lib/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 /* If we were threaded, we could wait for unlock, but we're not, so fail. */
  33 enum TDB_ERROR owner_conflict(struct tdb_context *tdb, const char *call)
  34 {
  35         return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
  36                           "%s: lock owned by another tdb in this process.",
  37                           call);
  38 }
  39
  40 /* If we fork, we no longer really own locks. */
  41 bool check_lock_pid(struct tdb_context *tdb, const char *call, bool log)
  42 {
  43         /* No locks?  No problem! */
  44         if (tdb->file->allrecord_lock.count == 0
  45             && tdb->file->num_lockrecs == 0) {
  46                 return true;
  47         }
  48
  49         /* No fork?  No problem! */
  50         if (tdb->file->locker == getpid()) {
  51                 return true;
  52         }
  53
  54         if (log) {
  55                 tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
  56                            "%s: fork() detected after lock acquisition!"
  57                            " (%u vs %u)", call, tdb->file->locker, getpid());
  58         }
  59         return false;
  60 }
  61
  62 int tdb_fcntl_lock(int fd, int rw, off_t off, off_t len, bool waitflag,
  63                    void *unused)
  64 {
  65         struct flock fl;
  66         int ret;
  67
  68         do {
  69                 fl.l_type = rw;
  70                 fl.l_whence = SEEK_SET;
  71                 fl.l_start = off;
  72                 fl.l_len = len;
  73
  74                 if (waitflag)
  75                         ret = fcntl(fd, F_SETLKW, &fl);
  76                 else
  77                         ret = fcntl(fd, F_SETLK, &fl);
  78         } while (ret != 0 && errno == EINTR);
  79         return ret;
  80 }
  81
  82 int tdb_fcntl_unlock(int fd, int rw, off_t off, off_t len, void *unused)
  83 {
  84         struct flock fl;
  85         int ret;
  86
  87         do {
  88                 fl.l_type = F_UNLCK;
  89                 fl.l_whence = SEEK_SET;
  90                 fl.l_start = off;
  91                 fl.l_len = len;
  92
  93                 ret = fcntl(fd, F_SETLKW, &fl);
  94         } while (ret != 0 && errno == EINTR);
  95         return ret;
  96 }
  97
  98 static int lock(struct tdb_context *tdb,
  99                       int rw, off_t off, off_t len, bool waitflag)
 100 {
 101         int ret;
 102         if (tdb->file->allrecord_lock.count == 0
 103             && tdb->file->num_lockrecs == 0) {
 104                 tdb->file->locker = getpid();
 105         }
 106
 107         tdb->stats.lock_lowlevel++;
 108         ret = tdb->lock_fn(tdb->file->fd, rw, off, len, waitflag,
 109                            tdb->lock_data);
 110         if (!waitflag) {
 111                 tdb->stats.lock_nonblock++;
 112                 if (ret != 0)
 113                         tdb->stats.lock_nonblock_fail++;
 114         }
 115         return ret;
 116 }
 117
 118 static int unlock(struct tdb_context *tdb, int rw, off_t off, off_t len)
 119 {
 120 #if 0 /* Check they matched up locks and unlocks correctly. */
 121         char line[80];
 122         FILE *locks;
 123         bool found = false;
 124
 125         locks = fopen("/proc/locks", "r");
 126
 127         while (fgets(line, 80, locks)) {
 128                 char *p;
 129                 int type, start, l;
 130
 131                 /* eg. 1: FLOCK  ADVISORY  WRITE 2440 08:01:2180826 0 EOF */
 132                 p = strchr(line, ':') + 1;
 133                 if (strncmp(p, " POSIX  ADVISORY  ", strlen(" POSIX  ADVISORY  ")))
 134                         continue;
 135                 p += strlen(" FLOCK  ADVISORY  ");
 136                 if (strncmp(p, "READ  ", strlen("READ  ")) == 0)
 137                         type = F_RDLCK;
 138                 else if (strncmp(p, "WRITE ", strlen("WRITE ")) == 0)
 139                         type = F_WRLCK;
 140                 else
 141                         abort();
 142                 p += 6;
 143                 if (atoi(p) != getpid())
 144                         continue;
 145                 p = strchr(strchr(p, ' ') + 1, ' ') + 1;
 146                 start = atoi(p);
 147                 p = strchr(p, ' ') + 1;
 148                 if (strncmp(p, "EOF", 3) == 0)
 149                         l = 0;
 150                 else
 151                         l = atoi(p) - start + 1;
 152
 153                 if (off == start) {
 154                         if (len != l) {
 155                                 fprintf(stderr, "Len %u should be %u: %s",
 156                                         (int)len, l, line);
 157                                 abort();
 158                         }
 159                         if (type != rw) {
 160                                 fprintf(stderr, "Type %s wrong: %s",
 161                                         rw == F_RDLCK ? "READ" : "WRITE", line);
 162                                 abort();
 163                         }
 164                         found = true;
 165                         break;
 166                 }
 167         }
 168
 169         if (!found) {
 170                 fprintf(stderr, "Unlock on %u@%u not found!",
 171                         (int)off, (int)len);
 172                 abort();
 173         }
 174
 175         fclose(locks);
 176 #endif
 177
 178         return tdb->unlock_fn(tdb->file->fd, rw, off, len, tdb->lock_data);
 179 }
 180
 181 /* a byte range locking function - return 0 on success
 182    this functions locks len bytes at the specified offset.
