vfs_cache.c revision 1.54 1 /* $NetBSD: vfs_cache.c,v 1.54 2003/09/01 12:13:55 yamt Exp $ */
2
3 /*
4 * Copyright (c) 1989, 1993
5 * The Regents of the University of California. All rights reserved.
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 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 * 3. Neither the name of the University nor the names of its contributors
16 * may be used to endorse or promote products derived from this software
17 * without specific prior written permission.
18 *
19 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 * SUCH DAMAGE.
30 *
31 * @(#)vfs_cache.c 8.3 (Berkeley) 8/22/94
32 */
33
34 #include <sys/cdefs.h>
35 __KERNEL_RCSID(0, "$NetBSD: vfs_cache.c,v 1.54 2003/09/01 12:13:55 yamt Exp $");
36
37 #include "opt_ddb.h"
38 #include "opt_revcache.h"
39
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/time.h>
43 #include <sys/mount.h>
44 #include <sys/vnode.h>
45 #include <sys/namei.h>
46 #include <sys/errno.h>
47 #include <sys/malloc.h>
48 #include <sys/pool.h>
49 #include <sys/lock.h>
50
51 /*
52 * Name caching works as follows:
53 *
54 * Names found by directory scans are retained in a cache
55 * for future reference. It is managed LRU, so frequently
56 * used names will hang around. Cache is indexed by hash value
57 * obtained from (dvp, name) where dvp refers to the directory
58 * containing name.
59 *
60 * For simplicity (and economy of storage), names longer than
61 * a maximum length of NCHNAMLEN are not cached; they occur
62 * infrequently in any case, and are almost never of interest.
63 *
64 * Upon reaching the last segment of a path, if the reference
65 * is for DELETE, or NOCACHE is set (rewrite), and the
66 * name is located in the cache, it will be dropped.
67 * The entry is dropped also when it was not possible to lock
68 * the cached vnode, either because vget() failed or the generation
69 * number has changed while waiting for the lock.
70 */
71
72 /*
73 * Structures associated with name cacheing.
74 */
75 LIST_HEAD(nchashhead, namecache) *nchashtbl;
76 u_long nchash; /* size of hash table - 1 */
77 long numcache; /* number of cache entries allocated */
78 #define NCHASH(cnp, dvp) \
79 (((cnp)->cn_hash ^ ((uintptr_t)(dvp) >> 3)) & nchash)
80
81 LIST_HEAD(ncvhashhead, namecache) *ncvhashtbl;
82 u_long ncvhash; /* size of hash table - 1 */
83 #define NCVHASH(vp) (((uintptr_t)(vp) >> 3) & ncvhash)
84
85 TAILQ_HEAD(, namecache) nclruhead; /* LRU chain */
86 struct nchstats nchstats; /* cache effectiveness statistics */
87
88 struct pool namecache_pool;
89
90 MALLOC_DEFINE(M_CACHE, "namecache", "Dynamically allocated cache entries");
91
92 int doingcache = 1; /* 1 => enable the cache */
93
94 /* A single lock to protect cache insertion, removal and lookup */
95 static struct simplelock namecache_slock = SIMPLELOCK_INITIALIZER;
96
97 static void cache_remove(struct namecache *);
98 static void cache_free(struct namecache *);
99
100 static void
101 cache_remove(struct namecache *ncp)
102 {
103
104 LOCK_ASSERT(simple_lock_held(&namecache_slock));
105
106 ncp->nc_dvp = NULL;
107 ncp->nc_vp = NULL;
108
109 TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
110 if (ncp->nc_hash.le_prev != NULL) {
111 LIST_REMOVE(ncp, nc_hash);
112 ncp->nc_hash.le_prev = NULL;
113 }
114 if (ncp->nc_vhash.le_prev != NULL) {
115 LIST_REMOVE(ncp, nc_vhash);
116 ncp->nc_vhash.le_prev = NULL;
117 }
118 if (ncp->nc_vlist.le_prev != NULL) {
119 LIST_REMOVE(ncp, nc_vlist);
120 ncp->nc_vlist.le_prev = NULL;
121 }
122 if (ncp->nc_dvlist.le_prev != NULL) {
123 LIST_REMOVE(ncp, nc_dvlist);
124 ncp->nc_dvlist.le_prev = NULL;
125 }
126 }
127
128 static void
129 cache_free(struct namecache *ncp)
130 {
131
132 pool_put(&namecache_pool, ncp);
133 numcache--; /* XXX MP */
134 }
135
136 /*
137 * Look for a the name in the cache. We don't do this
138 * if the segment name is long, simply so the cache can avoid
139 * holding long names (which would either waste space, or
140 * add greatly to the complexity).
