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 1.5 2003/08/07 21:17:40 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>
95 #include "coda_namecache.h"
98 #include "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.
282 cncp->namelen = namelen;
283 cncp->cred = crhold(cred);
285 bcopy(name, cncp->name, (unsigned)namelen);
287 /* Insert into the lru and hash chains. */
289 CODA_NC_LRUINS(cncp, &coda_nc_lru);
290 CODA_NC_HSHINS(cncp, &coda_nc_hash[hash]);
291 coda_nc_hash[hash].length++; /* Used for tuning */
293 CODA_NC_DEBUG(CODA_NC_PRINTCODA_NC, print_coda_nc(); )
297 * Find the (dir cnode, name) pair in the cache, if it's cred
298 * matches the input, return it, otherwise return 0
301 coda_nc_lookup(dcp, name, namelen, cred)
308 struct coda_cache *cncp;
310 if (coda_nc_use == 0) /* Cache is off */
311 return((struct cnode *) 0);
313 if (namelen > CODA_NC_NAMELEN) {
314 CODA_NC_DEBUG(CODA_NC_LOOKUP,
315 myprintf(("long name lookup %s\n",name));)
316 coda_nc_stat.long_name_lookups++; /* record stats */
317 return((struct cnode *) 0);
320 /* Use the hash function to locate the starting point,
321 then the search routine to go down the list looking for
325 hash = CODA_NC_HASH(name, namelen, dcp);
326 cncp = coda_nc_find(dcp, name, namelen, cred, hash);
327 if (cncp == (struct coda_cache *) 0) {
328 coda_nc_stat.misses++; /* record miss */
329 return((struct cnode *) 0);
334 /* put this entry at the end of the LRU */
335 CODA_NC_LRUREM(cncp);
336 CODA_NC_LRUINS(cncp, &coda_nc_lru);
338 /* move it to the front of the hash chain */
339 /* don't need to change the hash bucket length */
340 CODA_NC_HSHREM(cncp);
341 CODA_NC_HSHINS(cncp, &coda_nc_hash[hash]);
343 CODA_NC_DEBUG(CODA_NC_LOOKUP,
344 printf("lookup: dcp %p, name %s, cred %p = cp %p\n",
345 dcp, name, cred, cncp->cp); )
351 coda_nc_remove(cncp, dcstat)
352 struct coda_cache *cncp;
353 enum dc_status dcstat;
356 * remove an entry -- vrele(cncp->dcp, cp), crfree(cred),
357 * remove it from it's hash chain, and
358 * place it at the head of the lru list.
360 CODA_NC_DEBUG(CODA_NC_REMOVE,
361 myprintf(("coda_nc_remove %s from parent %lx.%lx.%lx\n",
362 cncp->name, (cncp->dcp)->c_fid.Volume,
363 (cncp->dcp)->c_fid.Vnode, (cncp->dcp)->c_fid.Unique));)
365 CODA_NC_HSHREM(cncp);
367 CODA_NC_HSHNUL(cncp); /* have it be a null chain */
368 if ((dcstat == IS_DOWNCALL) && (CTOV(cncp->dcp)->v_usecount == 1)) {
369 cncp->dcp->c_flags |= C_PURGING;
371 vrele(CTOV(cncp->dcp));
373 if ((dcstat == IS_DOWNCALL) && (CTOV(cncp->cp)->v_usecount == 1)) {
374 cncp->cp->c_flags |= C_PURGING;
376 vrele(CTOV(cncp->cp));
379 bzero(DATA_PART(cncp),DATA_SIZE);
381 /* Put the null entry just after the least-recently-used entry */
382 /* LRU_TOP adjusts the pointer to point to the top of the structure. */
383 CODA_NC_LRUREM(cncp);
384 CODA_NC_LRUINS(cncp, LRU_TOP(coda_nc_lru.lru_prev));
388 * Remove all entries with a parent which has the input fid.
391 coda_nc_zapParentfid(fid, dcstat)
393 enum dc_status dcstat;
395 /* To get to a specific fid, we might either have another hashing
396 function or do a sequential search through the cache for the
397 appropriate entries. The later may be acceptable since I don't
398 think callbacks or whatever Case 1 covers are frequent occurences.
400 struct coda_cache *cncp, *ncncp;
403 if (coda_nc_use == 0) /* Cache is off */
406 CODA_NC_DEBUG(CODA_NC_ZAPPFID,
407 myprintf(("ZapParent: fid 0x%lx, 0x%lx, 0x%lx \n",
408 fid->Volume, fid->Vnode, fid->Unique)); )
410 coda_nc_stat.zapPfids++;
412 for (i = 0; i < coda_nc_hashsize; i++) {
415 * Need to save the hash_next pointer in case we remove the
416 * entry. remove causes hash_next to point to itself.
