3 * Coda: an Experimental Distributed File System
6 * Copyright (c) 1987-1998 Carnegie Mellon University
9 * Permission to use, copy, modify and distribute this software and its
10 * documentation is hereby granted, provided that both the copyright
11 * notice and this permission notice appear in all copies of the
12 * software, derivative works or modified versions, and any portions
13 * thereof, and that both notices appear in supporting documentation, and
14 * that credit is given to Carnegie Mellon University in all documents
15 * and publicity pertaining to direct or indirect use of this code or its
18 * CODA IS AN EXPERIMENTAL SOFTWARE SYSTEM AND IS KNOWN TO HAVE BUGS,
19 * SOME OF WHICH MAY HAVE SERIOUS CONSEQUENCES. CARNEGIE MELLON ALLOWS
20 * FREE USE OF THIS SOFTWARE IN ITS "AS IS" CONDITION. CARNEGIE MELLON
21 * DISCLAIMS ANY LIABILITY OF ANY KIND FOR ANY DAMAGES WHATSOEVER
22 * RESULTING DIRECTLY OR INDIRECTLY FROM THE USE OF THIS SOFTWARE OR OF
23 * ANY DERIVATIVE WORK.
25 * Carnegie Mellon encourages users of this software to return any
26 * improvements or extensions that they make, and to grant Carnegie
27 * Mellon the rights to redistribute these changes without encumbrance.
29 * @(#) src/sys/coda/coda_namecache.c,v 1.1.1.1 1998/08/29 21:14:52 rvb Exp $
30 * $FreeBSD: src/sys/coda/coda_namecache.c,v 1.10 1999/08/28 00:40:53 peter Exp $
31 * $DragonFly: src/sys/vfs/coda/Attic/coda_namecache.c,v 1.3 2003/06/25 03:55:44 dillon Exp $
36 * Mach Operating System
37 * Copyright (c) 1990 Carnegie-Mellon University
38 * Copyright (c) 1989 Carnegie-Mellon University
39 * All rights reserved. The CMU software License Agreement specifies
40 * the terms and conditions for use and redistribution.
44 * This code was written for the Coda file system at Carnegie Mellon University.
45 * Contributers include David Steere, James Kistler, and M. Satyanarayanan.
49 * This module contains the routines to implement the CODA name cache. The
50 * purpose of this cache is to reduce the cost of translating pathnames
51 * into Vice FIDs. Each entry in the cache contains the name of the file,
52 * the vnode (FID) of the parent directory, and the cred structure of the
53 * user accessing the file.
55 * The first time a file is accessed, it is looked up by the local Venus
56 * which first insures that the user has access to the file. In addition
57 * we are guaranteed that Venus will invalidate any name cache entries in
58 * case the user no longer should be able to access the file. For these
59 * reasons we do not need to keep access list information as well as a
60 * cred structure for each entry.
62 * The table can be accessed through the routines cnc_init(), cnc_enter(),
63 * cnc_lookup(), cnc_rmfidcred(), cnc_rmfid(), cnc_rmcred(), and cnc_purge().
64 * There are several other routines which aid in the implementation of the
70 * 1. The name cache holds a reference to every vnode in it. Hence files can not be
71 * closed or made inactive until they are released.
72 * 2. coda_nc_name(cp) was added to get a name for a cnode pointer for debugging.
73 * 3. coda_nc_find() has debug code to detect when entries are stored with different
74 * credentials. We don't understand yet, if/how entries are NOT EQ but still
76 * 4. I wonder if this name cache could be replace by the vnode name cache.
77 * The latter has no zapping functions, so probably not.
80 #include <sys/param.h>
81 #include <sys/errno.h>
82 #include <sys/malloc.h>
83 #include <sys/ucred.h>
84 #include <sys/select.h>
87 #include <sys/systm.h>
91 #include <vm/vm_object.h>
93 #include <coda/coda.h>
94 #include <coda/cnode.h>
95 #include <coda/coda_namecache.h>
98 #include <coda/coda_vnops.h>
102 * Declaration of the name cache data structure.
