git.ozlabs.org Git - ccan/blob - 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 /* 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 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 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 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 (!(tdb->flags & TDB_VERSION1)
 350             && offset > (TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE
 351                          + tdb->file->map_size / 8)) {
 352                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 353                                   "tdb_nest_lock: invalid offset %zu ltype=%d",
 354                                   (size_t)offset, ltype);
 355         }
 356
 357         if (tdb->flags & TDB_NOLOCK)
 358                 return TDB_SUCCESS;
 359
 360         if (!check_lock_pid(tdb, "tdb_nest_lock", true)) {
 361                 return TDB_ERR_LOCK;
 362         }
 363
 364         tdb->stats.locks++;
 365
 366         new_lck = find_nestlock(tdb, offset, NULL);
 367         if (new_lck) {
 368                 if (new_lck->owner != tdb) {
 369                         return owner_conflict(tdb, "tdb_nest_lock");
 370                 }
 371
 372                 if (new_lck->ltype == F_RDLCK && ltype == F_WRLCK) {
 373                         return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 374                                           "tdb_nest_lock:"
 375                                           " offset %zu has read lock",
 376                                           (size_t)offset);
 377                 }
 378                 /* Just increment the struct, posix locks don't stack. */
 379                 new_lck->count++;
 380                 return TDB_SUCCESS;
 381         }
 382
 383 #if 0
 384         if (tdb->file->num_lockrecs
 385             && offset >= TDB_HASH_LOCK_START
 386             && offset < TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE) {
 387                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 388                                   "tdb_nest_lock: already have a hash lock?");
 389         }
 390 #endif
 391
 392         new_lck = (struct tdb_lock *)realloc(
 393                 tdb->file->lockrecs,
 394                 sizeof(*tdb->file->lockrecs) * (tdb->file->num_lockrecs+1));
 395         if (new_lck == NULL) {
 396                 return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
 397                                   "tdb_nest_lock:"
 398                                   " unable to allocate %zu lock struct",
 399                                   tdb->file->num_lockrecs + 1);
 400         }
 401         tdb->file->lockrecs = new_lck;
 402
 403         /* Since fcntl locks don't nest, we do a lock for the first one,
 404            and simply bump the count for future ones */
 405         ecode = tdb_brlock(tdb, ltype, offset, 1, flags);
 406         if (ecode != TDB_SUCCESS) {
 407                 return ecode;
 408         }
 409
 410         /* First time we grab a lock, perhaps someone died in commit? */
 411         if (!(flags & TDB_LOCK_NOCHECK)
 412             && tdb->file->num_lockrecs == 0) {
 413                 tdb_bool_err berr = tdb_needs_recovery(tdb);
 414                 if (berr != false) {
 415                         tdb_brunlock(tdb, ltype, offset, 1);
 416
 417                         if (berr < 0)
 418                                 return berr;
 419                         ecode = tdb_lock_and_recover(tdb);
 420                         if (ecode == TDB_SUCCESS) {
 421                                 ecode = tdb_brlock(tdb, ltype, offset, 1,
 422                                                    flags);
 423                         }
 424                         if (ecode != TDB_SUCCESS) {
 425                                 return ecode;
 426                         }
 427                 }
 428         }
 429
 430         tdb->file->lockrecs[tdb->file->num_lockrecs].owner = tdb;
 431         tdb->file->lockrecs[tdb->file->num_lockrecs].off = offset;
 432         tdb->file->lockrecs[tdb->file->num_lockrecs].count = 1;
 433         tdb->file->lockrecs[tdb->file->num_lockrecs].ltype = ltype;
 434         tdb->file->num_lockrecs++;
 435
 436         return TDB_SUCCESS;
 437 }
 438
 439 enum TDB_ERROR tdb_nest_unlock(struct tdb_context *tdb,
 440                                tdb_off_t off, int ltype)
 441 {
 442         struct tdb_lock *lck;
 443         enum TDB_ERROR ecode;
 444
 445         if (tdb->flags & TDB_NOLOCK)
 446                 return TDB_SUCCESS;
 447
 448         lck = find_nestlock(tdb, off, tdb);
 449         if ((lck == NULL) || (lck->count == 0)) {
 450                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 451                                   "tdb_nest_unlock: no lock for %zu",
 452                                   (size_t)off);
 453         }
 454
 455         if (lck->count > 1) {
 456                 lck->count--;
 457                 return TDB_SUCCESS;
 458         }
 459
 460         /*
 461          * This lock has count==1 left, so we need to unlock it in the
 462          * kernel. We don't bother with decrementing the in-memory array
 463          * element, we're about to overwrite it with the last array element
 464          * anyway.
