Wean off TDB_ERRCODE.
[ccan] / ccan / tdb / 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 /* Uses traverse lock: 0 = finish, -1 = error, other = record offset */
31 static int tdb_next_lock(struct tdb_context *tdb, struct tdb_traverse_lock *tlock,
32                          struct list_struct *rec)
33 {
34         int want_next = (tlock->off != 0);
35
36         /* Lock each chain from the start one. */
37         for (; tlock->hash < tdb->header.hash_size; tlock->hash++) {
38                 if (!tlock->off && tlock->hash != 0) {
39                         /* this is an optimisation for the common case where
40                            the hash chain is empty, which is particularly
41                            common for the use of tdb with ldb, where large
42                            hashes are used. In that case we spend most of our
43                            time in tdb_brlock(), locking empty hash chains.
44                            
45                            To avoid this, we do an unlocked pre-check to see
46                            if the hash chain is empty before starting to look
47                            inside it. If it is empty then we can avoid that
48                            hash chain. If it isn't empty then we can't believe
49                            the value we get back, as we read it without a
50                            lock, so instead we get the lock and re-fetch the
51                            value below.
52                            
53                            Notice that not doing this optimisation on the
54                            first hash chain is critical. We must guarantee
55                            that we have done at least one fcntl lock at the
56                            start of a search to guarantee that memory is
57                            coherent on SMP systems. If records are added by
58                            others during the search then thats OK, and we
59                            could possibly miss those with this trick, but we
60                            could miss them anyway without this trick, so the
61                            semantics don't change.
62                            
63                            With a non-indexed ldb search this trick gains us a
64                            factor of around 80 in speed on a linux 2.6.x
65                            system (testing using ldbtest).
66                         */
67                         tdb->methods->next_hash_chain(tdb, &tlock->hash);
68                         if (tlock->hash == tdb->header.hash_size) {
69                                 continue;
70                         }
71                 }
72
73                 if (tdb_lock(tdb, tlock->hash, tlock->lock_rw) == -1)
74                         return -1;
75
76                 /* No previous record?  Start at top of chain. */
77                 if (!tlock->off) {
78                         if (tdb_ofs_read(tdb, TDB_HASH_TOP(tlock->hash),
79                                      &tlock->off) == -1)
80                                 goto fail;
81                 } else {
82                         /* Otherwise unlock the previous record. */
83                         if (tdb_unlock_record(tdb, tlock->off) != 0)
84                                 goto fail;
85                 }
86
87                 if (want_next) {
88                         /* We have offset of old record: grab next */
89                         if (tdb_rec_read(tdb, tlock->off, rec) == -1)
90                                 goto fail;
91                         tlock->off = rec->next;
92                 }
93
94                 /* Iterate through chain */
95                 while( tlock->off) {
96                         tdb_off_t current;
97                         if (tdb_rec_read(tdb, tlock->off, rec) == -1)
98                                 goto fail;
99
100                         /* Detect infinite loops. From "Shlomi Yaakobovich" <Shlomi@exanet.com>. */
101                         if (tlock->off == rec->next) {
102                                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_next_lock: loop detected.\n"));
103                                 goto fail;
104                         }
105
106                         if (!TDB_DEAD(rec)) {
107                                 /* Woohoo: we found one! */
108                                 if (tdb_lock_record(tdb, tlock->off) != 0)
109                                         goto fail;
110                                 return tlock->off;
111                         }
112
113                         /* Try to clean dead ones from old traverses */
114                         current = tlock->off;
115                         tlock->off = rec->next;
116                         if (!(tdb->read_only || tdb->traverse_read) && 
117                             tdb_do_delete(tdb, current, rec) != 0)
118                                 goto fail;
119                 }
120                 tdb_unlock(tdb, tlock->hash, tlock->lock_rw);
121                 want_next = 0;
122         }
123         /* We finished iteration without finding anything */
124         tdb->ecode = TDB_SUCCESS;
125         return 0;
126
127  fail:
128         tlock->off = 0;
129         if (tdb_unlock(tdb, tlock->hash, tlock->lock_rw) != 0)
130                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_next_lock: On error unlock failed!\n"));
131         return -1;
132 }
133
134 /* traverse the entire database - calling fn(tdb, key, data) on each element.
