xref: /dflybsd-src/sys/kern/vfs_mount.c (revision cf6a53ca558fa4bbc637ee3949e2436254bcf4c2)
1 /*
2  * Copyright (c) 2004 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  * Copyright (c) 1989, 1993
35  *	The Regents of the University of California.  All rights reserved.
36  * (c) UNIX System Laboratories, Inc.
37  * All or some portions of this file are derived from material licensed
38  * to the University of California by American Telephone and Telegraph
39  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
40  * the permission of UNIX System Laboratories, Inc.
41  *
42  * Redistribution and use in source and binary forms, with or without
43  * modification, are permitted provided that the following conditions
44  * are met:
45  * 1. Redistributions of source code must retain the above copyright
46  *    notice, this list of conditions and the following disclaimer.
47  * 2. Redistributions in binary form must reproduce the above copyright
48  *    notice, this list of conditions and the following disclaimer in the
49  *    documentation and/or other materials provided with the distribution.
50  * 3. Neither the name of the University nor the names of its contributors
51  *    may be used to endorse or promote products derived from this software
52  *    without specific prior written permission.
53  *
54  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
64  * SUCH DAMAGE.
65  */
66 
67 /*
68  * External virtual filesystem routines
69  */
70 
71 #include <sys/param.h>
72 #include <sys/systm.h>
73 #include <sys/kernel.h>
74 #include <sys/malloc.h>
75 #include <sys/mount.h>
76 #include <sys/proc.h>
77 #include <sys/vnode.h>
78 #include <sys/buf.h>
79 #include <sys/eventhandler.h>
80 #include <sys/kthread.h>
81 #include <sys/sysctl.h>
82 
83 #include <machine/limits.h>
84 
85 #include <sys/buf2.h>
86 #include <sys/thread2.h>
87 #include <sys/sysref2.h>
88 
89 #include <vm/vm.h>
90 #include <vm/vm_object.h>
91 
92 struct mountscan_info {
93 	TAILQ_ENTRY(mountscan_info) msi_entry;
94 	int msi_how;
95 	struct mount *msi_node;
96 };
97 
98 struct vmntvnodescan_info {
99 	TAILQ_ENTRY(vmntvnodescan_info) entry;
100 	struct vnode *vp;
101 };
102 
103 struct vnlru_info {
104 	int	pass;
105 };
106 
107 static int vnlru_nowhere = 0;
108 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RD,
109 	    &vnlru_nowhere, 0,
110 	    "Number of times the vnlru process ran without success");
111 
112 
113 static struct lwkt_token mntid_token;
114 static struct mount dummymount;
115 
116 /* note: mountlist exported to pstat */
117 struct mntlist mountlist = TAILQ_HEAD_INITIALIZER(mountlist);
118 static TAILQ_HEAD(,mountscan_info) mountscan_list;
119 static struct lwkt_token mountlist_token;
120 
121 static TAILQ_HEAD(,bio_ops) bio_ops_list = TAILQ_HEAD_INITIALIZER(bio_ops_list);
122 
123 /*
124  * Called from vfsinit()
125  */
126 void
127 vfs_mount_init(void)
128 {
129 	lwkt_token_init(&mountlist_token, "mntlist");
130 	lwkt_token_init(&mntid_token, "mntid");
131 	TAILQ_INIT(&mountscan_list);
132 	mount_init(&dummymount);
133 	dummymount.mnt_flag |= MNT_RDONLY;
134 	dummymount.mnt_kern_flag |= MNTK_ALL_MPSAFE;
135 }
136 
137 /*
138  * Support function called to remove a vnode from the mountlist and
139  * deal with side effects for scans in progress.
140  *
141  * Target mnt_token is held on call.
142  */
143 static void
144 vremovevnodemnt(struct vnode *vp)
145 {
146         struct vmntvnodescan_info *info;
147 	struct mount *mp = vp->v_mount;
148 
149 	TAILQ_FOREACH(info, &mp->mnt_vnodescan_list, entry) {
150 		if (info->vp == vp)
151 			info->vp = TAILQ_NEXT(vp, v_nmntvnodes);
152 	}
153 	TAILQ_REMOVE(&vp->v_mount->mnt_nvnodelist, vp, v_nmntvnodes);
154 }
155 
156 /*
157  * Allocate a new vnode and associate it with a tag, mount point, and
158  * operations vector.
159  *
160  * A VX locked and refd vnode is returned.  The caller should setup the
161  * remaining fields and vx_put() or, if he wishes to leave a vref,
162  * vx_unlock() the vnode.
163  */
164 int
165 getnewvnode(enum vtagtype tag, struct mount *mp,
166 		struct vnode **vpp, int lktimeout, int lkflags)
167 {
168 	struct vnode *vp;
169 
170 	KKASSERT(mp != NULL);
171 
172 	vp = allocvnode(lktimeout, lkflags);
173 	vp->v_tag = tag;
174 	vp->v_data = NULL;
175 
176 	/*
177 	 * By default the vnode is assigned the mount point's normal
178 	 * operations vector.
179 	 */
180 	vp->v_ops = &mp->mnt_vn_use_ops;
181 
182 	/*
183 	 * Placing the vnode on the mount point's queue makes it visible.
184 	 * VNON prevents it from being messed with, however.
185 	 */
186 	insmntque(vp, mp);
187 
188 	/*
189 	 * A VX locked & refd vnode is returned.
190 	 */
191 	*vpp = vp;
192 	return (0);
193 }
194 
195 /*
196  * This function creates vnodes with special operations vectors.  The
197  * mount point is optional.
198  *
199  * This routine is being phased out but is still used by vfs_conf to
200  * create vnodes for devices prior to the root mount (with mp == NULL).
201  */
202 int
203 getspecialvnode(enum vtagtype tag, struct mount *mp,
204 		struct vop_ops **ops,
205 		struct vnode **vpp, int lktimeout, int lkflags)
206 {
207 	struct vnode *vp;
208 
209 	vp = allocvnode(lktimeout, lkflags);
210 	vp->v_tag = tag;
211 	vp->v_data = NULL;
212 	vp->v_ops = ops;
213 
214 	if (mp == NULL)
215 		mp = &dummymount;
216 
217 	/*
218 	 * Placing the vnode on the mount point's queue makes it visible.