 465          */
 466         ecode = tdb_brunlock(tdb, ltype, off, 1);
 467
 468         /*
 469          * Shrink the array by overwriting the element just unlocked with the
 470          * last array element.
 471          */
 472         *lck = tdb->file->lockrecs[--tdb->file->num_lockrecs];
 473
 474         return ecode;
 475 }
 476
 477 /*
 478   get the transaction lock
 479  */
 480 enum TDB_ERROR tdb_transaction_lock(struct tdb_context *tdb, int ltype)
 481 {
 482         return tdb_nest_lock(tdb, TDB_TRANSACTION_LOCK, ltype, TDB_LOCK_WAIT);
 483 }
 484
 485 /*
 486   release the transaction lock
 487  */
 488 void tdb_transaction_unlock(struct tdb_context *tdb, int ltype)
 489 {
 490         tdb_nest_unlock(tdb, TDB_TRANSACTION_LOCK, ltype);
 491 }
 492
 493 /* We only need to lock individual bytes, but Linux merges consecutive locks
 494  * so we lock in contiguous ranges. */
 495 enum TDB_ERROR tdb_lock_gradual(struct tdb_context *tdb,
 496                                 int ltype, enum tdb_lock_flags flags,
 497                                 tdb_off_t off, tdb_off_t len)
 498 {
 499         enum TDB_ERROR ecode;
 500         enum tdb_lock_flags nb_flags = (flags & ~TDB_LOCK_WAIT);
 501
 502         if (len <= 1) {
 503                 /* 0 would mean to end-of-file... */
 504                 assert(len != 0);
 505                 /* Single hash.  Just do blocking lock. */
 506                 return tdb_brlock(tdb, ltype, off, len, flags);
 507         }
 508
 509         /* First we try non-blocking. */
 510         ecode = tdb_brlock(tdb, ltype, off, len, nb_flags);
 511         if (ecode != TDB_ERR_LOCK) {
 512                 return ecode;
 513         }
 514
 515         /* Try locking first half, then second. */
 516         ecode = tdb_lock_gradual(tdb, ltype, flags, off, len / 2);
 517         if (ecode != TDB_SUCCESS)
 518                 return ecode;
 519
 520         ecode = tdb_lock_gradual(tdb, ltype, flags,
 521                                  off + len / 2, len - len / 2);
 522         if (ecode != TDB_SUCCESS) {
 523                 tdb_brunlock(tdb, ltype, off, len / 2);
 524         }
 525         return ecode;
 526 }
 527
 528 /* lock/unlock entire database.  It can only be upgradable if you have some
 529  * other way of guaranteeing exclusivity (ie. transaction write lock). */
 530 enum TDB_ERROR tdb_allrecord_lock(struct tdb_context *tdb, int ltype,
 531                                   enum tdb_lock_flags flags, bool upgradable)
 532 {
 533         enum TDB_ERROR ecode;
 534         tdb_bool_err berr;
 535
 536         if (tdb->flags & TDB_VERSION1) {
 537                 if (tdb1_allrecord_lock(tdb, ltype, flags, upgradable) == -1)
 538                         return tdb->last_error;
 539                 return TDB_SUCCESS;
 540         }
 541
 542         if (tdb->flags & TDB_NOLOCK)
 543                 return TDB_SUCCESS;
 544
 545         if (!check_lock_pid(tdb, "tdb_allrecord_lock", true)) {
 546                 return TDB_ERR_LOCK;
 547         }
 548
 549         if (tdb->file->allrecord_lock.count) {
 550                 if (tdb->file->allrecord_lock.owner != tdb) {
 551                         return owner_conflict(tdb, "tdb_allrecord_lock");
 552                 }
 553
 554                 if (ltype == F_RDLCK
 555                     || tdb->file->allrecord_lock.ltype == F_WRLCK) {
 556                         tdb->file->allrecord_lock.count++;
 557                         return TDB_SUCCESS;
 558                 }
 559
 560                 /* a global lock of a different type exists */
 561                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
 562                                   "tdb_allrecord_lock: already have %s lock",
 563                                   tdb->file->allrecord_lock.ltype == F_RDLCK
 564                                   ? "read" : "write");
 565         }
 566
 567         if (tdb_has_hash_locks(tdb)) {
 568                 /* can't combine global and chain locks */
 569                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
 570                                   "tdb_allrecord_lock:"
 571                                   " already have chain lock");
 572         }
 573
 574         if (upgradable && ltype != F_RDLCK) {
 575                 /* tdb error: you can't upgrade a write lock! */
 576                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 577                                   "tdb_allrecord_lock:"