105 int coda_nc_use = 1; /* Indicate use of CODA Name Cache */
106 int coda_nc_size = CODA_NC_CACHESIZE; /* size of the cache */
107 int coda_nc_hashsize = CODA_NC_HASHSIZE; /* size of the primary hash */
109 struct coda_cache *coda_nc_heap; /* pointer to the cache entries */
110 struct coda_hash *coda_nc_hash; /* hash table of coda_cache pointers */
111 struct coda_lru coda_nc_lru; /* head of lru chain */
113 struct coda_nc_statistics coda_nc_stat; /* Keep various stats */
116 * for testing purposes
118 int coda_nc_debug = 0;
121 * Entry points for the CODA Name Cache
123 static struct coda_cache *coda_nc_find(struct cnode *dcp, const char *name, int namelen,
124 struct ucred *cred, int hash);
125 static void coda_nc_remove(struct coda_cache *cncp, enum dc_status dcstat);
128 * Initialize the cache, the LRU structure and the Hash structure(s)
131 #define TOTAL_CACHE_SIZE (sizeof(struct coda_cache) * coda_nc_size)
132 #define TOTAL_HASH_SIZE (sizeof(struct coda_hash) * coda_nc_hashsize)
134 int coda_nc_initialized = 0; /* Initially the cache has not been initialized */
141 /* zero the statistics structure */
143 bzero(&coda_nc_stat, (sizeof(struct coda_nc_statistics)));
146 printf("CODA NAME CACHE: CACHE %d, HASH TBL %d\n", CODA_NC_CACHESIZE, CODA_NC_HASHSIZE);
148 CODA_ALLOC(coda_nc_heap, struct coda_cache *, TOTAL_CACHE_SIZE);
149 CODA_ALLOC(coda_nc_hash, struct coda_hash *, TOTAL_HASH_SIZE);
151 coda_nc_lru.lru_next =
152 coda_nc_lru.lru_prev = (struct coda_cache *)LRU_PART(&coda_nc_lru);
155 for (i=0; i < coda_nc_size; i++) { /* initialize the heap */
156 CODA_NC_LRUINS(&coda_nc_heap[i], &coda_nc_lru);
157 CODA_NC_HSHNUL(&coda_nc_heap[i]);
158 coda_nc_heap[i].cp = coda_nc_heap[i].dcp = (struct cnode *)0;
161 for (i=0; i < coda_nc_hashsize; i++) { /* initialize the hashtable */
162 CODA_NC_HSHNUL((struct coda_cache *)&coda_nc_hash[i]);
165 coda_nc_initialized++;
169 * Auxillary routines -- shouldn't be entry points
172 static struct coda_cache *
173 coda_nc_find(dcp, name, namelen, cred, hash)
181 * hash to find the appropriate bucket, look through the chain
182 * for the right entry (especially right cred, unless cred == 0)
184 struct coda_cache *cncp;
187 CODA_NC_DEBUG(CODA_NC_FIND,
188 myprintf(("coda_nc_find(dcp %p, name %s, len %d, cred %p, hash %d\n",
189 dcp, name, namelen, cred, hash));)
191 for (cncp = coda_nc_hash[hash].hash_next;
192 cncp != (struct coda_cache *)&coda_nc_hash[hash];
193 cncp = cncp->hash_next, count++)
196 if ((CODA_NAMEMATCH(cncp, name, namelen, dcp)) &&
197 ((cred == 0) || (cncp->cred == cred)))
199 /* compare cr_uid instead */
200 coda_nc_stat.Search_len += count;
204 else if (CODA_NAMEMATCH(cncp, name, namelen, dcp)) {
205 printf("coda_nc_find: name %s, new cred = %p, cred = %p\n",
206 name, cred, cncp->cred);
207 printf("nref %d, nuid %d, ngid %d // oref %d, ocred %d, ogid %d\n",
208 cred->cr_ref, cred->cr_uid, cred->cr_gid,
209 cncp->cred->cr_ref, cncp->cred->cr_uid, cncp->cred->cr_gid);
211 print_cred(cncp->cred);
216 return((struct coda_cache *)0);
220 * Enter a new (dir cnode, name) pair into the cache, updating the
221 * LRU and Hash as needed.
