| 1 | /*------------------------------------------------------------------------- |
| 2 | * |
| 3 | * lmgr.c |
| 4 | * POSTGRES lock manager code |
| 5 | * |
| 6 | * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group |
| 7 | * Portions Copyright (c) 1994, Regents of the University of California |
| 8 | * |
| 9 | * |
| 10 | * IDENTIFICATION |
| 11 | * src/backend/storage/lmgr/lmgr.c |
| 12 | * |
| 13 | *------------------------------------------------------------------------- |
| 14 | */ |
| 15 | |
| 16 | #include "postgres.h" |
| 17 | |
| 18 | #include "access/subtrans.h" |
| 19 | #include "access/transam.h" |
| 20 | #include "access/xact.h" |
| 21 | #include "catalog/catalog.h" |
| 22 | #include "commands/progress.h" |
| 23 | #include "miscadmin.h" |
| 24 | #include "pgstat.h" |
| 25 | #include "storage/lmgr.h" |
| 26 | #include "storage/procarray.h" |
| 27 | #include "storage/sinvaladt.h" |
| 28 | #include "utils/inval.h" |
| 29 | |
| 30 | |
| 31 | /* |
| 32 | * Per-backend counter for generating speculative insertion tokens. |
| 33 | * |
| 34 | * This may wrap around, but that's OK as it's only used for the short |
| 35 | * duration between inserting a tuple and checking that there are no (unique) |
| 36 | * constraint violations. It's theoretically possible that a backend sees a |
| 37 | * tuple that was speculatively inserted by another backend, but before it has |
| 38 | * started waiting on the token, the other backend completes its insertion, |
| 39 | * and then performs 2^32 unrelated insertions. And after all that, the |
| 40 | * first backend finally calls SpeculativeInsertionLockAcquire(), with the |
| 41 | * intention of waiting for the first insertion to complete, but ends up |
| 42 | * waiting for the latest unrelated insertion instead. Even then, nothing |
| 43 | * particularly bad happens: in the worst case they deadlock, causing one of |
| 44 | * the transactions to abort. |
| 45 | */ |
| 46 | static uint32 speculativeInsertionToken = 0; |
| 47 | |
| 48 | |
| 49 | /* |
| 50 | * Struct to hold context info for transaction lock waits. |
| 51 | * |
| 52 | * 'oper' is the operation that needs to wait for the other transaction; 'rel' |
| 53 | * and 'ctid' specify the address of the tuple being waited for. |
| 54 | */ |
| 55 | typedef struct XactLockTableWaitInfo |
| 56 | { |
| 57 | XLTW_Oper oper; |
| 58 | Relation rel; |
| 59 | ItemPointer ctid; |
| 60 | } XactLockTableWaitInfo; |
| 61 | |
| 62 | static void XactLockTableWaitErrorCb(void *arg); |
| 63 | |
| 64 | /* |
| 65 | * RelationInitLockInfo |
| 66 | * Initializes the lock information in a relation descriptor. |
| 67 | * |
| 68 | * relcache.c must call this during creation of any reldesc. |
| 69 | */ |
| 70 | void |
| 71 | RelationInitLockInfo(Relation relation) |
| 72 | { |
| 73 | Assert(RelationIsValid(relation)); |
| 74 | Assert(OidIsValid(RelationGetRelid(relation))); |
| 75 | |
| 76 | relation->rd_lockInfo.lockRelId.relId = RelationGetRelid(relation); |
| 77 | |
| 78 | if (relation->rd_rel->relisshared) |
| 79 | relation->rd_lockInfo.lockRelId.dbId = InvalidOid; |
| 80 | else |
| 81 | relation->rd_lockInfo.lockRelId.dbId = MyDatabaseId; |
| 82 | } |
| 83 | |
| 84 | /* |
| 85 | * SetLocktagRelationOid |
| 86 | * Set up a locktag for a relation, given only relation OID |
| 87 | */ |
| 88 | static inline void |
| 89 | SetLocktagRelationOid(LOCKTAG *tag, Oid relid) |
| 90 | { |
| 91 | Oid dbid; |
| 92 | |
| 93 | if (IsSharedRelation(relid)) |
| 94 | dbid = InvalidOid; |
| 95 | else |
| 96 | dbid = MyDatabaseId; |
| 97 | |
| 98 | SET_LOCKTAG_RELATION(*tag, dbid, relid); |
| 99 | } |
| 100 | |
| 101 | /* |
| 102 | * LockRelationOid |
| 103 | * |
| 104 | * Lock a relation given only its OID. This should generally be used |
| 105 | * before attempting to open the relation's relcache entry. |
| 106 | */ |
| 107 | void |
| 108 | LockRelationOid(Oid relid, LOCKMODE lockmode) |
| 109 | { |
| 110 | LOCKTAG tag; |
| 111 | LOCALLOCK *locallock; |
| 112 | LockAcquireResult res; |
| 113 | |
| 114 | SetLocktagRelationOid(&tag, relid); |
| 115 | |
| 116 | res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock); |
| 117 | |
| 118 | /* |
| 119 | * Now that we have the lock, check for invalidation messages, so that we |
| 120 | * will update or flush any stale relcache entry before we try to use it. |
| 121 | * RangeVarGetRelid() specifically relies on us for this. We can skip |
| 122 | * this in the not-uncommon case that we already had the same type of lock |
| 123 | * being requested, since then no one else could have modified the |
| 124 | * relcache entry in an undesirable way. (In the case where our own xact |
| 125 | * modifies the rel, the relcache update happens via |
| 126 | * CommandCounterIncrement, not here.) |
| 127 | * |
| 128 | * However, in corner cases where code acts on tables (usually catalogs) |
