2 * Copyright (c) 2004 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
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
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.
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
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.
42 * Redistribution and use in source and binary forms, with or without
43 * modification, are permitted provided that the following conditions
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. All advertising materials mentioning features or use of this software
51 * must display the following acknowledgement:
52 * This product includes software developed by the University of
53 * California, Berkeley and its contributors.
54 * 4. Neither the name of the University nor the names of its contributors
55 * may be used to endorse or promote products derived from this software
56 * without specific prior written permission.
58 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
59 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
60 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
61 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
62 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
63 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
64 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
65 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
66 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
67 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
72 * External virtual filesystem routines
75 #include <sys/param.h>
76 #include <sys/systm.h>
77 #include <sys/kernel.h>
78 #include <sys/malloc.h>
79 #include <sys/mount.h>
81 #include <sys/vnode.h>
83 #include <sys/eventhandler.h>
84 #include <sys/kthread.h>
85 #include <sys/sysctl.h>
87 #include <machine/limits.h>
90 #include <sys/thread2.h>
91 #include <sys/sysref2.h>
94 #include <vm/vm_object.h>
96 struct mountscan_info {
97 TAILQ_ENTRY(mountscan_info) msi_entry;
99 struct mount *msi_node;
102 struct vmntvnodescan_info {
103 TAILQ_ENTRY(vmntvnodescan_info) entry;
111 static int vnlru_nowhere = 0;
112 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RD,
114 "Number of times the vnlru process ran without success");
117 static struct lwkt_token mntid_token;
118 static struct mount dummymount;
120 /* note: mountlist exported to pstat */
121 struct mntlist mountlist = TAILQ_HEAD_INITIALIZER(mountlist);
122 static TAILQ_HEAD(,mountscan_info) mountscan_list;
123 static struct lwkt_token mountlist_token;
124 static TAILQ_HEAD(,vmntvnodescan_info) mntvnodescan_list;
125 struct lwkt_token mntvnode_token;
127 static TAILQ_HEAD(,bio_ops) bio_ops_list = TAILQ_HEAD_INITIALIZER(bio_ops_list);
130 * Called from vfsinit()
135 lwkt_token_init(&mountlist_token, "mntlist");
136 lwkt_token_init(&mntvnode_token, "mntvnode");
137 lwkt_token_init(&mntid_token, "mntid");
138 TAILQ_INIT(&mountscan_list);
139 TAILQ_INIT(&mntvnodescan_list);
140 mount_init(&dummymount);
141 dummymount.mnt_flag |= MNT_RDONLY;
142 dummymount.mnt_kern_flag |= MNTK_ALL_MPSAFE;
146 * Support function called with mntvnode_token held to remove a vnode
147 * from the mountlist. We must update any list scans which are in progress.
150 vremovevnodemnt(struct vnode *vp)
152 struct vmntvnodescan_info *info;
154 TAILQ_FOREACH(info, &mntvnodescan_list, entry) {
156 info->vp = TAILQ_NEXT(vp, v_nmntvnodes);
158 TAILQ_REMOVE(&vp->v_mount->mnt_nvnodelist, vp, v_nmntvnodes);
162 * Allocate a new vnode and associate it with a tag, mount point, and
165 * A VX locked and refd vnode is returned. The caller should setup the
166 * remaining fields and vx_put() or, if he wishes to leave a vref,
167 * vx_unlock() the vnode.
170 getnewvnode(enum vtagtype tag, struct mount *mp,
171 struct vnode **vpp, int lktimeout, int lkflags)
175 KKASSERT(mp != NULL);
177 vp = allocvnode(lktimeout, lkflags);
182 * By default the vnode is assigned the mount point's normal
185 vp->v_ops = &mp->mnt_vn_use_ops;
188 * Placing the vnode on the mount point's queue makes it visible.
189 * VNON prevents it from being messed with, however.
194 * A VX locked & refd vnode is returned.
201 * This function creates vnodes with special operations vectors. The
202 * mount point is optional.
204 * This routine is being phased out but is still used by vfs_conf to
205 * create vnodes for devices prior to the root mount (with mp == NULL).
208 getspecialvnode(enum vtagtype tag, struct mount *mp,
209 struct vop_ops **ops,
210 struct vnode **vpp, int lktimeout, int lkflags)
214 vp = allocvnode(lktimeout, lkflags);
223 * Placing the vnode on the mount point's queue makes it visible.
