Rename printf -> kprintf in sys/ and add some defines where necessary
[dragonfly.git] / sys / kern / vfs_mount.c
CommitLineData
5fd012e0
MD
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. 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.
57 *
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
68 * SUCH DAMAGE.
69 *
6ea70f76 70 * $DragonFly: src/sys/kern/vfs_mount.c,v 1.24 2006/12/23 00:35:04 swildner Exp $
5fd012e0
MD
71 */
72
73/*
74 * External virtual filesystem routines
75 */
76#include "opt_ddb.h"
77
78#include <sys/param.h>
79#include <sys/systm.h>
80#include <sys/kernel.h>
81#include <sys/malloc.h>
82#include <sys/mount.h>
83#include <sys/proc.h>
84#include <sys/vnode.h>
85#include <sys/buf.h>
86#include <sys/eventhandler.h>
87#include <sys/kthread.h>
88#include <sys/sysctl.h>
89
90#include <machine/limits.h>
91
92#include <sys/buf2.h>
93#include <sys/thread2.h>
94
95#include <vm/vm.h>
96#include <vm/vm_object.h>
97
861905fb
MD
98struct mountscan_info {
99 TAILQ_ENTRY(mountscan_info) msi_entry;
100 int msi_how;
101 struct mount *msi_node;
102};
103
be6c08cb
MD
104struct vmntvnodescan_info {
105 TAILQ_ENTRY(vmntvnodescan_info) entry;
106 struct vnode *vp;
107};
108
5fd012e0 109static int vnlru_nowhere = 0;
58552887 110SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RD,
5fd012e0
MD
111 &vnlru_nowhere, 0,
112 "Number of times the vnlru process ran without success");
113
114
115static struct lwkt_token mntid_token;
116
861905fb
MD
117static struct mntlist mountlist = TAILQ_HEAD_INITIALIZER(mountlist);
118static TAILQ_HEAD(,mountscan_info) mountscan_list;
119static struct lwkt_token mountlist_token;
be6c08cb 120static TAILQ_HEAD(,vmntvnodescan_info) mntvnodescan_list;
5fd012e0
MD
121struct lwkt_token mntvnode_token;
122
5fd012e0
MD
123/*
124 * Called from vfsinit()
125 */
126void
127vfs_mount_init(void)
128{
129 lwkt_token_init(&mountlist_token);
130 lwkt_token_init(&mntvnode_token);
131 lwkt_token_init(&mntid_token);
861905fb 132 TAILQ_INIT(&mountscan_list);
be6c08cb
MD
133 TAILQ_INIT(&mntvnodescan_list);
134}
135
136/*
137 * Support function called with mntvnode_token held to remove a vnode
138 * from the mountlist. We must update any list scans which are in progress.
139 */
140static void
141vremovevnodemnt(struct vnode *vp)
142{
143 struct vmntvnodescan_info *info;
144
145 TAILQ_FOREACH(info, &mntvnodescan_list, entry) {
146 if (info->vp == vp)
147 info->vp = TAILQ_NEXT(vp, v_nmntvnodes);
148 }
149 TAILQ_REMOVE(&vp->v_mount->mnt_nvnodelist, vp, v_nmntvnodes);
150}
151
152/*
153 * Support function called with mntvnode_token held to move a vnode to
154 * the end of the list.
155 */
156static void
157vmovevnodetoend(struct mount *mp, struct vnode *vp)
158{
159 vremovevnodemnt(vp);
160 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
5fd012e0
MD
161}
162
be6c08cb 163
5fd012e0
MD
164/*
165 * Allocate a new vnode and associate it with a tag, mount point, and
166 * operations vector.
167 *
168 * A VX locked and refd vnode is returned. The caller should setup the
169 * remaining fields and vx_put() or, if he wishes to leave a vref,
170 * vx_unlock() the vnode.
171 */
172int
6ddb7618
MD
173getnewvnode(enum vtagtype tag, struct mount *mp,
174 struct vnode **vpp, int lktimeout, int lkflags)
175{
176 struct vnode *vp;
177
178 KKASSERT(mp != NULL);
179
180 vp = allocvnode(lktimeout, lkflags);
181 vp->v_tag = tag;
182 vp->v_data = NULL;
183
184 /*
185 * By default the vnode is assigned the mount point's normal
186 * operations vector.
187 */
188 vp->v_ops = &mp->mnt_vn_use_ops;
189
190 /*
191 * Placing the vnode on the mount point's queue makes it visible.
192 * VNON prevents it from being messed with, however.
193 */
194 insmntque(vp, mp);
6ddb7618
MD
195
196 /*
197 * A VX locked & refd vnode is returned.
198 */
199 *vpp = vp;
200 return (0);
201}
202
203/*
204 * This function creates vnodes with special operations vectors. The
205 * mount point is optional.
206 *
207 * This routine is being phased out.
