lib/tdb: TDB_TRACE support (for developers)
[ira/wip.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 list_struct *rec)
36 {
37         int want_next = (tlock->off != 0);
38
39         /* Lock each chain from the start one. */
40         for (; tlock->hash < tdb->header.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->header.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         return TDB_ERRCODE(TDB_SUCCESS, 0);
129
130  fail:
131         tlock->off = 0;
132         if (tdb_unlock(tdb, tlock->hash, tlock->lock_rw) != 0)
133                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_next_lock: On error unlock failed!\n"));
134         return TDB_NEXT_LOCK_ERR;
135 }
136
137 /* traverse the entire database - calling fn(tdb, key, data) on each element.
138    return -1 on error or the record count traversed
139    if fn is NULL then it is not called
140    a non-zero return value from fn() indicates that the traversal should stop
141   */
142 static int tdb_traverse_internal(struct tdb_context *tdb, 
143                                  tdb_traverse_func fn, void *private_data,
144                                  struct tdb_traverse_lock *tl)
145 {
146         TDB_DATA key, dbuf;
147         struct list_struct rec;
148         int ret = 0, count = 0;
149         tdb_off_t off;
150
151         /* This was in the initializaton, above, but the IRIX compiler
152          * did not like it.  crh
153          */
154         tl->next = tdb->travlocks.next;
155
156         /* fcntl locks don't stack: beware traverse inside traverse */
157         tdb->travlocks.next = tl;
158
159         /* tdb_next_lock places locks on the record returned, and its chain */
160         while ((off = tdb_next_lock(tdb, tl, &rec)) != 0) {
161                 if (off == TDB_NEXT_LOCK_ERR) {
162                         ret = -1;
163                         goto out;
164                 }
165                 count++;
166                 /* now read the full record */
167                 key.dptr = tdb_alloc_read(tdb, tl->off + sizeof(rec), 
168                                           rec.key_len + rec.data_len);
169                 if (!key.dptr) {
170                         ret = -1;
171                         if (tdb_unlock(tdb, tl->hash, tl->lock_rw) != 0)
172                                 goto out;
173                         if (tdb_unlock_record(tdb, tl->off) != 0)
174                                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_traverse: key.dptr == NULL and unlock_record failed!\n"));
175                         goto out;
176                 }
177                 key.dsize = rec.key_len;
178                 dbuf.dptr = key.dptr + rec.key_len;
179                 dbuf.dsize = rec.data_len;
180
181                 tdb_trace_1rec_retrec(tdb, "traverse", key, dbuf);
182
183                 /* Drop chain lock, call out */
184                 if (tdb_unlock(tdb, tl->hash, tl->lock_rw) != 0) {
185                         ret = -1;
186                         SAFE_FREE(key.dptr);
187                         goto out;
188                 }
189                 if (fn && fn(tdb, key, dbuf, private_data)) {
190                         /* They want us to terminate traversal */
191                         tdb_trace_ret(tdb, "tdb_traverse_end", count);
192                         if (tdb_unlock_record(tdb, tl->off) != 0) {
193                                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_traverse: unlock_record failed!\n"));;
194                                 ret = -1;
195                         }
196                         SAFE_FREE(key.dptr);
197                         goto out;
198                 }
199                 SAFE_FREE(key.dptr);
200         }
201         tdb_trace(tdb, "tdb_traverse_end");
202 out:
203         tdb->travlocks.next = tl->next;
204         if (ret < 0)
205                 return -1;
206         else
207                 return count;
208 }
209
210
211 /*
212   a write style traverse - temporarily marks the db read only
213 */
214 int tdb_traverse_read(struct tdb_context *tdb, 
215                       tdb_traverse_func fn, void *private_data)
216 {
217         struct tdb_traverse_lock tl = { NULL, 0, 0, F_RDLCK };
218         int ret;
219
220         /* we need to get a read lock on the transaction lock here to
221            cope with the lock ordering semantics of solaris10 */
222         if (tdb_transaction_lock(tdb, F_RDLCK)) {
223                 return -1;
224         }
225
226         tdb->traverse_read++;
227         tdb_trace(tdb, "tdb_traverse_read_start");
228         ret = tdb_traverse_internal(tdb, fn, private_data, &tl);
229         tdb->traverse_read--;
230
231         tdb_transaction_unlock(tdb);
232
233         return ret;
234 }
235
236 /*
237   a write style traverse - needs to get the transaction lock to
238   prevent deadlocks
239
240   WARNING: The data buffer given to the callback fn does NOT meet the
241   alignment restrictions malloc gives you.
