lib/ntdb/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 ntdb
  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 <ccan/build_assert/build_assert.h>
  30
  31 /* If we were threaded, we could wait for unlock, but we're not, so fail. */
  32 enum NTDB_ERROR owner_conflict(struct ntdb_context *ntdb, const char *call)
  33 {
  34         return ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_USE_ERROR,
  35                           "%s: lock owned by another ntdb in this process.",
  36                           call);
  37 }
  38
  39 /* If we fork, we no longer really own locks. */
  40 bool check_lock_pid(struct ntdb_context *ntdb, const char *call, bool log)
  41 {
  42         /* No locks?  No problem! */
  43         if (ntdb->file->allrecord_lock.count == 0
  44             && ntdb->file->num_lockrecs == 0) {
  45                 return true;
  46         }
  47
  48         /* No fork?  No problem! */
  49         if (ntdb->file->locker == getpid()) {
  50                 return true;
  51         }
  52
  53         if (log) {
  54                 ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_USE_ERROR,
  55                            "%s: fork() detected after lock acquisition!"
  56                            " (%u vs %u)", call, ntdb->file->locker, getpid());
  57         }
  58         return false;
  59 }
  60
  61 int ntdb_fcntl_lock(int fd, int rw, off_t off, off_t len, bool waitflag,
  62                    void *unused)
  63 {
  64         struct flock fl;
  65         int ret;
  66
  67         do {
  68                 fl.l_type = rw;
  69                 fl.l_whence = SEEK_SET;
  70                 fl.l_start = off;
  71                 fl.l_len = len;
  72
  73                 if (waitflag)
  74                         ret = fcntl(fd, F_SETLKW, &fl);
  75                 else
  76                         ret = fcntl(fd, F_SETLK, &fl);
  77         } while (ret != 0 && errno == EINTR);
  78         return ret;
  79 }
  80
  81 int ntdb_fcntl_unlock(int fd, int rw, off_t off, off_t len, void *unused)
  82 {
  83         struct flock fl;
  84         int ret;
  85
  86         do {
  87                 fl.l_type = F_UNLCK;
  88                 fl.l_whence = SEEK_SET;
  89                 fl.l_start = off;
  90                 fl.l_len = len;
  91
  92                 ret = fcntl(fd, F_SETLKW, &fl);
  93         } while (ret != 0 && errno == EINTR);
  94         return ret;
  95 }
  96
  97 static int lock(struct ntdb_context *ntdb,
  98                       int rw, off_t off, off_t len, bool waitflag)
  99 {
 100         int ret;
 101         if (ntdb->file->allrecord_lock.count == 0
 102             && ntdb->file->num_lockrecs == 0) {
 103                 ntdb->file->locker = getpid();
 104         }
 105
 106         ntdb->stats.lock_lowlevel++;
 107         ret = ntdb->lock_fn(ntdb->file->fd, rw, off, len, waitflag,
 108                            ntdb->lock_data);
 109         if (!waitflag) {
 110                 ntdb->stats.lock_nonblock++;
 111                 if (ret != 0)
 112                         ntdb->stats.lock_nonblock_fail++;
 113         }
 114         return ret;
 115 }
 116
 117 static int unlock(struct ntdb_context *ntdb, int rw, off_t off, off_t len)
 118 {
 119 #if 0 /* Check they matched up locks and unlocks correctly. */
 120         char line[80];
 121         FILE *locks;
 122         bool found = false;
 123
 124         locks = fopen("/proc/locks", "r");
 125
 126         while (fgets(line, 80, locks)) {
 127                 char *p;
 128                 int type, start, l;
 129
 130                 /* eg. 1: FLOCK  ADVISORY  WRITE 2440 08:01:2180826 0 EOF */
 131                 p = strchr(line, ':') + 1;
 132                 if (strncmp(p, " POSIX  ADVISORY  ", strlen(" POSIX  ADVISORY  ")))
 133                         continue;
 134                 p += strlen(" FLOCK  ADVISORY  ");
 135                 if (strncmp(p, "READ  ", strlen("READ  ")) == 0)
 136                         type = F_RDLCK;
 137                 else if (strncmp(p, "WRITE ", strlen("WRITE ")) == 0)
 138                         type = F_WRLCK;
 139                 else
 140                         abort();
 141                 p += 6;
 142                 if (atoi(p) != getpid())
 143                         continue;
 144                 p = strchr(strchr(p, ' ') + 1, ' ') + 1;
 145                 start = atoi(p);
 146                 p = strchr(p, ' ') + 1;
 147                 if (strncmp(p, "EOF", 3) == 0)
 148                         l = 0;
 149                 else
 150                         l = atoi(p) - start + 1;
 151
 152                 if (off == start) {
 153                         if (len != l) {
 154                                 fprintf(stderr, "Len %u should be %u: %s",
 155                                         (int)len, l, line);
 156                                 abort();
 157                         }
 158                         if (type != rw) {
 159                                 fprintf(stderr, "Type %s wrong: %s",
 160                                         rw == F_RDLCK ? "READ" : "WRITE", line);
 161                                 abort();
 162                         }
 163                         found = true;
 164                         break;
 165                 }
 166         }
 167
 168         if (!found) {
 169                 fprintf(stderr, "Unlock on %u@%u not found!",
 170                         (int)off, (int)len);
 171                 abort();
 172         }
 173
 174         fclose(locks);
 175 #endif
 176
 177         return ntdb->unlock_fn(ntdb->file->fd, rw, off, len, ntdb->lock_data);
 178 }
 179
 180 /* a byte range locking function - return 0 on success
 181    this functions locks len bytes at the specified offset.
