kernel - Fix mount refs interactions and umount races
[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 *
67863d04 70 * $DragonFly: src/sys/kern/vfs_mount.c,v 1.37 2008/09/17 21:44:18 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>
3c37c940 94#include <sys/sysref2.h>
cd8ab232 95#include <sys/mplock2.h>
5fd012e0
MD
96
97#include <vm/vm.h>
98#include <vm/vm_object.h>
99
861905fb
MD
100struct mountscan_info {
101 TAILQ_ENTRY(mountscan_info) msi_entry;
102 int msi_how;
103 struct mount *msi_node;
104};
105
be6c08cb
MD
106struct vmntvnodescan_info {
107 TAILQ_ENTRY(vmntvnodescan_info) entry;
108 struct vnode *vp;
109};
110
0e8bd897
MD
111struct vnlru_info {
112 int pass;
113};
114
5fd012e0 115static int vnlru_nowhere = 0;
58552887 116SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RD,
5fd012e0
MD
117 &vnlru_nowhere, 0,
118 "Number of times the vnlru process ran without success");
119
120
121static struct lwkt_token mntid_token;
aac0aabd 122static struct mount dummymount;
5fd012e0 123
460426e6
MD
124/* note: mountlist exported to pstat */
125struct mntlist mountlist = TAILQ_HEAD_INITIALIZER(mountlist);
861905fb
MD
126static TAILQ_HEAD(,mountscan_info) mountscan_list;
127static struct lwkt_token mountlist_token;
be6c08cb 128static TAILQ_HEAD(,vmntvnodescan_info) mntvnodescan_list;
5fd012e0
MD
129struct lwkt_token mntvnode_token;
130
408357d8
MD
131static TAILQ_HEAD(,bio_ops) bio_ops_list = TAILQ_HEAD_INITIALIZER(bio_ops_list);
132
5fd012e0
MD
133/*
134 * Called from vfsinit()
135 */
136void
137vfs_mount_init(void)
138{
a3c18566
MD
139 lwkt_token_init(&mountlist_token, "mntlist");
140 lwkt_token_init(&mntvnode_token, "mntvnode");
141 lwkt_token_init(&mntid_token, "mntid");
861905fb 142 TAILQ_INIT(&mountscan_list);
be6c08cb 143 TAILQ_INIT(&mntvnodescan_list);
aac0aabd
MD
144 mount_init(&dummymount);
145 dummymount.mnt_flag |= MNT_RDONLY;
572a7ed9 146 dummymount.mnt_kern_flag |= MNTK_ALL_MPSAFE;
be6c08cb
MD
147}
148
149/*
150 * Support function called with mntvnode_token held to remove a vnode
151 * from the mountlist. We must update any list scans which are in progress.
152 */
153static void
154vremovevnodemnt(struct vnode *vp)
155{
156 struct vmntvnodescan_info *info;
157
158 TAILQ_FOREACH(info, &mntvnodescan_list, entry) {
159 if (info->vp == vp)
160 info->vp = TAILQ_NEXT(vp, v_nmntvnodes);
161 }
162 TAILQ_REMOVE(&vp->v_mount->mnt_nvnodelist, vp, v_nmntvnodes);
163}
164
165/*
5fd012e0
MD
166 * Allocate a new vnode and associate it with a tag, mount point, and
167 * operations vector.
168 *
169 * A VX locked and refd vnode is returned. The caller should setup the
170 * remaining fields and vx_put() or, if he wishes to leave a vref,
171 * vx_unlock() the vnode.
172 */
173int
6ddb7618
MD
174getnewvnode(enum vtagtype tag, struct mount *mp,
175 struct vnode **vpp, int lktimeout, int lkflags)
176{
177 struct vnode *vp;
178
179 KKASSERT(mp != NULL);
180
181 vp = allocvnode(lktimeout, lkflags);
182 vp->v_tag = tag;
183 vp->v_data = NULL;
184
185 /*
186 * By default the vnode is assigned the mount point's normal
187 * operations vector.
188 */
189 vp->v_ops = &mp->mnt_vn_use_ops;
190
191 /*
192 * Placing the vnode on the mount point's queue makes it visible.
193 * VNON prevents it from being messed with, however.
194 */
195 insmntque(vp, mp);
6ddb7618
MD
196
197 /*
198 * A VX locked & refd vnode is returned.
199 */
200 *vpp = vp;
201 return (0);
202}
203
204/*
205 * This function creates vnodes with special operations vectors. The
206 * mount point is optional.
207 *
aac0aabd
MD
208 * This routine is being phased out but is still used by vfs_conf to
209 * create vnodes for devices prior to the root mount (with mp == NULL).
6ddb7618
MD
210 */
211int
212getspecialvnode(enum vtagtype tag, struct mount *mp,
66a1ddf5 213 struct vop_ops **ops,
5fd012e0
MD
214 struct vnode **vpp, int lktimeout, int lkflags)
215{
216 struct vnode *vp;
217
218 vp = allocvnode(lktimeout, lkflags);
219 vp->v_tag = tag;
5fd012e0 220 vp->v_data = NULL;
66a1ddf5 221 vp->v_ops = ops;
5fd012e0 222
aac0aabd
MD
223 if (mp == NULL)
224 mp = &dummymount;
225
5fd012e0
MD
226 /*
227 * Placing the vnode on the mount point's queue makes it visible.
228 * VNON prevents it from being messed with, however.
229 */
230 insmntque(vp, mp);
5fd012e0
MD
231
232 /*
233 * A VX locked & refd vnode is returned.
234 */
235 *vpp = vp;
236 return (0);
237}
238
239/*
861905fb
MD
240 * Interlock against an unmount, return 0 on success, non-zero on failure.
241 *
242 * The passed flag may be 0 or LK_NOWAIT and is only used if an unmount
243 * is in-progress.
244 *
245 * If no unmount is in-progress LK_NOWAIT is ignored. No other flag bits
246 * are used. A shared locked will be obtained and the filesystem will not
247 * be unmountable until the lock is released.
