xref: /netbsd-src/sys/kern/vfs_cache.c (revision 5e4c038a45edbc7d63b7c2daa76e29f88b64a4e3)
1 /*	$NetBSD: vfs_cache.c,v 1.35 2002/03/08 20:48:42 thorpej 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 1.35 2002/03/08 20:48:42 thorpej 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(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp)
112 {
113 	struct namecache *ncp;
114 	struct nchashhead *ncpp;
115 	struct vnode *vp;
116 	int vpid, error;
117 
118 	if (!doingcache) {
119 		cnp->cn_flags &= ~MAKEENTRY;
120 		*vpp = NULL;
121 		return (-1);
122 	}
123 	if (cnp->cn_namelen > NCHNAMLEN) {
124 		nchstats.ncs_long++;
125 		cnp->cn_flags &= ~MAKEENTRY;
126 		*vpp = NULL;
127 		return (-1);
128 	}
129 	ncpp = &nchashtbl[NCHASH(cnp, dvp)];
130 	LIST_FOREACH(ncp, ncpp, nc_hash) {
131 		if (ncp->nc_dvp == dvp &&
132 		    ncp->nc_dvpid == dvp->v_id &&
133 		    ncp->nc_nlen == cnp->cn_namelen &&
134 		    !memcmp(ncp->nc_name, cnp->cn_nameptr, (u_int)ncp->nc_nlen))
135 			break;
136 	}
137 	if (ncp == 0) {
138 		nchstats.ncs_miss++;
139 		*vpp = NULL;
140 		return (-1);
141 	}
142 	if ((cnp->cn_flags & MAKEENTRY) == 0) {
143 		nchstats.ncs_badhits++;
144 		goto remove;
145 	} else if (ncp->nc_vp == NULL) {
146 		/*
147 		 * Restore the ISWHITEOUT flag saved earlier.
148 		 */
149 		cnp->cn_flags |= ncp->nc_vpid;
150 		if (cnp->cn_nameiop != CREATE ||
151 		    (cnp->cn_flags & ISLASTCN) == 0) {
152 			nchstats.ncs_neghits++;
153 			/*
154 			 * Move this slot to end of LRU chain,
155 			 * if not already there.
156 			 */
157 			if (ncp->nc_lru.tqe_next != 0) {
158 				TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
159 				TAILQ_INSERT_TAIL(&nclruhead, ncp, nc_lru);
160 			}
161 			return (ENOENT);
162 		} else {
163 			nchstats.ncs_badhits++;
164 			goto remove;
165 		}
166 	} else if (ncp->nc_vpid != ncp->nc_vp->v_id) {
167 		nchstats.ncs_falsehits++;
168 		goto remove;
169 	}
170 
171 	vp = ncp->nc_vp;
172 	vpid = vp->v_id;
173 	if (vp == dvp) {	/* lookup on "." */
174 		VREF(dvp);
175 		error = 0;
176 	} else if (cnp->cn_flags & ISDOTDOT) {
177 		VOP_UNLOCK(dvp, 0);
178 		cnp->cn_flags |= PDIRUNLOCK;
179 		error = vget(vp, LK_EXCLUSIVE);
180 		/*
181 		 * If the above vget() succeeded and both LOCKPARENT and
182 		 * ISLASTCN is set, lock the directory vnode as well.
183 		 */
184 		if (!error && (~cnp->cn_flags & (LOCKPARENT|ISLASTCN)) == 0) {
185 			if ((error = vn_lock(dvp, LK_EXCLUSIVE)) != 0) {
186 				vput(vp);
187 				return (error);
188 			}
189 			cnp->cn_flags &= ~PDIRUNLOCK;
190 		}
191 	} else {
192 		error = vget(vp, LK_EXCLUSIVE);
193 		/*
194 		 * If the above vget() failed or either of LOCKPARENT or
195 		 * ISLASTCN is set, unlock the directory vnode.
196 		 */
197 		if (error || (~cnp->cn_flags & (LOCKPARENT|ISLASTCN)) != 0) {
198 			VOP_UNLOCK(dvp, 0);
199 			cnp->cn_flags |= PDIRUNLOCK;
200 		}
201 	}
202 
203 	/*
204 	 * Check that the lock succeeded, and that the capability number did
205 	 * not change while we were waiting for the lock.
206 	 */
207 	if (error || vpid != vp->v_id) {
208 		if (!error) {
209 			vput(vp);
210 			nchstats.ncs_falsehits++;
211 		} else
212 			nchstats.ncs_badhits++;
213 		/*
214 		 * The parent needs to be locked when we return to VOP_LOOKUP().