141 *
142 * Lookup is called with ni_dvp pointing to the directory to search,
143 * ni_ptr pointing to the name of the entry being sought, ni_namelen
144 * tells the length of the name, and ni_hash contains a hash of
145 * the name. If the lookup succeeds, the vnode is locked, stored in ni_vp
146 * and a status of zero is returned. If the locking fails for whatever
147 * reason, the vnode is unlocked and the error is returned to caller.
148 * If the lookup determines that the name does not exist (negative cacheing),
149 * a status of ENOENT is returned. If the lookup fails, a status of -1
150 * is returned.
151 */
152 int
153 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp)
154 {
155 struct namecache *ncp;
156 struct nchashhead *ncpp;
157 struct vnode *vp;
158 int error;
159
160 if (!doingcache) {
161 cnp->cn_flags &= ~MAKEENTRY;
162 *vpp = NULL;
163 return (-1);
164 }
165
166 if (cnp->cn_namelen > NCHNAMLEN) {
167 /* XXXSMP - updating stats without lock; do we care? */
168 nchstats.ncs_long++;
169 cnp->cn_flags &= ~MAKEENTRY;
170 goto fail;
171 }
172 ncpp = &nchashtbl[NCHASH(cnp, dvp)];
173 simple_lock(&namecache_slock);
174 LIST_FOREACH(ncp, ncpp, nc_hash) {
175 if (ncp->nc_dvp == dvp &&
176 ncp->nc_nlen == cnp->cn_namelen &&
177 !memcmp(ncp->nc_name, cnp->cn_nameptr, (u_int)ncp->nc_nlen))
178 break;
179 }
180 if (ncp == 0) {
181 nchstats.ncs_miss++;
182 goto fail_wlock;
183 }
184 if ((cnp->cn_flags & MAKEENTRY) == 0) {
185 nchstats.ncs_badhits++;
186 goto remove;
187 } else if (ncp->nc_vp == NULL) {
188 /*
189 * Restore the ISWHITEOUT flag saved earlier.
190 */
191 cnp->cn_flags |= ncp->nc_flags;
192 if (cnp->cn_nameiop != CREATE ||
193 (cnp->cn_flags & ISLASTCN) == 0) {
194 nchstats.ncs_neghits++;
195 /*
196 * Move this slot to end of LRU chain,
197 * if not already there.
198 */
199 if (TAILQ_NEXT(ncp, nc_lru) != 0) {
200 TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
201 TAILQ_INSERT_TAIL(&nclruhead, ncp, nc_lru);
202 }
203 simple_unlock(&namecache_slock);
204 return (ENOENT);
205 } else {
206 nchstats.ncs_badhits++;
207 goto remove;
208 }
209 }
210
211 vp = ncp->nc_vp;
212
213 /*
214 * Move this slot to end of LRU chain, if not already there.
215 */
216 if (TAILQ_NEXT(ncp, nc_lru) != 0) {
217 TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
218 TAILQ_INSERT_TAIL(&nclruhead, ncp, nc_lru);
219 }
220
221 if (vp != dvp)
222 simple_lock(&vp->v_interlock);
223
224 /* Release the name cache mutex while we acquire vnode locks */
225 simple_unlock(&namecache_slock);
226
227 #ifdef DEBUG
228 /*
229 * since we released namecache_slock,
230 * we can't use this pointer any more.
231 */
232 ncp = NULL;
233 #endif /* DEBUG */
234
235 if (vp != dvp && __predict_false(vp->v_flag & VXLOCK)) {
236 /*
237 * this vnode is being cleaned out.
238 */
239 simple_unlock(&vp->v_interlock);
240 nchstats.ncs_falsehits++; /* XXX badhits? */
241 goto fail;
242 }
243
244 if (vp == dvp) { /* lookup on "." */
245 VREF(dvp);
246 error = 0;
247 } else if (cnp->cn_flags & ISDOTDOT) {
248 VOP_UNLOCK(dvp, 0);
249 cnp->cn_flags |= PDIRUNLOCK;
250 error = vget(vp, LK_EXCLUSIVE | LK_INTERLOCK);
251 /*
252 * If the above vget() succeeded and both LOCKPARENT and
253 * ISLASTCN is set, lock the directory vnode as well.