| 129 | * recursively, we might get here while still processing invalidation |
| 130 | * messages in some outer execution of this function or a sibling. The |
| 131 | * "cleared" status of the lock tells us whether we really are done |
| 132 | * absorbing relevant inval messages. |
| 133 | */ |
| 134 | if (res != LOCKACQUIRE_ALREADY_CLEAR) |
| 135 | { |
| 136 | AcceptInvalidationMessages(); |
| 137 | MarkLockClear(locallock); |
| 138 | } |
| 139 | } |
| 140 | |
| 141 | /* |
| 142 | * ConditionalLockRelationOid |
| 143 | * |
| 144 | * As above, but only lock if we can get the lock without blocking. |
| 145 | * Returns true iff the lock was acquired. |
| 146 | * |
| 147 | * NOTE: we do not currently need conditional versions of all the |
| 148 | * LockXXX routines in this file, but they could easily be added if needed. |
| 149 | */ |
| 150 | bool |
| 151 | ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode) |
| 152 | { |
| 153 | LOCKTAG tag; |
| 154 | LOCALLOCK *locallock; |
| 155 | LockAcquireResult res; |
| 156 | |
| 157 | SetLocktagRelationOid(&tag, relid); |
| 158 | |
| 159 | res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock); |
| 160 | |
| 161 | if (res == LOCKACQUIRE_NOT_AVAIL) |
| 162 | return false; |
| 163 | |
| 164 | /* |
| 165 | * Now that we have the lock, check for invalidation messages; see notes |
| 166 | * in LockRelationOid. |
| 167 | */ |
| 168 | if (res != LOCKACQUIRE_ALREADY_CLEAR) |
| 169 | { |
| 170 | AcceptInvalidationMessages(); |
| 171 | MarkLockClear(locallock); |
| 172 | } |
| 173 | |
| 174 | return true; |
| 175 | } |
| 176 | |
| 177 | /* |
| 178 | * UnlockRelationId |
| 179 | * |
| 180 | * Unlock, given a LockRelId. This is preferred over UnlockRelationOid |
| 181 | * for speed reasons. |
| 182 | */ |
| 183 | void |
| 184 | UnlockRelationId(LockRelId *relid, LOCKMODE lockmode) |
| 185 | { |
| 186 | LOCKTAG tag; |
| 187 | |
| 188 | SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId); |
| 189 | |
| 190 | LockRelease(&tag, lockmode, false); |
| 191 | } |
| 192 | |
| 193 | /* |
| 194 | * UnlockRelationOid |
| 195 | * |
| 196 | * Unlock, given only a relation Oid. Use UnlockRelationId if you can. |
| 197 | */ |
| 198 | void |
| 199 | UnlockRelationOid(Oid relid, LOCKMODE lockmode) |
| 200 | { |
| 201 | LOCKTAG tag; |
| 202 | |
| 203 | SetLocktagRelationOid(&tag, relid); |
| 204 | |
| 205 | LockRelease(&tag, lockmode, false); |
| 206 | } |
| 207 | |
| 208 | /* |
| 209 | * LockRelation |
| 210 | * |
| 211 | * This is a convenience routine for acquiring an additional lock on an |
| 212 | * already-open relation. Never try to do "relation_open(foo, NoLock)" |
| 213 | * and then lock with this. |
| 214 | */ |
| 215 | void |
| 216 | LockRelation(Relation relation, LOCKMODE lockmode) |
| 217 | { |
| 218 | LOCKTAG tag; |
| 219 | LOCALLOCK *locallock; |
| 220 | LockAcquireResult res; |
| 221 | |
| 222 | SET_LOCKTAG_RELATION(tag, |
| 223 | relation->rd_lockInfo.lockRelId.dbId, |
| 224 | relation->rd_lockInfo.lockRelId.relId); |
| 225 | |
| 226 | res = LockAcquireExtended(&tag, lockmode, false, false, true, &locallock); |
| 227 | |
| 228 | /* |
| 229 | * Now that we have the lock, check for invalidation messages; see notes |
| 230 | * in LockRelationOid. |
| 231 | */ |
| 232 | if (res != LOCKACQUIRE_ALREADY_CLEAR) |
| 233 | { |
| 234 | AcceptInvalidationMessages(); |
| 235 | MarkLockClear(locallock); |
| 236 | } |
| 237 | } |
| 238 | |
| 239 | /* |
| 240 | * ConditionalLockRelation |
| 241 | * |
| 242 | * This is a convenience routine for acquiring an additional lock on an |
| 243 | * already-open relation. Never try to do "relation_open(foo, NoLock)" |
| 244 | * and then lock with this. |
| 245 | */ |
| 246 | bool |
| 247 | ConditionalLockRelation(Relation relation, LOCKMODE lockmode) |
| 248 | { |
| 249 | LOCKTAG tag; |
| 250 | LOCALLOCK *locallock; |
| 251 | LockAcquireResult res; |
| 252 | |
| 253 | SET_LOCKTAG_RELATION(tag, |
| 254 | relation->rd_lockInfo.lockRelId.dbId, |
| 255 | relation->rd_lockInfo.lockRelId.relId); |
| 256 | |
| 257 | res = LockAcquireExtended(&tag, lockmode, false, true, true, &locallock); |
| 258 | |
| 259 | if (res == LOCKACQUIRE_NOT_AVAIL) |
| 260 | return false; |
| 261 | |
| 262 | /* |
| 263 | * Now that we have the lock, check for invalidation messages; see notes |
| 264 | * in LockRelationOid. |
| 265 | */ |
| 266 | if (res != LOCKACQUIRE_ALREADY_CLEAR) |
| 267 | { |
| 268 | AcceptInvalidationMessages(); |
| 269 | MarkLockClear(locallock); |
| 270 | } |
| 271 | |
| 272 | return true; |
| 273 | } |
| 274 | |
| 275 | /* |
| 276 | * UnlockRelation |
| 277 | * |
| 278 | * This is a convenience routine for unlocking a relation without also |
| 279 | * closing it. |
| 280 | */ |
| 281 | void |
| 282 | UnlockRelation(Relation relation, LOCKMODE lockmode) |