 182
 183    note that a len of zero means lock to end of file
 184 */
 185 static enum NTDB_ERROR ntdb_brlock(struct ntdb_context *ntdb,
 186                                  int rw_type, ntdb_off_t offset, ntdb_off_t len,
 187                                  enum ntdb_lock_flags flags)
 188 {
 189         int ret;
 190
 191         if (rw_type == F_WRLCK && (ntdb->flags & NTDB_RDONLY)) {
 192                 return ntdb_logerr(ntdb, NTDB_ERR_RDONLY, NTDB_LOG_USE_ERROR,
 193                                   "Write lock attempted on read-only database");
 194         }
 195
 196         if (ntdb->flags & NTDB_NOLOCK) {
 197                 return NTDB_SUCCESS;
 198         }
 199
 200         /* A 32 bit system cannot open a 64-bit file, but it could have
 201          * expanded since then: check here. */
 202         if ((size_t)(offset + len) != offset + len) {
 203                 return ntdb_logerr(ntdb, NTDB_ERR_IO, NTDB_LOG_ERROR,
 204                                   "ntdb_brlock: lock on giant offset %llu",
 205                                   (long long)(offset + len));
 206         }
 207
 208         ret = lock(ntdb, rw_type, offset, len, flags & NTDB_LOCK_WAIT);
 209         if (ret != 0) {
 210                 /* Generic lock error. errno set by fcntl.
 211                  * EAGAIN is an expected return from non-blocking
 212                  * locks. */
 213                 if (!(flags & NTDB_LOCK_PROBE)
 214                     && (errno != EAGAIN && errno != EINTR)) {
 215                         ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_ERROR,
 216                                    "ntdb_brlock failed (fd=%d) at"
 217                                    " offset %zu rw_type=%d flags=%d len=%zu:"
 218                                    " %s",
 219                                    ntdb->file->fd, (size_t)offset, rw_type,
 220                                    flags, (size_t)len, strerror(errno));
 221                 }
 222                 return NTDB_ERR_LOCK;
 223         }
 224         return NTDB_SUCCESS;
 225 }
 226
 227 static enum NTDB_ERROR ntdb_brunlock(struct ntdb_context *ntdb,
 228                                    int rw_type, ntdb_off_t offset, size_t len)
 229 {
 230         if (ntdb->flags & NTDB_NOLOCK) {
 231                 return NTDB_SUCCESS;
 232         }
 233
 234         if (!check_lock_pid(ntdb, "ntdb_brunlock", false))
 235                 return NTDB_ERR_LOCK;
 236
 237         if (unlock(ntdb, rw_type, offset, len) == -1) {
 238                 return ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_ERROR,
 239                                   "ntdb_brunlock failed (fd=%d) at offset %zu"
 240                                   " rw_type=%d len=%zu: %s",
 241                                   ntdb->file->fd, (size_t)offset, rw_type,
 242                                   (size_t)len, strerror(errno));
 243         }
 244         return NTDB_SUCCESS;
 245 }
 246
 247 /*
 248   upgrade a read lock to a write lock. This needs to be handled in a
 249   special way as some OSes (such as solaris) have too conservative
 250   deadlock detection and claim a deadlock when progress can be
 251   made. For those OSes we may loop for a while.
