Remove VOP_GETVOBJECT, VOP_DESTROYVOBJECT, and VOP_CREATEVOBJECT. Rearrange
[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 *
7540ab49 70 * $DragonFly: src/sys/kern/vfs_mount.c,v 1.13 2006/03/29 18:44:50 dillon 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,
211 struct vop_ops **ops_pp,
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;
6ddb7618 219 vp->v_ops = ops_pp;
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
861905fb 245vfs_busy(struct mount *mp, int flags, struct thread *td)
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 }
263 lkflags = LK_SHARED | LK_NOPAUSE;
861905fb 264 if (lockmgr(&mp->mnt_lock, lkflags, NULL, td))
5fd012e0
MD
265 panic("vfs_busy: unexpected lock failure");
266 return (0);
267}
268
269/*
270 * Free a busy filesystem.
271 */
272void
273vfs_unbusy(struct mount *mp, struct thread *td)
274{
275 lockmgr(&mp->mnt_lock, LK_RELEASE, NULL, td);
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{
287 struct thread *td = curthread; /* XXX */
288 struct vfsconf *vfsp;
289 struct mount *mp;
290
291 if (fstypename == NULL)
292 return (ENODEV);
293 for (vfsp = vfsconf; vfsp; vfsp = vfsp->vfc_next) {
294 if (!strcmp(vfsp->vfc_name, fstypename))
295 break;
296 }
297 if (vfsp == NULL)
298 return (ENODEV);
2281065e 299 mp = malloc(sizeof(struct mount), M_MOUNT, M_WAITOK);
5fd012e0 300 bzero((char *)mp, (u_long)sizeof(struct mount));
f2770c70 301 lockinit(&mp->mnt_lock, "vfslock", VLKTIMEOUT, LK_NOPAUSE);
861905fb 302 vfs_busy(mp, LK_NOWAIT, td);
5fd012e0
MD
303 TAILQ_INIT(&mp->mnt_nvnodelist);
304 TAILQ_INIT(&mp->mnt_reservedvnlist);
2281065e 305 TAILQ_INIT(&mp->mnt_jlist);
5fd012e0
MD
306 mp->mnt_nvnodelistsize = 0;
307 mp->mnt_vfc = vfsp;
308 mp->mnt_op = vfsp->vfc_vfsops;
309 mp->mnt_flag = MNT_RDONLY;
310 mp->mnt_vnodecovered = NULLVP;
311 vfsp->vfc_refcount++;
312 mp->mnt_iosize_max = DFLTPHYS;
313 mp->mnt_stat.f_type = vfsp->vfc_typenum;
314 mp->mnt_flag |= vfsp->vfc_flags & MNT_VISFLAGMASK;
315 strncpy(mp->mnt_stat.f_fstypename, vfsp->vfc_name, MFSNAMELEN);
069b825e 316 copystr(devname, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, 0);
5fd012e0
MD
317 *mpp = mp;
318 return (0);
319}
320
321/*
322 * Lookup a mount point by filesystem identifier.
323 */
324struct mount *
325vfs_getvfs(fsid_t *fsid)
326{
327 struct mount *mp;
328 lwkt_tokref ilock;
329
330 lwkt_gettoken(&ilock, &mountlist_token);
331 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
332 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
333 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
334 break;
335 }
336 }
337 lwkt_reltoken(&ilock);
338 return (mp);
339}
340
341/*
342 * Get a new unique fsid. Try to make its val[0] unique, since this value
343 * will be used to create fake device numbers for stat(). Also try (but
344 * not so hard) make its val[0] unique mod 2^16, since some emulators only
345 * support 16-bit device numbers. We end up with unique val[0]'s for the
346 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
347 *
348 * Keep in mind that several mounts may be running in parallel. Starting
349 * the search one past where the previous search terminated is both a
350 * micro-optimization and a defense against returning the same fsid to
351 * different mounts.
