tdb2: change to using a hash tree.
[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         if (waitflag)
44                 return fcntl(tdb->fd, F_SETLKW, &fl);
45         else
46                 return fcntl(tdb->fd, F_SETLK, &fl);
47 }
48
49 static int fcntl_unlock(struct tdb_context *tdb, int rw, off_t off, off_t len)
50 {
51         struct flock fl;
52 #if 0 /* Check they matched up locks and unlocks correctly. */
53         char line[80];
54         FILE *locks;
55         bool found = false;
56
57         locks = fopen("/proc/locks", "r");
58
59         while (fgets(line, 80, locks)) {
60                 char *p;
61                 int type, start, l;
62
63                 /* eg. 1: FLOCK  ADVISORY  WRITE 2440 08:01:2180826 0 EOF */
64                 p = strchr(line, ':') + 1;
65                 if (strncmp(p, " POSIX  ADVISORY  ", strlen(" POSIX  ADVISORY  ")))
66                         continue;
67                 p += strlen(" FLOCK  ADVISORY  ");
68                 if (strncmp(p, "READ  ", strlen("READ  ")) == 0)
69                         type = F_RDLCK;
70                 else if (strncmp(p, "WRITE ", strlen("WRITE ")) == 0)
71                         type = F_WRLCK;
72                 else
73                         abort();
74                 p += 6;
75                 if (atoi(p) != getpid())
76                         continue;
77                 p = strchr(strchr(p, ' ') + 1, ' ') + 1;
78                 start = atoi(p);
79                 p = strchr(p, ' ') + 1;
80                 if (strncmp(p, "EOF", 3) == 0)
81                         l = 0;
82                 else
83                         l = atoi(p) - start + 1;
84
85                 if (off == start) {
86                         if (len != l) {
87                                 fprintf(stderr, "Len %u should be %u: %s",
88                                         (int)len, l, line);
89                                 abort();
90                         }
91                         if (type != rw) {
92                                 fprintf(stderr, "Type %s wrong: %s",
93                                         rw == F_RDLCK ? "READ" : "WRITE", line);
94                                 abort();
95                         }
96                         found = true;
97                         break;
98                 }
99         }
100
101         if (!found) {
102                 fprintf(stderr, "Unlock on %u@%u not found!\n",
103                         (int)off, (int)len);
104                 abort();
105         }
106
107         fclose(locks);
108 #endif
109
110         fl.l_type = F_UNLCK;
111         fl.l_whence = SEEK_SET;
112         fl.l_start = off;
113         fl.l_len = len;
114         fl.l_pid = 0;
115
116         return fcntl(tdb->fd, F_SETLKW, &fl);
117 }
118
119 /* a byte range locking function - return 0 on success
120    this functions locks/unlocks 1 byte at the specified offset.
121
122    note that a len of zero means lock to end of file
123 */
124 static int tdb_brlock(struct tdb_context *tdb,
125                       int rw_type, tdb_off_t offset, tdb_off_t len,
126                       enum tdb_lock_flags flags)
127 {
128         int ret;
129
130         if (tdb->flags & TDB_NOLOCK) {
131                 return 0;
132         }
133
134         if (rw_type == F_WRLCK && tdb->read_only) {
135                 tdb->ecode = TDB_ERR_RDONLY;
136                 return -1;
137         }
138
139         /* A 32 bit system cannot open a 64-bit file, but it could have
140          * expanded since then: check here. */
141         if ((size_t)(offset + len) != offset + len) {
142                 tdb->ecode = TDB_ERR_IO;
143                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
144                          "tdb_brlock: lock on giant offset %llu\n",
145                          (long long)(offset + len));
146                 return -1;
147         }
148
149         do {
150                 ret = fcntl_lock(tdb, rw_type, offset, len,
151                                  flags & TDB_LOCK_WAIT);
152         } while (ret == -1 && errno == EINTR);
153
154         if (ret == -1) {
155                 tdb->ecode = TDB_ERR_LOCK;
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->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
161                                  "tdb_brlock failed (fd=%d) at"
162                                  " offset %llu rw_type=%d flags=%d len=%llu\n",
163                                  tdb->fd, (long long)offset, rw_type,
164                                  flags, (long long)len);
165                 }
166                 return -1;
167         }
168         return 0;
169 }
170
171 static int tdb_brunlock(struct tdb_context *tdb,
172                         int rw_type, tdb_off_t offset, size_t len)
173 {
174         int ret;
175
176         if (tdb->flags & TDB_NOLOCK) {
177                 return 0;
178         }
179
180         do {
181                 ret = fcntl_unlock(tdb, rw_type, offset, len);
182         } while (ret == -1 && errno == EINTR);
183
184         if (ret == -1) {
185                 tdb->log(tdb, TDB_DEBUG_TRACE, tdb->log_priv,
186                          "tdb_brunlock failed (fd=%d) at offset %llu"
187                          " rw_type=%d len=%llu\n",
188                          tdb->fd, (long long)offset, rw_type, (long long)len);
189         }
190         return ret;
191 }
192
193 #if 0
194 /*
195   upgrade a read lock to a write lock. This needs to be handled in a
196   special way as some OSes (such as solaris) have too conservative
197   deadlock detection and claim a deadlock when progress can be
198   made. For those OSes we may loop for a while.  