 252 */
 253 enum NTDB_ERROR ntdb_allrecord_upgrade(struct ntdb_context *ntdb, off_t start)
 254 {
 255         int count = 1000;
 256
 257         if (!check_lock_pid(ntdb, "ntdb_transaction_prepare_commit", true))
 258                 return NTDB_ERR_LOCK;
 259
 260         if (ntdb->file->allrecord_lock.count != 1) {
 261                 return ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_ERROR,
 262                                   "ntdb_allrecord_upgrade failed:"
 263                                   " count %u too high",
 264                                   ntdb->file->allrecord_lock.count);
 265         }
 266
 267         if (ntdb->file->allrecord_lock.off != 1) {
 268                 return ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_ERROR,
 269                                   "ntdb_allrecord_upgrade failed:"
 270                                   " already upgraded?");
 271         }
 272
 273         if (ntdb->file->allrecord_lock.owner != ntdb) {
 274                 return owner_conflict(ntdb, "ntdb_allrecord_upgrade");
 275         }
 276
 277         while (count--) {
 278                 struct timeval tv;
 279                 if (ntdb_brlock(ntdb, F_WRLCK, start, 0,
 280                                NTDB_LOCK_WAIT|NTDB_LOCK_PROBE) == NTDB_SUCCESS) {
 281                         ntdb->file->allrecord_lock.ltype = F_WRLCK;
 282                         ntdb->file->allrecord_lock.off = 0;
 283                         return NTDB_SUCCESS;
 284                 }
 285                 if (errno != EDEADLK) {
 286                         break;
 287                 }
 288                 /* sleep for as short a time as we can - more portable than usleep() */
 289                 tv.tv_sec = 0;
 290                 tv.tv_usec = 1;
 291                 select(0, NULL, NULL, NULL, &tv);
 292         }
 293
 294         if (errno != EAGAIN && errno != EINTR)
 295                 ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_ERROR,
 296                            "ntdb_allrecord_upgrade failed");
 297         return NTDB_ERR_LOCK;
 298 }
 299
 300 static struct ntdb_lock *find_nestlock(struct ntdb_context *ntdb, ntdb_off_t offset,
 301                                       const struct ntdb_context *owner)
 302 {
 303         unsigned int i;
 304
 305         for (i=0; i<ntdb->file->num_lockrecs; i++) {
 306                 if (ntdb->file->lockrecs[i].off == offset) {
 307                         if (owner && ntdb->file->lockrecs[i].owner != owner)
 308                                 return NULL;
 309                         return &ntdb->file->lockrecs[i];
 310                 }
 311         }
 312         return NULL;
 313 }
 314
 315 enum NTDB_ERROR ntdb_lock_and_recover(struct ntdb_context *ntdb)
 316 {
 317         enum NTDB_ERROR ecode;
 318
 319         if (!check_lock_pid(ntdb, "ntdb_transaction_prepare_commit", true))
 320                 return NTDB_ERR_LOCK;
 321
 322         ecode = ntdb_allrecord_lock(ntdb, F_WRLCK, NTDB_LOCK_WAIT|NTDB_LOCK_NOCHECK,
 323                                    false);
 324         if (ecode != NTDB_SUCCESS) {
 325                 return ecode;
 326         }
 327
 328         ecode = ntdb_lock_open(ntdb, F_WRLCK, NTDB_LOCK_WAIT|NTDB_LOCK_NOCHECK);
 329         if (ecode != NTDB_SUCCESS) {
 330                 ntdb_allrecord_unlock(ntdb, F_WRLCK);
 331                 return ecode;
 332         }
 333         ecode = ntdb_transaction_recover(ntdb);
 334         ntdb_unlock_open(ntdb, F_WRLCK);
 335         ntdb_allrecord_unlock(ntdb, F_WRLCK);
 336
 337         return ecode;
 338 }
 339
 340 /* lock an offset in the database. */
 341 static enum NTDB_ERROR ntdb_nest_lock(struct ntdb_context *ntdb,
 342                                     ntdb_off_t offset, int ltype,
 343                                     enum ntdb_lock_flags flags)
 344 {
 345         struct ntdb_lock *new_lck;
 346         enum NTDB_ERROR ecode;
 347
 348         assert(offset <= (NTDB_HASH_LOCK_START + (1 << ntdb->hash_bits)
 349                           + ntdb->file->map_size / 8));
 350
 351         if (ntdb->flags & NTDB_NOLOCK)
 352                 return NTDB_SUCCESS;
 353
 354         if (!check_lock_pid(ntdb, "ntdb_nest_lock", true)) {
 355                 return NTDB_ERR_LOCK;
 356         }
 357
 358         ntdb->stats.locks++;
 359
 360         new_lck = find_nestlock(ntdb, offset, NULL);
 361         if (new_lck) {
 362                 if (new_lck->owner != ntdb) {