224 * VNON prevents it from being messed with, however.
229 * A VX locked & refd vnode is returned.
236 * Interlock against an unmount, return 0 on success, non-zero on failure.
238 * The passed flag may be 0 or LK_NOWAIT and is only used if an unmount
241 * If no unmount is in-progress LK_NOWAIT is ignored. No other flag bits
242 * are used. A shared locked will be obtained and the filesystem will not
243 * be unmountable until the lock is released.
246 vfs_busy(struct mount *mp, int flags)
250 atomic_add_int(&mp->mnt_refs, 1);
251 if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
252 if (flags & LK_NOWAIT) {
253 atomic_add_int(&mp->mnt_refs, -1);
256 /* XXX not MP safe */
257 mp->mnt_kern_flag |= MNTK_MWAIT;
259 * Since all busy locks are shared except the exclusive
260 * lock granted when unmounting, the only place that a
261 * wakeup needs to be done is at the release of the
262 * exclusive lock at the end of dounmount.
264 tsleep((caddr_t)mp, 0, "vfs_busy", 0);
265 atomic_add_int(&mp->mnt_refs, -1);
269 if (lockmgr(&mp->mnt_lock, lkflags))
270 panic("vfs_busy: unexpected lock failure");
275 * Free a busy filesystem.
277 * Decrement refs before releasing the lock so e.g. a pending umount
278 * doesn't give us an unexpected busy error.
281 vfs_unbusy(struct mount *mp)
283 atomic_add_int(&mp->mnt_refs, -1);
284 lockmgr(&mp->mnt_lock, LK_RELEASE);
288 * Lookup a filesystem type, and if found allocate and initialize
289 * a mount structure for it.
291 * Devname is usually updated by mount(8) after booting.
294 vfs_rootmountalloc(char *fstypename, char *devname, struct mount **mpp)
296 struct vfsconf *vfsp;
299 if (fstypename == NULL)
302 vfsp = vfsconf_find_by_name(fstypename);
305 mp = kmalloc(sizeof(struct mount), M_MOUNT, M_WAITOK | M_ZERO);
307 lockinit(&mp->mnt_lock, "vfslock", VLKTIMEOUT, 0);
311 mp->mnt_op = vfsp->vfc_vfsops;
312 vfsp->vfc_refcount++;
313 mp->mnt_stat.f_type = vfsp->vfc_typenum;
314 mp->mnt_flag |= MNT_RDONLY;
315 mp->mnt_flag |= vfsp->vfc_flags & MNT_VISFLAGMASK;
316 strncpy(mp->mnt_stat.f_fstypename, vfsp->vfc_name, MFSNAMELEN);
317 copystr(devname, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, 0);
323 * Basic mount structure initialization
326 mount_init(struct mount *mp)
328 lockinit(&mp->mnt_lock, "vfslock", 0, 0);
329 lwkt_token_init(&mp->mnt_token, "permnt");
331 TAILQ_INIT(&mp->mnt_nvnodelist);
332 TAILQ_INIT(&mp->mnt_reservedvnlist);
333 TAILQ_INIT(&mp->mnt_jlist);
334 mp->mnt_nvnodelistsize = 0;
336 mp->mnt_iosize_max = MAXPHYS;
340 * Lookup a mount point by filesystem identifier.
343 vfs_getvfs(fsid_t *fsid)
347 lwkt_gettoken(&mountlist_token);
348 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
349 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
350 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
354 lwkt_reltoken(&mountlist_token);
359 * Get a new unique fsid. Try to make its val[0] unique, since this value
360 * will be used to create fake device numbers for stat(). Also try (but
361 * not so hard) make its val[0] unique mod 2^16, since some emulators only
362 * support 16-bit device numbers. We end up with unique val[0]'s for the
363 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
365 * Keep in mind that several mounts may be running in parallel. Starting
366 * the search one past where the previous search terminated is both a
367 * micro-optimization and a defense against returning the same fsid to
371 vfs_getnewfsid(struct mount *mp)
373 static u_int16_t mntid_base;
377 lwkt_gettoken(&mntid_token);
378 mtype = mp->mnt_vfc->vfc_typenum;
379 tfsid.val[1] = mtype;
380 mtype = (mtype & 0xFF) << 24;
382 tfsid.val[0] = makeudev(255,
383 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
385 if (vfs_getvfs(&tfsid) == NULL)
388 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
389 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
390 lwkt_reltoken(&mntid_token);
394 * Set the FSID for a new mount point to the template. Adjust
395 * the FSID to avoid collisions.