419 for (cncp = coda_nc_hash[i].hash_next;
420 cncp != (struct coda_cache *)&coda_nc_hash[i];
422 ncncp = cncp->hash_next;
423 if ((cncp->dcp->c_fid.Volume == fid->Volume) &&
424 (cncp->dcp->c_fid.Vnode == fid->Vnode) &&
425 (cncp->dcp->c_fid.Unique == fid->Unique)) {
426 coda_nc_hash[i].length--; /* Used for tuning */
427 coda_nc_remove(cncp, dcstat);
435 * Remove all entries which have the same fid as the input
438 coda_nc_zapfid(fid, dcstat)
440 enum dc_status dcstat;
442 /* See comment for zapParentfid. This routine will be used
443 if attributes are being cached.
445 struct coda_cache *cncp, *ncncp;
448 if (coda_nc_use == 0) /* Cache is off */
451 CODA_NC_DEBUG(CODA_NC_ZAPFID,
452 myprintf(("Zapfid: fid 0x%lx, 0x%lx, 0x%lx \n",
453 fid->Volume, fid->Vnode, fid->Unique)); )
455 coda_nc_stat.zapFids++;
457 for (i = 0; i < coda_nc_hashsize; i++) {
458 for (cncp = coda_nc_hash[i].hash_next;
459 cncp != (struct coda_cache *)&coda_nc_hash[i];
461 ncncp = cncp->hash_next;
462 if ((cncp->cp->c_fid.Volume == fid->Volume) &&
463 (cncp->cp->c_fid.Vnode == fid->Vnode) &&
464 (cncp->cp->c_fid.Unique == fid->Unique)) {
465 coda_nc_hash[i].length--; /* Used for tuning */
466 coda_nc_remove(cncp, dcstat);
473 * Remove all entries which match the fid and the cred
476 coda_nc_zapvnode(fid, cred, dcstat)
479 enum dc_status dcstat;
481 /* See comment for zapfid. I don't think that one would ever
482 want to zap a file with a specific cred from the kernel.
483 We'll leave this one unimplemented.
485 if (coda_nc_use == 0) /* Cache is off */
488 CODA_NC_DEBUG(CODA_NC_ZAPVNODE,
489 myprintf(("Zapvnode: fid 0x%lx, 0x%lx, 0x%lx cred %p\n",
490 fid->Volume, fid->Vnode, fid->Unique, cred)); )
495 * Remove all entries which have the (dir vnode, name) pair
498 coda_nc_zapfile(dcp, name, namelen)
503 /* use the hash function to locate the file, then zap all
504 entries of it regardless of the cred.
506 struct coda_cache *cncp;
509 if (coda_nc_use == 0) /* Cache is off */
512 CODA_NC_DEBUG(CODA_NC_ZAPFILE,
513 myprintf(("Zapfile: dcp %p name %s \n",
516 if (namelen > CODA_NC_NAMELEN) {
517 coda_nc_stat.long_remove++; /* record stats */
521 coda_nc_stat.zapFile++;
523 hash = CODA_NC_HASH(name, namelen, dcp);
524 cncp = coda_nc_find(dcp, name, namelen, 0, hash);
527 coda_nc_hash[hash].length--; /* Used for tuning */
529 coda_nc_remove(cncp, NOT_DOWNCALL);
530 cncp = coda_nc_find(dcp, name, namelen, 0, hash);
535 * Remove all the entries for a particular user. Used when tokens expire.
536 * A user is determined by his/her effective user id (id_uid).
539 coda_nc_purge_user(uid, dcstat)
541 enum dc_status dcstat;
544 * I think the best approach is to go through the entire cache
545 * via HASH or whatever and zap all entries which match the
546 * input cred. Or just flush the whole cache. It might be
547 * best to go through on basis of LRU since cache will almost
548 * always be full and LRU is more straightforward.
551 struct coda_cache *cncp, *ncncp;
554 if (coda_nc_use == 0) /* Cache is off */
557 CODA_NC_DEBUG(CODA_NC_PURGEUSER,
558 myprintf(("ZapDude: uid %x\n", uid)); )
559 coda_nc_stat.zapUsers++;
561 for (cncp = CODA_NC_LRUGET(coda_nc_lru);
562 cncp != (struct coda_cache *)(&coda_nc_lru);
564 ncncp = CODA_NC_LRUGET(*cncp);
566 if ((CODA_NC_VALID(cncp)) &&
567 ((cncp->cred)->cr_uid == uid)) {
568 /* Seems really ugly, but we have to decrement the appropriate
569 hash bucket length here, so we have to find the hash bucket
571 hash = CODA_NC_HASH(cncp->name, cncp->namelen, cncp->dcp);
572 coda_nc_hash[hash].length--; /* For performance tuning */
574 coda_nc_remove(cncp, dcstat);
580 * Flush the entire name cache. In response to a flush of the Venus cache.
583 coda_nc_flush(dcstat)
584 enum dc_status dcstat;
586 /* One option is to deallocate the current name cache and
587 call init to start again. Or just deallocate, then rebuild.
588 Or again, we could just go through the array and zero the
593 * Go through the whole lru chain and kill everything as we go.
