tdb2: catch errors in tdb1_needs_recovery()
[ccan] / ccan / tdb2 / tdb1_lock.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 /* list -1 is the alloc list, otherwise a hash chain. */
31 static tdb1_off_t lock_offset(int list)
32 {
33         return TDB1_FREELIST_TOP + 4*list;
34 }
35
36 /* a byte range locking function - return 0 on success
37    this functions locks/unlocks 1 byte at the specified offset.
38
39    On error, errno is also set so that errors are passed back properly
40    through tdb1_open().
41
42    note that a len of zero means lock to end of file
43 */
44 int tdb1_brlock(struct tdb_context *tdb,
45                int rw_type, tdb1_off_t offset, size_t len,
46                enum tdb_lock_flags flags)
47 {
48         enum TDB_ERROR ecode = tdb_brlock(tdb, rw_type, offset, len, flags
49                                           | TDB_LOCK_NOCHECK);
50         if (ecode == TDB_SUCCESS)
51                 return 0;
52         tdb->last_error = ecode;
53         return -1;
54 }
55
56 int tdb1_brunlock(struct tdb_context *tdb,
57                  int rw_type, tdb1_off_t offset, size_t len)
58 {
59         enum TDB_ERROR ecode = tdb_brunlock(tdb, rw_type, offset, len);
60         if (ecode == TDB_SUCCESS)
61                 return 0;
62         tdb->last_error = ecode;
63         return -1;
64 }
65
66 int tdb1_allrecord_upgrade(struct tdb_context *tdb)
67 {
68         enum TDB_ERROR ecode = tdb_allrecord_upgrade(tdb, TDB1_FREELIST_TOP);
69         if (ecode == TDB_SUCCESS)
70                 return 0;
71         tdb->last_error = ecode;
72         return -1;
73 }
74
75 static struct tdb_lock *tdb1_find_nestlock(struct tdb_context *tdb,
76                                            tdb1_off_t offset)
77 {
78         unsigned int i;
79
80         for (i=0; i<tdb->file->num_lockrecs; i++) {
81                 if (tdb->file->lockrecs[i].off == offset) {
82                         return &tdb->file->lockrecs[i];
83                 }
84         }
85         return NULL;
86 }
87
88 /* lock an offset in the database. */
89 int tdb1_nest_lock(struct tdb_context *tdb, uint32_t offset, int ltype,
90                   enum tdb_lock_flags flags)
91 {
92         enum TDB_ERROR ecode;
93
94         if (offset >= lock_offset(tdb->tdb1.header.hash_size)) {
95                 tdb->last_error = tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
96                                         "tdb1_lock: invalid offset %u for"
97                                         " ltype=%d",
98                                         offset, ltype);
99                 return -1;
100         }
101
102         ecode = tdb_nest_lock(tdb, offset, ltype, flags | TDB_LOCK_NOCHECK);
103         if (unlikely(ecode != TDB_SUCCESS)) {
104                 tdb->last_error = ecode;
105                 return -1;
106         }
107         return 0;
108 }
109
110 static int tdb1_lock_and_recover(struct tdb_context *tdb)
111 {
112         int ret;
113
114         /* We need to match locking order in transaction commit. */
115         if (tdb1_brlock(tdb, F_WRLCK, TDB1_FREELIST_TOP, 0,
116                         TDB_LOCK_WAIT|TDB_LOCK_NOCHECK)) {
117                 return -1;
118         }
119
120         if (tdb1_brlock(tdb, F_WRLCK, TDB1_OPEN_LOCK, 1,
121                         TDB_LOCK_WAIT|TDB_LOCK_NOCHECK)) {
122                 tdb1_brunlock(tdb, F_WRLCK, TDB1_FREELIST_TOP, 0);
123                 return -1;
124         }
125
126         ret = tdb1_transaction_recover(tdb);
127
128         tdb1_brunlock(tdb, F_WRLCK, TDB1_OPEN_LOCK, 1);
129         tdb1_brunlock(tdb, F_WRLCK, TDB1_FREELIST_TOP, 0);
130
131         return ret;
132 }
133
134 static bool have_data_locks(const struct tdb_context *tdb)
135 {
136         unsigned int i;
137
138         for (i = 0; i < tdb->file->num_lockrecs; i++) {
139                 if (tdb->file->lockrecs[i].off >= lock_offset(-1))
140                         return true;
141         }
142         return false;
143 }
144
145 static int tdb1_lock_list(struct tdb_context *tdb, int list, int ltype,
146                          enum tdb_lock_flags waitflag)
147 {
148         int ret;
149         bool check = false;
150
151         /* a allrecord lock allows us to avoid per chain locks */
152         if (tdb->file->allrecord_lock.count &&
153             (ltype == tdb->file->allrecord_lock.ltype || ltype == F_RDLCK)) {
154                 return 0;
155         }
156
157         if (tdb->file->allrecord_lock.count) {
158                 tdb->last_error = TDB_ERR_LOCK;
159                 ret = -1;
160         } else {
161                 /* Only check when we grab first data lock. */
162                 check = !have_data_locks(tdb);
163                 ret = tdb1_nest_lock(tdb, lock_offset(list), ltype, waitflag);
164
165                 if (ret == 0 && check) {
166                         tdb_bool_err berr = tdb1_needs_recovery(tdb);
167
168                         if (berr < 0) {
169                                 return -1;
170                         }
171                         if (berr == true) {
