xref: /netbsd-src/sys/kern/vfs_cache.c (revision 06be8101a16cc95f40783b3cb7afd12112103a9a)
1 /*	$NetBSD: vfs_cache.c,v 1.32 2001/11/12 15:25:36 lukem 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. All advertising materials mentioning features or use of this software
16  *    must display the following acknowledgement:
17  *	This product includes software developed by the University of
18  *	California, Berkeley and its contributors.
19  * 4. Neither the name of the University nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  *
35  *	@(#)vfs_cache.c	8.3 (Berkeley) 8/22/94
36  */
37 
38 #include <sys/cdefs.h>
39 __KERNEL_RCSID(0, "$NetBSD: vfs_cache.c,v 1.32 2001/11/12 15:25:36 lukem Exp $");
40 
41 #include "opt_ddb.h"
42 #include "opt_revcache.h"
43 
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/time.h>
47 #include <sys/mount.h>
48 #include <sys/vnode.h>
49 #include <sys/namei.h>
50 #include <sys/errno.h>
51 #include <sys/malloc.h>
52 #include <sys/pool.h>
53 
54 /*
55  * Name caching works as follows:
56  *
57  * Names found by directory scans are retained in a cache
58  * for future reference.  It is managed LRU, so frequently
59  * used names will hang around.  Cache is indexed by hash value
60  * obtained from (dvp, name) where dvp refers to the directory
61  * containing name.
62  *
63  * For simplicity (and economy of storage), names longer than
64  * a maximum length of NCHNAMLEN are not cached; they occur
65  * infrequently in any case, and are almost never of interest.
66  *
67  * Upon reaching the last segment of a path, if the reference
68  * is for DELETE, or NOCACHE is set (rewrite), and the
69  * name is located in the cache, it will be dropped.
70  * The entry is dropped also when it was not possible to lock
71  * the cached vnode, either because vget() failed or the generation
72  * number has changed while waiting for the lock.
73  */
74 
75 /*
76  * Structures associated with name cacheing.
77  */
78 LIST_HEAD(nchashhead, namecache) *nchashtbl;
79 u_long	nchash;				/* size of hash table - 1 */
80 long	numcache;			/* number of cache entries allocated */
81 #define NCHASH(cnp, dvp)	(((cnp)->cn_hash ^ (dvp)->v_id) & nchash)
82 
83 LIST_HEAD(ncvhashhead, namecache) *ncvhashtbl;
84 u_long	ncvhash;			/* size of hash table - 1 */
85 #define NCVHASH(vp)		((vp)->v_id & ncvhash)
86 
87 TAILQ_HEAD(, namecache) nclruhead;		/* LRU chain */
88 struct	nchstats nchstats;		/* cache effectiveness statistics */
89 
90 struct pool namecache_pool;
91 
92 int doingcache = 1;			/* 1 => enable the cache */
93 
94 /*
95  * Look for a the name in the cache. We don't do this
96  * if the segment name is long, simply so the cache can avoid
97  * holding long names (which would either waste space, or
98  * add greatly to the complexity).
99  *
100  * Lookup is called with ni_dvp pointing to the directory to search,
101  * ni_ptr pointing to the name of the entry being sought, ni_namelen
102  * tells the length of the name, and ni_hash contains a hash of
103  * the name. If the lookup succeeds, the vnode is locked, stored in ni_vp
104  * and a status of zero is returned. If the locking fails for whatever
105  * reason, the vnode is unlocked and the error is returned to caller.
106  * If the lookup determines that the name does not exist (negative cacheing),
107  * a status of ENOENT is returned. If the lookup fails, a status of -1
108  * is returned.
