lock =	vop_stdlock,	/* lock */
3026 	.vop_unlock =	vop_stdunlock,	/* unlock */
3027 	.vop_islocked =	vop_stdislocked,	/* islocked */
3028 };
3029 
3030 /*
3031  * Create a new filesystem syncer vnode for the specified mount point.
3032  */
3033 int
3034 vfs_allocate_syncvnode(struct mount *mp)
3035 {
3036 	struct vnode *vp;
3037 	static long start, incr, next;
3038 	int error;
3039 
3040 	/* Allocate a new vnode */
3041 	if ((error = getnewvnode("syncer", mp, &sync_vnodeops, &vp)) != 0) {
3042 		mp->mnt_syncer = NULL;
3043 		return (error);
3044 	}
3045 	vp->v_type = VNON;
3046 	error = insmntque(vp, mp);
3047 	if (error != 0)
3048 		panic("vfs_allocate_syncvnode: insmntque failed");
3049 	/*
3050 	 * Place the vnode onto the syncer worklist. We attempt to
3051 	 * scatter them about on the list so that they will go off
3052 	 * at evenly distributed times even if all the filesystems
3053 	 * are mounted at once.
3054 	 */
3055 	next += incr;
3056 	if (next == 0 || next > syncer_maxdelay) {
3057 		start /= 2;
3058 		incr /= 2;
3059 		if (start == 0) {
3060 			start = syncer_maxdelay / 2;
3061 			incr = syncer_maxdelay;
3062 		}
3063 		next = start;
3064 	}
3065 	VI_LOCK(vp);
3066 	vn_syncer_add_to_worklist(&vp->v_bufobj,
3067 	    syncdelay > 0 ? next % syncdelay : 0);
3068 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
3069 	mtx_lock(&sync_mtx);
3070 	sync_vnode_count++;
3071 	mtx_unlock(&sync_mtx);
3072 	VI_UNLOCK(vp);
3073 	mp->mnt_syncer = vp;
3074 	return (0);
3075 }
3076 
3077 /*
3078  * Do a lazy sync of the filesystem.
3079  */
3080 static int
3081 sync_fsync(struct vop_fsync_args *ap)
3082 {
3083 	struct vnode *syncvp = ap->a_vp;
3084 	struct mount *mp = syncvp->v_mount;
3085 	struct thread *td = ap->a_td;
3086 	int error;
3087 	struct bufobj *bo;
3088 
3089 	/*
3090 	 * We only need to do something if this is a lazy evaluation.
3091 	 */
3092 	if (ap->a_waitfor != MNT_LAZY)
3093 		return (0);
3094 
3095 	/*
3096 	 * Move ourselves to the back of the sync list.
3097 	 */
3098 	bo = &syncvp->v_bufobj;
3099 	BO_LOCK(bo);
3100 	vn_syncer_add_to_worklist(bo, syncdelay);
3101 	BO_UNLOCK(bo);
3102 
3103 	/*
3104 	 * Walk the list of vnodes pushing all that are dirty and
3105 	 * not already on the sync list.
3106 	 */
3107 	mtx_lock(&mountlist_mtx);
3108 	if (vfs_busy(mp, LK_EXCLUSIVE | LK_NOWAIT, &mountlist_mtx, td) != 0) {
3109 		mtx_unlock(&mountlist_mtx);
3110 		return (0);
3111 	}
3112 	if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
3113 		vfs_unbusy(mp, td);
3114 		return (0);
3115 	}
3116 	MNT_ILOCK(mp);
3117 	mp->mnt_noasync++;
3118 	mp->mnt_kern_flag &= ~MNTK_ASYNC;
3119 	MNT_IUNLOCK(mp);
3120 	vfs_msync(mp, MNT_NOWAIT);
3121 	error = VFS_SYNC(mp, MNT_LAZY, td);
3122 	MNT_ILOCK(mp);
3123 	mp->mnt_noasync--;
3124 	if ((mp->mnt_flag & MNT_ASYNC) != 0 && mp->mnt_noasync == 0)
3125 		mp->mnt_kern_flag |= MNTK_ASYNC;
3126 	MNT_IUNLOCK(mp);
3127 	vn_finished_write(mp);
3128 	vfs_unbusy(mp, td);
3129 	return (error);
3130 }
3131 
3132 /*
3133  * The syncer vnode is no referenced.
