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 static int fcntl_lock(struct tdb_context *tdb,
  33                       int rw, off_t off, off_t len, bool waitflag)
  34 {
  35         struct flock fl;
  36
  37         fl.l_type = rw;
  38         fl.l_whence = SEEK_SET;
  39         fl.l_start = off;
  40         fl.l_len = len;
  41         fl.l_pid = 0;
  42
  43         if (waitflag)
  44                 return fcntl(tdb->fd, F_SETLKW, &fl);
  45         else
  46                 return fcntl(tdb->fd, F_SETLK, &fl);
  47 }
  48
  49 static int fcntl_unlock(struct tdb_context *tdb, int rw, off_t off, off_t len)
  50 {
  51         struct flock fl;
  52 #if 0 /* Check they matched up locks and unlocks correctly. */
  53         char line[80];
  54         FILE *locks;
  55         bool found = false;
  56
  57         locks = fopen("/proc/locks", "r");
  58
  59         while (fgets(line, 80, locks)) {
  60                 char *p;
  61                 int type, start, l;
  62
  63                 /* eg. 1: FLOCK  ADVISORY  WRITE 2440 08:01:2180826 0 EOF */
  64                 p = strchr(line, ':') + 1;
  65                 if (strncmp(p, " POSIX  ADVISORY  ", strlen(" POSIX  ADVISORY  ")))
  66                         continue;
  67                 p += strlen(" FLOCK  ADVISORY  ");
  68                 if (strncmp(p, "READ  ", strlen("READ  ")) == 0)
  69                         type = F_RDLCK;
  70                 else if (strncmp(p, "WRITE ", strlen("WRITE ")) == 0)
  71                         type = F_WRLCK;
  72                 else
  73                         abort();
  74                 p += 6;
  75                 if (atoi(p) != getpid())
  76                         continue;
  77                 p = strchr(strchr(p, ' ') + 1, ' ') + 1;
  78                 start = atoi(p);
  79                 p = strchr(p, ' ') + 1;
  80                 if (strncmp(p, "EOF", 3) == 0)
  81                         l = 0;
  82                 else
  83                         l = atoi(p) - start + 1;
  84
  85                 if (off == start) {
  86                         if (len != l) {
  87                                 fprintf(stderr, "Len %u should be %u: %s",
  88                                         (int)len, l, line);
  89                                 abort();
  90                         }
  91                         if (type != rw) {
  92                                 fprintf(stderr, "Type %s wrong: %s",
  93                                         rw == F_RDLCK ? "READ" : "WRITE", line);
  94                                 abort();
  95                         }
  96                         found = true;
  97                         break;
  98                 }
  99         }
 100
 101         if (!found) {
 102                 fprintf(stderr, "Unlock on %u@%u not found!\n",
 103                         (int)off, (int)len);
 104                 abort();
 105         }
 106
 107         fclose(locks);
 108 #endif
 109
 110         fl.l_type = F_UNLCK;
 111         fl.l_whence = SEEK_SET;
 112         fl.l_start = off;
 113         fl.l_len = len;
 114         fl.l_pid = 0;
 115
 116         return fcntl(tdb->fd, F_SETLKW, &fl);
 117 }
 118
 119 /* a byte range locking function - return 0 on success
 120    this functions locks/unlocks 1 byte at the specified offset.
 121
 122    note that a len of zero means lock to end of file
 123 */
 124 static int tdb_brlock(struct tdb_context *tdb,
 125                       int rw_type, tdb_off_t offset, tdb_off_t len,
 126                       enum tdb_lock_flags flags)
 127 {
 128         int ret;
 129
 130         if (tdb->flags & TDB_NOLOCK) {
 131                 return 0;
 132         }
 133
 134         if (rw_type == F_WRLCK && tdb->read_only) {
 135                 tdb->ecode = TDB_ERR_RDONLY;
 136                 return -1;
 137         }
 138
 139         /* A 32 bit system cannot open a 64-bit file, but it could have
 140          * expanded since then: check here. */
 141         if ((size_t)(offset + len) != offset + len) {
 142                 tdb->ecode = TDB_ERR_IO;
 143                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 144                          "tdb_brlock: lock on giant offset %llu\n",
 145                          (long long)(offset + len));
 146                 return -1;
 147         }
 148
 149         do {
 150                 ret = fcntl_lock(tdb, rw_type, offset, len,
 151                                  flags & TDB_LOCK_WAIT);
 152         } while (ret == -1 && errno == EINTR);
 153
 154         if (ret == -1) {
 155                 tdb->ecode = TDB_ERR_LOCK;
 156                 /* Generic lock error. errno set by fcntl.
