87e233b47fe3fa4551b11db6adee922fc242978e
[ccan] / ccan / tdb2 / tdb1_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 "tdb1_private.h"
29
30 #define TDB1_NEXT_LOCK_ERR ((tdb1_off_t)-1)
31
32 /* Uses traverse lock: 0 = finish, TDB1_NEXT_LOCK_ERR = error,
33    other = record offset */
34 static tdb1_off_t tdb1_next_lock(struct tdb_context *tdb, struct tdb1_traverse_lock *tlock,
35                          struct tdb1_record *rec)
36 {
37         int want_next = (tlock->off != 0);
38
39         /* Lock each chain from the start one. */
40         for (; tlock->hash < tdb->tdb1.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 tdb1_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->tdb1.io->next_hash_chain(tdb, &tlock->hash);
71                         if (tlock->hash == tdb->tdb1.header.hash_size) {
72                                 continue;
73                         }
74                 }
75
76                 if (tdb1_lock(tdb, tlock->hash, tlock->lock_rw) == -1)
77                         return TDB1_NEXT_LOCK_ERR;
78
79                 /* No previous record?  Start at top of chain. */
80                 if (!tlock->off) {
81                         if (tdb1_ofs_read(tdb, TDB1_HASH_TOP(tlock->hash),
82                                      &tlock->off) == -1)
83                                 goto fail;
84                 } else {
85                         /* Otherwise unlock the previous record. */
86                         if (tdb1_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 (tdb1_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                         tdb1_off_t current;
100                         if (tdb1_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->last_error = tdb_logerr(tdb, TDB_ERR_CORRUPT,
106                                                         TDB_LOG_ERROR,
107                                                         "tdb1_next_lock:"
108                                                         " loop detected.");
109                                 goto fail;
110                         }
111
112                         if (!TDB1_DEAD(rec)) {
113                                 /* Woohoo: we found one! */
114                                 if (tdb1_lock_record(tdb, tlock->off) != 0)
115                                         goto fail;
116                                 return tlock->off;
117                         }
118
119                         /* Try to clean dead ones from old traverses */
120                         current = tlock->off;
121                         tlock->off = rec->next;
122                         if (!((tdb->flags & TDB_RDONLY) || tdb->tdb1.traverse_read) &&
123                             tdb1_do_delete(tdb, current, rec) != 0)
124                                 goto fail;
125                 }
126                 tdb1_unlock(tdb, tlock->hash, tlock->lock_rw);
127                 want_next = 0;
128         }
129         /* We finished iteration without finding anything */
130         tdb->last_error = TDB_SUCCESS;
131         return 0;
132
133  fail:
134         tlock->off = 0;
135         if (tdb1_unlock(tdb, tlock->hash, tlock->lock_rw) != 0)
136                 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
137                            "tdb1_next_lock: On error unlock failed!");
138         return TDB1_NEXT_LOCK_ERR;
139 }
140
141 /* traverse the entire database - calling fn(tdb, key, data) on each element.