398 vfs_setfsid(struct mount *mp, fsid_t *template)
402 bzero(&mp->mnt_stat.f_fsid, sizeof(mp->mnt_stat.f_fsid));
404 if (vfs_getvfs(template) == NULL)
409 mp->mnt_stat.f_fsid = *template;
414 * This routine is called when we have too many vnodes. It attempts
415 * to free <count> vnodes and will potentially free vnodes that still
416 * have VM backing store (VM backing store is typically the cause
417 * of a vnode blowout so we want to do this). Therefore, this operation
418 * is not considered cheap.
420 * A number of conditions may prevent a vnode from being reclaimed.
421 * the buffer cache may have references on the vnode, a directory
422 * vnode may still have references due to the namei cache representing
423 * underlying files, or the vnode may be in active use. It is not
424 * desireable to reuse such vnodes. These conditions may cause the
425 * number of vnodes to reach some minimum value regardless of what
426 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low.
430 * This is a quick non-blocking check to determine if the vnode is a good
431 * candidate for being (eventually) vgone()'d. Returns 0 if the vnode is
432 * not a good candidate, 1 if it is.
435 vmightfree(struct vnode *vp, int page_count, int pass)
437 if (vp->v_flag & VRECLAIMED)
440 if ((vp->v_flag & VFREE) && TAILQ_EMPTY(&vp->v_namecache))
443 if (sysref_isactive(&vp->v_sysref))
445 if (vp->v_object && vp->v_object->resident_page_count >= page_count)
449 * XXX horrible hack. Up to four passes will be taken. Each pass
450 * makes a larger set of vnodes eligible. For now what this really
451 * means is that we try to recycle files opened only once before
452 * recycling files opened multiple times.
454 switch(vp->v_flag & (VAGE0 | VAGE1)) {
474 * The vnode was found to be possibly vgone()able and the caller has locked it
475 * (thus the usecount should be 1 now). Determine if the vnode is actually
476 * vgone()able, doing some cleanups in the process. Returns 1 if the vnode
477 * can be vgone()'d, 0 otherwise.
479 * Note that v_auxrefs may be non-zero because (A) this vnode is not a leaf
480 * in the namecache topology and (B) this vnode has buffer cache bufs.
481 * We cannot remove vnodes with non-leaf namecache associations. We do a
482 * tentitive leaf check prior to attempting to flush out any buffers but the
483 * 'real' test when all is said in done is that v_auxrefs must become 0 for
484 * the vnode to be freeable.
486 * We could theoretically just unconditionally flush when v_auxrefs != 0,
487 * but flushing data associated with non-leaf nodes (which are always
488 * directories), just throws it away for no benefit. It is the buffer
489 * cache's responsibility to choose buffers to recycle from the cached
490 * data point of view.
493 visleaf(struct vnode *vp)
495 struct namecache *ncp;
497 spin_lock(&vp->v_spin);
498 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
499 if (!TAILQ_EMPTY(&ncp->nc_list)) {
500 spin_unlock(&vp->v_spin);
504 spin_unlock(&vp->v_spin);
509 * Try to clean up the vnode to the point where it can be vgone()'d, returning
510 * 0 if it cannot be vgone()'d (or already has been), 1 if it can. Unlike
511 * vmightfree() this routine may flush the vnode and block. Vnodes marked
512 * VFREE are still candidates for vgone()ing because they may hold namecache
513 * resources and could be blocking the namecache directory hierarchy (and
514 * related vnodes) from being freed.
517 vtrytomakegoneable(struct vnode *vp, int page_count)
519 if (vp->v_flag & VRECLAIMED)
521 if (vp->v_sysref.refcnt > 1)
523 if (vp->v_object && vp->v_object->resident_page_count >= page_count)
525 if (vp->v_auxrefs && visleaf(vp)) {
526 vinvalbuf(vp, V_SAVE, 0, 0);
528 kprintf((vp->v_auxrefs ? "vrecycle: vp %p failed: %s\n" :
529 "vrecycle: vp %p succeeded: %s\n"), vp,
530 (TAILQ_FIRST(&vp->v_namecache) ?
