2 Copyright (c) 2010 Frank Lahm <franklahm@gmail.com>
4 This program is free software; you can redistribute it and/or modify
5 it under the terms of the GNU General Public License as published by
6 the Free Software Foundation; either version 2 of the License, or
7 (at your option) any later version.
9 This program is distributed in the hope that it will be useful,
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 GNU General Public License for more details.
17 #endif /* HAVE_CONFIG_H */
26 #include <atalk/util.h>
27 #include <atalk/cnid.h>
28 #include <atalk/logger.h>
29 #include <atalk/volume.h>
30 #include <atalk/directory.h>
31 #include <atalk/queue.h>
32 #include <atalk/bstrlib.h>
33 #include <atalk/bstradd.h>
36 #include "directory.h"
44 * Cache files and directories in a LRU cache.
46 * The directory cache caches directories and files(!). The main reason for having the cache
47 * is avoiding recursive walks up the path, querying the CNID database each time, when
48 * we have to calculate the location of eg directory with CNID 30, which is located in a dir with
49 * CNID 25, next CNID 20 and then CNID 2 (the volume root as per AFP spec).
50 * If all these dirs where in the cache, each database look up can be avoided. Additionally there's
51 * the element "fullpath" in struct dir, which is used to avoid the recursion in any case. Wheneveer
52 * a struct dir is initialized, the fullpath to the directory is stored there.
54 * In order to speed up the CNID query for files too, which eg happens when a directory is enumerated,
55 * files are stored too in the dircache. In order to differentiate between files and dirs, we re-use
56 * the element fullpath, which for files is always NULL.
58 * The most frequent codepatch that leads to caching is directory enumeration (cf enumerate.c):
59 * - if a element is a directory:
60 * (1) the cache is searched by dircache_search_by_name()
61 * (2) if it wasn't found a new struct dir is created and cached both from within dir_add()
62 * - for files the caching happens a little bit down the call chain:
63 * (3) first getfilparams() is called, which calls
64 * (4) getmetadata() where the cache is searched with dircache_search_by_name()
65 * (5) if the element is not found
66 * (6) get_id() queries the CNID from the database
67 * (7) then a struct dir is initialized via dir_new() (note the fullpath arg is NULL)
68 * (8) finally added to the cache with dircache_add()
69 * (2) of course does contain the steps 6,7 and 8.
71 * The dircache is a LRU cache, whenever it fills up we call dircache_evict internally which removes
72 * DIRCACHE_FREE_QUANTUM elements from the cache.
74 * There is only one cache for all volumes, so of course we use the volume id in hashing calculations.
76 * In order to avoid cache poisoning, we store the cached entries st_ctime from stat in
77 * struct dir.ctime_dircache. Later when we search the cache we compare the stored
78 * value with the result of a fresh stat. If the times differ, we remove the cached
79 * entry and return "no entry found in cache".
80 * A elements ctime changes when
81 * 1) the element is renamed
82 * (we loose the cached entry here, but it will expire when the cache fills)
83 * 2) its a directory and an object has been created therein
84 * 3) the element is deleted and recreated under the same name
85 * Using ctime leads to cache eviction in case 2) where it wouldn't be necessary, because
86 * the dir itself (name, CNID, ...) hasn't changed, but there's no other way.
91 * The maximum dircache size is:
92 * max(DEFAULT_MAX_DIRCACHE_SIZE, min(size, MAX_POSSIBLE_DIRCACHE_SIZE)).
93 * It is a hashtable which we use to store "struct dir"s in. If the cache get full, oldest
94 * entries are evicted in chunks of DIRCACHE_FREE.
96 * We have/need two indexes:
97 * - a DID/name index on the main dircache, another hashtable
98 * - a queue index on the dircache, for evicting the oldest entries
103 * Sending SIGHUP to a afpd child causes it to dump the dircache to a file "/tmp/dircache.PID".