 578                                   " can't upgrade a write lock");
 579         }
 580
 581         tdb->stats.locks++;
 582 again:
 583         /* Lock hashes, gradually. */
 584         ecode = tdb_lock_gradual(tdb, ltype, flags, TDB_HASH_LOCK_START,
 585                                  TDB_HASH_LOCK_RANGE);
 586         if (ecode != TDB_SUCCESS)
 587                 return ecode;
 588
 589         /* Lock free tables: there to end of file. */
 590         ecode = tdb_brlock(tdb, ltype,
 591                            TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE,
 592                            0, flags);
 593         if (ecode != TDB_SUCCESS) {
 594                 tdb_brunlock(tdb, ltype, TDB_HASH_LOCK_START,
 595                              TDB_HASH_LOCK_RANGE);
 596                 return ecode;
 597         }
 598
 599         tdb->file->allrecord_lock.owner = tdb;
 600         tdb->file->allrecord_lock.count = 1;
 601         /* If it's upgradable, it's actually exclusive so we can treat
 602          * it as a write lock. */
 603         tdb->file->allrecord_lock.ltype = upgradable ? F_WRLCK : ltype;
 604         tdb->file->allrecord_lock.off = upgradable;
 605
 606         /* Now check for needing recovery. */
 607         if (flags & TDB_LOCK_NOCHECK)
 608                 return TDB_SUCCESS;
 609
 610         berr = tdb_needs_recovery(tdb);
 611         if (likely(berr == false))
 612                 return TDB_SUCCESS;
 613
 614         tdb_allrecord_unlock(tdb, ltype);
 615         if (berr < 0)
 616                 return berr;
 617         ecode = tdb_lock_and_recover(tdb);
 618         if (ecode != TDB_SUCCESS) {
 619                 return ecode;
 620         }
 621         goto again;
 622 }
 623
 624 enum TDB_ERROR tdb_lock_open(struct tdb_context *tdb,
 625                              int ltype, enum tdb_lock_flags flags)
 626 {
 627         return tdb_nest_lock(tdb, TDB_OPEN_LOCK, ltype, flags);
 628 }
 629
 630 void tdb_unlock_open(struct tdb_context *tdb, int ltype)
 631 {
 632         tdb_nest_unlock(tdb, TDB_OPEN_LOCK, ltype);
 633 }
 634
 635 bool tdb_has_open_lock(struct tdb_context *tdb)
 636 {
 637         return !(tdb->flags & TDB_NOLOCK)
 638                 && find_nestlock(tdb, TDB_OPEN_LOCK, tdb) != NULL;
 639 }
 640
 641 enum TDB_ERROR tdb_lock_expand(struct tdb_context *tdb, int ltype)
 642 {
 643         /* Lock doesn't protect data, so don't check (we recurse if we do!) */
 644         return tdb_nest_lock(tdb, TDB_EXPANSION_LOCK, ltype,
 645                              TDB_LOCK_WAIT | TDB_LOCK_NOCHECK);
 646 }
 647
 648 void tdb_unlock_expand(struct tdb_context *tdb, int ltype)
 649 {
 650         tdb_nest_unlock(tdb, TDB_EXPANSION_LOCK, ltype);
 651 }
 652
 653 /* unlock entire db */
 654 void tdb_allrecord_unlock(struct tdb_context *tdb, int ltype)
 655 {
 656         if (tdb->flags & TDB_VERSION1) {
 657                 tdb1_allrecord_unlock(tdb, ltype);
 658                 return;
 659         }
 660
 661         if (tdb->flags & TDB_NOLOCK)
 662                 return;
 663
 664         if (tdb->file->allrecord_lock.count == 0) {
 665                 tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
 666                            "tdb_allrecord_unlock: not locked!");
 667                 return;
 668         }
 669
 670         if (tdb->file->allrecord_lock.owner != tdb) {
 671                 tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
 672                            "tdb_allrecord_unlock: not locked by us!");
 673                 return;
 674         }
 675
 676         /* Upgradable locks are marked as write locks. */
 677         if (tdb->file->allrecord_lock.ltype != ltype
 678             && (!tdb->file->allrecord_lock.off || ltype != F_RDLCK)) {
 679                 tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 680                            "tdb_allrecord_unlock: have %s lock",
 681                            tdb->file->allrecord_lock.ltype == F_RDLCK
 682                            ? "read" : "write");
 683                 return;
 684         }
 685
 686         if (tdb->file->allrecord_lock.count > 1) {
 687                 tdb->file->allrecord_lock.count--;
 688                 return;
 689         }
 690
 691         tdb->file->allrecord_lock.count = 0;
 692         tdb->file->allrecord_lock.ltype = 0;
 693