352 */
353void
354vfs_getnewfsid(struct mount *mp)
355{
356 static u_int16_t mntid_base;
357 lwkt_tokref ilock;
358 fsid_t tfsid;
359 int mtype;
360
361 lwkt_gettoken(&ilock, &mntid_token);
362 mtype = mp->mnt_vfc->vfc_typenum;
363 tfsid.val[1] = mtype;
364 mtype = (mtype & 0xFF) << 24;
365 for (;;) {
366 tfsid.val[0] = makeudev(255,
367 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
368 mntid_base++;
369 if (vfs_getvfs(&tfsid) == NULL)
370 break;
371 }
372 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
373 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
374 lwkt_reltoken(&ilock);
375}
376
377/*
378 * This routine is called when we have too many vnodes. It attempts
379 * to free <count> vnodes and will potentially free vnodes that still
380 * have VM backing store (VM backing store is typically the cause
381 * of a vnode blowout so we want to do this). Therefore, this operation
382 * is not considered cheap.
383 *
384 * A number of conditions may prevent a vnode from being reclaimed.
385 * the buffer cache may have references on the vnode, a directory
386 * vnode may still have references due to the namei cache representing
387 * underlying files, or the vnode may be in active use. It is not
388 * desireable to reuse such vnodes. These conditions may cause the
389 * number of vnodes to reach some minimum value regardless of what
390 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low.
391 */
392
393/*
58552887
MD
394 * This is a quick non-blocking check to determine if the vnode is a good
395 * candidate for being (eventually) vgone()'d. Returns 0 if the vnode is
396 * not a good candidate, 1 if it is.
397 *
398 * vnodes marked VFREE are already on the free list, but may still need
399 * to be recycled due to eating namecache resources and potentially blocking
400 * the namecache directory chain and related vnodes from being freed.
5fd012e0
MD
401 */
402static __inline int
250d127c 403vmightfree(struct vnode *vp, int page_count)
5fd012e0 404{
58552887
MD
405 if (vp->v_flag & VRECLAIMED)
406 return (0);
407 if ((vp->v_flag & VFREE) && TAILQ_EMPTY(&vp->v_namecache))
5fd012e0 408 return (0);
250d127c 409 if (vp->v_usecount != 0)
5fd012e0
MD
410 return (0);
411 if (vp->v_object && vp->v_object->resident_page_count >= page_count)
412 return (0);
413 return (1);
414}
415
250d127c 416/*
58552887 417 * The vnode was found to be possibly vgone()able and the caller has locked it
250d127c 418 * (thus the usecount should be 1 now). Determine if the vnode is actually
58552887
MD
419 * vgone()able, doing some cleanups in the process. Returns 1 if the vnode
420 * can be vgone()'d, 0 otherwise.
250d127c
MD
421 *
422 * Note that v_holdcnt may be non-zero because (A) this vnode is not a leaf
423 * in the namecache topology and (B) this vnode has buffer cache bufs.
424 * We cannot remove vnodes with non-leaf namecache associations. We do a
425 * tentitive leaf check prior to attempting to flush out any buffers but the
426 * 'real' test when all is said in done is that v_holdcnt must become 0 for
427 * the vnode to be freeable.
428 *
429 * We could theoretically just unconditionally flush when v_holdcnt != 0,
430 * but flushing data associated with non-leaf nodes (which are always
431 * directories), just throws it away for no benefit. It is the buffer
432 * cache's responsibility to choose buffers to recycle from the cached
433 * data point of view.
434 */
435static int
436visleaf(struct vnode *vp)
437{
438 struct namecache *ncp;
439
440 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
441 if (!TAILQ_EMPTY(&ncp->nc_list))
442 return(0);
443 }
444 return(1);
445}
446
58552887
MD
447/*
448 * Try to clean up the vnode to the point where it can be vgone()'d, returning
449 * 0 if it cannot be vgone()'d (or already has been), 1 if it can. Unlike
450 * vmightfree() this routine may flush the vnode and block. Vnodes marked
451 * VFREE are still candidates for vgone()ing because they may hold namecache
452 * resources and could be blocking the namecache directory hierarchy (and
453 * related vnodes) from being freed.
