tdb: Don't malloc for every record in traverse
[samba.git] / lib / tdb / common / traverse.c
1  /*
2    Unix SMB/CIFS implementation.
3
4    trivial database library
5
6    Copyright (C) Andrew Tridgell              1999-2005
7    Copyright (C) Paul `Rusty' Russell              2000
8    Copyright (C) Jeremy Allison                    2000-2003
9
10      ** NOTE! The following LGPL license applies to the tdb
11      ** library. This does NOT imply that all of Samba is released
12      ** under the LGPL
13
14    This library is free software; you can redistribute it and/or
15    modify it under the terms of the GNU Lesser General Public
16    License as published by the Free Software Foundation; either
17    version 3 of the License, or (at your option) any later version.
18
19    This library is distributed in the hope that it will be useful,
20    but WITHOUT ANY WARRANTY; without even the implied warranty of
21    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
22    Lesser General Public License for more details.
23
24    You should have received a copy of the GNU Lesser General Public
25    License along with this library; if not, see <http://www.gnu.org/licenses/>.
26 */
27
28 #include "tdb_private.h"
29
30 #define TDB_NEXT_LOCK_ERR ((tdb_off_t)-1)
31
32 /* Uses traverse lock: 0 = finish, TDB_NEXT_LOCK_ERR = error,
33    other = record offset */
34 static tdb_off_t tdb_next_lock(struct tdb_context *tdb, struct tdb_traverse_lock *tlock,
35                          struct tdb_record *rec)
36 {
37         int want_next = (tlock->off != 0);
38
39         /* Lock each chain from the start one. */
40         for (; tlock->hash < tdb->hash_size; tlock->hash++) {
41                 if (!tlock->off && tlock->hash != 0) {
42                         /* this is an optimisation for the common case where
43                            the hash chain is empty, which is particularly
44                            common for the use of tdb with ldb, where large
45                            hashes are used. In that case we spend most of our
46                            time in tdb_brlock(), locking empty hash chains.
47
48                            To avoid this, we do an unlocked pre-check to see
49                            if the hash chain is empty before starting to look
50                            inside it. If it is empty then we can avoid that
51                            hash chain. If it isn't empty then we can't believe
52                            the value we get back, as we read it without a
53                            lock, so instead we get the lock and re-fetch the
54                            value below.
55
56                            Notice that not doing this optimisation on the
57                            first hash chain is critical. We must guarantee
58                            that we have done at least one fcntl lock at the
59                            start of a search to guarantee that memory is
60                            coherent on SMP systems. If records are added by
61                            others during the search then thats OK, and we
62                            could possibly miss those with this trick, but we
63                            could miss them anyway without this trick, so the
64                            semantics don't change.
65
66                            With a non-indexed ldb search this trick gains us a
67                            factor of around 80 in speed on a linux 2.6.x
68                            system (testing using ldbtest).
69                         */
70                         tdb->methods->next_hash_chain(tdb, &tlock->hash);
71                         if (tlock->hash == tdb->hash_size) {
72                                 continue;
73                         }
74                 }
75
76                 if (tdb_lock(tdb, tlock->hash, tlock->lock_rw) == -1)
77                         return TDB_NEXT_LOCK_ERR;
78
79                 /* No previous record?  Start at top of chain. */
80                 if (!tlock->off) {
81                         if (tdb_ofs_read(tdb, TDB_HASH_TOP(tlock->hash),
82                                      &tlock->off) == -1)
83                                 goto fail;
84                 } else {
85                         /* Otherwise unlock the previous record. */
86                         if (tdb_unlock_record(tdb, tlock->off) != 0)
87                                 goto fail;
88                 }
89
90                 if (want_next) {
91                         /* We have offset of old record: grab next */