215 		 * The `.' case here should be extremely rare (if it can happen
216 		 * at all), so we don't bother optimizing out the unlock/relock.
217 		 */
218 		if (vp == dvp ||
219 		    error || (~cnp->cn_flags & (LOCKPARENT|ISLASTCN)) != 0) {
220 			if ((error = vn_lock(dvp, LK_EXCLUSIVE)) != 0)
221 				return (error);
222 			cnp->cn_flags &= ~PDIRUNLOCK;
223 		}
224 		*vpp = NULL;
225 		return (-1);
226 	}
227 
228 	nchstats.ncs_goodhits++;
229 	/*
230 	 * Move this slot to end of LRU chain, if not already there.
231 	 */
232 	if (ncp->nc_lru.tqe_next != 0) {
233 		TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
234 		TAILQ_INSERT_TAIL(&nclruhead, ncp, nc_lru);
235 	}
236 	*vpp = vp;
237 	return (0);
238 
239 remove:
240 	/*
241 	 * Last component and we are renaming or deleting,
242 	 * the cache entry is invalid, or otherwise don't
243 	 * want cache entry to exist.
244 	 */
245 	TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
246 	LIST_REMOVE(ncp, nc_hash);
247 	ncp->nc_hash.le_prev = NULL;
248 	if (ncp->nc_vhash.le_prev != NULL) {
249 		LIST_REMOVE(ncp, nc_vhash);
250 		ncp->nc_vhash.le_prev = NULL;
251 	}
252 	TAILQ_INSERT_HEAD(&nclruhead, ncp, nc_lru);
253 	*vpp = NULL;
254 	return (-1);
255 }
256 
257 /*
258  * Scan cache looking for name of directory entry pointing at vp.
259  *
260  * Fill in dvpp.
261  *
262  * If bufp is non-NULL, also place the name in the buffer which starts
263  * at bufp, immediately before *bpp, and move bpp backwards to point
264  * at the start of it.  (Yes, this is a little baroque, but it's done
265  * this way to cater to the whims of getcwd).
266  *
267  * Returns 0 on success, -1 on cache miss, positive errno on failure.
268  */
269 int
270 cache_revlookup(struct vnode *vp, struct vnode **dvpp, char **bpp, char *bufp)
271 {
272 	struct namecache *ncp;
273 	struct vnode *dvp;
274 	struct ncvhashhead *nvcpp;
275 	char *bp;
276 
277 	if (!doingcache)
278 		goto out;
279 
280 	nvcpp = &ncvhashtbl[NCVHASH(vp)];
281 
282 	LIST_FOREACH(ncp, nvcpp, nc_vhash) {
283 		if (ncp->nc_vp == vp &&
284 		    ncp->nc_vpid == vp->v_id &&
285 		    (dvp = ncp->nc_dvp) != NULL &&
286 		    dvp != vp && 		/* avoid pesky . entries.. */
287 		    dvp->v_id == ncp->nc_dvpid) {
288 
289 #ifdef DIAGNOSTIC
290 			if (ncp->nc_nlen == 1 &&
291 			    ncp->nc_name[0] == '.')
292 				panic("cache_revlookup: found entry for .");
293 
294 			if (ncp->nc_nlen == 2 &&
295 			    ncp->nc_name[0] == '.' &&
296 			    ncp->nc_name[1] == '.')
297 				panic("cache_revlookup: found entry for ..");
298 #endif
299 			nchstats.ncs_revhits++;
300 
301 			if (bufp) {
302 				bp = *bpp;
303 				bp -= ncp->nc_nlen;
304 				if (bp <= bufp) {
305 					*dvpp = NULL;
306 					return (ERANGE);
307 				}
308 				memcpy(bp, ncp->nc_name, ncp->nc_nlen);
309 				*bpp = bp;
310 			}
311 
312 			/* XXX MP: how do we know dvp won't evaporate? */
313 			*dvpp = dvp;
314 			return (0);
315 		}
316 	}
317 	nchstats.ncs_revmiss++;
318  out:
319 	*dvpp = NULL;
320 	return (-1);
321 }
322 
323 /*
324  * Add an entry to the cache
325  */
326 void
327 cache_enter(struct vnode *dvp, struct vnode *vp, struct componentname *cnp)
328 {
329 	struct namecache *ncp;
330 	struct nchashhead *ncpp;
331 	struct ncvhashhead *nvcpp;
332 
333 #ifdef DIAGNOSTIC
334 	if (cnp->cn_namelen > NCHNAMLEN)
335 		panic("cache_enter: name too long");
336 #endif
337 	if (!doingcache)
338 		return;
339 	/*
340 	 * Free the cache slot at head of lru chain.
