git.ozlabs.org Git - ccan/blob - ccan/tdb2/tdb1_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 "tdb1_private.h"
  29
  30 /* list -1 is the alloc list, otherwise a hash chain. */
  31 static tdb1_off_t lock_offset(int list)
  32 {
  33         return TDB1_FREELIST_TOP + 4*list;
  34 }
  35
  36 /* a byte range locking function - return 0 on success
  37    this functions locks/unlocks 1 byte at the specified offset.
  38
  39    On error, errno is also set so that errors are passed back properly
  40    through tdb1_open().
  41
  42    note that a len of zero means lock to end of file
  43 */
  44 int tdb1_brlock(struct tdb1_context *tdb,
  45                int rw_type, tdb1_off_t offset, size_t len,
  46                enum tdb_lock_flags flags)
  47 {
  48         enum TDB_ERROR ecode = tdb_brlock(tdb, rw_type, offset, len, flags
  49                                           | TDB_LOCK_NOCHECK);
  50         if (ecode == TDB_SUCCESS)
  51                 return 0;
  52         tdb->last_error = ecode;
  53         return -1;
  54 }
  55
  56 int tdb1_brunlock(struct tdb1_context *tdb,
  57                  int rw_type, tdb1_off_t offset, size_t len)
  58 {
  59         enum TDB_ERROR ecode = tdb_brunlock(tdb, rw_type, offset, len);
  60         if (ecode == TDB_SUCCESS)
  61                 return 0;
  62         tdb->last_error = ecode;
  63         return -1;
  64 }
  65
  66 int tdb1_allrecord_upgrade(struct tdb1_context *tdb)
  67 {
  68         enum TDB_ERROR ecode = tdb_allrecord_upgrade(tdb, TDB1_FREELIST_TOP);
  69         if (ecode == TDB_SUCCESS)
  70                 return 0;
  71         tdb->last_error = ecode;
  72         return -1;
  73 }
  74
  75 static struct tdb_lock *tdb1_find_nestlock(struct tdb1_context *tdb,
  76                                            tdb1_off_t offset)
  77 {
  78         unsigned int i;
  79
  80         for (i=0; i<tdb->file->num_lockrecs; i++) {
  81                 if (tdb->file->lockrecs[i].off == offset) {
  82                         return &tdb->file->lockrecs[i];
  83                 }
  84         }
  85         return NULL;
  86 }
  87
  88 /* lock an offset in the database. */
  89 int tdb1_nest_lock(struct tdb1_context *tdb, uint32_t offset, int ltype,
  90                   enum tdb_lock_flags flags)
  91 {
  92         enum TDB_ERROR ecode;
  93
  94         if (offset >= lock_offset(tdb->header.hash_size)) {
  95                 tdb->last_error = tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
  96                                         "tdb1_lock: invalid offset %u for"
  97                                         " ltype=%d",
  98                                         offset, ltype);
  99                 return -1;
 100         }
 101
 102         ecode = tdb_nest_lock(tdb, offset, ltype, flags | TDB_LOCK_NOCHECK);
 103         if (unlikely(ecode != TDB_SUCCESS)) {
 104                 tdb->last_error = ecode;
 105                 return -1;
 106         }
 107         return 0;
 108 }
 109
 110 static int tdb1_lock_and_recover(struct tdb1_context *tdb)
 111 {
 112         int ret;
 113
 114         /* We need to match locking order in transaction commit. */
 115         if (tdb1_brlock(tdb, F_WRLCK, TDB1_FREELIST_TOP, 0,
 116                         TDB_LOCK_WAIT|TDB_LOCK_NOCHECK)) {
 117                 return -1;
 118         }
 119
 120         if (tdb1_brlock(tdb, F_WRLCK, TDB1_OPEN_LOCK, 1,
 121                         TDB_LOCK_WAIT|TDB_LOCK_NOCHECK)) {
 122                 tdb1_brunlock(tdb, F_WRLCK, TDB1_FREELIST_TOP, 0);
 123                 return -1;
 124         }
 125
 126         ret = tdb1_transaction_recover(tdb);
 127
 128         tdb1_brunlock(tdb, F_WRLCK, TDB1_OPEN_LOCK, 1);
 129         tdb1_brunlock(tdb, F_WRLCK, TDB1_FREELIST_TOP, 0);
 130
 131         return ret;
 132 }
 133
 134 static bool have_data_locks(const struct tdb1_context *tdb)
 135 {
 136         unsigned int i;
 137
 138         for (i = 0; i < tdb->file->num_lockrecs; i++) {
 139                 if (tdb->file->lockrecs[i].off >= lock_offset(-1))