5fd012e0
MD
248 */
249int
f9642f56 250vfs_busy(struct mount *mp, int flags)
5fd012e0
MD
251{
252 int lkflags;
253
412f07cd 254 atomic_add_int(&mp->mnt_refs, 1);
5fd012e0 255 if (mp->mnt_kern_flag & MNTK_UNMOUNT) {
412f07cd
MD
256 if (flags & LK_NOWAIT) {
257 atomic_add_int(&mp->mnt_refs, -1);
5fd012e0 258 return (ENOENT);
412f07cd 259 }
861905fb 260 /* XXX not MP safe */
5fd012e0
MD
261 mp->mnt_kern_flag |= MNTK_MWAIT;
262 /*
263 * Since all busy locks are shared except the exclusive
264 * lock granted when unmounting, the only place that a
265 * wakeup needs to be done is at the release of the
266 * exclusive lock at the end of dounmount.
5fd012e0
MD
267 */
268 tsleep((caddr_t)mp, 0, "vfs_busy", 0);
412f07cd 269 atomic_add_int(&mp->mnt_refs, -1);
5fd012e0
MD
270 return (ENOENT);
271 }
ab6f251b 272 lkflags = LK_SHARED;
df4f70a6 273 if (lockmgr(&mp->mnt_lock, lkflags))
5fd012e0
MD
274 panic("vfs_busy: unexpected lock failure");
275 return (0);
276}
277
278/*
279 * Free a busy filesystem.
412f07cd
MD
280 *
281 * Decrement refs before releasing the lock so e.g. a pending umount
282 * doesn't give us an unexpected busy error.
5fd012e0
MD
283 */
284void
f9642f56 285vfs_unbusy(struct mount *mp)
5fd012e0 286{
412f07cd 287 atomic_add_int(&mp->mnt_refs, -1);
df4f70a6 288 lockmgr(&mp->mnt_lock, LK_RELEASE);
5fd012e0
MD
289}
290
291/*
292 * Lookup a filesystem type, and if found allocate and initialize
293 * a mount structure for it.
294 *
295 * Devname is usually updated by mount(8) after booting.
296 */
297int
298vfs_rootmountalloc(char *fstypename, char *devname, struct mount **mpp)
299{
5fd012e0
MD
300 struct vfsconf *vfsp;
301 struct mount *mp;
302
303 if (fstypename == NULL)
304 return (ENODEV);
2613053d
MN
305
306 vfsp = vfsconf_find_by_name(fstypename);
5fd012e0
MD
307 if (vfsp == NULL)
308 return (ENODEV);
e7b4468c 309 mp = kmalloc(sizeof(struct mount), M_MOUNT, M_WAITOK | M_ZERO);
aac0aabd 310 mount_init(mp);
ab6f251b 311 lockinit(&mp->mnt_lock, "vfslock", VLKTIMEOUT, 0);
aac0aabd 312
fd4faaba 313 vfs_busy(mp, 0);
5fd012e0
MD
314 mp->mnt_vfc = vfsp;
315 mp->mnt_op = vfsp->vfc_vfsops;
5fd012e0 316 vfsp->vfc_refcount++;
5fd012e0 317 mp->mnt_stat.f_type = vfsp->vfc_typenum;
aac0aabd 318 mp->mnt_flag |= MNT_RDONLY;
5fd012e0
MD
319 mp->mnt_flag |= vfsp->vfc_flags & MNT_VISFLAGMASK;
320 strncpy(mp->mnt_stat.f_fstypename, vfsp->vfc_name, MFSNAMELEN);
069b825e 321 copystr(devname, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, 0);
5fd012e0
MD
322 *mpp = mp;
323 return (0);
324}
325
326/*
aac0aabd
MD
327 * Basic mount structure initialization
328 */
329void
330mount_init(struct mount *mp)
331{
332 lockinit(&mp->mnt_lock, "vfslock", 0, 0);
a3c18566 333 lwkt_token_init(&mp->mnt_token, "permnt");
aac0aabd
MD
334
335 TAILQ_INIT(&mp->mnt_nvnodelist);
336 TAILQ_INIT(&mp->mnt_reservedvnlist);
337 TAILQ_INIT(&mp->mnt_jlist);
338 mp->mnt_nvnodelistsize = 0;
339 mp->mnt_flag = 0;
340 mp->mnt_iosize_max = DFLTPHYS;
341}
342
343/*
5fd012e0
MD
344 * Lookup a mount point by filesystem identifier.
345 */
346struct mount *
347vfs_getvfs(fsid_t *fsid)
348{
349 struct mount *mp;
5fd012e0 350
3b998fa9 351 lwkt_gettoken(&mountlist_token);
5fd012e0
MD
352 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
353 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
354 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
355 break;
02dede15 356 }
5fd012e0 357 }
3b998fa9 358 lwkt_reltoken(&mountlist_token);
5fd012e0
MD
359 return (mp);
360}
361
362/*
363 * Get a new unique fsid. Try to make its val[0] unique, since this value
364 * will be used to create fake device numbers for stat(). Also try (but
365 * not so hard) make its val[0] unique mod 2^16, since some emulators only
366 * support 16-bit device numbers. We end up with unique val[0]'s for the
367 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
368 *
369 * Keep in mind that several mounts may be running in parallel. Starting
370 * the search one past where the previous search terminated is both a
371 * micro-optimization and a defense against returning the same fsid to
372 * different mounts.
373 */
374void
375vfs_getnewfsid(struct mount *mp)
376{
377 static u_int16_t mntid_base;
5fd012e0
MD
378 fsid_t tfsid;
379 int mtype;
380
3b998fa9 381 lwkt_gettoken(&mntid_token);
5fd012e0
MD
382 mtype = mp->mnt_vfc->vfc_typenum;
383 tfsid.val[1] = mtype;
384 mtype = (mtype & 0xFF) << 24;
385 for (;;) {
386 tfsid.val[0] = makeudev(255,
387 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
388 mntid_base++;
389 if (vfs_getvfs(&tfsid) == NULL)
390 break;
391 }
392 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
393 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
3b998fa9 394 lwkt_reltoken(&mntid_token);
5fd012e0
MD
395}
396
397/*
67863d04
MD
398 * Set the FSID for a new mount point to the template. Adjust
399 * the FSID to avoid collisions.