254 */
255 if (!error && (~cnp->cn_flags & (LOCKPARENT|ISLASTCN)) == 0) {
256 if ((error = vn_lock(dvp, LK_EXCLUSIVE)) != 0) {
257 vput(vp);
258 return (error);
259 }
260 cnp->cn_flags &= ~PDIRUNLOCK;
261 }
262 } else {
263 error = vget(vp, LK_EXCLUSIVE | LK_INTERLOCK);
264 /*
265 * If the above vget() failed or either of LOCKPARENT or
266 * ISLASTCN is set, unlock the directory vnode.
267 */
268 if (error || (~cnp->cn_flags & (LOCKPARENT|ISLASTCN)) != 0) {
269 VOP_UNLOCK(dvp, 0);
270 cnp->cn_flags |= PDIRUNLOCK;
271 }
272 }
273
274 /*
275 * Check that the lock succeeded.
276 */
277 if (error) {
278 /* XXXSMP - updating stats without lock; do we care? */
279 nchstats.ncs_badhits++;
280
281 /*
282 * The parent needs to be locked when we return to VOP_LOOKUP().
283 * The `.' case here should be extremely rare (if it can happen
284 * at all), so we don't bother optimizing out the unlock/relock.
285 */
286 if (vp == dvp ||
287 error || (~cnp->cn_flags & (LOCKPARENT|ISLASTCN)) != 0) {
288 if ((error = vn_lock(dvp, LK_EXCLUSIVE)) != 0)
289 return (error);
290 cnp->cn_flags &= ~PDIRUNLOCK;
291 }
292 *vpp = NULL;
293 return (-1);
294 }
295
296 /* XXXSMP - updating stats without lock; do we care? */
297 nchstats.ncs_goodhits++;
298 *vpp = vp;
299 return (0);
300
301 remove:
302 /*
303 * Last component and we are renaming or deleting,
304 * the cache entry is invalid, or otherwise don't
305 * want cache entry to exist.
306 */
307 TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
308 LIST_REMOVE(ncp, nc_hash);
309 ncp->nc_hash.le_prev = NULL;
310 if (ncp->nc_vhash.le_prev != NULL) {
311 LIST_REMOVE(ncp, nc_vhash);
312 ncp->nc_vhash.le_prev = NULL;
313 }
314 TAILQ_INSERT_HEAD(&nclruhead, ncp, nc_lru);
315
316 fail_wlock:
317 simple_unlock(&namecache_slock);
318 fail:
319 *vpp = NULL;
320 return (-1);
321 }
322
323 /*
324 * Scan cache looking for name of directory entry pointing at vp.
325 *
326 * Fill in dvpp.
327 *
328 * If bufp is non-NULL, also place the name in the buffer which starts
329 * at bufp, immediately before *bpp, and move bpp backwards to point
330 * at the start of it. (Yes, this is a little baroque, but it's done
331 * this way to cater to the whims of getcwd).
332 *
333 * Returns 0 on success, -1 on cache miss, positive errno on failure.
334 */
335 int
336 cache_revlookup(struct vnode *vp, struct vnode **dvpp, char **bpp, char *bufp)
337 {
338 struct namecache *ncp;
339 struct vnode *dvp;
340 struct ncvhashhead *nvcpp;
341 char *bp;
342
343 if (!doingcache)
344 goto out;
345
346 nvcpp = &ncvhashtbl[NCVHASH(vp)];
347
348 simple_lock(&namecache_slock);
349 LIST_FOREACH(ncp, nvcpp, nc_vhash) {
350 if (ncp->nc_vp == vp &&
351 (dvp = ncp->nc_dvp) != NULL &&
352 dvp != vp) { /* avoid pesky . entries.. */
353
354 #ifdef DIAGNOSTIC
355 if (ncp->nc_nlen == 1 &&
356 ncp->nc_name[0] == '.')
357 panic("cache_revlookup: found entry for .");
358
359 if (ncp->nc_nlen == 2 &&
360 ncp->nc_name[0] == '.' &&
361 ncp->nc_name[1] == '.')