208 */
209int
210getspecialvnode(enum vtagtype tag, struct mount *mp,
66a1ddf5 211 struct vop_ops **ops,
5fd012e0
MD
212 struct vnode **vpp, int lktimeout, int lkflags)
213{
214 struct vnode *vp;
215
216 vp = allocvnode(lktimeout, lkflags);
217 vp->v_tag = tag;
5fd012e0 218 vp->v_data = NULL;
66a1ddf5 219 vp->v_ops = ops;
5fd012e0
MD
220
221 /*
222 * Placing the vnode on the mount point's queue makes it visible.
223 * VNON prevents it from being messed with, however.
224 */
225 insmntque(vp, mp);
5fd012e0
MD
226
227 /*
228 * A VX locked & refd vnode is returned.
229 */
230 *vpp = vp;
231 return (0);
232}
233
234/*
861905fb
MD
235 * Interlock against an unmount, return 0 on success, non-zero on failure.
236 *
237 * The passed flag may be 0 or LK_NOWAIT and is only used if an unmount
238 * is in-progress.
239 *
240 * If no unmount is in-progress LK_NOWAIT is ignored. No other flag bits
241 * are used. A shared locked will be obtained and the filesystem will not
242 * be unmountable until the lock is released.
5fd012e0
MD
243 */
244int
f9642f56 245vfs_busy(struct mount *mp, int flags)
5fd012e0
MD
246{
247 int lkflags;
248
249 if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
250 if (flags & LK_NOWAIT)
251 return (ENOENT);
861905fb 252 /* XXX not MP safe */
5fd012e0
MD
253 mp->mnt_kern_flag |= MNTK_MWAIT;
254 /*
255 * Since all busy locks are shared except the exclusive
256 * lock granted when unmounting, the only place that a
257 * wakeup needs to be done is at the release of the
258 * exclusive lock at the end of dounmount.
5fd012e0
MD
259 */
260 tsleep((caddr_t)mp, 0, "vfs_busy", 0);
261 return (ENOENT);
262 }
ab6f251b 263 lkflags = LK_SHARED;
df4f70a6 264 if (lockmgr(&mp->mnt_lock, lkflags))
5fd012e0
MD
265 panic("vfs_busy: unexpected lock failure");
266 return (0);
267}
268
269/*
270 * Free a busy filesystem.
271 */
272void
f9642f56 273vfs_unbusy(struct mount *mp)
5fd012e0 274{
df4f70a6 275 lockmgr(&mp->mnt_lock, LK_RELEASE);
5fd012e0
MD
276}
277
278/*
279 * Lookup a filesystem type, and if found allocate and initialize
280 * a mount structure for it.
281 *
282 * Devname is usually updated by mount(8) after booting.
283 */
284int
285vfs_rootmountalloc(char *fstypename, char *devname, struct mount **mpp)
286{
5fd012e0
MD
287 struct vfsconf *vfsp;
288 struct mount *mp;
289
290 if (fstypename == NULL)
291 return (ENODEV);
292 for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) {
293 if (!strcmp(vfsp->vfc_name, fstypename))
294 break;
295 }
296 if (vfsp == NULL)
297 return (ENODEV);
efda3bd0 298 mp = kmalloc(sizeof(struct mount), M_MOUNT, M_WAITOK);
5fd012e0 299 bzero((char *)mp, (u_long)sizeof(struct mount));
ab6f251b 300 lockinit(&mp->mnt_lock, "vfslock", VLKTIMEOUT, 0);
f9642f56 301 vfs_busy(mp, LK_NOWAIT);
5fd012e0
MD
302 TAILQ_INIT(&mp->mnt_nvnodelist);
303 TAILQ_INIT(&mp->mnt_reservedvnlist);
2281065e 304 TAILQ_INIT(&mp->mnt_jlist);
5fd012e0
MD
305 mp->mnt_nvnodelistsize = 0;
306 mp->mnt_vfc = vfsp;
307 mp->mnt_op = vfsp->vfc_vfsops;
308 mp->mnt_flag = MNT_RDONLY;
5fd012e0
MD
309 vfsp->vfc_refcount++;
310 mp->mnt_iosize_max = DFLTPHYS;
311 mp->mnt_stat.f_type = vfsp->vfc_typenum;
312 mp->mnt_flag |= vfsp->vfc_flags & MNT_VISFLAGMASK;
313 strncpy(mp->mnt_stat.f_fstypename, vfsp->vfc_name, MFSNAMELEN);
069b825e 314 copystr(devname, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, 0);
5fd012e0
MD
315 *mpp = mp;
316 return (0);
317}
318
319/*
320 * Lookup a mount point by filesystem identifier.