400 */
401int
402vfs_setfsid(struct mount *mp, fsid_t *template)
403{
404 int didmunge = 0;
405
406 bzero(&mp->mnt_stat.f_fsid, sizeof(mp->mnt_stat.f_fsid));
407 for (;;) {
408 if (vfs_getvfs(template) == NULL)
409 break;
410 didmunge = 1;
411 ++template->val[1];
412 }
413 mp->mnt_stat.f_fsid = *template;
414 return(didmunge);
415}
416
417/*
5fd012e0
MD
418 * This routine is called when we have too many vnodes. It attempts
419 * to free <count> vnodes and will potentially free vnodes that still
420 * have VM backing store (VM backing store is typically the cause
421 * of a vnode blowout so we want to do this). Therefore, this operation
422 * is not considered cheap.
423 *
424 * A number of conditions may prevent a vnode from being reclaimed.
425 * the buffer cache may have references on the vnode, a directory
426 * vnode may still have references due to the namei cache representing
427 * underlying files, or the vnode may be in active use. It is not
428 * desireable to reuse such vnodes. These conditions may cause the
429 * number of vnodes to reach some minimum value regardless of what
430 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low.
431 */
432
433/*
58552887
MD
434 * This is a quick non-blocking check to determine if the vnode is a good
435 * candidate for being (eventually) vgone()'d. Returns 0 if the vnode is
436 * not a good candidate, 1 if it is.
5fd012e0
MD
437 */
438static __inline int
0e8bd897 439vmightfree(struct vnode *vp, int page_count, int pass)
5fd012e0 440{
58552887
MD
441 if (vp->v_flag & VRECLAIMED)
442 return (0);
44b1cf3d 443#if 0
58552887 444 if ((vp->v_flag & VFREE) && TAILQ_EMPTY(&vp->v_namecache))
5fd012e0 445 return (0);
44b1cf3d 446#endif
3c37c940 447 if (sysref_isactive(&vp->v_sysref))
5fd012e0
MD
448 return (0);
449 if (vp->v_object && vp->v_object->resident_page_count >= page_count)
450 return (0);
0e8bd897
MD
451
452 /*
453 * XXX horrible hack. Up to four passes will be taken. Each pass
454 * makes a larger set of vnodes eligible. For now what this really
455 * means is that we try to recycle files opened only once before
456 * recycling files opened multiple times.
457 */
458 switch(vp->v_flag & (VAGE0 | VAGE1)) {
459 case 0:
460 if (pass < 3)
461 return(0);
462 break;
463 case VAGE0:
464 if (pass < 2)
465 return(0);
466 break;
467 case VAGE1:
468 if (pass < 1)
469 return(0);
470 break;
471 case VAGE0 | VAGE1:
472 break;
473 }
5fd012e0
MD
474 return (1);
475}
476
250d127c 477/*
58552887 478 * The vnode was found to be possibly vgone()able and the caller has locked it
250d127c 479 * (thus the usecount should be 1 now). Determine if the vnode is actually
58552887
MD
480 * vgone()able, doing some cleanups in the process. Returns 1 if the vnode
481 * can be vgone()'d, 0 otherwise.
250d127c 482 *
3c37c940 483 * Note that v_auxrefs may be non-zero because (A) this vnode is not a leaf
250d127c
MD
484 * in the namecache topology and (B) this vnode has buffer cache bufs.
485 * We cannot remove vnodes with non-leaf namecache associations. We do a
486 * tentitive leaf check prior to attempting to flush out any buffers but the
3c37c940 487 * 'real' test when all is said in done is that v_auxrefs must become 0 for
250d127c
MD
488 * the vnode to be freeable.
489 *
3c37c940 490 * We could theoretically just unconditionally flush when v_auxrefs != 0,
250d127c
MD
491 * but flushing data associated with non-leaf nodes (which are always
492 * directories), just throws it away for no benefit. It is the buffer
493 * cache's responsibility to choose buffers to recycle from the cached
494 * data point of view.
495 */
496static int
497visleaf(struct vnode *vp)
498{
499 struct namecache *ncp;
500
b12defdc 501 spin_lock(&vp->v_spin);
250d127c 502 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) {
f63911bf 503 if (!TAILQ_EMPTY(&ncp->nc_list)) {
b12defdc 504 spin_unlock(&vp->v_spin);
250d127c 505 return(0);
f63911bf 506 }
250d127c 507 }
b12defdc 508 spin_unlock(&vp->v_spin);
250d127c
MD
509 return(1);
510}
511
58552887
MD
512/*
513 * Try to clean up the vnode to the point where it can be vgone()'d, returning
514 * 0 if it cannot be vgone()'d (or already has been), 1 if it can. Unlike
515 * vmightfree() this routine may flush the vnode and block. Vnodes marked
516 * VFREE are still candidates for vgone()ing because they may hold namecache
517 * resources and could be blocking the namecache directory hierarchy (and
518 * related vnodes) from being freed.
519 */
250d127c 520static int
58552887 521vtrytomakegoneable(struct vnode *vp, int page_count)
250d127c 522{
58552887 523 if (vp->v_flag & VRECLAIMED)
250d127c 524 return (0);
3c37c940 525 if (vp->v_sysref.refcnt > 1)
250d127c
MD
526 return (0);
527 if (vp->v_object && vp->v_object->resident_page_count >= page_count)
528 return (0);
3c37c940 529 if (vp->v_auxrefs && visleaf(vp)) {
87de5057 530 vinvalbuf(vp, V_SAVE, 0, 0);
250d127c 531#if 0 /* DEBUG */
3c37c940 532 kprintf((vp->v_auxrefs ? "vrecycle: vp %p failed: %s\n" :
250d127c
MD
533 "vrecycle: vp %p succeeded: %s\n"), vp,
534 (TAILQ_FIRST(&vp->v_namecache) ?
535 TAILQ_FIRST(&vp->v_namecache)->nc_name : "?"));
536#endif
537 }
5c6c3cac
MD
538
539 /*
540 * This sequence may seem a little strange, but we need to optimize
541 * the critical path a bit. We can't recycle vnodes with other
542 * references and because we are trying to recycle an otherwise
543 * perfectly fine vnode we have to invalidate the namecache in a
544 * way that avoids possible deadlocks (since the vnode lock is being
545 * held here). Finally, we have to check for other references one
546 * last time in case something snuck in during the inval.
547 */
548 if (vp->v_sysref.refcnt > 1 || vp->v_auxrefs != 0)
549 return (0);
550 if (cache_inval_vp_nonblock(vp))
551 return (0);
552 return (vp->v_sysref.refcnt <= 1 && vp->v_auxrefs == 0);
250d127c 553}
5fd012e0 554
58552887
MD
555/*
556 * Reclaim up to 1/10 of the vnodes associated with a mount point. Try
557 * to avoid vnodes which have lots of resident pages (we are trying to free
861905fb
MD
558 * vnodes, not memory).