 363                         return owner_conflict(ntdb, "ntdb_nest_lock");
 364                 }
 365
 366                 if (new_lck->ltype == F_RDLCK && ltype == F_WRLCK) {
 367                         return ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_ERROR,
 368                                           "ntdb_nest_lock:"
 369                                           " offset %zu has read lock",
 370                                           (size_t)offset);
 371                 }
 372                 /* Just increment the struct, posix locks don't stack. */
 373                 new_lck->count++;
 374                 return NTDB_SUCCESS;
 375         }
 376
 377 #if 0
 378         if (ntdb->file->num_lockrecs
 379             && offset >= NTDB_HASH_LOCK_START
 380             && offset < NTDB_HASH_LOCK_START + NTDB_HASH_LOCK_RANGE) {
 381                 return ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_ERROR,
 382                                   "ntdb_nest_lock: already have a hash lock?");
 383         }
 384 #endif
 385         if (ntdb->file->lockrecs == NULL) {
 386                 new_lck = ntdb->alloc_fn(ntdb->file, sizeof(*ntdb->file->lockrecs),
 387                                      ntdb->alloc_data);
 388         } else {
 389                 new_lck = (struct ntdb_lock *)ntdb->expand_fn(
 390                         ntdb->file->lockrecs,
 391                         sizeof(*ntdb->file->lockrecs)
 392                         * (ntdb->file->num_lockrecs+1),
 393                         ntdb->alloc_data);
 394         }
 395         if (new_lck == NULL) {
 396                 return ntdb_logerr(ntdb, NTDB_ERR_OOM, NTDB_LOG_ERROR,
 397                                   "ntdb_nest_lock:"
 398                                   " unable to allocate %zu lock struct",
 399                                   ntdb->file->num_lockrecs + 1);
 400         }
 401         ntdb->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 = ntdb_brlock(ntdb, ltype, offset, 1, flags);
 406         if (ecode != NTDB_SUCCESS) {
 407                 return ecode;
 408         }
 409
 410         /* First time we grab a lock, perhaps someone died in commit? */
 411         if (!(flags & NTDB_LOCK_NOCHECK)
 412             && ntdb->file->num_lockrecs == 0) {
 413                 ntdb_bool_err berr = ntdb_needs_recovery(ntdb);
 414                 if (berr != false) {
 415                         ntdb_brunlock(ntdb, ltype, offset, 1);
 416
 417                         if (berr < 0)
 418                                 return NTDB_OFF_TO_ERR(berr);
 419                         ecode = ntdb_lock_and_recover(ntdb);
 420                         if (ecode == NTDB_SUCCESS) {
 421                                 ecode = ntdb_brlock(ntdb, ltype, offset, 1,
 422                                                    flags);
 423                         }
 424                         if (ecode != NTDB_SUCCESS) {
 425                                 return ecode;
 426                         }
 427                 }
 428         }
 429
 430         ntdb->file->lockrecs[ntdb->file->num_lockrecs].owner = ntdb;
 431         ntdb->file->lockrecs[ntdb->file->num_lockrecs].off = offset;
 432         ntdb->file->lockrecs[ntdb->file->num_lockrecs].count = 1;
 433         ntdb->file->lockrecs[ntdb->file->num_lockrecs].ltype = ltype;
 434         ntdb->file->num_lockrecs++;
 435
 436         return NTDB_SUCCESS;
 437 }
 438
 439 static enum NTDB_ERROR ntdb_nest_unlock(struct ntdb_context *ntdb,
 440                                       ntdb_off_t off, int ltype)
 441 {
 442         struct ntdb_lock *lck;
 443         enum NTDB_ERROR ecode;
 444
 445         if (ntdb->flags & NTDB_NOLOCK)
 446                 return NTDB_SUCCESS;
 447
 448         lck = find_nestlock(ntdb, off, ntdb);
 449         if ((lck == NULL) || (lck->count == 0)) {
 450                 return ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_ERROR,
 451                                   "ntdb_nest_unlock: no lock for %zu",
 452                                   (size_t)off);
 453         }
 454
 455         if (lck->count > 1) {
 456                 lck->count--;
 457                 return NTDB_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 = ntdb_brunlock(ntdb, ltype, off, 1);
 467
 468         /*
 469          * Shrink the array by overwriting the element just unlocked with the
 470          * last array element.