321 */
322struct mount *
323vfs_getvfs(fsid_t *fsid)
324{
325 struct mount *mp;
326 lwkt_tokref ilock;
327
328 lwkt_gettoken(&ilock, &mountlist_token);
329 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
330 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
331 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
332 break;
333 }
334 }
335 lwkt_reltoken(&ilock);
336 return (mp);
337}
338
339/*
340 * Get a new unique fsid. Try to make its val[0] unique, since this value
341 * will be used to create fake device numbers for stat(). Also try (but
342 * not so hard) make its val[0] unique mod 2^16, since some emulators only
343 * support 16-bit device numbers. We end up with unique val[0]'s for the
344 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
345 *
346 * Keep in mind that several mounts may be running in parallel. Starting
347 * the search one past where the previous search terminated is both a
348 * micro-optimization and a defense against returning the same fsid to
349 * different mounts.
350 */
351void
352vfs_getnewfsid(struct mount *mp)
353{
354 static u_int16_t mntid_base;
355 lwkt_tokref ilock;
356 fsid_t tfsid;
357 int mtype;
358
359 lwkt_gettoken(&ilock, &mntid_token);
360 mtype = mp->mnt_vfc->vfc_typenum;
361 tfsid.val[1] = mtype;
362 mtype = (mtype & 0xFF) << 24;
363 for (;;) {
364 tfsid.val[0] = makeudev(255,
365 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
366 mntid_base++;
367 if (vfs_getvfs(&tfsid) == NULL)
368 break;
369 }
370 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
371 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
372 lwkt_reltoken(&ilock);
373}
374
375/*
376 * This routine is called when we have too many vnodes. It attempts
377 * to free <count> vnodes and will potentially free vnodes that still
378 * have VM backing store (VM backing store is typically the cause
379 * of a vnode blowout so we want to do this). Therefore, this operation
380 * is not considered cheap.
381 *
382 * A number of conditions may prevent a vnode from being reclaimed.
383 * the buffer cache may have references on the vnode, a directory
384 * vnode may still have references due to the namei cache representing
385 * underlying files, or the vnode may be in active use. It is not
386 * desireable to reuse such vnodes. These conditions may cause the
387 * number of vnodes to reach some minimum value regardless of what
388 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low.
389 */
390
391/*
58552887
MD
392 * This is a quick non-blocking check to determine if the vnode is a good
393 * candidate for being (eventually) vgone()'d. Returns 0 if the vnode is
394 * not a good candidate, 1 if it is.
395 *
44b1cf3d
MD
396 * Note that a vnode can be marked VFREE without really being free, so
397 * we don't use the flag for any tests.
5fd012e0
MD
398 */
399static __inline int
250d127c 400vmightfree(struct vnode *vp, int page_count)
5fd012e0 401{
58552887
MD
402 if (vp->v_flag & VRECLAIMED)
403 return (0);
44b1cf3d 404#if 0
58552887 405 if ((vp->v_flag & VFREE) && TAILQ_EMPTY(&vp->v_namecache))
5fd012e0 406 return (0);
44b1cf3d 407#endif
250d127c 408 if (vp->v_usecount != 0)
5fd012e0
MD
409 return (0);
410 if (vp->v_object && vp->v_object->resident_page_count >= page_count)
411 return (0);
412 return (1);
413}
414
250d127c 415/*
58552887 416 * The vnode was found to be possibly vgone()able and the caller has locked it
250d127c 417 * (thus the usecount should be 1 now). Determine if the vnode is actually
58552887
MD
418 * vgone()able, doing some cleanups in the process. Returns 1 if the vnode
419 * can be vgone()'d, 0 otherwise.
250d127c
MD
420 *
421 * Note that v_holdcnt may be non-zero because (A) this vnode is not a leaf
422 * in the namecache topology and (B) this vnode has buffer cache bufs.
423 * We cannot remove vnodes with non-leaf namecache associations. We do a
424 * tentitive leaf check prior to attempting to flush out any buffers but the
425 * 'real' test when all is said in done is that v_holdcnt must become 0 for
426 * the vnode to be freeable.
427 *
428 * We could theoretically just unconditionally flush when v_holdcnt != 0,
429 * but flushing data associated with non-leaf nodes (which are always
430 * directories), just throws it away for no benefit. It is the buffer
431 * cache's responsibility to choose buffers to recycle from the cached
432 * data point of view.
433 */
434static int
435visleaf(struct vnode *vp)
436{
437 struct namecache *ncp;
438
439 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
440 if (!TAILQ_EMPTY(&ncp->nc_list))
441 return(0);
442 }
443 return(1);
444}
445
58552887
MD
446/*
447 * Try to clean up the vnode to the point where it can be vgone()'d, returning
448 * 0 if it cannot be vgone()'d (or already has been), 1 if it can. Unlike
449 * vmightfree() this routine may flush the vnode and block. Vnodes marked
450 * VFREE are still candidates for vgone()ing because they may hold namecache
451 * resources and could be blocking the namecache directory hierarchy (and
452 * related vnodes) from being freed.