106 /********************************************************
107 * Local funcs and variables
108 ********************************************************/
110 /*****************************
113 static hash_t *dircache; /* The actual cache */
114 static unsigned int dircache_maxsize; /* cache maximum size */
116 static struct dircache_stat {
117 unsigned long long lookups;
118 unsigned long long hits;
119 unsigned long long misses;
120 unsigned long long added;
121 unsigned long long removed;
122 unsigned long long expunged;
123 unsigned long long evicted;
127 static hash_val_t hash_vid_did(const void *key)
129 const struct dir *k = (const struct dir *)key;
130 hash_val_t hash = 2166136261;
132 hash ^= k->d_vid >> 8;
137 hash ^= k->d_did >> 24;
139 hash ^= (k->d_did >> 16) & 0xff;
141 hash ^= (k->d_did >> 8) & 0xff;
143 hash ^= (k->d_did >> 0) & 0xff;
149 static int hash_comp_vid_did(const void *key1, const void *key2)
151 const struct dir *k1 = key1;
152 const struct dir *k2 = key2;
154 return !(k1->d_did == k2->d_did && k1->d_vid == k2->d_vid);
157 /**************************************************
158 * DID/name index on dircache (another hashtable) */
160 static hash_t *index_didname;
163 #if (defined(__GNUC__) && defined(__i386__)) || defined(__WATCOMC__) \
164 || defined(_MSC_VER) || defined (__BORLANDC__) || defined (__TURBOC__)
165 #define get16bits(d) (*((const uint16_t *) (d)))
168 #if !defined (get16bits)
169 #define get16bits(d) ((((uint32_t)(((const uint8_t *)(d))[1])) << 8) \
170 +(uint32_t)(((const uint8_t *)(d))[0]) )
173 static hash_val_t hash_didname(const void *p)
175 const struct dir *key = (const struct dir *)p;
176 const unsigned char *data = key->d_u_name->data;
177 int len = key->d_u_name->slen;
178 hash_val_t hash = key->d_pdid + key->d_vid;
185 for (;len > 0; len--) {
186 hash += get16bits (data);
187 tmp = (get16bits (data+2) << 11) ^ hash;
188 hash = (hash << 16) ^ tmp;
189 data += 2*sizeof (uint16_t);
193 /* Handle end cases */
195 case 3: hash += get16bits (data);
197 hash ^= data[sizeof (uint16_t)] << 18;
200 case 2: hash += get16bits (data);
204 case 1: hash += *data;
209 /* Force "avalanching" of final 127 bits */
220 static int hash_comp_didname(const void *k1, const void *k2)
222 const struct dir *key1 = (const struct dir *)k1;
223 const struct dir *key2 = (const struct dir *)k2;
225 return ! (key1->d_vid == key2->d_vid
226 && key1->d_pdid == key2->d_pdid
227 && (bstrcmp(key1->d_u_name, key2->d_u_name) == 0) );
230 /***************************
231 * queue index on dircache */
233 static q_t *index_queue; /* the index itself */
234 static unsigned long queue_count;
237 * @brief Remove a fixed number of (oldest) entries from the cache and indexes
239 * The default is to remove the 256 oldest entries from the cache.
240 * 1. Get the oldest entry
241 * 2. If it's in use ie open forks reference it or it's curdir requeue it,
243 * 3. Remove the dir from the main cache and the didname index
244 * 4. Free the struct dir structure and all its members
246 static void dircache_evict(void)
248 int i = DIRCACHE_FREE_QUANTUM;
251 LOG(log_debug, logtype_afpd, "dircache: {starting cache eviction}");
254 if ((dir = (struct dir *)dequeue(index_queue)) == NULL) { /* 1 */
256 AFP_PANIC("dircache_evict");
260 if (curdir == dir) { /* 2 */
261 if ((dir->qidx_node = enqueue(index_queue, dir)) == NULL) {
263 AFP_PANIC("dircache_evict");
269 dircache_remove(NULL, dir, DIRCACHE | DIDNAME_INDEX); /* 3 */
270 dir_free(dir); /* 4 */
273 AFP_ASSERT(queue_count == dircache->hash_nodecount);
274 dircache_stat.evicted += DIRCACHE_FREE_QUANTUM;
275 LOG(log_debug, logtype_afpd, "dircache: {finished cache eviction}");
279 /********************************************************
281 ********************************************************/
284 * @brief Search the dircache via a CNID for a directory
286 * Found cache entries are expunged if both the parent directory st_ctime and the objects
287 * st_ctime are modified.
288 * This func builds on the fact, that all our code only ever needs to and does search
289 * the dircache by CNID expecting directories to be returned, but not files.