341 	 */
342 	if (numcache < numvnodes) {
343 		ncp = pool_get(&namecache_pool, PR_WAITOK);
344 		memset(ncp, 0, sizeof(*ncp));
345 		numcache++;
346 	} else if ((ncp = TAILQ_FIRST(&nclruhead)) != NULL) {
347 		TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
348 		if (ncp->nc_hash.le_prev != NULL) {
349 			LIST_REMOVE(ncp, nc_hash);
350 			ncp->nc_hash.le_prev = NULL;
351 		}
352 		if (ncp->nc_vhash.le_prev != NULL) {
353 			LIST_REMOVE(ncp, nc_vhash);
354 			ncp->nc_vhash.le_prev = NULL;
355 		}
356 	} else
357 		return;
358 	/* Grab the vnode we just found. */
359 	ncp->nc_vp = vp;
360 	if (vp)
361 		ncp->nc_vpid = vp->v_id;
362 	else {
363 		/*
364 		 * For negative hits, save the ISWHITEOUT flag so we can
365 		 * restore it later when the cache entry is used again.
366 		 */
367 		ncp->nc_vpid = cnp->cn_flags & ISWHITEOUT;
368 	}
369 	/* Fill in cache info. */
370 	ncp->nc_dvp = dvp;
371 	ncp->nc_dvpid = dvp->v_id;
372 	ncp->nc_nlen = cnp->cn_namelen;
373 	memcpy(ncp->nc_name, cnp->cn_nameptr, (unsigned)ncp->nc_nlen);
374 	TAILQ_INSERT_TAIL(&nclruhead, ncp, nc_lru);
375 	ncpp = &nchashtbl[NCHASH(cnp, dvp)];
376 	LIST_INSERT_HEAD(ncpp, ncp, nc_hash);
377 
378 	ncp->nc_vhash.le_prev = NULL;
379 	ncp->nc_vhash.le_next = NULL;
380 
381 	/*
382 	 * Create reverse-cache entries (used in getcwd) for directories.
383 	 */
384 	if (vp != NULL &&
385 	    vp != dvp &&
386 #ifndef NAMECACHE_ENTER_REVERSE
387 	    vp->v_type == VDIR &&
388 #endif
389 	    (ncp->nc_nlen > 2 ||
390 	    (ncp->nc_nlen > 1 && ncp->nc_name[1] != '.') ||
391 	    (/* ncp->nc_nlen > 0 && */ ncp->nc_name[0] != '.'))) {
392 		nvcpp = &ncvhashtbl[NCVHASH(vp)];
393 		LIST_INSERT_HEAD(nvcpp, ncp, nc_vhash);
394 	}
395 }
396 
397 /*
398  * Name cache initialization, from vfs_init() when we are booting
399  */
400 void
401 nchinit(void)
402 {
403 
404 	TAILQ_INIT(&nclruhead);
405 	nchashtbl =
406 	    hashinit(desiredvnodes, HASH_LIST, M_CACHE, M_WAITOK, &nchash);
407 	ncvhashtbl =
408 #ifdef NAMECACHE_ENTER_REVERSE
409 	    hashinit(desiredvnodes, HASH_LIST, M_CACHE, M_WAITOK, &ncvhash);
410 #else
411 	    hashinit(desiredvnodes/8, HASH_LIST, M_CACHE, M_WAITOK, &ncvhash);
412 #endif
413 	pool_init(&namecache_pool, sizeof(struct namecache), 0, 0, 0,
414 	    "ncachepl", &pool_allocator_nointr);
415 }
416 
417 /*
418  * Name cache reinitialization, for when the maximum number of vnodes increases.