109  */
110 int
111 cache_lookup(dvp, vpp, cnp)
112 	struct vnode *dvp;
113 	struct vnode **vpp;
114 	struct componentname *cnp;
115 {
116 	struct namecache *ncp;
117 	struct nchashhead *ncpp;
118 	struct vnode *vp;
119 	int vpid, error;
120 
121 	if (!doingcache) {
122 		cnp->cn_flags &= ~MAKEENTRY;
123 		*vpp = 0;
124 		return (-1);
125 	}
126 	if (cnp->cn_namelen > NCHNAMLEN) {
127 		nchstats.ncs_long++;
128 		cnp->cn_flags &= ~MAKEENTRY;
129 		*vpp = 0;
130 		return (-1);
131 	}
132 	ncpp = &nchashtbl[NCHASH(cnp, dvp)];
133 	LIST_FOREACH(ncp, ncpp, nc_hash) {
134 		if (ncp->nc_dvp == dvp &&
135 		    ncp->nc_dvpid == dvp->v_id &&
136 		    ncp->nc_nlen == cnp->cn_namelen &&
137 		    !memcmp(ncp->nc_name, cnp->cn_nameptr, (u_int)ncp->nc_nlen))
138 			break;
139 	}
140 	if (ncp == 0) {
141 		nchstats.ncs_miss++;
142 		*vpp = 0;
143 		return (-1);
144 	}
145 	if ((cnp->cn_flags & MAKEENTRY) == 0) {
146 		nchstats.ncs_badhits++;
147 		goto remove;
148 	} else if (ncp->nc_vp == NULL) {
149 		/*
150 		 * Restore the ISWHITEOUT flag saved earlier.
151 		 */
152 		cnp->cn_flags |= ncp->nc_vpid;
153 		if (cnp->cn_nameiop != CREATE ||
154 		    (cnp->cn_flags & ISLASTCN) == 0) {
155 			nchstats.ncs_neghits++;
156 			/*
157 			 * Move this slot to end of LRU chain,
158 			 * if not already there.
159 			 */
160 			if (ncp->nc_lru.tqe_next != 0) {
161 				TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
162 				TAILQ_INSERT_TAIL(&nclruhead, ncp, nc_lru);
163 			}
164 			return (ENOENT);
165 		} else {
166 			nchstats.ncs_badhits++;
167 			goto remove;
168 		}
169 	} else if (ncp->nc_vpid != ncp->nc_vp->v_id) {
170 		nchstats.ncs_falsehits++;
171 		goto remove;
172 	}
173 
174 	vp = ncp->nc_vp;
175 	vpid = vp->v_id;
176 	if (vp == dvp) {	/* lookup on "." */
177 		VREF(dvp);
178 		error = 0;
179 	} else if (cnp->cn_flags & ISDOTDOT) {
180 		VOP_UNLOCK(dvp, 0);
181 		cnp->cn_flags |= PDIRUNLOCK;
182 		error = vget(vp, LK_EXCLUSIVE);
183 		/* if the above vget() succeeded and both LOCKPARENT and
184 		 * ISLASTCN is set, lock the directory vnode as well */
185 		if (!error && (~cnp->cn_flags & (LOCKPARENT|ISLASTCN)) == 0) {
186 			if ((error = vn_lock(dvp, LK_EXCLUSIVE)) != 0) {
187 				vput(vp);
188 				return (error);
189 			}
190 			cnp->cn_flags &= ~PDIRUNLOCK;
191 		}
192 	} else {
193 		error = vget(vp, LK_EXCLUSIVE);
194 		/* if the above vget() failed or either of LOCKPARENT or
195 		 * ISLASTCN is set, unlock the directory vnode */
196 		if (error || (~cnp->cn_flags & (LOCKPARENT|ISLASTCN)) != 0) {
197 			VOP_UNLOCK(dvp, 0);
198 			cnp->cn_flags |= PDIRUNLOCK;
199 		}
200 	}
201 
202 	/*
203 	 * Check that the lock succeeded, and that the capability number did
204 	 * not change while we were waiting for the lock.
205 	 */
206 	if (error || vpid != vp->v_id) {
207 		if (!error) {
208 			vput(vp);
209 			nchstats.ncs_falsehits++;
210 		} else
211 			nchstats.ncs_badhits++;
212 		/*
213 		 * The parent needs to be locked when we return to VOP_LOOKUP().
214 		 * The `.' case here should be extremely rare (if it can happen
215 		 * at all), so we don't bother optimizing out the unlock/relock.