531 TAILQ_FIRST(&vp->v_namecache)->nc_name : "?"));
536 * This sequence may seem a little strange, but we need to optimize
537 * the critical path a bit. We can't recycle vnodes with other
538 * references and because we are trying to recycle an otherwise
539 * perfectly fine vnode we have to invalidate the namecache in a
540 * way that avoids possible deadlocks (since the vnode lock is being
541 * held here). Finally, we have to check for other references one
542 * last time in case something snuck in during the inval.
544 if (vp->v_sysref.refcnt > 1 || vp->v_auxrefs != 0)
546 if (cache_inval_vp_nonblock(vp))
548 return (vp->v_sysref.refcnt <= 1 && vp->v_auxrefs == 0);
552 * Reclaim up to 1/10 of the vnodes associated with a mount point. Try
553 * to avoid vnodes which have lots of resident pages (we are trying to free
554 * vnodes, not memory).
556 * This routine is a callback from the mountlist scan. The mount point
557 * in question will be busied.
559 * NOTE: The 1/10 reclamation also ensures that the inactive data set
560 * (the vnodes being recycled by the one-time use) does not degenerate
561 * into too-small a set. This is important because once a vnode is
562 * marked as not being one-time-use (VAGE0/VAGE1 both 0) that vnode
563 * will not be destroyed EXCEPT by this mechanism. VM pages can still
564 * be cleaned/freed by the pageout daemon.
567 vlrureclaim(struct mount *mp, void *data)
569 struct vnlru_info *info = data;
575 int trigger_mult = vnlru_nowhere;
578 * Calculate the trigger point for the resident pages check. The
579 * minimum trigger value is approximately the number of pages in
580 * the system divded by the number of vnodes. However, due to
581 * various other system memory overheads unrelated to data caching
582 * it is a good idea to double the trigger (at least).
584 * trigger_mult starts at 0. If the recycler is having problems
585 * finding enough freeable vnodes it will increase trigger_mult.
586 * This should not happen in normal operation, even on machines with
587 * low amounts of memory, but extraordinary memory use by the system
588 * verses the amount of cached data can trigger it.
590 usevnodes = desiredvnodes;
593 trigger = vmstats.v_page_count * (trigger_mult + 2) / usevnodes;
596 lwkt_gettoken(&mntvnode_token);
597 count = mp->mnt_nvnodelistsize / 10 + 1;
599 while (count && mp->mnt_syncer) {
601 * Next vnode. Use the special syncer vnode to placemark
602 * the LRU. This way the LRU code does not interfere with
605 vp = TAILQ_NEXT(mp->mnt_syncer, v_nmntvnodes);
606 TAILQ_REMOVE(&mp->mnt_nvnodelist, mp->mnt_syncer, v_nmntvnodes);
608 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp,
609 mp->mnt_syncer, v_nmntvnodes);
611 TAILQ_INSERT_HEAD(&mp->mnt_nvnodelist, mp->mnt_syncer,
613 vp = TAILQ_NEXT(mp->mnt_syncer, v_nmntvnodes);
621 * The VP will stick around while we hold mntvnode_token,
622 * at least until we block, so we can safely do an initial
623 * check, and then must check again after we lock the vnode.
625 if (vp->v_type == VNON || /* syncer or indeterminant */
626 !vmightfree(vp, trigger, info->pass) /* critical path opt */
633 * VX get the candidate vnode. If the VX get fails the
634 * vnode might still be on the mountlist. Our loop depends
635 * on us at least cycling the vnode to the end of the
638 if (vx_get_nonblock(vp) != 0) {
644 * Since we blocked locking the vp, make sure it is still
645 * a candidate for reclamation. That is, it has not already
646 * been reclaimed and only has our VX reference associated
649 if (vp->v_type == VNON || /* syncer or indeterminant */
650 (vp->v_flag & VRECLAIMED) ||
652 !vtrytomakegoneable(vp, trigger) /* critical path opt */
660 * All right, we are good, move the vp to the end of the
661 * mountlist and clean it out. The vget will have returned
662 * an error if the vnode was destroyed (VRECLAIMED set), so we
663 * do not have to check again. The vput() will move the
664 * vnode to the free list if the vgone() was successful.