199 */
200 int tdb_allrecord_upgrade(struct tdb_context *tdb)
201 {
202         int count = 1000;
203
204         if (tdb->allrecord_lock.count != 1) {
205                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
206                          "tdb_allrecord_upgrade failed: count %u too high\n",
207                          tdb->allrecord_lock.count);
208                 return -1;
209         }
210
211         if (tdb->allrecord_lock.off != 1) {
212                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
213                          "tdb_allrecord_upgrade failed: already upgraded?\n");
214                 return -1;
215         }
216
217         while (count--) {
218                 struct timeval tv;
219                 if (tdb_brlock(tdb, F_WRLCK,
220                                TDB_HASH_LOCK_START
221                                + (1ULL << tdb->header.v.hash_bits), 0,
222                                TDB_LOCK_WAIT|TDB_LOCK_PROBE) == 0) {
223                         tdb->allrecord_lock.ltype = F_WRLCK;
224                         tdb->allrecord_lock.off = 0;
225                         return 0;
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         tdb->log(tdb, TDB_DEBUG_WARNING, tdb->log_priv,
236                  "tdb_allrecord_upgrade failed\n");
237         return -1;
238 }
239 #endif
240
241 static struct tdb_lock_type *find_nestlock(struct tdb_context *tdb,
242                                            tdb_off_t offset)
243 {
244         unsigned int i;
245
246         for (i=0; i<tdb->num_lockrecs; i++) {
247                 if (tdb->lockrecs[i].off == offset) {
248                         return &tdb->lockrecs[i];
249                 }
250         }
251         return NULL;
252 }
253
254 /* lock an offset in the database. */
255 static int tdb_nest_lock(struct tdb_context *tdb, tdb_off_t offset, int ltype,
256                          enum tdb_lock_flags flags)
257 {
258         struct tdb_lock_type *new_lck;
259
260         if (offset >= TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE + tdb->map_size / 8) {
261                 tdb->ecode = TDB_ERR_LOCK;
262                 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
263                          "tdb_lock: invalid offset %llu ltype=%d\n",
264                          (long long)offset, ltype);
265                 return -1;
266         }
267
268         if (tdb->flags & TDB_NOLOCK)
269                 return 0;
270
271         new_lck = find_nestlock(tdb, offset);
272         if (new_lck) {
273                 /*
274                  * Just increment the in-memory struct, posix locks
275                  * don't stack.