216 		 */
217 		if (vp == dvp ||
218 		    error || (~cnp->cn_flags & (LOCKPARENT|ISLASTCN)) != 0) {
219 			if ((error = vn_lock(dvp, LK_EXCLUSIVE)) != 0)
220 				return (error);
221 			cnp->cn_flags &= ~PDIRUNLOCK;
222 		}
223 		*vpp = NULL;
224 		return (-1);
225 	}
226 
227 	nchstats.ncs_goodhits++;
228 	/*
229 	 * move this slot to end of LRU chain, if not already there
230 	 */
231 	if (ncp->nc_lru.tqe_next != 0) {
232 		TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
233 		TAILQ_INSERT_TAIL(&nclruhead, ncp, nc_lru);
234 	}
235 	*vpp = vp;
236 	return (0);
237 
238 remove:
239 	/*
240 	 * Last component and we are renaming or deleting,
241 	 * the cache entry is invalid, or otherwise don't
242 	 * want cache entry to exist.
243 	 */
244 	TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
245 	LIST_REMOVE(ncp, nc_hash);
246 	ncp->nc_hash.le_prev = 0;
247 	if (ncp->nc_vhash.le_prev != NULL) {
248 		LIST_REMOVE(ncp, nc_vhash);
249 		ncp->nc_vhash.le_prev = 0;
250 	}
251 	TAILQ_INSERT_HEAD(&nclruhead, ncp, nc_lru);
252 	*vpp = 0;
253 	return (-1);
254 }
255 
256 /*
257  * Scan cache looking for name of directory entry pointing at vp.
258  *
259  * Fill in dvpp.
260  *
261  * If bufp is non-NULL, also place the name in the buffer which starts
262  * at bufp, immediately before *bpp, and move bpp backwards to point
263  * at the start of it.  (Yes, this is a little baroque, but it's done
264  * this way to cater to the whims of getcwd).
265  *
266  * Returns 0 on success, -1 on cache miss, positive errno on failure.
267  */
268 int
269 cache_revlookup (vp, dvpp, bpp, bufp)
270 	struct vnode *vp, **dvpp;
271 	char **bpp;
272 	char *bufp;
273 {
274 	struct namecache *ncp;
275 	struct vnode *dvp;
276 	struct ncvhashhead *nvcpp;
277 
278 	if (!doingcache)
279 		goto out;
280 
281 	nvcpp = &ncvhashtbl[NCVHASH(vp)];
282 
283 	LIST_FOREACH(ncp, nvcpp, nc_vhash) {
284 		if ((ncp->nc_vp == vp) &&
285 		    (ncp->nc_vpid == vp->v_id) &&
286 		    ((dvp = ncp->nc_dvp) != 0) &&
287 		    (dvp != vp) && 		/* avoid pesky . entries.. */
288 		    (dvp->v_id == ncp->nc_dvpid))
289 		{
290 			char *bp;
291 
292 #ifdef DIAGNOSTIC
293 			if ((ncp->nc_nlen == 1) &&
294 			    (ncp->nc_name[0] == '.'))
295 				panic("cache_revlookup: found entry for .");
296 
297 			if ((ncp->nc_nlen == 2) &&
298 			    (ncp->nc_name[0] == '.') &&
299 			    (ncp->nc_name[1] == '.'))
300 				panic("cache_revlookup: found entry for ..");
301 #endif
302 			nchstats.ncs_revhits++;
303 
304 			if (bufp) {
305 				bp = *bpp;
306 				bp -= ncp->nc_nlen;
307 				if (bp <= bufp) {
308 					*dvpp = 0;
309 					return ERANGE;
310 				}
311 				memcpy(bp, ncp->nc_name, ncp->nc_nlen);
312 				*bpp = bp;
313 			}
314 
315 			/* XXX MP: how do we know dvp won't evaporate? */
316 			*dvpp = dvp;
317 			return 0;
318 		}
319 	}
320 	nchstats.ncs_revmiss++;
321  out:
322 	*dvpp = 0;
323 	return -1;
324 }
325 
326 /*
327  * Add an entry to the cache
328  */
329 void
330 cache_enter(dvp, vp, cnp)
331 	struct vnode *dvp;
332 	struct vnode *vp;
333 	struct componentname *cnp;
334 {
335 	struct namecache *ncp;
336 	struct nchashhead *ncpp;
337 	struct ncvhashhead *nvcpp;
338 
339 #ifdef DIAGNOSTIC
340 	if (cnp->cn_namelen > NCHNAMLEN)
341 		panic("cache_enter: name too long");
342 #endif
343 	if (!doingcache)
344 		return;
345 	/*
346 	 * Free the cache slot at head of lru chain.