3134  */
3135 static int
3136 sync_inactive(struct vop_inactive_args *ap)
3137 {
3138 
3139 	vgone(ap->a_vp);
3140 	return (0);
3141 }
3142 
3143 /*
3144  * The syncer vnode is no longer needed and is being decommissioned.
3145  *
3146  * Modifications to the worklist must be protected by sync_mtx.
3147  */
3148 static int
3149 sync_reclaim(struct vop_reclaim_args *ap)
3150 {
3151 	struct vnode *vp = ap->a_vp;
3152 	struct bufobj *bo;
3153 
3154 	VI_LOCK(vp);
3155 	bo = &vp->v_bufobj;
3156 	vp->v_mount->mnt_syncer = NULL;
3157 	if (bo->bo_flag & BO_ONWORKLST) {
3158 		mtx_lock(&sync_mtx);
3159 		LIST_REMOVE(bo, bo_synclist);
3160  		syncer_worklist_len--;
3161 		sync_vnode_count--;
3162 		mtx_unlock(&sync_mtx);
3163 		bo->bo_flag &= ~BO_ONWORKLST;
3164 	}
3165 	VI_UNLOCK(vp);
3166 
3167 	return (0);
3168 }
3169 
3170 /*
3171  * Check if vnode represents a disk device
3172  */
3173 int
3174 vn_isdisk(struct vnode *vp, int *errp)
3175 {
3176 	int error;
3177 
3178 	error = 0;
3179 	dev_lock();
3180 	if (vp->v_type != VCHR)
3181 		error = ENOTBLK;
3182 	else if (vp->v_rdev == NULL)
3183 		error = ENXIO;
3184 	else if (vp->v_rdev->si_devsw == NULL)
3185 		error = ENXIO;
3186 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
3187 		error = ENOTBLK;
3188 	dev_unlock();
3189 	if (errp != NULL)
3190 		*errp = error;
3191 	return (error == 0);
3192 }
3193 
3194 /*
3195  * Common filesystem object access control check routine.  Accepts a
3196  * vnode's type, "mode", uid and gid, requested access mode, credentials,
3197  * and optional call-by-reference privused argument allowing vaccess()
3198  * to indicate to the caller whether privilege was used to satisfy the
3199  * request (obsoleted).  Returns 0 on success, or an errno on failure.
3200  *
3201  * The ifdef'd CAPABILITIES version is here for reference, but is not
3202  * actually used.
3203  */
3204 int
3205 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
3206     mode_t acc_mode, struct ucred *cred, int *privused)
3207 {
3208 	mode_t dac_granted;
3209 	mode_t priv_granted;
3210 
3211 	/*
3212 	 * Look for a normal, non-privileged way to access the file/directory
3213 	 * as requested.  If it exists, go with that.
3214 	 */
3215 
3216 	if (privused != NULL)
3217 		*privused = 0;
3218 
3219 	dac_granted = 0;
3220 
3221 	/* Check the owner. */
3222 	if (cred->cr_uid == file_uid) {
3223 		dac_granted |= VADMIN;
3224 		if (file_mode & S_IXUSR)
3225 			dac_granted |= VEXEC;
3226 		if (file_mode & S_IRUSR)
3227 			dac_granted |= VREAD;
3228 		if (file_mode & S_IWUSR)
3229 			dac_granted |= (VWRITE | VAPPEND);
3230 
3231 		if ((acc_mode & dac_granted) == acc_mode)
3232 			return (0);
3233 
3234 		goto privcheck;
3235 	}
3236 
3237 	/* Otherwise, check the groups (first match) */
3238 	if (groupmember(file_gid, cred)) {
3239 		if (file_mode & S_IXGRP)
3240 			dac_granted |= VEXEC;
3241 		if (file_mode & S_IRGRP)
3242 			dac_granted |= VREAD;
3243 		if (file_mode & S_IWGRP)
3244 			dac_granted |= (VWRITE | VAPPEND);
3245 
3246 		if ((acc_mode & dac_granted) == acc_mode)
3247 			return (0);
3248 
3249 		goto privcheck;
3250 	}
3251 
3252 	/* Otherwise, check everyone else. */
3253 	if (file_mode & S_IXOTH)
3254 		dac_granted |= VEXEC;
3255 	if (file_mode & S_IROTH)
3256 		dac_granted |= VREAD;
3257 	if (file_mode & S_IWOTH)
3258 		dac_granted |= (VWRITE | VAPPEND);
3259 	if ((acc_mode & dac_granted) == acc_mode)
3260 		return (0);
3261 
3262 privcheck:
3263 	/*
3264 	 * Build a privilege mask to determine if the set of privileges
3265 	 * satisfies the requirements when combined with the granted mask
3266 	 * from above.  For each privilege, if the privilege is required,
3267 	 * bitwise or the request type onto the priv_granted mask.