| 283 | { |
| 284 | LOCKTAG tag; |
| 285 | |
| 286 | SET_LOCKTAG_RELATION(tag, |
| 287 | relation->rd_lockInfo.lockRelId.dbId, |
| 288 | relation->rd_lockInfo.lockRelId.relId); |
| 289 | |
| 290 | LockRelease(&tag, lockmode, false); |
| 291 | } |
| 292 | |
| 293 | /* |
| 294 | * CheckRelationLockedByMe |
| 295 | * |
| 296 | * Returns true if current transaction holds a lock on 'relation' of mode |
| 297 | * 'lockmode'. If 'orstronger' is true, a stronger lockmode is also OK. |
| 298 | * ("Stronger" is defined as "numerically higher", which is a bit |
| 299 | * semantically dubious but is OK for the purposes we use this for.) |
| 300 | */ |
| 301 | bool |
| 302 | CheckRelationLockedByMe(Relation relation, LOCKMODE lockmode, bool orstronger) |
| 303 | { |
| 304 | LOCKTAG tag; |
| 305 | |
| 306 | SET_LOCKTAG_RELATION(tag, |
| 307 | relation->rd_lockInfo.lockRelId.dbId, |
| 308 | relation->rd_lockInfo.lockRelId.relId); |
| 309 | |
| 310 | if (LockHeldByMe(&tag, lockmode)) |
| 311 | return true; |
| 312 | |
| 313 | if (orstronger) |
| 314 | { |
| 315 | LOCKMODE slockmode; |
| 316 | |
| 317 | for (slockmode = lockmode + 1; |
| 318 | slockmode <= MaxLockMode; |
| 319 | slockmode++) |
| 320 | { |
| 321 | if (LockHeldByMe(&tag, slockmode)) |
| 322 | { |
| 323 | #ifdef NOT_USED |
| 324 | /* Sometimes this might be useful for debugging purposes */ |
| 325 | elog(WARNING, "lock mode %s substituted for %s on relation %s" , |
| 326 | GetLockmodeName(tag.locktag_lockmethodid, slockmode), |
| 327 | GetLockmodeName(tag.locktag_lockmethodid, lockmode), |
| 328 | RelationGetRelationName(relation)); |
| 329 | #endif |
| 330 | return true; |
| 331 | } |
| 332 | } |
| 333 | } |
| 334 | |
| 335 | return false; |
| 336 | } |
| 337 | |
| 338 | /* |
| 339 | * LockHasWaitersRelation |
| 340 | * |
| 341 | * This is a function to check whether someone else is waiting for a |
| 342 | * lock which we are currently holding. |
| 343 | */ |
| 344 | bool |
| 345 | LockHasWaitersRelation(Relation relation, LOCKMODE lockmode) |
| 346 | { |
| 347 | LOCKTAG tag; |
| 348 | |
| 349 | SET_LOCKTAG_RELATION(tag, |
| 350 | relation->rd_lockInfo.lockRelId.dbId, |
| 351 | relation->rd_lockInfo.lockRelId.relId); |
| 352 | |
| 353 | return LockHasWaiters(&tag, lockmode, false); |
| 354 | } |
| 355 | |
| 356 | /* |
| 357 | * LockRelationIdForSession |
| 358 | * |
| 359 | * This routine grabs a session-level lock on the target relation. The |
| 360 | * session lock persists across transaction boundaries. It will be removed |
| 361 | * when UnlockRelationIdForSession() is called, or if an ereport(ERROR) occurs, |
| 362 | * or if the backend exits. |
| 363 | * |
| 364 | * Note that one should also grab a transaction-level lock on the rel |
| 365 | * in any transaction that actually uses the rel, to ensure that the |
| 366 | * relcache entry is up to date. |
| 367 | */ |
| 368 | void |
| 369 | LockRelationIdForSession(LockRelId *relid, LOCKMODE lockmode) |
| 370 | { |
| 371 | LOCKTAG tag; |
| 372 | |
| 373 | SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId); |
| 374 | |
| 375 | (void) LockAcquire(&tag, lockmode, true, false); |
| 376 | } |
| 377 | |
| 378 | /* |
| 379 | * UnlockRelationIdForSession |
| 380 | */ |
| 381 | void |
| 382 | UnlockRelationIdForSession(LockRelId *relid, LOCKMODE lockmode) |
| 383 | { |
| 384 | LOCKTAG tag; |
| 385 | |
| 386 | SET_LOCKTAG_RELATION(tag, relid->dbId, relid->relId); |
| 387 | |
| 388 | LockRelease(&tag, lockmode, true); |
| 389 | } |
| 390 | |
| 391 | /* |
| 392 | * LockRelationForExtension |
| 393 | * |
| 394 | * This lock tag is used to interlock addition of pages to relations. |
| 395 | * We need such locking because bufmgr/smgr definition of P_NEW is not |
| 396 | * race-condition-proof. |
| 397 | * |
| 398 | * We assume the caller is already holding some type of regular lock on |
| 399 | * the relation, so no AcceptInvalidationMessages call is needed here. |
| 400 | */ |
| 401 | void |
| 402 | LockRelationForExtension(Relation relation, LOCKMODE lockmode) |
| 403 | { |
| 404 | LOCKTAG tag; |
| 405 | |
| 406 | SET_LOCKTAG_RELATION_EXTEND(tag, |
| 407 | relation->rd_lockInfo.lockRelId.dbId, |
| 408 | relation->rd_lockInfo.lockRelId.relId); |
| 409 | |
| 410 | (void) LockAcquire(&tag, lockmode, false, false); |
| 411 | } |
| 412 | |
| 413 | /* |
| 414 | * ConditionalLockRelationForExtension |
| 415 | * |
| 416 | * As above, but only lock if we can get the lock without blocking. |
| 417 | * Returns true iff the lock was acquired. |
| 418 | */ |
| 419 | bool |
| 420 | ConditionalLockRelationForExtension(Relation relation, LOCKMODE lockmode) |
| 421 | { |
| 422 | LOCKTAG tag; |
| 423 | |
| 424 | SET_LOCKTAG_RELATION_EXTEND(tag, |
| 425 | relation->rd_lockInfo.lockRelId.dbId, |
| 426 | relation->rd_lockInfo.lockRelId.relId); |
| 427 | |