559 *
560 * This routine is a callback from the mountlist scan. The mount point
561 * in question will be busied.
0e8bd897
MD
562 *
563 * NOTE: The 1/10 reclamation also ensures that the inactive data set
564 * (the vnodes being recycled by the one-time use) does not degenerate
565 * into too-small a set. This is important because once a vnode is
566 * marked as not being one-time-use (VAGE0/VAGE1 both 0) that vnode
567 * will not be destroyed EXCEPT by this mechanism. VM pages can still
568 * be cleaned/freed by the pageout daemon.
58552887 569 */
5fd012e0 570static int
861905fb 571vlrureclaim(struct mount *mp, void *data)
5fd012e0 572{
0e8bd897 573 struct vnlru_info *info = data;
5fd012e0 574 struct vnode *vp;
5fd012e0
MD
575 int done;
576 int trigger;
577 int usevnodes;
578 int count;
861905fb 579 int trigger_mult = vnlru_nowhere;
5fd012e0
MD
580
581 /*
58552887
MD
582 * Calculate the trigger point for the resident pages check. The
583 * minimum trigger value is approximately the number of pages in
584 * the system divded by the number of vnodes. However, due to
585 * various other system memory overheads unrelated to data caching
586 * it is a good idea to double the trigger (at least).
587 *
588 * trigger_mult starts at 0. If the recycler is having problems
589 * finding enough freeable vnodes it will increase trigger_mult.
590 * This should not happen in normal operation, even on machines with
591 * low amounts of memory, but extraordinary memory use by the system
592 * verses the amount of cached data can trigger it.
5fd012e0
MD
593 */
594 usevnodes = desiredvnodes;
595 if (usevnodes <= 0)
596 usevnodes = 1;
58552887 597 trigger = vmstats.v_page_count * (trigger_mult + 2) / usevnodes;
5fd012e0
MD
598
599 done = 0;
3b998fa9 600 lwkt_gettoken(&mntvnode_token);
5fd012e0 601 count = mp->mnt_nvnodelistsize / 10 + 1;
0e8bd897 602
2ec4b00d
MD
603 while (count && mp->mnt_syncer) {
604 /*
605 * Next vnode. Use the special syncer vnode to placemark
606 * the LRU. This way the LRU code does not interfere with
607 * vmntvnodescan().
608 */
609 vp = TAILQ_NEXT(mp->mnt_syncer, v_nmntvnodes);
610 TAILQ_REMOVE(&mp->mnt_nvnodelist, mp->mnt_syncer, v_nmntvnodes);
611 if (vp) {
612 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp,
613 mp->mnt_syncer, v_nmntvnodes);
614 } else {
615 TAILQ_INSERT_HEAD(&mp->mnt_nvnodelist, mp->mnt_syncer,
616 v_nmntvnodes);
617 vp = TAILQ_NEXT(mp->mnt_syncer, v_nmntvnodes);
618 if (vp == NULL)
619 break;
620 }
621
5fd012e0
MD
622 /*
623 * __VNODESCAN__
624 *
625 * The VP will stick around while we hold mntvnode_token,
626 * at least until we block, so we can safely do an initial
627 * check, and then must check again after we lock the vnode.
628 */
986e7cda 629 if (vp->v_type == VNON || /* syncer or indeterminant */
0e8bd897 630 !vmightfree(vp, trigger, info->pass) /* critical path opt */
5fd012e0 631 ) {
5fd012e0
MD
632 --count;
633 continue;
634 }
635
636 /*
637 * VX get the candidate vnode. If the VX get fails the
638 * vnode might still be on the mountlist. Our loop depends
639 * on us at least cycling the vnode to the end of the
640 * mountlist.
641 */
642 if (vx_get_nonblock(vp) != 0) {
5fd012e0
MD
643 --count;
644 continue;
645 }
646
647 /*
648 * Since we blocked locking the vp, make sure it is still
649 * a candidate for reclamation. That is, it has not already
650 * been reclaimed and only has our VX reference associated
651 * with it.
652 */
986e7cda 653 if (vp->v_type == VNON || /* syncer or indeterminant */
5fd012e0
MD
654 (vp->v_flag & VRECLAIMED) ||
655 vp->v_mount != mp ||
58552887 656 !vtrytomakegoneable(vp, trigger) /* critical path opt */
5fd012e0 657 ) {
5fd012e0
MD
658 --count;
659 vx_put(vp);
660 continue;
661 }
662
663 /*
664 * All right, we are good, move the vp to the end of the
665 * mountlist and clean it out. The vget will have returned
666 * an error if the vnode was destroyed (VRECLAIMED set), so we
667 * do not have to check again. The vput() will move the
668 * vnode to the free list if the vgone() was successful.
669 */
670 KKASSERT(vp->v_mount == mp);
3c37c940 671 vgone_vxlocked(vp);
5fd012e0
MD
672 vx_put(vp);
673 ++done;
674 --count;
675 }
3b998fa9 676 lwkt_reltoken(&mntvnode_token);
5fd012e0
MD
677 return (done);
678}
679
680/*
681 * Attempt to recycle vnodes in a context that is always safe to block.
682 * Calling vlrurecycle() from the bowels of file system code has some
683 * interesting deadlock problems.
684 */
685static struct thread *vnlruthread;
686static int vnlruproc_sig;
687
688void
689vnlru_proc_wait(void)
690{
1dbcfdac 691 tsleep_interlock(&vnlruproc_sig, 0);
5fd012e0
MD
692 if (vnlruproc_sig == 0) {
693 vnlruproc_sig = 1; /* avoid unnecessary wakeups */
694 wakeup(vnlruthread);
695 }
1dbcfdac 696 tsleep(&vnlruproc_sig, PINTERLOCKED, "vlruwk", hz);
5fd012e0
MD
697}
698
699static void
700vnlru_proc(void)
701{
5fd012e0 702 struct thread *td = curthread;
0e8bd897 703 struct vnlru_info info;
861905fb 704 int done;
5fd012e0
MD
705
706 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, td,
cd8ab232 707 SHUTDOWN_PRI_FIRST);
5fd012e0 708
cd8ab232 709 get_mplock();
e43a034f 710 crit_enter();
cd8ab232 711
5fd012e0
MD
712 for (;;) {
713 kproc_suspend_loop();
3c37c940
MD
714
715 /*
716 * Try to free some vnodes if we have too many
717 */
718 if (numvnodes > desiredvnodes &&
719 freevnodes > desiredvnodes * 2 / 10) {
720 int count = numvnodes - desiredvnodes;
721
722 if (count > freevnodes / 100)
723 count = freevnodes / 100;
724 if (count < 5)
725 count = 5;
726 freesomevnodes(count);
727 }
728
729 /*
730 * Nothing to do if most of our vnodes are already on
731 * the free list.