291 * (1) if we find a file (d_fullpath == NULL) for a given CNID we
292 * (1a) remove it from the cache
293 * (1b) return NULL indicating nothing found
294 * (2) we can then use d_fullpath to stat the directory
296 * @param vol (r) pointer to struct vol
297 * @param cnid (r) CNID of the directory to search
299 * @returns Pointer to struct dir if found, else NULL
301 struct dir *dircache_search_by_did(const struct vol *vol, cnid_t cnid)
303 struct dir *cdir = NULL;
309 AFP_ASSERT(ntohl(cnid) >= CNID_START);
311 dircache_stat.lookups++;
312 key.d_vid = vol->v_vid;
314 if ((hn = hash_lookup(dircache, &key)))
315 cdir = hnode_get(hn);
318 if (cdir->d_fullpath == NULL) { /* (1) */
319 LOG(log_debug, logtype_afpd, "dircache(cnid:%u): {not a directory:\"%s\"}",
320 ntohl(cnid), cfrombstr(cdir->d_u_name));
321 (void)dir_remove(vol, cdir); /* (1a) */
322 dircache_stat.expunged++;
323 return NULL; /* (1b) */
326 if (lstat(cfrombstr(cdir->d_fullpath), &st) != 0) {
327 LOG(log_debug, logtype_afpd, "dircache(cnid:%u): {missing:\"%s\"}",
328 ntohl(cnid), cfrombstr(cdir->d_fullpath));
329 (void)dir_remove(vol, cdir);
330 dircache_stat.expunged++;
333 if (cdir->ctime_dircache != st.st_ctime) {
334 LOG(log_debug, logtype_afpd, "dircache(cnid:%u): {modified:\"%s\"}",
335 ntohl(cnid), cfrombstr(cdir->d_u_name));
336 (void)dir_remove(vol, cdir);
337 dircache_stat.expunged++;
340 LOG(log_debug, logtype_afpd, "dircache(cnid:%u): {cached: path:\"%s\"}",
341 ntohl(cnid), cfrombstr(cdir->d_fullpath));
342 dircache_stat.hits++;
344 LOG(log_debug, logtype_afpd, "dircache(cnid:%u): {not in cache}", ntohl(cnid));
345 dircache_stat.misses++;
352 * @brief Search the cache via did/name hashtable
354 * Found cache entries are expunged if both the parent directory st_ctime and the objects
355 * st_ctime are modified.
357 * @param vol (r) volume
358 * @param dir (r) directory
359 * @param name (r) name (server side encoding)
360 * @parma len (r) strlen of name
361 * @param ctime (r) current st_ctime from stat
363 * @returns pointer to struct dir if found in cache, else NULL
365 struct dir *dircache_search_by_name(const struct vol *vol,
366 const struct dir *dir,
371 struct dir *cdir = NULL;
375 static_bstring uname = {-1, len, (unsigned char *)name};
380 AFP_ASSERT(len == strlen(name));
381 AFP_ASSERT(len < 256);
383 dircache_stat.lookups++;
384 LOG(log_debug, logtype_afpd, "dircache_search_by_name(did:%u, \"%s\")",
385 ntohl(dir->d_did), name);
387 if (dir->d_did != DIRDID_ROOT_PARENT) {
388 key.d_vid = vol->v_vid;
389 key.d_pdid = dir->d_did;
390 key.d_u_name = &uname;
392 if ((hn = hash_lookup(index_didname, &key)))
393 cdir = hnode_get(hn);
397 if (cdir->ctime_dircache != ctime) {
398 LOG(log_debug, logtype_afpd, "dircache(did:%u,\"%s\"): {modified}",
399 ntohl(dir->d_did), name);
400 (void)dir_remove(vol, cdir);
401 dircache_stat.expunged++;
404 LOG(log_debug, logtype_afpd, "dircache(did:%u,\"%s\"): {found in cache}",
405 ntohl(dir->d_did), name);
406 dircache_stat.hits++;
408 LOG(log_debug, logtype_afpd, "dircache(did:%u,\"%s\"): {not in cache}",
409 ntohl(dir->d_did), name);
410 dircache_stat.misses++;
417 * @brief create struct dir from struct path
419 * Add a struct dir to the cache and its indexes.