 183
 184    note that a len of zero means lock to end of file
 185 */
 186 static enum TDB_ERROR tdb_brlock(struct tdb_context *tdb,
 187                                  int rw_type, tdb_off_t offset, tdb_off_t len,
 188                                  enum tdb_lock_flags flags)
 189 {
 190         int ret;
 191
 192         if (tdb->flags & TDB_NOLOCK) {
 193                 return TDB_SUCCESS;
 194         }
 195
 196         if (rw_type == F_WRLCK && (tdb->flags & TDB_RDONLY)) {
 197                 return tdb_logerr(tdb, TDB_ERR_RDONLY, TDB_LOG_USE_ERROR,
 198                                   "Write lock attempted on read-only database");
 199         }
 200
 201         /* A 32 bit system cannot open a 64-bit file, but it could have
 202          * expanded since then: check here. */
 203         if ((size_t)(offset + len) != offset + len) {
 204                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 205                                   "tdb_brlock: lock on giant offset %llu",
 206                                   (long long)(offset + len));
 207         }
 208
 209         ret = lock(tdb, rw_type, offset, len, flags & TDB_LOCK_WAIT);
 210         if (ret != 0) {
 211                 /* Generic lock error. errno set by fcntl.
 212                  * EAGAIN is an expected return from non-blocking
 213                  * locks. */
 214                 if (!(flags & TDB_LOCK_PROBE)
 215                     && (errno != EAGAIN && errno != EINTR)) {
 216                         tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 217                                    "tdb_brlock failed (fd=%d) at"
 218                                    " offset %zu rw_type=%d flags=%d len=%zu:"
 219                                    " %s",
 220                                    tdb->file->fd, (size_t)offset, rw_type,
 221                                    flags, (size_t)len, strerror(errno));
 222                 }
 223                 return TDB_ERR_LOCK;
 224         }
 225         return TDB_SUCCESS;
 226 }
 227
 228 static enum TDB_ERROR tdb_brunlock(struct tdb_context *tdb,
 229                                    int rw_type, tdb_off_t offset, size_t len)
 230 {
 231         if (tdb->flags & TDB_NOLOCK) {
 232                 return TDB_SUCCESS;
 233         }
 234
 235         if (!check_lock_pid(tdb, "tdb_brunlock", true))
 236                 return TDB_ERR_LOCK;
 237
 238         if (unlock(tdb, rw_type, offset, len) == -1) {
 239                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 240                                   "tdb_brunlock failed (fd=%d) at offset %zu"
 241                                   " rw_type=%d len=%zu: %s",
 242                                   tdb->file->fd, (size_t)offset, rw_type,
 243                                   (size_t)len, strerror(errno));
 244         }
 245         return TDB_SUCCESS;
 246 }
 247
 248 /*
 249   upgrade a read lock to a write lock. This needs to be handled in a
 250   special way as some OSes (such as solaris) have too conservative
 251   deadlock detection and claim a deadlock when progress can be
 252   made. For those OSes we may loop for a while.