594 * I don't use remove since that would rebuild the lru chain
595 * as it went and that seemed unneccesary.
597 struct coda_cache *cncp;
600 if (coda_nc_use == 0) /* Cache is off */
603 coda_nc_stat.Flushes++;
605 for (cncp = CODA_NC_LRUGET(coda_nc_lru);
606 cncp != (struct coda_cache *)&coda_nc_lru;
607 cncp = CODA_NC_LRUGET(*cncp)) {
608 if (CODA_NC_VALID(cncp)) {
610 CODA_NC_HSHREM(cncp); /* only zero valid nodes */
611 CODA_NC_HSHNUL(cncp);
612 if ((dcstat == IS_DOWNCALL)
613 && (CTOV(cncp->dcp)->v_usecount == 1))
615 cncp->dcp->c_flags |= C_PURGING;
617 vrele(CTOV(cncp->dcp));
619 if (CTOV(cncp->cp)->v_flag & VTEXT) {
620 if (coda_vmflush(cncp->cp))
621 CODADEBUG(CODA_FLUSH,
622 myprintf(("coda_nc_flush: (%lx.%lx.%lx) busy\n", cncp->cp->c_fid.Volume, cncp->cp->c_fid.Vnode, cncp->cp->c_fid.Unique)); )
625 if ((dcstat == IS_DOWNCALL)
626 && (CTOV(cncp->cp)->v_usecount == 1))
628 cncp->cp->c_flags |= C_PURGING;
630 vrele(CTOV(cncp->cp));
633 bzero(DATA_PART(cncp),DATA_SIZE);
637 for (i = 0; i < coda_nc_hashsize; i++)
638 coda_nc_hash[i].length = 0;
646 * This routine should print out all the hash chains to the console.
652 struct coda_cache *cncp;
654 for (hash = 0; hash < coda_nc_hashsize; hash++) {
655 myprintf(("\nhash %d\n",hash));
657 for (cncp = coda_nc_hash[hash].hash_next;
658 cncp != (struct coda_cache *)&coda_nc_hash[hash];
659 cncp = cncp->hash_next) {
660 myprintf(("cp %p dcp %p cred %p name %s\n",
662 cncp->cred, cncp->name));
668 coda_nc_gather_stats(void)
670 int i, max = 0, sum = 0, temp, zeros = 0, ave, n;
672 for (i = 0; i < coda_nc_hashsize; i++) {
673 if (coda_nc_hash[i].length) {
674 sum += coda_nc_hash[i].length;
679 if (coda_nc_hash[i].length > max)
680 max = coda_nc_hash[i].length;
684 * When computing the Arithmetic mean, only count slots which
685 * are not empty in the distribution.
687 coda_nc_stat.Sum_bucket_len = sum;
688 coda_nc_stat.Num_zero_len = zeros;
689 coda_nc_stat.Max_bucket_len = max;
691 if ((n = coda_nc_hashsize - zeros) > 0)
697 for (i = 0; i < coda_nc_hashsize; i++) {
698 if (coda_nc_hash[i].length) {
699 temp = coda_nc_hash[i].length - ave;
703 coda_nc_stat.Sum2_bucket_len = sum;
707 * The purpose of this routine is to allow the hash and cache sizes to be
708 * changed dynamically. This should only be used in controlled environments,
709 * it makes no effort to lock other users from accessing the cache while it
710 * is in an improper state (except by turning the cache off).
713 coda_nc_resize(hashsize, heapsize, dcstat)
714 int hashsize, heapsize;
715 enum dc_status dcstat;
717 if ((hashsize % 2) || (heapsize % 2)) { /* Illegal hash or cache sizes */
721 coda_nc_use = 0; /* Turn the cache off */
723 coda_nc_flush(dcstat); /* free any cnodes in the cache */
725 /* WARNING: free must happen *before* size is reset */
726 CODA_FREE(coda_nc_heap,TOTAL_CACHE_SIZE);
727 CODA_FREE(coda_nc_hash,TOTAL_HASH_SIZE);
729 coda_nc_hashsize = hashsize;
730 coda_nc_size = heapsize;
732 coda_nc_init(); /* Set up a cache with the new size */
734 coda_nc_use = 1; /* Turn the cache back on */
739 char coda_nc_name_buf[CODA_MAXNAMLEN+1];
742 coda_nc_name(struct cnode *cp)
744 struct coda_cache *cncp, *ncncp;
747 if (coda_nc_use == 0) /* Cache is off */
750 for (i = 0; i < coda_nc_hashsize; i++) {
751 for (cncp = coda_nc_hash[i].hash_next;
752 cncp != (struct coda_cache *)&coda_nc_hash[i];
754 ncncp = cncp->hash_next;
755 if (cncp->cp == cp) {
756 bcopy(cncp->name, coda_nc_name_buf, cncp->namelen);
757 coda_nc_name_buf[cncp->namelen] = 0;
758 printf(" is %s (%p,%p)@%p",
759 coda_nc_name_buf, cncp->cp, cncp->dcp, cncp);