| 428 | return (LockAcquire(&tag, lockmode, false, true) != LOCKACQUIRE_NOT_AVAIL); |
| 429 | } |
| 430 | |
| 431 | /* |
| 432 | * RelationExtensionLockWaiterCount |
| 433 | * |
| 434 | * Count the number of processes waiting for the given relation extension lock. |
| 435 | */ |
| 436 | int |
| 437 | RelationExtensionLockWaiterCount(Relation relation) |
| 438 | { |
| 439 | LOCKTAG tag; |
| 440 | |
| 441 | SET_LOCKTAG_RELATION_EXTEND(tag, |
| 442 | relation->rd_lockInfo.lockRelId.dbId, |
| 443 | relation->rd_lockInfo.lockRelId.relId); |
| 444 | |
| 445 | return LockWaiterCount(&tag); |
| 446 | } |
| 447 | |
| 448 | /* |
| 449 | * UnlockRelationForExtension |
| 450 | */ |
| 451 | void |
| 452 | UnlockRelationForExtension(Relation relation, LOCKMODE lockmode) |
| 453 | { |
| 454 | LOCKTAG tag; |
| 455 | |
| 456 | SET_LOCKTAG_RELATION_EXTEND(tag, |
| 457 | relation->rd_lockInfo.lockRelId.dbId, |
| 458 | relation->rd_lockInfo.lockRelId.relId); |
| 459 | |
| 460 | LockRelease(&tag, lockmode, false); |
| 461 | } |
| 462 | |
| 463 | /* |
| 464 | * LockPage |
| 465 | * |
| 466 | * Obtain a page-level lock. This is currently used by some index access |
| 467 | * methods to lock individual index pages. |
| 468 | */ |
| 469 | void |
| 470 | LockPage(Relation relation, BlockNumber blkno, LOCKMODE lockmode) |
| 471 | { |
| 472 | LOCKTAG tag; |
| 473 | |
| 474 | SET_LOCKTAG_PAGE(tag, |
| 475 | relation->rd_lockInfo.lockRelId.dbId, |
| 476 | relation->rd_lockInfo.lockRelId.relId, |
| 477 | blkno); |
| 478 | |
| 479 | (void) LockAcquire(&tag, lockmode, false, false); |
| 480 | } |
| 481 | |
| 482 | /* |
| 483 | * ConditionalLockPage |
| 484 | * |
| 485 | * As above, but only lock if we can get the lock without blocking. |
| 486 | * Returns true iff the lock was acquired. |
| 487 | */ |
| 488 | bool |
| 489 | ConditionalLockPage(Relation relation, BlockNumber blkno, LOCKMODE lockmode) |
| 490 | { |
| 491 | LOCKTAG tag; |
| 492 | |
| 493 | SET_LOCKTAG_PAGE(tag, |
| 494 | relation->rd_lockInfo.lockRelId.dbId, |
| 495 | relation->rd_lockInfo.lockRelId.relId, |
| 496 | blkno); |
| 497 | |
| 498 | return (LockAcquire(&tag, lockmode, false, true) != LOCKACQUIRE_NOT_AVAIL); |
| 499 | } |
| 500 | |
| 501 | /* |
| 502 | * UnlockPage |
| 503 | */ |
| 504 | void |
| 505 | UnlockPage(Relation relation, BlockNumber blkno, LOCKMODE lockmode) |
| 506 | { |
| 507 | LOCKTAG tag; |
| 508 | |
| 509 | SET_LOCKTAG_PAGE(tag, |
| 510 | relation->rd_lockInfo.lockRelId.dbId, |
| 511 | relation->rd_lockInfo.lockRelId.relId, |
| 512 | blkno); |
| 513 | |
| 514 | LockRelease(&tag, lockmode, false); |
| 515 | } |
| 516 | |
| 517 | /* |
| 518 | * LockTuple |
| 519 | * |
| 520 | * Obtain a tuple-level lock. This is used in a less-than-intuitive fashion |
| 521 | * because we can't afford to keep a separate lock in shared memory for every |
| 522 | * tuple. See heap_lock_tuple before using this! |
| 523 | */ |
| 524 | void |
| 525 | LockTuple(Relation relation, ItemPointer tid, LOCKMODE lockmode) |
| 526 | { |
| 527 | LOCKTAG tag; |
| 528 | |
| 529 | SET_LOCKTAG_TUPLE(tag, |
| 530 | relation->rd_lockInfo.lockRelId.dbId, |
| 531 | relation->rd_lockInfo.lockRelId.relId, |
| 532 | ItemPointerGetBlockNumber(tid), |
| 533 | ItemPointerGetOffsetNumber(tid)); |
| 534 | |
| 535 | (void) LockAcquire(&tag, lockmode, false, false); |
| 536 | } |
| 537 | |
| 538 | /* |
| 539 | * ConditionalLockTuple |
| 540 | * |
| 541 | * As above, but only lock if we can get the lock without blocking. |
| 542 | * Returns true iff the lock was acquired. |
| 543 | */ |
| 544 | bool |
| 545 | ConditionalLockTuple(Relation relation, ItemPointer tid, LOCKMODE lockmode) |
| 546 | { |
| 547 | LOCKTAG tag; |
| 548 | |
| 549 | SET_LOCKTAG_TUPLE(tag, |
| 550 | relation->rd_lockInfo.lockRelId.dbId, |
| 551 | relation->rd_lockInfo.lockRelId.relId, |
| 552 | ItemPointerGetBlockNumber(tid), |
| 553 | ItemPointerGetOffsetNumber(tid)); |
| 554 | |
| 555 | return (LockAcquire(&tag, lockmode, false, true) != LOCKACQUIRE_NOT_AVAIL); |
| 556 | } |
| 557 | |
| 558 | /* |
| 559 | * UnlockTuple |
| 560 | */ |
| 561 | void |
| 562 | UnlockTuple(Relation relation, ItemPointer tid, LOCKMODE lockmode) |
| 563 | { |
| 564 | LOCKTAG tag; |
| 565 | |
| 566 | SET_LOCKTAG_TUPLE(tag, |
| 567 | relation->rd_lockInfo.lockRelId.dbId, |
| 568 | relation->rd_lockInfo.lockRelId.relId, |
| 569 | ItemPointerGetBlockNumber(tid), |
| 570 | ItemPointerGetOffsetNumber(tid)); |
| 571 | |
| 572 | LockRelease(&tag, lockmode, false); |
| 573 | } |
| 574 | |
| 575 | /* |
| 576 | * XactLockTableInsert |
| 577 | * |
| 578 | * Insert a lock showing that the given transaction ID is running --- |
| 579 | * this is done when an XID is acquired by a transaction or subtransaction. |
| 580 | * The lock can then be used to wait for the transaction to finish. |
| 581 | */ |
| 582 | void |