732 */
5fd012e0
MD
733 if (numvnodes - freevnodes <= desiredvnodes * 9 / 10) {
734 vnlruproc_sig = 0;
735 wakeup(&vnlruproc_sig);
1dbcfdac 736 tsleep(vnlruthread, 0, "vlruwt", hz);
5fd012e0
MD
737 continue;
738 }
65870584 739 cache_hysteresis();
0e8bd897
MD
740
741 /*
742 * The pass iterates through the four combinations of
743 * VAGE0/VAGE1. We want to get rid of aged small files
744 * first.
745 */
746 info.pass = 0;
747 done = 0;
748 while (done == 0 && info.pass < 4) {
749 done = mountlist_scan(vlrureclaim, &info,
750 MNTSCAN_FORWARD);
751 ++info.pass;
752 }
58552887
MD
753
754 /*
755 * The vlrureclaim() call only processes 1/10 of the vnodes
756 * on each mount. If we couldn't find any repeat the loop
757 * at least enough times to cover all available vnodes before
758 * we start sleeping. Complain if the failure extends past
759 * 30 second, every 30 seconds.
760 */
5fd012e0
MD
761 if (done == 0) {
762 ++vnlru_nowhere;
5fd012e0 763 if (vnlru_nowhere % 10 == 0)
1dbcfdac 764 tsleep(vnlruthread, 0, "vlrup", hz * 3);
58552887 765 if (vnlru_nowhere % 100 == 0)
6ea70f76 766 kprintf("vnlru_proc: vnode recycler stopped working!\n");
58552887
MD
767 if (vnlru_nowhere == 1000)
768 vnlru_nowhere = 900;
5fd012e0
MD
769 } else {
770 vnlru_nowhere = 0;
771 }
772 }
cd8ab232 773
e43a034f 774 crit_exit();
cd8ab232 775 rel_mplock();
5fd012e0
MD
776}
777
861905fb
MD
778/*
779 * MOUNTLIST FUNCTIONS
780 */
781
782/*
783 * mountlist_insert (MP SAFE)
784 *
785 * Add a new mount point to the mount list.
786 */
787void
788mountlist_insert(struct mount *mp, int how)
789{
3b998fa9 790 lwkt_gettoken(&mountlist_token);
861905fb
MD
791 if (how == MNTINS_FIRST)
792 TAILQ_INSERT_HEAD(&mountlist, mp, mnt_list);
793 else
794 TAILQ_INSERT_TAIL(&mountlist, mp, mnt_list);
3b998fa9 795 lwkt_reltoken(&mountlist_token);
861905fb
MD
796}
797
798/*
799 * mountlist_interlock (MP SAFE)
800 *
801 * Execute the specified interlock function with the mountlist token
802 * held. The function will be called in a serialized fashion verses
803 * other functions called through this mechanism.
804 */
805int
806mountlist_interlock(int (*callback)(struct mount *), struct mount *mp)
807{
861905fb
MD
808 int error;
809
3b998fa9 810 lwkt_gettoken(&mountlist_token);
861905fb 811 error = callback(mp);
3b998fa9 812 lwkt_reltoken(&mountlist_token);
861905fb
MD
813 return (error);
814}
815
816/*
817 * mountlist_boot_getfirst (DURING BOOT ONLY)
818 *
819 * This function returns the first mount on the mountlist, which is
820 * expected to be the root mount. Since no interlocks are obtained
821 * this function is only safe to use during booting.
822 */
823
824struct mount *
825mountlist_boot_getfirst(void)
826{
827 return(TAILQ_FIRST(&mountlist));
828}
829
830/*
831 * mountlist_remove (MP SAFE)
832 *
833 * Remove a node from the mountlist. If this node is the next scan node
834 * for any active mountlist scans, the active mountlist scan will be
835 * adjusted to skip the node, thus allowing removals during mountlist
836 * scans.
837 */
838void
839mountlist_remove(struct mount *mp)
840{
841 struct mountscan_info *msi;
861905fb 842
3b998fa9 843 lwkt_gettoken(&mountlist_token);
861905fb
MD
844 TAILQ_FOREACH(msi, &mountscan_list, msi_entry) {
845 if (msi->msi_node == mp) {
846 if (msi->msi_how & MNTSCAN_FORWARD)
847 msi->msi_node = TAILQ_NEXT(mp, mnt_list);
848 else
849 msi->msi_node = TAILQ_PREV(mp, mntlist, mnt_list);
850 }
851 }
852 TAILQ_REMOVE(&mountlist, mp, mnt_list);
3b998fa9 853 lwkt_reltoken(&mountlist_token);
861905fb
MD
854}
855
856/*
e6a2de88
FT
857 * mountlist_exists (MP SAFE)
858 *
859 * Checks if a node exists in the mountlist.
860 * This function is mainly used by VFS accounting code to check if a
861 * cached nullfs struct mount pointer is still valid at use time
862 */
863int
864mountlist_exists(struct mount *mp)
865{
866 int node_exists = 0;
867 struct mountscan_info *msi;
868
869 lwkt_gettoken(&mountlist_token);
870 TAILQ_FOREACH(msi, &mountscan_list, msi_entry) {
871 if (msi->msi_node == mp) {
872 node_exists = 1;
873 }
874 }
875 lwkt_reltoken(&mountlist_token);
876 return(node_exists);
877}
878
879/*
861905fb
MD
880 * mountlist_scan (MP SAFE)
881 *
882 * Safely scan the mount points on the mount list. Unless otherwise
883 * specified each mount point will be busied prior to the callback and
884 * unbusied afterwords. The callback may safely remove any mount point
885 * without interfering with the scan. If the current callback
886 * mount is removed the scanner will not attempt to unbusy it.