454 */
250d127c 455static int
58552887 456vtrytomakegoneable(struct vnode *vp, int page_count)
250d127c 457{
58552887 458 if (vp->v_flag & VRECLAIMED)
250d127c
MD
459 return (0);
460 if (vp->v_usecount != 1)
461 return (0);
462 if (vp->v_object && vp->v_object->resident_page_count >= page_count)
463 return (0);
464 if (vp->v_holdcnt && visleaf(vp)) {
465 vinvalbuf(vp, V_SAVE, NULL, 0, 0);
466#if 0 /* DEBUG */
467 printf((vp->v_holdcnt ? "vrecycle: vp %p failed: %s\n" :
468 "vrecycle: vp %p succeeded: %s\n"), vp,
469 (TAILQ_FIRST(&vp->v_namecache) ?
470 TAILQ_FIRST(&vp->v_namecache)->nc_name : "?"));
471#endif
472 }
473 return(vp->v_usecount == 1 && vp->v_holdcnt == 0);
474}
5fd012e0 475
58552887
MD
476/*
477 * Reclaim up to 1/10 of the vnodes associated with a mount point. Try
478 * to avoid vnodes which have lots of resident pages (we are trying to free
861905fb
MD
479 * vnodes, not memory).
480 *
481 * This routine is a callback from the mountlist scan. The mount point
482 * in question will be busied.
58552887 483 */
5fd012e0 484static int
861905fb 485vlrureclaim(struct mount *mp, void *data)
5fd012e0
MD
486{
487 struct vnode *vp;
488 lwkt_tokref ilock;
489 int done;
490 int trigger;
491 int usevnodes;
492 int count;
861905fb 493 int trigger_mult = vnlru_nowhere;
5fd012e0
MD
494
495 /*
58552887
MD
496 * Calculate the trigger point for the resident pages check. The
497 * minimum trigger value is approximately the number of pages in
498 * the system divded by the number of vnodes. However, due to
499 * various other system memory overheads unrelated to data caching
500 * it is a good idea to double the trigger (at least).
501 *
502 * trigger_mult starts at 0. If the recycler is having problems
503 * finding enough freeable vnodes it will increase trigger_mult.
504 * This should not happen in normal operation, even on machines with
505 * low amounts of memory, but extraordinary memory use by the system
506 * verses the amount of cached data can trigger it.
5fd012e0
MD
507 */
508 usevnodes = desiredvnodes;
509 if (usevnodes <= 0)
510 usevnodes = 1;
58552887 511 trigger = vmstats.v_page_count * (trigger_mult + 2) / usevnodes;
5fd012e0
MD
512
513 done = 0;
514 lwkt_gettoken(&ilock, &mntvnode_token);
515 count = mp->mnt_nvnodelistsize / 10 + 1;
516 while (count && (vp = TAILQ_FIRST(&mp->mnt_nvnodelist)) != NULL) {
517 /*
518 * __VNODESCAN__
519 *
520 * The VP will stick around while we hold mntvnode_token,
521 * at least until we block, so we can safely do an initial
522 * check, and then must check again after we lock the vnode.
523 */
524 if (vp->v_type == VNON || /* XXX */
525 vp->v_type == VBAD || /* XXX */
250d127c 526 !vmightfree(vp, trigger) /* critical path opt */
5fd012e0 527 ) {
be6c08cb 528 vmovevnodetoend(mp, vp);
5fd012e0
MD
529 --count;
530 continue;
531 }
532
533 /*
534 * VX get the candidate vnode. If the VX get fails the
535 * vnode might still be on the mountlist. Our loop depends
536 * on us at least cycling the vnode to the end of the
537 * mountlist.
538 */
539 if (vx_get_nonblock(vp) != 0) {
be6c08cb
MD
540 if (vp->v_mount == mp)
541 vmovevnodetoend(mp, vp);
5fd012e0
MD
542 --count;
543 continue;
544 }
545
546 /*
547 * Since we blocked locking the vp, make sure it is still
548 * a candidate for reclamation. That is, it has not already
549 * been reclaimed and only has our VX reference associated
550 * with it.