| 583 | XactLockTableInsert(TransactionId xid) |
| 584 | { |
| 585 | LOCKTAG tag; |
| 586 | |
| 587 | SET_LOCKTAG_TRANSACTION(tag, xid); |
| 588 | |
| 589 | (void) LockAcquire(&tag, ExclusiveLock, false, false); |
| 590 | } |
| 591 | |
| 592 | /* |
| 593 | * XactLockTableDelete |
| 594 | * |
| 595 | * Delete the lock showing that the given transaction ID is running. |
| 596 | * (This is never used for main transaction IDs; those locks are only |
| 597 | * released implicitly at transaction end. But we do use it for subtrans IDs.) |
| 598 | */ |
| 599 | void |
| 600 | XactLockTableDelete(TransactionId xid) |
| 601 | { |
| 602 | LOCKTAG tag; |
| 603 | |
| 604 | SET_LOCKTAG_TRANSACTION(tag, xid); |
| 605 | |
| 606 | LockRelease(&tag, ExclusiveLock, false); |
| 607 | } |
| 608 | |
| 609 | /* |
| 610 | * XactLockTableWait |
| 611 | * |
| 612 | * Wait for the specified transaction to commit or abort. If an operation |
| 613 | * is specified, an error context callback is set up. If 'oper' is passed as |
| 614 | * None, no error context callback is set up. |
| 615 | * |
| 616 | * Note that this does the right thing for subtransactions: if we wait on a |
| 617 | * subtransaction, we will exit as soon as it aborts or its top parent commits. |
| 618 | * It takes some extra work to ensure this, because to save on shared memory |
| 619 | * the XID lock of a subtransaction is released when it ends, whether |
| 620 | * successfully or unsuccessfully. So we have to check if it's "still running" |
| 621 | * and if so wait for its parent. |
| 622 | */ |
| 623 | void |
| 624 | XactLockTableWait(TransactionId xid, Relation rel, ItemPointer ctid, |
| 625 | XLTW_Oper oper) |
| 626 | { |
| 627 | LOCKTAG tag; |
| 628 | XactLockTableWaitInfo info; |
| 629 | ErrorContextCallback callback; |
| 630 | bool first = true; |
| 631 | |
| 632 | /* |
| 633 | * If an operation is specified, set up our verbose error context |
| 634 | * callback. |
| 635 | */ |
| 636 | if (oper != XLTW_None) |
| 637 | { |
| 638 | Assert(RelationIsValid(rel)); |
| 639 | Assert(ItemPointerIsValid(ctid)); |
| 640 | |
| 641 | info.rel = rel; |
| 642 | info.ctid = ctid; |
| 643 | info.oper = oper; |
| 644 | |
| 645 | callback.callback = XactLockTableWaitErrorCb; |
| 646 | callback.arg = &info; |
| 647 | callback.previous = error_context_stack; |
| 648 | error_context_stack = &callback; |
| 649 | } |
| 650 | |
| 651 | for (;;) |
| 652 | { |
| 653 | Assert(TransactionIdIsValid(xid)); |
| 654 | Assert(!TransactionIdEquals(xid, GetTopTransactionIdIfAny())); |
| 655 | |
| 656 | SET_LOCKTAG_TRANSACTION(tag, xid); |
| 657 | |
| 658 | (void) LockAcquire(&tag, ShareLock, false, false); |
| 659 | |
| 660 | LockRelease(&tag, ShareLock, false); |
| 661 | |
| 662 | if (!TransactionIdIsInProgress(xid)) |
| 663 | break; |
| 664 | |
| 665 | /* |
| 666 | * If the Xid belonged to a subtransaction, then the lock would have |
| 667 | * gone away as soon as it was finished; for correct tuple visibility, |
| 668 | * the right action is to wait on its parent transaction to go away. |
| 669 | * But instead of going levels up one by one, we can just wait for the |
| 670 | * topmost transaction to finish with the same end result, which also |
| 671 | * incurs less locktable traffic. |
| 672 | * |
| 673 | * Some uses of this function don't involve tuple visibility -- such |
| 674 | * as when building snapshots for logical decoding. It is possible to |
| 675 | * see a transaction in ProcArray before it registers itself in the |
| 676 | * locktable. The topmost transaction in that case is the same xid, |
| 677 | * so we try again after a short sleep. (Don't sleep the first time |
| 678 | * through, to avoid slowing down the normal case.) |
| 679 | */ |
| 680 | if (!first) |
| 681 | pg_usleep(1000L); |
| 682 | first = false; |
| 683 | xid = SubTransGetTopmostTransaction(xid); |
| 684 | } |
| 685 | |
| 686 | if (oper != XLTW_None) |
| 687 | error_context_stack = callback.previous; |
| 688 | } |
| 689 | |
| 690 | /* |
| 691 | * ConditionalXactLockTableWait |
| 692 | * |
| 693 | * As above, but only lock if we can get the lock without blocking. |
| 694 | * Returns true if the lock was acquired. |
| 695 | */ |
| 696 | bool |
| 697 | ConditionalXactLockTableWait(TransactionId xid) |
| 698 | { |
| 699 | LOCKTAG tag; |
| 700 | bool first = true; |
| 701 | |
| 702 | for (;;) |
| 703 | { |
| 704 | Assert(TransactionIdIsValid(xid)); |
| 705 | Assert(!TransactionIdEquals(xid, GetTopTransactionIdIfAny())); |
| 706 | |
| 707 | SET_LOCKTAG_TRANSACTION(tag, xid); |
| 708 | |
| 709 | if (LockAcquire(&tag, ShareLock, false, true) == LOCKACQUIRE_NOT_AVAIL) |
| 710 | return false; |
| 711 | |
| 712 | LockRelease(&tag, ShareLock, false); |
| 713 | |
| 714 | if (!TransactionIdIsInProgress(xid)) |
| 715 | break; |
| 716 | |
| 717 | /* See XactLockTableWait about this case */ |