887 *
888 * If a mount node cannot be busied it is silently skipped.
889 *
890 * The callback return value is aggregated and a total is returned. A return
891 * value of < 0 is not aggregated and will terminate the scan.
892 *
893 * MNTSCAN_FORWARD - the mountlist is scanned in the forward direction
894 * MNTSCAN_REVERSE - the mountlist is scanned in reverse
895 * MNTSCAN_NOBUSY - the scanner will make the callback without busying
896 * the mount node.
897 */
898int
899mountlist_scan(int (*callback)(struct mount *, void *), void *data, int how)
900{
901 struct mountscan_info info;
861905fb 902 struct mount *mp;
861905fb
MD
903 int count;
904 int res;
905
3b998fa9 906 lwkt_gettoken(&mountlist_token);
861905fb
MD
907
908 info.msi_how = how;
909 info.msi_node = NULL; /* paranoia */
910 TAILQ_INSERT_TAIL(&mountscan_list, &info, msi_entry);
911
912 res = 0;
861905fb
MD
913
914 if (how & MNTSCAN_FORWARD) {
915 info.msi_node = TAILQ_FIRST(&mountlist);
916 while ((mp = info.msi_node) != NULL) {
917 if (how & MNTSCAN_NOBUSY) {
918 count = callback(mp, data);
f9642f56 919 } else if (vfs_busy(mp, LK_NOWAIT) == 0) {
861905fb
MD
920 count = callback(mp, data);
921 if (mp == info.msi_node)
f9642f56 922 vfs_unbusy(mp);
861905fb
MD
923 } else {
924 count = 0;
925 }
926 if (count < 0)
927 break;
928 res += count;
929 if (mp == info.msi_node)
930 info.msi_node = TAILQ_NEXT(mp, mnt_list);
931 }
932 } else if (how & MNTSCAN_REVERSE) {
933 info.msi_node = TAILQ_LAST(&mountlist, mntlist);
934 while ((mp = info.msi_node) != NULL) {
935 if (how & MNTSCAN_NOBUSY) {
936 count = callback(mp, data);
f9642f56 937 } else if (vfs_busy(mp, LK_NOWAIT) == 0) {
861905fb
MD
938 count = callback(mp, data);
939 if (mp == info.msi_node)
f9642f56 940 vfs_unbusy(mp);
861905fb
MD
941 } else {
942 count = 0;
943 }
944 if (count < 0)
945 break;
946 res += count;
947 if (mp == info.msi_node)
948 info.msi_node = TAILQ_PREV(mp, mntlist, mnt_list);
949 }
950 }
951 TAILQ_REMOVE(&mountscan_list, &info, msi_entry);
3b998fa9 952 lwkt_reltoken(&mountlist_token);
861905fb
MD
953 return(res);
954}
955
956/*
957 * MOUNT RELATED VNODE FUNCTIONS
958 */
959
5fd012e0
MD
960static struct kproc_desc vnlru_kp = {
961 "vnlru",
962 vnlru_proc,
963 &vnlruthread
964};
965SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp)
966
967/*
968 * Move a vnode from one mount queue to another.
2247fe02
MD
969 *
970 * MPSAFE
5fd012e0
MD
971 */
972void
973insmntque(struct vnode *vp, struct mount *mp)
974{
3b998fa9 975 lwkt_gettoken(&mntvnode_token);
5fd012e0
MD
976 /*
977 * Delete from old mount point vnode list, if on one.
978 */
979 if (vp->v_mount != NULL) {
980 KASSERT(vp->v_mount->mnt_nvnodelistsize > 0,
981 ("bad mount point vnode list size"));
be6c08cb 982 vremovevnodemnt(vp);
5fd012e0
MD
983 vp->v_mount->mnt_nvnodelistsize--;
984 }
985 /*
986 * Insert into list of vnodes for the new mount point, if available.
2ec4b00d 987 * The 'end' of the LRU list is the vnode prior to mp->mnt_syncer.
5fd012e0
MD
988 */
989 if ((vp->v_mount = mp) == NULL) {
3b998fa9 990 lwkt_reltoken(&mntvnode_token);
5fd012e0
MD
991 return;
992 }
2ec4b00d
MD
993 if (mp->mnt_syncer) {
994 TAILQ_INSERT_BEFORE(mp->mnt_syncer, vp, v_nmntvnodes);
995 } else {
996 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
997 }
5fd012e0 998 mp->mnt_nvnodelistsize++;
3b998fa9 999 lwkt_reltoken(&mntvnode_token);
5fd012e0
MD
1000}
1001
1002
1003/*
be6c08cb
MD
1004 * Scan the vnodes under a mount point and issue appropriate callbacks.
1005 *
1006 * The fastfunc() callback is called with just the mountlist token held
1007 * (no vnode lock). It may not block and the vnode may be undergoing
1008 * modifications while the caller is processing it. The vnode will
1009 * not be entirely destroyed, however, due to the fact that the mountlist
1010 * token is held. A return value < 0 skips to the next vnode without calling
1011 * the slowfunc(), a return value > 0 terminates the loop.
1012 *
1013 * The slowfunc() callback is called after the vnode has been successfully
1014 * locked based on passed flags. The vnode is skipped if it gets rearranged
1015 * or destroyed while blocking on the lock. A non-zero return value from
1016 * the slow function terminates the loop. The slow function is allowed to
1017 * arbitrarily block. The scanning code guarentees consistency of operation
1018 * even if the slow function deletes or moves the node, or blocks and some
1019 * other thread deletes or moves the node.
aed76ef1
MD
1020 *
1021 * NOTE: We hold vmobj_token to prevent a VM object from being destroyed
1022 * out from under the fastfunc()'s vnode test. It will not prevent
1023 * v_object from getting NULL'd out but it will ensure that the
1024 * pointer (if we race) will remain stable.