276                  */
277                 new_lck->count++;
278                 return 0;
279         }
280
281         new_lck = (struct tdb_lock_type *)realloc(
282                 tdb->lockrecs,
283                 sizeof(*tdb->lockrecs) * (tdb->num_lockrecs+1));
284         if (new_lck == NULL) {
285                 tdb->ecode = TDB_ERR_OOM;
286                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
287                          "tdb_lock: unable to allocate %llu lock structure",
288                          (long long)(tdb->num_lockrecs + 1));
289                 errno = ENOMEM;
290                 return -1;
291         }
292         tdb->lockrecs = new_lck;
293
294         /* Since fcntl locks don't nest, we do a lock for the first one,
295            and simply bump the count for future ones */
296         if (tdb_brlock(tdb, ltype, offset, 1, flags)) {
297                 return -1;
298         }
299
300         tdb->lockrecs[tdb->num_lockrecs].off = offset;
301         tdb->lockrecs[tdb->num_lockrecs].count = 1;
302         tdb->lockrecs[tdb->num_lockrecs].ltype = ltype;
303         tdb->num_lockrecs++;
304
305         return 0;
306 }
307
308 static int tdb_lock_and_recover(struct tdb_context *tdb)
309 {
310 #if 0 /* FIXME */
311
312         int ret;
313
314         /* We need to match locking order in transaction commit. */
315         if (tdb_brlock(tdb, F_WRLCK, FREELIST_TOP, 0, TDB_LOCK_WAIT)) {
316                 return -1;
317         }
318
319         if (tdb_brlock(tdb, F_WRLCK, OPEN_LOCK, 1, TDB_LOCK_WAIT)) {
320                 tdb_brunlock(tdb, F_WRLCK, FREELIST_TOP, 0);
321                 return -1;
322         }
323
324         ret = tdb_transaction_recover(tdb);
325
326         tdb_brunlock(tdb, F_WRLCK, OPEN_LOCK, 1);
327         tdb_brunlock(tdb, F_WRLCK, FREELIST_TOP, 0);
328
329         return ret;
330 #else
331         abort();
332         return -1;
333 #endif
334 }
335
336 static bool tdb_needs_recovery(struct tdb_context *tdb)
337 {
338         /* FIXME */
339         return false;
340 }
341
342 static int tdb_nest_unlock(struct tdb_context *tdb, tdb_off_t off, int ltype)
343 {
344         int ret = -1;
345         struct tdb_lock_type *lck;
346
347         if (tdb->flags & TDB_NOLOCK)
348                 return 0;
349
350         lck = find_nestlock(tdb, off);
351         if ((lck == NULL) || (lck->count == 0)) {
352                 tdb->ecode = TDB_ERR_LOCK;
353                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
354                          "tdb_unlock: no lock for %llu\n", (long long)off);
355                 return -1;
356         }
357
358         if (lck->count > 1) {
359                 lck->count--;
360                 return 0;
361         }
362
363         /*
364          * This lock has count==1 left, so we need to unlock it in the
365          * kernel. We don't bother with decrementing the in-memory array
366          * element, we're about to overwrite it with the last array element
367          * anyway.
368          */
369         ret = tdb_brunlock(tdb, ltype, off, 1);
370
371         /*
372          * Shrink the array by overwriting the element just unlocked with the
373          * last array element.
374          */
375         *lck = tdb->lockrecs[--tdb->num_lockrecs];
376
377         return ret;
378 }
379
380 #if 0
381 /*
382   get the transaction lock
383  */
384 int tdb_transaction_lock(struct tdb_context *tdb, int ltype,
385                          enum tdb_lock_flags lockflags)
386 {
387         return tdb_nest_lock(tdb, TRANSACTION_LOCK, ltype, lockflags);
388 }
389
390 /*
391   release the transaction lock
392  */
393 int tdb_transaction_unlock(struct tdb_context *tdb, int ltype)
394 {
395         return tdb_nest_unlock(tdb, TRANSACTION_LOCK, ltype, false);
396 }
397 #endif
398
399 /* We only need to lock individual bytes, but Linux merges consecutive locks
400  * so we lock in contiguous ranges. */
401 static int tdb_lock_gradual(struct tdb_context *tdb,
402                             int ltype, enum tdb_lock_flags flags,
403                             tdb_off_t off, tdb_off_t len)
404 {
405         int ret;
406         enum tdb_lock_flags nb_flags = (flags & ~TDB_LOCK_WAIT);
407
408         if (len <= 1) {
409                 /* 0 would mean to end-of-file... */
410                 assert(len != 0);
411                 /* Single hash.  Just do blocking lock. */
412                 return tdb_brlock(tdb, ltype, off, len, flags);
413         }
414
415         /* First we try non-blocking. */
416         ret = tdb_brlock(tdb, ltype, off, len, nb_flags);
417         if (ret == 0) {
418                 return 0;
419         }
420
421         /* Try locking first half, then second. */
422         ret = tdb_lock_gradual(tdb, ltype, flags, off, len / 2);
423         if (ret == -1)
424                 return -1;
425
426         ret = tdb_lock_gradual(tdb, ltype, flags,
427                                     off + len / 2, len - len / 2);
428         if (ret == -1) {
429                 tdb_brunlock(tdb, ltype, off, len / 2);
430                 return -1;
431         }
432         return 0;
433 }
434
435 /* lock/unlock entire database.  It can only be upgradable if you have some
436  * other way of guaranteeing exclusivity (ie. transaction write lock).