 157                  * EAGAIN is an expected return from non-blocking
 158                  * locks. */
 159                 if (!(flags & TDB_LOCK_PROBE) && errno != EAGAIN) {
 160                         tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 161                                  "tdb_brlock failed (fd=%d) at"
 162                                  " offset %llu rw_type=%d flags=%d len=%llu\n",
 163                                  tdb->fd, (long long)offset, rw_type,
 164                                  flags, (long long)len);
 165                 }
 166                 return -1;
 167         }
 168         return 0;
 169 }
 170
 171 static int tdb_brunlock(struct tdb_context *tdb,
 172                         int rw_type, tdb_off_t offset, size_t len)
 173 {
 174         int ret;
 175
 176         if (tdb->flags & TDB_NOLOCK) {
 177                 return 0;
 178         }
 179
 180         do {
 181                 ret = fcntl_unlock(tdb, rw_type, offset, len);
 182         } while (ret == -1 && errno == EINTR);
 183
 184         if (ret == -1) {
 185                 tdb->log(tdb, TDB_DEBUG_TRACE, tdb->log_priv,
 186                          "tdb_brunlock failed (fd=%d) at offset %llu"
 187                          " rw_type=%d len=%llu\n",
 188                          tdb->fd, (long long)offset, rw_type, (long long)len);
 189         }
 190         return ret;
 191 }
 192
 193 #if 0
 194 /*
 195   upgrade a read lock to a write lock. This needs to be handled in a
 196   special way as some OSes (such as solaris) have too conservative
 197   deadlock detection and claim a deadlock when progress can be
 198   made. For those OSes we may loop for a while.
 199 */
 200 int tdb_allrecord_upgrade(struct tdb_context *tdb)
 201 {
 202         int count = 1000;
 203
 204         if (tdb->allrecord_lock.count != 1) {
 205                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 206                          "tdb_allrecord_upgrade failed: count %u too high\n",
 207                          tdb->allrecord_lock.count);
 208                 return -1;
 209         }
 210
 211         if (tdb->allrecord_lock.off != 1) {
 212                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 213                          "tdb_allrecord_upgrade failed: already upgraded?\n");
 214                 return -1;
 215         }
 216
 217         while (count--) {
 218                 struct timeval tv;
 219                 if (tdb_brlock(tdb, F_WRLCK,
 220                                TDB_HASH_LOCK_START
 221                                + (1ULL << tdb->header.v.hash_bits), 0,
 222                                TDB_LOCK_WAIT|TDB_LOCK_PROBE) == 0) {
 223                         tdb->allrecord_lock.ltype = F_WRLCK;
 224                         tdb->allrecord_lock.off = 0;
 225                         return 0;
 226                 }
 227                 if (errno != EDEADLK) {
 228                         break;
 229                 }
 230                 /* sleep for as short a time as we can - more portable than usleep() */
 231                 tv.tv_sec = 0;
 232                 tv.tv_usec = 1;
 233                 select(0, NULL, NULL, NULL, &tv);
 234         }
 235         tdb->log(tdb, TDB_DEBUG_WARNING, tdb->log_priv,
 236                  "tdb_allrecord_upgrade failed\n");
 237         return -1;
 238 }
 239 #endif
 240
 241 static struct tdb_lock_type *find_nestlock(struct tdb_context *tdb,
 242                                            tdb_off_t offset)
 243 {
 244         unsigned int i;
 245
 246         for (i=0; i<tdb->num_lockrecs; i++) {
 247                 if (tdb->lockrecs[i].off == offset) {
 248                         return &tdb->lockrecs[i];
 249                 }
 250         }
 251         return NULL;
 252 }
 253
 254 /* lock an offset in the database. */
 255 static int tdb_nest_lock(struct tdb_context *tdb, tdb_off_t offset, int ltype,
 256                          enum tdb_lock_flags flags)
 257 {
 258         struct tdb_lock_type *new_lck;
 259
 260         if (offset > TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE + tdb->map_size / 8) {
 261                 tdb->ecode = TDB_ERR_LOCK;
 262                 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 263                          "tdb_nest_lock: invalid offset %llu ltype=%d\n",