347 	 */
348 	if (numcache < numvnodes) {
349 		ncp = pool_get(&namecache_pool, PR_WAITOK);
350 		memset(ncp, 0, sizeof(*ncp));
351 		numcache++;
352 	} else if ((ncp = TAILQ_FIRST(&nclruhead)) != NULL) {
353 		TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
354 		if (ncp->nc_hash.le_prev != 0) {
355 			LIST_REMOVE(ncp, nc_hash);
356 			ncp->nc_hash.le_prev = 0;
357 		}
358 		if (ncp->nc_vhash.le_prev != 0) {
359 			LIST_REMOVE(ncp, nc_vhash);
360 			ncp->nc_vhash.le_prev = 0;
361 		}
362 	} else
363 		return;
364 	/* grab the vnode we just found */
365 	ncp->nc_vp = vp;
366 	if (vp)
367 		ncp->nc_vpid = vp->v_id;
368 	else {
369 		/*
370 		 * For negative hits, save the ISWHITEOUT flag so we can
371 		 * restore it later when the cache entry is used again.
372 		 */
373 		ncp->nc_vpid = cnp->cn_flags & ISWHITEOUT;
374 	}
375 	/* fill in cache info */
376 	ncp->nc_dvp = dvp;
377 	ncp->nc_dvpid = dvp->v_id;
378 	ncp->nc_nlen = cnp->cn_namelen;
379 	memcpy(ncp->nc_name, cnp->cn_nameptr, (unsigned)ncp->nc_nlen);
380 	TAILQ_INSERT_TAIL(&nclruhead, ncp, nc_lru);
381 	ncpp = &nchashtbl[NCHASH(cnp, dvp)];
382 	LIST_INSERT_HEAD(ncpp, ncp, nc_hash);
383 
384 	ncp->nc_vhash.le_prev = 0;
385 	ncp->nc_vhash.le_next = 0;
386 
387 	/*
388 	 * Create reverse-cache entries (used in getcwd) for directories.
389 	 */
390 	if (vp &&
391 	    (vp != dvp) &&
392 #ifndef NAMECACHE_ENTER_REVERSE
393 	    (vp->v_type == VDIR) &&
394 #endif
395 	    ((ncp->nc_nlen > 2) ||
396 	     ((ncp->nc_nlen == 2) && (ncp->nc_name[0] != '.') && (ncp->nc_name[1] != '.')) ||
397 	     ((ncp->nc_nlen == 1) && (ncp->nc_name[0] != '.'))))
398 	{
399 		nvcpp = &ncvhashtbl[NCVHASH(vp)];
400 		LIST_INSERT_HEAD(nvcpp, ncp, nc_vhash);
401 	}
402 
403 }
404 
405 /*
406  * Name cache initialization, from vfs_init() when we are booting
407  */
408 void
409 nchinit()
410 {
411 
412 	TAILQ_INIT(&nclruhead);
413 	nchashtbl =
414 	    hashinit(desiredvnodes, HASH_LIST, M_CACHE, M_WAITOK, &nchash);
415 	ncvhashtbl =
416 #ifdef NAMECACHE_ENTER_REVERSE
417 	    hashinit(desiredvnodes, HASH_LIST, M_CACHE, M_WAITOK, &ncvhash);
418 #else
419 	    hashinit(desiredvnodes/8, HASH_LIST, M_CACHE, M_WAITOK, &ncvhash);
420 #endif
421 	pool_init(&namecache_pool, sizeof(struct namecache), 0, 0, 0,
422 	    "ncachepl", 0, pool_page_alloc_nointr, pool_page_free_nointr,
423 	    M_CACHE);
424 }
425 
426 /*
427  * Name cache reinitialization, for when the maximum number of vnodes increases.