224 coda_nc_enter(dcp, name, namelen, cred, cp)
231 struct coda_cache *cncp;
234 if (coda_nc_use == 0) /* Cache is off */
237 CODA_NC_DEBUG(CODA_NC_ENTER,
238 myprintf(("Enter: dcp %p cp %p name %s cred %p \n",
239 dcp, cp, name, cred)); )
241 if (namelen > CODA_NC_NAMELEN) {
242 CODA_NC_DEBUG(CODA_NC_ENTER,
243 myprintf(("long name enter %s\n",name));)
244 coda_nc_stat.long_name_enters++; /* record stats */
248 hash = CODA_NC_HASH(name, namelen, dcp);
249 cncp = coda_nc_find(dcp, name, namelen, cred, hash);
250 if (cncp != (struct coda_cache *) 0) {
251 coda_nc_stat.dbl_enters++; /* duplicate entry */
255 coda_nc_stat.enters++; /* record the enters statistic */
257 /* Grab the next element in the lru chain */
258 cncp = CODA_NC_LRUGET(coda_nc_lru);
260 CODA_NC_LRUREM(cncp); /* remove it from the lists */
262 if (CODA_NC_VALID(cncp)) {
263 /* Seems really ugly, but we have to decrement the appropriate
264 hash bucket length here, so we have to find the hash bucket
266 coda_nc_hash[CODA_NC_HASH(cncp->name, cncp->namelen, cncp->dcp)].length--;
268 coda_nc_stat.lru_rm++; /* zapped a valid entry */
269 CODA_NC_HSHREM(cncp);
270 vrele(CTOV(cncp->dcp));
271 vrele(CTOV(cncp->cp));
276 * Put a hold on the current vnodes and fill in the cache entry.
283 cncp->namelen = namelen;
286 bcopy(name, cncp->name, (unsigned)namelen);
288 /* Insert into the lru and hash chains. */
290 CODA_NC_LRUINS(cncp, &coda_nc_lru);
291 CODA_NC_HSHINS(cncp, &coda_nc_hash[hash]);
292 coda_nc_hash[hash].length++; /* Used for tuning */
294 CODA_NC_DEBUG(CODA_NC_PRINTCODA_NC, print_coda_nc(); )
298 * Find the (dir cnode, name) pair in the cache, if it's cred
299 * matches the input, return it, otherwise return 0
302 coda_nc_lookup(dcp, name, namelen, cred)
309 struct coda_cache *cncp;
311 if (coda_nc_use == 0) /* Cache is off */
312 return((struct cnode *) 0);
314 if (namelen > CODA_NC_NAMELEN) {
315 CODA_NC_DEBUG(CODA_NC_LOOKUP,
316 myprintf(("long name lookup %s\n",name));)
317 coda_nc_stat.long_name_lookups++; /* record stats */
318 return((struct cnode *) 0);
321 /* Use the hash function to locate the starting point,
322 then the search routine to go down the list looking for
326 hash = CODA_NC_HASH(name, namelen, dcp);
327 cncp = coda_nc_find(dcp, name, namelen, cred, hash);
328 if (cncp == (struct coda_cache *) 0) {
329 coda_nc_stat.misses++; /* record miss */
330 return((struct cnode *) 0);
335 /* put this entry at the end of the LRU */
336 CODA_NC_LRUREM(cncp);
337 CODA_NC_LRUINS(cncp, &coda_nc_lru);
339 /* move it to the front of the hash chain */
340 /* don't need to change the hash bucket length */
341 CODA_NC_HSHREM(cncp);
342 CODA_NC_HSHINS(cncp, &coda_nc_hash[hash]);
344 CODA_NC_DEBUG(CODA_NC_LOOKUP,
345 printf("lookup: dcp %p, name %s, cred %p = cp %p\n",
346 dcp, name, cred, cncp->cp); )
352 coda_nc_remove(cncp, dcstat)
353 struct coda_cache *cncp;
354 enum dc_status dcstat;
357 * remove an entry -- vrele(cncp->dcp, cp), crfree(cred),
358 * remove it from it's hash chain, and
359 * place it at the head of the lru list.
