tdb2: move mmap into struct tdb_file
[ccan] / ccan / tdb2 / 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 "private.h"
29 #include <assert.h>
30 #include <ccan/build_assert/build_assert.h>
31
32 static int fcntl_lock(struct tdb_context *tdb,
33                       int rw, off_t off, off_t len, bool waitflag)
34 {
35         struct flock fl;
36
37         fl.l_type = rw;
38         fl.l_whence = SEEK_SET;
39         fl.l_start = off;
40         fl.l_len = len;
41         fl.l_pid = 0;
42
43         add_stat(tdb, lock_lowlevel, 1);
44         if (waitflag)
45                 return fcntl(tdb->file->fd, F_SETLKW, &fl);
46         else {
47                 add_stat(tdb, lock_nonblock, 1);
48                 return fcntl(tdb->file->fd, F_SETLK, &fl);
49         }
50 }
51
52 static int fcntl_unlock(struct tdb_context *tdb, int rw, off_t off, off_t len)
53 {
54         struct flock fl;
55 #if 0 /* Check they matched up locks and unlocks correctly. */
56         char line[80];
57         FILE *locks;
58         bool found = false;
59
60         locks = fopen("/proc/locks", "r");
61
62         while (fgets(line, 80, locks)) {
63                 char *p;
64                 int type, start, l;
65
66                 /* eg. 1: FLOCK  ADVISORY  WRITE 2440 08:01:2180826 0 EOF */
67                 p = strchr(line, ':') + 1;
68                 if (strncmp(p, " POSIX  ADVISORY  ", strlen(" POSIX  ADVISORY  ")))
69                         continue;
70                 p += strlen(" FLOCK  ADVISORY  ");
71                 if (strncmp(p, "READ  ", strlen("READ  ")) == 0)
72                         type = F_RDLCK;
73                 else if (strncmp(p, "WRITE ", strlen("WRITE ")) == 0)
74                         type = F_WRLCK;
75                 else
76                         abort();
77                 p += 6;
78                 if (atoi(p) != getpid())
79                         continue;
80                 p = strchr(strchr(p, ' ') + 1, ' ') + 1;
81                 start = atoi(p);
82                 p = strchr(p, ' ') + 1;
83                 if (strncmp(p, "EOF", 3) == 0)
84                         l = 0;
85                 else
86                         l = atoi(p) - start + 1;
87
88                 if (off == start) {
89                         if (len != l) {
90                                 fprintf(stderr, "Len %u should be %u: %s",
91                                         (int)len, l, line);
92                                 abort();
93                         }
94                         if (type != rw) {
95                                 fprintf(stderr, "Type %s wrong: %s",
96                                         rw == F_RDLCK ? "READ" : "WRITE", line);
97                                 abort();
98                         }
99                         found = true;
100                         break;
101                 }
102         }
103
104         if (!found) {
105                 fprintf(stderr, "Unlock on %u@%u not found!",
106                         (int)off, (int)len);
107                 abort();
108         }
109
110         fclose(locks);
111 #endif
112
113         fl.l_type = F_UNLCK;
114         fl.l_whence = SEEK_SET;
115         fl.l_start = off;
116         fl.l_len = len;
117         fl.l_pid = 0;
118
119         return fcntl(tdb->file->fd, F_SETLKW, &fl);
120 }
121
122 /* a byte range locking function - return 0 on success
123    this functions locks len bytes at the specified offset.
124
125    note that a len of zero means lock to end of file
126 */
127 static enum TDB_ERROR tdb_brlock(struct tdb_context *tdb,
128                                  int rw_type, tdb_off_t offset, tdb_off_t len,
129                                  enum tdb_lock_flags flags)
130 {
131         int ret;
132
133         if (tdb->flags & TDB_NOLOCK) {
134                 return TDB_SUCCESS;
135         }
136
137         if (rw_type == F_WRLCK && tdb->read_only) {
138                 return tdb_logerr(tdb, TDB_ERR_RDONLY, TDB_LOG_USE_ERROR,
139                                   "Write lock attempted on read-only database");
140         }
141
142         /* A 32 bit system cannot open a 64-bit file, but it could have
143          * expanded since then: check here. */
144         if ((size_t)(offset + len) != offset + len) {
145                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
146                                   "tdb_brlock: lock on giant offset %llu",
147                                   (long long)(offset + len));
148         }
149
150         do {
151                 ret = fcntl_lock(tdb, rw_type, offset, len,
152                                  flags & TDB_LOCK_WAIT);
153         } while (ret == -1 && errno == EINTR);
154
155         if (ret == -1) {
156                 /* Generic lock error. errno set by fcntl.