453 */
250d127c 454static int
58552887 455vtrytomakegoneable(struct vnode *vp, int page_count)
250d127c 456{
58552887 457 if (vp->v_flag & VRECLAIMED)
250d127c
MD
458 return (0);
459 if (vp->v_usecount != 1)
460 return (0);
461 if (vp->v_object && vp->v_object->resident_page_count >= page_count)
462 return (0);
463 if (vp->v_holdcnt && visleaf(vp)) {
87de5057 464 vinvalbuf(vp, V_SAVE, 0, 0);
250d127c 465#if 0 /* DEBUG */
6ea70f76 466 kprintf((vp->v_holdcnt ? "vrecycle: vp %p failed: %s\n" :
250d127c
MD
467 "vrecycle: vp %p succeeded: %s\n"), vp,
468 (TAILQ_FIRST(&vp->v_namecache) ?
469 TAILQ_FIRST(&vp->v_namecache)->nc_name : "?"));
470#endif
471 }
472 return(vp->v_usecount == 1 && vp->v_holdcnt == 0);
473}
5fd012e0 474
58552887
MD
475/*
476 * Reclaim up to 1/10 of the vnodes associated with a mount point. Try
477 * to avoid vnodes which have lots of resident pages (we are trying to free
861905fb
MD
478 * vnodes, not memory).
479 *
480 * This routine is a callback from the mountlist scan. The mount point
481 * in question will be busied.
58552887 482 */
5fd012e0 483static int
861905fb 484vlrureclaim(struct mount *mp, void *data)
5fd012e0
MD
485{
486 struct vnode *vp;
487 lwkt_tokref ilock;
488 int done;
489 int trigger;
490 int usevnodes;
491 int count;
861905fb 492 int trigger_mult = vnlru_nowhere;
5fd012e0
MD
493
494 /*
58552887
MD
495 * Calculate the trigger point for the resident pages check. The
496 * minimum trigger value is approximately the number of pages in
497 * the system divded by the number of vnodes. However, due to
498 * various other system memory overheads unrelated to data caching
499 * it is a good idea to double the trigger (at least).
500 *
501 * trigger_mult starts at 0. If the recycler is having problems
502 * finding enough freeable vnodes it will increase trigger_mult.
503 * This should not happen in normal operation, even on machines with
504 * low amounts of memory, but extraordinary memory use by the system
505 * verses the amount of cached data can trigger it.
5fd012e0
MD
506 */
507 usevnodes = desiredvnodes;
508 if (usevnodes <= 0)
509 usevnodes = 1;
58552887 510 trigger = vmstats.v_page_count * (trigger_mult + 2) / usevnodes;
5fd012e0
MD
511
512 done = 0;
513 lwkt_gettoken(&ilock, &mntvnode_token);
514 count = mp->mnt_nvnodelistsize / 10 + 1;
515 while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) {
516 /*
517 * __VNODESCAN__
518 *
519 * The VP will stick around while we hold mntvnode_token,
520 * at least until we block, so we can safely do an initial
521 * check, and then must check again after we lock the vnode.
522 */
523 if (vp->v_type == VNON || /* XXX */
524 vp->v_type == VBAD || /* XXX */
250d127c 525 !vmightfree(vp, trigger) /* critical path opt */
5fd012e0 526 ) {
be6c08cb 527 vmovevnodetoend(mp, vp);
5fd012e0
MD
528 --count;
529 continue;
530 }
531
532 /*
533 * VX get the candidate vnode. If the VX get fails the
534 * vnode might still be on the mountlist. Our loop depends
535 * on us at least cycling the vnode to the end of the
536 * mountlist.
537 */
538 if (vx_get_nonblock(vp) != 0) {
be6c08cb
MD
539 if (vp->v_mount == mp)
540 vmovevnodetoend(mp, vp);
5fd012e0
MD
541 --count;
542 continue;
543 }
544
545 /*
546 * Since we blocked locking the vp, make sure it is still
547 * a candidate for reclamation. That is, it has not already
548 * been reclaimed and only has our VX reference associated
549 * with it.
550 */
551 if (vp->v_type == VNON || /* XXX */
552 vp->v_type == VBAD || /* XXX */
553 (vp->v_flag & VRECLAIMED) ||
554 vp->v_mount != mp ||
58552887 555 !vtrytomakegoneable(vp, trigger) /* critical path opt */
5fd012e0 556 ) {
be6c08cb
MD
557 if (vp->v_mount == mp)
558 vmovevnodetoend(mp, vp);
5fd012e0
MD
559 --count;
560 vx_put(vp);
561 continue;
562 }
563
564 /*
565 * All right, we are good, move the vp to the end of the
566 * mountlist and clean it out. The vget will have returned
567 * an error if the vnode was destroyed (VRECLAIMED set), so we
568 * do not have to check again. The vput() will move the
569 * vnode to the free list if the vgone() was successful.