666 KKASSERT(vp->v_mount == mp);
672 lwkt_reltoken(&mntvnode_token);
677 * Attempt to recycle vnodes in a context that is always safe to block.
678 * Calling vlrurecycle() from the bowels of file system code has some
679 * interesting deadlock problems.
681 static struct thread *vnlruthread;
682 static int vnlruproc_sig;
685 vnlru_proc_wait(void)
687 tsleep_interlock(&vnlruproc_sig, 0);
688 if (vnlruproc_sig == 0) {
689 vnlruproc_sig = 1; /* avoid unnecessary wakeups */
692 tsleep(&vnlruproc_sig, PINTERLOCKED, "vlruwk", hz);
698 struct thread *td = curthread;
699 struct vnlru_info info;
702 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, td,
706 kproc_suspend_loop();
709 * Try to free some vnodes if we have too many
711 if (numvnodes > desiredvnodes &&
712 freevnodes > desiredvnodes * 2 / 10) {
713 int count = numvnodes - desiredvnodes;
715 if (count > freevnodes / 100)
716 count = freevnodes / 100;
719 freesomevnodes(count);
723 * Nothing to do if most of our vnodes are already on
726 if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) {
728 wakeup(&vnlruproc_sig);
729 tsleep(vnlruthread, 0, "vlruwt", hz);
735 * The pass iterates through the four combinations of
736 * VAGE0/VAGE1. We want to get rid of aged small files
741 while (done == 0 && info.pass < 4) {
742 done = mountlist_scan(vlrureclaim, &info,
748 * The vlrureclaim() call only processes 1/10 of the vnodes
749 * on each mount. If we couldn't find any repeat the loop
750 * at least enough times to cover all available vnodes before
751 * we start sleeping. Complain if the failure extends past
752 * 30 second, every 30 seconds.
756 if (vnlru_nowhere % 10 == 0)
757 tsleep(vnlruthread, 0, "vlrup", hz * 3);
758 if (vnlru_nowhere % 100 == 0)
759 kprintf("vnlru_proc: vnode recycler stopped working!\n");
760 if (vnlru_nowhere == 1000)
769 * MOUNTLIST FUNCTIONS
773 * mountlist_insert (MP SAFE)
775 * Add a new mount point to the mount list.
778 mountlist_insert(struct mount *mp, int how)
780 lwkt_gettoken(&mountlist_token);
781 if (how == MNTINS_FIRST)
782 TAILQ_INSERT_HEAD(&mountlist, mp, mnt_list);
784 TAILQ_INSERT_TAIL(&mountlist, mp, mnt_list);
785 lwkt_reltoken(&mountlist_token);
789 * mountlist_interlock (MP SAFE)
791 * Execute the specified interlock function with the mountlist token
792 * held. The function will be called in a serialized fashion verses
793 * other functions called through this mechanism.
796 mountlist_interlock(int (*callback)(struct mount *), struct mount *mp)
800 lwkt_gettoken(&mountlist_token);
801 error = callback(mp);
802 lwkt_reltoken(&mountlist_token);
807 * mountlist_boot_getfirst (DURING BOOT ONLY)
809 * This function returns the first mount on the mountlist, which is
810 * expected to be the root mount. Since no interlocks are obtained
811 * this function is only safe to use during booting.
815 mountlist_boot_getfirst(void)
817 return(TAILQ_FIRST(&mountlist));
821 * mountlist_remove (MP SAFE)
823 * Remove a node from the mountlist. If this node is the next scan node
824 * for any active mountlist scans, the active mountlist scan will be
825 * adjusted to skip the node, thus allowing removals during mountlist
829 mountlist_remove(struct mount *mp)
831 struct mountscan_info *msi;
833 lwkt_gettoken(&mountlist_token);
834 TAILQ_FOREACH(msi, &mountscan_list, msi_entry) {
835 if (msi->msi_node == mp) {
836 if (msi->msi_how & MNTSCAN_FORWARD)
837 msi->msi_node = TAILQ_NEXT(mp, mnt_list);
839 msi->msi_node = TAILQ_PREV(mp, mntlist, mnt_list);
842 TAILQ_REMOVE(&mountlist, mp, mnt_list);
843 lwkt_reltoken(&mountlist_token);
847 * mountlist_exists (MP SAFE)
849 * Checks if a node exists in the mountlist.