157                  * EAGAIN is an expected return from non-blocking
158                  * locks. */
159                 if (!(flags & TDB_LOCK_PROBE) && errno != EAGAIN) {
160                         tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
161                                    "tdb_brlock failed (fd=%d) at"
162                                    " offset %zu rw_type=%d flags=%d len=%zu:"
163                                    " %s",
164                                    tdb->file->fd, (size_t)offset, rw_type,
165                                    flags, (size_t)len, strerror(errno));
166                 }
167                 return TDB_ERR_LOCK;
168         }
169         return TDB_SUCCESS;
170 }
171
172 static enum TDB_ERROR tdb_brunlock(struct tdb_context *tdb,
173                                    int rw_type, tdb_off_t offset, size_t len)
174 {
175         int ret;
176
177         if (tdb->flags & TDB_NOLOCK) {
178                 return TDB_SUCCESS;
179         }
180
181         do {
182                 ret = fcntl_unlock(tdb, rw_type, offset, len);
183         } while (ret == -1 && errno == EINTR);
184
185         if (ret == -1) {
186                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
187                                   "tdb_brunlock failed (fd=%d) at offset %zu"
188                                   " rw_type=%d len=%zu",
189                                   tdb->file->fd, (size_t)offset, rw_type,
190                                   (size_t)len);
191         }
192         return TDB_SUCCESS;
193 }
194
195 /*
196   upgrade a read lock to a write lock. This needs to be handled in a
197   special way as some OSes (such as solaris) have too conservative
198   deadlock detection and claim a deadlock when progress can be
199   made. For those OSes we may loop for a while.
200 */
201 enum TDB_ERROR tdb_allrecord_upgrade(struct tdb_context *tdb)
202 {
203         int count = 1000;
204
205         if (tdb->file->allrecord_lock.count != 1) {
206                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
207                                   "tdb_allrecord_upgrade failed:"
208                                   " count %u too high",
209                                   tdb->file->allrecord_lock.count);
210         }
211
212         if (tdb->file->allrecord_lock.off != 1) {
213                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
214                                   "tdb_allrecord_upgrade failed:"
215                                   " already upgraded?");
216         }
217
218         while (count--) {
219                 struct timeval tv;
220                 if (tdb_brlock(tdb, F_WRLCK,
221                                TDB_HASH_LOCK_START, 0,
222                                TDB_LOCK_WAIT|TDB_LOCK_PROBE) == TDB_SUCCESS) {
223                         tdb->file->allrecord_lock.ltype = F_WRLCK;
224                         tdb->file->allrecord_lock.off = 0;
225                         return TDB_SUCCESS;
226                 }
227                 if (errno != EDEADLK) {
228                         break;
229                 }
230                 /* sleep for as short a time as we can - more portable than usleep() */
231                 tv.tv_sec = 0;
232                 tv.tv_usec = 1;
233                 select(0, NULL, NULL, NULL, &tv);
234         }
235         return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
236                           "tdb_allrecord_upgrade failed");
237 }
238
239 static struct tdb_lock_type *find_nestlock(struct tdb_context *tdb,
240                                            tdb_off_t offset)
241 {
242         unsigned int i;
243
244         for (i=0; i<tdb->file->num_lockrecs; i++) {
245                 if (tdb->file->lockrecs[i].off == offset) {
246                         return &tdb->file->lockrecs[i];
247                 }
248         }
249         return NULL;
250 }
251
252 enum TDB_ERROR tdb_lock_and_recover(struct tdb_context *tdb)
253 {
254         enum TDB_ERROR ecode;
255
256         ecode = tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK,
257                                    false);
258         if (ecode != TDB_SUCCESS) {
259                 return ecode;
260         }
261
262         ecode = tdb_lock_open(tdb, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK);
263         if (ecode != TDB_SUCCESS) {
264                 tdb_allrecord_unlock(tdb, F_WRLCK);
265                 return ecode;
266         }
267         ecode = tdb_transaction_recover(tdb);
268         tdb_unlock_open(tdb);
269         tdb_allrecord_unlock(tdb, F_WRLCK);
270
271         return ecode;
272 }
273
274 /* lock an offset in the database. */
275 static enum TDB_ERROR tdb_nest_lock(struct tdb_context *tdb,
276                                     tdb_off_t offset, int ltype,