361 CODA_NC_DEBUG(CODA_NC_REMOVE,
362 myprintf(("coda_nc_remove %s from parent %lx.%lx.%lx\n",
363 cncp->name, (cncp->dcp)->c_fid.Volume,
364 (cncp->dcp)->c_fid.Vnode, (cncp->dcp)->c_fid.Unique));)
366 CODA_NC_HSHREM(cncp);
368 CODA_NC_HSHNUL(cncp); /* have it be a null chain */
369 if ((dcstat == IS_DOWNCALL) && (CTOV(cncp->dcp)->v_usecount == 1)) {
370 cncp->dcp->c_flags |= C_PURGING;
372 vrele(CTOV(cncp->dcp));
374 if ((dcstat == IS_DOWNCALL) && (CTOV(cncp->cp)->v_usecount == 1)) {
375 cncp->cp->c_flags |= C_PURGING;
377 vrele(CTOV(cncp->cp));
380 bzero(DATA_PART(cncp),DATA_SIZE);
382 /* Put the null entry just after the least-recently-used entry */
383 /* LRU_TOP adjusts the pointer to point to the top of the structure. */
384 CODA_NC_LRUREM(cncp);
385 CODA_NC_LRUINS(cncp, LRU_TOP(coda_nc_lru.lru_prev));
389 * Remove all entries with a parent which has the input fid.
392 coda_nc_zapParentfid(fid, dcstat)
394 enum dc_status dcstat;
396 /* To get to a specific fid, we might either have another hashing
397 function or do a sequential search through the cache for the
398 appropriate entries. The later may be acceptable since I don't
399 think callbacks or whatever Case 1 covers are frequent occurences.
401 struct coda_cache *cncp, *ncncp;
404 if (coda_nc_use == 0) /* Cache is off */
407 CODA_NC_DEBUG(CODA_NC_ZAPPFID,
408 myprintf(("ZapParent: fid 0x%lx, 0x%lx, 0x%lx \n",
409 fid->Volume, fid->Vnode, fid->Unique)); )
411 coda_nc_stat.zapPfids++;
413 for (i = 0; i < coda_nc_hashsize; i++) {
416 * Need to save the hash_next pointer in case we remove the
417 * entry. remove causes hash_next to point to itself.
420 for (cncp = coda_nc_hash[i].hash_next;
421 cncp != (struct coda_cache *)&coda_nc_hash[i];
423 ncncp = cncp->hash_next;
424 if ((cncp->dcp->c_fid.Volume == fid->Volume) &&
425 (cncp->dcp->c_fid.Vnode == fid->Vnode) &&
426 (cncp->dcp->c_fid.Unique == fid->Unique)) {
427 coda_nc_hash[i].length--; /* Used for tuning */
428 coda_nc_remove(cncp, dcstat);
436 * Remove all entries which have the same fid as the input
439 coda_nc_zapfid(fid, dcstat)
441 enum dc_status dcstat;
443 /* See comment for zapParentfid. This routine will be used
444 if attributes are being cached.
446 struct coda_cache *cncp, *ncncp;
449 if (coda_nc_use == 0) /* Cache is off */
452 CODA_NC_DEBUG(CODA_NC_ZAPFID,
453 myprintf(("Zapfid: fid 0x%lx, 0x%lx, 0x%lx \n",
454 fid->Volume, fid->Vnode, fid->Unique)); )
456 coda_nc_stat.zapFids++;
458 for (i = 0; i < coda_nc_hashsize; i++) {
459 for (cncp = coda_nc_hash[i].hash_next;
460 cncp != (struct coda_cache *)&coda_nc_hash[i];
462 ncncp = cncp->hash_next;
463 if ((cncp->cp->c_fid.Volume == fid->Volume) &&
464 (cncp->cp->c_fid.Vnode == fid->Vnode) &&
465 (cncp->cp->c_fid.Unique == fid->Unique)) {
466 coda_nc_hash[i].length--; /* Used for tuning */
467 coda_nc_remove(cncp, dcstat);
474 * Remove all entries which match the fid and the cred
477 coda_nc_zapvnode(fid, cred, dcstat)
480 enum dc_status dcstat;
482 /* See comment for zapfid. I don't think that one would ever
483 want to zap a file with a specific cred from the kernel.