 471          */
 472         *lck = ntdb->file->lockrecs[--ntdb->file->num_lockrecs];
 473
 474         return ecode;
 475 }
 476
 477 /*
 478   get the transaction lock
 479  */
 480 enum NTDB_ERROR ntdb_transaction_lock(struct ntdb_context *ntdb, int ltype)
 481 {
 482         return ntdb_nest_lock(ntdb, NTDB_TRANSACTION_LOCK, ltype, NTDB_LOCK_WAIT);
 483 }
 484
 485 /*
 486   release the transaction lock
 487  */
 488 void ntdb_transaction_unlock(struct ntdb_context *ntdb, int ltype)
 489 {
 490         ntdb_nest_unlock(ntdb, NTDB_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 static enum NTDB_ERROR ntdb_lock_gradual(struct ntdb_context *ntdb,
 496                                        int ltype, enum ntdb_lock_flags flags,
 497                                        ntdb_off_t off, ntdb_off_t len)
 498 {
 499         enum NTDB_ERROR ecode;
 500         enum ntdb_lock_flags nb_flags = (flags & ~NTDB_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 ntdb_brlock(ntdb, ltype, off, len, flags);
 507         }
 508
 509         /* First we try non-blocking. */
 510         ecode = ntdb_brlock(ntdb, ltype, off, len, nb_flags);
 511         if (ecode != NTDB_ERR_LOCK) {
 512                 return ecode;
 513         }
 514
 515         /* Try locking first half, then second. */
 516         ecode = ntdb_lock_gradual(ntdb, ltype, flags, off, len / 2);
 517         if (ecode != NTDB_SUCCESS)
 518                 return ecode;
 519
 520         ecode = ntdb_lock_gradual(ntdb, ltype, flags,
 521                                  off + len / 2, len - len / 2);
 522         if (ecode != NTDB_SUCCESS) {
 523                 ntdb_brunlock(ntdb, 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 NTDB_ERROR ntdb_allrecord_lock(struct ntdb_context *ntdb, int ltype,
 531                                   enum ntdb_lock_flags flags, bool upgradable)
 532 {
 533         enum NTDB_ERROR ecode;
 534         ntdb_bool_err berr;
 535
 536         if (ntdb->flags & NTDB_NOLOCK) {
 537                 return NTDB_SUCCESS;
 538         }
 539
 540         if (!check_lock_pid(ntdb, "ntdb_allrecord_lock", true)) {
 541                 return NTDB_ERR_LOCK;
 542         }
 543
 544         if (ntdb->file->allrecord_lock.count) {
 545                 if (ntdb->file->allrecord_lock.owner != ntdb) {
 546                         return owner_conflict(ntdb, "ntdb_allrecord_lock");
 547                 }
 548
 549                 if (ltype == F_RDLCK
 550                     || ntdb->file->allrecord_lock.ltype == F_WRLCK) {
 551                         ntdb->file->allrecord_lock.count++;
 552                         return NTDB_SUCCESS;
 553                 }
 554
 555                 /* a global lock of a different type exists */
 556                 return ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_USE_ERROR,
 557                                   "ntdb_allrecord_lock: already have %s lock",
 558                                   ntdb->file->allrecord_lock.ltype == F_RDLCK
 559                                   ? "read" : "write");
 560         }
 561
 562         if (ntdb_has_hash_locks(ntdb)) {
 563                 /* can't combine global and chain locks */
 564                 return ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_USE_ERROR,
 565                                   "ntdb_allrecord_lock:"
 566                                   " already have chain lock");
 567         }
 568
 569         if (upgradable && ltype != F_RDLCK) {
 570                 /* ntdb error: you can't upgrade a write lock! */
 571                 return ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_ERROR,
 572                                   "ntdb_allrecord_lock:"
 573                                   " can't upgrade a write lock");
 574         }
 575
 576         ntdb->stats.locks++;
 577 again:
 578         /* Lock hashes, gradually. */
 579         ecode = ntdb_lock_gradual(ntdb, ltype, flags, NTDB_HASH_LOCK_START,
 580                                   1 << ntdb->hash_bits);
 581         if (ecode != NTDB_SUCCESS)
 582                 return ecode;
 583