5fd012e0
MD
1025 */
1026int
1027vmntvnodescan(
1028 struct mount *mp,
1029 int flags,
1030 int (*fastfunc)(struct mount *mp, struct vnode *vp, void *data),
1031 int (*slowfunc)(struct mount *mp, struct vnode *vp, void *data),
1032 void *data
1033) {
be6c08cb 1034 struct vmntvnodescan_info info;
5fd012e0
MD
1035 struct vnode *vp;
1036 int r = 0;
f0eb7cee 1037 int maxcount = mp->mnt_nvnodelistsize * 2;
19b97e01 1038 int stopcount = 0;
9fe8385f 1039 int count = 0;
5fd012e0 1040
3b998fa9 1041 lwkt_gettoken(&mntvnode_token);
aed76ef1 1042 lwkt_gettoken(&vmobj_token);
5fd012e0 1043
19b97e01
MD
1044 /*
1045 * If asked to do one pass stop after iterating available vnodes.
1046 * Under heavy loads new vnodes can be added while we are scanning,
1047 * so this isn't perfect. Create a slop factor of 2x.
1048 */
1049 if (flags & VMSC_ONEPASS)
f0eb7cee 1050 stopcount = mp->mnt_nvnodelistsize;
19b97e01 1051
be6c08cb
MD
1052 info.vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
1053 TAILQ_INSERT_TAIL(&mntvnodescan_list, &info, entry);
1054 while ((vp = info.vp) != NULL) {
f0eb7cee
MD
1055 if (--maxcount == 0) {
1056 kprintf("Warning: excessive fssync iteration\n");
1057 maxcount = mp->mnt_nvnodelistsize * 2;
1058 }
be6c08cb 1059
2ec4b00d
MD
1060 /*
1061 * Skip if visible but not ready, or special (e.g.
1062 * mp->mnt_syncer)
1063 */
1064 if (vp->v_type == VNON)
be6c08cb 1065 goto next;
5fd012e0
MD
1066 KKASSERT(vp->v_mount == mp);
1067
1068 /*
1069 * Quick test. A negative return continues the loop without
1070 * calling the slow test. 0 continues onto the slow test.
1071 * A positive number aborts the loop.
1072 */
1073 if (fastfunc) {
c66c09cf
MD
1074 if ((r = fastfunc(mp, vp, data)) < 0) {
1075 r = 0;
be6c08cb 1076 goto next;
c66c09cf 1077 }
5fd012e0
MD
1078 if (r)
1079 break;
1080 }
1081
1082 /*
1083 * Get a vxlock on the vnode, retry if it has moved or isn't
1084 * in the mountlist where we expect it.
1085 */
1086 if (slowfunc) {
1087 int error;
1088
19b97e01 1089 switch(flags & (VMSC_GETVP|VMSC_GETVX|VMSC_NOWAIT)) {
5fd012e0 1090 case VMSC_GETVP:
87de5057 1091 error = vget(vp, LK_EXCLUSIVE);
5fd012e0
MD
1092 break;
1093 case VMSC_GETVP|VMSC_NOWAIT:
87de5057 1094 error = vget(vp, LK_EXCLUSIVE|LK_NOWAIT);
5fd012e0
MD
1095 break;
1096 case VMSC_GETVX:
e3332475
MD
1097 vx_get(vp);
1098 error = 0;
5fd012e0 1099 break;
5fd012e0
MD
1100 default:
1101 error = 0;
1102 break;
1103 }
1104 if (error)
be6c08cb
MD
1105 goto next;
1106 /*
1107 * Do not call the slow function if the vnode is
1108 * invalid or if it was ripped out from under us
1109 * while we (potentially) blocked.
1110 */
1111 if (info.vp == vp && vp->v_type != VNON)
1112 r = slowfunc(mp, vp, data);
1113
1114 /*
1115 * Cleanup
1116 */
19b97e01 1117 switch(flags & (VMSC_GETVP|VMSC_GETVX|VMSC_NOWAIT)) {
5fd012e0
MD
1118 case VMSC_GETVP:
1119 case VMSC_GETVP|VMSC_NOWAIT:
1120 vput(vp);
1121 break;
1122 case VMSC_GETVX:
1123 vx_put(vp);
1124 break;
5fd012e0
MD
1125 default:
1126 break;
1127 }
1128 if (r != 0)
1129 break;
1130 }
be6c08cb 1131
9fe8385f
SS
1132next:
1133 /*
1134 * Yield after some processing. Depending on the number
1135 * of vnodes, we might wind up running for a long time.
1136 * Because threads are not preemptable, time critical
1137 * userland processes might starve. Give them a chance
1138 * now and then.
1139 */
1140 if (++count == 10000) {
5a96e837
SS
1141 /* We really want to yield a bit, so we simply sleep a tick */
1142 tsleep(mp, 0, "vnodescn", 1);
9fe8385f
SS
1143 count = 0;
1144 }
1145
be6c08cb 1146 /*
19b97e01
MD
1147 * If doing one pass this decrements to zero. If it starts
1148 * at zero it is effectively unlimited for the purposes of
1149 * this loop.
1150 */
1151 if (--stopcount == 0)
1152 break;
1153
1154 /*
be6c08cb
MD
1155 * Iterate. If the vnode was ripped out from under us
1156 * info.vp will already point to the next vnode, otherwise
1157 * we have to obtain the next valid vnode ourselves.
1158 */
be6c08cb
MD
1159 if (info.vp == vp)
1160 info.vp = TAILQ_NEXT(vp, v_nmntvnodes);
5fd012e0 1161 }
be6c08cb 1162 TAILQ_REMOVE(&mntvnodescan_list, &info, entry);
aed76ef1 1163 lwkt_reltoken(&vmobj_token);
3b998fa9 1164 lwkt_reltoken(&mntvnode_token);
5fd012e0
MD
1165 return(r);
1166}
1167
1168/*
1169 * Remove any vnodes in the vnode table belonging to mount point mp.
1170 *
1171 * If FORCECLOSE is not specified, there should not be any active ones,
1172 * return error if any are found (nb: this is a user error, not a
1173 * system error). If FORCECLOSE is specified, detach any active vnodes
1174 * that are found.
1175 *
1176 * If WRITECLOSE is set, only flush out regular file vnodes open for
1177 * writing.
1178 *
1179 * SKIPSYSTEM causes any vnodes marked VSYSTEM to be skipped.
1180 *
1181 * `rootrefs' specifies the base reference count for the root vnode
1182 * of this filesystem. The root vnode is considered busy if its
3c37c940 1183 * v_sysref.refcnt exceeds this value. On a successful return, vflush()
5fd012e0
MD
1184 * will call vrele() on the root vnode exactly rootrefs times.
1185 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
1186 * be zero.