 253 */
 254 enum TDB_ERROR tdb_allrecord_upgrade(struct tdb_context *tdb, off_t start)
 255 {
 256         int count = 1000;
 257
 258         if (!check_lock_pid(tdb, "tdb_transaction_prepare_commit", true))
 259                 return TDB_ERR_LOCK;
 260
 261         if (tdb->file->allrecord_lock.count != 1) {
 262                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 263                                   "tdb_allrecord_upgrade failed:"
 264                                   " count %u too high",
 265                                   tdb->file->allrecord_lock.count);
 266         }
 267
 268         if (tdb->file->allrecord_lock.off != 1) {
 269                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 270                                   "tdb_allrecord_upgrade failed:"
 271                                   " already upgraded?");
 272         }
 273
 274         if (tdb->file->allrecord_lock.owner != tdb) {
 275                 return owner_conflict(tdb, "tdb_allrecord_upgrade");
 276         }
 277
 278         while (count--) {
 279                 struct timeval tv;
 280                 if (tdb_brlock(tdb, F_WRLCK, start, 0,
 281                                TDB_LOCK_WAIT|TDB_LOCK_PROBE) == TDB_SUCCESS) {
 282                         tdb->file->allrecord_lock.ltype = F_WRLCK;
 283                         tdb->file->allrecord_lock.off = 0;
 284                         return TDB_SUCCESS;
 285                 }
 286                 if (errno != EDEADLK) {
 287                         break;
 288                 }
 289                 /* sleep for as short a time as we can - more portable than usleep() */
 290                 tv.tv_sec = 0;
 291                 tv.tv_usec = 1;
 292                 select(0, NULL, NULL, NULL, &tv);
 293         }
 294
 295         if (errno != EAGAIN && errno != EINTR)
 296                 tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 297                            "tdb_allrecord_upgrade failed");
 298         return TDB_ERR_LOCK;
 299 }
 300
 301 static struct tdb_lock *find_nestlock(struct tdb_context *tdb, tdb_off_t offset,
 302                                       const struct tdb_context *owner)
 303 {
 304         unsigned int i;
 305
 306         for (i=0; i<tdb->file->num_lockrecs; i++) {
 307                 if (tdb->file->lockrecs[i].off == offset) {
 308                         if (owner && tdb->file->lockrecs[i].owner != owner)
 309                                 return NULL;
 310                         return &tdb->file->lockrecs[i];
 311                 }
 312         }
 313         return NULL;
 314 }
 315
 316 enum TDB_ERROR tdb_lock_and_recover(struct tdb_context *tdb)
 317 {
 318         enum TDB_ERROR ecode;
 319
 320         if (!check_lock_pid(tdb, "tdb_transaction_prepare_commit", true))
 321                 return TDB_ERR_LOCK;
 322
 323         ecode = tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK,
 324                                    false);
 325         if (ecode != TDB_SUCCESS) {
 326                 return ecode;
 327         }
 328
 329         ecode = tdb_lock_open(tdb, F_WRLCK, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK);
 330         if (ecode != TDB_SUCCESS) {
 331                 tdb_allrecord_unlock(tdb, F_WRLCK);
 332                 return ecode;
 333         }
 334         ecode = tdb_transaction_recover(tdb);
 335         tdb_unlock_open(tdb, F_WRLCK);
 336         tdb_allrecord_unlock(tdb, F_WRLCK);
 337
 338         return ecode;
 339 }
 340
 341 /* lock an offset in the database. */
 342 static enum TDB_ERROR tdb_nest_lock(struct tdb_context *tdb,
 343                                     tdb_off_t offset, int ltype,
 344                                     enum tdb_lock_flags flags)
 345 {
 346         struct tdb_lock *new_lck;
 347         enum TDB_ERROR ecode;
 348
 349         if (offset > (TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE
 350                       + tdb->file->map_size / 8)) {
 351                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 352                                   "tdb_nest_lock: invalid offset %zu ltype=%d",
 353                                   (size_t)offset, ltype);
 354         }
 355
 356         if (tdb->flags & TDB_NOLOCK)
 357                 return TDB_SUCCESS;
 358
 359         if (!check_lock_pid(tdb, "tdb_nest_lock", true)) {
 360                 return TDB_ERR_LOCK;
 361         }
 362
 363         tdb->stats.locks++;
 364
 365         new_lck = find_nestlock(tdb, offset, NULL);