135    return -1 on error or the record count traversed
136    if fn is NULL then it is not called
137    a non-zero return value from fn() indicates that the traversal should stop
138   */
139 static int tdb_traverse_internal(struct tdb_context *tdb, 
140                                  tdb_traverse_func fn, void *private_data,
141                                  struct tdb_traverse_lock *tl)
142 {
143         TDB_DATA key, dbuf;
144         struct list_struct rec;
145         int ret, count = 0;
146
147         /* This was in the initializaton, above, but the IRIX compiler
148          * did not like it.  crh
149          */
150         tl->next = tdb->travlocks.next;
151
152         /* fcntl locks don't stack: beware traverse inside traverse */
153         tdb->travlocks.next = tl;
154
155         /* tdb_next_lock places locks on the record returned, and its chain */
156         while ((ret = tdb_next_lock(tdb, tl, &rec)) > 0) {
157                 count++;
158                 /* now read the full record */
159                 key.dptr = tdb_alloc_read(tdb, tl->off + sizeof(rec), 
160                                           rec.key_len + rec.data_len);
161                 if (!key.dptr) {
162                         ret = -1;
163                         if (tdb_unlock(tdb, tl->hash, tl->lock_rw) != 0)
164                                 goto out;
165                         if (tdb_unlock_record(tdb, tl->off) != 0)
166                                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_traverse: key.dptr == NULL and unlock_record failed!\n"));
167                         goto out;
168                 }
169                 key.dsize = rec.key_len;
170                 dbuf.dptr = key.dptr + rec.key_len;
171                 dbuf.dsize = rec.data_len;
172
173                 tdb_trace_1rec_retrec(tdb, "traverse", key, dbuf);
174
175                 /* Drop chain lock, call out */
176                 if (tdb_unlock(tdb, tl->hash, tl->lock_rw) != 0) {
177                         ret = -1;
178                         SAFE_FREE(key.dptr);
179                         goto out;
180                 }
181                 if (fn && fn(tdb, key, dbuf, private_data)) {
182                         /* They want us to terminate traversal */
183                         tdb_trace_ret(tdb, "tdb_traverse_end", count);
184                         ret = count;
185                         if (tdb_unlock_record(tdb, tl->off) != 0) {
186                                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_traverse: unlock_record failed!\n"));;
187                                 ret = -1;
188                         }
189                         SAFE_FREE(key.dptr);
190                         goto out;
191                 }
192                 SAFE_FREE(key.dptr);
193         }
194         tdb_trace(tdb, "tdb_traverse_end");
195 out:
196         tdb->travlocks.next = tl->next;
197         if (ret < 0)
198                 return -1;
199         else
200                 return count;
201 }
202
203
204 /*
205   a write style traverse - temporarily marks the db read only
206 */
207 int tdb_traverse_read(struct tdb_context *tdb, 
208                       tdb_traverse_func fn, void *private_data)
209 {
210         struct tdb_traverse_lock tl = { NULL, 0, 0, F_RDLCK };
211         int ret;
212
213         /* we need to get a read lock on the transaction lock here to
214            cope with the lock ordering semantics of solaris10 */
215         if (tdb_transaction_lock(tdb, F_RDLCK)) {
216                 return -1;
217         }
218
219         tdb->traverse_read++;
220         tdb_trace(tdb, "tdb_traverse_read_start");
221         ret = tdb_traverse_internal(tdb, fn, private_data, &tl);
222         tdb->traverse_read--;
223
224         tdb_transaction_unlock(tdb);
225
226         return ret;
227 }
228
229 /*
230   a write style traverse - needs to get the transaction lock to
231   prevent deadlocks
232
233   WARNING: The data buffer given to the callback fn does NOT meet the
234   alignment restrictions malloc gives you.