437  * Note that we don't lock the free chains: currently noone can get those locks
438  * without a hash chain lock first. */
439 int tdb_allrecord_lock(struct tdb_context *tdb, int ltype,
440                        enum tdb_lock_flags flags, bool upgradable)
441 {
442         /* FIXME: There are no locks on read-only dbs */
443         if (tdb->read_only) {
444                 tdb->ecode = TDB_ERR_LOCK;
445                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
446                          "tdb_allrecord_lock: read-only\n");
447                 return -1;
448         }
449
450         if (tdb->allrecord_lock.count && tdb->allrecord_lock.ltype == ltype) {
451                 tdb->allrecord_lock.count++;
452                 return 0;
453         }
454
455         if (tdb->allrecord_lock.count) {
456                 /* a global lock of a different type exists */
457                 tdb->ecode = TDB_ERR_LOCK;
458                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
459                          "tdb_allrecord_lock: already have %s lock\n",
460                          tdb->allrecord_lock.ltype == F_RDLCK
461                          ? "read" : "write");
462                 return -1;
463         }
464
465         if (tdb_has_locks(tdb)) {
466                 /* can't combine global and chain locks */
467                 tdb->ecode = TDB_ERR_LOCK;
468                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
469                          "tdb_allrecord_lock: already have chain lock\n");
470                 return -1;
471         }
472
473         if (upgradable && ltype != F_RDLCK) {
474                 /* tdb error: you can't upgrade a write lock! */
475                 tdb->ecode = TDB_ERR_LOCK;
476                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
477                          "tdb_allrecord_lock: can't upgrade a write lock\n");
478                 return -1;
479         }
480
481 again:
482         if (tdb_lock_gradual(tdb, ltype, flags, TDB_HASH_LOCK_START,
483                              TDB_HASH_LOCK_RANGE)) {
484                 if (!(flags & TDB_LOCK_PROBE)) {
485                         tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
486                                  "tdb_lockall hashes failed (%s)\n",
487                                  strerror(errno));
488                 }
489                 return -1;
490         }
491
492         tdb->allrecord_lock.count = 1;
493         /* If it's upgradable, it's actually exclusive so we can treat
494          * it as a write lock. */
495         tdb->allrecord_lock.ltype = upgradable ? F_WRLCK : ltype;
496         tdb->allrecord_lock.off = upgradable;
497
498         /* Now check for needing recovery. */
499         if (unlikely(tdb_needs_recovery(tdb))) {
500                 tdb_allrecord_unlock(tdb, ltype);
501                 if (tdb_lock_and_recover(tdb) == -1) {
502                         return -1;
503                 }               
504                 goto again;
505         }
506
507         return 0;
508 }
509
510 int tdb_lock_open(struct tdb_context *tdb)
511 {
512         return tdb_nest_lock(tdb, TDB_OPEN_LOCK, F_WRLCK, TDB_LOCK_WAIT);
513 }
514
515 void tdb_unlock_open(struct tdb_context *tdb)
516 {
517         tdb_nest_unlock(tdb, TDB_OPEN_LOCK, F_WRLCK);
518 }
519
520 int tdb_lock_expand(struct tdb_context *tdb, int ltype)
521 {
522         return tdb_nest_lock(tdb, TDB_EXPANSION_LOCK, ltype, TDB_LOCK_WAIT);
523 }
524
525 void tdb_unlock_expand(struct tdb_context *tdb, int ltype)
526 {
527         tdb_nest_unlock(tdb, TDB_EXPANSION_LOCK, ltype);
528 }
529
530 /* unlock entire db */
531 int tdb_allrecord_unlock(struct tdb_context *tdb, int ltype)
532 {
533         /* FIXME: There are no locks on read-only dbs */
534         if (tdb->read_only) {
535                 tdb->ecode = TDB_ERR_LOCK;
536                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
537                          "tdb_allrecord_unlock: read-only\n");
538                 return -1;
539         }
540
541         if (tdb->allrecord_lock.count == 0) {
542                 tdb->ecode = TDB_ERR_LOCK;
543                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
544                          "tdb_allrecord_unlock: not locked!\n");
545                 return -1;
546         }
547
548         /* Upgradable locks are marked as write locks. */
549         if (tdb->allrecord_lock.ltype != ltype
550             && (!tdb->allrecord_lock.off || ltype != F_RDLCK)) {
551                 tdb->ecode = TDB_ERR_LOCK;