551 */
552 if (vp->v_type == VNON || /* XXX */
553 vp->v_type == VBAD || /* XXX */
554 (vp->v_flag & VRECLAIMED) ||
555 vp->v_mount != mp ||
58552887 556 !vtrytomakegoneable(vp, trigger) /* critical path opt */
5fd012e0 557 ) {
be6c08cb
MD
558 if (vp->v_mount == mp)
559 vmovevnodetoend(mp, vp);
5fd012e0
MD
560 --count;
561 vx_put(vp);
562 continue;
563 }
564
565 /*
566 * All right, we are good, move the vp to the end of the
567 * mountlist and clean it out. The vget will have returned
568 * an error if the vnode was destroyed (VRECLAIMED set), so we
569 * do not have to check again. The vput() will move the
570 * vnode to the free list if the vgone() was successful.
571 */
572 KKASSERT(vp->v_mount == mp);
be6c08cb 573 vmovevnodetoend(mp, vp);
5fd012e0
MD
574 vgone(vp);
575 vx_put(vp);
576 ++done;
577 --count;
578 }
579 lwkt_reltoken(&ilock);
580 return (done);
581}
582
583/*
584 * Attempt to recycle vnodes in a context that is always safe to block.
585 * Calling vlrurecycle() from the bowels of file system code has some
586 * interesting deadlock problems.
587 */
588static struct thread *vnlruthread;
589static int vnlruproc_sig;
590
591void
592vnlru_proc_wait(void)
593{
594 if (vnlruproc_sig == 0) {
595 vnlruproc_sig = 1; /* avoid unnecessary wakeups */
596 wakeup(vnlruthread);
597 }
598 tsleep(&vnlruproc_sig, 0, "vlruwk", hz);
599}
600
601static void
602vnlru_proc(void)
603{
5fd012e0 604 struct thread *td = curthread;
861905fb 605 int done;
5fd012e0
MD
606
607 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, td,
608 SHUTDOWN_PRI_FIRST);
609
e43a034f 610 crit_enter();
5fd012e0
MD
611 for (;;) {
612 kproc_suspend_loop();
613 if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) {
614 vnlruproc_sig = 0;
615 wakeup(&vnlruproc_sig);
616 tsleep(td, 0, "vlruwt", hz);
617 continue;
618 }
62d0f1f0 619 cache_cleanneg(0);
861905fb 620 done = mountlist_scan(vlrureclaim, NULL, MNTSCAN_FORWARD);
58552887
MD
621
622 /*
623 * The vlrureclaim() call only processes 1/10 of the vnodes
624 * on each mount. If we couldn't find any repeat the loop
625 * at least enough times to cover all available vnodes before
626 * we start sleeping. Complain if the failure extends past
627 * 30 second, every 30 seconds.
628 */
5fd012e0
MD
629 if (done == 0) {
630 ++vnlru_nowhere;
5fd012e0 631 if (vnlru_nowhere % 10 == 0)
58552887
MD
632 tsleep(td, 0, "vlrup", hz * 3);
633 if (vnlru_nowhere % 100 == 0)
5fd012e0 634 printf("vnlru_proc: vnode recycler stopped working!\n");
58552887
MD
635 if (vnlru_nowhere == 1000)
636 vnlru_nowhere = 900;
5fd012e0
MD
637 } else {
638 vnlru_nowhere = 0;
639 }
640 }
e43a034f 641 crit_exit();
5fd012e0
MD
642}
643
861905fb
MD
644/*
645 * MOUNTLIST FUNCTIONS
646 */
647
648/*
649 * mountlist_insert (MP SAFE)
650 *
651 * Add a new mount point to the mount list.