3268 	 */
3269 	priv_granted = 0;
3270 
3271 	if (type == VDIR) {
3272 		/*
3273 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
3274 		 * requests, instead of PRIV_VFS_EXEC.
3275 		 */
3276 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3277 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP, SUSER_ALLOWJAIL))
3278 			priv_granted |= VEXEC;
3279 	} else {
3280 		if ((acc_mode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
3281 		    !priv_check_cred(cred, PRIV_VFS_EXEC, SUSER_ALLOWJAIL))
3282 			priv_granted |= VEXEC;
3283 	}
3284 
3285 	if ((acc_mode & VREAD) && ((dac_granted & VREAD) == 0) &&
3286 	    !priv_check_cred(cred, PRIV_VFS_READ, SUSER_ALLOWJAIL))
3287 		priv_granted |= VREAD;
3288 
3289 	if ((acc_mode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
3290 	    !priv_check_cred(cred, PRIV_VFS_WRITE, SUSER_ALLOWJAIL))
3291 		priv_granted |= (VWRITE | VAPPEND);
3292 
3293 	if ((acc_mode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
3294 	    !priv_check_cred(cred, PRIV_VFS_ADMIN, SUSER_ALLOWJAIL))
3295 		priv_granted |= VADMIN;
3296 
3297 	if ((acc_mode & (priv_granted | dac_granted)) == acc_mode) {
3298 		/* XXX audit: privilege used */
3299 		if (privused != NULL)
3300 			*privused = 1;
3301 		return (0);
3302 	}
3303 
3304 	return ((acc_mode & VADMIN) ? EPERM : EACCES);
3305 }
3306 
3307 /*
3308  * Credential check based on process requesting service, and per-attribute
3309  * permissions.
3310  */
3311 int
3312 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
3313     struct thread *td, int access)
3314 {
3315 
3316 	/*
3317 	 * Kernel-invoked always succeeds.
3318 	 */
3319 	if (cred == NOCRED)
3320 		return (0);
3321 
3322 	/*
3323 	 * Do not allow privileged processes in jail to directly manipulate
3324 	 * system attributes.
3325 	 */
3326 	switch (attrnamespace) {
3327 	case EXTATTR_NAMESPACE_SYSTEM:
3328 		/* Potentially should be: return (EPERM); */
3329 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0));
3330 	case EXTATTR_NAMESPACE_USER:
3331 		return (VOP_ACCESS(vp, access, cred, td));
3332 	default:
3333 		return (EPERM);
3334 	}
3335 }
3336 
3337 #ifdef DEBUG_VFS_LOCKS
3338 /*
3339  * This only exists to supress warnings from unlocked specfs accesses.  It is
3340  * no longer ok to have an unlocked VFS.