92                         if (tdb_rec_read(tdb, tlock->off, rec) == -1)
93                                 goto fail;
94                         tlock->off = rec->next;
95                 }
96
97                 /* Iterate through chain */
98                 while( tlock->off) {
99                         tdb_off_t current;
100                         if (tdb_rec_read(tdb, tlock->off, rec) == -1)
101                                 goto fail;
102
103                         /* Detect infinite loops. From "Shlomi Yaakobovich" <Shlomi@exanet.com>. */
104                         if (tlock->off == rec->next) {
105                                 tdb->ecode = TDB_ERR_CORRUPT;
106                                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_next_lock: loop detected.\n"));
107                                 goto fail;
108                         }
109
110                         if (!TDB_DEAD(rec)) {
111                                 /* Woohoo: we found one! */
112                                 if (tdb_lock_record(tdb, tlock->off) != 0)
113                                         goto fail;
114                                 return tlock->off;
115                         }
116
117                         /* Try to clean dead ones from old traverses */
118                         current = tlock->off;
119                         tlock->off = rec->next;
120                         if (!(tdb->read_only || tdb->traverse_read) &&
121                             tdb_do_delete(tdb, current, rec) != 0)
122                                 goto fail;
123                 }
124                 tdb_unlock(tdb, tlock->hash, tlock->lock_rw);
125                 want_next = 0;
126         }
127         /* We finished iteration without finding anything */
128         tdb->ecode = TDB_SUCCESS;
129         return 0;
130
131  fail:
132         tlock->off = 0;
133         if (tdb_unlock(tdb, tlock->hash, tlock->lock_rw) != 0)
134                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_next_lock: On error unlock failed!\n"));
135         return TDB_NEXT_LOCK_ERR;
136 }
137
138 /* traverse the entire database - calling fn(tdb, key, data) on each element.
139    return -1 on error or the record count traversed
140    if fn is NULL then it is not called
141    a non-zero return value from fn() indicates that the traversal should stop
142   */
143 static int tdb_traverse_internal(struct tdb_context *tdb,
144                                  tdb_traverse_func fn, void *private_data,
145                                  struct tdb_traverse_lock *tl)
146 {
147         TDB_DATA key, dbuf;
148         struct tdb_record rec;
149         int ret = 0, count = 0;
150         tdb_off_t off;
151         size_t recbuf_len;
152
153         recbuf_len = 4096;
154         key.dptr = malloc(recbuf_len);
155         if (key.dptr == NULL) {
156                 return -1;
157         }
158
159         /* This was in the initialization, above, but the IRIX compiler
160          * did not like it.  crh
161          */
162         tl->next = tdb->travlocks.next;
163
164         /* fcntl locks don't stack: beware traverse inside traverse */
165         tdb->travlocks.next = tl;
166
167         /* tdb_next_lock places locks on the record returned, and its chain */
168         while ((off = tdb_next_lock(tdb, tl, &rec)) != 0) {
169                 tdb_len_t full_len = rec.key_len + rec.data_len;
170                 int nread;
171
172                 if (full_len > recbuf_len) {
173                         recbuf_len = full_len;
174
175                         /*
176                          * No realloc, we don't need the old data and thus can
177                          * do without the memcpy
178                          */
179                         free(key.dptr);
180                         key.dptr = malloc(recbuf_len);
181
182                         if (key.dptr == NULL) {
183                                 ret = -1;
184                                 if (tdb_unlock(tdb, tl->hash, tl->lock_rw)
185                                     != 0) {
186                                         goto out;
187                                 }
188                                 if (tdb_unlock_record(tdb, tl->off) != 0) {
189                                         TDB_LOG((tdb, TDB_DEBUG_FATAL,
190                                                  "tdb_traverse: malloc "
191                                                  "failed and unlock_record "
192                                                  "failed!\n"));
193                                 }
194                                 goto out;