652 */
653void
654mountlist_insert(struct mount *mp, int how)
655{
656 lwkt_tokref ilock;
657
658 lwkt_gettoken(&ilock, &mountlist_token);
659 if (how == MNTINS_FIRST)
660 TAILQ_INSERT_HEAD(&mountlist, mp, mnt_list);
661 else
662 TAILQ_INSERT_TAIL(&mountlist, mp, mnt_list);
663 lwkt_reltoken(&ilock);
664}
665
666/*
667 * mountlist_interlock (MP SAFE)
668 *
669 * Execute the specified interlock function with the mountlist token
670 * held. The function will be called in a serialized fashion verses
671 * other functions called through this mechanism.
672 */
673int
674mountlist_interlock(int (*callback)(struct mount *), struct mount *mp)
675{
676 lwkt_tokref ilock;
677 int error;
678
679 lwkt_gettoken(&ilock, &mountlist_token);
680 error = callback(mp);
681 lwkt_reltoken(&ilock);
682 return (error);
683}
684
685/*
686 * mountlist_boot_getfirst (DURING BOOT ONLY)
687 *
688 * This function returns the first mount on the mountlist, which is
689 * expected to be the root mount. Since no interlocks are obtained
690 * this function is only safe to use during booting.
691 */
692
693struct mount *
694mountlist_boot_getfirst(void)
695{
696 return(TAILQ_FIRST(&mountlist));
697}
698
699/*
700 * mountlist_remove (MP SAFE)
701 *
702 * Remove a node from the mountlist. If this node is the next scan node
703 * for any active mountlist scans, the active mountlist scan will be
704 * adjusted to skip the node, thus allowing removals during mountlist
705 * scans.
706 */
707void
708mountlist_remove(struct mount *mp)
709{
710 struct mountscan_info *msi;
711 lwkt_tokref ilock;
712
713 lwkt_gettoken(&ilock, &mountlist_token);
714 TAILQ_FOREACH(msi, &mountscan_list, msi_entry) {
715 if (msi->msi_node == mp) {
716 if (msi->msi_how & MNTSCAN_FORWARD)
717 msi->msi_node = TAILQ_NEXT(mp, mnt_list);
718 else
719 msi->msi_node = TAILQ_PREV(mp, mntlist, mnt_list);
720 }
721 }
722 TAILQ_REMOVE(&mountlist, mp, mnt_list);
723 lwkt_reltoken(&ilock);
724}
725
726/*
727 * mountlist_scan (MP SAFE)
728 *
729 * Safely scan the mount points on the mount list. Unless otherwise
730 * specified each mount point will be busied prior to the callback and
731 * unbusied afterwords. The callback may safely remove any mount point
732 * without interfering with the scan. If the current callback
733 * mount is removed the scanner will not attempt to unbusy it.
734 *
735 * If a mount node cannot be busied it is silently skipped.
736 *
737 * The callback return value is aggregated and a total is returned. A return
738 * value of < 0 is not aggregated and will terminate the scan.
739 *
740 * MNTSCAN_FORWARD - the mountlist is scanned in the forward direction
741 * MNTSCAN_REVERSE - the mountlist is scanned in reverse
742 * MNTSCAN_NOBUSY - the scanner will make the callback without busying
743 * the mount node.