 584         /* Lock free tables: there to end of file. */
 585         ecode = ntdb_brlock(ntdb, ltype,
 586                             NTDB_HASH_LOCK_START + (1 << ntdb->hash_bits),
 587                             0, flags);
 588         if (ecode != NTDB_SUCCESS) {
 589                 ntdb_brunlock(ntdb, ltype, NTDB_HASH_LOCK_START,
 590                               1 << ntdb->hash_bits);
 591                 return ecode;
 592         }
 593
 594         ntdb->file->allrecord_lock.owner = ntdb;
 595         ntdb->file->allrecord_lock.count = 1;
 596         /* If it's upgradable, it's actually exclusive so we can treat
 597          * it as a write lock. */
 598         ntdb->file->allrecord_lock.ltype = upgradable ? F_WRLCK : ltype;
 599         ntdb->file->allrecord_lock.off = upgradable;
 600
 601         /* Now check for needing recovery. */
 602         if (flags & NTDB_LOCK_NOCHECK)
 603                 return NTDB_SUCCESS;
 604
 605         berr = ntdb_needs_recovery(ntdb);
 606         if (likely(berr == false))
 607                 return NTDB_SUCCESS;
 608
 609         ntdb_allrecord_unlock(ntdb, ltype);
 610         if (berr < 0)
 611                 return NTDB_OFF_TO_ERR(berr);
 612         ecode = ntdb_lock_and_recover(ntdb);
 613         if (ecode != NTDB_SUCCESS) {
 614                 return ecode;
 615         }
 616         goto again;
 617 }
 618
 619 enum NTDB_ERROR ntdb_lock_open(struct ntdb_context *ntdb,
 620                              int ltype, enum ntdb_lock_flags flags)
 621 {
 622         return ntdb_nest_lock(ntdb, NTDB_OPEN_LOCK, ltype, flags);
 623 }
 624
 625 void ntdb_unlock_open(struct ntdb_context *ntdb, int ltype)
 626 {
 627         ntdb_nest_unlock(ntdb, NTDB_OPEN_LOCK, ltype);
 628 }
 629
 630 bool ntdb_has_open_lock(struct ntdb_context *ntdb)
 631 {
 632         return !(ntdb->flags & NTDB_NOLOCK)
 633                 && find_nestlock(ntdb, NTDB_OPEN_LOCK, ntdb) != NULL;
 634 }
 635
 636 enum NTDB_ERROR ntdb_lock_expand(struct ntdb_context *ntdb, int ltype)
 637 {
 638         /* Lock doesn't protect data, so don't check (we recurse if we do!) */
 639         return ntdb_nest_lock(ntdb, NTDB_EXPANSION_LOCK, ltype,
 640                              NTDB_LOCK_WAIT | NTDB_LOCK_NOCHECK);
 641 }
 642
 643 void ntdb_unlock_expand(struct ntdb_context *ntdb, int ltype)
 644 {
 645         ntdb_nest_unlock(ntdb, NTDB_EXPANSION_LOCK, ltype);
 646 }
 647
 648 /* unlock entire db */
 649 void ntdb_allrecord_unlock(struct ntdb_context *ntdb, int ltype)
 650 {
 651         if (ntdb->flags & NTDB_NOLOCK)
 652                 return;
 653
 654         if (ntdb->file->allrecord_lock.count == 0) {
 655                 ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_USE_ERROR,
 656                            "ntdb_allrecord_unlock: not locked!");
 657                 return;
 658         }
 659
 660         if (ntdb->file->allrecord_lock.owner != ntdb) {
 661                 ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_USE_ERROR,
 662                            "ntdb_allrecord_unlock: not locked by us!");
 663                 return;
 664         }
 665
 666         /* Upgradable locks are marked as write locks. */
 667         if (ntdb->file->allrecord_lock.ltype != ltype
 668             && (!ntdb->file->allrecord_lock.off || ltype != F_RDLCK)) {
 669                 ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_ERROR,
 670                            "ntdb_allrecord_unlock: have %s lock",
 671                            ntdb->file->allrecord_lock.ltype == F_RDLCK
 672                            ? "read" : "write");
 673                 return;
 674         }
 675
 676         if (ntdb->file->allrecord_lock.count > 1) {
 677                 ntdb->file->allrecord_lock.count--;
 678                 return;
 679         }
 680
 681         ntdb->file->allrecord_lock.count = 0;
 682         ntdb->file->allrecord_lock.ltype = 0;
 683
 684         ntdb_brunlock(ntdb, ltype, NTDB_HASH_LOCK_START, 0);
 685 }
 686
 687 bool ntdb_has_expansion_lock(struct ntdb_context *ntdb)
 688 {
 689         return find_nestlock(ntdb, NTDB_EXPANSION_LOCK, ntdb) != NULL;