195                         }
196                 }
197
198                 if (off == TDB_NEXT_LOCK_ERR) {
199                         ret = -1;
200                         goto out;
201                 }
202                 count++;
203                 /* now read the full record */
204                 nread = tdb->methods->tdb_read(tdb, tl->off + sizeof(rec),
205                                                key.dptr, full_len, 0);
206                 if (nread == -1) {
207                         ret = -1;
208                         if (tdb_unlock(tdb, tl->hash, tl->lock_rw) != 0)
209                                 goto out;
210                         if (tdb_unlock_record(tdb, tl->off) != 0)
211                                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_traverse: key.dptr == NULL and unlock_record failed!\n"));
212                         goto out;
213                 }
214                 key.dsize = rec.key_len;
215                 dbuf.dptr = key.dptr + rec.key_len;
216                 dbuf.dsize = rec.data_len;
217
218                 tdb_trace_1rec_retrec(tdb, "traverse", key, dbuf);
219
220                 /* Drop chain lock, call out */
221                 if (tdb_unlock(tdb, tl->hash, tl->lock_rw) != 0) {
222                         ret = -1;
223                         goto out;
224                 }
225                 if (fn && fn(tdb, key, dbuf, private_data)) {
226                         /* They want us to terminate traversal */
227                         tdb_trace_ret(tdb, "tdb_traverse_end", count);
228                         if (tdb_unlock_record(tdb, tl->off) != 0) {
229                                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_traverse: unlock_record failed!\n"));;
230                                 ret = -1;
231                         }
232                         goto out;
233                 }
234         }
235         tdb_trace(tdb, "tdb_traverse_end");
236 out:
237         SAFE_FREE(key.dptr);
238         tdb->travlocks.next = tl->next;
239         if (ret < 0)
240                 return -1;
241         else
242                 return count;
243 }
244
245
246 /*
247   a read style traverse - temporarily marks the db read only
248 */
249 _PUBLIC_ int tdb_traverse_read(struct tdb_context *tdb,
250                       tdb_traverse_func fn, void *private_data)
251 {
252         struct tdb_traverse_lock tl = { NULL, 0, 0, F_RDLCK };
253         int ret;
254
255         /* we need to get a read lock on the transaction lock here to
256            cope with the lock ordering semantics of solaris10 */
257         if (tdb_transaction_lock(tdb, F_RDLCK, TDB_LOCK_WAIT)) {
258                 return -1;
259         }
260
261         tdb->traverse_read++;
262         tdb_trace(tdb, "tdb_traverse_read_start");
263         ret = tdb_traverse_internal(tdb, fn, private_data, &tl);
264         tdb->traverse_read--;
265
266         tdb_transaction_unlock(tdb, F_RDLCK);
267
268         return ret;
269 }
270
271 /*
272   a write style traverse - needs to get the transaction lock to
273   prevent deadlocks
274
275   WARNING: The data buffer given to the callback fn does NOT meet the
276   alignment guarantees malloc gives you.
277 */
278 _PUBLIC_ int tdb_traverse(struct tdb_context *tdb,
279                  tdb_traverse_func fn, void *private_data)
280 {
281         struct tdb_traverse_lock tl = { NULL, 0, 0, F_WRLCK };
282         enum tdb_lock_flags lock_flags;
283         int ret;
284
285         if (tdb->read_only || tdb->traverse_read) {
286                 return tdb_traverse_read(tdb, fn, private_data);
287         }
288
289         lock_flags = TDB_LOCK_WAIT;
290
291         if (tdb->allrecord_lock.count != 0) {
292                 /*
293                  * This avoids a deadlock between tdb_lockall() and
294                  * tdb_traverse(). See
295                  * https://bugzilla.samba.org/show_bug.cgi?id=11381
296                  */
297                 lock_flags = TDB_LOCK_NOWAIT;
298         }
299
300         if (tdb_transaction_lock(tdb, F_WRLCK, lock_flags)) {
301                 return -1;
302         }
303
304         tdb->traverse_write++;
305         tdb_trace(tdb, "tdb_traverse_start");
306         ret = tdb_traverse_internal(tdb, fn, private_data, &tl);
307         tdb->traverse_write--;
308
309         tdb_transaction_unlock(tdb, F_WRLCK);
310
311         return ret;
312 }
313
314
315 /* find the first entry in the database and return its key */
316 _PUBLIC_ TDB_DATA tdb_firstkey(struct tdb_context *tdb)
317 {
318         TDB_DATA key;
319         struct tdb_record rec;