 366         if (new_lck) {
 367                 if (new_lck->owner != tdb) {
 368                         return owner_conflict(tdb, "tdb_nest_lock");
 369                 }
 370
 371                 if (new_lck->ltype == F_RDLCK && ltype == F_WRLCK) {
 372                         return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 373                                           "tdb_nest_lock:"
 374                                           " offset %zu has read lock",
 375                                           (size_t)offset);
 376                 }
 377                 /* Just increment the struct, posix locks don't stack. */
 378                 new_lck->count++;
 379                 return TDB_SUCCESS;
 380         }
 381
 382 #if 0
 383         if (tdb->file->num_lockrecs
 384             && offset >= TDB_HASH_LOCK_START
 385             && offset < TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE) {
 386                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 387                                   "tdb_nest_lock: already have a hash lock?");
 388         }
 389 #endif
 390
 391         new_lck = (struct tdb_lock *)realloc(
 392                 tdb->file->lockrecs,
 393                 sizeof(*tdb->file->lockrecs) * (tdb->file->num_lockrecs+1));
 394         if (new_lck == NULL) {
 395                 return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
 396                                   "tdb_nest_lock:"
 397                                   " unable to allocate %zu lock struct",
 398                                   tdb->file->num_lockrecs + 1);
 399         }
 400         tdb->file->lockrecs = new_lck;
 401
 402         /* Since fcntl locks don't nest, we do a lock for the first one,
 403            and simply bump the count for future ones */
 404         ecode = tdb_brlock(tdb, ltype, offset, 1, flags);
 405         if (ecode != TDB_SUCCESS) {
 406                 return ecode;
 407         }
 408
 409         /* First time we grab a lock, perhaps someone died in commit? */
 410         if (!(flags & TDB_LOCK_NOCHECK)
 411             && tdb->file->num_lockrecs == 0) {
 412                 tdb_bool_err berr = tdb_needs_recovery(tdb);
 413                 if (berr != false) {
 414                         tdb_brunlock(tdb, ltype, offset, 1);
 415
 416                         if (berr < 0)
 417                                 return TDB_OFF_TO_ERR(berr);
 418                         ecode = tdb_lock_and_recover(tdb);
 419                         if (ecode == TDB_SUCCESS) {
 420                                 ecode = tdb_brlock(tdb, ltype, offset, 1,
 421                                                    flags);
 422                         }
 423                         if (ecode != TDB_SUCCESS) {
 424                                 return ecode;
 425                         }
 426                 }
 427         }
 428
 429         tdb->file->lockrecs[tdb->file->num_lockrecs].owner = tdb;
 430         tdb->file->lockrecs[tdb->file->num_lockrecs].off = offset;
 431         tdb->file->lockrecs[tdb->file->num_lockrecs].count = 1;
 432         tdb->file->lockrecs[tdb->file->num_lockrecs].ltype = ltype;
 433         tdb->file->num_lockrecs++;
 434
 435         return TDB_SUCCESS;
 436 }
 437
 438 static enum TDB_ERROR tdb_nest_unlock(struct tdb_context *tdb,
 439                                       tdb_off_t off, int ltype)
 440 {
 441         struct tdb_lock *lck;
 442         enum TDB_ERROR ecode;
 443
 444         if (tdb->flags & TDB_NOLOCK)
 445                 return TDB_SUCCESS;
 446
 447         lck = find_nestlock(tdb, off, tdb);
 448         if ((lck == NULL) || (lck->count == 0)) {
 449                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 450                                   "tdb_nest_unlock: no lock for %zu",
 451                                   (size_t)off);
 452         }
 453
 454         if (lck->count > 1) {
 455                 lck->count--;
 456                 return TDB_SUCCESS;
 457         }
 458
 459         /*
 460          * This lock has count==1 left, so we need to unlock it in the
 461          * kernel. We don't bother with decrementing the in-memory array
 462          * element, we're about to overwrite it with the last array element
 463          * anyway.
 464          */
 465         ecode = tdb_brunlock(tdb, ltype, off, 1);
 466
 467         /*
 468          * Shrink the array by overwriting the element just unlocked with the
 469          * last array element.