 690 }
 691
 692 bool ntdb_has_hash_locks(struct ntdb_context *ntdb)
 693 {
 694         unsigned int i;
 695
 696         for (i=0; i<ntdb->file->num_lockrecs; i++) {
 697                 if (ntdb->file->lockrecs[i].off >= NTDB_HASH_LOCK_START
 698                     && ntdb->file->lockrecs[i].off < (NTDB_HASH_LOCK_START
 699                                                       + (1 << ntdb->hash_bits)))
 700                         return true;
 701         }
 702         return false;
 703 }
 704
 705 static bool ntdb_has_free_lock(struct ntdb_context *ntdb)
 706 {
 707         unsigned int i;
 708
 709         if (ntdb->flags & NTDB_NOLOCK)
 710                 return false;
 711
 712         for (i=0; i<ntdb->file->num_lockrecs; i++) {
 713                 if (ntdb->file->lockrecs[i].off
 714                     > NTDB_HASH_LOCK_START + (1 << ntdb->hash_bits))
 715                         return true;
 716         }
 717         return false;
 718 }
 719
 720 enum NTDB_ERROR ntdb_lock_hash(struct ntdb_context *ntdb,
 721                                unsigned int h,
 722                                int ltype)
 723 {
 724         unsigned l = NTDB_HASH_LOCK_START + h;
 725
 726         assert(h < (1 << ntdb->hash_bits));
 727
 728         /* a allrecord lock allows us to avoid per chain locks */
 729         if (ntdb->file->allrecord_lock.count) {
 730                 if (!check_lock_pid(ntdb, "ntdb_lock_hashes", true))
 731                         return NTDB_ERR_LOCK;
 732
 733                 if (ntdb->file->allrecord_lock.owner != ntdb)
 734                         return owner_conflict(ntdb, "ntdb_lock_hashes");
 735                 if (ltype == ntdb->file->allrecord_lock.ltype
 736                     || ltype == F_RDLCK) {
 737                         return NTDB_SUCCESS;
 738                 }
 739
 740                 return ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_USE_ERROR,
 741                                   "ntdb_lock_hashes:"
 742                                   " already have %s allrecordlock",
 743                                   ntdb->file->allrecord_lock.ltype == F_RDLCK
 744                                   ? "read" : "write");
 745         }
 746
 747         if (ntdb_has_free_lock(ntdb)) {
 748                 return ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_ERROR,
 749                                   "ntdb_lock_hashes: already have free lock");
 750         }
 751
 752         if (ntdb_has_expansion_lock(ntdb)) {
 753                 return ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_ERROR,
 754                                   "ntdb_lock_hashes:"
 755                                   " already have expansion lock");
 756         }
 757
 758         return ntdb_nest_lock(ntdb, l, ltype, NTDB_LOCK_WAIT);
 759 }
 760
 761 enum NTDB_ERROR ntdb_unlock_hash(struct ntdb_context *ntdb,
 762                                  unsigned int h, int ltype)
 763 {
 764         unsigned l = NTDB_HASH_LOCK_START + (h & ((1 << ntdb->hash_bits)-1));
 765
 766         if (ntdb->flags & NTDB_NOLOCK)
 767                 return 0;
 768
 769         /* a allrecord lock allows us to avoid per chain locks */
 770         if (ntdb->file->allrecord_lock.count) {
 771                 if (ntdb->file->allrecord_lock.ltype == F_RDLCK
 772                     && ltype == F_WRLCK) {
 773                         return ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_ERROR,
 774                                           "ntdb_unlock_hashes RO allrecord!");
 775                 }
 776                 if (ntdb->file->allrecord_lock.owner != ntdb) {
 777                         return ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_USE_ERROR,
 778                                           "ntdb_unlock_hashes:"
 779                                           " not locked by us!");
 780                 }
 781                 return NTDB_SUCCESS;
 782         }
 783
 784         return ntdb_nest_unlock(ntdb, l, ltype);
 785 }
 786
 787 /* Hash locks use NTDB_HASH_LOCK_START + <number of hash entries>..