570 */
571 KKASSERT(vp->v_mount == mp);
be6c08cb 572 vmovevnodetoend(mp, vp);
5fd012e0
MD
573 vgone(vp);
574 vx_put(vp);
575 ++done;
576 --count;
577 }
578 lwkt_reltoken(&ilock);
579 return (done);
580}
581
582/*
583 * Attempt to recycle vnodes in a context that is always safe to block.
584 * Calling vlrurecycle() from the bowels of file system code has some
585 * interesting deadlock problems.
586 */
587static struct thread *vnlruthread;
588static int vnlruproc_sig;
589
590void
591vnlru_proc_wait(void)
592{
593 if (vnlruproc_sig == 0) {
594 vnlruproc_sig = 1; /* avoid unnecessary wakeups */
595 wakeup(vnlruthread);
596 }
597 tsleep(&vnlruproc_sig, 0, "vlruwk", hz);
598}
599
600static void
601vnlru_proc(void)
602{
5fd012e0 603 struct thread *td = curthread;
861905fb 604 int done;
5fd012e0
MD
605
606 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, td,
607 SHUTDOWN_PRI_FIRST);
608
e43a034f 609 crit_enter();
5fd012e0
MD
610 for (;;) {
611 kproc_suspend_loop();
612 if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) {
613 vnlruproc_sig = 0;
614 wakeup(&vnlruproc_sig);
615 tsleep(td, 0, "vlruwt", hz);
616 continue;
617 }
62d0f1f0 618 cache_cleanneg(0);
861905fb 619 done = mountlist_scan(vlrureclaim, NULL, MNTSCAN_FORWARD);
58552887
MD
620
621 /*
622 * The vlrureclaim() call only processes 1/10 of the vnodes
623 * on each mount. If we couldn't find any repeat the loop
624 * at least enough times to cover all available vnodes before
625 * we start sleeping. Complain if the failure extends past
626 * 30 second, every 30 seconds.
627 */
5fd012e0
MD
628 if (done == 0) {
629 ++vnlru_nowhere;
5fd012e0 630 if (vnlru_nowhere % 10 == 0)
58552887
MD
631 tsleep(td, 0, "vlrup", hz * 3);
632 if (vnlru_nowhere % 100 == 0)
6ea70f76 633 kprintf("vnlru_proc: vnode recycler stopped working!\n");
58552887
MD
634 if (vnlru_nowhere == 1000)
635 vnlru_nowhere = 900;
5fd012e0
MD
636 } else {
637 vnlru_nowhere = 0;
638 }
639 }
e43a034f 640 crit_exit();
5fd012e0
MD
641}
642
861905fb
MD
643/*
644 * MOUNTLIST FUNCTIONS
645 */
646
647/*
648 * mountlist_insert (MP SAFE)
649 *
650 * Add a new mount point to the mount list.
651 */
652void
653mountlist_insert(struct mount *mp, int how)
654{
655 lwkt_tokref ilock;
656
657 lwkt_gettoken(&ilock, &mountlist_token);
658 if (how == MNTINS_FIRST)
659 TAILQ_INSERT_HEAD(&mountlist, mp, mnt_list);
660 else
661 TAILQ_INSERT_TAIL(&mountlist, mp, mnt_list);
662 lwkt_reltoken(&ilock);
663}
664
665/*
666 * mountlist_interlock (MP SAFE)
667 *
668 * Execute the specified interlock function with the mountlist token
669 * held. The function will be called in a serialized fashion verses
670 * other functions called through this mechanism.
671 */
672int
673mountlist_interlock(int (*callback)(struct mount *), struct mount *mp)
674{
675 lwkt_tokref ilock;
676 int error;
677
678 lwkt_gettoken(&ilock, &mountlist_token);
679 error = callback(mp);
680 lwkt_reltoken(&ilock);
681 return (error);
682}
683
684/*
685 * mountlist_boot_getfirst (DURING BOOT ONLY)
686 *
687 * This function returns the first mount on the mountlist, which is
688 * expected to be the root mount. Since no interlocks are obtained
689 * this function is only safe to use during booting.
690 */
691
692struct mount *
693mountlist_boot_getfirst(void)
694{
695 return(TAILQ_FIRST(&mountlist));
696}
697
698/*
699 * mountlist_remove (MP SAFE)
700 *
701 * Remove a node from the mountlist. If this node is the next scan node
702 * for any active mountlist scans, the active mountlist scan will be
703 * adjusted to skip the node, thus allowing removals during mountlist
704 * scans.