 264                          (long long)offset, ltype);
 265                 return -1;
 266         }
 267
 268         if (tdb->flags & TDB_NOLOCK)
 269                 return 0;
 270
 271         new_lck = find_nestlock(tdb, offset);
 272         if (new_lck) {
 273                 if (new_lck->ltype == F_RDLCK && ltype == F_WRLCK) {
 274                         tdb->ecode = TDB_ERR_LOCK;
 275                         tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 276                                  "tdb_nest_lock: offset %llu has read lock\n",
 277                                  (long long)offset);
 278                         return -1;
 279                 }
 280                 /* Just increment the struct, posix locks don't stack. */
 281                 new_lck->count++;
 282                 return 0;
 283         }
 284
 285         if (tdb->num_lockrecs
 286             && offset >= TDB_HASH_LOCK_START
 287             && offset < TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE) {
 288                 tdb->ecode = TDB_ERR_LOCK;
 289                 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 290                          "tdb_nest_lock: already have a hash lock?\n");
 291                 return -1;
 292         }
 293
 294         new_lck = (struct tdb_lock_type *)realloc(
 295                 tdb->lockrecs,
 296                 sizeof(*tdb->lockrecs) * (tdb->num_lockrecs+1));
 297         if (new_lck == NULL) {
 298                 tdb->ecode = TDB_ERR_OOM;
 299                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 300                          "tdb_nest_lock: unable to allocate %llu lock struct",
 301                          (long long)(tdb->num_lockrecs + 1));
 302                 errno = ENOMEM;
 303                 return -1;
 304         }
 305         tdb->lockrecs = new_lck;
 306
 307         /* Since fcntl locks don't nest, we do a lock for the first one,
 308            and simply bump the count for future ones */
 309         if (tdb_brlock(tdb, ltype, offset, 1, flags)) {
 310                 return -1;
 311         }
 312
 313         tdb->lockrecs[tdb->num_lockrecs].off = offset;
 314         tdb->lockrecs[tdb->num_lockrecs].count = 1;
 315         tdb->lockrecs[tdb->num_lockrecs].ltype = ltype;
 316         tdb->num_lockrecs++;
 317
 318         return 0;
 319 }
 320
 321 static int tdb_lock_and_recover(struct tdb_context *tdb)
 322 {
 323 #if 0 /* FIXME */
 324
 325         int ret;
 326
 327         /* We need to match locking order in transaction commit. */
 328         if (tdb_brlock(tdb, F_WRLCK, FREELIST_TOP, 0, TDB_LOCK_WAIT)) {
 329                 return -1;
 330         }
 331
 332         if (tdb_brlock(tdb, F_WRLCK, OPEN_LOCK, 1, TDB_LOCK_WAIT)) {
 333                 tdb_brunlock(tdb, F_WRLCK, FREELIST_TOP, 0);
 334                 return -1;
 335         }
 336
 337         ret = tdb_transaction_recover(tdb);
 338
 339         tdb_brunlock(tdb, F_WRLCK, OPEN_LOCK, 1);
 340         tdb_brunlock(tdb, F_WRLCK, FREELIST_TOP, 0);
 341
 342         return ret;
 343 #else
 344         abort();
 345         return -1;
 346 #endif
 347 }
 348
 349 static bool tdb_needs_recovery(struct tdb_context *tdb)
 350 {
 351         /* FIXME */
 352         return false;
 353 }
 354
 355 static int tdb_nest_unlock(struct tdb_context *tdb, tdb_off_t off, int ltype)
 356 {
 357         int ret = -1;
 358         struct tdb_lock_type *lck;
 359
 360         if (tdb->flags & TDB_NOLOCK)
 361                 return 0;
 362
 363         lck = find_nestlock(tdb, off);
 364         if ((lck == NULL) || (lck->count == 0)) {
 365                 tdb->ecode = TDB_ERR_LOCK;
 366                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 367                          "tdb_nest_unlock: no lock for %llu\n", (long long)off);
 368                 return -1;
 369         }
 370
 371         if (lck->count > 1) {
 372                 lck->count--;
 373                 return 0;
 374         }
 375
 376         /*
 377          * This lock has count==1 left, so we need to unlock it in the
 378          * kernel. We don't bother with decrementing the in-memory array
 379          * element, we're about to overwrite it with the last array element
 380          * anyway.