 694         tdb_brunlock(tdb, ltype, TDB_HASH_LOCK_START, 0);
 695 }
 696
 697 bool tdb_has_expansion_lock(struct tdb_context *tdb)
 698 {
 699         return find_nestlock(tdb, TDB_EXPANSION_LOCK, tdb) != NULL;
 700 }
 701
 702 bool tdb_has_hash_locks(struct tdb_context *tdb)
 703 {
 704         unsigned int i;
 705
 706         for (i=0; i<tdb->file->num_lockrecs; i++) {
 707                 if (tdb->file->lockrecs[i].off >= TDB_HASH_LOCK_START
 708                     && tdb->file->lockrecs[i].off < (TDB_HASH_LOCK_START
 709                                                      + TDB_HASH_LOCK_RANGE))
 710                         return true;
 711         }
 712         return false;
 713 }
 714
 715 static bool tdb_has_free_lock(struct tdb_context *tdb)
 716 {
 717         unsigned int i;
 718
 719         if (tdb->flags & TDB_NOLOCK)
 720                 return false;
 721
 722         for (i=0; i<tdb->file->num_lockrecs; i++) {
 723                 if (tdb->file->lockrecs[i].off
 724                     > TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE)
 725                         return true;
 726         }
 727         return false;
 728 }
 729
 730 enum TDB_ERROR tdb_lock_hashes(struct tdb_context *tdb,
 731                                tdb_off_t hash_lock,
 732                                tdb_len_t hash_range,
 733                                int ltype, enum tdb_lock_flags waitflag)
 734 {
 735         /* FIXME: Do this properly, using hlock_range */
 736         unsigned l = TDB_HASH_LOCK_START
 737                 + (hash_lock >> (64 - TDB_HASH_LOCK_RANGE_BITS));
 738
 739         /* a allrecord lock allows us to avoid per chain locks */
 740         if (tdb->file->allrecord_lock.count) {
 741                 if (!check_lock_pid(tdb, "tdb_lock_hashes", true))
 742                         return TDB_ERR_LOCK;
 743
 744                 if (tdb->file->allrecord_lock.owner != tdb)
 745                         return owner_conflict(tdb, "tdb_lock_hashes");
 746                 if (ltype == tdb->file->allrecord_lock.ltype
 747                     || ltype == F_RDLCK) {
 748                         return TDB_SUCCESS;
 749                 }
 750
 751                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
 752                                   "tdb_lock_hashes:"
 753                                   " already have %s allrecordlock",
 754                                   tdb->file->allrecord_lock.ltype == F_RDLCK
 755                                   ? "read" : "write");
 756         }
 757
 758         if (tdb_has_free_lock(tdb)) {
 759                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 760                                   "tdb_lock_hashes: already have free lock");
 761         }
 762
 763         if (tdb_has_expansion_lock(tdb)) {
 764                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 765                                   "tdb_lock_hashes:"
 766                                   " already have expansion lock");
 767         }
 768
 769         return tdb_nest_lock(tdb, l, ltype, waitflag);
 770 }
 771
 772 enum TDB_ERROR tdb_unlock_hashes(struct tdb_context *tdb,
 773                                  tdb_off_t hash_lock,
 774                                  tdb_len_t hash_range, int ltype)
 775 {
 776         unsigned l = TDB_HASH_LOCK_START
 777                 + (hash_lock >> (64 - TDB_HASH_LOCK_RANGE_BITS));
 778
 779         if (tdb->flags & TDB_NOLOCK)
 780                 return 0;
 781
 782         /* a allrecord lock allows us to avoid per chain locks */
 783         if (tdb->file->allrecord_lock.count) {
 784                 if (tdb->file->allrecord_lock.ltype == F_RDLCK
 785                     && ltype == F_WRLCK) {
 786                         return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 787                                           "tdb_unlock_hashes RO allrecord!");
 788                 }
 789                 if (tdb->file->allrecord_lock.owner != tdb) {
 790                         return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
 791                                           "tdb_unlock_hashes:"
 792                                           " not locked by us!");
 793                 }
 794                 return TDB_SUCCESS;
 795         }
 796
 797         return tdb_nest_unlock(tdb, l, ltype);
 798 }
 799
 800 /* Hash locks use TDB_HASH_LOCK_START + the next 30 bits.