362 panic("cache_revlookup: found entry for ..");
363 #endif
364 nchstats.ncs_revhits++;
365
366 if (bufp) {
367 bp = *bpp;
368 bp -= ncp->nc_nlen;
369 if (bp <= bufp) {
370 *dvpp = NULL;
371 simple_unlock(&namecache_slock);
372 return (ERANGE);
373 }
374 memcpy(bp, ncp->nc_name, ncp->nc_nlen);
375 *bpp = bp;
376 }
377
378 /* XXX MP: how do we know dvp won't evaporate? */
379 *dvpp = dvp;
380 simple_unlock(&namecache_slock);
381 return (0);
382 }
383 }
384 nchstats.ncs_revmiss++;
385 simple_unlock(&namecache_slock);
386 out:
387 *dvpp = NULL;
388 return (-1);
389 }
390
391 /*
392 * Add an entry to the cache
393 */
394 void
395 cache_enter(struct vnode *dvp, struct vnode *vp, struct componentname *cnp)
396 {
397 struct namecache *ncp;
398 struct nchashhead *ncpp;
399 struct ncvhashhead *nvcpp;
400
401 #ifdef DIAGNOSTIC
402 if (cnp->cn_namelen > NCHNAMLEN)
403 panic("cache_enter: name too long");
404 #endif
405 if (!doingcache)
406 return;
407 /*
408 * Free the cache slot at head of lru chain.
409 */
410 simple_lock(&namecache_slock);
411 if (numcache < numvnodes) {
412 numcache++;
413 simple_unlock(&namecache_slock);
414 ncp = pool_get(&namecache_pool, PR_WAITOK);
415 memset(ncp, 0, sizeof(*ncp));
416 simple_lock(&namecache_slock);
417 } else if ((ncp = TAILQ_FIRST(&nclruhead)) != NULL) {
418 cache_remove(ncp);
419 } else {
420 simple_unlock(&namecache_slock);
421 return;
422 }
423 /* Grab the vnode we just found. */
424 ncp->nc_vp = vp;
425 if (vp == NULL) {
426 /*
427 * For negative hits, save the ISWHITEOUT flag so we can
428 * restore it later when the cache entry is used again.
429 */
430 ncp->nc_flags = cnp->cn_flags & ISWHITEOUT;
431 }
432 /* Fill in cache info. */
433 ncp->nc_dvp = dvp;
434 LIST_INSERT_HEAD(&dvp->v_dnclist, ncp, nc_dvlist);
435 if (vp)
436 LIST_INSERT_HEAD(&vp->v_nclist, ncp, nc_vlist);
437 ncp->nc_nlen = cnp->cn_namelen;
438 memcpy(ncp->nc_name, cnp->cn_nameptr, (unsigned)ncp->nc_nlen);
439 TAILQ_INSERT_TAIL(&nclruhead, ncp, nc_lru);
440 ncpp = &nchashtbl[NCHASH(cnp, dvp)];
441 LIST_INSERT_HEAD(ncpp, ncp, nc_hash);
442
443 ncp->nc_vhash.le_prev = NULL;
444 ncp->nc_vhash.le_next = NULL;
445
446 /*
447 * Create reverse-cache entries (used in getcwd) for directories.
448 */
449 if (vp != NULL &&
450 vp != dvp &&
451 #ifndef NAMECACHE_ENTER_REVERSE
452 vp->v_type == VDIR &&
453 #endif
454 (ncp->nc_nlen > 2 ||
455 (ncp->nc_nlen > 1 && ncp->nc_name[1] != '.') ||
456 (/* ncp->nc_nlen > 0 && */ ncp->nc_name[0] != '.'))) {
457 nvcpp = &ncvhashtbl[NCVHASH(vp)];
458 LIST_INSERT_HEAD(nvcpp, ncp, nc_vhash);
459 }
460 simple_unlock(&namecache_slock);
461 }
462
463 /*
464 * Name cache initialization, from vfs_init() when we are booting
465 */
466 void
467 nchinit(void)
468 {
469
470 TAILQ_INIT(&nclruhead);
471 nchashtbl =
472 hashinit(desiredvnodes, HASH_LIST, M_CACHE, M_WAITOK, &nchash);
473 ncvhashtbl =
474 #ifdef NAMECACHE_ENTER_REVERSE
475 hashinit(desiredvnodes, HASH_LIST, M_CACHE, M_WAITOK, &ncvhash);
476 #else
477 hashinit(desiredvnodes/8, HASH_LIST, M_CACHE, M_WAITOK, &ncvhash);
478 #endif
479 pool_init(&namecache_pool, sizeof(struct namecache), 0, 0, 0,
480 "ncachepl", &pool_allocator_nointr);
481 }
482
483 /*
484 * Name cache reinitialization, for when the maximum number of vnodes increases.