 381          */
 382         ret = tdb_brunlock(tdb, ltype, off, 1);
 383
 384         /*
 385          * Shrink the array by overwriting the element just unlocked with the
 386          * last array element.
 387          */
 388         *lck = tdb->lockrecs[--tdb->num_lockrecs];
 389
 390         return ret;
 391 }
 392
 393 #if 0
 394 /*
 395   get the transaction lock
 396  */
 397 int tdb_transaction_lock(struct tdb_context *tdb, int ltype,
 398                          enum tdb_lock_flags lockflags)
 399 {
 400         return tdb_nest_lock(tdb, TRANSACTION_LOCK, ltype, lockflags);
 401 }
 402
 403 /*
 404   release the transaction lock
 405  */
 406 int tdb_transaction_unlock(struct tdb_context *tdb, int ltype)
 407 {
 408         return tdb_nest_unlock(tdb, TRANSACTION_LOCK, ltype, false);
 409 }
 410 #endif
 411
 412 /* We only need to lock individual bytes, but Linux merges consecutive locks
 413  * so we lock in contiguous ranges. */
 414 static int tdb_lock_gradual(struct tdb_context *tdb,
 415                             int ltype, enum tdb_lock_flags flags,
 416                             tdb_off_t off, tdb_off_t len)
 417 {
 418         int ret;
 419         enum tdb_lock_flags nb_flags = (flags & ~TDB_LOCK_WAIT);
 420
 421         if (len <= 1) {
 422                 /* 0 would mean to end-of-file... */
 423                 assert(len != 0);
 424                 /* Single hash.  Just do blocking lock. */
 425                 return tdb_brlock(tdb, ltype, off, len, flags);
 426         }
 427
 428         /* First we try non-blocking. */
 429         ret = tdb_brlock(tdb, ltype, off, len, nb_flags);
 430         if (ret == 0) {
 431                 return 0;
 432         }
 433
 434         /* Try locking first half, then second. */
 435         ret = tdb_lock_gradual(tdb, ltype, flags, off, len / 2);
 436         if (ret == -1)
 437                 return -1;
 438
 439         ret = tdb_lock_gradual(tdb, ltype, flags,
 440                                     off + len / 2, len - len / 2);
 441         if (ret == -1) {
 442                 tdb_brunlock(tdb, ltype, off, len / 2);
 443                 return -1;
 444         }
 445         return 0;
 446 }
 447
 448 /* lock/unlock entire database.  It can only be upgradable if you have some
 449  * other way of guaranteeing exclusivity (ie. transaction write lock). */
 450 int tdb_allrecord_lock(struct tdb_context *tdb, int ltype,
 451                        enum tdb_lock_flags flags, bool upgradable)
 452 {
 453         /* FIXME: There are no locks on read-only dbs */
 454         if (tdb->read_only) {
 455                 tdb->ecode = TDB_ERR_LOCK;
 456                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 457                          "tdb_allrecord_lock: read-only\n");
 458                 return -1;
 459         }
 460
 461         if (tdb->allrecord_lock.count
 462             && (ltype == F_RDLCK || tdb->allrecord_lock.ltype == F_WRLCK)) {
 463                 tdb->allrecord_lock.count++;
 464                 return 0;
 465         }
 466
 467         if (tdb->allrecord_lock.count) {
 468                 /* a global lock of a different type exists */
 469                 tdb->ecode = TDB_ERR_LOCK;
 470                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 471                          "tdb_allrecord_lock: already have %s lock\n",
 472                          tdb->allrecord_lock.ltype == F_RDLCK
 473                          ? "read" : "write");
 474                 return -1;
 475         }
 476
 477         if (tdb_has_locks(tdb)) {
 478                 /* can't combine global and chain locks */
 479                 tdb->ecode = TDB_ERR_LOCK;
 480                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 481                          "tdb_allrecord_lock: already have chain lock\n");
 482                 return -1;
 483         }
 484
 485         if (upgradable && ltype != F_RDLCK) {
 486                 /* tdb error: you can't upgrade a write lock! */
 487                 tdb->ecode = TDB_ERR_LOCK;