172                                 tdb1_nest_unlock(tdb, lock_offset(list), ltype);
173
174                                 if (tdb1_lock_and_recover(tdb) == -1) {
175                                         return -1;
176                                 }
177                                 return tdb1_lock_list(tdb, list, ltype,
178                                                       waitflag);
179                         }
180                 }
181         }
182         return ret;
183 }
184
185 /* lock a list in the database. list -1 is the alloc list */
186 int tdb1_lock(struct tdb_context *tdb, int list, int ltype)
187 {
188         int ret;
189
190         ret = tdb1_lock_list(tdb, list, ltype, TDB_LOCK_WAIT);
191         /* Don't log for EAGAIN and EINTR: they could have overridden lock fns */
192         if (ret && errno != EAGAIN && errno != EINTR) {
193                 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
194                            "tdb1_lock failed on list %d "
195                            "ltype=%d (%s)",  list, ltype, strerror(errno));
196         }
197         return ret;
198 }
199
200 int tdb1_nest_unlock(struct tdb_context *tdb, uint32_t offset, int ltype)
201 {
202         enum TDB_ERROR ecode;
203
204         /* Sanity checks */
205         if (offset >= lock_offset(tdb->tdb1.header.hash_size)) {
206                 tdb->last_error = tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
207                                         "tdb1_unlock: offset %u invalid (%d)",
208                                         offset, tdb->tdb1.header.hash_size);
209                 return -1;
210         }
211
212         ecode = tdb_nest_unlock(tdb, offset, ltype);
213         if (unlikely(ecode != TDB_SUCCESS)) {
214                 tdb->last_error = ecode;
215                 return -1;
216         }
217         return 0;
218 }
219
220 int tdb1_unlock(struct tdb_context *tdb, int list, int ltype)
221 {
222         /* a global lock allows us to avoid per chain locks */
223         if (tdb->file->allrecord_lock.count &&
224             (ltype == tdb->file->allrecord_lock.ltype || ltype == F_RDLCK)) {
225                 return 0;
226         }
227
228         if (tdb->file->allrecord_lock.count) {
229                 tdb->last_error = TDB_ERR_LOCK;
230                 return -1;
231         }
232
233         return tdb1_nest_unlock(tdb, lock_offset(list), ltype);
234 }
235
236 /*
237   get the transaction lock
238  */
239 int tdb1_transaction_lock(struct tdb_context *tdb, int ltype,
240                          enum tdb_lock_flags lockflags)
241 {
242         return tdb1_nest_lock(tdb, TDB1_TRANSACTION_LOCK, ltype, lockflags);
243 }
244
245 /*
246   release the transaction lock
247  */
248 int tdb1_transaction_unlock(struct tdb_context *tdb, int ltype)
249 {
250         return tdb1_nest_unlock(tdb, TDB1_TRANSACTION_LOCK, ltype);
251 }
252
253 /* lock/unlock entire database.  It can only be upgradable if you have some
254  * other way of guaranteeing exclusivity (ie. transaction write lock).
255  * We do the locking gradually to avoid being starved by smaller locks. */
256 int tdb1_allrecord_lock(struct tdb_context *tdb, int ltype,
257                        enum tdb_lock_flags flags, bool upgradable)
258 {
259         enum TDB_ERROR ecode;
260         tdb_bool_err berr;
261
262         /* tdb_lock_gradual() doesn't know about tdb->tdb1.traverse_read. */
263         if (tdb->tdb1.traverse_read && !(tdb->flags & TDB_NOLOCK)) {
264                 tdb->last_error = tdb_logerr(tdb, TDB_ERR_LOCK,
265                                              TDB_LOG_USE_ERROR,
266                                              "tdb1_allrecord_lock during"
267                                              " tdb1_read_traverse");
268                 return -1;
269         }
270
271         if (tdb->file->allrecord_lock.count
272             && tdb->file->allrecord_lock.ltype == ltype) {
273                 tdb->file->allrecord_lock.count++;
274                 return 0;
275         }
276
277         if (tdb1_have_extra_locks(tdb)) {
278                 /* can't combine global and chain locks */
279                 tdb->last_error = tdb_logerr(tdb, TDB_ERR_LOCK,
280                                              TDB_LOG_USE_ERROR,
281                                              "tdb1_allrecord_lock holding"
282                                              " other locks");
283                 return -1;
284         }
285
286         if (upgradable && ltype != F_RDLCK) {
287                 /* tdb error: you can't upgrade a write lock! */
288                 tdb->last_error = tdb_logerr(tdb, TDB_ERR_LOCK,
289                                              TDB_LOG_ERROR,
290                                              "tdb1_allrecord_lock cannot"
291                                              " have upgradable write lock");
292                 return -1;
293         }
294
295         /* We cover two kinds of locks:
296          * 1) Normal chain locks.  Taken for almost all operations.