242 */
243 int tdb_traverse(struct tdb_context *tdb, 
244                  tdb_traverse_func fn, void *private_data)
245 {
246         struct tdb_traverse_lock tl = { NULL, 0, 0, F_WRLCK };
247         int ret;
248
249         if (tdb->read_only || tdb->traverse_read) {
250                 return tdb_traverse_read(tdb, fn, private_data);
251         }
252
253         if (tdb_transaction_lock(tdb, F_WRLCK)) {
254                 return -1;
255         }
256
257         tdb->traverse_write++;
258         tdb_trace(tdb, "tdb_traverse_start");
259         ret = tdb_traverse_internal(tdb, fn, private_data, &tl);
260         tdb->traverse_write--;
261
262         tdb_transaction_unlock(tdb);
263
264         return ret;
265 }
266
267
268 /* find the first entry in the database and return its key */
269 TDB_DATA tdb_firstkey(struct tdb_context *tdb)
270 {
271         TDB_DATA key;
272         struct list_struct rec;
273         tdb_off_t off;
274
275         /* release any old lock */
276         if (tdb_unlock_record(tdb, tdb->travlocks.off) != 0)
277                 return tdb_null;
278         tdb->travlocks.off = tdb->travlocks.hash = 0;
279         tdb->travlocks.lock_rw = F_RDLCK;
280
281         /* Grab first record: locks chain and returned record. */
282         off = tdb_next_lock(tdb, &tdb->travlocks, &rec);
283         if (off == 0 || off == TDB_NEXT_LOCK_ERR) {
284                 tdb_trace_retrec(tdb, "tdb_firstkey", tdb_null);
285                 return tdb_null;
286         }
287         /* now read the key */
288         key.dsize = rec.key_len;
289         key.dptr =tdb_alloc_read(tdb,tdb->travlocks.off+sizeof(rec),key.dsize);
290
291         tdb_trace_retrec(tdb, "tdb_firstkey", key);
292
293         /* Unlock the hash chain of the record we just read. */
294         if (tdb_unlock(tdb, tdb->travlocks.hash, tdb->travlocks.lock_rw) != 0)
295                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_firstkey: error occurred while tdb_unlocking!\n"));
296         return key;
297 }
298
299 /* find the next entry in the database, returning its key */
300 TDB_DATA tdb_nextkey(struct tdb_context *tdb, TDB_DATA oldkey)
301 {
302         uint32_t oldhash;
303         TDB_DATA key = tdb_null;
304         struct list_struct rec;
305         unsigned char *k = NULL;
306         tdb_off_t off;
307
308         /* Is locked key the old key?  If so, traverse will be reliable. */
309         if (tdb->travlocks.off) {
310                 if (tdb_lock(tdb,tdb->travlocks.hash,tdb->travlocks.lock_rw))
311                         return tdb_null;
312                 if (tdb_rec_read(tdb, tdb->travlocks.off, &rec) == -1
313                     || !(k = tdb_alloc_read(tdb,tdb->travlocks.off+sizeof(rec),
314                                             rec.key_len))
315                     || memcmp(k, oldkey.dptr, oldkey.dsize) != 0) {
316                         /* No, it wasn't: unlock it and start from scratch */
317                         if (tdb_unlock_record(tdb, tdb->travlocks.off) != 0) {
318                                 tdb_trace_1rec_retrec(tdb, "tdb_nextkey",
319                                                       oldkey, tdb_null);
320                                 SAFE_FREE(k);
321                                 return tdb_null;
322                         }
323                         if (tdb_unlock(tdb, tdb->travlocks.hash, tdb->travlocks.lock_rw) != 0) {
324                                 SAFE_FREE(k);
325                                 return tdb_null;
326                         }
327                         tdb->travlocks.off = 0;
328                 }
329
330                 SAFE_FREE(k);
331         }
332
333         if (!tdb->travlocks.off) {
334                 /* No previous element: do normal find, and lock record */
335                 tdb->travlocks.off = tdb_find_lock_hash(tdb, oldkey, tdb->hash_fn(&oldkey), tdb->travlocks.lock_rw, &rec);
336                 if (!tdb->travlocks.off) {
337                         tdb_trace_1rec_retrec(tdb, "tdb_nextkey", oldkey, tdb_null);
338                         return tdb_null;
339                 }
340                 tdb->travlocks.hash = BUCKET(rec.full_hash);
341                 if (tdb_lock_record(tdb, tdb->travlocks.off) != 0) {
342                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_nextkey: lock_record failed (%s)!\n", strerror(errno)));
343                         return tdb_null;
344                 }
345         }
346         oldhash = tdb->travlocks.hash;
347
348         /* Grab next record: locks chain and returned record,
349            unlocks old record */
350         off = tdb_next_lock(tdb, &tdb->travlocks, &rec);
351         if (off != TDB_NEXT_LOCK_ERR && off != 0) {
352                 key.dsize = rec.key_len;
353                 key.dptr = tdb_alloc_read(tdb, tdb->travlocks.off+sizeof(rec),
354                                           key.dsize);
355                 /* Unlock the chain of this new record */
356                 if (tdb_unlock(tdb, tdb->travlocks.hash, tdb->travlocks.lock_rw) != 0)
357                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_nextkey: WARNING tdb_unlock failed!\n"));
358         }
359         /* Unlock the chain of old record */
360         if (tdb_unlock(tdb, BUCKET(oldhash), tdb->travlocks.lock_rw) != 0)
361                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_nextkey: WARNING tdb_unlock failed!\n"));
362         tdb_trace_1rec_retrec(tdb, "tdb_nextkey", oldkey, key);
363         return key;
364 }
365