277                                     enum tdb_lock_flags flags)
278 {
279         struct tdb_lock_type *new_lck;
280         enum TDB_ERROR ecode;
281
282         if (offset > (TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE
283                       + tdb->file->map_size / 8)) {
284                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
285                                   "tdb_nest_lock: invalid offset %zu ltype=%d",
286                                   (size_t)offset, ltype);
287         }
288
289         if (tdb->flags & TDB_NOLOCK)
290                 return TDB_SUCCESS;
291
292         add_stat(tdb, locks, 1);
293
294         new_lck = find_nestlock(tdb, offset);
295         if (new_lck) {
296                 if (new_lck->ltype == F_RDLCK && ltype == F_WRLCK) {
297                         return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
298                                           "tdb_nest_lock:"
299                                           " offset %zu has read lock",
300                                           (size_t)offset);
301                 }
302                 /* Just increment the struct, posix locks don't stack. */
303                 new_lck->count++;
304                 return TDB_SUCCESS;
305         }
306
307         if (tdb->file->num_lockrecs
308             && offset >= TDB_HASH_LOCK_START
309             && offset < TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE) {
310                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
311                                   "tdb_nest_lock: already have a hash lock?");
312         }
313
314         new_lck = (struct tdb_lock_type *)realloc(
315                 tdb->file->lockrecs,
316                 sizeof(*tdb->file->lockrecs) * (tdb->file->num_lockrecs+1));
317         if (new_lck == NULL) {
318                 return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
319                                   "tdb_nest_lock:"
320                                   " unable to allocate %zu lock struct",
321                                   tdb->file->num_lockrecs + 1);
322         }
323         tdb->file->lockrecs = new_lck;
324
325         /* Since fcntl locks don't nest, we do a lock for the first one,
326            and simply bump the count for future ones */
327         ecode = tdb_brlock(tdb, ltype, offset, 1, flags);
328         if (ecode != TDB_SUCCESS) {
329                 return ecode;
330         }
331
332         /* First time we grab a lock, perhaps someone died in commit? */
333         if (!(flags & TDB_LOCK_NOCHECK)
334             && tdb->file->num_lockrecs == 0) {
335                 tdb_bool_err berr = tdb_needs_recovery(tdb);
336                 if (berr != false) {
337                         tdb_brunlock(tdb, ltype, offset, 1);
338
339                         if (berr < 0)
340                                 return berr;
341                         ecode = tdb_lock_and_recover(tdb);
342                         if (ecode == TDB_SUCCESS) {
343                                 ecode = tdb_brlock(tdb, ltype, offset, 1,
344                                                    flags);
345                         }
346                         if (ecode != TDB_SUCCESS) {
347                                 return ecode;
348                         }
349                 }
350         }
351
352         tdb->file->lockrecs[tdb->file->num_lockrecs].off = offset;
353         tdb->file->lockrecs[tdb->file->num_lockrecs].count = 1;
354         tdb->file->lockrecs[tdb->file->num_lockrecs].ltype = ltype;
355         tdb->file->num_lockrecs++;
356
357         return TDB_SUCCESS;
358 }
359
360 static enum TDB_ERROR tdb_nest_unlock(struct tdb_context *tdb,
361                                       tdb_off_t off, int ltype)
362 {
363         struct tdb_lock_type *lck;
364         enum TDB_ERROR ecode;
365
366         if (tdb->flags & TDB_NOLOCK)
367                 return TDB_SUCCESS;
368
369         lck = find_nestlock(tdb, off);
370         if ((lck == NULL) || (lck->count == 0)) {
371                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
372                                   "tdb_nest_unlock: no lock for %zu",
373                                   (size_t)off);
374         }
375
376         if (lck->count > 1) {
377                 lck->count--;
378                 return TDB_SUCCESS;
379         }
380
381         /*
382          * This lock has count==1 left, so we need to unlock it in the
383          * kernel. We don't bother with decrementing the in-memory array
384          * element, we're about to overwrite it with the last array element
385          * anyway.
386          */
387         ecode = tdb_brunlock(tdb, ltype, off, 1);
388
389         /*
390          * Shrink the array by overwriting the element just unlocked with the
391          * last array element.