 488                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 489                          "tdb_allrecord_lock: can't upgrade a write lock\n");
 490                 return -1;
 491         }
 492
 493 again:
 494         /* Lock hashes, gradually. */
 495         if (tdb_lock_gradual(tdb, ltype, flags, TDB_HASH_LOCK_START,
 496                              TDB_HASH_LOCK_RANGE)) {
 497                 if (!(flags & TDB_LOCK_PROBE)) {
 498                         tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 499                                  "tdb_allrecord_lock hashes failed (%s)\n",
 500                                  strerror(errno));
 501                 }
 502                 return -1;
 503         }
 504
 505         /* Lock free lists: there to end of file. */
 506         if (tdb_brlock(tdb, ltype, TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE,
 507                        0, flags)) {
 508                 if (!(flags & TDB_LOCK_PROBE)) {
 509                         tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 510                                  "tdb_allrecord_lock freelist failed (%s)\n",
 511                                  strerror(errno));
 512                 }
 513                 tdb_brunlock(tdb, ltype, TDB_HASH_LOCK_START,
 514                              TDB_HASH_LOCK_RANGE);
 515                 return -1;
 516         }
 517
 518         tdb->allrecord_lock.count = 1;
 519         /* If it's upgradable, it's actually exclusive so we can treat
 520          * it as a write lock. */
 521         tdb->allrecord_lock.ltype = upgradable ? F_WRLCK : ltype;
 522         tdb->allrecord_lock.off = upgradable;
 523
 524         /* Now check for needing recovery. */
 525         if (unlikely(tdb_needs_recovery(tdb))) {
 526                 tdb_allrecord_unlock(tdb, ltype);
 527                 if (tdb_lock_and_recover(tdb) == -1) {
 528                         return -1;
 529                 }
 530                 goto again;
 531         }
 532
 533         return 0;
 534 }
 535
 536 int tdb_lock_open(struct tdb_context *tdb)
 537 {
 538         return tdb_nest_lock(tdb, TDB_OPEN_LOCK, F_WRLCK, TDB_LOCK_WAIT);
 539 }
 540
 541 void tdb_unlock_open(struct tdb_context *tdb)
 542 {
 543         tdb_nest_unlock(tdb, TDB_OPEN_LOCK, F_WRLCK);
 544 }
 545
 546 int tdb_lock_expand(struct tdb_context *tdb, int ltype)
 547 {
 548         return tdb_nest_lock(tdb, TDB_EXPANSION_LOCK, ltype, TDB_LOCK_WAIT);
 549 }
 550
 551 void tdb_unlock_expand(struct tdb_context *tdb, int ltype)
 552 {
 553         tdb_nest_unlock(tdb, TDB_EXPANSION_LOCK, ltype);
 554 }
 555
 556 /* unlock entire db */
 557 int tdb_allrecord_unlock(struct tdb_context *tdb, int ltype)
 558 {
 559         /* FIXME: There are no locks on read-only dbs */
 560         if (tdb->read_only) {
 561                 tdb->ecode = TDB_ERR_LOCK;
 562                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 563                          "tdb_allrecord_unlock: read-only\n");
 564                 return -1;
 565         }
 566
 567         if (tdb->allrecord_lock.count == 0) {
 568                 tdb->ecode = TDB_ERR_LOCK;
 569                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 570                          "tdb_allrecord_unlock: not locked!\n");
 571                 return -1;
 572         }
 573
 574         /* Upgradable locks are marked as write locks. */
 575         if (tdb->allrecord_lock.ltype != ltype
 576             && (!tdb->allrecord_lock.off || ltype != F_RDLCK)) {
 577                 tdb->ecode = TDB_ERR_LOCK;
 578                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 579                          "tdb_allrecord_unlock: have %s lock\n",
 580                          tdb->allrecord_lock.ltype == F_RDLCK
 581                          ? "read" : "write");
 582                 return -1;
 583         }
 584
 585         if (tdb->allrecord_lock.count > 1) {
 586                 tdb->allrecord_lock.count--;
 587                 return 0;