219 	 * VNON prevents it from being messed with, however.
220 	 */
221 	insmntque(vp, mp);
222 
223 	/*
224 	 * A VX locked & refd vnode is returned.
225 	 */
226 	*vpp = vp;
227 	return (0);
228 }
229 
230 /*
231  * Interlock against an unmount, return 0 on success, non-zero on failure.
232  *
233  * The passed flag may be 0 or LK_NOWAIT and is only used if an unmount
234  * is in-progress.
235  *
236  * If no unmount is in-progress LK_NOWAIT is ignored.  No other flag bits
237  * are used.  A shared locked will be obtained and the filesystem will not
238  * be unmountable until the lock is released.
239  */
240 int
241 vfs_busy(struct mount *mp, int flags)
242 {
243 	int lkflags;
244 
245 	atomic_add_int(&mp->mnt_refs, 1);
246 	lwkt_gettoken(&mp->mnt_token);
247 	if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
248 		if (flags & LK_NOWAIT) {
249 			lwkt_reltoken(&mp->mnt_token);
250 			atomic_add_int(&mp->mnt_refs, -1);
251 			return (ENOENT);
252 		}
253 		/* XXX not MP safe */
254 		mp->mnt_kern_flag |= MNTK_MWAIT;
255 		/*
256 		 * Since all busy locks are shared except the exclusive
257 		 * lock granted when unmounting, the only place that a
258 		 * wakeup needs to be done is at the release of the
259 		 * exclusive lock at the end of dounmount.
260 		 */
261 		tsleep((caddr_t)mp, 0, "vfs_busy", 0);
262 		lwkt_reltoken(&mp->mnt_token);
263 		atomic_add_int(&mp->mnt_refs, -1);
264 		return (ENOENT);
265 	}
266 	lkflags = LK_SHARED;
267 	if (lockmgr(&mp->mnt_lock, lkflags))
268 		panic("vfs_busy: unexpected lock failure");
269 	lwkt_reltoken(&mp->mnt_token);
270 	return (0);
271 }
272 
273 /*
274  * Free a busy filesystem.
275  *
276  * Decrement refs before releasing the lock so e.g. a pending umount
277  * doesn't give us an unexpected busy error.
278  */
279 void
280 vfs_unbusy(struct mount *mp)
281 {
282 	atomic_add_int(&mp->mnt_refs, -1);
283 	lockmgr(&mp->mnt_lock, LK_RELEASE);
284 }
285 
286 /*
287  * Lookup a filesystem type, and if found allocate and initialize
288  * a mount structure for it.
289  *
290  * Devname is usually updated by mount(8) after booting.
291  */
292 int
293 vfs_rootmountalloc(char *fstypename, char *devname, struct mount **mpp)
294 {
295 	struct vfsconf *vfsp;
296 	struct mount *mp;
297 
298 	if (fstypename == NULL)
299 		return (ENODEV);
300 
301 	vfsp = vfsconf_find_by_name(fstypename);
302 	if (vfsp == NULL)
303 		return (ENODEV);
304 	mp = kmalloc(sizeof(struct mount), M_MOUNT, M_WAITOK | M_ZERO);
305 	mount_init(mp);
306 	lockinit(&mp->mnt_lock, "vfslock", VLKTIMEOUT, 0);
307 
308 	vfs_busy(mp, 0);
309 	mp->mnt_vfc = vfsp;
310 	mp->mnt_op = vfsp->vfc_vfsops;
311 	vfsp->vfc_refcount++;
312 	mp->mnt_stat.f_type = vfsp->vfc_typenum;
313 	mp->mnt_flag |= MNT_RDONLY;
314 	mp->mnt_flag |= vfsp->vfc_flags & MNT_VISFLAGMASK;
315 	strncpy(mp->mnt_stat.f_fstypename, vfsp->vfc_name, MFSNAMELEN);
316 	copystr(devname, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, 0);
317 	*mpp = mp;
318 	return (0);
319 }
320 
321 /*
322  * Basic mount structure initialization
323  */
324 void
325 mount_init(struct mount *mp)
326 {
327 	lockinit(&mp->mnt_lock, "vfslock", hz*5, 0);
328 	lwkt_token_init(&mp->mnt_token, "permnt");
329 
330 	TAILQ_INIT(&mp->mnt_vnodescan_list);
331 	TAILQ_INIT(&mp->mnt_nvnodelist);
332 	TAILQ_INIT(&mp->mnt_reservedvnlist);
333 	TAILQ_INIT(&mp->mnt_jlist);
334 	mp->mnt_nvnodelistsize = 0;
335 	mp->mnt_flag = 0;
336 	mp->mnt_iosize_max = MAXPHYS;
337 	vn_syncer_thr_create(mp);
338 }
339 
340 /*
341  * Lookup a mount point by filesystem identifier.
342  */
343 struct mount *
344 vfs_getvfs(fsid_t *fsid)
345 {
346 	struct mount *mp;
347 
348 	lwkt_gettoken(&mountlist_token);
349 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
350 		if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
351 		    mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
352 			break;
353 		}
354 	}
355 	lwkt_reltoken(&mountlist_token);
356 	return (mp);
357 }
358 
359 /*
360  * Get a new unique fsid.  Try to make its val[0] unique, since this value
361  * will be used to create fake device numbers for stat().  Also try (but
362  * not so hard) make its val[0] unique mod 2^16, since some emulators only
363  * support 16-bit device numbers.  We end up with unique val[0]'s for the
364  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
365  *
366  * Keep in mind that several mounts may be running in parallel.  Starting
367  * the search one past where the previous search terminated is both a
368  * micro-optimization and a defense against returning the same fsid to
369  * different mounts.
370  */
371 void
372 vfs_getnewfsid(struct mount *mp)
373 {
374 	static u_int16_t mntid_base;
375 	fsid_t tfsid;
376 	int mtype;
377 
378 	lwkt_gettoken(&mntid_token);
379 	mtype = mp->mnt_vfc->vfc_typenum;
380 	tfsid.val[1] = mtype;
381 	mtype = (mtype & 0xFF) << 24;
382 	for (;;) {
383 		tfsid.val[0] = makeudev(255,
384 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
385 		mntid_base++;
386 		if (vfs_getvfs(&tfsid) == NULL)
387 			break;
388 	}
389 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
390 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
391 	lwkt_reltoken(&mntid_token);
392 }
393 
394 /*
395  * Set the FSID for a new mount point to the template.  Adjust
396  * the FSID to avoid collisions.