 788  * Then we begin; bucket offsets are sizeof(ntdb_len_t) apart, so we divide.
 789  * The result is that on 32 bit systems we don't use lock values > 2^31 on
 790  * files that are less than 4GB.
 791  */
 792 static ntdb_off_t free_lock_off(const struct ntdb_context *ntdb,
 793                                 ntdb_off_t b_off)
 794 {
 795         return NTDB_HASH_LOCK_START + (1 << ntdb->hash_bits)
 796                 + b_off / sizeof(ntdb_off_t);
 797 }
 798
 799 enum NTDB_ERROR ntdb_lock_free_bucket(struct ntdb_context *ntdb, ntdb_off_t b_off,
 800                                     enum ntdb_lock_flags waitflag)
 801 {
 802         assert(b_off >= sizeof(struct ntdb_header));
 803
 804         if (ntdb->flags & NTDB_NOLOCK)
 805                 return 0;
 806
 807         /* a allrecord lock allows us to avoid per chain locks */
 808         if (ntdb->file->allrecord_lock.count) {
 809                 if (!check_lock_pid(ntdb, "ntdb_lock_free_bucket", true))
 810                         return NTDB_ERR_LOCK;
 811
 812                 if (ntdb->file->allrecord_lock.owner != ntdb) {
 813                         return owner_conflict(ntdb, "ntdb_lock_free_bucket");
 814                 }
 815
 816                 if (ntdb->file->allrecord_lock.ltype == F_WRLCK)
 817                         return 0;
 818                 return ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_ERROR,
 819                                   "ntdb_lock_free_bucket with"
 820                                   " read-only allrecordlock!");
 821         }
 822
 823 #if 0 /* FIXME */
 824         if (ntdb_has_expansion_lock(ntdb)) {
 825                 return ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_ERROR,
 826                                   "ntdb_lock_free_bucket:"
 827                                   " already have expansion lock");
 828         }
 829 #endif
 830
 831         return ntdb_nest_lock(ntdb, free_lock_off(ntdb, b_off), F_WRLCK,
 832                               waitflag);
 833 }
 834
 835 void ntdb_unlock_free_bucket(struct ntdb_context *ntdb, ntdb_off_t b_off)
 836 {
 837         if (ntdb->file->allrecord_lock.count)
 838                 return;
 839
 840         ntdb_nest_unlock(ntdb, free_lock_off(ntdb, b_off), F_WRLCK);
 841 }
 842
 843 _PUBLIC_ enum NTDB_ERROR ntdb_lockall(struct ntdb_context *ntdb)
 844 {
 845         return ntdb_allrecord_lock(ntdb, F_WRLCK, NTDB_LOCK_WAIT, false);
 846 }
 847
 848 _PUBLIC_ void ntdb_unlockall(struct ntdb_context *ntdb)
 849 {
 850         ntdb_allrecord_unlock(ntdb, F_WRLCK);
 851 }
 852
 853 _PUBLIC_ enum NTDB_ERROR ntdb_lockall_read(struct ntdb_context *ntdb)
 854 {
 855         return ntdb_allrecord_lock(ntdb, F_RDLCK, NTDB_LOCK_WAIT, false);
 856 }
 857
 858 _PUBLIC_ void ntdb_unlockall_read(struct ntdb_context *ntdb)
 859 {
 860         ntdb_allrecord_unlock(ntdb, F_RDLCK);
 861 }
 862
 863 void ntdb_lock_cleanup(struct ntdb_context *ntdb)
 864 {
 865         unsigned int i;
 866
 867         /* We don't want to warn: they're allowed to close ntdb after fork. */
 868         if (!check_lock_pid(ntdb, "ntdb_close", false))
 869                 return;
 870
 871         while (ntdb->file->allrecord_lock.count
 872                && ntdb->file->allrecord_lock.owner == ntdb) {
 873                 ntdb_allrecord_unlock(ntdb, ntdb->file->allrecord_lock.ltype);
 874         }
 875
 876         for (i=0; i<ntdb->file->num_lockrecs; i++) {
 877                 if (ntdb->file->lockrecs[i].owner == ntdb) {
 878                         ntdb_nest_unlock(ntdb,
 879                                         ntdb->file->lockrecs[i].off,
 880                                         ntdb->file->lockrecs[i].ltype);
 881                         i--;
 882                 }
 883         }
 884 }