484 We'll leave this one unimplemented.
486 if (coda_nc_use == 0) /* Cache is off */
489 CODA_NC_DEBUG(CODA_NC_ZAPVNODE,
490 myprintf(("Zapvnode: fid 0x%lx, 0x%lx, 0x%lx cred %p\n",
491 fid->Volume, fid->Vnode, fid->Unique, cred)); )
496 * Remove all entries which have the (dir vnode, name) pair
499 coda_nc_zapfile(dcp, name, namelen)
504 /* use the hash function to locate the file, then zap all
505 entries of it regardless of the cred.
507 struct coda_cache *cncp;
510 if (coda_nc_use == 0) /* Cache is off */
513 CODA_NC_DEBUG(CODA_NC_ZAPFILE,
514 myprintf(("Zapfile: dcp %p name %s \n",
517 if (namelen > CODA_NC_NAMELEN) {
518 coda_nc_stat.long_remove++; /* record stats */
522 coda_nc_stat.zapFile++;
524 hash = CODA_NC_HASH(name, namelen, dcp);
525 cncp = coda_nc_find(dcp, name, namelen, 0, hash);
528 coda_nc_hash[hash].length--; /* Used for tuning */
530 coda_nc_remove(cncp, NOT_DOWNCALL);
531 cncp = coda_nc_find(dcp, name, namelen, 0, hash);
536 * Remove all the entries for a particular user. Used when tokens expire.
537 * A user is determined by his/her effective user id (id_uid).
540 coda_nc_purge_user(uid, dcstat)
542 enum dc_status dcstat;
545 * I think the best approach is to go through the entire cache
546 * via HASH or whatever and zap all entries which match the
547 * input cred. Or just flush the whole cache. It might be
548 * best to go through on basis of LRU since cache will almost
549 * always be full and LRU is more straightforward.
552 struct coda_cache *cncp, *ncncp;
555 if (coda_nc_use == 0) /* Cache is off */
558 CODA_NC_DEBUG(CODA_NC_PURGEUSER,
559 myprintf(("ZapDude: uid %x\n", uid)); )
560 coda_nc_stat.zapUsers++;
562 for (cncp = CODA_NC_LRUGET(coda_nc_lru);
563 cncp != (struct coda_cache *)(&coda_nc_lru);
565 ncncp = CODA_NC_LRUGET(*cncp);
567 if ((CODA_NC_VALID(cncp)) &&
568 ((cncp->cred)->cr_uid == uid)) {
569 /* Seems really ugly, but we have to decrement the appropriate
570 hash bucket length here, so we have to find the hash bucket
572 hash = CODA_NC_HASH(cncp->name, cncp->namelen, cncp->dcp);
573 coda_nc_hash[hash].length--; /* For performance tuning */
575 coda_nc_remove(cncp, dcstat);
581 * Flush the entire name cache. In response to a flush of the Venus cache.
584 coda_nc_flush(dcstat)
585 enum dc_status dcstat;
587 /* One option is to deallocate the current name cache and
588 call init to start again. Or just deallocate, then rebuild.
589 Or again, we could just go through the array and zero the
594 * Go through the whole lru chain and kill everything as we go.
595 * I don't use remove since that would rebuild the lru chain
596 * as it went and that seemed unneccesary.