392          */
393         *lck = tdb->file->lockrecs[--tdb->file->num_lockrecs];
394
395         return ecode;
396 }
397
398 /*
399   get the transaction lock
400  */
401 enum TDB_ERROR tdb_transaction_lock(struct tdb_context *tdb, int ltype)
402 {
403         return tdb_nest_lock(tdb, TDB_TRANSACTION_LOCK, ltype, TDB_LOCK_WAIT);
404 }
405
406 /*
407   release the transaction lock
408  */
409 void tdb_transaction_unlock(struct tdb_context *tdb, int ltype)
410 {
411         tdb_nest_unlock(tdb, TDB_TRANSACTION_LOCK, ltype);
412 }
413
414 /* We only need to lock individual bytes, but Linux merges consecutive locks
415  * so we lock in contiguous ranges. */
416 static enum TDB_ERROR tdb_lock_gradual(struct tdb_context *tdb,
417                                        int ltype, enum tdb_lock_flags flags,
418                                        tdb_off_t off, tdb_off_t len)
419 {
420         enum TDB_ERROR ecode;
421         enum tdb_lock_flags nb_flags = (flags & ~TDB_LOCK_WAIT);
422
423         if (len <= 1) {
424                 /* 0 would mean to end-of-file... */
425                 assert(len != 0);
426                 /* Single hash.  Just do blocking lock. */
427                 return tdb_brlock(tdb, ltype, off, len, flags);
428         }
429
430         /* First we try non-blocking. */
431         if (tdb_brlock(tdb, ltype, off, len, nb_flags) == TDB_SUCCESS) {
432                 return TDB_SUCCESS;
433         }
434
435         /* Try locking first half, then second. */
436         ecode = tdb_lock_gradual(tdb, ltype, flags, off, len / 2);
437         if (ecode != TDB_SUCCESS)
438                 return ecode;
439
440         ecode = tdb_lock_gradual(tdb, ltype, flags,
441                                  off + len / 2, len - len / 2);
442         if (ecode != TDB_SUCCESS) {
443                 tdb_brunlock(tdb, ltype, off, len / 2);
444         }
445         return ecode;
446 }
447
448 /* lock/unlock entire database.  It can only be upgradable if you have some
449  * other way of guaranteeing exclusivity (ie. transaction write lock). */
450 enum TDB_ERROR tdb_allrecord_lock(struct tdb_context *tdb, int ltype,
451                                   enum tdb_lock_flags flags, bool upgradable)
452 {
453         enum TDB_ERROR ecode;
454         tdb_bool_err berr;
455
456         if (tdb->file->allrecord_lock.count
457             && (ltype == F_RDLCK
458                 || tdb->file->allrecord_lock.ltype == F_WRLCK)) {
459                 tdb->file->allrecord_lock.count++;
460                 return TDB_SUCCESS;
461         }
462
463         if (tdb->file->allrecord_lock.count) {
464                 /* a global lock of a different type exists */
465                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
466                                   "tdb_allrecord_lock: already have %s lock",
467                                   tdb->file->allrecord_lock.ltype == F_RDLCK
468                                   ? "read" : "write");
469         }
470
471         if (tdb_has_hash_locks(tdb)) {
472                 /* can't combine global and chain locks */
473                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
474                                   "tdb_allrecord_lock:"
475                                   " already have chain lock");
476         }
477
478         if (upgradable && ltype != F_RDLCK) {
479                 /* tdb error: you can't upgrade a write lock! */
480                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
481                                   "tdb_allrecord_lock:"
482                                   " can't upgrade a write lock");
483         }
484
485         add_stat(tdb, locks, 1);
486 again:
487         /* Lock hashes, gradually. */
488         ecode = tdb_lock_gradual(tdb, ltype, flags, TDB_HASH_LOCK_START,
489                                  TDB_HASH_LOCK_RANGE);
490         if (ecode != TDB_SUCCESS) {
491                 if (!(flags & TDB_LOCK_PROBE)) {
492                         tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
493                                    "tdb_allrecord_lock hashes failed");
494                 }
495                 return ecode;
496         }
497
498         /* Lock free tables: there to end of file. */
499         ecode = tdb_brlock(tdb, ltype,
500                            TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE,
501                            0, flags);