419  */
420 void
421 nchreinit(void)
422 {
423 	struct namecache *ncp;
424 	struct nchashhead *oldhash1, *hash1;
425 	struct ncvhashhead *oldhash2, *hash2;
426 	u_long oldmask1, oldmask2, mask1, mask2;
427 	int i;
428 
429 	hash1 = hashinit(desiredvnodes, HASH_LIST, M_CACHE, M_WAITOK, &mask1);
430 	hash2 =
431 #ifdef NAMECACHE_ENTER_REVERSE
432 	    hashinit(desiredvnodes, HASH_LIST, M_CACHE, M_WAITOK, &mask2);
433 #else
434 	    hashinit(desiredvnodes/8, HASH_LIST, M_CACHE, M_WAITOK, &mask2);
435 #endif
436 	oldhash1 = nchashtbl;
437 	oldmask1 = nchash;
438 	nchashtbl = hash1;
439 	nchash = mask1;
440 	oldhash2 = ncvhashtbl;
441 	oldmask2 = ncvhash;
442 	ncvhashtbl = hash2;
443 	ncvhash = mask2;
444 	for (i = 0; i <= oldmask1; i++) {
445 		while ((ncp = LIST_FIRST(&oldhash1[i])) != NULL) {
446 			LIST_REMOVE(ncp, nc_hash);
447 			ncp->nc_hash.le_prev = NULL;
448 		}
449 	}
450 	for (i = 0; i <= oldmask2; i++) {
451 		while ((ncp = LIST_FIRST(&oldhash2[i])) != NULL) {
452 			LIST_REMOVE(ncp, nc_vhash);
453 			ncp->nc_vhash.le_prev = NULL;
454 		}
455 	}
456 	hashdone(oldhash1, M_CACHE);
457 	hashdone(oldhash2, M_CACHE);
458 }
459 
460 /*
461  * Cache flush, a particular vnode; called when a vnode is renamed to
462  * hide entries that would now be invalid
463  */
464 void
465 cache_purge(struct vnode *vp)
466 {
467 	struct namecache *ncp;
468 	struct nchashhead *ncpp;
469 	static u_long nextvnodeid;
470 
471 	vp->v_id = ++nextvnodeid;
472 	if (nextvnodeid != 0)
473 		return;
474 	for (ncpp = &nchashtbl[nchash]; ncpp >= nchashtbl; ncpp--) {
475 		LIST_FOREACH(ncp, ncpp, nc_hash) {
476 			ncp->nc_vpid = 0;
477 			ncp->nc_dvpid = 0;
478 		}
479 	}
480 	vp->v_id = ++nextvnodeid;
481 }
482 
483 /*
484  * Cache flush, a whole filesystem; called when filesys is umounted to
485  * remove entries that would now be invalid.
486  */
487 void
488 cache_purgevfs(struct mount *mp)
489 {
490 	struct namecache *ncp, *nxtcp;
491 
492 	for (ncp = TAILQ_FIRST(&nclruhead); ncp != NULL; ncp = nxtcp) {
493 		nxtcp = TAILQ_NEXT(ncp, nc_lru);
494 		if (ncp->nc_dvp == NULL || ncp->nc_dvp->v_mount != mp) {
495 			continue;
496 		}
497 		/* Free the resources we had. */
498 		ncp->nc_vp = NULL;
499 		ncp->nc_dvp = NULL;
500 		TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
501 		if (ncp->nc_hash.le_prev != NULL) {
502 			LIST_REMOVE(ncp, nc_hash);
503 			ncp->nc_hash.le_prev = NULL;
504 		}
505 		if (ncp->nc_vhash.le_prev != NULL) {
506 			LIST_REMOVE(ncp, nc_vhash);
507 			ncp->nc_vhash.le_prev = NULL;
508 		}
509 		TAILQ_INSERT_HEAD(&nclruhead, ncp, nc_lru);
510 	}
511 }
512 
513 #ifdef DDB
514 void
515 namecache_print(struct vnode *vp, void (*pr)(const char *, ...))
516 {
517 	struct vnode *dvp = NULL;
518 	struct namecache *ncp;
519 
520 	TAILQ_FOREACH(ncp, &nclruhead, nc_lru) {
521 		if (ncp->nc_vp == vp && ncp->nc_vpid == vp->v_id) {
522 			(*pr)("name %.*s\n", ncp->nc_nlen, ncp->nc_name);
523 			dvp = ncp->nc_dvp;
524 		}
525 	}
526 	if (dvp == NULL) {
527 		(*pr)("name not found\n");
528 		return;
529 	}
530 	vp = dvp;
531 	TAILQ_FOREACH(ncp, &nclruhead, nc_lru) {
532 		if (ncp->nc_vp == vp && ncp->nc_vpid == vp->v_id) {
533 			(*pr)("parent %.*s\n", ncp->nc_nlen, ncp->nc_name);
534 		}
535 	}
536 }
537 #endif
538