 140                         return true;
 141         }
 142         return false;
 143 }
 144
 145 static int tdb1_lock_list(struct tdb1_context *tdb, int list, int ltype,
 146                          enum tdb_lock_flags waitflag)
 147 {
 148         int ret;
 149         bool check = false;
 150
 151         /* a allrecord lock allows us to avoid per chain locks */
 152         if (tdb->file->allrecord_lock.count &&
 153             (ltype == tdb->file->allrecord_lock.ltype || ltype == F_RDLCK)) {
 154                 return 0;
 155         }
 156
 157         if (tdb->file->allrecord_lock.count) {
 158                 tdb->last_error = TDB_ERR_LOCK;
 159                 ret = -1;
 160         } else {
 161                 /* Only check when we grab first data lock. */
 162                 check = !have_data_locks(tdb);
 163                 ret = tdb1_nest_lock(tdb, lock_offset(list), ltype, waitflag);
 164
 165                 if (ret == 0 && check && tdb1_needs_recovery(tdb)) {
 166                         tdb1_nest_unlock(tdb, lock_offset(list), ltype);
 167
 168                         if (tdb1_lock_and_recover(tdb) == -1) {
 169                                 return -1;
 170                         }
 171                         return tdb1_lock_list(tdb, list, ltype, waitflag);
 172                 }
 173         }
 174         return ret;
 175 }
 176
 177 /* lock a list in the database. list -1 is the alloc list */
 178 int tdb1_lock(struct tdb1_context *tdb, int list, int ltype)
 179 {
 180         int ret;
 181
 182         ret = tdb1_lock_list(tdb, list, ltype, TDB_LOCK_WAIT);
 183         /* Don't log for EAGAIN and EINTR: they could have overridden lock fns */
 184         if (ret && errno != EAGAIN && errno != EINTR) {
 185                 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
 186                            "tdb1_lock failed on list %d "
 187                            "ltype=%d (%s)",  list, ltype, strerror(errno));
 188         }
 189         return ret;
 190 }
 191
 192 int tdb1_nest_unlock(struct tdb1_context *tdb, uint32_t offset, int ltype)
 193 {
 194         enum TDB_ERROR ecode;
 195
 196         /* Sanity checks */
 197         if (offset >= lock_offset(tdb->header.hash_size)) {
 198                 tdb->last_error = tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
 199                                         "tdb1_unlock: offset %u invalid (%d)",
 200                                         offset, tdb->header.hash_size);
 201                 return -1;
 202         }
 203
 204         ecode = tdb_nest_unlock(tdb, offset, ltype);
 205         if (unlikely(ecode != TDB_SUCCESS)) {
 206                 tdb->last_error = ecode;
 207                 return -1;
 208         }
 209         return 0;
 210 }
 211
 212 int tdb1_unlock(struct tdb1_context *tdb, int list, int ltype)
 213 {
 214         /* a global lock allows us to avoid per chain locks */
 215         if (tdb->file->allrecord_lock.count &&
 216             (ltype == tdb->file->allrecord_lock.ltype || ltype == F_RDLCK)) {
 217                 return 0;
 218         }
 219
 220         if (tdb->file->allrecord_lock.count) {
 221                 tdb->last_error = TDB_ERR_LOCK;
 222                 return -1;
 223         }
 224
 225         return tdb1_nest_unlock(tdb, lock_offset(list), ltype);
 226 }
 227
 228 /*
 229   get the transaction lock
 230  */
 231 int tdb1_transaction_lock(struct tdb1_context *tdb, int ltype,
 232                          enum tdb_lock_flags lockflags)
 233 {
 234         return tdb1_nest_lock(tdb, TDB1_TRANSACTION_LOCK, ltype, lockflags);
 235 }
 236
 237 /*
 238   release the transaction lock
 239  */
 240 int tdb1_transaction_unlock(struct tdb1_context *tdb, int ltype)
 241 {
 242         return tdb1_nest_unlock(tdb, TDB1_TRANSACTION_LOCK, ltype);
 243 }
 244
 245 /* lock/unlock entire database.  It can only be upgradable if you have some
 246  * other way of guaranteeing exclusivity (ie. transaction write lock).