485 */
486 void
487 nchreinit(void)
488 {
489 struct namecache *ncp;
490 struct nchashhead *oldhash1, *hash1;
491 struct ncvhashhead *oldhash2, *hash2;
492 u_long i, oldmask1, oldmask2, mask1, mask2;
493
494 hash1 = hashinit(desiredvnodes, HASH_LIST, M_CACHE, M_WAITOK, &mask1);
495 hash2 =
496 #ifdef NAMECACHE_ENTER_REVERSE
497 hashinit(desiredvnodes, HASH_LIST, M_CACHE, M_WAITOK, &mask2);
498 #else
499 hashinit(desiredvnodes/8, HASH_LIST, M_CACHE, M_WAITOK, &mask2);
500 #endif
501 simple_lock(&namecache_slock);
502 oldhash1 = nchashtbl;
503 oldmask1 = nchash;
504 nchashtbl = hash1;
505 nchash = mask1;
506 oldhash2 = ncvhashtbl;
507 oldmask2 = ncvhash;
508 ncvhashtbl = hash2;
509 ncvhash = mask2;
510 for (i = 0; i <= oldmask1; i++) {
511 while ((ncp = LIST_FIRST(&oldhash1[i])) != NULL) {
512 LIST_REMOVE(ncp, nc_hash);
513 ncp->nc_hash.le_prev = NULL;
514 }
515 }
516 for (i = 0; i <= oldmask2; i++) {
517 while ((ncp = LIST_FIRST(&oldhash2[i])) != NULL) {
518 LIST_REMOVE(ncp, nc_vhash);
519 ncp->nc_vhash.le_prev = NULL;
520 }
521 }
522 simple_unlock(&namecache_slock);
523 hashdone(oldhash1, M_CACHE);
524 hashdone(oldhash2, M_CACHE);
525 }
526
527 /*
528 * Cache flush, a particular vnode; called when a vnode is renamed to
529 * hide entries that would now be invalid
530 */
531 void
532 cache_purge(struct vnode *vp)
533 {
534 struct namecache *ncp, *ncnext;
535
536 simple_lock(&namecache_slock);
537 for (ncp = LIST_FIRST(&vp->v_nclist); ncp != NULL; ncp = ncnext) {
538 ncnext = LIST_NEXT(ncp, nc_vlist);
539 cache_remove(ncp);
540 cache_free(ncp);
541 }
542 for (ncp = LIST_FIRST(&vp->v_dnclist); ncp != NULL; ncp = ncnext) {
543 ncnext = LIST_NEXT(ncp, nc_dvlist);
544 cache_remove(ncp);
545 cache_free(ncp);
546 }
547 simple_unlock(&namecache_slock);
548 }
549
550 /*
551 * Cache flush, a whole filesystem; called when filesys is umounted to
552 * remove entries that would now be invalid.
553 */
554 void
555 cache_purgevfs(struct mount *mp)
556 {
557 struct namecache *ncp, *nxtcp;
558
559 simple_lock(&namecache_slock);
560 for (ncp = TAILQ_FIRST(&nclruhead); ncp != NULL; ncp = nxtcp) {
561 nxtcp = TAILQ_NEXT(ncp, nc_lru);
562 if (ncp->nc_dvp == NULL || ncp->nc_dvp->v_mount != mp) {
563 continue;
564 }
565 /* Free the resources we had. */
566 cache_remove(ncp);
567 cache_free(ncp);
568 }
569 simple_unlock(&namecache_slock);
570 }
571
572 #ifdef DDB
573 void
574 namecache_print(struct vnode *vp, void (*pr)(const char *, ...))
575 {
576 struct vnode *dvp = NULL;
577 struct namecache *ncp;
578
579 TAILQ_FOREACH(ncp, &nclruhead, nc_lru) {
580 if (ncp->nc_vp == vp) {
581 (*pr)("name %.*s\n", ncp->nc_nlen, ncp->nc_name);
582 dvp = ncp->nc_dvp;
583 }
584 }
585 if (dvp == NULL) {
586 (*pr)("name not found\n");
587 return;
588 }
589 vp = dvp;
590 TAILQ_FOREACH(ncp, &nclruhead, nc_lru) {
591 if (ncp->nc_vp == vp) {
592 (*pr)("parent %.*s\n", ncp->nc_nlen, ncp->nc_name);
593 }
594 }
595 }
596 #endif
597