502         if (ecode != TDB_SUCCESS) {
503                 if (!(flags & TDB_LOCK_PROBE)) {
504                         tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
505                                  "tdb_allrecord_lock freetables failed");
506                 }
507                 tdb_brunlock(tdb, ltype, TDB_HASH_LOCK_START,
508                              TDB_HASH_LOCK_RANGE);
509                 return ecode;
510         }
511
512         tdb->file->allrecord_lock.count = 1;
513         /* If it's upgradable, it's actually exclusive so we can treat
514          * it as a write lock. */
515         tdb->file->allrecord_lock.ltype = upgradable ? F_WRLCK : ltype;
516         tdb->file->allrecord_lock.off = upgradable;
517
518         /* Now check for needing recovery. */
519         if (flags & TDB_LOCK_NOCHECK)
520                 return TDB_SUCCESS;
521
522         berr = tdb_needs_recovery(tdb);
523         if (likely(berr == false))
524                 return TDB_SUCCESS;
525
526         tdb_allrecord_unlock(tdb, ltype);
527         if (berr < 0)
528                 return berr;
529         ecode = tdb_lock_and_recover(tdb);
530         if (ecode != TDB_SUCCESS) {
531                 return ecode;
532         }
533         goto again;
534 }
535
536 enum TDB_ERROR tdb_lock_open(struct tdb_context *tdb, enum tdb_lock_flags flags)
537 {
538         return tdb_nest_lock(tdb, TDB_OPEN_LOCK, F_WRLCK, flags);
539 }
540
541 void tdb_unlock_open(struct tdb_context *tdb)
542 {
543         tdb_nest_unlock(tdb, TDB_OPEN_LOCK, F_WRLCK);
544 }
545
546 bool tdb_has_open_lock(struct tdb_context *tdb)
547 {
548         return !(tdb->flags & TDB_NOLOCK)
549                 && find_nestlock(tdb, TDB_OPEN_LOCK) != NULL;
550 }
551
552 enum TDB_ERROR tdb_lock_expand(struct tdb_context *tdb, int ltype)
553 {
554         /* Lock doesn't protect data, so don't check (we recurse if we do!) */
555         return tdb_nest_lock(tdb, TDB_EXPANSION_LOCK, ltype,
556                              TDB_LOCK_WAIT | TDB_LOCK_NOCHECK);
557 }
558
559 void tdb_unlock_expand(struct tdb_context *tdb, int ltype)
560 {
561         tdb_nest_unlock(tdb, TDB_EXPANSION_LOCK, ltype);
562 }
563
564 /* unlock entire db */
565 void tdb_allrecord_unlock(struct tdb_context *tdb, int ltype)
566 {
567         if (tdb->file->allrecord_lock.count == 0) {
568                 tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
569                            "tdb_allrecord_unlock: not locked!");
570                 return;
571         }
572
573         /* Upgradable locks are marked as write locks. */
574         if (tdb->file->allrecord_lock.ltype != ltype
575             && (!tdb->file->allrecord_lock.off || ltype != F_RDLCK)) {
576                 tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
577                            "tdb_allrecord_unlock: have %s lock",
578                            tdb->file->allrecord_lock.ltype == F_RDLCK
579                            ? "read" : "write");
580                 return;
581         }
582
583         if (tdb->file->allrecord_lock.count > 1) {
584                 tdb->file->allrecord_lock.count--;
585                 return;
586         }
587
588         tdb->file->allrecord_lock.count = 0;
589         tdb->file->allrecord_lock.ltype = 0;
590
591         tdb_brunlock(tdb, ltype, TDB_HASH_LOCK_START, 0);
592 }
593
594 bool tdb_has_expansion_lock(struct tdb_context *tdb)
595 {
596         return find_nestlock(tdb, TDB_EXPANSION_LOCK) != NULL;
597 }
598
599 bool tdb_has_hash_locks(struct tdb_context *tdb)
600 {
601         unsigned int i;
602
603         for (i=0; i<tdb->file->num_lockrecs; i++) {
604                 if (tdb->file->lockrecs[i].off >= TDB_HASH_LOCK_START
605                     && tdb->file->lockrecs[i].off < (TDB_HASH_LOCK_START
606                                                      + TDB_HASH_LOCK_RANGE))
607                         return true;
608         }
609         return false;
610 }
611
612 static bool tdb_has_free_lock(struct tdb_context *tdb)
613 {
614         unsigned int i;
615
616         if (tdb->flags & TDB_NOLOCK)
617                 return false;
618
619         for (i=0; i<tdb->file->num_lockrecs; i++) {
620                 if (tdb->file->lockrecs[i].off
621                     > TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE)
622                         return true;
623         }
624         return false;