705 */
706void
707mountlist_remove(struct mount *mp)
708{
709 struct mountscan_info *msi;
710 lwkt_tokref ilock;
711
712 lwkt_gettoken(&ilock, &mountlist_token);
713 TAILQ_FOREACH(msi, &mountscan_list, msi_entry) {
714 if (msi->msi_node == mp) {
715 if (msi->msi_how & MNTSCAN_FORWARD)
716 msi->msi_node = TAILQ_NEXT(mp, mnt_list);
717 else
718 msi->msi_node = TAILQ_PREV(mp, mntlist, mnt_list);
719 }
720 }
721 TAILQ_REMOVE(&mountlist, mp, mnt_list);
722 lwkt_reltoken(&ilock);
723}
724
725/*
726 * mountlist_scan (MP SAFE)
727 *
728 * Safely scan the mount points on the mount list. Unless otherwise
729 * specified each mount point will be busied prior to the callback and
730 * unbusied afterwords. The callback may safely remove any mount point
731 * without interfering with the scan. If the current callback
732 * mount is removed the scanner will not attempt to unbusy it.
733 *
734 * If a mount node cannot be busied it is silently skipped.
735 *
736 * The callback return value is aggregated and a total is returned. A return
737 * value of < 0 is not aggregated and will terminate the scan.
738 *
739 * MNTSCAN_FORWARD - the mountlist is scanned in the forward direction
740 * MNTSCAN_REVERSE - the mountlist is scanned in reverse
741 * MNTSCAN_NOBUSY - the scanner will make the callback without busying
742 * the mount node.
743 */
744int
745mountlist_scan(int (*callback)(struct mount *, void *), void *data, int how)
746{
747 struct mountscan_info info;
748 lwkt_tokref ilock;
749 struct mount *mp;
750 thread_t td;
751 int count;
752 int res;
753
754 lwkt_gettoken(&ilock, &mountlist_token);
755
756 info.msi_how = how;
757 info.msi_node = NULL; /* paranoia */
758 TAILQ_INSERT_TAIL(&mountscan_list, &info, msi_entry);
759
760 res = 0;
761 td = curthread;
762
763 if (how & MNTSCAN_FORWARD) {
764 info.msi_node = TAILQ_FIRST(&mountlist);
765 while ((mp = info.msi_node) != NULL) {
766 if (how & MNTSCAN_NOBUSY) {
767 count = callback(mp, data);
f9642f56 768 } else if (vfs_busy(mp, LK_NOWAIT) == 0) {
861905fb
MD
769 count = callback(mp, data);
770 if (mp == info.msi_node)
f9642f56 771 vfs_unbusy(mp);
861905fb
MD
772 } else {
773 count = 0;
774 }
775 if (count < 0)
776 break;
777 res += count;
778 if (mp == info.msi_node)
779 info.msi_node = TAILQ_NEXT(mp, mnt_list);
780 }
781 } else if (how & MNTSCAN_REVERSE) {
782 info.msi_node = TAILQ_LAST(&mountlist, mntlist);
783 while ((mp = info.msi_node) != NULL) {
784 if (how & MNTSCAN_NOBUSY) {
785 count = callback(mp, data);
f9642f56 786 } else if (vfs_busy(mp, LK_NOWAIT) == 0) {
861905fb
MD
787 count = callback(mp, data);
788 if (mp == info.msi_node)
f9642f56 789 vfs_unbusy(mp);
861905fb
MD
790 } else {
791 count = 0;
792 }
793 if (count < 0)
794 break;
795 res += count;
796 if (mp == info.msi_node)
797 info.msi_node = TAILQ_PREV(mp, mntlist, mnt_list);
798 }
799 }
800 TAILQ_REMOVE(&mountscan_list, &info, msi_entry);
801 lwkt_reltoken(&ilock);
802 return(res);
803}
804
805/*
806 * MOUNT RELATED VNODE FUNCTIONS
807 */
808
5fd012e0
MD
809static struct kproc_desc vnlru_kp = {
810 "vnlru",
811 vnlru_proc,
812 &vnlruthread
813};
814SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
815
816/*
817 * Move a vnode from one mount queue to another.
818 */
819void
820insmntque(struct vnode *vp, struct mount *mp)
821{
822 lwkt_tokref ilock;
823
824 lwkt_gettoken(&ilock, &mntvnode_token);
825 /*
826 * Delete from old mount point vnode list, if on one.
827 */
828 if (vp->v_mount != NULL) {
829 KASSERT(vp->v_mount->mnt_nvnodelistsize > 0,
830 ("bad mount point vnode list size"));
be6c08cb 831 vremovevnodemnt(vp);
5fd012e0
MD
832 vp->v_mount->mnt_nvnodelistsize--;
833 }
834 /*
835 * Insert into list of vnodes for the new mount point, if available.
836 */
837 if ((vp->v_mount = mp) == NULL) {
838 lwkt_reltoken(&ilock);
839 return;
840 }
841 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
842 mp->mnt_nvnodelistsize++;
843 lwkt_reltoken(&ilock);
844}
845
846
847/*
be6c08cb
MD
848 * Scan the vnodes under a mount point and issue appropriate callbacks.