3341  */
3342 #define	IGNORE_LOCK(vp) ((vp)->v_type == VCHR || (vp)->v_type == VBAD)
3343 
3344 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
3345 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, "");
3346 
3347 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
3348 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, "");
3349 
3350 int vfs_badlock_print = 1;	/* Print lock violations. */
3351 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, "");
3352 
3353 #ifdef KDB
3354 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
3355 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, "");
3356 #endif
3357 
3358 static void
3359 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
3360 {
3361 
3362 #ifdef KDB
3363 	if (vfs_badlock_backtrace)
3364 		kdb_backtrace();
3365 #endif
3366 	if (vfs_badlock_print)
3367 		printf("%s: %p %s\n", str, (void *)vp, msg);
3368 	if (vfs_badlock_ddb)
3369 		kdb_enter("lock violation");
3370 }
3371 
3372 void
3373 assert_vi_locked(struct vnode *vp, const char *str)
3374 {
3375 
3376 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
3377 		vfs_badlock("interlock is not locked but should be", str, vp);
3378 }
3379 
3380 void
3381 assert_vi_unlocked(struct vnode *vp, const char *str)
3382 {
3383 
3384 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
3385 		vfs_badlock("interlock is locked but should not be", str, vp);
3386 }
3387 
3388 void
3389 assert_vop_locked(struct vnode *vp, const char *str)
3390 {
3391 
3392 	if (vp && !IGNORE_LOCK(vp) && VOP_ISLOCKED(vp, NULL) == 0)
3393 		vfs_badlock("is not locked but should be", str, vp);
3394 }
3395 
3396 void
3397 assert_vop_unlocked(struct vnode *vp, const char *str)
3398 {
3399 
3400 	if (vp && !IGNORE_LOCK(vp) &&
3401 	    VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE)
3402 		vfs_badlock("is locked but should not be", str, vp);
3403 }
3404 
3405 void
3406 assert_vop_elocked(struct vnode *vp, const char *str)
3407 {
3408 
3409 	if (vp && !IGNORE_LOCK(vp) &&
3410 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLUSIVE)
3411 		vfs_badlock("is not exclusive locked but should be", str, vp);
3412 }
3413 
3414 #if 0
3415 void
3416 assert_vop_elocked_other(struct vnode *vp, const char *str)
3417 {
3418 
3419 	if (vp && !IGNORE_LOCK(vp) &&
3420 	    VOP_ISLOCKED(vp, curthread) != LK_EXCLOTHER)
3421 		vfs_badlock("is not exclusive locked by another thread",
3422 		    str, vp);
3423 }
3424 
3425 void
3426 assert_vop_slocked(struct vnode *vp, const char *str)
3427 {
3428 
3429 	if (vp && !IGNORE_LOCK(vp) &&
3430 	    VOP_ISLOCKED(vp, curthread) != LK_SHARED)
3431 		vfs_badlock("is not locked shared but should be", str, vp);
3432 }
3433 #endif /* 0 */
3434 #endif /* DEBUG_VFS_LOCKS */
3435 
3436 void
3437 vop_rename_pre(void *ap)
3438 {
3439 	struct vop_rename_args *a = ap;
3440 
3441 #ifdef DEBUG_VFS_LOCKS
3442 	if (a->a_tvp)
3443 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
3444 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
3445 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
3446 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
3447 
3448 	/* Check the source (from). */
3449 	if (a->a_tdvp != a->a_fdvp && a->a_tvp != a->a_fdvp)
3450 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
3451 	if (a->a_tvp != a->a_fvp)
3452 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
3453 
3454 	/* Check the target. */
3455 	if (a->a_tvp)
3456 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
3457 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
3458 #endif
3459 	if (a->a_tdvp != a->a_fdvp)
3460 		vhold(a->a_fdvp);
3461 	if (a->a_tvp != a->a_fvp)
3462 		vhold(a->a_fvp);
3463 	vhold(a->a_tdvp);
3464 	if (a->a_tvp)
3465 		vhold(a->a_tvp);
3466 }
3467 
3468 void
3469 vop_strategy_pre(void *ap)
3470 {
3471 #ifdef DEBUG_VFS_LOCKS
3472 	struct vop_strategy_args *a;
3473 	struct buf *bp;
3474 
3475 	a = ap;
3476 	bp = a->a_bp;
3477 
3478 	/*
3479 	 * Cluster ops lock their component buffers but not the IO container.