850 * This function is mainly used by VFS quota code to check if a
851 * cached nullfs struct mount pointer is still valid at use time
853 * FIXME: there is no warranty the mp passed to that function
854 * will be the same one used by VFS_ACCOUNT() later
857 mountlist_exists(struct mount *mp)
862 lwkt_gettoken(&mountlist_token);
863 TAILQ_FOREACH(lmp, &mountlist, mnt_list) {
869 lwkt_reltoken(&mountlist_token);
874 * mountlist_scan (MP SAFE)
876 * Safely scan the mount points on the mount list. Unless otherwise
877 * specified each mount point will be busied prior to the callback and
878 * unbusied afterwords. The callback may safely remove any mount point
879 * without interfering with the scan. If the current callback
880 * mount is removed the scanner will not attempt to unbusy it.
882 * If a mount node cannot be busied it is silently skipped.
884 * The callback return value is aggregated and a total is returned. A return
885 * value of < 0 is not aggregated and will terminate the scan.
887 * MNTSCAN_FORWARD - the mountlist is scanned in the forward direction
888 * MNTSCAN_REVERSE - the mountlist is scanned in reverse
889 * MNTSCAN_NOBUSY - the scanner will make the callback without busying
893 mountlist_scan(int (*callback)(struct mount *, void *), void *data, int how)
895 struct mountscan_info info;
900 lwkt_gettoken(&mountlist_token);
903 info.msi_node = NULL; /* paranoia */
904 TAILQ_INSERT_TAIL(&mountscan_list, &info, msi_entry);
908 if (how & MNTSCAN_FORWARD) {
909 info.msi_node = TAILQ_FIRST(&mountlist);
910 while ((mp = info.msi_node) != NULL) {
911 if (how & MNTSCAN_NOBUSY) {
912 count = callback(mp, data);
913 } else if (vfs_busy(mp, LK_NOWAIT) == 0) {
914 count = callback(mp, data);
915 if (mp == info.msi_node)
923 if (mp == info.msi_node)
924 info.msi_node = TAILQ_NEXT(mp, mnt_list);
926 } else if (how & MNTSCAN_REVERSE) {
927 info.msi_node = TAILQ_LAST(&mountlist, mntlist);
928 while ((mp = info.msi_node) != NULL) {
929 if (how & MNTSCAN_NOBUSY) {
930 count = callback(mp, data);
931 } else if (vfs_busy(mp, LK_NOWAIT) == 0) {
932 count = callback(mp, data);
933 if (mp == info.msi_node)
941 if (mp == info.msi_node)
942 info.msi_node = TAILQ_PREV(mp, mntlist, mnt_list);
945 TAILQ_REMOVE(&mountscan_list, &info, msi_entry);
946 lwkt_reltoken(&mountlist_token);
951 * MOUNT RELATED VNODE FUNCTIONS
954 static struct kproc_desc vnlru_kp = {
959 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
962 * Move a vnode from one mount queue to another.
967 insmntque(struct vnode *vp, struct mount *mp)
969 lwkt_gettoken(&mntvnode_token);
971 * Delete from old mount point vnode list, if on one.
973 if (vp->v_mount != NULL) {
974 KASSERT(vp->v_mount->mnt_nvnodelistsize > 0,
975 ("bad mount point vnode list size"));
977 vp->v_mount->mnt_nvnodelistsize--;
980 * Insert into list of vnodes for the new mount point, if available.
981 * The 'end' of the LRU list is the vnode prior to mp->mnt_syncer.
983 if ((vp->v_mount = mp) == NULL) {
984 lwkt_reltoken(&mntvnode_token);
987 if (mp->mnt_syncer) {
988 TAILQ_INSERT_BEFORE(mp->mnt_syncer, vp, v_nmntvnodes);
990 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
992 mp->mnt_nvnodelistsize++;
993 lwkt_reltoken(&mntvnode_token);
998 * Scan the vnodes under a mount point and issue appropriate callbacks.
1000 * The fastfunc() callback is called with just the mountlist token held
1001 * (no vnode lock). It may not block and the vnode may be undergoing
1002 * modifications while the caller is processing it. The vnode will
1003 * not be entirely destroyed, however, due to the fact that the mountlist
1004 * token is held. A return value < 0 skips to the next vnode without calling
1005 * the slowfunc(), a return value > 0 terminates the loop.