421 * @param dir (r) pointer to parrent directory
423 * @returns 0 on success, -1 on error which should result in an abort
425 int dircache_add(const struct vol *vol,
428 struct dir *cdir = NULL;
433 AFP_ASSERT(ntohl(dir->d_pdid) >= 2);
434 AFP_ASSERT(ntohl(dir->d_did) >= CNID_START);
435 AFP_ASSERT(dir->d_u_name);
436 AFP_ASSERT(dir->d_vid);
437 AFP_ASSERT(dircache->hash_nodecount <= dircache_maxsize);
439 /* Check if cache is full */
440 if (dircache->hash_nodecount == dircache_maxsize)
444 * Make sure we don't add duplicates
447 /* Search primary cache by CNID */
448 key.d_vid = dir->d_vid;
449 key.d_did = dir->d_did;
450 if ((hn = hash_lookup(dircache, &key))) {
451 /* Found an entry with the same CNID, delete it */
452 dir_remove(vol, hnode_get(hn));
453 dircache_stat.expunged++;
455 key.d_vid = vol->v_vid;
456 key.d_pdid = dir->d_did;
457 key.d_u_name = dir->d_u_name;
458 if ((hn = hash_lookup(index_didname, &key))) {
459 /* Found an entry with the same DID/name, delete it */
460 dir_remove(vol, hnode_get(hn));
461 dircache_stat.expunged++;
464 /* Add it to the main dircache */
465 if (hash_alloc_insert(dircache, dir, dir) == 0) {
470 /* Add it to the did/name index */
471 if (hash_alloc_insert(index_didname, dir, dir) == 0) {
476 /* Add it to the fifo queue index */
477 if ((dir->qidx_node = enqueue(index_queue, dir)) == NULL) {
484 dircache_stat.added++;
485 LOG(log_debug, logtype_afpd, "dircache(did:%u,'%s'): {added}",
486 ntohl(dir->d_did), cfrombstr(dir->d_u_name));
488 AFP_ASSERT(queue_count == index_didname->hash_nodecount
489 && queue_count == dircache->hash_nodecount);
495 * @brief Remove an entry from the dircache
497 * Callers outside of dircache.c should call this with
498 * flags = QUEUE_INDEX | DIDNAME_INDEX | DIRCACHE.
500 void dircache_remove(const struct vol *vol _U_, struct dir *dir, int flags)
505 AFP_ASSERT((flags & ~(QUEUE_INDEX | DIDNAME_INDEX | DIRCACHE)) == 0);
507 if (flags & QUEUE_INDEX) {
508 /* remove it from the queue index */
509 dequeue(dir->qidx_node->prev); /* this effectively deletes the dequeued node */
513 if (flags & DIDNAME_INDEX) {
514 if ((hn = hash_lookup(index_didname, dir)) == NULL) {
515 LOG(log_error, logtype_afpd, "dircache_remove(%u,\"%s\"): not in didname index",
516 ntohl(dir->d_did), cfrombstr(dir->d_u_name));
518 AFP_PANIC("dircache_remove");
520 hash_delete_free(index_didname, hn);
523 if (flags & DIRCACHE) {
524 if ((hn = hash_lookup(dircache, dir)) == NULL) {
525 LOG(log_error, logtype_afpd, "dircache_remove(%u,\"%s\"): not in dircache",
526 ntohl(dir->d_did), cfrombstr(dir->d_u_name));
528 AFP_PANIC("dircache_remove");
530 hash_delete_free(dircache, hn);
533 LOG(log_debug, logtype_afpd, "dircache(did:%u,\"%s\"): {removed}",
534 ntohl(dir->d_did), cfrombstr(dir->d_u_name));
536 dircache_stat.removed++;
537 AFP_ASSERT(queue_count == index_didname->hash_nodecount
538 && queue_count == dircache->hash_nodecount);
542 * @brief Initialize the dircache and indexes
544 * This is called in child afpd initialisation. The maximum cache size will be
545 * max(DEFAULT_MAX_DIRCACHE_SIZE, min(size, MAX_POSSIBLE_DIRCACHE_SIZE)).
546 * It initializes a hashtable which we use to store a directory cache in.