428  */
429 void
430 nchreinit()
431 {
432 	struct namecache *ncp;
433 	struct nchashhead *oldhash1, *hash1;
434 	struct ncvhashhead *oldhash2, *hash2;
435 	u_long oldmask1, oldmask2, mask1, mask2;
436 	int i;
437 
438 	hash1 = hashinit(desiredvnodes, HASH_LIST, M_CACHE, M_WAITOK, &mask1);
439 	hash2 =
440 #ifdef NAMECACHE_ENTER_REVERSE
441 	    hashinit(desiredvnodes, HASH_LIST, M_CACHE, M_WAITOK, &mask2);
442 #else
443 	    hashinit(desiredvnodes/8, HASH_LIST, M_CACHE, M_WAITOK, &mask2);
444 #endif
445 	oldhash1 = nchashtbl;
446 	oldmask1 = nchash;
447 	nchashtbl = hash1;
448 	nchash = mask1;
449 	oldhash2 = ncvhashtbl;
450 	oldmask2 = ncvhash;
451 	ncvhashtbl = hash2;
452 	ncvhash = mask2;
453 	for (i = 0; i <= oldmask1; i++) {
454 		while ((ncp = LIST_FIRST(&oldhash1[i])) != NULL) {
455 			LIST_REMOVE(ncp, nc_hash);
456 			ncp->nc_hash.le_prev = NULL;
457 		}
458 	}
459 	for (i = 0; i <= oldmask2; i++) {
460 		while ((ncp = LIST_FIRST(&oldhash2[i])) != NULL) {
461 			LIST_REMOVE(ncp, nc_vhash);
462 			ncp->nc_vhash.le_prev = NULL;
463 		}
464 	}
465 	hashdone(oldhash1, M_CACHE);
466 	hashdone(oldhash2, M_CACHE);
467 }
468 
469 /*
470  * Cache flush, a particular vnode; called when a vnode is renamed to
471  * hide entries that would now be invalid
472  */
473 void
474 cache_purge(vp)
475 	struct vnode *vp;
476 {
477 	struct namecache *ncp;
478 	struct nchashhead *ncpp;
479 	static u_long nextvnodeid;
480 
481 	vp->v_id = ++nextvnodeid;
482 	if (nextvnodeid != 0)
483 		return;
484 	for (ncpp = &nchashtbl[nchash]; ncpp >= nchashtbl; ncpp--) {
485 		LIST_FOREACH(ncp, ncpp, nc_hash) {
486 			ncp->nc_vpid = 0;
487 			ncp->nc_dvpid = 0;
488 		}
489 	}
490 	vp->v_id = ++nextvnodeid;
491 }
492 
493 /*
494  * Cache flush, a whole filesystem; called when filesys is umounted to
495  * remove entries that would now be invalid.
496  */
497 void
498 cache_purgevfs(mp)
499 	struct mount *mp;
500 {
501 	struct namecache *ncp, *nxtcp;
502 
503 	for (ncp = TAILQ_FIRST(&nclruhead); ncp != NULL; ncp = nxtcp) {
504 		nxtcp = TAILQ_NEXT(ncp, nc_lru);
505 		if (ncp->nc_dvp == NULL || ncp->nc_dvp->v_mount != mp) {
506 			continue;
507 		}
508 		/* free the resources we had */
509 		ncp->nc_vp = NULL;
510 		ncp->nc_dvp = NULL;
511 		TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
512 		if (ncp->nc_hash.le_prev != 0) {
513 			LIST_REMOVE(ncp, nc_hash);
514 			ncp->nc_hash.le_prev = 0;
515 		}
516 		if (ncp->nc_vhash.le_prev != 0) {
517 			LIST_REMOVE(ncp, nc_vhash);
518 			ncp->nc_vhash.le_prev = 0;
519 		}
520 		TAILQ_INSERT_HEAD(&nclruhead, ncp, nc_lru);
521 	}
522 }
523 
524 #ifdef DDB
525 void
526 namecache_print(struct vnode *vp, void (*pr)(const char *, ...))
527 {
528 	struct vnode *dvp = NULL;
529 	struct namecache *ncp;
530 
531 	TAILQ_FOREACH(ncp, &nclruhead, nc_lru) {
532 		if (ncp->nc_vp == vp && ncp->nc_vpid == vp->v_id) {
533 			(*pr)("name %.*s\n", ncp->nc_nlen, ncp->nc_name);
534 			dvp = ncp->nc_dvp;
535 		}
536 	}
537 	if (dvp == NULL) {
538 		(*pr)("name not found\n");
539 		return;
540 	}
541 	vp = dvp;
542 	TAILQ_FOREACH(ncp, &nclruhead, nc_lru) {
543 		if (ncp->nc_vp == vp && ncp->nc_vpid == vp->v_id) {
544 			(*pr)("parent %.*s\n", ncp->nc_nlen, ncp->nc_name);
545 		}
546 	}
547 }
548 #endif
549