 470          */
 471         *lck = tdb->file->lockrecs[--tdb->file->num_lockrecs];
 472
 473         return ecode;
 474 }
 475
 476 /*
 477   get the transaction lock
 478  */
 479 enum TDB_ERROR tdb_transaction_lock(struct tdb_context *tdb, int ltype)
 480 {
 481         return tdb_nest_lock(tdb, TDB_TRANSACTION_LOCK, ltype, TDB_LOCK_WAIT);
 482 }
 483
 484 /*
 485   release the transaction lock
 486  */
 487 void tdb_transaction_unlock(struct tdb_context *tdb, int ltype)
 488 {
 489         tdb_nest_unlock(tdb, TDB_TRANSACTION_LOCK, ltype);
 490 }
 491
 492 /* We only need to lock individual bytes, but Linux merges consecutive locks
 493  * so we lock in contiguous ranges. */
 494 static enum TDB_ERROR tdb_lock_gradual(struct tdb_context *tdb,
 495                                        int ltype, enum tdb_lock_flags flags,
 496                                        tdb_off_t off, tdb_off_t len)
 497 {
 498         enum TDB_ERROR ecode;
 499         enum tdb_lock_flags nb_flags = (flags & ~TDB_LOCK_WAIT);
 500
 501         if (len <= 1) {
 502                 /* 0 would mean to end-of-file... */
 503                 assert(len != 0);
 504                 /* Single hash.  Just do blocking lock. */
 505                 return tdb_brlock(tdb, ltype, off, len, flags);
 506         }
 507
 508         /* First we try non-blocking. */
 509         ecode = tdb_brlock(tdb, ltype, off, len, nb_flags);
 510         if (ecode != TDB_ERR_LOCK) {
 511                 return ecode;
 512         }
 513
 514         /* Try locking first half, then second. */
 515         ecode = tdb_lock_gradual(tdb, ltype, flags, off, len / 2);
 516         if (ecode != TDB_SUCCESS)
 517                 return ecode;
 518
 519         ecode = tdb_lock_gradual(tdb, ltype, flags,
 520                                  off + len / 2, len - len / 2);
 521         if (ecode != TDB_SUCCESS) {
 522                 tdb_brunlock(tdb, ltype, off, len / 2);
 523         }
 524         return ecode;
 525 }
 526
 527 /* lock/unlock entire database.  It can only be upgradable if you have some
 528  * other way of guaranteeing exclusivity (ie. transaction write lock). */
 529 enum TDB_ERROR tdb_allrecord_lock(struct tdb_context *tdb, int ltype,
 530                                   enum tdb_lock_flags flags, bool upgradable)
 531 {
 532         enum TDB_ERROR ecode;
 533         tdb_bool_err berr;
 534
 535         if (tdb->flags & TDB_NOLOCK)
 536                 return TDB_SUCCESS;
 537
 538         if (!check_lock_pid(tdb, "tdb_allrecord_lock", true)) {
 539                 return TDB_ERR_LOCK;
 540         }
 541
 542         if (tdb->file->allrecord_lock.count) {
 543                 if (tdb->file->allrecord_lock.owner != tdb) {
 544                         return owner_conflict(tdb, "tdb_allrecord_lock");
 545                 }
 546
 547                 if (ltype == F_RDLCK
 548                     || tdb->file->allrecord_lock.ltype == F_WRLCK) {
 549                         tdb->file->allrecord_lock.count++;
 550                         return TDB_SUCCESS;
 551                 }
 552
 553                 /* a global lock of a different type exists */
 554                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
 555                                   "tdb_allrecord_lock: already have %s lock",
 556                                   tdb->file->allrecord_lock.ltype == F_RDLCK
 557                                   ? "read" : "write");
 558         }
 559
 560         if (tdb_has_hash_locks(tdb)) {
 561                 /* can't combine global and chain locks */
 562                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
 563                                   "tdb_allrecord_lock:"
 564                                   " already have chain lock");
 565         }
 566
 567         if (upgradable && ltype != F_RDLCK) {
 568                 /* tdb error: you can't upgrade a write lock! */
 569                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 570                                   "tdb_allrecord_lock:"
 571                                   " can't upgrade a write lock");
 572         }
 573
 574         tdb->stats.locks++;
 575 again:
 576         /* Lock hashes, gradually. */