297          * 3) Individual records locks.  Taken after normal or free
298          *    chain locks.
299          *
300          * It is (1) which cause the starvation problem, so we're only
301          * gradual for that. */
302         ecode = tdb_lock_gradual(tdb, ltype, flags | TDB_LOCK_NOCHECK,
303                                  TDB1_FREELIST_TOP, tdb->tdb1.header.hash_size * 4);
304         if (ecode != TDB_SUCCESS) {
305                 tdb->last_error = ecode;
306                 return -1;
307         }
308
309         /* Grab individual record locks. */
310         if (tdb1_brlock(tdb, ltype, lock_offset(tdb->tdb1.header.hash_size), 0,
311                        flags) == -1) {
312                 tdb1_brunlock(tdb, ltype, TDB1_FREELIST_TOP,
313                              tdb->tdb1.header.hash_size * 4);
314                 return -1;
315         }
316
317         /* FIXME: Temporary cast. */
318         tdb->file->allrecord_lock.owner = (void *)(struct tdb_context *)tdb;
319         tdb->file->allrecord_lock.count = 1;
320         /* If it's upgradable, it's actually exclusive so we can treat
321          * it as a write lock. */
322         tdb->file->allrecord_lock.ltype = upgradable ? F_WRLCK : ltype;
323         tdb->file->allrecord_lock.off = upgradable;
324
325         berr = tdb1_needs_recovery(tdb);
326         if (berr < 0) {
327                 return -1;
328         }
329
330         if (berr == true) {
331                 tdb1_allrecord_unlock(tdb, ltype);
332                 if (tdb1_lock_and_recover(tdb) == -1) {
333                         return -1;
334                 }
335                 return tdb1_allrecord_lock(tdb, ltype, flags, upgradable);
336         }
337
338         return 0;
339 }
340
341
342
343 /* unlock entire db */
344 int tdb1_allrecord_unlock(struct tdb_context *tdb, int ltype)
345 {
346         /* Don't try this during r/o traversal! */
347         if (tdb->tdb1.traverse_read) {
348                 tdb->last_error = TDB_ERR_LOCK;
349                 return -1;
350         }
351
352         if (tdb->file->allrecord_lock.count == 0) {
353                 tdb->last_error = TDB_ERR_LOCK;
354                 return -1;
355         }
356
357         /* Upgradable locks are marked as write locks. */
358         if (tdb->file->allrecord_lock.ltype != ltype
359             && (!tdb->file->allrecord_lock.off || ltype != F_RDLCK)) {
360                 tdb->last_error = TDB_ERR_LOCK;
361                 return -1;
362         }
363
364         if (tdb->file->allrecord_lock.count > 1) {
365                 tdb->file->allrecord_lock.count--;
366                 return 0;
367         }
368
369         if (tdb1_brunlock(tdb, ltype, TDB1_FREELIST_TOP, 0)) {
370                 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
371                            "tdb1_unlockall failed (%s)", strerror(errno));
372                 return -1;
373         }
374
375         tdb->file->allrecord_lock.count = 0;
376         tdb->file->allrecord_lock.ltype = 0;
377
378         return 0;
379 }
380
381 /* lock/unlock one hash chain. This is meant to be used to reduce
382    contention - it cannot guarantee how many records will be locked */
383 int tdb1_chainlock(struct tdb_context *tdb, TDB_DATA key)
384 {
385         int ret = tdb1_lock(tdb,
386                             TDB1_BUCKET(tdb_hash(tdb, key.dptr, key.dsize)),
387                             F_WRLCK);
388         return ret;
389 }
390
391 int tdb1_chainunlock(struct tdb_context *tdb, TDB_DATA key)
392 {
393         return tdb1_unlock(tdb, TDB1_BUCKET(tdb_hash(tdb, key.dptr, key.dsize)),
394                            F_WRLCK);
395 }
396
397 int tdb1_chainlock_read(struct tdb_context *tdb, TDB_DATA key)
398 {
399         int ret;
400         ret = tdb1_lock(tdb, TDB1_BUCKET(tdb_hash(tdb, key.dptr, key.dsize)),
401                         F_RDLCK);
402         return ret;
403 }
404
405 int tdb1_chainunlock_read(struct tdb_context *tdb, TDB_DATA key)
406 {
407         return tdb1_unlock(tdb, TDB1_BUCKET(tdb_hash(tdb, key.dptr, key.dsize)),
408                            F_RDLCK);
409 }
410
411 /* record lock stops delete underneath */
412 int tdb1_lock_record(struct tdb_context *tdb, tdb1_off_t off)
413 {
414         if (tdb->file->allrecord_lock.count) {
415                 return 0;
416         }
417         return off ? tdb1_brlock(tdb, F_RDLCK, off, 1, TDB_LOCK_WAIT) : 0;