397  */
398 int
399 vfs_setfsid(struct mount *mp, fsid_t *template)
400 {
401 	int didmunge = 0;
402 
403 	bzero(&mp->mnt_stat.f_fsid, sizeof(mp->mnt_stat.f_fsid));
404 	for (;;) {
405 		if (vfs_getvfs(template) == NULL)
406 			break;
407 		didmunge = 1;
408 		++template->val[1];
409 	}
410 	mp->mnt_stat.f_fsid = *template;
411 	return(didmunge);
412 }
413 
414 /*
415  * This routine is called when we have too many vnodes.  It attempts
416  * to free <count> vnodes and will potentially free vnodes that still
417  * have VM backing store (VM backing store is typically the cause
418  * of a vnode blowout so we want to do this).  Therefore, this operation
419  * is not considered cheap.
420  *
421  * A number of conditions may prevent a vnode from being reclaimed.
422  * the buffer cache may have references on the vnode, a directory
423  * vnode may still have references due to the namei cache representing
424  * underlying files, or the vnode may be in active use.   It is not
425  * desireable to reuse such vnodes.  These conditions may cause the
426  * number of vnodes to reach some minimum value regardless of what
427  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
428  */
429 
430 /*
431  * This is a quick non-blocking check to determine if the vnode is a good
432  * candidate for being (eventually) vgone()'d.  Returns 0 if the vnode is
433  * not a good candidate, 1 if it is.
434  */
435 static __inline int
436 vmightfree(struct vnode *vp, int page_count, int pass)
437 {
438 	if (vp->v_flag & VRECLAIMED)
439 		return (0);
440 #if 0
441 	if ((vp->v_flag & VFREE) && TAILQ_EMPTY(&vp->v_namecache))
442 		return (0);
443 #endif
444 	if (sysref_isactive(&vp->v_sysref))
445 		return (0);
446 	if (vp->v_object && vp->v_object->resident_page_count >= page_count)
447 		return (0);
448 
449 	/*
450 	 * XXX horrible hack.  Up to four passes will be taken.  Each pass
451 	 * makes a larger set of vnodes eligible.  For now what this really
452 	 * means is that we try to recycle files opened only once before
453 	 * recycling files opened multiple times.
454 	 */
455 	switch(vp->v_flag & (VAGE0 | VAGE1)) {
456 	case 0:
457 		if (pass < 3)
458 			return(0);
459 		break;
460 	case VAGE0:
461 		if (pass < 2)
462 			return(0);
463 		break;
464 	case VAGE1:
465 		if (pass < 1)
466 			return(0);
467 		break;
468 	case VAGE0 | VAGE1:
469 		break;
470 	}
471 	return (1);
472 }
473 
474 /*
475  * The vnode was found to be possibly vgone()able and the caller has locked it
476  * (thus the usecount should be 1 now).  Determine if the vnode is actually
477  * vgone()able, doing some cleanups in the process.  Returns 1 if the vnode
478  * can be vgone()'d, 0 otherwise.
479  *
480  * Note that v_auxrefs may be non-zero because (A) this vnode is not a leaf
481  * in the namecache topology and (B) this vnode has buffer cache bufs.
482  * We cannot remove vnodes with non-leaf namecache associations.  We do a
483  * tentitive leaf check prior to attempting to flush out any buffers but the
484  * 'real' test when all is said in done is that v_auxrefs must become 0 for
485  * the vnode to be freeable.
486  *
487  * We could theoretically just unconditionally flush when v_auxrefs != 0,
488  * but flushing data associated with non-leaf nodes (which are always
489  * directories), just throws it away for no benefit.  It is the buffer
490  * cache's responsibility to choose buffers to recycle from the cached
491  * data point of view.
492  */
493 static int
494 visleaf(struct vnode *vp)
495 {
496 	struct namecache *ncp;
497 
498 	spin_lock(&vp->v_spin);
499 	TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
500 		if (!TAILQ_EMPTY(&ncp->nc_list)) {
501 			spin_unlock(&vp->v_spin);
502 			return(0);
503 		}
504 	}
505 	spin_unlock(&vp->v_spin);
506 	return(1);
507 }
508 
509 /*
510  * Try to clean up the vnode to the point where it can be vgone()'d, returning
511  * 0 if it cannot be vgone()'d (or already has been), 1 if it can.  Unlike
512  * vmightfree() this routine may flush the vnode and block.  Vnodes marked
513  * VFREE are still candidates for vgone()ing because they may hold namecache
514  * resources and could be blocking the namecache directory hierarchy (and
515  * related vnodes) from being freed.
516  */
517 static int
518 vtrytomakegoneable(struct vnode *vp, int page_count)
519 {
520 	if (vp->v_flag & VRECLAIMED)
521 		return (0);
522 	if (vp->v_sysref.refcnt > 1)
523 		return (0);
524 	if (vp->v_object && vp->v_object->resident_page_count >= page_count)
525 		return (0);
526 	if (vp->v_auxrefs && visleaf(vp)) {
527 		vinvalbuf(vp, V_SAVE, 0, 0);
528 #if 0	/* DEBUG */
529 		kprintf((vp->v_auxrefs ? "vrecycle: vp %p failed: %s\n" :
530 			"vrecycle: vp %p succeeded: %s\n"), vp,
531 			(TAILQ_FIRST(&vp->v_namecache) ?
532 			    TAILQ_FIRST(&vp->v_namecache)->nc_name : "?"));
533 #endif
534 	}
535 
536 	/*
537 	 * This sequence may seem a little strange, but we need to optimize
538 	 * the critical path a bit.  We can't recycle vnodes with other
539 	 * references and because we are trying to recycle an otherwise
540 	 * perfectly fine vnode we have to invalidate the namecache in a
541 	 * way that avoids possible deadlocks (since the vnode lock is being
542 	 * held here).  Finally, we have to check for other references one
543 	 * last time in case something snuck in during the inval.