 577         ecode = tdb_lock_gradual(tdb, ltype, flags, TDB_HASH_LOCK_START,
 578                                  TDB_HASH_LOCK_RANGE);
 579         if (ecode != TDB_SUCCESS)
 580                 return ecode;
 581
 582         /* Lock free tables: there to end of file. */
 583         ecode = tdb_brlock(tdb, ltype,
 584                            TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE,
 585                            0, flags);
 586         if (ecode != TDB_SUCCESS) {
 587                 tdb_brunlock(tdb, ltype, TDB_HASH_LOCK_START,
 588                              TDB_HASH_LOCK_RANGE);
 589                 return ecode;
 590         }
 591
 592         tdb->file->allrecord_lock.owner = tdb;
 593         tdb->file->allrecord_lock.count = 1;
 594         /* If it's upgradable, it's actually exclusive so we can treat
 595          * it as a write lock. */
 596         tdb->file->allrecord_lock.ltype = upgradable ? F_WRLCK : ltype;
 597         tdb->file->allrecord_lock.off = upgradable;
 598
 599         /* Now check for needing recovery. */
 600         if (flags & TDB_LOCK_NOCHECK)
 601                 return TDB_SUCCESS;
 602
 603         berr = tdb_needs_recovery(tdb);
 604         if (likely(berr == false))
 605                 return TDB_SUCCESS;
 606
 607         tdb_allrecord_unlock(tdb, ltype);
 608         if (berr < 0)
 609                 return TDB_OFF_TO_ERR(berr);
 610         ecode = tdb_lock_and_recover(tdb);
 611         if (ecode != TDB_SUCCESS) {
 612                 return ecode;
 613         }
 614         goto again;
 615 }
 616
 617 enum TDB_ERROR tdb_lock_open(struct tdb_context *tdb,
 618                              int ltype, enum tdb_lock_flags flags)
 619 {
 620         return tdb_nest_lock(tdb, TDB_OPEN_LOCK, ltype, flags);
 621 }
 622
 623 void tdb_unlock_open(struct tdb_context *tdb, int ltype)
 624 {
 625         tdb_nest_unlock(tdb, TDB_OPEN_LOCK, ltype);
 626 }
 627
 628 bool tdb_has_open_lock(struct tdb_context *tdb)
 629 {
 630         return !(tdb->flags & TDB_NOLOCK)
 631                 && find_nestlock(tdb, TDB_OPEN_LOCK, tdb) != NULL;
 632 }
 633
 634 enum TDB_ERROR tdb_lock_expand(struct tdb_context *tdb, int ltype)
 635 {
 636         /* Lock doesn't protect data, so don't check (we recurse if we do!) */
 637         return tdb_nest_lock(tdb, TDB_EXPANSION_LOCK, ltype,
 638                              TDB_LOCK_WAIT | TDB_LOCK_NOCHECK);
 639 }
 640
 641 void tdb_unlock_expand(struct tdb_context *tdb, int ltype)
 642 {
 643         tdb_nest_unlock(tdb, TDB_EXPANSION_LOCK, ltype);
 644 }
 645
 646 /* unlock entire db */
 647 void tdb_allrecord_unlock(struct tdb_context *tdb, int ltype)
 648 {
 649         if (tdb->flags & TDB_NOLOCK)
 650                 return;
 651
 652         if (tdb->file->allrecord_lock.count == 0) {
 653                 tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
 654                            "tdb_allrecord_unlock: not locked!");
 655                 return;
 656         }
 657
 658         if (tdb->file->allrecord_lock.owner != tdb) {
 659                 tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
 660                            "tdb_allrecord_unlock: not locked by us!");
 661                 return;
 662         }
 663
 664         /* Upgradable locks are marked as write locks. */
 665         if (tdb->file->allrecord_lock.ltype != ltype
 666             && (!tdb->file->allrecord_lock.off || ltype != F_RDLCK)) {
 667                 tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 668                            "tdb_allrecord_unlock: have %s lock",
 669                            tdb->file->allrecord_lock.ltype == F_RDLCK
 670                            ? "read" : "write");
 671                 return;
 672         }
 673
 674         if (tdb->file->allrecord_lock.count > 1) {
 675                 tdb->file->allrecord_lock.count--;
 676                 return;
 677         }
 678
 679         tdb->file->allrecord_lock.count = 0;
 680         tdb->file->allrecord_lock.ltype = 0;
 681
 682         tdb_brunlock(tdb, ltype, TDB_HASH_LOCK_START, 0);
 683 }
 684
 685 bool tdb_has_expansion_lock(struct tdb_context *tdb)
 686 {
 687         return find_nestlock(tdb, TDB_EXPANSION_LOCK, tdb) != NULL;