418 }
419
420 /*
421   Write locks override our own fcntl readlocks, so check it here.
422   Note this is meant to be F_SETLK, *not* F_SETLKW, as it's not
423   an error to fail to get the lock here.
424 */
425 int tdb1_write_lock_record(struct tdb_context *tdb, tdb1_off_t off)
426 {
427         struct tdb1_traverse_lock *i;
428         for (i = &tdb->tdb1.travlocks; i; i = i->next)
429                 if (i->off == off)
430                         return -1;
431         if (tdb->file->allrecord_lock.count) {
432                 if (tdb->file->allrecord_lock.ltype == F_WRLCK) {
433                         return 0;
434                 }
435                 return -1;
436         }
437         return tdb1_brlock(tdb, F_WRLCK, off, 1, TDB_LOCK_NOWAIT|TDB_LOCK_PROBE);
438 }
439
440 int tdb1_write_unlock_record(struct tdb_context *tdb, tdb1_off_t off)
441 {
442         if (tdb->file->allrecord_lock.count) {
443                 return 0;
444         }
445         return tdb1_brunlock(tdb, F_WRLCK, off, 1);
446 }
447
448 /* fcntl locks don't stack: avoid unlocking someone else's */
449 int tdb1_unlock_record(struct tdb_context *tdb, tdb1_off_t off)
450 {
451         struct tdb1_traverse_lock *i;
452         uint32_t count = 0;
453
454         if (tdb->file->allrecord_lock.count) {
455                 return 0;
456         }
457
458         if (off == 0)
459                 return 0;
460         for (i = &tdb->tdb1.travlocks; i; i = i->next)
461                 if (i->off == off)
462                         count++;
463         return (count == 1 ? tdb1_brunlock(tdb, F_RDLCK, off, 1) : 0);
464 }
465
466 bool tdb1_have_extra_locks(struct tdb_context *tdb)
467 {
468         unsigned int extra = tdb->file->num_lockrecs;
469
470         /* A transaction holds the lock for all records. */
471         if (!tdb->tdb1.transaction && tdb->file->allrecord_lock.count) {
472                 return true;
473         }
474
475         /* We always hold the active lock if CLEAR_IF_FIRST. */
476         if (tdb1_find_nestlock(tdb, TDB1_ACTIVE_LOCK)) {
477                 extra--;
478         }
479
480         /* In a transaction, we expect to hold the transaction lock */
481         if (tdb->tdb1.transaction
482             && tdb1_find_nestlock(tdb, TDB1_TRANSACTION_LOCK)) {
483                 extra--;
484         }
485
486         return extra;
487 }
488
489 /* The transaction code uses this to remove all locks. */
490 void tdb1_release_transaction_locks(struct tdb_context *tdb)
491 {
492         unsigned int i, active = 0;
493
494         if (tdb->file->allrecord_lock.count != 0) {
495                 tdb1_brunlock(tdb, tdb->file->allrecord_lock.ltype, TDB1_FREELIST_TOP, 0);
496                 tdb->file->allrecord_lock.count = 0;
497         }
498
499         for (i=0;i<tdb->file->num_lockrecs;i++) {
500                 struct tdb_lock *lck = &tdb->file->lockrecs[i];
501
502                 /* Don't release the active lock!  Copy it to first entry. */
503                 if (lck->off == TDB1_ACTIVE_LOCK) {
504                         tdb->file->lockrecs[active++] = *lck;
505                 } else {
506                         tdb1_brunlock(tdb, lck->ltype, lck->off, 1);
507                 }
508         }
509         tdb->file->num_lockrecs = active;
510         if (tdb->file->num_lockrecs == 0) {
511                 SAFE_FREE(tdb->file->lockrecs);
512         }
513 }