3480 	 */
3481 	if ((bp->b_flags & B_CLUSTER) != 0)
3482 		return;
3483 
3484 	if (BUF_REFCNT(bp) < 1) {
3485 		if (vfs_badlock_print)
3486 			printf(
3487 			    "VOP_STRATEGY: bp is not locked but should be\n");
3488 		if (vfs_badlock_ddb)
3489 			kdb_enter("lock violation");
3490 	}
3491 #endif
3492 }
3493 
3494 void
3495 vop_lookup_pre(void *ap)
3496 {
3497 #ifdef DEBUG_VFS_LOCKS
3498 	struct vop_lookup_args *a;
3499 	struct vnode *dvp;
3500 
3501 	a = ap;
3502 	dvp = a->a_dvp;
3503 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3504 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3505 #endif
3506 }
3507 
3508 void
3509 vop_lookup_post(void *ap, int rc)
3510 {
3511 #ifdef DEBUG_VFS_LOCKS
3512 	struct vop_lookup_args *a;
3513 	struct vnode *dvp;
3514 	struct vnode *vp;
3515 
3516 	a = ap;
3517 	dvp = a->a_dvp;
3518 	vp = *(a->a_vpp);
3519 
3520 	ASSERT_VI_UNLOCKED(dvp, "VOP_LOOKUP");
3521 	ASSERT_VOP_LOCKED(dvp, "VOP_LOOKUP");
3522 
3523 	if (!rc)
3524 		ASSERT_VOP_LOCKED(vp, "VOP_LOOKUP (child)");
3525 #endif
3526 }
3527 
3528 void
3529 vop_lock_pre(void *ap)
3530 {
3531 #ifdef DEBUG_VFS_LOCKS
3532 	struct _vop_lock_args *a = ap;
3533 
3534 	if ((a->a_flags & LK_INTERLOCK) == 0)
3535 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3536 	else
3537 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
3538 #endif
3539 }
3540 
3541 void
3542 vop_lock_post(void *ap, int rc)
3543 {
3544 #ifdef DEBUG_VFS_LOCKS
3545 	struct _vop_lock_args *a = ap;
3546 
3547 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
3548 	if (rc == 0)
3549 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
3550 #endif
3551 }
3552 
3553 void
3554 vop_unlock_pre(void *ap)
3555 {
3556 #ifdef DEBUG_VFS_LOCKS
3557 	struct vop_unlock_args *a = ap;
3558 
3559 	if (a->a_flags & LK_INTERLOCK)
3560 		ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
3561 	ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
3562 #endif
3563 }
3564 
3565 void
3566 vop_unlock_post(void *ap, int rc)
3567 {
3568 #ifdef DEBUG_VFS_LOCKS
3569 	struct vop_unlock_args *a = ap;
3570 
3571 	if (a->a_flags & LK_INTERLOCK)
3572 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
3573 #endif
3574 }
3575 
3576 void
3577 vop_create_post(void *ap, int rc)
3578 {
3579 	struct vop_create_args *a = ap;
3580 
3581 	if (!rc)
3582 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3583 }
3584 
3585 void
3586 vop_link_post(void *ap, int rc)
3587 {
3588 	struct vop_link_args *a = ap;
3589 
3590 	if (!rc) {
3591 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
3592 		VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
3593 	}
3594 }
3595 
3596 void
3597 vop_mkdir_post(void *ap, int rc)
3598 {
3599 	struct vop_mkdir_args *a = ap;
3600 
3601 	if (!rc)
3602 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
3603 }
3604 
3605 void
3606 vop_mknod_post(void *ap, int rc)
3607 {
3608 	struct vop_mknod_args *a = ap;
3609 
3610 	if (!rc)
3611 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3612 }
3613 
3614 void
3615 vop_remove_post(void *ap, int rc)
3616 {
3617 	struct vop_remove_args *a = ap;
3618 
3619 	if (!rc) {
3620 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3621 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
3622 	}
3623 }
3624 
3625 void
3626 vop_rename_post(void *ap, int rc)
3627 {
3628 	struct vop_rename_args *a = ap;
3629 
3630 	if (!rc) {
3631 		VFS_KNOTE_UNLOCKED(a->a_fdvp, NOTE_WRITE);
3632 		VFS_KNOTE_UNLOCKED(a->a_tdvp, NOTE_WRITE);
3633 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
3634 		if (a->a_tvp)
3635 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
3636 	}
3637 	if (a->a_tdvp != a->a_fdvp)
3638 		vdrop(a->a_fdvp);
3639 	if (a->a_tvp != a->a_fvp)
3640 		vdrop(a->a_fvp);
3641 	vdrop(a->a_tdvp);
3642 	if (a->a_tvp)
3643 		vdrop(a->a_tvp);
3644 }
3645 
3646 void
