xref: /netbsd-src/sys/kern/vfs_cache.c (revision 3b01aba77a7a698587faaae455bbfe740923c1f5)
1 /*	$NetBSD: vfs_cache.c,v 1.29 2001/03/29 22:39:23 fvdl 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 "opt_ddb.h"
39 #include "opt_revcache.h"
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/time.h>
44 #include <sys/mount.h>
45 #include <sys/vnode.h>
46 #include <sys/namei.h>
47 #include <sys/errno.h>
48 #include <sys/malloc.h>
49 #include <sys/pool.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 
79 LIST_HEAD(ncvhashhead, namecache) *ncvhashtbl;
80 u_long	ncvhash;			/* size of hash table - 1 */
81 
82 TAILQ_HEAD(, namecache) nclruhead;		/* LRU chain */
83 struct	nchstats nchstats;		/* cache effectiveness statistics */
84 
85 struct pool namecache_pool;
86 
87 int doingcache = 1;			/* 1 => enable the cache */
88 
89 /*
90  * Look for a the name in the cache. We don't do this
91  * if the segment name is long, simply so the cache can avoid
92  * holding long names (which would either waste space, or
93  * add greatly to the complexity).
94  *
95  * Lookup is called with ni_dvp pointing to the directory to search,
96  * ni_ptr pointing to the name of the entry being sought, ni_namelen
97  * tells the length of the name, and ni_hash contains a hash of
98  * the name. If the lookup succeeds, the vnode is locked, stored in ni_vp
99  * and a status of zero is returned. If the locking fails for whatever
100  * reason, the vnode is unlocked and the error is returned to caller.
101  * If the lookup determines that the name does not exist (negative cacheing),
102  * a status of ENOENT is returned. If the lookup fails, a status of -1
103  * is returned.
104  */
105 int
106 cache_lookup(dvp, vpp, cnp)
107 	struct vnode *dvp;
108 	struct vnode **vpp;
109 	struct componentname *cnp;
110 {
111 	struct namecache *ncp;
112 	struct nchashhead *ncpp;
113 	struct vnode *vp;
114 	int vpid, error;
115 
116 	if (!doingcache) {
117 		cnp->cn_flags &= ~MAKEENTRY;
118 		*vpp = 0;
119 		return (-1);
120 	}
121 	if (cnp->cn_namelen > NCHNAMLEN) {
122 		nchstats.ncs_long++;
123 		cnp->cn_flags &= ~MAKEENTRY;
124 		*vpp = 0;
125 		return (-1);
126 	}
127 	ncpp = &nchashtbl[(cnp->cn_hash ^ dvp->v_id) & nchash];
128 	LIST_FOREACH(ncp, ncpp, nc_hash) {
129 		if (ncp->nc_dvp == dvp &&
130 		    ncp->nc_dvpid == dvp->v_id &&
131 		    ncp->nc_nlen == cnp->cn_namelen &&
132 		    !memcmp(ncp->nc_name, cnp->cn_nameptr, (u_int)ncp->nc_nlen))
133 			break;
134 	}
135 	if (ncp == 0) {
136 		nchstats.ncs_miss++;
137 		*vpp = 0;
138 		return (-1);
139 	}
140 	if ((cnp->cn_flags & MAKEENTRY) == 0) {
141 		nchstats.ncs_badhits++;
142 		goto remove;
143 	} else if (ncp->nc_vp == NULL) {
144 		/*
145 		 * Restore the ISWHITEOUT flag saved earlier.
146 		 */
147 		cnp->cn_flags |= ncp->nc_vpid;
148 		if (cnp->cn_nameiop != CREATE ||
149 		    (cnp->cn_flags & ISLASTCN) == 0) {
150 			nchstats.ncs_neghits++;
151 			/*
152 			 * Move this slot to end of LRU chain,
153 			 * if not already there.