 688 }
 689
 690 bool tdb_has_hash_locks(struct tdb_context *tdb)
 691 {
 692         unsigned int i;
 693
 694         for (i=0; i<tdb->file->num_lockrecs; i++) {
 695                 if (tdb->file->lockrecs[i].off >= TDB_HASH_LOCK_START
 696                     && tdb->file->lockrecs[i].off < (TDB_HASH_LOCK_START
 697                                                      + TDB_HASH_LOCK_RANGE))
 698                         return true;
 699         }
 700         return false;
 701 }
 702
 703 static bool tdb_has_free_lock(struct tdb_context *tdb)
 704 {
 705         unsigned int i;
 706
 707         if (tdb->flags & TDB_NOLOCK)
 708                 return false;
 709
 710         for (i=0; i<tdb->file->num_lockrecs; i++) {
 711                 if (tdb->file->lockrecs[i].off
 712                     > TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE)
 713                         return true;
 714         }
 715         return false;
 716 }
 717
 718 enum TDB_ERROR tdb_lock_hashes(struct tdb_context *tdb,
 719                                tdb_off_t hash_lock,
 720                                tdb_len_t hash_range,
 721                                int ltype, enum tdb_lock_flags waitflag)
 722 {
 723         /* FIXME: Do this properly, using hlock_range */
 724         unsigned l = TDB_HASH_LOCK_START
 725                 + (hash_lock >> (64 - TDB_HASH_LOCK_RANGE_BITS));
 726
 727         /* a allrecord lock allows us to avoid per chain locks */
 728         if (tdb->file->allrecord_lock.count) {
 729                 if (!check_lock_pid(tdb, "tdb_lock_hashes", true))
 730                         return TDB_ERR_LOCK;
 731
 732                 if (tdb->file->allrecord_lock.owner != tdb)
 733                         return owner_conflict(tdb, "tdb_lock_hashes");
 734                 if (ltype == tdb->file->allrecord_lock.ltype
 735                     || ltype == F_RDLCK) {
 736                         return TDB_SUCCESS;
 737                 }
 738
 739                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
 740                                   "tdb_lock_hashes:"
 741                                   " already have %s allrecordlock",
 742                                   tdb->file->allrecord_lock.ltype == F_RDLCK
 743                                   ? "read" : "write");
 744         }
 745
 746         if (tdb_has_free_lock(tdb)) {
 747                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 748                                   "tdb_lock_hashes: already have free lock");
 749         }
 750
 751         if (tdb_has_expansion_lock(tdb)) {
 752                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 753                                   "tdb_lock_hashes:"
 754                                   " already have expansion lock");
 755         }
 756
 757         return tdb_nest_lock(tdb, l, ltype, waitflag);
 758 }
 759
 760 enum TDB_ERROR tdb_unlock_hashes(struct tdb_context *tdb,
 761                                  tdb_off_t hash_lock,
 762                                  tdb_len_t hash_range, int ltype)
 763 {
 764         unsigned l = TDB_HASH_LOCK_START
 765                 + (hash_lock >> (64 - TDB_HASH_LOCK_RANGE_BITS));
 766
 767         if (tdb->flags & TDB_NOLOCK)
 768                 return 0;
 769
 770         /* a allrecord lock allows us to avoid per chain locks */
 771         if (tdb->file->allrecord_lock.count) {
 772                 if (tdb->file->allrecord_lock.ltype == F_RDLCK
 773                     && ltype == F_WRLCK) {
 774                         return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 775                                           "tdb_unlock_hashes RO allrecord!");
 776                 }
 777                 if (tdb->file->allrecord_lock.owner != tdb) {
 778                         return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
 779                                           "tdb_unlock_hashes:"
 780                                           " not locked by us!");
 781                 }
 782                 return TDB_SUCCESS;
 783         }
 784
 785         return tdb_nest_unlock(tdb, l, ltype);
 786 }
 787
 788 /* Hash locks use TDB_HASH_LOCK_START + the next 30 bits.