544 	 */
545 	if (vp->v_sysref.refcnt > 1 || vp->v_auxrefs != 0)
546 		return (0);
547 	if (cache_inval_vp_nonblock(vp))
548 		return (0);
549 	return (vp->v_sysref.refcnt <= 1 && vp->v_auxrefs == 0);
550 }
551 
552 /*
553  * Reclaim up to 1/10 of the vnodes associated with a mount point.  Try
554  * to avoid vnodes which have lots of resident pages (we are trying to free
555  * vnodes, not memory).
556  *
557  * This routine is a callback from the mountlist scan.  The mount point
558  * in question will be busied.
559  *
560  * NOTE: The 1/10 reclamation also ensures that the inactive data set
561  *	 (the vnodes being recycled by the one-time use) does not degenerate
562  *	 into too-small a set.  This is important because once a vnode is
563  *	 marked as not being one-time-use (VAGE0/VAGE1 both 0) that vnode
564  *	 will not be destroyed EXCEPT by this mechanism.  VM pages can still
565  *	 be cleaned/freed by the pageout daemon.
566  */
567 static int
568 vlrureclaim(struct mount *mp, void *data)
569 {
570 	struct vnlru_info *info = data;
571 	struct vnode *vp;
572 	int done;
573 	int trigger;
574 	int usevnodes;
575 	int count;
576 	int trigger_mult = vnlru_nowhere;
577 
578 	/*
579 	 * Calculate the trigger point for the resident pages check.  The
580 	 * minimum trigger value is approximately the number of pages in
581 	 * the system divded by the number of vnodes.  However, due to
582 	 * various other system memory overheads unrelated to data caching
583 	 * it is a good idea to double the trigger (at least).
584 	 *
585 	 * trigger_mult starts at 0.  If the recycler is having problems
586 	 * finding enough freeable vnodes it will increase trigger_mult.
587 	 * This should not happen in normal operation, even on machines with
588 	 * low amounts of memory, but extraordinary memory use by the system
589 	 * verses the amount of cached data can trigger it.
590 	 *
591 	 * (long) -> deal with 64 bit machines, intermediate overflow
592 	 */
593 	usevnodes = desiredvnodes;
594 	if (usevnodes <= 0)
595 		usevnodes = 1;
596 	trigger = (long)vmstats.v_page_count * (trigger_mult + 2) / usevnodes;
597 
598 	done = 0;
599 	lwkt_gettoken(&mp->mnt_token);
600 	count = mp->mnt_nvnodelistsize / 10 + 1;
601 
602 	while (count && mp->mnt_syncer) {
603 		/*
604 		 * Next vnode.  Use the special syncer vnode to placemark
605 		 * the LRU.  This way the LRU code does not interfere with
606 		 * vmntvnodescan().
607 		 */
608 		vp = TAILQ_NEXT(mp->mnt_syncer, v_nmntvnodes);
609 		TAILQ_REMOVE(&mp->mnt_nvnodelist, mp->mnt_syncer, v_nmntvnodes);
610 		if (vp) {
611 			TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp,
612 					   mp->mnt_syncer, v_nmntvnodes);
613 		} else {
614 			TAILQ_INSERT_HEAD(&mp->mnt_nvnodelist, mp->mnt_syncer,
615 					  v_nmntvnodes);
616 			vp = TAILQ_NEXT(mp->mnt_syncer, v_nmntvnodes);
617 			if (vp == NULL)
618 				break;
619 		}
620 
621 		/*
622 		 * __VNODESCAN__
623 		 *
624 		 * The VP will stick around while we hold mnt_token,
625 		 * at least until we block, so we can safely do an initial
626 		 * check, and then must check again after we lock the vnode.
627 		 */
628 		if (vp->v_type == VNON ||	/* syncer or indeterminant */
629 		    !vmightfree(vp, trigger, info->pass) /* critical path opt */
630 		) {
631 			--count;
632 			continue;
633 		}
634 
635 		/*
636 		 * VX get the candidate vnode.  If the VX get fails the
637 		 * vnode might still be on the mountlist.  Our loop depends
638 		 * on us at least cycling the vnode to the end of the
639 		 * mountlist.
640 		 */
641 		if (vx_get_nonblock(vp) != 0) {
642 			--count;
643 			continue;
644 		}
645 
646 		/*
647 		 * Since we blocked locking the vp, make sure it is still
648 		 * a candidate for reclamation.  That is, it has not already
649 		 * been reclaimed and only has our VX reference associated
650 		 * with it.
651 		 */
652 		if (vp->v_type == VNON ||	/* syncer or indeterminant */
653 		    (vp->v_flag & VRECLAIMED) ||
654 		    vp->v_mount != mp ||
655 		    !vtrytomakegoneable(vp, trigger)	/* critical path opt */
656 		) {
657 			--count;
658 			vx_put(vp);
659 			continue;
660 		}
661 
662 		/*
663 		 * All right, we are good, move the vp to the end of the
664 		 * mountlist and clean it out.  The vget will have returned
665 		 * an error if the vnode was destroyed (VRECLAIMED set), so we
666 		 * do not have to check again.  The vput() will move the
667 		 * vnode to the free list if the vgone() was successful.
668 		 */
669 		KKASSERT(vp->v_mount == mp);
670 		vgone_vxlocked(vp);
671 		vx_put(vp);
672 		++done;
673 		--count;
674 	}
675 	lwkt_reltoken(&mp->mnt_token);
676 	return (done);
677 }
678 
679 /*
680  * Attempt to recycle vnodes in a context that is always safe to block.
681  * Calling vlrurecycle() from the bowels of file system code has some
682  * interesting deadlock problems.
683  */
684 static struct thread *vnlruthread;
685 
686 static void
687 vnlru_proc(void)
688 {
689 	struct thread *td = curthread;
690 	struct vnlru_info info;
691 	int done;
692 
693 	EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, td,
694 			      SHUTDOWN_PRI_FIRST);
695 
696 	for (;;) {
697 		kproc_suspend_loop();
698 
699 		/*
700 		 * Do some opportunistic roving.