598 struct coda_cache *cncp;
601 if (coda_nc_use == 0) /* Cache is off */
604 coda_nc_stat.Flushes++;
606 for (cncp = CODA_NC_LRUGET(coda_nc_lru);
607 cncp != (struct coda_cache *)&coda_nc_lru;
608 cncp = CODA_NC_LRUGET(*cncp)) {
609 if (CODA_NC_VALID(cncp)) {
611 CODA_NC_HSHREM(cncp); /* only zero valid nodes */
612 CODA_NC_HSHNUL(cncp);
613 if ((dcstat == IS_DOWNCALL)
614 && (CTOV(cncp->dcp)->v_usecount == 1))
616 cncp->dcp->c_flags |= C_PURGING;
618 vrele(CTOV(cncp->dcp));
620 if (CTOV(cncp->cp)->v_flag & VTEXT) {
621 if (coda_vmflush(cncp->cp))
622 CODADEBUG(CODA_FLUSH,
623 myprintf(("coda_nc_flush: (%lx.%lx.%lx) busy\n", cncp->cp->c_fid.Volume, cncp->cp->c_fid.Vnode, cncp->cp->c_fid.Unique)); )
626 if ((dcstat == IS_DOWNCALL)
627 && (CTOV(cncp->cp)->v_usecount == 1))
629 cncp->cp->c_flags |= C_PURGING;
631 vrele(CTOV(cncp->cp));
634 bzero(DATA_PART(cncp),DATA_SIZE);
638 for (i = 0; i < coda_nc_hashsize; i++)
639 coda_nc_hash[i].length = 0;
647 * This routine should print out all the hash chains to the console.
653 struct coda_cache *cncp;
655 for (hash = 0; hash < coda_nc_hashsize; hash++) {
656 myprintf(("\nhash %d\n",hash));
658 for (cncp = coda_nc_hash[hash].hash_next;
659 cncp != (struct coda_cache *)&coda_nc_hash[hash];
660 cncp = cncp->hash_next) {
661 myprintf(("cp %p dcp %p cred %p name %s\n",
663 cncp->cred, cncp->name));
669 coda_nc_gather_stats(void)
671 int i, max = 0, sum = 0, temp, zeros = 0, ave, n;
673 for (i = 0; i < coda_nc_hashsize; i++) {
674 if (coda_nc_hash[i].length) {
675 sum += coda_nc_hash[i].length;
680 if (coda_nc_hash[i].length > max)
681 max = coda_nc_hash[i].length;
685 * When computing the Arithmetic mean, only count slots which
686 * are not empty in the distribution.
688 coda_nc_stat.Sum_bucket_len = sum;
689 coda_nc_stat.Num_zero_len = zeros;
690 coda_nc_stat.Max_bucket_len = max;
692 if ((n = coda_nc_hashsize - zeros) > 0)
698 for (i = 0; i < coda_nc_hashsize; i++) {
699 if (coda_nc_hash[i].length) {
700 temp = coda_nc_hash[i].length - ave;
704 coda_nc_stat.Sum2_bucket_len = sum;
708 * The purpose of this routine is to allow the hash and cache sizes to be
709 * changed dynamically. This should only be used in controlled environments,
710 * it makes no effort to lock other users from accessing the cache while it
711 * is in an improper state (except by turning the cache off).
714 coda_nc_resize(hashsize, heapsize, dcstat)
715 int hashsize, heapsize;
716 enum dc_status dcstat;
718 if ((hashsize % 2) || (heapsize % 2)) { /* Illegal hash or cache sizes */
722 coda_nc_use = 0; /* Turn the cache off */
724 coda_nc_flush(dcstat); /* free any cnodes in the cache */
726 /* WARNING: free must happen *before* size is reset */
727 CODA_FREE(coda_nc_heap,TOTAL_CACHE_SIZE);
728 CODA_FREE(coda_nc_hash,TOTAL_HASH_SIZE);
730 coda_nc_hashsize = hashsize;
731 coda_nc_size = heapsize;
733 coda_nc_init(); /* Set up a cache with the new size */
735 coda_nc_use = 1; /* Turn the cache back on */
740 char coda_nc_name_buf[CODA_MAXNAMLEN+1];
743 coda_nc_name(struct cnode *cp)
745 struct coda_cache *cncp, *ncncp;
748 if (coda_nc_use == 0) /* Cache is off */
751 for (i = 0; i < coda_nc_hashsize; i++) {
752 for (cncp = coda_nc_hash[i].hash_next;
753 cncp != (struct coda_cache *)&coda_nc_hash[i];
755 ncncp = cncp->hash_next;
756 if (cncp->cp == cp) {
757 bcopy(cncp->name, coda_nc_name_buf, cncp->namelen);
758 coda_nc_name_buf[cncp->namelen] = 0;
759 printf(" is %s (%p,%p)@%p",
760 coda_nc_name_buf, cncp->cp, cncp->dcp, cncp);