547 * It also initializes two indexes:
548 * - a DID/name index on the main dircache
549 * - a queue index on the dircache
551 * @param size (r) requested maximum size from afpd.conf
553 * @return 0 on success, -1 on error
555 int dircache_init(int reqsize)
557 dircache_maxsize = DEFAULT_MAX_DIRCACHE_SIZE;
559 /* Initialize the main dircache */
560 if (reqsize > DEFAULT_MAX_DIRCACHE_SIZE && reqsize < MAX_POSSIBLE_DIRCACHE_SIZE) {
561 while ((dircache_maxsize < MAX_POSSIBLE_DIRCACHE_SIZE) && (dircache_maxsize < reqsize))
562 dircache_maxsize *= 2;
564 if ((dircache = hash_create(dircache_maxsize, hash_comp_vid_did, hash_vid_did)) == NULL)
567 LOG(log_debug, logtype_afpd, "dircache_init: done. max dircache size: %u", dircache_maxsize);
569 /* Initialize did/name index hashtable */
570 if ((index_didname = hash_create(dircache_maxsize, hash_comp_didname, hash_didname)) == NULL)
573 /* Initialize index queue */
574 if ((index_queue = queue_init()) == NULL)
579 /* Initialize index queue */
580 if ((invalid_dircache_entries = queue_init()) == NULL)
583 /* As long as directory.c hasn't got its own initializer call, we do it for it */
584 rootParent.d_did = DIRDID_ROOT_PARENT;
585 rootParent.d_fullpath = bfromcstr("ROOT_PARENT");
586 rootParent.d_m_name = bfromcstr("ROOT_PARENT");
587 rootParent.d_u_name = rootParent.d_m_name;
593 * Log dircache statistics
595 void log_dircache_stat(void)
597 LOG(log_info, logtype_afpd, "dircache statistics: "
598 "entries: %lu, lookups: %llu, hits: %llu, misses: %llu, added: %llu, removed: %llu, expunged: %llu, evicted: %llu",
600 dircache_stat.lookups,
602 dircache_stat.misses,
604 dircache_stat.removed,
605 dircache_stat.expunged,
606 dircache_stat.evicted);
610 * @brief Dump dircache to /tmp/dircache.PID
612 void dircache_dump(void)
616 qnode_t *n = index_queue->next;
619 const struct dir *dir;
622 LOG(log_warning, logtype_afpd, "Dumping directory cache...");
624 sprintf(tmpnam, "/tmp/dircache.%u", getpid());
625 if ((dump = fopen(tmpnam, "w+")) == NULL) {
626 LOG(log_error, logtype_afpd, "dircache_dump: %s", strerror(errno));
631 fprintf(dump, "Number of cache entries in LRU queue: %lu\n", queue_count);
632 fprintf(dump, "Configured maximum cache size: %u\n\n", dircache_maxsize);
634 fprintf(dump, "Primary CNID index:\n");
635 fprintf(dump, " VID DID CNID STAT PATH\n");
636 fprintf(dump, "====================================================================\n");
637 hash_scan_begin(&hs, dircache);
639 while ((hn = hash_scan_next(&hs))) {
641 fprintf(dump, "%05u: %3u %6u %6u %s %s\n",
646 dir->d_fullpath ? "d" : "f",
647 cfrombstr(dir->d_u_name));
650 fprintf(dump, "\nSecondary DID/name index:\n");
651 fprintf(dump, " VID DID CNID STAT PATH\n");
652 fprintf(dump, "====================================================================\n");
653 hash_scan_begin(&hs, index_didname);
655 while ((hn = hash_scan_next(&hs))) {
657 fprintf(dump, "%05u: %3u %6u %6u %s %s\n",
662 dir->d_fullpath ? "d" : "f",
663 cfrombstr(dir->d_u_name));
666 fprintf(dump, "\nLRU Queue:\n");
667 fprintf(dump, " VID DID CNID STAT PATH\n");
668 fprintf(dump, "====================================================================\n");
670 for (i = 1; i <= queue_count; i++) {
671 if (n == index_queue)
673 dir = (struct dir *)n->data;
674 fprintf(dump, "%05u: %3u %6u %6u %s %s\n",
679 dir->d_fullpath ? "d" : "f",
680 cfrombstr(dir->d_u_name));