 588         }
 589
 590         tdb->allrecord_lock.count = 0;
 591         tdb->allrecord_lock.ltype = 0;
 592
 593         return tdb_brunlock(tdb, ltype, TDB_HASH_LOCK_START, 0);
 594 }
 595
 596 bool tdb_has_expansion_lock(struct tdb_context *tdb)
 597 {
 598         return find_nestlock(tdb, TDB_EXPANSION_LOCK) != NULL;
 599 }
 600
 601 bool tdb_has_locks(struct tdb_context *tdb)
 602 {
 603         return tdb->allrecord_lock.count || tdb->num_lockrecs;
 604 }
 605
 606 #if 0
 607 /* lock entire database with write lock */
 608 int tdb_lockall(struct tdb_context *tdb)
 609 {
 610         tdb_trace(tdb, "tdb_lockall");
 611         return tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT, false);
 612 }
 613
 614 /* lock entire database with write lock - nonblocking varient */
 615 int tdb_lockall_nonblock(struct tdb_context *tdb)
 616 {
 617         int ret = tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_NOWAIT, false);
 618         tdb_trace_ret(tdb, "tdb_lockall_nonblock", ret);
 619         return ret;
 620 }
 621
 622 /* unlock entire database with write lock */
 623 int tdb_unlockall(struct tdb_context *tdb)
 624 {
 625         tdb_trace(tdb, "tdb_unlockall");
 626         return tdb_allrecord_unlock(tdb, F_WRLCK);
 627 }
 628
 629 /* lock entire database with read lock */
 630 int tdb_lockall_read(struct tdb_context *tdb)
 631 {
 632         tdb_trace(tdb, "tdb_lockall_read");
 633         return tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, false);
 634 }
 635
 636 /* lock entire database with read lock - nonblock varient */
 637 int tdb_lockall_read_nonblock(struct tdb_context *tdb)
 638 {
 639         int ret = tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_NOWAIT, false);
 640         tdb_trace_ret(tdb, "tdb_lockall_read_nonblock", ret);
 641         return ret;
 642 }
 643
 644 /* unlock entire database with read lock */
 645 int tdb_unlockall_read(struct tdb_context *tdb)
 646 {
 647         tdb_trace(tdb, "tdb_unlockall_read");
 648         return tdb_allrecord_unlock(tdb, F_RDLCK);
 649 }
 650 #endif
 651
 652 static bool tdb_has_free_lock(struct tdb_context *tdb)
 653 {
 654         unsigned int i;
 655
 656         for (i=0; i<tdb->num_lockrecs; i++) {
 657                 if (tdb->lockrecs[i].off
 658                     > TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE)
 659                         return true;
 660         }
 661         return false;
 662 }
 663
 664 int tdb_lock_hashes(struct tdb_context *tdb,
 665                     tdb_off_t hash_lock,
 666                     tdb_len_t hash_range,
 667                     int ltype, enum tdb_lock_flags waitflag)
 668 {
 669         /* FIXME: Do this properly, using hlock_range */
 670         unsigned lock = TDB_HASH_LOCK_START
 671                 + (hash_lock >> (64 - TDB_HASH_LOCK_RANGE_BITS));
 672
 673         /* a allrecord lock allows us to avoid per chain locks */
 674         if (tdb->allrecord_lock.count &&
 675             (ltype == tdb->allrecord_lock.ltype || ltype == F_RDLCK)) {
 676                 return 0;
 677         }
 678
 679         if (tdb->allrecord_lock.count) {
 680                 tdb->ecode = TDB_ERR_LOCK;
 681                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 682                          "tdb_lock_hashes: have %s allrecordlock\n",
 683                          tdb->allrecord_lock.ltype == F_RDLCK
 684                          ? "read" : "write");
 685                 return -1;
 686         }
 687
 688         if (tdb_has_free_lock(tdb)) {
 689                 tdb->ecode = TDB_ERR_LOCK;
 690                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 691                          "tdb_lock_hashes: have free lock already\n");
 692                 return -1;
 693         }
 694
 695         if (tdb_has_expansion_lock(tdb)) {
 696                 tdb->ecode = TDB_ERR_LOCK;
 697                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 698                          "tdb_lock_hashes: have expansion lock already\n");