3647 vop_rmdir_post(void *ap, int rc)
3648 {
3649 	struct vop_rmdir_args *a = ap;
3650 
3651 	if (!rc) {
3652 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
3653 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
3654 	}
3655 }
3656 
3657 void
3658 vop_setattr_post(void *ap, int rc)
3659 {
3660 	struct vop_setattr_args *a = ap;
3661 
3662 	if (!rc)
3663 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
3664 }
3665 
3666 void
3667 vop_symlink_post(void *ap, int rc)
3668 {
3669 	struct vop_symlink_args *a = ap;
3670 
3671 	if (!rc)
3672 		VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
3673 }
3674 
3675 static struct knlist fs_knlist;
3676 
3677 static void
3678 vfs_event_init(void *arg)
3679 {
3680 	knlist_init(&fs_knlist, NULL, NULL, NULL, NULL);
3681 }
3682 /* XXX - correct order? */
3683 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
3684 
3685 void
3686 vfs_event_signal(fsid_t *fsid, u_int32_t event, intptr_t data __unused)
3687 {
3688 
3689 	KNOTE_UNLOCKED(&fs_knlist, event);
3690 }
3691 
3692 static int	filt_fsattach(struct knote *kn);
3693 static void	filt_fsdetach(struct knote *kn);
3694 static int	filt_fsevent(struct knote *kn, long hint);
3695 
3696 struct filterops fs_filtops =
3697 	{ 0, filt_fsattach, filt_fsdetach, filt_fsevent };
3698 
3699 static int
3700 filt_fsattach(struct knote *kn)
3701 {
3702 
3703 	kn->kn_flags |= EV_CLEAR;
3704 	knlist_add(&fs_knlist, kn, 0);
3705 	return (0);
3706 }
3707 
3708 static void
3709 filt_fsdetach(struct knote *kn)
3710 {
3711 
3712 	knlist_remove(&fs_knlist, kn, 0);
3713 }
3714 
3715 static int
3716 filt_fsevent(struct knote *kn, long hint)
3717 {
3718 
3719 	kn->kn_fflags |= hint;
3720 	return (kn->kn_fflags != 0);
3721 }
3722 
3723 static int
3724 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
3725 {
3726 	struct vfsidctl vc;
3727 	int error;
3728 	struct mount *mp;
3729 
3730 	error = SYSCTL_IN(req, &vc, sizeof(vc));
3731 	if (error)
3732 		return (error);
3733 	if (vc.vc_vers != VFS_CTL_VERS1)
3734 		return (EINVAL);
3735 	mp = vfs_getvfs(&vc.vc_fsid);
3736 	if (mp == NULL)
3737 		return (ENOENT);
3738 	/* ensure that a specific sysctl goes to the right filesystem. */
3739 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
3740 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
3741 		vfs_rel(mp);
3742 		return (EINVAL);
3743 	}
3744 	VCTLTOREQ(&vc, req);
3745 	error = VFS_SYSCTL(mp, vc.vc_op, req);
3746 	vfs_rel(mp);
3747 	return (error);
3748 }
3749 
3750 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLFLAG_WR,
3751         NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid");
3752 
3753 /*
3754  * Function to initialize a va_filerev field sensibly.
3755  * XXX: Wouldn't a random number make a lot more sense ??
3756  */
3757 u_quad_t
3758 init_va_filerev(void)
3759 {
3760 	struct bintime bt;
3761 
3762 	getbinuptime(&bt);
3763 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
3764 }
3765 
3766 static int	filt_vfsread(struct knote *kn, long hint);
3767 static int	filt_vfswrite(struct knote *kn, long hint);
3768 static int	filt_vfsvnode(struct knote *kn, long hint);
3769 static void	filt_vfsdetach(struct knote *kn);
3770 static struct filterops vfsread_filtops =
3771 	{ 1, NULL, filt_vfsdetach, filt_vfsread };
3772 static struct filterops vfswrite_filtops =
3773 	{ 1, NULL, filt_vfsdetach, filt_vfswrite };
3774 static struct filterops vfsvnode_filtops =
3775 	{ 1, NULL, filt_vfsdetach, filt_vfsvnode };
3776 
3777 static void
3778 vfs_knllock(void *arg)
3779 {
3780 	struct vnode *vp = arg;
3781 
3782 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, curthread);
3783 }
3784 
3785 static void
3786 vfs_knlunlock(void *arg)
3787 {
3788 	struct vnode *vp = arg;
3789 
3790 	VOP_UNLOCK(vp, 0, curthread);
3791 }
3792 
3793 static int
3794 vfs_knllocked(void *arg)
3795 {
