Fix traverse nesting unlock bug.
[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         return TDB_ERRCODE(TDB_SUCCESS, 0);
125
126  fail:
127         tlock->off = 0;
128         if (tdb_unlock(tdb, tlock->hash, tlock->lock_rw) != 0)
129                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_next_lock: On error unlock failed!\n"));
130         return -1;
131 }
132
133 /* traverse the entire database - calling fn(tdb, key, data) on each element.
134    return -1 on error or the record count traversed
135    if fn is NULL then it is not called
136    a non-zero return value from fn() indicates that the traversal should stop
137   */
138 static int tdb_traverse_internal(struct tdb_context *tdb, 
139                                  tdb_traverse_func fn, void *private_data,
140                                  struct tdb_traverse_lock *tl)
141 {
142         TDB_DATA key, dbuf;
143         struct list_struct rec;
144         int ret, count = 0;
145
146         /* This was in the initializaton, above, but the IRIX compiler
147          * did not like it.  crh
148          */
149         tl->next = tdb->travlocks.next;
150
151         /* fcntl locks don't stack: beware traverse inside traverse */
152         tdb->travlocks.next = tl;
153
154         /* tdb_next_lock places locks on the record returned, and its chain */
155         while ((ret = tdb_next_lock(tdb, tl, &rec)) > 0) {
156                 count++;
157                 /* now read the full record */
158                 key.dptr = tdb_alloc_read(tdb, tl->off + sizeof(rec), 
159                                           rec.key_len + rec.data_len);
160                 if (!key.dptr) {
161                         ret = -1;
162                         if (tdb_unlock(tdb, tl->hash, tl->lock_rw) != 0)
163                                 goto out;
164                         if (tdb_unlock_record(tdb, tl->off) != 0)
165                                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_traverse: key.dptr == NULL and unlock_record failed!\n"));
166                         goto out;
167                 }
168                 key.dsize = rec.key_len;
169                 dbuf.dptr = key.dptr + rec.key_len;
170                 dbuf.dsize = rec.data_len;
171
172                 tdb_trace(tdb, "traverse ");
173                 tdb_trace_record(tdb, key);
174                 tdb_trace_record(tdb, dbuf);
175                 tdb_trace(tdb, "\n");
176
177                 /* Drop chain lock, call out */
178                 if (tdb_unlock(tdb, tl->hash, tl->lock_rw) != 0) {
179                         ret = -1;
180                         SAFE_FREE(key.dptr);
181                         goto out;
182                 }
183                 if (fn && fn(tdb, key, dbuf, private_data)) {
184                         /* They want us to terminate traversal */
185                         tdb_trace(tdb, "tdb_traverse_end = %i\n", count);
186                         ret = count;
187                         if (tdb_unlock_record(tdb, tl->off) != 0) {
188                                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_traverse: unlock_record failed!\n"));;
189                                 ret = -1;
190                         }
191                         SAFE_FREE(key.dptr);
192                         goto out;
193                 }
194                 SAFE_FREE(key.dptr);
195         }
196         tdb_trace(tdb, "tdb_traverse_end\n");
197 out:
198         tdb->travlocks.next = tl->next;
199         if (ret < 0)
200                 return -1;
201         else
202                 return count;
203 }
204
205
206 /*
207   a write style traverse - temporarily marks the db read only
208 */
209 int tdb_traverse_read(struct tdb_context *tdb, 
210                       tdb_traverse_func fn, void *private_data)
211 {
212         struct tdb_traverse_lock tl = { NULL, 0, 0, F_RDLCK };
213         int ret;
214
215         /* we need to get a read lock on the transaction lock here to
216            cope with the lock ordering semantics of solaris10 */
217         if (tdb->traverse_read == 0 && tdb_transaction_lock(tdb, F_RDLCK)) {
218                 return -1;
219         }
220
221         tdb->traverse_read++;
222         tdb_trace(tdb, "tdb_traverse_read_start\n");
223         ret = tdb_traverse_internal(tdb, fn, private_data, &tl);
224         tdb->traverse_read--;
225
226         if (tdb->traverse_read == 0) {
227                 tdb_transaction_unlock(tdb);
228         }
229
230         return ret;
231 }
232
233 /*
234   a write style traverse - needs to get the transaction lock to
235   prevent deadlocks
236
237   WARNING: The data buffer given to the callback fn does NOT meet the
238   alignment restrictions malloc gives you.