701 		 */
702 		if (numvnodes > 100000)
703 			vnode_free_rover_scan(50);
704 		else if (numvnodes > 10000)
705 			vnode_free_rover_scan(20);
706 		else
707 			vnode_free_rover_scan(5);
708 
709 		/*
710 		 * Try to free some vnodes if we have too many
711 		 *
712 		 * (long) -> deal with 64 bit machines, intermediate overflow
713 		 */
714 		if (numvnodes > desiredvnodes &&
715 		    freevnodes > desiredvnodes * 2 / 10) {
716 			int count = numvnodes - desiredvnodes;
717 
718 			if (count > freevnodes / 100)
719 				count = freevnodes / 100;
720 			if (count < 5)
721 				count = 5;
722 			freesomevnodes(count);
723 		}
724 
725 		/*
726 		 * Do non-critical-path (more robust) cache cleaning,
727 		 * even if vnode counts are nominal, to try to avoid
728 		 * having to do it in the critical path.
729 		 */
730 		cache_hysteresis(0);
731 
732 		/*
733 		 * Nothing to do if most of our vnodes are already on
734 		 * the free list.
735 		 */
736 		if (numvnodes - freevnodes <= (long)desiredvnodes * 9 / 10) {
737 			tsleep(vnlruthread, 0, "vlruwt", hz);
738 			continue;
739 		}
740 
741 		/*
742 		 * The pass iterates through the four combinations of
743 		 * VAGE0/VAGE1.  We want to get rid of aged small files
744 		 * first.
745 		 */
746 		info.pass = 0;
747 		done = 0;
748 		while (done == 0 && info.pass < 4) {
749 			done = mountlist_scan(vlrureclaim, &info,
750 					      MNTSCAN_FORWARD);
751 			++info.pass;
752 		}
753 
754 		/*
755 		 * The vlrureclaim() call only processes 1/10 of the vnodes
756 		 * on each mount.  If we couldn't find any repeat the loop
757 		 * at least enough times to cover all available vnodes before
758 		 * we start sleeping.  Complain if the failure extends past
759 		 * 30 second, every 30 seconds.
760 		 */
761 		if (done == 0) {
762 			++vnlru_nowhere;
763 			if (vnlru_nowhere % 10 == 0)
764 				tsleep(vnlruthread, 0, "vlrup", hz * 3);
765 			if (vnlru_nowhere % 100 == 0)
766 				kprintf("vnlru_proc: vnode recycler stopped working!\n");
767 			if (vnlru_nowhere == 1000)
768 				vnlru_nowhere = 900;
769 		} else {
770 			vnlru_nowhere = 0;
771 		}
772 	}
773 }
774 
775 /*
776  * MOUNTLIST FUNCTIONS
777  */
778 
779 /*
780  * mountlist_insert (MP SAFE)
781  *
782  * Add a new mount point to the mount list.
783  */
784 void
785 mountlist_insert(struct mount *mp, int how)
786 {
787 	lwkt_gettoken(&mountlist_token);
788 	if (how == MNTINS_FIRST)
789 	    TAILQ_INSERT_HEAD(&mountlist, mp, mnt_list);
790 	else
791 	    TAILQ_INSERT_TAIL(&mountlist, mp, mnt_list);
792 	lwkt_reltoken(&mountlist_token);
793 }
794 
795 /*
796  * mountlist_interlock (MP SAFE)
797  *
798  * Execute the specified interlock function with the mountlist token
799  * held.  The function will be called in a serialized fashion verses
800  * other functions called through this mechanism.
801  */
802 int
803 mountlist_interlock(int (*callback)(struct mount *), struct mount *mp)
804 {
805 	int error;
806 
807 	lwkt_gettoken(&mountlist_token);
808 	error = callback(mp);
809 	lwkt_reltoken(&mountlist_token);
810 	return (error);
811 }
812 
813 /*
814  * mountlist_boot_getfirst (DURING BOOT ONLY)
815  *
816  * This function returns the first mount on the mountlist, which is
817  * expected to be the root mount.  Since no interlocks are obtained
818  * this function is only safe to use during booting.
819  */
820 
821 struct mount *
822 mountlist_boot_getfirst(void)
823 {
824 	return(TAILQ_FIRST(&mountlist));
825 }
826 
827 /*
828  * mountlist_remove (MP SAFE)
829  *
830  * Remove a node from the mountlist.  If this node is the next scan node
831  * for any active mountlist scans, the active mountlist scan will be
832  * adjusted to skip the node, thus allowing removals during mountlist
833  * scans.
834  */
835 void
836 mountlist_remove(struct mount *mp)
837 {
838 	struct mountscan_info *msi;
839 
840 	lwkt_gettoken(&mountlist_token);
841 	TAILQ_FOREACH(msi, &mountscan_list, msi_entry) {
842 		if (msi->msi_node == mp) {
843 			if (msi->msi_how & MNTSCAN_FORWARD)
844 				msi->msi_node = TAILQ_NEXT(mp, mnt_list);
845 			else
846 				msi->msi_node = TAILQ_PREV(mp, mntlist, mnt_list);
847 		}
848 	}
849 	TAILQ_REMOVE(&mountlist, mp, mnt_list);
850 	lwkt_reltoken(&mountlist_token);
851 }
852 
853 /*
854  * mountlist_exists (MP SAFE)
855  *
856  * Checks if a node exists in the mountlist.
857  * This function is mainly used by VFS quota code to check if a
858  * cached nullfs struct mount pointer is still valid at use time
859  *
860  * FIXME: there is no warranty the mp passed to that function
861  * will be the same one used by VFS_ACCOUNT() later
862  */
863 int
864 mountlist_exists(struct mount *mp)
865 {
866 	int node_exists = 0;
867 	struct mount* lmp;
868 
869 	lwkt_gettoken(&mountlist_token);
870 	TAILQ_FOREACH(lmp, &mountlist, mnt_list) {
871 		if (lmp == mp) {
872 			node_exists = 1;
873 			break;
874 		}
875 	}
876 	lwkt_reltoken(&mountlist_token);
877 	return(node_exists);
878 }
879 
880 /*
881  * mountlist_scan (MP SAFE)
882  *
883  * Safely scan the mount points on the mount list.  Unless otherwise
884  * specified each mount point will be busied prior to the callback and
885  * unbusied afterwords.  The callback may safely remove any mount point
886  * without interfering with the scan.  If the current callback
887  * mount is removed the scanner will not attempt to unbusy it.
