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