142    return -1 on error or the record count traversed
143    if fn is NULL then it is not called
144    a non-zero return value from fn() indicates that the traversal should stop
145   */
146 static int tdb1_traverse_internal(struct tdb_context *tdb,
147                                  tdb1_traverse_func fn, void *private_data,
148                                  struct tdb1_traverse_lock *tl)
149 {
150         TDB_DATA key, dbuf;
151         struct tdb1_record rec;
152         int ret = 0, count = 0;
153         tdb1_off_t off;
154
155         /* This was in the initializaton, above, but the IRIX compiler
156          * did not like it.  crh
157          */
158         tl->next = tdb->tdb1.travlocks.next;
159
160         /* fcntl locks don't stack: beware traverse inside traverse */
161         tdb->tdb1.travlocks.next = tl;
162
163         /* tdb1_next_lock places locks on the record returned, and its chain */
164         while ((off = tdb1_next_lock(tdb, tl, &rec)) != 0) {
165                 if (off == TDB1_NEXT_LOCK_ERR) {
166                         ret = -1;
167                         goto out;
168                 }
169                 count++;
170                 /* now read the full record */
171                 key.dptr = tdb1_alloc_read(tdb, tl->off + sizeof(rec),
172                                           rec.key_len + rec.data_len);
173                 if (!key.dptr) {
174                         ret = -1;
175                         if (tdb1_unlock(tdb, tl->hash, tl->lock_rw) != 0)
176                                 goto out;
177                         if (tdb1_unlock_record(tdb, tl->off) != 0)
178                                 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
179                                            "tdb1_traverse: key.dptr == NULL and"
180                                            " unlock_record failed!");
181                         goto out;
182                 }
183                 key.dsize = rec.key_len;
184                 dbuf.dptr = key.dptr + rec.key_len;
185                 dbuf.dsize = rec.data_len;
186
187                 /* Drop chain lock, call out */
188                 if (tdb1_unlock(tdb, tl->hash, tl->lock_rw) != 0) {
189                         ret = -1;
190                         SAFE_FREE(key.dptr);
191                         goto out;
192                 }
193                 if (fn && fn(tdb, key, dbuf, private_data)) {
194                         /* They want us to terminate traversal */
195                         if (tdb1_unlock_record(tdb, tl->off) != 0) {
196                                 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
197                                            "tdb1_traverse:"
198                                            " unlock_record failed!");
199                                 ret = -1;
200                         }
201                         SAFE_FREE(key.dptr);
202                         goto out;
203                 }
204                 SAFE_FREE(key.dptr);
205         }
206 out:
207         tdb->tdb1.travlocks.next = tl->next;
208         if (ret < 0)
209                 return -1;
210         else
211                 return count;
212 }
213
214
215 /*
216   a write style traverse - temporarily marks the db read only
217 */
218 int tdb1_traverse_read(struct tdb_context *tdb,
219                       tdb1_traverse_func fn, void *private_data)
220 {
221         struct tdb1_traverse_lock tl = { NULL, 0, 0, F_RDLCK };
222         int ret;
223
224         /* we need to get a read lock on the transaction lock here to
225            cope with the lock ordering semantics of solaris10 */
226         if (tdb1_transaction_lock(tdb, F_RDLCK, TDB_LOCK_WAIT)) {
227                 return -1;
228         }
229
230         tdb->tdb1.traverse_read++;
231         ret = tdb1_traverse_internal(tdb, fn, private_data, &tl);
232         tdb->tdb1.traverse_read--;
233
234         tdb1_transaction_unlock(tdb, F_RDLCK);
235
236         return ret;
237 }
238
239 /*
240   a write style traverse - needs to get the transaction lock to
241   prevent deadlocks
242
243   WARNING: The data buffer given to the callback fn does NOT meet the
244   alignment restrictions malloc gives you.
245 */
246 int tdb1_traverse(struct tdb_context *tdb,
247                  tdb1_traverse_func fn, void *private_data)
248 {
249         struct tdb1_traverse_lock tl = { NULL, 0, 0, F_WRLCK };
250         int ret;
251
252         if ((tdb->flags & TDB_RDONLY) || tdb->tdb1.traverse_read) {
253                 return tdb1_traverse_read(tdb, fn, private_data);
254         }
255
256         if (tdb1_transaction_lock(tdb, F_WRLCK, TDB_LOCK_WAIT)) {
257                 return -1;
258         }
259
260         tdb->tdb1.traverse_write++;
261         ret = tdb1_traverse_internal(tdb, fn, private_data, &tl);
262         tdb->tdb1.traverse_write--;
263