888  *
889  * If a mount node cannot be busied it is silently skipped.
890  *
891  * The callback return value is aggregated and a total is returned.  A return
892  * value of < 0 is not aggregated and will terminate the scan.
893  *
894  * MNTSCAN_FORWARD	- the mountlist is scanned in the forward direction
895  * MNTSCAN_REVERSE	- the mountlist is scanned in reverse
896  * MNTSCAN_NOBUSY	- the scanner will make the callback without busying
897  *			  the mount node.
898  */
899 int
900 mountlist_scan(int (*callback)(struct mount *, void *), void *data, int how)
901 {
902 	struct mountscan_info info;
903 	struct mount *mp;
904 	int count;
905 	int res;
906 
907 	lwkt_gettoken(&mountlist_token);
908 
909 	info.msi_how = how;
910 	info.msi_node = NULL;	/* paranoia */
911 	TAILQ_INSERT_TAIL(&mountscan_list, &info, msi_entry);
912 
913 	res = 0;
914 
915 	if (how & MNTSCAN_FORWARD) {
916 		info.msi_node = TAILQ_FIRST(&mountlist);
917 		while ((mp = info.msi_node) != NULL) {
918 			if (how & MNTSCAN_NOBUSY) {
919 				count = callback(mp, data);
920 			} else if (vfs_busy(mp, LK_NOWAIT) == 0) {
921 				count = callback(mp, data);
922 				if (mp == info.msi_node)
923 					vfs_unbusy(mp);
924 			} else {
925 				count = 0;
926 			}
927 			if (count < 0)
928 				break;
929 			res += count;
930 			if (mp == info.msi_node)
931 				info.msi_node = TAILQ_NEXT(mp, mnt_list);
932 		}
933 	} else if (how & MNTSCAN_REVERSE) {
934 		info.msi_node = TAILQ_LAST(&mountlist, mntlist);
935 		while ((mp = info.msi_node) != NULL) {
936 			if (how & MNTSCAN_NOBUSY) {
937 				count = callback(mp, data);
938 			} else if (vfs_busy(mp, LK_NOWAIT) == 0) {
939 				count = callback(mp, data);
940 				if (mp == info.msi_node)
941 					vfs_unbusy(mp);
942 			} else {
943 				count = 0;
944 			}
945 			if (count < 0)
946 				break;
947 			res += count;
948 			if (mp == info.msi_node)
949 				info.msi_node = TAILQ_PREV(mp, mntlist, mnt_list);
950 		}
951 	}
952 	TAILQ_REMOVE(&mountscan_list, &info, msi_entry);
953 	lwkt_reltoken(&mountlist_token);
954 	return(res);
955 }
956 
957 /*
958  * MOUNT RELATED VNODE FUNCTIONS
959  */
960 
961 static struct kproc_desc vnlru_kp = {
962 	"vnlru",
963 	vnlru_proc,
964 	&vnlruthread
965 };
966 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
967 
968 /*
969  * Move a vnode from one mount queue to another.
970  *
971  * MPSAFE
972  */
973 void
974 insmntque(struct vnode *vp, struct mount *mp)
975 {
976 	struct mount *omp;
977 
978 	/*
979 	 * Delete from old mount point vnode list, if on one.
980 	 */
981 	if ((omp = vp->v_mount) != NULL) {
982 		lwkt_gettoken(&omp->mnt_token);
983 		KKASSERT(omp == vp->v_mount);
984 		KASSERT(omp->mnt_nvnodelistsize > 0,
985 			("bad mount point vnode list size"));
986 		vremovevnodemnt(vp);
987 		omp->mnt_nvnodelistsize--;
988 		lwkt_reltoken(&omp->mnt_token);
989 	}
990 
991 	/*
992 	 * Insert into list of vnodes for the new mount point, if available.
993 	 * The 'end' of the LRU list is the vnode prior to mp->mnt_syncer.
994 	 */
995 	if (mp == NULL) {
996 		vp->v_mount = NULL;
997 		return;
998 	}
999 	lwkt_gettoken(&mp->mnt_token);
1000 	vp->v_mount = mp;
1001 	if (mp->mnt_syncer) {
1002 		TAILQ_INSERT_BEFORE(mp->mnt_syncer, vp, v_nmntvnodes);
1003 	} else {
1004 		TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1005 	}
1006 	mp->mnt_nvnodelistsize++;
1007 	lwkt_reltoken(&mp->mnt_token);
1008 }
1009 
1010 
1011 /*
1012  * Scan the vnodes under a mount point and issue appropriate callbacks.
1013  *
1014  * The fastfunc() callback is called with just the mountlist token held
1015  * (no vnode lock).  It may not block and the vnode may be undergoing
1016  * modifications while the caller is processing it.  The vnode will
1017  * not be entirely destroyed, however, due to the fact that the mountlist
1018  * token is held.  A return value < 0 skips to the next vnode without calling
1019  * the slowfunc(), a return value > 0 terminates the loop.
1020  *
1021  * The slowfunc() callback is called after the vnode has been successfully
1022  * locked based on passed flags.  The vnode is skipped if it gets rearranged
1023  * or destroyed while blocking on the lock.  A non-zero return value from
1024  * the slow function terminates the loop.  The slow function is allowed to
1025  * arbitrarily block.  The scanning code guarentees consistency of operation
1026  * even if the slow function deletes or moves the node, or blocks and some
1027  * other thread deletes or moves the node.
1028  *
1029  * NOTE: We hold vmobj_token to prevent a VM object from being destroyed
1030  *	 out from under the fastfunc()'s vnode test.  It will not prevent
1031  *	 v_object from getting NULL'd out but it will ensure that the
1032  *	 pointer (if we race) will remain stable.  Only needed when
1033  *	 fastfunc is non-NULL.
1034  */
1035 int
1036 vmntvnodescan(
1037     struct mount *mp,
1038     int flags,
1039     int (*fastfunc)(struct mount *mp, struct vnode *vp, void *data),
1040     int (*slowfunc)(struct mount *mp, struct vnode *vp, void *data),
1041     void *data
1042 ) {
1043 	struct vmntvnodescan_info info;
1044 	struct vnode *vp;
1045 	int r = 0;
1046 	int maxcount = mp->mnt_nvnodelistsize * 2;
1047 	int stopcount = 0;
1048 	int count = 0;
1049 
1050 	lwkt_gettoken(&mp->mnt_token);
1051 	if (fastfunc)
1052 		lwkt_gettoken(&vmobj_token);
1053 
1054 	/*
1055 	 * If asked to do one pass stop after iterating available vnodes.