 801  * Then we begin; bucket offsets are sizeof(tdb_len_t) apart, so we divide.
 802  * The result is that on 32 bit systems we don't use lock values > 2^31 on
 803  * files that are less than 4GB.
 804  */
 805 static tdb_off_t free_lock_off(tdb_off_t b_off)
 806 {
 807         return TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE
 808                 + b_off / sizeof(tdb_off_t);
 809 }
 810
 811 enum TDB_ERROR tdb_lock_free_bucket(struct tdb_context *tdb, tdb_off_t b_off,
 812                                     enum tdb_lock_flags waitflag)
 813 {
 814         assert(b_off >= sizeof(struct tdb_header));
 815
 816         if (tdb->flags & TDB_NOLOCK)
 817                 return 0;
 818
 819         /* a allrecord lock allows us to avoid per chain locks */
 820         if (tdb->file->allrecord_lock.count) {
 821                 if (!check_lock_pid(tdb, "tdb_lock_free_bucket", true))
 822                         return TDB_ERR_LOCK;
 823
 824                 if (tdb->file->allrecord_lock.owner != tdb) {
 825                         return owner_conflict(tdb, "tdb_lock_free_bucket");
 826                 }
 827
 828                 if (tdb->file->allrecord_lock.ltype == F_WRLCK)
 829                         return 0;
 830                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 831                                   "tdb_lock_free_bucket with"
 832                                   " read-only allrecordlock!");
 833         }
 834
 835 #if 0 /* FIXME */
 836         if (tdb_has_expansion_lock(tdb)) {
 837                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 838                                   "tdb_lock_free_bucket:"
 839                                   " already have expansion lock");
 840         }
 841 #endif
 842
 843         return tdb_nest_lock(tdb, free_lock_off(b_off), F_WRLCK, waitflag);
 844 }
 845
 846 void tdb_unlock_free_bucket(struct tdb_context *tdb, tdb_off_t b_off)
 847 {
 848         if (tdb->file->allrecord_lock.count)
 849                 return;
 850
 851         tdb_nest_unlock(tdb, free_lock_off(b_off), F_WRLCK);
 852 }
 853
 854 enum TDB_ERROR tdb_lockall(struct tdb_context *tdb)
 855 {
 856         return tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT, false);
 857 }
 858
 859 void tdb_unlockall(struct tdb_context *tdb)
 860 {
 861         tdb_allrecord_unlock(tdb, F_WRLCK);
 862 }
 863
 864 enum TDB_ERROR tdb_lockall_read(struct tdb_context *tdb)
 865 {
 866         return tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, false);
 867 }
 868
 869 void tdb_unlockall_read(struct tdb_context *tdb)
 870 {
 871         tdb_allrecord_unlock(tdb, F_RDLCK);
 872 }
 873
 874 void tdb_lock_cleanup(struct tdb_context *tdb)
 875 {
 876         unsigned int i;
 877
 878         /* We don't want to warn: they're allowed to close tdb after fork. */
 879         if (!check_lock_pid(tdb, "tdb_close", false))
 880                 return;
 881
 882         while (tdb->file->allrecord_lock.count
 883                && tdb->file->allrecord_lock.owner == tdb) {
 884                 tdb_allrecord_unlock(tdb, tdb->file->allrecord_lock.ltype);
 885         }
 886
 887         for (i=0; i<tdb->file->num_lockrecs; i++) {
 888                 if (tdb->file->lockrecs[i].owner == tdb) {
 889                         tdb_nest_unlock(tdb,
 890                                         tdb->file->lockrecs[i].off,
 891                                         tdb->file->lockrecs[i].ltype);
 892                         i--;
 893                 }
 894         }
 895 }