552                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
553                          "tdb_allrecord_unlock: have %s lock\n",
554                          tdb->allrecord_lock.ltype == F_RDLCK
555                          ? "read" : "write");
556                 return -1;
557         }
558
559         if (tdb->allrecord_lock.count > 1) {
560                 tdb->allrecord_lock.count--;
561                 return 0;
562         }
563
564         tdb->allrecord_lock.count = 0;
565         tdb->allrecord_lock.ltype = 0;
566
567         return tdb_brunlock(tdb, ltype, TDB_HASH_LOCK_START,
568                             TDB_HASH_LOCK_RANGE);
569 }
570
571 bool tdb_has_expansion_lock(struct tdb_context *tdb)
572 {
573         return find_nestlock(tdb, TDB_EXPANSION_LOCK) != NULL
574                 || (tdb->flags & TDB_NOLOCK);
575 }
576
577 bool tdb_has_locks(struct tdb_context *tdb)
578 {
579         return tdb->allrecord_lock.count || tdb->num_lockrecs;
580 }
581
582 #if 0
583 /* lock entire database with write lock */
584 int tdb_lockall(struct tdb_context *tdb)
585 {
586         tdb_trace(tdb, "tdb_lockall");
587         return tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT, false);
588 }
589
590 /* lock entire database with write lock - nonblocking varient */
591 int tdb_lockall_nonblock(struct tdb_context *tdb)
592 {
593         int ret = tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_NOWAIT, false);
594         tdb_trace_ret(tdb, "tdb_lockall_nonblock", ret);
595         return ret;
596 }
597
598 /* unlock entire database with write lock */
599 int tdb_unlockall(struct tdb_context *tdb)
600 {
601         tdb_trace(tdb, "tdb_unlockall");
602         return tdb_allrecord_unlock(tdb, F_WRLCK);
603 }
604
605 /* lock entire database with read lock */
606 int tdb_lockall_read(struct tdb_context *tdb)
607 {
608         tdb_trace(tdb, "tdb_lockall_read");
609         return tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, false);
610 }
611
612 /* lock entire database with read lock - nonblock varient */
613 int tdb_lockall_read_nonblock(struct tdb_context *tdb)
614 {
615         int ret = tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_NOWAIT, false);
616         tdb_trace_ret(tdb, "tdb_lockall_read_nonblock", ret);
617         return ret;
618 }
619
620 /* unlock entire database with read lock */
621 int tdb_unlockall_read(struct tdb_context *tdb)
622 {
623         tdb_trace(tdb, "tdb_unlockall_read");
624         return tdb_allrecord_unlock(tdb, F_RDLCK);
625 }
626 #endif
627
628 int tdb_lock_hashes(struct tdb_context *tdb,
629                     tdb_off_t hash_lock,
630                     tdb_len_t hash_range,
631                     int ltype, enum tdb_lock_flags waitflag)
632 {
633         /* FIXME: Do this properly, using hlock_range */
634         unsigned lock = TDB_HASH_LOCK_START
635                 + (hash_lock >> (64 - TDB_HASH_LOCK_RANGE_BITS));
636
637         /* a allrecord lock allows us to avoid per chain locks */
638         if (tdb->allrecord_lock.count &&
639             (ltype == tdb->allrecord_lock.ltype || ltype == F_RDLCK)) {
640                 return 0;
641         }
642
643         if (tdb->allrecord_lock.count) {
644                 tdb->ecode = TDB_ERR_LOCK;
645                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
646                          "tdb_lock_list: have %s allrecordlock\n",
647                          tdb->allrecord_lock.ltype == F_RDLCK
648                          ? "read" : "write");
649                 return -1;
650         }
651
652         return tdb_nest_lock(tdb, lock, ltype, waitflag);
653 }
654
655 int tdb_unlock_hashes(struct tdb_context *tdb,
656                       tdb_off_t hash_lock,
657                       tdb_len_t hash_range, int ltype)
658 {
659         unsigned lock = TDB_HASH_LOCK_START
660                 + (hash_lock >> (64 - TDB_HASH_LOCK_RANGE_BITS));
661
662         /* a allrecord lock allows us to avoid per chain locks */
663         if (tdb->allrecord_lock.count) {
664                 if (tdb->allrecord_lock.ltype == F_RDLCK
665                     && ltype == F_WRLCK) {
666                         tdb->ecode = TDB_ERR_LOCK;
667                         tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
668                                  "tdb_unlock_list RO allrecord!\n");
669                         return -1;
670                 }
671                 return 0;
672         }
673
674         return tdb_nest_unlock(tdb, lock, ltype);
675 }
676
677 /* Hash locks use TDB_HASH_LOCK_START + the next 30 bits.