1056 	 * Under heavy loads new vnodes can be added while we are scanning,
1057 	 * so this isn't perfect.  Create a slop factor of 2x.
1058 	 */
1059 	if (flags & VMSC_ONEPASS)
1060 		stopcount = mp->mnt_nvnodelistsize;
1061 
1062 	info.vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
1063 	TAILQ_INSERT_TAIL(&mp->mnt_vnodescan_list, &info, entry);
1064 
1065 	while ((vp = info.vp) != NULL) {
1066 		if (--maxcount == 0) {
1067 			kprintf("Warning: excessive fssync iteration\n");
1068 			maxcount = mp->mnt_nvnodelistsize * 2;
1069 		}
1070 
1071 		/*
1072 		 * Skip if visible but not ready, or special (e.g.
1073 		 * mp->mnt_syncer)
1074 		 */
1075 		if (vp->v_type == VNON)
1076 			goto next;
1077 		KKASSERT(vp->v_mount == mp);
1078 
1079 		/*
1080 		 * Quick test.  A negative return continues the loop without
1081 		 * calling the slow test.  0 continues onto the slow test.
1082 		 * A positive number aborts the loop.
1083 		 */
1084 		if (fastfunc) {
1085 			if ((r = fastfunc(mp, vp, data)) < 0) {
1086 				r = 0;
1087 				goto next;
1088 			}
1089 			if (r)
1090 				break;
1091 		}
1092 
1093 		/*
1094 		 * Get a vxlock on the vnode, retry if it has moved or isn't
1095 		 * in the mountlist where we expect it.
1096 		 */
1097 		if (slowfunc) {
1098 			int error;
1099 
1100 			switch(flags & (VMSC_GETVP|VMSC_GETVX|VMSC_NOWAIT)) {
1101 			case VMSC_GETVP:
1102 				error = vget(vp, LK_EXCLUSIVE);
1103 				break;
1104 			case VMSC_GETVP|VMSC_NOWAIT:
1105 				error = vget(vp, LK_EXCLUSIVE|LK_NOWAIT);
1106 				break;
1107 			case VMSC_GETVX:
1108 				vx_get(vp);
1109 				error = 0;
1110 				break;
1111 			default:
1112 				error = 0;
1113 				break;
1114 			}
1115 			if (error)
1116 				goto next;
1117 			/*
1118 			 * Do not call the slow function if the vnode is
1119 			 * invalid or if it was ripped out from under us
1120 			 * while we (potentially) blocked.
1121 			 */
1122 			if (info.vp == vp && vp->v_type != VNON)
1123 				r = slowfunc(mp, vp, data);
1124 
1125 			/*
1126 			 * Cleanup
1127 			 */
1128 			switch(flags & (VMSC_GETVP|VMSC_GETVX|VMSC_NOWAIT)) {
1129 			case VMSC_GETVP:
1130 			case VMSC_GETVP|VMSC_NOWAIT:
1131 				vput(vp);
1132 				break;
1133 			case VMSC_GETVX:
1134 				vx_put(vp);
1135 				break;
1136 			default:
1137 				break;
1138 			}
1139 			if (r != 0)
1140 				break;
1141 		}
1142 
1143 next:
1144 		/*
1145 		 * Yield after some processing.  Depending on the number
1146 		 * of vnodes, we might wind up running for a long time.
1147 		 * Because threads are not preemptable, time critical
1148 		 * userland processes might starve.  Give them a chance
1149 		 * now and then.
1150 		 */
1151 		if (++count == 10000) {
1152 			/*
1153 			 * We really want to yield a bit, so we simply
1154 			 * sleep a tick
1155 			 */
1156 			tsleep(mp, 0, "vnodescn", 1);
1157 			count = 0;
1158 		}
1159 
1160 		/*
1161 		 * If doing one pass this decrements to zero.  If it starts
1162 		 * at zero it is effectively unlimited for the purposes of
1163 		 * this loop.
1164 		 */
1165 		if (--stopcount == 0)
1166 			break;
1167 
1168 		/*
1169 		 * Iterate.  If the vnode was ripped out from under us
1170 		 * info.vp will already point to the next vnode, otherwise
1171 		 * we have to obtain the next valid vnode ourselves.
1172 		 */
1173 		if (info.vp == vp)
1174 			info.vp = TAILQ_NEXT(vp, v_nmntvnodes);
1175 	}
1176 
1177 	TAILQ_REMOVE(&mp->mnt_vnodescan_list, &info, entry);
1178 	if (fastfunc)
1179 		lwkt_reltoken(&vmobj_token);
1180 	lwkt_reltoken(&mp->mnt_token);
1181 	return(r);
1182 }
1183 
1184 /*
1185  * Remove any vnodes in the vnode table belonging to mount point mp.
1186  *
1187  * If FORCECLOSE is not specified, there should not be any active ones,
1188  * return error if any are found (nb: this is a user error, not a
1189  * system error). If FORCECLOSE is specified, detach any active vnodes
1190  * that are found.
1191  *
1192  * If WRITECLOSE is set, only flush out regular file vnodes open for
1193  * writing.
1194  *
1195  * SKIPSYSTEM causes any vnodes marked VSYSTEM to be skipped.
1196  *
1197  * `rootrefs' specifies the base reference count for the root vnode
1198  * of this filesystem. The root vnode is considered busy if its
1199  * v_sysref.refcnt exceeds this value. On a successful return, vflush()
1200  * will call vrele() on the root vnode exactly rootrefs times.
1201  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
1202  * be zero.