1007 * The slowfunc() callback is called after the vnode has been successfully
1008 * locked based on passed flags. The vnode is skipped if it gets rearranged
1009 * or destroyed while blocking on the lock. A non-zero return value from
1010 * the slow function terminates the loop. The slow function is allowed to
1011 * arbitrarily block. The scanning code guarentees consistency of operation
1012 * even if the slow function deletes or moves the node, or blocks and some
1013 * other thread deletes or moves the node.
1015 * NOTE: We hold vmobj_token to prevent a VM object from being destroyed
1016 * out from under the fastfunc()'s vnode test. It will not prevent
1017 * v_object from getting NULL'd out but it will ensure that the
1018 * pointer (if we race) will remain stable.
1024 int (*fastfunc)(struct mount *mp, struct vnode *vp, void *data),
1025 int (*slowfunc)(struct mount *mp, struct vnode *vp, void *data),
1028 struct vmntvnodescan_info info;
1031 int maxcount = mp->mnt_nvnodelistsize * 2;
1035 lwkt_gettoken(&mntvnode_token);
1036 lwkt_gettoken(&vmobj_token);
1039 * If asked to do one pass stop after iterating available vnodes.
1040 * Under heavy loads new vnodes can be added while we are scanning,
1041 * so this isn't perfect. Create a slop factor of 2x.
1043 if (flags & VMSC_ONEPASS)
1044 stopcount = mp->mnt_nvnodelistsize;
1046 info.vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
1047 TAILQ_INSERT_TAIL(&mntvnodescan_list, &info, entry);
1048 while ((vp = info.vp) != NULL) {
1049 if (--maxcount == 0) {
1050 kprintf("Warning: excessive fssync iteration\n");
1051 maxcount = mp->mnt_nvnodelistsize * 2;
1055 * Skip if visible but not ready, or special (e.g.
1058 if (vp->v_type == VNON)
1060 KKASSERT(vp->v_mount == mp);
1063 * Quick test. A negative return continues the loop without
1064 * calling the slow test. 0 continues onto the slow test.
1065 * A positive number aborts the loop.
1068 if ((r = fastfunc(mp, vp, data)) < 0) {
1077 * Get a vxlock on the vnode, retry if it has moved or isn't
1078 * in the mountlist where we expect it.
1083 switch(flags & (VMSC_GETVP|VMSC_GETVX|VMSC_NOWAIT)) {
1085 error = vget(vp, LK_EXCLUSIVE);
1087 case VMSC_GETVP|VMSC_NOWAIT:
1088 error = vget(vp, LK_EXCLUSIVE|LK_NOWAIT);
1101 * Do not call the slow function if the vnode is
1102 * invalid or if it was ripped out from under us
1103 * while we (potentially) blocked.
1105 if (info.vp == vp && vp->v_type != VNON)
1106 r = slowfunc(mp, vp, data);
1111 switch(flags & (VMSC_GETVP|VMSC_GETVX|VMSC_NOWAIT)) {
1113 case VMSC_GETVP|VMSC_NOWAIT:
1128 * Yield after some processing. Depending on the number
1129 * of vnodes, we might wind up running for a long time.
1130 * Because threads are not preemptable, time critical
1131 * userland processes might starve. Give them a chance
1134 if (++count == 10000) {
1135 /* We really want to yield a bit, so we simply sleep a tick */
1136 tsleep(mp, 0, "vnodescn", 1);
1141 * If doing one pass this decrements to zero. If it starts
1142 * at zero it is effectively unlimited for the purposes of
1145 if (--stopcount == 0)
1149 * Iterate. If the vnode was ripped out from under us
1150 * info.vp will already point to the next vnode, otherwise
1151 * we have to obtain the next valid vnode ourselves.
1154 info.vp = TAILQ_NEXT(vp, v_nmntvnodes);
1156 TAILQ_REMOVE(&mntvnodescan_list, &info, entry);
1157 lwkt_reltoken(&vmobj_token);
1158 lwkt_reltoken(&mntvnode_token);
1163 * Remove any vnodes in the vnode table belonging to mount point mp.
1165 * If FORCECLOSE is not specified, there should not be any active ones,
1166 * return error if any are found (nb: this is a user error, not a
1167 * system error). If FORCECLOSE is specified, detach any active vnodes
1170 * If WRITECLOSE is set, only flush out regular file vnodes open for
1173 * SKIPSYSTEM causes any vnodes marked VSYSTEM to be skipped.