320         tdb_off_t off;
321
322         /* release any old lock */
323         if (tdb_unlock_record(tdb, tdb->travlocks.off) != 0)
324                 return tdb_null;
325         tdb->travlocks.off = tdb->travlocks.hash = 0;
326         tdb->travlocks.lock_rw = F_RDLCK;
327
328         /* Grab first record: locks chain and returned record. */
329         off = tdb_next_lock(tdb, &tdb->travlocks, &rec);
330         if (off == 0 || off == TDB_NEXT_LOCK_ERR) {
331                 tdb_trace_retrec(tdb, "tdb_firstkey", tdb_null);
332                 return tdb_null;
333         }
334         /* now read the key */
335         key.dsize = rec.key_len;
336         key.dptr =tdb_alloc_read(tdb,tdb->travlocks.off+sizeof(rec),key.dsize);
337
338         tdb_trace_retrec(tdb, "tdb_firstkey", key);
339
340         /* Unlock the hash chain of the record we just read. */
341         if (tdb_unlock(tdb, tdb->travlocks.hash, tdb->travlocks.lock_rw) != 0)
342                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_firstkey: error occurred while tdb_unlocking!\n"));
343         return key;
344 }
345
346 /* find the next entry in the database, returning its key */
347 _PUBLIC_ TDB_DATA tdb_nextkey(struct tdb_context *tdb, TDB_DATA oldkey)
348 {
349         uint32_t oldhash;
350         TDB_DATA key = tdb_null;
351         struct tdb_record rec;
352         unsigned char *k = NULL;
353         tdb_off_t off;
354
355         /* Is locked key the old key?  If so, traverse will be reliable. */
356         if (tdb->travlocks.off) {
357                 if (tdb_lock(tdb,tdb->travlocks.hash,tdb->travlocks.lock_rw))
358                         return tdb_null;
359                 if (tdb_rec_read(tdb, tdb->travlocks.off, &rec) == -1
360                     || !(k = tdb_alloc_read(tdb,tdb->travlocks.off+sizeof(rec),
361                                             rec.key_len))
362                     || memcmp(k, oldkey.dptr, oldkey.dsize) != 0) {
363                         /* No, it wasn't: unlock it and start from scratch */
364                         if (tdb_unlock_record(tdb, tdb->travlocks.off) != 0) {
365                                 tdb_trace_1rec_retrec(tdb, "tdb_nextkey",
366                                                       oldkey, tdb_null);
367                                 SAFE_FREE(k);
368                                 return tdb_null;
369                         }
370                         if (tdb_unlock(tdb, tdb->travlocks.hash, tdb->travlocks.lock_rw) != 0) {
371                                 SAFE_FREE(k);
372                                 return tdb_null;
373                         }
374                         tdb->travlocks.off = 0;
375                 }
376
377                 SAFE_FREE(k);
378         }
379
380         if (!tdb->travlocks.off) {
381                 /* No previous element: do normal find, and lock record */
382                 tdb->travlocks.off = tdb_find_lock_hash(tdb, oldkey, tdb->hash_fn(&oldkey), tdb->travlocks.lock_rw, &rec);
383                 if (!tdb->travlocks.off) {
384                         tdb_trace_1rec_retrec(tdb, "tdb_nextkey", oldkey, tdb_null);
385                         return tdb_null;
386                 }
387                 tdb->travlocks.hash = BUCKET(rec.full_hash);
388                 if (tdb_lock_record(tdb, tdb->travlocks.off) != 0) {
389                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_nextkey: lock_record failed (%s)!\n", strerror(errno)));
390                         return tdb_null;
391                 }
392         }
393         oldhash = tdb->travlocks.hash;
394
395         /* Grab next record: locks chain and returned record,
396            unlocks old record */
397         off = tdb_next_lock(tdb, &tdb->travlocks, &rec);
398         if (off != TDB_NEXT_LOCK_ERR && off != 0) {
399                 key.dsize = rec.key_len;
400                 key.dptr = tdb_alloc_read(tdb, tdb->travlocks.off+sizeof(rec),
401                                           key.dsize);
402                 /* Unlock the chain of this new record */
403                 if (tdb_unlock(tdb, tdb->travlocks.hash, tdb->travlocks.lock_rw) != 0)
404                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_nextkey: WARNING tdb_unlock failed!\n"));
405         }
406         /* Unlock the chain of old record */
407         if (tdb_unlock(tdb, BUCKET(oldhash), tdb->travlocks.lock_rw) != 0)
408                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_nextkey: WARNING tdb_unlock failed!\n"));
409         tdb_trace_1rec_retrec(tdb, "tdb_nextkey", oldkey, key);
410         return key;
411 }
412