849 *
850 * The fastfunc() callback is called with just the mountlist token held
851 * (no vnode lock). It may not block and the vnode may be undergoing
852 * modifications while the caller is processing it. The vnode will
853 * not be entirely destroyed, however, due to the fact that the mountlist
854 * token is held. A return value < 0 skips to the next vnode without calling
855 * the slowfunc(), a return value > 0 terminates the loop.
856 *
857 * The slowfunc() callback is called after the vnode has been successfully
858 * locked based on passed flags. The vnode is skipped if it gets rearranged
859 * or destroyed while blocking on the lock. A non-zero return value from
860 * the slow function terminates the loop. The slow function is allowed to
861 * arbitrarily block. The scanning code guarentees consistency of operation
862 * even if the slow function deletes or moves the node, or blocks and some
863 * other thread deletes or moves the node.
5fd012e0
MD
864 */
865int
866vmntvnodescan(
867 struct mount *mp,
868 int flags,
869 int (*fastfunc)(struct mount *mp, struct vnode *vp, void *data),
870 int (*slowfunc)(struct mount *mp, struct vnode *vp, void *data),
871 void *data
872) {
be6c08cb 873 struct vmntvnodescan_info info;
5fd012e0 874 lwkt_tokref ilock;
5fd012e0
MD
875 struct vnode *vp;
876 int r = 0;
be6c08cb 877 int maxcount = 1000000;
5fd012e0
MD
878
879 lwkt_gettoken(&ilock, &mntvnode_token);
5fd012e0 880
be6c08cb
MD
881 info.vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
882 TAILQ_INSERT_TAIL(&mntvnodescan_list, &info, entry);
883 while ((vp = info.vp) != NULL) {
884 if (--maxcount == 0)
885 panic("maxcount reached during vmntvnodescan");
886
5fd012e0 887 if (vp->v_type == VNON) /* visible but not ready */
be6c08cb 888 goto next;
5fd012e0
MD
889 KKASSERT(vp->v_mount == mp);
890
891 /*
892 * Quick test. A negative return continues the loop without
893 * calling the slow test. 0 continues onto the slow test.
894 * A positive number aborts the loop.
895 */
896 if (fastfunc) {
897 if ((r = fastfunc(mp, vp, data)) < 0)
be6c08cb 898 goto next;
5fd012e0
MD
899 if (r)
900 break;
901 }
902
903 /*
904 * Get a vxlock on the vnode, retry if it has moved or isn't
905 * in the mountlist where we expect it.
906 */
907 if (slowfunc) {
908 int error;
909
910 switch(flags) {
911 case VMSC_GETVP:
87de5057 912 error = vget(vp, LK_EXCLUSIVE);
5fd012e0
MD
913 break;
914 case VMSC_GETVP|VMSC_NOWAIT:
87de5057 915 error = vget(vp, LK_EXCLUSIVE|LK_NOWAIT);
5fd012e0
MD
916 break;
917 case VMSC_GETVX:
e3332475
MD
918 vx_get(vp);
919 error = 0;
5fd012e0 920 break;
5fd012e0
MD
921 default:
922 error = 0;
923 break;
924 }
925 if (error)
be6c08cb
MD
926 goto next;
927 /*
928 * Do not call the slow function if the vnode is
929 * invalid or if it was ripped out from under us
930 * while we (potentially) blocked.
931 */
932 if (info.vp == vp && vp->v_type != VNON)
933 r = slowfunc(mp, vp, data);
934
935 /*
936 * Cleanup
937 */
5fd012e0
MD
938 switch(flags) {
939 case VMSC_GETVP:
940 case VMSC_GETVP|VMSC_NOWAIT:
941 vput(vp);
942 break;
943 case VMSC_GETVX:
944 vx_put(vp);
945 break;
5fd012e0
MD
946 default:
947 break;
948 }
949 if (r != 0)
950 break;
951 }
be6c08cb
MD
952
953 /*
954 * Iterate. If the vnode was ripped out from under us
955 * info.vp will already point to the next vnode, otherwise
956 * we have to obtain the next valid vnode ourselves.
957 */
958next:
959 if (info.vp == vp)
960 info.vp = TAILQ_NEXT(vp, v_nmntvnodes);
5fd012e0 961 }
be6c08cb 962 TAILQ_REMOVE(&mntvnodescan_list, &info, entry);
5fd012e0
MD
963 lwkt_reltoken(&ilock);
964 return(r);
965}
966
967/*
968 * Remove any vnodes in the vnode table belonging to mount point mp.
969 *
970 * If FORCECLOSE is not specified, there should not be any active ones,
971 * return error if any are found (nb: this is a user error, not a
972 * system error). If FORCECLOSE is specified, detach any active vnodes
973 * that are found.