 247  * We do the locking gradually to avoid being starved by smaller locks. */
 248 int tdb1_allrecord_lock(struct tdb1_context *tdb, int ltype,
 249                        enum tdb_lock_flags flags, bool upgradable)
 250 {
 251         enum TDB_ERROR ecode;
 252
 253         /* tdb_lock_gradual() doesn't know about tdb->traverse_read. */
 254         if (tdb->traverse_read && !(tdb->flags & TDB_NOLOCK)) {
 255                 tdb->last_error = tdb_logerr(tdb, TDB_ERR_LOCK,
 256                                              TDB_LOG_USE_ERROR,
 257                                              "tdb1_allrecord_lock during"
 258                                              " tdb1_read_traverse");
 259                 return -1;
 260         }
 261
 262         if (tdb->file->allrecord_lock.count
 263             && tdb->file->allrecord_lock.ltype == ltype) {
 264                 tdb->file->allrecord_lock.count++;
 265                 return 0;
 266         }
 267
 268         if (tdb1_have_extra_locks(tdb)) {
 269                 /* can't combine global and chain locks */
 270                 tdb->last_error = tdb_logerr(tdb, TDB_ERR_LOCK,
 271                                              TDB_LOG_USE_ERROR,
 272                                              "tdb1_allrecord_lock holding"
 273                                              " other locks");
 274                 return -1;
 275         }
 276
 277         if (upgradable && ltype != F_RDLCK) {
 278                 /* tdb error: you can't upgrade a write lock! */
 279                 tdb->last_error = tdb_logerr(tdb, TDB_ERR_LOCK,
 280                                              TDB_LOG_ERROR,
 281                                              "tdb1_allrecord_lock cannot"
 282                                              " have upgradable write lock");
 283                 return -1;
 284         }
 285
 286         /* We cover two kinds of locks:
 287          * 1) Normal chain locks.  Taken for almost all operations.
 288          * 3) Individual records locks.  Taken after normal or free
 289          *    chain locks.
 290          *
 291          * It is (1) which cause the starvation problem, so we're only
 292          * gradual for that. */
 293         ecode = tdb_lock_gradual(tdb, ltype, flags | TDB_LOCK_NOCHECK,
 294                                  TDB1_FREELIST_TOP, tdb->header.hash_size * 4);
 295         if (ecode != TDB_SUCCESS) {
 296                 tdb->last_error = ecode;
 297                 return -1;
 298         }
 299
 300         /* Grab individual record locks. */
 301         if (tdb1_brlock(tdb, ltype, lock_offset(tdb->header.hash_size), 0,
 302                        flags) == -1) {
 303                 tdb1_brunlock(tdb, ltype, TDB1_FREELIST_TOP,
 304                              tdb->header.hash_size * 4);
 305                 return -1;
 306         }
 307
 308         /* FIXME: Temporary cast. */
 309         tdb->file->allrecord_lock.owner = (void *)(struct tdb1_context *)tdb;
 310         tdb->file->allrecord_lock.count = 1;
 311         /* If it's upgradable, it's actually exclusive so we can treat
 312          * it as a write lock. */
 313         tdb->file->allrecord_lock.ltype = upgradable ? F_WRLCK : ltype;
 314         tdb->file->allrecord_lock.off = upgradable;
 315
 316         if (tdb1_needs_recovery(tdb)) {
 317                 tdb1_allrecord_unlock(tdb, ltype);
 318                 if (tdb1_lock_and_recover(tdb) == -1) {
 319                         return -1;
 320                 }
 321                 return tdb1_allrecord_lock(tdb, ltype, flags, upgradable);
 322         }
 323
 324         return 0;
 325 }
 326
 327
 328