1175 * `rootrefs' specifies the base reference count for the root vnode
1176 * of this filesystem. The root vnode is considered busy if its
1177 * v_sysref.refcnt exceeds this value. On a successful return, vflush()
1178 * will call vrele() on the root vnode exactly rootrefs times.
1179 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
1183 static int busyprt = 0; /* print out busy vnodes */
1184 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
1187 static int vflush_scan(struct mount *mp, struct vnode *vp, void *data);
1189 struct vflush_info {
1196 vflush(struct mount *mp, int rootrefs, int flags)
1198 struct thread *td = curthread; /* XXX */
1199 struct vnode *rootvp = NULL;
1201 struct vflush_info vflush_info;
1204 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
1205 ("vflush: bad args"));
1207 * Get the filesystem root vnode. We can vput() it
1208 * immediately, since with rootrefs > 0, it won't go away.
1210 if ((error = VFS_ROOT(mp, &rootvp)) != 0) {
1211 if ((flags & FORCECLOSE) == 0)
1214 /* continue anyway */
1220 vflush_info.busy = 0;
1221 vflush_info.flags = flags;
1222 vflush_info.td = td;
1223 vmntvnodescan(mp, VMSC_GETVX, NULL, vflush_scan, &vflush_info);
1225 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
1227 * If just the root vnode is busy, and if its refcount
1228 * is equal to `rootrefs', then go ahead and kill it.
1230 KASSERT(vflush_info.busy > 0, ("vflush: not busy"));
1231 KASSERT(rootvp->v_sysref.refcnt >= rootrefs, ("vflush: rootrefs"));
1232 if (vflush_info.busy == 1 && rootvp->v_sysref.refcnt == rootrefs) {
1234 vgone_vxlocked(rootvp);
1236 vflush_info.busy = 0;
1239 if (vflush_info.busy)
1241 for (; rootrefs > 0; rootrefs--)
1247 * The scan callback is made with an VX locked vnode.
1250 vflush_scan(struct mount *mp, struct vnode *vp, void *data)
1252 struct vflush_info *info = data;
1256 * Skip over a vnodes marked VSYSTEM.
1258 if ((info->flags & SKIPSYSTEM) && (vp->v_flag & VSYSTEM)) {
1263 * If WRITECLOSE is set, flush out unlinked but still open
1264 * files (even if open only for reading) and regular file
1265 * vnodes open for writing.
1267 if ((info->flags & WRITECLOSE) &&
1268 (vp->v_type == VNON ||
1269 (VOP_GETATTR(vp, &vattr) == 0 &&
1270 vattr.va_nlink > 0)) &&
1271 (vp->v_writecount == 0 || vp->v_type != VREG)) {
1276 * If we are the only holder (refcnt of 1) or the vnode is in
1277 * termination (refcnt < 0), we can vgone the vnode.
1279 if (vp->v_sysref.refcnt <= 1) {
1285 * If FORCECLOSE is set, forcibly destroy the vnode and then move
1286 * it to a dummymount structure so vop_*() functions don't deref
1289 if (info->flags & FORCECLOSE) {
1292 if (vp->v_mount == NULL)
1293 insmntque(vp, &dummymount);
1299 vprint("vflush: busy vnode", vp);
1306 add_bio_ops(struct bio_ops *ops)
1308 TAILQ_INSERT_TAIL(&bio_ops_list, ops, entry);
1312 rem_bio_ops(struct bio_ops *ops)
1314 TAILQ_REMOVE(&bio_ops_list, ops, entry);
1318 * This calls the bio_ops io_sync function either for a mount point
1321 * WARNING: softdeps is weirdly coded and just isn't happy unless
1322 * io_sync is called with a NULL mount from the general syncing code.
1325 bio_ops_sync(struct mount *mp)
1327 struct bio_ops *ops;
1330 if ((ops = mp->mnt_bioops) != NULL)
1333 TAILQ_FOREACH(ops, &bio_ops_list, entry) {
1340 * Lookup a mount point by nch
1343 mount_get_by_nc(struct namecache *ncp)
1345 struct mount *mp = NULL;
1347 lwkt_gettoken(&mountlist_token);
1348 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
1349 if (ncp == mp->mnt_ncmountpt.ncp)
1352 lwkt_reltoken(&mountlist_token);