1187 */
1188#ifdef DIAGNOSTIC
1189static int busyprt = 0; /* print out busy vnodes */
1190SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "");
1191#endif
1192
1193static int vflush_scan(struct mount *mp, struct vnode *vp, void *data);
1194
1195struct vflush_info {
1196 int flags;
1197 int busy;
1198 thread_t td;
1199};
1200
1201int
1202vflush(struct mount *mp, int rootrefs, int flags)
1203{
1204 struct thread *td = curthread; /* XXX */
1205 struct vnode *rootvp = NULL;
1206 int error;
1207 struct vflush_info vflush_info;
1208
1209 if (rootrefs > 0) {
1210 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
1211 ("vflush: bad args"));
1212 /*
1213 * Get the filesystem root vnode. We can vput() it
1214 * immediately, since with rootrefs > 0, it won't go away.
1215 */
d9adbeaf
MD
1216 if ((error = VFS_ROOT(mp, &rootvp)) != 0) {
1217 if ((flags & FORCECLOSE) == 0)
1218 return (error);
1219 rootrefs = 0;
1220 /* continue anyway */
1221 }
1222 if (rootrefs)
1223 vput(rootvp);
5fd012e0
MD
1224 }
1225
1226 vflush_info.busy = 0;
1227 vflush_info.flags = flags;
1228 vflush_info.td = td;
1229 vmntvnodescan(mp, VMSC_GETVX, NULL, vflush_scan, &vflush_info);
1230
1231 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
1232 /*
1233 * If just the root vnode is busy, and if its refcount
1234 * is equal to `rootrefs', then go ahead and kill it.
1235 */
1236 KASSERT(vflush_info.busy > 0, ("vflush: not busy"));
3c37c940
MD
1237 KASSERT(rootvp->v_sysref.refcnt >= rootrefs, ("vflush: rootrefs"));
1238 if (vflush_info.busy == 1 && rootvp->v_sysref.refcnt == rootrefs) {
e3332475 1239 vx_lock(rootvp);
3c37c940 1240 vgone_vxlocked(rootvp);
e3332475
MD
1241 vx_unlock(rootvp);
1242 vflush_info.busy = 0;
5fd012e0
MD
1243 }
1244 }
1245 if (vflush_info.busy)
1246 return (EBUSY);
1247 for (; rootrefs > 0; rootrefs--)
1248 vrele(rootvp);
1249 return (0);
1250}
1251
1252/*
1253 * The scan callback is made with an VX locked vnode.
1254 */
1255static int
1256vflush_scan(struct mount *mp, struct vnode *vp, void *data)
1257{
1258 struct vflush_info *info = data;
1259 struct vattr vattr;
1260
1261 /*
1262 * Skip over a vnodes marked VSYSTEM.
1263 */
1264 if ((info->flags & SKIPSYSTEM) && (vp->v_flag & VSYSTEM)) {
1265 return(0);
1266 }
1267
1268 /*
1269 * If WRITECLOSE is set, flush out unlinked but still open
1270 * files (even if open only for reading) and regular file
1271 * vnodes open for writing.
1272 */
1273 if ((info->flags & WRITECLOSE) &&
1274 (vp->v_type == VNON ||
87de5057 1275 (VOP_GETATTR(vp, &vattr) == 0 &&
5fd012e0
MD
1276 vattr.va_nlink > 0)) &&
1277 (vp->v_writecount == 0 || vp->v_type != VREG)) {
1278 return(0);
1279 }
1280
1281 /*
3c37c940
MD
1282 * If we are the only holder (refcnt of 1) or the vnode is in
1283 * termination (refcnt < 0), we can vgone the vnode.
5fd012e0 1284 */
3c37c940
MD
1285 if (vp->v_sysref.refcnt <= 1) {
1286 vgone_vxlocked(vp);
5fd012e0
MD
1287 return(0);
1288 }
1289
1290 /*
1e49b98c
MD
1291 * If FORCECLOSE is set, forcibly destroy the vnode and then move
1292 * it to a dummymount structure so vop_*() functions don't deref
1293 * a NULL pointer.
5fd012e0
MD
1294 */
1295 if (info->flags & FORCECLOSE) {
1e49b98c 1296 vhold(vp);
aec8eea4 1297 vgone_vxlocked(vp);
1e49b98c
MD
1298 if (vp->v_mount == NULL)
1299 insmntque(vp, &dummymount);
1300 vdrop(vp);
5fd012e0
MD
1301 return(0);
1302 }
1303#ifdef DIAGNOSTIC
1304 if (busyprt)
1305 vprint("vflush: busy vnode", vp);
1306#endif
1307 ++info->busy;
1308 return(0);
1309}
1310
408357d8
MD
1311void
1312add_bio_ops(struct bio_ops *ops)
1313{
1314 TAILQ_INSERT_TAIL(&bio_ops_list, ops, entry);
1315}
1316
1317void
1318rem_bio_ops(struct bio_ops *ops)
1319{
1320 TAILQ_REMOVE(&bio_ops_list, ops, entry);
1321}
1322
1323/*
1324 * This calls the bio_ops io_sync function either for a mount point
1325 * or generally.
1326 *
1327 * WARNING: softdeps is weirdly coded and just isn't happy unless
1328 * io_sync is called with a NULL mount from the general syncing code.
1329 */
1330void
1331bio_ops_sync(struct mount *mp)
1332{
1333 struct bio_ops *ops;
1334
1335 if (mp) {
1336 if ((ops = mp->mnt_bioops) != NULL)
1337 ops->io_sync(mp);
1338 } else {
1339 TAILQ_FOREACH(ops, &bio_ops_list, entry) {
1340 ops->io_sync(NULL);
1341 }
1342 }
1343}
1344
5b4cfb7e
AH
1345/*
1346 * Lookup a mount point by nch
1347 */
1348struct mount *
1349mount_get_by_nc(struct namecache *ncp)
1350{
1351 struct mount *mp = NULL;
5b4cfb7e 1352
3b998fa9 1353 lwkt_gettoken(&mountlist_token);
5b4cfb7e
AH
1354 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
1355 if (ncp == mp->mnt_ncmountpt.ncp)
1356 break;
1357 }
3b998fa9 1358 lwkt_reltoken(&mountlist_token);
5b4cfb7e
AH
1359 return (mp);
1360}
1361