154 			 */
155 			if (ncp->nc_lru.tqe_next != 0) {
156 				TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
157 				TAILQ_INSERT_TAIL(&nclruhead, ncp, nc_lru);
158 			}
159 			return (ENOENT);
160 		} else {
161 			nchstats.ncs_badhits++;
162 			goto remove;
163 		}
164 	} else if (ncp->nc_vpid != ncp->nc_vp->v_id) {
165 		nchstats.ncs_falsehits++;
166 		goto remove;
167 	}
168 
169 	vp = ncp->nc_vp;
170 	vpid = vp->v_id;
171 	if (vp == dvp) {	/* lookup on "." */
172 		VREF(dvp);
173 		error = 0;
174 	} else if (cnp->cn_flags & ISDOTDOT) {
175 		VOP_UNLOCK(dvp, 0);
176 		cnp->cn_flags |= PDIRUNLOCK;
177 		error = vget(vp, LK_EXCLUSIVE);
178 		/* if the above vget() succeeded and both LOCKPARENT and
179 		 * ISLASTCN is set, lock the directory vnode as well */
180 		if (!error && (~cnp->cn_flags & (LOCKPARENT|ISLASTCN)) == 0) {
181 			if ((error = vn_lock(dvp, LK_EXCLUSIVE)) != 0) {
182 				vput(vp);
183 				return (error);
184 			}
185 			cnp->cn_flags &= ~PDIRUNLOCK;
186 		}
187 	} else {
188 		error = vget(vp, LK_EXCLUSIVE);
189 		/* if the above vget() failed or either of LOCKPARENT or
190 		 * ISLASTCN is set, unlock the directory vnode */
191 		if (error || (~cnp->cn_flags & (LOCKPARENT|ISLASTCN)) != 0) {
192 			VOP_UNLOCK(dvp, 0);
193 			cnp->cn_flags |= PDIRUNLOCK;
194 		}
195 	}
196 
197 	/*
198 	 * Check that the lock succeeded, and that the capability number did
199 	 * not change while we were waiting for the lock.
200 	 */
201 	if (error || vpid != vp->v_id) {
202 		if (!error) {
203 			vput(vp);
204 			nchstats.ncs_falsehits++;
205 		} else
206 			nchstats.ncs_badhits++;
207 		/*
208 		 * The parent needs to be locked when we return to VOP_LOOKUP().
209 		 * The `.' case here should be extremely rare (if it can happen
210 		 * at all), so we don't bother optimizing out the unlock/relock.
211 		 */
212 		if (vp == dvp ||
213 		    error || (~cnp->cn_flags & (LOCKPARENT|ISLASTCN)) != 0) {
214 			if ((error = vn_lock(dvp, LK_EXCLUSIVE)) != 0)
215 				return (error);
216 			cnp->cn_flags &= ~PDIRUNLOCK;
217 		}
218 		goto remove;
219 	}
220 
221 	nchstats.ncs_goodhits++;
222 	/*
223 	 * move this slot to end of LRU chain, if not already there
224 	 */
225 	if (ncp->nc_lru.tqe_next != 0) {
226 		TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
227 		TAILQ_INSERT_TAIL(&nclruhead, ncp, nc_lru);
228 	}
229 	*vpp = vp;
230 	return (0);
231 
232 remove:
233 	/*
234 	 * Last component and we are renaming or deleting,
235 	 * the cache entry is invalid, or otherwise don't
236 	 * want cache entry to exist.
237 	 */
238 	TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
239 	LIST_REMOVE(ncp, nc_hash);
240 	ncp->nc_hash.le_prev = 0;
241 	if (ncp->nc_vhash.le_prev != NULL) {
242 		LIST_REMOVE(ncp, nc_vhash);
243 		ncp->nc_vhash.le_prev = 0;
244 	}
245 	TAILQ_INSERT_HEAD(&nclruhead, ncp, nc_lru);
246 	*vpp = 0;
247 	return (-1);
248 }
249 
250 /*
251  * Scan cache looking for name of directory entry pointing at vp.
252  *
253  * Fill in dvpp.
254  *
255  * If bufp is non-NULL, also place the name in the buffer which starts
256  * at bufp, immediately before *bpp, and move bpp backwards to point
257  * at the start of it.  (Yes, this is a little baroque, but it's done
258  * this way to cater to the whims of getcwd).