 699                 return -1;
 700         }
 701
 702         return tdb_nest_lock(tdb, lock, ltype, waitflag);
 703 }
 704
 705 int tdb_unlock_hashes(struct tdb_context *tdb,
 706                       tdb_off_t hash_lock,
 707                       tdb_len_t hash_range, int ltype)
 708 {
 709         unsigned lock = TDB_HASH_LOCK_START
 710                 + (hash_lock >> (64 - TDB_HASH_LOCK_RANGE_BITS));
 711
 712         /* a allrecord lock allows us to avoid per chain locks */
 713         if (tdb->allrecord_lock.count) {
 714                 if (tdb->allrecord_lock.ltype == F_RDLCK
 715                     && ltype == F_WRLCK) {
 716                         tdb->ecode = TDB_ERR_LOCK;
 717                         tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 718                                  "tdb_unlock_hashes RO allrecord!\n");
 719                         return -1;
 720                 }
 721                 return 0;
 722         }
 723
 724         return tdb_nest_unlock(tdb, lock, ltype);
 725 }
 726
 727 /* Hash locks use TDB_HASH_LOCK_START + the next 30 bits.
 728  * Then we begin; bucket offsets are sizeof(tdb_len_t) apart, so we divide.
 729  * The result is that on 32 bit systems we don't use lock values > 2^31 on
 730  * files that are less than 4GB.
 731  */
 732 static tdb_off_t free_lock_off(tdb_off_t b_off)
 733 {
 734         return TDB_HASH_LOCK_START + TDB_HASH_LOCK_RANGE
 735                 + b_off / sizeof(tdb_off_t);
 736 }
 737
 738 int tdb_lock_free_bucket(struct tdb_context *tdb, tdb_off_t b_off,
 739                          enum tdb_lock_flags waitflag)
 740 {
 741         assert(b_off >= sizeof(struct tdb_header));
 742
 743         /* a allrecord lock allows us to avoid per chain locks */
 744         if (tdb->allrecord_lock.count) {
 745                 if (tdb->allrecord_lock.ltype == F_WRLCK)
 746                         return 0;
 747                 tdb->ecode = TDB_ERR_LOCK;
 748                 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 749                          "tdb_lock_free_bucket with RO allrecordlock!\n");
 750                 return -1;
 751         }
 752
 753 #if 0 /* FIXME */
 754         if (tdb_has_expansion_lock(tdb)) {
 755                 tdb->ecode = TDB_ERR_LOCK;
 756                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 757                          "tdb_lock_free_bucket: have expansion lock already\n");
 758                 return -1;
 759         }
 760 #endif
 761
 762         return tdb_nest_lock(tdb, free_lock_off(b_off), F_WRLCK, waitflag);
 763 }
 764
 765 void tdb_unlock_free_bucket(struct tdb_context *tdb, tdb_off_t b_off)
 766 {
 767         if (tdb->allrecord_lock.count)
 768                 return;
 769
 770         tdb_nest_unlock(tdb, free_lock_off(b_off), F_WRLCK);
 771 }
 772
 773 #if 0
 774 /* lock/unlock one hash chain, non-blocking. This is meant to be used
 775    to reduce contention - it cannot guarantee how many records will be
 776    locked */
 777 int tdb_chainlock_nonblock(struct tdb_context *tdb, TDB_DATA key)
 778 {
 779         return chainlock(tdb, &key, F_WRLCK, TDB_LOCK_NOWAIT,
 780                          "tdb_chainlock_nonblock");
 781 }
 782
 783 int tdb_chainlock_read(struct tdb_context *tdb, TDB_DATA key)
 784 {
 785         return chainlock(tdb, &key, F_RDLCK, TDB_LOCK_WAIT,
 786                          "tdb_chainlock_read");
 787 }
 788
 789 int tdb_chainunlock_read(struct tdb_context *tdb, TDB_DATA key)
 790 {
 791         uint64_t h = tdb_hash(tdb, key.dptr, key.dsize);
 792         tdb_trace_1rec(tdb, "tdb_chainunlock_read", key);
 793         return tdb_unlock_list(tdb, h & ((1ULL << tdb->header.v.hash_bits)-1),
 794                                F_RDLCK);
 795 }
 796
 797 /* record lock stops delete underneath */
 798 int tdb_lock_record(struct tdb_context *tdb, tdb_off_t off)
 799 {
 800         if (tdb->allrecord_lock.count) {
 801                 return 0;
 802         }
 803         return off ? tdb_brlock(tdb, F_RDLCK, off, 1, TDB_LOCK_WAIT) : 0;