678  * Then we begin; bucket offsets are sizeof(tdb_len_t) apart, so we divide.
679  * The result is that on 32 bit systems we don't use lock values > 2^31 on
680  * files that are less than 4GB.
681  */
682 static tdb_off_t free_lock_off(tdb_off_t b_off)
683 {
684         return TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE + b_off / sizeof(tdb_off_t);
685 }
686
687 int tdb_lock_free_bucket(struct tdb_context *tdb, tdb_off_t b_off,
688                          enum tdb_lock_flags waitflag)
689 {
690         assert(b_off >= sizeof(struct tdb_header));
691
692         /* You're supposed to have a hash lock first! */
693         if (!(tdb->flags & TDB_NOLOCK) && !tdb_has_locks(tdb)) {
694                 tdb->ecode = TDB_ERR_LOCK;
695                 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
696                          "tdb_lock_free_list without lock!\n");
697                 return -1;
698         }
699
700         /* a allrecord lock allows us to avoid per chain locks */
701         if (tdb->allrecord_lock.count) {
702                 if (tdb->allrecord_lock.ltype == F_WRLCK)
703                         return 0;
704                 tdb->ecode = TDB_ERR_LOCK;
705                 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
706                          "tdb_lock_free_list with RO allrecordlock!\n");
707                 return -1;
708         }
709
710         return tdb_nest_lock(tdb, free_lock_off(b_off), F_WRLCK, waitflag);
711 }
712
713 void tdb_unlock_free_bucket(struct tdb_context *tdb, tdb_off_t b_off)
714 {
715         if (tdb->allrecord_lock.count)
716                 return;
717
718         tdb_nest_unlock(tdb, free_lock_off(b_off), F_WRLCK);
719 }
720
721 /* Even if the entry isn't in this hash bucket, you'd have to lock this
722  * bucket to find it. */
723 static int chainlock(struct tdb_context *tdb, const TDB_DATA *key,
724                      int ltype, enum tdb_lock_flags waitflag,
725                      const char *func)
726 {
727         int ret;
728         uint64_t h = tdb_hash(tdb, key->dptr, key->dsize);
729
730         ret = tdb_lock_hashes(tdb, h, 1, ltype, waitflag);
731         tdb_trace_1rec(tdb, func, *key);
732         return ret;
733 }
734
735 /* lock/unlock one hash chain. This is meant to be used to reduce
736    contention - it cannot guarantee how many records will be locked */
737 int tdb_chainlock(struct tdb_context *tdb, TDB_DATA key)
738 {
739         return chainlock(tdb, &key, F_WRLCK, TDB_LOCK_WAIT, "tdb_chainlock");
740 }
741
742 int tdb_chainunlock(struct tdb_context *tdb, TDB_DATA key)
743 {
744         uint64_t h = tdb_hash(tdb, key.dptr, key.dsize);
745         tdb_trace_1rec(tdb, "tdb_chainunlock", key);
746         return tdb_unlock_hashes(tdb, h, 1, F_WRLCK);
747 }
748
749 #if 0
750 /* lock/unlock one hash chain, non-blocking. This is meant to be used
751    to reduce contention - it cannot guarantee how many records will be
752    locked */
753 int tdb_chainlock_nonblock(struct tdb_context *tdb, TDB_DATA key)
754 {
755         return chainlock(tdb, &key, F_WRLCK, TDB_LOCK_NOWAIT,
756                          "tdb_chainlock_nonblock");
757 }
758
759 int tdb_chainlock_read(struct tdb_context *tdb, TDB_DATA key)
760 {
761         return chainlock(tdb, &key, F_RDLCK, TDB_LOCK_WAIT,
762                          "tdb_chainlock_read");
763 }
764
765 int tdb_chainunlock_read(struct tdb_context *tdb, TDB_DATA key)
766 {
767         uint64_t h = tdb_hash(tdb, key.dptr, key.dsize);
768         tdb_trace_1rec(tdb, "tdb_chainunlock_read", key);
769         return tdb_unlock_list(tdb, h & ((1ULL << tdb->header.v.hash_bits)-1),
770                                F_RDLCK);
771 }
772
773 /* record lock stops delete underneath */
774 int tdb_lock_record(struct tdb_context *tdb, tdb_off_t off)
775 {
776         if (tdb->allrecord_lock.count) {
777                 return 0;
778         }
779         return off ? tdb_brlock(tdb, F_RDLCK, off, 1, TDB_LOCK_WAIT) : 0;