3796 	struct vnode *vp = arg;
3797 
3798 	return (VOP_ISLOCKED(vp, curthread) == LK_EXCLUSIVE);
3799 }
3800 
3801 int
3802 vfs_kqfilter(struct vop_kqfilter_args *ap)
3803 {
3804 	struct vnode *vp = ap->a_vp;
3805 	struct knote *kn = ap->a_kn;
3806 	struct knlist *knl;
3807 
3808 	switch (kn->kn_filter) {
3809 	case EVFILT_READ:
3810 		kn->kn_fop = &vfsread_filtops;
3811 		break;
3812 	case EVFILT_WRITE:
3813 		kn->kn_fop = &vfswrite_filtops;
3814 		break;
3815 	case EVFILT_VNODE:
3816 		kn->kn_fop = &vfsvnode_filtops;
3817 		break;
3818 	default:
3819 		return (EINVAL);
3820 	}
3821 
3822 	kn->kn_hook = (caddr_t)vp;
3823 
3824 	if (vp->v_pollinfo == NULL)
3825 		v_addpollinfo(vp);
3826 	if (vp->v_pollinfo == NULL)
3827 		return (ENOMEM);
3828 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
3829 	knlist_add(knl, kn, 0);
3830 
3831 	return (0);
3832 }
3833 
3834 /*
3835  * Detach knote from vnode
3836  */
3837 static void
3838 filt_vfsdetach(struct knote *kn)
3839 {
3840 	struct vnode *vp = (struct vnode *)kn->kn_hook;
3841 
3842 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
3843 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
3844 }
3845 
3846 /*ARGSUSED*/
3847 static int
3848 filt_vfsread(struct knote *kn, long hint)
3849 {
3850 	struct vnode *vp = (struct vnode *)kn->kn_hook;
3851 	struct vattr va;
3852 
3853 	/*
3854 	 * filesystem is gone, so set the EOF flag and schedule
3855 	 * the knote for deletion.
3856 	 */
3857 	if (hint == NOTE_REVOKE) {
3858 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3859 		return (1);
3860 	}
3861 
3862 	if (VOP_GETATTR(vp, &va, curthread->td_ucred, curthread))
3863 		return (0);
3864 
3865 	kn->kn_data = va.va_size - kn->kn_fp->f_offset;
3866 	return (kn->kn_data != 0);
3867 }
3868 
3869 /*ARGSUSED*/
3870 static int
3871 filt_vfswrite(struct knote *kn, long hint)
3872 {
3873 	/*
3874 	 * filesystem is gone, so set the EOF flag and schedule
3875 	 * the knote for deletion.
3876 	 */
3877 	if (hint == NOTE_REVOKE)
3878 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
3879 
3880 	kn->kn_data = 0;
3881 	return (1);
3882 }
3883 
3884 static int
3885 filt_vfsvnode(struct knote *kn, long hint)
3886 {
3887 	if (kn->kn_sfflags & hint)
3888 		kn->kn_fflags |= hint;
3889 	if (hint == NOTE_REVOKE) {
3890 		kn->kn_flags |= EV_EOF;
3891 		return (1);
3892 	}
3893 	return (kn->kn_fflags != 0);
3894 }
3895 
3896 int
3897 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
3898 {
3899 	int error;
3900 
3901 	if (dp->d_reclen > ap->a_uio->uio_resid)
3902 		return (ENAMETOOLONG);
3903 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
3904 	if (error) {
3905 		if (ap->a_ncookies != NULL) {
3906 			if (ap->a_cookies != NULL)
3907 				free(ap->a_cookies, M_TEMP);
3908 			ap->a_cookies = NULL;
3909 			*ap->a_ncookies = 0;
3910 		}
3911 		return (error);
3912 	}
3913 	if (ap->a_ncookies == NULL)
3914 		return (0);
3915 
3916 	KASSERT(ap->a_cookies,
3917 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
3918 
3919 	*ap->a_cookies = realloc(*ap->a_cookies,
3920 	    (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
3921 	(*ap->a_cookies)[*ap->a_ncookies] = off;
3922 	return (0);
3923 }
3924 
3925 /*
3926  * Mark for update the access time of the file if the filesystem
3927  * supports VA_MARK_ATIME.  This functionality is used by execve
3928  * and mmap, so we want to avoid the synchronous I/O implied by
3929  * directly setting va_atime for the sake of efficiency.
3930  */
3931 void
3932 vfs_mark_atime(struct vnode *vp, struct thread *td)
3933 {
3934 	struct vattr atimeattr;
3935 
3936 	if ((vp->v_mount->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) {
3937 		VATTR_NULL(&atimeattr);
3938 		atimeattr.va_vaflags |= VA_MARK_ATIME;
3939 		(void)VOP_SETATTR(vp, &atimeattr, td->td_ucred, td);
3940 	}
3941 }
3942