 804 }
 805
 806 /*
 807   Write locks override our own fcntl readlocks, so check it here.
 808   Note this is meant to be F_SETLK, *not* F_SETLKW, as it's not
 809   an error to fail to get the lock here.
 810 */
 811 int tdb_write_lock_record(struct tdb_context *tdb, tdb_off_t off)
 812 {
 813         struct tdb_traverse_lock *i;
 814         for (i = &tdb->travlocks; i; i = i->next)
 815                 if (i->off == off)
 816                         return -1;
 817         if (tdb->allrecord_lock.count) {
 818                 if (tdb->allrecord_lock.ltype == F_WRLCK) {
 819                         return 0;
 820                 }
 821                 return -1;
 822         }
 823         return tdb_brlock(tdb, F_WRLCK, off, 1, TDB_LOCK_NOWAIT|TDB_LOCK_PROBE);
 824 }
 825
 826 int tdb_write_unlock_record(struct tdb_context *tdb, tdb_off_t off)
 827 {
 828         if (tdb->allrecord_lock.count) {
 829                 return 0;
 830         }
 831         return tdb_brunlock(tdb, F_WRLCK, off, 1);
 832 }
 833
 834 /* fcntl locks don't stack: avoid unlocking someone else's */
 835 int tdb_unlock_record(struct tdb_context *tdb, tdb_off_t off)
 836 {
 837         struct tdb_traverse_lock *i;
 838         uint32_t count = 0;
 839
 840         if (tdb->allrecord_lock.count) {
 841                 return 0;
 842         }
 843
 844         if (off == 0)
 845                 return 0;
 846         for (i = &tdb->travlocks; i; i = i->next)
 847                 if (i->off == off)
 848                         count++;
 849         return (count == 1 ? tdb_brunlock(tdb, F_RDLCK, off, 1) : 0);
 850 }
 851
 852 /* The transaction code uses this to remove all locks. */
 853 void tdb_release_transaction_locks(struct tdb_context *tdb)
 854 {
 855         unsigned int i;
 856
 857         if (tdb->allrecord_lock.count != 0) {
 858                 tdb_off_t hash_size, free_size;
 859
 860                 hash_size = (1ULL << tdb->header.v.hash_bits)
 861                         * sizeof(tdb_off_t);
 862                 free_size = tdb->header.v.free_zones
 863                         * (tdb->header.v.free_buckets + 1) * sizeof(tdb_off_t);
 864
 865                 tdb_brunlock(tdb, tdb->allrecord_lock.ltype,
 866                              tdb->header.v.hash_off, hash_size);
 867                 tdb_brunlock(tdb, tdb->allrecord_lock.ltype,
 868                              tdb->header.v.free_off, free_size);
 869                 tdb->allrecord_lock.count = 0;
 870                 tdb->allrecord_lock.ltype = 0;
 871         }
 872
 873         for (i = 0; i<tdb->num_lockrecs; i++) {
 874                 struct tdb_lock_type *lck = &tdb->lockrecs[i];
 875
 876                 tdb_brunlock(tdb, lck->ltype, lck->off, 1);
 877         }
 878         tdb->num_lockrecs = 0;
 879         SAFE_FREE(tdb->lockrecs);
 880         tdb->header_uptodate = false;
 881 }
 882 #endif
 883
 884 void tdb_lock_init(struct tdb_context *tdb)
 885 {
 886         tdb->num_lockrecs = 0;
 887         tdb->lockrecs = NULL;
 888         tdb->allrecord_lock.count = 0;
 889 }