744 */
745int
746mountlist_scan(int (*callback)(struct mount *, void *), void *data, int how)
747{
748 struct mountscan_info info;
749 lwkt_tokref ilock;
750 struct mount *mp;
751 thread_t td;
752 int count;
753 int res;
754
755 lwkt_gettoken(&ilock, &mountlist_token);
756
757 info.msi_how = how;
758 info.msi_node = NULL; /* paranoia */
759 TAILQ_INSERT_TAIL(&mountscan_list, &info, msi_entry);
760
761 res = 0;
762 td = curthread;
763
764 if (how & MNTSCAN_FORWARD) {
765 info.msi_node = TAILQ_FIRST(&mountlist);
766 while ((mp = info.msi_node) != NULL) {
767 if (how & MNTSCAN_NOBUSY) {
768 count = callback(mp, data);
769 } else if (vfs_busy(mp, LK_NOWAIT, td) == 0) {
770 count = callback(mp, data);
771 if (mp == info.msi_node)
772 vfs_unbusy(mp, td);
773 } else {
774 count = 0;
775 }
776 if (count < 0)
777 break;
778 res += count;
779 if (mp == info.msi_node)
780 info.msi_node = TAILQ_NEXT(mp, mnt_list);
781 }
782 } else if (how & MNTSCAN_REVERSE) {
783 info.msi_node = TAILQ_LAST(&mountlist, mntlist);
784 while ((mp = info.msi_node) != NULL) {
785 if (how & MNTSCAN_NOBUSY) {
786 count = callback(mp, data);
787 } else if (vfs_busy(mp, LK_NOWAIT, td) == 0) {
788 count = callback(mp, data);
789 if (mp == info.msi_node)
790 vfs_unbusy(mp, td);
791 } else {
792 count = 0;
793 }
794 if (count < 0)
795 break;
796 res += count;
797 if (mp == info.msi_node)
798 info.msi_node = TAILQ_PREV(mp, mntlist, mnt_list);
799 }
800 }
801 TAILQ_REMOVE(&mountscan_list, &info, msi_entry);
802 lwkt_reltoken(&ilock);
803 return(res);
804}
805
806/*
807 * MOUNT RELATED VNODE FUNCTIONS
808 */
809
5fd012e0
MD
810static struct kproc_desc vnlru_kp = {
811 "vnlru",
812 vnlru_proc,
813 &vnlruthread
814};
815SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
816
817/*
818 * Move a vnode from one mount queue to another.
819 */
820void
821insmntque(struct vnode *vp, struct mount *mp)
822{
823 lwkt_tokref ilock;
824
825 lwkt_gettoken(&ilock, &mntvnode_token);
826 /*
827 * Delete from old mount point vnode list, if on one.
828 */
829 if (vp->v_mount != NULL) {
830 KASSERT(vp->v_mount->mnt_nvnodelistsize > 0,
831 ("bad mount point vnode list size"));
be6c08cb 832 vremovevnodemnt(vp);
5fd012e0
MD
833 vp->v_mount->mnt_nvnodelistsize--;
834 }
835 /*
836 * Insert into list of vnodes for the new mount point, if available.
837 */
838 if ((vp->v_mount = mp) == NULL) {
839 lwkt_reltoken(&ilock);
840 return;
841 }
842 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
843 mp->mnt_nvnodelistsize++;
844 lwkt_reltoken(&ilock);
845}
846
847
848/*
be6c08cb
MD
849 * Scan the vnodes under a mount point and issue appropriate callbacks.
850 *
851 * The fastfunc() callback is called with just the mountlist token held
852 * (no vnode lock). It may not block and the vnode may be undergoing
853 * modifications while the caller is processing it. The vnode will
854 * not be entirely destroyed, however, due to the fact that the mountlist
855 * token is held. A return value < 0 skips to the next vnode without calling
856 * the slowfunc(), a return value > 0 terminates the loop.
857 *
858 * The slowfunc() callback is called after the vnode has been successfully
859 * locked based on passed flags. The vnode is skipped if it gets rearranged
860 * or destroyed while blocking on the lock. A non-zero return value from
861 * the slow function terminates the loop. The slow function is allowed to
862 * arbitrarily block. The scanning code guarentees consistency of operation
863 * even if the slow function deletes or moves the node, or blocks and some
864 * other thread deletes or moves the node.