235 */
236 int tdb_traverse(struct tdb_context *tdb, 
237                  tdb_traverse_func fn, void *private_data)
238 {
239         struct tdb_traverse_lock tl = { NULL, 0, 0, F_WRLCK };
240         int ret;
241
242         if (tdb->read_only || tdb->traverse_read) {
243                 return tdb_traverse_read(tdb, fn, private_data);
244         }
245
246         if (tdb_transaction_lock(tdb, F_WRLCK)) {
247                 return -1;
248         }
249
250         tdb->traverse_write++;
251         tdb_trace(tdb, "tdb_traverse_start");
252         ret = tdb_traverse_internal(tdb, fn, private_data, &tl);
253         tdb->traverse_write--;
254
255         tdb_transaction_unlock(tdb);
256
257         return ret;
258 }
259
260
261 /* find the first entry in the database and return its key */
262 TDB_DATA tdb_firstkey(struct tdb_context *tdb)
263 {
264         TDB_DATA key;
265         struct list_struct rec;
266
267         /* release any old lock */
268         if (tdb_unlock_record(tdb, tdb->travlocks.off) != 0)
269                 return tdb_null;
270         tdb->travlocks.off = tdb->travlocks.hash = 0;
271         tdb->travlocks.lock_rw = F_RDLCK;
272
273         /* Grab first record: locks chain and returned record. */
274         if (tdb_next_lock(tdb, &tdb->travlocks, &rec) <= 0) {
275                 tdb_trace_retrec(tdb, "tdb_firstkey", tdb_null);
276                 return tdb_null;
277         }
278         /* now read the key */
279         key.dsize = rec.key_len;
280         key.dptr =tdb_alloc_read(tdb,tdb->travlocks.off+sizeof(rec),key.dsize);
281
282         tdb_trace_retrec(tdb, "tdb_firstkey", key);
283
284         /* Unlock the hash chain of the record we just read. */
285         if (tdb_unlock(tdb, tdb->travlocks.hash, tdb->travlocks.lock_rw) != 0)
286                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_firstkey: error occurred while tdb_unlocking!\n"));
287         return key;
288 }
289
290 /* find the next entry in the database, returning its key */
291 TDB_DATA tdb_nextkey(struct tdb_context *tdb, TDB_DATA oldkey)
292 {
293         uint32_t oldhash;
294         TDB_DATA key = tdb_null;
295         struct list_struct rec;
296         unsigned char *k = NULL;
297
298         /* Is locked key the old key?  If so, traverse will be reliable. */
299         if (tdb->travlocks.off) {
300                 if (tdb_lock(tdb,tdb->travlocks.hash,tdb->travlocks.lock_rw))
301                         return tdb_null;
302                 if (tdb_rec_read(tdb, tdb->travlocks.off, &rec) == -1
303                     || !(k = tdb_alloc_read(tdb,tdb->travlocks.off+sizeof(rec),
304                                             rec.key_len))
305                     || memcmp(k, oldkey.dptr, oldkey.dsize) != 0) {
306                         /* No, it wasn't: unlock it and start from scratch */
307                         if (tdb_unlock_record(tdb, tdb->travlocks.off) != 0) {
308                                 tdb_trace_1rec_retrec(tdb, "tdb_nextkey",
309                                                       oldkey, tdb_null);
310                                 SAFE_FREE(k);
311                                 return tdb_null;
312                         }
313                         if (tdb_unlock(tdb, tdb->travlocks.hash, tdb->travlocks.lock_rw) != 0) {
314                                 SAFE_FREE(k);
315                                 return tdb_null;
316                         }
317                         tdb->travlocks.off = 0;
318                 }
319
320                 SAFE_FREE(k);
321         }
322
323         if (!tdb->travlocks.off) {
324                 /* No previous element: do normal find, and lock record */
325                 tdb->travlocks.off = tdb_find_lock_hash(tdb, oldkey, tdb->hash_fn(&oldkey), tdb->travlocks.lock_rw, &rec);
326                 if (!tdb->travlocks.off) {
327                         tdb_trace_1rec_retrec(tdb, "tdb_nextkey", oldkey, tdb_null);
328                         return tdb_null;
329                 }
330                 tdb->travlocks.hash = BUCKET(rec.full_hash);
331                 if (tdb_lock_record(tdb, tdb->travlocks.off) != 0) {
332                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_nextkey: lock_record failed (%s)!\n", strerror(errno)));
333                         return tdb_null;
334                 }
335         }
336         oldhash = tdb->travlocks.hash;
337
338         /* Grab next record: locks chain and returned record,
339            unlocks old record */
340         if (tdb_next_lock(tdb, &tdb->travlocks, &rec) > 0) {
341                 key.dsize = rec.key_len;
342                 key.dptr = tdb_alloc_read(tdb, tdb->travlocks.off+sizeof(rec),
343                                           key.dsize);
344                 /* Unlock the chain of this new record */
345                 if (tdb_unlock(tdb, tdb->travlocks.hash, tdb->travlocks.lock_rw) != 0)
346                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_nextkey: WARNING tdb_unlock failed!\n"));
347         }
348         /* Unlock the chain of old record */
349         if (tdb_unlock(tdb, BUCKET(oldhash), tdb->travlocks.lock_rw) != 0)
350                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_nextkey: WARNING tdb_unlock failed!\n"));
351         tdb_trace_1rec_retrec(tdb, "tdb_nextkey", oldkey, key);
352         return key;
353 }
354