974 *
975 * If WRITECLOSE is set, only flush out regular file vnodes open for
976 * writing.
977 *
978 * SKIPSYSTEM causes any vnodes marked VSYSTEM to be skipped.
979 *
980 * `rootrefs' specifies the base reference count for the root vnode
981 * of this filesystem. The root vnode is considered busy if its
982 * v_usecount exceeds this value. On a successful return, vflush()
983 * will call vrele() on the root vnode exactly rootrefs times.
984 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
985 * be zero.
986 */
987#ifdef DIAGNOSTIC
988static int busyprt = 0; /* print out busy vnodes */
989SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
990#endif
991
992static int vflush_scan(struct mount *mp, struct vnode *vp, void *data);
993
994struct vflush_info {
995 int flags;
996 int busy;
997 thread_t td;
998};
999
1000int
1001vflush(struct mount *mp, int rootrefs, int flags)
1002{
1003 struct thread *td = curthread; /* XXX */
1004 struct vnode *rootvp = NULL;
1005 int error;
1006 struct vflush_info vflush_info;
1007
1008 if (rootrefs > 0) {
1009 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
1010 ("vflush: bad args"));
1011 /*
1012 * Get the filesystem root vnode. We can vput() it
1013 * immediately, since with rootrefs > 0, it won't go away.
1014 */
1015 if ((error = VFS_ROOT(mp, &rootvp)) != 0)
1016 return (error);
1017 vput(rootvp);
1018 }
1019
1020 vflush_info.busy = 0;
1021 vflush_info.flags = flags;
1022 vflush_info.td = td;
1023 vmntvnodescan(mp, VMSC_GETVX, NULL, vflush_scan, &vflush_info);
1024
1025 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
1026 /*
1027 * If just the root vnode is busy, and if its refcount
1028 * is equal to `rootrefs', then go ahead and kill it.
1029 */
1030 KASSERT(vflush_info.busy > 0, ("vflush: not busy"));
1031 KASSERT(rootvp->v_usecount >= rootrefs, ("vflush: rootrefs"));
1032 if (vflush_info.busy == 1 && rootvp->v_usecount == rootrefs) {
e3332475
MD
1033 vx_lock(rootvp);
1034 vgone(rootvp);
1035 vx_unlock(rootvp);
1036 vflush_info.busy = 0;
5fd012e0
MD
1037 }
1038 }
1039 if (vflush_info.busy)
1040 return (EBUSY);
1041 for (; rootrefs > 0; rootrefs--)
1042 vrele(rootvp);
1043 return (0);
1044}
1045
1046/*
1047 * The scan callback is made with an VX locked vnode.
1048 */
1049static int
1050vflush_scan(struct mount *mp, struct vnode *vp, void *data)
1051{
1052 struct vflush_info *info = data;
1053 struct vattr vattr;
1054
1055 /*
1056 * Skip over a vnodes marked VSYSTEM.
1057 */
1058 if ((info->flags & SKIPSYSTEM) && (vp->v_flag & VSYSTEM)) {
1059 return(0);
1060 }
1061
1062 /*
1063 * If WRITECLOSE is set, flush out unlinked but still open
1064 * files (even if open only for reading) and regular file
1065 * vnodes open for writing.
1066 */
1067 if ((info->flags & WRITECLOSE) &&
1068 (vp->v_type == VNON ||
87de5057 1069 (VOP_GETATTR(vp, &vattr) == 0 &&
5fd012e0
MD
1070 vattr.va_nlink > 0)) &&
1071 (vp->v_writecount == 0 || vp->v_type != VREG)) {
1072 return(0);
1073 }
1074
1075 /*
1076 * With v_usecount == 0, all we need to do is clear out the
1077 * vnode data structures and we are done.
1078 */
1079 if (vp->v_usecount == 1) {
1080 vgone(vp);
1081 return(0);
1082 }
1083
1084 /*
1085 * If FORCECLOSE is set, forcibly close the vnode. For block
1086 * or character devices, revert to an anonymous device. For
1087 * all other files, just kill them.
1088 */
1089 if (info->flags & FORCECLOSE) {
1090 if (vp->v_type != VBLK && vp->v_type != VCHR) {
1091 vgone(vp);
1092 } else {
e3332475
MD
1093 spin_lock_wr(&vp->v_spinlock);
1094 vclean_interlocked(vp, 0);
1095 /* spinlock unlocked */
66a1ddf5 1096 vp->v_ops = &spec_vnode_vops_p;
5fd012e0
MD
1097 insmntque(vp, NULL);
1098 }
1099 return(0);
1100 }
1101#ifdef DIAGNOSTIC
1102 if (busyprt)
1103 vprint("vflush: busy vnode", vp);
1104#endif
1105 ++info->busy;
1106 return(0);
1107}
1108