259  *
260  * Returns 0 on success, -1 on cache miss, positive errno on failure.
261  */
262 int
263 cache_revlookup (vp, dvpp, bpp, bufp)
264 	struct vnode *vp, **dvpp;
265 	char **bpp;
266 	char *bufp;
267 {
268 	struct namecache *ncp;
269 	struct vnode *dvp;
270 	struct ncvhashhead *nvcpp;
271 
272 	if (!doingcache)
273 		goto out;
274 
275 	nvcpp = &ncvhashtbl[(vp->v_id & ncvhash)];
276 
277 	LIST_FOREACH(ncp, nvcpp, nc_vhash) {
278 		if ((ncp->nc_vp == vp) &&
279 		    (ncp->nc_vpid == vp->v_id) &&
280 		    ((dvp = ncp->nc_dvp) != 0) &&
281 		    (dvp != vp) && 		/* avoid pesky . entries.. */
282 		    (dvp->v_id == ncp->nc_dvpid))
283 		{
284 			char *bp;
285 
286 #ifdef DIAGNOSTIC
287 			if ((ncp->nc_nlen == 1) &&
288 			    (ncp->nc_name[0] == '.'))
289 				panic("cache_revlookup: found entry for .");
290 
291 			if ((ncp->nc_nlen == 2) &&
292 			    (ncp->nc_name[0] == '.') &&
293 			    (ncp->nc_name[1] == '.'))
294 				panic("cache_revlookup: found entry for ..");
295 #endif
296 			nchstats.ncs_revhits++;
297 
298 			if (bufp) {
299 				bp = *bpp;
300 				bp -= ncp->nc_nlen;
301 				if (bp <= bufp) {
302 					*dvpp = 0;
303 					return ERANGE;
304 				}
305 				memcpy(bp, ncp->nc_name, ncp->nc_nlen);
306 				*bpp = bp;
307 			}
308 
309 			/* XXX MP: how do we know dvp won't evaporate? */
310 			*dvpp = dvp;
311 			return 0;
312 		}
313 	}
314 	nchstats.ncs_revmiss++;
315  out:
316 	*dvpp = 0;
317 	return -1;
318 }
319 
320 /*
321  * Add an entry to the cache
322  */
323 void
324 cache_enter(dvp, vp, cnp)
325 	struct vnode *dvp;
326 	struct vnode *vp;
327 	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 != 0) {
349 			LIST_REMOVE(ncp, nc_hash);
350 			ncp->nc_hash.le_prev = 0;
351 		}
352 		if (ncp->nc_vhash.le_prev != 0) {
353 			LIST_REMOVE(ncp, nc_vhash);
354 			ncp->nc_vhash.le_prev = 0;
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[(cnp->cn_hash ^ dvp->v_id) & nchash];
376 	LIST_INSERT_HEAD(ncpp, ncp, nc_hash);
377 
378 	ncp->nc_vhash.le_prev = 0;
379 	ncp->nc_vhash.le_next = 0;
380 
381 	/*
382 	 * Create reverse-cache entries (used in getcwd) for directories.
383 	 */
384 	if (vp &&
385 	    (vp != dvp) &&
386 #ifndef NAMECACHE_ENTER_REVERSE
387 	    (vp->v_type == VDIR) &&
388 #endif
389 	    ((ncp->nc_nlen > 2) ||
390 	     ((ncp->nc_nlen == 2) && (ncp->nc_name[0] != '.') && (ncp->nc_name[1] != '.')) ||
391 	     ((ncp->nc_nlen == 1) && (ncp->nc_name[0] != '.'))))