 329 /* unlock entire db */
 330 int tdb1_allrecord_unlock(struct tdb1_context *tdb, int ltype)
 331 {
 332         /* Don't try this during r/o traversal! */
 333         if (tdb->traverse_read) {
 334                 tdb->last_error = TDB_ERR_LOCK;
 335                 return -1;
 336         }
 337
 338         if (tdb->file->allrecord_lock.count == 0) {
 339                 tdb->last_error = TDB_ERR_LOCK;
 340                 return -1;
 341         }
 342
 343         /* Upgradable locks are marked as write locks. */
 344         if (tdb->file->allrecord_lock.ltype != ltype
 345             && (!tdb->file->allrecord_lock.off || ltype != F_RDLCK)) {
 346                 tdb->last_error = TDB_ERR_LOCK;
 347                 return -1;
 348         }
 349
 350         if (tdb->file->allrecord_lock.count > 1) {
 351                 tdb->file->allrecord_lock.count--;
 352                 return 0;
 353         }
 354
 355         if (tdb1_brunlock(tdb, ltype, TDB1_FREELIST_TOP, 0)) {
 356                 tdb_logerr(tdb, tdb->last_error, TDB_LOG_ERROR,
 357                            "tdb1_unlockall failed (%s)", strerror(errno));
 358                 return -1;
 359         }
 360
 361         tdb->file->allrecord_lock.count = 0;
 362         tdb->file->allrecord_lock.ltype = 0;
 363
 364         return 0;
 365 }
 366
 367 /* lock entire database with write lock */
 368 int tdb1_lockall(struct tdb1_context *tdb)
 369 {
 370         return tdb1_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT, false);
 371 }
 372
 373 /* unlock entire database with write lock */
 374 int tdb1_unlockall(struct tdb1_context *tdb)
 375 {
 376         return tdb1_allrecord_unlock(tdb, F_WRLCK);
 377 }
 378
 379 /* lock entire database with read lock */
 380 int tdb1_lockall_read(struct tdb1_context *tdb)
 381 {
 382         return tdb1_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, false);
 383 }
 384
 385 /* unlock entire database with read lock */
 386 int tdb1_unlockall_read(struct tdb1_context *tdb)
 387 {
 388         return tdb1_allrecord_unlock(tdb, F_RDLCK);
 389 }
 390
 391 /* lock/unlock one hash chain. This is meant to be used to reduce
 392    contention - it cannot guarantee how many records will be locked */
 393 int tdb1_chainlock(struct tdb1_context *tdb, TDB_DATA key)
 394 {
 395         int ret = tdb1_lock(tdb,
 396                             TDB1_BUCKET(tdb_hash(tdb, key.dptr, key.dsize)),
 397                             F_WRLCK);
 398         return ret;
 399 }
 400
 401 int tdb1_chainunlock(struct tdb1_context *tdb, TDB_DATA key)
 402 {
 403         return tdb1_unlock(tdb, TDB1_BUCKET(tdb_hash(tdb, key.dptr, key.dsize)),
 404                            F_WRLCK);
 405 }
 406
 407 int tdb1_chainlock_read(struct tdb1_context *tdb, TDB_DATA key)
 408 {
 409         int ret;
 410         ret = tdb1_lock(tdb, TDB1_BUCKET(tdb_hash(tdb, key.dptr, key.dsize)),
 411                         F_RDLCK);
 412         return ret;
 413 }
 414
 415 int tdb1_chainunlock_read(struct tdb1_context *tdb, TDB_DATA key)
 416 {
 417         return tdb1_unlock(tdb, TDB1_BUCKET(tdb_hash(tdb, key.dptr, key.dsize)),
 418                            F_RDLCK);
 419 }
 420
 421 /* record lock stops delete underneath */
 422 int tdb1_lock_record(struct tdb1_context *tdb, tdb1_off_t off)
 423 {
 424         if (tdb->file->allrecord_lock.count) {
 425                 return 0;
 426         }
 427         return off ? tdb1_brlock(tdb, F_RDLCK, off, 1, TDB_LOCK_WAIT) : 0;
 428 }
 429
 430 /*
 431   Write locks override our own fcntl readlocks, so check it here.
 432   Note this is meant to be F_SETLK, *not* F_SETLKW, as it's not
 433   an error to fail to get the lock here.