 789  * Then we begin; bucket offsets are sizeof(tdb_len_t) apart, so we divide.
 790  * The result is that on 32 bit systems we don't use lock values > 2^31 on
 791  * files that are less than 4GB.
 792  */
 793 static tdb_off_t free_lock_off(tdb_off_t b_off)
 794 {
 795         return TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE
 796                 + b_off / sizeof(tdb_off_t);
 797 }
 798
 799 enum TDB_ERROR tdb_lock_free_bucket(struct tdb_context *tdb, tdb_off_t b_off,
 800                                     enum tdb_lock_flags waitflag)
 801 {
 802         assert(b_off >= sizeof(struct tdb_header));
 803
 804         if (tdb->flags & TDB_NOLOCK)
 805                 return 0;
 806
 807         /* a allrecord lock allows us to avoid per chain locks */
 808         if (tdb->file->allrecord_lock.count) {
 809                 if (!check_lock_pid(tdb, "tdb_lock_free_bucket", true))
 810                         return TDB_ERR_LOCK;
 811
 812                 if (tdb->file->allrecord_lock.owner != tdb) {
 813                         return owner_conflict(tdb, "tdb_lock_free_bucket");
 814                 }
 815
 816                 if (tdb->file->allrecord_lock.ltype == F_WRLCK)
 817                         return 0;
 818                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 819                                   "tdb_lock_free_bucket with"
 820                                   " read-only allrecordlock!");
 821         }
 822
 823 #if 0 /* FIXME */
 824         if (tdb_has_expansion_lock(tdb)) {
 825                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 826                                   "tdb_lock_free_bucket:"
 827                                   " already have expansion lock");
 828         }
 829 #endif
 830
 831         return tdb_nest_lock(tdb, free_lock_off(b_off), F_WRLCK, waitflag);
 832 }
 833
 834 void tdb_unlock_free_bucket(struct tdb_context *tdb, tdb_off_t b_off)
 835 {
 836         if (tdb->file->allrecord_lock.count)
 837                 return;
 838
 839         tdb_nest_unlock(tdb, free_lock_off(b_off), F_WRLCK);
 840 }
 841
 842 _PUBLIC_ enum TDB_ERROR tdb_lockall(struct tdb_context *tdb)
 843 {
 844         return tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT, false);
 845 }
 846
 847 _PUBLIC_ void tdb_unlockall(struct tdb_context *tdb)
 848 {
 849         tdb_allrecord_unlock(tdb, F_WRLCK);
 850 }
 851
 852 _PUBLIC_ enum TDB_ERROR tdb_lockall_read(struct tdb_context *tdb)
 853 {
 854         return tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, false);
 855 }
 856
 857 _PUBLIC_ void tdb_unlockall_read(struct tdb_context *tdb)
 858 {
 859         tdb_allrecord_unlock(tdb, F_RDLCK);
 860 }
 861
 862 void tdb_lock_cleanup(struct tdb_context *tdb)
 863 {
 864         unsigned int i;
 865
 866         /* We don't want to warn: they're allowed to close tdb after fork. */
 867         if (!check_lock_pid(tdb, "tdb_close", false))
 868                 return;
 869
 870         while (tdb->file->allrecord_lock.count
 871                && tdb->file->allrecord_lock.owner == tdb) {
 872                 tdb_allrecord_unlock(tdb, tdb->file->allrecord_lock.ltype);
 873         }
 874
 875         for (i=0; i<tdb->file->num_lockrecs; i++) {
 876                 if (tdb->file->lockrecs[i].owner == tdb) {
 877                         tdb_nest_unlock(tdb,
 878                                         tdb->file->lockrecs[i].off,
 879                                         tdb->file->lockrecs[i].ltype);
 880                         i--;
 881                 }
 882         }
 883 }