239 */
240 int tdb_traverse(struct tdb_context *tdb, 
241                  tdb_traverse_func fn, void *private_data)
242 {
243         struct tdb_traverse_lock tl = { NULL, 0, 0, F_WRLCK };
244         int ret;
245
246         if (tdb->read_only || tdb->traverse_read) {
247                 return tdb_traverse_read(tdb, fn, private_data);
248         }
249
250         /* Nested traversals: transaction lock doesn't nest. */
251         if (tdb->traverse_write == 0 && tdb_transaction_lock(tdb, F_WRLCK)) {
252                 return -1;
253         }
254
255         tdb->traverse_write++;
256         tdb_trace(tdb, "tdb_traverse_start\n");
257         ret = tdb_traverse_internal(tdb, fn, private_data, &tl);
258         tdb->traverse_write--;
259
260         if (tdb->traverse_write == 0) {
261                 tdb_transaction_unlock(tdb);
262         }
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
274         /* release any old lock */
275         if (tdb_unlock_record(tdb, tdb->travlocks.off) != 0)
276                 return tdb_null;
277         tdb->travlocks.off = tdb->travlocks.hash = 0;
278         tdb->travlocks.lock_rw = F_RDLCK;
279
280         /* Grab first record: locks chain and returned record. */
281         if (tdb_next_lock(tdb, &tdb->travlocks, &rec) <= 0) {
282                 tdb_trace(tdb, "tdb_firstkey = ENOENT\n");
283                 return tdb_null;
284         }
285         /* now read the key */
286         key.dsize = rec.key_len;
287         key.dptr =tdb_alloc_read(tdb,tdb->travlocks.off+sizeof(rec),key.dsize);
288
289         tdb_trace(tdb, "tdb_firstkey = ");
290         tdb_trace_record(tdb, key);
291         tdb_trace(tdb, "\n");
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
307         /* Is locked key the old key?  If so, traverse will be reliable. */
308         if (tdb->travlocks.off) {
309                 if (tdb_lock(tdb,tdb->travlocks.hash,tdb->travlocks.lock_rw))
310                         return tdb_null;
311                 if (tdb_rec_read(tdb, tdb->travlocks.off, &rec) == -1
312                     || !(k = tdb_alloc_read(tdb,tdb->travlocks.off+sizeof(rec),
313                                             rec.key_len))
314                     || memcmp(k, oldkey.dptr, oldkey.dsize) != 0) {
315                         /* No, it wasn't: unlock it and start from scratch */
316                         if (tdb_unlock_record(tdb, tdb->travlocks.off) != 0) {
317                                 tdb_trace(tdb, "tdb_nextkey ");
318                                 tdb_trace_record(tdb, oldkey);
319                                 tdb_trace(tdb, "= ENOENT\n");
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(tdb, "tdb_nextkey ");
338                         tdb_trace_record(tdb, oldkey);
339                         tdb_trace(tdb, "= ENOENT\n");
340                         return tdb_null;
341                 }
342                 tdb->travlocks.hash = BUCKET(rec.full_hash);
343                 if (tdb_lock_record(tdb, tdb->travlocks.off) != 0) {
344                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_nextkey: lock_record failed (%s)!\n", strerror(errno)));
345                         return tdb_null;
346                 }
347         }
348         oldhash = tdb->travlocks.hash;
349
350         /* Grab next record: locks chain and returned record,
351            unlocks old record */
352         if (tdb_next_lock(tdb, &tdb->travlocks, &rec) > 0) {
353                 key.dsize = rec.key_len;
354                 key.dptr = tdb_alloc_read(tdb, tdb->travlocks.off+sizeof(rec),
355                                           key.dsize);
356                 /* Unlock the chain of this new record */
357                 if (tdb_unlock(tdb, tdb->travlocks.hash, tdb->travlocks.lock_rw) != 0)
358                         TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_nextkey: WARNING tdb_unlock failed!\n"));
359         }
360         /* Unlock the chain of old record */
361         if (tdb_unlock(tdb, BUCKET(oldhash), tdb->travlocks.lock_rw) != 0)
362                 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_nextkey: WARNING tdb_unlock failed!\n"));
363         tdb_trace(tdb, "tdb_nextkey ");
364         tdb_trace_record(tdb, oldkey);
365         tdb_trace(tdb, "= ");
366         tdb_trace_record(tdb, key);
367         tdb_trace(tdb, "\n");
368         return key;
369 }
370