 434 */
 435 int tdb1_write_lock_record(struct tdb1_context *tdb, tdb1_off_t off)
 436 {
 437         struct tdb1_traverse_lock *i;
 438         for (i = &tdb->travlocks; i; i = i->next)
 439                 if (i->off == off)
 440                         return -1;
 441         if (tdb->file->allrecord_lock.count) {
 442                 if (tdb->file->allrecord_lock.ltype == F_WRLCK) {
 443                         return 0;
 444                 }
 445                 return -1;
 446         }
 447         return tdb1_brlock(tdb, F_WRLCK, off, 1, TDB_LOCK_NOWAIT|TDB_LOCK_PROBE);
 448 }
 449
 450 int tdb1_write_unlock_record(struct tdb1_context *tdb, tdb1_off_t off)
 451 {
 452         if (tdb->file->allrecord_lock.count) {
 453                 return 0;
 454         }
 455         return tdb1_brunlock(tdb, F_WRLCK, off, 1);
 456 }
 457
 458 /* fcntl locks don't stack: avoid unlocking someone else's */
 459 int tdb1_unlock_record(struct tdb1_context *tdb, tdb1_off_t off)
 460 {
 461         struct tdb1_traverse_lock *i;
 462         uint32_t count = 0;
 463
 464         if (tdb->file->allrecord_lock.count) {
 465                 return 0;
 466         }
 467
 468         if (off == 0)
 469                 return 0;
 470         for (i = &tdb->travlocks; i; i = i->next)
 471                 if (i->off == off)
 472                         count++;
 473         return (count == 1 ? tdb1_brunlock(tdb, F_RDLCK, off, 1) : 0);
 474 }
 475
 476 bool tdb1_have_extra_locks(struct tdb1_context *tdb)
 477 {
 478         unsigned int extra = tdb->file->num_lockrecs;
 479
 480         /* A transaction holds the lock for all records. */
 481         if (!tdb->transaction && tdb->file->allrecord_lock.count) {
 482                 return true;
 483         }
 484
 485         /* We always hold the active lock if CLEAR_IF_FIRST. */
 486         if (tdb1_find_nestlock(tdb, TDB1_ACTIVE_LOCK)) {
 487                 extra--;
 488         }
 489
 490         /* In a transaction, we expect to hold the transaction lock */
 491         if (tdb->transaction
 492             && tdb1_find_nestlock(tdb, TDB1_TRANSACTION_LOCK)) {
 493                 extra--;
 494         }
 495
 496         return extra;
 497 }
 498
 499 /* The transaction code uses this to remove all locks. */
 500 void tdb1_release_transaction_locks(struct tdb1_context *tdb)
 501 {
 502         unsigned int i, active = 0;
 503
 504         if (tdb->file->allrecord_lock.count != 0) {
 505                 tdb1_brunlock(tdb, tdb->file->allrecord_lock.ltype, TDB1_FREELIST_TOP, 0);
 506                 tdb->file->allrecord_lock.count = 0;
 507         }
 508
 509         for (i=0;i<tdb->file->num_lockrecs;i++) {
 510                 struct tdb_lock *lck = &tdb->file->lockrecs[i];
 511
 512                 /* Don't release the active lock!  Copy it to first entry. */
 513                 if (lck->off == TDB1_ACTIVE_LOCK) {
 514                         tdb->file->lockrecs[active++] = *lck;
 515                 } else {
 516                         tdb1_brunlock(tdb, lck->ltype, lck->off, 1);
 517                 }
 518         }
 519         tdb->file->num_lockrecs = active;
 520         if (tdb->file->num_lockrecs == 0) {
 521                 SAFE_FREE(tdb->file->lockrecs);
 522         }
 523 }