| 718 | if (!first) |
| 719 | pg_usleep(1000L); |
| 720 | first = false; |
| 721 | xid = SubTransGetTopmostTransaction(xid); |
| 722 | } |
| 723 | |
| 724 | return true; |
| 725 | } |
| 726 | |
| 727 | /* |
| 728 | * SpeculativeInsertionLockAcquire |
| 729 | * |
| 730 | * Insert a lock showing that the given transaction ID is inserting a tuple, |
| 731 | * but hasn't yet decided whether it's going to keep it. The lock can then be |
| 732 | * used to wait for the decision to go ahead with the insertion, or aborting |
| 733 | * it. |
| 734 | * |
| 735 | * The token is used to distinguish multiple insertions by the same |
| 736 | * transaction. It is returned to caller. |
| 737 | */ |
| 738 | uint32 |
| 739 | SpeculativeInsertionLockAcquire(TransactionId xid) |
| 740 | { |
| 741 | LOCKTAG tag; |
| 742 | |
| 743 | speculativeInsertionToken++; |
| 744 | |
| 745 | /* |
| 746 | * Check for wrap-around. Zero means no token is held, so don't use that. |
| 747 | */ |
| 748 | if (speculativeInsertionToken == 0) |
| 749 | speculativeInsertionToken = 1; |
| 750 | |
| 751 | SET_LOCKTAG_SPECULATIVE_INSERTION(tag, xid, speculativeInsertionToken); |
| 752 | |
| 753 | (void) LockAcquire(&tag, ExclusiveLock, false, false); |
| 754 | |
| 755 | return speculativeInsertionToken; |
| 756 | } |
| 757 | |
| 758 | /* |
| 759 | * SpeculativeInsertionLockRelease |
| 760 | * |
| 761 | * Delete the lock showing that the given transaction is speculatively |
| 762 | * inserting a tuple. |
| 763 | */ |
| 764 | void |
| 765 | SpeculativeInsertionLockRelease(TransactionId xid) |
| 766 | { |
| 767 | LOCKTAG tag; |
| 768 | |
| 769 | SET_LOCKTAG_SPECULATIVE_INSERTION(tag, xid, speculativeInsertionToken); |
| 770 | |
| 771 | LockRelease(&tag, ExclusiveLock, false); |
| 772 | } |
| 773 | |
| 774 | /* |
| 775 | * SpeculativeInsertionWait |
| 776 | * |
| 777 | * Wait for the specified transaction to finish or abort the insertion of a |
| 778 | * tuple. |
| 779 | */ |
| 780 | void |
| 781 | SpeculativeInsertionWait(TransactionId xid, uint32 token) |
| 782 | { |
| 783 | LOCKTAG tag; |
| 784 | |
| 785 | SET_LOCKTAG_SPECULATIVE_INSERTION(tag, xid, token); |
| 786 | |
| 787 | Assert(TransactionIdIsValid(xid)); |
| 788 | Assert(token != 0); |
| 789 | |
| 790 | (void) LockAcquire(&tag, ShareLock, false, false); |
| 791 | LockRelease(&tag, ShareLock, false); |
| 792 | } |
| 793 | |
| 794 | /* |
| 795 | * XactLockTableWaitErrorContextCb |
| 796 | * Error context callback for transaction lock waits. |
| 797 | */ |
| 798 | static void |
| 799 | XactLockTableWaitErrorCb(void *arg) |
| 800 | { |
| 801 | XactLockTableWaitInfo *info = (XactLockTableWaitInfo *) arg; |
| 802 | |
| 803 | /* |
| 804 | * We would like to print schema name too, but that would require a |
| 805 | * syscache lookup. |
| 806 | */ |
| 807 | if (info->oper != XLTW_None && |
| 808 | ItemPointerIsValid(info->ctid) && RelationIsValid(info->rel)) |
| 809 | { |
| 810 | const char *cxt; |
| 811 | |
| 812 | switch (info->oper) |
| 813 | { |
| 814 | case XLTW_Update: |
| 815 | cxt = gettext_noop("while updating tuple (%u,%u) in relation \"%s\"" ); |
| 816 | break; |
| 817 | case XLTW_Delete: |
| 818 | cxt = gettext_noop("while deleting tuple (%u,%u) in relation \"%s\"" ); |
| 819 | break; |
| 820 | case XLTW_Lock: |
| 821 | cxt = gettext_noop("while locking tuple (%u,%u) in relation \"%s\"" ); |
| 822 | break; |
| 823 | case XLTW_LockUpdated: |
| 824 | cxt = gettext_noop("while locking updated version (%u,%u) of tuple in relation \"%s\"" ); |
| 825 | break; |
| 826 | case XLTW_InsertIndex: |
| 827 | cxt = gettext_noop("while inserting index tuple (%u,%u) in relation \"%s\"" ); |
| 828 | break; |
| 829 | case XLTW_InsertIndexUnique: |
| 830 | cxt = gettext_noop("while checking uniqueness of tuple (%u,%u) in relation \"%s\"" ); |
| 831 | break; |
| 832 | case XLTW_FetchUpdated: |
| 833 | cxt = gettext_noop("while rechecking updated tuple (%u,%u) in relation \"%s\"" ); |
| 834 | break; |
| 835 | case XLTW_RecheckExclusionConstr: |
| 836 | cxt = gettext_noop("while checking exclusion constraint on tuple (%u,%u) in relation \"%s\"" ); |
| 837 | break; |
| 838 | |
| 839 | default: |
| 840 | return; |
| 841 | } |
| 842 | |
| 843 | errcontext(cxt, |
| 844 | ItemPointerGetBlockNumber(info->ctid), |
| 845 | ItemPointerGetOffsetNumber(info->ctid), |
| 846 | RelationGetRelationName(info->rel)); |
| 847 | } |
| 848 | } |
| 849 | |
| 850 | /* |
| 851 | * WaitForLockersMultiple |
| 852 | * Wait until no transaction holds locks that conflict with the given |
| 853 | * locktags at the given lockmode. |
| 854 | * |
| 855 | * To do this, obtain the current list of lockers, and wait on their VXIDs |
| 856 | * until they are finished. |
| 857 | * |
| 858 | * Note we don't try to acquire the locks on the given locktags, only the VXIDs |