625 }
626
627 enum TDB_ERROR tdb_lock_hashes(struct tdb_context *tdb,
628                                tdb_off_t hash_lock,
629                                tdb_len_t hash_range,
630                                int ltype, enum tdb_lock_flags waitflag)
631 {
632         /* FIXME: Do this properly, using hlock_range */
633         unsigned lock = TDB_HASH_LOCK_START
634                 + (hash_lock >> (64 - TDB_HASH_LOCK_RANGE_BITS));
635
636         /* a allrecord lock allows us to avoid per chain locks */
637         if (tdb->file->allrecord_lock.count &&
638             (ltype == tdb->file->allrecord_lock.ltype || ltype == F_RDLCK)) {
639                 return TDB_SUCCESS;
640         }
641
642         if (tdb->file->allrecord_lock.count) {
643                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
644                                   "tdb_lock_hashes:"
645                                   " already have %s allrecordlock",
646                                   tdb->file->allrecord_lock.ltype == F_RDLCK
647                                   ? "read" : "write");
648         }
649
650         if (tdb_has_free_lock(tdb)) {
651                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
652                                   "tdb_lock_hashes: already have free lock");
653         }
654
655         if (tdb_has_expansion_lock(tdb)) {
656                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
657                                   "tdb_lock_hashes:"
658                                   " already have expansion lock");
659         }
660
661         return tdb_nest_lock(tdb, lock, ltype, waitflag);
662 }
663
664 enum TDB_ERROR tdb_unlock_hashes(struct tdb_context *tdb,
665                                  tdb_off_t hash_lock,
666                                  tdb_len_t hash_range, int ltype)
667 {
668         unsigned lock = TDB_HASH_LOCK_START
669                 + (hash_lock >> (64 - TDB_HASH_LOCK_RANGE_BITS));
670
671         if (tdb->flags & TDB_NOLOCK)
672                 return 0;
673
674         /* a allrecord lock allows us to avoid per chain locks */
675         if (tdb->file->allrecord_lock.count) {
676                 if (tdb->file->allrecord_lock.ltype == F_RDLCK
677                     && ltype == F_WRLCK) {
678                         return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
679                                           "tdb_unlock_hashes RO allrecord!");
680                 }
681                 return TDB_SUCCESS;
682         }
683
684         return tdb_nest_unlock(tdb, lock, ltype);
685 }
686
687 /* Hash locks use TDB_HASH_LOCK_START + the next 30 bits.
688  * Then we begin; bucket offsets are sizeof(tdb_len_t) apart, so we divide.
689  * The result is that on 32 bit systems we don't use lock values > 2^31 on
690  * files that are less than 4GB.
691  */
692 static tdb_off_t free_lock_off(tdb_off_t b_off)
693 {
694         return TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE
695                 + b_off / sizeof(tdb_off_t);
696 }
697
698 enum TDB_ERROR tdb_lock_free_bucket(struct tdb_context *tdb, tdb_off_t b_off,
699                                     enum tdb_lock_flags waitflag)
700 {
701         assert(b_off >= sizeof(struct tdb_header));
702
703         if (tdb->flags & TDB_NOLOCK)
704                 return 0;
705
706         /* a allrecord lock allows us to avoid per chain locks */
707         if (tdb->file->allrecord_lock.count) {
708                 if (tdb->file->allrecord_lock.ltype == F_WRLCK)
709                         return 0;
710                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
711                                   "tdb_lock_free_bucket with"
712                                   " read-only allrecordlock!");
713         }
714
715 #if 0 /* FIXME */
716         if (tdb_has_expansion_lock(tdb)) {
717                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
718                                   "tdb_lock_free_bucket:"
719                                   " already have expansion lock");
720         }
721 #endif
722
723         return tdb_nest_lock(tdb, free_lock_off(b_off), F_WRLCK, waitflag);
724 }
725
726 void tdb_unlock_free_bucket(struct tdb_context *tdb, tdb_off_t b_off)
727 {
728         if (tdb->file->allrecord_lock.count)
729                 return;
730
731         tdb_nest_unlock(tdb, free_lock_off(b_off), F_WRLCK);
732 }