1203  */
1204 #ifdef DIAGNOSTIC
1205 static int busyprt = 0;		/* print out busy vnodes */
1206 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
1207 #endif
1208 
1209 static int vflush_scan(struct mount *mp, struct vnode *vp, void *data);
1210 
1211 struct vflush_info {
1212 	int flags;
1213 	int busy;
1214 	thread_t td;
1215 };
1216 
1217 int
1218 vflush(struct mount *mp, int rootrefs, int flags)
1219 {
1220 	struct thread *td = curthread;	/* XXX */
1221 	struct vnode *rootvp = NULL;
1222 	int error;
1223 	struct vflush_info vflush_info;
1224 
1225 	if (rootrefs > 0) {
1226 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
1227 		    ("vflush: bad args"));
1228 		/*
1229 		 * Get the filesystem root vnode. We can vput() it
1230 		 * immediately, since with rootrefs > 0, it won't go away.
1231 		 */
1232 		if ((error = VFS_ROOT(mp, &rootvp)) != 0) {
1233 			if ((flags & FORCECLOSE) == 0)
1234 				return (error);
1235 			rootrefs = 0;
1236 			/* continue anyway */
1237 		}
1238 		if (rootrefs)
1239 			vput(rootvp);
1240 	}
1241 
1242 	vflush_info.busy = 0;
1243 	vflush_info.flags = flags;
1244 	vflush_info.td = td;
1245 	vmntvnodescan(mp, VMSC_GETVX, NULL, vflush_scan, &vflush_info);
1246 
1247 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
1248 		/*
1249 		 * If just the root vnode is busy, and if its refcount
1250 		 * is equal to `rootrefs', then go ahead and kill it.
1251 		 */
1252 		KASSERT(vflush_info.busy > 0, ("vflush: not busy"));
1253 		KASSERT(rootvp->v_sysref.refcnt >= rootrefs, ("vflush: rootrefs"));
1254 		if (vflush_info.busy == 1 && rootvp->v_sysref.refcnt == rootrefs) {
1255 			vx_lock(rootvp);
1256 			vgone_vxlocked(rootvp);
1257 			vx_unlock(rootvp);
1258 			vflush_info.busy = 0;
1259 		}
1260 	}
1261 	if (vflush_info.busy)
1262 		return (EBUSY);
1263 	for (; rootrefs > 0; rootrefs--)
1264 		vrele(rootvp);
1265 	return (0);
1266 }
1267 
1268 /*
1269  * The scan callback is made with an VX locked vnode.
1270  */
1271 static int
1272 vflush_scan(struct mount *mp, struct vnode *vp, void *data)
1273 {
1274 	struct vflush_info *info = data;
1275 	struct vattr vattr;
1276 	int flags = info->flags;
1277 
1278 	/*
1279 	 * Skip over a vnodes marked VSYSTEM.
1280 	 */
1281 	if ((flags & SKIPSYSTEM) && (vp->v_flag & VSYSTEM)) {
1282 		return(0);
1283 	}
1284 
1285 	/*
1286 	 * Do not force-close VCHR or VBLK vnodes
1287 	 */
1288 	if (vp->v_type == VCHR || vp->v_type == VBLK)
1289 		flags &= ~(WRITECLOSE|FORCECLOSE);
1290 
1291 	/*
1292 	 * If WRITECLOSE is set, flush out unlinked but still open
1293 	 * files (even if open only for reading) and regular file
1294 	 * vnodes open for writing.
1295 	 */
1296 	if ((flags & WRITECLOSE) &&
1297 	    (vp->v_type == VNON ||
1298 	    (VOP_GETATTR(vp, &vattr) == 0 &&
1299 	    vattr.va_nlink > 0)) &&
1300 	    (vp->v_writecount == 0 || vp->v_type != VREG)) {
1301 		return(0);
1302 	}
1303 
1304 	/*
1305 	 * If we are the only holder (refcnt of 1) or the vnode is in
1306 	 * termination (refcnt < 0), we can vgone the vnode.
1307 	 */
1308 	if (vp->v_sysref.refcnt <= 1) {
1309 		vgone_vxlocked(vp);
1310 		return(0);
1311 	}
1312 
1313 	/*
1314 	 * If FORCECLOSE is set, forcibly destroy the vnode and then move
1315 	 * it to a dummymount structure so vop_*() functions don't deref
1316 	 * a NULL pointer.
1317 	 */
1318 	if (flags & FORCECLOSE) {
1319 		vhold(vp);
1320 		vgone_vxlocked(vp);
1321 		if (vp->v_mount == NULL)
1322 			insmntque(vp, &dummymount);
1323 		vdrop(vp);
1324 		return(0);
1325 	}
1326 	if (vp->v_type == VCHR || vp->v_type == VBLK)
1327 		kprintf("vflush: Warning, cannot destroy busy device vnode\n");
1328 #ifdef DIAGNOSTIC
1329 	if (busyprt)
1330 		vprint("vflush: busy vnode", vp);
1331 #endif
1332 	++info->busy;
1333 	return(0);
1334 }
1335 
1336 void
1337 add_bio_ops(struct bio_ops *ops)
1338 {
1339 	TAILQ_INSERT_TAIL(&bio_ops_list, ops, entry);
1340 }
1341 
1342 void
1343 rem_bio_ops(struct bio_ops *ops)
1344 {
1345 	TAILQ_REMOVE(&bio_ops_list, ops, entry);
1346 }
1347 
1348 /*
1349  * This calls the bio_ops io_sync function either for a mount point
1350  * or generally.
1351  *
1352  * WARNING: softdeps is weirdly coded and just isn't happy unless
1353  * io_sync is called with a NULL mount from the general syncing code.
1354  */
1355 void
1356 bio_ops_sync(struct mount *mp)
1357 {
1358 	struct bio_ops *ops;
1359 
1360 	if (mp) {
1361 		if ((ops = mp->mnt_bioops) != NULL)
1362 			ops->io_sync(mp);
1363 	} else {
1364 		TAILQ_FOREACH(ops, &bio_ops_list, entry) {
1365 			ops->io_sync(NULL);
1366 		}
1367 	}
1368 }
1369 
1370 /*
1371  * Lookup a mount point by nch
1372  */
1373 struct mount *
1374 mount_get_by_nc(struct namecache *ncp)
1375 {
1376 	struct mount *mp = NULL;
1377 
1378 	lwkt_gettoken(&mountlist_token);
1379 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
1380 		if (ncp == mp->mnt_ncmountpt.ncp)
1381 			break;
1382 	}
1383 	lwkt_reltoken(&mountlist_token);
1384 	return (mp);
1385 }
1386 
1387