780 }
781
782 /*
783   Write locks override our own fcntl readlocks, so check it here.
784   Note this is meant to be F_SETLK, *not* F_SETLKW, as it's not
785   an error to fail to get the lock here.
786 */
787 int tdb_write_lock_record(struct tdb_context *tdb, tdb_off_t off)
788 {
789         struct tdb_traverse_lock *i;
790         for (i = &tdb->travlocks; i; i = i->next)
791                 if (i->off == off)
792                         return -1;
793         if (tdb->allrecord_lock.count) {
794                 if (tdb->allrecord_lock.ltype == F_WRLCK) {
795                         return 0;
796                 }
797                 return -1;
798         }
799         return tdb_brlock(tdb, F_WRLCK, off, 1, TDB_LOCK_NOWAIT|TDB_LOCK_PROBE);
800 }
801
802 int tdb_write_unlock_record(struct tdb_context *tdb, tdb_off_t off)
803 {
804         if (tdb->allrecord_lock.count) {
805                 return 0;
806         }
807         return tdb_brunlock(tdb, F_WRLCK, off, 1);
808 }
809
810 /* fcntl locks don't stack: avoid unlocking someone else's */
811 int tdb_unlock_record(struct tdb_context *tdb, tdb_off_t off)
812 {
813         struct tdb_traverse_lock *i;
814         uint32_t count = 0;
815
816         if (tdb->allrecord_lock.count) {
817                 return 0;
818         }
819
820         if (off == 0)
821                 return 0;
822         for (i = &tdb->travlocks; i; i = i->next)
823                 if (i->off == off)
824                         count++;
825         return (count == 1 ? tdb_brunlock(tdb, F_RDLCK, off, 1) : 0);
826 }
827
828 /* The transaction code uses this to remove all locks. */
829 void tdb_release_transaction_locks(struct tdb_context *tdb)
830 {
831         unsigned int i;
832
833         if (tdb->allrecord_lock.count != 0) {
834                 tdb_off_t hash_size, free_size;
835
836                 hash_size = (1ULL << tdb->header.v.hash_bits)
837                         * sizeof(tdb_off_t);
838                 free_size = tdb->header.v.free_zones 
839                         * (tdb->header.v.free_buckets + 1) * sizeof(tdb_off_t);
840
841                 tdb_brunlock(tdb, tdb->allrecord_lock.ltype,
842                              tdb->header.v.hash_off, hash_size);
843                 tdb_brunlock(tdb, tdb->allrecord_lock.ltype,
844                              tdb->header.v.free_off, free_size);
845                 tdb->allrecord_lock.count = 0;
846                 tdb->allrecord_lock.ltype = 0;
847         }
848
849         for (i = 0; i<tdb->num_lockrecs; i++) {
850                 struct tdb_lock_type *lck = &tdb->lockrecs[i];
851
852                 tdb_brunlock(tdb, lck->ltype, lck->off, 1);
853         }
854         tdb->num_lockrecs = 0;
855         SAFE_FREE(tdb->lockrecs);
856         tdb->header_uptodate = false;
857 }
858 #endif
859
860 void tdb_lock_init(struct tdb_context *tdb)
861 {
862         tdb->num_lockrecs = 0;
863         tdb->lockrecs = NULL;
864         tdb->allrecord_lock.count = 0;
865 }