| 859 | * of its lock holders; if somebody grabs a conflicting lock on the objects |
| 860 | * after we obtained our initial list of lockers, we will not wait for them. |
| 861 | */ |
| 862 | void |
| 863 | WaitForLockersMultiple(List *locktags, LOCKMODE lockmode, bool progress) |
| 864 | { |
| 865 | List *holders = NIL; |
| 866 | ListCell *lc; |
| 867 | int total = 0; |
| 868 | int done = 0; |
| 869 | |
| 870 | /* Done if no locks to wait for */ |
| 871 | if (list_length(locktags) == 0) |
| 872 | return; |
| 873 | |
| 874 | /* Collect the transactions we need to wait on */ |
| 875 | foreach(lc, locktags) |
| 876 | { |
| 877 | LOCKTAG *locktag = lfirst(lc); |
| 878 | int count; |
| 879 | |
| 880 | holders = lappend(holders, |
| 881 | GetLockConflicts(locktag, lockmode, |
| 882 | progress ? &count : NULL)); |
| 883 | if (progress) |
| 884 | total += count; |
| 885 | } |
| 886 | |
| 887 | if (progress) |
| 888 | pgstat_progress_update_param(PROGRESS_WAITFOR_TOTAL, total); |
| 889 | |
| 890 | /* |
| 891 | * Note: GetLockConflicts() never reports our own xid, hence we need not |
| 892 | * check for that. Also, prepared xacts are not reported, which is fine |
| 893 | * since they certainly aren't going to do anything anymore. |
| 894 | */ |
| 895 | |
| 896 | /* Finally wait for each such transaction to complete */ |
| 897 | foreach(lc, holders) |
| 898 | { |
| 899 | VirtualTransactionId *lockholders = lfirst(lc); |
| 900 | |
| 901 | while (VirtualTransactionIdIsValid(*lockholders)) |
| 902 | { |
| 903 | /* |
| 904 | * If requested, publish who we're going to wait for. This is not |
| 905 | * 100% accurate if they're already gone, but we don't care. |
| 906 | */ |
| 907 | if (progress) |
| 908 | { |
| 909 | PGPROC *holder = BackendIdGetProc(lockholders->backendId); |
| 910 | |
| 911 | pgstat_progress_update_param(PROGRESS_WAITFOR_CURRENT_PID, |
| 912 | holder->pid); |
| 913 | } |
| 914 | VirtualXactLock(*lockholders, true); |
| 915 | lockholders++; |
| 916 | |
| 917 | if (progress) |
| 918 | pgstat_progress_update_param(PROGRESS_WAITFOR_DONE, ++done); |
| 919 | } |
| 920 | } |
| 921 | if (progress) |
| 922 | { |
| 923 | const int index[] = { |
| 924 | PROGRESS_WAITFOR_TOTAL, |
| 925 | PROGRESS_WAITFOR_DONE, |
| 926 | PROGRESS_WAITFOR_CURRENT_PID |
| 927 | }; |
| 928 | const int64 values[] = { |
| 929 | 0, 0, 0 |
| 930 | }; |
| 931 | |
| 932 | pgstat_progress_update_multi_param(3, index, values); |
| 933 | } |
| 934 | |
| 935 | list_free_deep(holders); |
| 936 | } |
| 937 | |
| 938 | /* |
| 939 | * WaitForLockers |
| 940 | * |
| 941 | * Same as WaitForLockersMultiple, for a single lock tag. |
| 942 | */ |
| 943 | void |
| 944 | WaitForLockers(LOCKTAG heaplocktag, LOCKMODE lockmode, bool progress) |
| 945 | { |
| 946 | List *l; |
| 947 | |
| 948 | l = list_make1(&heaplocktag); |
| 949 | WaitForLockersMultiple(l, lockmode, progress); |
| 950 | list_free(l); |
| 951 | } |
| 952 | |
| 953 | |
| 954 | /* |
| 955 | * LockDatabaseObject |
| 956 | * |
| 957 | * Obtain a lock on a general object of the current database. Don't use |
| 958 | * this for shared objects (such as tablespaces). It's unwise to apply it |
| 959 | * to relations, also, since a lock taken this way will NOT conflict with |
| 960 | * locks taken via LockRelation and friends. |
| 961 | */ |
| 962 | void |
| 963 | LockDatabaseObject(Oid classid, Oid objid, uint16 objsubid, |
| 964 | LOCKMODE lockmode) |
| 965 | { |
| 966 | LOCKTAG tag; |
| 967 | |
| 968 | SET_LOCKTAG_OBJECT(tag, |
| 969 | MyDatabaseId, |
| 970 | classid, |
| 971 | objid, |
| 972 | objsubid); |
| 973 | |
| 974 | (void) LockAcquire(&tag, lockmode, false, false); |
| 975 | |
| 976 | /* Make sure syscaches are up-to-date with any changes we waited for */ |
| 977 | AcceptInvalidationMessages(); |
| 978 | } |
| 979 | |
| 980 | /* |
| 981 | * UnlockDatabaseObject |
| 982 | */ |
| 983 | void |
| 984 | UnlockDatabaseObject(Oid classid, Oid objid, uint16 objsubid, |
| 985 | LOCKMODE lockmode) |
| 986 | { |
| 987 | LOCKTAG tag; |
| 988 | |
| 989 | SET_LOCKTAG_OBJECT(tag, |
| 990 | MyDatabaseId, |
| 991 | classid, |
| 992 | objid, |
| 993 | objsubid); |
| 994 | |
| 995 | LockRelease(&tag, lockmode, false); |
| 996 | } |
| 997 | |
| 998 | /* |
| 999 | * LockSharedObject |
| 1000 | * |
| 1001 | * Obtain a lock on a shared-across-databases object. |
| 1002 | */ |
| 1003 | void |
| 1004 | LockSharedObject(Oid classid, Oid objid, uint16 objsubid, |
| 1005 | LOCKMODE lockmode) |
| 1006 | { |
| 1007 | LOCKTAG tag; |
| 1008 | |
| 1009 | SET_LOCKTAG_OBJECT(tag, |
| 1010 | InvalidOid, |
| 1011 | classid, |
| 1012 | objid, |
| 1013 | objsubid); |
| 1014 | |
| 1015 | (void) LockAcquire(&tag, lockmode, false, false); |
| 1016 | |
| 1017 | /* Make sure syscaches are up-to-date with any changes we waited for */ |
| 1018 | AcceptInvalidationMessages(); |
| 1019 | } |
| 1020 | |
