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