392 	{
393 		nvcpp = &ncvhashtbl[(vp->v_id & ncvhash)];
394 		LIST_INSERT_HEAD(nvcpp, ncp, nc_vhash);
395 	}
396 
397 }
398 
399 /*
400  * Name cache initialization, from vfs_init() when we are booting
401  */
402 void
403 nchinit()
404 {
405 
406 	TAILQ_INIT(&nclruhead);
407 	nchashtbl =
408 	    hashinit(desiredvnodes, HASH_LIST, M_CACHE, M_WAITOK, &nchash);
409 	ncvhashtbl =
410 #ifdef NAMECACHE_ENTER_REVERSE
411 	    hashinit(desiredvnodes, HASH_LIST, M_CACHE, M_WAITOK, &ncvhash);
412 #else
413 	    hashinit(desiredvnodes/8, HASH_LIST, M_CACHE, M_WAITOK, &ncvhash);
414 #endif
415 	pool_init(&namecache_pool, sizeof(struct namecache), 0, 0, 0,
416 	    "ncachepl", 0, pool_page_alloc_nointr, pool_page_free_nointr,
417 	    M_CACHE);
418 }
419 
420 /*
421  * Cache flush, a particular vnode; called when a vnode is renamed to
422  * hide entries that would now be invalid
423  */
424 void
425 cache_purge(vp)
426 	struct vnode *vp;
427 {
428 	struct namecache *ncp;
429 	struct nchashhead *ncpp;
430 	static u_long nextvnodeid;
431 
432 	vp->v_id = ++nextvnodeid;
433 	if (nextvnodeid != 0)
434 		return;
435 	for (ncpp = &nchashtbl[nchash]; ncpp >= nchashtbl; ncpp--) {
436 		LIST_FOREACH(ncp, ncpp, nc_hash) {
437 			ncp->nc_vpid = 0;
438 			ncp->nc_dvpid = 0;
439 		}
440 	}
441 	vp->v_id = ++nextvnodeid;
442 }
443 
444 /*
445  * Cache flush, a whole filesystem; called when filesys is umounted to
446  * remove entries that would now be invalid.
447  */
448 void
449 cache_purgevfs(mp)
450 	struct mount *mp;
451 {
452 	struct namecache *ncp, *nxtcp;
453 
454 	for (ncp = TAILQ_FIRST(&nclruhead); ncp != NULL; ncp = nxtcp) {
455 		nxtcp = TAILQ_NEXT(ncp, nc_lru);
456 		if (ncp->nc_dvp == NULL || ncp->nc_dvp->v_mount != mp) {
457 			continue;
458 		}
459 		/* free the resources we had */
460 		ncp->nc_vp = NULL;
461 		ncp->nc_dvp = NULL;
462 		TAILQ_REMOVE(&nclruhead, ncp, nc_lru);
463 		if (ncp->nc_hash.le_prev != 0) {
464 			LIST_REMOVE(ncp, nc_hash);
465 			ncp->nc_hash.le_prev = 0;
466 		}
467 		if (ncp->nc_vhash.le_prev != 0) {
468 			LIST_REMOVE(ncp, nc_vhash);
469 			ncp->nc_vhash.le_prev = 0;
470 		}
471 		TAILQ_INSERT_HEAD(&nclruhead, ncp, nc_lru);
472 	}
473 }
474 
475 #ifdef DDB
476 void
477 namecache_print(struct vnode *vp, void (*pr)(const char *, ...))
478 {
479 	struct vnode *dvp = NULL;
480 	struct namecache *ncp;
481 
482 	TAILQ_FOREACH(ncp, &nclruhead, nc_lru) {
483 		if (ncp->nc_vp == vp && ncp->nc_vpid == vp->v_id) {
484 			(*pr)("name %.*s\n", ncp->nc_nlen, ncp->nc_name);
485 			dvp = ncp->nc_dvp;
486 		}
487 	}
488 	if (dvp == NULL) {
489 		(*pr)("name not found\n");
490 		return;
491 	}
492 	vp = dvp;
493 	TAILQ_FOREACH(ncp, &nclruhead, nc_lru) {
494 		if (ncp->nc_vp == vp && ncp->nc_vpid == vp->v_id) {
495 			(*pr)("parent %.*s\n", ncp->nc_nlen, ncp->nc_name);
496 		}
497 	}
498 }
499 #endif
500