5fd012e0
MD
865 */
866int
867vmntvnodescan(
868 struct mount *mp,
869 int flags,
870 int (*fastfunc)(struct mount *mp, struct vnode *vp, void *data),
871 int (*slowfunc)(struct mount *mp, struct vnode *vp, void *data),
872 void *data
873) {
be6c08cb 874 struct vmntvnodescan_info info;
5fd012e0 875 lwkt_tokref ilock;
5fd012e0
MD
876 struct vnode *vp;
877 int r = 0;
be6c08cb 878 int maxcount = 1000000;
5fd012e0
MD
879
880 lwkt_gettoken(&ilock, &mntvnode_token);
5fd012e0 881
be6c08cb
MD
882 info.vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
883 TAILQ_INSERT_TAIL(&mntvnodescan_list, &info, entry);
884 while ((vp = info.vp) != NULL) {
885 if (--maxcount == 0)
886 panic("maxcount reached during vmntvnodescan");
887
5fd012e0 888 if (vp->v_type == VNON) /* visible but not ready */
be6c08cb 889 goto next;
5fd012e0
MD
890 KKASSERT(vp->v_mount == mp);
891
892 /*
893 * Quick test. A negative return continues the loop without
894 * calling the slow test. 0 continues onto the slow test.
895 * A positive number aborts the loop.
896 */
897 if (fastfunc) {
898 if ((r = fastfunc(mp, vp, data)) < 0)
be6c08cb 899 goto next;
5fd012e0
MD
900 if (r)
901 break;
902 }
903
904 /*
905 * Get a vxlock on the vnode, retry if it has moved or isn't
906 * in the mountlist where we expect it.
907 */
908 if (slowfunc) {
909 int error;
910
911 switch(flags) {
912 case VMSC_GETVP:
913 error = vget(vp, LK_EXCLUSIVE, curthread);
914 break;
915 case VMSC_GETVP|VMSC_NOWAIT:
916 error = vget(vp, LK_EXCLUSIVE|LK_NOWAIT,
917 curthread);
918 break;
919 case VMSC_GETVX:
920 error = vx_get(vp);
921 break;
922 case VMSC_REFVP:
923 vref(vp);
924 /* fall through */
925 default:
926 error = 0;
927 break;
928 }
929 if (error)
be6c08cb
MD
930 goto next;
931 /*
932 * Do not call the slow function if the vnode is
933 * invalid or if it was ripped out from under us
934 * while we (potentially) blocked.
935 */
936 if (info.vp == vp && vp->v_type != VNON)
937 r = slowfunc(mp, vp, data);
938
939 /*
940 * Cleanup
941 */
5fd012e0
MD
942 switch(flags) {
943 case VMSC_GETVP:
944 case VMSC_GETVP|VMSC_NOWAIT:
945 vput(vp);
946 break;
947 case VMSC_GETVX:
948 vx_put(vp);
949 break;
950 case VMSC_REFVP:
951 vrele(vp);
952 /* fall through */
953 default:
954 break;
955 }
956 if (r != 0)
957 break;
958 }
be6c08cb
MD
959
960 /*
961 * Iterate. If the vnode was ripped out from under us
962 * info.vp will already point to the next vnode, otherwise
963 * we have to obtain the next valid vnode ourselves.
964 */
965next:
966 if (info.vp == vp)
967 info.vp = TAILQ_NEXT(vp, v_nmntvnodes);
5fd012e0 968 }
be6c08cb 969 TAILQ_REMOVE(&mntvnodescan_list, &info, entry);
5fd012e0
MD
970 lwkt_reltoken(&ilock);
971 return(r);
972}
973
974/*
975 * Remove any vnodes in the vnode table belonging to mount point mp.
976 *
977 * If FORCECLOSE is not specified, there should not be any active ones,
978 * return error if any are found (nb: this is a user error, not a
979 * system error). If FORCECLOSE is specified, detach any active vnodes
980 * that are found.
981 *
982 * If WRITECLOSE is set, only flush out regular file vnodes open for
983 * writing.
984 *
985 * SKIPSYSTEM causes any vnodes marked VSYSTEM to be skipped.
986 *
987 * `rootrefs' specifies the base reference count for the root vnode
988 * of this filesystem. The root vnode is considered busy if its
989 * v_usecount exceeds this value. On a successful return, vflush()
990 * will call vrele() on the root vnode exactly rootrefs times.
991 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
992 * be zero.