| 1021 | /* |
| 1022 | * UnlockSharedObject |
| 1023 | */ |
| 1024 | void |
| 1025 | UnlockSharedObject(Oid classid, Oid objid, uint16 objsubid, |
| 1026 | LOCKMODE lockmode) |
| 1027 | { |
| 1028 | LOCKTAG tag; |
| 1029 | |
| 1030 | SET_LOCKTAG_OBJECT(tag, |
| 1031 | InvalidOid, |
| 1032 | classid, |
| 1033 | objid, |
| 1034 | objsubid); |
| 1035 | |
| 1036 | LockRelease(&tag, lockmode, false); |
| 1037 | } |
| 1038 | |
| 1039 | /* |
| 1040 | * LockSharedObjectForSession |
| 1041 | * |
| 1042 | * Obtain a session-level lock on a shared-across-databases object. |
| 1043 | * See LockRelationIdForSession for notes about session-level locks. |
| 1044 | */ |
| 1045 | void |
| 1046 | LockSharedObjectForSession(Oid classid, Oid objid, uint16 objsubid, |
| 1047 | LOCKMODE lockmode) |
| 1048 | { |
| 1049 | LOCKTAG tag; |
| 1050 | |
| 1051 | SET_LOCKTAG_OBJECT(tag, |
| 1052 | InvalidOid, |
| 1053 | classid, |
| 1054 | objid, |
| 1055 | objsubid); |
| 1056 | |
| 1057 | (void) LockAcquire(&tag, lockmode, true, false); |
| 1058 | } |
| 1059 | |
| 1060 | /* |
| 1061 | * UnlockSharedObjectForSession |
| 1062 | */ |
| 1063 | void |
| 1064 | UnlockSharedObjectForSession(Oid classid, Oid objid, uint16 objsubid, |
| 1065 | LOCKMODE lockmode) |
| 1066 | { |
| 1067 | LOCKTAG tag; |
| 1068 | |
| 1069 | SET_LOCKTAG_OBJECT(tag, |
| 1070 | InvalidOid, |
| 1071 | classid, |
| 1072 | objid, |
| 1073 | objsubid); |
| 1074 | |
| 1075 | LockRelease(&tag, lockmode, true); |
| 1076 | } |
| 1077 | |
| 1078 | |
| 1079 | /* |
| 1080 | * Append a description of a lockable object to buf. |
| 1081 | * |
| 1082 | * Ideally we would print names for the numeric values, but that requires |
| 1083 | * getting locks on system tables, which might cause problems since this is |
| 1084 | * typically used to report deadlock situations. |
| 1085 | */ |
| 1086 | void |
| 1087 | DescribeLockTag(StringInfo buf, const LOCKTAG *tag) |
| 1088 | { |
| 1089 | switch ((LockTagType) tag->locktag_type) |
| 1090 | { |
| 1091 | case LOCKTAG_RELATION: |
| 1092 | appendStringInfo(buf, |
| 1093 | _("relation %u of database %u" ), |
| 1094 | tag->locktag_field2, |
| 1095 | tag->locktag_field1); |
| 1096 | break; |
| 1097 | case LOCKTAG_RELATION_EXTEND: |
| 1098 | appendStringInfo(buf, |
| 1099 | _("extension of relation %u of database %u" ), |
| 1100 | tag->locktag_field2, |
| 1101 | tag->locktag_field1); |
| 1102 | break; |
| 1103 | case LOCKTAG_PAGE: |
| 1104 | appendStringInfo(buf, |
| 1105 | _("page %u of relation %u of database %u" ), |
| 1106 | tag->locktag_field3, |
| 1107 | tag->locktag_field2, |
| 1108 | tag->locktag_field1); |
| 1109 | break; |
| 1110 | case LOCKTAG_TUPLE: |
| 1111 | appendStringInfo(buf, |
| 1112 | _("tuple (%u,%u) of relation %u of database %u" ), |
| 1113 | tag->locktag_field3, |
| 1114 | tag->locktag_field4, |
| 1115 | tag->locktag_field2, |
| 1116 | tag->locktag_field1); |
| 1117 | break; |
| 1118 | case LOCKTAG_TRANSACTION: |
| 1119 | appendStringInfo(buf, |
| 1120 | _("transaction %u" ), |
| 1121 | tag->locktag_field1); |
| 1122 | break; |
| 1123 | case LOCKTAG_VIRTUALTRANSACTION: |
| 1124 | appendStringInfo(buf, |
| 1125 | _("virtual transaction %d/%u" ), |
| 1126 | tag->locktag_field1, |
| 1127 | tag->locktag_field2); |
| 1128 | break; |
| 1129 | case LOCKTAG_SPECULATIVE_TOKEN: |
| 1130 | appendStringInfo(buf, |
| 1131 | _("speculative token %u of transaction %u" ), |
| 1132 | tag->locktag_field2, |
| 1133 | tag->locktag_field1); |
| 1134 | break; |
| 1135 | case LOCKTAG_OBJECT: |
| 1136 | appendStringInfo(buf, |
| 1137 | _("object %u of class %u of database %u" ), |
| 1138 | tag->locktag_field3, |
| 1139 | tag->locktag_field2, |
| 1140 | tag->locktag_field1); |
| 1141 | break; |
| 1142 | case LOCKTAG_USERLOCK: |
| 1143 | /* reserved for old contrib code, now on pgfoundry */ |
| 1144 | appendStringInfo(buf, |
| 1145 | _("user lock [%u,%u,%u]" ), |
| 1146 | tag->locktag_field1, |
| 1147 | tag->locktag_field2, |
| 1148 | tag->locktag_field3); |
| 1149 | break; |
| 1150 | case LOCKTAG_ADVISORY: |
| 1151 | appendStringInfo(buf, |
| 1152 | _("advisory lock [%u,%u,%u,%u]" ), |
| 1153 | tag->locktag_field1, |
| 1154 | tag->locktag_field2, |
| 1155 | tag->locktag_field3, |
| 1156 | tag->locktag_field4); |
| 1157 | break; |
| 1158 | default: |
| 1159 | appendStringInfo(buf, |
| 1160 | _("unrecognized locktag type %d" ), |
| 1161 | (int) tag->locktag_type); |
| 1162 | break; |
| 1163 | } |
| 1164 | } |
| 1165 | |
| 1166 | /* |
| 1167 | * GetLockNameFromTagType |
| 1168 | * |
| 1169 | * Given locktag type, return the corresponding lock name. |
| 1170 | */ |
| 1171 | const char * |
| 1172 | GetLockNameFromTagType(uint16 locktag_type) |
| 1173 | { |
| 1174 | if (locktag_type > LOCKTAG_LAST_TYPE) |
| 1175 | return "???" ; |
| 1176 | return LockTagTypeNames[locktag_type]; |
| 1177 | } |
| 1178 | |