264         tdb1_transaction_unlock(tdb, F_WRLCK);
265
266         return ret;
267 }
268
269
270 /* find the first entry in the database and return its key */
271 TDB_DATA tdb1_firstkey(struct tdb_context *tdb)
272 {
273         TDB_DATA key;
274         struct tdb1_record rec;
275         tdb1_off_t off;
276
277         /* release any old lock */
278         if (tdb1_unlock_record(tdb, tdb->tdb1.travlocks.off) != 0)
279                 return tdb1_null;
280         tdb->tdb1.travlocks.off = tdb->tdb1.travlocks.hash = 0;
281         tdb->tdb1.travlocks.lock_rw = F_RDLCK;
282
283         /* Grab first record: locks chain and returned record. */
284         off = tdb1_next_lock(tdb, &tdb->tdb1.travlocks, &rec);
285         if (off == 0 || off == TDB1_NEXT_LOCK_ERR) {
286                 return tdb1_null;
287         }
288         /* now read the key */
289         key.dsize = rec.key_len;
290         key.dptr =tdb1_alloc_read(tdb,tdb->tdb1.travlocks.off+sizeof(rec),key.dsize);
291
292         /* Unlock the hash chain of the record we just read. */
293         if (tdb1_unlock(tdb, tdb->tdb1.travlocks.hash, tdb->tdb1.travlocks.lock_rw) != 0)
294                 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
295                            "tdb1_firstkey:"
296                            " error occurred while tdb1_unlocking!");
297         return key;
298 }
299
300 /* find the next entry in the database, returning its key */
301 TDB_DATA tdb1_nextkey(struct tdb_context *tdb, TDB_DATA oldkey)
302 {
303         uint32_t oldhash;
304         TDB_DATA key = tdb1_null;
305         struct tdb1_record rec;
306         unsigned char *k = NULL;
307         tdb1_off_t off;
308
309         /* Is locked key the old key?  If so, traverse will be reliable. */
310         if (tdb->tdb1.travlocks.off) {
311                 if (tdb1_lock(tdb,tdb->tdb1.travlocks.hash,tdb->tdb1.travlocks.lock_rw))
312                         return tdb1_null;
313                 if (tdb1_rec_read(tdb, tdb->tdb1.travlocks.off, &rec) == -1
314                     || !(k = tdb1_alloc_read(tdb,tdb->tdb1.travlocks.off+sizeof(rec),
315                                             rec.key_len))
316                     || memcmp(k, oldkey.dptr, oldkey.dsize) != 0) {
317                         /* No, it wasn't: unlock it and start from scratch */
318                         if (tdb1_unlock_record(tdb, tdb->tdb1.travlocks.off) != 0) {
319                                 SAFE_FREE(k);
320                                 return tdb1_null;
321                         }
322                         if (tdb1_unlock(tdb, tdb->tdb1.travlocks.hash, tdb->tdb1.travlocks.lock_rw) != 0) {
323                                 SAFE_FREE(k);
324                                 return tdb1_null;
325                         }
326                         tdb->tdb1.travlocks.off = 0;
327                 }
328
329                 SAFE_FREE(k);
330         }
331
332         if (!tdb->tdb1.travlocks.off) {
333                 /* No previous element: do normal find, and lock record */
334                 tdb->tdb1.travlocks.off = tdb1_find_lock_hash(tdb, oldkey, tdb_hash(tdb, oldkey.dptr, oldkey.dsize), tdb->tdb1.travlocks.lock_rw, &rec);
335                 if (!tdb->tdb1.travlocks.off) {
336                         return tdb1_null;
337                 }
338                 tdb->tdb1.travlocks.hash = TDB1_BUCKET(rec.full_hash);
339                 if (tdb1_lock_record(tdb, tdb->tdb1.travlocks.off) != 0) {
340                         tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
341                                    "tdb1_nextkey: lock_record failed (%s)!",
342                                    strerror(errno));
343                         return tdb1_null;
344                 }
345         }
346         oldhash = tdb->tdb1.travlocks.hash;
347
348         /* Grab next record: locks chain and returned record,
349            unlocks old record */
350         off = tdb1_next_lock(tdb, &tdb->tdb1.travlocks, &rec);
351         if (off != TDB1_NEXT_LOCK_ERR && off != 0) {
352                 key.dsize = rec.key_len;
353                 key.dptr = tdb1_alloc_read(tdb, tdb->tdb1.travlocks.off+sizeof(rec),
354                                           key.dsize);
355                 /* Unlock the chain of this new record */
356                 if (tdb1_unlock(tdb, tdb->tdb1.travlocks.hash, tdb->tdb1.travlocks.lock_rw) != 0)
357                         tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
358                                    "tdb1_nextkey: WARNING tdb1_unlock failed!");
359         }
360         /* Unlock the chain of old record */
361         if (tdb1_unlock(tdb, TDB1_BUCKET(oldhash), tdb->tdb1.travlocks.lock_rw) != 0)
362                 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
363                            "tdb1_nextkey: WARNING tdb1_unlock failed!");
364         return key;
365 }