993 */
994#ifdef DIAGNOSTIC
995static int busyprt = 0; /* print out busy vnodes */
996SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
997#endif
998
999static int vflush_scan(struct mount *mp, struct vnode *vp, void *data);
1000
1001struct vflush_info {
1002 int flags;
1003 int busy;
1004 thread_t td;
1005};
1006
1007int
1008vflush(struct mount *mp, int rootrefs, int flags)
1009{
1010 struct thread *td = curthread; /* XXX */
1011 struct vnode *rootvp = NULL;
1012 int error;
1013 struct vflush_info vflush_info;
1014
1015 if (rootrefs > 0) {
1016 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
1017 ("vflush: bad args"));
1018 /*
1019 * Get the filesystem root vnode. We can vput() it
1020 * immediately, since with rootrefs > 0, it won't go away.
1021 */
1022 if ((error = VFS_ROOT(mp, &rootvp)) != 0)
1023 return (error);
1024 vput(rootvp);
1025 }
1026
1027 vflush_info.busy = 0;
1028 vflush_info.flags = flags;
1029 vflush_info.td = td;
1030 vmntvnodescan(mp, VMSC_GETVX, NULL, vflush_scan, &vflush_info);
1031
1032 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
1033 /*
1034 * If just the root vnode is busy, and if its refcount
1035 * is equal to `rootrefs', then go ahead and kill it.
1036 */
1037 KASSERT(vflush_info.busy > 0, ("vflush: not busy"));
1038 KASSERT(rootvp->v_usecount >= rootrefs, ("vflush: rootrefs"));
1039 if (vflush_info.busy == 1 && rootvp->v_usecount == rootrefs) {
1040 if (vx_lock(rootvp) == 0) {
1041 vgone(rootvp);
1042 vx_unlock(rootvp);
1043 vflush_info.busy = 0;
1044 }
1045 }
1046 }
1047 if (vflush_info.busy)
1048 return (EBUSY);
1049 for (; rootrefs > 0; rootrefs--)
1050 vrele(rootvp);
1051 return (0);
1052}
1053
1054/*
1055 * The scan callback is made with an VX locked vnode.
1056 */
1057static int
1058vflush_scan(struct mount *mp, struct vnode *vp, void *data)
1059{
1060 struct vflush_info *info = data;
1061 struct vattr vattr;
1062
1063 /*
1064 * Skip over a vnodes marked VSYSTEM.
1065 */
1066 if ((info->flags & SKIPSYSTEM) && (vp->v_flag & VSYSTEM)) {
1067 return(0);
1068 }
1069
1070 /*
1071 * If WRITECLOSE is set, flush out unlinked but still open
1072 * files (even if open only for reading) and regular file
1073 * vnodes open for writing.
1074 */
1075 if ((info->flags & WRITECLOSE) &&
1076 (vp->v_type == VNON ||
1077 (VOP_GETATTR(vp, &vattr, info->td) == 0 &&
1078 vattr.va_nlink > 0)) &&
1079 (vp->v_writecount == 0 || vp->v_type != VREG)) {
1080 return(0);
1081 }
1082
1083 /*
1084 * With v_usecount == 0, all we need to do is clear out the
1085 * vnode data structures and we are done.
1086 */
1087 if (vp->v_usecount == 1) {
1088 vgone(vp);
1089 return(0);
1090 }
1091
1092 /*
1093 * If FORCECLOSE is set, forcibly close the vnode. For block
1094 * or character devices, revert to an anonymous device. For
1095 * all other files, just kill them.
1096 */
1097 if (info->flags & FORCECLOSE) {
1098 if (vp->v_type != VBLK && vp->v_type != VCHR) {
1099 vgone(vp);
1100 } else {
1101 vclean(vp, 0, info->td);
6ddb7618 1102 vp->v_ops = &spec_vnode_vops;
5fd012e0
MD
1103 insmntque(vp, NULL);
1104 }
1105 return(0);
1106 }
1107#ifdef DIAGNOSTIC
1108 if (busyprt)
1109 vprint("vflush: busy vnode", vp);
1110#endif
1111 ++info->busy;
1112 return(0);
1113}
1114